Lecture Notes in Networks and Systems 225

Leonard Barolli
Isaac Woungang
Tomoya Enokido Editors

Advanced
Information
Networking and
Applications
Proceedings of the 35th International
Conference on Advanced Information
Networking and Applications
(AINA-2021), Volume 1
Lecture Notes in Networks and Systems

Volume 225

Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland

Advisory Editors
Fernando Gomide, Department of Computer Engineering and Automation—DCA,
School of Electrical and Computer Engineering—FEEC, University of Campinas—
UNICAMP, São Paulo, Brazil
Okyay Kaynak, Department of Electrical and Electronic Engineering,
Bogazici University, Istanbul, Turkey
Derong Liu, Department of Electrical and Computer Engineering, University
of Illinois at Chicago, Chicago, USA; Institute of Automation, Chinese Academy
of Sciences, Beijing, China
Witold Pedrycz, Department of Electrical and Computer Engineering,
University of Alberta, Alberta, Canada; Systems Research Institute,
Polish Academy of Sciences, Warsaw, Poland
Marios M. Polycarpou, Department of Electrical and Computer Engineering,
KIOS Research Center for Intelligent Systems and Networks, University of Cyprus,
Nicosia, Cyprus
Imre J. Rudas, Óbuda University, Budapest, Hungary
Jun Wang, Department of Computer Science, City University of Hong Kong,
Kowloon, Hong Kong
The series “Lecture Notes in Networks and Systems” publishes the latest
developments in Networks and Systems—quickly, informally and with high quality.
Original research reported in proceedings and post-proceedings represents the core
of LNNS.
Volumes published in LNNS embrace all aspects and subfields of, as well as new
challenges in, Networks and Systems.
The series contains proceedings and edited volumes in systems and networks,
spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor
Networks, Control Systems, Energy Systems, Automotive Systems, Biological
Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems,
Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems,
Robotics, Social Systems, Economic Systems and other. Of particular value to both
the contributors and the readership are the short publication timeframe and the
world-wide distribution and exposure which enable both a wide and rapid
dissemination of research output.
The series covers the theory, applications, and perspectives on the state of the art
and future developments relevant to systems and networks, decision making, control,
complex processes and related areas, as embedded in the fields of interdisciplinary
and applied sciences, engineering, computer science, physics, economics, social, and
life sciences, as well as the paradigms and methodologies behind them.
Indexed by SCOPUS, INSPEC, WTI Frankfurt eG, zbMATH, SCImago.
All books published in the series are submitted for consideration in Web of
Science.

More information about this series at http://www.springer.com/series/15179

Leonard Barolli Isaac Woungang
• •

Tomoya Enokido
Editors

Advanced Information
Networking and Applications
Proceedings of the 35th International
Conference on Advanced Information
Networking and Applications (AINA-2021),
Volume 1

123
Editors
Leonard Barolli Isaac Woungang
Department of Information Department of Computer Science
and Communication Engineering Ryerson University
Fukuoka Institute of Technology Toronto, ON, Canada
Fukuoka, Japan

Tomoya Enokido
Faculty of Business Administration
Rissho University
Tokyo, Japan

ISSN 2367-3370 ISSN 2367-3389 (electronic)

Lecture Notes in Networks and Systems
ISBN 978-3-030-75099-2 ISBN 978-3-030-75100-5 (eBook)
https://doi.org/10.1007/978-3-030-75100-5
© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Welcome Message from AINA-2021 Organizers

Welcome to the 35th International Conference on Advanced Information

Networking and Applications (AINA-2021). On behalf of AINA-2021 Organizing
Committee, we would like to express to all participants our cordial welcome and
high respect.
AINA is an international forum, where scientists and researchers from academia
and industry working in various scientific and technical areas of networking and
distributed computing systems can demonstrate new ideas and solutions in dis-
tributed computing systems. AINA was born in Asia, but it is now an international
conference with high quality thanks to the great help and cooperation of many
international friendly volunteers. AINA is a very open society and is always wel-
coming international volunteers from any country and any area in the world.
AINA international conference is a forum for sharing ideas and research work in
the emerging areas of information networking and their applications. The area of
advanced networking has grown very rapidly, and the applications have experi-
enced an explosive growth especially in the areas of pervasive and mobile appli-
cations, wireless sensor networks, wireless ad hoc networks, vehicular networks,
multimedia computing and social networking, semantic collaborative systems, as
well as Grid, P2P, IoT, Big Data and Cloud Computing. This advanced networking
revolution is transforming the way people live, work and interact with each other
and is impacting the way business, education, entertainment and health care are
operating. The papers included in the proceedings cover theory, design and
application of computer networks, distributed computing and information systems.
Each year AINA receives a lot of paper submissions from all around the world. It
has maintained high-quality accepted papers and is aspiring to be one of the main
international conferences on the information networking in the world.
We are very proud and honored to have two distinguished keynote talks by Dr.
Flora Amato, University of Naples “Federico II”, Italy and Prof. Shahrokh Valaee,
University of Toronto, Canada, who will present their recent work and will give
new insights and ideas to the conference participants.

v
vi Welcome Message from AINA-2021 Organizers

An international conference of this size requires the support and help of many
people. A lot of people have helped and worked hard to produce a successful
AINA-2021 technical program and conference proceedings. First, we would like to
thank all authors for submitting their papers, the session chairs and distinguished
keynote speakers. We are indebted to program track co-chairs, program committee
members and reviewers, who carried out the most difficult work of carefully
evaluating the submitted papers.
We would like to thank AINA-2021 general co-chairs, PC co-chairs, workshop
co-chairs for their great efforts to make AINA-2021 a very successful event. We
have special thanks to the finance chair and web administrator co-chairs.
We do hope that you will enjoy the conference proceedings and readings.

Leonard Barolli
Makoto Takizawa
AINA Steering Committee Co-chairs
Isaac Woungang
Markus Aleksy
Farookh Hussain
AINA-2021 General Co-chairs
Glaucio Carvalho
Tomoya Enokido
Flora Amato
AINA-2021 Program Committee Co-chairs
AINA-2021 Organizing Committee

General Co-chairs

Isaac Woungang Ryerson University, Canada

Markus Aleksy ABB Corporate Research Center, Germany
Farookh Hussain University of Technology, Sydney, Australia

Program Committee Co-chairs

Glaucio Carvalho Sheridan College, Canada
Tomoya Enokido Rissho University, Japan
Flora Amato University of Naples “Federico II”, Italy

Workshops Co-chairs
Kin Fun Li University of Victoria, Canada
Omid Ameri Sianaki Victoria University, Australia
Yi-Jen Su Shu-Te University, Taiwan

International Journals Special Issues Co-chairs

Fatos Xhafa Technical University of Catalonia, Spain
David Taniar Monash University, Australia

Award Co-chairs
Marek Ogiela AGH University of Science and Technology,
Poland
Arjan Durresi Indiana University Purdue University
in Indianapolis (IUPUI), USA
Fang-Yie Leu Tunghai University, Taiwan

vii
viii AINA-2021 Organizing Committee

Publicity Co-chairs
Lidia Ogiela Pedagogical University of Cracow, Poland
Minoru Uehara Toyo University, Japan
Hsing-Chung Chen Asia University, Taiwan

International Liaison Co-chairs

Akio Koyama Yamagata University, Japan
Nadeem Javaid COMSATS University Islamabad, Pakistan
Wenny Rahayu La Trobe University, Australia

Local Arrangement Co-chairs

Mehrdad Tirandazian Ryerson University, Canada
Glaucio Carvalho Sheridan College, Canada

Finance Chair
Makoto Ikeda Fukuoka Institute of Technology, Japan

Web Co-chairs
Phudit Ampririt Fukuoka Institute of Technology, Japan
Kevin Bylykbashi Fukuoka Institute of Technology, Japan
Ermioni Qafzezi Fukuoka Institute of Technology, Japan

Steering Committee Chairs

Leonard Barolli Fukuoka Institute of Technology, Japan
Makoto Takizawa Hosei University, Japan

Tracks and Program Committee Members

1. Network Protocols and Applications
Track Co-chairs
Makoto Ikeda Fukuoka Institute of Technology, Japan
Sanjay Kumar Dhurandher Netaji Subhas University of Technology,
New Delhi, India
Bhed Bahadur Bista Iwate Prefectural University, Japan
AINA-2021 Organizing Committee ix

TPC Members
Elis Kulla Okayama University of Science, Japan
Keita Matsuo Fukuoka Institute of Technology, Japan
Shinji Sakamoto Seikei University, Japan
Akio Koyama Yamagata University, Japan
Evjola Spaho Polytechnic University of Tirana, Albania
Jiahong Wang Iwate Prefectural University, Japan
Shigetomo Kimura University of Tsukuba, Japan
Chotipat Pornavalai King Mongkut’s Institute of Technology
Ladkrabang, Thailand
Danda B. Rawat Howard University, USA
Akio Koyama Yamagata University, Japan
Amita Malik Deenbandhu Chhotu Ram University of Science
and Technology, India
R. K. Pateriya Maulana Azad National Institute of Technology,
India
Vinesh Kumar University of Delhi, India
Petros Nicopolitidis Aristotle University of Thessaloniki, Greece
Satya Jyoti Borah North Eastern Regional Institute of Science and
Technology, India

2. Next Generation Wireless Networks

Track Co-chairs
Christos J. Bouras University of Patras, Greece
Tales Heimfarth Universidade Federal de Lavras, Brazil
Leonardo Mostarda University of Camerino, Italy

TPC Members
Fadi Al-Turjman Near East University, Nicosia, Cyprus
Alfredo Navarra University of Perugia, Italy
Purav Shah Middlesex University London, UK
Enver Ever Middle East Technical University,
Northern Cyprus Campus, Cyprus
Rosario Culmone University of Camerino, Camerino, Italy
Antonio Alfredo F. Loureiro Federal University of Minas Gerais, Brazil
Holger Karl University of Paderborn, Germany
Daniel Ludovico Guidoni Federal University of São João Del-Rei, Brazil
João Paulo Carvalho Lustosa Hamm-Lippstadt University of Applied Sciences,
da Costa Germany
Jorge Sá Silva University of Coimbra, Portugal
Apostolos Gkamas University Ecclesiastical Academy of Vella,
Ioannina, Greece
x AINA-2021 Organizing Committee

Zoubir Mammeri University Paul Sabatier, France

Eirini Eleni Tsiropoulou University of New Mexico, USA
Raouf Hamzaoui De Montfort University, UK
Miroslav Voznak University of Ostrava, Czech Republic

3. Multimedia Systems and Applications

Track Co-chairs
Markus Aleksy ABB Corporate Research Center, Germany
Francesco Orciuoli University of Salerno, Italy
Tomoyuki Ishida Fukuoka Institute of Technology, Japan

TPC Members
Tetsuro Ogi Keio University, Japan
Yasuo Ebara Osaka Electro-Communication University, Japan
Hideo Miyachi Tokyo City University, Japan
Kaoru Sugita Fukuoka Institute of Technology, Japan
Akio Doi Iwate Prefectural University, Japan
Hadil Abukwaik ABB Corporate Research Center, Germany
Monique Duengen Robert Bosch GmbH, Germany
Thomas Preuss Brandenburg University of Applied Sciences,
Germany
Peter M. Rost NOKIA Bell Labs, Germany
Lukasz Wisniewski inIT, Germany
Hadil Abukwaik ABB Corporate Research Center, Germany
Monique Duengen Robert Bosch GmbH, Germany
Peter M. Rost NOKIA Bell Labs, Germany
Lukasz Wisniewski inIT, Germany
Angelo Gaeta University of Salerno, Italy
Graziano Fuccio University of Salerno, Italy
Giuseppe Fenza University of Salerno, Italy
Maria Cristina University of Salerno, Italy
Alberto Volpe University of Salerno, Italy

4. Pervasive and Ubiquitous Computing

Track Co-chairs
Chih-Lin Hu National Central University, Taiwan
Vamsi Paruchuri University of Central Arkansas, USA
Winston Seah Victoria University of Wellington, New Zealand
AINA-2021 Organizing Committee xi

TPC Members
Hong Va Leong Hong Kong Polytechnic University, Hong Kong
Ling-Jyh Chen Academia Sinica, Taiwan
Jiun-Yu Tu Southern Taiwan University of Science
and Technology, Taiwan
Jiun-Long Huang National Chiao Tung University, Taiwan
Thitinan Tantidham Mahidol University, Thailand
Tanapat Anusas-amornkul King Mongkut’s University of Technology North
Bangkok, Thailand
Xin-Mao Huang Aletheia University, Taiwan
Hui Lin Tamkang University, Taiwan
Eugen Dedu Universite de Franche-Comte, France
Peng Huang Sichuan Agricultural University, China
Wuyungerile Li Inner Mongolia University, China
Adrian Pekar Budapest University of Technology
and Economics, Hungary
Jyoti Sahni Victoria University of Technology, New Zealand
Normalia Samian Universiti Putra Malaysia, Malaysia
Sriram Chellappan University of South Florida, USA
Yu Sun University of Central Arkansas, USA
Qiang Duan Penn State University, USA
Han-Chieh Wei Dallas Baptist University, USA

5. Web-Based and E-Learning Systems

Track Co-chairs
Santi Caballe Open University of Catalonia, Spain
Kin Fun Li University of Victoria, Canada
Nobuo Funabiki Okayama University, Japan

TPC Members
Jordi Conesa Open University of Catalonia, Spain
Joan Casas Open University of Catalonia, Spain
David Gañán Open University of Catalonia, Spain
Nicola Capuano University of Basilicata, Italy
Antonio Sarasa Complutense University of Madrid, Spain
Chih-Peng Fan National Chung Hsing University, Taiwan
Nobuya Ishihara Okayama University, Japan
Sho Yamamoto Kindai University, Japan
Khin Khin Zaw Yangon Technical University, Myanmar
Kaoru Fujioka Fukuoka Women’s University, Japan
Kosuke Takano Kanagawa Institute of Technology, Japan
Shengrui Wang University of Sherbrooke, Canada
xii AINA-2021 Organizing Committee

Darshika Perera University of Colorado at Colorado Spring, USA

Carson Leung University of Manitoba, Canada

6. Distributed and Parallel Computing

Track Co-chairs
Naohiro Hayashibara Kyoto Sangyo University, Japan
Minoru Uehara Toyo University, Japan
Tomoya Enokido Rissho University, Japan

TPC Members
Eric Pardede La Trobe University, Australia
Lidia Ogiela Pedagogical University of Cracow, Poland
Evjola Spaho Polytechnic University of Tirana, Albania
Akio Koyama Yamagata University, Japan
Omar Hussain University of New South Wales, Australia
Hideharu Amano Keio University, Japan
Ryuji Shioya Toyo University, Japan
Ji Zhang The University of Southern Queensland
Lucian Prodan Universitatea Politehnica Timisoara, Romania
Ragib Hasan The University of Alabama at Birmingham, USA
Young-Hoon Park Sookmyung Women’s University, Korea

7. Data Mining, Big Data Analytics and Social Networks

Track Co-chairs
Eric Pardede La Trobe University, Australia
Alex Thomo University of Victoria, Canada
Flora Amato University of Naples “Frederico II”, Italy

TPC Members
Ji Zhang University of Southern Queensland, Australia
Salimur Choudhury Lakehead University, Canada
Xiaofeng Ding Huazhong University of Science and
Technology, China
Ronaldo dos Santos Mello Universidade Federal de Santa Catarina, Brasil
Irena Holubova Charles University, Czech Republic
Lucian Prodan Universitatea Politehnica Timisoara, Romania
Alex Tomy La Trobe University, Australia
Dhomas Hatta Fudholi Universitas Islam Indonesia, Indonesia
AINA-2021 Organizing Committee xiii

Saqib Ali Sultan Qaboos University, Oman

Ahmad Alqarni Al Baha University, Saudi Arabia
Alessandra Amato University of Naples “Frederico II”, Italy
Luigi Coppolino Parthenope University, Italy
Giovanni Cozzolino University of Naples “Frederico II”, Italy
Giovanni Mazzeo Parthenope University, Italy
Francesco Mercaldo Italian National Research Council, Italy
Francesco Moscato University of Salerno, Italy
Vincenzo Moscato University of Naples “Frederico II”, Italy
Francesco Piccialli University of Naples “Frederico II”, Italy

8. Internet of Things and Cyber-Physical Systems

Track Co-chairs
Euripides G. M. Petrakis Technical University of Crete (TUC), Greece
Tomoki Yoshihisa Osaka University, Japan
Mario Dantas Federal University of Juiz de Fora (UFJF), Brazil

TPC Members
Akihiro Fujimoto Wakayama University, Japan
Akimitsu Kanzaki Shimane University, Japan
Kawakami Tomoya University of Fukui, Japan
Lei Shu University of Lincoln, UK
Naoyuki Morimoto Mie University, Japan
Yusuke Gotoh Okayama University, Japan
Vasilis Samolada Technical University of Crete (TUC), Greece
Konstantinos Tsakos Technical University of Crete (TUC), Greece
Aimilios Tzavaras Technical University of Crete (TUC), Greece
Spanakis Manolis Foundation for Research and Technology Hellas
(FORTH), Greece
Katerina Doka National Technical University of Athens
(NTUA), Greece
Giorgos Vasiliadis Foundation for Research and Technology Hellas
(FORTH), Greece
Stefan Covaci Technische Universität Berlin, Berlin (TUB),
Germany
Stelios Sotiriadis University of London, UK
Stefano Chessa University of Pisa, Italy
Jean-Francois Méhaut Université Grenoble Alpes, France
Michael Bauer University of Western Ontario, Canada
xiv AINA-2021 Organizing Committee

9. Intelligent Computing and Machine Learning

Track Co-chairs
Takahiro Uchiya Nagoya Institute of Technology, Japan
Omar Hussain UNSW, Australia
Nadeem Javaid COMSATS University Islamabad, Pakistan

TPC Members
Morteza Saberi University of Technology, Sydney, Australia
Abderrahmane Leshob University of Quebec in Montreal, Canada
Adil Hammadi Curtin University, Australia
Naeem Janjua Edith Cowan University, Australia
Sazia Parvin Melbourne Polytechnic, Australia
Kazuto Sasai Ibaraki University, Japan
Shigeru Fujita Chiba Institute of Technology, Japan
Yuki Kaeri Mejiro University, Japan
Zahoor Ali Khan HCT, UAE
Muhammad Imran King Saud University, Saudi Arabia
Ashfaq Ahmad The University of Newcastle, Australia
Syed Hassan Ahmad JMA Wireless, USA
Safdar Hussain Bouk Daegu Gyeongbuk Institute of Science and
Technology, Korea
Jolanta Mizera-Pietraszko Military University of Land Forces, Poland

10. Cloud and Services Computing

Track Co-chairs
Asm Kayes La Trobe University, Australia
Salvatore Venticinque University of Campania “Luigi Vanvitelli”, Italy
Baojiang Cui Beijing University of Posts and
Telecommunications, China

TPC Members
Shahriar Badsha University of Nevada, USA
Abdur Rahman Bin Shahid Concord University, USA
Iqbal H. Sarker Chittagong University of Engineering
and Technology, Bangladesh
Jabed Morshed Chowdhury La Trobe University, Australia
Alex Ng La Trobe University, Australia
AINA-2021 Organizing Committee xv

Indika Kumara Jheronimus Academy of Data Science,

Netherlands
Tarique Anwar Macquarie University and CSIRO’s Data61,
Australia
Giancarlo Fortino University of Calabria, Italy
Massimiliano Rak University of Campania “Luigi Vanvitelli”, Italy
Jason J. Jung Chung-Ang University, Korea
Dimosthenis Kyriazis University of Piraeus, Greece
Geir Horn University of Oslo, Norway
Gang Wang Nankai University, China
Shaozhang Niu Beijing University of Posts and
Telecommunications, China
Jianxin Wang Beijing Forestry University, China
Jie Cheng Shandong University, China
Shaoyin Cheng University of Science and Technology of China,
China

11. Security, Privacy and Trust Computing

Track Co-chairs
Hiroaki Kikuchi Meiji University, Japan
Xu An Wang Engineering University of PAP, China
Lidia Ogiela Pedagogical University of Cracow, Poland

TPC Members
Takamichi Saito Meiji University, Japan
Kouichi Sakurai Kyushu University, Japan
Kazumasa Omote University of Tsukuba, Japan
Shou-Hsuan Stephen Huang University of Houston, USA
Masakatsu Nishigaki Shizuoka University, Japan
Mingwu Zhang Hubei University of Technology, China
Caiquan Xiong Hubei University of Technology, China
Wei Ren China University of Geosciences, China
Peng Li Nanjing University of Posts and
Telecommunications, China
Guangquan Xu Tianjing University, China
Urszula Ogiela Pedagogical University of Cracow, Poland
Hoon Ko Chosun University, Korea
Goreti Marreiros Institute of Engineering of Polytechnic of Porto,
Portugal
Chang Choi Gachon University, Korea
Libor Měsíček J.E. Purkyně University, Czech Republic
xvi AINA-2021 Organizing Committee

12. Software-Defined Networking and Network Virtualization

Track Co-chairs
Flavio de Oliveira Silva Federal University of Uberlândia, Brazil
Ashutosh Bhatia Birla Institute of Technology and Science, Pilani,
India
Alaa Allakany Kyushu University, Japan

TPC Members
Yaokai Feng Kyushu University, Japan
Chengming Li Chinese Academy of Science (CAS), China
Othman Othman An-Najah National University (ANNU), Palestine
Nor-masri Bin-sahri University Technology of MARA, Malaysia
Sanouphab Phomkeona National University of Laos, Laos
Haribabu K. BITS Pilani, India
Shekhavat, Virendra BITS Pilani, India
Makoto Ikeda Fukuoka Institute of Technology, Japan
Farookh Hussain University of Technology Sydney, Australia
Keita Matsuo Fukuoka Institute of Technology, Japan
AINA-2021 Organizing Committee xvii

AINA-2021 Reviewers
Admir Barolli Feroz Zahid
Adrian Pekar Flavio Silva
Ahmed Elmokashfi Flora Amato
Akihiro Fujihara Francesco Orciuoli
Akihiro Fujimoto Francesco Piccialli
Akimitsu Kanzaki Gang Wang
Akio Koyama Geir Horn
Alaa Allakany Giancarlo Fortino
Alberto Volpe Giorgos Vasiliadis
Alex Ng Giuseppe Fenza
Alex Thomo Guangquan Xu
Alfredo Navarra Hadil Abukwaik
Aneta Poniszewska-Maranda Hideharu Amano
Angelo Gaeta Hiroaki Kikuchi
Anne Kayem Hiroshi Maeda
Antonio Loureiro Hiroyuki Fujioka
Apostolos Gkamas Holger Karl
Arjan Durresi Hong Va Leong
Ashfaq Ahmad Huey-Ing Liu
Ashutosh Bhatia Hyunhee Park
Asm Kayes Indika Kumara
Baojiang Cui Isaac Woungang
Beniamino Di Martino Jabed Chowdhury
Bhed Bista Jana Nowaková
Carson Leung Jason Jung
Christos Bouras Jawwad Shamsi
Danda Rawat Jesús Escudero-Sahuquillo
Darshika Perera Ji Zhang
David Taniar Jiun-Long Huang
Dimitris Apostolou Jolanta Mizera-Pietraszko
Dimosthenis Kyriazis Jordi Conesa
Eirini Eleni Tsiropoulou Jörg Domaschka
Emmanouil Spanakis Jorge Sá Silva
Enver Ever Juggapong Natwichai
Eric Pardede Jyoti Sahni
Ernst Gran K. Haribabu
Eugen Dedu Katerina Doka
Euripides Petrakis Kazumasa Omote
Fadi Al-Turjman Kazuto Sasai
Farhad Daneshgar Keita Matsuo
Farookh Hussain Kin Fun Li
Fatos Xhafa Kiyotaka Fujisaki
Feilong Tang Konstantinos Tsakos
xviii AINA-2021 Organizing Committee

Kyriakos Kritikos Salimur Choudhury

Lei Shu Salvatore Venticinque
Leonard Barolli Sanjay Dhurandher
Leonardo Mostarda Santi Caballé
Libor Mesicek Shahriar Badsha
Lidia Ogiela Shigeru Fujita
Lin Hui Shigetomo Kimura
Ling-Jyh Chen Sriram Chellappan
Lucian Prodan Stefan Covaci
Makoto Ikeda Stefano Chessa
Makoto Takizawa Stelios Sotiriadis
Marek Ogiela Stephane Maag
Mario Dantas Takahiro Uchiya
Markus Aleksy Takamichi Saito
Masakatsu Nishigaki Tarique Anwar
Masaki Kohana Thitinan Tantidham
Massimiliano Rak Thomas Dreibholz
Massimo Ficco Thomas Preuss
Michael Bauer Tomoki Yoshihisa
Minoru Uehara Tomoya Enokido
Morteza Saberi Tomoyuki Ishida
Nadeem Javaid Vamsi Paruchuri
Naeem Janjua Vasilis Samoladas
Naohiro Hayashibara Vinesh Kumar
Nicola Capuano Virendra Shekhawat
Nobuo Funabiki Wang Xu An
Omar Hussain Wei Ren
Omid Ameri Sianaki Wenny Rahayu
Paresh Saxena Wuyungerile Li
Purav Shah Xin-Mao Huang
Qiang Duan Xing Zhou
Quentin Jacquemart Yaokai Feng
Rajesh Pateriya Yiannis Verginadis
Ricardo Rodríguez Jorge Yoshihiro Okada
Ronaldo Mello Yusuke Gotoh
Rosario Culmone Zahoor Khan
Ryuji Shioya Zia Ullah
Safdar Hussain Bouk Zoubir Mammeri
AINA-2021 Keynote Talks
The Role of Artificial Intelligence
in the Industry 4.0

Flora Amato

University of Naples “Federico II”, Naples, Italy

Abstract. Artificial intelligence (AI) deals with the ability of machines to simulate
human mental competences. The AI can effectively boost the manufacturing sector,
changing the strategies used to implement and tune productive processes by
exploiting information acquired at real time. Industry 4.0 integrates critical tech-
nologies of control and computing. In this talk is discussed the integration of
knowledge representation, ontology modeling with deep learning technology with
the aim of optimizing orchestration and dynamic management of resources. We
review AI techniques used in Industry 4.0 and show an adaptable and extensible
contextual model for creating context-aware computing infrastructures in Internet of
Things (IoT). We also address deep learning techniques for optimizing manufac-
turing resources, assets management and dynamic scheduling. The application of
this model ranges from small embedded devices to high-end service platforms. The
presented deep learning techniques are designed to solve numerous critical chal-
lenges in industrial and IoT intelligence, such as application adaptation, interop-
erability, automatic code verification and generation of a device-specific intelligent
interface.

xxi
Localization in 6G

Shahrokh Valaee

University of Toronto, Toronto, Canada

Abstract. The next generation of wireless systems will employ networking

equipment mounted on mobile platforms, unmanned air vehicles (UAVs) and
low-orbit satellites. As a result, the topology of the sixth-generation (6G) wireless
technology will extend to the three-dimensional (3D) vertical networking. With its
extended service, 6G will also give rise to new challenges which include the
introduction of intelligent reflective surfaces (IRS), the mmWave spectrum, the
employment of massive MIMO systems and the agility of networks. Along with the
advancement in networking technology, the user devices are also evolving rapidly
with the emergence of highly capable cellphones, smart IoT equipments and
wearable devices. One of the key elements of 6G technology is the need for
accurate positioning information. The accuracy of today’s positioning systems is
not acceptable for many applications of future, especially in smart environments. In
this talk, we will discuss how positioning can be a key enabler of 6G and what
challenges the next generation of localization technology will face when integrated
within the new wireless networks.

xxiii
Contents

Optimization of Task Allocations in Cloud to Fog Environment

with Application to Intelligent Transportation Systems . . . . . . . . . . . . . 1
Fatos Xhafa, Alhassan Aly, and Angel A. Juan
User Allocation in 5G Networks Using Machine Learning Methods
for Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Christos Bouras and Rafail Kalogeropoulos
An Efficient Cross-Layer Design for Multi-hop Broadcast
of Emergency Warning Messages in Vehicular Networks . . . . . . . . . . . 25
Abir Rebei, Fouzi Boukhalfa, Haifa Touati, Mohamed Hadded,
and Paul Muhlethaler
Time Series Forecasting for the Number
of Firefighters Interventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Roxane Elias Mallouhy, Christophe Guyeux, Chady Abou Jaoude,
and Abdallah Makhoul
Stock Prediction Based on Adaptive Gradient Descent
Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Bo Li and Li-feng Li
A QoS-Aware Fuzzy-Based System for Assessment of Edge
Computing Resources in SDN-VANETs . . . . . . . . . . . . . . . . . . . . . . . . . 63
Ermioni Qafzezi, Kevin Bylykbashi, Phudit Ampririt, Makoto Ikeda,
Keita Matsuo, and Leonard Barolli
A Comparison Study of Constriction and Random Inertia Weight
Router Replacement Methods for WMNs by WMN-PSOHC Hybrid
Intelligent Simulation System Considering Normal Distribution
of Mesh Clients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Shinji Sakamoto, Leonard Barolli, and Shusuke Okamoto

xxv
xxvi Contents

Performance Evaluation of RIWM and LDVM Router Replacement

Methods for WMNs by WMN-PSOSA-DGA Considering
Chi-Square Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Admir Barolli, Shinji Sakamoto, Phudit Ampririt, Leonard Barolli,
and Makoto Takizawa
An Adaptive Anti-packet Recovery Method for Vehicular DTN
Considering Message Possession Rate . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Shota Uchimura, Masaya Azuma, Yoshiki Tada, Makoto Ikeda,
and Leonard Barolli
From Incomplete to Complete Networks in Asynchronous
Byzantine Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Michel Raynal and Jiannong Cao
Towards the Performance Evaluation of a Trust Based Routing
Protocol for VANET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Amira Kchaou, Ryma Abassi, and Sihem Guemara El Fatmi
Design and Implementation of the TBOI (Time-Based Operation
Interruption) Protocol to Prevent Late Information Flow in the IoT . . . 125
Shigenari Nakamura, Tomoya Enokido, and Makoto Takizawa
How Far Should We Look Back to Achieve Effective Real-Time
Time-Series Anomaly Detection? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Ming-Chang Lee, Jia-Chun Lin, and Ernst Gunnar Gran
A Novel Hybrid Recommendation Approach Based on Correlation
and Co-occurrence Between Activities Within Social
Learning Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Sonia Souabi, Asmaâ Retbi, Mohammed Khalidi Idrissi,
and Samir Bennani
Sentiment-Driven Breakdown Detection Model Using Contextual
Embedding ElMo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Ebtesam Hussain Almansor and Farookh Khadeer Hussain
The Improved Redundant Active Time-Based (IRATB) Algorithm
for Process Replication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Tomoya Enokido, Dilawaer Duolikun, and Makoto Takizawa
The Actual Cost of Programmable SmartNICs: Diving
into the Existing Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Pablo B. Viegas, Ariel G. de Castro, Arthur F. Lorenzon, Fábio D. Rossi,
and Marcelo C. Luizelli
On the Adaptation of Bitcoin-Like Blockchains for the Secure
Storage of Traffic-Related Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
El-hacen Diallo, Omar DIB, and Khaldoun Al Agha
Contents xxvii

Trustworthy Acceptance: A New Metric for Trustworthy

Artificial Intelligence Used in Decision Making
in Food–Energy–Water Sectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Suleyman Uslu, Davinder Kaur, Samuel J. Rivera, Arjan Durresi,
Mimoza Durresi, and Meghna Babbar-Sebens
Classifying Holes, Voids, Negative Objects and Nothing . . . . . . . . . . . . 220
Katrina Hooper, Alex Ferworn, and Fatima Hussain
Empirical Study of Software Adoption Process
in the Bitcoin Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Mitsuyoshi Imamura and Kazumasa Omote
Investigation and Analysis of Features in Decentralized Network
Management of Minor Cryptocurrencies . . . . . . . . . . . . . . . . . . . . . . . . 245
Mitsuyoshi Imamura and Kazumasa Omote
DynAgentX Framework: Extending the Internet Management
to the VCPE in the SD-WAN Context . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Maxwell E. Monteiro
k-Level Contact Tracing Using Mesh Block-Based Trajectories
for Infectious Diseases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Kiki Adhinugraha, Wenny Rahayu, and David Taniar
Architecture for Real-Time Visualizing Arabic Words with Diacritics
Using Augmented Reality for Visually Impaired People . . . . . . . . . . . . . 285
Imene Ouali, Mohamed Saifeddine Hadj Sassi, Mohamed Ben Halima,
and Ali Wali
Progressive Mobile Web Application Subresource Tampering During
Penetration Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Tobiasz Wróbel, Michał Kędziora, Michał Szczepanik, Piotr P. Jóźwiak,
Alicja M. Jóźwiak, and Jolanta Mizera–Pietraszko
CO-BPG: A Centralized Optimizer for Routing in BGP-Based Data
Center Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Chen Duan, Wei Peng, and Baosheng Wang
Joint Channel and Power Allocation Algorithm for Flying Ad Hoc
Networks Based on Bayesian Optimization . . . . . . . . . . . . . . . . . . . . . . 320
Pengcheng Wang, Wei Peng, Wenxin Zhang, and Gaofeng Lv
Cluster-Based Distributed Self-reconfiguration Algorithm
for Modular Robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
Mohamad Moussa, Benoit Piranda, Abdallah Makhoul,
and Julien Bourgeois
xxviii Contents

Prediction of the Response Time in Web Services Based

on Matrix Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
Maksymilian Iwanow and Jolanta Wrzuszczak-Noga
BO-RL: Buffer Optimization in Data Plane Using Reinforcement
Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
M. Sneha, Kotaro Kataoka, and G. Shobha
Formal Modeling, Verification, and Analysis of a Distributed
Task Execution Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
Amar Nath and Rajdeep Niyogi
Deep Learning Library Performance Analysis on Raspberry
(IoT Device) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
Luciana P. Oliveira, José Humberto da S. S. Santos,
Edson L. V. de Almeida, José R. Barbosa, Arley W. N. da Silva,
Lucas P. de Azevedo, and Maria V. L. da Silva
Network Topology-Traceable Fault Recovery Framework
with Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
Tatsuji Miyamoto, Genichi Mori, Yusuke Suzuki, and Tomohiro Otani
Mobile Sensing Data Analysis for Road State Decision
in Snow Country . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
Yoshitaka Shibata, Akira Sakuraba, Yoshikazu Arai, Yoshiya Saito,
and Jun Hakura
Towards Industrial 5G: An LTE+ Case Study Report . . . . . . . . . . . . . . 414
Alexander Gogolev and Markus Aleksy
Thermal Environment Simulator for Office Building
with Multi-agent-Based Simulation Technique . . . . . . . . . . . . . . . . . . . . 424
Kodai Murakami, Toshihiro Mega, and Noriyuki Kushiro
Message Ferry Routing Based on Nomadic Lévy Walk in Wireless
Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
Koichiro Sugihara and Naohiro Hayashibara
Image Similarity Based Data Reduction Technique in Wireless Video
Sensor Networks for Smart Agriculture . . . . . . . . . . . . . . . . . . . . . . . . . 448
Christian Salim and Nathalie Mitton
Using Clusters for Approximate Continuous k-Nearest
Neighbour Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
Wendy Osborn
Contents xxix

Enhancement of Medical Information Security Through Lightweight

Cryptography and Steganography in the Platform
of Cloud-IoT Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
Humayra Binte Arfan, Zinia Anzum Tonni, Saria Jahin Taluckder,
and Rashed Mazumder
Blockchain-Based Trust Management Approach for IoV . . . . . . . . . . . . 483
Amal Hbaieb, Samiha Ayed, and Lamia Chaari
Improved Visual Analytic Process Under Cognitive Aspects . . . . . . . . . 494
Samar Bouazizi and Hela Ltifi
Secure Distribution of Factor Analysis of Mixed Data (FAMD)
and Its Application to Personalized Medicine
of Transplanted Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
Sirine Sayadi, Estelle Geffard, Mario Südholt, Nicolas Vince,
and Pierre-Antoine Gourraud
Comparative Study of Traditional Techniques for Unsupervised
Autonomous Intruder Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
Anik Alvi, Tarem Ahmed, and Mohammad Faisal Uddin
A Novel Dynamic Authentication Method for Sybil Attack Detection
and Prevention in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . 531
Wadii Jlassi, Rim Haddad, Ridha Bouallegue, and Raed Shubair
An Effective Location-Based Alert Messages Dissemination
Scheme for Software Defined Vehicular Networks . . . . . . . . . . . . . . . . . 542
Raoua Chakroun, Slim Abdellatif, and Thierry Villemur
Energy-Efficient Fuzzy Geocast Routing Protocol
for Opportunistic Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
Khuram Khalid, Isaac Woungang, Sanjay Kumar Dhurandher,
and Jagdeep Singh
Time Synchronization for Automotive Prototyping
and Testing Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
Esraa Elelimy and Andrew Azmy
Localization of Wireless Sensor Networks Using Negative Knowledge
Bilateration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Saroja Kanchi
UAVs Route Planning in Sea Emergencies . . . . . . . . . . . . . . . . . . . . . . . 588
Nicholas Formica, Leonardo Mostarda, and Alfredo Navarra
SQUAB: A Virtualized Infrastructure for Experiments
on BGP and its Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
Naoki Umeda, Naoto Yanai, Tatsuya Takemura, Masayuki Okada,
Jason Paul Cruz, and Shingo Okamura
xxx Contents

Scalability Support with Future Internet in Mobile

Crowdsourcing Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
Peron Sousa and Antonio A. de A. Rocha
HyDiLLEch: A WSN-Based Distributed Leak Detection
and Localisation in Crude Oil Pipelines . . . . . . . . . . . . . . . . . . . . . . . . . 626
Safuriyawu Ahmed, Frédéric Le Mouël, Nicolas Stouls,
and Gislain Lipeme Kouyi
Tactful Opportunistic Forwarding: What Human Routines
and Cooperation Can Improve? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
Rafael Lima Costa, Aline Carneiro Viana, Artur Ziviani,
and Leobino Nascimento Sampaio
Cooperative and Coordinated Mobile Femtocells Technology
in High-Speed Vehicular Environments . . . . . . . . . . . . . . . . . . . . . . . . . 653
Rand Raheem, Aboubaker Lasebae, Miltos Petridis, and Ali Raheem
Resource Allocation for Millimeter Wave D2D Communications
in Presence of Static Obstacles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 667
Rathindra Nath Dutta and Sasthi C. Ghosh
Bandwidth Slicing with Reservation Capability and Application
Priority Awareness for Future Vehicular Networks . . . . . . . . . . . . . . . . 681
Abdullah A. Al-khatib, Abdelmajid Khelil, and Mohammed Balfaqih
Proposal of a Disaster Response AI System . . . . . . . . . . . . . . . . . . . . . . 692
Tomoyuki Ishida and Tenma Hirose
An Efficient Image Codec Using Spline-Based Directional Lifting
Wavelet Transform and an Optimized EZW Algorithm . . . . . . . . . . . . 701
Rania Boujelbene and Yousra Ben Jemaa
An Innovative Monocular Mobile Object Self-localization Approach
Based on Ceiling Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
Alfredo Cuzzocrea, Luca Camilotti, and Enzo Mumolo
Analysis of Visualized Bioindicators Related to Activities
of Daily Living . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
Tomokazu Matsui, Kosei Onishi, Shinya Misaki, Hirohiko Suwa,
Manato Fujimoto, Teruhiro Mizumoto, Wataru Sasaki, Aki Kimura,
Kiyoyasu Maruyama, and Keiichi Yasumoto
Non-contact Person Identification by Piezoelectric-Based Gait
Vibration Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745
Keisuke Umakoshi, Tomokazu Matsui, Makoto Yoshida, Hyuckjin Choi,
Manato Fujimoto, Hirohiko Suwa, and Keiichi Yasumoto
Contents xxxi

Design of Healthcare Information Sharing Methods

Using Range-Based Information Disclosure Incentives . . . . . . . . . . . . . . 758
Miku Honda, Joji Toshima, Takuo Suganuma, and Akiko Takahashi

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771

 Optimization of Task Allocations in Cloud
to Fog Environment with Application
to Intelligent Transportation Systems

Fatos Xhafa1(B) , Alhassan Aly1 , and Angel A. Juan2

1
Universitat Politècnica de Catalunya, Barcelona, Spain
[email protected]
2
Universitat Oberta de Catalunya, Barcelona, Spain
[email protected]

Abstract. Fog and Edge computing are opening up new opportunities

to implement novel features of mobility, edge intelligence and end-user
support. The successful implementation and deployment of Fog layers,
as part of Cloud-to-thing-computing, largely depends on optimized allo-
cation of tasks and applications to Fog and Edge nodes. Similarly as
in other large scale distributed systems, the optimization problems that
arise are computationally hard to solve. Such problems become even more
challenging due to the need of application scenarios for larger computing
capacity, beyond those of single nodes, requiring thus efficient resource
grouping. In this paper we present some clustering techniques for creat-
ing virtual computing nodes from Fog/Edge nodes by combining seman-
tic description of resources with semantic clustering techniques. Then,
we use such clusters for optimal allocation (via heuristics and Integer
Linear Programming) of applications to virtual computing nodes. Sim-
ulation results are reported to support the feasibility of the model and
efficacy of the proposed approach. Applications of allocation methods to
Intelligent Transportation Systems are also discussed.

1 Introduction

A number of computing paradigms and technologies have emerged after IoT

and Cloud computing. Indeed, the view of IoT Cloud, in which IoT devices
are directly connected to Cloud platforms and Data centers has shifted to a
continuum view, where thereby various computing layers sit in between of IoT
and Cloud. These layers referred to as Edge and Fog computing aim to retain
the processing of data close to the source where the data is generated. OpenFog
Consortium defines Fog computing as “A horizontal, system-level architecture
that distributes computing, storage, control and networking functions closer to
the users along a Cloud-to-thing continuum”. Besides significantly alleviating
the computing burden to the Cloud, Edge and Fog layers enable faster round-
trip processing and better support to end users. Most importantly, by processing
the data close to the data source and to end-users, real time decision making
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 1–12, 2021.
https://doi.org/10.1007/978-3-030-75100-5_1
2 F. Xhafa et al.

can be designed to support a variety of use cases. This is made possible by

achieving the computing and data fabric view of Edge computing [18], where
different computing granularity devices can collaboratively support both real
time processing and data storage requirements.
Cloud-to-thing-continuum brings the advantage of both real time processing
and in-depth processing of IoT data. Real time processing enables, in particular,
the detection of anomalies or abnormal events [4,19] regarding transportation
fleet, etc., which would require immediate intervention by managers and local
stakeholders. Further, massive offline processing enables computing of advanced
descriptive, predictive and prescriptive analytics to support decision making and
business intelligence and big data solutions. Cloud-to-thing-continuum enables
the implementation of a plethora of application scenarios including smart logis-
tics [8] and intelligent transportation systems [13,21]. Indeed, with computing
layers of Fog and Edge computing, the premise is that latency and round trip
time can be drastically reduced providing thus support to real time applications,
mobility, etc.
In case of intelligent transportation systems (ITS), a new generation of ITS
is being envisaged and deployed based on OBU–On Board Units– and RSU –
Road Site Units– technologies deployment in road transport networks [5,11,15].
The information gathered by OBU in vehicles and RSU units in roads enables
information gathering and can be used to evaluate in real time the traffic density
and to predict it for short term periods as well as to compute more accurate maps
of road traffics density. Through cooperative sensing and collective intelligence
improved traffic efficiency can be achieved [9].
The implementation of applications in the Cloud-to-thing-continuum, how-
ever, comes with a number of challenges, among which we could distinguish the
high degree of heterogeneity of resources, the variability in computing in the
continuum, the scale and reliability. In this context, the allocation of computing
resources becomes more complex than in traditional distributed systems. Many
applications may require more computing resources than single Fog/Edge nodes
can provide. In this paper, we address this issue by using semantic clustering of
Fog resources into virtual nodes of larger computing capacity, to which applica-
tions can be allocated. Semantic categorization of resources enables a coherent
clustering of resources and thus building clusters both statically and dynam-
ically to cope with application requirements. Then, such clusters are used as
input information to optimization procedures (heuristic methods and Integer
Linear Programming) to optimize the allocation. Feasibility and efficacy of the
proposed approach is evaluated via simulation.
The rest of the paper is structured as follows. In Sect. 2, we present the model
for clustering of small nodes into larger virtual nodes; an application scenario
from traffic load prediction is also given. In Sect. 3 we present methods for opti-
mizing task/application allocation to clusters of nodes. Then, in in Sect. 4, we
report some experimental results obtained via simulation of the Fog/Edge model
and various application profiles. We end the paper in Sect. 5 with conclusions
and future work.
 Optimization of Task Allocations in Cloud to Fog Environment 3

2 Modeling of C2F Environment and Semantic Clustering

We consider the architecture depicted in Fig. 1. Fog nodes could be composed
of one or more layers, however in our implementation we consider just one layer
between the Cloud and the edge devices.

Fig. 1. Architecture overview

As shown in Fig. 1 we first have a Cloud layer which contains the Cloud
infrastructure, then the Fog layer containing the Fog nodes which are connected
to the Cloud through the internet. The connections between the Cloud and the
Fog nodes pass through several hops. The physical distance between the Cloud
data centers, and the Fog nodes affects data transfer rate, since it increases both
latency and potential packet loss. The Fog nodes are directly connected to the
edge devices i.e. one hop away, which makes the data transfer faster, more stable
and reliable.
The Cloud node has abundant memory and processing power, while the Fog
nodes are rather limited in comparison. The edge devices in our system will just
have one application each that they need to run using certain specifications.
4 F. Xhafa et al.

2.1 Semantic Categorization of Resources, Clustering and

Offloading

The need to create groups of devices in Edge lies in exploiting capabilities of

several devices simultaneously to carry out a computing intensive task collabo-
ratively and efficiently or because the Edge devices separately cannot complete
the task under the requirements of efficiency, latency, etc. The devices that are
part of a group can share network computing resources among themselves to
meet the requirements of the tasks/applications. To do this, we first propose to
apply techniques of semantic categorization of resources at the Fog/Edge layer.
Through semantic characterization, for example through RDF, the information
of Edge devices such as their computational capacity (CPU, RAM, bandwidth)
and storage capacity, the specificity of each resource (heterogeneity) and enriched
with contextual information such as location, information on mobility, security,
trust, etc.
Virtual Clusters
The clusters of devices in Edge, referred to as virtual nodes, can be formed
based on semantic information of machine-readable format according to sev-
eral criteria: (1) similarity of the devices (homogeneity); (2) device location; (3)
computational and reliability performance analytics of a device and (4) device
behavior analytics in terms of security, privacy and trust [6]. The first criterion
allows similar devices in a cluster in terms of their computational characteristics
in order to avoid disparity between devices and that some of them could become
bottlenecks in the collaborative performance of a computational task. The sec-
ond criterion allows to obtain device clusters “close to” an end user or to a data
source and thus explore the “user-proximity” or “data-locality” depending on
the requirements of the application. The third criterion can be taken into account
when assigning tasks to groups simply to distribute the computing load. Finally,
the fourth criterion would be applicable when it comes to processing sensitive
data.
Static Clusters
We can calculate a priori static clusters, for example using the semantic sim-
ilarity criterion (semantic clustering) or clustering based on the location (geo-
clustering), which can be useful for certain types of applications whose require-
ments are known a priori, for example to process data from a stream of fixed
cameras on a highway. However, static clusters would not be suitable for all
types of applications and scenarios, because finding a priori a suitable cluster
size that fits various applications and scenarios is very difficult. In particular,
static clusters cannot incorporate the current state of the device workload at
run-time.
Dynamic Clusters
To satisfy the computational demands of the applications without knowing a
priori their requirements, that is, at run-time, it is necessary to create clusters
of devices dynamically and taking into account the computational load of the
 Optimization of Task Allocations in Cloud to Fog Environment 5

devices. To do this, it is proposed to formulate optimization problems that cal-

culate clusters of Edge devices to be assigned to computational tasks (either
workflow or bag-of-tasks) in order to optimize the execution of the tasks and
optimally use the Edge resources.
Offloading Algorithms
Offloading algorithms is a class of algorithms that allow to delegate or unload
the computation or performance of tasks between parts of the same system or
between different computer systems. The reasons for downloading the compu-
tation from an entity/system A to an entity/system B can be diverse (one or
more at a time) but all of them try to take advantage of the B resources or fea-
tures to achieve system performance, resource utilization, alleviating the burden
of computations, etc.
Offloading algorithms arise very naturally in a Cloud-to-thing continuum sys-
tem. In fact, the offloading algorithms are the very essence of such systems [14].

2.2 Application Scenario: Prediction of the Volume of Traffic by

Images Of Mobile Crowd-Sensing
This application scenario aims to process a data flow of images sent by smart
phones (mobile crowd-sensing) to predict the volume of traffic in real time in
large cities. Traffic jams in large cities constitute a social problem and have
significant economic costs to transportation systems. An alternative to calculate
the congestion map of a city in real time and to make the prediction of traffic vol-
ume is to apply crowd-sensing techniques. Prediction of the volume of traffic can
be very useful to optimize performance parameters in Intelligent Transportation
Systems.
Specifically, thousands of drivers can send images of the state of the roads
in a city in real time, forming a stream of images and corresponding positions
whose real-time processing can provide valuable information on the current state
and predict traffic in the city. The premise is in the large number of images that
allows to achieve the prediction with high precision. This problem is defined
as predicting real-time spatial information from data collected by mobile IoT
sensors.
Therefore mobile crowd-sensing constitutes an alternative to the costly
deployment and maintenance of video-camera infrastructures throughout a city,
especially in large cities where the cost of deployment and maintenance can be
very high. However, crowd-sensing has intrinsic problems and challenges:
• Driver participation is voluntary and mass participation is not guaranteed.
• The quality of the images may be poor or just fake, which could affect the
accuracy of the prediction.
• The images can reveal information about the license plates of the cars and
the information received must be anonymized.
The volunteer participation of drivers can be addressed by implementing
a participation incentives scheme (a problem of interest to Vehicular Net-
works [10,17]). Similarly, the second point can be addressed by implementing
6 F. Xhafa et al.

a classifier based an image quality model of image features that allows deducing
the contribution of an image in predicting traffic volume and thus to be able
to implement an efficient classifier of the images. Finally, for the third problem,
we can apply the state of the art to anonymize the license plates of cars in the
images. The reader is referred to [16] for more details.

3 Optimization of Task/Application Allocation

The responsibility of the solver in our system is to perform the mapping between
the tasks in an application to Fog nodes within a cluster. Its function is to find
which Fog node is going to execute which tasks of an application. Thus, mapping
could be considered as an optimization problem. The problem could be described
as a variant to the bin-packing problem [2], where we have a number of items
that we want to place in a number of bins, with a common capacity and minimize
the number of used bins. Unlike the traditional problem, we will have different
capacities to represent the memory of the Fog nodes. We first apply ILP –Integer
Linear Programming– to get the optimum solution, then we solve the problem
using a First Fit Heuristic Algorithm (FFHA) [20] approach to decrease the
run-time of the optimization.
The objective is minimizing both the number of Fog nodes used, and the
processing time by allocating as many tasks to Fog nodes with higher CPU as
possible.
We formalize the problem as follows. Let us denote by f –a Fog node, F –the
set of Fog nodes, fpt –the processing time of node f and f used defined as equal
to 1 if node f is used, 0 otherwise.

min fpt × f used
f ∈F

subject to:

∀t ∈ T :

xtf = 1
f ∈F

∀f ∈ F :

t size × xtf ≤ fRAM × f
t∈T

where:

• xtf : equals 1 if task t is placed in Fog node f, 0 otherwise.

• t size: Size of task t
• fRAM : RAM capacity of f.
 Optimization of Task Allocations in Cloud to Fog Environment 7

The constraints of the ILP ensure that a task can only exist in one Fog node,
and that the sum of task sizes cannot exceed the memory capacity of the Fog
node they are placed in respectively.
For the ILP method, we use the PuLP Python package to solve the optimiza-
tion problem. PuLP can generate Mathematical Programming System (MPS)
or Linear Programming (LP) files and call GLPK [12], COIN CLP/CBC [3],
CPLEX [1], and GUROBI [7] to solve linear problems. PuLP uses Coin-or Branch
and Cut (CBC) as its default solver, which is an open-source mixed integer pro-
gramming solver written in C++.
For the FFHA, we define our own algorithm to solve the specific problem
at hand, where we have a variable number to represent Fog node’s RAM. The
pseudo-code of the FFHA is shown in Algorithm 1.

Algorithm 1. First Fit Heuristic Algorithm

for F ogs = 1, 2, . . . do
for index = 1, 2, . . . , inRangeF ogSizes.length do
Assign taskSize in sizes to sizes[index][1]
if f illedF ogselectedF ogSize . . . then
if taskSize <= (F ogSize − f illedF ogSize) then
f illedF og + taskSize
else
break
end if
else
return to loop
end if
end for
Return assignment results
end for

As mentioned before, every time we run the simulation, the Fog nodes and
the edge devices are generated in different positions with random attributes.

4 Experimental Evaluation

To assess the performance of our implementation, we benchmark the overall pro-

cess by running the simulation with different sets of numbers for the Fog nodes
and the devices. Since the semantic reasoning and clustering directly affects the
overall performance, we can test how does the system scale by increasing the
number of the Fog nodes and edge devices. Moreover, we compare between the
two solvers we used by focusing on the differences in run-times, and the appli-
cation response times obtained from the varied possible scenarios of how the
application is being sent.
8 F. Xhafa et al.

Various scenarios are considered, namely, the application is directly sent to

the Cloud, the application is sent to the Cloud through a Fog node, the appli-
cation is sent to a single Fog node, the application is sent to an In-range cluster
and the application is sent to a neighbor cluster.
We evaluate the utility of the two proposed clustering techniques, by compar-
ing and the response times obtained from each method of sending applications.

4.1 Experimental Setting and Results

The machine specifications used in the simulations to measure the performance

of the system are as follows.
All tests where executed on a single machine with the following specifications:
Windows 10 Pro, Intel Core i7-4710HQ @ 2.5 GHz, base Clock Speed up to 3.5
GHz Turbo Speed, 6 MB L3 cache, 16 GB DDR3 @ 1600 MHz, Dual channel
and 500 GB NAND Sata SSD.
Since the simulation generates random attributes for the Fog nodes, the edge
devices and the connections, we run the simulation for number of times and
calculate the mean values to acquire statistically significant results. On the one
hand, for bench-marking and comparison between the run-times of the solvers,
we run the simulation using different sets of numbers for the Fog nodes and edge
devices. For each set of numbers, we run the simulation 50 times with the ILP
solver and 50 times with the FFHA. On the other hand, for the response time
analysis, we run the simulation with the same number of Fog nodes and edge
devices 50 times. In that manner, we focus on how the clustering techniques
could affect the response times.

4.2 System Performance

To evaluate the optimization techniques used to solve the task allocation prob-
lem, we focus on the response times of the applications in different scenarios,
for which we measure the average run-time for all the scenarios. We run the
simulation with the following sets of parameters 50 times with the ILP solver
and 50 times with the FFHA.

Fig. 2. Response times for 5 Fogs and 250 devices (ILP - FFHA)
 Optimization of Task Allocations in Cloud to Fog Environment 9

The results show that the ILP solver acquires slightly better response times
than that of the FFHA, in the case of running the simulation with 5 Fog nodes
and 250 devices as shown in Fig. 2.

Fig. 3. Response times for 10 Fogs and 250 devices (ILP - FFHA)

When running the simulation with 10 Fog nodes and the same sets of devices,
the results show that in some cases, the FFHA gets better response times com-
pared to the ILP solver as shown in Fig. 3.

Fig. 4. Response times for 20 Fogs and 250 devices (ILP - FFHA)

The Fig. 4 shows the results when running the simulation with 20 Fog nodes.
We can see that the results are similar to running the simulation with 10 Fog
nodes. The ILP solver performs slightly better than the FFHA, yet the difference
is not significant.
Overall, regarding the application response times, given the different methods
of sending the application. The results gathered from running the simulation
with the aforementioned sets of numbers for the Fog nodes and the devices, we
can deduce that the ILP solver has a slight edge over the FFHA. However, the
difference is rather minor. That is due to the fact that we perform the memory
checks, as an initial filter, then the solvers optimize where the tasks are processed
to get the least possible response time.
In contrast, when reviewing the run-times of the simulations, there is a sub-
stantial difference between the performance of the ILP solver and the FFHA,
where the FFHA performs much faster than the ILP solver as shown in Fig. 5.
10 F. Xhafa et al.

Fig. 5. Run-times (ILP vs FFHA)

On one hand, increasing the number of edge devices linearly affects the run-
time. On the other hand, we can see an exponential increase in run-time when
increasing the number of Fog nodes, which is unsurprising, since this increases
the number combinations when forming either an in-range or neighbor cluster.
To measure the consistency of the optimization techniques, we calculate the
standard deviation, of both the response times and the run-times by conducting
another experiment, using an instance size of 10 Fog nodes and 100 devices,
and 100 iterations. Additionally, we calculated the number of Fog nodes in the
clusters from the same instance, to give insight on the size of the clusters. Tables 1
and 2 show the average response times and run-times with their corresponding
Standard Deviations (STD), when running the simulation using the ILP solver
and the FFHA respectively.

Table 1. ILP - Averages and standard deviations of response times & run-time

ILP
Avg response times Response times STD
Cloud direct 17.1991722 8.71893098
Cloud through fog 2.18383078 0.23542268
Single fog 0.43523073 0.10711002
In-range cluster 1.09988548 0.21959514
Neighbor cluster 0.88225381 0.21159424
Avg run-time Run-time STD
3.82993240 0.81419498
 Optimization of Task Allocations in Cloud to Fog Environment 11

Table 2. FFHA - Averages and standard deviations of response times & run-time

FFHA
Avg response times Response times STD
Cloud direct 18.2210832 10.5720824
Cloud through fog 2.19942492 0.26223306
Single fog 0.45561651 0.11327142
In-range cluster 1.09464597 0.19415273
Neighbor cluster 0.93392979 0.22757457
Avg run-time Run-time STD
0.22792500 0.09110830

When comparing the run-times STD, we can see that the FFHA behaves
more consistently than the ILP solver, with less deviation.

5 Conclusions and Future Work

In this paper we presented a model for Cloud to Fog computing (C2F) and
allocation of tasks and applications in such environment. The aim is to evaluate
the benefits of using semantic categorization of resources, semantic clustering
and optimizing task allocation to such clusters. Our approach shows that there
is a massive gain in response times when utilizing the Fog computing paradigm,
compared to just relying on the Cloud. Based on the results, it is apparent
that the application response times are understandably faster when utilizing Fog
nodes or clusters. Integer Linear Programming (ILP) and an heuristic method
are used for optimization purposes in this work. Various settings are considered
to evaluate the scalability of the solvers. Although various application scenarios
have been used in the study, more application profiles would be needed to achieve
conclusive results on the allocation methods.
In our future work we would like to evaluate the proposed approach on a real-
istic scenario of transportation systems with V2X communication and Intelligent
Road Side Units. Scenarios of safety, decreased trip times and energy savings will
be considered. Also, we plan to use the iFog simulator to evaluate whether the
proposed heuristic can give adequate response to mobility of Edge nodes.

Acknowledgements. This work is supported by Research Project, “Efficient &

Sustainable Transport Systems in Smart Cities: Internet of Things, Transport
Analytics, and Agile Algorithms” (TransAnalytics) PID2019-111100RB-C21/AEI/
10.13039/501100011033, Ministerio de Ciencia e Innovación, Spain.

References
1. Ceselli, A., Fiore, M., Premoli, M., Secci, S.: Optimized assignment patterns in
Mobile Edge Cloud networks. Comput. Oper. Res. 106, 246–259 (2019)
12 F. Xhafa et al.

2. Coffman Jr., E.G., Csirik, J., Galambos, G., Martello, S., Vigo, D.: Bin packing
approximation algorithms: survey and classification. In: Pardalos, P., Du, D.Z.,
Graham, R. (eds.) Handbook of Combinatorial Optimization. Springer (2013)
3. COIN-OR (Common Infrastructure for Operations Research). http://www.coin-
or.org
4. Corizzo, R., Ceci, M., Japkowicz, N.: Anomaly detection and repair for accurate
predictions in geo-distributed big data. Big Data Res. 16, 18–35 (2019)
5. Fouchal, H., Bourdy, E., Wilhelm, G., Ayaida, M.: A validation tool for cooperative
intelligent transport systems. J. Comput. Sci. 22, 283–288 (2017)
6. Gago, M.C.F., Moyano, F., López, J.: Modelling trust dynamics in the Internet of
Things. Inf. Sci. 396, 72–82 (2017)
7. Gurobi Optimization, LLC., Gurobi Optimizer Reference Manual (2020). http://
www.gurobi.com
8. Faulin, J., Scott S.E., Juan, A.A., Hirsch, P.: Sustainable Transportation and Smart
Logistics. Decision-Making Models and Solutions. Elsevier (2019)
9. Ferreira, J., et al.: Cooperative sensing for improved traffic efficiency: the highway
field trial. Comput. Netw. 143, 82–97 (2018)
10. He, Z., Cao, J., Liu, X.: High quality participant recruitment in vehicle-based
crowdsourcing using predictable mobility. In: 2015 IEEE Conference on Computer
Communications, INFOCOM 2015, Hong Kong, 2015, pp. 2542–2550 (2015)
11. Mfenjou, M.L., et al.: Control points deployment in an Intelligent Transportation
System for monitoring inter-urban network roadway. J. King Saud Univ. Comput.
Inf. Sci. In Press. https://doi.org/10.1016/j.jksuci.2019.10.005
12. Makhorin, A.: GLPK (GNU Linear Programming Kit). http://www.gnu.org/
software/glpk/glpk.html
13. Hussain, M.M., Alam, M.S., Sufyan Beg, S.: Vehicular fog computing-planning and
design. Procedia Comput. Sci. 167, 2570–2580 (2020)
14. Santa, J., Fernández, P.J., Sanchez-Iborra, R., Murillo, J.O., Skarmeta, A.F.:
Offloading positioning onto network edge. Wirel. Commun. Mobile Comput. 2018,
7868796:1–7868796:13 (2018)
15. Santa, J., Bernal-Escobedo, L.: On-board unit to connect personal mobility vehicles
to the IoT. FNC/MobiSPC, Ramon Sanchez-Iborra, pp. 173–180 (2020)
16. Shinkuma, R., Takagi, T., Inagaki, Y., Oki, E., Xhafa, F.: Incentive mechanism for
mobile crowdsensing in spatial information prediction using machine learning. In:
AINA, pp. 792–803 (2020)
17. Wang, X., Wu, W., Qi, D.: Mobility-aware participant recruitment for vehicle-based
mobile crowdsensing. IEEE Trans. Veh. Technol. 67(5), 4415–4426 (2018)
18. Xhafa, F.: The vision of edges of internet as a compute fabric. In: Advances in Edge
Computing: Massive Parallel Processing and Applications. Book Series: Advances
in Parallel Computing Series, Chapter 1. IOS Press (2019)
19. Xhafa, F., Kilic, B., Krause, P.: Evaluation of IoT stream processing at edge com-
puting layer for semantic data enrichment. Future Gener. Comput. Syst. 105,
730–736 (2020)
20. Yesodha, R., Amudha, T.: A comparative study on heuristic procedures to solve
bin packing problems. Int. J. Found. Comput. Sci. Technol. 2, 37–49 (2012)
21. Zhang, H., Lu, X.: Vehicle communication network in intelligent transportation
system based on Internet of Things. Comput. Commun. 160, 779–789 (2020).
Special Issue on Internet of Things and Augmented Reality in the age of 5G
 User Allocation in 5G Networks Using
Machine Learning Methods for Clustering

Christos Bouras(B) and Rafail Kalogeropoulos

Computer Engineering and Informatics Department, Patras, Greece

[email protected], [email protected]

Abstract. The rapid increase in the volume of devices connected to

mobile networks poses unprecedented demands on existing network-
ing infrastructures. Machine Learning is a promising technique, already
applied in various sectors of our everyday lives. It enables decision mak-
ing not with the use of traditional programming but rather by using data
to train models to cope with various problems without explicit program-
ming on how to do so. The integration of Machine Learning techniques is
deemed necessary in as many processes as possible to help the network
face congestion and enable efficient real time decision making. In this
paper we present two Machine Learning based mechanisms for improv-
ing real time user allocation on the network as well as predicting the
best positioning scheme for Smallcell Base Stations to provide effective
utilization of the network’s resources.

1 Introduction

The introduction of more capable mobile devices has set the pace for a complete
redesing of all cellular networks and their infrastructures. These new devices
are able to sustain a plethora of applications. Their computing power always
allows for the development of new apps while the establishment of communica-
tion standards allows them to connect with a multitude of sensors and devices.
This motivates the integration of new devices such as mobile phones and sensors
into the network. They can be used for a variety of purposes such as medical [9]
or agricultural use [2].
Wireless networks as a result face an unprecedented rise in the number of
connected devices that seek to utilize their resources, while the volume of data
transfered through the network’s infrastructure increases as well. In contrast to
previous generations of networks that mostly consisted of a Base Stations (BSs)
of the same type, namely Macro cell Base Stations (McBSs), in latest generations
we saw the introduction of Smallcell Base Stations (SBSs), smaller BSs that can
be further categorized based on their transmit power.
In previous generations, McBSs prevailed due to a higher transmit power,
ensuring high Signal to Interference plus Noise Ratio (SINR), especially in the
Downlink (DL) direction, while performance in the Uplink (UL) direction was

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 13–24, 2021.
https://doi.org/10.1007/978-3-030-75100-5_2
14 C. Bouras and R. Kalogeropoulos

deemed secondary. All McBSs feature the same characteristics throughout the
entire network and are also capable of supporting a similar amount of users
[12]. As a result, User Equipment (UE) were associated on the same BS for both
directions, based solely on DL performance. These networks, called Homogeneous
Networks were applicable since the required throughput for DL was tradition-
ally higher than UL, resulting in an asymmetric utilization from the two traffic
directions [10]. This approach is now completely dated, with the plethora of new
devices uploading in the UL direction.
The fifth generation, considers the DL and UL direction as separate networks
and does not demand users to connect to the same BS for both of them [6]. Such
a technique is called Downlink and Uplink Decoupling (DUD). These networks
are called heterogeneous networks (HetNets). Such networks consist of prominent
McBSs and various SBSs that are scattered among a McBS’s vicinity [8]. They
are necessary to ensure accessibility for all connected devices and guarantee a
respectable Quality of Service (QoS) in real-time.
In HetNets, SBSs become more prevalent since more traffic is generated
on the UL direction. Most network connected mobile devices are battery pow-
ered with identical transmit power, though with different demands. For example
autonomous cars rely on sensors that produce data needed for autonomous driv-
ing. They are extremely dependent on accuracy to function properly and demand
high QoS while other applications such as music streaming usually have less QoS
demands. This increases dramatically the complexity of the UL direction.
The use of Machine Learning (ML) offers a significant opportunity to refine
existent network applied methods and it can even work supplementary to them to
improve real time performance. So far network application of ML techniques con-
tains the work of [1], where the authors explore an application of ML techniques
to explore unknown guest user dynamics for a network of users with different
demands. Self-organizing maps applied to cellular networks have been proposed
to dynamically shape the connectivity of the network to the prospective demands
in [4]. The survey in [3], discusses multiple open issues on ML applications for
computer networks and potential problems that arise with them.
In this paper we will propose two ML based mechanisms. Our first proposal
is based on Decision Trees, that can be trained on data produced by already
deployed networks, pinpoint patterns on optimal user allocation and produce
promising results without the need to assess all these metrics that are necessary
for traditional user allocation techniques. Our second proposal can function sup-
plementary to our first proposal, and predict the optimal positioning of SBSs
based on user distribution across the network. This mechanism utilizes the k-
means clustering algorithm to produce cluster centers, which we consider as
candidates for positioning SBSs. Our proposals can be utilized for Efficient user
allocation in networks and can be proven extremely helpful in research that aims
to provide an efficient scheme for enriching network infrastructures.
In remaining sectors, we present our full proposal. Sect. 2 focuses on the
system model. In Sect. 3 we present our two ML mechanisms. Section 4 presents
our simulation parameters while Sect. 5 presents the simulation results. Finally,
 User Allocation in 5G Networks Using Machine Learning Methods 15

in Sect. 6 we draw our conclusions and we make suggestions on how ML can be

introduced in future research.

2 System Model
With the emerge of new technologies like the Internet of Things (IoT), we expect
radical changes in computer networks, such as a massive rise in the number of
UEs that try to utilize the network’s resources. As Small Cells become smaller we
expect their deployment to be massive in current and next generation networks.
These two cases create the need for solutions on how to manage such an increase
in the number of connected UEs and how to position and utilize new SBSs to
optimally match a UE with the appropriate BS for its UL/DL needs.
Regarding the basic network layout, we consider a Heterogeneous Network.
Our network consists of BSs (McBSs and SBSs) and users (UEs). McBSs are
denoted as M (M=1,...,|M |), SBSs are denoted as S (S=1,...,|S |) and the set
of UEs is denoted as U (U=1,....,|U |). Network users attempt to transmit and
receive data. Traffic can be split into two networks for transmitting and receiving
data (UL Network and DL network). In HetNets, they are considered as separate
channels. All BSs have limited resources, meaning they have a bound on the
maximum number of users they can serve simultaneously. All BSs of the same
type are considered to have the same resources.
To compute the number of RBs that a user (suppose user j) demands from
a specific BS, for achieving their desired DR, we will use the following equation:
Tj
RBj,i = [ ], (1)
BRB ∗ log2 (1 + SIN Rj,i )
where Tj denotes the UE throughput demands and DRi the desired Data
Rate for the user j, BRB is the bandwidth of a RB and SIN Rj,i is the SINR
between a BS and an associated user.
Next we shall define the equations for calculating Pathloss (PL) and Signal
to Interference and Noise Ratio (SINR). We use the distance dependent Path
loss model, to calculate pathloss, and thus we have two different equations, one
for McBSs and one for SBSs. PL is used to measure the signal loss we can expect
from a signal as the user receives it, relatively to when it was emitted from the
BS. The equations are given below, both for McBSs and SBSs:

P LM = 128.1 + 37.6 ∗ log10 d, (2)

P LS = 140.7 + 36.7 ∗ log10 d, (3)

where d is the distance with a user and its serving BS.

Next we shall define SINR for DL and UL. SIN Rj,i corresponds to the SINR
between a user j and its corresponding BS i. This metric represents a ratio,
namely the strength of the signal received by the receiving antenna and can be
calculated as:
UL PBS
SIN Ri,j = , (4)
N +I
16 C. Bouras and R. Kalogeropoulos

UL Pue
SIN Ri,j = , (5)
N +I
Here, Pue corresponds to Transmit Power (TP) of the UE, while PBS is the
TP of the BS, whether it is a McBS or a SBS. Regarding noise and interferences,
N corresponds to the Noise power while I corresponds to the total interferences
[7]. We try to simulate a typical metropolitan area network scenario. Considering
that most real life networks suffer from great Non Line Of Sight (NLOS) issues,
we assume that our network features more SBSs compared to McBSs. These
SBSs aim to alleviate congestion in areas with high user density.
For our simulation, the main focus of applying ML is associating users and
BSs, minimizing the amount of calculations necessary. Our goal is to reduce time
complexity and enable better real time decision making. Network load should be
efficiently distributed among all BSs, retaining an acceptable QoS level for all
users. For our simulations we utilize two ML techniques that are quite different,
yet incorporating them into the network can yield promising results. The two
techniques are:

2.1 Decision Trees

A decision tree is a decision tool that creates a tree like presentation to model
decisions and their possible outcomes. The decision tree consists a rooted tree,
since it has a node called “root” with zero incoming edges while all other nodes
feature exactly one incoming edge. Nodes with no outgoing edges are called
leaves (and in this case they can also be called decision nodes). In a decision
tree, based on a designated function each internal node splits the instance space
into two or more sub-spaces.
Each leaf may be assigned to one class or it may hold a vector indicating the
probability of the target attribute assigned to a certain value. The classification
is produced by navigating from the tree root all the way down to a leaf [5]. In
other words, following this path we can see all decisions made by the Decision
Tree to come up with the depicted result (depicted on the leaf).

2.2 K-Means Clustering

Clustering is the process of grouping a set of points into “clusters”, where points
with a small distance belong to the same cluster, where points with a large
distance are placed in different clusters. K-means algorithm assumes a Euclidean
space and a predifined number of clusters. These two characteristics produce two
of the most prominent issues with k-means clustering [11]. It is not always easy
to know in advance the number of clusters, while it also creates similarly shaped
clusters (ball-shaped), since the clusters the based on Euclidean distance.
 User Allocation in 5G Networks Using Machine Learning Methods 17

For the algorithm itself, there are several ways to define the starting cluster
centers. They can be random or we can select points from the dataset we want to
process. All the points that we want to classify need to be assigned to the nearest
cluster, meaning that for each point we need to calculate its Euclidean distance
to the cluster center. As new points are added to the clusters, the cluster centers
are constantly re-evaluated. The process can stop when we have no more points
to be clustered. As an optional step, in the end we can fix the cluster centers
and we can re-examine all points. Euclidean distance can be calculated as:

 n

D = ||X − Z|| =  (xi − zi )2 , (6)
n=1

where xi and zi , are the coordinates of the points (user and BS) in question.

3 The Mechanisms

In this sector we present our two proposed mechanisms. These mechanisms can
stand alone or they can work collectively to provide a prediction mechanism that
aims to refine user allocation on BSs and provide an adequate BS positioning
scheme that takes into consideration user distribution across the network.

3.1 Mechanism for Predicting User Allocation on BSs

The first mechanism is based on Decision Trees. We expect our Decision tree-
based model to able to predict the best matched BS for each user based on a
specific metric. We begin by modeling a network where all users are distributed
uniformally across the network and allocated to a certain BS using SINR as
the preferred allocation metric. For this scenario, when the network model has
decided the best matching BS for each user based on our metric, the allocation
results are saved on a dataset. The dataset features the coordinates of each user
in the network, as well as their matching BS both on the DL as well as the UL
direction. This dataset is then utilized by our ML based model for training and
testing.
The size of the dataset varies with the number of users that are deployed on
the network. We test our mechanisms for various dataset sizes. The produced
dataset is split into two different datasets. The first one being the training dataset
and the second one is the test dataset. The training dataset is fitted into a
ML model that is based on the Decision Trees technique. After the model has
been trained, we use the test dataset to assess its capabilities and calculate its
precision. We will evaluate the model precision for a variable number of users to
see how the dataset size affects the model performance.
18 C. Bouras and R. Kalogeropoulos

Algorithm 1. Pseudocode for the first mechanism

U: Denoting the number of users
for i = 1 to U do
Calculate SINR for all BSs on DL and UL;
Create BS preference list for DL and UL over SINR;
Associate to best matched BS for DL and UL;
end for
Produce dataset with User coordinates and DL,UL associated BS
while Precision is low do
Split produced dataset into training dataset and test dataset
Fit training dataset into Decision trees model
Use test dataset to predict associations and calculate model precision
end while

3.2 Mechanism for Efficient Placement of SBSs

The second proposed mechanism is based on K-means Clustering and aims to

provide a reliable method to optimally position BSs across the network so that
we can accommodate as many users as possible to avoid congestion escalation
in the network. The network consists of a set of users, distributed uniformly
across the network to ensure a realistic deployment scenario. These users can be
classified, based on their positions where users that are closer together, will be
classified in the same class.
To complete this task we use the k-means clustering algorithm. This algo-
rithm starts with a set of empty clusters (we begin with 42 clusters - the same
number as the number of BSs in our network) where users will be allocated. The
algorithm then places users into clusters based on their distance from the cluster
center. As a new user is allocated into a cluster, the coordinates of the cluster
center are calculated again to account for the new user that is now part of the
cluster. It is essential that all users must connect to some cluster, so that we
can produce accurate coordinates for the cluster centers. When the procedure
is completed, for each of these final cluster centers, we calculate their distance
from all BSs.
We then identify the minimum distance for all clusters and any BS. Cluster
centers that share a minimum distance with McBSs, are ignored while the coor-
dinates of cluster centers that are placed near SBSs, are considered to be the
optimal coordinates for placing the SBSs in our network. Finally we compare
the network performance using random coordinates for SBS positioning vs the
positioning scheme that we proposed. The coordinates of cluster centers that are
closer to McBSs are deleted, since McBSs are stationary and cannot be moved
and thus we cannot propose a cost-efficient repositioning scheme for them.
 User Allocation in 5G Networks Using Machine Learning Methods 19

Algorithm 2. Pseudocode for the second Mechanism

U: Denoting the number of users
Start with XX empty clusters
for i = 1 to U do
Associate user with a specific cluster
Calculate new cluster center
end for
Calculate cluster centers distance with all BSs
Delete cluster centers closer to McBSs
for clusters centers that are closer to SBSs do
Use calculated cluster centers as the new SBS coordinates
end for

4 Simulation Setup
In this sector we will present the parameters used to model our 5G network. All
network simulations and the proposed mechanisms were produced using Python.
We used Python because it incorporates functions for a plethora of ML tech-
niques, making it a powerful tool for ML centric simulations. Our produced
network, can be seen on Fig. 1. This network consists of 13 McBS, 29 SBSs and
200 UEs. McBS are presented as big triangles placed at the center of the net-
work hexagons, SBSs are presented as “Y” figures, while all users are presented
as bullets.

Fig. 1. The Python produced network, consisting of 13 MBSs, 29 SBs and 200 users.

In both of our simulations we have the same number of users, and they
are distributed along the network using the uniform distribution to ensure a
uniform distribution across the entirety of our network. This ensures a realistic
deployment with users wanting to exploit the entirety of our network. In the
second mechanism, we aim to produce user clusters. Using the Euclidean distance
20 C. Bouras and R. Kalogeropoulos

as the distance metric for the k-means algorithm, all produced clusters seem to
revolve around the cluster centers. Our mechanism is provided with the number
of wanted clusters, that is equal to the number of BSs in the network, a feature
that we can utilize if we want to expand the network with new BSs. All simulation
parameters can be seen on Table 1.

Table 1. Simulation parameters

Parameter Setting
Macro cell transmition power 50 dbm
Macro cell transmition power 24 dbm
User equipment transmition power 20 dbm
McBS Pathloss exponent 4
SBS Pathloss exponent 3.6
Network deployment 13 Mcells and 29 Scells
Number of users 200, 500 and 1000
Stationary user distribution Uniform distribution

5 Results

In this section we will present the results produced from our two mechanisms.
In both cases, we consider a basic network where users are static. For our first
scenario, we used a dataset containing the association results from the network
model based on the SINR metric, to train a ML model, based on the “Decision
Trees” technique.
From the results presented on Fig. 2, we see that the model can be quite
precise on the produced predictions. The dataset used is split on two parts, a
training set, used to train the model and a test set, used to test its capabilities.
Considering a fixed test dataset that is 20% of the size of the original dataset
produced from the network model, we test our precision value. Starting from
200 users, we get a precision value of 0.625 for the UL direction and a value
of 0.85 for the DL direction. When we increase the number of network users,
our produced dataset size increases as well. With a dataset of 500 users, we see
an improvement on the produced accuracy that now reaches 0.73 for the UL
direction and 0.92 for the DL direction. Finally with a dataset consisting of the
results on the allocation of 1000 users we see the accuracy rise to 0.84 for the
UL direction and to 0.935 for the DL direction.
These results, show that using a ML technique can be proven extremely
efficient in predicting user allocation on the network. Allocating users on BSs
can be quite a complicated procedure, since it has to take into account multiple
parameters for the association. As we can see from the produced results, the
predictions for the DL direction are far superior to the predictions on the UL
 User Allocation in 5G Networks Using Machine Learning Methods 21

Fig. 2. Prediction accuracy.

direction. This is to be expected, since most users for the DL direction are
allocated to McBSs. That means that in the produced dataset, most BSs are
McBSs, making it easier for the model to predict the associated BS for each
user, since it has to choose from a limited pool of BSs.
It is important though to mention that the size of our training dataset (and
the test dataset) massively impacts performance. For complicated models like
these, it is important that we fit them with a respectable size training dataset, so
that the model can pinpoint all connections between the parameters and produce
accurate results. The size of the training set should be appropriate, according to
our needs, in order to avoid cases of overfitting and underfitting.
The allocation techniques that exist are quite extensive and very accurate
on selecting the BS that best matches each user. This means that they can
be used as a pretty good source for creating accurate datasets to use for ML
models training. Especially in our case, the produced model can predict with
high accuracy the best allocation for each user. In real world scenarios, this
can lead to a massive improvement in real time performance for the network.
Classic user allocation models take into account multiple metrics and network
parameters and as a result they suffer from high complexity. Utilizing ML based
models that are trained on datasets produced from classic allocation models,
results in better real time decision making since a simpler model features much
better time complexity, that can produce reliable results even in cases with
limited network processing resources.
For our second scenario, the produced user clusters can be seen on Fig. 3,
along with their respective cluster centers. The produced cluster centers are
depicted as black bullets, while users that consist a cluster are depicted as bullets
22 C. Bouras and R. Kalogeropoulos

that share the same color. In our network we have defined 42 BSs and we also
produce 42 user clusters. The position of the McBSs cannot be changed, while
the positioning of the SBSs is not definite.

Fig. 3. Produced user clusters

Using the cluster centers as the new coordinates for the SBSs we ran our simu-
lations. We can conclude that our proposal results in smaller distances between
cluster centers and the closest available SBS. Smaller distances, result in less
pathloss and signal deterioration. This means that our users will be able to enjoy
higher DRs. Traditionally, UL direction demands are smaller than DL demands
and are satisfied from SBSs. In our simulations, cluster centers that are close to
McBSs, result in the same distance from McBSs in both results since McBSs are
stationary.
Our proposal yields impressive results for the UL direction. As seen on Fig. 4,
after improving the positioning scheme of SBSs, the number of associated users
in the UL direction remains high and is usually better than the results with a
random positioning of SBSs. This is subject to small changes, considering we
can never be certain about user spawn points, and their RB needs. Since most of
the associations in the DL direction are with McBSs (which remain stationary
across both simulations), DL association results are the same on both cases. The
difference varies with the UL direction.
All these users are matched with the BS that offers them the best SINR,
meaning that they enjoy the best available QoS. Our results show that our sug-
gestion can improve the network performance, but they ensure that our proposal
remains a trustworthy method for pinpointing the best locations for placing new
SBSs if we want to enrich our network’s infrastructures with new SBSs. That will
result in a higher offered QoS for the UL direction as well as the DL direction
while the improved number of associated users results in improved total network
throughput.
 User Allocation in 5G Networks Using Machine Learning Methods 23

Fig. 4. Associated users with best matched BS

Considering that both pathloss and SINR are indeed heavily dependent on
distance, users will see increased DRs that suffer less reduction (due to the lower
pathloss). The smaller distance also means that the network resources will be
better allocated. Users will enjoy better DRs, but will need less RBs from their
matching BSs. That means that less RBs will be consumed by a single user,
while more users will be satisfied using the same resources. As result, with our
proposal we will see a massive improvement on the achieved channel throughput.
Our two proposals show that using ML we can begin to see important benefits
on both user allocation as well as resource management. Our first proposal can
allocate users on their best matching BSs, using minimal network resources,
ensuring better real time decision making for the network. With our second
proposal, we can predict the optimal positioning for SBSs. This supplements
our first proposal, ensuring that more users are satisfied (especially in the UL
direction) with higher DRs. This improves network performance as well as the
perceived QoS, since now more users can utilize the network’s resources.

6 Conclusion and Future Work

With the rise of IoT, we expect networks to face an unprecedented rise in the
number of UEs that seek to use their resources. This is a crucial issue since
it dictates a fruitful utilization of the existent infrastructure and the avail-
able resources. With the number of UEs expected, real time performance on
networks can also be affected since massive calculations should be completed
for optimal user allocation. With our two proposals, we proved that applying
ML techniques can help with real time performance by predicting the optimal
association between users and BSs, but it can also be used to predict the best
24 C. Bouras and R. Kalogeropoulos

positioning system for SBSs. This will significantly improve the existing net-
work performance but can also complement the results produced from our first
proposal.
Many ML techniques have been introduced to various scientific sectors and its
introduction to networking is deemed necessary. This introduces an opportunity
for massively improving existing mechanisms. In the future prediction mecha-
nisms can be introduced to resource management (like frequency allocation),
while clustering can also be used to improve the network scalability by finding
an optimal system for positioning new BSs to reinforce current infrastructures.

References
1. Yang, J., Wang, C., Wang, X., Shen, C.: A Machine Learning Approach to User
Association in Enterprise Small Cell Networks, pp. 850–854 (2018). https://doi.
org/10.1109/ICCChina.2018.8641148
2. Mahbub, M.: A smart farming concept based on smart embedded electronics,
internet of things and wireless sensor network. Internet of Things 9, 1–30 (2020).
https://doi.org/10.1016/j.iot.2020.100161
3. Sun, Y., Peng, M., Zhou, Y., Huang, Y., Mao, S.: Application of machine learning in
wireless networks: key techniques and open issues. IEEE Commun. Surv. Tutorials.
21(4), 3072–3108 (2019). https://doi.org/10.1109/COMST.2019.2924243
4. Mom, J.M., Ani, C.: Application of self-organizing map to intelligent analysis of
cellular networks. ARPN J. Eng. Appl. Sci. 8, 407–412 (2013)
5. Rokach, L., Maimon, O.: Data mining with decision trees. Theory and Applications
(2008). https://doi.org/10.1142/9789812771728-0001
6. Shi, M., Yang, K., Xing, C., Fan, R.: Decoupled heterogeneous networks with mil-
limeter wave small cells. IEEE Trans. Wireless Commun. 17(9), 5871–5884 (2017).
https://doi.org/10.1109/TWC.2018.2850897
7. Elshaer, H., Boccardi, F., Dohler, M., Irmer, R.: Downlink and uplink decoupling:
a disruptive architectural design for 5G networks (2014). https://doi.org/10.1109/
GLOCOM.2014.7037069
8. Feng, Z., Li, W., Chen, W.: Downlink and uplink splitting user association in
two-tier heterogeneous cellular networks. In: 2014 IEEE Global Communications
Conference (GLOBECOM 2014), pp. 4659–4664 (2015). https://doi.org/10.1109/
GLOCOM.2014.7037543
9. Ghosh, A., Raha, A., Mukherjee, A.: Energy-efficient IoT-health monitoring system
using approximate computing. Internet of Things. 9, 100166 (2020). https://doi.
org/10.1016/j.iot.2020.100166
10. Sun, S., Adachi, K., Tan, P.H., Zhou, Y., Joung, J., Ho, C.K.: Heterogeneous
network: an evolutionary path to 5G, pp. 174–178 (2015) https://doi.org/10.1109/
APCC.2015.7412506
11. Cheung, Y.: K*-means: a new generalized k-means clustering algorithm.
Pattern Recogn. Lett. 24, 2883–2893 (2003). https://doi.org/10.1016/S0167-
8655(03)00146-6
12. Boostanimehr, H., Bhargava, V.: Unified and distributed QoS-driven cell associ-
ation algorithms in heterogeneous networks. In: IEEE Transactions on Wireless
Communications. vol. 14 (2014). https://doi.org/10.1109/TWC.2014.2371465
 An Efficient Cross-Layer Design for Multi-hop
Broadcast of Emergency Warning Messages
in Vehicular Networks

Abir Rebei1(B) , Fouzi Boukhalfa2(B) , Haifa Touati3(B) , Mohamed Hadded1(B) ,

and Paul Muhlethaler4(B)
1IResCoMath, Gabes, Tunisia
[email protected]
2 VEDECOM, Versailles, France

[email protected]
3 IResCoMath, Gabes, Tunisia

[email protected]
4 Inria, Rennes, France

[email protected]

Abstract. The main objective of Vehicular ad hoc networks (VANETs) is to

make road transportation systems more intelligent in order to anticipate and avoid
dangerous, potentially life-threatening situations. Due to its promising safety
applications, this type of network has attracted a lot of attention in the research
community. The dissemination of warning messages, such as DENMs (Decen-
tralized Environmental Notification Messages), requirse an efficient and robust
routing protocol. In previous studies, the active signaling mechanism has shown
its ability to prevent collisions between users trying to allocate the same resource.
In this paper, we propose an original message forwarding strategy based on the
active signaling mechanism. Our proposal disseminates warning messages from
a source vehicle to the rest of the network while minimizing the access delay
and the number of relay nodes. For this purpose, a special time slot is dedicated
to forwarding emergency warning messages. To avoid access collisions on this
slot, the active signaling scheme we propose favours the selection of the furthest
node as the forwarder. We carry out a number of simulations and comparisons to
evaluate the performances of the scheme.

Keywords: VANETs · Broadcast protocol · Cross-layer design · TDMA ·

Active signaling · Multi-hop communication · DENM · Network simulation

1 Introduction and Motivation

One of the most prominent technologies in the field of intelligent transportation systems
(ITS) that has drawn significant interest from researchers and industry is Vehicular Ad-
hoc Networks (VANETs). VANETS are essentially an extension of the unerlying prin-
ciples of Mobile Ad-hoc Networks (MANETs), while featuring a number of alterations.
VANETs provide a wide range of applications covering both safety and non-safety
applications. These applications aim at improving road safety traffic and assisting
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 25–38, 2021.
https://doi.org/10.1007/978-3-030-75100-5_3
26 A. Rebei et al.

drivers. For example, they may warn of dangerous situations and accidents that may
occur on the roads and provide internet access and infotainment services. To provide
the transfer of information between a group of vehicles within range of each other or
between vehicles and the infrastructure, VANETs allow for Vehicle-to-Vehicle (V2V)
and Vehicle-to-Infrastructure (V2I) communications [2, 3].
Safety applications impose stringent requirements in terms of Quality of Service
(QoS), given the need to achieve bounded delays and reliable communications. Guar-
anteeing high QoS is an major challenge in VANETs due to the high mobility of nodes,
frequent changes in network topology and lack of a central control. In order to satisfy
these requirements, therefore, it is necessary to have a QoS model provisioning. More-
over, different layers should cooperate to make correct routing decisions. An efficient
Medium Access Control (MAC) is required to provide a broadcast service that respects
the necessary requirements. Contention-free MAC protocols based on the Time Divi-
sion Multiple Access (TDMA) technique provide considerable benefits, such as col-
lision prevention [4]. Since central coordination is absent in a VANET topology, the
propagation of a safety messages to the set of vehicles must be relayed through multi-
ple relay vehicles. Furthermore, building a new multi-hop broadcast scheme of emer-
gency messages seems difficult due to the very nature of VANETs. Many MAC-aware
routing and broadcasting protocols have been designed in order to provide multi-hop
communications and disseminate safety messages in a timely manner. In this paper, we
present a cross-layer design for multi-hop broadcast of emergency warning messages
called AS-DTMAC Multihop. Our approach mixes MAC and routing layers together
for an efficient dissemination by selecting the best relay based on the AS-DTMAC [5, 6]
protocol, which uses an active signaling technique to eliminate the problem of access
collisions during the slot assignment process. In AS-DTMAC Multihop, we adjusted
the active signaling process to let the furthest vehicle acquire the dedicated forwarding
slot and thus to speed up the dissemination of warning messages.
The remainder of this paper is organized in 4 sections. The next section discusses
relevant related work. In Sect. 3, we present our message broadcast cross-layer solution
by presenting first the proposed MAC strategy and then the forwarding strategy. In
Sect. 4, we evaluate the performance of our protocol by presenting simulations results.
Finally, in Sect. 5, we conclude the paper and outline future work.

2 Related Work
Recently, several MAC-aware routing protocols have been proposed in the literature to
efficiently support multi-hop communication and disseminate safety messages in vehic-
ular networks in a timely manner. These proposals make use of different parameters
from the MAC layer, like transmission time slot allocation, channel state, and colli-
sion probability, to improve data dissemination in VANETs. A detailed review of these
cross-layer routing approaches for VANET is given in [25]. In this section, we focus on
solutions dedicated to multi-hop emergency messages broadcast.
One of the earliest approaches to disseminating broadcast packets is called OB-
VAN [22]. It is an opportunistic routing protocol that uses a modified 802.11 MAC
layer. OB-VAN uses an acknowledgement scheme to choose relay nodes. Choosing the
 An Efficient Cross-Layer Design for Multi-hop Broadcast of EmergencyWarning 27

best relay node is performed by using an active signaling technique. Nodes that have
captured the packet, transmit a short acknowledgement made up of signaling bursts,
calculated based on the distance criterion just after receiving the packet. This scheme is
a generalized CSMA/CA where the backoff technique is replaced by the active signal-
ing technique. To prohibit interference on signaling bursts, OB-VAN uses the CDMA
spreading code. Signaling bursts can be presented by 0 or 1. 0 denotes a listening inter-
val and 1 denotes a transmission interval. This binary sequence is composed of two
parts. The first part, is dedicated to optimizing the criterion for the best relay selection
while the second is used to discriminate between nodes and permit the winner to relay
the data packet.
Another TDMA-based routing protocol designed for warning message dissemina-
tion on bi-directional highways is proposed in [23]. This approach, called Priority-
Based Direction-Aware Media Access Control (PDMAC), classifies nodes as either
cluster heads (CHs) or ordinary vehicles (OVs).
To disseminate warning messages, PDMAC develops a three-tier priority assign-
ment process. The first tier is Direction-Based Relay Selection. A source disseminates
to its neighbors a request message (REQ) that indicates its direction, destination, etc.
and reserves all available time slots in this frame for itself. Neighbors respond with
an acknowledgment message (ACK) that contains all free time slots and the slot to be
assigned for the transmission of the message according to the severity level of the mes-
sage. The node selected as the best relay is the one that is closest in distance to the
destination and is moving in the direction towards it. The second tier is the Priority on
the Basis of Message Type. PDMAC prioritizes warning messages over non warning
messages by adding a bit in the message header to indicate the type of the message.
Finally, the third tier is Priority on the Basis of Severity Levels to differentiate between
different warning messages depending on their severity levels by computing the colli-
sion probability. In this case, warning messages are classified into 3 levels. In the case
of a lowest priority message, the sender should wait for a free time slot to send. If it is a
second level priority, it requests the release of a slot of another non-warning or warning
message with lower priority. Otherwise, in the case of a highest priority level message,
it is mandatory to release on the time slot of a non-warning or a lower-priority message.
A recent protocol called Multi-Channel Token Ring Protocol (MCTRP) is pre-
sented in [24]. MCTRP employs the multi-channel structure defined in IEEE 802.11p.
The network is composed of multiple virtual rings. Nodes are classified into 5 types:
Ring Founder Node (RFN), Token Holder Node (THN), Ring Member Node (RMN),
Dissociative Node (DN), and Semi-Dissociative Node (SDN). There are 2 types of
radio: Radio-I and Radio-II. A DN uses only Radio-I since it does not belong to any
ring, but the other nodes use both of them. Also, the time system is partitioned into a
control period and a data period.
The MCTRP protocol follows 3 sub-protocols. The first sub-protocol is the Ring
Coordination Protocol, which manages rings and nodes and schedules Service CHan-
nels (SCH) for each ring. First, the Ring Initialization Process consists of sending a
Ring Founding Message (RFM) that includes a selected SCH number for intra-ring
data communications and waiting for an invitation. After establishing a ring, a Joining
Invitation Message (JIM), which includes some information such as the SCH num-
ber, the speed, etc., will be broadcasted by the RFN to the DNs. The DN will reply to
28 A. Rebei et al.

the RFN with a Joining Acknowledgement Message (JAM) if the difference between
its moving speed and that of the RFN is smaller than a predefined speed threshold.
Other messages will be exchanged between RFN, DN and RMN such as Connection
Notification Messages (CNMs), Connecting Successor Messages (CSMs), etc. using
the contention-based CSMA/CA scheme. The second sub-protocol is the Emergency
message exchange protocol. To efficiently deliver emergency messages, MCTRP uses
Radio-I or Radio-II, depending on the case. This can be done through 4 steps. Firstly,
when an RMN detects an accident, it sends an emergency message to its RFN by adopt-
ing CSMA/CA and using Radio-II. Secondly, the RFN node replies with an acknowl-
edgement to the RMN, and then broadcasts the emergency message to all its RMNs
using Radio-II. Thirdly, it also broadcasts the message to its neighboring DNs, SDNs,
RFNs using Radio-I. Finally, neighboring RFNs rebroadcast the emergency message
again to their RMNs using Radio-II.
The third sub-protocol is the Data Exchange Protocol. Two types of data commu-
nications exist: inter-ring data communications where packets are transmitted using
CSMA/CA and intra-ring data communications where data packets are transmitted
using a token based mechanism. After receiving a token, a node can transmit data during
a token holding time and then pass the token to its successor.

3 Cross-Layer Solution for Emergency Messages Multi-hop

Dissemination
Cross-Layer design is an emerging proposal to support flexible layer approaches in
VANETs [25]. As described above, the recent ongoing research has shown increased
interest in protocols that rely on interactions between different layers. In this paper,
we propose to exploit the relation between MAC and network layers, in an effort to
improve the performance of the Multi-hop Broadcast of DENM in Vehicular Networks.
We propose the idea of combining an approach based on a MAC layer protocol named
AS-DTMAC with a forwarding strategy. Figure 1 represents the general architecture of
our message broadcast cross-layer solution. In this section, we first focus on the MAC
protocol principle and then we describe the forwarding strategy adopted.

3.1 Mac Strategy

Time Division Multiple Access (TDMA) is a contention-free MAC1 protocol with

scheduled channel. It is mostly used for safety applications in order to satisfy real-
time constraints. Nevertheless, while using TDMA with distributed schemes, the access
collision2 problem can occur (see [2] for more details). Very many protocols are sus-
ceptible to this problem, including: ADHOC MAC [15], VeMAC [18], DTMAC [4],

1 According to the control scheme used to access the channel, MAC random access protocols
are categorized into: contention-based or contention-free [2].
2 The access collision problem is can occur when using distributed schemes. It happens when
two or more vehicles within the same two-hop neighborhood attempt to access the same avail-
able time slot [2].
 An Efficient Cross-Layer Design for Multi-hop Broadcast of EmergencyWarning 29

Fig. 1. AS-DTMAC Multihop Architecture

R-ALOHA [17]. To solve this problem, we have designed in [5, 8–10], an efficient
mechanism named active signaling and we apply it in the DTMAC’s random slot selec-
tion process.
As shown in Fig. 2, frame composition in AS-DTMAC Multihop is slightly different
from that of AS-DTMAC [5]. In AS-DTMAC Multihop, each frame is composed of 100
slots and it is divided into three sets denoted as S0, S1 and S2, corresponding to three
successive zones. AS-DTMAC Multihop defines two types of slots: Normal Slot and
Emergency Message Forwarding Slot or EMF Slot. Normal slots are used by each
vehicle to send data, as in AS-DTMAC. However, EMF Slots are special slots dedi-
cated, by AS-DTMAC Multihop, to forwarding emergency messages. In each frame,
three slots, namely the first slot of each set Si (i.e. Slot0 , Slot34 and Slot67 ), are defined
as EMF Slots. This choice can be explained by the fact that emergency messages are
time-sensitive, hence choosing the first slot in each set to forward them will speed up
warning message dissemination.

Fig. 2. AS-DTMAC Multihop Frame

30 A. Rebei et al.

As illustrated in Fig. 3, whether normal or EMF, a slot is usually formed by two

time intervals. The first one, is dedicated to the selection process. However, the sec-
ond is held by the winner of the competition to transmit its payload packet. In the first
interval, a random binary key is generated by every node to compete for a slot. Hope-
fully, at the end of the time interval, only one node will remain as the winner. The
key is a succession of bits (0,1). ‘1’ means that the vehicle with a packet to send will
transmits during the signaling bursts. ‘0’ means that the vehicle with a packet to send
senses the channel during this mini-slot. When a vehicle selects a listening period and
senses a transmission, the competition to get the slot is over. For instance, a vehicle
that draws the key ‘01001110’ will listen during the first mini-slot and if no competing
transmission is sensed during this mini-slot, it will transmit during the next mini-slot.
The following two steps of the selection process in our example, will be two listening
periods. The selection process continues using the same rule until the key is completely
used up.

Fig. 3. Structure of the AS-DTMAC slot

3.2 Forwarding Strategy

Given the high sensitivity of emergency messages to delay constraints and collisions, it
is crucial to propose an efficient multi-hop dissemination scheme that avoids these prob-
lems. Since the active signaling technique solves the collision problem, our forwarding
strategy is based on an adapted version of the Active Signaling-DTMAC protocol.
As shown in Fig. 4, when a source node, SN , detects a warning event it has to
broadcast an emergency message. Vehicle SN sends the message during its reserved
Normal Slot. All nodes that receive the message and that are located behind the sender:
situated in the sender’s zone or in the adjacent one (in the sender’s transmission range),
will compete to forward the received emergency message to other vehicles during the
EMF Slot. To avoid access collision on this slot, each vehicle generates a binary key
based on the distance that separates it from the sender. The key is composed of mini-
slots. As explained in the previous section, these mini-slots take the value ‘1’ or ‘0’.
‘1’ means that the node is in a transmission mode and ‘0’ means that the node is in a
listening mode. The forwarding strategy consists of selecting the best next relay from
the list of vehicles that have generated keys. The winner node will forward the message
to the rest of the nodes situated in the opposite direction of the sender vehicle in order
to propagate the message as far as possible.
 An Efficient Cross-Layer Design for Multi-hop Broadcast of EmergencyWarning 31

In practice, the winner is always the furthest vehicle from the transmitter. This is
counted as a benefit in terms of packet propagation since the emergency message will
be quickly broadcasted and the danger will be avoided.
The transmitter sends a message in its own slot reserved in the set dedicated for the
zone to which it belongs. So, the forwarding will take place in the first slot of the next
set.
To give a clearer idea of our forwarding strategy, we consider the example shown
in Fig. 5. In this example, the sender, which is the red vehicle, sends an emergency
message during its slot (slot 8). In this case, the competition and the first relay of the
message will take place in the slot 34. The black vehicle (which is the furthest vehicle)
is the winner of the competition and it will forward the message to the other nodes. The
same process will be repeated in the next hop until all vehicles have been informed of
the potentially dangerous situation.

Fig. 4. Forwarding Strategy Flowchart.

Fig. 5. Forwarding Strategy

32 A. Rebei et al.

4 Performance Evaluation
• Latency: this defines the time between the first broadcast of the message and the
time of its reception by the last vehicle on the road (see Figs. 6 and 7).
• Packet Loss: as shown below in the equation, the packet loss percentage defines
the number of lost packets that are not received, divided by the total number of
packets that should be received. The formula to compute the packet loss is defined
as follows:
Losses Packets
Packet Loss(%) =
Total Number o f Packets
• Number of Forwarders: this metric defines the number of forwarders (relays)
needed to relay a message.
• Used Bandwidth: the used bandwidth metric represents the total number of packets
received by vehicles.

0.7
ASDTMAC-multihop
0.6 Flooding

0.5
Latency(s)

0.4

0.3

0.2

0.1

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
OA

Fig. 6. Latency versus channel occupancy for AS-DTMAC Multi-hop and Flooding with SSD =
20 and error bar (95% confidence interval).

Fig. 7. Latency versus channel occupancy for AS-DTMAC Multi-hop and Flooding with SSD =
30 and error bar (95% confidence interval).
 An Efficient Cross-Layer Design for Multi-hop Broadcast of EmergencyWarning 33

Now, We move on to the Packet Loss metric. Figure 8 presents the packet loss ver-
sus OA in percent. The error bars are for a 95% confidence interval. It is clear that AS-
DTMAC Multihop has 0% of packet losses, whereas in Flooding we find a considerable
packet loss rate that can reach more than 80% in high traffic level density conditions. As
we have explained above, in the Flooding mechanism, every vehicle that has received
the message will attempt to forward it and this will cause a high interference in the
Flooding scheme. As a result, many packets will be lost.

100

80
Packet Loss (%)

60

40

20
SSD α = 20: ASDTMAC Multihop
SSD α = 20: Flooding
SSD α = 30: ASDTMAC Multihop
SSD α = 30: Flooding
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
OA

Fig. 8. Packet Loss versus channel occupancy for AS-DTMAC Multi-hop and Flooding with error
bar (95% confidence interval).

We now evaluate the Number of Forwarders metric. In Figs. 9 and 10, we show the
number of forwarders needed for each approach (AS-DTMAC Multihop and flooding)
to warn of an emergency event versus OA. Figure 9 represents results with a SSD equal
to 20 (km/h) while Fig. 10 with a SSD equal to 30 (km/h). As we can see, AS-DTMAC
Multihop requires fewer resources, this will reduce the channel occupancy. In contrast,
the flooding technique uses more resources because all the vehicles that received the
packet will attempt to forward it.
We move on to the Used Bandwidth metric results. In Figs. 11 and 12, we plot the
used bandwidth versus OA for both AS-DTMAC Multihop and flooding. This metric,
represents the total number of received packets. In flooding, we notice that it provides a
high value compared to AS-DTMAC Multihop. In our approach, at every hop, only one
winner vehicle will relay the packet to its neighbors, whereas in flooding every receiver
will relay the packet and thus vehicles could receive the packet several times.
Finally, we compare the dissemination delay achieved by AS-DTMAC Multi-
hop and Flooding to the estimated delay. We begin by deriving an analytic expres-
sion of the estimated delay. As illustrated in Fig. 13, to deliver an emergency mes-
sage from V1 to V4 , one frame is sufficient. In fact, the message is relayed 3 times
(n0 + n1 + n2 = τ slots = one f rame). Based on this information, we derived an analyti-
cal formula to compute the estimated delay needed to deliver a message from a source
i to a vehicle j separated by such a distance. As defined in [7], the ED is defined in the
equation as following:
34 A. Rebei et al.

30
ASDTMAC-multihop
Flooding
25

Number of forwarders
20

15

10

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
OA

Fig. 9. Number of forwarders versus channel occupancy for AS-DTMAC Multi-hop and Flooding
with SSD = 20 and error bar (95% confidence interval).

30
ASDTMAC-multihop
Flooding
25
Number of forwarders

20

15

10

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
OA

Fig. 10. Number of forwarders versus channel occupancy for AS-DTMAC Multi-hop and Flood-
ing with SSD = 30 and error bar (95% confidence interval).

Disti, j
∗ τ ∗ sd
ED = (1)
3∗R
where i, j, R, τ , Disti, j and sd are respectively the sender, the receiver, the transmission
range, the length of the frame, the distance between the sender and the receiver and the
duration of the slot.
Figure 14 presents the estimated delay to propagate a packet from a source to a
receiver versus the distance between them.
We can observe from this figure that AS-DTMAC Multihop can provide a shorter
delay than the estimated one. This can be explained by the fact that the forwarder will
send in one of the reserved slots of forwarding (Slot0 or Slot34 or Slot67 ). The delay will
depend on the vehicle’s position (following the AS-DMAC Multihop scheme: vehicles
 An Efficient Cross-Layer Design for Multi-hop Broadcast of EmergencyWarning 35

700
ASDTMAC Multihop
600 Flooding

Used Bandwidth
500

400

300

200

100

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
OA

Fig. 11. Used Bandwidth versus channel occupancy for AS-DTMAC Multi-hop and Flooding
with SSD = 20 and error bar (95% confidence interval).

700
ASDTMAC Multihop
600 Flooding
Used Bandwidth

500

400

300

200

100

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
OA

Fig. 12. Used Bandwidth versus channel occupancy for AS-DTMAC Multi-hop and Flooding
with SSD = 30 and error bar (95% confidence interval).

Fig. 13. Message propagation based on TDMA slot information.

36 A. Rebei et al.

Fig. 14. Estimated Delay versus distance for AS-DTMAC Multi-hop and Flooding with SSD =
30 and error bar (95% confidence interval).

can access only the set of slot of their area): they can reserve at the beginning, at the
middle or at the end. However, due to the interference in the flooding approach, the
forwarding operation may not occur in the same frame. If this happens, a vehicle will
relay on the next frame. This hypothesis can explain the results obtained.

5 Conclusion

In this paper, we have proposed to enhance DTMAC protocol by integrating active

signaling. The simulation results show that AS-DTMAC drastically reduces the access
collision rate and allocates slots to all the vehicles in the network in half the time it takes
DTMAC to do so. We also presented a use case in the V2V for urgent and high priority
traffic message like DENM, that can help to avoid an accident, all these new features
are very important for the future technology described in the beginning of this paper.
As future work, we will do additional simulations to compare with the standard used in
V2V (IEEE 802.11p) and we plan to develop an analytical model for AS-DTMAC as
well as to investigate further advanced access features that could be provided using the
active signaling scheme.

References
1. Hu., F.: Security and Privacy in Internet of Things (IoTs): Models, Algorithms, and Imple-
mentations. ISBN 978-1-4987-2318-3
2. Hadded, M., Muhlethaler, P., Laouiti, A., Zagrouba, R., Saidane, L.A.: TDMA-based MAC
protocols for vehicular ad hoc networks a survey, qualitative analysis and open research
issues. IEEE Commun. Surv. Tutorials 17(4), 2461–2492 (2015)
3. Laouiti, A., Qayyum, A., Saad, M.-N.-M.: Vehicular Ad-Hoc networks for smart cities. In:
Proceedings of the Second International Workshop of Advances in Intelligent Systems and
Computing, vol. 548, Springer (2017)
 An Efficient Cross-Layer Design for Multi-hop Broadcast of EmergencyWarning 37

4. Hadded, M., Laouiti, A., Muhlethaler, P., Saidane, L.A.: An infrastructure-free slot assign-
ment algorithm for reliable broadcast of periodic messages in vehicular Ad hoc networks.
In: Vehicular Technology Conference VTC-Fall. Montreal, Canada (2016)
5. Boukhalfa, F., Hadded, M., Muhlethaler, P., Shagdar, O.: An active signaling mechanism to
reduce access collisions in a distributed TDMA based MAC protocol for vehicular networks.
In: Advanced Information Networking and Applications (AINA-2019), Kunibiki Messe,
Matsue, Japan (2019)
6. Boukhalfa, F., Hadded, M., Muhlethaler, P., Shagdar, O.: An analytical model for perfor-
mance analysis of an active signaling-based MAC protocol for vehicular networks. In: Vehic-
ular Technology Conference VTC-FALL. Honolulu, Hawaii, USA (2019)
7. Hadded, M., Muhlethaler, P., Laouiti, A.: Performance evaluation of a TDMA-based multi-
hop communication scheme for reliable delivery of warning messages in vehicular networks.
In: Proceedings of Wireless Communication and Mobile Computing Conference (IWCMC),
pp. 1029–1034 (2017)
8. Boukhalfa, F., Adjih, C., Muhlethaler, P., Hadded, M., Shagdar, O.: Physical and MAC layer
design for active signaling schemes in vehicular networks. In: WiMob, pp. 86–92. Thessa-
loniki, Greece (2020)
9. Boukhalfa, F., Hadded, M., Muhlethaler, P., Shagdar, O.: Coexistence of IEEE 802.11p and
the TDMA-based AS-DTMAC Protocol. In: SoftCOM, pp. 1–6. Split, Croatia (2020)
10. Boukhalfa, F., Hadded, M., Muhlethaler, P., Shagdar, O.: Performance evaluation of an active
signaling based time-slot scheduling scheme for connected vehicles. In: Annals of Telecom-
munications, pp. 1–12 (2020)
11. Abate, J., Whitt, W.: Numerical inversion of Laplace transforms of probability distributions.
ORSA J. Comput. 7(1), 38–43 (1995)
12. Ye, F., Yim, R., Zhang, J., Roy, S.: Congestion control to achieve optimal broadcast effi-
ciency in vanets. In: IEEE International Conference on Communications (ICC), pp. 1–5.
South Africa, May, Cape Town (2010)
13. 802.11p-2010, IEEE standard for information technology - Telecommunications and infor-
mation exchange between systems - local and metropolitan area networks - specific require-
ments part 11 : Wireless LAN medium access control (MAC) and physical layer (PHY) and
physical layer (PHY) specifications amendment 6 : Wireless access in vehicular environ-
ments std (2010)
14. Uzcategui, R., Acosta-Marum, G.: Wave: a tutorial. IEEE Commun. Mag. 47(5), 126–133
(2009)
15. Borgonovo, F., Capone, A., Cesana, M., Fratta, L.: Adhoc MAC: new MAC architecture
for ad hoc networks providing efficient and reliable point-to-point and broadcast services.
Wireless Netw. 10(4), 359–366 (2004)
16. Naik, G., Choudhury, B., Park, J.M.: IEEE 802.11 BD & 5G NR V2X: evolution of radio
access technologies for V2X communications. IEEE Access J. 7, 70169–70184 (2019)
17. Borgonovo, F., Capone, A., Cesana, M., Fratta, L.: RR-ALOHA, a reliable R-ALOHA
broadcast channel for Ad-Hoc intervehicle communication networks. In: IEEE IFIP Annual
Mediterranean Ad Hoc Networking Workshop (Med-Hoc-Net). Baia Chia, Italy (2002)
18. Zhuang, W., Omar, H.A., Lio, L.: VeMAC: A novel multichannel MAC protocol for vehicu-
lar ad hoc networks. In: IEEE Conference on Computer Communications Workshops (INFO-
COM WKSHPS), pp. 413–418. China, Aug, Shanghai (2011)
19. Weidong, Y., Pan, L., Yan, L., Hongsong, Z.: Adaptive TDMA slot assignment protocol for
vehicular ad-hoc networks. J. China Univ. Posts Telecommun. 20(1), 11–18 (2013)
20. Ke, W., Weidong, Y., Pan, L., Hongsong, Z.: A decentralized adaptive TDMA schedul-
ing strategy for VANETs. In: IEEE Wireless Communications and Networking Conference
Workshops (WCNCW), pp. 216–221. China, Apr, Shanghai (2013)
38 A. Rebei et al.

21. Karnadi, F., Mo, Z., Lan, K.C.: Rapid generation of realistic mobility models for VANET.
In: IEEE WCNC, pp. 2506–2511. Hong Kong, China, Mar (2007)
22. Blaszczyszyn, B., Laouiti, A., Muhlethaler, P., Toor, Y.: Opportunistic broadcast in VANETs
(OB-VAN) using active signaling for relays selection. In: IEEE International Conference on
ITS Telecommunications (ITST 2008), pp. 384–389, Oct (2008)
23. Abbas, G., Abbas, Z.H., Haider, S., Baker, T., Boudjit, S., Muhammad, F.: PDMAC: A
priority-based enhanced TDMA protocol for warning message dissemination in VANETs.
Sensors 20, 45 (2020)
24. Bi, Y., Liu, K.H., Cai, L.X., Shen, X.: A multi-channel token ring protocol for QoS provi-
sioning in inter-vehicle communications. IEEE Trans. Wireless Commun. 8(11), 5621–5631
(2009)
25. Rebei, A., Hadded, M., Touati, H., Boukhalfa, F., Muhlethaler, P.: MAC-aware routing pro-
tocols for vehicular ad hoc networks: a survey. In: 2020 International Conference on Soft-
ware, Telecommunications and Computer Networks (SoftCOM), pp. 1–6. Split, Hvar, Croa-
tia (2020). https://doi.org/10.23919/SoftCOM50211.2020.9238249
 Time Series Forecasting for the Number
of Firefighters Interventions

Roxane Elias Mallouhy1(B) , Christophe Guyeux2 , Chady Abou Jaoude3 ,

and Abdallah Makhoul2
1 Prince Mohammad Bin Fahd University, Khobar, Kingdom of Saudi Arabia
[email protected]
2 University of Bourgogne Franche-Comté, Belfort, France

{christophe.guyeux,abdallah.makhoul}@univ-fcomte.fr
3 Antonine University, Baabda, Lebanon

[email protected]

Abstract. Time series forecasting is one of the most attractive analysis of dataset
that involves a time component to extract meaningful results in economy, biology,
meteorology, civil protection services, retail, etc. This paper aims to study three
different time series forecasting algorithms and compare them to other models
applied in previous researchers’ work as well as an application of Prophet tool
launched by Facebook. This work relies on an hourly real dataset of firefighters’
interventions registered from 2006 till 2017 in the region of Doubs-France by the
fire and rescue department. Each algorithm is explained with best fit parameters,
statistical features are calculated and then compared between applied models on
the same dataset.

1 Introduction
Many studies show that reaching good forecasts is vital in many activities such as indus-
tries, commerce, economy, and science. The fact of gathering a collection of observa-
tions over time will provide predictions of new observations in the future and extract
meaningful characteristics of the data and statistics in different time intervals: hours,
days, weeks, months, and years. The usage of data science, machine learning, and time
series forecasting are feasible in the prediction of firefighters’ interventions since it is
logical to assume that firefighters’ interventions are affected somehow by temporal,
climatic, and other events such as new year’s eve, snowing weather, traffic peak time,
fires in summer, holidays, etc. Due to the French economic crisis (closure of small
hospitals, population growth, etc.), the impact of optimizing the number of human
interventions leads directly to a reduction and better control in the financial, human,
and material resources. The goal is to size the number of firefighters according to the
need and demand: a greater number of firefighters should be available when they are
mostly used. Indeed, the number of guards available should be related to the location,
number, and type of the intervention. For example, during the weekend when acci-
dents indeed increase, the number of firefighters ready to serve the society should be
greater than a regular working weekday where most of the people reside in their offices.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 39–50, 2021.
https://doi.org/10.1007/978-3-030-75100-5_4
40 R. Elias Mallouhy et al.

Fig. 1. Number of firefighter interventions from 2006 until 2017.

A trusted result and high prediction are affected by many factors: the algorithm used
to train and test the dataset plays a big role as well as the chosen parameters. In this
work, three different well-known time series algorithms have been implemented: Auto
Regression (AR), Moving Average (MA), and Autoregressive Integrated Moving Aver-
age (ARIMA). Each algorithm is explained and detailed in Sect. 3. The dataset used
carries information about firefighters’ interventions in the region of Doubs-France from
01/01/2006 00:00:00” until 31/12/2017 23:00:00” . All the data were registered by the
fire and rescue department SDIS 25 by blocks of one hour [23]. An overview of the
number of firefighters interventions through the years is shown in Fig. 1.
This paper concentrates on statistical parameters calculated of three different
Machine Learning algorithms to predict the number of firefighters’ interventions,
prophet, and comparison between the applied algorithms and related work using the
same dataset.The remaining of this paper is structured as follows: literature review pro-
viding an overview about related work done by researchers on the same topic; Machine
Learning algorithms for predicting firefighters’ interventions explaining the three ML
models used; parameters chosen showing the values used for the corresponding algo-
rithms; building data with Prophet by applying this Facebook tool on the firefighters
dataset; obtained results; results interpretation, and conclusion.

2 Literature Review
Many influential works on time series forecasting have been published in the latest years
with enormous progress in this field. The idea of forecasting starts with Gardner and
Snyder who boost two subsequent papers in the same year (1985) in the area of time
series forecasting about exponential smoothing methods. Gardner provided a review
of all existing work done to that date and extended his research to include a damped
trend [2] After Gardner’s work, Snyder demonstrated that simple exponential smooth-
ing could arise from state space model innovation [3]. Most of the researches since 1985
has focused on empirical properties [10], forecasts evaluations [9], and proposition of
new methods for initialization and estimation [8]. Many studies have stimulated the use
of exponential smoothing methods in various areas such as air passengers [16], com-
puter components [5], and production planning [12]. Later on, numerous variations on
 Time Series Forecasting for the Number of Firefighters Interventions 41

original methods have been proposed to deal with continuities, constraints, and renor-
malization at each period time [13, 15]. Multivariate simple exponential smoothing was
used for processing the control charts by introducing a moving average technique [21].
Moreover, Taylor [7] and Hyndman et al. [20] have extended basic methods and have
included all 15 different exponential methods. Moreover, they have proposed models
that correspond to multiplicative error cases and additive errors. However, these meth-
ods were not unique since it has been known that ARIMA models give equivalent results
in forecasting, but the innovation in their work was that statistical models can lead to
non-linear exponential smoothing methods.
Early studies of time series forecasting in the nineteenth century were globally
based on the idea that every single time series can be seen as a realization of a stochas-
tic process. Based on this simple proposal, many time series methods since then have
been developed. Workers such as Walker et al. [22] formulated the concept of moving
average MA model and autoregressive AR models. The conception of linear forecasting
happened by the decomposition theorem. After that, many studies have appeared deal-
ing with parameter identification, forecasting estimation, and model checking. Time
series forecasting in this paper is specifically studied by running a real dataset of fire-
fighters’ interventions. This area of research in machine learning is new some-how
where only a few articles are targeting this topic. A work achieved by C. Guyeux et
al. [24] started by collecting a list of interventions and preparing the set for learn-
ing, validating, and testing. Then, by using supercomputers, the learning was carried
out on an ad hoc Multi-Layer Perception. The study ends up by applying the neural
architecture on a real case study with mature and encouraging results. Another work
done by Couchot et al. [23] has shown that a machine learning tool can provide accu-
rate results by deploying a learning process based on real and anonymized data using
extreme gradient boosting to guess an accurate behavior. Ñahuis et al. [25, 27], using the
same firefighter’s dataset demonstrated that machine learning is mature enough to make
feasible predictions for critical events such as natural disasters. They used LSTM and
XGBoost approaches to predict the number of firefighter’s interventions. These inves-
tigations have been deepened following a feature-based machine learning approach in
various directions [24, 26, 28], but never by considering the time series alone.

3 Machine Learning Algorithms for Predicting Firefighters’

Interventions

Introductory books on time-series algorithms and analysis include Davis and Brock-
well [19], Diggle [4], Swift and Janacek [11], Wei et al. [6].

3.1 Auto-Regressive Model

Auto regression is a statistical time series model that predicts an output for the near-
future (number of houses sold, price of something, number of interventions, . . . ) based
on past values. It was originated in 1920 by Udny Yule, Eugen Slutsky, and Ragnar
Frisch [14]. For instance, to predict today’s value based on yesterday, last week, last
42 R. Elias Mallouhy et al.

month, or last year one. AR models are also called Markov models, conditional mod-
els, or transition ones. Regression uses external factors which are independent as an
explanatory variable for the dependent values. Autoregression model is conditioned by
the product of certain lagged variables and coefficients allowing inference to be made.
In reality, AR works hardly if the future predictors are unknown because it requires a
set of predictor variables. On the other hand, AR is capable of adjusting the regression
coefficient β and violating the assumption of uncorrelated error since the independent
observations are time lagged values for the dependent observations.
In an AR model, the value of the predicted outcome variable (Y) at some time t is
Yt = β0 + β1Yt−1 + εt where the parameters β0 + β1Yt−1 rely on the past and εt which is
the white noise could not be predictable from the past. It is important to mention that
knowing the previous lagged values of Yt−2 ,Yt−3 does not affect the prediction of Yt
because as shown in the formula, Yt is affected only by Yt−1 .

3.2 Moving Average Model

Moving Average MA is a model introduced in 1921 by Hooker who considers multi-

ple period averages to predict future output and event [1]. It is an effective and naive
technique in time series forecasting used for data prediction, data preparation or fea-
ture engineering. It uses the most recent historical data values to generate a forecast.
MA removed the fine-grained variation between time steps, to expose the signal. This
method uses the average data periods’ number. The term “moving” indicates the up
and down moves of the time series done to calculate the average of a fixed number
of observations. On the other hand, the process of averaging relies on the overlapping
observations that create averages. Moving Average method can be used for both lin-
ear and non-linear trends. However, it is not applicable for short time series forecasting
fluctuation because the trend obtained by applying the model is neither a standard curve
nor a straight line. Besides, trend values are not available for some intervals at the start
and the end values of time series.
The outcome value in the MA(q) model, a moving average model of order q, is
presented as the following:
yt = c + εt + θ1 εt−1 + θ2 εt−2 +...+θq εt−q where εt is the white noise. This technique
involves creating a new time series with compromised values of row observations and
average in the original data set time series. Moreover, it relies on past forecast errors.

3.3 AutoRegressive Integrated Moving Average

ARIMA, also called Box-Jenkins, is a model proposed by George Box and Gwilym
Jenkins in 1970 by using a mathematical approach to describe changes in the time
series forecasting [18]. ARIMA is an integration of auto regression and moving average
methods that use a dependent relationship between an observation and some number of
lagged observations by differencing between raw observations. It subtracts an observa-
tion from the previous time step and takes into consideration the residual error. ARIMA
is a powerful model as it takes into consideration history as an explanatory variable,
 Time Series Forecasting for the Number of Firefighters Interventions 43

but in such model, the data cost is usually high due to the large observations require-
ment needed to build it properly. A standard notation for ARIMA being used is ARIMA
(p,d,q) where:

• p: is the auto-regressive part of the model, which means the number of lag observa-
tions that are included in the model. It helps to incorporate the effect of past values
of the model. In other terms, it is logical to state that it is likely to need 5 firefighters
tomorrow if the number of interventions was 5 for the past 4 d. A stationary series
with autocorrelation can be corrected by adding enough AutoRegression terms.
• d: is the integrated part of the model. It shows the degree of differencing the number
of times that the raw observations have been differenced. This is similar to state that
if in the last 4 d the difference in the number of interventions has been very small, it
is likely to be the same tomorrow. The order of differencing required is the minimum
order needed to get a near-stationary series.
• q: order of moving average which is the size of the MA window. Autocorrelation
graph shows the error of the lagged forecast. The ACF shows the number of MA
terms required to remove autocorrelation in the stationaries series.

4 Parameters Chosen

Auto Regression algorithm does not need any parameters to be chosen or modified
as explained in the formula in the previous section. Nevertheless, different parameters
were tested and registered for Moving Average and ARIMA;

4.1 Window Size for Moving Average

After trying different values of window sizes for different hours and days, the best size
chosen is 3 h having the minimal values of MAE and RMSE. On the other hand, to select
the values of ARIMA, the following parameters should be taken into consideration:

1. p: the order of AR term was basically taken to be equal to the number of lags that
crosses the significance limit in the Autocorrelation Fig. 3a. It is observed that the
ACF lag 4 is quite significant. Then, p was fixed to 4.
2. d: let us use the Augmented Dickey Fuller (ADF) test to see if the number of inter-
ventions is stationary. The p value found is 5.12e−28 , which is lower than the signif-
icant level of 0.05. This means that no differencing is needed. Let’s fix d to 0.
3. q: one lag above the significance level was found, thus q = 1 (Fig. 3b).

Figure 2 shows the actual number of interventions versus the predicted ones after
applying a moving average transformation overlaid by 3 h (Table 1).
44 R. Elias Mallouhy et al.

Table 1. Different window sizes for Moving Average Algorithm.

Window Size MAE

1h 1.477
2h 1.360
3h 1.349
4h 1.359
5h 1.387
10 h 1.541

Fig. 2. Actual versus predicted number of interventions.

4.2 P, D and Q Parameters for ARIMA

After determining the values of p, d and q, ARIMA model is fitted by using order
(4,0,1).

(a) Autocorrelation plot (b) Partial autocorrelation plot

Fig. 3. p and q parameters.

5 Building Data with Prophet

A prophet forecasting model is an open-source algorithm designed by Facebook in
2017 [17] for time series having common features and intuitive parameters, where
 Time Series Forecasting for the Number of Firefighters Interventions 45

experts and non-experts in statistics and time series forecasting can use it. Prophet is
based on time series models and relies on four main components: (1) Yearly and weekly
seasonality, (2) Non linear trend, (3) Holidays and (4) Error. Prophet fits very well for
data that have at least one year of historical inputs with daily periodicity. It is very fast
in terms of fitting the model, working without converting data into time-series objects,
and being robust to missing values. In addition, prophet is simpler compared to other
time series forecasting algorithms, because it requires less number of parameters and
models. Prophet works as following: y(t) = g(t) + s(t) + h(t) + εt where:

• g(t): describes the increase or decrease trends in the long-term data.

• s(t): represents the impact of seasonal factors over the year on the time-series data.
• h(t): models how large events and holidays affect the data.
• εt : shows the non reducible error term.
For prophet’s preparation, a new dataframe should be found: a new column is added
to the data that emerges years, months, days, and hours. Then, this column is renamed to
‘ds’ and the predicted output presented in the data under the name of nbinterventions is
renamed as ‘y’. Figure 4 helps to visualize the forecast of the dataframe where the black
dots display actual measurements, the blue line indicates Prophet’s forecast and the light
blue shaded line shows uncertainty intervals. It looks like the number of interventions
was increasing over the years slightly (Table 2).

Table 2. New dataframe for the dataset.

Index y ds
0 1 2006-01-01 00:00:00
1 1 2006-01-01 01:00:00
2 0 2006-01-01 02:00:00
3 2 2006-01-01 03:00:00
4 4 2006-01-01 04:00:00

Fig. 4. Forecast for Prophet algorithm.

46 R. Elias Mallouhy et al.

6 Obtained Results
Considering having 24 h per day, 7 d per week, 30 d per month, and 365 d per year, three
algorithms have been implemented: AR, MA and ARIMA on the dataset in addition to
the prophet. Statistical features of firemen predictions for every year from 2006 until
2017 have been registered in Table 4 and graphs that gather all statistical features are
shown in Fig. 5. On the other hand, let us overview the result of the forecast by illus-
trating a breakdown of the former elements (Fig. 6) for daily, weekly and yearly trends
using the prophet tool (see Table 3). The number of interventions during a trimmed time
slot is shown in Fig. 7 where:

• yellow plot represents the actual number of interventions y.

• purple plot indicates the prediction yhat.
• blue and red plots show the upper and lower bound of prediction respectively.

Table 3. Statistical features using AutoRegression for different time slots.

Time MAE MSE RMSE

1h 0.307 0.094 0.307
2h 0.403 0.171 0.414
12 h 1.205 1.932 1.390
1 day 1.288 2.128 1.459
5d 1.168 1.822 1.350
1 week 1.209 2.057 1.434
1 month 1.213 2.123 1.457

(a) AR (b) MA (c) ARIMA

Fig. 5. Statistical features over many years using various models.

 Time Series Forecasting for the Number of Firefighters Interventions 47

Table 4. Statistical features using AR, MA , ARIMA and prophet from 2006 until 2017.

AR MA ARIMA Prophet
2006 1.481 2.046 1.349 1.86 1.018 1.307 2.55 3.53
2007 1.601 2.064 1.429 1.924 1.376 1.822 1.95 2.97
2008 1.496 1.952 1.385 1.868 1.263 1.644 1.31 1.63
2009 2.374 3.35 1.854 2.787 1.414 1.904 3.25 5.69
2010 2.161 3.058 1.847 2.716 1.629 2.154 2.00 2.23
2011 2.574 3.676 2.09 2.922 1.699 2.247 6.00 11.00
2012 1.99 2.5 1.84 2.415 1.642 2.031 2.44 2.87
2013 1.972 2.478 1.83 2.392 1.682 2.14 2.33 2.66
2014 2.04 2.545 1.874 2.451 1.504 1.843 2.44 2.90
2015 2.145 2.678 1.939 2.525 1.829 2.43 2.73 3.15
2016 2.223 3.137 2.026 2.807 1.898 2.31 2.45 2.99
2017 2.359 2.917 2.111 2.738 1.941 2.544 2.86 3.49
2018 2.552 3.216 2.24 2.929 2.075 2.742 1.85 2.60

(a) Yearly and daily prediction (b) Trend and weekly prediction

Fig. 6. Breakdown of the forecast using prophet.

Fig. 7. Number of firefighters interventions using prophet.

48 R. Elias Mallouhy et al.

7 Results Interpretation
Mean Absolute Error for AR, MA, ARIMA and prophet are compared in Fig. 8. As
shown in the previous section, the best algorithm in term of fewer errors in most cases
is ARIMA represented by the gray line. Among most of the years, starting from 2006
until 2018, this gray line has reached the minimum mean absolute error and root mean
squared error, comparing to autoregression, moving average algorithms and prophet
tool. Thus, the second, third, and fourth places are reserved respectively for Moving
Average, Autoregressive, and Prophet. Generally, since ARIMA(p,d,q) stands for Auto
Regressive Integrated Moving Average, it is logical to conclude that it combines AR
(parameter p) and MA (parameter q) models. Other than that, ARIMA ensures the
stationarity of the model (parameter d), unlike AR and MA. Therefore, by applying
the components of these two models together, the probability of making errors will be
reduced as shown in the experiments. It is important to mention that ARIMA is more
complex than applied algorithms since it requires more time to identify the excessive
number of parameters p,d and q. In contrast, when comparing eXtreme Gradient Boost-
ing (XGBoost) and Long Short-Term Memory (LSTM) algorithms applied and exper-
imented in [28] together with ARIMA from 2006 to 2014, it seems that ARIMA has
the lowest root mean squared error values. However, XGBoost has the minimal RMSE
values for 2015,2016 and 2017. This result reflects that XGBoost is better for long term
forecast usage.
On the other side, by analyzing the prophet result, it is very clear that the number of
interventions of firefighters increments highly during the weekend (Saturday and Sun-
day) and reaches the minimum during the middle of the week (Wednesday). This inter-
pretation corresponds to official days off in France. Also, the number of interventions
increments slightly starting the month of May and reaches the maximum in July, then
decreases gradually till November. The fluctuation of interventions per month reflects
that during summer incidents are more likely to happen than other seasons. On the other
hand, the daily seasonality illustrates that the number of interventions increases during
the morning starting around 5:00 am and reaches higher values between 11:00 am and
5:00 pm. It’s very logical to link this variation with the departure and return time from
work.

Fig. 8. Mean absotule error comparison for AR, MA and ARIMA.

 Time Series Forecasting for the Number of Firefighters Interventions 49

8 Conclusion and Future Work

In this paper, Autoregression, Moving Average and Auto Regressive Integrated Moving
Average has been implemented as well as a Facebook tool for time series forecasting
called Prophet. For each model, statistical parameters have been calculated and com-
pared between each other then compared between other results previously done. In
general, as many researchers agreed, no hypothesis or rule electing a better algorithm
among all-time series forecasting. The choice of the technique used depends on the
specific prediction problem taking into account trends, seasonality, variables, size of
the dataset, etc. In this paper, the statistical metrics indicate that ARIMA is the best
model comparing to AR and MA as it combines first, the characteristics of these two
algorithms and second the stationarity of the model. In contrast, XGBoost fits better
than ARIMA for long term prediction. An extension to this work would be to apply
different time series forecasting models to firefighters dataset.

Acknowledgement. This work has been supported by the EIPHI Graduate School (contract
ANR-17-EURE-0002) and is partially funded with support from the Hubert Curien CEDRE pro-
gramme n 46543ZD.

References
1. Hooker, R.H.: Forecasting the crops from the weather. Q. J. R. Meteor. Soc. 47(198), 75–100
(1921)
2. Gardner Jr., E.S.: Exponential smoothing: the state of the art. J. forecast. 4(1), 1–28 (1985)
3. Snyder, R.D.: Recursive estimation of dynamic linear models. J. R. Stat. Soc. Ser B (Method-
ological), 272–276 (1985)
4. Diggle, P.J.: Time series; a biostatistical introduction (1990) (No. 04; QA280, D5.)
5. Gardner Jr., E.S.: Forecasting the failure of component parts in computer systems: a case
study. Int. J. Forecast. 9(2), 245–253 (1993)
6. Chatfield, C.: Time-series forecasting. CRC Press, Boca Raton, USA (2000)
7. Taylor, J.W.: Exponential smoothing with a damped multiplicative trend. Int. J. Forecast.
19(4), 715–725 (2003)
8. Ledolter, J., Abraham, B.: Some comments on the initialization of exponential smoothing. J.
Forecast. 3(1), 79–84 (1984)
9. Sweet, A.L., Wilson, J.R.: Pitfalls in simulation-based evaluation of forecast monitoring
schemes. Int. J. Forecast. 4(4), 573–579 (1988)
10. Bartolomei, S.M., Sweet, A.L.: A note on a comparison of exponential smoothing methods
for forecasting seasonal series. Int. J. Forecast. 5(1), 111–116 (1989)
11. Janacek, G.J., Swift, L.: Time series: forecasting, simulation, applications. Ellis Horwood
Limited, Chichester, England (1993)
12. Miller, T., Liberatore, M.: Seasonal exponential smoothing with damped trends: an applica-
tion for production planning. Int. J. Forecast. 9(4), 509–515 (1993)
13. Rosas, A.L., Guerrero, V.M.: Restricted forecasts using exponential smoothing techniques.
Int. J. Forecast. 10(4), 515–527 (1994)
14. Klein, J.L., Klein, D.: Statistical visions in time: a history of time series analysis, pp. 1662–
1938. Cambridge University Press, Cambridge, UK (1997)
15. Williams, D.W., Miller, D.: Level-adjusted exponential smoothing for modeling planned dis-
continuities. Int. J. Forecast. 15(3), 273–289 (1999)
50 R. Elias Mallouhy et al.

16. Grubb, H., Mason, A.: Long lead-time forecasting of UK air passengers by Holt-Winters
methods with damped trend. Int. J. Forecast. 17(1), 71–82 (2001)
17. Taylor, S.J., Letham, B.: Forecasting at scale. Am. Stat. 72(1), 37–45 (2018)
18. Box, G.E., Jenkins, G.M. and Reinsel, G.: Time series analysis: forecasting and control
Holden-day San Francisco. BoxTime Series Analysis: Forecasting and Control Holden Day
(1970)
19. Brockwell, P.J., Davis, R.A., Calder, M.V.: Introduction to time series and forecasting, vol.
2, pp. 3118–3121. Springer, New York (2002)
20. Hyndman, R.J., Koehler, A.B., Snyder, R.D., Grose, S.: A state space framework for auto-
matic forecasting using exponential smoothing methods. Int. J. Forecast. 18(3), 439–454
(2002)
21. Gang, T.T., Yang, J., Zhao, Y.: Multivariate control chart based on the highest possibility
region. J. Appl. Stat. 40(8), 1673–1681 (2013)
22. Chen, S., Lan, X., Hu, Y., Liu, Q., Deng, Y.: The time series forecasting: from the aspect of
network (2014). arXiv preprint arXiv:1403.1713
23. Couchot, J.F., Guyeux, C., Royer, G.: Anonymously forecasting the number and nature of
firefighting operations. In: Proceedings of the 23rd International Database Applications &
Engineering Symposium, pp. 1–8 (June 2019)
24. Guyeux, C., et al.: Firemen prediction by using neural networks: a real case study. In: Pro-
ceedings of SAI Intelligent Systems Conference, pp. 541–552. Springer, Cham (September
2019)
25. Nahuis, S.L.C., Guyeux, C., Arcolezi, H.H., Couturier, R., Royer, G. and Lotufo, A.D.P.:
Long short-term memory for predicting firemen interventions. In: 2019 6th International
Conference on Control, Decision and Information Technologies (CoDIT), pp. 1132–1137.
IEEE (April 2019)
26. Arcolezi, H.H., Couchot, J.F., Cerna, S., Guyeux, C., Royer, G., Al Bouna, B., Xiao, X.:
Forecasting the Number of Firefighters Interventions per Region with Local-Differential-
Privacy-Based Data. Computers & Security, p.101888 (2020)
27. Cerna, S., Guyeux, C., Arcolezi, H.H., Couturier, R., Royer, G.: A comparison of LSTM and
XGBoost for predicting firemen interventions. In: World Conference on Information Systems
and Technologies, pp. 424-434. Springer, Cham (April 2020)
28. Cerna, S., Guyeux, C., Arcolezi, H.H., Royer, G.: Boosting Methods for Predicting Firemen
Interventions. In 2020 11th International Conference on Information and Communication
Systems (ICICS), pp. 001-006. IEEE (April 2020)
 Stock Prediction Based on Adaptive Gradient
Descent Deep Learning

Bo Li1(B) and Li-feng Li2

1 No. 32, Chengfang Street, Xicheng District, Beijing, China
[email protected]
2 20th Floor, Block A, No. 8 Building, No. 5 Anding Road, Beijing, China

Abstract. Historical data is a permanent collection that allows people to make

predictions about the future trends of stocks. With the development of machine
learning especially deep learning, the prediction accuracy raised up. In this paper,
we proposed feature extraction of stock data through CNN, and a stock predic-
tion algorithm based on deep learning combined with traditional LSTM is con-
structed. At the same time, the third-order moment is used to improve the tradi-
tional stochastic gradient descent, and obtain a dynamic one. Experiments show
that compared with other methods it has better accuracy which provides a solu-
tion for stock trend prediction by deep learning.

1 Introduction

People have diversified demands for financial management. Stocks are one of the most
important financial choices for a large number of people. The aim of investments is
making assets more valuable. Therefor investors prefer to buy high and sell low. But
the stock market if full of uncertainty, and the return is proportional to the risk. Few
investors and institutions can accurately predict stock movements, but historical data is
a permanent collection that allows people to make predictions about the future trends
of stocks.
Stocks and investments themselves belong to the economics. So most stock trend
prediction algorithms of the early stage are based on the knowledge of economics,
among which the knowledge of statistical probability theory is the most important.
From the perspective of mathematics, statistically analyzing the historical trading mar-
ket of stocks to obtain the laws, and then predict based on the statistical laws. However
based on the category of mathematical statistics, when it comes to hypothesis testing,
the difference between the predicted value and true is too large, which leads to the
decline of the prediction accuracy.
With the present of machine learning especially deep learning which integrate the
internal relations of training samples and build models based on them, the prediction
accuracy raised up. Deep learning can not only improve many defects of traditional pre-
diction ways but also avoid the influence of many human factors on the results because
of the strong nonlinear segmentation performance of the algorithm. Most of the hun-
dreds of senior managers in the financial field believe that machine learning and deep

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 51–62, 2021.
https://doi.org/10.1007/978-3-030-75100-5_5
52 B. Li and L. Li

learning can have a substantial impact on financial industry, especially in financial risk,
credit evaluation, and portfolio management.
The source of deep learning is a further exploration of neural network. For example,
the framework of deep learning includes a multilayer perceptron with multiple hidden
layers. In fact, deep learning is to find the internal rules among a large number of sam-
ples, with the ultimate goal of thinking and distinguish just like human beings. CNN
(Convolutional Neural Network) is a feedforward neural network whose artificial neu-
rons can respond to a part of the coverage of the surrounding units.
In this paper,we proposed feature extraction of stock data through CNN, and a stock
prediction algorithm based on deep learning combined with traditional LSTM. At the
same time, the third-order moment is used to improve the traditional stochastic gradient
descent, and obtain a dynamic adaptive stochastic gradient descent algorithm. On this
basis, experiments are carried out with actual dataset and compared with other meth-
ods. The results shows the new method is better than other algorithms which provides
a solution for stock trend prediction by deep learning.

2 Literature Review
2.1 Summary of Feature Selection

The first and the most important step of stocks prediction is feature selection. If we
can find the feature that affects fluctuation of stock most which other machine learning
problems faced, we can model algorithm using related factors and get the target val-
ues. A Rao and YK Kwon built mathematical analysis on history stock data and used
relatively mature statistical methods to get the correlation of technical indicators as a
basis for stock trend prediction [1, 2]. The most popular researches are on the transac-
tion data itself, such as the mathematical statistics of the trend and relevant analysis of
financial statements, which means the historical data of securities trading is taken as
the research object to analyze the future trend of securities [3, 4]. So most of researches
would firstly select features and then confirm the orientation. Such as CF Tsai carried
out a study by integrating various screening methods of impact factors [5] and JNK Liu
used a completely different screening method of impact factors [6].

2.2 Summary of Deep Learning

In the last century, basic regression models were very popular, and some scholars also
used classification models to predict stock trends such as Decision Tree [7], LR [8],
SVM [9]. In fact some scholars started exploring securities trading by neural network.
With the development of artificial intelligence, deep learning become popular and a
large of researches of stock prediction based on deep learning came up. Ozbayoglu,
A. M. and Bahadir, I. showed the different performance of securities trading between
bayes estimation and deep learning, and came to a conclusion that the accuracy of deep
learning model was better [10].
When using neural network method, the efficiency is seriously affected by the large
amount of calculation, and the quality of the optimization algorithm also determines
 Stock Prediction Based on Adaptive Gradient Descent Deep Learning 53

whether it can converge to the optimal solution. As an important aspect of deep learn-
ing, stochastic optimization algorithm affects the operating efficiency and final results.
Adaptive optimization methods, such as Adagrad, RMSprop, and Adam are proposed
for implementation a rapid training process, in which the learning rate can be flexibly
adjusted. However, instability and extreme learning rates may not converge to opti-
mal solutions (or critical points in non-convex Settings). With the development of deep
learning, a variety of first-order stochastic optimization algorithm came out, of which
the most widely algorithm is a stochastic gradient descent (SGD) [11]. Despite the
stochastic gradient descent is relatively simple, in actual it perform well and is still one
of the most common algorithm. But because of the fixed learning rate, SGD can only
uniformly scaling coordinates gradient in all directions to iteratively updating parame-
ters, which may leads slow convergence or results in loss function fluctuating around
the local minimum and even divergence.
In order to solve this problem, there are many adaptive stochastic optimization
method, one of the general solution is exponential moving averages (EMA) used by
Momentum and Adam [12], which helps to accelerate and reduce the gradient oscil-
lation on the gradient descent direction. And a variety of adaptive methods have been
proposed, including Adagrad, AdaDelta, RMSprop, Adam et al. These methods scale
the gradient by calculating the average value of the past gradient square, and due to their
fast training speed, they have been successfully extended in the field of deep learning.
Although these methods have been successfully adopted in several practical applica-
tions, it has also been observed that they cannot converge in some environments. Reddi
et al. provided some counter examples to prove that Adam could not converge to the
optimal solution, and in some cases, the generalization ability of invisible data is not
better than the non-adaptive method [13]. One of the reasons for this non-convergence
is EMA. Using EMA to estimate the proportion of gradient cannot guarantee that the
second-order is positively correlated with time, which will lead to the non-convergence
of Adam.
Reddi et al. proposed a new Adam variant AMSGrad to solve the problem of non-
convergence of the Adam, but at the expense of efficiency [13]. Zhou et al. put forward
a new explanation for the problem of non-convergence [14]. An abnormal association
exists between the gradient and Adam second-order moment, which may lead to the
step size imbalance. Because small gradients may have larger step sizes, which can
lead to larger updates, while large gradients may have smaller step sizes. Therefore, if
always updates in the wrong direction, it may lead to differences.

3 Proposed Method
3.1 CNN and LSTM
Except the input layer and output layer of normal neural network, the convolutional
neural network (CNN) has convolution layer, pooling layer and full connection layer.
The convolution layer is mainly used to perform the convolution operation on samples
by convolution verifies and then gets the input values of next layer. Pooling layer is
important for CNN, its funtion is to reduce the connection inside the convolution layer
and reduce the computational complexity.
54 B. Li and L. Li

The operation of convolution layer is the top priority of the whole CNN. Through a
example can help us to understand the CNN. LeNet proposed by Y.LeCun [15], which
structure as Fig. 1.

Fig. 1. Structure of LeNet

The input on the left is a hand-drawn image. In the middle, there are two pairs of
pooling layers and multiple fully connected hidden layers. The number of layers C1 is
20, and layer C2 is 50 different convolution templates. The number of neurons in the
next layer is 500. LeNet used only two convolutional layers at that time which is very
shallow, but its effect is good. Through the experiment on MNIST dataset, the error rate
of the final result is even lower than 0.95%, thus it can be seen that the depth of neural
network does not need to be very deep to achieve a good effect. The model adopts
gradient descent method to modify each parameter in the model layer by layer, and
minimizes the loss function through constant training, so as to improve the accuracy of
the model to the prediction target.

Fig. 2. LSTM neual unit

LSTM proposed by Sepp hochreiter and Jurgen Schmidhuber in 1997 [16]. It is a

special RNN (Recurrent Neural Network) [17]. The main advantage of RNN is taking
place in the context of information integration, and applicating in the process of the
input and output. But it still has gradient disappear, gradient explosion, etc.To solve
these problem, LSTM arosed. LSTM can capture long-term dependence better. By
introducing a memory unit to remember values in any time interval, and simultaneously
using input gate, output gate and forgetting gate to adjust the process of information
entering and leaving the memory unit, the gradient disappearance or explosion problem
existing in general RNN is effectively solved. The structure of each memory unit in the
LSTM network is shown in Fig. 2.
 Stock Prediction Based on Adaptive Gradient Descent Deep Learning 55

Firstly, the forgetting gate determines the process of losing information. According
to the output ht−1 of the previous memory unit and the input data xt , the forgetting gate
generates a value between 0 and 1 to determine how much information is lost in the
previous long-term state Ct−1 .

ft = σ (W f · [ht−1 , xt ] + b f ) , (1)

The input gate determines which information is added. ht−1 and xt are used to deter-
mine the update information through the input gate, and meanwhile, a tanh layer is used

to obtain the new candidate memory unit information Ct .

it = σ (Wi · [ht−1 , xt ] + bi ) , (2)

Ct = tanh(WC · [ht−1 , xt ] + bC ) , (3)

The previous long-term state Ct−1 is updated to Ct through the operation of the
forgetting gate and the input gate.

Ct = ft ∗Ct−1 + it ∗Ct , (4)

Finally, the judgment condition is obtained from the output gate, and then a value
between −1 and 1 is obtained through a tanh layer. This value is multiplied by the
judgment condition to determine which state characteristics of the current memory unit
to output.

ot = σ (Wo [ht−1 , xt ] + bo ) , (5)

ht = ot ∗ tanh(Ct ) , (6)

3.2 Improved Adam

We proposed an improving algorithm based on Adam in which applying third-order
moment. First-order moment optimization algorithm: SGDM algorithm introduces the
idea of first-order moment to help accelerate and reduce the gradient oscillation in the
descending direction of gradient. The first-order moment is the exponential moving
average (EMA) of the gradient. It’s equal to the average of the sum of the last 1/(1− β1 )
gradient vectors. The mark of the era of “adaptive learning rate” optimization algorithm
is the birth of the second-order moment. The second-order moment is the sum of the
√
squares of all gradients so far. For example, compared with the learning rate α / vt in
SGD, Adagrad effectively
 uses the idea of moderate learning rate attenuation, and the
learning rate is α / ∑ti=1 g2i . It has a good convergence effect in case of the gradient is
sparse. We usually add a small smoothing term to the denominator, otherwise there will
be no sense such as the denominator is 0, so the denominator is always greater than 0.
According to the definition of second-order moment, we can know that the faster the
56 B. Li and L. Li

parameter changes, the greater the second-order moment is, and then the learning rate
decreases. When the data distribution is sparse, the effect of this method is significant.
√
But at the moment, we find that increasing vt will cause the learning rate decreas-
ing rapidly and approach to zero infinitely. Leading to the problem of ending the train-
ing process in advance, and there is no way to learn well for the subsequent data,
which means giving up the later learning process directly. AdaDelt focus only on the
descent gradient in the past sliding window. Previous algorithms often take all cumula-
tive gradients into account. In this case, the second-order moment cumulative momen-
tum vt = β2 vt−1 + (1 − β2 ) is the gradient average of the past. The advantage of this
method is effectively preventing the second-order moment from increasing all the time
which result in the premature stop of the training process.
Combining the advantages of all the methods described above, Adam is very pop-
ular in the field of deep learning which is to make full use of the first-order moment
mt = β1 mt−1 + (1 − β1 ) · gt2 and the second-order moment vt = β2 vt−1 + (1 − β2 ) · gt2 .
So it can not only fully consider the acceleration and reduce the gradient oscillation in
the direction of gradient descent, but also avoid the rapid decline of learning rate and
end the training in advance. Although Adam is an advanced optimization algorithm,
many scholars have mentioned that sometimes the performance of adaptive methods
is even worse than SGD [18]. Reddi et al. pointed out the shortcomings of Adam and
provided some counter examples to prove that Adam cannot converge to the optimal
solution, and in some cases, the generalization ability of invisible data is not better than
non-adaptive method [13].
We propose a new variant of the optimization method, which applies the third-order
moment to Adam. Our goal is to propose a new method that will benefit from adap-
tive methods such as Adam and adadelta, as well as the excellent final performance of
SGD. Based on the previous discussion, we conclude that the main reason that Adam
may not be able to converge may be the oscillatory learning speed. We use the third-
order moment to reduce the fluctuation of learning rate. We first propose the following
definition of the third moment.
Definition 1. (skewness) The skewness of random variable x is defined as the third-
order moment γ1 defined as follows:
  
X −μ 3 μ3 E[(X − μ )3 ]
γ1 = E = =
σ σ3 (E[(X − μ )2 ]) 23
Where μ is the mean, σ is the standard deviation and E is the expected operator.
The concept of skewness is actually used to describe the asymmetry of random vari-
ables. We know that the probability distribution of random variables is usually irregular
and asymmetric, and it will shift to a certain direction in the actual graph. So the skew-
ness can be positive or negative. In some special cases, it can even be undefined. Nega-
tive skewness generally means that most of the values in the probability distribution are
greater than the average, which results in the tail on the left side of the function longer
than that on the right side. The conclusion of positive skewness is opposite to that of
negative skewness.
√
As we all know that α / vt is the learning rate and vt = β2 vt−1 + (1 − β2 ) · gt2 . Now,
√ √
let’s set the learning rate to α /( vt + 3 st , in order to make our method behave as
 Stock Prediction Based on Adaptive Gradient Descent Deep Learning 57

adaptive method in the early stage of training and behave like SGD algorithm at the end
of training.
Based on the third-order moment, we gave the pseudo code of adaptive gradient
method.

Algorithm 1:
Input:
α :step size
β1 , β2 , β3 ∈ [0, 1):exponential decay rates for the moment estimates.
η1 :lower bound function,ηu :upper bound function
f (θ ):Stochastic objective function with parameters
θ0 :Initial parameter vector
m0 ← 0:initialize 1st moment vector
v0 ← 0:initialize 2nd moment vector
t ← 0:initialize time step
While θt not converged do
t ← t +1
gt ← ∇θ ft (θt−1 ) stochastic objective gradients at time t
mt ← β2 vt−1 + (1 − β1 ) · gt update biased first moment estimate
vt ← β2 vt−1 + (1 − β2 ) · gt2 update biased second raw moment estimate
st ← β3 st−1 + (1 − β3 ) · (gt − mt )3 update third moment estimate
mt ← mt /(1 − β2t ) compute bias-corrected first moment estimate
vt ← vt /(1 − β2t ) compute bias-corrected second raw moment estimate
vt ← max( vt−1 , vt ) make sure | vt | > |
vt−1 |
α
ηt ← clip(√ , η1 (t), ηu (t)) clip the learn rates
[vt +sqrt[3][st ]
θt ← θt−1 − m
t · ηt update parameters
End while

4 Experiments
In the experiment, we will verifie the above methods. The experiment will input the
original data, predict the rise and fall of stocks in the future through CNN+LSTM
structure. CNN adopts the LeNet structure mentioned before. The LSTM uses three cir-
culating neural network units. At the same time, dynamic adaptive stochastic gradient
descent is adopted in the training, and compared with the traditional stochastic gradient
descent. This experiment uses deep learning framework Keras and Python to implement
the model. The operating system of the experimental environment was Ubuntu, and the
hardware environment was Xeon E5-2630v4 CPU and Nvidia RTX 2080TI GPU.
For data, we choose yahoo finance [19] from 2000 to 2019, a total of 19 years of
the Hang Seng Index. Indexes are a better reflection of stock market volatility than
individual stocks as a result of policies or specific economic conditions, at the same
time Hang Seng Index has enough volatility to verifies the algorithm we proposed. We
obtained the complete non-empty data of the 4439 groups where 1300 will be set as a
test set, the rest is used as training set. The input data used in this paper has the following
dimensions:
58 B. Li and L. Li

Table 1. Part of test data

Date Open High Low Close

2000-3-27 17827.71094 18350.60938 17815.67989 18292.85938
2000-3-28 18272.80078 18397.57031 18188.26953 18301.68945
2000-3-29 18203.13086 18249.26953 17950.19922 18096.36914
2000-3-30 18011.14063 18038.47070 17456.56055 17467.15039
2000-3-31 17383.63080 17537.13086 17107.43945 17406.53906
2000-4-3 17444.48047 17458.06055 16872.82031 16892.92969
2000-4-5 16599.07031 16736.22070 16246.53027 16318.44043
2000-4-6 16416.48047 16528.58008 16283.20020 16491.39063
2000-4-7 16657.83984 16992.52930 16657.83984 16941.67969
2000-4-10 17038.22070 17083.26953 16676.07031 16850.74023

The opening, lowest, highest, closing price, trading volume and KDJ, RSI, DMI,
BIAS, total 9 dimensional indicators.
The historical data of the first 25 d are used to predict the situation of the following
day. Therefore, the input of the experiment is a 25 × 9 matrix, some of which are shown
in the Table 1 below. The objective of the experiment is to predict the closing price. The
historical data of the experiment is shown in the following Fig. 3 In the preprocessing of
experimental data, in addition to deleting and replacing the missing data, the following
formula 7 is used to standardize. The standardized data is easy to calculate and shorten
the running time of the program. It is also more conducive to model training and model
generalization.
nt − mean(nt )
n̄t = , (7)
std(nt )
Use formula 8 as historical input data to predict the close price of the following day.
Mean Absolute Error (MAE) and Root-mean-square error (RMSE) are used to evaluate
the training results of three models. MAE is the average of the absolute error between
the observed value and the true value. RMSE refers to the square root of the ratio of
the squared deviation between the observed value and the truth value to the number of
observations m. The smaller the MAE and RMSE are, the better prediction model is.
Training set and test set are selected randomly by Gaussian distribution. Model evalua-
tion index formula are formula 9 and formula 10.
x(t) = (x(t−25) , x(t−24) , ..., x(2) , x(1) )T , (8)

i=1 |(x − x̄)|

∑m
MAE = , (9)
m

i=1 (x − x̄)
∑m 2
RMSE = , (10)
m
 Stock Prediction Based on Adaptive Gradient Descent Deep Learning 59

Fig. 3. Historical closing price chart

5 Discussion
SimpleRNN model and LSTM model of Keras are used for training and testing. We
realizes three model under the different conditions. The first is a stock prediction model
using LSTM alone, which uses traditional stochastic gradient descent in training. The
second uses traditional stochastic gradient descent training and CNN+LSTM model.
The third is the CNN+LSTM model using dynamic adaptive stochastic gradient descent
for training. For gradient descent, we select the initial learning rate as 0.01. The Fig. 4
shows the predicted results of the above three algorithms for the rise and fall of the
data set, in which the blue line represents the target value and the yellow line represents
the predicted value. At the same time, in order to reflect the trend of our model more
intuitively, we randomly selected a certain section of data in each experiment to predict
the trend. The relevant results are shown in the Fig. 4.
Through the above experiments, we can find that for each model, the number of
times that the yellow line coincides with the blue line in the pattern trend is significantly
more than not coincide. At the same time, it can be seen from the trend prediction dia-
gram that CNN + LSTM is obviously superior to the traditional LSTM under the con-
dition of traditional stochastic gradient descent. But for CNN + LSTM using dynamic
adaptive stochastic gradient descent training, it is obviously better than the traditional
stochastic gradient descent training model. Next, we calculate MAE and RMSE under
three model conditions, as shown in the Table 2.

Table 2. MAE and RMSE in there model

LSTM under CNN and LSTM under CNN + LSTM under

traditional SGD traditional SGD Dynamic adaptive SGD
MAE 0.1092 0.0279 0.0063
RMSE 0.1413 0.0396 0.0083

According to the results in the above table, the prediction errors of the three mod-
els are decreasing and the prediction accuracy is improving. The prediction accuracy
of the adaptive stochastic gradient descent model is greatly improved. The MAE and
RMSE of LSTM prediction model with traditional SGD training are 0.1092 and 0.1413
60 B. Li and L. Li

Fig. 4. Randomly selected several section of data to predict the trend

respectively. CNN +LSTM reduces the MAE by 0.0813 and the RMSE by 0.1017. After
using dynamic adaptive stochastic gradient descent, the prediction MAE is reduced by
0.1029 and the RMSE is reduced by 0.133 compared with the cyclic neural network.
By comparing the MAE and RMSE of the three prediction models, the prediction effect
of CNN + LSTM using the adaptive stochastic gradient descent is better than that of
the conventional stochastic gradient descent. The CNN + LSTM using traditional ran-
dom gradient descent is superior to the single LSTM network using traditional random
gradient descent. As a conclusion, the CNN + LSTM model with adaptive stochastic
gradient descent has higher accuracy and more predictive value in practice.
 Stock Prediction Based on Adaptive Gradient Descent Deep Learning 61

6 Conclusions
In this paper, feature extraction of stock data through CNN is proposed, and a stock pre-
diction algorithm based on deep learning is constructed in combination with traditional
LSTM. At the same time, the third-order moment is used to improve the traditional
stochastic gradient descent, and obtain a dynamic adaptive stochastic gradient descent
algorithm. On this basis, the experiment is carried out with actual data and compared
with other methods, and the results are better than other algorithms. It provides a solu-
tion for stock trend prediction by deep learning.

References
1. Shaikh, H., Rao, A., Hule, S.: Survey: market prediction using statistical computational
methodologies and artificial neural networks (2015)
2. Kwon, Y.K., Choi, S.S., Moon, B.: Stock prediction based on financial correlation. pp. 2061–
2066, June 2005
3. Lahmiri, S.: Article: a comparison of PNN and SVM for stock market trend prediction using
economic and technical information. Int. J. Comput. Appl. 29(3), 24–30 (2011)
4. Gui, B., Wei, X., Shen, Q., Qi, J., Guo, L.: Financial time series forecasting using support
vector machine. In: 2014 Tenth International Conference on Computational Intelligence and
Security, pp. 39–43 (2014)
5. Akita, R., Yoshihara, A., Matsubara, T., Uehara., K.: Deep learning for stock prediction
using numerical and textual information. In: 2016 IEEE/ACIS 15th International Confer-
ence on Computer and Information Science (ICIS), Los Alamitos, CA, USA, pp. 1–6. IEEE
Computer Society, June 2016
6. Tsai, C.F., Hsiao, Y.C.: Combining multiple feature selection methods for stock prediction:
union, intersection, and multi-intersection approaches. Decis. Support Syst. 50(1), 258–269
(2011)
7. Hunt, E.B., Marin, J., Stone, P.J.: Experiments in induction. Am. J. Psychol. 80, 4 (1966)
8. Cramer, J.S.: The origins of logistic regression. Tinbergen Institute Discussion Papers (2002)
9. Vapnik, V.N., Chervonenkis, A.: A note on one class of perceptrons. Autom. Remote Control
25(1) (1964)
10. Ozbayoglu, A.M., Bahadir, I.: comparison of Bayesian estimation and neural network model
in stock market trading. Intelligent Engineering Systems Through Artificial Neural Networks
(2008)
11. Robbins, H., Monro, S.: A stochastic approximation method. Ann. Math. Stat. 22(3), 400–
407 (1951)
12. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts, August
2016
13. Sashank, J., Reddi, S.K., Sanjiv K.: On the convergence of Adam & beyond, May 2018
14. Zhou, Z., Zhang, Q., Lu, G., Wang, H., Zhang, W., Yu, Y.: Adashift: decorrelation and con-
vergence of adaptive learning rate methods, September 2018
15. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document
recognition. Proc. IEEE 86(11), 2278–2324 (1998)
16. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8), 1735–1780
(1997)
62 B. Li and L. Li

17. Le Guillarme, N., Lerouvreur, X.: Unsupervised extraction of knowledge from S-AIS data
for maritime situational awareness. In: 2013 16th International Conference on Information
Fusion (FUSION) (2013)
18. Keskar, N.S., Socher, R.: Improving generalization performance by switching from Adam to
SGD (2017)
19. Finance yahoo. http://table.finance.yahoo.com/
 A QoS-Aware Fuzzy-Based System
for Assessment of Edge Computing
Resources in SDN-VANETs

Ermioni Qafzezi1(B) , Kevin Bylykbashi1 , Phudit Ampririt1 , Makoto Ikeda2 ,

Keita Matsuo2 , and Leonard Barolli2
1
Graduate School of Engineering, Fukuoka Institute of Technology (FIT),
3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan
2
Department of Information and Communication Engineering, Fukuoka Institute
of Technology (FIT), 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan
[email protected], {kt-matsuo,barolli}@fit.ac.jp

Abstract. In this paper, we propose a fuzzy-based system to assess the

available edge computing resources in Software Defined Vehicular Ad
hoc Networks (SDN-VANETs) by considering Quality of Service (QoS)
of the communication link among vehicles as a new parameter. Our pro-
posed system, called Fuzzy System for Assessment of Neighboring Vehicle
Processing Capability (FS-ANVPC), determines if a neighbor vehicle is
capable of helping out to complete certain tasks in terms of data process-
ing. FS-ANVPC is implemented in vehicles and in SDN Controllers. It
is used from vehicles which need additional resources to run their appli-
cations or process different application data related (but not limited) to
VANETs. However, for that to happen, the vehicles should remain within
the communication range of each other for a while and the communica-
tion link should satisfy certain requirements. In addition, the available
resources of the neighbor vehicle are also considered by the system. The
output decides the processing capability of that neighbor. We evaluate
FS-ANVPC by computer simulations. A neighbor is decided as helpful
when a communication session with the vehicle is maintained for a long
time over a reliable link, while possessing a moderate amount of available
resources.

1 Introduction

The long distances separating homes and workplaces/facilities/schools as well

as the traffic present in these distances make people spend a significant amount
of time in vehicles. Thus, it is important to offer drivers and passengers ease of
driving, convenience, efficiency and safety. This has led to the emerging of Vehic-
ular Ad hoc Networks (VANETs), where vehicles are able to communicate and
share important information among them. VANETs are a relevant component
of Intelligent Transportation System (ITS), which offers more safety and better
transportation.

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 63–72, 2021.
https://doi.org/10.1007/978-3-030-75100-5_6
64 E. Qafzezi et al.

VANETs are capable to offer numerous services such as road safety, enhanced
traffic management, as well as travel convenience and comfort. To achieve road
safety, emergency messages must be transmitted in real-time, which stands also
for the actions that should be taken accordingly in order to avoid potential acci-
dents. Thus, it is important for the vehicles to always have available connections
to infrastructure and to other vehicles on the road. On the other hand, traffic
efficiency is achieved by managing traffic dynamically according to the situation
and by avoiding congested roads, whereas comfort is attained by providing in-car
infotainment services.
The advances in vehicle technology have made possible for the vehicles to be
equipped with various forms of smart cameras and sensors, wireless communi-
cation modules, storage and computational resources. While more and more of
these smart cameras and sensors are incorporated in vehicles, massive amounts
of data are generated from monitoring the on-road and in-board status. This
exponential growth of generated vehicular data, together with the boost of the
number of vehicles and the increasing data demands from in-vehicle users, has
led to a tremendous amount of data in VANETs [8]. Moreover, applications
like autonomous driving require even more storage capacity and complex com-
putational capability. As a result, traditional VANETs face huge challenges in
meeting such essential demands of the ever-increasing advancement of VANETs.
The integration of Cloud-Fog-Edge Computing in VANETs is the solution
to handle complex computation, provide mobility support, low latency and high
bandwidth. Each of them serves different functions, but also complements each-
other in order to enhance the performance of VANETs. Even though the inte-
gration of Cloud, Fog and Edge Computing in VANETs solves significant chal-
lenges, this architecture lacks mechanisms needed for resource and connectivity
management because the network is controlled in a decentralized manner. The
prospective solution to solve these problems is the augmentation of Software
Defined Networking (SDN) in this architecture.
The SDN is a promising choice in managing complex networks with minimal
cost and providing optimal resource utilization. SDN offers a global knowledge
of the network with a programmable architecture which simplifies network man-
agement in such extremely complicated and dynamic environments like VANETs
[7]. In addition, it will increase flexibility and programmability in the network
by simplifying the development and deployment of new protocols and by bring-
ing awareness into the system, so that it can adapt to changing conditions and
requirements, i.e., emergency services [3]. This awareness allows SDN-VANET to
make better decisions based on the combined information from multiple sources,
not just individual perception from each node.
In a previous work [6], we have proposed an intelligent approach to man-
age the cloud-fog-edge resources in SDN-VANETs using fuzzy logic. We have
presented a cloud-fog-edge layered architecture which is coordinated by an intel-
ligent system that decides the appropriate resources to be used by a particular
vehicle (hereafter will be referred as the vehicle) in need of additional comput-
ing resources. The main objective is to achieve a better management of these
resources.
 A QoS-Aware Fuzzy-Based System for Assessment of Edge Resources 65

In this work we focus only on the edge layer resources. We include the Quality
of Service (QoS) of the communication link among vehicles to better assess their
processing capability. The proposed Fuzzy System for Assessment of Neighboring
Vehicle Processing Capability (FS-ANVPC) determines the processing capability
for each neighboring vehicle separately, hence helpful neighboring vehicles could
be discovered and a better assessment of available edge resources can be made.
We see the effect of QoS on the evaluation of the neighbor vehicles capability to
help the vehicle accomplishing different tasks.
The remainder of the paper is as follows. In Sect. 2, we present an overview
of Cloud-Fog-Edge SDN-VANETs. In Sect. 3, we describe the proposed fuzzy-
based system. In Sect. 4, we discuss the simulation results. Finally, conclusions
and future work are given in Sect. 5.

2 Cloud-Fog-Edge SDN-VANETs
While cloud, fog and edge computing in VANETs offer scalable access to stor-
age, networking and computing resources, SDN provides higher flexibility, pro-
grammability, scalability and global knowledge. In Fig. 1, we give a detailed
structure of this novel VANET architecture. It includes the topology structure,
its logical structure and the content distribution on the network. As it is shown,
it consists of Cloud Computing data centers, fog servers with SDNCs, road-
side units (RSUs), RSU Controllers (RSUCs), Base Stations and vehicles. We
also illustrate the infrastructure-to-infrastructure (I2I), vehicle-to-infrastructure
(V2I), and vehicle-to-vehicle (V2V) communication links. The fog devices (such
as fog servers and RSUs) are located between vehicles and the data centers of
the main cloud environments.
The safety applications data generated through in-board and on-road sensors
are processed first in the vehicles as they require real-time processing. If more
storing and computing resources are needed, the vehicle can request to use those
of the other adjacent vehicles, assuming a connection can be established and
maintained between them for a while. With the vehicles having created multi-
ple virtual machines on other vehicles, the virtual machine migration must be
achievable in order to provide continuity as one/some vehicle may move out of
the communication range. However, to set-up virtual machines on the nearby
vehicles, multiple requirements must be met and when these demands are not
satisfied, the fog servers are used.
Cloud servers are used as a repository for software updates, control policies
and for the data that need long-term analytics and are not delay-sensitive. On
the other side, SDN modules which offer flexibility and programmability, are
used to simplify the network management by offering mechanisms that improve
the network traffic control and coordination of resources. The implementation
of this architecture promises to enable and improve the VANET applications
such as road and vehicle safety services, traffic optimization, video surveillance,
telematics, commercial and entertainment applications.
66 E. Qafzezi et al.

Fig. 1. Logical architecture of cloud-fog-edge SDN-VANET with content distribution.

3 Proposed Fuzzy-Based System

In this section, we present our proposed fuzzy based system. A vehicle that needs
storage and computing resources for a particular application can use those of
neighboring vehicles, fog servers or cloud data centers based on the application
requirements. For instance, for a temporary application that needs real-time pro-
cessing, the vehicle can use the resources of adjacent vehicles if the requirements
to realize such operations are fulfilled. Otherwise, it will use the resources of fog
servers, which offer low latency as well. Whereas real-time applications require
the usage of edge and fog layer resources, for delay tolerant applications, vehicles
can use the cloud resources as these applications do not require low latency.
The proposed system is implemented in the SDNC and in the vehicles which
are equipped with SDN modules. If a vehicle does not have an SDN module, it
sends the information to SDNC which sends back its decision. The system uses
the beacon messages received from the adjacent vehicles to extract information
such as their current position, velocity, direction, available computing power,
available storage, and based on the received data, the processing capability of
each adjacent vehicle is decided.
The structure of the proposed system is shown in Fig. 2. For the implemen-
tation of our system, we consider four input parameters: Available Computing
Power (APC), Available Storage (AS), Predicted Contact Duration (PCD) and
Quality of Service (QoS) to determine the Neighbor i Processing Capability
(NiPC).
 A QoS-Aware Fuzzy-Based System for Assessment of Edge Resources 67

Fig. 2. Proposed system structure.

ACP : Vehicles might be using their computing power for their own applica-
tions but a reserved amount can be allocated to help other vehicles in need to
complete certain tasks. Vehicles let their neighbors know that they are willing
to share their resources and how much they want to share. In other words, they
decide the amount of physical processor cores and the amount of memory that
a particular vehicle can use.
AS : The neighbors should have a specific amount of storage so the vehicle can
run the virtual machines. This storage is used also to store data after completing
specific tasks of all the tasks these neighbors are asked to accomplish.
PCD: In a V2V communication, the duration of the communication session
is important since it determines the amount of data to be exchanged and the
services that can be performed. A vehicle which needs additional resources will
create virtual machines on the neighbors that are willing to lend their resources,
therefore the contact duration becomes even more important since much more
time is needed to accomplish these tasks than just performing a data exchange.
QoS : Different VANET applications have different requirements in terms of
delay, bandwidth, throughput and so on. Safety applications, for example, have
strong QoS demands as they require a real time communication and a reliable
link to reach their goals, that is reducing the car accidents. However, it is a
challenging task to satisfy at all times the QoS requirements given the highly
dynamic topology and physical obstacles and interferences on these networks.
NiPC : The output parameter values consist of values between 0 and 1, with
the value 0.5 working as a border to determine if a neighbor is capable of helping
out the vehicle. A NiPC ≥ 0.5 means that this neighbor i has the required
conditions to help the vehicle to complete its tasks.
We consider fuzzy logic to implement the proposed system because our sys-
tem parameters are not correlated with each other. Having three or more param-
eters which are not correlated with each other results in a non-deterministic
polynomial-time hard (NP-hard) problem and fuzzy logic can deal with these
problems. Moreover, we want our system to make decisions in real time and
68 E. Qafzezi et al.

fuzzy systems can give very good results in decision making and control prob-
lems [1,2,4,5,9,10].

Fig. 3. Membership functions.

The input parameters are fuzzified using the membership functions shown in
Figs. 3(a), 3(b), 3(c) and 3(d). In Fig. 3(e) are shown the membership functions
used for the output parameter. We use triangular and trapezoidal membership
functions because they are suitable for real-time operation. The term sets for
each linguistic parameter are shown in Table 1. We decided the number of term
sets by carrying out many simulations. In Table 2, we show the Fuzzy Rule
Base (FRB) of our proposed system, which consists of 81 rules. The control
rules have the form: IF “conditions” THEN “control action”. For instance, for
Rule 1: “IF ACP is Sm, AS is S, PCD is Sh and QoS is Lw, THEN NiPC is
ELPC” or for Rule 50: “IF ACP is Me, AS is B, PCD is Md and QoS is Mo
THEN NiPC is LPC”.
 A QoS-Aware Fuzzy-Based System for Assessment of Edge Resources 69

Table 1. System parameters and their term sets.

Parameters Term sets

Available Computing Power (ACP) Small (Sm), Medium (Me), Large
(La)
Available Storage (AS) Small (S), Medium (M), Big (B)
Predicted Contact Duration (PCD) Short (Sh), Medium (Md), Long
(Lo)
Quality of Service (QoS) Low (Lw), Moderate (Mo), High
(Hi)
Neighbor i Processing Capability (NiPC) Extremely Low Processing
Capability (ELPC), Very Low
Processing Capability (VLPC),
Low Processing Capability (LPC),
Moderate Processing Capability
(MPC), High Processing Capability
(HPC), Very High Processing
Capability (VHPC), Extremely
High Processing Capability
(EHPC)

4 Simulation Results

The simulations were conducted using FuzzyC and the results are shown for
three scenarios: in Fig. 4 is shown the scenario where the available computing
power is small, Fig. 5 shows the results for medium available computing power
and Fig. 6 the results for large available computing power. We show the relation
between NiPC and QoS for different PCD while keeping constant ACP and AS
values.
In Fig. 4(a), the neighboring vehicles have small processing capability and
small available storage. Due to the low processing resources, we see that none of
the vehicles is considered as helpful, even when a long predicted contact duration
between vehicles is anticipated. However, with the increase of AS, we see that
vehicles with long predicted contact duration and the highest value of QoS are
considered as helpful. Although these neighbors have small available computing
power, it is worth taking these vehicles into consideration as they can become
potential neighbors in the case of an increase of the ACP, even if this increase
might not be a big one. We see this effect in Fig. 5 where we consider the scenario
with neighbors that have moderate amounts of ACP.
70 E. Qafzezi et al.

Table 2. FRB of FS-ANVPC.

No ACP AS PCD QoS NiPC No ACP AS PCD QoS NiPC No ACP AS PCD QoS NiPC
1 Sm S Sh Lw ELPC 28 Me S Sh Lw ELPC 55 La S Sh Lw ELPC
2 Sm S Sh Mo ELPC 29 Me S Sh Mo ELPC 56 La S Sh Mo VLPC
3 Sm S Sh Hi ELPC 30 Me S Sh Hi VLPC 57 La S Sh Hi LPC
4 Sm S Md Lw ELPC 31 Me S Md Lw VLPC 58 La S Md Lw VLPC
5 Sm S Md Mo ELPC 32 Me S Md Mo VLPC 59 La S Md Mo LPC
6 Sm S Md Hi VLPC 33 Me S Md Hi LPC 60 La S Md Hi MPC
7 Sm S Lo Lw ELPC 34 Me S Lo Lw VLPC 61 La S Lo Lw LPC
8 Sm S Lo Mo ELPC 35 Me S Lo Mo LPC 62 La S Lo Mo MPC
9 Sm S Lo Hi LPC 36 Me S Lo Hi MPC 63 La S Lo Hi VHPC
10 Sm M Sh Lw ELPC 37 Me M Sh Lw ELPC 64 La M Sh Lw VLPC
11 Sm M Sh Mo ELPC 38 Me M Sh Mo ELPC 65 La M Sh Mo LPC
12 Sm M Sh Hi ELPC 39 Me M Sh Hi LPC 66 La M Sh Hi HPC
13 Sm M Md Lw ELPC 40 Me M Md Lw VLPC 67 La M Md Lw LPC
14 Sm M Md Mo ELPC 41 Me M Md Mo VLPC 68 La M Md Mo MPC
15 Sm M Md Hi VLPC 42 Me M Md Hi MPC 69 La M Md Hi VHPC
16 Sm M Lo Lw ELPC 43 Me M Lo Lw LPC 70 La M Lo Lw MPC
17 Sm M Lo Mo VLPC 44 Me M Lo Mo MPC 71 La M Lo Mo VHPC
18 Sm M Lo Hi LPC 45 Me M Lo Hi HPC 72 La M Lo Hi EHPC
19 Sm B Sh Lw ELPC 46 Me B Sh Lw ELPC 73 La B Sh Lw LPC
20 Sm B Sh Mo ELPC 47 Me B Sh Mo VLPC 74 La B Sh Mo MPC
21 Sm B Sh Hi ELPC 48 Me B Sh Hi LPC 75 La B Sh Hi HPC
22 Sm B Md Lw ELPC 49 Me B Md Lw VLPC 76 La B Md Lw MPC
23 Sm B Md Mo VLPC 50 Me B Md Mo LPC 77 La B Md Mo VHPC
24 Sm B Md Hi LPC 51 Me B Md Hi MPC 78 La B Md Hi VHPC
25 Sm B Lo Lw VLPC 52 Me B Lo Lw LPC 79 La B Lo Lw HPC
26 Sm B Lo Mo LPC 53 Me B Lo Mo HPC 80 La B Lo Mo EHPC
27 Sm B Lo Hi MPC 54 Me B Lo Hi HPC 81 La B Lo Hi EHPC

Fig. 4. Simulation results for ACP = 0.1.

 A QoS-Aware Fuzzy-Based System for Assessment of Edge Resources 71

Fig. 5. Simulation results for ACP = 0.5.

Fig. 6. Simulation results for ACP = 0.9.

When vehicles are willing to lend a large processing power but small storage
(see Fig. 6(a)), we see that only the ones that will stay inside the communication
range for a long period of time and offer at least a moderate value of QoS are
able to help. Although these neighbors have small available storage, they can
process the data while retransmitting the processed data step by step after each
single sub-task/task has finished, thus successfully accomplishing the required
processing tasks. However, for this to happen, the communication link between
these neighbors and the vehicle, must assure a good value of QoS so that they
could be able to retransmit the output after each completed subtask. On the
other hand, from Fig. 6(b), we see that when a neighbor vehicle offers big avail-
able resources in terms of both computing power and storage, it is considered
as helpful if the communication session is expected to last for a while regardless
the assured QoS. In case of a bad QoS, these neighbors can store the processed
data in their storage and transmit them when an improved Qos is provided.

5 Conclusions
In this paper, we proposed a fuzzy-based system to assess the available edge com-
puting resources in a layered Cloud-Fog-Edge architecture for SDN-VANETs.
72 E. Qafzezi et al.

Our proposed system decides if a neighboring vehicle is capable to help a vehicle

that lacks the appropriate resources to accomplish certain tasks based on ACP,
AS, PCD and QoS. After assessing the processing capability for each neighbor,
our previous proposed Fuzzy System for Resource Management [6] can select
the appropriate layer in terms of data processing. We evaluated our proposed
system by computer simulations. From the simulations results, we conclude as
follows.

• The neighboring vehicles which have small ACP and small AS are not capable
to help other vehicles in need regardless enabling a reliable communication
link between them.
• A neighboring vehicle with a medium amount of ACP is considered as helpful,
when the communication session with the vehicle in need is long and reliable.
• The highest value of NiPC is achieved when the neighboring vehicle has a
large ACP, big AS, long PCD and a moderate/high QoS.

In the future, we would like to make extensive simulations to evaluate the

proposed system and compare the performance with other systems.

References
1. Kandel, A.: Fuzzy expert systems. CRC Press Inc., Boca Raton, FL, USA (1992)
2. Klir, G.J., Folger, T.A.: Fuzzy sets, uncertainty, and information. Prentice Hall,
Upper Saddle River, NJ, USA (1988)
3. Ku, I., Lu, Y., Gerla, M., Gomes, R.L., Ongaro, F., Cerqueira, E.: Towards
software-defined VANET: Architecture and services. In: 13th Annual Mediter-
ranean Ad Hoc Networking Workshop (MED-HOC-NET), pp. 103–110 (2014)
4. McNeill, F.M., Thro, E.: Fuzzy logic: a practical approach. Academic Press Pro-
fessional Inc., San Diego, CA, USA (1994)
5. Munakata, T., Jani, Y.: Fuzzy systems: an overview. Commun. ACM 37(3), 69–77
(1994)
6. Qafzezi, E., Bylykbashi, K., Ikeda, M., Matsuo, K., Barolli, L.: Coordination and
management of cloud, fog and edge resources in SDN-VANETS using fuzzy logic:
a comparison study for two fuzzy-based systems. Internet of Things 11, 100169
(2020)
7. Truong, N.B., Lee, G.M., Ghamri-Doudane, Y.: Software defined networking-based
vehicular adhoc network with fog computing. In: 2015 IFIP/IEEE International
Symposium on Integrated Network Management (IM), pp. 1202–1207 (2015)
8. Xu, W., Zhou, H., Cheng, N., Lyu, F., Shi, W., Chen, J., Shen, X.: Internet of
vehicles in big data era. IEEE/CAA J. Automatica Sinica 5(1), 19–35 (2018)
9. Zadeh, L.A., Kacprzyk, J.: Fuzzy logic for the management of uncertainty. John
Wiley & Sons Inc., New York, NY, USA (1992)
10. Zimmermann, H.J.: Fuzzy control. In: Fuzzy Set Theory and Its Applications, pp.
203–240. Springer (1996)
 A Comparison Study of Constriction
and Random Inertia Weight Router
Replacement Methods for WMNs
by WMN-PSOHC Hybrid Intelligent
Simulation System Considering Normal
Distribution of Mesh Clients

Shinji Sakamoto1(B) , Leonard Barolli2 , and Shusuke Okamoto1

1
Department of Computer and Information Science, Seikei University,
3-3-1 Kichijoji-Kitamachi, Musashino-shi 180-8633, Tokyo, Japan
[email protected]
2
Department of Information and Communication Engineering,
Fukuoka Institute of Technology, 3-30-1 Wajiro-Higashi,
Higashi-Ku, Fukuoka 811-0295, Japan
[email protected]

Abstract. Wireless Mesh Networks (WMNs) are low-cost and have easy
maintenance. However, WMNs have some problems e.g. node placement
problems, security, transmission power and so on. To solve node place-
ment problem in WMNs, we have implemented a hybrid simulation sys-
tem based on PSO and HC called WMN-PSOHC. In this paper, we
present the performance evaluation of WMNs by using WMN-PSOHC
intelligent system considering Constriction Method (CM) and Rondom
Inertia Weight Method (RIWM). The simulation results show that a
better performance is achieved for CM compared with the RIWM.

1 Introduction

The wireless networks and devices are becoming increasingly popular and they
provide users access to information and communication anytime and any-
where [2–4,7,8,10]. Wireless Mesh Networks (WMNs) are gaining a lot of atten-
tion because of their low cost nature that makes them attractive for providing
wireless Internet connectivity. A WMN is dynamically self-organized and self-
configured, with the nodes in the network automatically establishing and main-
taining mesh connectivity among them-selves (creating, in effect, an ad hoc net-
work). This feature brings many advantages to WMNs such as low up-front cost,
easy network maintenance, robustness and reliable service coverage [1]. More-
over, such infrastructure can be used to deploy community networks, metropoli-
tan area networks, municipal and corporative networks, and to support applica-
tions for urban areas, medical, transport and surveillance systems.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 73–81, 2021.
https://doi.org/10.1007/978-3-030-75100-5_7
74 S. Sakamoto et al.

In this work, we deal with node placement problem in WMNs. We consider

the version of the mesh router nodes placement problem in which we are given
a grid area where to deploy a number of mesh router nodes and a number of
mesh client nodes of fixed positions (of an arbitrary distribution) in the grid
area. The objective is to find a location assignment for the mesh routers to
the cells of the grid area that maximizes the network connectivity and client
coverage. Network connectivity is measured by Size of Giant Component (SGC)
of the resulting WMN graph, while the user coverage is simply the number
of mesh client nodes that fall within the radio coverage of at least one mesh
router node and is measured by Number of Covered Mesh Clients (NCMC). Node
placement problems are known to be computationally hard to solve [21]. In some
previous works, intelligent algorithms have been recently investigated [9,15,16].
We already implemented a Particle Swarm Optimization (PSO) based simulation
system, called WMN-PSO [13]. Also, we implemented a simulation system based
on Hill Climbing (HC) for solving node placement problem in WMNs, called
WMN-HC [12].
In our previous work [13,14], we presented a hybrid intelligent simulation sys-
tem based on PSO and HC. We called this system WMN-PSOHC. In this paper,
we analyze the performance of Constriction Method (CM) and Random Iner-
tia Weight Method (RIWM) by WMN-PSOHC simulation system considering
Normal distribution of mesh clients.
The rest of the paper is organized as follows. We present our designed and
implemented hybrid simulation system in Sect. 2. In Sect. 3, we introduce WMN-
PSOHC Web GUI tool. The simulation results are given in Sect. 4. Finally, we
give conclusions and future work in Sect. 5.

2 Proposed and Implemented Simulation System

2.1 Particle Swarm Optimization
In Particle Swarm Optimization (PSO) algorithm, a number of simple entities
(the particles) are placed in the search space of some problem or function and
each evaluates the objective function at its current location. The objective func-
tion is often minimized and the exploration of the search space is not through
evolution [11]. However, following a widespread practice of borrowing from the
evolutionary computation field, in this work, we consider the bi-objective func-
tion and fitness function interchangeably. Each particle then determines its move-
ment through the search space by combining some aspect of the history of its
own current and best (best-fitness) locations with those of one or more members
of the swarm, with some random perturbations. The next iteration takes place
after all particles have been moved. Eventually the swarm as a whole, like a flock
of birds collectively foraging for food, is likely to move close to an optimum of
the fitness function.
Each individual in the particle swarm is composed of three D-dimensional
vectors, where D is the dimensionality of the search space. These are the current
position xi , the previous best position pi and the velocity vi .
 A Comparison Study of CM and RIWM Router Replacement Methods 75

The particle swarm is more than just a collection of particles. A particle by

itself has almost no power to solve any problem; progress occurs only when the
particles interact. Problem solving is a population-wide phenomenon, emerging
from the individual behaviors of the particles through their interactions. In any
case, populations are organized according to some sort of communication struc-
ture or topology, often thought of as a social network. The topology typically
consists of bidirectional edges connecting pairs of particles, so that if j is in i’s
neighborhood, i is also in j’s. Each particle communicates with some other par-
ticles and is affected by the best point found by any member of its topological
neighborhood. This is just the vector pi for that best neighbor, which we will
denote with pg . The potential kinds of population “social networks” are hugely
varied, but in practice certain types have been used more frequently.
In the PSO process, the velocity of each particle is iteratively adjusted so
that the particle stochastically oscillates around pi and pg locations.

2.2 Hill Climbing

Hill Climbing (HC) algorithm is a heuristic algorithm. The idea of HC is simple.

In HC, the solution s is accepted as the new current solution if δ ≤ 0 holds,
where δ = f (s ) − f (s). Here, the function f is called the fitness function. The
fitness function gives points to a solution so that the system can evaluate the
next solution s and the current solution s.
The most important factor in HC is to define effectively the neighbor solution.
The definition of the neighbor solution affects HC performance directly. In our
WMN-PSOHC system, we use the next step of particle-pattern positions as the
neighbor solutions for the HC part.

2.3 WMN-PSOHC System Description

In following, we present the initialization, client distributions, particle-pattern,

fitness function and router replacement methods.
Initialization
Our proposed system starts by generating an initial solution randomly, by ad hoc
methods [22]. We decide the velocity of particles by a random process considering
instance, when√the area size is W × H, the velocity is decided
the area size. For √
randomly from − W 2 + H 2 to W 2 + H 2 .
Client Distributions
Our system can generate many client distributions. In this paper, we consider
Normal distribution of mesh clients. In Normal distribution, clients are located
around the center of considered area as shown in Fig. 1
Particle-Pattern
A particle is a mesh router. A fitness value of a particle-pattern is computed
by combination of mesh routers and mesh clients positions. In other words,
76 S. Sakamoto et al.

Fig. 1. Normal distribution of mesh clients.

each particle-pattern is a solution as shown is Fig. 2. Therefore, the number of

particle-patterns is a number of solutions.
Fitness Function
One of most important thing is to decide the determination of an appropriate
objective function and its encoding. In our case, each particle-pattern has an
own fitness value and compares other particle-patterns fitness value in order to
share information of global solution. The fitness function follows a hierarchical
approach in which the main objective is to maximize the SGC in WMN. Thus,
we use α and β weight-coefficients for the fitness function and the fitness function
of this scenario is defined as:
Fitness = α × SGC(xij , y ij ) + β × NCMC(xij , y ij ).

Router Replacement Methods

A mesh router has x, y positions and velocity. Mesh routers are moved based on
velocities. There are many router replacement methods in PSO field [6,18–20].
In this paper, we consider CM and RIWM.
Constriction Method (CM)
CM is a method which PSO parameters are set to a week stable region (ω =
0.729, C1 = C2 = 1.4955) based on analysis of PSO by M. Clerc et al. [4–6].
Random Inertia Weight Method (RIWM)
In RIWM, the ω parameter is changing randomly from 0.5 to 1.0. The C1
and C2 are kept 2.0. The ω can be estimated by the week stable region. The
average of ω is 0.75 [17,19].

3 WMN-PSOHC Web GUI Tool

The Web application follows a standard Client-Server architecture and is imple-
mented using LAMP (Linux + Apache + MySQL + PHP) technology (see
 A Comparison Study of CM and RIWM Router Replacement Methods 77

Fig. 2. Relationship among global solution, particle-patterns and mesh routers.

Fig. 3). We show the WMN-PSOHC Web GUI tool in Fig. 4. Remote users
(clients) submit their requests by completing first the parameter setting. The
parameter values to be provided by the user are classified into three groups, as
follows.

Fig. 3. System structure for web interface.

• Parameters related to the problem instance: These include parameter values

that determine a problem instance to be solved and consist of number of
router nodes, number of mesh client nodes, client mesh distribution, radio
coverage interval and size of the deployment area.
• Parameters of the resolution method: Each method has its own parameters.
• Execution parameters: These parameters are used for stopping condition of
the resolution methods and include number of iterations and number of inde-
pendent runs. The former is provided as a total number of iterations and
depending on the method is also divided per phase (e.g., number of iterations
in a exploration). The later is used to run the same configuration for the same
problem instance and parameter configuration a certain number of times.
78 S. Sakamoto et al.

Fig. 4. WMN-PSOHC Web GUI Tool.

4 Simulation Results
In this section, we show simulation results using WMN-PSOHC system. In this
work, we consider Normal distribution of mesh clients. The number of mesh
routers is considered 16 and the number of mesh clients 48. We consider the
number of particle-patterns 9. We conducted simulations 100 times, in order to
avoid the effect of randomness and create a general view of results. The total
number of iterations is considered 800 and the iterations per phase is considered
4. We show the parameter setting for WMN-PSOHC in Table 1.

Table 1. Parameter settings.

Parameters Values
Clients distribution Normal distribution
Area size 32.0 × 32.0
Number of mesh routers 16
Number of mesh clients 48
Total iterations 800
Iteration per phase 4
Number of particle-patterns 9
Radius of a mesh router 2.0
Fitness function weight-coefficients (α, β) 0.7, 0.3
Replacement method CM, RIWM

We show the simulation results in Figs. 5 and 6. For SGC, both replacement
methods reach the maximum (100%). This means that all mesh routers are
connected to each other. We see that RIWM converges faster than CM for SGC.
 A Comparison Study of CM and RIWM Router Replacement Methods 79

Fig. 5. Simulation results of WMN-PSOHC for SGC.

Fig. 6. Simulation results of WMN-PSOHC for NCMC.

Also, for the NCMC, RIWM performs better than CM. Therefore, we conclude
that the performance of RIWM is better than CM.

5 Conclusions

In this work, we evaluated the performance of CM and RIWM router replace-

ment methods for WMNs by WMN-PSOHC hybrid intelligent simulation sys-
tem. Simulation results show that the performance of CM is better compared
with RIWM.
In our future work, we would like to evaluate the performance of the proposed
system for different parameters and scenarios.

References
1. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Comput.
Netw. 47(4), 445–487 (2005)
2. Barolli, A., Sakamoto, S., Barolli, L., Takizawa, M.: A hybrid simulation sys-
tem based on particle swarm optimization and distributed genetic algorithm for
wmns: performance evaluation considering normal and uniform distribution of
mesh clients. In: International Conference on Network-Based Information Systems,
pp. 42–55. Springer (2018)
80 S. Sakamoto et al.

3. Barolli, A., Sakamoto, S., Barolli, L., Takizawa, M.: Performance analysis of simula-
tion system based on particle swarm optimization and distributed genetic algorithm
for WMNs considering different distributions of mesh clients. In: International Con-
ference on Innovative Mobile and Internet Services in Ubiquitous Computing, pp
32–45. Springer (2018)
4. Barolli, A., Sakamoto, S., Barolli, L., Takizawa, M.: Performance evaluation of
WMN-PSODGA system for node placement problem in WMNs considering four
different crossover methods. In: The 32nd IEEE International Conference on
Advanced Information Networking and Applications (AINA-2018), pp 850–857.
IEEE (2018)
5. Barolli, A., Sakamoto, S., Durresi, H., Ohara, S., Barolli, L., Takizawa, M.: A
Comparison Study of Constriction and Linearly Decreasing Vmax Replacement
Methods for Wireless Mesh Networks by WMN-PSOHC-DGA Simulation System.
In: International Conference on P2P, pp. 26–34. Parallel, Grid, Cloud and Internet
Computing, Springer (2019)
6. Clerc, M., Kennedy, J.: The particle swarm-explosion, stability, and convergence in
a multidimensional complex space. IEEE Trans. Evol. Comput. 6(1), 58–73 (2002)
7. Matsuo, K., Sakamoto, S., Oda, T., Barolli, A., Ikeda, M., Barolli, L.: Performance
analysis of WMNs by WMN-GA simulation system for two WMN architectures
and different TCP congestion-avoidance algorithms and client distributions. Int.
J. Commun. Networks Distrib. Syst. 20(3), 335–351 (2018)
8. Ohara, S., Barolli, A., Sakamoto, S., Barolli, L.: Performance analysis of WMNs
by WMN-PSODGA simulation system considering load balancing and client uni-
form distribution. In: International Conference on Innovative Mobile and Internet
Services in Ubiquitous Computing, pp. 25–38. Springer (2019)
9. Ozera, K., Bylykbashi, K., Liu, Y., Barolli, L.: A fuzzy-based approach for cluster
management in VANETs: performance evaluation for two fuzzy-based systems.
Internet Things 3, 120–133 (2018)
10. Ozera, K., Inaba, T., Bylykbashi, K., Sakamoto, S., Ikeda, M., Barolli, L.: A WLAN
Triage Testbed based on fuzzy logic and its performance evaluation for different
number of clients and throughput parameter. Int. J. Grid Util. Comput. 10(2),
168–178 (2019)
11. Poli, R., Kennedy, J., Blackwell, T.: Particle swarm optimization. Swarm Intell.
1(1), 33–57 (2007)
12. Sakamoto, S., Lala, A., Oda, T., Kolici, V., Barolli, L., Xhafa, F.: Analysis of
WMN-HC simulation system data using friedman test. In: The Ninth International
Conference on Complex, Intelligent, and Software Intensive Systems (CISIS-2015),
pp 254–259. IEEE (2015)
13. Sakamoto, S., Oda, T., Ikeda, M., Barolli, L., Xhafa, F.: Implementation and eval-
uation of a simulation system based on particle swarm optimisation for node place-
ment problem in wireless mesh networks. Int. J. Commun. Networks Distrib. Syst.
17(1), 1–13 (2016)
14. Sakamoto, S., Ozera, K., Ikeda, M., Barolli, L.: Implementation of intelligent hybrid
systems for node placement problem in WMNs considering particle swarm opti-
mization, hill climbing and simulated annealing. Mob. Networks Appl. 23(1), 27–33
(2018)
15. Sakamoto, S., Barolli, A., Barolli, L., Okamoto, S.: Implementation of a web inter-
face for hybrid intelligent systems. Int. J. Web Inf. Syst. 15(4), 420–431 (2019)
16. Sakamoto, S., Barolli, L., Okamoto, S.: WMN-PSOSA: an intelligent hybrid simu-
lation system for WMNs and its performance evaluations. Int. J. Web Grid Serv.
15(4), 353–366 (2019)
 A Comparison Study of CM and RIWM Router Replacement Methods 81

17. Sakamoto, S., Ohara, S., Barolli, L., Okamoto, S.: Performance evaluation of
WMNs by WMN-PSOHC system considering random inertia weight and linearly
decreasing vmax replacement methods. In: International Conference on Network-
Based Information Systems, pp. 27–36. Springer (2019)
18. Schutte, J.F., Groenwold, A.A.: A study of global optimization using particle
swarms. J. Global Optim. 31(1), 93–108 (2005)
19. Shi, Y.: Particle swarm optimization. IEEE Connections 2(1), 8–13 (2004)
20. Shi, Y., Eberhart, R.C.: Parameter selection in particle swarm optimization. Evo-
lutionary programming VII, pp. 591–600 (1998)
21. Wang, J., Xie, B., Cai, K., Agrawal, D.P.: Efficient mesh router placement in wire-
less mesh networks. In: Proceedings of IEEE Internatonal Conference on Mobile
Adhoc and Sensor Systems (MASS-2007), pp. 1–9 (2007)
22. Xhafa, F., Sanchez, C., Barolli, L.: Ad hoc and neighborhood search methods for
placement of mesh routers in wireless mesh networks. In: Proceedings of 29th IEEE
International Conference on Distributed Computing Systems Workshops (ICDCS-
2009) pp. 400–405 (2009)
 Performance Evaluation of RIWM
and LDVM Router Replacement Methods
for WMNs by WMN-PSOSA-DGA
Considering Chi-Square Distribution

Admir Barolli1(B) , Shinji Sakamoto2 , Phudit Ampririt3 , Leonard Barolli4 ,

and Makoto Takizawa5
1
Department of Information Technology, Aleksander Moisiu University of Durres,
L.1, Rruga e Currilave, Durres, Albania
2
Department of Computer and Information Science, Seikei University,
3-3-1 Kichijoji-Kitamachi, Musashino-shi, Tokyo 180-8633, Japan
3
Graduate School of Engineering, Fukuoka Institute of Technology,
3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan
4
Department of Information and Communication Engineering,
Fukuoka Institute of Technology, 3-30-1 Wajiro-Higashi, Higashi-Ku,
Fukuoka 811-0295, Japan
[email protected]
5
Department of Advanced Sciences, Faculty of Science and Engineering,
Hosei University, 3-7-2 Kajino-Cho, Koganei-Shi, Tokyo 184-8584, Japan
[email protected]

Abstract. Wireless Mesh Networks (WMNs) are gaining a lot of atten-

tion from researchers due to their advantages such as easy maintenance,
low upfront cost, and high robustness. Connectivity and stability directly
affect the performance of WMNs. However, WMNs have some problems
such as node placement problem, hidden terminal problem and so on.
In our previous work, we implemented a simulation system to solve the
node placement problem in WMNs considering Particle Swarm Opti-
mization (PSO), Simulated Annealing (SA) and Distributed Genetic
Algorithm (DGA), called WMN-PSOSA-DGA. In this paper, we eval-
uate the performance of Random Inertia Weight Method (RIWM) and
Linearly Decreasing Vmax Method (LDVM) for WMNs using WMN-
PSOSA-DGA hybrid simulation system considering Chi-square distribu-
tion of mesh clients. Simulation results show that a good performance is
achieved for LDVM compared with the case of RIWM.

1 Introduction
The wireless networks and devices are becoming increasingly popular and they
provide users access to information and communication anytime and any-
where [2,11,13]. Wireless Mesh Networks (WMNs) are gaining a lot of attention
because of their low cost nature that makes them attractive for providing wireless
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 82–91, 2021.
https://doi.org/10.1007/978-3-030-75100-5_8
Performance Evaluation of RIWM and LDVM Router Replacement Methods 83

Internet connectivity. A WMN is dynamically self-organized and self-configured,

with the nodes in the network automatically establishing and maintaining mesh
connectivity among them-selves (creating, in effect, an ad hoc network). This
feature brings many advantages to WMNs such as low up-front cost, easy net-
work maintenance, robustness and reliable service coverage [1].
Mesh node placement in WMN can be seen as a family of problems, which
are shown to be computationally hard to solve for most of the formulations [21].
We consider the version of the mesh router nodes placement problem in which
we are given a grid area where to deploy a number of mesh router nodes and a
number of mesh client nodes of fixed positions (of an arbitrary distribution) in
the grid area. The objective is to find a location assignment for the mesh routers
to the cells of the grid area that maximizes the network connectivity and client
coverage. Network connectivity is measured by Size of Giant Component (SGC)
of the resulting WMN graph, while the user coverage is simply the number
of mesh client nodes that fall within the radio coverage of at least one mesh
router node and is measured by Number of Covered Mesh Clients (NCMC).
Node placement problems are known to be computationally hard to solve [9,22].
In previous works, some intelligent algorithms have been investigated for node
placement problem [4,12].
In [15], we implemented a Particle Swarm Optimization (PSO) and Sim-
ulated Annealing (SA) based simulation system, called WMN-PSOSA. Also,
we implemented another simulation system based on Genetic Algorithm (GA),
called WMN-GA [4,10], for solving node placement problem in WMNs. Then,
we designed a hybrid intelligent system based on PSO, SA and DGA, called
WMN-PSOSA-DGA [14].
In this paper, we evaluate the performance of Random Inertia Weight Method
(RIWM) and Linearly Decreasing Vmax Method (LDVM) for WMNs using
WMN-PSOSA-DGA simulation system considering Chi-square distribution of
mesh clients.
The rest of the paper is organized as follows. We present our designed and
implemented hybrid simulation system in Sect. 2. The simulation results are
given in Sect. 3. Finally, we give conclusions and future work in Sect. 4.

2 Proposed and Implemented Simulation System

Distributed Genetic Algorithms (DGAs) are capable of producing solutions with
higher efficiency (in terms of time) and efficacy (in terms of better quality solu-
tions). They have shown their usefulness for the resolution of many computa-
tionally hard combinatorial optimization problems. Also, Particle Swarm Opti-
mization (PSO) and Simulated Annealing (SA) are suitable for solving NP-hard
problems.

2.1 Velocities and Positions of Particles

WMN-PSOSA-DGA decides the velocity of particles by a random process con-
sidering the area size. For instance, when the area size is W × H, the velocity
84 A. Barolli et al.
√ √
is decided randomly from − W 2 + H 2 to W 2 + H 2 . Each particle’s velocities
are updated by simple rule.
For SA mechanism, next positions of each particle are used for neighbor
solution s . The fitness function f gives points to the current solution s. If f (s )
is larger than f (s), the s is better than s so the s is updated to s . However, if
f (s) is not larger
 than f (s), the s may be updated by using the probability of

exp f (s )−f
T
(s)
. Where T is called the “Temperature value” which is decreased
with the computation so that the probability to update will be decreased. This
mechanism of SA is called a cooling schedule and the next Temperature value of
computation is calculated as Tn+1 = α × Tn . In this paper, we set the starting
temperature, ending temperature and number of iterations. We calculate α as
 1.0/number of iterations
SA ending temperature
α= .
SA starting temperature

It should be noted that the positions are not updated but the velocities are
updated in the case when the solusion s is not updated.

2.2 Routers Replacement Methods

A mesh router has x, y positions and velocity. Mesh routers are moved based on
velocities. There are many router replacement methods. In this paper, we use
RIWM and LDVM.

Constriction Method (CM).

CM is a method which PSO parameters are set to a week stable region (ω =
0.729, C1 = C2 = 1.4955) based on analysis of PSO by M. Clerc et al. [5,8,17].
Random Inertia Weight Method (RIWM).
In RIWM, the ω parameter is changing randomly from 0.5 to 1.0. The C1
and C2 are kept 2.0. The ω can be estimated by the week stable region. The
average of ω is 0.75 [7,19].
Linearly Decreasing Inertia Weight Method (LDIWM)
In LDIWM, C1 and C2 are set to 2.0, constantly. On the other hand, the ω
parameter is changed linearly from unstable region (ω = 0.9) to stable region
(ω = 0.4) with increasing of iterations of computations [6,20].
Linearly Decreasing Vmax Method (LDVM)
In LDVM, PSO parameters are set to unstable region (ω = 0.9, C1 = C2 =
2.0). A value of Vmax which is maximum velocity of particles is considered.
With increasing of iteration of computations, the Vmax is kept decreasing
linearly [7,16,18].
Rational Decrement of Vmax Method (RDVM)
In RDVM, PSO parameters are set to unstable region (ω = 0.9, C1 = C2 =
2.0). The Vmax is kept decreasing with the increasing of iterations as
 T −x
Vmax (x) = W 2 + H2 × .
x
Performance Evaluation of RIWM and LDVM Router Replacement Methods 85

Where, W and H are the width and the height of the considered area, respec-
tively. Also, T and x are the total number of iterations and a current number
of iteration, respectively.

2.3 DGA Operations

Population of individuals: Unlike local search techniques that construct a path

in the solution space jumping from one solution to another one through local
perturbations, DGA use a population of individuals giving thus the search a
larger scope and chances to find better solutions. This feature is also known
as “exploration” process in difference to “exploitation” process of local search
methods.
Selection: The selection of individuals to be crossed is another important
aspect in DGA as it impacts on the convergence of the algorithm. Several selec-
tion schemes have been proposed in the literature for selection operators trying
to cope with premature convergence of DGA. There are many selection methods
in GA. In our system, we implement 2 selection methods: Random method and
Roulette wheel method.
Crossover operators: Use of crossover operators is one of the most impor-
tant characteristics. Crossover operator is the means of DGA to transmit best
genetic features of parents to offsprings during generations of the evolution pro-
cess. Many methods for crossover operators have been proposed such as Blend
Crossover (BLX-α), Unimodal Normal Distribution Crossover (UNDX), Simplex
Crossover (SPX).
Mutation operators: These operators intend to improve the individuals of a
population by small local perturbations. They aim to provide a component of
randomness in the neighborhood of the individuals of the population. In our
system, we implemented two mutation methods: uniformly random mutation
and boundary mutation.
Escaping from local optimal: GA itself has the ability to avoid falling prema-
turely into local optimal and can eventually escape from them during the search
process. DGA has one more mechanism to escape from local optimal by consid-
ering some islands. Each island computes GA for optimizing and they migrate
its gene to provide the ability to avoid from local optimal.
Convergence: The convergence of the algorithm is the mechanism of DGA
to reach to good solutions. A premature convergence of the algorithm would
cause that all individuals of the population be similar in their genetic features
and thus the search would result ineffective and the algorithm getting stuck
into local optimal. Maintaining the diversity of the population is therefore very
important to this family of evolutionary algorithms.
In following, we present fitness function, migration function, particle pattern
and gene coding.
86 A. Barolli et al.

2.4 Fitness and Migration Functions

The determination of an appropriate fitness function, together with the chromo-
some encoding are crucial to the performance. Therefore, one of most important
thing is to decide the determination of an appropriate objective function and its
encoding. In our case, each particle-pattern and gene has an own fitness value
which is comparable and compares it with other fitness value in order to share
information of global solution. The fitness function follows a hierarchical app-
roach in which the main objective is to maximize the SGC in WMN. Thus, the
fitness function of this scenario is defined as

Fitness = 0.7 × SGC(xij , y ij ) + 0.3 × NCMC(xij , y ij ).

Our implemented simulation system uses Migration function as shown in

Fig. 1. The Migration function swaps solutions between PSOSA part and DGA
part.

Fig. 1. Model of WMN-PSOSA-DGA migration.

2.5 Particle-Pattern and Gene Coding

In order to swap solutions, we design particle-patterns and gene coding carefully.
A particle is a mesh router. Each particle has position in the considered area
and velocities. A fitness value of a particle-pattern is computed by combination
of mesh routers and mesh clients positions. In other words, each particle-pattern
is a solution as shown is Fig. 2.

Fig. 2. Relationship among global solution, particle-patterns and mesh routers in

PSOSA part.
Performance Evaluation of RIWM and LDVM Router Replacement Methods 87

A gene describes a WMN. Each individual has its own combination of mesh
nodes. In other words, each individual has a fitness value. Therefore, the combi-
nation of mesh nodes is a solution.

3 Simulation Results
In this section, we show simulation results. In this work, we analyze the perfor-
mance of RIWM and LDVM router replacement methods for WMNs by WMN-
PSOSA-DGA hybrid intelligent simulation system considering Chi-square client
distribution as shown in Fig. 3 [3].

Fig. 3. Chi-square distribution.

We carried out the simulations 10 times in order to avoid the effect of ran-
domness and create a general view of results. We show the parameter setting for
WMN-PSOSA-DGA in Table 1.
We show simulation results in Figs. 4 and 5. We see that for SGC, the perfor-
mance is almost the same for both replacement method. However, for NCMC,
LDVM performs better than RIWM. The visualized simulation results are shown
in Fig. 6. For both replacement methods, all mesh routers are connected, but
some clients are not covered. We can see the number of covered mesh clients is
larger for LDVM compared with the case of RIWM.
88 A. Barolli et al.

Table 1. WMN-PSOSA-DGA parameters.

Parameters Values
Clients distribution Chi-Square
Area size 32.0 × 32.0
Number of mesh routers 16
Number of mesh clients 48
Number of GA islands 16
Number of Particle-patterns 32
Number of migrations 200
Evolution steps 320
Radius of a mesh router 2.0 − 3.5
Selection method Roulette wheel method
Crossover method SPX
Mutation method Boundary mutation
Crossover rate 0.8
Mutation rate 0.2
SA Starting value 10.0
SA Ending value 0.01
Total number of iterations 64000
Replacement method RIWM, LDVM

Fig. 4. Simulation results of WMN-PSOSA-DGA for SGC.

Fig. 5. Simulation results of WMN-PSOSA-DGA for NCMC.

Performance Evaluation of RIWM and LDVM Router Replacement Methods 89

Fig. 6. Visualized simulation results of WMN-PSOSA-DGA for different replacement

methods.

4 Conclusions
In this work, we evaluated the performance of RIWM and LDVM replacement
methods for WMNs using a hybrid simulation system based on PSO, SA and
DGA (called WMN-PSOSA-DGA) considering Chi-square distribution of mesh
clients. Simulation results show that a good performance was achieved for LDVM
compared with the case of RIWM.
In our future work, we would like to evaluate the performance of the proposed
system for different parameters and patterns.

References
1. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Comput.
Networks 47(4), 445–487 (2005)
2. Barolli, A., Sakamoto, S., Barolli, L., Takizawa, M.: A hybrid simulation sys-
tem based on particle swarm optimization and distributed genetic algorithm for
wmns: performance evaluation considering normal and uniform distribution of
mesh clients. In: International Conference on Network-Based Information Systems,
pp. 42–55. Springer (2018)
3. Barolli, A., Sakamoto, S., Barolli, L., Takizawa, M.: Performance analysis of simula-
tion system based on particle swarm optimization and distributed genetic algorithm
for WMNs considering different distributions of mesh clients. In: International Con-
ference on Innovative Mobile and Internet Services in Ubiquitous Computing, pp
32–45. Springer (2018)
4. Barolli, A., Sakamoto, S., Ozera, K., Barolli, L., Kulla, E., Takizawa, M.: Design
and Implementation of a Hybrid Intelligent System Based on Particle Swarm
Optimization and Distributed Genetic Algorithm. In: International Conference on
Emerging Internetworking, pp. 79–93. Springer, Data & Web Technologies (2018)
5. Barolli, A., Sakamoto, S., Durresi, H., Ohara, S., Barolli, L., Takizawa, M.: A com-
parison study of constriction and linearly decreasing vmax replacement methods
for wireless mesh networks by WMN-PSOHC-DGA simulation system. In: Inter-
national Conference on P2P Parallel, Grid, Cloud and Internet Computing, pp.
26–34. Springer (2019)
90 A. Barolli et al.

6. Barolli, A., Sakamoto, S., Ohara, S., Barolli, L., Takizawa, M.: Performance anal-
ysis of WMNs by WMN-PSOHC-DGA simulation system considering linearly
decreasing inertia weight and linearly decreasing vmax replacement methods. In:
International Conference on Intelligent Networking and Collaborative Systems, pp.
14–23. Springer (2019)
7. Barolli, A., Sakamoto, S., Ohara, S., Barolli, L., Takizawa, M.: Performance analy-
sis of WMNs by WMN-PSOHC-DGA simulation system considering random iner-
tia weight and linearly decreasing vmax router replacement methods. In: Confer-
ence on Complex, Intelligent, and Software Intensive Systems, pp. 13–21. Springer
(2019)
8. Clerc, M., Kennedy, J.: The particle swarm-explosion, stability, and convergence in
a multidimensional complex space. IEEE Trans. Evol. Comput. 6(1), 58–73 (2002)
9. Maolin, T., et al.: Gateways placement in backbone wireless mesh networks. Int.
J. Commun. Network Syst. Sci. 2(1), 44 (2009)
10. Matsuo, K., Sakamoto, S., Oda, T., Barolli, A., Ikeda, M., Barolli, L.: Performance
analysis of WMNs by WMN-GA simulation system for two WMN architectures
and different TCP congestion-avoidance algorithms and client distributions. Int.
J. Commun. Networks Distributed Syst. 20(3), 335–351 (2018)
11. Ohara, S., Barolli, A., Sakamoto, S., Barolli, L.: Performance analysis of WMNs
by WMN-PSODGA simulation system considering load balancing and client uni-
form distribution. In: International Conference on Innovative Mobile and Internet
Services in Ubiquitous Computing, pp. 25–38. Springer (2019)
12. Ozera, K., Sakamoto, S., Elmazi, D., Bylykbashi, K., Ikeda, M., Barolli, L.: A fuzzy
approach for clustering in manets: performance evaluation for different parameters.
Int. J. Space-Based Situated Comput. 7(3), 166–176 (2017)
13. Ozera, K., Inaba, T., Bylykbashi, K., Sakamoto, S., Ikeda, M., Barolli, L.: A WLAN
triage testbed based on fuzzy logic and its performance evaluation for different
number of clients and throughput parameter. Int. J. Grid Utility Comput. 10(2),
168–178 (2019)
14. Sakamoto, S., Barolli, A., Barolli, L., Takizawa, M.: Design and implementation of
a hybrid intelligent system based on particle swarm optimization, hill climbing and
distributed genetic algorithm for node placement problem in WMNs: a compari-
son study. In: The 32nd IEEE International Conference on Advanced Information
Networking and Applications (AINA-2018), pp 678–685. IEEE (2018)
15. Sakamoto, S., Ozera, K., Ikeda, M., Barolli, L.: Implementation of intelligent hybrid
systems for node placement problem in WMNs considering particle swarm opti-
mization, hill climbing and simulated annealing. Mob. Networks Appl. 23(1), 27–33
(2018)
16. Sakamoto, S., Ohara, S., Barolli, L., Okamoto, S.: Performance Evaluation of
WMNs by WMN-PSOHC system considering random inertia weight and linearly
decreasing vmax replacement methods. In: International Conference on Network-
Based Information Systems, pp. 27–36. Springer (2019)
17. Sakamoto, S., Ohara, S., Barolli, L., Okamoto, S.: Performance Evaluation of
WMNs WMN-PSOHC System Considering Constriction and Linearly Decreasing
Inertia Weight Replacement Methods. In: International Conference on Broadband
and Wireless Computing Communication and Applications, pp. 22–31. Springer
(2019)
18. Schutte, J.F., Groenwold, A.A.: A study of global optimization using particle
swarms. J. Global Optim. 31(1), 93–108 (2005)
19. Shi, Y.: Particle swarm optimization. IEEE Connections 2(1), 8–13 (2004)
Performance Evaluation of RIWM and LDVM Router Replacement Methods 91

20. Shi, Y., Eberhart, R.C.: Parameter Selection in Particle Swarm Optimization. Evo-
lutionary programming VII, pp. 591–600 (1998)
21. Vanhatupa, T., Hannikainen, M., Hamalainen, T.: Genetic algorithm to optimize
node placement and configuration for WLAN panning. In: The 4th IEEE Interna-
tional Symposium on Wireless Communication Systems, pp. 612–616 (2007)
22. Wang, J., Xie, B., Cai, K., Agrawal, D.P.: Efficient mesh router placement in wire-
less mesh networks. In: Proceedings of IEEE International Conference on Mobile
Adhoc and Sensor Systems (MASS-2007), pp 1–9
 An Adaptive Anti-packet Recovery
Method for Vehicular DTN Considering
Message Possession Rate

Shota Uchimura1 , Masaya Azuma1 , Yoshiki Tada1 , Makoto Ikeda2(B) ,

and Leonard Barolli2
1
Graduate School of Engineering, Fukuoka Institute of Technology,
3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan
[email protected]
2
Department of Information and Communication Engineering,
Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka
811-0295, Japan
[email protected]

Abstract. In this paper, we propose an Adaptive Anti-packet Recovery

(AAR) method based on epidemic protocol. Our method can be used
in Vehicular Delay Tolerant Networking (DTN). We use conventional
Epidemic with recovery function and the proposed AAR method as the
message delivery protocols. From the simulation results, we observed that
the delivery ratio of proposed AAR is good for dense networks.

Keywords: DTN · Adaptive anti-packet · Recovery method ·

Epidemic

1 Introduction

During recent years, with the development of the vehicle and information indus-
tries, we see many different services via the Internet. Therefore, it is important
to build a strong network that is resilient to disasters. The network should have
a higher QoE, which means that it can quickly switch to an alternate link in
case of a communication disconnection or long delay [10]. This research focuses
on vehicular networks. We consider the vehicle as an actor in different network
situations. Delay-/Disruption-/Disconnection-Tolerant Networking (DTN) tech-
nology [5] has attracted much attention as a vehicular-based communication
method. However, due to the overhead, the storage resources are a problem for
the conventional DTN approach. Therefore, hybrid and recovery methods have
been proposed to address these problems [4,6,11].
In [4], the authors have proposed a hybrid DTN routing method which con-
siders the Epidemic-based routing method with many replications and the SpW-
based routing method with few replications. They consider storage state of the
vehicles to choice the routing method.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 92–101, 2021.
https://doi.org/10.1007/978-3-030-75100-5_9
 An Adaptive Anti-packet Recovery Method for Vehicular DTN 93

In [11], we have evaluated a message relaying method with Enhanced

Dynamic Timer (EDT) considering disaster situations. The EDT is one of the
recovery method for Epidemic-based routing method [6].
In this paper, we propose an Adaptive Anti-packet Recovery (AAR) method
for Vehicular DTN. For evaluation, we use Epidemic with conventional anti-
packet and proposed AAR method as the message delivery protocols.
The structure of the paper is as follows. In Sect. 2, we give the message
relaying methods for Vehicular DTN. In Sect. 3 is described the adaptive anti-
packet recovery method for Vehicular DTN. In Sect. 4, we provide the evaluation
system and the simulation results. Finally, conclusions and future work are given
in Sect. 5.

2 Message Relaying Methods for Vehicular DTN

DTN can provide a reliable internet-working for space tasks [3,8,13]. The space
networks have possibly large latency, frequent link disruption and frequent dis-
connection. In Vehicular DTN, the intermediate vehicles stored messages in their
storage and then transmit to others. The network architecture is specified in RFC
4838 [2].
The famous DTN protocol is Epidemic routing [7,12]. Epidemic routing is
performed by using two control messages to replicate a bundle message. Each
vehicle periodically broadcasts a Summary Vector (SV) in the network. The
SV contains a list of stored messages of each vehicle. When the vehicles receive
the SV, they compare received SV to their SV. The vehicle sends a REQUEST
message if the received SV contains an unknown message.
In Epidemic routing, consumption of network resources and storage sate
become a critical problem, because the vehicles replicate messages to neighbors
in their communication range. Therefore, received messages remain in the stor-
age and the messages are continuously replicated even if the end-point receives
the messages. However, recovery schemes such as timer may delete the replicate
messages in the network. Then, the anti-packet deletes replicate messages too
late.
In the case of the conventional timer, messages have a lifetime. The messages
are punctually deleted when the lifetime of the messages is expired. However,
the setting of a suitable lifetime is difficult. In the case of the conventional
anti-packet, the end-point broadcasts the anti-packet, which contains the list
of messages that are received by the end-point. Vehicles delete the messages
according to the anti-packet. Then, the vehicles replicate the anti-packet to other
vehicles. However, network resources are affected by anti-packet.

3 Adaptive Anti-packet Recovery Method

In this section, we describe the proposed AAR method for Vehicular DTN. For
the conventional anti-packet recovery method, end-point broadcasts the anti-
packet, which contains the list of messages that are received by the end-point.
94 S. Uchimura et al.

Then, the vehicles replicate the anti-packet to other vehicles. In our proposed
AAR method, intermediate vehicles generate the anti-packet even if the message
does not reach the end-point. The flowchart of the proposed AAR method is
shown in Fig. 1.

Fig. 1. Flowchart of proposed message relaying method with AAR.

In our approach, after the vehicle receives a SV, the vehicle counts the number
of received SVs and measures the number of neighbors for each message. The
AAR method considers the number of neighboring vehicles (Count Node) and
calculates the number of Count Same Messages by comparing the Threshold.
The AAR method selects the recovery mode by the condition in Eq. (1):

Count Same MESSAGE

≥ Treshould. (1)
Count Node
If the Eq. (1) is satisfied, an anti-packet is generated. On the other hand, if it is
not satisfied, the message is stored in storage. In addition, the vehicle will also
replicate the anti-packet continually as the conventional Epidemic.
 An Adaptive Anti-packet Recovery Method for Vehicular DTN 95

4 Evaluation of Proposed Method

In this section, we evaluate the AAR method in Vehicular DTN. We implemented
the proposed AAR method on the Scenargie simulator [9].

4.1 Evaluation Setting

We consider a grid road model for two types of scenarios (see Fig. 2). The
red square in Fig. 2 indicates a vehicle. This figure shows an example of 50
vehicles/km2 . The following two cases are prepared for this evaluation.
• Case A: The message starting-point and end-point are fixed.
• Case B: Both starting-point and end-point are mobile.
• Common: All vehicles continue to move on the roads based on the map-based
random way-point mobility model.

Fig. 2. Grid road model (50 vehicles).

For this evaluation, we consider four vehicular densities with 50, 100, 150
and 200 vehicles/km2 . Table 1 shows the simulation parameters used for the
network simulator. For both scenarios, the start-point replicates 40 kinds of
bundle messages to the relay vehicles and deliver the bundle message to the
end-point. The simulation time is 600 s. The ITU-R P.1411 propagation model
is used in this simulation [1]. We consider the interference from obstacles at
5.9 GHz radio frequency.
96 S. Uchimura et al.

We evaluate the performance of delivery ratio, overhead and delay for differ-
ent densities of vehicles. The delivery ratio indicates the value of the generated
bundle messages divided by the delivered bundle messages to the end-point. The
overhead indicates the number of times for sending replicate bundles. The delay
indicates the transmission delay of the bundle to reach the end-point.

Table 1. Simulation parameters.

Parameter Value
Simulation time (Tmax ) 600 [s]
Area dimensions 1, 000 [m] × 1, 000 [m]
Number of vehicles 50, 100, 150, 200 [vehicles/km2 ]
Minimum speed 8.333 [m/s]
Maximum speed 16.666 [m/s]
Message start and end time 10–400 [sec]
Message generation interval 10 [sec]
AAR: Threshold 0.1–1.0
SV transmission interval 1 [sec]
Message size 500 [bytes]
PHY model IEEE 802.11p
Propagation model ITU-R P.1411
Radio frequency 5.9 GHz
Antenna model Omni-directional

4.2 Simulation Results

We evaluate the proposed AAR method compared to Epidemic with anti-packet

for Case A and Case B. In all cases, we evaluate the network performance by
changing the value of Threshold from 0.1 to 1.0.
We present the simulation results of delivery ratio in Fig. 3. For 50 vehicles,
the delivery ratio of Case B is better than Case A. Regardless of the number
of vehicles, the results of delivery ratio improve with increasing numbers of
thresholds. We have also observed that the message delivery ratio increases faster
with increasing the vehicle densities. In Case A, the difference is large due to the
effect of distance from starting-point to end-point. However, we found that the
difference becomes smaller in Case B. This is because of the anti-packet reached
the relay vehicle before the message reached the end-point.
 An Adaptive Anti-packet Recovery Method for Vehicular DTN 97

0.8

Delivery Rate
0.6

0.4

0.2 50[Vehicles]
100[Vehicles]
150[Vehicles]
200[Vehicles]
0
Th

Th

Th

Th

Th

Th

Th

Th

Th

Th

Ep
re

re

re

re

re

re

re

re

re

re

id
em
sh

sh

sh

sh

sh

sh

sh

sh

sh

sh
ol

ol

ol

ol

ol

ol

ol

ol

ol

ol

ic
d=

d=

d=

d=

d=

d=

d=

d=

d=

d=

+A
0.

0.

0.

0.

0.

0.

0.

0.

0.

1.

nt
1

i-P
ac
ke
t
(a) Case A
1

0.8
Delivery Rate

0.6

0.4

0.2 50[Vehicles]
100[Vehicles]
150[Vehicles]
200[Vehicles]
0
Th

Th

Th

Th

Th

Th

Th

Th

Th

Th

Ep
re

re

re

re

re

re

re

re

re

re

id
em
sh

sh

sh

sh

sh

sh

sh

sh

sh

sh
ol

ol

ol

ol

ol

ol

ol

ol

ol

ol

ic
d=

d=

d=

d=

d=

d=

d=

d=

d=

d=

+A
0.

0.

0.

0.

0.

0.

0.

0.

0.

1.

nt
1

i-P
ac
ke
t

(b) Case B

Fig. 3. Delivery ratio for Case A and Case B.

We present the simulation results of the overhead in Fig. 4. For 50 vehicles, the
overheads of threshold 1.0 are reduced compared to the conventional Epidemic
with anti-packet. For 200 vehicles, the results of overhead for Case A are higher
than Case B. This is due to the high frequency of replication in some relay
vehicles. For both cases with AAR method, we observed that the results of
overhead are decreased with decreasing the threshold.
98 S. Uchimura et al.

50[Vehicles]
100[Vehicles]
150[Vehicles]
200[Vehicles]
7000

6000

5000
Overhead

4000

3000

2000

1000

0
Th

Th

Th

Th

Th

Th

Th

Th

Th

Th

Ep
re

re

re

re

re

re

re

re

re

re

id
em
sh

sh

sh

sh

sh

sh

sh

sh

sh

sh
ol

ol

ol

ol

ol

ol

ol

ol

ol

ol

ic
d=

d=

d=

d=

d=

d=

d=

d=

d=

d=

+A
0.

0.

0.

0.

0.

0.

0.

0.

nt
.2

.0
1

i-P
ac
ke
t
(a) Case A
50[Vehicles]
100[Vehicles]
150[Vehicles]
200[Vehicles]
6000

5000

4000
Overhead

3000

2000

1000

0
Th

Th

Th

Th

Th

Th

Th

Th

Th

Th

Ep
re

re

re

re

re

re

re

re

re

re

id
em
sh

sh

sh

sh

sh

sh

sh

sh

sh

sh
ol

ol

ol

ol

ol

ol

ol

ol

ol

ol

ic
d=

d=

d=

d=

d=

d=

d=

d=

d=

d=

+A
0.

0.

0.

0.

0.

0.

0.

0.

0.

1.

nt
1

i-P
ac
ke
t

(b) Case B

Fig. 4. Overhead for Case A and Case B.

We present the simulation results of delay in Fig. 5. There is little difference

in the delay results compared to the conventional Epidemic with anti-packet.
For Case A, the results of delay for 150 vehicles is the shortest. While for Case
B, the results of delay decrease by increasing the number of vehicles. The delay
of Case B is shorter than Case A. This is due to the starting-point and end-point
are moving, which reduces the duration to reach the message.
 An Adaptive Anti-packet Recovery Method for Vehicular DTN 99

50[Vehicles]
100[Vehicles]
150[Vehicles]
200[Vehicles]
30

25

20
Delay[Sec]

15

10

0
Th

Th

Th

Th

Th

Th

Th

Th

Th

Th

Ep
re

re

re

re

re

re

re

re

re

re

id
em
sh

sh

sh

sh

sh

sh

sh

sh

sh

sh
ol

ol

ol

ol

ol

ol

ol

ol

ol

ol

ic
d=

d=

d=

d=

d=

d=

d=

d=

d=

d=

+A
0.

0.

0.

0.

0.

0.

1.

nt
.2

.5

.9
1

i-P
ac
ke
t
(a) Case A
50[Vehicles]
100[Vehicles]
150[Vehicles]
200[Vehicles]
8

5
Delay[Sec]

0
Th

Th

Th

Th

Th

Th

Th

Th

Th

Th

Ep
re

re

re

re

re

re

re

re

re

re

id
em
sh

sh

sh

sh

sh

sh

sh

sh

sh

sh
ol

ol

ol

ol

ol

ol

ol

ol

ol

ol

ic
d=

d=

d=

d=

d=

d=

d=

d=

d=

d=

+A
0.

0.

0.

0.

0.

0.

0.

0.

0.

1.

nt
1

i-P
ac
ke
t

(b) Case B

Fig. 5. Delay for Case A and Case B.

5 Conclusions

In this paper, we proposed an anti-packet generation method and storage man-

agement method to improve the message delivery rate in vehicle DTN. We eval-
uated the network performance considering the cases when starting-point and
end-point are fixed and mobile. For evaluation, we used Epidemic with conven-
tional anti-packet and AAR method as the message delivery protocols. From the
simulation results, we found that the delivery ratio of AAR is good for dense
networks. For sparse network, some anti-packets reached the relay vehicles before
the messages reached the end-point.
100 S. Uchimura et al.

In future work, we would like to investigate the impact of storage usage and
the other parameters.

Acknowledgments. This work has been partially funded by the research project
from Comprehensive Research Organization at Fukuoka Institute of Technology (FIT),
Japan.

References
1. Rec. ITU-R P.1411-7: Propagation data and prediction methods for the planning
of short-range outdoor radiocommunication systems and radio local area networks
in the frequency range 300 MHz to 100 GHz. ITU (2013)
2. Cerf, V., Burleigh, S., Hooke, A., Torgerson, L., Durst, R., Scott, K., Fall, K., Weiss,
H.: Delay-tolerant networking architecture. IETF RFC 4838 (Informational), April
2007
3. Fall, K.: A delay-tolerant network architecture for challenged Internets. In: Pro-
ceedings of the International Conference on Applications, Technologies, Architec-
tures, and Protocols for Computer Communications, SIGCOMM 2003, pp. 27–34
(2003)
4. Henmi, K., Koyama, A.: Hybrid type DTN routing protocol considering storage
capacity. In: Proceedings of the 8th International Conference on Emerging Internet,
Data and Web Technologies (EIDWT 2020). pp. 491–502, February 2020
5. Kawabata, N., Yamasaki, Y., Ohsaki, H.: Hybrid cellular-DTN for vehicle volume
data collection in rural areas. In: Proceedings of the IEEE 43rd Annual Computer
Software and Applications Conference (COMPSAC-2019), vol. 2, pp. 276–284, July
2019
6. Nakasaki, S., Ikeda, M., Barolli, L.: A message relaying method with enhanced
dynamic timer considering decrease rate of neighboring nodes for Vehicular-DTN.
In: Proceedings of the 14th International Conference on Broad-Band Wireless Com-
puting, Communication and Applications (BWCCA-2019), pp. 711–720, November
2019
7. Ramanathan, R., Hansen, R., Basu, P., Hain, R.R., Krishnan, R.: Prioritized epi-
demic routing for opportunistic networks. In: Proceedings of the 1st International
MobiSys Workshop on Mobile Opportunistic Networking (MobiOpp 2007), pp.
62–66 (2007)
8. Rüsch, S., Schürmann, D., Kapitza, R., Wolf, L.: Forward secure delay-tolerant
networking. In: Proceedings of the 12th Workshop on Challenged Networks
(CHANTS-2017), pp. 7–12, October 2017
9. Scenargie: Space-time engineering, LLC http://www.spacetime-eng.com/
10. Solpico, D., Tan, M.I., Manalansan, E.J., Zagala, F.A., Leceta, J.A., Lanuza, D.F.,
Bernal, J., Ramos, R.D., Villareal, R.J., Cruz, X.M., dela Cruz, J.A., Lagazo,
D.J., Honrado, J.L., Abrajano, G., Libatique, N.J., Tangonan, G.: Application
of the V-HUB standard using LoRa beacons, mobile cloud, UAVs, and DTN for
disaster-resilient communications. In: Proceedings of the IEEE Global Humanitar-
ian Technology Conference (GHTC-2019), pp. 1–8, October 2019
11. Tada, Y., Ikeda, M., Barolli, L.: Performance evaluation of a message relaying
method for resilient disaster networks. In: Proceedings of the 15th International
Conference on Broadband and Wireless Computing, Communication and Applica-
tions (BWCCA-2020), pp. 1–10, October 2020
 An Adaptive Anti-packet Recovery Method for Vehicular DTN 101

12. Vahdat, A., Becker, D.: Epidemic routing for partially-connected ad hoc networks.
Duke University, Technical report (2000)
13. Wyatt, J., Burleigh, S., Jones, R., Torgerson, L., Wissler, S.: Disruption toler-
ant networking flight validation experiment on NASA’s EPOXI mission. In: Pro-
ceedings of the 1st International Conference on Advances in Satellite and Space
Communications (SPACOMM-2009), pp. 187–196, July 2009
 From Incomplete to Complete Networks
in Asynchronous Byzantine Systems

Michel Raynal1,2(B) and Jiannong Cao2

1
Univ Rennes IRISA, CNRS, Inria, Rennes, France
[email protected]
2
Department of Computing, Polytechnic University, Hung Hom, Hong Kong

Abstract. This paper presents a simple broadcast operation suited to

n-process asynchronous message-passing systems in which (i) up to t
processes may commit Byzantine faults, and (2), while the underlying
communication network is connected (any pair of processes is connected
by a path), not all the pairs of processes are directly connected by a com-
munication channel. The algorithm proposed to implement this operation
assumes (i) t < n/3 and (ii) (2t+1)-vertex connectivity of the underlying
network (each pair of processes is connected by at least (2t + 1) disjoint
paths), requirements which are shown to be necessary. When considering
incomplete networks, this abstraction can be used as the first level of a
software stack on top of which, without any modifications, Byzantine-
tolerant broadcast and agreement abstractions designed for fully con-
nected networks can directly be used. The paper has also a short survey
flavor from a “failure-tolerant broadcast abstractions” point of view.

Keywords: Asynchronous system · Broadcast abstraction · Byzantine

process · Distributed software engineering · Fault-tolerance ·
Incomplete network · k-vertex connectivity · Message-passing ·
Modularity · Simplicity

1 Preliminary Definitions

Asynchronous Processes. The system is made up of a finite set Π of n > 1

asynchronous sequential processes, namely Π = {p1 , . . . , pn }. “Asynchronous”
means that each process proceeds at its own speed, which can vary arbitrarily and
always remains unknown to the other processes. The index i of pi is considered
to be the identity of pi , and when there no ambiguity pi and i are regarded as
synonymous.

Communication Network (Graph). From a structural point of view, the sys-

tem can be represented by an undirected graph whose vertices are the pro-
cesses and whose edges are the process-to-process communication channels. From
an operational point of view, the processes communicate by exchanging mes-
sages through an asynchronous network made up of bi-directional channels,
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 102–112, 2021.
https://doi.org/10.1007/978-3-030-75100-5_10
From Incomplete to Complete Networks in Asynchronous Byzantine Systems 103

each channel connecting two processes. The channels are asynchronous and reli-
able. “Asynchronous” means that each message that has been sent is eventually
received by its destination process, but there is no bound on message transfer
delays. “Reliable” means that a channel neither loose, duplicate, modify, nor
create messages.
The network is connected, i.e., there is a path connecting any pair of distinct
processes. It can be complete or incomplete. “Complete” (or fully connected)
means that each pair of processes is connected by a channel. “Incomplete” means
that not all the pairs for processes are connected by a channel.
A process pi sends a message to a process pj by invoking the primitive “send
tag(m) to pj ”, where pj is a neighbor of pi , tag is the type of the message and m
its content. A process receives a message by executing the primitive “receive()”.
If there is a channel connecting the processes pi and pj , we assume that if
pi receives a message on this channel, it knows the message is from pj . Said
differently, no process can impersonate another process.
An incomplete network (not fully connected graph) is “x-vertex connected”
if it has no set of (x − 1) vertices (processes) whose removal disconnects it. Due
to Menger’s Theorem [15] if the vertex connectivity of the network is x, there
exist at least x disjoint paths connecting any pair of processes1 .

Process Failure Models. A fault occurs when a process does not follow its behav-
ior. A process that commits faults is said to be faulty. Otherwise, it is correct.
The model parameter t defines an upper bound on the number of processes that
can be faulty.
We consider two types of faults: crash and Byzantine behavior. A process
fails by crashing when it stops executing before terminating its algorithm. The
Byzantine fault model [14,19] is more severe. A Byzantine process is a process
that behaves arbitrarily: it can crash, fail to send or receive messages, send arbi-
trary messages, start in an arbitrary state, perform arbitrary state transitions,
etc., in short a Byzantine process does not behave as specified by its algorithm.
Such a behavior can be intentional (also called malicious) or the result of a
transient fault which altered the content of local variables of a process, thereby
modifying its intended behavior in an unpredictable way. As a simple example,
a Byzantine process, which is assumed to send the same message m to all the
processes, can send a message m1 to some processes, a different message m2 to
another subset of processes, and no message at all to the other processes. More-
over, Byzantine processes can collude to “pollute” the computation. They can
also control the network in the sense that they can re-order message deliveries at
correct processes. It is however assumed that a Byzantine process cannot send
an infinite number of messages.

1
“Disjoint” means that the only vertices (processes) shared by any two of these paths
are their first and last processes.
104 M. Raynal and J. Cao

2 Broadcast in Fully Connected Crash-Prone Systems: A

Brief Survey
Basic Broadcast. In a fully point-to-point connected network, the basic broadcast
operation, denoted here “ba broadcast tag(m)”, is implemented by the following
loop:
for j ∈ {1, · · · , n} do send tag(m) to pj end for. (1)
This non-atomic macro-operation constitutes a trivial broadcast operation. It is
important to see that this operation is not reliable: if the sender crashes while
executing ba broadcast(), it is possible that only an arbitrary subset of non-
crashed processes receive the message. This type of broadcast (sometimes called
best effort broadcast) is clearly not reliable.

Reliable Broadcast. The previous unreliability has motivated the definition of

the Reliable broadcast (RB), which is a communication abstraction central to
fault-tolerant distributed systems. It allows each process to broadcast messages
to all processes despite failures. More precisely, it guarantees that the correct
processes deliver the same set M of messages. This set M includes at least all
the messages they broadcast, and can also contain messages broadcast by faulty
processes. A faulty process receives a subset of the messages in M .
The fundamental property of reliable broadcast lies in the fact that no two
correct processes deliver different sets of messages. This communication abstrac-
tion is a basic building block used in a lot fault-tolerant distributed applications
(see e.g. [2,5,17,20]).
Interestingly, RB abstraction can be built on top the unreliable broadcast
operation ba broadcast. Moreover, without enriching the model with additional
computability power, RB can be extended to ensure FIFO delivery order (on
each channel messages are received in their sending order) and causal delivery
order (messages are delivered in their causal sending order [13]). The interested
reader will find in [20] (pages 21–74) families of algorithms implementing these
broadcast abstractions in fully connected asynchronous systems where up to
t < n processes may crash.

Total Order Broadcast. Extending RB so that all the messages be delivered in the
same order at each process is impossible. This comes from the fact that, to build
a common total order on message deliveries, the processes need to agree, and
the kind of agreement they have to obtain requires to solve consensus, which is
known to be impossible in the presence of asynchrony and (even a single) process
crash [10]2 .
This means that the system model (full connectivity, asynchrony, process
crashes) is computationally too weak, and must be enriched with additional com-
putational power. This can be done, for example, by adding synchrony assump-
tions [9], providing processes with information on failure [6], or providing the
processes with randomization [1].
2
Actually total order broadcast and consensus are equivalent problems in the sense
any of them can be solved on top of the other one [6].
From Incomplete to Complete Networks in Asynchronous Byzantine Systems 105

3 Byzantine Reliable Broadcast in Fully Connected

Systems
Definition. Reliable broadcast has been studied in the context of Byzantine
failures since the eighties. This communication abstraction (denoted BRB-
broadcast) provides each process with two operations, denoted brb broadcast()
and brb deliver(). As in [11], we use the following terminology: when a process
invokes brb broadcast(), we say that it “brb-broadcasts a message”, and when it
executes brb deliver(), we say that it “brb-delivers a message”. BRB-broadcast is
defined by the following properties. We assume (for the moment) that a process
brb-broadcasts at most one message.

– BRB-Validity. If a correct process brb-delivers a message m from a correct

process pi , then pi brb-broadcast the message m.
– BRB-Integrity. No correct process brb-delivers twice the same message from
a (correct or Byzantine) process pi .
– BRB-Termination-1. If a correct process brb-broadcasts a message m, all
correct processes eventually brb-deliver m.
– BRB-Termination-2. If a correct process brb-delivers a message m from pi
(possibly faulty) then all correct processes eventually brb-deliver m.

On the Safety Properties’ Side. BRB-validity relates the output (messages brb-
delivered) to the inputs (messages brb-broadcast). BRB-integrity states that
there is no duplication. BRB-agreement states that there is no duplicity [20,21]:
be the sender pi correct or not, it is not possible for a correct process to brb-
deliver m from pi while another correct process brb-delivers m = m.

On the Liveness Properties’ Side. The BRB-Termination properties state guar-

antees on message brb-deliveries. BRB-Termination-1 states that a message brb-
broadcast by a correct process is brb-delivered by all correct processes. BRB-
Termination-2 is a strong delivery property, namely, be the sender correct or
Byzantine, every message brb-delivered by a correct process is brb-delivered by
all correct processes.
It follows that all correct processes brb-deliver the same set of messages, and
this set contains at least all the messages they brb-broadcast.
The previous definition considers that each correct process RB-broadcasts at
most one message. Using sequence numbers, it is easy to extend it to the case
where a correct process RB-broadcasts a sequence of messages.

Byzantine-Tolerant Reliable Broadcast Algorithms. In the context of fully con-

nected n-process asynchronous systems, a signature-free reliable broadcast algo-
rithm has been proposed G. Bracha in [3]. This algorithm requires that at most
t < n/3 processes be Byzantine, which is shown to be an upper bound on the
number of Byzantine processes that can be tolerated to implement Byzantine-
tolerant reliable broadcast on top of asynchronous fully connected reliable net-
works [4,20]. Hence, Bracha’s algorithm is t-resilience optimal. This algorithm
106 M. Raynal and J. Cao

is based on a “double echo” mechanism of the value broadcast by the sender

process. For each application message3 it uses three types of protocol messages,
requires three consecutive communication steps (one for each message type), and
generates (n − 1)(2n + 1) protocol messages.
A more recent signature-free Byzantine reliable broadcast algorithm has been
introduced by D. Imbs and M. Raynal in [12]. This algorithm implements the reli-
able broadcast abstraction of an application message with only two consecutive
communication steps (“single echo”), two message types, and n2 − 1 protocol
messages. The price to pay for this gain in efficiency is a weaker t-resilience,
namely t < n/5. Hence, these two algorithms differ in their trade-offs between
t-resilience and message/time efficiency.

Observation. Both the previous reliable broadcast algorithms are built on top of
the ba broadcast operation used in a Byzantine context. Hence, when considering
the “for” loop implementing ba broadcast, while a correct process sends the same
message to processes, a Byzantine sender can send different fake messages, or no
message at all, to different processes.

BRB-Broadcasting a Sequence of Message. The previous version of BRB-

broadcast is one-shot in the sense a process invokes at most once the ba broadcast.
In this case, the identity of the message can be reduced to the identity of the
invoking process. A simple way to obtain a multi-shot BBR-broadcast abstrac-
tion consists in associating an appropriate identity with each invocation of
brb broadcast() by a process. This can be easily realized with sequence numbers,
a process pi now invokes brb broadcast(sni , m), where sni is the next sequence
number value. This instance is then identified by the pair i, sni .

4 Byzantine Basic Broadcast in an Incomplete Network

4.1 Aim and Broadcast Operation

Considering an n-process system including up to t Byzantine processes and where

the communication network is incomplete, our aim is to build a broadcast oper-
ation, denoted bbai broadcast() (Byzantine basic incomplete) that allows the
processes to execute as if the underlying communication network was fully con-
nected.
As shown below, this operation allows to simulate a fully connected network
on top of a n-process system, where up to t < n/3 processes can be Byzantine,
and the underlying network is (2t + 1)-vertex connected. It follows that it is pos-
sible to build a software stack as depicted in Fig. 1. More precisely, replacing the
underlying ba broadcast operation, used at the upper level by Byzantine-tolerant
algorithms designed for a fully connected network, with the bbai broadcast()

3
An application message is a message sent by the reliable broadcast abstraction, while
a protocol message is a message used to implement reliable broadcast.
From Incomplete to Complete Networks in Asynchronous Byzantine Systems 107

Consensus, Total Order Broadcast, ...

Adding computational power (e.g., randomization, synchrony assumptions, ...)

Byzantine reliable broadcast, ...

n processes, t < n/3 Byzantine processes, fully connected

Algorithm 1: bbai_broadcast()

n processes, t < n/3 Byzantine processes, (2t + 1)-vertex connectivity

Fig. 1. A distributed software stack

operation, it follows that, without modifications, these Byzantine-tolerant algo-

rithms can be used on top of an incomplete network. This is the case not only
for the reliable algorithms presented previously [3,12], but also for consensus
and total order algorithms when the computing model is enriched with addi-
tional computability power such as synchrony assumptions or randomization. An
example, replacing the ba broadcast operation it uses with the bbai broadcast(),
the randomized Byzantine binary consensus algorithm described in [16] works
(without modifications) in an incomplete (2t + 1)-vertex connected network4 .

Byzantine Basic Broadcast for Incomplete Network. This (unreliable) broadcast

is denoted BBaI-broadcast. Designed for incomplete networks, it provides the
processes with two operations denoted bbai broadcast() and bbai deliver(). It is
defined by the following properties. Without loss of generality, it is assumed
that all the messages sent by a (correct) process bbai-broadcasts at most one
message. It is easy to obtain a multi-shot version using the same mechanism as
the one used for the multi-shot version of BRB-broadcast (see the last paragraph
of Sect. 3), the identity of a message is not only the identity the sending process
but a pair composed a sender identity plus a local sequence number.
– BBaI-Validity. If a correct process bbai-delivers a message m from a correct
process pi , then pi bbai-broadcast the message m.
– BBaI-Integrity. A correct process bbai-delivers at most one message m from
a (correct or Byzantine) process pk .
– BBaI-Termination. If a correct process bbai-broadcasts a message m, then all
correct processes bbai-deliver m.
Let us observe that bbai broadcast() is pretty weak. It is nothing but
ba broadcast() in an incomplete network. As an example, if the sender pi is
4
Deterministic algorithms implementing Byzantine multivalued consensus on top of
Byzantine binary consensus in asynchronous fully connected networks are described
in [2, 18, 20].
108 M. Raynal and J. Cao

Byzantine, it is possible that distinct sets of correct processes bbai-deliver differ-

ent message from pi , while other correct processes do not bbai-deliver a message
from pi .

4.2 Simulating a Complete Network on Top of an Incomplete

Network
The proposed algorithm is a simple extension to asynchronous systems of a
synchronous algorithm introduced by D. Dolev in [7,8]. It is assumed that t <
n/3, the network is (2t + 1)-connected, and each process knows its structure.
(Let us notice that the assumption that each process knows the communication
graph is implicit in fully connected systems.)

Notations. A path of the network from a process pi to a process pk is repre-

sented by the sequence path = i, . . . , k of the process identities that compose it.
Assuming i ∈ path is not its last identity, the function next(i, path) returns the
identity that follows i in path.
Local variables at a process pi
– The set neighbors i contains the identities of the neighbors of pi .
– The array dis paths[1..n] is such that, for each j, dis paths[j] is a set con-
taining (2t + 1) disjoint paths from pi to pj (“disjoint” except for the first
and the last processes pi and pj repectively). This implicitly means that a
process knows the network (to be able to compute, for each other process pj ,
the (2t + 1) disjoint paths it needs to communicate with pj ).
While m denotes an application message, the protocol messages used to
implement bbai broadcast are tagged impl. The ideas behind the design of the
algorithm are well-known and pretty simple. The first one can be summarized
as “discard or forward”, while the second one consists in delivering a message
only when enough copies of it have been received on disjoint paths.

Broadcast of an Application Message m. When a process invokes

bbai broadcast(m), it considers each destination process pj (line 1), and for each
of the (2t + 1) disjoint paths from it to pj denoted path (line 2), it sends the
protocol message impl(m, i, path) to its successor p on this path path (line 3).
Hence, path is the path, defined by the sender pi , that the message m has to
follow from pi to its destination pj . Let us notice that if a process pi is Byzan-
tine, it can send different messages m, m , etc., or no message at all, to different
subset of processes.

Reception of a Message. When it receives a message impl(m, k, path) from its

neighbor pj , a correct process pi first checks if the protocol message is well-formed
(line 6). “Well-formed” means that, for its receiver pi , the message appears as it
was initially bbai-broadcast by pk and up to pi it traveled along the path path
of the network. More precisely, pk must be the first process in path, which must
From Incomplete to Complete Networks in Asynchronous Byzantine Systems 109

operation bbai broadcast(m) is code for pi

(1) for each j ∈ {1, · · · , n} \ {i} do
(2) for each path ∈ dis paths[j] do
(3) let  be next(i, path); send impl(m, i, path) to p
(4) end for
(5) end for

when impl(m, k, path) is received from pj ∈ neighborsi do

(6) if ¬ well formed(i, j, k, path)
(7) then discard impl(m, k, path)
(8) else if i is not the last identity in path
(9) then let  be next(i, path); send impl(m, k, path) to p
(10) else if (m, k, −) received on (t + 1) disjoint paths)
∧ ((m, k) not yet bbai-delivered)
(11) then bbai deliver(m, k)
(12) end if
(13) end if
(14) end if.

 formed(i, j, k, path) is
function well
(15) return (k is the first identity in path)
∧ (path is an elementary path of the network)
∧ (j is an immediate predecessor of i in path) .

Algorithm 1: From an incomplete network to a fully connected network

be an elementary path of the network, and in which pj is the predecessor of pi

(line 15). If the message is not well-formed, pi discards it (line 7).
If the received message impl(m, k, path) is well-formed, and pi is not the
destination process (i.e., it is not the last process in path, line 8), it forwards the
message impl(m, k, path) to its successor in path line 9). If it is the destination
of the message, pi checks if it has received messages impl(m, k, −) on (t + 1)
disjoint paths (from pk to itself, line 10). If it is the case and additionally pi has
not bbai-delivered the pair (m, k), it does it line 11.

An Improvement. Let consider the case where the message impl(m, k, path)
received by a process pi is such that pk is the immediate predecessor of pi on path.
Hence, pk executed “send impl(m, k, path) to pi ” (entailed by the invocation of
bbai broadcast(m) if pk is correct) or possibly bypassing this invocation if pk
is Byzantine. Whatever the case pi can bbai-delivers the pair (m, k) when it
receives the message impl(m, k, path). This is due to the fact that, as pk and
pi are neighbors, there is no Byzantine process between them which could have
altered impl(m, k, path). The algorithm must be slightly modified to take into
account this observation.
110 M. Raynal and J. Cao

4.3 Proof of the Algorithm

Theorem 1. Algorithm 1 implements BBaI-broadcast.

Proof. Let us first show that, while they can collude, the Byzantine processes
cannot direct a correct process pj to bbai-deliver a pair (m , k) while pk is cor-
rect but has never invoked bbai broadcast(m ). To this end, let px a Byzantine
process that knows that the path path x = k, . . . , x, . . . , j, is a path used by pk
to send a message to pj . Process pk can forge the well-formed protocol message
impl(m , k, path x ) and send it to its successor in path x as done in line 9. More-
over, this can be done by each Byzantine process py on a well-formed path path y
defined similarly to path x . It then follows from the predicate of line 10 that pj
can receive at most t distinct messages impl(m , k, path x ), ..., impl(m , k, path y ).
Consequently, pj will never bbai-deliver the pair (m , k).
The BBAI-Integrity property follows from the last predicate of line 7. To
prove the BBAI-Termination property, let pk be a correct process that invoked
bbai broadcast(m). So, pk executed lines 1–5. Let pi a correct process. As there
are at least (2t + 1) disjoint paths from pk to pi , and at most t Byzantine pro-
cesses, it follows that at least (t + 1) of these disjoint paths are made up of
correct processes only. According to the forwarding mechanism used by the cor-
rect processes, it follows that pi eventually receives (t + 1) well-formed messages
impl(m, k, −) and the predicate of line 5 becomes satisfied, which entails the
bbai-delivered of msg by pi .
To prove the BBAI-Validity property, let pi a process that bbai-delivers the
pair (m, k) where pk is a correct process. In the worst case, among the (t + 1)
paths from pk to pi that make true the predicate of line 5 at pi , at most t can
include Byzantine processes. It follows that at least one of these (t + 1) paths
is composed of correct process only, from which we conclude that pk invoked
bbai broadcast(m). 


Theorem 2. The condition t < n/3 and the (2t + 1)-vertex connectivity are
both necessary and sufficient to implement BBaI-broadcast.

Proof. The sufficiency results from Theorem 1. As far the necessity is concerned,
we have the following.
– Let us first consider the condition t < n/3. On the one hand, Theorem 1 has
shown that the bbai broadcast operation simulates the basic unreliable broad-
cast ba broadcast() operation used in a fully connected network. On the other
hand, it has been shown (see e.g., [3,12]) that the BRB-broadcast abstrac-
tion can be implemented in a fully connected network on top of ba broadcast()
when t < n/3. It consequently follows that BRB-broadcast can be built on
top of bbai broadcast in an incomplete network.
On another side, it is shown in [4,20] that t < n/3 is a necessary condition for
BRB-broadcast in a Byzantine fully connected asynchronous message-passing
system. It follows from these two observations that there is no algorithm that
implements bbai broadcast when t ≥ n/3.
From Incomplete to Complete Networks in Asynchronous Byzantine Systems 111

– As far as the (2t + 1)-vertex connectivity condition is concerned we have

the following. Assuming the vertex connectivity is less than (2t + 1), let us
restrict the asynchronous model to its synchronous executions and consider
a synchronous consensus algorithm executed on top of it (see the software
stack in Fig. 1). Such an algorithm cannot exist. This is due to the fact that
it is proved in [8] that, in a synchronous system where up to t processes can
be Byzantine, the previous connectivity condition is a necessary requirement
for consensus in the presence of up to t Byzantine processes. As synchronous
systems are computationally stronger than asynchronous systems, it follows
that the (2t + 1)-vertex connectivity condition is a necessary requirement for
asynchronous systems which are not fully connected.




5 Conclusion

This paper has presented a simple operation that provides the processes with
a broadcast that simulates an unreliable broadcast on top of an (2t + 1)-vertex
connected network in which up to t < n/3 processes may be Byzantine. The
algorithm implementing this operation is a simple extension of a synchronous
algorithm described in [8]. Its main interest lies in the fact that it allows any
asynchronous algorithm (i) tolerating up to t < n/3 Byzantine processes, and
(ii) designed to work on an underlying fully connected network, to work also
on incomplete (2t + 1)-vertex connected networks. It has also been shown that
both these conditions (t < n/3 and (2t + 1)-vertex connectivity) are necessary
and sufficient, not only for synchronous systems as shown in [8], but also for
asynchronous distributed systems. Last but not least, as shown by Fig. 1, the
simplicity of the bbai broadcast operation favors a modular approach when one
has to cope with Byzantine processes and incomplete networks. Finding a more
efficient algorithm however remains an open challenging non-trivial problem.

Acknowledgments. The authors thank the reviewers for their constructive com-
ments. This work was partially supported by the French ANR projects DESCARTES
(16-CE40-0023-03) devoted to layered and modular structures in distributed comput-
ing, and ByBLosS (20-CE25-0002-01) devoted to the design of modular building blocks
for large-scale trustless multi-users applications.

References
1. Ben-Or, M.: Another advantage of free choice: completely asynchronous agreement
protocols. In: Proceedings 2nd ACM Symposium on Principles of Distributed Com-
puting (PODC’83), pp. 27–30. ACM Press (1983)
2. Ben-Or, M., Kelmer, B., Rabin, T.: Asynchronous secure computations with opti-
mal resilience. In: Proceedings of 13th ACM Symposium on Principles of Dis-
tributed Computing (PODC’94), pp. 183–192. ACM Press (1994)
112 M. Raynal and J. Cao

3. Bracha, G.: Asynchronous Byzantine agreement protocols. Inf. Comput. 75(2),

130–143 (1987)
4. Bracha, G., Toueg, S.: Asynchronous consensus and broadcast protocols. J. ACM
32(4), 824–840 (1985)
5. Cachin, Ch., Guerraoui, R., Rodrigues, L.: Reliable and Secure Distributed Pro-
gramming, 367 p. Springer, ISBN 978-3-642-15259-7 (2011)
6. Chandra, T.D., Toueg, S.: Unreliable failure detectors for reliable distributed sys-
tems. J. ACM 43(2), 225–267 (1996)
7. Dolev, D.: Unanimity in an unknown and unreliable environment. In: Proceed-
ings of 22nd ACM Symposium on Foundations of Computer Science (FOSC’81),
pp. 159–168. ACM Press (1981)
8. Dolev, D.: The Byzantine general strike again. J. Algorithms 3, 14–30 (1982)
9. Dolev, D., Dwork, C., Stockmeyer, L.: On the minimal synchronism needed for
distributed consensus. J. ACM 34(1), 77–97 (1987)
10. Fischer, M.J., Lynch, N.A., Paterson, M.S.: Impossibility of distributed consensus
with one faulty process. JACM 32(2), 374–382 (1985)
11. Hadzilacos, V., Toueg, S.: A Modular Approach to Fault-Tolerant Broadcasts and
Related Problems. Technical report 94-1425, 83 p. Cornell University (1994)
12. Imbs, D., Raynal, M.: Trading t-resilience for efficiency in asynchronous Byzantine
reliable broadcast. Parallel Process. Lett. 26(4), 8 (2016)
13. Lamport, L.: Time, clocks, and the ordering of events in a distributed system.
Commun. ACM 21(7), 558–565 (1978)
14. Lamport L., Shostack R., Pease M.: The Byzantine generals problem. ACM Trans.
Programm. Lang. Syst. 4(3), 382–401 (1982)
15. Menger, K.: Zur allgemeinen kurventheorie. Fundamentae Mathematicae 10, 96–
115 (1927)
16. Mostéfaoui, A., Moumen, H., Raynal, M., Signature-free asynchronous binary
Byzantine consensus with t<n/3, O(n2 ) messages, and O(1) expected time. J.
ACM 62(4), Article 31, 21 pages (2015)
17. Mostéfaoui, A., Raynal, M.: Intrusion-tolerant broadcast and agreement abstrac-
tions in the presence of Byzantine processes. IEEE Trans. Parallel Distrib. Syst.
27(4), 1085–1098 (2016)
18. Mostéfaoui, A., Raynal, M.: Signature-free asynchronous Byzantine systems: from
multivalued to binary consensus with t<n/3, O(n2 ) messages, and constant time.
Acta Informatica 54(5), 501–520 (2017)
19. Pease, M., Shostak, R., Lamport, L.: Reaching agreement in the presence of faults.
J. ACM 27, 228–234 (1980)
20. Raynal, M.: Fault-tolerant message-passing distributed systems: an algorithmic
approach. Springer, 480 p, ISBN: 978-3-319-94140-0 (2018)
21. Toueg, S.: Randomized Byzantine agreement. In: Proceedings of 3rd Annual ACM
Symposium on Principles of Distributed Computing (PODC’84), pp. 163–178.
ACM Press (1984)
 Towards the Performance Evaluation of a
Trust Based Routing Protocol for VANET

Amira Kchaou(B) , Ryma Abassi, and Sihem Guemara El Fatmi

Digital Security Research Lab, Higher School of Communication of Tunis,

University of Carthage, Tunis, Tunisia
{Amira.Kchaou,Ryma.Abassi,Sihem.Guemara}@supcom.tn

Abstract. Vehicular Ad-hoc NETwork (VANET) is a decentralized

dynamic network of vehicles that move rapidly and communicate with
one another to disseminate various types of information and deliver data.
However, dishonest vehicles may present in the route and may lead for
changing or dropping the content of delivered messages. This latter is a
challenge issue in VANET to find a trustful route for reliable data dis-
semination. In a previous work, we have proposed a routing protocol for
VANETs based on trust (TRPV) to find an optimal route depending on
the trustworthiness of the path and the number of hops. Thus, in this
paper, simulations are achieved to evaluate the performance of TRPV
in terms of the throughput, the Packet Data Ratio (PDR), the routing
overhead, and the average end-to-end delay.

1 Introduction

Vehicular Ad-hoc NETwork (VANET) is a decentralized wireless communica-

tion network which provides internet connection and multimedia services such
as video/audio streaming, and also safety services such as traffic management
and road safety to avoid dangerous situations beside weather or positioning ser-
vices, etc. [1]. These services depend on a timely transmission of data through
messages within the network, which can achieved by the use of vehicular com-
munications [1].
Thus, the communication links between vehicles are unreliable due to the vehi-
cle highmobility, dynamic network, the existence of dishonest vehicles and others
VANET characteristics. However, the routing is a challenging task in the VANET
to find an efficient and secure route to successfully transmit the data from the
source to the destination vehicles. To resolve these problems, the proposed trust
and clustering mechanism are introduced in the VANET to form clusters having
similar vehicle velocity, electing ClusterHeads (CHs) using the common vehicle
between the neighbors of each vehicle, and to successfully transform information
[2].
In previous work [3], we have proposed a trust based routing protocol for the
VANET baptized TRPV through trusted CHs (i.e. CH having a trust between it

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 113–124, 2021.
https://doi.org/10.1007/978-3-030-75100-5_11
114 A. Kchaou et al.

and other vehicle greater than a minimum threshold value τ ) that facilitates the
information transmission through messages between the vehicle and the outside
of its cluster.
The proposed TRPV finds a trustful route depending on the trustworthiness
of the path from the source to the destination vehicles, as well as the number
of hops for reliable information and message dissemination to the destination.
In fact, the proposed TRPV is used for intra-cluster and inter-cluster route
discovery. During the intra-cluster route discovery, the source vehicle sends a
message to its corresponding CH in order to check the correctness of the message
by calculating the message credibility. Afterwards, the CH directly forwards the
message to the destination vehicle. On the other hand, the source vehicle sends a
route request packet (RREQ) to its trusted CH. Next, the CH pursues the route
discovery process via the CHs existing at the level two (i.e. having all CHs)
until the routes are received. In addition, the CH forwards the routes to the
requesting vehicle. Therefore, the source vehicle chooses the trustful and optimal
route depending on the lowest number of hops, as well as the trustworthiness of
the path.
In this paper, we evaluate the performance of TRPV through simulations of
the throughput, the routing overhead, the Packet Data Ratio (PDR) and the
end to end delay, and compare it to the AODV routing protocol.
The remainder of this paper is organized as follows. Section 2 reviews some
related trust based routing protocols. Section 3 introduces the proposed TRPV.
Section 4 presents the simulation results showing some TRPV performances com-
pared to AODV the protocol. Finally, Sect. 5 concludes this paper.

2 Related Works

This section presents some existing works dealing with trust based routing pro-
tocol in VANETs.
In [4], the authors proposed a reliability model based routing protocol for
the vehicular network, called R-AODV. The aim of the proposed is to facili-
tate the reliable routing between vehicles calculating the link reliability value on
the basis of the location, direction and the velocity of vehicles along the route.
However, R-AODV is not used trust for calculating the link reliability. In addi-
tion, R-AODV is not suited due to high payload transmission and frequent link
breakage. In [5], Sahoo et al. proposed a trust based clustering with ant colony
routing in VANET, called TACR in order to provide an efficient vehicular com-
munication. In fact, VANET is organized into clusters having the RSUs as CHs
and vehicles as mobile nodes. Then, CH collects the opinions of members on
the vehicle, and takes a decision about the message. Thereafter, the proposed
used the ant colony algorithm in order to select the efficient route among clusters
using boundaries vehicles. However, TACR has two weaknesses: the slow sending
decision and the using of a static CH. In [6], Kerrache et al. introduced a Trusted
ROUting protocol for urban Vehicular Environment, called TROUVE in order
to establish a reliable and secured route to the destination vehicle. In fact, each
 Towards the Performance Evaluation of a Trust Based Routing Protocol 115

vehicle maintains the neighborhoods information and broadcasts it via Cooper-

ative Awareness Messages (CAM). Next, it broadcasts the data routing through
the shortest and secured route. However, TROUVE address unicast data of traf-
fic management in urban environments. In [7], the authors proposed a trusted
routing protocol for VANET on the basis of trusted metrics (the velocity, the
direction, the distance, and the trust value) using trusted computing algorithm
in order to provide a secured route selection and prevent malicious vehicles. How-
ever, the proposed neglects the high traffic density and the route maintenance
process. In [8], the authors proposed a trust based distributed authentication
method (TDA) in VANET for avoiding the collision attack, ensuring secure
information transmission, and selecting the optimal path with minimal energy.
In fact, honest vehicles are identified according to the channel state and the
on-board unit (OBU) energy of the vehicle for establishing seamless communi-
cation. However, the proposed is focused on less energy and not based on the less
number of hops to select the shortest route including in the optimal route. In
[9], the authors proposed a routing protocol based on trust and minimum delay
for VANET in order to select the optimal path with highest trust and minimal
delay. In fact, the proposed computes the trust calculation and delay of vehicles
on the basis of the participation vehicles in path using the present and previous
knowledge of vehicles along with risk threshold of messages to be sent. Then, the
optimal path is selected with highest trust and less delay. However, the proposed
is not focused on number of hops in the selection of optimal path.
Given that the existing works benefit from trust management and routing, in
this paper, we propose a new trust based routing protocol for VANET to provide
an efficient and trustful route. This protocol will use the trustworthiness of the
whole path from the source vehicle to the destination based the reputation of
the vehicles, and the less number of hops.

3 TRPV: Trust Based Routing Protocol for VANET

This section deals with the Trust based Routing Protocol VANET baptized
TRPV based on the number of hops (Hop Count) and the trustworthiness of
the path (TP ) in order to select the optimal route in the network.

3.1 Basic Properties

The TRPV has several properties:

• Each CH has a CH list containing all the other CHs in the network and
its reputation value, a blacklist including the dishonest vehicle (which has a
reputation value equaled to -3), and a REPUTATION TABLE containing
the members of the cluster and its reputation value.
• Each CH stores the TP in its routing table,which is updated for a certain
time β.
116 A. Kchaou et al.

• The CH having a trust between it and a vehicle higher than a predefined

threshold τ is considered as trusted CH.
• The routing discovery process for inter-clusters can be achieved through
trusted CHs.

3.2 TRPV Description

In VANET, each CH computes the TP to the destination vd using the trust
between vehicles along the route. The TP is the product of trust between the
vehicles (vehicle vi and vehicle vj ) Tij along the route for finding the trustful
route.
The TP is formalized by the following formula (1).
k

TP = Tij (1)
i=1,j=1

Where the Tij is the trust link between vehicle vi and vehicle vj along the route.
The Tij is formalized by the formula (2).

Tij = REPi ∗ REPj vi, vj ∈ route (2)

Where the Tij is equal to the product between vehicle vi and vehicle vj reputation
values in the route.
When a source vehicle vs wants to transfer the data to destination vd, the
process of the route discovery will be achieved according to two scenarios: the
route discovery for intra-cluster and the route discovery for inter-cluster.

3.3 Route Discovery for Intra-clusters

When the source vehicle vs and the destination vehicle vd belong to the same
cluster, the process of the route discovery for intra-clusters is achieved, as pre-
sented below:

• The source vehicle vs sends an URGENT MSG to its CH to check the cor-
rectness of the message.
• After calculating the message credibility, the CH consults its
CH MEMBER TABLE to find destination vd.
• If the destination vd exists, the CH forwards to destination vd an
URGENT MSG, and to the source vehicle vs an ACK MSG, otherwise go
to next scenario;
 Towards the Performance Evaluation of a Trust Based Routing Protocol 117

Table 1. Routing table

IDvd Dest seq num Routes list

Route ID Next Hop Hop Count TP LifeTime (LT)

3.4 Route Discovery for Inter-clusters

When the source vehicle vs and the destination vehicle vd do not belong to the
same cluster, the process of the route discovery for inter-clusters will be achieved.
In a previous work, when the source vehicle vs sends a RREQ to a trusted
CH having TvsCH higher than τ . If there is no trusted CH with an appropriate β
then, the source vehicle vs resends the RREQ. Next, the CH consults its CH list,
continues the route discovery through CHs existing at level two (i.e. having all
CHs while routes are received) and forwards the RREQs to the requesting vehi-
cle. Thus, the route discovery process prohibits RREQ flooding in the VANET.
The RREQ packet header is changed as follows :

RREQ( Trusted CH, RREQ ID, Hop Count, ID vd ,

Dest seq num, ID vs , S seq num, τ )

At level two, vehicle vi receives a RREQ packet from vehicle vj, it looks up
in its routing table for a valid route until destination vd. The routing table
structure is depicted in Table 1. Where the route list refers to several feasible
routes and to the destination IDvd classified in an ascending order of Hop Count,
like Hop Count1 ≤ Hop Count2 ≤ ... ≤ Hop Countn in the routing table. In fact,
every route has a Next Hop corresponding to the vehicle of Next Hop in this
latter. TP refers to the trust value of the overall route from the current vehicle
to the destination IDvd . The Lifetime (LT ) represents the time of the validity.
Then, after consulting the routing table, if vehicle vi does not get a valid
route to the destination vd, then it adds its address IDvi to the RREQ packet,
increases the Hop Count by one, and sends it to the trusted CHs existing at
level two excepting the vehicle vj.
Next, vehicle vi updates the reverse route entry in the routing table where
the Next Hop must be vehicle vj and modifies the TP of vehicle vj.
Moreover, if there is no transaction between vehicle vi and vehicle vj then,
the vehicle vi forms an entry for the vehicle vj in its own local routing table.
Therefore, this process will continue to destination vd.
Else if vehicle vi finds the valid route to destination vd, the dest seq num
in the RREQ rises by one, then vehicle vi adds the route parameters: the
Hop Count and TP in the route reply (RREP), and sends it reverse to the
previous vehicle vj.
Else if vehicle vj receives the RREP, then it calculates the TP to destination
vd, and sends back the Current TP dest and Hop Count to its previous vehicle vk
using the reverse route up to the source vehicle vs. Then, the latter receives the
118 A. Kchaou et al.

Fig. 1. Flow chart of the route discovery process for the inter-cluster.

RREP and calculates the overall TP up to destination vd. The RREP structure
is depicted as follows:
RREP (Hop Count, Current TP dest, ID vd , Dest seq num,
ID vs , LT )
Where the Current TP dest represents the current TP value to IDvd.
Else if the Next Hop of vehicle vi represents destination vd, then the
dest seq num in the RREQ increases by one and vehicle vi checks if the Tvivd of
 Towards the Performance Evaluation of a Trust Based Routing Protocol 119

destination vd is higher than τ , then the Hop Count in the RREQ rises by one.
Moreover, vehicle vi sends the RREQ to destination vd and updates the reverse
route entry in the local routing table where the Next Hop, which is destination
vd, alters the TP of destination vd. Otherwise, vehicle vi sends the route error
(RERR) to vehicle vj.
After that, the vehicle source vs determines the optimal route depending on
two metrics: the TP and the Hop Count. If there are various routes having the
same TPs, then the route having the lowest Hop Count is elected as the opti-
mal and trustful route. However, since several routes have the same Hop Count,
the route having the highest TP is elected as the optimal and trustful route.
Thus, the source vehicle vs saves the other discovered routes to be used in some
situations like broken link, the routing failure, route errors, and unreachable des-
tinations. After a period of time, the source vehicle vs waits an for an ACK MSG
from the destination vehicle vd or for a RERR packet from the CHs.
In each step of the route discovery process, the RERR is sent back to the
source vs when it cannot reach the destination vehicle vd. Then, the source
vehicle vs and the intermediate vehicles delete the destination entry in their
routing table upon receiving the RERR packet.
The RRER structure is depicted as follows :

RERR (Dest Count, Unreachable ID vd 1,

Unreachable dest seq num 1, Unreachable ID vd 2,
Unreachable dest seq num 2 )

Where the Dest Count corresponds to the computation of unreachable vehicles.

With the AODV routing protocol, if the RREP packet in our proposal is not
received after a period time β, then the source vs will broadcast a new RREQ.
Else if the failure remains, then the next RREQ will hold for 2*β seconds to be
broadcast.
In this case, this process will be pursued until three broadcast RREQ packet
is reached. Hence, the flow chart is illustrated in Fig. 1.

4 Simulation and Results

The aim of this section is to evaluate the performance of TRPV proposal.

4.1 Simulation Environment

We have conduct the simulations using the vehicular network simulation (Veins),
the network simulator (Omnet++) and the traffic simulator: the Simulation of
Urban MObility (SUMO). In our simulation, we have 50 vehicles in the network
randomly placed in 2000 m * 2000 m. Additionally, each vehicle has 250 m range
and moves with a maximum velocity of 30 m/s. Then, the simulation duration
is 400 s and each situation is repeated 20 times for taking the average value.
Table 2 illustrates the simulation parameters.
120 A. Kchaou et al.

Table 2. Simulation parameters of VANET

Parameter Value
Simulation area 2000 m * 2000 m
Mobility model Random waypoint
Mac protocol IEEE 802.11p
Number of vehicles 50
Number of RSUs 1
Velocity 5 m/s–30 m/s
Transmission range 250 m
Trafic type CBR
Packet size 512 bytes
Packet rate 4 packets/s
Pause time 10 s
Bandwidth 2 Mbps
β 2.8 s
τ 0.7
Routing protocol AODV, TRPV
Simulation time 400 s

To evaluate the performance of the proposed TRPV, we consider the follow-

ing metrics:

• The throughput: is the ratio of the number of packets received by the desti-
nation vehicle per unit of time.
• The routing overhead: is the ratio between the number of the routing packets
(RREQs, RERRs, RREPs and the updated packets) and the total amount of
data and the routing packets.
• The Packet Data Ratio (PDR): is the ratio between the number of RREQ
packets received by the destination node and the number of RREQ packets
transmitted by the source node. In fact, it represents the reliability of packet
transmissions in a network.
• The average end-to-end delay: is the delay experienced by the successfully
forwarded packets to reach their destinations.

4.2 Simulation Results and Analysis

We assess the performance of TRPV to different velocity of vehicles. To show

the performance of TRPV, we used the following metrics: throughput, routing
overhead, average end-to-end delay, and Packet Data Ratio (PDR).
 Towards the Performance Evaluation of a Trust Based Routing Protocol 121

4.2.1 Throughput
‘Figure 2 shows the variation of the throughput when varying the velocity of the
vehicles for the proposed TRPV with the AODV protocol in the network. We
notice that the throughput decreases as the velocity of the vehicles increases. In
fact, when the vehicle velocity increases (5-10-15-20-25-30-35), the packet loss
increases on the communication links as the possibility of the reliable and optimal
path discovery decreases, and the network throughput should be degraded. This
problem is severe in the AODV. The obtained results show that the proposed
TRPV gets better throughput than AODV protocol. This can be explained by
the fact that the proposed TRPV involves stable clustering and provides trustful
route that improves by maximizing the selected route TP, and minimizing the
number of hops.

Fig. 2. Throughput according to the velocity of vehicles in the network.

4.2.2 Routing Overhead

Figure 3 shows the evolution of the routing overhead for different vehicle veloci-
ties (5-10-15-20-25-30-35). The obtained results show that the proposed TRPV
is lower than AODV protocol. In fact, at high velocities, the connection between
the vehicles frequently fails. However, in the proposed TRPV, the control pack-
ets (RREQ and RREP) are generated in CH configurations in order to discover
and recover the route. Therefore, the proposed TRPV is achieved among CHs
and avoids the routing discovery among all vehicles in the network. Moreover,
the control packets are sent to qualified routes in order to meet trust in the
proposed TRPV, meanwhile, the trust is not treated in the AODV protocol.
122 A. Kchaou et al.

Fig. 3. Routing overhead according to the velocity of vehicles in the network.

4.2.3 Packet Data Ratio (PDR)

Figure 4 shows the variation of the PDR for different vehicle velocities (5-10-15-
20-25-30-35). The obtained results show that the proposed TRPV is higher than
AODV protocol. In fact, using the clustering mechanism in the proposed RREQ
flooding is prevented. Moreover, the proposed TRPV excludes the malicious
vehicles from the participation in network transactions as a result of malicious
vehicle detection.

4.2.4 End to End Delay

Figure 5 shows the evolution of the End to End delay when varying the velocity
of the vehicles for the proposed TRPV and AODV protocol in the network.
The obtained results show that the proposed TRPV is lower than the AODV
protocol. This can be explained by the fact that the proposed TRPV can handle
alternative routes and the later alternative route can be handled when receiving
an RERR packet. However, the AODV protocol has one route in routing table.

Fig. 4. Packet data ratio according to the velocity of vehicles in the network.
 Towards the Performance Evaluation of a Trust Based Routing Protocol 123

Fig. 5. End to End delay according to the velocity of vehicles in the network.

When the route failure happens, the routing step should be repeated. Thus, the
packets encounter more latency when the vehicles move rapidly.

5 Conclusion

In VANET, the vehicles exchange various types of information and transmit it

from the source to the destination. However, warning messages exchanged and
unreliable links presented in the route pose road security problems. To miti-
gate this problem, we proposed a routing protocol based trust for the VANET
through trusted CHs in order to find a trustful route depending on all vehicles
trustworthiness along the route and the number of hops. In fact, the path trust-
worthiness is computed by the evaluating trust of link that must be greater than
τ . In this paper, the simulation results presented the performances of the trust
routing protocol for the VANET compared to the AODV in terms of throughput,
Packet Data Ratio (PDR), routing overhead, and average end-to-end delay by
varying the velocity of the vehicles. Hence, the proposal selects the optimal and
trustful route and improve the data transmission.

References
1. Shafi, S., Ratnam, D.V.: Mobility-aware cross layer based reliable routing protocol
for improved data dissemination in VANETs (2019)
2. Kchaou, A., Abassi, R., Guemara, S.: Towards a secured clustering mechanism for
messages exchange in VANET. In: 2018 32nd International Conference on Advanced
Information Networking and Applications Workshops (WAINA), pp. 88–93. IEEE
(2018)
3. Kchaou, A., Abassi, R., El Fatmi, S.G.: A new trust based routing protocol for
VANETs. In: 2018 Seventh International Conference on Communications and Net-
working (ComNet), pp. 1–6. IEEE (2018)
124 A. Kchaou et al.

4. Eiza, M.H., Ni, Q., Owens, T., Min, G.: Investigation of routing reliability of vehic-
ular ad hoc networks. EURASIP J. Wirel. Commun. Network. 2013(1), 179 (2013).
https://doi.org/10.1186/1687-1499-2013-179
5. Sahoo, R.R., Panda, R., Behera, D.K., Naskar, M.K.: A trust based clustering with
ant colony routing in VANET. In 2012 Third International Conference on Comput-
ing, Communication and Networking Technologies (ICCCNT 2012), pp. 1–8. IEEE
(2012)
6. Kerrache, C.A., Lagraa, N., Calafate, C.T., Lakas, A.: TROUVE: a trusted rout-
ing protocol for urban vehicular environments. In: 2015 IEEE 11th International
Conference on Wireless and Mobile Computing, Networking and Communications
(WiMob), pp. 260–267. IEEE (2015)
7. Jeyaprakash, T., Mukesh, R.: A new trusted routing protocol for vehicular ad hoc
networks using trusted metrics. IJ Netw. Secur. 19(4), 537–545 (2017)
8. Tolba, A.M.R.: Trust-based distributed authentication method for collision attack
avoidance in VANETs. IEEE Access 6, 62747–62755 (2018)
9. Sataraddi, M.J., Kakkasageri, M.S.: Trust and delay based routing for VANETs. In:
2019 IEEE International Conference on Advanced Networks and Telecommunica-
tions Systems (ANTS), pp. 1–6. IEEE (2019)
 Design and Implementation of the TBOI
(Time-Based Operation Interruption) Protocol
to Prevent Late Information Flow in the IoT

Shigenari Nakamura1(B) , Tomoya Enokido2 , and Makoto Takizawa3

1 Tokyo Metropolitan Industrial Technology Research Institute, Tokyo, Japan
2 Rissho University, Tokyo, Japan

[email protected]
3 Hosei University, Tokyo, Japan

[email protected]

Abstract. In the CBAC (Capability-Based Access Control) model proposed for

the IoT (Internet of Things), authorizers, e.g. device owners, issue capability
tokens, i.e. a set of access rights on resource objects in devices, to subjects. Since
data of a device are stored in another device through communication among sub-
jects and devices, a subject sb can get data of a resource object r1 by accessing
another resource object r2 even if the subject sb is not allowed to get the data from
the resource object r1 . Here, an illegal information flow from the resource object
r1 to the subject sb occurs. In addition, each capability token is valid only for
some time period. Suppose, data generated at time τ flow to a subject sb at time
τ  (> τ ). If the subject sb is allowed to get data at time τ  but not at time τ , the sub-
ject sb should not receive the data. Here, the information flow is referred to as late.
In our previous studies, the OI (Operation Interruption) and TBOI (Time-Based
OI) protocols are proposed to prevent operations implying only illegal informa-
tion flow and both illegal and late information flows, respectively. In this paper,
we discuss the design and implementation of a device supporting the TBOI pro-
tocol and evaluate an authorization process of the TBOI protocol in terms of the
execution time. In the evaluation, we show the late information flow is prevented
in addition to the illegal one in the TBOI protocol although the execution time is
almost the same as the OI protocol.

Keywords: IoT (Internet of Things) · Device security · CBAC

(Capability-Based Access Control) model · Information flow control · Late
information flow · TBOI (Time-Based Operation Interruption) protocol

1 Introduction
It is widely recognized that access control models [3] are useful to make information
systems like database systems secure. For the IoT (Internet of Things) [6], the CBAC
(Capability-Based Access Control) model named “CapBAC model” is proposed [5, 7].
Here, subjects are issued capability tokens which are collections of access rights. Only
the authorized subjects can manipulate resource objects of devices only in the autho-
rized operations. Data are exchanged among subjects and resource objects through
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 125–135, 2021.
https://doi.org/10.1007/978-3-030-75100-5_12
126 S. Nakamura et al.

manipulating resource objects of devices. Therefore, a subject might obtain data via
other subjects and resource objects even if the subject is granted no access right to
obtain the data, i.e. illegal information flow might occur. The illegal information flow is
a critical issue for not only the IoT but also other various types of information systems.
In database systems with the RBAC (Role-Based Access Control) model, protocols to
prevent illegal information flow are proposed [4, 8].
In order to solve this illegal information flow problem in the IoT, the OI (Operation
Interruption) protocol is proposed in our previous studies [10]. Here, operations imply-
ing illegal information flow are interrupted, i.e. not performed, at devices. Since illegal
information flow may occur in the CBAC model, each device can check whether or not
each operation is illegal. The OI protocol is implemented on a device and evaluated in
terms of the execution time to authorize access requests [11].
A capability token issued to a subject is valid only for some time period. In the OI
protocol, although no illegal information flow occurs, a subject might get data generated
out of validity period of a capability token to get the data. Suppose a subject sb is issued
a capability token including an access right r, get, i.e. the subject sb is allowed to get
data from the resource r. The subject sb tries to get data generated at time τ from the
resource r. Here, even if the time τ is not within the validity period of the capability
token, the subject sb can get the data. The data are older than the subject sb expects to
get, i.e. information comes to the subject sb late.
In our previous studies [9], the TBOI (Time-Based OI) protocol is proposed to pre-
vent both illegal and late information flows. Here, operations implying not only illegal
information flow but also late information flow are interrupted. The late information
flow is defined based on the generation time of data of resources and validity peri-
ods of capability tokens issued to subjects. Each device can check whether or not each
operation is late by using the time information. The TBOI protocol is evaluated in a
simulation. The number of operations interrupted in the TBOI protocol is larger than
the OI protocol.
In this paper, we design and implement a device supporting the TBOI protocol.
Here, a Raspberry Pi3 Model B+ with Raspbian [2] is used as an IoT device. The device
and a subject are a CoAP server and a CoAP client, respectively, in CoAPthon3 [14].
In the evaluation, we show that both illegal and late information flows are prevented in
the TBOI protocol although the execution time of an authorization process is almost the
same as the OI protocol.
In Sect. 2, we present the system model. In Sect. 3, we discuss types of information
flow relations to occur in the CBAC model. In Sect. 4, we overview the TBOI protocol.
In Sect. 5, we implement the TBOI protocol on a Raspberry Pi 3 Model B+. In Sect. 6,
we evaluate the authorization process of the TBOI protocol in terms of the execution
time.

2 System Model

In this paper, we assume that an IoT includes a pair of sets D and SB composed of the
number dn of devices d1 , . . . , ddn (dn ≥ 1) and the number sbn of subjects sb1 , . . . , sbsbn
(sbn ≥ 1), respectively. Each device dk holds a set dk .R composed of the number nrk
 Design and Implementation of the TBOI 127

(≥ 1) of resource objects r1k , . . . , rnr

k . Notations r k and r denote some resource object
k
k
rm in a device dk and some resource object rm k in some device d , respectively. In order
k
to make the IoT secure, the CBAC (Capability-Based Access Control) model is pro-
posed [5, 7]. Here, authorizers like device owners issue capability tokens to subjects to
be able to access resource objects in the device. Each subject sbi obtains a set CAPi
composed of the number nci of capability tokens capi1 , . . . , capinci issued. A capability
token includes access rights. An access right such as a pair r, op indicates a subject
granted the access right is allowed to manipulate a resource r in an operation op. When
a subject sbi tries to manipulate data of a resource rk in a device dk in an operation op,
an access request and a capability token capi including an access right rk , op are sent
to the device dk from the subject sbi . If the capability token capi is valid, the access
request is accepted, i.e. the operation op is performed on the resource rk . Otherwise,
the message which indicates the access request is rejected is sent to the subject sbi from
the device dk . Since the device dk just checks the capability token capi to authorize
the subject sbi , it is widely recognized that the CBAC model is more suitable for the
IoT than other access control models based on the ACL (Access Control List) such
as RBAC (Role-Based Access Control) [12] and ABAC (Attribute-Based Access Con-
trol) [16] models. However, it is more difficult to change capability tokens, i.e. access
rights already granted to subjects, because these access rights are already distributed to
subjects.
Each capability token is only valid for some period. Suppose a subject sbi is issued
a capability token capi including an access right r, op on a resource object r which is
specified to be valid for time t1 to time t2 . The subject sbi can only access the resource
object r in the operation op only in the validity period for time t1 to time t2 . A capability
token capi to get data from a resource r is valid from time gt i .st(r) to time gt i .et(r).
Data in a resource object r are generated at some time which is referred to a generation
time. Data in a resource object r may flow to a subject and another resource object. Let
minRTmk (r) be the earliest generation time of data in a resource object r which flow to
another resource object rm k . In addition, minSBT i (r) is the earliest generation time of

data in a resource object r which flow into a subject sbi .

There are three types of devices, sensor, actuator, and hybrid devices. A sensor
device just collects data by sensing events occurring around itself. An actuator device
acts according to data and action requests sent from subjects. Subjects get and put data
from sensor devices and to actuator devices, respectively. A hybrid device both collects
data and acts such as cars and robots. Hence, subjects issue both get and put operations
to hybrid devices.

3 Information Flow Relations

Through manipulating data of resource objects in devices, data are exchanged among
entities, i.e. resource objects and subjects. For example, a subject sb gets data from a
resource object r1 in a device d1 and puts the data to a resource object r2 . Here, the data
of the resource object r1 flow into the subject sb and the resource object r2 . In this paper,
resource objects whose data flow into entities such as resource objects and subjects are
referred to as source resource objects for these entities. Let r.sR and sb.sR are sets of
128 S. Nakamura et al.

source resource objects for a resource object r and a subject sb, respectively, which are
initially φ . In this example, sb.sR = r2 .sR = {r1 }. In another example, a sensor device
d1 collects data by sensing events and stores the data in its resource object r1 . Here,
r1 .sR = {r1 }.
Let IN(sb) be a set of resource objects whose data a subject sb can get, i.e. IN(sbi ) =
{r | r, get ∈ capi ∧ capi ∈ CAPi }. In this section, we define types of information flow
relations on resource objects and subjects based on the CBAC model as follows [9, 10]:
k flows to a subject sb (r k → sb ) iff (if and only if)
Definition 1. A resource object rm i m i
rm .sR = φ and rm ∈ IN(sbi ).
k k

k legally flows to a subject sb (r k ⇒ sb ) iff r k → sb

Definition 2. A resource object rm i m i m i
k .sR ⊆ IN(sb ).
and rm i

k illegally flows to a subject sb (r k → sb ) iff r k → sb

Definition 3. A resource object rm i m i m i
k .sR  IN(sb ).
and rm i

If rmk → sb holds, data flow from the resource object r k to the subject sb . Otherwise, no
i m i
data flow from the resource object rm k into the subject sb . In the legal information flow
i
relation (⇒), the condition “rm k .sR ⊆ IN(sb )” means that only data which the subject
i
sbi is allowed to get flow from the resource object rm k into the subject sb . In the illegal
i
information flow relation (→), the condition “rm k .sR  IN(sb )” means that data which
i
the subject sbi is not allowed to get flow from the resource object rm k into the subject

sbi .

Definition 4. A resource object rm k timely flows to a subject sb (r k ⇒ sb ) iff r k ⇒ sb

i m t i m i
and ∀r ∈ rm
k .sR(gt i .st(r) ≤ minRT k (r) ≤ gt i .et(r)).
m

Definition 5. A resource object rm k flows late to a subject sb (r k → sb ) iff r k ⇒ sb

i m l i m i
and ∃r ∈ rm
k .sR(¬(gt i .st(r) ≤ minRT k (r) ≤ gt i .et(r))).
m

In the timely information flow relation (⇒t ), the condition “∀r ∈ rm k .sR(gt i .st(r) ≤

minRTmk (r) ≤ gt i .et(r))” means that only the data generated while the subject sbi
is allowed to get flow from the resource object rm k into the subject sb . In the late
i
information flow relation (→l ), the condition “∃r ∈ rm k .sR(¬(gt i .st(r) ≤ minRT k (r) ≤
m
gt i .et(r))” means that data generated while the subject sbi is not allowed to get flow
from the resource object rm k into the subject sb . This means, even if the data are not
i
generated while the subject sbi is allowed to get the data, the subject sbi can get the
data.

4 Protocols to Prevent Illegal and Late Information Flow

4.1 OI (Operation Interruption) Protocol
In the CBAC model for the IoT, capability tokens including access rights are issued
to subjects. Data are exchanged among subjects and resource objects through manip-
ulating resource objects in devices. If data of a resource object r1 flowed into another
resource object r2 are furthermore brought into a subject sb which is not allowed to get
 Design and Implementation of the TBOI 129

the data of the resource object r1 , the data of the resource r1 illegally flow into the sub-
ject sb. In Sect. 3, the illegal information flow relation (→) among subjects and resource
objects is defined based on the CBAC model. In order to prevent the illegal information
flow, the OI (Operation Interruption) protocol is proposed in our previous studies [10].
Here, operations implying illegal information flow are interrupted at devices. The OI
protocol is shown as follows:
[OI (Operation Interruption) protocol]
1. A device dk gets data by sensing events occurring around the device dk .
a. The device dk stores the sensor data in the resource object rm k and r k .sR =
m
rm .sR ∪ {rm };
k k

2. A subject sbi issues a get operation on a resource rmk to a device d .

k
k ⇒ sb , the subject sb gets data from the resource r k and sb .sR = sb .sR ∪
a. If rm i i m i i
k .sR;
rm
b. Otherwise, the get operation is interrupted at the device dk ;
3. A subject sbi issues a put operation on a resource rmk to a device d .
k
k
a. The subject sbi puts the data to the resource rm and rm k .sR = sb .sR;
i

rmk .sR and sb .sR are sets of source resource objects for a resource object r k and
i m
a subject sbi , respectively. Although rm k .sR and sb .sR are initially φ , they are updated
i
through manipulating resource objects. IN(sbi ) is a set of resource objects whose data
a subject sbi can get. The illegal information flow relation is defined based on these
sets, rmk .sR, sb .sR, and IN(sb ), as shown in the Definition 3. If data which the subject
i i
sbi is not allowed to get flow from the resource object rm k into the subject sb in a get
i
operation of the subject sbi , an illegal information flow (rm k → sb ) occurs. Hence, the
i
get operation is interrupted in the OI protocol.

4.2 TBOI (Time-Based OI) Protocol

In the OI protocol, although every illegal information flow is prevented, subjects might
get data generated out of validity period of a capability token to get the data. For exam-
ple, a subject sb1 sends an access request and a capability token cap1 to get data gen-
erated at time τ from the resource r. Even if the time τ is not within the validity period
of the capability token cap1 , the subject sb1 can get the data. Here, the data are older
than the subject sb1 expects to get, i.e. information comes to the subject sb1 late. The
late information flow relation is defined based on the CBAC model as shown in Sect. 3.
In order to prevent both illegal and late types of information flows, the TBOI (Time-
Based OI) protocol is proposed in our previous studies [9]. Here, operations implying
not only illegal information flow but also late one are interrupted. The TBOI protocol
is shown as follows:
[TBOI (Time-Based OI) protocol]
1. A device dk gets data by sensing events occurring around the device dk at time τ .
k and r k .sR = r k .sR ∪
a. The device dk stores the sensor data in a resource object rm m m
{rm };
k
k ) = NULL, minRT k (r k ) = τ ;
b. If minRTmk (rm m m
130 S. Nakamura et al.

k to a device d .
2. A subject sbi issues a get operation on a resource object rm k
k ⇒ sb ,
a. If rm t i
i. For each resource object r such that r ∈ (sbi .sR ∩ rm k .sR), minSBT i (r) =

min(minSBT (r), minRTm (r));

i k

/ i .sR but r ∈ rm
ii. For each resource object r such that r∈sb k .sR, minSBT i (r) =

minRTm (r);
k

iii. The subject sbi gets data from the resource object rmk and sb .sR = sb .sR ∪
i i
k
rm .sR;
b. Otherwise, the get operation is interrupted at the device dk ;
k to a device d .
3. A subject sbi issues a put operation on a resource object rm k
a. For each resource object r such that r ∈ sbi .sR, minRTmk (r) = minSBT i (r);
k and r k .sR = sb .sR;
b. The subject sbi puts the data to the resource object rm m i

minRTmk (r) and minSBT i (r) are a pair of the earliest generation times of data which
are in a resource object r and flow into a resource object rm k and into a subject sb ,
i
respectively. These generation times are updated through manipulating resource objects.
If a subject sb1 issues a get operation on a resource object r1 to a device d1 , the
device d1 checks whether or not every data which flow from the resource object r1 to the
subject sb1 are generated within the validity period of a capability token to get the data.
The condition “gt 1 .st(r) ≤ minRT 1 (r) ≤ gt 1 .et(r)” means that all data of the resource
object r are generated within the validity period. Here, the existence of resource objects
which do not hold this condition, i.e. ∃r ∈ r1 .sR(¬(gt 1 .st(r) ≤ minRT 1 (r) ≤ gt 1 .et(r))),
means that an late information flow (r1 →l sb1 ) occurs from the resource object r1 to
the subject sb1 . Hence, the get operation is interrupted in the TBOI protocol.

Example 1. An example of the TBOI protocol is shown in Fig. 1. Suppose a pair of

devices d1 and d2 and a pair of subjects sb1 and sb2 exist. Here, d1 and d2 are sensor
and hybrid devices, respectively. There are resource objects r1 and r2 in devices d1
and d2 , respectively. The capability token issued a subject sb1 is composed of a pair
of access rights r1 , get and r2 , put and valid from time t1 to time t2 . In addition,
another capability token issued to a subject sb2 includes a pair of access rights r1 , get
and r2 , get and is valid from time t3 to time t4 . Here, IN(sb1 ) = {r1 } and IN(sb2 ) =
{r1 , r2 }.
First, the sensor d1 gets data by sensing events occurring around itself at time τ and
stores the date in a resource object r1 , i.e. r1 .sR = {r1 } and minRT 1 (r1 ) = τ . Next, the
subject sb1 gets data from the resource object r1 in the device d1 because r1 ⇒t sb1
holds. Hence, sb1 .sR = sb1 .sR(= φ ) ∪ r1 .sR(= {r1 }) = {r1 } and minSBT 1 (r1 ) =
minRT 1 (r1 ) = τ . Then, the hybrid device d2 gets data by sensing events occurring
around itself at time τ  and stores the data in a resource object r2 , i.e. r2 .sR = {r2 }
and minRT 2 (r2 ) = τ  . After that, the subject sb1 puts the data to the resource object
r2 in the hybrid device d2 . Here, r2 .sR = r2 .sR(= {r2 }) ∪ sb1 .sR(= {r1 }) = {r1 , r2 }
and minRT 2 (r1 ) = minSBT 1 (r1 ) = τ . Finally, the subject sb2 tries to get the data from
the resource object r2 . Here, a late information flow relation r2 →l sb2 holds because
minRT 2 (r1 )(= τ ) ≤ gt 2 .st(r1 )(= t3 ). Hence, the get operation from the subject sb2 is
interrupted at the hybrid device d2 to prevent late information flow.
 Design and Implementation of the TBOI 131

Fig. 1. TBOI protocol.

5 Implementation of the TBOI Protocol

We implement the TBOI protocol on a hybrid device d1 which is realized on a Rasp-

berry Pi 3 Model B+ [1] whose operating system is Raspbian [2]. A subject sb1 is also
implemented in Python on the hybrid device d1 . The hybrid device d1 and the subject
sb1 are realized as a CoAP server and a CoAP client in CoAPthon3 [14], respectively.

5.1 Capability Token

A capability token is designed to be composed of the following fields referring to a

capability token example shown in the paper [7].

• Identifier (variable size): identifier of the capability token.

• Issued-time (10 bytes): time at which the capability token is issued.
• Issuer (64 bytes): issuer of the capability token, i.e. owner of a device.
• Subject (64 bytes): subject granted the access rights in the capability token.
• Device (variable size): device whose access rights are in the capability token.
• Access rights: set of access rights granted to the subject. Each access right is com-
posed of the following fields.
– Operation (variable size): operation allowed, which is any CoAP (Constrained
Application Protocol) [13] method, e.g. GET and PUT.
– Resource (variable size): resource object of the device which is allowed to
manipulate in the operation which is designated in the previous field.
• Not before (10 bytes): time before which the capability token must not be valid.
132 S. Nakamura et al.

• Not after (10 bytes): time after which the capability token must not be valid.
• Signature (64 bytes): signature of the capability token.
A capability token and a signature of the subject sb1 are included in the payload field of
a CoAP request. An “Issuer” field indicates a public key of the issuer of the capability
token. A “Signature” field shows a signature of the issuer. A “Subject” field contains
a public key of the subject sb1 granted the capability token. Keys and signatures of
issuers and subjects are generated in the ECDSA algorithm implemented as a Python
package [15]. Values of “Issuer”, “Signature”, and “Subject” fields are encoded into
Base64 form. The capability token is valid from time “Not before” to time “Not after”.
“Issued-time”, “Not before”, and “Not after” fields denote Unix time, i.e. the number
of seconds elapsed since 00:00:00 UTC on Jan. 1, 1970.

5.2 Authorization Process

In the TBOI protocol, operations implying illegal or late information flow are inter-
rupted at devices. For this aim, the following authorization process is performed at the
hybrid device d1 if the subject sb1 sends a CoAP request with capability tokens to the
hybrid device d1 . In this paper, we consider a pair of operations, get and put on devices.
If the operation op is put, the subject sb1 includes its set sb1 .sR of source resource
objects in the payload field of the CoAP request.
Step 1. The hybrid device d1 checks if the subject sb1 is allowed to manipulate a
resource object r1 or not in the operation op. If the subject sb1 is issued a capa-
bility token cap1 which satisfies the following conditions, the subject sb1 is allowed
to issue the operation op. Otherwise, the operation op is interrupted.
• The operation op matches “Operation” of the capability token cap1 .
• The destination and the Resource-URI option of the CoAP request match
“Device” and “Resource” of the capability token cap1 , respectively.
• The current time is larger than “Not before” and smaller than “Not after” of the
capability token cap1 .
Step 2. If the operation op is get, the hybrid device d1 checks if the subject sb1 is allowed
to get data from every resource object in r1 .sR or not. If at least one resource object
from which the subject sb1 is not allowed to get data exists in r1 .sR, i.e. r1 → sb1
holds, the operation op is interrupted.
Step 3. If the operation op is get, the hybrid device d1 checks if only the data generated
while the subject sb1 is allowed to get the data flow into the subject sb1 . If at least one
resource object whose data flow late to the subject sb1 exists in r1 .sR, i.e. r1 →l sb1
holds, the operation op is interrupted.
Step 4. A public key of the subject sb1 is in each capability token as a “Subject” filed. A
signature of the subject sb1 is in the payload field of the CoAP request. The hybrid
device d1 checks if the signature of the subject sb1 is valid or not. If the signature is
not valid, the operation op is interrupted.
Step 5. A public key and a signature of its issuer in each capability token are a “Issuer”
and a “Signature” fields, respectively. The hybrid device d1 checks if the signature
of every capability token used in the above steps 1, 2 and 3 is valid or not. If at least
one capability token which is not valid exists, the operation op is interrupted.
 Design and Implementation of the TBOI 133

After the authorization process, sets r1 .sR and sb1 .sR of source resource objects are
updated. If the authorization process is completed for a put operation, the hybrid device
d1 updates a set r1 .sR of its source resource objects with sb1 .sR, i.e. r1 .sR = sb1 .sR. On
the other hand, if the authorization process is completed for a get operation, the hybrid
device d1 sends a CoAP response to the subject sb1 . Here, the hybrid device d1 includes
its set r1 .sR of source resource objects in the payload field of the CoAP response. The
subject sb1 adds the resource objects in the set r1 .sR to its set sb1 .sR of source resource
objects, i.e. sb1 .sR = sb1 .sR ∪ r1 .sR.

6 Evaluation
In the evaluation, first, a subject sb1 puts data to the resource object r1 in the hybrid
device d1 . Here, the data are randomly generated. The data size is also randomly decided
from 80 to 120 [Bytes]. Next, the subject sb1 gets data in the resource object r1 in the
hybrid device d1 . We iterate the above measurement ten times, i.e. a pair of put and
get operations are issued repeatedly ten times, and measure the execution time of every
authorization process. Here, we generate a scenario so that every authorization process
is completed because the total execution time is longest in this case. If any illegal or late
information flow occurs, an authorization process aborts in the middle and the execution
time is always shorter. After the measurements, we calculate the average execution time
of an authorization process.

TBOI
OI
CBAC

CBAC

1 2 3 4 5
sb1

Fig. 2. Execution time for a get operation.

Figure 2 shows the execution time of an authorization process for a get operation.
The horizontal axis is the number of capability tokens issued to the subject sb1 , i.e.
|CAP1 |. In this evaluation, we assume that data are put to the resource object r1 before
134 S. Nakamura et al.

each get operation such that r1 .sR = {rm | m = 1, . . . , |CAP1 |}. Next, we generate CAP1
such that CAP1 = {cap1m | rm , get ∈ cap1m (m = 1, . . . , |CAP1 |)} to complete the autho-
rization process and the subject sb1 issues a get operation on the resource object r1 to
the hybrid device d1 .
The label “CBAC” shows the execution time of an authorization process where only
the CBAC model is adopted, i.e. no information flow control is supported. Here, only
steps 1, 4, and 5 are performed. In the OI protocol, illegal information flow is prevented.
Here, the step 2 is performed in addition to the authorization process of the CBAC
model. In the TBOI protocol, both illegal and late information flows are prevented.
Here, all the steps are performed. As a result, the execution time of the authorization
process in the CBAC model is shorter than the OI and TBOI protocols. On the other
hand, the execution time of the authorization process in the TBOI protocol is almost
the same as the OI protocol. This means, in the TBOI protocol, late information flow
is prevented in addition to illegal information flow although the execution time of the
authorization process is almost the same as the OI protocol.

7 Concluding Remarks
The CBAC (Capability-Based Access Control) model is proposed to make the IoT
(Internet of Things) secure. Here, capability tokens are issued to subjects as access
rights. Since data are exchanged among subjects and resource objects through manip-
ulating resource objects, a subject might obtain data via other subjects and resource
objects even if the subject is not allowed to get the data, i.e. an illegal information
flow occurs. In the OI (Operation Interruption) protocol proposed in our previous stud-
ies, operations implying illegal information flow are interrupted. However, a subject
might get data generated out of validity period of a capability token to get the data.
Here, the data come to the subject late. Hence, the TBOI (Time-Based OI) protocol
was also proposed in our previous studies to interrupt not only operations implying ille-
gal information flow but also operations implying late information flow. In this paper,
we implemented the TBOI protocol on a hybrid device which is realized in Raspberry
Pi 3 Model B+. We evaluated the authorization process of the TBOI protocol in terms of
the execution time. In the evaluation, we showed the late information flow is prevented
in addition to the illegal information flow in the TBOI protocol although the execution
time is almost the same as the OI protocol.

Acknowledgements. This work was supported by Japan Society for the Promotion of Science
(JSPS) KAKENHI Grant Number JP20K23336.

References
1. Raspberry pi 3 model b+. https://www.raspberrypi.org/products/raspberry-pi-3-model-b-
plus/
2. Raspbian, version 10.3, 13 February 2020. https://www.raspbian.org/ (visited on 11 March
2020)
3. Denning, D.E.R.: Cryptography and Data Security. Addison Wesley, Boston (1982)
 Design and Implementation of the TBOI 135

4. Enokido, T., Barolli, V., Takizawa, M.: A legal information flow (LIF) scheduler based on
role-based access control model. Int. J. Comput. Stand. Interfaces 31(5), 906–912 (2009)
5. Gusmeroli, S., Piccione, S., Rotondi, D.: A capability-based security approach to manage
access control in the internet of things. Math. Comput. Model. 58(5–6), 1189–1205 (2013)
6. Hanes, D., Salgueiro, G., Grossetete, P., Barton, R., Henry, J.: IoT Fundamentals: Network-
ing Technologies, Protocols, and Use Cases for the Internet of Things. Cisco Press, Indi-
anapolis (2018)
7. Hernández-Ramos, J.L., Jara, A.J., Marı́n, L., Skarmeta, A.F.: Distributed capability-based
access control for the internet of things. J. Internet Serv. Inf. Secur. 3(3/4), 1–16 (2013)
8. Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: A read-write abortion protocol to
prevent illegal information flow in role-based access control systems. Int. J. Space-Based
Situated Comput. 6(1), 43–53 (2016)
9. Nakamura, S., Enokido, T., Takizawa, M.: Time-based legality of information flow in the
capability-based access control model for the internet of things. Concurrency Comput. Pract.
Experience https://doi.org/10.1002/cpe.5944
10. Nakamura, S., Enokido, T., Takizawa, M.: Information flow control based on the CapBAC
(capability-based access control) model in the IoT. Int. J. Mob. Comput. Multimed. Commun.
10(4), 13–25 (2019)
11. Nakamura, S., Enokido, T., Ogiela, L., Takizawa, M.: Implementation of a device adopt-
ing the OI (operation interruption) protocol to prevent illegal information flow in the IoT.
In: Proceedings of the 9th International Conference on Emerging Internet, Data, and Web
Technologies, pp. 168–179 (2021)
12. Sandhu, R.S., Coyne, E.J., Feinstein, H.L., Youman, C.E.: Role-based access control models.
IEEE Comput. 29(2), 38–47 (1996)
13. Shelby, Z., Hartke, K., Bormann, C.: Constrained application protocol (CoAP). IFTF
Internet-draft, http://tools.ietf.org/html/draft-ietf-core-coap-18 (2013)
14. Tanganelli, G., Vallati, C., Mingozzi, E.: CoAPthon: Easy development of CoAP-based IoT
applications with python. In: IEEE 2nd World Forum on Internet of Things (WF-IoT 2015),
pp. 63–68 (2015)
15. Warner, B.: python-ecdsa-0.11, 11 March 2014. https://github.com/ecdsa/python-ecdsa (vis-
ited on 7 July 2020)
16. Yuan, E., Tong, J.: Attributed based access control (ABAC) for web services. In: Proceed-
ings of the IEEE International Conference on Web Services (ICWS 2005). IEEE Computer
Society (2005)
 How Far Should We Look Back to Achieve
Effective Real-Time Time-Series Anomaly
Detection?

Ming-Chang Lee1(B) , Jia-Chun Lin1 , and Ernst Gunnar Gran1,2

1 Department of Information Security and Communication Technology, Norwegian University
of Science and Technology, 2815 Gjøvik, Norway
{ming-chang.lee,jia-chun.lin,ernst.g.gran}@ntnu.no
2 Simula Research Laboratory, 1364 Fornebu, Norway

Abstract. Anomaly detection is the process of identifying unexpected events or

abnormalities in data, and it has been applied in many different areas such as
system monitoring, fraud detection, healthcare, intrusion detection, etc. Providing
real-time, lightweight, and proactive anomaly detection for time series with nei-
ther human intervention nor domain knowledge could be highly valuable since it
reduces human effort and enables appropriate countermeasures to be undertaken
before a disastrous event occurs. To our knowledge, RePAD (Real-time Proactive
Anomaly Detection algorithm) is a generic approach with all above-mentioned
features. To achieve real-time and lightweight detection, RePAD utilizes Long
Short-Term Memory (LSTM) to detect whether or not each upcoming data point
is anomalous based on short-term historical data points. However, it is unclear that
how different amounts of historical data points affect the performance of RePAD.
Therefore, in this paper, we investigate the impact of different amounts of his-
torical data on RePAD by introducing a set of performance metrics that cover
novel detection accuracy measures, time efficiency, readiness, and resource con-
sumption, etc. Empirical experiments based on real-world time series datasets are
conducted to evaluate RePAD in different scenarios, and the experimental results
are presented and discussed.

Keywords: Anomaly detection · Time series · Unsupervised learning · Machine

learning · LSTM · Look-Back and Predict-Forward strategy

1 Introduction

Anomaly detection refers to any method or technique designed to identify unexpected

events or anomalies in data, and it has been used to discover anomalies in a time series,
which is a series of data points generated continuously and evenly indexed in time order.
During the past decade, a number of approaches and methods have been introduced
for time series anomaly detections such as [2–8]. Many of them require either domain
knowledge (e.g., data patterns, data distributions, pre-labeled training data) or human

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 136–148, 2021.
https://doi.org/10.1007/978-3-030-75100-5_13
 How Far Should We Look Back to Achieve Effective Real-Time 137

intervention (e.g., collecting sufficient training data, pre-labelling training data, config-
uring appropriate hyperparameters, determining proper detection thresholds, etc.). These
requirements consequently limit the applicability of these approaches.
It could be highly valuable to provide an anomaly detection approach that requires
neither human intervention nor domain knowledge because such an approach can signif-
icantly reduce human effort and that it can be easily and immediately applied to detect
anomalies in any time series. In addition, offering real-time and proactive anomaly detec-
tion might be also highly appreciated since it enables appropriate actions and counter-
measures to be undertaken as early as possible. Last but not least, it would be beneficial
to offer a lightweight anomaly detection approach because such an approach can be
deployed on any commodity machine, such as desktops, laptops, and mobile phones.
To our knowledge, RePAD [1] is an approach with all above-mentioned features.
RePAD is a real-time proactive anomaly detection algorithm for streaming time series
based on a special type of recurrent neural networks called Long Short-Term Memory
(LSTM). As soon as RePAD is employed to detect anomalies in a time series, it trains
a LSTM model with short-term historical data points and automatically calculates a
detection threshold to determine if each upcoming data point is anomalous or not. RePAD
also decides if it should retrain its LSTM model based on the most recent short-term
data points so as to adapt to minor pattern change in the time series. By utilizing LSTM
with a simple network structure and training the LSTM model with short-term historical
data points, RePAD demonstrates good, cost-effective, and real-time anomaly detection
performance according to [1]. However, it is unclear from [1] that how different amounts
of historical data points impact the performance of RePAD and that how far RePAD
should look back such that it is able to provide the most effective anomaly detection. By
the most effective, we mean that:

• Anomaly detect can soon start without lengthy preprocessing.

• Anomalies can be detected as early as possible.
• Low false positives and low false negatives.
• Infrequent LSTM model retraining.

In this paper, we conduct a comprehensive study to investigate the performance of

RePAD. The study starts by introducing a set of metrics which cover novel detection
accuracy measures that are appropriate and suitable for proactive detection approaches.
The metrics also cover readiness and time efficiency since achieving lightweight and
real-time anomaly detection as mentioned earlier are essential and desirable. We design
two empirical experiments based on real-world time series datasets provided by the
Numenta Anomaly Benchmark (NAB) [9] to evaluate RePAD in different scenarios.
The experiment results show that RePAD based on long-term historical data points leads
to a long preparation period, frequent LSTM retraining, and long detection time. Most
importantly, RePAD loses its capability to accurately detect anomalies. On the contrary,
choosing short-term historical data points is highly recommended for RePAD since it
enables RePAD to provide the most effective anomaly detection.
The rest of the paper is organized as follows: Sect. 2 introduces the design of RePAD.
Section 3 presents all metrics used to evaluate the performance of RePAD. Section 4
138 M.-C. Lee et al.

presents and discusses evaluation results based on empirical experiments and real-world
time series datasets. In Sect. 5, we conclude this paper and outline future work.

2 RePAD

RePAD uses LSTM with a simple network structure (e.g., one hidden layer with only ten
hidden units) to predict the values of the data points arriving at the next time point based
on the data values observed at the past b continuous time points. Note that b is called
the Look-Back parameter, which is set to 3 in [1]. Figure 1 illustrates the algorithm of
RePAD. Let t be the current time point and t starts from 0, which is the time point when
RePAD is launched. Since b is set to 3, RePAD has to collect three data points observed
at time points 0, 1, and 2, denoted by v0 , v1 , and v2 , respectively. With these three data
points, RePAD at time point 2 is able to train a LSTM model (called M ) and uses M to


predict the next data value, denoted by v3 .

When t separately advances to 3 and 4, implying that the real data-point values v3
and v4 have been observed respectively, RePAD is able to calculate the corresponding
AARE values (i.e., AARE 3 and AARE 4 ) based on Eq. 1:
t  
 vy − vy 


1
AARE t = · (1)
t−b+1 vy
y=b


where vy is the observed data value at time point y, and vy is the forecast
 
data value for
 

v3 −v3 
time point y, and y = b, b + 1, . . . , t. It is clear that AARE 3 = v3 and AARE 4 =
   


  

v3 −v3  v4 −v4 
2· + v4
1
v3 . Recall that AARE stands for Average Absolute Relative Error.

A low AARE value indicates that the predicted values are close to the observed values.
After deriving AARE 3 and AARE 4 , RePAD replaces M with another LSTM model trained
 

with the most three recent data points so as to predict v4 and v5 , respectively.
When t = 5 (i.e., t = 2b − 1), RePAD calculates AARE 5 . At this moment, RePAD
is able to officially detect anomalies since it has already b AARE values (i.e., AARE 3 ,
AARE 4 , and AARE 5 ), which are sufficient to calculate detection threshold thd . In other
words, RePAD requires a preparation period of 2b − 1 time points to get its detection
started. Equation 2 shows how thd is calculated based on the Three-Sigma Rule [2],
which is commonly used for anomaly detection.

thd = μAARE + 3 · σ (2)


where μAARE = t−b+1
1
· tx=b AARE x (i.e., the average AARE of all previously
derived
 AARE values), and σ is the corresponding standard deviation, i.e., σ =
t
x=b (AAREx −μAARE )
2
t−b+1 . If AARE5 is lower than or equal to thd (see line 13 of Fig. 1),
RePAD does not considers v5 anomalous. Hence, it keeps using the same LSTM model


(i.e., M ) to predict the next data point v6 . However, if AARE5 is higher than thd (i.e.,
line 15 holds), it means that either the data pattern of the time series has changed or an
 How Far Should We Look Back to Achieve Effective Real-Time 139

Fig. 1. The algorithm of RePAD.

anomaly happens. To be sure, RePAD retrains an LSTM model with the most b recent


data points (i.e., v2 , v3 , v4 ) to re-predict v5 and re-calculate AARE5 .

If the new AARE5 is lower than or equal to thd (see line 19), RePAD confirms that
the data pattern of the time series has changed and that v5 is not anomalous. In this case,
RePAD replaces M with this new LSTM model to accommodate the change.
On the contrary, if the new AARE5 is still higher than thd (see line 22), RePAD
considers v5 anomalous (see line 23) since the LSTM trained with the most recent data
points is still unable to accurately predict v5 . At this time point, RePAD immediately
reports v5 as an anomaly. The above detection process will repeat over and over again
as time advances.
It is clear that the Look-Back parameter (i.e., b) is a critical parameter to RePAD since
it controls the length of the preparation period and determines the detection performance
of RePAD. In the next section, we investigate how this parameter impacts RePAD.

3 Evaluation Methodology

To comprehensively study the impact of the Look-Back parameter on RePAD, we

consider the following six performance metrics:
140 M.-C. Lee et al.

1. Precision = TP+FP
TP

2. Recall = TP+FN
TP

3. F-score =2 × Precision×Recall
Precision+Recall
4. PP (=2b − 1 time intervals)
5. LSTM retraining ratio = NR
   
N N
t (Tt −μd )
2
t Tt
6. Average detection time μd = N −t+1 and standard deviation σd = N −t+1

Recall that RePAD is capable of proactive anomaly detection, so using traditional

well-known point-wise metrics (such as precision, recall, and F-score) to measure
RePAD is inappropriate. Therefore, we propose an evaluation strategy based on [10] to
capture both proactive detection performance and lasting detection performance. More
specifically, if any anomaly occurring at time point T can be detected within a detection
time window ranging from time point T − K to time point T + K, we say this anomaly
is correctly detected, where K  T . We adopt the above-mentioned strategy together
with precision, recall, and F-score to evaluate the detection performance of RePAD.
In the equations of precision and recall, TP, FP, and FN represent true positive, false
positive, and false negative, respectively. Precision refers to positive predictive value,
i.e., the proportion of positive/anomalous results that truly are positive/anomalous within
the corresponding detection time windows. On the other hand, recall refers to the true
positive rate or sensitivity, which is the ability to correctly identify positive/anomalous
results within the corresponding detection time windows.
The F-score is a measure of a test’s accuracy. It is defined as the weighted harmonic
mean of the precision and recall of the test, i.e., F-score = Recall−1 +Precision
2
−1 =
2 × Precision×Recall
Precision+Recall . The F-score reaches the best value at a value of 1, meaning perfect
precision and recall. The worst F-score would be a value of 0, implying the lowest
precision and the lowest recall.
PP, stands for Preparation Period, is a period of time required by RePAD to derive
the required amount of AARE values for determining the detection threshold. More
specifically, PP is defined as a time period starting when RePAD is launched and ending
when RePAD officially starts its detection function. As mentioned earlier, RePAD starts
at time point 0 and is able to derive the threshold at time point 2b − 1, implying that PP
would be 2b − 1 time intervals. As b increases, the corresponding PP will accordingly
prolong, meaning that RePAD will need more time before it can start detecting anomalies
in the target time series.
Recall that whenever time advances to the next time point, RePAD might need to
retrain its LSTM model. If the retraining is required, such a process will prolong the
time taken by RePAD to decide whether or not the data point collected at that moment is
anomalous. In addition, RePAD might require more computation resources to retrain its
LSTM model. Therefore, the LSTM-retraining ratio (= NR ) is another important indicator
showing the impact of Look-Back parameter on RePAD, where N is the total number of
time points in the target time series, R is the total number of time points at which RePAD
has to retrain its LSTM, and R ≤ N . If the ratio is low, it indicates several things. Firstly,
 How Far Should We Look Back to Achieve Effective Real-Time 141

it means that RePAD does not need frequent LSTM retraining since its LSTM model
is able to accurately predict future data points. Secondly, it further means that RePAD
is able to quickly make prediction and detection whenever time advances. Thirdly, it
implies that RePAD does not need extra computational resources to retrain its LSTM.
Therefore, the lower the ratio, the better efficiency RePAD is able to provide.
Since the goal of RePAD is to offer anomaly detection in real-time, it is essential to
evaluate how soon RePAD is able to decide the anomalousness of every single data point
in the target time series. Consequently, measuring average detection time μd required
by RePAD and the corresponding

standard deviation σd are necessary. Hence, μd can
N
t tT
be derived by equation N −t+1 , where Tt is the time required by RePAD to detect if
vt (i.e., the data point arrived at time point t) is anomalous or not, and N as stated
earlier is the total number of time points in the target time series. More specifically,
Tt is defined as a time period starting when time has advanced to t and ending when
RePAD has determined the anomalousness of vt . Note that if RePAD needs to retrain
its LSTM model at time point t, Tt will also include the required LSTM retraining time.
Apparently, if μd and σd are both low, it means that RePAD is able to make detection
in real-time.
In the next section, we investigate how the Look-Back parameter impacts RePAD
based on all the above-mentioned metrics.

4 Evaluation Results

To investigate how the Look-Back parameter impacts RePAD, we designed four sce-
narios as listed in Table 1. In the first scenario, the Look-Back parameter is set to 3,
following the setting in [1]. In the rest three scenarios, we separately increase the value
to 30, 60, and 90. Note that we did not evaluate RePAD in scenarios where the value of
the Look-Back parameter is less than 3 since RePAD under the settings will not have
sufficient historical data points to learn the pattern of the target time series, leading to
an extremely high LSTM retraining rate according to our experience. In addition, the
network structure of LSTMs in all the four considered scenarios are kept as simple as
possible, i.e., one hidden layer with 10 hidden units.
Two empirical experiments were conducted based on two real-world time series
datasets from the Numenta Anomaly Benchmark [11]. These datasets are called ec2-
cpu-utilization-825cc2 and rds-cpu-utilization-cc0c53. They are abbreviated as CPU-cc2
and CPU-c53 in this paper, respectively. Table 2 lists the details of these two datasets.
Note that the interval time between data points in all these datasets is 5 min. For this
reason, RePAD in all the scenarios also followed the same time interval to perform
anomaly detection. The two experiments were separately performed on a laptop running
MacOS 10.15.1 with 2.6 GHz 6-Core Intel Core i7 and 16 GB DDR4 SDRAM. Note
that we followed [12] and set K to 7 so as to determine the length of the detection time
window. In other words, the detection time window is from time point T − 7 to time
point T + 7 where T is recalled as the time point at which an anomaly occurs. The
performance metrics presented in Sect. 3 were all used in the experiments.
142 M.-C. Lee et al.

Table 1. RePAD in four scenarios. Note that b is the look-back parameter.

Scenario 1 2 3 4
The value of b 3 30 60 90
Number of hidden layers 1 1 1 1
Number of hidden units 10 10 10 10

Table 2. Two real-world time-series datasets used in the experiments.

Dataset Time period # of data points

CPU-cc2 From 2014-04-10, 00:04 to 2014-04-24, 00:09 4032
CPU-c53 From 2014-02-14, 14:30 to 2014-02-28, 14:30 4032

4.1 Experiment 1
Figures 2 and 3 show the detection results of RePAD on the CPU-cc2 dataset and a
close-up result, respectively. Note that this dataset contains two anomalies labeled by
human experts, and these two anomalies are marked as red hollow circles in both figures.
It is clear that RePAD in Scenario 1 (i.e., b = 3) is the only one that is able to detect the
first anomaly on time and the second anomaly proactively. In the rest of the scenarios,
RePAD could not detect the first anomaly at all. In addition, RePAD in Scenario 1 has
more true positives and less false negatives than RePAD in the other scenarios. This
phenomenon can be clearly observed from Fig. 3, and this is the reason why the recall
of RePAD in Scenario 1 reached the best value of 1 (as listed in Table 3). If we further
take precision and F-score into consideration (see Table 3 as well), we can see that
RePAD in Scenario 1 outperforms RePAD in the other scenarios, implying that setting
the Look-Back parameter to be 3 seems appropriate for RePAD.
Table 4 lists the PP required by RePAD on the CPU-cc2 dataset. In scenario 1 (i.e.,
b = 3), the required PP consists of only 5 (=2 ∗ 3 − 1) time intervals according to
PP = 2b − 1. It means that after 5 time intervals, RePAD can officially start detecting
anomalies in the time series. When we increased the value of the Look-Back parameter,
the corresponding PP proportionally prolonged, implying that RePAD needed more time
to get itself ready. In other words, RePAD loses its capability for providing readiness
anomaly detection.
Table 5 summaries the retraining performance of RePAD on the CPU-cc2 dataset.
We can see that RePAD in a scenario 1 only needs to retrain its LSTM model 83 times.
The corresponding LSTM retraining ratio is very low since it is around 2% (=83/4027).
However, when we increased the value of the Look-Back parameter, the number of
LSTM retraining required by RePAD increased. Apparently, increasing the value of the
Look-Back parameter leads to more frequent LSTM retraining. In other words, training
the LSTM of RePAD with long-term historical data points does not help RePAD in
learning the data pattern of the CPU-cc2 dataset or making accurate predictions. In
addition, as Table 6 shows, the average detection time required by RePAD to decide the
 How Far Should We Look Back to Achieve Effective Real-Time 143

Fig. 2. The detection results of RePAD on the CPU-cc2 dataset. Note that this dataset has two
labeled anomalies, marked as red hollow circles.

Fig. 3. A close-up of the detection results on the CPU-cc2 dataset. The two anomalies were
marked as red hollow circles.

anomalousness of each data point increases when the value of the Look-Back parameter
increases. The main reason is that the detection time includes both the corresponding
LSTM retraining time (if the retraining is required) and the corresponding detection
time. Since RePAD in Scenario 1 requires the least number of LSTM retraining and
utilizes the least amount of training data (i.e., 3 data points), it leads to the shortest
average detection time.
Based on the above results, we conclude that RePAD in Scenario 1 outperforms
RePAD in the other three scenarios since it provides the best detection accuracy and
time efficiency.
144 M.-C. Lee et al.

Table 3. The detection accuracy of RePAD on the CPU-cc2 dataset.

Scenario Precision Recall F-score

1 0.4259 1 0.5973
2 0.0919 0.5 0.1553
3 0.0523 0.5 0.0947
4 0.0667 0.5 0.1176

Table 4. The PP required by RePAD in different scenarios.

Scenario 1 2 3 4
PP (preparation period) 5 59 119 179

Table 5. The LSTM retraining performance of RePAD on the CPU-cc2 dataset.

Scenario # of required LSTM LSTM retraining ratio

retraining
1 83 2%
2 119 2.97%
3 199 5.01%
4 257 6.5%

Table 6. The time consumption of RePAD on the CPU-cc2 dataset.

Scenario Average detection time Standard deviation (sec)

(sec)
1 0.015 0.027
2 0.188 0.339
3 0.385 0.691
4 0.494 0.846

4.2 Experiment 2
In the second experiment, we employ RePAD in the four scenarios to detect anomalies
on the CPU-c53 dataset. The detection results and a close-up result are shown in Figs. 4
and 5, respectively. Note that this dataset contains two anomalies labeled by human
experts and marked as red hollow circles. Table 7 summaries the detection accuracy of
RePAD. Apparently, RePAD in Scenario 1 is the only one that is able to detect the two
anomalies within the detection time windows without generating many false positives,
 How Far Should We Look Back to Achieve Effective Real-Time 145

consequently leading to higher precision, recall, and F-score than RePAD in the rest of
the scenarios. Note that both Scenarios 3 and 4 led to zero precision and recall, so it is
impossible to calculate the corresponding F-scores.

Fig. 4. The detection results of RePAD on the CPU-c53 dataset. Note that the two anomalies are
marked as red hollow circles.

Fig. 5. A close-up of the detection results on the CPU-c53 dataset.

When it comes to PP, RePAD on the CPU-c53 dataset requires exactly the same PP
as those as listed in Table 4. This is because the value of the Look-Back parameter is the
only factor determining PP (i.e., 2b − 1 time intervals). It is worth noting that the LSTM
retraining ratio in this experiment does not always increase as the value of the Look-Back
146 M.-C. Lee et al.

parameter increases. As listed in Table 8, RePAD in Scenario 3 requires the least number
of LSTM retraining, whereas RePAD in Scenario 2 requires the highest number of LSTM
retraining. In other words, the Look-Back parameter is not the only factor determining
LSTM retraining ratio. The pattern of the target dataset is also influential. Nevertheless,
we can see from Table 9 that the average detection time of RePAD increases when the
value of the Look-Back parameter increases since the average detection time depends
on not only LSTM retraining ratio, but also the amount of data points used to retrain the
LSTM.
Overall speaking, Scenario 1 leads to the shortest PP (Please see Table 4), the second
least amount of LSTM retraining (see Table 8), and the shortest average detection time
with the least standard deviation (see Table 9). In other words, configuring the Look-
Back parameter to be 3 enables RePAD to promptly start anomaly detection and detect
anomalies in real-time.

Table 7. The detection accuracy of RePAD on the CPU-c53 dataset.

Scenario Precision Recall F-score

1 0.457 1 0.627
2 0.1569 0.5 0.2388
3 0 0 n/a
4 0 0 n/a

Table 8. The LSTM retraining performance of RePAD on the CPU-c53 dataset.

Scenario # of required LSTM LSTM retraining ratio

retraining
1 59 1.46%
2 82 2.05%
3 52 1.31%
4 73 1.85%
 How Far Should We Look Back to Achieve Effective Real-Time 147

Table 9. The time consumption of RePAD on the CPU-c53 dataset.

Scenario Average detection time Standard deviation (sec)

(sec)
1 0.015 0.024
2 0.122 0.199
3 0.213 0.304
4 0.321 0.522

5 Conclusion and Future Work

In this paper, we have investigated how the Look-Back parameter influences the per-
formance of RePAD by introducing a set of performance metrics that not only evaluate
detection accuracy for novel proactive anomaly detection approaches, but also evaluate
readiness and time efficiency. Regardless of which real-world dataset we utilized in our
experiments, setting a lower value for the Look-Back parameter always leads to bet-
ter performance on all the six considered metrics. More specifically, taking short-term
historic data points as online training data is an appropriate strategy for RePAD since
RePAD is able to promptly learn and adapt to the changing patterns of the target time
series. Furthermore, using short-term historic data points also enables RePAD to soon
start detecting anomalies in the target time series and detect potential anomalies in real-
time. In all the scenarios discussed in this paper, scenario 1 (i.e., setting the Look-Back
parameter to 3) is the most recommended since it enables RePAD to provide real-time,
lightweight, and proactive anomaly detection for time series without requiring human
intervention or domain knowledge.
As future work, we plan to improve RePAD by further reducing false positives.
Furthermore, we would like to extend RePAD to detect anomalies on time series data
observed from the eX3 HPC cluster [13] by referring to [14, 15, 16] and designing the
methodology in a parallel and distributed way.

Acknowledgments. This work was supported by the project eX3 - Experimental Infrastructure
for Exploration of Exascale Computing funded by the Research Council of Norway under contract
270053 and the scholarship under project number 80430060 supported by Norwegian University
of Science and Technology.

References
1. Lee, M.-C., Lin, J.-C., Gran, E.G.: RePAD: real-time proactive anomaly detection for time
series. In: Proceedings of the 34th International Conference on Advanced Information
Networking and Applications (AINA 2020), pp. 1291–1302 (2020)
2. Hochenbaum, J., Vallis, O.S., Kejariwal, A.: Automatic anomaly detection in the cloud via
statistical learning. arXiv preprint arXiv:1704.07706 (2017)
148 M.-C. Lee et al.

3. Aggarwal, C.C., Yu, P.S.: Outlier detection with uncertain data. In: Proceedings of the
2008 SIAM International Conference on Data Mining. Society for Industrial and Applied
Mathematics, pp. 483–493 (2008)
4. Xu, J., Shelton, C.R.: Intrusion detection using continuous time Bayesian networks. J. Artif.
Intell. Res. 39, 745–774 (2010)
5. Fisher, W.D., Camp, T.K., Krzhizhanovskaya, V.V.: Crack detection in earth dam and levee
passive seismic data using support vector machines. Procedia Comput. Sci. 80, 577–586
(2016)
6. Wu, J., Zeng, W., Yan, F.: Hierarchical temporal memory method for time-series-based
anomaly detection. Neurocomputing 273, 535–546 (2018)
7. Staudemeyer, R.C.: Applying long short-term memory recurrent neural networks to intrusion
detection. South African Comput. J. 56(1), 136–154 (2015)
8. Bontemps, L., McDermott, J., Le-Khac, N.A.: Collective anomaly detection based on long
short-term memory recurrent neural networks. In: International Conference on Future Data
and Security Engineering, pp. 141–152. Springer, Cham, November 2016
9. Lavin, A., Ahmad, S.: Evaluating real-time anomaly detection algorithms – the numenta
anomaly benchmark. In: 14th International Conference on Machine Learning and Applica-
tions (2015).
10. Xu, H., et al.: Unsupervised anomaly detection via variational auto-encoder for seasonal KPIs
in web applications. In: Proceedings of the 2018 World Wide Web Conference, pp. 187–196
(2018)
11. numenta/NAB.: The Numenta Anomaly Benchmark [Online code repository]. https://github.
com/numenta/NAB Accessed 03 Mar 2021
12. Ren, H., et al.: Time-series anomaly detection service at microsoft. In: Proceedings of the
25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,
pp. 3009–3017 (2019)
13. Simula Research Laboratory: the eX3 research infrastructure. https://www.ex3.simula.no
Accessed 03 Mar 2021
14. Lin, J.-C., Lee, M.-C.: Performance evaluation of job schedulers under hadoop YARN.
Concurrency Computat. Pract. Exper. (CCPE) 28(9), 2711–2728 (2016)
15. Lee, M.-C., Lin, J.-C., Yahyapour, R.: Hybrid job-driven scheduling for virtual mapreduce
clusters. IEEE Trans. Parallel Distrib. Syst. (TPDS) 27(6), 1687–1699 (2016)
16. Lee, M.-C., Lin, J.-C., Gran, E.G.: ReRe: a lightweight real-time ready-to-go anomaly detec-
tion approach for time series. In: Proceedings of the 44th IEEE Computer Society Signa-
ture Conference on Computers, Software, and Applications (COMPSAC 2020), pp. 322–327
(2020)
 A Novel Hybrid Recommendation Approach
Based on Correlation and Co-occurrence
Between Activities Within Social Learning
Network

Sonia Souabi(B) , Asmaâ Retbi, Mohammed Khalidi Idrissi, and Samir Bennani

RIME TEAM-Networking, Modelling and e-Learning Team, MASI Laboratory, Engineering.3S

Research Center, Mohammadia School of Engineers (EMI), Mohammed V University,
Rabat, Morocco
[email protected], {retbi,khalidi,
sbennani}@emi.ac.ma

Abstract. Social learning is considered one of the most prevalent disciplines

within e-learning. To handle learning resources within social environments, rec-
ommendation systems are gaining tremendous prominence based on a series of
criteria such as the rate of learner interaction with the learning environment. On the
basis of this, we highlight an overriding issue focusing on integrating the variety of
activities carried out by learners within the learning environment. The challenge
with most recommendation systems is that they do not address multiple activi-
ties performed by learners which may significantly affect recommendations. In
our study, we focus on this point while proposing a hybrid system combining the
two parameters of correlation and co-occurrence. The objective is to pinpoint the
degree of connectivity between activities and their influence on recommendations.
We carried out our study on an available database carried out by a Chinese team.
We thus compare the results of the hybrid system with the non-hybrid system based
solely on co-occurrence by measuring the performance of the two approaches. It is
demonstrated that the proposed hybrid system produces highly gratifying results
compared to the non-hybrid system.

Keywords: Social learning · Recommendation systems · Hybrid system ·

Correlation · Co-occurrence

1 Introduction
To assist the user in the information research, recommendation systems are designed
to efficiently find and obtain relevant information [1]. In terms of social learning, it is
noteworthy to mention that a recommendation system plays a highly significant role in
the management of learning resources for learners [2]. Several approaches have been dis-
cussed: Collaborative filtering approach, hybrid approaches, content-based approaches
(…) [3]. Generally, recommendation systems tend to manifest themselves in several
forms: recommendation systems based on explicit data and recommendation systems

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 149–162, 2021.
https://doi.org/10.1007/978-3-030-75100-5_14
150 S. Souabi et al.

based on implicit data. Explicit data are collected on the basis of learner-defined annota-
tions, while implicit data reflect the real traces of learners by collecting the frequency of
events performed, as well as actions not performed [4]. Thus, implicit data are relevant
in terms of recommendations as they consider the actual interactions of learners with the
learning environment. In the context of learning, a recommendation approach has been
advocated based on implicit learner attributes [5]. Other approaches merge implicit and
explicit learner data to design a new recommendation system [6]. However, the implicit
data has not been adequately addressed in the e-learning recommendation systems. Most
approaches proposed still rely on the ratings assigned by learners to generate recommen-
dations, whereas ratings are still insufficient to generate relevant recommendations. On
the other hand, most recommendation systems proposed in online learning do not focus
on the overall activities of learners to generate more appropriate recommendations. The
focus is limited to a specific activity in many cases, except for the ratings provided by
the learners. There is therefore a need to address this issue and to integrate the majority
of the activities carried out by the learners into the generation of recommendations.
In order to highlight a very critical aspect that has not been addressed, we tackle
an insightful issue where we can study the impact of carrying out multiple actions
on the various recommendations. Implicit data approaches do not handle this point in
relation to the effect of the activities undertaken on the recommendations. To further
illustrate this point, we would identify the link between the mastery activity related to
the recommendations and the other secondary activities of the learners recorded within
the learning environment. This link is configured by two parameters, the first one, which
is the correlation parameter, addresses the frequency of each action independently of
another action, and the second one, the co-occurrence parameter, expresses how many
times two activities appears simultaneously. Instead of considering only one parameter,
we focus on both simultaneously, covering the totality of aspects that may exist between
two actions performed. To prove how the combination of correlation and co-occurrence
is influencing the performance of the recommendation system, we drew a comparison
between the results obtained by the hybrid system and the results obtained by the non-
hybrid system. The objective is to warrant that the integration of correlation and co-
occurrence leads to better accuracy in terms of recommendations. For a recommendation
system to be adapted for a learning social network, the dynamism of the learners must be
included as well as all the actions carried out. In order for a recommendation system to
be suitable for the genuine needs of learners, it is important to pay attention to the degree
of learner interactivity within the learning community, including which activities were
performed the most, in which categories learners are more engaged, in which categories
are less engaged or not engaged at all. All these questions will play a very prominent
position in generating recommendations appropriate to learners’ requirements. In this
respect, we propose a recommendation system that focuses on implicit data and endure
a hybrid character. Two far-reaching factors are involved: correlation and co-occurrence
to ensure that the calculation of recommendations includes the two factors. We carry out
the test based on an online database containing a large number of activities recorded in
the learning environment.
The remaining part of the document is organized as follows: Sect. 2 outlines high-
lights some recommendation systems proposed in the literature; Sect. 3 outlines the
 A Novel Hybrid Recommendation Approach Based on Correlation 151

materials and methods we used including the recommendation strategy we proposed,

Sect. 4 focuses on analysing the evaluation results and Sect. 5 draws the conclusion in
association with future work.

2 Background
2.1 Social Recommender Systems

Shokeen and Rana [7] first describe social recommendation systems. They define this
type of systems as the integration of recommendation systems within social media and
networks. Recommendation systems, therefore, play a major role in improving the qual-
ity of recommendations. On the other hand, the article introduces the different charac-
teristics of a social recommendation system, including: context, trust, tag, group, cross
social media, temporal dynamics, heterogeneous social connections, semantic filtering.
Guy [8] provide a general vision of social recommendation systems. Several types of
recommendations can be defined within social media and networks: group recommenda-
tions allowing to generate recommendations for a group of people instead of individually,
people recommendation, people recommendation to follow. Ahmadian et al. [9] propose
a social recommendation system based on reliable implicit relationships. The proposed
recommendation system consists of several steps. First, a network between users is gen-
erated based on annotations. Then, a link prediction approach is adopted to predict the
implicit relationships between users, with an evaluation of prediction quality. Finally,
the new predicted data are used to reconstruct a network with new neighbors. Rezvanian
et al. [10] suggest analyzing two types of existing recommendations: CALA-OptMF
and LTRS. The study reveals that each approach has its own advantages. CALA-OptMF
improves the accuracy of recommendations in fuzzy recommendation systems. As for
the LTRS, the results confirm that it can effectively manage scoring problems for users
of scarce data and cold start. V. R. Revathy and Anitha S. Pillai [11] discuss the appli-
cation and classification of recommendation systems within several known social net-
works. Social recommendation systems have been classified into two broad categories:
entity-based and audience-based recommendation systems. Entity-based recommenda-
tion systems can be based on tags, context, communities, social influence. For public
recommendation systems, they are generally based on friendships and groups.

2.2 Related Previous Studies

Several recommendation approaches were adapted regarding online learning, including
collaborative filtering and hybrid approaches. The main purpose is to ensure the avail-
ability of pedagogical material to learners while responding to their specific needs. R.
Turnip et al. [12] draw on the history of learners’ behaviour and activities in order to
provide them with a set of recommendations appropriate to their needs. S. P. Perumal
et al. [13] propose a recommendation system based on the learner’s profile based on the
k-means algorithm. M. H. Mohamed et al. [14] aim to provide learners with the right
courses based on the application of association rules to all transactions within the learn-
ing environment. J. Wintrode et al. [15] evaluated recommendation system techniques
152 S. Souabi et al.

within the social learning platform through user-centred and data-centred evaluations.
They relied on several algorithms, namely KNN, graph based algorithm and matrix fac-
torization. N. Mustafa et al. [16] develop a recommendation system based on the study
of graphs by taking into consideration the social interactions between learners. Chiefly,
studies led by researchers concerning recommendation systems focus particularly on
learners’ profiles, without looking closely at the actions performed by learners as an
entity. These studies are focusing especially on explicit data more than implicit data. In
our proposition, we intend to measure the influence of activities carried out by learners
on the recommendations generated and to integrate the correlation and the co-occurrence
between events for calculating recommendations.

3 Our Recommendation Approach

Our recommendation system is summarized in the following figure (Fig. 1). The main
components of the model are:

• The history of the learners: These are the activities and actions carried out by learners
within the learning platform. These activities are divided into two categories:
• The primary activity: which is directly associated with the recommendations gen-
erated, i.e. the indicator conveying the learners’ interest towards the learning
objects.
• Secondary activities: which are related to the recommendations, although indirectly.
These are also indicators representing the learners’ interest towards the learning items,
however they come in second position in respect to the primary activity.
• Data restructuring: This part is dedicated to data restructuring. Indeed, the data are
structured by learner, i.e. each data expresses an activity carried out by a certain
learner. The objective is to convert the data structure per learner into a data structure
per learning object. The occurrence frequency of each activity will be collected for
each learning object from the original database.
• Recommendation part: After restructuring our database to suit our requirements, we
will generate the recommendations based on two notions:

• The correlation between the activities performed by the learners.

• The existing co-occurrence between the activities performed by the learners.

The top recommendations will therefore be generated from the recommendation

scores based on the correlation score and the co-occurrence score.
To calculate the total recommendation score for each item, we calculate the correla-
tion score on the one hand and the co-occurrence score on the other hand. The correlation
score is obtained from the correlation matrix, while the co-occurrence score is obtained
from the co-occurrence matrix.
To explain the two parameters of correlation and co-occurrence, we shall consider
the following data:

M1 : The activity or primary action

 A Novel Hybrid Recommendation Approach Based on Correlation 153

Fig. 1. The hybrid recommendation model

{M2 , M3 , . . . , Mi }: The other activities or actions (secondary activities)

{c1 , c2 , . . . , cn }:Items to recommend
s1j , s2j , . . . , sij : The history of the activities or the number of times the activities
carried
 out for each  learning object
p1j , p2j , . . . , pij : The frequency of activity appearance

• Correlation: The correlation between two variables or two activities in our con-
text is a notion of connection that contradicts independence. The more dependent
two variables are, the higher the correlation coefficient between. We choose the
Spearman correlation because it’s the most appropriate measure for our case which
describes a non-normal distribution.
• Co-occurrence: This notion of connection makes it possible to ascertain how many
times two activities have appeared together.
154 S. Souabi et al.

Thus, to calculate the total recommendation score, we combine the correlation score
(Eq. 1) and the co-occurrence score (Eq. 2) so as to obtain the pursuing results (Eq. 3):

{j1 , j2 , . . . , jn } : Learning objects to recommend .

{o1 , o2 , . . . , om } : Activities performed by learners such as o1 is the primary
activity and {o2 , . .. , om } are
  the secondaryactivities
pj1 o1 , . . . , pj1 om , . . . , pjn o1 , . . . , pjn om : Learners historical activities towards
each
 learning object.   
Pj1 o1 , . . . , Pj1 om , . . . , Pjn o1 , . . . , Pjn om : Binary history of activities performed
by learners for each learning object.

⎡ ⎤
pj1 o1 pj1 o2 . . . pj1 om
⎢ ⎥
Correlation score matrix = ⎣ pj2 o1 pj2 o2 . . . pj2 om ⎦ × 1 cor(o1 , o2 ) . . . cor(o1 , om )
.. .. ..
. . .
⎛ ⎞
correlation score(j1 )
⎜ .. ⎟
=⎝ . ⎠ (1)
correlation score(jn )
⎡ ⎤
Pj1 o1 Pj1 o2 . . . Pj1 om
⎢ ⎥
Co − occurrence score matrix = ⎣ Pj2 o1 Pj2 o2 . . . Pj2 om ⎦
.. .. ..
. . .
× 1 co − occ(o1 , o2 ) . . . co − occ(o1 , om )
⎛ ⎞
co − occurrence score(j1 )
⎜ .. ⎟
=⎝ . ⎠ (2)
co − occurrence score(jn )

Total score matrix = co − occurrence score matrix + correlation score matrix

⎛ ⎞
correlation score(j1 ) + co − occurrence score(j1 )
⎜ .. ⎟
= ⎝ . ⎠ (3)
correlation score(jn ) + co − occurrence score(jn )
 A Novel Hybrid Recommendation Approach Based on Correlation 155

3.1 The Hybrid Recommendation Algorithm

In order to generate the top recommendations for learners based on the correlation and co-
occurrence between the activities carried out by the learners, we embrace the following
algorithm step by step (Algorithm 1).

4 Experiments and Evaluations

Experiments were conducted to evaluate the performance of our proposed recommen-
dation system for recommending learning objects.

4.1 Experimental Setup

The database we will focus on in our experiment is a dataset extracted from a video-based
educational experience using a social and collaborative platform1 . The interdisciplinary
learning activity is carried out between students in computer engineering and media
and communication. The collaborative social network is divided into groups, with each
group incorporating computer engineering and media and communication students. We
opted for this database as it fits perfectly with our context and our expectations, and
it handles all the activities carried out by the learners within the social network while
providing them with several materials, including: documents, videos, presentations. The
social network is, therefore, harnessed for a purely educational context supporting the
assessment process. Students have a workspace where they can share files, images, and
various resources, as well as messaging to interact with other students. The exchange,
therefore, consists of sharing several types of educational materials (videos, ppt, word,
images). The study involves 111 students.
1 https://bera-journals.onlinelibrary.wiley.com/doi/full/10.1111/bjet.12318.
156 S. Souabi et al.

As our hybrid recommendation system is part of educational social networks, we

will carry out our study on this already available database in a different perspective,
namely that of recommendation systems. Our recommendation approach consists of
including activities carried out by learners within a social learning network. Therefore,
it is important to identify the activities included before proceeding with the operation of
calculating recommendations based on correlation and co-occurrence.
In order to enhance the value of our work and based on the available data, we focus
our study on the recommendations of videos contained in blogs. We will, thus, look at
any action or activity directly concerned by the videos:

• The evaluation of videos by learners (fivestar).

• Creating a comment in a video (generic_comment).

Many activities have been recorded in this database. We restricted our work to rel-
evant activities for recommending videos. Indeed, the video is a very practical support
to illustrate certain notions. It is one of the strongest techniques for learning as it is
supported by images and sound, and these two elements fully catch a learner’s attention.
Since the correlation between the primary activity associated with the recommendations
and the other secondary activities, as well as co-occurrence, must be included, the pri-
mary activity must be identified in addition to the secondary activities whose relevance
comes afterwards (Table 1).

Table 1. Primary activity vs secondary activities

Primary activity Secondary activities

The evaluation of videos by learners (fivestar) Creating a comment in a video

4.2 Experimental Case

• List of learning objects

Several learning elements are available within the database to be analyzed (Table 2).
 A Novel Hybrid Recommendation Approach Based on Correlation 157

Table 2. Learning objects

Learning objects ID
1 3983
2 4024
3 3916
4 3887
.. ..
. .

• Original matrix

The original matrix is the matrix outlining the interaction of the different learning objects
with the activities performed by learners. Here is an extract from this matrix (Table 3):

Table 3. Learning objects actions interaction

Learning object ID Update fivestar Create generic_comment

3983 41 48
4024 24 43
3916 29 43
3887 11 38

• Training model

Firstly, we are supposed to find the recommendation model based on our approach refer-
ring to correlation and co-occurrence. After performing the corresponding calculations,
we obtain the following recommendation model (Eq. 4):

Total scores = H1 × M1 + h1 × m1 (4)

Such as:

H1 : history matrix
h1 : binary history matrix  
1
M1 = correlation vector =
0, 84
 
1
m1 = co − occurrence vector =
0, 8
158 S. Souabi et al.

• Comparing training results to validation results

As the training model is obtained, the validation part is used to measure the performance
of our recommendation system. On the latter, we compare the results obtained by the
training model to the results obtained by the validation model, and then we use the
required measurements. Here is an extract of the validation results obtained (Table 4).
In the following section, we discuss the evaluation process under consideration.

Table 4. Total scores of some learning objects to recommend

Learning objects Total score

Learning object 1 83,09
Learning object 2 61,9
Learning object 3 66,89
Learning object 4 44,7
Learning object 5 60,86
Learning object 6 51,37
Learning object 7 3

4.3 Experimental Results

We choose the online evaluation method based on the basic truths deduced. Thus, the
basic truths deduced reveal the real preferences and orientations of a learner towards
different object materials. We rely on this type of assessment to determine the point
reached by our recommendation system in predicting the real preferences of the learners.
In our tests, we compare the results of the assessment measures between the hybrid
recommendation system and the non-hybrid recommendation system based solely on
co-occurrence.

• Evaluation method
Our analysis consists of dividing the database into two parts:

• Part 1 (80% of the data): To test the hybrid recommendation approach by generating
group recommendations for learners.
• Part 2 (20% of the data): To validate our recommendation system by comparing
the results generated in part 1 to the actual preferences of the learners in part 2.
• Performance experiments

To calculate performance (Eq. 5), we will not include the recommendation ranking,
simply whether or not the generated recommendations appear in the list of real learner
preferences identified in the 20%. The more the recommendations generated in the
 A Novel Hybrid Recommendation Approach Based on Correlation 159

first part of the data overlap with those in the second part of the data, the higher the
performance of the recommendation system will be. Accuracy measures the ability of
the system to predict learners’ real preferences (Table 5).

Table 5. Interpretation of accuracy parameters

Recommended Not recommended

Preferred True positive (TP) False positive (FP)
Not preferred False negative (FN) True negative (TN)

The system performance is expressed by the F1 measurement. This measurement is

derived from precision and recall using the following formula (Eq. 5):
Precision × recall
Performance = F1 − measure = 2 ×
Precision + recall
= Precision(Precision = recall) (5)

The accuracy is equal to the recall and the F1 measurement, so only one measurement
needs to be calculated to identify the values of the other parameters. In this case we
consider the traditional formula of the precision (Eq. 6):

TP Number of correct items recommended

Precision = = (6)
TP + FP Number of recommendations

Table 6. Precision according to number of recommendations for hybrid and non-hybrid

recommender systems

Number of recommendations Precision of hybrid approach Precision of non-hybrid

approach
N=4 0,75 0,5
N=6 1 0,8
N=5 0,83 0,66

On this basis, we obtain the following graph for the two recommendation systems
(Fig. 2):
Similarly, we realize that the performance of the hybrid recommendation system
outperforms that of the fully co-occurrence-based approach. The performance associated
with the proposed recommendation system reaches values between 75% and 100%, while
the performance associated with the non-hybrid approach is limited to 80% with a lowest
value of 50% (Table 6).
160 S. Souabi et al.

Fig. 2. Precision of the hybrid recommender system vs the co-occurrence approach based on the
number of recommendations.

5 Discussion
After evaluating the results, it turns out that the performance of the hybrid recommen-
dation system outperforms the performance of the non-hybrid recommendation system
based solely on co-occurrence. This amounts to the integration of the correlation concept
between the activities carried out by learners. The discrepancy between the performance
of the hybrid recommendation system and the performance of the non-hybrid recom-
mendation system stands between 25% and 30%, which reveals a significant gap in
recommendation systems. Moreover, the maximal performance of the hybrid recom-
mendation system reaches 100%, while the maximum value of the recommendation
system based solely on co-occurrence is limited to 80%.
Our recommendation system achieves an accuracy rate between 75% and 100%,
meaning that the recommended videos meet the real preferences of the learners with
a percentage between 75% and 100%, unlike the non-hybrid recommendation system,
which reaches a lower level of satisfaction than the hybrid system. Correlation allows
the recommendation system to be improved and its added value to be enhanced.
This entails integrating the two matrices together: the correlation matrix and the co-
occurrence matrix. On the one hand, we measure the correlation between the different
activities within the learning environment, and then we ascertain how often the activities
appear together from the historical data. In the literature, the hybrid recommendation
system is generally still the predominant one, as it produces prominent results than
non-hybrid approaches. The calculation time may be higher, but the results are actually
better. Depending on the learners’ activity within a social learning platform, we will be
led to monitor the learner’s behavior and interactivity with the pedagogical resources
and in discussion forums with other learners. As the number of activities increases, the
recommendation system gains efficiency. When learners are more interactive, it helps us
to propose more appropriate recommendations for the learner profile and needs. We are
witnessing a major technological evolution over the years, so we need to improve the
quality of the recommendations provided to users, especially learners, to improve their
level of knowledge and engagement. The more satisfied the learner is with the recommen-
dations, the more engaged they will be, leading to an improvement in their level. In the
 A Novel Hybrid Recommendation Approach Based on Correlation 161

literature, most research has focused on a single activity to generate recommendations.

In turn, we embedded several activities simultaneously into a single recommendation
system. On the other hand, we measured two key parameters to identify the connectivity
between activities. Correlation is not the unique factor that can measure the relationship
between two activities, but also co-occurrence, which indicates how many times two
activities were performed together.
In our case, we considered only a few activities. If learners are more interactive
and other activities are registered within the learning environment, our proposed rec-
ommendation system will always be worthwhile regardless of the number of activities
registered.

6 Conclusion

In e-learning, particularly in social learning via social networks, recommendation sys-

tems play a very important role in managing the large amount of data available to
learners. However, the recommendation systems already proposed in the literature do
not support the integration of activities carried out by learners to generate more appro-
priate recommendations. In our approach, we propose a hybrid system that includes
all learner-performed activities and combines correlation and co-occurrence between
the different activities to measure the connection strength between these activities and
adapt recommendations to learners’ level of interaction. Experimental results show that
our hybrid recommendation system generates higher accuracy for both universities com-
pared to the non-hybrid system, proving the effectiveness of a hybrid system that merges
correlation and co-occurrence into a unified system. Thus, it was possible to also inte-
grate several types of learner activities in order to provide recommendations according
to learner needs. This recommendation system therefore brings several advantages in
terms of e-learning as well as the engineering education. Our future work consists of:

• Testing our recommendation system within our university.

• Integrating other learner activities to generate prominent recommendations.

References
1. Kulkarni, P.V., Rai, S., Kale, R.: Recommender system in eLearning: a survey. In: Bhalla,
S., Kwan, P., Bedekar, M., Phalnikar, R., Sirsikar, S. (eds.) Proceeding of International
Conference on Computational Science and Applications, pp. 119–126. Springer, Singapore
(2020)
2. Abbas, M., Riaz, M.U., Rauf, A., Khan, M.T., Khalid, S.: Context-aware Youtube recom-
mender system. In: 2017 International Conference on Information and Communication Tech-
nologies (ICICT), Karachi, December 2017, pp. 161–164 (2017). https://doi.org/10.1109/
ICICT.2017.8320183
3. Lund, J., Ng, Y.-K.: Movie recommendations using the deep learning approach. In: 2018
IEEE International Conference on Information Reuse and Integration (IRI), Salt Lake City,
UT, July 2018, pp. 47–54 (2018). https://doi.org/10.1109/IRI.2018.00015
162 S. Souabi et al.

4. Wang, X., Wang, C.: Recommendation system of ecommerce based on improved collaborative
filtering algorithm. In: 2017 8th IEEE International Conference on Software Engineering and
Service Science (2017)
5. Pujahari, A., Padmanabhan, V.: An approach to content based recommender systems using
decision list based classification with k-DNF rule set. In: 2014 International Conference on
Information Technology, Bhubaneswar, India, December 2014, pp. 260–263 (2014). https://
doi.org/10.1109/ICIT.2014.13
6. Zhang, R., Liu, Q., Chun-Gui, Wei, J.-X., Huiyi-Ma: Collaborative filtering for recom-
mender systems. In: 2014 Second International Conference on Advanced Cloud and Big
Data, Huangshan, China, November 2014, pp. 301 308 (2014). https://doi.org/10.1109/CBD.
2014.47
7. Shokeen, J., Rana, C.: A study on features of social recommender systems. Artif Intell Rev
53, 965–988 (2020). https://doi.org/10.1007/s10462-019-09684-w
8. Guy, I.: Social recommender systems. In: Ricci, F., Rokach, L., Shapira, B. (eds.) Recom-
mender Systems Handbook, pp. 511–543. Springer, Boston (2015). https://doi.org/10.1007/
978-1-4899-7637-6_15
9. Ahmadian, S., Joorabloo, N., Jalili, M., Ren, Y., Meghdadi, M., Afsharchi, M.: A social rec-
ommender system based on reliable implicit relationships. Knowl.-Based Syst. 192, 105371
(2020). https://doi.org/10.1016/j.knosys.2019.105371
10. Rezvanian, A., Moradabadi, B., Ghavipour, M., Daliri Khomami, M.M., Meybodi, M.R.:
Social recommender systems. In: Learning Automata Approach for Social Networks. Studies
in Computational Intelligence, pp. 281–313. Springer, Cham (2019). https://doi.org/10.1007/
978-3-030-10767-3_8
11. Revathy, V.R., Pillai, A.S.: Classification and applications of social recommender systems.
In: Abraham, A., Cherukuri, A., Madureira, A., Muda, A. (eds.) Proceedings of the Eighth
International Conference on Soft Computing and Pattern Recognition (SoCPaR 2016). SoC-
PaR 2016. Advances in Intelligent Systems and Computing, vol. 614. Springer, Cham (2018).
https://doi.org/10.1007/978-3-319-60618-7_70
12. Turnip, R., Nurjanah, D., Kusumo, D.S.: Hybrid recommender system for learning mate-
rial using content-based filtering and collaborative filtering with good learners’ rating. In:
2017 IEEE Conference on e-Learning, e-Management and e-Services (IC3e), Miri, Sarawak,
Malaysia, November 2017, pp. 61–66 (2017). https://doi.org/10.1109/IC3e.2017.8409239
13. Perumal, S.P., Arputharaj, K., Sannasi, G.: Fuzzy family tree similarity based effective e-
learning recommender system. In: 2016 Eighth International Conference on Advanced Com-
puting (ICoAC), Chennai, India, January 2017, pp. 146–150 (2017). https://doi.org/10.1109/
ICoAC.2017.7951760
14. Mohamed, M.H., Khafagy, M.H., Ibrahim, M.H.: Recommender systems challenges and
solutions survey. In: 2019 International Conference on Innovative Trends in Computer Engi-
neering (ITCE), Aswan, Egypt, February 2019, pp. 149–155 (2019). https://doi.org/10.1109/
ITCE.2019.8646645
15. Wintrode, J., Sell, G., Jansen, A., Fox, M., Garcia-Romero, D., McCree, A.: Content-based
recommender systems for spoken documents. In: 2015 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), South Brisbane, Queensland, Australia,
April 2015, pp. 5201–5205 (2015). https://doi.org/10.1109/ICASSP.2015.7178963
16. Mustafa, N., Ibrahim, A.O., Ahmed, A., Abdullah, A.: Collaborative filtering: techniques
and applications. In: 2017 International Conference on Communication, Control, Computing
and Electronics Engineering (ICCCCEE), Khartoum, Sudan, January 2017, pp. 1–6 (2017).
https://doi.org/10.1109/ICCCCEE.2017.7867668
 Sentiment-Driven Breakdown Detection
Model Using Contextual
Embedding ElMo

Ebtesam Hussain Almansor1,2(B) and Farookh Khadeer Hussain1

1
School of Computer Science, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, Australia
[email protected]
2
Community College, Najran University, Najran, Saudi Arabia
http://www.elsevier.com

Abstract. The recent rapid developments in neural networks have stim-

ulated significant performance progress in detecting breakdown in dia-
logue. The existing research focuses on detecting breakdown from the
data label using different features such as word similarity and topic
transition; however, these features are insufficient to capture contextual
features. Therefore, we used ELMo language model to extract the contex-
tual feature to help to detect the breakdown which faces a challenge due
to the basis of human opinion. Another crucial challenge facing break-
down detection is the lack annotated dataset. To take this challenge we
used sentiment as the data label with contextualized ELMo embedding
model to detect the breakdown in dialogue. The model was evaluated
with state-of-the-art models and the results shows the proposed model
outperforms other models.

Keywords: Chatbot · Breakdown detection · Contextualized ELMo

embedding model · Sentiment

1 Introduction

More recently, conversation agents (CAs) including chatbots, voice control inter-
faces and personal assistants are gaining attention in both academia and indus-
try. The conversational agents are designed to convince humans that they are
having a normal conversation with a machine or computer. Conversation agents
that are based on textual data are called chatbots. Focusing on the chat-oriented
system, the vital objective of the chatbot is to respond to the user’s request.
While the primary purpose of the chatbot is to provide a proper response to
the user, sometimes the chatbot fails to provide an appropriate response pro-
vides a poor response [1]. The reason for providing an inappropriate response is

Fully documented templates are available in the elsarticle package on CTAN.

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 163–171, 2021.
https://doi.org/10.1007/978-3-030-75100-5_15
164 E. H. Almansor and F. K. Hussain

that the system could not understand user’s utterance. Detecting the inappro-
priate responses from chatbots could enhance user experience and increase the
consumer’s trust in the chatbots.
More recently, a selected few studies have been focused on detecting an inap-
propriate response from the chatbot [2]. The point at which the responses from
the chatbots are inaccurate is known as the breakdown point. The existing stud-
ies have focused on detecting the breakdown point based on chatbot response
using hand features such as topic and word similarity measures. However, these
feature are insufficient to capture contextual features.
Another important automated features learning process is called word embed-
ding which uses representations of the words as vectors and uses them as an input
for machine learning models. Word embedding has been used in many natural
language processing tasks such as classification [3]. There are several types of
word embedding such as word2vec [4], Glove [5], fastText [6]. While standard
word embedding models such as word2vec are fast to train, they do not capture
the context. ElMo (Embedding from language model) is an embedding model
that consider the context of the words. ELMo has been used in several domains
such as question answer [7] sentiment analysis [8] and name entity [9]. Due to
the effectiveness of the ElMo model in NLP tasks, in this paper, we build logistic
regression with contextual embedding model (ELMo) to detect breakdown.
The rest of the paper is structured as follows: Sect. 2 presents a related litera-
ture review and discussion of existing studies. In Sect. 3, the proposed framework
is explained. Section 4 explains the experimental results. Section 5 presents the
Conclusion and future work

2 Related Works

Recently, detecting responses that cause dialogue breakdown between a human

user and a chatbot has received a lot of attention [10,11]. The key advantage
of detecting a breakdown in conversation is to increase a user’s trust and sat-
isfaction. Previous studies on breakdown detection used features such as word
similarities and topic differences. For example, Sugiyama et al. [12] proposed
a breakdown detector model based on features such as dialogue act, sentence
length and word-based similarity which improved the performance. The model
used the Extra Trees Regressor to estimate the evaluation distribution. Similarly,
Takayama et al. [2] developed a detection method based on an ensemble of detec-
tors which trained data using different sets of annotators. The proposed model
clustered the training data by the annotation distribution for each annotator and
then built a detector for each cluster. The neural network was utilized to encode
the utterance to their label. The model consists of two long short-term memory
(LSTM) and two convolutional neural network encoders. They use the DBCD
dataset which contains user-system dialogue which is annotated with three labels
(O: Not a breakdown, T: Possible breakdown and X: Breakdown). Some anno-
tated tasks are predictably subjective. Therefore, the annotated label could be
biased due to the annotators. This model was evaluated in Dialogue Breakdown
 Sentiment-Driven Breakdown Detection Model 165

Detection Challenge 3(DBDC3) and achieved 63.6% F-measure, outperforming

the baseline by 5.6% which confirmed the effectiveness of considering the biases
which occur when using annotators [13]. Also, Xie et al. [14] utilized hierarchi-
cal Bi-Direction LSTMs which are trained to map sequences of words to their
breakdown labels. The model was evaluated against three dialogue models that
were provided by DBDC3, and it outperformed the CRF baseline and achieved
high accuracy using both classification and distributed evaluated metrics. Lee
et al. [15] developed a detector model LSTM that encodes temporal utterance
with memory attention. Almansor et al. [16] have proposed a word embedding
model that is measuring the chatbot quality of response.
Other studies focused on inappropriate responses as a breakdown point but
they did not focus on detecting these responses. For example, Higashinaka [17]
proposed a taxonomy of errors in the chatbot-oriented system and classified the
errors into four categories, namely, utterance-level, response-level, context-level,
and environment-level errors. Furthermore, an evaluation metrics was developed
for a breakdown detector [1]. These evaluation metrics were used in DBDCs
which are a gold label that used classification metrics and gold distribution
metrics.
While existing literature in detecting breakdown uses several models, no stud-
ies consider the context of word embedding. Therefore, in this work we take one
step further and build a model that use sentiment as data label and detects the
breakdown using a hybrid model that combines logistic regression and ElMo.
The next section elaborates on the proposed framework.

3 Proposed Model
We proposed a prediction model to predict the breakdown in dialogue based on
sentiment score. Our main aim of this model is to overcome the limitation of
the lack of label dataset. Also, the proposed model introduces the effectiveness
of using ELMo as a contextual language model with logistic regression. The
structure of the proposed model is shown in Fig. 1. It contains three main steps
including data preparation, feature extraction and prediction model. Each step
is explained in the next sub-section in more details.

3.1 Data Preparation

This step helps to prepare the data for the next step which is the extraction of
features and it contains three sub-steps to clean the datasets including prepress-
ing, lemmatization and data labelling.
Pre-processing: in this step we remove the URL links, the punctuation marks,
numbers, and white spaces and convert the utterances into lower case.
Lemmatization: this step helps to reduce and returns words to their base form.
Data Labeling: we apply the Valence Aware Dictionary and sEntiment Rea-
soner (VADER) sentiment analyzer to score the responses. We selected VADER
166 E. H. Almansor and F. K. Hussain

Fig. 1. The structure of the proposed model

because the literature shows it is the most accurate sentiment analyzer for social
media datasets [18]. The VADER sentiment analyzer is used to score the sen-
timent of each response and to label it. VADER is a lexicon and rule-based
sentiment analysis tool which is attuned to sentiment expression in social media
[18]. VADER uses a sentiment lexicon which scores words as either positive or
negative. The sentiment score is measured on a scale [−1, 1], so a response with
a score [ 0.05] is positive; if a response has a score between -0.05 and 0.05 it is
neutral; and a response with a score [ 0.05] is negative.

3.2 Feature Extraction Using ELMO

ELMo provides word vectors that compute on top of two-layer bidirectional lan-
guage model (biLM). The bidirectional model has two layer stacked together;
each layers has two passes, one is the forward pass and the other one is the back-
ward pass. Figure 2 shows the architecture of ElMo. This model uses character
level convolutional neural network to represent the word of the data string into
raw word vectors and pass on these vectors as an input to the first layer of biLM.
The forward pass has information about the word and the context before this
 Sentiment-Driven Breakdown Detection Model 167

word and the backward pass has information about the word and the context
after that word. The information from the forward and backward pass customs
the intermediate word vectors. The intermediate word vectors fed into the next
layer of biLM. The final representation is weighted sum of the raw words vectors
and the two intermediate word vectors, so the input to the ELMo is computed
from character level instead of words which means it captures the inner structure
of the words.

Fig. 2. The structure of ELMo model

After the preprocessing step we import the Elmo model and extract the Elmo
vectors from the cleaned utterances for both train and test datasets. By this step
we have ElMo vectors models for the train and test data that are ready to be
used by the detection model in the next step.

3.2.1 Prediction Model

In this step we build the prediction model that is based on logistic regression
with the ELMo model. So we use the vectors model to train the logistic model
and make the predictions on the test data.
Logistic Regression has been used in the biological sciences and in several
applications of social science [19]. Logistic regression uses a logistic function
which is also known as a sigmoid function. The logistic function takes the real-
value numbers and maps them between 0–1 and represents this as an S-shaped
curve [19]. This function is based on Eq. 1:
1
(1)
1 + e−x
168 E. H. Almansor and F. K. Hussain

where x represents any number that is mapped between 0–1. The main objec-
tive behind the logistic function is to predict y for the given input x . Logistics
regression predicts the probability for the first or default class based on Eq. 2:
T
eβ ·x
p(x) = (2)
1 + eβ T ·x
where β process is followed the parameter vector and x represents the train-
ing data. The logistic regression uses an estimation algorithm such as maximum-
likelihood to estimate and learn the parameter from the training dataset. The
aim is to reduce the error in prediction probability in the training phase.

4 Experimental Results
In this section, we detail the experiment setup that is used to train and test
the classifiers and describe their performance in automatically detecting the
quality of responses. The Python programing language has been use for this
study for programming and building the classifiers and prediction model. For all
the experiments, we divided the data into training 85% and testing 15%. Also,
we use four evaluation metrics, namely accuracy, recall, precision and F1 score.

4.1 Datasets

In this experiment we used DCDB3 datasets which is the common dataset that
is used for the breakdown task. This dataset has four datasets: TKTK-150,
IRIS-150, CIC-165 and YI-150 including training and testing [15].

4.2 Evaluation Metrics

For evaluation purposes, we used the standard evaluation metrics for classifica-
tion tasks, which are accuracy, F1-score, recall, and precision.
TP + TN + TN
Accuracy = (3)
TP + TN + TN + FN + FN + FN + FN + FP + FP
TN
PrecisionN = (4)
TN + FN
TN
RecallN = (5)
TN + FP
2 × Precision × Recall
F − measure = (6)
Precision + Recall
 Sentiment-Driven Breakdown Detection Model 169

5 Results and Discussion

We compare the proposed model with state-of-the-art models in detection break-
down in dialogue. Table 1 shows the comparison of the proposed model with
state-of-the art models. The comparison was based on data label so we applied
the proposed model with the dataset with the original data label which was done
by a human and we refer to this as the proposed model (O). Also, we apply the
proposed model on the same datasets with sentiment score as data label and we
refer to this as the proposed model (S). From Table 1 we can see the accuracy
score by using the proposed model (O) which outperforms the state-of-the-art
models by 0.67. Also, using sentiment with the proposed model obtains better
results in the accuracy by around 0.84 compared with the other models.

Table 1. Comparison results of the proposed ElMo embedding model against the
breakdown detect models on DBDC3 dataset

Models Method Accuracy F1-score(B,NB)

CRF CRF 0.42 0.76
Majority baseline – 0.37 0.89
Encoded vectors LSTM 0.47 0.74
from LSTM [15]
The proposed Logistic 0.67 0.80
model (O) regression with
ELMo
The proposed Logistic 0.84 0.73
model (S) regression with
ELMo with
sentiment score

From these results we can see using logistic regression with ELMo model has
improved the accuracy of breakdown detection model. Also, using sentiment has
a positive effect as the data label.

Table 2. comparison results of the proposed Elmo embedding model against other
ElMo embedding model

Datasets Model Accuracy score

DBDC3 Elmo embedding with kersa 0.84
Embedding model with logistic regression 0.86
MELD Elmo embedding with kersa 0.78
Embedding model with logistic regression 0.80
170 E. H. Almansor and F. K. Hussain

Table 2 show comparison results of Elmo embedding with kersa and with
logistic regression. The accuracy results of Elmo embedding model with logistic
outperform Elmo with kersa on both DBDC3 and MELD datasets as it obtain
0.86 and 0.80 accuracy score respectively.

6 Conclusion and Future Work

With rapid growth of the importance of detection breakdown in dialogue, we

develop sentiment driven breakdown detection using ELMo as the contextual
embedding model. The model showed that it outperformed the state-of-the-art
model. From this paper, we found that sentiment has improved the accuracy of
the proposed model and it has a positive effect as the data label to detect the
breakdown.

References
1. Tsunomori, Y., Higashinaka, R., Takahashi, T., Inaba, M.: Evaluating dialogue
breakdown detection in chat-oriented dialogue systems (2018)
2. Takayama, J., Nomoto, E., Arase, Y.: Dialogue breakdown detection considering
annotation biases. In: Proceedings of the Dialog System Technology Challenge 6
Workshop (DSTC6) (2017)
3. Ge, L., Moh, T.-S.: Improving text classification with word embedding. In: 2017
IEEE International Conference on Big Data (Big Data), pp. 1796–1805. IEEE
(2017)
4. Mikolov, T., Le, Q.V., Sutskever, I.: Exploiting similarities among languages for
machine translation. arXiv preprint arXiv:1309.4168 (2013)
5. Pennington, J., Socher, R., Manning, C.D.: GloVe: global vectors for word repre-
sentation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural
Language Processing (EMNLP), pp. 1532–1543 (2014)
6. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with
subword information. Trans. Assoc. Comput. Linguist. 5, 135–146 (2017)
7. Peters, M.E., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K.,
Zettlemoyer, L.: Deep contextualized word representations. arXiv preprint
arXiv:1802.05365 (2018)
8. Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C.D., Ng, A.Y., Potts, C.:
Recursive deep models for semantic compositionality over a sentiment treebank.
In: Proceedings of the 2013 Conference on Empirical Methods in Natural Language
Processing, pp. 1631–1642 (2013)
9. Sang, E.F., De Meulder, F.: Introduction to the CoNLL-2003 shared task: language-
independent named entity recognition, arXiv preprint arXiv:cs/0306050 (2003)
10. Sugiyama, H.: Dialogue breakdown detection based on estimating appropriateness
of topic transition. In: Dialog System Technology Challenges (DSTC6) (2017)
11. Horii, T., Mori, H., Araki, M.: Breakdown detector for chat-oriented dialogue, pp.
119–127. Springer, Singapore (2017). https://doi.org/10.1007/978-981-10-2585-3 9
12. Sugiyama, H.: Empirical feature analysis for dialogue breakdown detection. Com-
put. Speech Lang. 54, 140–150 (2019)
 Sentiment-Driven Breakdown Detection Model 171

13. Takayama, J., Nomoto, E., Arase, Y.: Dialogue breakdown detection robust to
variations in annotators and dialogue systems. Comput. Speech Lang. 54, 31–43
(2019)
14. Xie, Z., Ling, G.: Dialogue breakdown detection using hierarchical bi-directional
LSTMs. In: Proceedings of the Dialog System Technology Challenges Workshop
(DSTC6) (2017)
15. Lee, S., Lee, D., Hooshyar, D., Jo, J., Lim, H.: Integrating breakdown detection
into dialogue systems to improve knowledge management: encoding temporal utter-
ances with memory attention. Inf. Technol. Manag. 21(1), 51–59 (2020)
16. Almansor, E.H., Hussain, F.K.: Modeling the chatbot quality of services (CQoS)
using word embedding to intelligently detect inappropriate responses. In: Inter-
national Conference on Advanced Information Networking and Applications, pp.
60–70. Springer (2020)
17. Higashinaka, R., Funakoshi, K., Araki, M., Tsukahara, H., Kobayashi, Y.,
Mizukami, M.: Towards taxonomy of errors in chat-oriented dialogue systems. In:
Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse
and Dialogue, pp. 87–95 (2015)
18. Hutto, C.J., Gilbert, E.: VADER: a parsimonious rule-based model for sentiment
analysis of social media text. In: Eighth International AAAI Conference on Weblogs
and Social Media (2014)
19. Brownlee, J.: Master Machine Learning Algorithms: discover how they work and
implement them from scratch (2016)
 The Improved Redundant Active
Time-Based (IRATB) Algorithm
for Process Replication

Tomoya Enokido1(B) , Dilawaer Duolikun2 , and Makoto Takizawa2

1
Faculty of Business Administration, Rissho University, 4-2-16, Osaki,
Shinagawa-ku, Tokyo 141-8602, Japan
[email protected]
2
Department of Advanced Sciences, Faculty of Science and Engineering,
Hosei University, 3-7-2, Kajino-cho, Koganei-shi, Tokyo 184-8584, Japan
[email protected], [email protected]

Abstract. Process replication can provide reliable and available dis-

tributed application services. On the other hand, a large amount of elec-
tric energy is consumed to provide replicated processes since multiple
replicas of each application process have to be performed on multiple
virtual machines in a server cluster system like cloud computing systems.
In this paper, an improved redundant active time-based (IRATB) algo-
rithm is newly proposed to realize energy-efficient server cluster systems
for redundantly performing computation type application processes.

Keywords: Energy-efficient load balancing · Server cluster · Virtual

machine · Improved redundant active time-based (IRATB) algorithm

1 Introduction
In current information systems, a large volume of data is gathered from various
types of smart devices to provide various kinds of distributed application ser-
vices. In order to realize distributed application services, fault-tolerant, scalable,
and high performance computing systems like server cluster systems [1–4] are
required since a large volume of data is manipulated and application services have
to be reliable and available. Virtual machines [5] supported by server cluster sys-
tems like cloud computing systems [6] are widely used to more efficiently utilize
the computation resources of a physical server. Reliable and available distributed
application services can be provided by replicating [7] each application process
on multiple virtual machines in a server cluster system. On the other hand,
the more number of replicas of application processes are performed on multiple
virtual machines than non-replication approaches. As a results, a server cluster
system consumes a large amount of electric energy to perform multiple replicas
of each application process on multiple virtual machines. Hence, it is necessary
to realize energy-efficient server cluster systems to provide reliable and available
distributed application services as discussed in Green computing [6].
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 172–180, 2021.
https://doi.org/10.1007/978-3-030-75100-5_16
 The IRATB Algorithm 173

In our previous studies, the redundant active time-based (RAT B) algo-

rithm [7] is proposed to select multiple virtual machines for each computation
type application process (computation process) so that the total electric energy
consumption of a server cluster and the average response time of each compu-
tation process can be reduced. However, the RATB algorithm assumes only one
thread on a CPU is allocated to each virtual machine in each physical server
even if some thread is not used in the physical server. Therefore, each physi-
cal server consumes more electric energy to perform computation processes on
virtual machines since it takes longer time to perform computation processes if
more number of processes are performed on each virtual machine.
In this paper, the Improved RATB (IRATB) algorithm is newly proposed to
furthermore reduce the total electric energy consumption of a server cluster and
the response time of each process for redundantly performing each computation
process. In the IRATB algorithm, the total electric energy consumption of a
server cluster and the average response time of each computation process can
be reduced by allocating idle threads to active virtual machines which are per-
forming computation processes in each physical server if the total electric energy
consumption of the physical server does not increase. As a result, the response
time of each computation process and the total electric energy consumption of
a server cluster can be more reduced in the IRATB algorithm than the RATB
algorithm. The evaluation results show the total electric energy consumption of
a server cluster and the response time of each computation process can be more
reduced in the IRATB algorithm than the RATB algorithm.
In Sect. 2, we discuss the system model of this paper. In Sect. 3, we discuss
the IRATB algorithm. In Sect. 4, we evaluate the IRATB algorithm compared
with the RATB algorithm.

2 System Model
2.1 Server Cluster System
Multiple physical servers s1 , ..., sn (n ≥ 1) construct a server cluster S, S =
{s1 , ..., sn }. Let nct be the total number of cores in a server st (nct ≥ 1) and
Ct be a set of homogeneous cores c1t , ..., cnct t in the server st . Let ctt be the
number of threads [8] on each core cht in a server st and ntt (ntt ≥ 1) be the
total number of threads in a server st . The total number ntt of threads in a
server st is nct · ctt . Each server st holds a set T Ht of threads th1t , ..., thntt t
and threads th(h−1)·ctt +1 , ..., thh·ctt (1 ≤ h ≤ nct ) are bounded to a core cht .
Each server st supports a set Vt of virtual machines V M1t , ..., V Mntt t . A virtual
machine V Mvt is performed on a set V Tvt (τ ) of threads allocated to the virtual
machine V Mvt at time τ . nV Tvt (τ ) = |V Tvt (τ )| and 1 ≤ nV Tvt (τ ) ≤ ntt .
Computation processes which mainly consume CPU resources are performed
on each virtual machine V Mvt . A term process stands for a computation process
in this paper. Each time a load balancer K receives a request process pi from
a client cli , the load balancer K selects a subset V M S i of virtual machines
in a server cluster S to redundantly perform the request process pi . The load
174 T. Enokido et al.

balancer K forwards the process pi to every virtual machine V Mvt in the subset
V M S i . A notation N F denotes the maximum number of physical servers which
concurrently stop by fault in the cluster S. Let rdi be the redundancy of a
process pi , i.e. number of replicas of a process pi . We assume N F + 1 ≤ rdi =
|V M S i | ≤ n. On receipt of a request process pi , a virtual machine V Mvt creates
and performs a replica pivt of the process pi . On termination of the process pivt ,
i
the virtual machine V Mvt sends a reply rvt to the load balancer K. The load
i
balancer K takes only the first reply rvt and ignores every other reply. Hence, a
i
load balancer K can receive at least one reply rvt from a virtual machine V Mvt
even if N F servers stop by fault in a server cluster S. A virtual machine V Mvt is
active iff (if and only if) at least one replica is performed on the virtual machine
V Mvt . Otherwise, the virtual machine V Mvt is idle. A thread thkt is active iff
at least one virtual machine V Mvt is active on the thread thkt . Otherwise, the
thread thkt is idle. A core cht is active iff at least one thread thkt is active on the
core cht . Otherwise, the core cht is idle. In this paper, we assume each thread
thkt in a server st is not allocated to multiple virtual machines at time τ .

2.2 Computation Model

A notation ptikt stands for a replica of a process pi performed on a thread thkt
allocated to a virtual machine V Mvt in a server st . Let T hikt be the total com-
putation time [msec] of a replica ptikt where the replica ptikt is performed on a
thread thkt . A notation minT hikt shows the minimum computation time of a
replica ptikt where the replica ptikt is exclusively performed on one thread thkt on
a core cht and the other threads on the core cht are idle in a server st . We assume
minT hi1t = minT hi2t = · · · = minT hirt in a server st . minT hi = minT hikt on the
fastest server st . We assume one virtual computation step [vs] is performed for
one time unit [tu] on one thread thkt in the fastest server st . The maximum com-
putation rate M axfkt of a thread thkt on a core cht in the fastest server st where
only the thread thkt is active on the core cht is 1 [vs/tu]. M axfku ≤ M axfkt
on the slower server su . We assume M axf1t = M axf2t = · · · = M axfntt t in
a server st . M axf = max(M axfk1 , ..., M axfkn ) which is defined to be one. A
replica ptikt is considered to be composed of V Skt i
virtual computation steps.
i i i
V Skt = minT hkt · M axf = minT hkt [vs]. The maximum computation rate
i
maxfkt of a replica ptikt is V Skt
i
/minT hikt (0 ≤ maxfkt i
≤ 1) where the replica
i
pkt is exclusively performed on a thread thkt on a core cht and only the thread
thkt is active the a core cht in a server st .
The computation rate F Tkt (τ ) of a thread thkt on a core cht in a server st
at time τ is given as follows:

F Tkt (τ ) = M axfkt · βkt (atkt (τ )). (1)

Here, atkt (τ ) is the number of active threads on the core cht at time τ where
the thread thkt is bounded to the core cht . Let βkt (atkt (τ )) be the performance
degradation ratio of a thread thkt on a core cht at time τ (0 ≤ βkt (atkt (τ )) ≤ 1)
where multiple threads are active on the same core cht . βkt (atkt (τ )) = 1 if
atkt (τ ) = 1. βkt (atkt (τ1 )) ≤ βkt (atkt (τ2 )) ≤ 1 if atkt (τ1 ) ≥ atkt (τ2 ).
 The IRATB Algorithm 175

Suppose a virtual machine V Mvt is performed on a set V Tvt (τ ) of threads

in a server st at time τ . The computation rate F Vvt (τ ) of the virtual machine
V Mvt at time τ is given as follows:

F Vvt (τ ) = F Tkt (τ ). (2)
thkt ∈V Tvt (τ )

In this paper, each replica pivt

of a request process pi is performed on a virtual
machine V Mvt installed in each server st . Replicas which are being performed on
a virtual machine V Mvt at time τ are current. A notation CPvt (τ ) shows a set of
current processes on a virtual machine V Mvt at time τ and N Cvt (τ ) = |CPvt (τ )|.
In this paper, we assume the computation rate F Vvt (τ ) of a virtual machine
V Mvt at time τ is uniformly allocated to every current replica on the virtual
machine V Mvt .
i
The computation rate fvt (τ ) of a replica pivt performed on a virtual machine
V Mvt at time τ is given as follows:


i αvt (τ ) · F Vvt (τ ) / N Cvt (τ ), if N Cvt (τ ) > nV Tvt (τ ),
fvt (τ ) = (3)
F Tkt (τ ), otherwise.

Here, αvt (τ ) is the computation degradation ratio of a virtual machine V Mvt

at time τ (0 ≤ αvt (τ ) ≤ 1). αvt (τ1 ) ≤ αvt (τ2 ) ≤ 1 if N Cvt (τ1 ) ≥ N Cvt (τ2 ).
N C (τ )−1
αvt (τ ) = 1 if N Cvt (τ ) ≤ 1. αvt (τ ) is assumed to be εvt vt where 0 ≤ εvt ≤ 1.
Suppose that a replica pivt starts and terminates on a virtual machine V Mvt
at time stivt and etivt , respectively. Let Tvt i
be the total computation time of a
i i
replica pvt performed on a virtual machine V Mvt . Here, Tvt = etivt − stikt and
etivt i i
τ =sti
fvt (τ ) = V Svt . At time stivt a replica pivt starts, the computation laxity
vt
lcivt (τ ) is V S i
vtτ [vs]. The computation laxity lcivt (τ ) [vs] of a replica pivt at time
i i
τ is V Svt - x=sti fvt (x).
vt

2.3 Power Consumption Model of a Server

A notation Et (τ ) denotes the electric power [W] of a server st at time τ to
perform replicas of computation processes on multiple virtual machines. Let
maxEt and minEt be the maximum and minimum electric power [W] of a server
st , respectively. Let act (τ ) be the number of active cores in a server st at time
τ and minCt be the electric power [W] where at least one core cht is active on
a server st . Let cEt be the electric power [W] consumed by a server st to make
one core active.
The power consumption model of a server with virtual machines (PCSV
model ) [9] to perform computation processes on virtual machines is proposed.
According to the PCSV model, the electric power Et (τ ) [W] of a server st to
perform replicas of computation processes on virtual machines at time τ is given
as follows [9]:

Et (τ ) = minEt + σt (τ ) · (minCt + act (τ ) · cEt ). (4)

176 T. Enokido et al.

Here, σt (τ ) = 1 if at least one core cht is active on a server st at time τ .

Otherwise, σt (τ ) = 0. The electric power Et (τ ) of a server st depends on the
number of active cores at time τ .
The processing power P Et (τ ) [W] is Et (τ ) - minEt at time τ in a server st .
The total
processing electric energy T P Et (τ1 , τ2 ) [J] of a server st from time τ1
τ
to τ2 is τ2=τ1 P Et (τ ).

3 The Energy-Efficient Process Replication

3.1 The Redundant Active Time-Based (RATB) Algorithm
In our previous studies, the redundant active time-based (RAT B) algorithm [7]
is proposed to reduce the total electric energy consumption of a server cluster
S and the average response time of each process for redundantly performing
each computation process. The RATB algorithm assumes only one thread thkt
is allocated to each virtual machine V Mvt in a server st even if some other
threads are idle in the server st . Let dKt be the delay time [msec] between a load
i
balancer K and a server st . The minimum response time minRTvt of a replica
i i
pvt where the replica pvt is exclusively performed on a virtual machine V Mvt
i
and only V Mvt is active on a core cht in a server st is calculated as minRTvt =
i
2dKt + minTvt · M axf /M axfvt = 2dKt + 1 · 1/M axfvt = 2dKt + 1/M axfvt
[msec]. The RATB algorithm estimates the increased active time iACTht (τ ) of
i
each core cht in a server st at time τ based on the response time RTvt of each
replica pivt performed on each virtual machine V Mvt in the server st . The total
processing electric energy laxity tpelt (τ ) [J] shows how much electric energy a
server st has to consume to perform every current replica on every active virtual
machine in the server st at time τ . In the PCSV model, the electric power Et (τ )
[W] of a server st at time τ depends on the number act (τ ) of active cores in
the server st at time τ as shown in Eq. (4). In the RATB algorithm, the total
processing electric energy laxity tpelt (τ ) of a server st at time τ is estimated by
the following Eq. (5):
nct

tpelt (τ ) = minCt + (iAT Cht (τ ) · cEt ). (5)
h=1

Let T P ELivt (τ ) be the total processing electric energy laxity of a server clus-
ter S where a replica pivt of a request process pi is allocated to a virtual machine
V Mvt performed on a server st at time τ . In the RATB algorithm, a virtual
machine V Mvt where the total processing electric energy laxity T P ELivt (τ ) of
a server cluster S is the minimum at time τ is selected for a replica pivt of a
request process pi .

3.2 The Improved RATB (IRATB) Algorithm

In the RATB algorithm [7], only one thread is allocated to a virtual machine in a
server. Hence, some threads on a same core may be idle while the other threads
 The IRATB Algorithm 177

are active to perform replicas. As a result, the more number of replicas are
performed on each virtual machine, the more electric energy a server consumes
since it takes longer time to perform replicas and the active time of each core
in the server increases. In this paper, we newly propose an Improved RATB
(IRATB) algorithm to furthermore reduce the total processing electric energy
consumption of a server cluster S for redundantly performing each process. In the
PCSV model [9], the electric power consumption Et (τ ) of a server st depends on
the number act (τ ) of active cores in the server st at time τ . At time τ , the electric
power consumption Et (τ ) of the server st where current replicas are performed
on a virtual machine V Mvt with one thread thkt on a core cht is the same as
the electric power where current replicas are performed on the virtual machine
V Mvt with multiple threads on the same core cht since act (τ ) = 1. In the IRATB
algorithm, the total processing electric energy consumption of a server cluster S
and the average response time of each process can be reduced by allocating idle
threads to active virtual machines in each server st . In the IRATB algorithm,
if a thread thk t on a core cht allocated to a virtual machine V Mv t is idle, the
thread thk t is used for another active virtual machine V Mvt performed on the
same core cht in each server st at time τ . Then, the computation time of each
replica performed on the virtual machine V Mvt can be reduced in the IRATB
algorithm than the RATB algorithm since the computation rate F Vvt (τ ) of the
virtual machine V Mvt increases. As a result, the total processing electric energy
consumption of each server st can be more reduced in the IRATB algorithm
than the RATB algorithm since the active time of each core can be reduced.
Let idleht (τ ) be a set of idle threads on a core cht in a server st at time τ and
avmht (τ ) be a set of active virtual machines on a core cht at time τ . At time
τ , idle threads on a core cht are allocated to each active virtual machine V Mvt
performed on the core cht in a server st by the THREAD ALLOC procedure as
shown in Algorithm 1.

4 Evaluation

The IRATB algorithm is evaluated in terms of the total processing electric energy
consumption [KJ] of a homogeneous server cluster S and the response time of
each process pi compared with the RATB algorithm [7]. A homogeneous server
cluster S is composed of three servers s1 , s2 , s3 (n = 3). Every server st (1 ≤ t ≤
3) and every virtual machine V Mvt follows the same computation model and the
same power consumption model as shown in Tables 1 and 2. The parameters of
each server st are obtained from the experiment [9]. Every server st is equipped
with a dual-core CPU (nct = 2). Two threads are bounded for each core in a
server st , i.e. ctt = 2. The number ntt of threads in each server st is four. The
total number of threads in the server cluster S is twelve. Initially, each thread
thkt is allocated to one virtual machine V Mvt in a server st (k, v = 1, ..., 4 and
t = 1, ..., 3). Hence, there are twelve virtual machines in the server cluster S.
We assume the fault probability f rt for every server st is the same f r = 0.1.
178 T. Enokido et al.

Algorithm 1. Thread Alloc procedure

Input: st , τ .
Output: V Tvt .
procedure Thread Alloc(st , τ )
for each core cht in a server st do
avmht (τ ) = a set of active VMs on a core cht at time τ ;
if |idleht (τ )| ≥ 1 and |avmht (τ )| ≥ 1 then
while |idleht (τ )| > 0 do
th = thkt ∈ idleht (τ );
V Mvt = a virtual machine where CPvt (τ ) is the maximum in avmht (τ );
V Tvt (τ ) = V Tvt (τ ) ∪ {th};
idleht (τ ) = idleht (τ ) - {th};
avmht (τ ) = avmht (τ ) - {V Mvt };
if avmht (τ ) = φ then
avmht (τ ) = a set of active VMs on cht at time τ ;
end if
end while
end if
end for
end procedure

Table 1. Homogeneous cluster S.

Server nct ctt ntt minEt minCt cEt maxEt

st 2 2 4 14.8 [W] 6.3 [W] 3.9 [W] 33.8 [W]

The number m of processes p1 , ..., pm (0 ≤ m ≤ 8, 000) are issued. The

starting time sti of each process pi is randomly selected in a unit of one millisec-
i
ond between 1 and 3600 [ms]. The minimum computation time minTvt of every
i
replica pvt is assumed to be 1 [ms]. The delay time dKt of every pair of a load
balancer K and every server st is 1 [msec] in the server cluster S. The minimum
i
response time minRTvt of every replica pivt is 2dKt + minTvti
= 2·1+1 = 3
[msec].

4.1 Response Time

Figure 1 shows the average response time of each process in the IRATB and
RATB algorithms. In Fig. 1, IRATB(rd) and RATB(rd) stand for the aver-
age response time of each process in the IRATB and RATB algorithms with
redundancy rd (= 1, 2, 3), respectively. The average response time of each

Table 2. Parameters of virtual machine.

Virtual machine M axfvt εvt βvt (1) βvt (2)

V Mvt 1 [vs/msec] 1 1 0.6
 The IRATB Algorithm 179

process increases as the number m of processes increases in the IRATB(rd)

and RATB(rd) algorithms. The average response time in the IRATB algorithm
can be more reduced than the RATB algorithm. In the IRATB algorithm, idle
threads in each server st are allocated to active virtual machines in the server
st if the total processing electric energy consumption of the server st does not
increase. Then, the computation rate of each active virtual machine to perform
replicas of each process increases. As a result, the average response time of each
process can be more reduced in the IRATB algorithm than the RATB algorithm
since the computation resources in the server cluster S can be more efficiently
utilized in the IRATB algorithm than the RATB algorithm.

4.2 Total Energy Consumption of a Server Cluster

Figure 2 shows the average total processing electric energy consumption AT EC
of the server cluster S to perform the number m of processes in the IRATB and
RATB algorithms. In Fig. 2, IRATB(rd) and RATB(rd) stand for the average
total processing electric energy consumption of the server cluster S in the IRATB
and RATB algorithms with redundancy rd (= 1, 2, 3), respectively. In the IRATB
algorithm, the response time of each replica can be more reduced than the RATB
algorithm as shown in Fig. 1 since idle threads in each server st are allocated to
active virtual machines in the server st . Then, the active time of each virtual
machine can be more reduced in the IRATB algorithm than the RATB algorithm.
As a result, the average total processing electric energy consumption of the
server cluster S can be more reduced in the IRATB algorithm than the RATB
algorithm. Following the evaluation, the IRATB algorithm is more useful than
the RATB algorithm.

2000 180
Total electric energy consumption [KJ]

1800 IRATB(1)
IRATB(1) 160
Average response time [msec]

IRATB(2)
1600 IRATB(2) 140 IRATB(3)
IRATB(3)
1400 120 RATB(1)
1200 RATB(1) RATB(2)
100 RATB(3)
1000 RATB(2)
RATB(3) 80
800
60
600
40
400
200 20
3 0
0 1000 2000 3000 4000 5000 6000 7000 8000 0 1000 2000 3000 4000 5000 6000 7000 8000
Number m of processes Number m of processes

Fig. 1. Average response time. Fig. 2. Total electric energy consump-

tion.

5 Concluding Remarks
In this paper, the IRATB algorithm was newly proposed to reduce the total
electric energy consumption of a server cluster and the response time of each
180 T. Enokido et al.

process for redundantly performing computation processes by allocating idle

threads in each server st to active virtual machines. We showed the total electric
energy consumption of a server cluster and the response time of each process
can be more reduced in the IRATB algorithm than the RATB algorithm. The
evaluation results showed the IRATB algorithm is more useful than the RATB
algorithm.

Acknowledgements. This work was supported by the Japan Society for the Promo-
tion of Science (JSPS) KAKENHI Grant Number 19K11951.

References
1. Enokido, T., Aikebaier, A., Takizawa, M.: Process allocation algorithms for saving
power consumption in peer-to-peer systems. IEEE Trans. Industr. Electron. 58(6),
2097–2105 (2011)
2. Enokido, T., Aikebaier, A., Takizawa, M.: A model for reducing power consumption
in peer-to-peer systems. IEEE Syst. J. 4(2), 221–229 (2010)
3. Enokido, T., Aikebaier, A., Takizawa, M.: An extended simple power consump-
tion model for selecting a server to perform computation type processes in digital
ecosystems. IEEE Trans. Industr. Inform. 10(2), 1627–1636 (2014)
4. Enokido, T., Takizawa, M.: Integrated power consumption model for distributed
systems. IEEE Trans. Industr. Electron. 60(2), 824–836 (2013)
5. KVM: Main Page - KVM (Kernel Based Virtual Machine) (2015). http://www.
linux-kvm.org/page/Mainx Page
6. Natural Resources Defense Council (NRDS): Data center efficiency assessment
- scaling up energy efficiency across the data center industry: Evaluating
key drivers and barriers (2014). http://www.nrdc.org/energy/files/data-center-
efficiency-assessment-IP.pdf
7. Enokido, T., Duolikun, D., Takizawa, M.: An energy-efficient process replication
algorithm based on the active time of cores. In: Proceedings of the 32nd IEEE
International Conference on Advanced Information Networking and Applications
(AINA 2018), pp. 165–172 (2018)
8. Intel: Intel Xeon Processor 5600 Series: The Next Generation of Intelligent
Server Processors (2010). http://www.intel.com/content/www/us/en/processors/
xeon/xeon-5600-brief.html
9. Enokido, T., Takizawa, M.: Power consumption and computation models of virtual
machines to perform computation type application processes. In: Proceedings of
the 9th International Conference on Complex, Intelligent and Software Intensive
Systems (CISIS 2015), pp. 126–133 (2015)
 The Actual Cost of Programmable
SmartNICs: Diving into the Existing
Limits

Pablo B. Viegas1 , Ariel G. de Castro1 , Arthur F. Lorenzon1 ,

Fábio D. Rossi2 , and Marcelo C. Luizelli1(B)
1
Federal University of Pampa (UNIPAMPA), Alegrete, Brazil
{pabloviegas.aluno,arielcastro.aluno,arthurlorenzon,
marceloluizelli}@unipampa.edu.br
2
Federal Institute Farroupilha (IFFAR), Alegrete, Brazil
[email protected]

Abstract. Programmable Data Planes is a novel paradigm that enables

efficient offloading of network applications. An important enabler for this
paradigm is the current available SmartNICs, which should satisfy rigid
network requirements such as high throughput and low latency. Despite
recent research in this field, not much attention was given to understand
the costs and limitations of offloading network applications into Smart-
NIC devices. Existing offloading approaches either neglect the existing
limitations of SmartNICs or assume that as a fixed cost – leading, there-
fore, to sub-optimal offloading approaches. In this work, we conduct a
comprehensive evaluation of SmartNICs in order to quantify existing
performance limitations. We provide insights on network performance
metrics such as throughput and packet latency while considering differ-
ent key building blocks of complex P4 programs (e.g., registers, cryp-
tography functions, or packet recirculation). Results show that line-rate
throughput can degrade up to 8x, while latency can increase as much as
80x when performing memory-intensive operations in the data plane.

1 Introduction
Programmable Data Planes (PDP) is a mainstream technology that has been
recently redesigning the networking domain. PDP allows to (re)define the behav-
ior of network devices (e.g., programmable routers and SmartNICs), allowing to
deliver specialized packet processing mechanisms [2]. Recent advances in data
plane programmability have enabled offloading typical control plane applica-
tions to the data plane (e.g., machine learning algorithms [19], routing [12], or
network monitoring [4,8,9]). On shifting the operation of these applications to
the data plane, it brings the benefit to process every single packet and react to
network conditions in the order of nanoseconds, with minimum control plane
intervention. Despite that, data plane operation might become complex – and
the complexity comes at a price: lower throughput and higher latency.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 181–194, 2021.
https://doi.org/10.1007/978-3-030-75100-5_17
182 P. B. Viegas et al.

Current SmartNIC architectures (e.g., [11]) do not limit the number of opera-
tions performed by the data plane in a single pipeline stage. For example, a PDP
application could trigger the read and write of an unbounded number of regis-
ters in a given stage of the packet processing pipeline. Yet, the same application
could recirculate the ingress packet multiple times in order to mimic a loop-based
mechanism. These are straightforward examples of simple operations commonly
used by more complex PDP applications (e.g., in-network clustering [19]). There-
fore, understanding the current performance limitations of existing SmartNICs
is paramount to the design of efficient PDP applications.
A recent study [7] has made the first effort to understand the existing lim-
itations of SmartNICs. Harkous et al. [7] have focused on evaluating the per-
formance of general PDP metrics such as parsers, control blocks, and header
modifications in P4 programs. Despite this effort, no study has yet thoroughly
evaluated key building blocks of complex P4 programs (e.g., registers, cryptog-
raphy functions, or packet recirculation) – which are essential for most recent P4
applications. To fill in this gap, we perform an extensive performance evaluation
of SmartNICs to understand and quantify PDP application primitives and exist-
ing limitations. We focus our evaluation on measuring the performance in terms
of latency and throughput for a variety of packet sizes (from 64B to 1500B) when
(i) operating multiple registers of different sizes/widths (e.g., used to implement
bloom filter alike structures [1]), (ii) matching on multiple tables in the ingress
and egress pipeline (e.g., used to implement machine-learning algorithms [19]),
(iii) performing packet recirculation (e.g., used to implement IoT data desegre-
gation [17]), and (iv) using cryptography functions and arithmetic operations.
Results show that network throughput can degrade up to 8X, while latency can
increase as much as 80X when performing memory-intensive operations in the
data plane. The main contributions of this paper can be summarized as:

• an in-depth performance evaluation of SmartNICs;

• a discussion of current limitations in SmartNIC architectures; and
• an open-source code of all experiments in order to foster reproducibility.

The remainder of this paper is organized as follows. In Sect. 2, we describe

the SmartNIC architecture used in this work. In Sect. 3, we overview the recent
literature regarding PDP applications and performance evaluation. In Sect. 4, we
present and discuss the obtained results and, in Sect. 5, we conclude this paper
with final remarks.

2 SmartNIC Architecture
The hardware techniques used to deliver high-speed network packet processing
require context switching on the order of nanoseconds and very high degrees
of processing parallelism to scale the performance of P4-based programs to
Gb/s of throughput and beyond [11]. This is particularly challenging given the
ever-increasing complexity of offloaded codes to network adapters. Current pro-
grammable NICs (also referred to as SmartNICs) rely their architectures on
 The Actual Cost of Programmable SmartNICs 183

multi-threaded, multicore flow processor units to cope with this increasing and
stringent demand. Next, we focus our discussion on the general architectural
details of the Netronome SmartNIC architecture [11] – which is used later in our
experiments.

FPC Island
ME0 ME1 ME7
Cluster
Code Code Code Target
Data Data Data Memory

CPP - Command Push Pull

ARM PCIe

Ingress Ingress Egress Egress

Distributed Switch Fabric
MAC PPC PPC MAC

Crypto IMU EMU

Fig. 1. An overview of the Netronome SmartNIC architecture.

The SmartNIC Netronome NFP4000 architecture is illustrated in Fig. 1 and

organizes its flow processing cores (FPC) in multiple islands. FPC is a 32-bit
machine, and therefore, all of its internal registers and local memory consists
of 32-bit words. Each FPC contains eight Micro Engines (MEs) – a separate
processor with its own instruction store (code) and local memory (data). Because
of that, every ME can run in parallel with all other MEs. Each ME has 8 threads
that can be used for co-operative multithreading, in the way that at most one
thread is executing code from the same program at any given moment. Hence,
each FPC runs at most 8 parallel threads at 1.2 GHz (one thread per ME). FPCs
follow a Harvard Architecture, and therefore code and data occupy different
memories – 4K bytes are shared between threads for data and private memory,
while 8K instructions for the coding store.
In each FPC, local memory is composed of a set of 32-bit registers, shared
between all 8 threads. These registers are divided into: (i ) general-purpose reg-
isters (256 32-bits registers) – used by default to store any register of up to
32-bits size; (ii ) transfer registers (512 32-bits registers) – used for copying reg-
ister over the interconnection bus (e.g., from or to other FPCs or memories);
(iii ) next-neighbor registers (128 32-bits registers) – used mainly to communi-
cate with neighboring FPCs; and last (iv ) local memory (1024 32-bit registers)
– which is a little bit slower than general register (3 cycles access, instead of
just 1 cycle). When there is a need for more memory than available space in
local FPC registers, variables are automatically and statically allocated to other
in-chip memory hierarchies.
There are four other kinds of memory which are available to FPCs: (i ) Cluster
Local Scratch (CLS) (20–50 cycles); (ii ) Cluster Target Memory (CTM) (50–
100 cycles); (iii ) Internal Memory (IMEM) (120–250 cycles); and (iv ) External
Memory (EMEM) (150–590 cycles). In summary, the local memory register is
184 P. B. Viegas et al.

used for data that is used in every packet. The CLS is used for data, which is
needed for most packets and small shared tables. The CTM is used for packet
headers and coordination between other sub-systems. Then, IMEM is used for
packet bodies and medium-sized shared tables. Finally, EMEM is used for large
shared tables.
As packets are received from the network, an FPC thread picks up the packet
from the Distributed Switch Fabric and processes it (i.e., on-path SmartNIC [6]).
Additional threads are allocated to new packets as they arrive. For instance, the
SmartNIC NFP-4000 supports up to 60 FPCs, each supporting up to 8 threads.
Then, the device is able to process up to 480 packets simultaneously.

3 Related Work
In this section, we discuss the most recent efforts towards P4 SmartNIC offload-
ing and the performance evaluation of them in PDP.
DAIET [15] is a network system that performs in-network data aggrega-
tion. It uses Machine Learning (ML) to judiciously decide which partition of the
application (e.g., MapReduce) is deployed into PDP to reduce network traffic
while maintaining the correctness of the overall computation. FairNIC [6] uti-
lizes SmartNICs to decrease CPU host utilization while providing performance
isolation in a multi-tenant environment. It provides an abstraction that allows
network applications to access NIC resources. In turn, Clara [13] provides per-
formance clarity for SmartNIC offloading. It analyzes network functions (NFs)
and predicts their performance when ported to SmartNIC targets. It uses a
logical SmartNIC model to capture SmartNIC architecture. Then, an interme-
diate representation identifies the code blocks and maps them onto the logi-
cal model while optimizing for a performance objective. Finally, it outputs the
performance profile for the original NF input on a particular workload. Simi-
larly, SmartChain [18] minimizes the redundant packet transmission by analyz-
ing service function chaining (SFC) forwarding paths and reducing the packet
transmissions between the SmartNIC and the host CPU. In the same direction,
iPipe [10] allows to offload distributed applications onto SmartNICs. At its core,
a scheduler combines first-come & first-serve strategy with deficient round-robin
to tolerate applications with variable execution costs.
In addition to the efforts mentioned above towards enabling efficient offload-
ing of network applications to SmartNICs, there are also recent initiatives
to offload ML algorithms to PDPs. In this context, N2Net [16] and Sanvito
et al. [14] represent the first steps toward in-network inference. N2Net [16] is
a compiler that generates a P4 program configuration for an RMT-like switch
pipeline [3] based on a binary neural network (BNN) model description while
Sanvito et al. [14] introduce BaNaNa SPLIT, a system capable of offloading
BNNs from CPUs to SmartNICs through a quantization process that transforms
the NN model into a format that can be appropriately executed on PDPs. More
recently, Xiong et al. [19] propose to deploy trained ML classification algorithms
into PDPs. The proposed approach relies on multiple match-action tables and,
therefore, is portable between different PDP targets.
 The Actual Cost of Programmable SmartNICs 185

As one can observe, most of the existing efforts are still restricted to offload-
ing mechanisms to PDP. However, to the best of our knowledge, these studies
do not take into account the current limitations of existing SmartNICs on the
offloading process. Therefore, these solutions might either lead to infeasible solu-
tions (e.g., using more resources than available) or suffer high penalties on the
expected performance. One noticeable exception is the recent study conducted by
Harkous et al. [7]. They evaluate different P4 programs and their impact on the
packet processing latency. They gradually increase the complexity of a Smart-
NIC pipeline (i.e., including parser, control blocks, and deparser) to identify the
most influential variables for predicting packet latency. Despite this effort, the
work conducted by [7] is still full of gaps considering key building blocks of com-
plex P4 programs (e.g., registers, cryptography functions, packet recirculation,
or multiple tables) – which are essential in P4 applications (especially ML appli-
cations). In this work, we take a step further into thoroughly understanding the
performance of SmartNICs and their existing limitations.

4 Deployment Evaluation
In this section, we evaluate the performance of Netronome SmartNIC for P4
programs concerning their properties and achieved throughput and latency. We
start describing our environment setup and performance metrics, followed by the
discussion of results.

4.1 Setup

Our environment setup consists of two high-end servers. Each server has an Intel
Xeon 4214R processor with 32 GB RAM. One server is our Device Under Test
(DUT) – i.e., the server in which P4 programs are loaded – and the other is our
traffic generator. Both servers have a Netronome SmartNIC Agilio CX 10 Gbit/s
network device with two network interfaces, which are physically connected. We
use MoonGen [5] as our DPDK1 traffic generator. We instruct MoonGen using
the Netronome Packet Generator2 . In our experiments, we send IPv4 packets at
line rate (i.e., 10Gbit/s) with random source and destination prefixes. For our
evaluation, we varied the packet size from 64B to 1500B. All experiments were
run at least 30 times to ensure a confidence level higher than 90%.

4.2 Metrics

In our evaluation, we aim to measure the performance of P4 programs with

respect to the achieved throughput and latency, and identify existing hardware
limitations. We evaluate the impact on those metrics regarding (i ) the number
of operations on registers, (ii ) the number of access to match+action tables,

1
https://www.dpdk.org/.
2
https://github.com/emmericp/MoonGen/tree/master/examples/netronome-
packetgen.
186 P. B. Viegas et al.

(iii ) the number of packet recirculation, (iv ) the number of applied cryptography
functions, and (v ) the number of performed arithmetic operations. To evaluate
these metrics, we automatically generate P4 codes with the properties to be
analyzed. All P4 codes have at least one match+action table used to perform
IP forwarding between physical ports. After generating the P4 source codes, we
compiled them using the Netronome compiler and loaded the generated firmware
into the physical SmartNIC. Then, we pump network traffic with MoonGen and
collect the obtained throughput and latency. To measure data plane latency (i.e.,
the amount of time a packet stays on PDP), all P4 programs have at least one
register which keeps that information (i.e., the difference between the ingress
and egress timestamps). During our experiments, we read that register data
using the Netronome CLI. The obtained throughput (measured in packets per
second) is obtained directly from MoonGen. In order to foster reproducibility,
our experimental codes are public available3 . It is important to note that other
SmartNICs and compilers can be easily adapted to our experimental codes.

4.3 Results
4.3.1 The Cost of Reading and Writing at Multiple Registers
We start by analyzing the cost of performing multiple register operations in the
P4 pipeline. Register operations are one of the main building blocks of recent
P4 applications (e.g., [1,17,19]). In the experiments, we varied the number of
register operations performed sequentially by the P4 program from 10 to 200
registers while varying the register width from 32- to 512-bit words. We consider
that registers are placed in the ingress pipeline and are either read, write, or read
& write. Our goal is to quantify the impact on throughput and latency, as well as
to quantify the existing limitation of the current architecture. Figure 2 illustrates
the measured throughput (in packets per second) and latency (in nanoseconds).

Packet Intense Line-Rate Network Throughput. Figures 2(a) and 2(b)

depict the measured throughput and latency, respectively, for small packets (64
Bytes) and register width of 32 bits. As the number of registers increases (and
consequently the number of P4 instructions), there is a sharp performance degra-
dation on observed throughput and latency. The line rate for 10 Gbit/s (i.e.,
14.88 million packets per second – Mpps) can be sustained for reading oper-
ations to only 10 registers (Fig. 2(a)). Even with 10 registers, the bandwidth
degradation for writing operations is 30% (and it reaches 50% for read & write)
– as they demand more micro engine cycles to be performed. After, we observe a
linear decrease of up to 87% (i.e., 2 Mpps) considering 200 registers. In Fig. 2(b),
the latency increases linearly as the number of operations performed by each pro-
cessed packet in the pipeline. For a small number of registers (i.e., 10), there is
an acceptable overhead of 8650 ns (reading), 20296 ns (writing), and 35839 ns
(both operations). However, this overhead can be as high as 0.12 and 0.24 ms
for 50 and 200 registers, respectively.
3
https://github.com/mcluizelli/performanceSmartNIC.
 The Actual Cost of Programmable SmartNICs 187

15Mpps 250000
Read Read
Write Write
12Mpps 200000

(in nanoseconds)
Read/Write Read/Write

Throughput
(in pps)
150000

Latency
8Mpps
100000
4Mpps
50000
2Mpps
1Mpps 0
10 20 30 40 50 60 70 80 90 100 150 200 10 20 30 40 50 60 70 80 90 100 150 200
# of registers # of registers

(a) Measured throughput for different (b) Measured latency for different
register operations (packet size 64B). register operations (packet size 64B).
10Mpps 350000
64B 64B
128B 300000 128B
8Mpps

(in nanoseconds)
256B 256B
512B 250000 512B
Throughput

1024B 1024B
(in pps)

Latency
1500B 200000 1500B
4Mpps 150000
100000
2Mpps
50000
1Mpps 0
10 20 30 40 50 60 70 80 90 100 150 200 10 20 30 40 50 60 70 80 90 100 150 200
# of registers # of registers

(c) Measured throughput for different (d) Measured latency for different
packet sizes. packet sizes.
15Mpps 500000
32 bits 32 bits
64 bits 450000 64 bits
12Mpps 400000
(in nanoseconds)

128 bits 128 bits

256 bits 350000 256 bits
Throughput

512 bits 512 bits

(in pps)

300000
Latency

8Mpps 250000
200000
4Mpps 150000
100000
2Mpps 50000
1Mpps 0
10 20 30 40 50 60 70 80 90 100 150 200 10 20 30 40 50 60 70 80 90 100 150 200
# of registers # of registers

(e) Measured throughput for different (f) Measured latency for different reg-
register width. ister width.

Fig. 2. Measured throughput and latency for register operations.

Line Rate Network Throughput for Different Packet Sizes. Figures 2(c)
and 2(d) depict the measured throughput and latency, respectively, for different
packet sizes (from 64B to 1500B). We fixed the register width to 32 bits and the
register operation to read & write (as it is the most resource-consuming one).
Observe that the larger the packet size (and, consequently, the less the number
of packets per second to achieve 10Gbit/s), the more register operations can be
sustained at line rate. For instance, to 512B-size packets, the line rate through-
put can be sustained up to 60 registers being read and write (2Mppps). For
1024B-size packets and higher, there is no throughput degradation – even for 200
registers. Although there is no throughput degradation, Fig. 2(d) illustrates that
per-packet latency increases substantially. For instance, for 1500B-size packet,
the latency increases up to 3X (from 7983 ns to 22172 ns), while for small packet
sizes (and, consequently, higher packet throughput), this latency overhead can
be as high as 0.25 ms per packet (i.e., a 34X increase).
188 P. B. Viegas et al.

Line Rate Network Throughput for Different Register Width.

Figure 2(e) and Fig. 2(f) illustrate the throughput and the latency for a varying
width of registers. We fixed the packet size to 64B and the operation to read
– as the goal is to quantify the performance degradation with respect to the
achieved line rate. As one can observe, the line-rate operation is kept for a reg-
ister width of up to 128 bits (and 10 registers). Larger register width demands
more cycles to fetch the data from memory. As discussed in Sect. 2, Netronome
SmartNIC follows a 32-bit architecture. Therefore, any register width wider than
that requires extra cycles to be fetched. In addition to that, it is important to
mention the memory hierarchy. The more register is needed, the more external
memories are used – which directly affects the time to fetch data.

Current Limitations. Is there any limitation on operating registers in a

SmartNIC? In our experiments, we were able to define at most six arrays of 32-
bit registers, each having 130M positions. This limitation is due to the amount
of memory available on the board (see Section 2). Although we could instan-
tiate such a large number of registers, we could not access all of them in a
single pipeline pass because micro engines have a limited code space to store
the instruction set (i.e., at most 8K instruction). Differently from traditional
languages, P4 does not have go-to primitives, and therefore all instructions are
defined at compiling time. For that reason, we were able to operate at most 200
registers in a single P4 program (considering that our P4 program has also other
instructions to perform the forwarding).

4.3.2 Impact of Packet Recirculation

Next, we evaluate the impact of performing packet recirculation in the P4
pipeline. As previously mentioned, the P4 language does not support iteration-
based structures, and therefore, packet recirculation has been used as a way to
circumvent such limitations. In short, packet recirculation consists of sending a
packet back to the ingress pipeline after processing it (and therefore, mimic a
loop-based structure). In the experiments, we varied the number of packet recir-
culation made (from 0 to 50) in each packet while varying the packet size (from
64B to 1500B). We consider that our P4 program is just forwarding network
traffic from physical interfaces. Figure 3 illustrates the measured throughput
and latency.

15Mpps
64B 64B
128B 1M 128B
12Mpps
(in nanoseconds)

256B 256B
512B 800K 512B
Throughput

1024B 1024B
(in pps)

Latency

8Mpps 1500B 600K 1500B

400K
4Mpps
200K
2Mpps
1Mpps
0 1 2 3 4 5 6 7 8 9 1011121314151617181920304050 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 30 40 50

# of packet recirculations # of packet recirculations

(a) Measured throughput. (b) Measured latency.

Fig. 3. Analysing the effect of packet recirculation in the P4 pipeline.

 The Actual Cost of Programmable SmartNICs 189

Figure 3(a) illustrates the throughput behavior as the number of packet recir-
culations increases (it decreases in a super-linear manner). For small-packet net-
work traffic, we observe that fewer packet recirculations are sustained at line
rate (e.g., three packet recirculations for 128B size packets). In contrast, for
larger packets (e.g., 1024B), it can maintain line rate even for up to two dozen
packet recirculations. As the packets are recirculated, more packets are being
pushed into the data plane – which enqueue them, and eventually dropped occurs
(reducing throughput). In turn, Fig. 3(b) depicts the observed latency. As one
can observe, there is a non-negligible increase in latency as packets recirculate.
For instance, even for large packets (e.g., 1500B) – in which we observe little or
no throughput degradation, the per-packet latency doubles on performing just 3
packet recirculations (from 5610 ns to 10688 ns). This overhead is even sharper as
the number of recirculations increases. For example, there is a 6x latency increase
(32762 ns) for doing 10 recirculations, and an 80x latency increase (450062 ns)
for 50 recirculations. As one can observe, this overhead is even greater for small
packets.

Current Limitations. In our experiments, we observe that the Netronome

architecture does not allow for recirculating custom-made metadata structure.
This limits substantially the ability to write complex P4 programs – especially
the ones using packet recirculation to circumvent the lack of iteration-based
structure.

4.3.3 Impact of Using Multiple Tables

We evaluate the impact of using multiple match-action tables. Unlike traditional
forwarding devices that use match+action tables exclusively for routing (i.e., to
look up network addresses), the P4 language has opened up new possibilities for
this construction type. For instance, Xiong et al. [19] have used multiple tables
to implement data plane clustering approaches (e.g., k-means with a table per
cluster). Similar to the work conducted by Harkous et al. [7], we also observe
that the performance of P4 programs is not affected by the size of tables – as a
hash-based data structure implements them. Usually, large match+action tables
are already placed on larger and slow memories (e.g., DRAM). Here, instead, we
aim at analyzing the impact of using multiple match+action tables at different
stages of the pipeline. In the experiments, we varied the number of existing
tables in our P4 programs (from 1 to 10), and we ensure that every packet is
always matched sequentially on all tables. An action is invoked to read a single
32-bit data from the table and store it in a metadata structure on a packet
matching. We varied the packet size (from 64B to 1500B) and the number of
tables per pipeline (either on the ingress or egress pipeline). Figure 4 illustrates
the measured throughput and latency.
Figure 4(a) illustrates the measured throughput for an increasing number of
match+action tables. As observed, the throughput for 64B packets (most packet
intensive network traffic) is almost negligible for up to 5 match+action tables
(i.e., it keeps the line rate). However, we observe an abrupt decay after 5 tables,
190 P. B. Viegas et al.

followed by a constant throughput behavior (up to 10). In the Netronome archi-

tecture, a P4 program can only have 5 tables in each pipeline (ingress/egress).
Therefore, tables 1–5 are located in the ingress pipeline, while 6–10 in the egress.
As all the memory is statically allocated for a P4 program, the Netronome com-
piler tends to use faster, closer available memory to micro engines to allocate
ingress tables. Even when defining tables only in the egress pipeline, the compiler
tends not to use faster memory. We empirically show this behavior in Fig. 4(c).
We incrementally place 5 tables either in the ingress or egress pipeline. We
observe that the ingress pipeline is always faster to use available tables (w.r.t.
latency) – even in the cases where no tables are used in the ingress. Last, Fig. 4(b)
illustrates the per-packet latency for an increasing number of tables (both in
the ingress/egress pipeline). On average, there in an increase of 40–50% in the
latency in the ingress pipeline (between 1 and 5 tables).

11000 80000
64B 256B 1024B 128B Egress pipeline
128B 512B 1500B 10000 256B 70000 Ingress pipeline
15Mpps
(in nanoseconds)

(in nanoseconds)
512B
9000 1024B 60000
Throughput

12Mpps 1500B
(in pps)

Latency

Latency
8000 50000
8Mpps 7000 40000
6000 30000
4Mpps
2Mpps 5000 20000
1Mpps 4000 10000
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5
# of tables # of tables # of tables

(a) Measured throughput. (b) Measured latency. (c) Measured latency

in the egress/ingress
pipeline.

Fig. 4. Analysing the effect of using multiple tables on the P4 pipeline.

Current Limitations. As mentioned, the Netronome architecture poses a limit

on the number of match+action tables one can use in a P4 program (i.e., at most
5 in each pipeline). That limits the applicability of more complex algorithms
(e.g., [7]) in SmartNICs. Further, we also observe performance differences w.r.t.
to the incurred latency (i.e., tables placed in the ingress behave 20% faster).

4.3.4 Impact of Using Cryptography Functions

and Arithmetic Operations
Next, we evaluate the impact of using cryptography functions in our P4 pro-
grams. Bloom filters or hash-based data structures widely use them (e.g., to
identify heavy-hitters [1]). Cryptography functions are target-dependent (i.e.,
the implementation depends on the hardware), and in the Netronome archi-
tecture, they are implemented by specific micro engines (named Crypto in
Fig. 1). Netronome implements eight hash functions: (i ) crc32, (ii ) crc16, (iii )
identity, (iv ) csum16, (v ) crc32custom, (vi ) crc16custom, (vii ) random, and
(viii ) xor16. In the experiments, we analyse the impact of applying consecu-
tive calls to these cryptography functions (from 0 to 30) in each packet being
processed. For the purpose of this experiment, we keep the packet size in 64B
and consider that our P4 program just forward network traffic between physical
interfaces. Figure 5 illustrates the measured throughput and latency.
 The Actual Cost of Programmable SmartNICs 191

Figures 5(a) and 5(b) illustrate the throughput and latency, respectively.
As one can observe, only three out of the eight cryptography functions
(crc32-custom, random, and csum16) do not lead to performance degradation
w.r.t. throughput and latency when increasing the number of calls to them.
The remaining ones (i.e., crc32, crc16, identity, crc16-custom, and xor16)
lead to some performance degradation from applying 10 cryptography functions
(e.g., 23% of throughput degradation for applying xor16). This overhead is even
higher for 30 cryptography functions (up to 55% overhead for cryptography func-
tion xor16). In turn, Fig. 5(b) illustrates the incurred per-packet latency. For
up to 10 cryptography functions, the per-packet latency remains acceptable (i.e.,
below 10000 ns). However, on applying higher number of cryptography functions
(e.g., from 15–20 and on), the latency cost grows exponentially. For instance,
the xor16 function reaches up to 7X higher latency (applying 30 functions) in
comparison to the simple forwarding (the case of 0 cryptography functions). We
further evaluate the impact of using arithmetic operations in our P4 programs
(i.e., +, −, ∗, /, %, and << (bit-shifting)). In the experiments, we analyse the
impact of applying consecutive arithmetic operations (varying from 10 to 10000
operations) for intensive packet processing (i.e., 64B packets). We also consider
that our P4 program just forwarding network traffic between physical interfaces.
In this experiment, we do not observe any statistically significant latency or
throughput degradation.

45000
crc32 crc32−custom crc32
crc16 crc16−custom 40000 crc16
(in nanoseconds)

csum16 random 35000 csum16

identity xor16 identity
Throughput

15Mpps 30000 crc32−custom

(in pps)

Latency

12Mpps crc16−custom
25000 random
8Mpps 20000 xor16
15000
4Mpps
2Mpps 10000
1Mpps 5000
0 1 5 10 15 20 25 30 0 1 5 10 15 20 25 30

# of hash functions # of hash functions

(a) Measured throughput. (b) Measured latency.

Fig. 5. Analysing the effect of applying multiple hash functions on the P4 pipeline.

Current Limitations. Netronome architecture poses a few limitations restrict-

ing the design of more complex P4 programs to its SmartNICs. For instance, mul-
tiplication and division are only performed over an integer. Further, the archi-
tecture restricts multiplication and division operations of fixed-size operands
(i.e., at most 32 bits operands). This limits, for instance, the implementation
of precise fixed-point representation for real numbers. Another limitation is the
bit shifting operation. The current architecture requires bit-shifting to be static-
compiled with predefined values – limiting its applicability. Last, the number of
arithmetic operations is limited to 10000 operations per pipeline – related to the
number of instructions a micro engine can store.
192 P. B. Viegas et al.

Operations = 0 Operations = 40 Operations = 0 Operations = 40

Operations = 20 Operations = 60 Operations = 20 Operations = 60

15Mpps 110000

(in nanoseconds)
Throughput 12Mpps 90000
(in pps) 70000

Latency
8Mpps
50000
4Mpps 30000
2Mpps 10000
1Mpps
10 20 30 40 50 60 10 20 30 40 50 60
# Micro Engines (ME) # Micro Engines (ME)
(a) Measured throughput. (b) Measured latency.

Fig. 6. Analysing the effect of using multiple Micro Engines (CPU cores) to process
network traffic.

4.3.5 Analysing Used Cores and Energy Consumption

Last, we evaluate the impact of varying the number of micro engines used by the
SmartNIC (from 10 to 60 ME). The goal is to verify whether or not it affects
the obtained performance. We keep the packet size at 64B for this experiment
and consider that our P4 program forwards network traffic between physical
interfaces. We varied the number of reading operations (from 0 to 60) in existing
32-bit registers to stress out the hardware. Figure 6(a) illustrates the measured
throughput. As expected, the more ME is available, the more throughput is
achieved (in general). However, we can observe that the SmartNIC does not need
all ME working in parallel for the evaluated workload. For instance, for 0 read
operations, the line rate throughput is achieved using 40 ME. When considering
the case of 60 read operations, we observe that more than 20 ME does not affect
the performance. Yet, we also observe that allocating a higher number of ME
is not always the best strategy. In some cases (e.g., for 20 read operations), the
performance is worsened by increasing the number of MEs from 40 to 50–60.
Figure 6(b) depicts the measured latency. As one can observe, there is always
a latency reduction when increasing the available ME (even when there is no
improvement in the throughput). Finally, we evaluated the energy consumption.
In our experiments, the energy consumption varied 0.2 W between using 10 ME
and 60 ME.

5 Final Remarks

In this paper, we performed an extensive performance evaluation of SmartNICs

to understand and quantify existing limitations. We focus our evaluation on
measuring the performance in terms of latency and throughput for a plethora
of packet memory-intensive scenarios. We showed that the line-rate throughput
is bounded by (i) the number of register operations (up to 10 operations), (ii)
the number of multiple match+action tables user in the pipeline (up to 5), (iii)
the number of cryptography operations (up to 10). As future work, we intend to
build an analytical model that can accurately estimate the performance of P4
applications executing on SmartNICs.
 The Actual Cost of Programmable SmartNICs 193

Acknowledgements. This work was partially funded by National Council for

Scientific and Technological Development (CNPq) (grant 427814/2018-9), São
Paulo Research Foundation (FAPESP) (grant 2018/23092-1), Rio Grande do
Sul Research Foundation (FAPERGS) (grants 19/2551-0001266-7,20/2551-000483-0,
19/2551-0001224-1).

References
1. Ben Basat, R., Einziger, G., Friedman, R., Luizelli, M.C., Waisbard, E.: Constant
time updates in hierarchical heavy hitters. In: Proceedings of the ACM SIGCOMM,
SIGCOMM 2017, pp. 127–140. ACM, New York (2017)
2. Bosshart, P., Daly, D., Gibb, G., Izzard, M., McKeown, N., Rexford, J., Schlesinger,
C., Talayco, D., Vahdat, A., Varghese, G., Walker, D.: P4: programming protocol-
independent packet processors. ACM SIGCOMM 44(3), 87–95 (2014)
3. Bosshart, P., Gibb, G., Kim, H.S., Varghese, G., McKeown, N., Izzard, M., Mujica,
F., Horowitz, M.: Forwarding metamorphosis: fast programmable match-action
processing in hardware for SDN. ACM SIGCOMM CCR 43(4), 99–110 (2013)
4. Castro, A.G., Lorenzon, A.F., Rossi, F.D., Da Costa Filho, R.I.T., Ramos, F.M.V.,
Rothenberg, C.E., Luizelli, M.C.: Near-optimal probing planning for in-band
network telemetry. IEEE Commun. Lett. 1 (2021). https://ieeexplore.ieee.org/
document/9330755/keywords#keywords
5. Emmerich, P., Gallenmüller, S., Raumer, D., Wohlfart, F., Carle, G.: MoonGen:
a scriptable high-speed packet generator. In: Proceedings of the ACM IMC, IMC
2015, pp. 275–287. ACM, New York (2015)
6. Grant, S., Yelam, A., Bland, M., Snoeren, A.C.: SmartNIC performance isolation
with FairNIC: programmable networking for the cloud. In: Proceedings of the ACM
SIGCOMM, pp. 681–693 (2020)
7. Harkous, H., Jarschel, M., He, M., Priest, R., Kellerer, W.: Towards understanding
the performance of p4 programmable hardware. In: ACM/IEEE Symposium on
Architectures for Networking and Communications Systems, pp. 1–6. IEEE (2019)
8. Hohemberger, R., Castro, A.G., Vogt, F.G., Mansilha, R.B., Lorenzon, A.F., Rossi,
F.D., Luizelli, M.C.: Orchestrating in-band data plane telemetry with machine
learning. IEEE Commun. Lett. 23(12), 2247–2251 (2019)
9. Hohemberger, R., Lorenzon, A.F., Rossi, F.D., Luizelli, M.C.: A heuristic app-
roach for large-scale orchestration of the in-band data plane telemetry problem.
In: Barolli, L., Amato, F., Moscato, F., Enokido, T., Takizawa, M. (eds.) Advanced
Information Networking and Applications, pp. 381–392. Springer International
Publishing (2020)
10. Liu, M., Cui, T., Schuh, H., Krishnamurthy, A., Peter, S., Gupta, K.: Offloading
distributed applications onto SmartNICs using iPipe. In: Proceedings of the ACM
Special Interest Group on Data Communication, pp. 318–333 (2019)
11. Netronome: Internet (2020). https://www.netronome.com/static/app/img/
products/silicon-solutions/WP NFP4000 TOO.pdf
12. Pizzutti, M., Schaeffer-Filho, A.E.: Adaptive multipath routing based on hybrid
data and control plane operation. In: IEEE INFOCOM, pp. 730–738 (2019)
13. Qiu, Y., Kang, Q., Liu, M., Chen, A.: Clara: performance clarity for SmartNIC
offloading. In: Proceedings of the ACM Hot Topics in Networks, pp. 16–22 (2020)
14. Sanvito, D., Siracusano, G., Bifulco, R.: Can the network be the AI accelerator?
In: Proceedings of the Workshop on In-Network Computing, pp. 20–25 (2018)
194 P. B. Viegas et al.

15. Sapio, A., Abdelaziz, I., Aldilaijan, A., Canini, M., Kalnis, P.: In-network com-
putation is a dumb idea whose time has come. In: Proceedings of the 16th ACM
Workshop on Hot Topics in Networks, pp. 150–156 (2017)
16. Siracusano, G., Bifulco, R.: In-network neural networks. arXiv preprint
arXiv:1801.05731 (2018)
17. Wang, S.Y., Wu, C.M., Lin, Y.B., Huang, C.C.: High-speed data-plane packet
aggregation and disaggregation by p4 switches. J. Netw. Comput. Appl. 142, 98–
110 (2019)
18. Wang, S., Meng, Z., Sun, C., Wang, M., Xu, M., Bi, J., Yang, T., Huang, Q.,
Hu, H.: SmartChain: enabling high-performance service chain partition between
SmartNIC and CPU. In: IEEE International Conference on Communications, pp.
1–7. IEEE (2020)
19. Xiong, Z., Zilberman, N.: Do switches dream of machine learning? Toward in-
network classification. In: Proceedings of the 18th ACM Workshop on Hot Topics
in Networks, pp. 25–33 (2019)
 On the Adaptation of Bitcoin-Like
Blockchains for the Secure Storage
of Traffic-Related Events

El-hacen Diallo1(B) , Omar DIB2 , and Khaldoun Al Agha1

1
CNRS, Laboratoire Interdisciplinaire des Sciences du Numérique,
Université Paris-Saclay, 91190 Gif-sur-Yvette, France
{diallo,alagha}@lisn.fr
2
Department of Computer Science, Wenzhou-Kean University, Wenzhou, China
[email protected]

Abstract. This paper investigates Bitcoin-like blockchains enabled traf-

fic efficiency applications in VANETs. Prior works do not provide a thor-
ough study on the adaptation of the Proof-of-Work (PoW) mechanism
in the context of traffic records management, nor a detailed study of per-
formance. Therefore, in this work, we study a PoW based blockchain for
secure traffic-related records in the VANETs’ context. We introduce a
blockchain simulator to assess our protocol’s performance (i.e., through-
put and latency). Results show that the proposed scheme ensures the
integrity and traceability of the shared traffic-related records and point
out trade-offs between security and performance.

1 Introduction
The rapid development of modern devices, and the evolution of wireless tech-
nology, have extended Vehicular Ad-hoc NETworks’ (VANETs) applications to
the next level. Autonomous vehicles are equipped with smart devices such as
OnBoard Units (OBUs), sensors, Global Positioning System (GPS) receivers.
They are also supplied with WAVE (Wireless Access in Vehicular Environments)
module, easing their interaction with nearby vehicles as well as with infrastruc-
tures deployed on the road, i.e., Road-Side-Units (RSUs). A huge amount of
traffic-related data is collected throughout vehicles. And this data is used to
enhance traffic safety and efficiency by minimizing accidents and providing opti-
mized trip planning, parking services, real-time streaming, etc. Traditionally,
collected traffic records are uploaded and managed relying on a central cloud [1].
However, such centralized architectures face privacy issues and high bandwidth
and storage cost which can lead to severe network congestion [2]; they are also
highly prone to a single point of failure vulnerability.
The Edge computing paradigm is a potential solution to cope with the afore-
mentioned drawbacks of centralized schemes. The notion of Edge computing
consists of splitting the cloud to Edge devices and bringing it to the edge of the
network [2], in other words, bringing facilities closer to vehicles. For example,
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 195–207, 2021.
https://doi.org/10.1007/978-3-030-75100-5_18
196 E. Diallo et al.

Edges nodes could be embedded on RSUs. Doing so will minimize network con-
gestion, bandwidth usage and, more importantly, help support real time data
processing. As a result, the QoS related to traffic management will be improved.
Nevertheless, although the Edge cloud paradigm will offer better QoS by
tackling latency issues in cloud-centric systems, it has brought along daunting
security challenges. Ensuring secure and transparent communication between
Edge devices is a major challenge because of the Edge computing’s distributed
nature. Moreover, the network has become more vulnerable as it is difficult to
protect Edge devices from intrusion, unauthorized access, and various attacks [3].
We believe that blockchain technology, and its exciting capacity of maintain-
ing a decentralized and distributed database between untrusted entities without
relying on a central authority, can meet the aforementioned security issues. The
Blockchain technology consists of a peer-to-peer network exploiting a consen-
sus algorithm to agree on a distributed database state. From a data structure
point of view, a blockchain is a linked list of blocks, where each block stores the
hash of the previous block. As the core of any blockchain platform, the consen-
sus mechanism is the main component that differentiates blockchain platforms.
Introduced as the underlying consensus mechanism with Bitcoin [4], the first
blockchain platform, the Poof-of-Work (PoW) consists of dedicated processors
trying to solve a difficult but easy-to-verify cryptographic puzzle. PoW based
blockchains may witness inconsistencies that cause branches in the chain. This
phenomenon is known as forks, where more than one valid block is broadcasted
to the network simultaneously [5]. In Bitcoin, the most extended branch is con-
sidered as the main chain. And blocks that do not belong to the main chain are
called stale blocks.
In this work, we take advantage of the decentralization, data integrity, and
public variability properties of the blockchain, aiming to secure related traffic
records. We focus on investigating and studying the adoption of the Proof-of-
Work consensus for efficient traffic-related records management.

2 Related Work
A considerable number of applications in VANET’s context relied on the Proof-
of-Work mechanism. Such PoW-based applications can be categorized as follows:

Vehicles as Blockchain Nodes: In [6], authors proposed a blockchain-based

anonymous reputation system (BARS), aiming to prevent the distribution of
forged messages. The global consensus mechanism was provided by the vehi-
cles assuming they have enough amount of resources to act as miners. Also,
in [7], basic concepts of blockchain were used for data sharing among intelligent
vehicles while relying on the PoW algorithm. Unfortunately, no details were pro-
vided about the blockchain design, nor the protocol’s performance metrics, thus
making the assessment of their work very hard.
Indeed, the idea of using vehicles as blockchain nodes seems interesting from a
decentralization point of view. However, such a scheme faces challenges regarding
 On the Adaptation of Bitcoin-Like Blockchains for the Secure Storage 197

the scalability, latency, and global blockchain update [8]. Because of the storage,
communication, and computing load resulting from the usage of the PoW as con-
sensus protocol between vehicles, guaranteeing the reliability of latency-sensitive
applications (e.g., safety messages dissemination) is a challenging task [9]. To
mitigate the storage overhead on vehicles, in this work [10], the authors sug-
gested a light blockchain. However, a scalability evaluation is still needed.

RSUs as Blockchain Nodes: On the other hand, RSUs have been used as
blockchain miners. For example, in [11] the PoW mechanism is executed by the
RSUs to build a secure and trustworthy event message delivery in VANETs. Also,
in [12] RSUs were relied on as blockchain nodes, which compute to solve the PoW
puzzle. The authors’ goal is to propose a blockchain scheme for secure announce-
ment messages sharing in IoV (Internet of Vehicles). Furthermore, in [13], the
authors proposed a blockchain-based protocol for distributed trust management,
where a fixed number of RSUs validate events using fuzzy logic and participate
in the block creation process using PoW.

3 Limitations of Related Works

All the works mentioned above lack metrics related to their system performance.
Indeed, VANETs applications are usually delay-sensitive; it is, therefore, crucial
to provide metrics such as message confirmation time and the number of trans-
actions that can be processed by the blockchain in a unit of time.
Furthermore, data validation in VANETs applications differs from Bitcoin
and other coin based blockchains. The latter relies only on the history (i.e.,
committed data) for transaction validation. In contrast, traffic-related records
need to be approved by the witnesses of the concerned events. Therefore, it is
essential to investigate and study the PoW mechanism’s suitability in VANETs
applications; and particularly for implementing a secure traffic-related records
sharing system.
Despite PoW, other consensus algorithms were introduced in the context of
VANETs. We cite: Proof-of-Event (PoE) [14] and Proof-of-Driving (PoD) [7];
however, further discussions regarding their security models are required so that
such protocols can be applied reliably and securely. PBFT (Practical Byzan-
tine Fault Tolerance) [15], and Proof-of-Authority (PoA) [16] were also adopted;
nevertheless, such protocols seem to be centralized since a pre-defined subgroup
of RSUs is designated by a central trust authority to maintain the blockchain.
Furthermore, stake-based protocols like Proof-of-Stake (PoS) [17] was also intro-
duced. Although this approach mitigates the CPU demands, new issues arise that
were not present in PoW-based blockchains, such as Nothing at stake attack and
Grinding attack [18].

4 Contributions
To the best of our knowledge, the state-of-the-art lacks a more throughout study
about adoption of PoW based blockchain in the context of traffic-related records
198 E. Diallo et al.

management, followed by a rigorous performance evaluation. The main contri-

butions of this paper are as follows:
• We detail how the Proof-of-Work (PoW) can be adapted to secure the traffic
messages in the VANETs context.
• We analyze and discuss PoW contribution regarding traffic-events validation.
• We present a benchmark framework whereby we assess the performance of
PoW enabled traffic events management in terms of throughput (i.e., number
of confirmed events per second) and latency (event confirmation delay).
• We also assess the impact of the PoW puzzle difficulty on the performance of
the system (i.e., throughput and latency).
• Finally, we highlight and discuss trade-offs between security and performance.
The remainder of this article is organized as follows. In Sect. 5, we present the
proposed architecture. Section 6 describes the protocol, which is a blockchain-
based scheme for traffic records storage. And Sect. 7 describes the implemented
simulation environment. In Sect. 8, the performance of the proposed system is
assessed. Finally, Sect. 10 concludes this paper and discusses future works.

5 System Model
In this section, we describe the system model, where a blockchain is enabled
traffic records management. We first describe the system architecture, then high-
light VANETs applications and finally discuss traffic events validation techniques
(Fig. 1).

Fig. 1. System architecture

 On the Adaptation of Bitcoin-Like Blockchains for the Secure Storage 199

5.1 System Components

The proposed architecture can be separated into three layers. The first layer rep-
resents the Ad-hoc domain, which represents intelligent vehicles. Through their
dedicated communication modules and embedded sensors, cars record relevant
traffic events (e.g., accidents, traffic congestion data, safety warnings). These
records are signed by vehicles with their private keys to ensure the communica-
tions’ authenticity and reliability before being uploaded to RSUs.
The second layer of the proposed architecture illustrates RSUs connections,
which form a peer-to-peer network. Each RSU is supplied with an Edge node,
providing the required resources (i.e., computation and storage) to maintain a
blockchain and solve the PoW puzzle. Received traffic records are stored into
RSU’s mempool under a pending state. Pending events wait for a threshold of
confirmations before turning to valid.
The last layer represents the Traffic Management Authority (TMA); it is a
trusted organization generating cryptographic credentials to newly joining vehi-
cles and RSUs. The provided certificates must be anonymous to guarantee vehicle
privacy. Besides, the TMA monitors the blockchain and inflicts punishment to
misbehavior actions.
We recall here that the paper’s scope is to investigate and study the adoption
of PoW towards traffic-related records management.

5.2 Traffic-Related Events Validation

Fig. 2. RSU-centric events validation Fig. 3. Vehicle-centric events valida-

tion

Messages validation is the first step toward trustworthy traffic records sharing
and storage since data issuers could be reporting fictitious events. We classify
existing VANETs protocols for events validation into two main groups: Rsu-
centric events validation and Vehicle-centric events validation methodologies.

RSU-Centric Events Validation Protocol: In this category, RSUs validate

messages uploaded by vehicles. More precisely, after receiving a message report-
ing a traffic event, a RSU waits for a threshold of confirmation messages from
vehicles before considering the event to be plausible. If necessary, RSUs would
request confirmations from nearby vehicles as illustrated in Fig. 2.
200 E. Diallo et al.

Vehicle-Centric Events Validation: Differently, in this approach, events are

validated and signed by vehicles through threshold of signatures or a consensus
algorithm as suggested in [12]. As we can notice, no need for RSUs intervention
as illustrated in Fig. 3. However, reaching a consensus between highly mobile
vehicles could be an obstacle. Indeed, such a scheme may add a substantial load
to the vehicle-to-vehicle communications (V2V), hence engendering a significant
delay to the event validation time.
Therefore, in this paper, we will be considering the RSU-centric events vali-
dation methodology when designing our model. More precisely, in our protocol,
the RSUs will act as blockchain nodes; they will receive traffic events from vehi-
cles and use a PoW to validate the events according to a threshold and then
update their maintained blockchain nodes accordingly. As such, a RSU-based
blockchain will be built to achieve a decentralized, transparent, immutable, and
secure traffic-related system in the VANETs context. In the following, we will
describe in detail how the blockchain is enabled traffic message storage.

6 Protocol Description

Fig. 4. Model flow

This section describes the proposed scheme for secure traffic-related data valida-
tion, storage, and sharing. In Fig. 4, we present a diagram flow of the proposed
model, which consist of 5 steps. We also provide pseudo-codes describing the
block creation and confirmation processes.
The first step of our protocol is the collection of traffic-related messages. This
phase consists of vehicles reporting hazardous traffic events such as accidents,
attacks performed by other vehicles, traffic jams, ice on the road, etc. Collected
data are used to optimize road-trips, traffic flow, and hence reduce the CO2
emissions and fuel consumption; they can also help report vehicle’s misbehavior
and help reduce accidents caused by traffic jams. The immutability property of
the blockchain ensures received data integrity. Nevertheless, collected messages
have to be validated in the first place. So, in steps 2 and 3 of the protocol, RSUs
 On the Adaptation of Bitcoin-Like Blockchains for the Secure Storage 201

will gather enough confirmations from surrounding vehicles. After a threshold is

reached, the corresponding event will be valid. Step 4 consists of block mining,
whereas RSUs take advantage of their CPUs to solve the Proof-of-Work puzzle.
Algorithm 1 depicts how RSUs mine blocks. Next, in step 5, the block is spread
through the network. Finally, step 6 consists of mined block verification. When
a RSU receives a new block, it verifies if the block has been correctly signed, the
PoW was correctly solved, and finally, if the required threshold of confirmations
was reached for each event within the block. After verification, the RSU will stop
trying to mine the same block (in PoW, RSUs are parallelly trying to mine the
next block) and add the block to its local blockchain. See Algorithm 2.

Algorithm 1. Block Mining

Input : mempool : mempool;
blockchain : chain ;
peers : peers;
block time out : blocktime ;
PoW difficulty : d
1: b ← createBlock()
2: nonce ← 0
3: tries ← 10000
4: repeat
5: b.setBlockN once(nonce)
6: b.calculateBlockHash();
7: if b.getHash().getLeadingZeros() ≥ d then
8: break
9: else if chain.getHeight() ≤ b.getHeight() then
10: break
11: end if
12: nonce ← nonce + 1
13: until nonce ≤ tries
14: if nonce ≤ tries then
15: b.shuf f leT ransactions()
16: else
17: chain.addBlock(b)
18: mempool.update()
19: end if
20: wait(blocktime ) % wait for block spreading delay %
21: Repeat() % Repeat the mining process %

6.1 Security Model

PoW security lies down on the global computing power within the system; for
example, Bitcoin is considered secure as long as adversary miners do not hold
51% of the total CPU power [4]. In Bitcoin, over 1 million miners maintain the
202 E. Diallo et al.

Algorithm 2. Receive block

Input : mempool : mempool;
blockchain : chain ;
block : block;
peers : peers;
PoW difficulty : d
1: if chain.HasBlock(block.getHash()) then
2: return
3: end if
4: height ← block.getHeight()
5: if isValidBlock(block) then
6: stopM inig(height)
7: chain.addBlock()
8: chain.update()
9: sendT o(peers)
10: M ine(height + 1) % Start mining the next block %
11: end if
12: function isValidBlock(block)
13: hash ← block.concAttributes() % Concatenate all block atrributes except
block hash %
14: if SHA256(hash) <> hash then
15: return F ALSE
16: else if hash.getLeadingZeros() < d then
17: return F ALSE
18: else if notblock.hasT hreshold() then % Test if has enough confirmations from
vehicles %
19: return F ALSE
20: end if
21: return T RU E
22: end function

blockchain, hence the challenge of getting the majority of the computation power.
In our context, the blockchain might be for a given zone, where the number of
RSUs is a lot lower than Bitcoin. Miners are RSUs; therefore, it is easier to
obtain the majority of RSUs global CPUs. However, RSUs are authenticated
in advance; thus, only known RSUs can join the network. Hence, the network
is less subject to external attacks or RSUs conspiring to attack the system.
Besides the 51% attack, forks increase the advantage of the adversary in the
network [19]. Essentially, the more the number of forks is important, the less
secure the blockchain becomes. Therefore, in the evaluation section, we assess
the impact of forks on the overall performance of the adopted PoW consensus
protocol.
 On the Adaptation of Bitcoin-Like Blockchains for the Secure Storage 203

7 Simulation Environment
We have designed our simulation environment using NS-3 [20], a discrete events
simulator. RSUs represent nodes in NS-3; each of them runs an instance of
the blockchain. The positioning of RSUs is settled following a grid topology.
RSUs are connected by a point to point channel to form a peer-to-peer network.
The
√ average distance between RSUs is 8–11 km. Each RSU has between 2 to
N connections (peers) generated randomly, where N designate the number of
RSUs. And a TCP connection is settled between RSUs. The network connection
speed is 100 Mbps. Moreover, communications are encrypted using Schnorr sig-
nature [21] implemented using the OpenSSL library. Finally, SHA-256 is used as
hash function.
The presented simulator is used to picture the life cycle of traffic-events
storage in VANETs, from the collection, reception, and validation of such events
to the blockchain confirmation and final persistence in the ledger.
The evaluated metrics are the average number of events that can be confirmed
per second, i.e., the throughput of the blockchain (event/s), the average delay
between traffic event creation time to event confirmation, i.e., the latency (s).
Also, we assess the impact of the PoW puzzle difficulty (d), and the number of
RSUs involved in the consensus protocol (N ), as well as events arrival rate, i.e.,
f (event/s) on the performance of the system (i.e., throughput and latency).
We suppose that the traffic is dense enough and that the fixed threshold of
needed confirmation messages to validate an event will be reached in less than
λ = 500 ms. Some parameters of the simulator have been fixed, such as the event
size is set to 800 bytes as defined in [15,17], and the period to generate events
is set to 60 s (i.e., if the event arrival rate (f ) is equal to 50 event/s, there will
be 50 × 60 generated events).
The above-mentioned parameters include real-world configurations. For
instance, we have analyzed a dataset of alert messages from San Franciso’s city
for the year 2019 [22]. The traffic warnings arrival rate is 60 events per day, which
is not that much. Indeed, the frequency of traffic alert arrival will increase with
the number of autonomous vehicles. Therefore, when running simulations, we
vary the events’ arrival rate from 1–500 event/s to capture respectively low and
heavy traffics. On the other hand, regarding the number of RSUs and the network
speed. The RSUs can be communicating using WiMAX [23], as the converge rate
can reach 15 km with a communication bandwidth of 100 Mbps [23]. Thus, less
than 24 RSUs is enough to cover San Francisco, whose area is 121,4 km2 .

8 Performance Evaluation
This section evaluates various instances of the PoW enabled traffic-related data
management protocol using a server with the following settings: Dell R640 server,
Intel(R) Xeon(R) Silver 4112; CPU 2.60 GHz; 8 core CPU; 64 GB RAM, and
running Ubuntu 18.04. The plotted results represent end-to-end measurements
from all RSUs. The throughput, i.e., the number of confirmed events per second
204 E. Diallo et al.

(event/s), is examined by dividing the total confirmed events by the simulation

time. As for the latency, we subtract the creation timestamp from the generated
event’s confirmation timestamp for each event; then, the average delay is com-
puted. Each experiment is repeated for 10× with different seeds, and the mean
is plotted with errors, using the 95% confidence interval.

Table 1. PoW solving delay.

PoW difficulty (leading zeros) Delay (s)

2 0.0014
4 0.2635
6 91.5003

We evaluate the PoW puzzle-solving delay by launching multiple instances of

PoW with various difficulties (2, 4, and 6). For each difficulty (d), we repeat the
simulation 100× to obtain enough granularity. All those 100 PoW puzzle-solving
delays are stored in a file through which the RSUs read the PoW delays. Table 1
represents the mean of the filtered results from the file mentioned above. Results
shows that PoW delay increases exponentially with d.

Fig. 5. Performance (throughput, Fig. 6. Performance (throughput,

latency) vs. Events arrival rate - f latency) vs. Number of RSUs

Figure 5 depicts the system’ performance of with the increasing events arrival
rate (f ). Results show that reducing the PoW difficulty (d) from 6 to 4 enhances
the performance. When d is set to 4, the system can process up to 300 event/s
out of 500 event/s generated, and on average, the event confirmation delay does
not exceed 2 s. Lower performance is depicted when d = 6 as the throughput
does not exceed 106 event/s and the latency can attain 18 s. This significant
difference of performance comes from PoW solving’ high delay when d = 6.
 On the Adaptation of Bitcoin-Like Blockchains for the Secure Storage 205

Figure 6 shows the impact of the number of RSUs on the system’s perfor-
mance. As can be seen, the performance decreases with the increasing number
of RSUs. For example, in the worst case, if d = 4, increasing the number of
RSUs from 4 to 24 decreases the throughput by 141 event/s. This drop in per-
formance is partially due to the block spreading delay, which increases with the
network size; nevertheless, it is not the leading cause. Because when the number
of RSUs is important, there are more collisions, i.e., RSUs solving the PoW puzzle
simultaneously; hence, worsening the system performance. As shown in Fig. 7,
the number of stale blocks increases with the number of RSUs. For instance,
increasing the number of RSUs from 4 to 24 induces 39% of mined blocks to be
stale. Hence, affecting the system performance.

Fig. 7. Stale blocks Fig. 8. Stale blocks (%) Fig. 9. Forks size vs.
(%) vs. Number of vs. Events arrival rate-f . Events arrival rate-f .
RSUs

In Figure 8, the rate of stale blocks (rs) versus events arrival rate (f ) is
assessed. Results show that if d = 4, rs can attain 39%. i.e., RSUs waste 39%
of their resources (i.e., computation and bandwidth) on processing blocks that
won’t be part of the blockchain. That’s because, with low difficulty, many RSUs
will be likely to solve the puzzle at the same time. However, with d = 6, no
resources are wasted. Nevertheless, in return, the gained time will be exhausted
on PoW solving.
To determine the required number of blocks before considering a block to be
persistent in the blockchain, in Fig. 9, we measure the most extended fork size,
i.e., the longest branch that does not belong to the blockchain. The obtained
results show that implicitly reducing PoW puzzle difficulty (d) increases the
forks’ size because the RSUs are likely to solve the PoW at the same time. In
case d = 4, the most extended fork is equal to 2 blocks, while for d = 6, there
are no forks.

9 Discussion
Although reducing the Proof-of-Work performance would minimize the events’
confirmation delay, it may imply a significant number of forks, which disturbs the
blockchain’s immutability. Moreover, forks cause a huge waste of computation
206 E. Diallo et al.

and bandwidth. On the other hand, a greater PoW difficulty induces a consid-
erable delay due to the PoW solving task. Hence, raising a trade-off between
security and performance.

10 Conclusion
We investigated and adapted PoW-of-Work (PoW) based blockchain for traffic-
related events securing in VANETs’ context. We showed how the PoW adap-
tation differs from the Bitcoin-like use cases. We also presented a benchmark
framework whereby we evaluated different instances of the studied protocol; the
assessed metrics are the number of events processed by the blockchain (through-
put) and events’ confirmation time (latency). Results show that PoW can be
adapted to validate and track traffic-messages in VANET’s. From a performance
point of view, it was shown that reducing PoW difficulty engenders a significant
number of stale blocks; thus, weakening the security and inducing extra resource
costs (communication and computation load). In future works, we have planned
to study the impact of vehicle mobility on blockchain performance and to assess-
ing the efficiency of the protocol under malicious vehicles.

References
1. Lee, E., Lee, E.-K., Gerla, M., Oh, S.Y.: Vehicular cloud networking: architecture
and design principles. IEEE Commun. Mag. 52(2), 148–155 (2014)
2. Kaur, K., Garg, S., Aujla, G.S., Kumar, N., Rodrigues, J.J.P.C., Guizani, M.:
Edge computing in the industrial internet of things environment: software-defined-
networks-based edge-cloud interplay. IEEE Commun. Mag. 56(2), 44–51 (2018)
3. Khan, W.Z., Ahmed, E., Hakak, S., Yaqoob, I., Ahmed, A.: Edge computing: a
survey. Future Gener. Comput. Syst. 97, 219–235 (2019)
4. Nakamoto, S., et al.: Bitcoin: A peer-to-peer electronic cash system (2008)
5. Shahsavari, Y., Zhang, K., Talhi, C.: A theoretical model for fork analysis in the bit-
coin network. In: 2019 IEEE International Conference on Blockchain (Blockchain),
pp. 237–244. IEEE (2019)
6. Zhaojun, L., Liu, W., Wang, Q., Gang, Q., Liu, Z.: A privacy-preserving trust
model based on blockchain for VANETs. IEEE Access 6, 45655–45664 (2018)
7. Singh, M., Kim, S.: Blockchain based intelligent vehicle data sharing framework.
arXiv preprint arXiv:1708.09721 (2017)
8. Mendiboure, L., Chalouf, M.A., Krief, F.: Survey on blockchain-based applications
in internet of vehicles. Comput. Electr. Eng. 84, 106646 (2020)
9. Elagin, V., Spirkina, A., Buinevich, M., Vladyko, A.: Technological aspects of
blockchain application for vehicle-to-network. Information 11(10), 465 (2020)
10. Mostafa, A.: VANET blockchain: a general framework for detecting malicious vehi-
cles. J. Commun. 14(5), 356–362 (2019)
11. Shrestha, R., Bajracharya, R., Shrestha, A.P., Nam, S.Y.: A new-type of blockchain
for secure message exchange in VANET. Digit. Commun. Netw. 6, 177–186 (2019)
12. Zhang, L., Luo, M., Li, J., Au, M.H., Choo, K.-K.R., Chen, T., Tian, S.: Blockchain
based secure data sharing system for internet of vehicles: a position paper. Veh.
Commun. 16, 85–93 (2019)
 On the Adaptation of Bitcoin-Like Blockchains for the Secure Storage 207

13. Kchaou, A., Abassi, R., Guemara, S.: Toward a distributed trust management
scheme for VANET. In: Proceedings of the 13th International Conference on Avail-
ability, Reliability and Security, p. 53. ACM (2018)
14. Yang, Y., Chou, L., Tseng, C., Tseng, F., Liu, C.: Blockchain-based traffic event
validation and trust verification for VANETs. IEEE Access 7, 30868–30877 (2019)
15. Zhang, X., Chen, X.: Data security sharing and storage based on a consortium
blockchain in a vehicular Adhoc network. IEEE Access 7, 58241–58254 (2019)
16. De Angelis, S., Aniello, L., Baldoni, R., Lombardi, F., Margheri, A., Sassone,
V.: PBFT vs proof-of-authority: applying the CAP theorem to permissioned
blockchain (2018)
17. Yang, Z., Yang, K., Lei, L., Zheng, K., Leung, V.C.M.: Blockchain-based decen-
tralized trust management in vehicular networks. IEEE Internet Things J. 6(2),
1495–1505 (2018)
18. Siim, J.: Proof-of-stake. In: Research Seminar in Cryptography (2017)
19. Gervais, A., Karame, G.O., Wüst, K., Glykantzis, V., Ritzdorf, H., Capkun, S.:
On the security and performance of proof of work blockchains. In: Proceedings of
the 2016 ACM SIGSAC Conference on Computer and Communications Security,
pp. 3–16 (2016)
20. https://www.nsnam.org . Network simulator ns-3
21. Schnorr, C.-P.: Efficient identification and signatures for smart cards. In: Confer-
ence on the Theory and Application of Cryptology, pp. 239–252. Springer (1989)
22. Moosavi, S., Samavatian, M.H., Nandi, A., Parthasarathy, S., Ramnath, R.: Short
and long-term pattern discovery over large-scale geo-spatiotemporal data. In: Pro-
ceedings of the 25th ACM SIGKDD International Conference on Knowledge Dis-
covery & Data Mining, pp. 2905–2913 (2019)
23. Vaughan-Nichols, S.J.: Achieving wireless broadband with WiMAX. IEEE Com-
put. 37(6), 10–13 (2004)
 Trustworthy Acceptance: A New Metric
for Trustworthy Artificial Intelligence
Used in Decision Making
in Food–Energy–Water Sectors

Suleyman Uslu1 , Davinder Kaur1 , Samuel J. Rivera2 , Arjan Durresi1(B) ,

Mimoza Durresi3 , and Meghna Babbar-Sebens2
1
Indiana University-Purdue University Indianapolis, Indianapolis, IN, USA
{suslu,davikaur}@iu.edu, [email protected]
2
Oregon State University, Corvallis, OR, USA
{sammy.rivera,meghna}@oregonstate.edu
3
European University of Tirana, Tirana, Albania
[email protected]

Abstract. We propose, for the first time, a trustworthy acceptance met-

ric and its measurement methodology to evaluate the trustworthiness of
AI-based systems used in decision making in Food Energy Water (FEW)
management. The proposed metric is a significant step forward in the
standardization process of AI systems. It is essential to standardize the
AI systems’ trustworthiness, but until now, the standardization efforts
remain at the level of high-level principles. The measurement methodol-
ogy of the proposed includes human experts in the loop, and it is based
on our trust management system. Our metric captures and quantifies
the system’s transparent evaluation by field experts on as many control
points as desirable by the users. We illustrate the trustworthy acceptance
metric and its measurement methodology using AI in decision-making
scenarios of Food-Energy-Water sectors. However, the proposed metric
and its methodology can be easily adapted to other fields of AI appli-
cations. We show that our metric successfully captures the aggregated
acceptance of any number of experts, can be used to do multiple mea-
surements on various points of the system, and provides confidence values
for the measured acceptance.

1 Introduction
Despite so many advantages of AI systems and their uses, these systems some-
times directly or indirectly harm the users and society. It has become essential
to make these systems safe, reliable, and trustworthy. Lately, trustworthy AI has
been gaining increasing attention from governments, organizations, and scien-
tific communities. So European Union (EU) has proposed ethnic guidelines and
laws [50] for trustworthy AI to govern and facilitate the development and working
of AI systems [13]. DARPA [15] also launched an XAI program, whose motive
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 208–219, 2021.
https://doi.org/10.1007/978-3-030-75100-5_19
 Trustworthy Acceptance: A New Metric for Trustworthy 209

was to make these AI systems explainable and trustworthy. Garter estimates

that 30% of all the digital products that use AI will require a trustworthy AI
framework by 2025 [6], and 86% of users will trust and remain loyal to companies
that use ethical AI principles [4]. So, developing AI systems using a trustwor-
thy framework has become a necessity for today’s society. Furthermore, there
are various recent studies on developing trustworthy and explainable algorithms
and AI [19,28,29,40–42,49,55]. Kaur et al. [23] surveyed similar approaches that
aimed to create trustworthy and explainable AI systems.
And as AI technologies mature, they have to follow the natural process that
all established technologies have gone through, standardization. Among many
advantages, standards allow manufacturers and users to speak the same lan-
guage, enable the users to check the quality of products on the market, reduce
the legal liability of manufacturers and providers, etc. Already ISO, an orga-
nization that deals with standardization, has presented different approaches to
establish trust in AI systems using fairness, transparency, accountability, and
controllability [20]. However, such works remain at the level of high-level prin-
ciples. What is needed is to develop metrics and measurement procedures to
establish standards for trustworthy AI [27]. We envision that in the future, vari-
ous agencies will use such metrics to certify AI-based solutions, similarly as FDA
certifies medications and treatments. For this reason, in this paper, for the first
time, we propose a concrete metric, trustworthy acceptance, and its measure-
ment methodology. Our metric captures and quantifies the system’s transparent
evaluation by field experts.
In this paper, we use the following definition: Trustworthy AI is a framework
to ensure that a system is worthy of being trusted concerning its stated require-
ments based on the evidence. It makes sure that the users’ and stakeholders’
expectations are met in a verifiable way [20]. Furthermore, for the time being, AI
lacks many aspects of human intelligence, including meaning, multidimensional
data beyond the set used for algorithm training, meaningful causality, ethics,
etc. Therefore, AI systems should complement and empower humans without
replacing them. This is the essential requirement to make AI trustworthy. And
for this reason, when developing the proposed metric and corresponding mea-
surement procedures, we include human experts in the loop. Our contributions
can be summarized as follows:

• We propose in Sect. 4 a trustworthy acceptance metric for the evaluation of

the acceptance of AI-based systems by field experts.
• The measurement procedure for the proposed metric is described in Sect. 4
and is based on the concept of a distance acceptance approach that is adapt-
able to a wide range of systems. In addition to the acceptance value, our
metric provides the confidence of the acceptance.
• Our metric utilizes the trust of the experts in the given context, managed by
our trust system, summarized in Sect. 3.
• Our metric can be measured in many points of the system in order to reach
an assessment of the whole system, as discussed in Sects. 4 and 5.
210 S. Uslu et al.

• Finally, in Sect. 5, we illustrate the application of our trustworthy acceptance

metric and its measurement methodology using three systems for environ-
mental decision making.

2 Related Work
As part of our daily lives, decision-making is also an essential element of pro-
cesses of the most significant fields such as economics, finance, healthcare, and
the environment. Kambiz [21] explained the importance of decision makings in
such fields by giving examples of the crucial results caused by erroneous or dis-
satisfactory decisions such as the world economic crisis. He indicated that the
difficulty of making such critical decisions relies on their complex nature and
the involvement of multiple stakeholders who can have different expertise, back-
ground, perspectives, or maybe even competing for objectives.
As it is possible and common in some areas to have experts as stakeholders,
there is a need to have a consensus mechanism for such decision makings. Dong
and Xu [11] proposed approaches to minimize the modifications to the solutions
that the experts propose at each round of the decision making. Furthermore,
Hegselmann and Krause [17] investigated the consensus reaching mechanisms of
decision makings involving agents with different behaviors in both mathematical
modelings and computer simulations. Similarly, Babbar-Sebens and Minsker [3]
proposed a design employing an algorithm for the utilization of expert feedback
for an improved optimization in the environmental field. These studies clearly
show a considerable need for methods and frameworks for advanced decision-
making, especially for the ones where humans and machines need to collaborate
for superior results.
Trust has long been a concept that is believed to be an essential part of the
decision-making process especially involving multiple stakeholders. It has been
shown that the utilization of trust contributed to the integrative behavior and
helped disruptive activities to decline [14,18,25]. There has been multiple stud-
ies [30–32,52] and surveys [10,33,53] discussing trust management and its frame-
works. In [39], a trust management framework based on measurement theory is
proposed for online social communities. There are several applications of this
framework, such as stock market prediction with Twitter data [37]. Other exam-
ples include trust management in social networks [54], cloud computing [34,35],
internet of things [36,38], healthcare [8,9], emergency communications [12], and
detection of crime [24] and fake users [22].
In the environmental field, Alfantoukh et al. [1,2] proposed a model for water
allocation problem and a more generic model for consensus reaching problem
involving trust. We introduced a trust-based decision support system for natu-
ral resource sharing problems in Food-Energy-Water sectors [48]. Also, we pre-
sented the enhanced versions of our system utilizing discrete and precomputed
solutions [43], game-theoretical approach [44], and scenarios with different eval-
uation criteria [45]. Furthermore, we presented the role of trust sensitivity of
actors in environmental trust-based decision-making scenarios [47].
 Trustworthy Acceptance: A New Metric for Trustworthy 211

When computers make decisions, the liability of the decisions becomes a

significant issue. Therefore, it requires a comprehensive testing process before
deploying and utilizing such systems, which brings the acceptance of a system
in the picture. Although there are studies to reach a desired level of acceptance,
proposals could be concentrated on revealing and satisfying the needs of the
users [26]. Also, there are studies to model users’ desire to use such AI-based
systems [16]. One of the most critical areas that we need the high acceptance of
is healthcare, where trust between doctors and computerized systems is highly
crucial [51]. In [24], it has been shown that in decision makings involving both
humans and computers, there is no exact winner when considering all scenarios
of crime detection, which opens the door for collaboration for more significant
results. Similarly, in [22], AI-based fake user detection gave better results when
it utilized the community’s preferences.

3 Trust Management Framework

We summarize here our measurement theory-based trust management framework
[39], because the new metric and its methodology are based on this framework.
Trust defined in this framework has two main parameters: impression, m, and
confidence, c. Impression represents the level of trust from one party to another,
while confidence is the degree of certainty of the impression. Impression calcu-
lation is done by averaging the measurements, as shown in Eq. 1 whereas the
confidence is calculated related to the standard error of the mean as shown in
Eq. 2. In these formulas, mA:B , cA:B , and riA:B are the impression, confidence,
and a measurement from A to B.
N A:B
r
m A:B
= i=1 i (1)
N

N A:B − r A:B )2
i=1 (m
cA:B = 1 − 2 i
(2)
N (N − 1)
Another essential feature of this framework is to anticipate the trust even
without communication between two parties. It supports the propagation meth-
ods, namely transition and aggregation, to assess trust between two entities
that are connected through third-party nodes. Ruan et al. [39] proposed several
trust propagation methods and provided their error propagation functions. For
aggregation, we selected the averaging method, as shown in Eq. 3, and provided
its error formula in Eq. 4. In these formulas, S represents the source, D is the
destination, and T s are the transitive nodes.

mSD SD
T1 + mT2
mSD SD
T1 ⊕ mT2 = (3)
2

1
eSD
T1 ⊕ eSD
T2 = ((eSD )2 + (eSD 2
T2 ) ) (4)
22 T1
212 S. Uslu et al.

4 Trustworthy Acceptance Metric and Its Methodology

This section describes the methodology of our trustworthy acceptance metric.
We assume that an AI based system generates sets of solutions that a group
of experts will evaluate. Each expert has its own set of solutions to evaluate.
Part of each set of solution is a reference, optimal solution and some sub-optimal
solutions, based on the trade-offs and the criteria applied by the system. Experts
might chose the optimal solution or another sub-optimal solutions, based on their
expertise, as shown in Fig. 1.

Fig. 1. Experts evaluate and select solutions from the set generated by the AI system.

Acceptance is based on and inversely related to the distance between the

proposals of two parties. Distance is measured using the Euclidean distance
formula, where each parameter of a proposed solution becomes a dimension. In
other words, solutions can be considered as vectors when measuring the distance.
After normalizing the each dimension, we also normalize the final distance where
the maximum distance becomes 1. Also, distances are always non-negative. An
example of the distance between n-dimensional solutions S and T , represented
as dTS , is shown in Eq. 5 where Si and Ti are the values of the dimension i of
each solution.  n
i=1 (Si − Ti )
2
dTS = (5)
n
Acceptance is measured as shown in Eq. 6 where Ae is the acceptance of the
expert e, S is the solution selected by the expert, and T is the reference point,
optimal solution. In this case, acceptance is also in the same range of distance
which is [0,1]. Although the distance is bidirectionally the same, we usually stress
the acceptance of the system evaluated by the expert. Therefore, if a system fails
to provide appropriate solution alternatives close enough to the reference point,
it is appropriate for experts to reject the system or rate it with lower acceptance.

Ae = 1 − dTS (6)
 Trustworthy Acceptance: A New Metric for Trustworthy 213

We consider each acceptance as a trust assessment. Since it is possible to

evaluate the acceptance of a system by multiple experts, we use our trust sys-
tem, see Sect. 3, to aggregate the individual acceptances of experts weighted by
their trust as shown in Eq. 7 where k is the number of experts. Similarly, we
find the confidence of the acceptance measurements by calculating the popula-
tion standard error of the mean as shown in Eqs. 8 and 9. We used the regular
standard deviation formula instead of the weighted one for simplicity.

Ae Te
T wA = e (7)
k

e (T wA − Ae )
2
SET wA = (8)
n

cT wA = 1 − 2(SET wA ) (9)
With these two values, the weighted mean acceptance and the confidence, we
have our trustworthy acceptance metric, (T wA , cT wA ). It is also possible to have
multiple sample measurements to evaluate a system’s acceptance by the same
group of experts. In this case, again we use our trust system for aggregation
and calculate the aggregated mean acceptance of the system as shown in Eq. 10
where n represents the sample size. Furthermore, we update the standard error
of the mean using Eq. 11 which is a generalized version of Eq. 4 that we use in
our trust measurements explained in detail in [5,39].

T wA
SystemT wA = n (10)
n

1 
SESystemT wA = (SET wA )2 (11)
n2 n

5 Simulated Measurement Results

To illustrate the evaluation of the proposed acceptance metric, we used the data
set that consists of environmental solutions for over 200 fields in a region as
explained in [47]. The genetic algorithmic system [7], generates near-optimal
solutions while having constraints on the resource usage. The tests were run
using three systems that generate solutions. System 1 is the default system where
there is no constraint on the number of solutions generated for a field. In System
2 and 3, we reduced the number of available solutions to 7 and 4, respectively.
Five agent experts were assigned to evaluate these systems by selecting the most
appropriate solution for the specific field which could represent a neighborhood
of farmers. Trust of such experts is assumed to be high because it makes sense
that the organization which is testing the system deploy top experts. However,
if needed, their trust can be dynamically adjusted by techniques described in
our previous work [45–47]. Furthermore, a reference solution is also determined
214 S. Uslu et al.

and assumed to be legitimate by the user. In our experiments, for simplicity, we

used only the environmental protection values in the distance function.
An evaluation of a system starts by assigning each expert a field and pre-
senting them the solutions prepared by the system for that specific field. After
we completed the assignment, the experts selected the fittest solution based
on their expertise. For an expert who didn’t select the reference solution, a
nonzero distance is measured between the expert and the reference solution,
which is calculated using Eq. 5. Then, we calculated the individual acceptance
rates using Eq. 6. After having individual acceptance rates, we averaged the
acceptance rates weighted by their trust, as in Eq. 7, and also calculated the
confidence of the acceptance, shown in Eq. 9. We call the pair of trust-weighted
average acceptance and its confidence as our trustworthy acceptance metric. To
increase the confidence in our metric, we repeated this process two more times
for each system. We named the selected fieldsets A, B, and C. Then, we aggre-
gated the trustworthy acceptance measurements for each fieldset using Eq. 10.
After completing the evaluation for selected samples and generating the system’s
trustworthy acceptance, we performed the same tasks for System 2 and 3 and
measured the acceptance of each system.
Figure 2 shows the trustworthy acceptance of fields A, B, and C for Systems
1, 2, and 3. Compared to System 1, the acceptance declined in System 2 because
of limiting the available solutions to 7 which eliminated some solutions closer
to the optimal point. Similarly, a stricter constraint, having only four available
solutions, resulted in even less acceptance of System 3. These results could be
evident for a person who understands and can compare the systems. For exam-
ple, such results could be used by USDA to certify only Systems 1 and 2 but not
System 3. Also, the aggregation of samples’ acceptances could be essential in real

Fig. 2. Trustworthy acceptance of system 1, 2, and 3 measured over the sampled fields
A, B, and C from the whole region is presented.
 Trustworthy Acceptance: A New Metric for Trustworthy 215

scenarios to reduce the bias of uneven sampling. As shown in Fig. 3, we aggre-

gated the results of fields A, B, and C and presented the trustworthy acceptance
of each system measured by our experts. As figures illustrate the acceptance
of each system, Table 1 shows both the acceptance values and their confidence
which together form our trustworthy acceptance metric.

Fig. 3. Measured acceptances for fields A, B, and C are aggregated to find the trust-
worthy acceptance of system 1, 2, and 3.

Table 1. Trustworthy acceptances, which consist of acceptance and its confidence, are
presented for fields A, B, and C and also their aggregation for system 1, 2, and 3.

System 1 Tw Acceptance Confidence

Field A 0.872 0.956
Field B 0.853 0.960
Field C 0.866 0.964
Tw Acceptance S1 0.863 0.977
System 2 Tw Acceptance Confidence
Field A 0.809 0.856
Field B 0.841 0.948
Field C 0.831 0.918
Tw Acceptance S2 0.827 0.942
System 3 Tw Acceptance Confidence
Field A 0.627 0.726
Field B 0.673 0.719
Field C 0.785 0.857
Tw Acceptance S3 0.695 0.861
216 S. Uslu et al.

6 Conclusions
We presented a new trustworthy acceptance metric and its measurement method-
ology for evaluating the approval of an AI system that generate solutions for
environmental decisions for a region. Our metric can be used for standardization
of Trustworthy AI. The human experts assigned for the evaluation selected the
most appropriate solutions presented by each system. We measured the distance
from these solutions to a optimal, reference point and calculated the trustworthy
acceptance of a system using our trust framework. Furthermore, using our trust
framework, we aggregated multiple measurements and provided the confidence
of the acceptance using error propagation methods. Finally, we calculated and
compared the trustworthy acceptance of each system measured by the assigned
experts.
Our trustworthy acceptance metric can be applied to many AI applications
that can use the concept of distance-based acceptance. Our approach to confi-
dence measurement is based on our trust system that considers trust assessments
as measurements. Lastly, the experts’ inclusion of trust helped build a metric
that can differentiate experts and is even more appropriate for scenarios where
there is a possibility to engage different groups of experts.

Acknowledgements. This work was partially supported by the National Science

Foundation under Grant No.1547411 and by the U.S. Department of Agriculture
(USDA) National Institute of Food and Agriculture (NIFA) (Award Number 2017-
67003-26057) via an interagency partnership between USDA-NIFA and the National
Science Foundation (NSF) on the research program Innovations at the Nexus of Food,
Energy and Water Systems.

References
1. Alfantoukh, L., Ruan, Y., Durresi, A.: Trust-based multi-stakeholder decision mak-
ing in water allocation system. In: International Conference on Broadband and
Wireless Computing, Communication and Applications, pp. 314–327. Springer,
Cham (2017)
2. Alfantoukh, L., Ruan, Y., Durresi, A.: Multi-stakeholder consensus decision-
making framework based on trust: a generic framework. In: 2018 IEEE 4th Interna-
tional Conference on Collaboration and Internet Computing (CIC), pp. 472–479.
IEEE (2018)
3. Babbar-Sebens, M., Minsker, B.S.: Interactive genetic algorithm with mixed ini-
tiative interaction for multi-criteria ground water monitoring design. Appl. Soft
Comput. 12(1), 182–195 (2012)
4. Barometer, E.T.: Edelman trust barometer global report. Edelman (2019).
https://www.edelman.com/sites/g/files/aatuss191/files/2019-02/2019 Edelman
Trust Barometer Global Report 2.pdf
5. Berendsen, H.J.: A Student’s Guide to Data and Error Analysis. Cambridge Uni-
versity Press, Cambridge (2011)
6. Burke, B., Cearley, D., Jones, N., Smith, D., Chandrasekaran, A., Lu, C., Panetta,
K.: Gartner top 10 strategic technology trends for 2020-smarter with gartner (2019)
 Trustworthy Acceptance: A New Metric for Trustworthy 217

7. Chipperfield, A., Fleming, P.: The MATLAB genetic algorithm toolbox (1995)
8. Chomphoosang, P., Durresi, A., Durresi, M., Barolli, L.: Trust management of
social networks in health care. In: 2012 15th International Conference on Network-
Based Information Systems, pp. 392–396. IEEE (2012)
9. Chomphoosang, P., Ruan, Y., Durresi, A., Durresi, M., Barolli, L.: Trust manage-
ment of health care information in social networks. In: Complex, Intelligent, and
Software Intensive Systems (CISIS), 2013 Seventh International Conference on, pp.
228–235. IEEE (2013)
10. Chomphoosang, P., Zhang, P., Durresi, A., Barolli, L.: Survey of trust based com-
munications in social networks. In: 2011 14th International Conference on Network-
Based Information Systems, pp. 663–666. IEEE (2011)
11. Dong, Y., Xu, J.: Consensus Building in Group Decision Making. Springer, Singa-
pore (2016)
12. Durresi, A., Durresi, M., Paruchuri, V., Barolli, L.: Trust management in emer-
gency networks. In: 2009 International Conference on Advanced Information Net-
working and Applications, pp. 167–174. IEEE (2009)
13. EC: Ethics guidelines for trustworthy AI (2018). https://ec.europa.eu/digital-
single-market/en/news/ethics-guidelines-trustworthy-ai
14. Gunia, B.C., Brett, J.M., Nandkeolyar, A.K., Kamdar, D.: Paying a price: culture,
trust, and negotiation consequences. J. Appl. Psychol. 96(4), 774 (2011)
15. Gunning, D.: Explainable artificial intelligence (xai). Defense Advanced Research
Projects Agency (DARPA), nd Web 2, 2 (2017)
16. Gursoy, D., Chi, O.H., Lu, L., Nunkoo, R.: Consumers acceptance of artificially
intelligent (AI) device use in service delivery. Int. J. Inf. Manag. 49, 157–169 (2019)
17. Hegselmann, R., Krause, U., et al.: Opinion dynamics and bounded confidence
models, analysis, and simulation. J. Artif. Soc. Soc. Simul. 5(3), 1–2 (2002)
18. Hüffmeier, J., Freund, P.A., Zerres, A., Backhaus, K., Hertel, G.: Being tough
or being nice? a meta-analysis on the impact of hard-and softline strategies in
distributive negotiations. J. Manag. 40(3), 866–892 (2014)
19. Hurlburt, G.: How much to trust artificial intelligence? IT Prof. 19(4), 7–11 (2017)
20. Information technology – Artificial intelligence – Overview of trustworthiness in
artificial intelligence. Standard, International Organization for Standardization,
Geneva, CH (2020)
21. Kambiz, M.: Multi-Stakeholder Decision Making for Complex Problems: A Systems
Thinking Approach with Cases. World Scientific, Singapore (2016)
22. Kaur, D., Uslu, S., Durresi, A.: Trust-based security mechanism for detecting clus-
ters of fake users in social networks. In: Workshops of the International Conference
on Advanced Information Networking and Applications, pp. 641–650. Springer,
Cham (2019)
23. Kaur, D., Uslu, S., Durresi, A.: Requirements for trustworthy artificial intelligence–
a review. In: International Conference on Network-Based Information Systems, pp.
105–115. Springer, Cham (2020)
24. Kaur, D., Uslu, S., Durresi, A., Mohler, G., Carter, J.G.: Trust-based human-
machine collaboration mechanism for predicting crimes. In: International Con-
ference on Advanced Information Networking and Applications, pp. 603–616.
Springer, Cham (2020)
25. Kimmel, M.J., Pruitt, D.G., Magenau, J.M., Konar-Goldband, E., Carnevale, P.J.:
Effects of trust, aspiration, and gender on negotiation tactics. J. Pers. Soc. Psychol.
38(1), 9 (1980)
218 S. Uslu et al.

26. Kocielnik, R., Amershi, S., Bennett, P.N.: Will you accept an imperfect AI? explor-
ing designs for adjusting end-user expectations of AI systems. In: Proceedings of
the 2019 CHI Conference on Human Factors in Computing Systems, pp. 1–14
(2019)
27. Lakkaraju, S., Adebayo, J.: Neurips (2020) tutorial. In: Tutorial: (track2) explain-
ing machine learning predictions: State-of-the-art, challenges, and opportunities
(2020)
28. Mueller, S.T., Hoffman, R.R., Clancey, W., Emrey, A., Klein, G.: Explanation in
human-AI systems: A literature meta-review, synopsis of key ideas and publica-
tions, and bibliography for explainable AI. arXiv preprint arXiv:1902.01876 (2019)
29. Rossi, F.: Building trust in artificial intelligence. J. Int. Aff. 72(1), 127–134 (2018)
30. Ruan, Y., Alfantoukh, L., Durresi, A.: Exploring stock market using Twitter trust
network. In: Advanced Information Networking and Applications (AINA), 2015
IEEE 29th International Conference on, pp. 428–433. IEEE (2015)
31. Ruan, Y., Alfantoukh, L., Fang, A., Durresi, A.: Exploring trust propagation
behaviors in online communities. In: Network-Based Information Systems (NBiS),
2014 17th International Conference on, pp. 361–367. IEEE (2014)
32. Ruan, Y., Durresi, A.: Trust management for social networks. In: Proceedings of the
14th Annual Information Security Symposium, p. 24. CERIAS-Purdue University
(2013)
33. Ruan, Y., Durresi, A.: A survey of trust management systems for online social
communities-trust modeling, trust inference and attacks. Knowl. Based Syst. 106,
150–163 (2016)
34. Ruan, Y., Durresi, A.: A trust management framework for cloud computing plat-
forms. In: Advanced Information Networking and Applications (AINA), 2017 IEEE
31st International Conference on, pp. 1146–1153. IEEE (2017)
35. Ruan, Y., Durresi, A.: A trust management framework for clouds. Comput. Com-
mun. 144, 124–131 (2019)
36. Ruan, Y., Durresi, A., Alfantoukh, L.: Trust management framework for internet
of things. In: Advanced Information Networking and Applications (AINA), 2016
IEEE 30th International Conference on, pp. 1013–1019. IEEE (2016)
37. Ruan, Y., Durresi, A., Alfantoukh, L.: Using Twitter trust network for stock market
analysis. Knowl. Based Syst. 145, 207–218 (2018)
38. Ruan, Y., Durresi, A., Uslu, S.: Trust assessment for internet of things in multi-
access edge computing. In: 2018 IEEE 32nd International Conference on Advanced
Information Networking and Applications (AINA), pp. 1155–1161. IEEE (2018)
39. Ruan, Y., Zhang, P., Alfantoukh, L., Durresi, A.: Measurement theory-based trust
management framework for online social communities. ACM Trans. Internet Tech-
nol. (TOIT) 17(2), 16 (2017)
40. Smith, C.J.: Designing trustworthy AI: A human-machine teaming framework to
guide development. arXiv preprint arXiv:1910.03515 (2019)
41. Smuha, N.A.: The EU approach to ethics guidelines for trustworthy artificial intel-
ligence. CRi-Comput. Law Rev. Int. 20(4), 97–106 (2019)
42. Sutrop, M.: Should we trust artificial intelligence? TRAMES: J. Humanit. Soc. Sci.
23(4), 499–522 (2019)
43. Uslu, S., Kaur, D., Rivera, S.J., Durresi, A., Babbar-Sebens, M.: Decision sup-
port system using trust planning among food-energy-water actors. In: International
Conference on Advanced Information Networking and Applications, pp. 1169–1180.
Springer, Cham (2019)
 Trustworthy Acceptance: A New Metric for Trustworthy 219

44. Uslu, S., Kaur, D., Rivera, S.J., Durresi, A., Babbar-Sebens, M.: Trust-based game-
theoretical decision making for food-energy-water management. In: International
Conference on Broadband and Wireless Computing, Communication and Applica-
tions, pp. 125–136. Springer, Cham (2019)
45. Uslu, S., Kaur, D., Rivera, S.J., Durresi, A., Babbar-Sebens, M.: Trust-based
decision making for food-energy-water actors. In: International Conference on
Advanced Information Networking and Applications, pp. 591–602. Springer, Cham
(2020)
46. Uslu, S., Kaur, D., Rivera, S.J., Durresi, A., Babbar-Sebens, M., Tilt, J.H.: Control
theoretical modeling of trust-based decision making in food-energy-water manage-
ment. In: Conference on Complex, Intelligent, and Software Intensive Systems, pp.
97–107. Springer, Cham (2020)
47. Uslu, S., Kaur, D., Rivera, S.J., Durresi, A., Babbar-Sebens, M., Tilt, J.H.: A trust-
worthy human-machine framework for collective decision making in food-energy-
water management: the role of trust sensitivity. Knowl. Based Syst. 213, 106683
(2021)
48. Uslu, S., Ruan, Y., Durresi, A.: Trust-based decision support system for plan-
ning among food-energy-water actors. In: Conference on Complex, Intelligent, and
Software Intensive Systems, pp. 440–451. Springer, Cham (2018)
49. Varshney, K.R.: Trustworthy machine learning and artificial intelligence. XRDS:
Crossroads, ACM Mag. Stud. 25(3), 26–29 (2019)
50. Wachter, S., Mittelstadt, B., Russell, C.: Counterfactual explanations without
opening the black box: automated decisions and the GDPR. Harv. JL Tech. 31,
841 (2017)
51. Wang, W., Siau, K.: Trusting artificial intelligence in healthcare (2018)
52. Zhang, P., Durresi, A.: Trust management framework for social networks. In: 2012
IEEE International Conference on Communications (ICC), pp. 1042–1047. IEEE
(2012)
53. Zhang, P., Durresi, A., Barolli, L.: Survey of trust management on various net-
works. In: Complex, Intelligent and Software Intensive Systems (CISIS), 2011 Inter-
national Conference on, pp. 219–226. IEEE (2011)
54. Zhang, P., Durresi, A., Ruan, Y., Durresi, M.: Trust based security mechanisms for
social networks. In: Broadband, Wireless Computing, Communication and Appli-
cations (BWCCA), 2012 Seventh International Conference on, pp. 264–270. IEEE
(2012)
55. Zhang, Y., Liao, Q.V., Bellamy, R.K.: Effect of confidence and explanation on
accuracy and trust calibration in AI-assisted decision making. arXiv preprint
arXiv:2001.02114 (2020)
 Classifying Holes, Voids, Negative
Objects and Nothing

Katrina Hooper(B) , Alex Ferworn, and Fatima Hussain

Ryerson University, Toronto, Canada

{katrina.hooper,aferworn}@ryerson.ca, [email protected]

Abstract. In the field of Urban Search and Rescue (USAR), searching

for people trapped in the rubble of a collapsed building largely assumes
the presence of many objects that are not the trapped victims, but may
provide clues as to the location of them. However, the lack of objects can
be just as informative. Unlike positive objects (POs), negative objects
extend into the surface of interest [1], and may contain people. These
negative objects (NOs) are commonly known as “holes” or “voids”.
In this paper we discuss the necessity for creating a lexicon and
classifier for negative objects or holes. We review current efforts to detect
and classify them, and finally begin building a vocabulary to effectively
discuss negative objects. This paper aims to spark a discussion about
negative object research and suggest a starting point for a novel research
area.

Keywords: Negative objects · Holes · Voids · USAR

1 Introduction
1.1 What Is a Hole?
Machine vision research has largely focused on detecting and classifying physical
objects [2]. Though, the lack of objects can be just as important as the presence
of them [3]. An example of this is the assumption in robotic path planning, that
the absence of an object means “safe space”. A similar argument can be made for
the operation of Unmanned Aerial Systems (UASs) where “up” is often assumed
to be safe because “down” often has objects restricting travel, like the ground.
This logic of positive objects as a guide, breaks down when a robot must
move through a space with no objects, i.e. negative objects. As NOs are clearly
involved in questions of traversal and detection, we argue, it would be useful to
have a lexicon describing them and their characteristics.

1.2 Types of Holes

The classic definitions of holes typically only emphasize absence, which can be
misleading. Holes require additional characteristics to be well defined and thus
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 220–232, 2021.
https://doi.org/10.1007/978-3-030-75100-5_20
 Classifying Holes, Voids, Negative Objects and Nothing 221

distinguishable from one another. For example, the hole in a donut is very dif-
ferent from the hole in a cup, both can be filled by a PO, but the size, shape,
surrounding material and functionality differ greatly. We include Table 1, provid-
ing definitions for a wide variety of holes discussed in various research works from
Sect. 2, where the absence of something is a key component of the definition as
interpreted by the researchers. We see that many NOs are identified by function
or by a PO that can occupy the space of a NO. For example, “manhole” defines
a hole that is designed to provide access for a human but uses the absence of a
human as a kind of abstraction based on filled space (space the size of a “man”)
rather than the characteristics of empty space–for which there is no lexicon.

1.3 Importance of Detection and Classification in USAR

In today’s society, the likelihood of a disaster is much higher than in the past due
to the effects of urban development, climate change and other factors making the
world more dangerous than it used to be [4,5]. In addition to natural calamity,
it is known that heavily populated areas are at risk of man-made threats like
industrial accidents [6] and terrorist attacks [7]. These threats make it more
likely that urban structures may collapse and trap people inside. It is the role of
Urban Search and Rescue (USAR) teams to respond to such events in a timely
manner. The purpose of their response is to stabilize the new structures caused
by the collapse of a building and locate, medically stabilize and extricate trapped
people as quickly as possible because any delay in extrication reduces the prob-
ability that these people will survive [8]. One of the characteristics of collapsing
buildings is that new openings are created throughout what will become the new
structure of rubble. Usually, the fastest method of access to the people trapped
inside rubble will be through new “holes”. Therefore, we argue, it is important
for first responders to be able to identify and classify the type of entry ways into
the rubble to determine which holes are more likely to lead to someone trapped.
The complexity of rubble piles is also an accessibility issue as rubble can be
comprised of many chaotically placed (and potentially unstable) materials of
random shape [9]. Being able to classify the structure of completely contained
and semi-contained NOs can provide responders better insight into the rubble
pile’s structural integrity, in addition to helping responders and structural engi-
neers devise a plan on how to access the internal spaces potentially containing
victims.
As the use of robotic traversal and interaction on and within disaster rubble
piles is an ongoing research pursuit to increase human safety [10], hole or NO
classification can help in decision-making with respect to autonomous navigation
tasks and with determining traversability. For USAR robots on the ground,
detecting NOs (especially at long range) is challenging because many sensors
are not in a position to detect the signs of a NO until the robot is very close
to it [3]. Imaging can be distorted by external lighting making this task even
harder [11]. A non-trivial problem is that NOs cannot be detected by sensors,
since sensors sense the presence of things. Therefore the presence of NOs must
be inferred from the lack of sensor data [2].
222 K. Hooper et al.

2 Related Work and Hole Specifications

In this section we will discuss methods researchers have used to identify broad
and specific types of NOs. We concentrate on how authors approached their
definitions of NOs to gain better insight into creating a lexicon for all NO types.

2.1 Potholes and Cracks

Eriksson et al. [12] develops an algorithm to differentiate between road imper-

fections using readings from GPS and accelerometer sensors. The authors aim
to classify potholes, pothole clusters, manholes, railroad crossings as well as
some POs. In this paper the authors define potholes as being areas of missing
pavement from the road surface, significantly sunken manhole covers, as well
as protruding manhole covers and other notable road surface irregularities [12],
shown in Fig. 1. This definition of pothole, a type of NO, holds some contradic-
tions: what is commonly known as a pothole does not include sunken-in manhole
covers and especially not protrusions in the road surface, shown in the centre
and right-most images in Fig. 1. It is within this definition that Eriksson et al.
combine the search for a NO with a PO.

Fig. 1. These are examples of road surface anomalies that generated a pothole detection
in the Pothole Patrol algorithm developed by [12]. The left picture is what Eriksson
et al. call a typical small pothole, the centre picture is a sunken in manhole cover and
the right picture is a bridge expansion joint.

Wang et al., [13], and Mednis et al., [14], propose computationally light-
weight, real-time algorithms for detecting and marking the location of potholes
using GPS and accelerometer sensors from a smartphone, by bettering the work
of Mohan et al. [15]. [14] expanded upon the accelerometer approach from [15], it
maintained the need for a restricted smartphone position and instead introduced
the combination of 4 detection functions based on available accelerometer data
to give an accurate final classification. The 4 functions are determined by if the
reading surpasses a vertical threshold (Z-THRESH ), a derivative threshold (Z-
DIFF ), has a standard deviation value threshold (STDEV-Z ), and when 3 read-
ings of G-force all reach around zero Gs (G-ZERO). [13] proposes an algorithm
 Classifying Holes, Voids, Negative Objects and Nothing 223

to reduce errors in all classifiers from [15], noting that Z-DIFF and STDEV-Z
are time and frequency dependent, and that G-ZERO and Z-THRESH would
be influenced by their respective peak measurements. The authors incorporate a
time reading between when Z-DIFF and STDEV-Z functions returned positive
readouts to ensure they weren’t false positives, and combined the Z-THRESH
and G-ZERO readings to also reduce false positives. In the above papers the
definition of a pothole is based on sensor readings and not explicitly defined.
At most, an example of a pothole is shown in Fig. 2 of the pothole [13] used as
a case study. Comparing this to Eriksson et al., it is clear that at minimum,
Wang et al. does not agree with the definition given in [12] and because of this
non-uniform definition we can question what [15] and [14] were detecting and
classifying.

Fig. 2. This is the case study pothole used in Wang et al. 2015 [13]

Lin et al., [16], uses a non-linear support vector machine (SVM) algorithm on
segmented gray-scale images to identify the “sink piece” of a pothole. The sink
piece is the region along the wall of the pothole where the composition is much
more granular than road top [16]. With the SVM sorting the image segments by
granularity to find if a sink piece is present, the authors can classify if a pothole
is in the image. The authors show the beginning of a clear definition of a pothole
in saying that it must contain a sink piece, though authors do not restrict the
size of a pothole and their algorithm could not classify a pothole that was filled
with muddy water. With the interference of weather conditions, we see that a
sink piece cannot be the only characteristic of a pothole. Since a pothole can
exist in all weather conditions, and further more it is possible that other types
of NOs may also contain a sink piece, such as the NO shown in Fig. 3. Here, we
see the need to refine the definition of a pothole from a visual perspective.
224 K. Hooper et al.

Fig. 3. This shows another type of NO that contains a sink piece but should not be
classified as a pothole. Image taken from The Denver Post [17].

Yu et al. propose a method for autonomous pothole detection using a ground

robot to collect the data. [18] use a strong laser to project a line perpendicular
to the direction of travel along the ground and a camera to capture the image
of the laser line on the target surface. The authors create a template of the
ground (potentially containing a NO) based on the deformation of the laser
beam. The authors then create a distress image based on depths calculated
from the deformation readings. Using parameters of depth distress, horizontal
distress, vertical distress and size, the authors classify potholes, vertical cracks
and transversal cracks, as well as give a rating to the severity of the pothole
after it has been classified. Since their algorithm is a threshold based algorithm,
there is no upper limit as to how large a pothole can be. As for the scale of
severity, if we were to compare this analysis of a pothole to one by [14], we
would expect discrepancies in the classification of a pothole corresponding to a
pothole that is small in size but has high severity by the algorithm in [18]. We
expect these discrepancies as running over a severe pothole will result in greater
accelerometer data and thus surpassing set thresholds and classifying it as a
larger pothole than it is.
Sy et al. use image processing to identify cracks in the pavement in real-time.
The authors define two criteria that cracks must meet, that the crack must be
photometrically darker than its surroundings as it is harder for light to reach its
inner portions. The authors also restrict cracks to being continuous regions in
the image [19]. We notice that among many algorithms using computer vision as
a means of data analysis, that photometric brightness is a widely agreed upon
quality of NOs.

2.2 USAR

Both [20] and [21], implement algorithms, proposed by [11], to analyze image
pairs with data types RGB and depth to determine if the full colour-depth image
contains a hole large enough to support a human, this is known as an access hole.
 Classifying Holes, Voids, Negative Objects and Nothing 225

Their algorithm checks if any part of the image surpasses all predefined thresh-
olds, representing the access hole qualities of image darkness, depth disparity and
aspect ratio. In addition, authors define an access hole as having clearly defined
edges on an RGB image, a noticeable depth change from its surroundings and
being poorly illuminated.
Hu et al. focus on isolating inner-structure voids in images produced by
ground penetrating radar (GPR) data, radargrams, on disaster rubble [22]. The
authors isolate that Lean-To Voids must have a triangular shape with clearly
marked boundaries. The authors explain that their algorithm can also classify V-
Shape and A-Frame Voids because they are similar in shape, since their algorithm
focuses more on the internal structure of the void as opposed to the components
making up the void itself. The authors also explain that voids with thickness’
much smaller than the depth scanned to or is around the same size as the material
surrounding the void becomes essentially undetectable due to noise. An example
of these undetectable voids are Pancake Voids, which are pockets created by
layers of heterogeneous material commonly found in rubble piles.

2.3 Robotic Navigation and Path Planning

[23] and [24] work on a broad idea of NOs, in which authors aim to identify
edges of any area of a ground-level path that is below path level. Both Coughlan
and Murarka use stereo cameras to create a depth disparity map to represent
the surrounding ground space around the robot. Coughlan uses discontinuities
in the disparity map to identify edges of NOs. The authors absorb everything
beyond the edges as a NO and deem the area as unsafe for travel [23], whereas
Murarka uses labeling within the disparity map to first classify all sections of the
surrounding area. The authors label sections based on if the surrounding ground
is level with its current position, if the ground is inclined, if there is a drop-off
expected, if there are areas that are “non-ground”-such as walls- or unknown if
the [24] doesn’t have enough data to classify within the other labels. Unknown
areas could be due to occlusions caused by NOs [24].
[25] presents a reversed outlook on the NO problem, where instead of disre-
garding incomplete information caused by a NO or simply labeling the area as
unsafe as a whole, the authors focus on trying to fill in the missing data. Using
a laser range scanner the authors produce a 3-D map representing the robot’s
surroundings. Using ray tracing to label which areas of the map are considered
to be occluded and those laying between the sensor and the occluded area are
labeled visible. Then, Heckman et al. create a 2-D map in the same shape as
the ground plane of the 3-D map used for marking where potential NOs lie. To
find NOs the authors search each 3-D column and mark if a transition between
visible and occluded occurs, if it is entirely occluded or if it is the end of the
data structure. Finally, based on the type of occlusion, the authors can infer
additional data and determine if there is a NO present. While this algorithm
does not categorize detected NOs, the authors bring to light the behaviour of
NOs from the perspective of unmanned ground vehicles (UGVs).
226 K. Hooper et al.

[2,3,25,26] use laser range sensors to identify both potential NOs and con-
firmed NOs. All authors approach the presence of a NO as being a gap in the
point cloud data mapping the UGV surroundings. The work to decipher the dif-
ference between sparse data, caused by non-uniform point spacing, and missing
data caused by occlusions that meet the criteria shown in [25]. Larson et al.
classify potential NOs based on how the angles of the points surrounding the
point cloud gap (PCG) relate to each other, recognizing that an occlusion will
skew the way the points are located on the ground. [3] uses the angle difference
to attempt to determine the edges of the NO. [26] and [2] focus more on finding
patterns in point density surrounding the PCG as opposed to vertical depth
changes.
[27] uses stereo cameras and computer vision tactics to classify the areas of a
UGV’s surroundings. Hu et al. establish characteristics of photometric appear-
ance, similar to what is determined for access holes from [11], and further the
condition to account for a brightness change based on distance and angle of
image approach. The authors also adjust the theory of occlusion from [25] to
work for stereo cameras. Like Hu et al., [28], uses stereo cameras for detecting
and classifying NOs, though Hergehelegiu et al. use the paired images to create
a 3-D full colour disparity map of the surroundings. Herghelegiu et al. extend
the conditions defined in [27], however the condition of photometric brightness is
used but only as an aid to the assumptions that the NOs will have disparity val-
ues lower than the ground plane or will not produce disparity values at all. [28]
furthers their classification to identifying if the NO has a gentler extension into
the PO of the ground or if it more abruptly cuts into ground thereby occluding
data points and causing data gaps.

3 Definitions

We notice the definition of pothole has many different interpretations and, in

some cases, even includes POs as part of its definition [12]. We also notice that
the definition of a pothole was implied but not explicitly defined in [14–16]. This
can cause contradictions as well as inaccuracies when comparing and extending
research.
Another example of this is shown in the definition of gaps, where it can be
taken as a type of NO [15,25] as well as a term to represent the lack of points in
a point cloud caused by a NO [11]. Since definitions are the backbone of object
recognition, we must work to create terms with clear and unique definitions to
standardize research and make accurate comparisons in algorithm performance.
 Classifying Holes, Voids, Negative Objects and Nothing 227

Table 1. This table contains terms representing sub-classifications of NOs and their
definitions as authors are used in research papers discussed in Sect. 2

Term Definitions Term Definitions

Pothole • [12]: Missing chunks of pavement, Access Hole • [11, 20, 21]:
severely sunk in or protruding manhole – defined through its intended use
covers, other significant road surface – sufficiently large to permit human
anomalies entry
• [13]: Not explicitly defined, example – potential for access into a collapsed
shown in Fig. 2 structure, i.e., its function
– a hole must be deeper in the interior
• [14]: Not explicitly defined, further
than the surrounding terrain
separated into large potholes and small
potholes, neither is explicitly defined – clearly marked edges
– strong change in depth
• [15]: Not explicitly defined
– strong change in pixel darkness
• [16]: Contains a sink piece
– a hole must be large enough to support
• [18]: If both depth, horizontal and
entry by a searcher, such as a human,
vertical distress measures have a large
dog, or robot
value, the distress, which indicates the
deviation is significant in both directions – has three attributes that characterize a
hole and allows us to perform access hole
detection: (i) depth disparity, (ii) hole
size, and (iii) photometric brightness
Pothole • [12]: a series of a minimum of k Manhole • [12]: Manhole covers and other
Cluster pothole detections with high enough equipment in the road that are nearly
confidence in a localized area flush with the road surface. Moderate
• [14]: Not explicitly defined cracking, sinking or bulging
Crack • [19] Drain Pits • [14]: Not explicitly defined
– region that is darker than its
surrounding
– is a continuous region
Transversal • [16]: Not explicitly defined Longitudinal • [16]: Not explicitly defined, sample
Crack • [18]: The vertical distress measure Crack image provided in paper
should have a low value, which indicates
that there is little deviation between any
of the columns for an image
• [19]: Not explicitly defined
Alligator • [16]: Not explicitly defined, sample Gaps • [14]: These objects are located on the
Crack image provided in paper main multi-lane street where the total
• [19]: Not explicitly defined road smoothness is better than average
and the transitions between several road
segments have only minor impact on the
suspension of the vehicles
• [27]: data missing that is not caused by
PO occlusion
• [2]: Not explicitly defined
• [3]: An absence of data
Pancake • [22]: A-Frame/Lean- • [22]: Have well defined boundaries
Void – are the most complicated void type To/V-shape
because of the complex geometry Void
– are often very small and covered by
several layers of heterogeneous rubble
Hazardous • [28] Areas below ground that emerge Man-made • [28]: Areas below ground purposely
negative due to natural or human factors, e.g., negative existing in the environment, e.g., edge of
obstacles holes in the ground due to cracks in the obstacles a railway platform, stairs down
pavement/asphalt, missing sewer cap
228 K. Hooper et al.

4 Framework for Moving Forward

What we hope to do in this paper is to create a framework for further developing
sub-classifications of NOs with clearly defined boundaries between them. Cre-
ating a lexicon for NOs will provide a method to have an informed and widely
understood discussion about NOs. In many fields such as USAR, mining and
autonomous path planning, discussions about NOs can be of great value.
In USAR, NOs house the areas in which trapped victims can be located
and they can also provide the location for potential instabilities. When a NO is
completely contained within collapsed rubble, it leaves the opportunity for a PO
to occupy the space. Lean-To Voids, a sub-classification of NOs, shown in Fig. 4,
is an example of a type of NO contained in a rubble pile. We notice that the
walls of the void precariously balanced against each other and when under the
weight and stress of building rubble and debris, it can cause this void to collapse
in on itself. Having an informed way to speak about the void and classify it
allows responders to quickly assess the stability of the void and how much the
collapse of the void can impact the rest of the rubble pile, which a vital part
of victim rescue and public safety precautions. NOs also provide access into the
rubble to get responders closer to voids that can contain people [11]. Making
the distinction between shallow NOs that rest on the surface of the rubble and
NOs that are large enough to support a responder and extend into the rubble to
provide a means of access to voids contained within the rubble is a crucial part
of the initial analysis of a rubble pile. Having a vocabulary to discuss aspects of
NOs and efficiently and effectively communicate features of targeted NOs may
greatly increase response team productivity.
Essentially, mining is the construction of NOs to safely contain humans and
withstand vibrations caused by heavy machinery, explosives and working within
these NOs to retrieve resources. Similar to USAR, understanding weak points in
the shape of the constructed NOs can help increase miner safety. Furthermore,
having the ability to effectively communicate about mine structure and necessary
updates to modify mine configuration (potentially transitioning to a different
type of NO) would increase the efficiency and safety of mine construction and
maintenance.
Autonomous path planning for robotics extends across all versions of
unmanned vehicles. Frequently in path planning, any ground with a lower depth
than the current path is labeled unsafe for travel. It would be beneficial for NOs
to be classified by features, as some may be traversable. For autonomous driving,
the ability to classify holes in the road (called potholes) and sunken manholes
could allow cities to track and record the locations that require road mainte-
nance. Further classification of potholes could allow for priority rankings based
on size1 , depth, impact on car, and frequency of being driven over, to increase
the efficiency of road repairs to maintain safe roads. NOs can also provide useful
information on choosing paths, instead of focusing on them as objects to avoid,
they can be used as guides towards targeted locations.
1
Size - area on the surface plane the NO rests in.
 Classifying Holes, Voids, Negative Objects and Nothing 229

Fig. 4. Example of the structure of a Lean-To Void. Taken from [22]

In [28], the authors work to develop headgear to aid visually impaired persons
with walking safety. We can suggest an example of a visually impaired person
looking for a cross-walk. On either side of the slope that connects a sidewalk to
a road is a type of NO with a strong depth change, commonly known as a curb.

5 Challenges
NO research is a challenging field as it is frequently based on the absence of data
or the inability to obtain the necessary data. We see in [11,20,21] how obtaining
the data and maintaining appearance constraints poses to be a challenge as
external lighting and image point of view skew the assumption that access holes
are poorly illuminated compared to its surroundings. Furthermore, in order to
accurately determine if the access hole leads into the rubble you would need to
be able to “trace the path” that the NO takes into the rubble and ensure there
are no blockages restricting further access and finally you need to determine a
threshold for how deep into the rubble an access hole must stretch in order to
be considered an access hole by the responders who created the term.
As discussed in Sect. 3, without having clear categories and definitions of
NOs, research centred around specific types of NOs becomes ambiguous. It is
necessary to explicitly define the criteria of the NO that is being studied, espe-
cially when using a direct comparison between different algorithms as a method
of verification. This is shown in [26], where the authors compare their method of
detecting NOs to several other algorithms including [2] and [3]. Zhong et al. pro-
pose an algorithm to detect NOs with respect to a small autonomous SUV while
Sinha et al. work on detecting NOs with respect to a small UGV. NOs detected
by [2], are the subset of NOs that are considered hazardous to the UGV. This
means that some of the NOs detected may be considered NOs in the algorithm
by Zhong et al., however due to the vehicle size difference, many NOs detected
by Sinha et al. would not be detected by Zhong et al. Furthermore, the vehicle’s
means of travel differs as one is a 4-wheeled UGV and the other is a tracked
230 K. Hooper et al.

UGV again affecting what is considered a NO to an algorithm. This stresses the

need for defining criteria for NOs so researchers can accurately evaluate their
algorithms.
Since sensors are limited to detecting POs and are frequently limited by
occlusion to collect further data, we are limited on gaining informative data
about NOs. The idea of combining sensors, such as those that are and are not
limited by occlusion to gain data around the external surrounding of the NO as
well as the internal structure of NOs. This idea offers to give more data about
the surface bounding the shape of the NO.

6 Conclusion
Developing a lexicon for NOs is the first step towards understanding and appro-
priately classifying them. NOs make up all of the space that is not taken up by
POs and in many fields that space is much more valuable. In all variations of
path planning it is labeled safe, unless it impacts traversability and in USAR
it is the location of instabilities in a rubble pile and the space that can enclose
trapped human beings.
Further research should include creating clear definitions for classes of NOs.
Classes could be determined by the types of POs that could fill the NO or by
way of defining new classes unique to NOs. Furthermore, a vocabulary of terms
to allow for effective discussions about NOs can be explored.
This research can be extended into the fields of USAR to classify NOs con-
tained in or on the surface of rubble piles and evaluating stability and weak
points in the POs surrounding the NO. It can also be extended into the field
of robotics for autonomous piloting of UASs and UGVs to further classify sur-
roundings to make better decisions regarding path planning. Finally, it can be
used to better communities by classifying road imperfections and prioritizing
maintenance to ensure safety and reduce the risk of emergency situations.

Glossary
Base - the region of the NO located the deepest inside of the NO
Depth - the average length of the wall creating the NO, measured perpendicular
to the surface the NO resides in
Hole - there exists at dimensionally most a 2-dimensional cross section such
that the NO is entirely surrounded
Negative Object (NO) - An object that can be occupied by a positive object
Negative Obstacle - A NO that cannot be traversed
Sink Piece - The region along the wall of the pothole where the composition is
much more granular than road top
Tunnel - A NO that extends entirely through a positive object (does not have
a base)
Wall - the surface of the NO shared with its bounding positive object
Void - NO completely contained inside an positive object
 Classifying Holes, Voids, Negative Objects and Nothing 231

References
1. Matthies, L. et al.: Obstacle detection for unmanned ground vehicles: a progress
report. In: Robotics Research, pp. 475–486 (1996). https://doi.org/10.1007/978-1-
4471-1021-7 52
2. Sinha, A., Papadakis, P.: Mind the gap: detection and traversability analysis of
terrain gaps using LIDAR for safe robot navigation. Robotica 31(7), 1085–1101
(2013). https://doi.org/10.1017/S0263574713000349. ISSN 02635747
3. Larson, J., Trivedi, M.: Lidar based off-road negative obstacle detection and anal-
ysis. In: IEEE Conference on Intelligent Transportation Systems, Proceedings,
ITSC, pp. 192–197 (2011). https://doi.org/10.1109/ITSC.2011.6083105
4. Anderson, J., Bausch, C.: Climate change and natural disasters: scientific evidence
of a possible relation between recent natural disasters and climate change. In:
Policy Department Economic and Scientific Policy 2 (2006)
5. Banholzer, S., Kossin, J., Donner, S.: The impact of climate change on natural
disasters. In: Reducing Disaster: Early Warning Systems for Cimate Change, pp.
21–49. Springer, Dordrecht (2014)
6. Joseph, G., Kaszniak, M., Long, L.: Lessons after Bhopal: CSB a catalyst for
change. J. Loss Prev. Process Ind. 18(4–6), 537–548 (2005). https://doi.org/10.
1016/j.jlp.2005.07.009. ISSN 09504230
7. Perry, W.L., et al.: Predicting Suicide Attacks: Integrating Spatial, Temporal,
and Social Features of Terrorist Attack Targets. Rand Corporation, Santa Monica
(2013)
8. Edwards, F.L., Steinhäusler, F.: NATO and Terrorism: On Scene: New Challenges
for First Responders and Civil Protection. Springer Science & Business Media,
Berlin (2007)
9. Ferworn, A., et al.: Interacting with a virtual destroyed environment constructed
from real disaster data. In: 12th IEEE International Symposium on Safety, Security
and Rescue Robotics, SSRR 2014 - Symposium Proceedings, pp. 24–27 (2014).
https://doi.org/10.1109/SSRR.2014.7017654
10. Krotkov, E., et al.: The DARPA robotics challenge finals: results and perspectives.
J. Field Rob. 34(2), 229–240 (2017)
11. Kong, C., et al.: Toward the automatic detection of access holes in disaster rubble.
In: IEEE International Symposium on Safety, Security, and Rescue Robotics, SSRR
2013 (2013). https://doi.org/10.1109/SSRR.2013.6719364
12. Eriksson, J., et al.: The pothole patrol : using a mobile sensor network for road
surface monitoring. In: Proceedings of the 6th International Conference on Mobile
Systems, Applications, and Services, pp. 29–39 (2008)
13. Wang, H.W., et al.: A real-time pothole detection approach for intelligent trans-
portation system. Mathematical Problems in Engineering 2015 (2015). https://doi.
org/10.1155/2015/869627. ISSN 15635147
14. Mednis, A., et al.: Real time pothole detection using android smartphones with
accelerometers. In: 2011 International Conference on Distributed Computing in
Sensor Systems and Workshops, DCOSS’11, pp. 1–6 (2011). https://doi.org/10.
1109/DCOSS.2011.5982206
15. Mohan, P., Padmanabhan, V.N., Ramjee, R.: Nericell: rich monitoring of road
and traffic conditions using mobile smartphones. In: Proceedings of the 6th ACM
Conference on Embedded Network Sensor Systems, pp. 323–336 (2008). https://
doi.org/10.1145/1460412.1460450
232 K. Hooper et al.

16. Lin, J., Liu, Y.: Potholes detection based on SVM in the pavement distress image.
In: Proceedings - 9th International Symposium on Distributed Computing and
Applications to Business, Engineering and Science, DCABES 2010, pp. 544–547.
IEEE (2010). https://doi.org/10.1109/DCABES.2010.115. ISBN 9780769541105
17. Parker, R.: Xcel refilling big holes in Mayfair streets, but may need to dig more.
In: The Denver Post (2012)
18. Yu, X., Salari, E.: Pavement pothole detection and severity measurement using
laser imaging. In: 2011 IEEE International Conference on Electro/Information
Technology, pp. 1–5. IEEE (2011). https://doi.org/10.1109/EIT.2011.5978573
19. Sy, N.T., et al.: Detection of defects in road surface by a vision system. In: Proceed-
ings of the Mediterranean Electrotechnical Conference - MELECON, pp. 847–851
(2008). https://doi.org/10.1109/MELCON.2008.4618541
20. Kong, C., et al.: What is a hole? discovering access holes in disaster rubble with
functional and photometric attributes. J. Field Rob. 33(6), 825–836 (2015). http://
ncart.scs.ryerson.ca/wp-content/uploads/2015/10/What-is-a-Hole-Discovering-
Access-Holes-in-Disaster-Rubblewith-Functional-and-Photometric-Attributes.pdf
21. Waismark, B., Ferwom, A., Tran, J.: Enhancing autonomous access hole detection.
In: IHTC 2017 - IEEE Canada International Humanitarian Technology Confer-
ence 2017, pp. 51–54 (2017). https://doi.org/10.1109/IHTC.2017.8058198. ISBN
9781509062645
22. Hu, D., et al.: Detecting, locating, and characterizing voids in disaster rubble for
search and rescue. Adv. Eng. Inf. 42(March) (2019). https://doi.org/10.1016/j.aei.
2019.100974. ISSN 14740346
23. Coughlan, J., Shen, H.: Terrain analysis for blind wheelchair users: computer vision
algorithms for finding curbs and other negative obstacles. In: CVHI (2007)
24. Murarka, A., Kuipers, B.: A stereo vision based mapping algorithm for detecting
inclines, drop-offs, and obstacles for safe local navigation. In: 2009 IEEE/RSJ
International Conference on Intelligent Robots and Systems, IROS 2009, pp. 1646–
1653 (2009). https://doi.org/10.1109/IROS.2009.5354253
25. Heckman, N., et al.: Potential negative obstacle detection by occlusion labeling. In:
IEEE International Conference on Intelligent Robots and Systems, pp. 2168–2173
(2007). https://doi.org/10.1109/IROS.2007.4398970
26. Zhong, Z., et al.: Robust negative obstacle detection in off-road environments using
multiple LiDARs. In: 2020 6th International Conference on Control, Automa-
tion and Robotics, ICCAR 2020, pp. 700–705 (2020). https://doi.org/10.1109/
ICCAR49639.2020.9108058
27. Hu, T., et al.: Negative obstacle detection from image sequences. In: Third Interna-
tional Conference on Digital Image Processing (ICDIP 2011), vol. 8009, p. 80090Y,
July 2011 (2011). https://doi.org/10.1117/12.896288. ISSN 0277786X
28. Herghelegiu, P., Burlacu, A., Caraiman, S.: Negative obstacle detection for wear-
able assistive devices for visually impaired. In: 2017 21st International Conference
on System Theory, Control and Computing, ICSTCC 2017, pp. 564–570 (2017).
https://doi.org/10.1109/ICSTCC.2017.8107095
 Empirical Study of Software Adoption
Process in the Bitcoin Network

Mitsuyoshi Imamura1(B) and Kazumasa Omote1,2

1
University of Tsukuba, Tennodai 1–1–1, Tsukuba 305-8573, Japan
[email protected], [email protected]
2
National Institute of Information and Communications Technology,
4-2-1 Nukui-Kitamachi, Koganei, Tokyo 184-8795, Japan

Abstract. Managing the flood of devices is a modern challenge in dis-

tributed networks as it is becoming increasingly difficult to manage the
software updates needed to protect the devices from malicious attacks.
Bitcoin, which is a blockchain-based system that relies on approximately
10,000 volunteer nodes to manage its ledger, faces similar problems. A
centralized system mitigates this problem by enforcing update policies
through one authority. In contrast, a distributed system with different
update policies for different users could enter a worst-case scenario of
sparse deployment due to forgotten updates or refusal to apply them.
Nevertheless, we find that Bitcoin peer-software updates have been sur-
prisingly successful in terms of evading this scenario. Therefore, we inves-
tigate the modeling of the impact caused by software version history and
the remaining challenges, such as the inability to work well with update
cycles. Consequently, we report that previous vulnerable software ver-
sions are still being used, and the share growth decelerates as the number
of versions increases. Moreover, we clarify that the probability of recov-
ering the adoption rate to previous levels is small, while surviving older
versions increases the vulnerability of the network.

1 Introduction
The research of Satoshi Nakamoto [7] is widely accepted as the starting point
of the academic movement of blockchains. Moreover, Satoshi Nakamoto is also
considered an open source software (OSS) developer, as he first proposed the
concept of Bitcoin and uploaded a prototype version to SOURCEFORGE1 on
January 9, 20092 . Since then, Bitcoin has been developed with contributions
from many anonymous volunteers and, as of August 2020, it has been released
as 20 major versions as Bitcoin Core. In other words, Nakamoto was the initial
developer of the well-known Bitcoin Core, software that has been continuously
maintained for over 10 years and has grown to become one of the most used
programs in the blockchain network. According to information from BitcoinN-
odes3 , approximately 10,000 nodes currently contribute to the maintenance of
1
https://sourceforge.net/.
2
https://www.mail-archive.com/[email protected]/msg10142.html.
3
https://bitnodes.io/.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 233–244, 2021.
https://doi.org/10.1007/978-3-030-75100-5_21
234 M. Imamura and K. Omote

the blockchain network, as of the author’s writing. Although the ledger in a

blockchain network is a critical component that all users must reference in a
trustless network, there is no economic incentive from sharing it to the network.
This suggests that there are still many volunteers in the network, despite the
growing storage costs of maintaining the ledger.
Hence, the interesting point of the Bitcoin network is that its success for
over 10 years and the maintenance of such a widely distributed network with no
economic incentives is being completed by volunteers.
Herein, we study the management entity of volunteers in the network, assum-
ing that the latest released software must be used to protect devices from
attackers and keep networks secure. It becomes increasingly challenging to sup-
port software updates as the number of managed devices in a distributed net-
work increases [8]. Furthermore, the update policy differs from one volunteer to
another in a blockchain network because of the decentralized management struc-
tures, whereas centralized systems have unified update policies. Unlike storage
costs, which are considered at launch, maintenance costs reduce the motivation
of volunteers who have no economic incentive because they continuously check
the latest release status and update the installed software. In this context, it is
unclear whether the release and widespread adoption of software are compatible
in the network.
Therefore, this research focuses on volunteers who maintain the ledger and
investigates the trends in software managed by volunteers in the Bitcoin network.
These trends in software update cycles and adoption rates provide insight into
volunteer management. The contributions of this work are summarized in the
following points:
• Empirical trends are identified by analyzing the software update cycles in
Bitcoin networks.
• Potential challenges caused by the usage of older versions are determined
based on the software adoption rate in the network.
• The software adoption rates in networks are modelled using the Gompertz
curve model.
The remainder of this paper is organized as follows. Section 2 introduces the
evolution of the major software of the Bitcoin network, which is the main subject
of this research. Section 3 presents previous works that report software trends in
the network and case studies related to releases in general software. In Sect. 4,
the adoption rate of software in the network is analyzed, while in Sect. 5, the
issues and impact of the adoption rate in the network are discussed on the basis
of the performed analysis. Finally, Sect. 6 concludes this paper.

2 Background
In this section, we review the evolution of the most major Bitcoin peer software,
Bitcoin Core. Herein, we introduce four generations of software based on its
historical background: 1st generation (from 0.1 to 0.3), 2nd generation (from 0.4
to 0.8), 3rd generation (from 0.9 to 0.15), and 4th generation (from 0.16 to 0.20).
 Empirical Study of Software Adoption Process in the Bitcoin Network 235

2.1 Version History and Implementation

2.1.1 1st Generation (Versions 0.1–0.3)

The first major version of Bitcoin Core, version 0.1, was released by Satoshi
Nakamoto on January 9, 2009. The implementation at that time mainly pro-
vided initial functionality and could support only Windows; Linux has become
available since version 0.2. As a classic countermeasure to the 51% attack on
the blockchain, 11111, 33333, and 68555 block hash values were hard-coded as
a checkpoint in version 0.3. Bitcoin Core versions 0.1 to 0.3.20 were released
before the community management was moved to bitcoin.org. From version
0.3.21 onwards, the community has moved from SOURCEFORGE to GitHub4 ,
which is currently used for version management.

2.1.2 2nd Generation (Versions 0.4–0.8)

Versions 0.4 to 0.8 started branding as Bitcoin-Qt from the Qt-based user inter-
face, which was implemented in version 0.5. Since initial implementation, the
UI/UX has been improved, growing closer to the current prototype. In this
generation, wallet encryption and multi-signature transactions on a preliminary
began being supported. Additionally, as it relates to network interfaces, sup-
port for the Tor hidden service was introduced. A specification change to switch
the database from Berkeley DB to Level DB was also implemented for storing
transactions and block indexes.

2.1.3 3rd Generation (Version 0.9–0.15)

Versions 0.9 to 0.15 started rebranding as Bitcoin Core and implemented
usability enhancements, including the introduction of automatic estimation to
improve hard-coded transaction fees. Additionally, the replace-by-fee feature,
which allows sent transactions to change to new transactions for an additional
fee before storage in the block, was implemented. Performance was improved by
the block file pruning mode, which automatically discards old block data and
transactions instead of keeping the complete block data. However, storage costs
were increased by changing the chain state DB management method used for
unspent transaction (UTXO) tracing from per transaction to per UTXO to speed
up the validation and re-indexing of the blockchain in the initial block download
(IBD). Other implementations included parallel block downloading after block
header synchronization and performance improvements to speed up verification
by changing OpenSSL to libsecp256k1 and retiring mempool memory. Finally,
Segregated Witness (SegWitt) was implemented in the test net.

2.1.4 4th Generation (Version 0.16–0.20)

Since the release of 0.16, SegWit has been introduced to improve scalability [10],
and the network block size limit has been gradually relaxed. Segwit fully supports

4
https://github.com/bitcoin/bitcoin.
236 M. Imamura and K. Omote

more transactions into blocks on the main network packs by removing the sig-
nature part from the transaction ID and signing it in a transaction-independent
area. This change was also associated with the failure to call mining protocols
that do not define SegWit rules, which were implemented to encourage SegWit
compatibility.
Nevertheless, the new features could be removed in some cases because of
operational issues. Based on the BPI61 proposal, reject messages were imple-
mented in version 0.9 to improve interoperability between peers. Reject messages
were sent directly to peers to inform them a block or transaction was rejected
or give feedback on errors in the protocol. However, this feature was removed
because it was mainly used for debugging and resulted in security vulnerabilities
from DoS attacks targeting memory pools, and bandwidth issues.

2.2 Handling Vulnerabilities

While Sect. 2.1 describes the implementation and modification of the main fea-
tures, it should be noted that fixes for known security issues were also introduced.
As the software was being updated, developers encountered not only dependency
library issues but also critical bugs in the software, as detailed in the unique com-
mon vulnerabilities and exposures (CVE) catalog of Bitcoin Core.
Vulnerabilities related to OpenSSL, such as CVE-2014-0076, CVE-2014-0160,
and CVE-2014-0224, were announced in minor updates of version 0.9, whereas
vulnerabilities related to the OS version, such as CVE-2017-8798 and CVE-
2017-1000494, were announced in version 0.19.0.1; these are not directly related
because they are not software-specific vulnerabilities.
Direct vulnerabilities include CVE-2011-4447, CVE-2012-3789, and CVE-
2013-4165. The former reported an issue with a private key leak in a wallet
encryption function, which was fixed in version 0.5. CVE-2012-3789 was related
to denial of service (DoS) attacks, which caused Bitcoin processes to become
unresponsive and was fixed in version 0.6.3. CVE-2013-4165 also reported prob-
lems with the Bloom filter processing code, which was fixed in version 0.8.4.
Further, CVE-2018-17144 was a particularly serious issue that not only
caused DoS attacks on full nodes stored by the ledger but also triggered coin
inflation issues. This issue was announced by the community soon after it was
found and affected versions 0.14.x, 0.15.x, and 0.16.x before versions 0.14.3,
0.15.2, and 0.16.3, respectively.
Besides the issues reported in the CVE, the recently released version 20 also
embedded an experimental implementation called Asmap to counter the large-
scale EREBUS attack, which was reported in previous research [9] as a potential
disruption to the Bitcoin network.

3 Related Work
In this section, we briefly review related works that perform software-based Bit-
coin network analysis and report general software release trends to understand
the empirical trends of software releases in the Bitcoin network.
 Empirical Study of Software Adoption Process in the Bitcoin Network 237

3.1 Network Analysis

Although many studies on Bitcoin network analysis showcase network informa-

tion, such as IP addresses, regions, and autonomous systems (AS), few related
works report software information.
Pappalardo et al. [5] included information on the software version used in
an investigation of the timing when transactions broadcasted to the Bitcoin
network are stored in blocks. Particularly, they reported on the type of software
and number of nodes in the process of analyzing the dynamics of transactions
and blocks in the network. Their observations from May 4 to May 11, 2016
demonstrate that users mostly used Bitcoin Core, Bitcoin Classic, Bitcoin XT,
BTCC, and BitcoinUnlimited. Regarding Bitcoin Core, versions 12 and 8 were
the most current and oldest versions available, respectively.
Park et al. [6] analyzed the differences between the recent Bitcoin network
structure in 2018 and that investigated in previous studies from 2013 to 2016
[5]. Their results highlighted the different characteristics between the observed
network and the various previously used software, taking into account that Bit-
coin Core version 15 had the majority share in the network. This study consid-
ered Bitcoin Core and Bitcoin ABC, BitcoinUnlimited, BUCash, Bitcoin Classic,
Bitcoin XT, and BTCC; however, users used only a few of the abovementioned
networks. The oldest version of Bitcoin Core, version 8, was still in use by a few
users.
Neudecker [4] reported on network characterization based on measurements
of the Bitcoin network between 2015 and 2018 and showed the long-term software
distribution in the Bitcoin network in the long term. The measurement results
suggested the dynamism of the distribution of Bitcoin Core and Bitcoin ABC, as
users preferred the most common software during the observation period from
April, 2016 to April, 2018.
Nevertheless, this study was limited to researching the characteristics of the
Bitcoin network, which is insufficient for demonstrating any correlation between
software release and adoption.

3.2 Software Releases

Bosu et al. [1] collected questionnaires from blockchain-based software developers

to understand the development motivations, challenges, and needs and investi-
gated the differences from other OSS. Their research points regarding the diffi-
culty of post-release updates in non-blockchain software can be easily upgraded
for functionality. Contrarily, blockchain software needs to be backward compat-
ible to allow the validation of previous transactions and anticipates worldwide
nodes for upgrading functionality. In the community level, the dissemination of
software releases to users was found to be difficult.
Dong et al. [2] studied the impact of release timing on user behavior. Par-
ticularly, they analyzed the impact between the initial software release rate, the
update rate at which new versions and patches are released, and the user’s inter-
est to download the OSS. Their results indicated that, although a high initial
238 M. Imamura and K. Omote

release or update speed is a factor in increasing user interest, too high of an

initial release or update speed may have the opposite effect.
The abovementioned studies suggest that promoting software in general can
be challenging and that the blockchain community has low expectations. The
analysis in Sect. 4 explores various points, such as the release timing.

4 Empirical Analysis

As mentioned in Sect. 3, the reality of software management in the network is

unclear because it is only considered in part of the network analysis. To address
this issue, in this section, we focus on the usage of Bitcoin software versions in
the network.

4.1 Data

First, the data used in this study include the containing node information of
the Bitcoin network collected by the Bitcoin Network Monitoring and Research
project, which is run by the DSN research group5 . The data are obtained using
a peering protocol feature and include the operating system version, transaction
arrival speed, geographic location, IP version, and autonomous system number
(ASN) of the nodes on the Bitcoin network. The software version data cover the
period from January 29, 2016 to January 11, 2021; the analysis uses data from
this observation period, unless otherwise noted. Corrupted data or data with
obvious gaps owing to maintenance actions or system failures, such as network
outages and system restarts, were excluded from the analysis. Our results also
face the same potential limitations as those of previous studies, such as the
inability to reach nodes hiding on top of firewalls and network address translation
(NAT), which is a common problem.
Next, the software version to be analyzed in this study was determined by
the identifier defined by the string in the payload of the version message in
the Bitcoin protocol. However, there are users in the network that use software
with modified identifiers, so the name identification process is necessary. The
preprocessing of the analysis conducted in this study is as follows:

• If the software does not have the same name as the first work, it is considered
a unique software version.
• Clustering is performed by the longest match from the top.
• A version that has been modified to a non-released version is considered differ-
ent software, even if the version control identifiers match (e.g., Satoshi:10.0.1).

Of course, this match of identifiers and protocol versions does not guarantee the
assumed software name. Users presumably do not modify the software version
unless there are special circumstances.

5
https://dsn.tm.kit.edu/bitcoin/index.html.
 Empirical Study of Software Adoption Process in the Bitcoin Network 239

Fig. 1. Software share in the Bitcoin network

4.2 Software Distribution in Networks

To investigate the usage of software version updates, it is important to know the

distribution of major software and consider its impact on the network. Therefore,
the position of the Bitcoin Core in the network in terms of usage should be
clarified before software version analysis is performed.
Figure 1 shows the distribution of software used by the nodes in the net-
work during the observation period. The colored areas corresponding to different
software represent the number of nodes, and the black line shows the share of
Bitcoin Core in the network. In addition to Bitcoin Core, we checked the dis-
tribution of Bitcoin Classic, BitCore, BTCC, Bitcoin ABC, BitcoinUnlimited,
and Bitcoin XT, which were introduced in previous research [5,6]. All other soft-
ware are classified as “Other”; Specifically, this category included BitcoinNodes,
dsn.tm.kit.edu [4], and the Falcon Relay Network [3]. A few emerging software
programs, such as btcd(btcwire)6 , have been updated recently, whereas others,
such as bitprim7 , have not been updated for several years.
Confirming the dynamics of the nodes on which the software depends, we
found that the overall number of nodes in the network nearly tripled between
2016 and 2018; in fact, it reached a peak in 2018 and held steady until 2020,
before declining slowly. We also confirmed the Bitcoin Core had the dominant
share over the observation period. Bitcoin ABC, which launched in the Bitcoin
Cash divide in late 2017, affected the share of Bitcoin Core for a short time,
but Bitcoin Core maintained a share of over 80% long-term. Over 90% of users
were using Bitcoin Core, making it the dominant software in the network and
the evident recent trend since 2018.
This implies that Bitcoin Core provides more information about installation,
is easier to install compared to newer software, enables easy first-time access, and
is highly recognized because it is maintained based on Satoshi’s original software.
Moreover, when we consider the overall reduction in the number of nodes, a
software with an existing major share of the market is more advantageous.
6
https://github.com/btcsuite/btcd.
7
https://github.com/bitprim/bitprim-core.
240 M. Imamura and K. Omote

Fig. 2. Trends in Bitcoin Core network share by version

4.3 Version Distribution

We assumed that the behavior of users in the network in terms of software

updating is reflected in Bitcoin Core, as it is the proprietary software of the
network. Therefore, we categorized the nodes that use Bitcoin Core based on
major versions to explore the corresponding changes in the share of each version.
Figure 2 illustrates the post-release share of each software version in the mar-
ket share of Bitcoin Core. Despite the fact that software updates in the network
depend on user behavior, new versions were found to consistently disseminate
in the network after release. This implies that users are more interested in the
implementation of new features and the security of the network than expected.
Additionally, we confirmed that previously released versions do not override the
market share of the latest version, and the network as a whole is not being
divided in user support for the software version.
While this is not the worst case of unregulated and unhealthy software
updates in the network, two concerns remain.
First, the unsupported scope is not at the same level as before because
growth decelerates with the release of new versions. To confirm the peak share
dynamism of each software version, Table 1 lists the maximum share on the net-
work from Satoshi:0.12 to Satoshi:0.19 released during the observation period.
The decreasing trend of maximum values has continued since Satoshi:0.14, which
was released March 8, 2017 and has since gained the largest share of the mar-
ket. The peak adoption percentage of the version of Satoshi:0.19 is below 50%,
which means that the majority of the network does not benefit from the software
update.
The second concern stems from the fact that a few users continue using very
old software versions, some of which are no longer supported. The oldest version
we found was Satoshi:0.6, which was released on March 30, 2012. Approximately
0.2% of users use versions from Satoshi:0.6 to Satoshi:0.9, while Satoshi:0.10 is
now considered the mainstream version.
 Empirical Study of Software Adoption Process in the Bitcoin Network 241

Table 1. Peak adoption rate and release interval in released software versions

Release version Peak adoption rate Peak reach days Released interval NewReleased - Peak
Satoshi:0.12 71.78% (N/A) 175 days 226 days 7 days
Satoshi:0.13 72.38% (+0.59) 188 days 182 days 9 days
Satoshi:0.14 79.71% (+7.34) 174 days 197 days 16 days
Satoshi:0.15 78.55% (−1.16) 157 days 190 days 8 days
Satoshi:0.16 64.26% (−14.29) 217 days 165 days −59 days
Satoshi:0.17 60.56% (−3.70) 256 days 158 days −45 days
Satoshi:0.18 59.74% (−0.82) 262 days 211 days −102 days
Satoshi:0.19 45.39% (−14.35) 295 days 160 days 2 days

4.4 Analysis of Adoption Process Model

Assuming that the decline in software adoption in the network follows the Gom-
pertz curve model, we introduce Eq. (1) using two contradictory factors: k, which
accelerates the progression of adoption as the adoption rate, S, increases; and
bt , which inhibits the progression of adoption over time. It should be noted that
k and b are constants, and 0 < b < 1.
dS
= kSbt (1)
dt
Equation (1) was solved for an adoption rate, S; then, assuming α = eβ ,
β = k/ log b, and C = S(0)/α. Equation (2) was obtained.
t
S(t) = Cαb (2)

In summary, S(t) equals C when t becomes ∞, and C in turn equals the

value of the adoption rate, Smax . Then, the inflection point, ti , at which the
adoption rate starts to decelerate, is computed as:
− ln [− ln α] C
ti = , S(ti ) = (3)
ln b e
Figure 3 shows an estimate of the acceleration constant value, k, and the
decay constant value, b, for versions 12 to 19, for which the peak values during
the observation period are clear. The blue solid, orange solid, and black dashed
lines correspond to the raw values, estimated values, and inflection points, respec-
tively. The x-axis and y-axis of each figure correspond to the number of days
and the adoption rate (%), respectively.
For versions 12 to 16, a trend of immediate adoption after release can be
observed, and the inflection point appears within two weeks to one month. In
contrast to this trend, for versions 17 to 19, the inflection point is delayed 3 to 5
months, as the adoption of the software starts some time after release. Finally,
comparing the decay constant values for each version, we notice that the decay
constant values after version 16 are larger than those before version 15, which
corresponds to the timing of the sharp drop in the market share peak.
242 M. Imamura and K. Omote

Fig. 3. Gompertz model fitting of the growing share of software versions

Fig. 4. Changes in the number of nodes containing vulnerabilities after the release of
a fixed version

5 Discussion

In this section, the security risks in the network are discussed based on the
results of the analysis described in Sect. 4.
Despite the improvements implemented in the latest version, some users in
the network continue to use older versions of Bitcoin Core, which, as described in
 Empirical Study of Software Adoption Process in the Bitcoin Network 243

Sect. 2, include security risks. Here, we focus on CVE-2018-17144, which affected

various versions, to assess software usage. Figure 4 shows the software usage for
versions 14.0 to 16.2, where vulnerability was reported, based on data collected
on September 18, 2018, when CVE-2018-17144 was published.
The results show that 539 (7%) Bitcoin Core users continue to use software
that poses security risks. These users presumably have different user character-
istics because they do not correlate with the overall adoption trend. The present
research classifies users according to their intentions as follows.
If users unintentionally use vulnerable versions, we assume that they are
unaware of the security risk information or have abandoned maintenance actions.
In any case, the cost of volunteering to run the node with no economic incentive
is found to be high. However, the reason for not quitting on running nodes is
unknown and needs to be investigated.
If users intentionally use vulnerable versions, we assume that they run a
honeypot in the network. Because this is a node crash-type vulnerability, it is
useful to set up a honeypot to verify that a malicious user actually exists in the
network.
We expect that these nodes are non-maintenance nodes abandoned for main-
tenance by users because of the large number of nodes and the historical trend
confirming the survival of old software versions.

6 Conclusion
In this research, we analyzed the software update cycles and adoption rates of
Bitcoin, which is a blockchain-based software, to determine the reality of volun-
teer management in a trustless and distributed volunteer-dependent network.
As a result, we found that software update cycles in the Bitcoin network cor-
respond to the release of new software; thus, the software adoption rate increases
steadily. Therefore, the assumption holds that once a version is released, it does
not override the prevalence of previous versions, and the network overall is not
divided in user support for the software version.
However, older versions of software remain in use. Software with vulnera-
bilities reported in the CVE affect not only the individual user but the whole
network, and should be considered a potential security risk. Furthermore, we
should note that the peak of software adoption has been declining every year.
This trend confutes the assumption that the currently established adoption rate
is not overridden. Additionally, by modeling empirical trends in adoption rates,
we found that there are greater time-dependent impediments to adoption com-
pared with past versions. This means if the latest software is released at constant
intervals and no fork is formed, an increase in the number of days to reach the
adoption rate peak consequently contributes to a reduced adoption rate peak.
In future work, we intend to clarify the mechanisms behind the adoption
model through detailed analysis. Although the release and adoption of software
operates better than expected, our research does not identify the factors and
mechanisms behind the observed behavior. Moreover, a simplified model was
244 M. Imamura and K. Omote

deployed in this study; thus, there is room for improvement. Future studies
should focus on improving the fit of a better growth curve model. In this paper,
we optimized the model for each software version, but the model can be further
optimized for all versions present in the network.

Acknowledgements. This work was supported by Grant-in-Aid for Scientific

Research (B) (19H04107).

References
1. Bosu, A., Iqbal, A., Shahriyar, R., Chakraborty, P.: Understanding the motivations,
challenges and needs of blockchain software developers: a survey. Empirical Softw.
Eng. 24(4), 2636–2673 (2019)
2. Dong, J.Q., Wu, W., Zhang, Y.S.: The faster the better? innovation speed and user
interest in open source software. Inf. Manage. 56(5), 669–680 (2019)
3. Gencer, A.E., Basu, S., Eyal, I., Van Renesse, R., Sirer, E.G.: Decentralization in
bitcoin and ethereum networks. In: International Conference on Financial Cryp-
tography and Data Security, pp. 439–457. Springer, Heidelberg (2018)
4. Neudecker, T.: Characterization of the bitcoin peer-to-peer network (2015–
2018) (2019). http://dsn.tm.kit.edu/bitcoin/publications/bitcoin network
characterization.pdf
5. Pappalardo, G., Di Matteo, T., Caldarelli, G., Aste, T.: Blockchain inefficiency in
the bitcoin peers network. EPJ Data Sci. 7(1), 30 (2018)
6. Park, S., Im, S., Seol, Y., Paek, J.: Nodes in the bitcoin network: comparative
measurement study and survey. IEEE Access 7, 57009–57022 (2019)
7. Satoshi, N.: Bitcoin: a peer-to-peer electronic cash system (2008). http://www.
bitcoin.org/bitcoin.pdf
8. Taivalsaari, A., Mikkonen, T.: A roadmap to the programmable world: software
challenges in the IoT era. IEEE Softw. 34(1), 72–80 (2017)
9. Tran, M., Choi, I., Moon, G.J., Vu, A.V., Kang, M.S.: A stealthier partitioning
attack against bitcoin peer-to-peer network. In: IEEE Symposium on Security and
Privacy (S&P) (2020)
10. Xie, J., Yu, F.R., Huang, T., Xie, R., Liu, J., Liu, Y.: A survey on the scalability
of blockchain systems. IEEE Netw. 33(5), 166–173 (2019)
 Investigation and Analysis of Features
in Decentralized Network Management
of Minor Cryptocurrencies

Mitsuyoshi Imamura1(B) and Kazumasa Omote1,2

1
University of Tsukuba, Tennodai 1–1–1, Tsukuba 305-8573, Japan
[email protected], [email protected]
2
National Institute of Information and Communications Technology,
4-2-1 Nukui-Kitamachi, Koganei, Tokyo 184-8795, Japan

Abstract. Blockchain is a trustless interoperability platform for users

and is managed in a decentralized structure. There are no controllers in
this platform, and service continuity depends on the users present in the
network. Surveys of such network have mostly focused on miners who
are authorized to write to the ledger; however, the non-mining nodes
that maintain the ledger and are operated by volunteers with no direct
incentive are also important indicators of stability. Although there are
several projects that are inspired by or forked from Bitcoin, it is not
intuitive to assume that projects not pertaining to Bitcoin are running
in the same manner. Compared with major currencies, networks running
minor currency nodes, despite having the same infrastructure, are more
challenging to focus owing to their unclear operational realities. Thus,
learning from the characteristics of failed projects is as important as
learning from successful ones. Therefore, in this study, we investigate
the unique features in the network by analyzing blockchain-based minor
currencies. We analyze 317 minor cryptocurrency nodes and report that
they contain locality, we also report on the issues of blacklisted nodes.

1 Introduction

Blockchain is a trustless interoperability platform for peer-to-peer (P2P) net-

works, and the infrastructure was used in the original paper published by Satoshi
Nakamoto [15]. One of the major differences from traditional systems is that it
has a decentralized structure with no management, and service continuity is
dependent on the users in the P2P network. Bitcoin has been developed by sev-
eral volunteers, and this aspect is well demonstrated in this technology, which has
been operating for more than 10 years since January 9, 20091 , since its released
to SOURCEFORGE2 . It provides a baseline for not only cryptocurrency but
also for technical discussions in the blockchain.

1
https://www.mail-archive.com/[email protected]/msg10142.html.
2
https://sourceforge.net/.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 245–258, 2021.
https://doi.org/10.1007/978-3-030-75100-5_22
246 M. Imamura and K. Omote

In an observational review focusing on the distributed network aspect of

blockchain, the volunteer nodes that contribute to ledger maintenance in the
network were found to be an interesting indicator. According to a previous
classification study [8], there are three types of activities that take place in
the network: P2P transactions, which constitute the transaction network, min-
ing, which records the transactions in a ledger, and decentralized management,
which maintains the ledger. Transaction networks and mining, which are activ-
ities authorized to write to the ledgers, are the most popular topics to focus
on in view of the attack on privacy and consensus, owing to their strong rela-
tionships with direct incentives. Decentralized management is different from the
other aforementioned activities in the sense that it is a volunteer activity with
no direct incentives. Therefore, the economic cost of storing the ledger can easily
lead to the user leaving the network and is an important indicator of network
stability, including its availability and robustness. The measurement methods
and network status for major cryptocurrencies, such as Bitcoin and Ethereum,
have been reported in previous studies [5,11].
Minor currencies, in addition to major ones, must also be analyzed for under-
standing the conditions during the provisioning of services. It is difficult to under-
stand the trends and instability in the network when a service is abandoned due
to project failure as well as when it moves toward convergence. Further, it has
been noted that Bitcoin and Ethereum bubbles burst in 2017; however, it is chal-
lenging to identify the signs of service termination, even in 2020. Nevertheless, we
assume that minor currencies cannot survive in the same history. Furthermore,
they would not have survived under the same conditions as those of Bitcoin and
Ethereum.
There are more than 2,000 projects inspired by or forked from the Bitcoin
initiative. However, it is not intuitive to assume that minor projects, other than
those that grow into major ones such as Ethereum, continue to run in the same
manner. They receive less attention than major currencies and their operational
realities are not as clear. Thus, we can expect that the networks operating the
nodes of these minor currencies have the same infrastructure and features that
successful projects do not have.
Therefore, in this study, we analyze the distribution of nodes in the networks
of these minor currencies to identify their similarities with the major currencies
and the characteristics of those who remain in the network at the end of the
service. We also investigate the status of nodes through security scans, assuming
that unpopular services potentially contain nodes that have either given up on
maintenance or have been abandoned.
The main contributions of this study can be summarized as follows:

• We identify the characteristics of minor currencies in the network that differ

from those of major currencies.
• We identify blacklisted nodes that are included in the operational nodes.
• We report on the empirical successes and availability risks in decentralized
structures, as nodes continue to operate despite the lack of economic incen-
tives.
 Investigation and Analysis of Features in Decentralized 247

The remainder of this paper is organized as follows. Section 2 presents related

works on dynamism in blockchain networks. Section 3 reports the data used in
this study and the results of the analysis. Finally, Sect. 4 concludes this paper.

2 Related Work
Observational research leads to the discovery of potential challenges in the net-
work, providing important feedback to improve the system and contributing to
modeling the empirical rules that historical changes clear the dynamism behind
the user’s behavior. In this section, we introduce related work on the well-known
Bitcoin and Ethereum, as well as other cryptocurrencies.
Donet et al. [5] reported an early observational study that provided meth-
ods and empirical measurements regarding data collection in a Bitcoin network.
They focused on the network information that was provided as a reference in pre-
vious research on transactions and block propagation [4] and presented detailed
information, such as the size of the P2P network, the geographic distribution of
nodes, network stability, and propagation time. They also reported on the trend
for networks with many nodes located in the United States and China; such
nodes were rare in sparsely populated areas and developing countries and evenly
distributed in Western Europe and the United States. Most previous studies
refer to this trend as a baseline trend.
Feld et al. [6] reported a network-level classification using autonomous system
numbers (ASNs), in addition to reporting a country-level classification using
geolocation and analyzing Bitcoin anonymity using IP addresses. In their work,
the networks of the most deployed nodes included AS7922 (Comcast), AS701
(Verizon), AS4134 (China Telecom), AS5089 (Virgin Media, UK), and AS22773
(Cox Communications Inc. US).
Gencer et al. [7] conducted an analysis of multiple infrastructures covering
not only Bitcoin, which is a major platform, but also Ethereum, which receives
considerable attention. They reported improved provisioning bandwidth com-
pared with that in previous research [2] and clusters of geographical proximity
between the nodes.
Park et al. [14] reported on the evolution of the Bitcoin network struc-
ture. The problem of decentralization in the Bitcoin network was highlighted
in another study [8,12].
As Bitcoin and Ethereum are widely known technologies, there is a significant
amount of historical background and several reported features of these networks.
In contrast, minor currencies have received less attention and only a few studies
have been conducted on this topic.
Padmavathi et al. [13] extended the previous investigation of Bitcoin to
include Litecoin, which was created via forking from the Bitcoin project. They
reported on the characteristics of the Litecoin network, such as the relatively
small number of routable nodes as compared with Bitcoin for alternative coins.
248 M. Imamura and K. Omote

Daniel et al. [3] proposed a topology inference model that relies on a passive
timing analysis of measured block arrival times for Zcash, which is a highly
anonymous cryptocurrency, according to network size and node distribution.
To characterize the network, Cao et al. [1] measured the size, node distri-
bution, and connectivity of Monero, a well-known anonymous cryptocurrency
similar to Zcash.
Kim et al. [10] investigated the distribution of nodes in the network to analyze
network security in Stellar.
These investigated cryptocurrencies are inspired by or forked from Bitcoin,
and they use enhanced versions of anonymity and consensus algorithms. Further,
the community behind these technologies is different from that of the initial one.
However, these are relatively well-known cryptocurrencies, and the features they
implement are too popular to be classified as minor. The focus of this study is
on currencies that have not been discussed in these related works.

3 Analysis

3.1 Data

First, we use the nodes on the network to analyze the network information for
317 cryptocurrencies listed on cryptoID3 and Trittium4 . We excluded from our
analysis 11 currencies that are among the top 200 currencies in terms of market
capitalization according to Coinmarketcap at the time of our observation because
our study targets only minor cryptocurrencies. We also excluded 7 currencies
that consist only of the onion router (Tor) servers. As a result, 301 currencies
are included in our analysis. We obtain the network information every 60 min and
summarize it as a daily statistic. Note that we merge and aggregate currencies
that are duplicated in the reference data site.
Next, we use VirusTotal5 , which is a well-known online scanning engine, to
investigate security risks after obtaining the node information. The risk results
are“Clean,” “Unrated,” “Malware,” “Phishing,” “Malicious,” “Suspicious,” and
“Spam.” Then, we consider results other than “Clean” and “Unrated” to be
security risks unless otherwise stated.
The observation period is from May 1 to August 31, 2020. There are potential
limitations to our results owing to our reliance on website information and devi-
ations from real-time information. Furthermore, it should be noted that because
the target cryptocurrencies are minor, there may be a sudden suspension of
distribution due to a decline in popularity or difficulty in survival. Therefore,
investigating minor cryptocurrencies of this type, unlike major cryptocurrencies,
makes it difficult to collect data on the network.

3
https://chainz.cryptoid.info/.
4
https://chains.trittium.cc/.
5
https://www.virustotal.com/gui/.
 Investigation and Analysis of Features in Decentralized 249

Fig. 1. Node and cumulative distributions of cryptocurrencies

3.2 Node Information

We summarize the basic information regarding the collected nodes of the cryp-
tocurrency to investigate the security risk.
First, in Fig. 1 we show the node and cumulative distributions per currency
obtained during the observation period. The green bars correspond to the num-
ber of unique nodes per currency, and the orange bars denote the number of
nodes overlapping between currencies. The red line denotes the cumulative dis-
tribution of the number of unique nodes. The distribution of the collected nodes
is color-coded according to the area in the order of 1000 or more, 100 or more,
and less than 100.
The number of unique nodes collected during the observation period is
140,385, of which 16,949 nodes (12.07%) are duplicates across currencies. If these
duplicate nodes are excluded, the total number of currencies with unique nodes
is estimated at 300. Unless otherwise noted, we calculate the distribution and
share ratio based on this value.
Second, we focus on the distribution of the number of currencies as well as the
cumulative distribution. We report that there are 33 currencies with more than
1,000 nodes, 140 currencies with more than 100 nodes, and 128 currencies with
less than 100 nodes. The number of unique nodes collected in approximately 90%
of the currencies is less than 1,000, classified roughly in half of the currencies
with a baseline of 100 nodes.
With regard to most of the collected node distributions, the intersection of
the cumulative distribution line and the 50% dotted line is located in the > 1000-
node area. In other words, a few currencies still tend to use IP addresses in the
network relatively well, even though they are minor currencies. In our analysis,
half of the nodes are collected in 33 currencies, and we find that the contribution
of a particular currency to the collected data is significant.
In terms of currency, the maximum number of unique nodes is 5,578 and
the maximum number of duplicate nodes is 1,236. As the number of nodes to
be discovered increases, the number of duplicate nodes also tends to increase;
however, because the identity of the nodes is not identified, it is necessary to
consider duplicate counting using dynamically assigned IP addresses.
250 M. Imamura and K. Omote

Fig. 2. Node distribution according to country

3.3 Regional and Network Areas

Next, we classify the regions where the nodes are built to identify trends in the
regional characteristics of the collected data. Figure 2 shows the number of nodes
distributed according to country. The countries are classified according to the
IP addresses assigned using the whois API6 and IP2Location7 . The colors in
the figure represent the number of nodes; regions with zero nodes are shown in
gray, and regions with only one node are shown in blue. Regions with more than
one node are indicated in red, with the higher number of nodes shown in darker
colors.
Ranking the number of nodes in Fig. 2, there are 164 countries with nodes.
The countries with more than 10,000 nodes, which account for a majority of
nodes in the network, are the United States (20,425 nodes) and Germany (18,821
nodes), which together account for approximately 31.79% of the total nodes in
the considered countries. Further, there are 21 countries with the next largest
number of nodes, in the range of 1,000–10,000 nodes; these countries include
European nations such as Russia (8,585 nodes), the United Kingdom (7,068
nodes), and France (3,373 nodes), and Asian nations such as China (8,051 nodes)
and Thailand (3,144 nodes), and Australia (2,742 nodes). Finally, 50 countries
are in the 100–1,000 node range, 73 countries are in the 1–100 node range, and
only 18 countries have only one node.
Our results are similar to those presented in most recent research on Bitcoin
and Ethereum networks [8,9,12,14]. Most nodes were found to be distributed in
North America and Europe, followed by a strong presence in Asia. One unique
trend that was noted in our study was that there are more nodes in countries such
as Nigeria (985 nodes) and South Africa (892 nodes), which are located in a minor
regions in terms of Bitcoin and Ethereum presence. Overall, the distribution
trend is the same as that of major currencies; however, we find different trends

6
https://whois.whoisxmlapi.com/.
7
https://lite.ip2location.com/.
 Investigation and Analysis of Features in Decentralized 251

Fig. 3. Number of currencies per number of countries

in some regions. We attribute this to the fact that minor currencies have different
user usage trends as compared to that of major currencies.
Moreover, half of the minor currencies were found to be distributed in a few
countries, whereas the major currencies are distributed in different areas. This
implies a locality property that is operationalized in a particular community
when it loses popularity.
Furthermore, the distribution of the number of currencies per number of
countries is shown in Fig. 3 to confirm the regional dispersion trend of the cur-
rencies. The bar is color-coded according to the number of nodes distribution,
as in Fig. 1. The vertical dotted line indicates the position where the number of
currencies is half.
There are two currencies with nodes distributed across 86 countries. In con-
trast, there are three currencies with nodes built in only one country. The largest
number of cryptocurrencies have nodes distributed in four countries. There are
20 currencies in total. In addition, half of the currencies are distributed in less
than 14 countries. Furthermore, there are 34 currencies that have more than 100
nodes and 111 currencies that have less than 100 nodes. Intuitively, we know
that a small number of nodes denotes a limited number of regions to distribute;
accordingly, this result provides a concrete criterion to classify minor currencies.
In addition, to verify the trend at a finer granularity, we confirm the distribu-
tion at the ASN level at which the IP addresses are managed. Figure 4 classifies
the nodes into ASNs by referring to the whoisAPI and IP2Location, similar to
the country classification, and shows the top 15 ASNs, which are approximately
30% of the total network nodes. The bar is color-coded according to the number
of nodes distribution, as in Fig. 1.
In the ranking of ASNs, telecommunication companies such as AS3320
(Deutsche Telekom) and AS4134 (China Telecom) are at the top. Some of the
most popular cloud providers reported in previous research [12], such as AS24940
(Hetzner), AS16276 (OVH), and AS14061 (DigitalOcean), are listed, whereas the
well-known service AS16509 (Amazon) is not the most popular and ranks 26th.
252 M. Imamura and K. Omote

Fig. 4. Top 15 ASNs in the distribution of ASN levels

Fig. 5. Number of currencies per number of ASNs

Then, we focus on the distribution of currencies, and the results identi-

fied that AS20473 (CHOOPA), AS24940 (Hetzner), AS14061 (DigitalOcean),
AS16276 (OVH), and AS51167 (Contabo) have an interesting feature. This fea-
ture is that all these ASNs have less than half of the currencies with more than
1000 nodes, contradictory to the case of the other top ASNs. Additionally, these
networks are provided by cloud providers that supply hosts with relatively low
prices.
This result fits with the intuition that the group of nodes that users eventu-
ally abandon are concentrated in cloud providers that have relatively low prices.
However, we note the contradictory behavior of users; that is, they are not eco-
nomically rational when it comes to managing nodes despite their unpopularity,
while they use cloud providers with relatively low prices.
The number of currencies per number of ASNs, as well as number of countries,
is shown in Fig. 5. The bar is color-coded according to the number of nodes
distribution, as in Fig. 1. The vertical dotted line indicates the position where
the number of currencies is half. This figure shows the distribution at a more
detailed granularity, as compared to the distribution per country shown in Fig. 3.
There is only one currency that was identified with the largest number of
nodes distributed in most ASNs, and this currency is distributed in 511 ASNs.
In contrast, there are three currencies that were identified as being distributed
in only one ASN. Further, it was found that the largest number of currencies is
distributed when the nodes are distributed in four ASNs, with 12 currencies. In
addition, half of the currencies were found to be distributed among lesser than
28 ASNs.
 Investigation and Analysis of Features in Decentralized 253

It is well known that users of nodes in major currencies tend to prefer those
nodes that are offered by major cloud providers, as these providers have low
maintenance costs. We find that the most recent reports [8,12] show that this
tendency is becoming increasing pronounced year by year. Therefore, we assume
that hosting nodes in a region without a maintenance cost advantage has a
different siting tendency for minor currencies than for major currencies, such as
being dependent on the community’s area of operation and the location of the
developer, similar to the analysis at the country level.
The distribution of individual currencies shows that half of the currencies
operate in networks with less than 30 ASNs. In other words, network availability
is extremely low and based on the national distribution results, we can see that
it is operated by a very narrow community.
In summary, we find that minor currencies are characterized by locality.
This is based on our observation that such currencies are managed in regions
and specific networks that are rare and therefore differ from those for major
currencies. Moreover, if we assume the scenario of users abandoning nodes, we
find that the results agree with the intuition that nodes that require a relatively
low maintenance price remain in the network. However, there are inconsistencies
in the rationality in the behavior of users.

3.4 Security Risk Scanning

In this section, we describe our investigation of the security risks of 140,385

nodes discovered during the observation period using the scanning engines via
the VirusTotal API. As a result, 5,203 unique nodes were reported to have
security risks, which is 3.71% of the total nodes that were identified.
Here, we consider the results of the scanning engine to reflect the type of
cryptocurrency, and the nodes are located in the region and network identified
in the previous section.
The nodes where security risks are reported by the scanning engine are cat-
egorized by cryptocurrency labels, and we find that they are distributed to 238
(79.07%) different cryptocurrencies. However, this classification includes nodes
that are duplicated in multiple cryptocurrencies, so we need to consider the
results of unique nodes to avoid overestimation. There are 1,242 nodes with
duplicate runs across currencies in nodes that security risks are reported by
scanning engines. Excluding these nodes, we confirm the cryptocurrency labels
of the nodes that run with unique cryptocurrencies. We find that there are 206
different cryptocurrency labels. This means that 68.44% of all minor cryptocur-
rencies have a security risk warning, excluding duplicate nodes. Although there
is a cryptocurrency bias in the collected nodes, we find that this result does not
reflect the specific cryptocurrency results that we were concerned about. This
conclusion remains the same even when the nodes located in different countries
and networks are investigated.
254 M. Imamura and K. Omote

Table 1. The number of nodes that the scanning engine predicted to have a security
risk

No. Engine Nodes No. Engine Nodes No. Engine Nodes

1 ADMINUSLabs 126 34 EonScope 0 67 Quttera 3
2 AICC 0 35 Feodo Tracker 0 68 Rising 0
(MONITORAPP)
3 AegisLab 25 36 Forcepoint 46 69 SCUMWARE.org 0
WebGuard ThreatSeeker
4 AlienVault 177 37 Fortinet 66 70 Sangfor 2
5 Antiy-AVL 0 38 FraudScore 0 71 SecureBrain 0
6 Armis 0 39 FraudSense 0 72 Segasec 0
7 Artists Against 0 40 G-Data 3 73 Snort IP sample 15
419 list
8 AutoShun 0 41 Google 0 74 Sophos 4
Safebrowsing
9 Avira 4 42 GreenSnow 464 75 Spam404 0
10 BADWARE.INFO 0 43 Hoplite 0 76 Spamhaus 3911
Industries
11 Baidu- 0 44 IPsum 292 77 StopBadware 0
International
12 BitDefender 59 45 K7AntiVirus 0 78 StopForumSpam 0
13 BlockList 663 46 Kaspersky 18 79 Sucuri 0
SiteCheck
14 Blueliv 1 47 Lumu 0 80 Tencent 0
15 Botvrij.eu 0 48 MalBeacon 0 81 ThreatHive 0
16 CINS Army 99 49 MalSilo 0 82 Threatsourcing 277
17 CLEAN MX 27 50 Malc0de 0 83 Trustwave 4
Database
18 CMC Threat 0 51 Malekal 0 84 URLQuery 0
Intelligence
19 CRDF 630 52 Malware 0 85 URLhaus 0
Domain
Blocklist
20 Certego 11 53 Malware 0 86 VX Vault 0
DomainList
21 Cisco Talos IP 96 54 MalwarePatrol 1 87 Virusdie 0
Blacklist External Site
Scan
22 Comodo Site 0 55 Malwarebytes 0 88 Web Security 0
Inspector hpHosts Guard
23 Comodo Valkyrie 1073 56 Malwared 0 89 Yandex 0
Verdict Safebrowsing
24 CyRadar 221 57 Netcraft 4 90 ZCloudsec 0
25 Cyan 0 58 NotMining 0 91 ZDB Zeus 0
26 CyberCrime 0 59 Nucleon 0 92 ZeroCERT 0
27 Cyren 0 60 OpenPhish 0 93 Zerofox 0
28 DNS8 0 61 Opera 0 94 ZeusTracker 0
29 Dr.Web 16 62 PREBYTES 0 95 desenmascara.me 0
30 ESET 5 63 PhishLabs 0 96 malwares.com 0
URL checker
31 ESTsecurity- 0 64 Phishing 1 97 securolytics 0
Threat Database
Inside
32 EmergingThreats 91 65 Phishtank 0 98 zvelo 5
33 Emsisoft 2 66 Quick Heal 32
 Investigation and Analysis of Features in Decentralized 255

All nodes, including the unique and duplicate nodes, are distributed in 114
countries, with the highest number of nodes being in the UK (960 nodes). Fur-
ther, the number of nodes in the United States does not reflect a distributional
bias. We also find that this number is not dependent on the regional location
and the scale of the network space, based on the result of 905 different ASe
distributions.
In summary, the scanning engine results are not weighted by the level of the
number of nodes collected, and we ignore the trends in the node distribution
and analyze them as characteristics of minor cryptocurrencies. In the following
sections, we will elaborate on the results of the scanning engine.
First, regarding the scanning engines, we find that not all engines reported
security risks, and that there exists a trend of engines that are more likely to
report a security risk. Table 1 presents the number of nodes classified as having
a security risk for the 98 different scanning engines.

Fig. 6. Number of scanning engines that reported a security risk per node

A total of 36 different scanning engines have reported security risks, and the
scanning engine with the greatest number of alerts is Spamhaus8 , which classified
3,911 nodes (75.17%). Other scanning engines that were classified as having
several nodes with security risks include Blocklist, Comodo Valkyrie Verdict9 ,
CRDF10 , and GreenSnow11 . Conversely, there are other scanning engines, such as
8
https://www.spamhaus.org/.
9
https://verdict.valkyrie.comodo.com/.
10
https://www.crdf.fr/.
11
https://greensnow.co/.
256 M. Imamura and K. Omote

the Blueliv12 , MalwarePatrol13 , and Phishing Database14 , that respond uniquely

to a small number of nodes.
Next, focusing on the node-level, we find a trend in the node distribution of
the scanning engine that reported security risks. Figure 6 shows the relationship
between the nodes and scanning engines that reported security risks. The top
panel depicts the number of nodes per scanning engine that reported a security
risk, the middle panel depicts the ratio of the number of nodes using the dis-
tribution of nodes per currency, and the bottom panel illustrates the number of
unique patterns of the scanning engines.
Our results show that 3,961 nodes have the highest number of alerts from a
single scanning engine, and a total of 1,242 nodes have reported security risks
from two or more scanning engines. As more scanning engines report security
risks, the number of nodes decreases. The maximum number of scanning engines
in this investigation is 14. However, with the increase in the number of alerts,
the ratio of the currencies with fewer and duplicate nodes also tends to increase.
There are a total of 397 different combinations of scanning engine response
patterns for the nodes. We find that the largest number of patterns represents
a combination of the three scanning engines, with 56 different patterns. We also
assume that as the number of combinations of scanning engines increases, the
nodes tend to have a unique security risk criterion, since the ratio of patterns to
the number of nodes approaches unity.
With regard to user behavior, we assume the establishment of a backing
community based on the trends that the nodes have locality in the country and
network area level presented in Sect. 3.3. In contrast, we assume that the vari-
ability that appears in the security of nodes reflects a microscopic trend for
management to depend on individuals. It also emphasizes a scenario in which
nodes that are not well managed by volunteers after operation remain as frag-
ments of P2P network.

4 Conclusion

We investigated the network of minor cryptocurrency nodes and analyzed their

running status in the network to understand the operational differences between
major and minor currencies. We surveyed 317 minor cryptocurrency nodes and
found that locality is a feature that differentiates them from major currencies.
We also identified features that support the overall trend of node abandonment,
that is, low-cost nodes tend to remain in the network, and over 75% of the
currencies have nodes that are blacklisted.
According to the results of this analysis, we assume that there are only certain
individuals and small communities that have worked toward building the network
of minor currencies in the last stage, and that it is a sham operation involving

12
https://www.blueliv.com/.
13
https://www.malwarepatrol.net/.
14
https://github.com/mitchellkrogza/Phishing.Database.
 Investigation and Analysis of Features in Decentralized 257

abandoned nodes. If the major currencies that are currently in the spotlight also
have the same network tendencies, they will essentially shift to minor currencies.
Furthermore, several nodes run without economic rationality, and we assume
that they are used by malicious users without the node owners being aware of
it. Although the countermeasure is a simple method that involves stopping the
nodes, it is necessary to wait for a natural shutdown in a distributed network
because the blockchain protocol cannot be stopped. This decentralized manage-
ment functions in the network until the service ends, similar to the case of major
currencies. It acts as a powerful feature when the system needs to be started;
however, it becomes a bottleneck if the need to shut down the system arises.
In the future, we must discuss the design of blockchains including the case
of stopping service. We believe that our analysis results could be useful for
developing countermeasures in the future.

Acknowledgements. This work was supported by Grant-in-Aid for Scientific

Research (B) (19H04107).

References
1. Cao, T., Yu, J., Decouchant, J., Luo, X., Verissimo, P.: Exploring the monero peer-
to-peer network. In: Financial Cryptography and Data Security 2020, Sabah, 10-14
Feb 2020. Springer, Cham (2020)
2. Croman, K., Decker, C., Eyal, I., Gencer, A.E., Juels, A., Kosba, A., Miller, A.,
Saxena, P., Shi, E., Sirer, E.G., et al.: On scaling decentralized blockchains. In:
International Conference on Financial Cryptography and Data Security, pp. 106–
125. Springer, Heidelberg (2016)
3. Daniel, E., Rohrer, E., Tschorsch, F.: Map-z: exposing the zcash network in times
of transition. In: 2019 IEEE 44th Conference on Local Computer Networks (LCN),
pp. 84–92. IEEE (2019)
4. Decker, C., Wattenhofer, R.: Information propagation in the bitcoin network. In:
IEEE P2P 2013 Proceedings, pp. 1–10. IEEE (2013)
5. Donet, J.A.D., Pérez-Sola, C., Herrera-Joancomartı́, J.: The bitcoin P2P network.
In: International Conference on Financial Cryptography and Data Security, pp.
87–102. Springer, Heidelberg (2014)
6. Feld, S., Schönfeld, M., Werner, M.: Traversing bitcoin’s P2P network: insights into
the structure of a decentralised currency. Int. J. Comput. Sci. Eng. 13(2), 122–131
(2016)
7. Gencer, A.E., Basu, S., Eyal, I., Van Renesse, R., Sirer, E.G.: Decentralization in
bitcoin and ethereum networks. In: International Conference on Financial Cryp-
tography and Data Security, pp. 439–457. Springer, Heidelberg (2018)
8. Imamura, M., Omote, K.: Difficulty of decentralized structure due to rational user
behavior on blockchain. In: International Conference on Network and System Secu-
rity, pp. 504–519. Springer, Cham (2019)
9. Imtiaz, M.A., Starobinski, D., Trachtenberg, A., Younis, N.: Churn in the bitcoin
network: characterization and impact. In: 2019 IEEE International Conference on
Blockchain and Cryptocurrency (ICBC), pp. 431–439. IEEE (2019)
10. Kim, M., Kwon, Y., Kim, Y.: Is stellar as secure as you think? In: 2019 IEEE
European Symposium on Security and Privacy Workshops (EuroS&PW), pp. 377–
385. IEEE (2019)
258 M. Imamura and K. Omote

11. Kim, S.K., Ma, Z., Murali, S., Mason, J., Miller, A., Bailey, M.: Measuring
ethereum network peers. In: Proceedings of the Internet Measurement Conference
2018, pp. 91–104 (2018)
12. Mariem, S.B., Casas, P., Romiti, M., Donnet, B., Stütz, R., Haslhofer, B.: All that
glitters is not bitcoin–unveiling the centralized nature of the BTC (IP) network. In:
NOMS 2020-2020 IEEE/IFIP Network Operations and Management Symposium,
pp. 1–9. IEEE (2020)
13. Padmavathi, M., Suresh, R.: Secure P2P intelligent network transaction using lite-
coin. Mobile Netw. Appl. 24(2), 318–326 (2019)
14. Park, S., Im, S., Seol, Y., Paek, J.: Nodes in the bitcoin network: comparative
measurement study and survey. IEEE Access 7, 57009–57022 (2019)
15. Satoshi, N.: Bitcoin: a peer-to-peer electronic cash system (2008). http://www.
bitcoin.org/bitcoin.pdf
 DynAgentX Framework: Extending
the Internet Management to the VCPE
in the SD-WAN Context

Maxwell E. Monteiro(B)

Federal Technology Institue of Espirito Santo, 6, 5KM Rd. ES-010, Serra, Brazil
[email protected]

Abstract. The virtual customer premises equipment (VCPE) is a vir-

tual Computing platform, capable of to host simultaneous and dynam-
ically provisioned virtual network functions (VNFs), the VCPE has
become an important element in SD-WAN and Edge Computing.
Notwithstanding the SDWAN points to new management solutions (such
as, gRPC-Network Management Interface, or RestConf), the basic man-
agement tasks, namely monitoring, and simple configuration, still the
same. In this interim of coexistence between traditional WAN and SD-
WAN, the maturity, the past investments, and dominance of Internet
management solutions, suggest that it would be quite interesting for
the operators to extend their legacy Internet management solutions till
VCPE and SD-WAN get their maturity. This paper presents the Dyna-
gentX, a VCPE internal management instrumentation’s architecture,
aiming to adapt the Internet management framework to the SD-WAN,
and edge computing scenario, focusing the VCPE. In this work I show
results that substantiate the proposal feasibility, performance, and secu-
rity.

Keywords: CPE · Management · Internet management framework ·

OSGi · Agent X

1 Introduction
The Software Defined WAN (SD-WAN) [12] has risen as a new paradigm for
Wide Area Networks. The idea of an overlay software defined network upon a
mix of well known WAN technologies was enthusiastically received by network
operators and vendors. Besides the network programmability, the network func-
tion virtualization (NFV) has become an important SD-WAN, and the edge com-
puting cornerstone [15]. SD-WAN has coined what has been called the Virtual
Customer Premises Equipment (vCPE) [10]. The VCPE is a virtual Computing
platform capable of to host simultaneous and dynamically provisioned virtual
network functions (nfvs). This innovative Wide Area Network interpretation
has imposed some conflicts with the legacy WAN technologies, among which

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 259–271, 2021.
https://doi.org/10.1007/978-3-030-75100-5_23
260 M. E. Monteiro

the most notorious is the Internet [2] management framework. Notwithstand-

ing the SD-WAN points to new management solutions (such as, gRPC-Network
Management Interface, or RestConf), the basic management tasks, namely mon-
itoring, and simple configuration, still the same. In this interim of coexistence
between traditional WAN and SD-WAN, the maturity, the past investments, and
the Internet management solutions dominance, suggest that it would be quite
interesting for the operators [10] to extend their legacy Internet management
solutions till VCPE and SD-WAN get their maturity, in terms of standards. This
is especially true for those operators that are migrating from traditional WAN
to SD-WAN technology. Unfortunately, due to the clear difference between to
manage a single function physical boxes (CPEs) and a dynamically provisioned
multi-functional virtual CPE (vCPE), the Internet framework in its basic oper-
ation mode seems to be an inadequate technology. The challenge resides on the
dynamism of the Virtual network functions deployed inside the VCPE. This
dynamic composition involves a variety of processing technologies, such as con-
tainers, virtual machines, and traditional network services, all of them enchained
to fulfill the costumer services contract. Furthermore, each virtual network func-
tion may be provided by different vendors. Besides, some services providers
have a mix of physical and virtual CPEs. The Network Management System
(NMS)has to treat the CPEs and VCPE as units, although the VCPE could be
conformed by a bunch of Virtual network Functions. For the legancy NMS is
extreamely important a certain abstraction from the internal VCPE structure,
in order to treat it as a unit.
Motivated by the above mentioned challenge and the advantages of extending
the Internet management framework life-cycle, while the SD-WAN and VCPE
come to maturity, we have investigated how to adapt the Internet management
framework to the VCPE scenarios.
As the result of the aforementioned investigation, this paper presents the
DynagentX an VCPE internal management’s instrumentantion architecture,
aiming to adapt the Internet management framework to the SD-WAN VCPE
scenario. Capturing pros and cons of Different VCPES internal structures and
virtualization technologies, we have figured out an approach based on the con-
servation, as much as possible, of the traditional Internet Network management
ecosystem. Which implies the use of the well-known and mature protocols and
technologies. DynagentX uses the standard AgentX architecture and function-
alities [5], along with the modern dynamically provisioned, and service oriented
OSGi technology. This combination adequates the Internet framework for the
VCPE management task in the SD-WAN context. This work does not advocate
that Internet management is the best solution for the VCPE management, but
an alternative to adapt the legacy Internet NMS to the facing the transition
from traditional WAN to SD-WAN. The comparison among SNMP, gRPCNMI,
and Restconf for the VCPE is out of the paper’s scope.
Concerning the paper organization, Sect. 2 presents our contribution claims,
Sect. 3 shows the related works, while Sect. 4 describe the fundamental con-
cepts and technologies involved in DynAgentX. Finally the Sect. 5 presents and
 DynAgentX Framework 261

describes the DynAgentX architecture and implementation. The Validation and

conclusion are respectively in Sects. 6 and 7.

2 Our Contributions

In our investigation and development aiming the adaptation of the Internet

Management solutions toward the new SD-WAN paradigm, we believe have
made contributions in three directions: (a) the vision of the most adequate
Internet Management standards for the Virtual Custumer Premise
Equipment. Having in mind the VCPE management challenge described in
the Introduction, we have investigated among the SNMP standard documents
those with the potential to face it. We have identified the AgentX (RFC2741)
[5] as the most relevant for the VCPE case, though its well-known implementa-
tion presents some practical issues for the VCPE scenario; (b) the design of
an adaptable and dynamic VCPE’s internal management instrumen-
tation architecture. Considering the practical limitations of the AGentX’s
well-known implementation the net-snmp [4] for the VCPE case, we had the
intuition that a technology like OSGi [1] would help to minimize them. Thus
we have designed the VCPE’s internal management instrumentation and com-
munication architecture, giving it the management flexibility and adaptability
needed; (c) the design and development of a OSGi-JAVA API, aiming
to reach the behavior demanded in contribution (b). Since there wasn’t
a well-known implementation of the agentX using OSGi technology.

3 Related Works

Once our contribution claims were divided into three directions, the related
works were organized in the same way:

– 1. SD-WAN and VCPE management proposals: most of the works

involving Software Defined WAN management are focused on the network
survivability [17], or on the distributed Control Plan performance [8]. To the
extent of our knowledge we couldn’t find works addressing the VCPE internal
management architecture, and technology.
– 2. Agent X, and its well-know implementations: As we have identified
the Agent X as the most relevant SNMP standard for the VCPE management.
Here we cite its well-known implementation (the net-snmp) [3]. The Agent X
description is provided in the Background section, as well as the motives for
which we consider net-snmp basic operation mode inadequate for the VCPE
case. Even alternative Agent X implementation such as PyAgentX, SNMP4J
and AgentC++ present similar limitations, since they either have limitations
to identify master and sub-agents process at run-time.
– 3. Implementations of Internet Management using OSGi: Finally as
we have identified the OSGi Technology as the dynamism provider for our
262 M. E. Monteiro

Agent X implementation. We have investigated OSGi SNMP agents imple-

mentations. The OSGI SNMP Agent implementation is not a new idea, as
can be seen in [14]. However, the main contribution of Merino’s et al. was a
SNMP agent aiming the OSGI platform management, not a network devices
or functions. Although the Merino’s agent internal architecture seems close
to the Agent X proposal, the work didn’t mention or implement the rfc2741.
This work distinguishes itself from those above by proposing a VCPE man-
agement instrumentation architecture, which allows a variety of VCPEs config-
uration behaving as a single device.

4 Background
4.1 The Software-Defined WAN (SD-WAN) and the vCPE
The Software defined Wide Area Network (SD-WAN), shown in Fig. 1, is a set of
technologies organized in order to orchestrate a overlay WAN upon a diversity
of communication technologies [12]. The software defined control allows the con-
struction of a complex, and dynamic network, supposedly increasing availability,
and performance, besides decreasing costs. An important component of the SD-
WAN is the Virtual Customer Primesi equipment (vCPE) [10], illustrated by
the black equipment at the branch office in the Fig. 1. First, it’s important to
emphasize that the VCPE hasn’t a standard or consolidated definition. Some
authors claim that VCPEs are based upon virtual entities, comprising a chain
of network functions (NFVs) [9] allocated in the Cloud or inside a bare metal
(X86 or ARM) box at the edge of the S-D WAN. These two VCPE’s definitions
open room for a third one, the hybrid, joining both.In this paper we consider
the three VCPE’s definitions, admitting that in all cases there is a main com-
puting Container either physical or virtual, wrapping the network functions. In

Fig. 1. S-D WAN and VCPE

 DynAgentX Framework 263

the Fig. 1 the Virtual Customer premise Equipment (VCPE) is shown as a bare
metal (X86 or ARM) box (black equipment) or as a virtual Computer (inside a
Cloud provider), in which Virtual network functions run.

4.2 AgentX
The SNMP extensible Agent Agent X is an IETF initiative, aiming the hability
of a SNMP agent to be composed by a group of dynamically aggregated auxiliary
agents. Thus, The RFC 2741 describes an architecture in which there are two
roles:
(a) Master Agent: The principal entity, which behaves as a single SNMP agent
for the external Network management System (NMS) an as a coordinator
for the internal auxiliary sub-agents.
(b) Sub Agents: They are Auxiliary agents which contribute with the instru-
mentation of part of Management information Base (MIB), but are not
directly accessible from the NMS. A sub agent is accessed via Master Agent,
only when a MIB object under its control is required.
The main Idea behind Agent X is the coordination among SNMP agents forming
a single point of access for multiple MIB instrumentation providers. The Master
Agent is the coordinator of the group, exposing a regular external SNMP access
port. A internal group control port is used by the sub-agents to register and
unregister their contributions to the MIB instrumentation.
Besides the roles, the Agent X standard defines a protocol to rule the rela-
tionship between Master Agent and Sub Agents. Among the Agent X protocol
primitives the most relevants are:
1. Open and Close Session: primitives used to organize the dialog of each
pair Master-SubAgent;
2. register and unregister MIB responsibility: primitives used by the sub
agents to inform the master agent, which part of the MIB they are responsible
for;
3. internal management data exchange (GET, GET BULK, GET
NExt, SET): primitives used to move management data between a sub
agent and the Master agent;
The most well-known implementation of AgentX is the net-snmp framework [4].
Unfortunately In its basic operating mode, net-snmp runs each agent (master
and sub-agents) as a instance of the snmpd deamon. This makes hard to some
one, at run-time, to know which snmpd is the master and which is a specific sub-
agent. This control must be done collecting the PID of each one upon the start
command. this PID collection procedure imposes tailor made scripts or software,
out of net-snmp API scope. Without this capacity of easily start and stop sub-
agents, net-snmp is not adequate to fuflil the requirements of a internal VCPE
management framework. It is important to remind that in a VCPE services and
functions are deployed, started and stoped dynamically, which requires from the
internal management framework the same capacity.
264 M. E. Monteiro

4.3 OSGi and Distributed OSGi

OSGI is a set of specifications, and reference implementations of a middleware
for the dynamically deployable java modules.

Fig. 2. OSGI and D OSGI reference architecture by Richard S. Hall

Figure 2(a) shows the standard OSGI system Stack. The OSGi middleware
runs on top of the JVM, offering a set of services to loc ate, and dynamically
deploy bundles (the term that OSGi uses to name its modules). A bundle is a
self-contained bunch of computing artifacts able to offer and consume services.
Typically a bundle is a .jar artifact. Bundles receive unique identities in the
OSGi middleware’s register service, turning possible to operate each one, indi-
vidually, at run-time. Furthermore, bundles can interact with the middleware,
requesting: the deployment of a new bundle, the stopping, or updating of the
active ones. Since the OSGi specification release 5, bundles upon OSGi middle-
ware platforms in different hosts can access services from each other, as presented
in Fig. 2(b). This feature is called Distributed OSGi (DOSGi) [13].

5 The DynagentX Framework

The DynagentX framework is an AgentX standard adaptation for fulfill the
requirements of the SD-WAN, or edge computing VCPE. The framework starts
from a VCPE internal management instrumentation architecture, and goes
through the OSGi agentX implementation. It is important to highlight the flexi-
bility and adaptability provided by the DynagentX, the main requirement of the
VCPE’s management.

5.1 The VCPE Internal Management Instrumentation and

Communication Architecture
The Fig. 3 shows the DynAgentX proposed architecture for a VCPE’s internal
management instrumentation. Its components are:
 DynAgentX Framework 265

Fig. 3. DynAgentX architecture

– The Main DynAgentX platform: It is the architecture’s foundation piece,

in which the main OSGi platform is instantiated. Upon this main OSGi plat-
form the Snmp AgentX and the auxiliary bundles are deployed.
– The OSGiSnmpMasterAgentX: This is the bundle playing the Master
AngentX role, which exposes the external SNMP (udp 161) interface, and
the internal AgentX control services. On this same main OSGi platform.
– The OSGiSnmpSubAgentX: These are the bundles that consume Mas-
ter’s internal services, and offer internal management data exchange services
(GET, GET BULK, GET NExt, SET). Each of them is bonded with a VNF
(either implemented as a container, or a process).
– The Remote DynagentX: It represents the secondary OSGi platforms
deployed into a separated and self-contained virtual machines or container. All
the OSGi Remote platforms connect to the Main DynAgentX OSGi Platform
via DistributedOSGi (DOSGi) interface. After DOSGi connection, OSGiS-
nmpSubAgentX bundles in remote OSGi platforms can interact with the
OSGiSnmpMasterAgentX as they would do if they were in the Main DynA-
gentX platform.

It’s fundamental to emphasize the dynamism provided by the OSGi platform,

allowing run-time deployment, and shutdown of the OSGiSnmpSubAgentX bun-
dles.

5.2 The OSGiSnmpAgentX Bundle Implementation

Figure 4(a) shows the OSGiSnmpAgentX Class diagram, in which can be noticed
that both Master and Sub-agents derive from the same superclass, which is a
extension of the SNMP4j library [7]. The SNMP4J was used intending to avoid
the insertion of bugs if we had to re-code the SNMP protocol from scratch.
Besides the OSGiSnmAgentX, the DynagentX has two other auxiliary bundles as
shown in Fig. 4(b). The Monitor Bundle: is responsible for to detect VCPE’s
266 M. E. Monteiro

new functionalities, and to load the OGSGiSnmpSubAgenX bundle associ-

ated with each OSGi platform has its own Monitor. The Topology Bundle:
is responsible for to keep track of all VCPE’s remote OSGi platforms. Only
the main OSGi platform has its Topology bundle. The remote OSGi platforms
update their connection status using the main platform topologyBundle’s ser-
vices. The topology bundle keeps a unified view of all OSGi platforms involved
in a VCPE’s instance. It also can be extended to accommodate VCPEs with a
more complex internal architectures.

(a) OSGIAgentX agent bundles (b) OSGiAuxilairy Bundles

Fig. 4. The implemented bundle UML class diagram

The interaction diagram shown in Fig. 5(a) (left side of the figure) shows
how a Network Management System (NMS) accesses a nfv sub-agent through
the Agent Master, while the (b) (right side of the figure) highlights how the
Monitor bundle starts a new OSGiSnmpSubAgentX bundle when a new virtual
network function is detected. As well as also shuts it down when the functionality
is no longer needed. Whenever a new OSGiSnmpSubAgentX is needed, the mon-
itor bundle requests the download from the secure https service-provider’s site.
A table associating nfvs and associated OSGiSnmpSubAgentX Bundle. Which is
regularly updated by the monitor bundle. This mechanism is relevant to DynA-
gentX’s self-adaptable behavior.

5.3 OSGiSnmpAgentX Implementation Notes

Security Note: DynAgentX’s security relies on two main points:
(i) As a SNMP’s derived technology, its security level depends on the utilized
protocol’s version 3.
(ii) Once the second external interaction happens between the Monitor bundle
and the service provider secure site. DynAgentX’s security depends on the
provider site’s security level.
It is capital to mark that, although possible, no other interactions are espec-
ified to the DynAgentX at this moment. So there are no other security risks
mapped to DynAgentX.
 DynAgentX Framework 267

Fig. 5. OSGi bundles interaction diagrams

Third Part Libraries Used: Our implementation has used some third part
open code presented bellow:

(i) the Apache karaf Cellar 4.1.5 running over the Apache Felix OSGi platform
1.18.0. Karaf Cellar provides all DOSGi features required by the DynA-
gentX.
(ii) We have used the SNMP4J2.2.0. within the JVM 1.8, as mentioned before

Developed Sub-agents and Respective MIBS: In order to permit some

validation setups we have developed the following OSGisnmpSubagents:

(a) for DNS, Vnf1 : DNS-Bind9 - MIB : DNS ServerMIB (rfcs 1611, and 1612)
(b) for DHCP, vnf2 : DHCP-dhcp3-serve -MIB : IETF DHCP-ServerMIB (draft)
(c) for HAProxy, vnf3 : HAProxy-1.8.- MIB : the HA Proxy MIB
(OID1.3.6.1.4.1.29385.106.1.0.0)

The OSGiSnmpSubAgentX for those vnfs were based on the respective MIBS.
Once developed, the subagents can be loaded on the Main Platform or on the
remote platform, depending on the VCPE composition and the orchestration
decision. Orchestration is beyond the scope of this paper.

6 Validation Tests and Results

The validation experiments were mainly designed to verify the DynAgentX feasi-
bility, and its performance behaviour. The DynagentX was functionally validated
in Three VCPE composition scenarios, as shown in Table 1. All of them were
setup on the bare metal VCPE, but only scenarios 1 and 2 were setup on the
Cloud VCPE. For the two tested setups we simulated the dynamic deployment
of vnfs described in Sect. 5.
268 M. E. Monteiro

Table 1. Validation scenarios

Scenario VCPE composition

1 Process based VNFs
2 Container based VNFs
3 Virtual machine based VNFs

6.1 The VCPE Experimental Setup

The Bare Metal Setup: for this setup we’ve used a typical X86 box, a i3
processor, 8GRAM, 500G HD. The VCPE’s HOST OS was the Ubuntu server
18.04.1 LTS. The Hypervisor was the qemu-kvm. The Container manager was the
Docker-ce 18.03.1. The NMS station was a notebook i3 4G, With peppermint9
OS, using the net-snmp as SNMP client. A python script was created in order
to lunch snmp PDUs and register the enlepsed time. The NMS station and
the VCPE was connected by a Ethernet switch. The Cloud based VCPE
Experimental Setup: for the VCPE Cloud based setup we’ve used the amazon
EC2 service. Aiming to create a as fair as possible comparison, we’ve used a
m4.large instance (2 vcpus, 8Gi RAM), which costed us 24 dollars cents per 24 h
of tests. The NMS station was the same used in the bare metal VCPE setup.

Fig. 6. OSGi Agent X JVM resource usage

The Performance Results: the performance test was made considering the
5 minuts EC2 monitoring data consolidation. The NMS station was configured
to send SNMP PDUS separated by 1 min intervals, during approximately 2 h.
The graph shown in Fig. 6 (left side) presents the worst case for JVM memory
usage in MBytes, among scenarios 1, 2, and 3 for the both setups. While the
Fig. 6 (right side) presents the worst case for the CPU (bare-metalVCPE), among
scenarios 1, 2, and 3; and VCPU usage (cloud-based VCPE), between scenarios
1 and 2. The response time Test Results Unfortunately We had not space
 DynAgentX Framework 269

to present The response time graph with the collected data. Even though, the
typical response time for Bare Metal and Cloud-based VCPEs (among scenarios
1, 2 and 3) was mean = 25 ms, min = 5 ms, max = 45 ms. In order to take
in account the Internet delay between our lab and the AWS instances, we’ve
measured the RTT from the NMS station to the VCPE Amazon instance, which
was mean = 145 ms, min = 140 ms, max = 156 ms. So the worse response time
for the Cloud-based yagentX Framework (considering the RTT was mean = 170
ms, min = 145 ms, max = 201 ms.
Note: The graphs presented in this section do not consider the memory amount
and CPU percentage used by the VNFs. The measures were made focusing the
resource usage by the DynagentX’s components. The performance tests have con-
firmed that DynagentX is not greedy for resources, being suitable to deployment
over low-end cloud virtual machines or low-end bare-metal boxes. Observing the
Dynagentx response time, it is possible to affirm that, for a few vnfs (three
in our tests) its internal infrastructure does not compromise its responsiveness.
Certainly, Dynagentx must be tested on a more computing intensive scenario,
where vnfs would be heavily demanded, provoking high competition for mem-
ory and cpu. The Dynamic SubAgent Deployment Tests have shown that the
DynagentX is prone to automation and migth be suitable to operate within an
autonomic SD-WAN system.
The Dynamic SubAgent Deploymnent Validation Results. Additionally
to performance tests, we conducted a limited validation to the dynamic deploy
of the following virtual network function: DNS and DHCP. In order to do this
validation, we programmed the Monitor Bundle to find the OSGiSNMPSub-
Agents at a https secure url. We have made a packet trace to verify de subagent
upload, but we did not collect performance measures for these sub-agents trans-
fer. We also did not register the elapsed time till the sub-agent has became fully
operational.

7 Conclusion
This paper presented the Dynagentx a framework aiming to extend
the Internet management technology into the SD-WAN and edge computing
scenarios.
The framework was based on OSGi Technology and its validation results
have proved its adequacy to the VCPE management, either in the X86 bare-
metal boxes, as well as in the cloud-based VCPE. The performance tests have
shown that Dynagentx framework is responsive, even though doesn’t consume
too much computing resources.
The D-OSGi proved to be the right technology to give the flexibility and
adaptability to the SNMP AgentX standard. The presented VCPE internal
management instrumentation and communication architecture is the framework
highlight, allowing different VNFs’ implementation technology scenarios.
270 M. E. Monteiro

As future works, we intend to evaluate the DynAgentX in a more complex

cloud-based VCPE internal-structure. Also we intend to find the minimum X86
bare metal hardware and EC2 instance for each VCPE scenario in Table 1. As
mentioned before, we intend to perform a broader performance test involving
a heavy vnfs demand. The VCPE orchestration and provisioning, not covered
by the DynagentX framework, seems to be the natural extension for this work.
Other interesting future path is to include the NetConf [6] into the presented
management instrumentation and communication architecture. Furthermore, a
massive security test is also on our radar. We think that a vCPE orchestration
service compatible with some Management platfors such as OSM [11] and ONAP
[16] would be helpful.

References
1. Alliance, O.: OSGI-the dynamic module system for Java (2009)
2. Case, J.D., Fedor, M., Schoffstall, M.L., Davin, J.: Simple network management
protocol (SNMP). Technical report (1990)
3. Chavan, S.S., Madanagopal, R.: Generic SNMP proxy agent framework for man-
agement of heterogeneous network elements. In: 2009 First International Commu-
nication Systems and Networks and Workshops. COMSNETS 2009, pp. 1–6 (2009)
4. group at CMU, N.G.: the cmu-snmp suite (2002). http://www.net-snmp.org/
5. Daniele, M., Wijnen, B., Ellison, M., Francisco, D.: Agent extensibility (AgentX)
protocol version 1. Technical report (1999)
6. Enns, R.: NETCONF configuration protocol. Technical report (2006)
7. Fock, F., Katz, J.: SNMP4J-The Object Oriented SNMP API for Java Managers
and Agents, August 2008. http://snmp4j.org/index.html
8. Görkemli, B., Parlakışık, A.M., Civanlar, S., Ulaş, A., Tekalp, A.M.: Dynamic
management of control plane performance in software-defined networks. In: NetSoft
Conference and Workshops (NetSoft), 2016 IEEE, pp. 68–72 (2016)
9. Herrera, J.G., Botero, J.F.: Resource allocation in NFV: a comprehensive survey.
IEEE Trans. Netw. Serv. Manage. 13(3), 518–532 (2016)
10. Kerpez, K., Ginis, G.: Software-defined access network (SDAN), pp. 1–6 (2014)
11. Kourtis, M.A., McGrath, M.J., Gardikis, G., Xilouris, G., Riccobene, V., Papadim-
itriou, P., Trouva, E., Liberati, F., Trubian, M., Batalle, J., et al.: T-NOVA: an
open-source MANO stack for NFV infrastructures. IEEE Trans. Netw. Serv. Man-
age. 14(3), 586–602 (2017)
12. Michel, O., Keller, E.: SDN in wide-area networks: a survey. IEEE Internet Things
J. 37–42 (2017). https://doi.org/10.1109/SDS.2017.7939138
13. Rellermeyer, J.S., Alonso, G., Roscoe, T.: R-OSGI: distributed applications
through software modularization. In: ACM/IFIP/USENIX International Confer-
ence on Distributed Systems Platforms and Open Distributed Processing, pp. 1–20
(2007)
14. Rodero-Merino, L., Vaquero, L.M., Caron, E., Muresan, A., Desprez, F.: Building
safe PaaS clouds: a survey on security in multitenant software platforms. Comput.
Secur. 31(1), 96–108 (2012)
15. Shi, W., Cao, J., Zhang, Q., Li, Y., Xu, L.: Edge computing: vision and challenges.
IEEE Internet Things J. 3(5), 637–646 (2016)
 DynAgentX Framework 271

16. Slim, F., Guillemin, F., Gravey, A., Hadjadj-Aoul, Y.: Towards a dynamic adaptive
placement of virtual network functions under ONAP. In: 2017 IEEE Conference
on Network Function Virtualization and Software Defined Networks (NFV-SDN),
pp. 210–215. IEEE (2017)
17. Vdovin, L., Likin, P., Vilchinskii, A.: Network utilization optimizer for SD-WAN.
In: 2014 First International Science and Technology Conference (Modern Network-
ing Technologies) (MoNeTeC), pp. 1–4 (2014)
 k-Level Contact Tracing Using Mesh
Block-Based Trajectories for Infectious
Diseases

Kiki Adhinugraha1(B) , Wenny Rahayu2 , and David Taniar3

1
Department of Computer Science and Information Technology,
La Trobe University, Melbourne, Australia
[email protected]
2
School of Engineering and Mathematical Sciences, La Trobe University,
Melbourne, Australia
[email protected]
3
Faculty of Information Technology, Monash University, Melbourne, Australia
[email protected]

Abstract. Contact tracing is a method of identifying people who may

have come into contact with an infected person during a virus incuba-
tion time. Standard contact tracing is performed to identify only the
first level of exposures, whereas in reality, exposures can occur across
multiple levels. Multi-level contact tracing can reveal a wider range of
objects which contribute to the spread of the virus. This paper proposes
a k-Level Contact Tracing Query (kL-CTQ) to reveal a broader range
of objects which have possibly been exposed to pathogens. To minimize
manual location tracing while preserving the user’s privacy, we propose
a mesh block sequence (MBS) method where the trajectories are trans-
formed into an MBS before being shared with health authorities. While
our simulation uses an Australian administrative region structure, this
method is applicable in countries which implement similar administrative
hierarchical building blocks.

1 Introduction
Pathogens that cause infectious diseases can spread directly from one person to
another through respiratory droplets. When a pathogen infects a person, the time
it takes for the symptoms to appear is called the incubation time or subclinical
stage. During this incubation time, infected hosts might spread the virus to the
nearby objects while travelling from one place to another [7]. Identifying these
newly exposed objects as a result of being in close contact with an infected
person is referred to as contact tracing [11].
Contact tracing has become critical in understanding how and where the
virus has spread. Several contact tracing strategies have been applied to assist
the government and health authorities to fight the COVID-19 pandemic such as
COVIDsafe [4], PEPP-PT [14], and SafePaths [16]. In Australia, despite the high
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 272–284, 2021.
https://doi.org/10.1007/978-3-030-75100-5_24
 k-Level Contact Tracing Using Mesh Block-Based Trajectories 273

Fig. 1. Pathogen transmission

number of COVID-19 cases, the COVIDsafe app has not provided much assis-
tance to health authorities, and the number of users who utilize this application
is still below the minimum required number [13].
Figure 1 shows how objects can be exposed to pathogens which can then
spread from one host to the broader community. There are two types of objects
which can possibly be exposed which need to be identified during the contact
tracing process. The first type is the spatiotemporal objects, and the second one
is the host objects h, as shown in Fig. 1a. The spatiotemporal objects compro-
mise two inseparable objects, which are location and time window as (p, t). The
spatiotemporal object is the location p where the infected host visited at a cer-
tain time window t. The spatiotemporal objects can be obtained by examining
the hosts’ historical trajectories.
When an infected host visits a certain location p at a specific time t, this
location is considered to be an exposed location. If a group of people visited
this exposed location (p, t), these people will have a high chance of having been
exposed to the virus and will be the newly infected hosts h . The location, time,
and new hosts which have had direct interaction with the infected host h are
considered to be the level-1 exposed objects. Common contact tracing queries
aim to identify the first level objects. A k-level exposure is a condition where
274 K. Adhinugraha et al.

the newly exposed hosts hn infect other spatiotemporal objects and hosts. As
shown in Fig. 1b, the newly exposed hosts hn+1 has one intermediate connection
to the original infected host h. As the level number increases, the number of
newly infected hosts hn and exposed locations increases.
It is important to understand the object’s movement in contact tracing
queries since the virus spreads with the urban movement [3]. In relation to
contact tracing, an infected host is considered to have a high probability of
spreading the virus if this host comes into close contact with other people while
walking. Identifying a close contact event requires two people or trajectories to
be located close to each other while walking, so there is no confusion between the
trajectories of people who are travelling in separate vehicles. A user is consid-
ered to be walking if the velocity of trajectory point V (x) is within the range of
walking velocity, which is around 5–8 km/h. Due to the issues of inaccuracy and
incompleteness in a trajectory and also the privacy issue, trajectory data is not
suitable for contact tracing queries without undertaking a prior preprocessing
step [2].
This paper proposes a k-Level Contact Tracing Query (kL-CTQ) to identify
possible spatiotemporal objects and newly exposed hosts on a certain level of
exposure using the users’ trajectories. The trajectories will be transformed into
a mesh block sequence (MBS) to minimisethe privacy issue [1]. In summary, we
describe the contribution of our works as follows:
1. We propose a k-Level Contact Tracing Query (kL-CTQ) to identify wider
object exposures.
2. We demonstrate the effectiveness of kL-CTQ in identifying wider object expo-
sure than common CTQ using crowdsourced trajectory data.
This paper is organized as follows: Sect. 2 discusses about the related works
and Sect. 3 discusses about the proposed model. The simulation and evaluation
are presented in Sect. 4 and Sect. 5 concludes our work.

2 Related Works
During a pandemic, contact tracing is the most critical step in identifying the
places which have been visited by an infected person and the people with whom
they have come in contact during the incubation period. The COVID-19 virus
can be transmitted to another person located within a 1.5 m–2 m range of
a COVID-19 infected person. Therefore, a proximity-based detection approach
using Bluetooth Low Energy (BLE) from a smartphone is utilized for many
contact tracing methods and apps [15]. Most proximity-based contact tracing
apps do not record the locations that have been visited due to privacy issues
[17].
Another contact tracing approach is to use the person’s historical trajectory
data extracted from their mobile phone. Trajectory data have been widely used
for spatial analysis problems, such as data mining for movement patterns [5,19],
traffic modelling [22] and contact tracing [8].
 k-Level Contact Tracing Using Mesh Block-Based Trajectories 275

One of the main aims of contact tracing is to identify all the places visited
by an infected person during a trip. The geolocation problem has attracted
many researchers who have proposed and developed new methods, especially in
mobility environments [6,21]. Since revealing a person’s geolocation for contact
tracing might raise privacy issues, several methods have been developed to hide
the person’s real identity or location such as FakeMask to generate fake context
[23], KAT to replace real locations with dummy locations [10], PLQP [9] and
Landmark Tree [18].

Fig. 2. Spatial object representation

In the contact tracing procedure, the coordinates of a trajectory point can

be substituted with the name of the area in which the point is located. The
area is considered an exposed spatiotemporal object instead of the coordinates
themselves, as shown in Fig. 2. A minimum bounding rectangle (MBR) is used
to represent a commercial building on a map in this figure. While an MBR
might appropriately represent a commercial place, MBR coverage is limited to
the object itself. Therefore, not all commercial locations will be considered. For
example, if the infected host visited a park, the location will not be recognized.
The gap problem in the MBR approach can be solved using a grid model,
where the idea is to split the area into uniform cells to cover the entire map
[12,20] as shown in Fig. 2b. In a grid model, the entire map is covered without
any visible gaps. However, an important location might be split into several cells.
A mesh block-based approach is superior since this method does not show leave
gaps or overlapping areas, and one crucial object is mapped into exactly one
mesh block.

3 k-Level Contact Tracing

3.1 Preliminaries

Trajectory or GPS traces are a collection of s user’s position x in a time inter-

val or speed interval where T = {x1 , x2 , ..., xn }. A trajectory can be obtained
276 K. Adhinugraha et al.

by utilizing a GPS-enabled device, such as a smartphone, to record movement

and store the information either in local or cloud storage. A trajectory file may
contain numerous trajectory points, where each point has several attributes or
tags to describe the geolocation condition1 . The most common tags are coor-
dinates (latitude, longitude), elevation and time. Some trajectories have a time
zone indicator in the tag such as +11:00 for AEST during daylight savings, while
some trajectories denote the time in UTC format.

Definition 1. Movement mode M (xi ) represents the estimated transporta-

tion mode used at a certain trajectory point xi based on V (xi ). The movement
mode is the walk point when V (xi ) is within the walking speed threshold vw .
The movement mode is vehicle point when V (xi ) exceeds vw .

W alkP oint(W P ), if 0 ≤ V (xi ) ≤ vw
M (xi ) = (1)
V ehicleP oint(V P ), otherwise

Definition 2. An MBS is the set of visited MBs that represent the trajectory
path in chronological order. The time a trajectory point enters an MB is called
entry time te while the time the last trajectory point stays in the same MB is
called exit time tx .

Definition 3. Reduced Mesh Block Sequence (RMBS) is a subset of

M BS where for each MB, the occurrences of W P must be higher than the

walk point number threshold σW P , where W P ≥ σW P .

Fig. 3. Trajectory to RMBS transformation

The process to transform trajectory to RMBS is shown in Fig. 3. The first

step is to identify the velocity point V (x) and movement mode M (x) for each
trajectory point in the trajectory file. Using the grid map layer, the number of
walk points W P and vehicle  points V P are calculated and kept in the MBS
table. The grids that have W P ≥ σW P will be removed from the MBS table,

1
https://www.topografix.com/GPX/1/1/.
 k-Level Contact Tracing Using Mesh Block-Based Trajectories 277

Fig. 4. Example of a mesh block trajectory

Table 1. Example of MBS to RMBS

 
MB te tx WP V P V (p)
 
MB te tx WP V P V (p)
20552100000 . . 7 2 4
20552910000 . . 2 24 28 20552100000 . . 7 2 4
20556522000 . . 3 1 12 20556523000 . . 16 6 3
... . . ... ... ... 20552813000 . . 8 8 17
20552813000 . . 8 8 17 (b) Reduced MBS table σW P = 5
(a) MBS table

and the final result will be the RMBS table. Since an RMBS only contains MBs
with a high number of walking points, RMBS is considered to be a list of exposed
spatiotemporal objects.
We demonstrate the transformation of MBS to RMBS in Table 1. As shown
in the example in Fig. 4, the MBS for trajectory T is the set of MBs in
chronological order as follows: M BS(T ) = {20552100000, 20552910000, 2055652
2000, 20556523000, 20556522000, 20551070000, 20552811000, 20552814000, 20552
813000}. We then populate the statistical data in each MB and present the
MBS in Table 1a. The average speed V (p) for each MB is set to km/h. We dis-
regard the entry time and exit time for the sake of simplicity. As shown in this
table, Mesh Block 20556522000 occurs twice as the user visits the same MB at
different times.

3.2 k-Level Contact Tracing Query

A contact tracing query is a method of obtaining exposed objects from the

infected host by observing the host’s historical trajectories at a particular time.
The k-Level Contact Tracing Query (kL-CTQ) is used to obtain all the exposed
objects at level k from an infected host.
278 K. Adhinugraha et al.

Fig. 5. Contact tracing query in RMBS

In kL-CTQ, there are two objectives that need to be achieved, which are:

1. To identify all exposed spatiotemporal objects.

2. To identify all new hosts and their trajectories.

Fig. 6. 2-level contact tracing

The kL-CTQ is based on MBS where the exposed spatiotemporal objects

are obtained from RMBS since the RMBS table only contains all mesh blocks
with a high number of walking points. The MBS-based CTQ is shown in Fig. 5.
The first step is to identify new hosts based on the infected RMBS. The infected
RMBS is a table that contains all the spatiotemporal objects for the infected
host. Once all newly exposed hosts have been identified, each exposed new host
generates RMBS from each host trajectory. For k-Level CTQ, this process is
repeated k − 1 times to obtain all k-level exposed objects.
Figure 6 shows the 2-level CTQ where the infected host is identified by the
red trajectory h. Let (p, t)1 be the level-1 exposed spatiotemporal object when
the trajectory from host h1 goes through this exposed spatiotemporal object.
This host becomes the level-1 exposed host. When host h1 could expose other
users, the location and time is called the spatiotemporal object level-2 (p, t)2 .
When the green trajectory from host h2 goes through this location and becomes
exposed due to host h1 , host h2 will become the newly exposed host level-2. The
2L-CTQ(h) will obtain (p, t)1 , (p, t)2 as two spatiotemporal objects, and exposed
hosts h1 , h2 .
 k-Level Contact Tracing Using Mesh Block-Based Trajectories 279

4 Experiments and Evaluation

This section evaluates the efficiency of kL-CTQ in identifying the k-level of
exposed objects, which can be spatiotemporal objects or other hosts. We modify
the GPX dataset from planet.gpx2 to ensure all the trajectories share a similar
travelling time to simulate the massive movements in the 850 km2 of the Mel-
bourne metropolitan area. For these experiments, we processed 300+ trajectories
from various contributors. We transformed the trajectories into RMBS for a day
trip. Our methods are implemented in a system running Java with PostGIS as
the database engine3 . For the visualization tool, we use QGIS4 . The system uses
Intel i7-8665U CPU with 16 GB of RAM.

4.1 Level-1 Object Exposure

In this evaluation, as an example, we assume trajectory number 000509760

belongs to an exposed host. As the average walking speed is around 5 Km/h,
we set the walking speed threshold to V (p) = 8 km/h to accommodate a faster
walking speed. Therefore, all the velocity points beyond this limit are considered
a non-walking mode, while all the points below this velocity limit are considered
a walking mode. Common contact tracing queries are focused on obtaining the
exposed level-1 objects.
The example of obtaining the 1st level of exposure from an infected host is
illustrated in Fig. 7. Figure 7a shows the trajectory of the infected host and the
exposed spatial-temporal objects (p, t)1 along the trajectory path indicated by
the red polygon, where the average trajectory speed is within the walking speed
of 8 km/h. Figure 7b shows the table of newly exposed hosts in a mesh block
20400630000, which has two-time windows at 8 AM and 10 AM. During time
windows, eight possible newly exposed hosts visited these exposed spatiotempo-
ral objects. The trajectory routes from these newly exposed hosts are shown in
Fig. 7c.
The number of newly exposed hosts from all the exposed spatiotemporal
objects from this infected host is shown in Fig. 8, where some exposed locations
might have no host exposure at all. There might be no other users or trajectories
at these spatiotemporal objects, or the trajectories moved faster than the walking
speed limit.

2
https://planet.openstreetmap.org/gps/.
3
https://postgis.net.
4
https://www.qgis.org.
280 K. Adhinugraha et al.

Fig. 7. 1st level exposure

4.2 Level-k Exposed Objects

For the level-k object exposure, we limit our evaluation to level-3 object expo-
sure. From the same infected host, the results of our evaluation are shown in
Fig. 9. The exposed locations or MBs are shown in Fig. 9a, where the number
of distinct mesh blocks increases significantly from level 1 to level 2. The same
trend can be seen for the number of spatiotemporal objects shown in Fig. 9b.
While the number of exposed locations and spatiotemporal objects increases sig-
nificantly from level 1 to level 2, the exposed area expansion goes from level 2
to level 3 as shown in Fig. 9c. The exposed coverage area has increased 10-fold,
from only 22 km2 to 229 km2 at level 3 as shown in Fig. 10.
 k-Level Contact Tracing Using Mesh Block-Based Trajectories 281

Fig. 8. Newly exposed hosts for each exposed MB

Fig. 9. k-level exposure

Although the exposed region is vast, the crowdsourced data does not provide
sufficient trajectories for contact tracing evaluation. Therefore, there is no sig-
nificant increase in the number of newly exposed hosts from level 1 to level 3, as
shown in Fig. 9d.
282 K. Adhinugraha et al.

Fig. 10. Exposed locations on level 2

5 Conclusion

In this paper, we proposed k-Level Contact Tracing Query (CTQ) with the aim
of identifying two objects, namely exposed spatiotemporal objects and newly
exposed hosts for a certain level. The value of k can show the relationship
between the exposed objects and the infected host, where k > 1 indicates there
are no direct relationships between the exposed objects and the original infected
host. Our simulation using a crowdsourced dataset in Melbourne shows that kL-
CTQ might play an essential role in identifying the potential spread of pathogens
for contagious diseases and prevent unnecessary lockdown during a pandemic.

References
1. Adhinugraha, K., Rahayu, W., Hara, T., Taniar, D.: Contact Tracing with Mesh
Block-based Trajectories. ACM Trans. Spat. Algorithms Syst. (submitted) (2020)
2. Adhinugraha, K., Rahayu, W., Hara, T., Taniar, D.: Dealing with noise in
crowdsourced GPS human trajectory logging data. In: Concurrency Computation,
September, pp. 1–18 (2020). https://doi.org/10.1002/cpe.6139
3. Alamri, S., Taniar, D., Safar, M.: A taxonomy for moving object queries in spatial
databases. Future Gener. Comput. Syst. 37, 232–242 (2014). https://doi.org/10.
1016/j.future.2014.02.007. http://dx.doi.org/10.1016/j.future.2014.02.007
4. Australia, D.O.H.: COVIDSafe. https://www.health.gov.au/resources/apps-and-
tools/covidsafe-app
5. Goh, J., Taniar, D.: Mining frequency pattern from mobile users. Lecture Notes
in Computer Science (including subseries Lecture Notes in Artificial Intelligence
and Lecture Notes in Bioinformatics), vol. 3215, 795–801 (2004). https://doi.org/
10.1007/978-3-540-30134-9 106
 k-Level Contact Tracing Using Mesh Block-Based Trajectories 283

6. Goh, J.Y., Taniar, D.: Mobile data mining by location dependencies. Lecture Notes
in Computer Science (including subseries Lecture Notes in Artificial Intelligence
and Lecture Notes in Bioinformatics), vol. 3177, 225–231 (2004). https://doi.org/
10.1007/978-3-540-28651-6 33
7. Hellewell, J., Abbott, S., Gimma, A., Bosse, N.I., Jarvis, C.I., Russell, T.W., Mun-
day, J.D., Kucharski, A.J., Edmunds, W.J., Sun, F., Flasche, S., Quilty, B.J.,
Davies, N., Liu, Y., Clifford, S., Klepac, P., Jit, M., Diamond, C., Gibbs, H.,
van Zandvoort, K., Funk, S., Eggo, R.M.: Feasibility of controlling COVID-19 out-
breaks by isolation of cases and contacts. Lancet Global Health 8(4), e488–e496
(2020). https://doi.org/10.1016/S2214-109X(20)30074-7
8. Kim, W., Lee, H., Chung, Y.D.: Safe contact tracing for COVID-19: a method
without privacy breach using functional encryption techniques based-on spatio-
temporal trajectory data, pp. 1–16 (2020). https://doi.org/10.1101/2020.06.29.
20143180
9. Li, X.Y., Jung, T.: Search me if you can: privacy-preserving location query ser-
vice. In: Proceedings - IEEE INFOCOM, pp. 2760–2768 (2013). https://doi.org/
10.1109/INFCOM.2013.6567085
10. Liao, D., Li, H., Sun, G., Anand, V.: Protecting user trajectory in location-based
services. In: 2015 IEEE Global Communications Conference, GLOBECOM 2015
(2015). https://doi.org/10.1109/GLOCOM.2014.7417512
11. Marcel, S., Christian, A.L., Richard, N., Silvia, S., Emma, H., Jacques, F., Marcel,
Z., Gabriela, S., Manuel, B., Annelies, W.S., Isabella, E., Matthias, E., Nicola, L.:
COVID-19 epidemic in Switzerland: on the importance of testing, contact tracing
and isolation. Swiss Med. Wkly. 150(11–12), 4–6 (2020). https://doi.org/10.4414/
smw.2020.20225
12. Mariescu-Istodor, R., Fränti, P.: Grid-based method for GPS route analysis for
retrieval. ACM Trans. Spat. Algorithms Syst. 3(3), 1–28 (2017). https://doi.org/
10.1145/3125634
13. Meixner, S.: How many people have downloaded the COVIDSafe app and
how central has it been to Australia’s coronavirus response? (2020). https://
www.abc.net.au/news/2020-06-02/coronavirus-covid19-covidsafe-app-how-many-
downloads-greg-hunt/12295130
14. PEPP-PT: Pan-European Privacy-Preserving Proximity Tracing (2020). https://
www.pepp-pt.org, https://github.com/pepp-pt/pepp-pt-documentation/blob/
master/PEPP-PT-high-level-overview.pdf
15. Pietrzak, K.: Delayed Authentication Replay and Relay Attacks on DP-3T, pp.
1–6 (2020). https://eprint.iacr.org/2020/418.pdf
16. Raskar, R.M.: Private Kit: Safe Paths; Privacy-by-Design (2020). https://
safepaths.mit.edu/
17. Reichert, L., Brack, S., Scheuermann, B.: Privacy-Preserving Contact Tracing of
COVID-19 Patients. IACR Cryptol, pp. 4–5 (2020). http://arxiv.org/abs/2004.
06818, https://eprint.iacr.org/2020/375.pdf
18. Shao, Z., Taniar, D., Adhinugraha, K.M.: Range-kNN queries with privacy protec-
tion in a mobile environment. Pervasive Mob. Comput. 24, 30–49 (2015). https://
doi.org/10.1016/j.pmcj.2015.05.004
19. Taniar, D., Goh, J.: On mining movement pattern from mobile users. Int. J. Distrib.
Sens. Netw. 3(1), 69–86 (2007). https://doi.org/10.1080/15501320601069499
20. Thomas, S.A., Gharge, S.: Enhanced security for military grid reference system
using visual cryptography. In: 2018 9th International Conference on Computing,
Communication and Networking Technologies, ICCCNT 2018 (2018). https://doi.
org/10.1109/ICCCNT.2018.8494139
284 K. Adhinugraha et al.

21. Waluyo, A.B., Srinivasan, B., Taniar, D.: Research on location-dependent queries
in mobile databases. Comput. Syst. Sci. Eng. 20(2), 79–95 (2005)
22. Xu, F., Huang, Z., Dai, H., Zhu, X., Wu, H., Zhang, J.: A real-time traffic index
model for expressways. Concurrency Comput. 31(24), 1–20 (2019). https://doi.
org/10.1002/cpe.5155
23. Zhang, L., Cai, Z., Wang, X.: FakeMask: a novel privacy preserving approach for
smartphones. IEEE Trans. Netw. Serv. Manage. 13(2), 335–348 (2016). https://
doi.org/10.1109/TNSM.2016.2559448
 Architecture for Real-Time Visualizing
Arabic Words with Diacritics Using
Augmented Reality for Visually
Impaired People

Imene Ouali1(B) , Mohamed Saifeddine Hadj Sassi2 , Mohamed Ben Halima1 ,

and Ali Wali1
1
REGIM-Lab: Research Groups in Intelligent Machines, National Engineering
School of Sfax, University of Sfax, BP 1173, 3038 Sfax, Tunisia
{mohamed.benhlima,ali.wali}@ieee.org
2
Digital Research Center of Sfax (CRNS), Laboratory of Signals, systeMs,
aRtificial Intelligence and neTworkS (SM@RTS), Sfax, Tunisia
[email protected]

Abstract. The number of visually impaired people is continuously

increasing around the world due to the aging of the population. Several
methods of communication are needed as solutions to perform their daily
activities as Arabic text which appear everywhere and contains several
diacritical marks such as i’jām, consonant score and tashkil. Accordingly,
several architectures have adopted different techniques to detect and rec-
ognize Arabic words without diacritics and few are capable of reading
the Arabic word with diacritics. In this context, a new architecture based
on Augmented Reality (AR) is proposed in order to detect and recognize
Arabic letters with diacritical marks in real-time. The proposed architec-
ture shows great potential for using AR engine to detect Arabic words
with diacritics within orientation, writing style and complex background.
In addition, it improves the visualization by reading the detected Ara-
bic words with diacritics through a created dataset. Our work aims to
improve the user’s experience and simplicity for partially sighted and
visually impaired people.

1 Introduction
Nowadays, the use of Immersive Virtual Technologies (IVT) leads to a more
efficient and creative future. Augmented Reality (AR) provides an ideal inter-
face to smart devices in order to show information such as the text displayed
on natural Arabic scenes in real time. Accordingly, text is used everywhere in
our environment like advertising posters, business cards and driving directions.
Thus, it is very important to understand it to interact with our environment.
However, this is not the case for visually impaired people. The information can
be treated incorrectly in case of unclear text. Visually impaired people may find
difficulties to understand texts in case of complex background or bizarre text
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 285–296, 2021.
https://doi.org/10.1007/978-3-030-75100-5_25
286 I. Ouali et al.

type or small size text. Moreover, the complex script of Arabic text requires big
effort and attention to ensure an accurate text detection. The Arabic script has
numerous diacritics, including i’jam, consonant pointing, and tashkil, supple-
mentary diacritics. Thus, the detection and recognition of words with diacritics
is a complex process. Even though several architectures exist for detection and
recognition of Arabic text based on deep learning (DL), few systems are imple-
mented in intelligent devices capable of detecting, recognizing and pronouncing
Arabic words with vowels. Hence, a new architecture has to be created in order
to detect and recognize Arabic words with diacritics. In this context, we propose
a new architecture based on AR for detection, recognition and reading of the
Arabic with diacritics and their vowels. Consequently, our architecture uses AR
should have a transition from supporting communication among text on natural
scene, and human to enabled by data visualization methods through speeches
for people with visually impaired. In the current work, our contributions can be
summarized as follows:
– Several works [1,26] related to text detection and recognition recently
published by the research community have been introduced. The problem
of detection and recognition has been studied for many years. Scientific
researchers indicate that the small fonts and the complex background and
adjustments like color contrasts are big problems facing detection and recog-
nition tasks. It is therefore necessary to continue testing and researching with
different technologies to resolve such problems. This will enhance the selec-
tion of appropriate technologies for our architecture that deal with needs of
visually impaired people.
– We provide an architecture of detection and recognition Arabic text with
vowels. Several systems based on Deep Learning (DL) for Arabic text detec-
tion. However, some of such systems focus on the issue related to detecting
Arabic vowels. For this reason, we have studied methods of text with vowels
detecting based on DL. The limitation of these works allows us to study other
technologies like AR in order to detect correctly Arabic text with vowels.
– We present the detection process method using AR engine and show its effec-
tiveness on detecting Arabic Text in natural scenes through different devices
cited in related work section. This can help future researcher to understand
how to choose the suitable engine.
The remainder of this paper is organized as follows: Sect. 2 defines several
related works. In Sect. 3, the framework of the proposed method and system
is described. Section 4 present the evaluation of our method. Section 5 present
the result and the discussion. Finally, Sect. 6 is reserved for the conclusion and
future work.

2 Related Work
In this section, we present related work focused on the development of new corre-
spondence methods and systems. Several works of text detection and recognition
 Architecture for Visualizing Arabic Words with Diacritics Using AR 287

from videos such as [6–8] and from images such [4,5,10–14,17] and [29] were pro-
posed. A comparative study of different methods and systems of detection and
recognition of text from images is presented in Table 1. Among these methods of
text detection and recognition, there are systems based on AR as a device such
as [1–3,16,18] and DL such as [16,19,20,23,24,27] and [26].

Table 1. Comparison between text detection and recognition systems

Ref Engine Device Databases Language

AR [1] Vuforia Smartphone or tablet Vuforia Word List (VWL) Arabic - English
[2] Unreal Android tablets Photo gallery of Jesi (Italy) Italy
Engine 4
[3] Vuforia Android smartphone Image data of Tajweeed Qur’an Arabic
or tablet
[16] Vuforia Android smartphone Vuforia Word List (VWL) Banjar - Indonesian
or tablet
[18] Vuforia Android smartphone Vuforia Word List (VWL) Aceh - Indonesian
DL [19] OCR Mobile android Data from Web Latin language
[20] OCR Smartphone Database created by authors Latin language
[24] OCR Microsoft HoloLens Database created by authors English
[9] OCR- FCN Computer ACTIV dataset Arabic
[23] OCR Optical glasses Database created by authors Latin language
[26] Google Vision Microsoft’s HoloLens Web Latin language
API - OCR

Mostafa, et al. [1] has proposed an intelligent educational application based

on AR technology. The proposed system is simple and useful and dedicated
mainly for children by teaching them the alphabets, words and numbers in two
languages (Arabic and English). It contains image recognition, text to speech
recognition, fun games and other ideas that are appealing to children. In addi-
tion, Mori, et al. [2] show a virtual interactive experience created for the photo
gallery of Jesi (Italy). Namely, three interactive works produced with Unreal
Engine 4 to give the viewer a greater immersion in the immortal images of
Lorenzo Lotto. The goal was achieved by creating three choreographies with
audio supervised by a historian, a recreation of the works with threedimen-
sional graphics and a soundtrack specially composed by Tecla Zorzi. Further-
more, Andriyandi, et al. [3] define an application of AR as one of the multi-
media technologies that can be used as interactive educational support to help
study the tajweed of the Koran. The method used in this research is Acceler-
ated segment test corner detection (FAST) features. Detecting Arabic text with
vowels consider as an open research issue since there are Arabic language have
numerous diacritics, including i’jam, consonant pointing, and tashkil, supple-
mentary diacritics. Moreover, Syahidi, et al. [16] present a mobile application
(DewataAR) providing information on tourist objects in Bali. This application
uses AR to display three-dimensional objects, video and audio information of
the temples on these tourist objects. DewataAR works by scanning the brochure
of a tourist object using an Android device such as: a tablet or a smartphone.
288 I. Ouali et al.

In addition, Ardian, et al. [18] describe a translation application of the Indone-

sian language Aceh (ARgot). This application is able to operate as a real-time
application without the need for an input medium such as text that must type
using a good quality camera and having the characteristics of the autofocus in
the camera. This research focuses on the process of detecting and tracing text
available in the company. This research also aims to see how AR technology
can be applied to unmarked text and have a different font. The result of the
search is an Android smartphone software, which enables real-time multimedia
translation. It can therefore be used almost in all types of smartphone. Based
on the test and the results of the respondent, this application is able to arouse
the interest of the user and can also be applied in real life.

3 System Architecture
In this section, we present a new architecture to detect, recognize and read
vowelized Arabic text. This work is inspired and based on our previous work [15].
We aim to detect the Arabic word using smart camera features such as a camera.
In addition, the data collected will be analyzed using a tool. It detects the word in
the image and verify it in a database. This tool establishes connections between
the letters stored in the Vuforia dataset and the audioclip dataset. In addition,
the analyzed data will be transmitted to an Audioclip data stored for reading and
viewing using appropriate techniques and technologies. This makes the results
of the Arabic text available to the user (Fig. 1). A detection and recognition
method based on ontology such as [28] is proposed as a Future work.

Fig. 1. Layers of our architectures

 Architecture for Visualizing Arabic Words with Diacritics Using AR 289

3.1 Mapping Between Schema Architecture and Existing

Technologies
Based on a knowledge of the data processing [15] and [25] within proposed exist-
ing AR technologies at the related work section, we map the system components
to more concrete oriented system components (Fig. 2). Firstly, we transmit col-
lected data from camera sensors of smart devices to structured data that can be
analyzed as this method in [21] and [22]. Accordingly, we define a tool that can
extract and recognize data from unstructured data. Android fulfill requirements
of the smart building system thanks to its availability, popularity, and porta-
bility. Android is not designed to be a platform only for mobile phones, it is a
comprehensive platform for smart building devices such as a smartwatch, smart
glasses, and so on. It can collect data from smart devices features that have capa-
bilities like Wi-Fi, camera. Besides, the output data from the tool will be stored
into a Comma Separated Values (CSV) file and loaded into a MySQL database
through JSON Web service. The study of IVT technologies in the second section
permit to select the appropriate AR toolkits and technologies for our architec-
ture. We have decided to use Vuforia AR software development kit. This engine
is more accessible for our study. Furthermore, we can use the local database of
Vuforia to store data of holders. Besides, this engine helps to recognize Arabic
words in a few seconds through the letter and vowelization in local data store
of Vuforia. Moreover, Vuforia AR rendering engine delivers reliable vision-based
experiences across a range of conditions and ensures that users have a great
UX. Furthermore, audioclip data store aims to analyse and read the detected
letters of the word. Finally, a smart speech reader for Arabic text available to
end-to-end user.

Fig. 2. Detailed level of our architecture

290 I. Ouali et al.

3.2 Used AR Engine and Database

Vuforia performs the detection and tracking of text elements by performing a

camera display scan called the Region of Interest (ROI). This technology uses
two different ROI, one used for detection and one used for tracking. Vuforia can
detect words included in the list of predefined words. You can also specify a list
of special words and filters that prevent word recognition. In our system, we add
a database for the Arabic language. Text recognition is useful for applications
that need to recognize individual words and word series. Text recognition may
be used as a stand-alone feature or in combination with a target.
Vuforia relies on UTF-8 character encoding standards and can recognize all
listed characters. It detects and tracks text elements by scanning the camera
display called the Region Of Interest (ROI). A word database containing the
words used to process the translator (Fig. 3). Word lists can be loaded from files
and can be extended with additional words specific to application use cases. It
is also possible to filter word lists using a filter list to exclude the detection of
certain words (using a black list filter) or allow only the detection of certain
words (using a white list filter). Word lists represent a basic set of words that
Vuforia can use as a reference to appropriate text elements appearing in the
camera’s field of view.

Fig. 3. Used databases

The system process contains three phases: detection, recognition and con-
catenation (Fig. 4). In the detection phase, each Arabic letter with vowels is
stored in the Vuforia dataset. A detection function for each letter is used in the
system background to identify each vowelized letter by it correspond Identity
(Id). Besides, an algorithm searches in the audio clip dataset on the correspond-
ing audio of each Id of a letter. The concatenation synthesis is based on the
concatenation of recorded speech segments from audio clip database. Thus, each
Id is concatenated in order to generate the sound of the detected word.
 Architecture for Visualizing Arabic Words with Diacritics Using AR 291

Fig. 4. Process of a word detecting and reading letters with vowels

3.3 Specification of Our Architecture

We have identified architectural requirements that are interdependent on the

existence of the word in the dataset. Therefore, in this section, we will illus-
trate an algorithm for finding words name in the Vuforia dataset BDv and the
audioclup dataset BDa . The new architecture is defined as:
length(name)

(id(l)i ) (1)
i=0

• name: the name of the word.

• i: between 0 and length word.
• l: the letter i of the word.
• id: the identifier of each letter of the word.

To facilitate the tasks and clarify the principle of our idea, we will define
this algorithm which allows to read Arabic word in the dataset. After giving the
specification of our architecture, we define an algorithm to find the word in BDv
and BDa .

Algorithm 1: Read Arabic word with vowels

res=""

For (i = 0, i ++, i < length(name))

For (j = 0, j ++, j < length((BD_v))
If (name[i] == (BD_v) [j])
For (k = 0, k ++, k < length((BD_a))
If (name[i] == (BD_a) [k])
res = res + id(name[i])
292 I. Ouali et al.

ENDIF
ENDFOR
ENDIF
ENDFOR
ENDFOR

return res

Algorithm 1 is used to read an Arabic word in BDv and BDa .

In this algorithm, we will look for the existence of the word (name) captured
by camera in BDv and BDa . The word goes segmented letter by letter. Each
letter (name[i]) has a specific id. The same letter has the same id for the BDv
and BDa . We will look for the existence of the word in the dataset, letter by
letter until the end of the word. If the id of the word found in BDv , we will look
for the same id in BDa . In this case, if we find the same id in BDv and BDa , we
will add this id in the result variable (res). This initialized variable was empty
before starting the execution of the algorithm.

4 Evaluation
In this section, we assess the contribution of our proposed approach. First, we
explain the experimental configuration for the evaluation of the proposed system.
Then, we focus on the evaluation of this system (Fig. 6). The purpose of this
evaluation is to test the potential of an AR device for the general public to
improve the functional vision of the visually impaired and people with almost
complete vision loss and to successfully detect, recognize and read Arabic word
with vowels. During the evaluation of the functionalities of our AR application
which exploit information from reality to facilitate large tasks, the participants
commented positively and confirmed their usefulness in real life. In this part,
we will detail the configuration steps to evaluate our proposed system. We have
tested our system on 25 participants. The latter are 9 women and 16 men of
different ages (between 20 and 50). All of participants were visually impaired
peoples. Participants were identified as P1–P25, respectively. Each participant
has tested the application and after answering a short questionnaire (Q1–Q4).
• Q1 (User Experience): How was your experience?
• Q2 (Satisfaction): Was the system comfort and easy to use?
• Q3 (Efficiency): Were the best results achieved for Arabic text visualization?
• Q4 (Effectiveness): Were the intended results achieved as error rates?
The provided mean value for Usability and User Experience of AR system
for Arabic text with diacritics are significantly higher than other as shown in
Fig. 6. Based on the test and the results of the respondent, this application is
able to arouse the interest of the user and can also be applied in real life. The
overall interface is considered to be powerful enough and flexible in order to
detect and recognize the Arabic words with vowels. Furthermore, the detection
method using the Vuforia engine with the ability to detect the word regardless
of their character size, writing style, complex background (Fig. 5).
 Architecture for Visualizing Arabic Words with Diacritics Using AR 293

Fig. 5. Arabic word detection based on size (a), orientation and writing style (b), and
complex background (c).

Fig. 6. Mean values per question for measuring usability and security on Likert scale

Table 2. Median standard deviation (SD) values per question for measuring usability
and UX on Likertscale.

Users of Median SD
UX Our system 3.6 0.6
System [9] 3.2 1.2
Satisfaction Our system 3.6 1
System [9] 3.5 1.1
Efficiency Our system 3.2 1
System [9] 4.2 0.2
Effectiveness Our system 3.5 1.2
System [9] 2.7 1.4
294 I. Ouali et al.

5 Result and Discussion

Based on the analysis performed on the test, the results confirm that our pro-
posed system has no problems on the different styles of Arabic word, as long as
the text still has a standard pattern on paper. The complex background, ori-
entation and the size are solved through the Vuforia engine. Arabic language
has 28 letters of the Arabic alphabet. Arabic has six vowel phonemes (three
short vowels and three long vowels). Most of systems such as proposed in the
work [9], cannot detect the Arabic language with vowels. We have created the
database that can treat all the letters and vowels in Arabic language. Vuforia
engine deals with the detection phase. We have created another database that
can deal with recognition phase and solve the issue of vowels. Our system based
on Android-based Smartphone in real-time unlike [16,19,20,23,24] systems. The
used architecture is similar to [3]. However, a different method is used in order
to recognize letters with vowels and add an audioclip dataset to read vowels.
As we have mentioned before, we have used two datasets that can handle the
different letters with vowels and solve the problem of writing style that can be
found in [1,16,18] (Table 2).

6 Conclusion and Future Work

In this paper, we have presented an architecture that helps the visually impaired
to read Arabic word with vowels. Hence, we have studied and analyzed the exist-
ing architectures under two categories based on the AR and DL technologies.
This helped us to define the appropriate engine, device and database for our
architecture. Furthermore, we have described the most important limitation of
existing works. Accordingly, we have designed an AR based architecture for Ara-
bic text detection to assist visually impaired people. Besides, we have mapped
between schema architecture and suitable existing technologies. This allows to
deal with problems that was discussed in the related works section. Our archi-
tecture provides a straightforward mobile application that uses Vuforia as an AR
engine to bring in a real-time element with intuitively comprehensible interfaces.
Our paper has focused on solving the limitations of existing works for Arabic
text detection and recognition by taking into consideration visually impaired
people needs. Several works have used deep learning to detect Arabic text with-
out vowels. However, our system can detect and recognize Arabic words with
vowels. In the future, we will try to explore the prospects of extending the data
store by improving the system for a bigger data store of audio-clip. In addition,
we intend to implement Microsoft Hololens display system. This could be used
to improve the AR experience to be more comfortable.
 Architecture for Visualizing Arabic Words with Diacritics Using AR 295

References
1. Mostafa, A., Elsayed, A., Ahmed, M., Mohamed, R., Adel, M., Ashraf, Y.: Smart
educational game based on augmented reality. Technical report, EasyChair (2020
2. Mori, B., Giovent‘u, C.: An augmented reality (ar) experience forlorenzo lotto. In:
Virtual and Augmented Reality in Education, Art, and Museums, pp. 324–332.
IGI Global (2020)
3. Andriyandi, A.P., Darmalaksana, W., Maylawati, D.S., Irwansyah, F.S., Mantoro,
T., Ramdhani, M.A.: Augmented reality using featuresaccelerated segment test for
learning ta. Telkomnika 18(1), 208–216 (2020)
4. Wei, Z.-H., Du, X.-P., Cao, L.: Research on the text detection and recognition
in natural scene images. In: Eleventh International Conference on Graphics and
Image Processing (ICGIP 2019), vol. 11373, p. 113730F. International Society for
Optics and Photonics (2020)
5. Ahmed, S.B., Razzak, M.I., Yusof, R.: Progress in cursive wild text recognition.
In: Cursive Script Text Recognition in Natural Scene Images, pp. 85–92. Springer,
Heidelberg (2020)
6. Elleuch, H., Wali, A., Samet, A., Alimi, A.M.: A real-time eye gesture recogni-
tion system based on fuzzy inference system for mobile devices monitoring. In:
International Conference on Advanced Concepts for Intelligent Vision Systems,
pp. 172–180. Springer, Heidelberg (2016)
7. Elleuch, H., Wali, A., Samet, A., Alimi, A.M.: Interacting with mobile devices by
fusion eye and hand gestures recognition systems based on decision tree approach.
In: Ninth International Conference on Machine Vision (ICMV 2016), vol. 10341,
p. 103410D. International Society for Optics and Photonics (2017)
8. Abdelhedi, S., Wali, A., Alimi, A.M.: Human activity recognition based on mid-
level representations in video surveillance applications. In: 2016 International Joint
Conference on Neural Networks (IJCNN), pp. 3984–3989. IEEE (2016)
9. Beltaief, I., Halima, M.B.: Deep fcn for Arabic scene text detection. In: 2018 IEEE
2nd International Workshop on Arabic and Derived Script Analysis and Recogni-
tion (ASAR), pp. 129–134. IEEE (2018)
10. Turki, H., Halima, M.B., Alimi, A.M.: Text detection in natural scene images
using two masks filtering. In: 2016 IEEE/ACS 13th International Conference of
Computer Systems and Applications (AICCSA), pp. 1–6. IEEE (2016)
11. Ayed, A.B., Halima, M.B., Alimi, A.M.: Map reduce based text detection in big
data natural scene videos. In: INNS Conference on BigData, pp. 216–223 (2015)
12. Turki, H., Halima, M.B., Alimi, A.M.: Scene text detection images with pyramid
image and mser enhanced. In: 2015 15th International Conference on Intelligent
Systems Design and Applications (ISDA), pp. 301–306. IEEE (2015)
13. Selmi, Z., Halima, M.B., Wali, A., Alimi, A.M.: A framework of text detection
and recognition from natural images for mobile device. In: Ninth International
Conference on Machine Vision (ICMV 2016), vol. 10341, p. 1034127. International
Society for Optics and Photonics (2017)
14. Turki, H., Halima, M.B., Alimi, A.M.: A hybrid method of natural scene text
detection using msers masks in hsv space color. In: Ninth International Conference
on Machine Vision (ICMV 2016), vol. 10341, p. 1034111. International Society for
Optics and Photonics (2017)
15. Sassi, M.S.H., Jedidi, F.G., Fourati, L.C.: A new architecture for cognitive internet
of things and big data. Procedia Comput. Sci. 159, 534–543 (2019)
296 I. Ouali et al.

16. Syahidi, A.A., Tolle, H., Supianto, A.A., Arai, K.: Bandoar: real-time text based
detection system using augmented reality for media translator banjar language
to Indonesian with smartphone. In: 2018 IEEE 5th International Conference on
Engineering Technologies and Applied Sciences (ICETAS), pp. 1–6. IEEE (2018)
17. Sassi, M.S.H., Fourati, L.C.: Architecture for visualizing indoor air quality data
with augmented reality based cognitive internet of things. In: International Con-
ference on Advanced Information Networking and Applications, pp. 405–418.
Springer, Cham (2020)
18. Ardian, Z., Santoso, P.I., Hantono, B.S.: Argot: text-based detection systems in real
time using augmented reality for media translator aceh-Indonesia with android-
based smartphones. In: Journal of Physics: Conference Series, vol. 1019, p. 012074.
IOP Publishing (2018)
19. Rajendran, P.S., Christian, I.S., Shedge, M.S.: Aredai augmented reality based
educational artificial intelligence system (2019)
20. Gupta, T., Sisodia, M., Fazulbhoy, S., Raju, M., Agrawal, S.: Improving acces-
sibility for dyslexic impairments using augmented reality. In: 2019 International
Conference on Computer Communication and Informatics (ICCCI), pp. 1–4. IEEE
(2019)
21. Sassi, M.S.H., Fourati, L.C., Jedidi, F.G.: Business information architecture for big
data and Internet of Things. In: 2019 15th International Wireless Communications
Mobile Computing Conference (IWCMC), pp. 1749–1756. IEEE (2019)
22. Sassi, M.S.H., Jedidi, F.G., Fourati, L.C.: A new architecture for cognitive internet
of things and big data. Procedia Comput. Sci. 159, 534–543 (2019)
23. Zhao, Y., Hu, M., Hashash, S., Azenkot, S.: Understanding low visionpeople’s visual
perception on commercial augmented reality glasses. In: Proceedings of the 2017
CHI Conference on Human Factors in Computing Systems, pp. 4170–4181 (2017)
24. Huang, J., Kinateder, M., Dunn, M.J., Jarosz, W., Yang, X.-D., Cooper, E.A.: An
augmented reality sign-reading assistant for users with reduced vision. PloS one
14(1), e0210630 (2019)
25. Sassi, M.S.H., Jedidi, F.G., Fourati, L.C.: Computer-aided software engineering
(CASE) tool for big data and IoT architecture. In: 2019 15th International Wireless
Communications Mobile Computing Conference (IWCMC), pp. 1403–1410. IEEE
(2019)
26. Huang, J.: A hololens application to aid people who are visually impaired in navi-
gation tasks (2017)
27. Sassi, M.S.H., Fourati, L.C.: Investigation on deep learning methods for privacy
and security challenges of cognitive IoV. In: 2020 International Wireless Commu-
nications and Mobile Computing (IWCMC), pp. 714–720. IEEE (2020)
28. Ouali, I., Ghozzi, F., Taktak, R., Sassi, M.S.H.: Ontology alignment using stable
matching. Procedia Comput. Sci. 159, 746–755 (2019)
29. Ouali, I., Sassi, M.S.H., Halima, M.B., Ali, W.: A new architecture based ar for
detection and recognition of objects and text to enhance navigation of visually
impaired people. Procedia Comput. Sci. 176, 602–611 (2020)
 Progressive Mobile Web Application
Subresource Tampering During Penetration
Testing

Tobiasz Wróbel1 , Michał K˛edziora1 , Michał Szczepanik1 , Piotr P. Jóźwiak1 ,

Alicja M. Jóźwiak2 , and Jolanta Mizera–Pietraszko3(B)
1 Faculty of Computer Science and Management, Wroclaw University of Science
and Technology, Wrocław, Poland
{tobiasz.wrobel,michal.kedziora,michal.szczepanik,
piotr.jozwiak}@pwr.edu.pl
2 Faculty of Architecture, Wroclaw University of Science and Technology, Wrocław, Poland
[email protected]
3 Department of Innovative Projects Management, Military University of Land Forces,
Wrocław, Poland
[email protected]

Abstract. Since the boost of mobile devices popularity, both operating systems
and mobile applications have become more complex which in turn transfers into
greater vulnerability to hackers’ attacks. Penetration testing is aimed at detection
of security gaps in mobile systems. On the other hand, Progressive Web uses Web
browser API to enhance the range of functionalities to cross-platform. Thus, this
paper focuses on mobile Web application penetration tests of Progressive Web.
First, some new functionalities were evaluated for vulnerabilities, then an in-depth
analysis of the Web push functionalities was carried out. External resources, which
deliver Web push services, were explored for the libraries security. Then, Man-
in-the-Middle attack on Subresource Integrity Mechanism (SIM) was analyzed to
exploit the vulnerabilities detected.

1 Introduction

Progressive Web Applications (PWA) are becoming a viable option when it comes
to targeting the mobile market. The main benefit from a business standpoint is cost
reduction. PWA is a type of application software delivered through the web and built
using common web technologies including HTML, CSS and JavaScript. It is intended
to work on any platform that uses a standards-compliant browser, including desktop and
mobile devices. Based on that this solution also helps to increase audience and improve
user experience on all popular platforms. Instead of creating separate applications for
each mobile operating system, one application can be used for all of them. All that
comes with almost no requirements of the user’s experience on the contrary to the native
applications. The user has the application’s icon on his device’s home screen, receives
notifications and the application is not fully relying on the internet connection. But

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 297–306, 2021.
https://doi.org/10.1007/978-3-030-75100-5_26
298 T. Wróbel et al.

since Progressive Web applications have become in fact the only Web applications in
usage with some extra functionalities, that modern Web browsers APIs implement, their
vulnerability to attacks is comparable to the native apps [1].
This paper performs analysis of the Progressive Web application standard function-
alities. We study Web push functionalities, external Web push services, Man-in-the-
Middle attacks, subresource integrity mechanism, and subresource tampering. Based
on it, we propose a tool to support penetration tests against the vulnerability detected.
The proposed solution was then implemented and tested both in an artificial simulated
environment and a real environment.

1.1 Progressive Web Applications

A Progressive Web application or PWA is a loose term that describes a type of Web
application built using Web technologies but adjusted to give the user a native app
[7]. There were three technical approaches that had a big impact on PWA as found an
innovative and very popular solutions. First of all, there was a need to store data on the
devices to make the applications less dependent on the constant network connection.
After Application Cache or AppCache has become a part of the HTML standard [17],
Websites and applications started informing the user’s browser to save the files needed.
They used a Manifest file, which contained a list of files, paths or URLs to be stored
on the device’s memory. Then, they can be accessible, even in case of the network
unavailability. However, the question is how to check the cached data for validity. This
is where Service Workers [6] bring their functionalities. They are event-driven JavaScript
processes that work separately from the main browser thread. Such that it enables them
to work in the background even if the user closes the browser. They can perform the
operations such as data synchronization, or they can handle caching resources, intercept
requests, receive updates for the application used by the service worker, and handle push
notifications. The last functionality was defined by the World Wide Web Consortium
in their recommendations [15]. The document describes an API that uses HTTP/2 to
display notifications outside the context of a Web page. Under such circumstances, even
in case when the user does not have his or her browser open, the notification will be in
the background displayed to the user (see Fig. 1).

Fig. 1. The application is submitting requests for a library – first step of MitM Attack
 Progressive Mobile Web Application Subresource Tampering 299

When developing a Web application, we usually need some standard functionalities

such as authentication, managing user accounts, logins, passwords, and access control.
We expect these functionalities to be reliable, work fast, and be secure. Just security is
the most crucial points. Diego Poza from Auth0, an authentication, and authorization
service provider, describes, why it is not advised to manage user identities on your own
[16]. Building the identity management component of an application may be a challenge.
Integrating the application with the various login options may cause problems even with
a very straightforward design. Another one is the user input validation. Custom user
identity management comes with a great risk. Any vulnerabilities pose a threat and any
data breach result in finding product unreliable. Apart from security issues, the service
must be reliable as well. For this reason, according to a W3Techs survey [9] only 23.8%
developers prefer to use frameworks and libraries, while others avoid them. Library usage
means acquiring the source code from the application’s server or accessing a content
delivery network. In both cases, the library is transported over HTTP that can be either
secure or insecure.

1.2 Subresource Tampering via MITMM

The transport phase of acquiring external resources might be vulnerable to a Man-in-

the-Middle attack. In computer security, it is an act of sniffing by connecting both the
victim and the destination, that the victim is trying to contact, and modifying messages
between them without their knowledge [3]. A successful all-out attack gives the hacker
a full knowledge about the content of the messages sent between the user and the server
and even the ability to modify the contents or to insert messages into the communication
channel. Man-in-the-Middle (MITM) is just a model of an attack and can be executed
on different communication layers and channels.
Some attacks performed on HTTPS [3], cellular network UMTS [14], or Bluetooth
[11] are reported in numerous resources. Nonetheless, the most dangerous out of these
three mentioned is the attack on HTTPS, because of the common belief, that the lock
symbol in the browser, which indicates a HTTPS connection to the destination, means
their communication is secure. In practice, the security relies on the validity of the
certificate used to encrypt the communication. The authors mention that a hacker may
successfully perform the attack using his own certificate to secure the communication
between him and the user, and the server’s certificate for the communication between the
attacker and the server simultaneously. Then the lock symbol misleads the user giving
him a false sense of security. After introducing the HTTP Strict Transport Security
[10] the encryption certificates are checked for trustworthiness that is whether or not
it has been issued by a valid certificate authority. If HSTS (HTTP Strict Transport
Security) is set in the website header, the connection is encrypted and consequently, any
other connection that uses unverifiable certificates SSL is blocked such that it cannot
be modified, and the user is alarmed about the incident. Once a hacker has successfully
placed himself between the victim and the server that it is trying to contact him, there’s
a lot of harm to the victim. One of the methods called libraries tampering is while the
victim’s application is working the data can be downloaded in the background without
the victim’s awareness. Tampering means inserting, changing, or deleting any contents
300 T. Wróbel et al.

Fig. 2. When receiving a resource that can be tampered, the attacker changes it and then returns
to the victim – final step of MitM Attack

of the library but only in the case when the resources reach the victim because the attacker
redirects the traffic to it. A simplified subresource tampering attack is shown in Fig. 2.
After the main HTML file of the Website has been downloaded onto the victim’s
device, the browser starts executing the code. Without the the victim’s awareness, the
browser starts requesting for any external resources he needs. A well-performed attack
does not have any impact on the application’s standard behavior. Instead, it performs
malicious actions in the background. For example, the authors of the “Pride and Prejudice
in Progressive Web Apps: Abusing Native App-like Features in Web Applications”
paper were able to change the URL of the third-party Web push server, that the attacked
application tried to use. They replaced it with a self-controlled server were able to send
self-crafted push notifications to the victim.

2 Related Research

With the rise of the popularity of Progressive Web Applications, the concerns about
security related to it have arisen [19, 20]. A quite thorough study was conducted by
a group of researchers at the Korea Advanced Institute of Science and Technology.
The results were presented at the 2018 ACM SIGSAC Conference on Computer and
Communications Security [12]. The authors presented how to abuse some of the new
PWA features. One of them is capturing the subscription object, an object that enables the
third-party Web-Push service to send push notifications to the application user. Given
that the notifications resemble the original, the user may not notice it and become a
victim of this attack. Even if the object is transported over a secure layer, an attacker
can still access it. Some third-party Web-Push services need inclusion of a library to
run the application service. Providing the library sent uses an insecure medium (plain
HTTP for example), it can easily be intercepted and changed. As described in “6.2
- Reflected Transmission of Subscription Objects over HTTPS” [12] some libraries
 Progressive Mobile Web Application Subresource Tampering 301

were holding the push service’s server address, which when redirected to a hacker-
controlled server, results in the user receiving attacker-crafted notifications. According
to the “Involvement of Third Par-ties” paragraph in the paper “How the Web Tangled
Itself: Uncovering the History of Client-Side Web (In)Security” [21] there are two major
risks for the applications with third-party libraries in-built. The first one is the usage of
some outdated versions having all security vulnerabilities identified. The second one is
an insecure library transfer to the application requesting this library. So is the role of SIM.
According to the recommendation from the W3 Consortium [2], in order to assure SIM of
the requested resources, application developers are recommended to use the “integrity”
attribute which contains a cryptographic hash representation of the resources. Upon
receiving the server response with the resources, the application computes the hash to
check for any inconsistency between the present and expected data.

3 Implementation of the Solution

The behavior of SIM attack can be approximated for testing purposes with a HTTP or
HTTPS proxy server. As described in the paper about the CERN httpd proxy server [13],
the server waits for requests from the clients. When he receives a request, he sends it to
the destination server. The response is handled by the proxy server, which then forwards
it directly to the client. Having the same rights as a MITM attacker the proxy performs
additional tasks. One of the key functionalities of httpd [13] was caching resources.
The server stores the resources to his own memory. Then, on receiving a request for a
resource stored in his memory, the server returns it immediately, which gives a significant
performance boost for the client and reduces the traffic on the channel to the destination
server. Another proxy function can monitor and filter the resources. Having access to
the destination URL, the proxy can manage transfer to the resource that the URL points
to.
The search for an extendable proxy starts at the beginning of the implementation
process. At first proxy.py [18] was considered. It does not need a lot of resource to work.
The author describes its background for mobile developers to test their applications with
server-side software in-built in another network. The proxy server allows to redirect the
traffic. It is released under the BSD 3-Clause license. Proxy.py uses custom plugins, but
the user cannot relate a particular request to a response. For this reason, the mitmproxy
[5] was chosen as our proxy server to build upon. It provides easy-to-use interfaces
for submitting requests and receiving responses, quite a straightforward filtering and a
content replacement in the traffic. In addition, it handles relations between the requests
and responses, so that it seems rather unlikely that the proxy user may receive the wrong
response. Creating the plugin was also simplified by the mitmproxy developers. Mitproxy
server provides an event-based interface that allows defining callbacks for many Web
events, like request, response, TCP, or Websocket connection. The solution proposed
is a plugin dedicated for PWA type of apps which provides an option to modify server
requests and responses in a way that can detect some typical for PWA vulnerabilities.
302 T. Wróbel et al.

3.1 Rule Based Replacements

The created tool (see Fig. 3) is an extension for the mitmproxy. The user can define
the content replacements in HTTP response messages. When the response event in
mitmproxy appears, it means that the whole HTTP response has been received by the
proxy server. From now on the user gets an access to the request and the response. Also
the replacements can be easily made.

Fig. 3. The user interface of the proxy tool

At first, the request URL is matched against a set of patterns defined by the user.
Each set is related to the pairs of a pattern and the replacement in case the pattern is
matched. After all the modifications have been made to the response’s content, it is
forwarded to the proxy user. This way the same set of rules on various functions of the
application being tested can be defined. After the first matching the URL pattern changes
the content of the message. However, the rule applies that the changes defined by the
user are minimal because another traffic is kept when received by the proxy and sent to
the its user.

3.2 Transfer Rules Solution

The rules defined by the user can be saved to a file. The JavaScript Object Notation
was chosen as the destination file format. In our opinion, it simplifies handling the read
& write commands from and to the json library for Python [8]. Also, it automatically
handles the mapping of the data from the file to a Python dictionary object being a set
of the key & value pairs. The text-based file with the set of the rules is integrated with
the version control systems. When comparing two versions of the rules file, the user
may have some preferences such as his favorite text editor, the integrated development
environment (IDE), or the version control system, because most of them implement an
interface for highlighting differences between the files. The fundamental structure of
the file containing the rules is shown in Fig. 4. The character sequence in line 2 is the
pattern that the request URL is matched against. In case the pattern matches the URL,
the tool starts processing the key value pairs that are related to this URL pattern.
 Progressive Mobile Web Application Subresource Tampering 303

4 Runtime Environment and Testing

In order to run the background proxy process, which is in fact mitmproxy’s mitmdump
[4], and the user interface, Python environment version at least 3.6 is required. To quickly
setup the environment we will also need pip, a package installer for Python.

Fig. 4. The resulting alert with the changed text message

To separate the environment, we use a virtual environment, which allows creating

individual Python environments with their own packages isolated from other environ-
ments or the operating system’s (OS) packages. Venv is a module that comes as standard
with Python. Before running the proxy we have to establish a set of rules to be in use.
As the tool uses regular expressions, we can define our own rules to be matched. After
that, we want to define the content’s regular expression and the character sequences for
replacements. On having that done the rules in the graphical user interface (GUI) are
saved to the JSON file. From then on, no work will be lost. After setting up all the URL
patterns and the replacement rules, we start the proxy server. By default, it runs on port
8899, but it can be changed in the “config.conf” file in order to filter and tamper only
the tested application’s traffic instead of our whole computer’s internet traffic.
The mock application is a simple Web application that has only a single button.
On clicking a button, it executes the “myFunction” method whose body is located in
the “lib.js” library, that the Web browser requests after it receives the HTML file. The
method displays an alert. In the test for the tool, we change the library content and make
the Web application display an alternative alert (see Fig. 4). We have to match against
“lib.js” in the request URL and then replace the strings.
For the regular expressions usage to match and replace character sequences, we use
“lib[.]js” which matches the desired request URL, as the dot is a special sign in regular
expressions representing any character. Adding the brackets indicates that we want to
match the dot character rather than any character. Then we put our strings for the content
pattern to match and replace it. After refreshing the page by running the proxy and
pressing the same button again, we can see an alert as in Fig. 5. The researchers behind
the “Pride and Prejudice in Progressive Web Apps: Abusing Native App-like Features in
304 T. Wróbel et al.

Web Applications” paper [12] have proved the “AFN.az” news portal to be vulnerable,
so we want to test our tool on it. When accessing the portal using the URL “afn.az”
we are redirected to a non-secure HTTP page. There is a HTTPS version, but the user
would have to explicitly type in the full URL including the right protocol. In this test,
we try to replace the standard footer of the dialog shown at startup, which asks the user
if he would like to receive notifications from the portal. After a brief research, we can
see that the library provided by SendPulse contains the JavaScript methods to display
the subscription dialog. To replace the footer message of the dialog, we identified the
part of the library responsible for it. We change the text to “Tampered by LibReplacer”
(see Fig. 5). After running the proxy and accessing the tested Website, we receive the
following dialog that encourages us to subscribe to the service.

Fig. 5. The tested Website after applying the replacement

As for quantity evaluation, we have tested the user inputs and data presented in this
solution on 132 Progressive Web Apps altogether. Our analysis indicated that around
83% of the server data were vulnerable and 41% of user inputs data were affected by
this attack, but they are usually additionally verified on server-side before processing.
Environment settings needed for performing this kind of attack decrease the possibility
that this kind of attack can be performed in real scenario secure environment on the
user’s site.

5 Conclusion and Further Work

The goal of this paper was to analyze some attacks on subresource integrity SIM using the
Man-in-the-Middle method. Vulnerabilities were assessed in the new Web application
standard, called Progressive Web application. After a more in-depth analysis of Web
push, a new functionality brought by the PWA standard, external Web push services
were assessed for their security. On detecting an attack against an insecure Web push
library transport, we created an attack vector. Man-in-the-Middle attack was analyzed
for its impact on the attacked application’s activity. A Man-in-the-Middle attack type
on insecure resource transport was chosen as the main attack vector that represents the
developed tool. After establishing the attack method, mitmproxy was chosen as the proxy
server to build the tool upon. We developed an extension for mitmproxy that allows the
user to modify the responses received by the proxy before they were relayed to the client.
The tool was tested on a mock Web application and a real news portal application. In
both cases the tool successfully modified the contents of the libraries required by the
Web applications, but it did not change their functionality nor impacted their stability.
 Progressive Mobile Web Application Subresource Tampering 305

Defense strategy against the Man-in-the-Middle based attack on subresource

integrity SIM consists of two stages. The first stage is a defense against a Man-in-
the-Middle attack, which is a matter of network security. A rule of thumb is to avoid
connections to a suspicious public Wi-Fi network, always encrypt the home Wi-Fi net-
work, never access sensitive information while being connected to a suspicious network,
also, to do his best to avoid being attacked by a Man-in-the-Middle hacker.
As for the second stage of defense, the security of the transport must be assured. The
Web application server should use at least HTTPS as the communication protocol to the
application users and subresource integrity SIM is a must. Subresource integrity serves
for validating whether or not the content of the requested resource, usually provided by
a third-party service, is what was expected. SIM is achieved by using a cryptographic
hash of the resource content. When the application receives the resource, it computes the
cryptographic hash of the received resources and compares it to the hash of the expected
one. Any violation means that the resource was somehow changed and may suggest a
security breach.
Our future work on the developed tool goes in two directions. One of them is to
study some altered HTTPS traffics in order to target the applications that implement
secure connections but not necessarily in a fully secure way. Another one is automation
of the detection. Currently, the tool allows using the knowledge about the application
through manual analysis. This part could be potentially done by the computer to allow
fully automatic detection of such vulnerabilities.

References
1. Chalupnik, R., Kedziora, M., Jozwiak, P., Jozwiak, I.: Correspondent sensitive encryption
standard (CSES) algorithm in insecure communication channel. In: International Conference
on Dependability and Complex Systems, pp. 90–98. Springer, Heidelberg (2019)
2. Braun, F., Akhawe, D., Weinberger, J., West, M.: Subresourceintegrity. In: W3C Working
Draft, vol. 3, pp. 3655–3682 (2014)
3. Callegati, F., Cerroni, W., Ramilli, M.: Man-in-the-Middle attack to the https protocol. IEEE
Secur. Priv. 7(1), 78–81 (2009)
4. Cortesi, A., Hils, M., Kriechbaumer, T.: Contributors. Mitmdump (2010)
5. Cortesi, A., Hils, M., Kriechbaumer, T.: Contributors. Mitmproxy: A free and open source
interactive HTTPS proxy, 2010–. [Version 5.0] (2010)
6. GoogleDevelopers.Introductiontoserviceworker. https://developers.google.com/Web/ilt/
pwa/introduction-to-service-worker, Accessed 19 Dec 2019
7. GoogleDevelopers.ProgressiveWebappchecklist. https://developers.google.com/Web/Progre
ssive-Web-apps/checklist, Accessed 17 Dec 2019
8. Python Software Foundation. json—json encoder and decoder. Accessed 23 Dec 2019
9. W3Techs. Usage of javascript libraries for Websites. https://w3techs.com/technologies/ove
rview/javascript_library, Accessed 20 Dec 2019
10. Hodges, J., Jackson, C., Barth, A.: Rfc 6797: Http strict transport security (hsts). IETF (2012).
https://tools.ietf.org/html/rfc6797
11. Kügler, D.: Man in the Middle attacks on bluetooth. In: International Conference on Financial
Cryptography, pp. 149–161. Springer, Heidelberg (2003)
12. Lee, J., Kim, H., Park, J., Shin, I., Son, S.: Pride and prejudice in progressive web apps: abusing
native app-like features in Web applications. In: Proceedings of the 2018 ACM SIGSAC
Conference on Computer and Communications Security, pp. 1731–1746. ACM (2018)
306 T. Wróbel et al.

13. Luotonen, A., Altis, K.: World-wide Web proxies. Comput. Netw. ISDN Syst. 27(2), 147–154
(1994)
14. Meyer, U., Wetzel, S.: A Man-in-the-middle attack on umts. In: Proceedings of the 3rd ACM
workshop on Wireless security, pp. 90–97. ACM (2004)
15. W3C. W3c recommendation (2015). https://www.w3.org/TR/2015/REC-notifications-201
51022/, Accessed 19 Dec 2019
16. Poza, D.: Why you should not Manage your users’ identities. https://auth0.com/blog/why-
you-should-not-Manage-your-users-identities/, Accessed 20 Dec 2019
17. W3C. Application cache as part of the w3c recommendation. https://html.spec.whatwg.org/
multipage/offline.html#appcache, Accessed 17 Dec 2019
18. Singh, A.: Proxy.py – a lightweight, single file http proxy server in python (2018). Accessed
23 Dec 2019
19. Kołaczek, G., Mizera-Pietraszko, J.: Security framework for dynamic service-oriented IT
systems. J. Inf. Telecommun. 2, 428–448 (2018). https://doi.org/10.1080/24751839.2018.
1479926
20. Kołaczek, G., Mizera-Pietraszko, J.: Analysis of dynamic service oriented systems for secu-
rity related problems detection. In: 2017 IEEE International Conference on Innovations in
Intelligent Systems and Applications (INISTA), Gdynia, pp. 472–477 (2017). https://doi.org/
10.1109/INISTA.2017.8001206
21. Stock, B., Johns, M., Steffens, M., Backes, M.: How the Web tangled itself: uncovering the
history of client-side Web (in) security. In: 26th {USENIX} Security Symposium ({USENIX}
Security 17), pp. 971–987 (2017)
 CO-BPG: A Centralized Optimizer for Routing
in BGP-Based Data Center Networks

Chen Duan, Wei Peng, and Baosheng Wang(B)

College of Computer, National University of Defense Technology,

Changsha 410073, Hunan, China
{duanchen,wpeng,bswang}@nudt.edu.cn

Abstract. Load-balancing routing is an important and challenging issue in data

center networks. Traditional internet routing protocols like OSPF and BGP can-
not provide satisfiable network throughput because they are traffic-unaware and
cannot utilize the rich connections and topology regularity in data center net-
works. New routing approaches proposed for data center networks can achieve
good load balance, but cannot deploy widely because they often introduced in
special switch/host hardware or software. In this paper, we propose a centralized
routing optimizer, CO-BGP, for data center networks using BGP as a base rout-
ing protocol. It provides routing optimization with minimal deployment cost with
off-the-shelf data center switches and servers. A routing optimization algorithm
is proposed to optimize the routing performance by taking into account of net-
work traffic load. The algorithm uses Bayes learning to cope with the dynamic of
network traffic. Experiment results have shown that the proposed algorithm has
better performance than the well-known equal cost multi-path (ECMP) approach.

1 Introduction
Data center networks have some important characteristics which differ from the Internet.
Firstly, they normally have a regular topology structure with rich connections like fat
tree, spine-leaf structure [1], which can provide full bisection bandwidth and multiple
paths between any source-destination pairs. Secondly, centralized control is feasible
since devices in a data center are normally concentrated in a local area. Thirdly, data
centers could offer really high bandwidth even upon 40 Gbps using current Ethernet
techniques.
Routing in data center networks is a well-known yet complicated research topic for
decades. It greatly influences the performance of a data center network [2]. There are two
main categories for routing in data center networks. One is to use the Internet routing
protocols as they are mature techniques. Another is to devise new routing schemes
considering the unique characteristics of data center networks.
Traditional Internet routing protocols include intra-domain routing protocols like
OSPF and IS-IS, and inter-domain routing protocol, e.g. BGP. It is argued that IS-IS and
OSPF work well if the network scale is small, but BGP is better for large-scale data center
networks [3]. These distributed routing protocols provide a single path for a destination
and are traffic-unaware, so they are inefficient and cannot fully take advantage of rich

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 307–319, 2021.
https://doi.org/10.1007/978-3-030-75100-5_27
308 C. Duan et al.

connections in data center networks. To alleviate the problem, the equal cost multi-path
(ECMP) technique has often been used to distribute data traffic upon multiple paths.
Some new routing protocols are specifically designed for data center networks. There
are two main categories. The first category adopts the software-defined networking
(SDN) framework which uses a central controller to collect both the network topology
and traffic load information, and then find the optimal flow table configuration. The
OpenFlow protocol is necessary for communication between the SDN controller and
SDN switches. The second category uses programmable switches or smart network
interface controllers (NICs) to realize customized load balancing functions based on
local decisions. Both categories rely on new techniques and do not take advantage of
cost-effective traditional switches.
In this paper, we try to achieve load balancing in a data center network with com-
modity switches running BGP routing protocol. A centralized optimizer, CO-BGP, is
proposed which collects routing and traffic information from switches and reroutes traf-
fic on congested links to avoid packet losses. The Bayes learning technique is used to
make rerouting decisions. The optimizer can be immediately deployed in data centers
without any hardware improvement. We have done experiments on 3-tier CLOS topol-
ogy to validate the performance of the proposed scheme. The experimental results have
shown that CO-BGP can achieve better performance than ECMP with least updating
cost on data center networks.
The rest of the paper is organized as follows. Section 2 introduces the related research
work. The problem formulation is described in Sect. 3. The design of the central optimizer
CO-BGP is illustrated in Sect. 4. Section 5 presents the experiment results and a brief
conclusion is given in Sect. 6.

2 Related Work

To realize load balancing and improve resource utilization ratio of data center networks,
ECMP is often applied as its principle is simple and easy to implement. In ECMP, packets
are forwarded to different interfaces based on the calculation of a hash value using some
selected fields of packet header. Despite of its simplicity, it is believed that ECMP has
three disadvantages. First, it is traffic-unaware, just like traditional routing protocols
used in the Internet. Secondly, it does not support using non-shortest paths. Thirdly, it
is not very efficient in the scenarios with asymmetric network topology or non-uniform
network traffic.
Within the SDN framework, Kannagevlu et al. [4] proposed an edge-to-edge re-
routing scheme with flow granularity for data center networks which have a mixing of
elephant flows and mice flows. Only elephant flows are rerouted to alternative paths
to achieve load balancing. Similarly, FastPass [5] is also a centralized approach which
can realize zero or extra-shallow queues in data center switches. It uses a centralized
controller to assign a transmission path and timeslots for each data flow.
CONGA [6] is a distributed congestion-aware load balancing scheme for data centers
which use global congestion information by feedback and make routing decisions on a
flowlet granularity on access switches. The load information of a path is collected by
piggybacking the information on packets traversing the path. HULA [7] is a hop-by-hop
 CO-BPG: A Centralized Optimizer for Routing in BGP-Based Data Center Networks 309

traffic-aware algorithm on flowlet granularity too. Each switch sends probing packets
to disseminate its load information periodically. Fan et al. [8] proposed a new method
based on them. It divides a network into multiple individual routing domains. Core
switches can carry load information of the links. Edge switches can make optimized
routing decisions.

3 Problem Formulation

3.1 Macro vs. Micro Routing Optimization

The main purpose of routing optimization in a data center network is to balance the
distribution of traffic in the network so that high network bandwidth utilization ratio,
low transmission delay and low packet loss rate can be achieved. To this end, a routing
optimization method needs to collect network states and make optimizing decisions.
Network states usually include the network topology and traffic distribution information.
In addition, the routing information can also be collected by the optimizer.
We classify the optimization into microscopic optimization and macroscopic opti-
mization based on the available network status information. Micro optimization provides
fine-grained traffic scheduling functions. It usually uses programmable switches to col-
lect information of flows [9, 10] or utilize servers’ buffer to predict the traffic amount of
flows [11].
Macro optimization schedule the traffic with coarse granularity. Commodity switches
count the transmitted bytes and discarded packets of the links. However, there is no
available method to calculate the precise traffic matrix of the server pairs. It is hard to
calculate the size of the traffic going to each destination. Researchers [12] adopted deep
learning like GNN to estimate the traffic matrix, but the effect is not good.
Current commodity data centers deploy traditional routing protocols and use ECMP.
They couldn’t afford the high deployment cost of updating switch hardware. Obviously,
in data center networks, there is a gap in routing optimization technology between
academic research and industry. We aim to analyze the macro optimization problem and
propose a deployable optimizer to fill in this gap.

3.2 Problem Definitions

Formally, a data center network is represented by a directed graph G = (V, E), where
V is the set of switch and server nodes, and E is the set of links connecting different
nodes. The set V consists of server nodes and switch nodes. The switch nodes are
commodity switches that support SNMP and run the BGP. E represents physical links
and each element in E could be identified by a switch’s interface or a tuple <u, v> which
bandwidth is denoted by B[v,i] or B(u, v).
In addition, a controller is deployed to compute the optimization decisions, and it
communicates with switch nodes through a separate control network. The controller also
plays as a dumb BGP router which establishes a BGP connection with each switch and
collect routing update messages from switches.
310 C. Duan et al.

Assumption 1. According to RFC 7938 [1], we assume all the switches running an
eBGP process and each switch is configured with a unique AS (Autonomous System)
number. A directed link in E also can be identified by [as1 , as2 ].
Assumption 2. The controller queries each switch about its interface statistics in
a period of T, including: 1) TX (Transmit) dropped count lossN. 2) TX packets count
txpktN. 3) TX bytes count txB.
The distributed routing protocol BGP selects the best path to a destination based
on routing update messages from neighbors. It means that for a switch v, there is only
one valid route item from s to t, which causes wasting of the bandwidth of idle links
and couldn’t utilize the inherent multi-path advantage in data centers. We aim to realize
the load balance by periodically rerouting some traffic from the saturated links to the
underutilized links. Specifically, we filter out the links needed optimize, and reroute part
of the traffic to available underutilized links by configuring the static routes.
Definition 1. In the n-th period, the loss rate of the interface i of node v, lossr(n,v,i)
is defined as formula 1. It also represents the loss rate of the link which connects with
interface i in the output direction and the value is in [0, 1].

lossr(n, v, i) = lossN (n, v, i) / demandN (n, v, i) (1)

demandN (n, v, i) = txN (n, v, i) + lossN (n, v, i) (2)

where demandN(n,v,i) represents the number of packets which are forwarded to

interface i. It may include packets to different destinations and from different input
interfaces. demandB(n,v,i) represent the total bytes of these packets.
Definition 2. The bandwidth utilization of the link in the n-th period is defined as
formula 3.
txB(n, v, i)
util(n, v, i) = . (3)
B[v,i]
Definition 3. In the n-th period, congestion links are the links whose loss rates are
greater than zero. Let Int(v) be the set of switch v’s interfaces, in the n-th period, the set
of congestion links L con (n) is defined as formula 4.

Lcon (n) = {[v, i]|loss(v, i) > 0, v ∈ Vswitch , i ∈ Int(v)} (4)

Definition 4. In the n-th period, average loss rate of network links reflects congestion
status. Because only a few links are congested, the lossr(n,v,i) of most links are equal
to zero which causing the average loss rate really small. We define the true average loss
rate which only calculate the average loss rate of congestion links by formula 5.
1 
TlossR(n) = lossr(n, v, i), n ∈ N + (5)
|Lcon | [v,i]∈Lcon

Definition 5. In the n-th period, the load deviation of the bandwidth utilization
among links could briefly reveal the load balance status in data center networks. The
load-deviation of the bandwidth utilization is defined as formula 6.
  
LB(n) = (1 Ninterface ) (1 + |util(n, v, i) − avgU (n)|)2 , n ∈ N+
v∈Vswitch i∈Int(v)
(6)
 CO-BPG: A Centralized Optimizer for Routing in BGP-Based Data Center Networks 311

  
avgU (n) = 1 (|Vswitch | ∗ n) ∗ util(n, v, i), n ∈ N+ (7)
v∈Vswitch i∈Int(v)

where avgU(n) is the average load of links, and N interface is the total number of links
which is equal to the twice of physical link number because edges are bidirectional. Due
to 0 <= |util(n, v, i) − avgU (n)| <= 1, we add an offset to make it greater than 1.
Obviously, when traffic loads become more unbalanced, LB(n) increases accordingly. In
addition, it could be proved that 1 <= LB(n) <= 2.25.
Definition 6. In the route table of a switch v, routeinfo(v,dst) represents the set of
available output interfaces to destination dst.

3.3 Problem Formulation

For a data centers network G = (V, E), in the n-th period, the routing optimization is a
multi-objective optimization problem. It is formulated as formula 8.

Minimize TlossR (n), LB(n), n ∈ N+

(8)
s.t. TlossR(n) ≥ 0, 1 ≤ LB(n) ≤ 2.25
These objective functions are non-deterministic relationship and difficult to perform
regression analysis. Therefore, we discuss how to transform this multi-objective opti-
mization problem to a single objective function. Given a link, the relation of loss rate
and bandwidth utilization is concluded as formula 9.

→ 1,loss(n, v, i)> 0
util(n, v, i) = (9)
< 1, loss(n, v, i)= 0
In other words, if packet loss happens, the link is saturated and the bandwidth uti-
lization tends to be 1. Under such condition, the multi-objective optimization problem
could be simplified as:

Minimize TlossR(n), if TlossR(n) > 0 (10)

Minimize LB(n), if TlossR(n) = 0 (11)

4 Design and Algorithm

4.1 CO-BGP Overview
Our goal is to design a centralized routing optimizer CO-BGP to alleviate network
congestion and realize load balancing. Figure 1 shows the architecture of CO-BGP,
which is implemented in C language and deployed on the controller. CO-BGP can be
immediately deployed in data centers without any hardware improvement. The controller
may be a Linux server and communicates with switches by separate control network. The
data networks carry business traffic. In this way, it provides high bandwidth and ultra-
low latency between the controller and switches which brings reliability and efficiency
to CO-BGP.
312 C. Duan et al.

Fig. 1. CO-BGP architecture.

Information Collection Module. It consists of SNMP mechanism and dumb BGP

router mechanism which collect interfaces statistics such as ifInDiscards, ifInUcastPkts
and ifInOctets and global routes forwarding tables. They are implemented by SNMP
protocol and a dumb BGP router that establishes connection with each switch by con-
trol network, and only sends HELLO and KEEPLIVE messages. So it learns the global
routes and hand these information over to route optimization module.

Route Optimization Module. According to the current network states uploaded by the
information collection module, it needs to search the unbalanced links and adjust routes
to achieve load balancing which are implemented by the heuristic algorithm and the
Bayes learning-based rerouting mechanism. The details are described in Sect. 4.2, 4.3.
This module hands over the routing update set to Configuration module.

Configure Module. It is responsible for updating static route entries in RIBs. It may
add or delete route entries which is performed by Netconf or Telnet protocol.

4.2 Routing Optimization Module

It is the main body of CO-BGP. It uses heuristic algorithm to divide the global optimiza-
tion problem into multiple sub-problems corresponding to unbalanced saturated links.
The main idea of it is rerouting part of traffic traversing saturated links to underutilized
links so as to satisfy the formula 12.
In the n-th period, the process of this module is described as follow.
 CO-BPG: A Centralized Optimizer for Routing in BGP-Based Data Center Networks 313

Input A connected network G=(V,E)

Initial traffic matrix txBMatrix
Optimize parameter opN, q
Ouput NewSet: Set of new route items
WithdrawSet: Set of route items to be withdrawed
step1 If TlossR(n)>0
sort congestion links based on the loss rate
else
sort unbalanced links based on the utilization
step2 Select the top opN links to optimize
2.1 Search the routes through the link, sort them
and select q ratio of them
2.2 call optimization algorithm to reroute the
selected routes
2.3 optimization algorithm return Newset WithdrawSet
step3 After the n-th period
3.1 calculate the optimization traffic matrix
3.2 update the opN, q for the heuristic algorithm

step1 it determines the optimization object according to the formula 10 and 11, and
sorts the links based on the network conditions.
Step2 it chooses the top opN links and for each link <u, v> or [asu , asv ], it finds out
the route, that is as-path [as1 , as2 , …, asn ], traversing this link by matching AS numbers.
It prefers to reroute routes that have a smaller impact on global network. Therefore,
it sorts them based on the index of asu in as-path which also represents the distance
between the source of as-path and the congestioned link. It just chooses the top q ratio
of these sroutes, then in step 2.2 it hands over the chosen links and routes to rerouting
mechanism. The rerouting mechanism outputs the routing updates.
step3 it collects the network information in the n+1-th period and updates the param-
eters as formula 12. The main idea of updating parameters is that the worse network con-
dition, the larger parameters. In addition, the updating policies of rerouting mechanism
are shown in Sect. 4.3.
⎧
⎨ (opNn−1 + 1, q ∗ 1.2) if TlossR(n) > TlossR(n − 1) ≥ 0
(opN , q) = (opNn−1 , q) if TlossR(n − 1) > TlossR(n) > 0
⎩
(opNn−1 − 1, q ∗ 0.8) if TlossR(n − 1) > TlossR(n) = 0

else if TloosR(n) = TLossR(n − 1) = 0

(opNn−1 + 1, q ∗ 1.2) if loadVar (n) > loadVar(n − 1)

(opN , q) = (12)
(opNn−1 − 1, q ∗ 0.8) if loadVar (n) ≤ loadVar(n − 1)

4.3 Bayes Learning-Based Rerouting Mechanism

The main idea of the mechanism is rerouting part of the traffic traversed congested links
to more appropriate paths, while the rerouting behavior wouldn’t cause congestion in
314 C. Duan et al.

new paths. Therefore, it is critical to know the size of traffic to each destination on
switches. In SDN, researchers tend to estimate the traffic matrix of end to end in flow
granularity. In commodity switches, the collected corse-grained statistics in Assume
2 are too rough to identify the flows, that is difficult to estimate, so macro routing
optimization cannot reserve bandwidth for each flow like FastPass [3]. None the less,
in order to make reasonable optimization decisions and avoid routing oscillation, we
propose a novel Bayes learning-based rerouting mechanism.
We use Bayes learning to build a mapping from statistic data matrix to the score of
each destination traffic routing on switches. We reroute the selected network prefix to the
path with highest score and according to the optimization result matrix fine tune scores
online. It could converge quickly and doesn’t need pre-trained models. The details of
the rerouting mechanism are shown below.
We use Bayes learning to estimate the traffic. The posterior probability is defined
as formula 13, where P(c|S) represents the probability of routing traffic to interface c in
the current network state. P(c) represents the prior probability and P(S|c) represents the
likelihood. P(S) could be regarded as the normalization factor. Then ignoring P(S) we
could infer that the formula 14 is true.

P(c|S) = P(c) P(S|c) / P(S) (13)

P(c|S) ∝ P(c) P(S|c) (14)

Because the Beta distribution could represent many kinds of distribution by changing
parameters, we assume that the likelihood P(S|c) submits to the Beta distribution. In
addition, the prior probability P(c) submits to the Binomial distribution obviously shown
in formula 15. Therefore, because of the conjugacy of the Beta distribution and the
Binomial distribution, it is easier to infer that the posterior probability also submits to
the Beta distribution in formula 16.

P(c) ∼ B(1, 1 / | routeinfo(v, dst) |) (15)

P(c|S) ∼ Beta(a, b) (16)

In the n-th period, given a route entry to destination dst on switch v, if the
|routeinfo(v,dst)|>1, rerouting mechanism would calculate the mean of each posterior
distribution and choose the largest one as the interface to reroute. The mean of the Beta
distribution is defined as formula 17.

u = a / (a + b) (17)

After the n-th period, the heuristic algorithm would generate optimization traffic
matrix as the initial traffic matrix in the n + 1-th period. It also updates the posterior
probability in rerouting mechanism as formula 18, where reward represents the gains of
rerouting and regret represents the loss. The detailed policy is shown in formula 19.

P(c|S) ∼ Beta(a + reward , b + regret) (18)

 CO-BPG: A Centralized Optimizer for Routing in BGP-Based Data Center Networks 315

⎧
⎪
⎪ (0, 2) if TlossR(n) > TlossR(n − 1) and LB(n) > LB(n − 1)
⎪
⎨
(1, 2) if TlossR(n) > TlossR(n − 1) and LB(n) < LB(n − 1)
(reward , regret) =
⎪
⎪ (2, 1) if TlossR(n) < TlossR(n − 1) and LB(n) > LB(n − 1)
⎪
⎩
(2, 0) if TlossR(n) < TlossR(n − 1) and LB(n) < LB(n − 1)
(19)

5 Evaluation
In this section, we evaluate CO-BGP’s performance with real TCP/UDP traffic as well as
large-scale simulations. The real traffic experiments are established on the testbed which
consists of docker containers. They illustrate CO-BGP’s good performance with realistic
workload. Our detailed simulations confirm that CO-BGP scales to large topologies and
give suggestions on the algorithm parameters. We compare our optimization algorithm
with the benchmark ECMP.

5.1 Experiment with Realistic Traffic

Experiment Settings. We evaluate the performance of proposed algorithms on the 3-

tier clos topology network in Fig. 2 and compare with ECMP. We use frr-routing to
simulate data centers switches and ubuntu containers to simulate servers. The switches
are connected by veth (Virtual Ethernet Device) and we modify the driver code of veth
to realize differentiated bandwidth. The optimization period of CO-BGP is 10 s.

Fig. 2. Experiment topology.

The traffic model is relatively stable and mixing with some burst traffic. We send
TCP or UDP flows at a rate of 5 Mbit/s across network, every time for 120 s and send
multiple times. The following metrics are used to measure the performance:

1) true average loss count: In order to show the result more directly, we use the absolute
value of the average of TX dropped count instead of the percentage of loss rate and
define it as TlossA.
316 C. Duan et al.

2) load-Deviation of bandwidth utilization: It shows the load balance status of the

global network. We calculate the load-deviation of bandwidth utilization of the links
according to formula 7.

Experimental Results. Figure 3 shows a sampling results of the optimization perfor-

mance in constant period. In fact, we tested it for many times and they all have significant
results.

Figure 3(a) shows the true average loss count and it could be partitioned into three
stages corresponding to three traffic periods. It is shown that in each stage the TlossA
of CO-BGP decreases significantly in three or five period, and it could reach to zero.
ECMP performs better than no-optimize but because of traffic-unaware, it keeps l oss
packets continuously and the TlossA is above the horizontal dotted line closed to 70.
Figure 3(b) shows the load-deviation of the bandwidth utilization in two traffic peri-
ods. It shows that the LB(n) of CO-BGP decreases significantly. Although ECMP per-
forms better than no-optimize, it oscillates between two horizontal dotted lines. It shows
that CO-BGP could achieve load balancing meanwhile achieve congestion control.
According to above experiment results, we could conclude that CO-BGP shows
significant effect. In the constant optimizing condition, it could achieve better result as
expected because of estimating the traffic.

Fig. 3. Comparison of two evaluation metrics.

5.2 Large-Scale Simulations

Our experimental results are for a comparatively small topology. In order to explore the
performance in large scale topology, we simulate the route optimization module on the
topology of 24 nodes (6 spine and 18 leaf), 120 nodes (10 spine and 110 leaf) and 256
nodes (16 spine and 240 leaf). We implement it in Python. We mainly test the parameter
space and the optimization performance on different scale topologies.

Parameter Space. The opN determines the number of links which needed to be opti-
mized. The initial parameter influence the convergence time of the algorithm. If it was
 CO-BPG: A Centralized Optimizer for Routing in BGP-Based Data Center Networks 317

set too small, the algorithms would need lots of rounds to converge. On the contrary, if
it was set to large, in each period, it would reroute too many flows and cause congestion
so that it may cause route oscillation in the following periods.
Therefore, we set it to 5%, 30%, 50% of the total number of links, and test them.
According to the simulation results in Fig. 4, we find that when setting opN to 30%
of the total number of links, no matter the true average loss count or load-Deviation of
bandwidth utilization it performs better than 5% and 50%.

Fig. 4. Comparison of different parameters in simulations.

Fig. 5. Comparison of two evaluation metrics in simulations.

Optimization Performance. According to the simulations above, we set opN to 30%

of the total number of links and test the optimization performance on different scale
topologies compared with ECMP. The results in Fig. 5 show that our algorithm could
318 C. Duan et al.

decrease the Tloss(n) and LB(n) rapidly in each topology. In most instances, after three
periods the Tloss A(n) and LB(n) of the proposed algorithm is less than ECMP.

We conclude that proposed algorithm could perform well in large scale topology,
and the optimal opN configuration should be around 30% of the total number of links.

6 Conclusion

In this paper, we analyze the centralized routing optimization problem. We survey the
related works and conclude our motivation. Then the routing optimization problem is
classified, formulated and the problem properties are analyzed. We proposed an easily
deployed centralized routing optimizer, CO-BGP, and a novel Bayes learning-based
rerouting mechanism to fill in the gap between academic research and industry. Finally,
the experiment and simulation results show that with minimal deployment cost with
off-the-shelf data center switches and servers, the CO-BGP also achieves significant
effect.

Acknowledgments. This work is supported by National Key Research and Development Program
of China (No. 2018YFB0204301) and a project from National Science Foundation of China (No.
61972412).

References
1. Bao, J., Dong, D., Zhao, B.: DETOUR: a large-scale non-blocking optical data center fabric.
In: Asian Conference on Supercomputing Frontiers, pp. 30–50. Springer, Cham (2018)
2. Wei, Z., Dong, D., Huang, S., Xiao, L.Q.: EC4: ECN and credit-reservation converged con-
gestion control. In: 2019 IEEE 25th International Conference on Parallel and Distributed
Systems (ICPADS), 209–216. IEEE (2019)
3. Internet Engineering Task Force. Use of BGP for Routing in Large-Scale Data Centers: RFC
7938. The Internet Society, New York (2016)
4. Kanagavelu, R., Mingjie, L.N., Mi, K.M., Lee, F.B.H.: OpenFlow based control for re-
routing with differentiated flows in Data Center Networks. In: 2012 18th IEEE International
Conference on Networks (ICON), Singapore, pp. 228–233 (2012)
5. Perry, J., Ousterhout, A., Balakrishnan, H., Shah, D., Fugal, H.: Fastpass: a centralized zero-
queue datacenter network. In: SIGCOMM, pp. 307–318 (2014)
6. Alizadeh, M., Edsall, T., Dharmapurikar, S., Vaidyanathan, R., Chu, K., Fingerhut, A., et al.:
CONGA: distributed congestion-aware load balancing for datacenters. In: Proceedings of the
ACM conference on SIGCOMM, pp. 503–514 (2014)
7. Katta, N., Hira, M., Kim, C., Sivaraman, A., Rexford, J.H.: Scalable load balancing using
programmable data planes. In: Proceedings of the Symposium on SDN Research, pp. 1–12
(2016)
8. Fan, F., Hu, B., Yeung, K.L.: Routing in black box: modularized load balancing for multipath
data center networks. In: INFOCOM on Computer Communications, pp. 1639–1647. IEEE
(2019)
9. Zhang, J., Ren, F., Huang, T., Tang, L., Liu, Y.: Congestion-aware adaptive forwarding in
datacenter networks. Comput. Commun. 62, 34–46 (2015)
 CO-BPG: A Centralized Optimizer for Routing in BGP-Based Data Center Networks 319

10. Ghorbani, S., Yang, Z., Godfrey, P.B., et al.: Drill: micro load balancing for low-latency data
center networks. In: Proceedings of the Conference of the ACM Special Interest Group on
Data Communication, pp. 225–238 (2017)
11. Li, Z., Zhang, Y., Li, D., Chen, K., Peng, Y.: OPTAS: decentralized flow monitoring and
scheduling for tiny tasks. In: IEEE INFOCOM 2016, pp. 1–9. IEEE (2016)
12. Rusek, K., Suarez-Varela, J., Almasan, P., Barlet-Ros, P., Cabellos-Aparicio, A.: RouteNet:
leveraging graph neural networks for network modeling and optimization in SDN. IEEE J.
Sel. Areas Commun. 38(10), 2260–2270 (2020)
 Joint Channel and Power Allocation Algorithm
for Flying Ad Hoc Networks Based on Bayesian
Optimization

Pengcheng Wang, Wei Peng(B) , Wenxin Zhang, and Gaofeng Lv

College of Computer Science, National University of Defense Technology,

Changsha 410073, China
{pcwang,wpeng}@nudt.edu.cn

Abstract. Flying ad hoc networks (FANETs) formed with unmanned aerial vehi-
cles (UAVs) have wide applications in military and civilian fields, including
patrolling, search and rescue operations, monitoring. Due to limited bandwidths
and the fragility of wireless communication, channel and power allocation has
become a key issue which significantly affects the performance of a FANET. This
problem is usually formulated as a mixed-integer nonlinear programming prob-
lem and some intelligent approximation algorithms like particle swarm optimiza-
tion algorithms have been proposed in wireless mesh network (WMN) and D2D
(Device-to-Device) network scenarios. In order to improve the overall network
throughput and spectrum utilization ratio, a joint channel selection and power
allocation algorithm based on Bayesian Optimization is proposed, which uses the
conjugacy property of Beta distribution to obtain approximate optimal allocation
results. Simulation experiment results have shown that the proposed algorithm has
better performance in network throughput and normalized network satisfaction
ratio than the compared algorithms.

Keywords: Channel allocation · Transmitting power control · Bayesian

Optimization · Flying ad hoc network

1 Introduction
Owing to the explosive expansion of wireless communication and UAV technologies,
UAVs have been widely used in civilian and military fields to carry out the missions
including patrolling, search and rescue operations, monitoring, and other military mis-
sions. The cooperative UAVs can be used as wireless relays and flying sensors covering
a wider range, enhancing the multitasking ability and scalability. In order to meet the
increasing network bandwidth and transmission throughout requirements, channel and
power allocation have become a key issue in FANETs.
In wireless mesh network (WMN), paper [1] proposed a swarm intelligence algorithm
named PSO to solve the channel allocation. And paper [2] proposed a joint channel and
power allocation algorithm based on graph theory. In the cognitive radio network, paper
[3] proposed a channel and power allocation algorithm based on random game theory. In

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 320–331, 2021.
https://doi.org/10.1007/978-3-030-75100-5_28
 Joint Channel and Power Allocation Algorithm for Flying 321

Device-To-Device (D2D) networks, paper [4] employs a secondary allocation strategy

to solve the joint channel and power allocation problem. Whether these algorithms are
applicable to FANETs is still a question that needs to be verified. Paper [6] proposed
a slot-based channel allocation and power optimization algorithm under the network
organization relying UAV ground station [5]. Paper [7] considers difference of priority
between UAVs, and makes tendentious allocation of network resources according to
the priority. Paper [8, 9] apply deep reinforcement learning to the channel selection of
wireless networks, but they have not taken care of the interference between different
channels.
Different from the above research work, this paper proposes a heterogeneous multi-
channel network structure (Fig. 1). In this network structure, each node has two sets
of radio frequency equipment, which work in a high frequency band and a low fre-
quency band respectively. Different frequency bands use different encoding, decoding
and modulation methods. The low frequency bands have strong anti-interference ability
called Highly-Protected Band. The high frequency bands called Primary Band here have
higher data rate. In the network, the core node can communicate with all nodes at dif-
ferent time slots in the Highly-Protected Band, and allocates the channel in the Primary
Band according to the communication demand generated by other nodes. We research
the power selection and channel allocation of the Primary Band under such a network
structure, and proposes an algorithm based on Bayesian Optimization. Depending on
the good conjugacy of the Beta Distribution, the probability of a better-performing allo-
cation strategy is continuously increased in the iterative process. Experimental results
show that the algorithm we proposed has better performance in network throughput
and normalized network satisfaction when compared with random allocation algorithm,
PSO, and greedy algorithm.

Fig. 1. Each UAV has two sets of radio frequency equipment, one works in the Highly Protected
Band, the other works in the Primary Band.
322 P. Wang et al.

2 Network Model and Problem Definition

2.1 Network Model

In a multi-node FANET network, we define N = {1,2, 3, …N} as the set of all nodes,
which are peer nodes. Each node has two sets of radio frequency equipment. One set of
radio frequency works in Highly-Protected Band, and all nodes share a channel. This
channel is the control channel that completes the communication between the core node
and the others. The other set of radio frequency works in Primary Band, which have k
channels to choose. These channels are called data channel. Define channel selection
strategy ci = {c1i , c2i , c3i , . . . cki }, when user i selects k channel cki is 1, otherwise it is 0.
The power of the control channel remains unchanged. Power P of the data channel can
be selected between levels {p1 , p2 , p3 …pn }.

2.2 Propagation Model

In this paper, we select the free space propagation model. The minimum transmit power
level of the free space propagation model is calculated by:

Rthresh (4π dij )2 Pmax

pij = [ / ] (1)
Gt Gr λ2 pn
Rthresh is the threshold at which the receiving node can normally receive signals,
dij is the Euclidean distance between nodes i and j, and Gt and Gr are the gains of the
transmitting antenna and the receiving antenna. The power received by the receiving
node is calculated by Eq. (2):

d(j, i)−θ
r,c,p t,c,p
Pij =P (2)
ij

t,c,p
 is the path loss coefficient, and usually the value is a constant between (2, 4), Pij
is the transmission power.

2.3 Interference Model

As shown in Fig. 2, different communication node pairs (such as node pairs A, B, and
C, D) select different channels to minimize interference and improve communication
quality. Relative to the node B, the communication between node C and D is a source
of interference. In the network, when node N i and node N j are in communication, the
interference that it receives can be defined according to the physical interference model
[10]:
 t,c,p
Iijc = Py d(j, y)−θ εj,y (y = i, j) (3)
y∈N

It is the sum of the received signals from other nodes at the receiving point N j . Where
is the transmission power of the node y, d(j, y)−θ is the path loss, d(j, y) is the
t,c,p
Py
 Joint Channel and Power Allocation Algorithm for Flying 323

Fig. 2. Relative to the node B, the communication between node C and D is a source of
interference.

Euclidean distance between nodes j and y. εjy is the weight of the interference which is
related to the channel selected and its value range is [0,1]. The closer the two channels
you choose are, the larger the value is.
Define the signal to interference noise ratio (SINR) based on the physical interference
model:
· d(j, i)−θ
t,c,p
c,p Pij
γij = (4)
PN + Iijc

Where d(j, y)−θ is the path loss, PN is the noise, and Iijc is the interference received from
other nodes.

2.4 Problem Definition

According to Shannon’s formula, the channel capacity is:
 
c,p
Cij = B · log2 1 + γij (5)

Where B is the channel bandwidth. When the channel capacity in the network com-
munication link is greater than the network constant packet rate v, Cij = v. The total
throughput rate in the network is the sum of all transmission link channel capacities:

U= Cij (6)
∀i,j∈N i=j
324 P. Wang et al.

Formally, the joint channel and power allocation problem can be described as:

x∗ = argmaxU
s.t. (1) − (6)
Solving the problem of maximizing U by channel and power allocation in FANET is a
mixed integer nonlinear programming problem, which is an NP problem.

3 Algorithm Description
3.1 Channel and Power Allocation Algorithm Based on Bayesian Optimization
Bayesian Optimization has attracted widespread attention in the field of artificial intelli-
gence by virtue of its good performance in hyperparameter optimization, and is currently
a hot research direction in the field of machine learning. Bayesian Optimization has a
wide range of applications in environmental monitoring and sensor networks, natural lan-
guage text processing, intelligent recommendation systems, combination optimization
and other fields. we use Bayesian Optimization to solve the problem of channel selection
and power allocation in FANETs. Usually optimization problems can be expressed by
the following equation:

x∗ = arg maxf (x) (7)

x∈χ

Where x represents the selected decision vector, χ represents the decision space,
and f(x) represents the objective function. Specific to the optimization problem in this
article, x represents the vector of the channel and power selected by the communication
pair in the network, and f(x) is the network throughput. The framework of Bayesian
optimization is composed of two parts: model which can be updated and queried to drive
optimization decisions and acquisition function. The model can be divided into a priori
probability and a posterior distribution. The prior probability refers to the prior setting
of the unknown problem before the observation. And the posterior distribution refers to
the result of Bayesian analysis, reflecting all our knowledge of the problem under the
given data and model. An appropriate acquisition function can balance the relationship
between exploration and exploitation while searching for the optimal decision. We can
infer the posterior distribution by the Bayes’ rule:
p(D | H )p(H )
p(H | D) = (8)
p(D)
Among them, p(H) is the prior distribution, p(D|H) represents the likelihood, which
reflects the credibility of a certain set of data under a given parameter, p(H|D) is the
posterior distribution, p (D) is the marginal likelihood, which refers to the average of
the probability of the specified observation value obtained under the condition of taking
all possible values of the model parameters [11]. We choose Beta-Bernoulli model to
match the prior and the likelihood. Under the good benefit of its conjugacy, we calculate
the posterior by the Eq. (9):

p(H | α, β) = Beta(Hi | α, β) (9)
 Joint Channel and Power Allocation Algorithm for Flying 325

After one interaction, we update the distribution:


p(H|D) = Beta(Hi |α + Reward, β + Regret) (10)

In the algorithm, we define Reward as the ratio of the current total network throughput
U n to the first throughput (U 0 ) after network initialization. Regret is the absolute value
of the difference between 1 and U n /U 0 . The specific update strategy is as follows:
Un
α = α +ρU0
if Un > Un−1 (11)
 
 Un 
β = β + ρ 1 − if Un < Un−1 (12)
U0 
Where ρ is the amplification factor, and the value is a constant greater than 1 to
improve the convergence speed of the algorithm. The algorithm is encapsulated in the
Algorithm 1.

Algorithm 1 Bayesian Optimization

1: Initialize each element of the hyperparameter variable
matrix α(i, j), β(i, j);
2: for i = 1 to T do
3: for i = 1 to m do
4: for j = 1 to (k + pn) do
5: samples the expected value of the channel or
power level of the link of i : θ(i, j) ∼ Beta(α(i, j),
β(i, j));
6: end for
7: end for
8: Choose the optimal channel and power allocation
strategy;
9: According to the selected channel and power
allocation strategy, calculate the network throughput,
and then calculate the Reward and Regret;
10: update α(i, j) and β(i, j);
11: end for
12: end

3.2 Other Algorithm Description

(1) Particle Swarm Optimization. The Particle Swarm Optimization is a typical
swarm intelligence algorithm. Similar to a flock of birds searching for food and notify-
ing each other to find the best place, PSO finds the optimal solution through continuous
exploration of particles iteratively. In the problem of channel selection and power alloca-
tion, PSO first initialize the solution vector Xi (x1, x2 . . . xn ) of each particle(the solution

vector here includes the channel and power level) and the velocity vector Vi v1, v2 . . . vn .
The iterative equation of the velocity vector is as follows:

Vi = ωVi + c1 r1 (Pbesti − Xi ) + c2 r2 (Gbest − Xi ) (13)

326 P. Wang et al.

Where ω is the weight of inertia, generally between [0,1]. The parameters c1 and c2
are called learning factors, which are random values between [0,1]. The iterative formula
of the solution vector is Xi = Xi + Vi . The specific algorithm is described as follows
according to paper [1]:

Algorithm 2 Particle Swarm Optimization

1: Initialize the particles, calculate the total network
throughput under the individual solution vector, select
the individual local optimal solution vector and the
global optimal of the population;
2: for 1 to gm (gm is the number of iterations)
3: Update the individual velocity vector according to
the velocity vector update equation;
4: Update the solution vector according to the
solution vector update equation;
5: Calculate the total network throughput under the
solution vector, update the local optimal solution vector
and the global optimal of the population;
6: end for
7: output the ;
8: end

(2) Greedy Algorithm. According to the joint power and channel allocation algorithm
based on secondary allocation mentioned in the paper [4], for the network model of this
paper, the algorithm is now described as follows:

Algorithm 3 Greedy Algorithm

1: Sort the communication node pairs in the network in
ascending order of distance;
2: for i in k (k represents the total number of
communication links)
3: Set the power of the transmitter to the maximum;
4: Calculate the total throughput of the network when
different channels are selected respectively, and choose
the one with the largest throughput as the final channel;
5: Calculate the total throughput of the network under
different power levels in turn, and choose the one with
the largest throughput as the final power level;
6: end for
7: end

(3) Random Channel Allocation Algorithm. In this allocation algorithm, the proba-
bility of the channel being selected is the same. According to the communication require-
ments in the network, the channel and power level are randomly assigned to each pair
of communication chains in an optional range with a uniform distribution model.

3.3 Algorithm Complexity Analysis

The time complexity of the algorithm in this paper is O(T*N*(k + pn )) while carrying
out channel selection and power allocation. Where T is the number of iterations, N is the
 Joint Channel and Power Allocation Algorithm for Flying 327

channel need to be allocated. k is the total number of channels that can be allocated, and
pn represents the total number of power levels. The time complexity when calculating
throughput is O(T*N*N 2 ), so the total time complexity of the algorithm in this paper is
O(T*N 3 ).

4 Evaluation
4.1 Environment Model
Assume that UAVs are distributed in the geographic area of H*W, where H is the length
of the area, and W is the width of the area. N communication nodes are randomly
deployed in the area. Supposing that the communication in the network is evenly dis-
tributed, there are k random communication pairs in the network at a certain moment.
The network is updated by random movement of the node position at regular intervals,
and the communication requirements are updated at the same time. The traffic of nodes
in the communication is assumed to be constant bit rate (CBR, Constant Bit Rate) traf-
fic, and the communication packet rate is assumed to be v. The parameters used in the
simulation are listed in Table 1.

Table 1. Simulation parameters

Symbol Meaning Value

H (m) Length of the area 1000
W (m) Width of the area 1000
N Communication nodes 40
v (KB/s) Communication packet rate 128
Rth (dBm) Receive power threshold −40

The evaluation indicators of algorithm performance include network throughput rate

(THt) and normalized network satisfaction (St). Calculate the channel capacity of each
link according to Shannon’s formula and compare it with the packet sending rate of the
communication nodes. If the link channel capacity is greater than the packet sending rate,
the packet sending rate is recorded as the throughput rate of the link, otherwise the actual
channel capacity is recorded. Calculate the total throughput rate of communication node
pairs in the network. The calculation is shown as Eq. (14).

Cij if Cij < v  

c,p
TH t = Cij = B · log2 1 + γij (14)
v if Cij > v
Calculate the percentage of the number greater than the packet rate v to the total
number of communication requirements, and record it as the normalized network
satisfaction.
mt
St = (15)
kt
328 P. Wang et al.

As the Eq. 15 shows, mt is the number of communication nodes greater than the
packet sending rate v until the time period t, and k t is the total number of communication
requirements.

4.2 Evaluation Result

(1) Evaluation of Different Algorithm

Network Throughput. Update the node position in the range of 1000 *1000 m, and then
randomly generate communication requirements, calculate the total network throughput
rate after 100 update cycles. The network throughput rate of each algorithm is shown in
Fig. 3:

Fig. 3. This picture shows the throughput of the network in each update period.

It can be seen from the Fig. 3 that the Bayesian Optimization algorithm is better
than the other three algorithms in most cases in the randomly generated network. We
accumulate the throughput obtained by these 100 network updates to obtain Fig. 4. It
can be seen from the Fig. 4 that the Bayesian Optimization algorithm is significantly
better than the other three algorithms.
Normalize Network Satisfaction. Count the number of communication links in the net-
work whose channel capacity is greater than the packet sending rate, and compare with
the number of communication links. The network satisfaction of various algorithms is
shown in Fig. 5. As can be seen, the bayesian network algorithms have higher satisfaction,
and significantly better than the other three algorithms.

(2) Evaluation of Different Network Settings

Number of Network Nodes. The number of nodes in the network increases from 40 to 90
sequentially, calculating the average throughput and the average normalized satisfaction
after 20 update period (Fig. 6).
 Joint Channel and Power Allocation Algorithm for Flying 329

Fig. 4. The sum of the network throughput in the 100 period is plotted in the figure.

Fig. 5. The cumulative normalize network satisfaction is shown above.

Fig. 6. With the increasing nodes, the average of the network throughput is rising, and the
normalize network satisfaction is descending.

According to the experimental results, the total network throughput shows an increas-
ing trend, but the interference increases, and the network satisfaction shows a down-
ward trend. Although when nodes increasing to a certain extent, the throughput rate of
330 P. Wang et al.

the greedy algorithm is higher than our algorithm, but our algorithm still has obvious
advantages in network satisfaction.
Communication Requirement. In a network with 40 nodes, the communication require-
ments are increased from 8 pairs to 19 pairs, and the average throughput and average
normalized network satisfaction are calculated after 20 update period.

Fig. 7. With the communication requirement increasing, the average of network throughput is
incremental, and the normalize network satisfaction is downward.

As can be seen from Fig. 7, as the number of communication nodes increases, the total
network throughput is still increasing, and the overall network satisfaction is decreasing.
The algorithm proposed in this paper still has a better performance.

5 Conclusion
This paper proposes a joint channel selection and power allocation algorithm based on
Bayesian optimization for heterogeneous multi-channel flying wireless ad hoc networks.
The algorithm uses the Beta distribution as the prior distribution, and iteratively changes
the distribution parameters to maximize the network throughput. The Bayesian opti-
mization algorithm is compared with the traditional swarm intelligence algorithm PSO
[1], greedy algorithm [4] and the simplest random allocation algorithm. Experimental
results show that the Bayesian Optimization algorithm is better than the other three algo-
rithms in terms of network throughput and normalized network satisfaction. The study
in this paper is based on a single-hop network model and proposes a centralized chan-
nel and power allocation algorithm. Realizing distributed channel and power allocation
algorithms in a multi-hop network model is the direction of future research.

Acknowledgments. This work was supported in part by the National Natural Science Founda-
tion of China (No. 61601483), Hunan Young Talents Grant (No. 2020RC3027) and the Training
Program for Excellent Young Innovators of Changsha (No. kq2009027).

References
1. Zhang, Y.C., Wang, Y.J, Yao, S.W.: A PSO-based channel assignment algorithm in wireless
mesh networks (in Chinese). J. Univ. Electron. Sci. Technol. China, 05(v.46), 90–95+108
(2017)
 Joint Channel and Power Allocation Algorithm for Flying 331

2. Chen, X., Zhang, Z.Y., Luo, H.Y.: Joint optimization of power control, channel assignment
andscheduling in wireless mesh network (in Chinese). J. Zhejiang Univ. (Eng. Sci.) 43(8),
1406–11 (2009)
3. Wang, Z.Y., Zhang, H.Y., Xu, N.: Channel assignmentand power control based on stochastic
learning game in cognitive radio networks (in Chinese). J. Acta Electronica Sinica 046(12),
2870–7 (2018)
4. Hsu, C.S., Chen, W.C.: Joint power control and channel assignment for green device-to-device
communication. In: DASC/PiCom/DataCom/CyberSciTech 2018, pp. 881–884 (2018)
5. Bacco, M., Pietro Cassarà, C.M., Gotta, A., Marchese, M., Patrone, F.: A survey on network
architectures and applications for nanosat and UAV swarms. In: WISATS 2017, pp. 75–85
(2017)
6. Xue, Z., Wang, J., Shi, Q., Ding, G., Wu, Q.: Time-frequency scheduling and power
optimization for reliable multiple uav communications. J. IEEE Access 6, 3992–4005 (2018)
7. Zhou, L., Chen, X., Hong, M., Jin, S., Shi, Q.: Joint channel assignment and power allocation
for multi-UAV communication. In: SPAWC 2020, pp. 1–5 (2020)
8. Naparstek, O., Cohen, K.: Deep multi-user reinforcement learning for dynamic spectrum
access in multichannel wireless networks. In: GLOBECOM 2017, pp. 1–7 (2017)
9. Wang, S., Liu, H., Gomes, P.H., Krishnamachari, B.: Deep Reinforcement Learning for
Dynamic Multichannel Access. IEEE Trans. Cogn. Commun. Netw. 4(2), 257–265 (2018)
10. Cardieri, P.: Modeling interference in wireless ad hoc networks. J. IEEE Commun. Surv.
Tutorials 12(4), 551–572 (2010)
11. Shahriari, B., Swersky, K., Wang, Z., Adams, R.P., De Freitas, N.: Taking the human out of
the loop: a review of bayesian optimization. Proc. IEEE 104(1), 148–175 (2016)
 Cluster-Based Distributed
Self-reconfiguration Algorithm
for Modular Robots

Mohamad Moussa, Benoit Piranda, Abdallah Makhoul(B) ,

and Julien Bourgeois

FEMTO-ST Institute, University Bourgogne Franche-Comté, CNRS,

Montbéliard, France
{mohamad.moussa,benoit.piranda,abdallah.makhoul,
julien.bourgeois}@univ-fcomte.fr

Abstract. Modular robots are automated modules that can change

their morphology self-sufficiently and progressively for control or recon-
figuration purposes. Self-reconfiguration is a very challenging problem
in modular robots systems. Existing algorithms are complex and not
suitable for low resources devices. In this paper, we propose a par-
allel and fully distributed cluster-based algorithm to convert a set of
blocks/modules in a new geometrical configuration. We proposed a clus-
ter based self-reconfiguration algorithm. The main idea is to study the
impact of clustering on the self-reconfiguration problem. The modules in
each cluster remain together and try to move in order to find the final
configuration. Based on this concept, our algorithm operates in parallel
in each cluster to fasten reconfiguration process and ultimately the set of
blocks will reach the desirable shape. We evaluate our algorithm on the
centimeter-scale sliding blocks, developed in the Smart Blocks project
in a simulator showing the entire reconfiguration process in real-time.
We show the effectiveness of our algorithm, especially the benefits of
clustering in self-reconfiguration task by comparing performance of both
approaches (with and without clustering) while varying the number of
clusters.

1 Introduction
Programmable matter consists of tactile materials that have physical form
dynamically modifiable. They are able to transform on its own, depending on
program controlled. It is essentially composed from materials and components
that have capabilities to be programmed and change their shape property on
demand. The main approach to implement programmable matter is to build a
huge self-reconfigurable robot consisting of a bunch of modules. Each of these

This work has been supported by the EIPHI Graduate School (contract ANR-17-
EURE-0002).
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 332–344, 2021.
https://doi.org/10.1007/978-3-030-75100-5_29
 Cluster-Based Distributed Self-reconfiguration Algorithm 333

modules has some computation, sensing and communication capabilities. The

modules can detach, move independently and connect to each other to form new
configurations.
Modular robots are distributed systems made up of identical connected mod-
ules which are able to coordinate together in order to perform complex tasks that
can revolutionize everything starting from surgical applications, to transporta-
tion applications or space exploration [1]. The main challenge in modular robots
is self-reconfiguration problem which enables the modules to modify their struc-
ture from one arbitrary initial shape to another (goal shape) [15]. This problem
is very challenging because of the very high number of possible configurations
and a fixed set of rules does not work for all possible situations. It poses several
challenges. Firstly, this process is very slow due to the limited number of modules
which can move simultaneously without interfering which results in limited paral-
lel possible motions and other system constraints. Secondly, self-reconfiguration
requires a coordination between modules during motion in order to avoid any
collision by adapting a set of rules that does not work for all situations since
some motions requires cooperation outfaced by motion algorithms [11]. However,
existing solutions for the self-reconfiguration problem propose several algorithms
using search-based and control-based techniques [5,14]. The basic concept is to
search for a goal configuration in the modular robots configuration space. These
solutions are still very complex and not suitable for modular robots with low
computation and energetic resources. One solution to fasten self-reconfiguration
process is clustering. Grouping modules into clusters and carrying out some tasks
in cluster-based approaches can improve the effectiveness in terms of execution
time and communication cost. Therefore, the partitioning of the modules into
clusters and the parallel displacement of the modules in each partition will accel-
erate the convergence of the modules to the goal shape [2]. In fact, nodes would
care only for executing tasks at intra-cluster level and will not be affected by
inter-cluster changes, thus reducing scope of inter-cluster interactions. Cluster-
ing can yield improvements to the execution of tasks related to modular robots’
programmable matter such as self-reconfiguration where modules forming spe-
cific parts of the initial shape can form one cluster. Therefore, it requires less
reconfiguration adjustments to attain a more or less similar parts of the goal
shape. Hence, here lies the importance and effectiveness of clustering.
Our objective in this paper is to show the impact and the efficiency of cluster-
ing and parallelization on the self-reconfiguration problem. We propose a cluster-
based distributed and parallel self-reconfiguration algorithm that scales to large
modular robot systems in order to speed up the reconfiguration of the modular
robot systems from an initial shape to a goal one. Our aim in this paper is not to
propose a clustering technique, but to provide a new cluster-based distributed
self-reconfiguration algorithm while using our clustering techniques as described
in [2]. The number of clusters is specified based on the goal configuration and
based on the density cut algorithm [13]. The proposed algorithm consists in slid-
ing the not well-placed robots in the interior the current configuration in order
to fill target free cells. For each not well-placed robot, the method compute a
334 M. Moussa et al.

gradient distance from free goal cells, then construct a path for each not well-
placed robot to a free cell which does not cross the path of another robot, and
then moves successively horizontal and vertical lines of robots from the end to
the begin of the path. These steps are repeated until all the goal cells are filled.
The remainder of this paper is organized as follows. In Sect. 2 we provide a short
description of the related work. Section 3 presents the system model and some
assumptions. Section 4 provides the details of our self-reconfiguration algorithm.
The experimental results are presented in Sect. 5 and Sect. 6 concludes our paper.

2 Related Works

In this paper, we consider modular self-reconfigurable robot made from

centimeter-size cubes (blocks) able to move [10]. Works with sliding cubes was
widely used in the literature [6–8]. In [6], it is showed that using homogeneous
Limited Sliding-Cube modules could be performed in quadratic time using meta-
modules and tunneling motions. The method’s advantage is a high degree of
motion parallelism. It was then further investigated in [7] and a novel approach
was proposed for heterogeneous modules. Moreover, a tunneling-based recon-
figuration algorithm was later proposed in [8] to remove the limitation on the
arrangement of start and goal configurations by employing tunneling techniques.
Self-reconfiguration is one of the most critical challenges for modular robots.
It may seem easy to solve, however, it is one of the most challenging issues.
Several solutions have been proposed in the literature and there is no general
solution has been appeared yet. For instance, in [4] the authors propose a self-
reconfiguration in two dimensions. They propose the idea of module substraction,
where unwanted modules remove themselves from the system without external
intervention to attend the desired configuration. The algorithm is composed
of two parts, sequential one for unwanted modules selection and parallel one
for modules movements. The algorithm has quadratic complexity and does not
work in all situations especially in hollow structures. In [11], the authors pro-
pose a distributed algorithm for reconfiguration of lattice-based modular robots
systems made from the same cubic modules used in this paper. This iterative
algorithm can transform various large modular robots, beginning from a filled
shape towards a target one by using motion rules and creating a train of mod-
ules made up of the border of the configuration, which will be moved gradually
in order to reach the target shape. In cite [3] the authors introduced geometric
algorithms for copying, rotating, scaling and reflecting a given polyomino and
they also provided an algorithm for constructing a bounding box surrounding a
polyomino.
Most of existing Self-reconfiguration algorithms are based on sequential
movements which results in dramatically increasing the reconfiguration process’s
duration. In fact, there exists an obvious trade-off between a high degree of
motion parallelism and full convergence to the goal shape due to collisions and
deadlocks that may appear with increasing number of modules moving simulta-
neously. In the literature, existing methods with parallel motions suffer from a
 Cluster-Based Distributed Self-reconfiguration Algorithm 335

limitation in the number of parallel motions depending on the approach used.

In [9], a scaffolding technique is proposed which allows this parallelism while
avoiding collisions. The contribution of this paper is to propose a fully par-
allel and distributed self-reconfiguration algorithm using clustering approaches
with increase in motion parallelism for accelerating self-reconfiguration that can
manage almost to build any goal shape given the initial shape divided into a
pre-defined number of clusters.

3 System Model and Assumptions

In this paper, we consider modular self-reconfigurable robot made from

centimeter-size cubes (blocks) able to move [10]. Blocks are equipped wih a
micro-electro-mechanical-system (MEMS) actuator array in the upper face to
move the objects and they are powered externally. A block can slide along the
border of one of its neighbors. Each block can connect to up to 4 neighbors posi-
tioned on each of its faces. They can communicate together by using neighbor-
to-neighbor message passing.
We assume that these blocks are placed in a simple 2D Cartesian coordinate
system along x and y axes. Each cell can either be occupied by a block or
empty. It is assumed that blocks can recognize their attached neighbors at any
given time and moving blocks cannot communicate with any other ones. We
assume that, at any time, a block can detect if there exists a surrounding cell
∃ f ∈ {East; W est; N orth; South} that should be filled.
Moreover, each cluster is given its goal shape G, also know by cluster mem-
bers. We consider that this target shape is fixed during self-reconfiguration pro-
cess, so no new modules can join or quit any cluster.
Furthermore, the number of blocks in each cluster must be sufficient enough
(greater or at least equal) in order to be able to build the cluster’s target shape.

4 Cluster Based Self-reconfiguration

4.1 Agents

Throughout this paper, we consider the Cluster Head agent for explaining the
algorithm. The Cluster Head is a special treatment to manage a cluster, the
module with this agent acts as a local coordinator guiding the transformation
process to form the target shape of its own cluster. The Cluster Head agent is
dynamic and can change throughout the algorithm. A new Cluster Head module
can be elected in each cluster during reconfiguration, it is defined as:

Definition 1. (Cluster Head Coordinator)

Let N (M, D) be the neighbor cell of M in the direction D, and G be the
goal configuration. M is a Cluster Head coordinator, if it exists: D verifying
(N (M, D) ∈ G) ∧ (N (M, D) is empty).
336 M. Moussa et al.

4.2 Principle

In previous work [10], a module could block all the others by the path it defined
to move to an empty cell. Our algorithm allows reconfiguring a set of connected
blocks divided into clusters considering that each cluster is self-reconfigured to
form its corresponding part of the overall goal shape. This algorithm consists
in exchanging messages with data transmission in the network made up of con-
nected blocks. Each block is considered as an independent robot equipped with
its own processing unit and memory able to execute the program instructions. A
first stage affects a position to each block in the grid and the goal shape for its
cluster. Algorithm 2 illustrates how cluster-based self-reconfiguration is handled
by the Cluster Head block.
At each iteration, one Cluster Head module is elected in each cluster among
potential ones. This block will trigger cluster building (see Algorithm 2) and then
propagate useful information, i.e. distance, to cluster member blocks in order
to start the motion and self-reconfigure into cluster’s goal shape. Also, it will
schedule activities in the cluster and will be in charge of handling intra-cluster
interactions. This Cluster Head block is not static during the self-reconfiguration
algorithm. It can change dynamically at the end of each blocks’ movement. It is
the case when the previous Cluster Head module does not verify the Definition 1
anymore, in other words it no longer has any empty surrounding cell to fill.
Cluster Head block will take in charge of finding the motion path starting from
the block not belonging to goal shape with maximum distance in order to start
self-reconfiguration into cluster’s target configuration.
The algorithm is detailed in the following. It allows to start the motion events
once all the computation is done on all blocks. Mainly, our algorithm consists of
3 major steps. As shown in Fig. 1, these three steps will be executed in parallel
and repeated in each cluster until the cluster’s goal shape is reached.

1. Set blocks distance in each cluster.

2. Find the block Bmax with maximum distance not in the goal shape.
3. Displacement of blocks by establishing motion path following decreasing dis-
tance order starting from Bmax .

4.3 Self-reconfiguration Algorithm

We use message passing system to communicate and transmit data between

connected blocks. As aforementioned, our reconfiguration approach follows three
main steps we detail bellow.

4.3.1 Defining Distances to Blocks in Each Cluster

• CLUSTER GO: Sent by Cluster Head block to its connected neighbors to

start construction of its own cluster’s tree while distributing distance. When
received by a cluster member blocks, it defines its distance as being that
received in the message plus 1.
 Cluster-Based Distributed Self-reconfiguration Algorithm 337

Algorithm 1: Distributed algorithm for the construction of a cluster’s

tree.
myCH // Cluster Id
parentCluster
subT reeSize // per cluster
neighbors // set of neighbours
childrenCluster // set of children
nbExpectedAnswers
initialized // boolean value
1 if isHead() then
2 TriggerCluster()
3 Msg Handler CLUSTER GO(msgCH, msgDistance):
4 initialized ← f alse
5 if myCH = msgCH then
6 distance ← msgDistance + 1
7 if parentCluster = null then
8 parentCluster ← sender
9 nbExpectedAnswers ← neighbours.size − 1
10 if nbExpectedAnswers = 0 then
11 subT reeSize ← 1
12 send CLUSTER BACK(myCH,1) to sender
13 else
14 foreach nId ∈ neighbours do
15 if nid = Mj .id then
16 send CLUSTER GO(myCH) to nId
17 else
18 send CLUSTER BACK(myCH,−1) to sender
19 else
20 send CLUSTER BACK(myCH,−1) to sender
21 Msg Handler CLUSTER BACK(msgCH, msgSize):
22 nbExpectedAnswers ← nbExpectedAnswers − 1
23 if msgSize = −1 and myCH = msgCH then
24 childrenCluster.add(sender)
25 subT reeSize ← subT reeSize + msgSize
26 if nbExpectedAnswers = 0 then
27 subT reeSize ← subT reeSize + 1
28 if parentCluster = null then
29 send CLUSTER BACK(myCH,subT reeSize) to sender
30 else
// head
31 foreach child ∈ childrenCluster do
32 send REQUEST MAX(myCH) to child

• CLUSTER BACK : Response to CLUSTER GO sent by a block to either

acknowledge receipt of distance and that all its neighbors have successfully
set their distance, or to notify the sender that it does not belong to the same
cluster.
338 M. Moussa et al.

Algorithm process in cluster

NO Select new
isHead
head

YES

Reset all variables Set distance = 1

send INIT(myCH)
Propagate distance
information in cluster

Find farthest block

Path elaboration

END
Start motion

Fig. 1. Algorithm of cluster-based reconfiguration.

Algorithm 2: Distributed control algorithm for the Cluster Head agent.

neighbours // maximum 4 neighbours
head ← true
bmaxId // farthest block’s id not in goal shape
bmaxDistance // bmaxId’s distance
myCH // Cluster Id
1 while head do
2 T riggerCluster()
3 isHead()
4 Function TriggerCluster():
5 setNeighbours()
6 distance ← 1
7 nbExpectedAnswers ← neighbours.size()
8 foreach nId ∈ neighbours do
9 send CLUSTER GO(myCH,distance) to nId
10 Event Handler TRANSLATION END:
11 setNeighbours()
12 nbExpectedAnswers ← neighbours.size
13 foreach nId ∈ neighbours do
14 reset all variables
15 send INIT(myCH) to nId
16 Function isHead():
17 return (∃emptyCell ∈ {W est, N orth, East, South} ∧ emptyCell ∈ goal)
 Cluster-Based Distributed Self-reconfiguration Algorithm 339

Once cluster tree is built, the block with maximum distance and not in goal
shape must be found, then the motion starts by following the decreasing path
order of distance. Algorithm 1 illustrates the first step in details. Cluster tree
construction is done by calling the helper function TriggerCluster. Starting from
a Cluster Head module in each cluster, a structure tree rooted at the Cluster
Head node is built with distance assignment using two messages CLUSTER GO
and CLUSTER BACK.

4.3.2 Finding Bm ax
Bmax is defined as the block with maximum distance and not belonging to clus-
ter’s goal shape. The root (Cluster Head ) in each cluster will then launch the
search for Bmax by sending REQUEST MAX to its children through its cluster
tree. This message will be transmitted locally across all cluster members. When
received by a block already in the cluster’s goal shape this message is propa-
gated to children blocks until leaf ones, otherwise block is not in cluster’s goal
shape and should send its id and distance back to tree’s root using SEND MAX
message which is progressively returned to the parent until it reaches the root.
Thus, the root will maintain the block’s id (bmaxId) not in goal shape having
the maximum distance (bmaxDistance) value.

4.3.3 Blocks Displacement

The motion path can be easily established. Starting from Bmax block and follow-
ing decreasing order distance by gradually crossing parentCluster blocks bene-
fiting from the tree structure. This way back leads us to the Cluster Head block
that has empty spot to fill, once Cluster Head is reached the displacement of the
blocks starts where each block has to replace and move to its parentCluster’s
position.
Using programming, the blocks are able to deal with state changes locally
and to handle different events, such as connection/disconnection of a new block,
or the end of a motion. When a Cluster Head block finishes motion, i.e., all the
blocks on the path succeeded in moving to its parentCluster’s position, it resets
all its variables and sends a message of type INIT(myCH) to all neighbors in
order to notify the blocks to reinitialize all their variables with a new iteration
taking place. Then, this Cluster Head must verify if it can still acts as Cluster
Head block or not. If it is the case, old Cluster Head triggers a new iteration and
repeat the process, otherwise a new Cluster Head takes over. This reconfiguration
algorithm is being repeated in each cluster, until all cluster’s blocks converge
towards goal shape, hence cluster’s target shape is well formed.

4.4 Message Evaluation

The proposed cluster-based self-reconfiguration algorithm operates by using a
message–passing model. Our algorithm involves only local and intra-cluster inter-
actions between blocks. The messages are used in the system for communication
purposes, where a block sends a message to a neighbor one asking for data. We
340 M. Moussa et al.

Algorithm 3: Message Handler for any module Mi .

wellP laced // module ∈ goal shape or not
1 Msg Handler REQUEST MAX(msgCH):
2 if ¬wellP laced and myCH = msgCH then
3 send SEND MAX(Mi .id,distance) to parentCluster
4 foreach child ∈ childrenCluster do
5 send REQUEST MAX(msgCH) to child
6 Msg Handler SEND MAX(msgId,msgDistance):
7 if parentCluster = null then
8 send SEND MAX (msgId,msgDistance) to parentCluster
9 else
10 if bmaxDistance < msgDistance then
11 bmaxDistance ← msgDistance
12 bmaxId ← msgId
13 start motion
14 Msg Handler INIT(msgCH):
15 if myCH = msgCH then
16 send INIT ACK() to sender
17 else
18 if !initialized then
19 reset all variables
20 toAnswer ← sender
21 nbExpectedAnswers ← neighbours.size − 1
22 if nbExpectedAnswers = 0 then
23 send INIT ACK () to sender
24 else
25 foreach nId ∈ neighbours do
26 send INIT(myCH) to nId
27 else
28 send INIT ACK () to sender
29 Msg Handler INIT ACK():
30 nbExpectedAnswers ← nbExpectedAnswers − 1
31 if nbExpectedAnswers = 0 and toAnswer = null then
32 send INIT ACK () to toAnswer
33 else
34 if isHead() then
35 TriggerCluster()
36 else
37 foreach nid in neighbours do
38 send NEW HEAD (id) to nid
39 Msg Handler NEW HEAD(msgId):
40 if Mi .id = msgId then
41 TriggerCluster(); // new Head
42 foreach nid ∈ neighbours do
43 if nid = Mj .id then
44 send NEW HEAD (msgId) to nid; // except sender
 Cluster-Based Distributed Self-reconfiguration Algorithm 341

note m the maximum number of connections a module can have, n the number
of blocks, and k the number of clusters. During the first step, except for the
Cluster Head , when a block receives a CLUSTER GO message, it will send it
back to all its neighbours except the sender hence the number of CLUSTER GO
messages exchanged is O(k × n × (m − 1)). Each reception of a CLUSTER GO is
followed by a response of CLUSTER BACK message. Therefore, the total num-
ber of messages exchanged during step 1 is O(2k × n(m − 1)). The total number
of messages exchanged during the last two steps, by using REQUEST MAX,
SEND MAX  and INIT, INIT ACK  is O(2k × m) at each step. Then, the new
Cluster Head selection requires O(k × m) messages. Therefore, the total number
of messages exchanged at each iteration’s execution is O(k × n × m).

5 Simulation and Results

We implemented our proposed algorithm in VisibleSim [12], a simulation and

development environment that supports multi-target large scale modular robot
systems (Smart Blocks, Blinky Blocks, Claytronics...). We conducted extensive
simulations to evaluate the performance of the proposed algorithm. In order
to visualize clustering impact, we focus on comparing cluster-based and with-
out clustering algorithms with varying number of clusters and ensemble sizes.
Two types of complex reconfiguration scenarios have been carried out on several
shapes while varying the number of modules in the ensemble. For each scenario,
we generated different versions of the goal configurations using different scales
ranging from hundreds to ten thousands of blocks. The first scenario formed by
the text: ‘smart’ with various number of blocks. ‘Smart’ word is composed from
five letters, one idea is to let each cluster form one letter as shown in Fig. 2a. In
fact, we have five clusters, each corresponding to one letter with a specific color.
We consider that blocks are initially divided into 5 clusters using [2]. Each block
belongs to a cluster and the cluster’s goal shape is stored in all cluster members.

(a) Smart word formed by 5 clusters. (b) Large comb formed by 2 clusters.

Fig. 2. Two self-reconfiguration scenarios.

The second scenario is to build a large comb. We test our algorithm to self-
reconfigure large networks of more than 10,000 blocks. In this experiment (cf.
Fig. 2a), a large set of Smart Blocks is assembled presenting very regular volumes
of blocks with symmetries. For each experiment, blocks can glow in different
colors depending on running program in order to show a state upon receipt of
342 M. Moussa et al.

Fig. 3. Number of messages exchanged for ‘smart’ word using 5 clusters and to build
a comb using 2,4 and 8 clusters.

a message, detection of an event or to indicate that it has successfully reached

the goal shape. They are also able to display value upon their front face, used
in our case to display useful information such as distance and cluster identifiers.
In Fig. 3, we studied the number of messages exchanged in the two scenarios
according to the size of the goal shape. Firstly, using five clusters to form the
smart’ word with various network sizes (up to 4,800 blocks) and compare it
with results obtained without clustering. The total messages exchanged by both
algorithms linearly increases with the number of blocks. Obviously, the number of
messages exchanged using 5 clusters is much lower without clustering. Secondly,
Fig. 3 also shows the total number of messages sent to construct a comb. It
displays the messages exchanged for a dividing into 2,4 and 8 clusters, versus
those exchanged without clustering. Without clustering approach requires the
highest number of messages exchanged due to the fact that the whole system is
considered as a single cluster, hence a higher number of connections and messages
exchanged between blocks. When no cluster is used, the number of messages
increases exponentially, while it decreases gradually and becomes linear with
the increase in the number of clusters. Hence, in each cluster a lot of messages
will be dropped, especially by modules at the boundaries between clusters. As
a consequence, this confirms that our algorithm only involves local intra-cluster
interactions as announced in previous sections. Fig. 3 enables one to easily notice
that with the increase of both number of clusters and ensemble size, also the gap
between the two approaches is largely increasing.

6 Conclusion and Future Work

In this paper, we proposed a cluster based self-reconfiguration algorithm in a
fully distributed fashion. We conducted two types of complex reconfiguration
experiments, and we evaluated the algorithm on several shapes while varying the
number of clusters and the number of modules in the ensemble. Our proposal
has shown to be effective, i.e., self-reconfiguration algorithms can be aided by
clustering approaches to speed up reconfiguration process. From obtained results,
one can easily notice that the speed of algorithm’s convergence with clustering is
 Cluster-Based Distributed Self-reconfiguration Algorithm 343

much way better than without clustering. The results also show that clustering
has many benefits including energy-efficiency by reducing communication load
(messages exchanged), scalability. As a future work, we aim to improve the
proposed algorithm by making many possible motions within each cluster while
avoiding collisions, thus achieving a high degree of motion parallelism per cluster.
This will surely affect the reconfiguration speed and reduce complexity, but it is
a complex task since it involves more interactions at the inter-cluster level.

References
1. Alattas, R.J., Patel, S., Sobh, T.M.: Evolutionary modular robotics: survey and
analysis. J. Intell. Rob. Syst. 95(3–4), 815–828 (2019)
2. Bassil, J., Moussa, M., Makhoul, A., Piranda, B., Bourgeois, J.: Linear dis-
tributed clustering algorithm for modular robots based programmable matter.
In: IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS
(2020)
3. Fekete, S.P., Gmyr, R., Hugo, S., Keldenich, P., Scheffer, C., Schmidt, A.: Algo-
rithmic aspects of manipulating programmable matter with finite automata. Algo-
rithmica 83, 1–26 (2020)
4. Hall, M.D., Ozdemir, A., Groß, R.: Self-reconfiguration in two-dimensions via
active subtraction with modular robots. In: Robotics: Science and System XVI
(2020)
5. Hauser, S., Mutlu, M., Léziart, P.-A., Khodr, H., Bernardino, A., Ijspeert, A.J.:
Roombots extended: challenges in the next generation of self-reconfigurable mod-
ular robots and their application in adaptive and assistive furniture. Rob. Auton.
Syst. 127, 103467 (2020)
6. Kawano, H.: Complete reconfiguration algorithm for sliding cube-shaped modular
robots with only sliding motion primitive. In: IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS), pp. 3276–3283 (2015)
7. Kawano, H.: Tunneling-based self-reconfiguration of heterogeneous sliding cube-
shaped modular robots in environments with obstacles. In: 2017 IEEE International
Conference on Robotics and Automation (ICRA), pp. 825–832 (2017)
8. Kawano, H.: Distributed Tunneling Reconfiguration of Sliding Cubic Modular
Robots in Severe Space Requirements, p. 14 (2018)
9. Lengiewicz, J., Holobut, P.: Efficient collective shape shifting and locomotion of
massively-modular robotic structures. Auton. Rob. 43(1), 97–122 (2019)
10. Piranda, B., Laurent, G., Bourgeois, J., et al.: A new concept of planar self-
reconfigurable modular robot for conveying microparts. Mechatronics 23(7), 906–
915 (2013)
11. Piranda, B., Bourgeois, J.: A distributed algorithm for reconfiguration of lattice-
based modular self-reconfigurable robots. In: 2016 24th Euromicro International
Conference on Parallel, Distributed, and Network-Based Processing (PDP), pp.
1–9. IEEE (2016)
12. Piranda, B., Fekete, S., Richa, A., Römer, K., Scheideler, C.: Your simulator for
programmable matter. In: Algorithmic Foundations of Programmable Matter, Vis-
iblesim (2016)
13. Shao, J., Yang, Q., Liu, J.,Kramer, S.: Graph clustering with density-cut. arXiv
preprint (2016)
344 M. Moussa et al.

14. Thalamy, P., Piranda, B., Bourgeois, J.: A survey of autonomous self-
reconfiguration methods for robot-based programmable matter. Rob. Auton. Syst.
120, 103242 (2019)
15. Yim, M., Shen, W.M., Salemi, B., Rus, D., Moll, M., Lipson, H., Klavins, E.,
Chirikjian, G.S.: Modular self-reconfigurable robot systems [grand challenges of
robotics]. IEEE Rob. Autom. Mag 14(1), 43–52 (2007)
 Prediction of the Response Time in Web Services
Based on Matrix Factorization

Maksymilian Iwanow1(B) and Jolanta Wrzuszczak-Noga2

1 Faculty of Management, Computer Science and Finance, Wroclaw University of Economics
and Business, ul. Komandorska 118-120, 53-334 Wrocław, Poland
[email protected]
2 Faculty of Computer Science and Management,
Wroclaw University of Science and Technology,
Wybrzeże Wyspiańskiego 27, 50-370 Wrocław, Poland
[email protected]

Abstract. This paper concerns the prediction of the response time of internet
resources depending on a group of various factors, such as agent location, web
browser and cookies, using algorithms related to matrix factorization techniques.
The experiment was based on testing ten university websites from five different
agent locations. The collected data was used to predict the load time of pages using
the SVD (Singular Value Decomposition), SVD++ (Singular Value Decomposi-
tion++) and NMF (Non-negative Matrix Factorization) algorithms. The quality of
the prediction was evaluated using RMSE (Root Mean Square Error), MSE (Mean
Square Error) and MAE (Mean Absolute Error) indexes.

1 Introduction

The Internet has a fundamental impact on people’s lives. It allows to communicate on

such a scale, that was not possible in the time before it was invented, regardless of gender,
nationality, race, religion and from every point in the world. The Internet gives new
opportunities to share sound, images and video. Due to the faster and faster technology
expansion, there is a need to download resources from various places in the world. Time
is money, so every second of the response time matters. It is related to the development
of new online enterprises, start-ups. It creates new e-business opportunities. It is clear
that further progress of e-services will continue making human life easier. Therefore,
load time issues will become more and more important.
To ensure the network ability to provide the necessary services to given traffic within
a certain technological framework, QoS – Quality of Service was introduced. There
are different criteria related to the QoS. The most important for users are metrics that
generally refer to the response time. Resources need to be available as fast as possible
– it is said, that a speed of HTTP (Hypertext Transfer Protocol) load should be less
than 2–3 s [1]. On the other hand, there is a server viewpoint. Providers should ensure
reasonable and precise QoS expectations – parameters are related to the latency, the
minimal number of failures, bit rate, throughput and many others [2].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 345–354, 2021.
https://doi.org/10.1007/978-3-030-75100-5_30
346 M. Iwanow and J. Wrzuszczak-Noga

This work aims to give insides referring to the prediction of the pages load time
using algorithms that are based on matrix factorization techniques. The word ‘predic-
tion’ relates to the performance, i.e. downloading resources according to a set of factors
like agent location, browser. Forecast approaches are correlated with the NMF (Non-
negative Matrix Factorization) and SVD (Singular Value Decomposition) factorization
techniques. Three prediction algorithms (SVD, SVD++, NMF) were evaluated by pre-
diction performance indexes like RMSE (Root Mean Square Error), MSE (Mean Square
Error) and MAE (Mean Absolute Error).
This paper provides the comparison of prediction algorithms based on matrix fac-
torization methods, used to predict load time for different agent locations, browsers and
cookies (caching) policy.
Our experiment contributes to the intelligent measurements, suggests which agent
(agent type, agent localization) and which forecasting method could be used to predict
the future behaviour of the monitored websites.
This paper is organized as follows. In the second section, we present the background
information, including related work. In the third section, we describe the implementation,
setup and scenarios of our real-world Internet experiment. In the fourth section, we
present the results of our study. The final section presents conclusions.

2 Background. Related Work

Prediction is discussed within many areas because it can provide the information that
complements the limited data. In telecommunication and energy networks forecasting
is very important in the context of serving customers as the aspects of the business
processes. There are many different methods used for predicting. Recently collaborative
filtering with flagship matrix factorization techniques has become very important. Based
on the literature, matrix factorization algorithms are usually applied in the QoS field
for predicting metrics related to Web Services, runtime service adaptations, predicting
distances in large-scale networks. Those methods are superior techniques for product
and social or community network recommendations as well.
In [3] there is presented a model for large-scale networks forecasting by matrix
factorization. The authors proposed a scalable system, called Internet Distance Estima-
tion Service (IDES), for predicting large numbers of network distances from limited
measurements. There were applied two algorithms for representing a matrix of network
distances: SVD and NMF. Simulations carried out on the real-world datasets confirmed
that IDES prediction was efficient and robust.
The power of matrix factorization was demonstrated by the Netflix Prize competition
[4]: it turned out, that such based models are superior techniques to formulate product
recommendations. Those techniques have become popular in recent years because of
combining a scalability aspect with predictive accuracy, they offer flexibility in modelling
real-life problems.
In [5] few matrix factorization based algorithms were employed to improve the
Social Recommendation. Nowadays such user-specific applications seem more and more
important. The authors describe the most inconvenient elements in such systems (data
sparsity, cold-start problems).
 Prediction of the Response Time in Web Services 347

In [6] an adaptive approach based on matrix factorization is proposed for performing

online QoS prediction of response time (RT) and throughput (TP). The presented model
is based on the stored data stream applying data transformation and online learning
techniques. The algorithm was evaluated within the prototype implementation for flight
orders. The results showed, that the online QoS prediction approach can be successfully
applied in the runtime service adaptation.
Another model related to the QoS forecast was proposed in the paper [7]. The time-
aware matrix factorization framework (TMF) integrates QoS time series to provide
two-phase QoS predictions in the field of cloud service recommendation.
One more idea of applying matrix factorization was offered in the paper [8] for the
links prediction. The framework based on NMF takes into account not only the graph
structure, but also both internal and external information (extracted from network). The
model solves the link prediction problems basically with local information about a node
attribute and global information related to the topological structure.
[9] deals with the community networks where clients exert random or nearest proxy
Internet access. The situation clearly showed that some of the proxies will be saturated
while the others will be idle. There was proposed a solution based on three matrix
factorization algorithms for estimating the response time: Probabilistic Matrix Factor-
ization, Non-negative Tensor Factorization, Adaptive Matrix Factorization. The results
were compared with the baseline algorithm.
More and more Web Services are published every day, therefore there is a need to
rate them basing on the non-functional features. The forecast of QoS metrics plays very
important role in the service selection field. In [10] an efficient approach is proposed that
allows to predict the missing QoS parameters values in terms of Location and Reputation
aware Matrix Factorization. They show that matrix factorization outperforms many other
existing prediction algorithms.
[11] presents very similar approach. It concerns the forecast of the response time
of downloading web resources, depending on the location of the client and the browser
with the use of the regression based on Fine Tree algorithm and the regression based on
Gaussian 5/2 Metric.
In [12] the application of data mining algorithms for the prediction of Web per-
formance is presented. Researchers evaluate four different predictive models by differ-
ent algorithms: neural networks, decision trees, time series, and regression algorithms.
Authors compare features of two general data mining systems, i.e. Microsoft SQL Server
and IBM Intelligent Miner.

3 Experimental Setup

The initial step of the experiment was to gather the real response time measurements
for the group of ten websites. The process of downloading resources lasted almost four
months. It involved launching five WebPageTest agents and gathering results to the text-
file-structured database. Collected data was exploited for prediction of the pages loading
performance for algorithms SVD, SVD++ and NMF. This section describes the details
of the experimental phase.
348 M. Iwanow and J. Wrzuszczak-Noga

3.1 Data Collection

Tests were conducted according to the following plan:

– 10 tested websites,
– 5 tested locations,
– tests requested 2 times per day (in the morning and in the evening, every website in
each agent location),
– first and repeat view.

The following websites were tested:

– Wrocław University of Science and Technology – https://pwr.edu.pl/

– University of Wrocław – https://uni.wroc.pl/
– AGH University of Science and Technology – https://www.agh.edu.pl/
– Jagiellonian University in Kraków – https://www.uj.edu.pl/
– Warsaw University of Technology – https://www.pw.edu.pl/
– Nanyang Technological University – https://www.ntu.edu.sg/Pages/home.aspx
– University of Tokyo – https://www.u-tokyo.ac.jp/en/index.html
– University of Oxford – https://www.ox.ac.uk/
– Harvard University – https://www.harvard.edu/
– Lomonosov Moscow State University – https://www.msu.ru/

The following agents were chosen:

– Dulles:Chrome – Dulles, USA, Chrome browser

– London_EC2:Chrome – East Central London, England, Chrome browser
– ec2-ap-northeast-1:Firefox – East Central Tokyo, Japan, Firefox browser
– ec2-eu-central-1:Firefox – East Central Frankfurt, Germany, Firefox browser
– ec2-ap-southeast-2:Chrome – East Central Sydney, Australia, Chrome browser

To gather the initial dataset, necessary for conducting the analysis process, there was
used Node.js platform and JavaScript files to reach WebPageTest API. The considered
data was divided into 5 files – each file per one agent (2.000 records per file).

3.2 Algorithms Implementation and Evaluation Details

For the prediction process, three matrix factorization-based methods were tested and
compared with the Normal Predictor and Nearest Neighbour algorithms: SVD, SVD++,
NMF. Prediction with simple SVD is calculated in the following way:


r ui = µ + bu + bi + qiT pu (1)

If the user u is unknown, then the bias bu and the factors pu are assumed to be zero. The
same applies to item i with bi and qi . The SVD++ extends the basic method with a new
 Prediction of the Response Time in Web Services 349

set of item factors that capture implicit ratings. The NMF predictions refer to the simple


equation r ui = qiT pu where both (user and item) factors are kept positive.
Algorithms were implemented in Python with the great help of Surprise Library. It
was designed for the prediction systems with support in the form of few packages like
prediction algorithm package, dataset module, accuracy module. All three algorithms:
SVD, SVD++ and NMF belonged to the Matrix Factorization Prediction algorithms
package, Normal Predictor was related to the Basic Prediction algorithms package and
KNN – to the KNN-inspired Prediction algorithms package.
Algorithms were ranked with three indexes:

• Mean squared error (MSE)

N  2 

i=1 yi − y i
MSE = (2)
N
• Mean absolute error (MAE)
N  
yi − y 


MAE = i (3)
i=1

• Root mean square error (RMSE)


N  2


i=1 yi − y i
RMSE = (4)
N


Where yi – actual value, yi – predicted value, N – number of observations.

The following scenarios were tested:

• first view measurements prediction

• repeat view measurements prediction
• all-measurements, browser prediction
• median load time for agents.

4 Results and Discussion

This paragraph presents the experiment results involving three matrix factorization pre-
diction algorithms: SVD, SVD++, NMF and two other popular ones: KNN and Normal
Predictor (Tables 1, 2 and 3), (Fig. 1). There were used three indexes for accuracy eval-
uation: RMSE, MSE and MAE. Results highlighted with bold font are considered as the
most significant.
350 M. Iwanow and J. Wrzuszczak-Noga

Results: Scenario 1 (First View Measurements Prediction)

Table 1. Quality prediction indexes for SVD, SVD ++, NMF, KNN and Normal Predictor for
five agents – first view measurements prediction.

Algorithm Index Dulles, USA London, Tokyo, Japan Frankfurt, Sydney,

England Germany Australia
SVD MSE 0.097 0.132 0.463 0.253 0.211
MAE 0.155 0.169 0.349 0.208 0.245
RMSE 0.308 0.364 0.678 0.498 0.460
RMSE-MAE 0.153 0.195 0.329 0.29 0.215
SVD++ MSE 0.159 0.136 0.452 0.257 0.217
MAE 0.248 0.174 0.347 0.210 0.256
RMSE 0.396 0.369 0.672 0.505 0.457
RMSE-MAE 0.148 0.195 0.325 0.295 0.201
NMF MSE 0.099 0.137 0.456 0.255 0.215
MAE 0.151 0.165 0.347 0.193 0.245
RMSE 0.314 0.366 0.675 0.495 0.459
RMSE-MAE 0.163 0.201 0.328 0.302 0.214
KNN MSE 0.097 0.133 0.464 0.250 0.214
MAE 0.155 0.171 0.351 0.195 0.252
RMSE 0.311 0.360 0.681 0.499 0.461
RMSE-MAE 0.156 0.189 0.33 0.304 0.209
Normal MSE 0.863 1.118 1.452 1.425 0.993
predictor MAE 0.730 0.809 0.908 0.891 0.779
RMSE 0.929 1.057 1.205 1.194 0.997
RMSE-MAE 0.199 0.248 0.297 0.303 0.218

Results: Scenario 2 (Repeat View Measurements Prediction)

Table 2. Quality prediction indexes for SVD, SVD ++, NMF, KNN and Normal Predictor for
five agents–repeat view measurements prediction.

Algorithm Index Dulles, USA London, Tokyo, Japan Frankfurt, Sydney,

England Germany Australia
SVD MSE 0.030 0.037 0.249 0.060 0.140
MAE 0.039 0.052 0.159 0.059 0.069
RMSE 0.170 0.187 0.499 0.243 0.374
RMSE-MAE 0.131 0.135 0.34 0.184 0.305
(continued)
 Prediction of the Response Time in Web Services 351

Table 2. (continued)

Algorithm Index Dulles, USA London, Tokyo, Japan Frankfurt, Sydney,

England Germany Australia
SVD++ MSE 0.032 0.035 0.252 0.061 0.141
MAE 0.051 0.042 0.152 0.058 0.063
RMSE 0.170 0.187 0.499 0.244 0.332
RMSE-MAE 0.119 0.145 0.347 0.186 0.269
NMF MSE 0.033 0.039 0.250 0.064 0.145
MAE 0.060 0.059 0.131 0.063 0.076
RMSE 0.180 0.195 0.499 0.247 0.381
RMSE-MAE 0.12 0.136 0.368 0.184 0.305
KNN MSE 0.030 0.036 0.250 0.061 0.143
MAE 0.040 0.047 0.152 0.056 0.074
RMSE 0.171 0.189 0.500 0.240 0.348
RMSE-MAE 0.131 0.142 0.348 0.184 0.274
Normal MSE 0.064 0.066 0.355 0.108 0.208
predictor MAE 0.156 0.152 0.332 0.187 0.233
RMSE 0.253 0.257 0.595 0.329 0.443
RMSE-MAE 0.097 0.105 0.263 0.142 0.21

Results: Scenario 3 (All-Measurements, Browser Prediction)

Table 3. Quality prediction indexes for SVD, SVD++, NMF, KNN and Normal Predictor due to
the browser: Chrome and Firefox.

Algorithm Index Chrome Firefox

SVD MSE 0.403 0.602
MAE 0.456 0.511
RMSE 0.628 0.771
RMSE-MAE 0.172 0.26
SVD++ MSE 0.409 0.598
MAE 0.459 0.511
RMSE 0.633 0.768
RMSE-MAE 0.174 0.257
NMF MSE 0.406 0.604
MAE 0.455 0.510
RMSE 0.631 0.772
RMSE-MAE 0.176 0.262
(continued)
352 M. Iwanow and J. Wrzuszczak-Noga

Table 3. (continued)

Algorithm Index Chrome Firefox

KNN MSE 0.414 0.612
MAE 0.461 0.514
RMSE 0.636 0.776
RMSE-MAE 0.175 0.262
Normal MSE 0.875 1.268
predictor MAE 0.699 0.815
RMSE 0.932 1.122
RMSE-MAE 0.233 0.307

Referring to the first view results (Table 1) the best outcomes provided the SVD
algorithm (minimal errors in most cases). Common out-turns delivered NMF and KNN,
a little bit worse was SVD++. In general, MSE and MAE index values for NP algorithm
were about 4 times worse comparing to SVD.
Repeat view outcomes (Table 2) succeeded, in general, similarly: the best was SVD,
a little bit worse: SVD++, NMF and KNN. MSE and MAE index values for NP were
about 2–4 worse comparing to the SVD algorithm. It is pretty clear that repeat view
results (Table 2) are around 3–4 times better (in case of MSE, MAE for SVD, SVD++,
NMF and KNN) that first view outcomes.
The Tokyo and Sydney agent locations delivered the biggest differences between
RMSE and MAE indexes providing the biggest fluctuation errors (Table 1 and Table 2).
The average inaccuracies for two tested browsers (Chrome, Firefox) are presented
in (Table 3). Firefox loses in every scenario with about 10–30% worse performance than
Chrome.
Results: Scenario 4 (Median Load Time)

Fig. 1. Median load time for five agents [s].

According to the median load time for all of the agents, Sydney is the one with the
biggest response time (~6 s), while London with the smallest one (~3 s) (Fig. 1). Tokyo
 Prediction of the Response Time in Web Services 353

is the worst after Sydney, but Dulles in that case lose even with Frankfurt. All in all,
Tokyo and Sydney statistically perform the worst in all prediction scenarios.

5 Conclusion
The experiment shows, that all tested factors like agent location, browser, cookies/cache
policy influence the prediction of resources response time.
The first, general remark is that in every scenario case, per every agent localization,
appropriate index values are in the one, similar result range for SVD, SVD++, NMF and
KNN algorithms, however Normal Predictor always deliver the biggest errors.
When analyzing the agent localization, Tokyo (Firefox browser) generates the biggest
errors. On the other hand, location that statistically appears the best, is Dulles (Chrome
browser). Very evident is the difference between prediction errors for the browsers. In
fact, in every case Chrome browser generates smaller mistakes than Firefox. According
to the median load time for all of the agents, Sydney is the one with the biggest response
time (~6 s), while London with the smallest one (~3 s).
Generally, the smallest differences between RMSE and MAE errors generate the
SVD++. NMF algorithm very often outputs quite similar to SVD.
The future work related to this experiment could be connected with the choice of
other prediction algorithms group, for example regression algorithms. Another idea is
to check the same locations with other browsers. There could be developed other factors
to compare the impact on the websites load time. As page loading is a process, there
might be a point in comparing other metrics that refers to it, for example start rendering
or first byte of server response.

References
1. Schwartz, B.: Google keep site speed below 2–3 seconds (2016). https://www.seroundtable.
com/google-site-speed-recommendation-23057.html, Accessed 10 Oct 2020
2. Cignoli, C.: Why managing QoS matters more than ever (2017). https://www.appneta.com/
blog/why-managing-qos-matters-more-than-ever/, Accessed 10 Oct 2020
3. Mao, Y., Saul, L.K.: Modeling distances in large-scale networks by matrix factorization. In:
Proceedings of the 4th ACM SIGCOMM conference on Internet Measurement, pp. 278–287
(2004)
4. Koren, Y., Bell, R., Volinsky, C.: Matrix factorization techniques for recommender systems.
Computer 42(8), 30–37 (2009)
5. https://labs.engineering.asu.edu/sarwat/wp-content/uploads/sites/26/2014/11/a17-forsati.
pdf, Accessed 30 Oct 2020
6. Zhu, J.M., He, P.J., Zheng, Z.B., Lyu, M.R.: Online QoS prediction for runtime service
adaptation via adaptive matrix factorization. IEEE Trans. Parallel Distrib. Syst. 28(10), 2911–
2924 (2017)
7. Shun, L., et al.: Time-aware QoS prediction for cloud service recommendation based on
matrix factorization. IEEE Access 6, 77716–77724 (2018)
8. Wang, W., Tang, M., Jiao, P.: A unified framework for link prediction based on
non-negative matrix factorization with coupling multivariate information. PloS one
13(11), e0208185 (2018). https://journals.plos.org/plosone/article?id=10.1371/journal.pone.
0208185, Accessed 30 Oct 2020
354 M. Iwanow and J. Wrzuszczak-Noga

9. Gómez-Sánchez, E., et al.: Estimation of web proxy response times in community networks
using matrix factorization algorithms. Electronics 9(1), 88 (2020)
10. Li, S., Wen, J., Luo, F., Cheng, T., Xiong, Q.: A location and reputationaware matrix factor-
ization approach for personalized quality of serviceprediction. In: Proceedings IEEE ICWS,
pp. 652–659 (2017)
11. Wrzuszczak-Noga, J., Borzemski, L.: Prediction of load time for different agent locations
and browsers. In: Workshops of the International Conference on Advanced Information
Networking and Applications, pp. 1298–1307. Springer Cham (2020)
12. Borzemski, L., Kliber, M., Nowak, Z.: Using data mining algorithms in web performance
prediction. Cybern. Syst. Int. J. 40(2), 176–187 (2009)
 BO-RL: Buffer Optimization in Data Plane
Using Reinforcement Learning

M. Sneha1(B) , Kotaro Kataoka2 , and G. Shobha1

1 R V College of Engineering, Bengaluru, Karnataka, India
[email protected], [email protected]
2 Indian Institute of Technology Hyderabad, Sangareddy, Telangana, India
[email protected]

Abstract. Fine tuning of the buffer size is well known technique to improve the
latency and throughput in the network. However, it is difficult to achieve because
the microscopic traffic pattern changes dynamically and affected by many factors
in the network, and fine grained information to predict the optimum buffer size for
the upcoming moment is difficult to calculate. To address this problem, this paper
proposes a new approach, Buffer Optimization using Reinforcement Learning
(BO-RL), which can dynamically adjust the buffer size of routers based on the
observations on the network environment including routers and end devices. The
proof of concept implementation was developed using NS-3, OpenAI Gym and
TensorFlow to integrate the Reinforcement Learning (RL) agent and the router to
dynamically adjust its buffer size. This paper reports the working of BO-RL, and
the results of preliminary experiments in a network topology with limited number
of nodes. Significant improvements are observed in the end to end delay and the
average throughput by applying BO-RL on a router.

1 Introduction
Low latency and high throughput are the global demands for any of modern networked
services. Many optimization techniques have been explored to improve latency and
throughput at the same time including Quality of Service (QoS) [1, 2] transport layer
protocol extensions [3, 4] and traffic engineering in controlled environment such as Data
Center Networks [5–8] dynamic buffer-size adjustment for packet forwarding in the
data plane of the network has also been actively explored [9–12]. Packet buffering in
the data plane determines how many packets can wait before they are transmitted to
the link or discarded. Consequently, it significantly affects the latency and packet loss
rate caused by the devices in the data plane. However, algorithm based solutions cannot
exhibit their best performance if the traffic pattern goes out of their model. When a new
destructive protocol emerges, such a protocol may introduce new factors and parameters
that are not supported by those algorithm based solutions.
In order to achieve the resilience to the dynamic changes of traffic trend on the
buffer size optimization in the data plane, this paper proposes Buffer Optimization using
Reinforcement Learning (BO-RL). BO-RL uses Deep Q Learning to implement the RL
agent, which continuously makes the decision to change the buffer size based on the
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 355–369, 2021.
https://doi.org/10.1007/978-3-030-75100-5_31
356 M. Sneha et al.

constant monitoring results of end to end latency and throughput in the network. This
paper developed a proof of concept (PoC) implementation using NS-3, OpenAI Gym,
and TensorFlow to optimize the buffer size of one router in the network. The experimen-
tation results showed the significant improvement on both latency and throughput, and
left some lessons for both further improvement and real-world deployment. The main
contributions of this research work are: 1) designing and implementing a practical use
case of RL to optimize the buffer size, 2) experimentally proving the effectiveness of
the proposed approach in the small and controlled environment, 3) sharing the lessons
for the real world deployment.
The rest of the paper is organized as follows: Related work is reviewed in Sect. 2.
BO-RL system design is described in Sect. 3 and system implementation in Sect. 4.
Section 5 shows the simulation results with limitations and potential improvements to
the study. Finally, the work is concluded in Sect. 6.

2 Related Work

Quang Tran Anh Pham et al. [1] used RL with DDPG (Deep Deterministic Policy Gra-
dient) algorithm for Software Defined Networks (SDN) which takes traffic patterns as
input and incorporates QoS into the reward function. This work explored a better rout-
ing configuration as the action of the agent, which also improves the latency. H. Yao
et al. [3] proposed load balance routing schemes for SDN considering queue utilization
(QU). The routing process was divided into three steps, the dimension reduction, the QU
prediction using Neural Networks, and the load balance routing. H. Yao et al. assumed
that if the worst throughput is higher, then the load balancing is better. Yiming Kong
et al. [4] proposed TCP Congestion Control using RL based Loss Predictors where the
static buffer is altered manually at router to check different scenarios. RL-TCP showed
a 7–8% decrease in RTT and 9% increase in throughput on an extremely under-buffered
bottleneck link.
M. Elsayed et al. [13] proposed a Deep Reinforcement Learning (DRL) based
dynamic resource allocation algorithm to reduce the latency of devices offering
mission-critical services for small cell networks. Both the base station and the end nodes
were running the DRL agent and the results showed 30% improvement in latency for
dense scenarios with 10% throughput reduction. N. N. Krishnan et al. [14] proposed
a DRL framework to learn the channel assignment for the Distributed Multiple Input
Multiple Output (D-MIMO) groups for WiFi network in order to maximize the user
throughput performance. They achieved 20% improvement in the user throughput per-
formance compared to other benchmarks.
N. Bouacida et al. [9] proposed LearnQueue which is an AQM (Active Queue Man-
agement) technique using ML. It dynamically tunes the buffer size of Wi-Fi Access
Point to decrease the time that packets spend in the queues and thus to reduce the
latency. N. Bouacida et al. observed from the experimental results that LearnQueue
reduces queuing latency drastically while maintaining the similar throughput in most of
the cases. Minsu Kim proposed [10] a DRL based AQM which selects a packet drop or
non-drop action at the packet departure stage depending on the current state consisting
of current queue length, dequeue rate, and queuing delay. After an action is selected,
 BO-RL: Buffer Optimization in Data Plane Using Reinforcement Learning 357

a reward is calculated based on the queuing delay and packet drop-rate. Dropping or
non-dropping action of a packet is not taken by the agent rather it is left to the queuing
discipline employed on the node.

3 System Design
3.1 The Overall Architecture

This paper proposes a system to improve end-to-end delay and throughput in a network
by optimizing the buffer size on the routers using RL algorithm as shown in Fig. 1. The
proposed architecture introduces RL Agent, which is an additional node in the network
to run the RL operation. Network nodes and end hosts are extended to interact with RL
Agent to allow it to give the instruction of buffer size adjustment on the network. There
are four different steps in the RL operation: 1) Observing 2) Training & Learning 3)
Taking the Action on the environment 4) Getting the reward. RL Agent observes the
network state from both router and end hosts from the environment, takes actions based
on the total reward earned for the previous action, gets the reward for current action.

3.2 Formulating the State Space st

Four parameters are considered to form a state space st ={x1 , x2 , x3 , x4 } as shown in
Fig. 1 The parameters x1 = the number of packets dropped because of queue length
and x2 = average qsize at the router are observed from the switch/router. The parame-
ters x3 =average per packet delay and x4 = average throughput are observed from the
destination node.

3.2.1 Observing the Router for the Parameters x1 and x2

At every time step t the agent observes the average queue size and the packets dropped
due to queue size from the router in the network environment. Here queue refers to the
QueueDisc [15] of traffic control layer in Linux systems. Figure 2(a) illustrates the flow
chart of the algorithm used for converting this network information from the router to
agent’s observation. CheckQueueDiscSize procedure is called every 0.1 ms to check the
average queue size and the number of packets dropped by exceeding the queue size
within the time step t.

3.2.2 Observing the Destination for the Parameters x3 and x4

At every time step t the agent observes the network environment to read the average
throughput and average per packet delay from the destination. Figure 2(b) illustrates
the procedure to calculate the average end-to-end delay and throughput within the time
step t as the observation parameters for the agent. Whenever a packet is received at
the destination node the FeedDelayThroughput procedure is called, to record the delay
experienced by that packet and the size of the packet (in bytes) received. The packet’s
transmission time is attached to the packet at the sender when it’s transmitted, which is
extracted at the destination node for calculating the delay.
358 M. Sneha et al.

Fig. 1. Overall architecture of the proposed system

3.3 Choosing the Algorithm for RL Agent

Once the state space is formulated, the agent algorithm must be identified, which would
best suit the requirements for BO-RL. The agent learning algorithms can be value-
based, policy-based or model based [16]. The first two categories are model free algo-
rithms and are suitable for networking research because the agent will not be having
access to a model of the network environment.
BO-RL implements the DQN algorithm proposed by Mnih et al. [18] with expe-
rience replay as the agent algorithm because; 1) DQN can accurately approximate the
desired value function for the environments with large state space. Since we have net-
work as the environment, it is impossible that the agent would observe a previously
observed state 2) DQN’s function approximation based value function can scale very
well, which is a natural choice for more complicated large networks.
In the value based algorithm, the goal of the algorithm is to build a value function
which eventually gets defined into a policy. The simplest algorithm in this category is
Q-Learning[17]. Q-learning builds a lookup table of Q(s,a) values, with one entry for
each state-action pair. Algorithm is well suited for the environments where the states
and the actions are small and finite. Q-learning algorithm cannot be used for BO-RL,
because the environment in the proposed system is a network and is highly dynamic in
nature.
The state as defined in the Sect. 3.B has 4 observation parameters and even a small
change in one of the parameters will result in a different state, increasing the number of
states to infinite. The Q-learning algorithm will start storing the Q-values for each such
 BO-RL: Buffer Optimization in Data Plane Using Reinforcement Learning 359

Fig. 2. (a) FlowChart for CheckQueueDiscSize(), and (b) FlowChart for FeedDelayThroughput()

state-action pair into the Q-table and the convergence will be very slow. The memory
requirement of the agent will also increase as the table size increases. But the only
limitation of the DQN is, because of the deep neural networks, which takes more time
to train compared to Q-Learning. The DQN algorithm leverages the advantages of Deep
Neural Networks (DNN) as a Q-function that takes the state as input and outputs Q-
values for each possible action.
These Q-values are optimized using mean squared error loss function (L), which is
defined as in equation below:
1
L = [r + γ maxa Q(s , a ) − Q(s, a)]2 . (1)
2
Where r is the total reward, s is current state, a is the action taken in current state,
s next state where the agent might move to by taking the action a and γ is the dis-
count factor to balance between immediate and future rewards: Every time, the network
predicts: the Q-values for all the actions, by feed forward pass for the current state s, cal-
culates the maxa Q(s , a ) by doing a feed forward to state s , sets the Q-value for action
a as r + γ maxa Q(s , a ), update the weights using the back propagation. The agent tries
to minimize this error by learning. But convergence is still an issue. To overcome this
problem the authors in [18] have proposed two methods: 1) Experience replay 2) Target
network. In experience replay method, instead of training the network and updating the
weights for every transition, agent stores the last N experiences defined as (s, a, r, s )in a
replay memory. When the number of experiences in the replay memory is exceeding the
minibatch size, agent selects the samples of the experiences randomly from the replay
memory which will help in taking proper actions. This helps in breaking the strong
correlation between consecutive samples and reduces the variance of the updates.
360 M. Sneha et al.

3.4 The Working of RL Agent and the Network Environment

The RL agent interacts with the Network environment E in a sequence of discrete
steps = 1,2,3.... At each time step t, the RL agent selects an action at ∈ A(s) from the
set of action space on the given state st ∈ S. Reward rt ∈ R is received after taking
the action at , and observes a new state st+1 . BO-RL implements DQN algorithm with
replay memory as the RL agent’s algorithm.
At each time step, the state along with the reward are inputs to the agent. As an
output, the agent returns one of the three possible actions (increase/decrease/maintain
same Qsize). The higher probability action is taken on the router under observation in
the network.

3.5 The Process of Training

The DNN is built using two hidden layers with 128 neurons for each layer. ReLU [19]
is used for the hidden layers and softmax is used for the output layer as activation
functions. The network weights are gradually learnt as the DNN gets trained. Adam
Optimizer [20], an optimization algorithm which is an extension to stochastic gradient
descent is used to make the learning faster. Different hyper parameter of the RL agent
are listed in Table 1.

Table 1. Hyper parameters of the RL agent

Parameter Brief description

ε Percentage you want to explore (selecting a random
action) or exploit (selecting an action with max
future reward)
Epsilon min Minimum ε value
Epsilon decay Rate at which ε is decreased from 1.0 to
Epsilon min
Batch size Replay memory batch size
Gamma (γ ) Discount factor

For every time step t the RL agent learns the queue size in n number of episodes
and each episode has m number of steps, so totally in n × m number of steps before
an action is taken. Game over condition should be defined, so that the RL agent stops
learning infinitely.

3.6 The Action at

The RL agent can take one of the actions( Increase/decrease/remain the same). The out-
put layer of the DNN assigns probabilities to different actions and the Softmax function
picks the action with higher probability. The selected action will be implemented back
on the router in the network environment.
 BO-RL: Buffer Optimization in Data Plane Using Reinforcement Learning 361

Every time the future actions which an RL agent takes are influenced by the reward
which it had earned by taking the previous action. While choosing the action RL agent
must decide whether to take the action randomly (exploration) or to take the same action
which had got better reward in the past (exploit). This is a difficult choice and depends
on the environment as well. Greedy policies can be used to balance between exploration
and exploitation. BO-RL implements decaying ε – greedy policy to choose between
exploration and exploitation while choosing the actions. After each time step t, ε is
updated using the equation below:

ε = ε × E psilon decay. (2)

3.7 The Reward rt

After the RL agent takes an action, it waits for the next state st+1 to see the impact on
the network. This impact is evaluated by a reward function to optimize the queue size
further to achieve better delay and throughput. The reward can be on a single parameter
or can be cumulative based on the requirement. A penalty will be given if the delay is
higher and the throughput is decreased. The RL agent always tries to achieve higher
rewards by which it automatically achieves better delay and throughput. The Gamma
(γ ), is used to balance between immediate and future reward; value near to 0.1 for
future rewards and value near to 0.9 for immediate rewards. Since BO-RL is an online
learning setting, only the immediate rewards have been considered for taking the action
decisions.

4 System Implementation
4.1 Implementation Details
BO-RL is implemented using four main components;Environment Network, Ope-
nAIGym, ns3-gym and RL agent as shown in Fig. 3.
Environment Network: Environment Network is implemented using ns3 network sim-
ulator [15]. The simulation scenario script contains the internet topology and serves as
the environment for the RL agent. The procedures (Queue size Monitoring Function and
Delay, throughput monitoring function ) in the script collects the observation parame-
ter data as defined in Sect. 3 and sends them to the RL agent through the Environment
gateway. The procedure Queue size Monitoring Function executes the received action
from the RL Agent. All the routers in the topology built, implement the DropTail queue
as the device queue. The traffic control layer between the netdevice and the IP layer
implements FIFO QueueDisc [15], which is manipulated by the RL Agent.
OpenAIGym: gym is a toolkit for developing RL algorithms. It is a collection of envi-
ronments, which can be directly used to implement RL algorithms. Custom environ-
ments can also be added to this library according to the requirements of the problem
definition. gym is independent of the structure of the RL agent implemented. The main
purpose of gym is to have a standardized interface allowing to access the state and
execute actions in the environment.
362 M. Sneha et al.

Fig. 3. System block diagram

ns3-gym Middleware: It interconnects ns-3 simulator with OpenAI Gym framework.

It provides communication between the components over the ZMQ sockets (ZeroMQ is
a library used to implement messaging and communication between two applications)
using Protocol Buffers (library for serialization of messages). ns3-gym has two parts;
Environment Gateway and Environment Proxy. The Environment Gateway created by
OpenGymGateway object resides inside the simulator and is responsible for:
1. Collecting environment state into structured numerical data as RL agent would
expect using callback functions GetObservationspace(), GetAction-Space(), GetO-
bservation(), GetReward, GetGameOver()
2. Translating the numerical actions received from the RL agent to proper function
calls with proper arguments using the callback function ExecuteActions()

Environment Proxy created by the standard Gym::make(‘ns3-v0’) receives environ-

ment state and exposes it towards the RL agent through the Gym API written in Python.
The Environment Proxy translates the gym function calls into messages and sends them
towards the Environment Gateway over ZMQ socket.
RL Agent: The RL agent is implemented with Python 3.5 using TensorFlow [21]
library for Deep Learning Networks.

4.2 Experimental Setup

Network Configuration and Flow Generation: In order to evaluate the effect of BO-
RL, network topologies are created on NS3. Table 2 summarises the parameters and
their values used for traffic generation and the RL agent implementation.
TCP flows are generated using the BulkSend application of ns3 with randomized
senders and receivers. The generation of flows follow a Poisson distribution while the
start time follows an exponential distribution. The choice of initial queue size is impor-
tant to allow the RL agent perform reasonably. The initial queue size was set to 300
packets that does not unnecessarily degrade the performance of the RL agent after per-
forming preliminary experiments with different values between the minimum of 100
packets and maximum of 500 packets. If the initial buffer size is too small, it causes
 BO-RL: Buffer Optimization in Data Plane Using Reinforcement Learning 363

Table 2. Simulation parameter setting

Parameter Values
Max size of the queue on all the routers 500 packets
Total size of data to transmit over TCP 1 MB
Total simulation time for testing 10 s
Agents observation time interval t 2 ms
Exploration rate ε 1.0
Epsilon min 0.01
Epsilon decay 0.999
Batch size 32
Discount factor Gamma(γ ) 0.7
Learning rate 0.001
Total number of training episodes n 200
Total number of steps m in each episode n 100

lots of packet drops and the throughput gets affected. If the initial queue size is too
large, it affects the delay.
RL Agent Configuration: The possible actions to be taken by the RL agent are:
1. increase the queue size by 1,
2. decrease the queue size by 1, or
3. maintain the same queue size as previous.
Initially the value of ε is 1.0 so that the RL agent will explore more by taking
random actions and as the time passes the ε is eventually decreased by the epsilon decay
factor to reach 0.01 where the RL agent tries to exploit by picking the actions from its
stored experiences. The game over condition is met, when the RL agent achieves 95%
of the minimum possible delay (link delay). The below equation is used to check the
game over condition:
delay = min delay ∗ 1.05. (3)
The reward function is designed to give more priority to delay and throughput
parameters than packet drops. RL agent gets cumulative rewards/penalty according to
the following conditions.
1. if the cur delay
a. is lesser than the prev delay a +ve reward
b. is greater than the prev delay a –ve reward
c. is same as the prev delay then a 0 reward
2. if the cur throughput
a. is lesser than the prev throughput a +ve reward
b. is greater than the prev throughput a –ve reward
c. is same as the prev throughput then a 0 reward
364 M. Sneha et al.

5 Evaluation
We evaluate the BO-RL proof of concept implementation based on the latency and
throughput of the end-to-end communication in NS-3 based testbed.

5.1 Choice of Reward Function

Figure 4 plots the CDF of average reward earned by the RL Agent for 200 episodes
during training. The average reward and the standard deviation for different delay and
throughput ratios were 1270 and 156 for 40:60, 1183 and 168 for 50:50, 1079 and 148
for 60:40 respectively. Hence 40:60 reward ratio was considered in the conduction of
all the experiments.

1.00
0.75
CDF

0.50 Ratios
40:60
0.25 50:50
60:40
0.00
600 800 1000 1200 1400 1600
Average Reward

Fig. 4. CDF of average reward for 200 training episodes

5.2 Buffer Optimization in Single Destination Environment

This experiment observes the baseline performance of the BO-RL RL by enabling it
on Router 1 (R1) in the network given in Fig. 5 on the basis of delay and throughput
of the flows from multiple sources to a single destination. We compare the results with
R1 without BO-RL, which works as an ordinary router same as R2. In this setup for
enabling BO-RL, the RL agent receives the observation parameters from R1 and Desti-
nation 1 (D1). A total of 16 flows were generated by the sources to D1 to transfer 1 MB
data in each flow.

Fig. 5. The network configuration where the link between R1 and R2 is supposed to be congested
due to the concurrent flows from 4 Sources to 1 Destination

Figure 6(a) plots the CDF of the average end-to-end delay of packets from the
sources to D1 through the corresponding routers, and also shows that R1 with BO-
RL successfully reduced the overall latency compared with R1 without it. The average
 BO-RL: Buffer Optimization in Data Plane Using Reinforcement Learning 365

delay and its standard deviation for packets through R1 with BO-RL were 53.43 ms and
9.780 ms respectively, which out perform those of R1 without BO-RL 64.264 ms and
19.347 ms. The maximum latency with BO-RL was also significantly shorter than the
case without BO-RL.
Figure 6(b) plots the CDF of the end-to-end throughput of flows observed in the
same setup. The average throughput and the standard deviation were 2.215 Mbps and
1.534 Mbps with BO-RL, and 1.940 Mbps and 1.113 Mbps without BO-RL respec-
tively.

1.00 1.00
0.75 0.75
CDF

CDF
0.50 0.50
Router Router
0.25 0.25
R1 with BO−RL R1 with BO−RL
0.00 R1 without BO−RL 0.00 R1 without BO−RL
50 60 70 80 90 0 1 2 3 4 5
Average Delay in Milliseconds Average Throughput in Mbps
(a) (b)

Fig. 6. (a) Comparison of CDF End-to-end Average delay at D1 for the flows routed through R1
without BO-RL and R1 with BO-RL, and (b) the corresponding throughput comparison

Figure 7(a) shows the queue size variation, and the number of packets dropped at
R1 because of queue limits exceeded, and Fig. 7(b) the corresponding rewards gained
by RL agent in the same test phase. Figure 7(a) shows that the queue size is decreas-
ing irrespective of packet drops observed at R1 since the throughput and delay is still
improving and the reward gained is increasing as shown in Fig. 7(b). This is because of
the choice of RL agent’s reward function is more biased towards improving the delay
and throughput rather than decreasing the packet drops.

(a) (b)

Fig. 7. (a) Dynamic adjustment of queue size Vs number of packet dropped at R1 by RL agent,
and (b) the corresponding reward gained
366 M. Sneha et al.

5.3 Performance of BO-RL with Multiple Destinations and Varying Number of

Feedback Sources
This evaluation item examines whether the performance of BO-RL improves if more
receivers provide feedback to the RL agent. RL agent receives the observation parame-
ters from R1 with BO-RL and more than one destinations, for dynamically optimizing
the queue size at R1 as shown in Fig. 8.
Tables 3 and Table 4 compare the results of average end to end delay and throughput
of flows between the R1 configuration with or without BO-RL. 64 flows to transfer
1 MB Data each were generated with randomized sources and destinations so that each
destination receives up to 16 flows.

Fig. 8. The network configuration with 4 Sources and 4 Destinations

In this configuration, the network congestion on the link between R1 and R2 is

severer than the previous experiment due to the increased number of flows. The end-
to-end delay has been significantly reduced regardless of the number of receivers to
provide feedback to the RL agent. However, this experience did not clarify the relation-
ship between the number of feedback receivers and the ratio of delay reduction. On the
other hand, average throughput slightly degraded in this experiment.

Table 3. Delay analysis with and without Table 4. Throughput analysis with and with-
BO-RL out BO-RL

R1 without The number of feedback sources R1 without The number of feedback sourcesin
BO-RL in addition to R1 with BO-RL BO-RL addition to R1 with BO-RL
1 2 3 4 1 2 3 4
StdDev 3.558 3.667 1.912 2.543 1.395 StdDev 0.459 0.281 0.364 0.374 0.328
Max 111.442 84.135 85.074 78.392 83.127 Max 1.937 1.048 1.454 1.657 1.430
Min 97.966 69.436 77.735 69.089 78.247 Min 0.092 0.120 0.107 0.145 0.125
Avg 105.084 74.055 81.641 74.584 81.354 Avg 0.394 0.380 0.363 0.344 0.363
Med 105.613 73.265 81.466 74.846 81.729 Med 0.230 0.262 0.230 0.254 0.253

5.4 Limitations and Potential Improvement

5.4.1 Network Topology and Traffic Pattern
The experiment was performed in a very small and simple network, and only TCP
traffic was generated. While BO-RL shown the benefit of dynamic buffer optimization,
it’s benefit should also be explored in larger network typologies with variety of traffic
patterns and applications. The number of BO-RL router can also be increased so that
the buffer size optimization can happen in many parts in the network.
 BO-RL: Buffer Optimization in Data Plane Using Reinforcement Learning 367

5.4.2 Number of Observation Parameters and Choice of RL Agent Algorithm

Our system used the following 4 parameters: the average queue size, the number of
packets dropped from queue, the average throughput, and the average latency were
used as observation parameters. Timestamp, which was used to measure the end-to-end
per-packet delay, was obtained using the feature of NS3. In the real world networking
scenario, precisely measuring one-way delay is not easy. However, there are some tech-
niques [22–24] can be used to measure the one-way delay in the real world deployment.
The observation parameters can be increased to feed more feedback to the RL agent,
such as packet size, inter-packet arrival time, 5-tuple information (source/destination
IP addresses and port numbers, protocol), etc., with DQN extension for the RL agent
algorithm like Double DQN [25], Prioritized Experience Replay [26], Dueling DQN
[27], Noisy Nets for Exploration [28].

5.4.3 Dependency on End User Feedback

In the current BO-RL implementation, the RL agent collects feedback from destination
hosts in the network. While this mechanism works in NS-3 environment, it’s not feasible
in the real network. In order to complete the BO-RL operation without depending on the
interaction with destination hosts, the network needs to support measurement, feedback
collection and reward calculation. Software Defined Networking (SDN) [29], where
SDN Controller can collect the observation parameters directly or indirectly from the
switches, is a promising approach.

6 Conclusion

This paper proposed BO-RL, a use case for Reinforcement Learning based technique on
dynamic buffer optimization in data plane aiming for improving both end-to-end delay
and throughput in the network. We conducted the preliminary evaluation of the PoC
implementation of BO-RL, and observed that the use of BO-RL introduced the signif-
icant improvement to both end-to-end delay and throughput by 17% and 12% respec-
tively compared with the router without BO-RL when the RL agent collects the feed-
back from the router and only one destination of the traffic. However, if the RL agent
collects feedback from multiple destinations, while a minimum of 22.30% improve-
ment was observed in the end to end delay, the throughput improvement was not so
significant. Considering the size of network and traffic patterns used in the experiment
setup, as well as the observation parameters using that the RL agent was developed,
BO-RL still has room for the further improvement as discussed in Sect. 5.4 for future
work.

References
1. Pham, T.A.Q., Hadjadj-Aoul, Y., Outtagarts, A.: Deep reinforcement learning based qos-
aware routing in knowledge-defined networking. In: Qshine 2018 - 14th EAI International
Conference on Heterogeneous Networking for Quality, Reliability, Security and Robustness,
Ho Chi Minh City, Vietnam (2018)
368 M. Sneha et al.

2. Chiariotti, F., Kucera, S., Zanella, A., Claussen, H.: Analysis and design of a latency control
protocol for multi-path data delivery with pre-defined QoS guarantees. IEEE/ACM Trans.
Netw. 27(3), 1165–1178 (2019)
3. Yao, H., Yuan, X., Zhang, P., Wang, J., Jiang, C., Guizani, M.: Machine learning aided
load balance routing scheme considering queue utilization. IEEE Trans. Veh. Technol 68(8),
7987–7999 (2019)
4. Kong, Y., Zang, H., Ma, X.: Improving TCP congestion control with machine intelligence.
In: NetAI 2018: ACM SIGCOMM 2018 Workshop on Network Meets AI & ML, Budapest,
Hungary, 24 August 2018 (2018)
5. Alizadeh, M., et al.: Data center TCP (DCTCP). ACM SIGCOMM Comput. Commun. Rev.
40(4), 63–74 (2010)
6. Vamanan, B., Hasan, J., Vijaykumar, T.: Deadline-aware datacenter TCP (D 2 TCP). ACM
SIGCOMM Comput. Commun. Rev. 42(4), 115–126 (2012)
7. Wu, H., Feng, Z., Guo, C., Zhang, Y.: ICTCP: incast congestion control for TCP. In: Pro-
ceedings ACM CoNEXT (2010)
8. Wilson, C., Ballani, H., Karagiannis, T., Rowtron, A.: Better never than late: Meeting dead-
lines in datacenter networks. In: Proceedings ACM SIGCOMM Conference (SIGCOMM),
pp. 50–61 (2011)
9. Bouacida, N., Shihada, B.: Practical and dynamic buffer sizing using learn queue. IEEE
Trans. Mob. Comput. 18(8), 1885–1897 (2019)
10. Kim, M.: Deep Reinforcement Learning based Active Queue Management for IoT Networks
(2019). https://doi.org/10.13140/RG.2.2.24361.65126
11. Radwan, A., To, H.T., Hwang, W.J.: Optimal control for bufferbloat queue management
using indirect method with parametric optimization. Sci. Programm. 2016, p. 10 (2016).
Article ID 4180817
12. Handley, M., Raiciu, C., Agache, A., Voinescu, A., Moore, A.W., Antichi, G., Wójcik, M.:
Re-architecting datacenter networks and stacks for low latency and high performance. In:
Proceedings of SIGCOMM 2017, Los Angeles, CA, USA, 21–25 August 2017, P. 14 (2017)
13. Elsayed, M., Erol-Kantarci, M.: Deep reinforcement learning for reducing latency in mission
critical services. In: IEEE Global Communications Conference (GLOBECOM), Abu Dhabi,
United Arab Emirates, vol. 2018, pp. 1–6 (2018)
14. Krishnan, N.N., Torkildson, E., Mandayam, N.B., Raychaudhuri, D., Rantala, E., Doppler,
K.: Optimizing throughput performance in distributed MIMO Wi-Fi networks using deep
reinforcement learning. In: IEEE Transactions on Cognitive Communications and Network-
ing (2019)
15. Nsnam. (n.d.). 3. Accessed 27 February 2021, https://www.nsnam.org/
16. François-Lavet, V., Henderson, P., Islam, R., Bellemare, M.G., Pineau, J.: An introduction to
deep reinforcement learning. In: CoRR (2018)
17. Watkins, C.J.C.H., Dayan, P.: Q-learning. Mach. Learn. 8, 279–292 (1992)
18. Mnih, V., Kavukcuoglu, K., Silver, D., et al.: Human-level control through deep reinforce-
ment learning. Nature 518, 529–533 (2015)
19. Nwankpa, C.E., Ijomah, W., Gachagan, A., Marshall, S.: Activation Functions: Comparison
of Trends in Practice and Research for Deep Learning (2018). arXiv:1811.03378 [cs.LG]
20. Kingma, D.P., Jimmy, B.: Adam: A Method for Stochastic Optimization (2014).
arXiv:1412.6980 [cs.LG]
21. Tensorflow. (n.d.). Acceessed 27 Feb 2021, https://www.tensorflow.org/
22. Hernandez, A., Magana, E.: One-way delay measurement and characterization. In: Third
International Conference on Networking and Services, ICNS 2007, Athens, pp. 114–114
(2007)
 BO-RL: Buffer Optimization in Data Plane Using Reinforcement Learning 369

23. Kumar, R., et al.: End-to-End network delay guarantees, for real-time systems using SDN.
In: IEEE Real-Time Systems Symposium (RTSS), Paris, pp. 231–242 (2017). https://doi.org/
10.1109/RTSS.2017.00029
24. Zhang, T., Liu, B.: Exposing end-to-end delay in software-defined networking. Int. J. Recon-
figurable Comput 2019, p. 12 (2019). Article ID 7363901, https://doi.org/10.1155/2019/
7363901
25. van Hasselt, H., Guez, A., Silver, D.: Deep reinforcement learning with double Q-learning
(2016)arXiv:1509.06461 [cs.LG]
26. Schaul, T., Quan, J., Antonoglou, I., Silver, D.: Prioritized experience replay (2016).
arXiv:1511.05952v4 [cs.LG], Accessed 25 Feb 2016
27. Wang, Z., Schaul, T., Hessel, M., van Hasselt, H., Lanctot, M., de Freitas, N.: Dueling Net-
work Architectures for Deep Reinforcement Learning (2016). arXiv:1511.06581v3 [cs.LG],
Accessed 5 Apr 2016
28. Fortunato, M., Azar, M.G., Piot, B., Menick, J., Osband, I., Graves, A., Mnih, V., Munos,
R., Hassabis, D., Pietquin, O., Blundell, C., Legg, S.: Noisy networks for exploration (2019).
arXiv:1706.10295v3 [cs.LG], Accesssed 9 July 2019
29. Farhady, H., Lee, H.Y., Nakao, A.: Software-defined networking: a survey. Comput. Netw.
81, pp. 79–95 (2015). ISSN 1389-1286, https://doi.org/10.1016/j.comnet.2015.02.014
 Formal Modeling, Verification,
and Analysis of a Distributed Task
Execution Algorithm

Amar Nath1(B) and Rajdeep Niyogi2

1
Sant Longowal Institute of Engineering and Technology, Deemed University,
Longowal, India
[email protected]
2
Indian Institute of Technology Roorkee, Roorkee, India
[email protected]

Abstract. Task execution in a dynamic environment is a challenging

and important problem in multi-robot systems. In such an environment,
the arrival time, location of a task, states and availability of robots are
not known in advance. Recently a distributed algorithm for task execu-
tion in such an environment has been suggested. The distributed algo-
rithm is supposed to be run on different components (robots) of the
distributed system (multi-robot system) in an asynchronous model. It
needs to be formally proved or verified that the algorithm behaves as
expected. In this paper, formal modeling, verification, and analysis of
the distributed algorithm are carried out. SPIN model checker is used
for modeling and verification of the algorithm. We identify some interest-
ing safety and liveness properties of the algorithm and perform extensive
simulations of the algorithm. The verification results ascertain that the
behavior of the algorithm is as expected.

1 Introduction
In an automated environment (e.g., office, workshop, and warehouse), a team
of mobile robots may be used to perform routine tasks like moving (carrying) a
heavy box from one location to another and lifting a heavy object. If a robot r
detects a task at some location, it attends the task and determines the size of
the team needed for its execution. Since r has only the local view of the envi-
ronment and does not know the current states and locations of other robots,
it cannot determine a team of robots available for the task. With such insuf-
ficient information, a team cannot be formed by r, and subsequently, the task
cannot be executed. Thus, the necessary information has to be acquired by r by
communicating with other robots. This necessitates the design of a distributed
algorithm for task execution in such a dynamic environment.
A distributed approach for task execution in a dynamic environment is pre-
sented in our previous work [1], where a communicating automata based model

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 370–382, 2021.
https://doi.org/10.1007/978-3-030-75100-5_32
 Formal Modeling, Verification, and Analysis of a Distributed Task 371

is used to capture the behavior of robots. We use a distributed algorithm corre-

sponding to the automata based model in this paper. A distributed algorithm is
supposed to be run on different components (robots) of the distributed system
(multi-robot system) in an asynchronous model. It needs to be formally proved
or verified that the algorithm behaves as expected.
Several model checking tools have been developed (e.g., SPIN [2],
UPPAAL [3], MCMAS [4], and PRISM [5]), to formally verify the system under
study. The tools verify certain properties and accept properties specification and
design models constructed using only specific languages. SPIN accepts proper-
ties specified in linear temporal logic (LTL), and the system design is modeled
usingPromela. MCMAS accepts properties specified using CTL (Computation
Tree Logic) and ATL (Alternating Time Logic) and the model is constructed
using ISPL (Interpreted Systems Programming Language).
There exists other distributed approaches [7–9] for team formation. However,
these works have not considered formal verification of the algorithms. SPIN
model checker has been suggested to find multi-agent plans in [10].
In this paper, formal modeling, verification, and analysis of the distributed
algorithm [1] are carried out using SPIN model checker [2]. Some important
safety and liveness properties of the algorithm [1] are identified and verified to
ensure the correctness of the algorithm.
The rest of the paper is structured as follows. The distributed algorithm for
task execution is given in Sect. 2. The formal modeling is carried out in Sect. 3.
The formal verification results and analysis are given in Sect. 4. Section 5 gives
the conclusion.

2 Distributed Algorithm for Task Execution

The distributed algorithm corresponding to the communicating automata based

model [1] is given below.
The algorithm consists of two functions send and receive shown in
Algorithm 1 and 2 respectively. The agents are assumed to be autonomous and
collaborative, and whenever possible they express their willingness to be part of
a team. The states of an agent are IDLE, ANALYZE, BUSY, and PROMISE. IDLE
state means that an agent is not engaged in any kind of activity; BUSY state
means that an agent is engaged in an activity; ANALYZE state means that an
agent has initiated the team formation process; and PROMISE state means that
an agent has expressed its willingness to be the part of a team. The agent who
initiates the team formation process is called initiator and the other agents avail-
able in the environment are referred to as non-initiator. The messages are of the
following types: Request, Engaged, Willing, Not-required, and Confirm. Whenever
an agent finds a complex task (that needs multiple agents) at some location l
and time t, it becomes the initiator and executes the send function.
In the send function, an initiator i broadcasts a Request message, and waits
for some time, say Δ. It is assumed that broadcast would be within a cer-
tain communication range which need not necessarily cover the entire working
372 A. Nath and R. Niyogi

environment and Request would be delivered only to the agents present in the
environment at that moment in time. A non-initiator j, who has the necessary
skills, will send either a Willing or an Engaged message if its state is IDLE or
PROMISE respectively. Otherwise, it will ignore the Request message. The ini-
tiator increases its counter c when it receives a Willing message. After Δ time
has elapsed, i checks if there are enough agents available to form a team (c ≥
k − 1). If yes, i selects the team with minimum cost and sends Confirm message
to the members of the team and sends a Not-required message to (c − (k − 1))
agents, if any (line 5–9, Algorithm 1). The selection of a team is done using a cost
function [1]. If no, i sends a Not-required message to all c agents who expressed
their willingness to help. Also, i will change its state from ANALYZE to IDLE or
PROMISE depending on its queue status.

Fig. 1. Send function of i

The receive function of an agent is given in Algorithm 2. The computations

are done based on the current state that may be IDLE (line no. 17–24), PROMISE
(line no. 25–42), BUSY (line no. 12–16 and 43–47), and ANALYZE (line no. 2–
10). Within a state, the type of message is checked and appropriate actions
are taken. For example, when state is IDLE, if a Request message is received,
it becomes PROMISE, the identifier of the sender is enqueued, and f lag is set
 Formal Modeling, Verification, and Analysis of a Distributed Task 373

Fig. 2. Receive function of j

to true; all these actions are done atomically (denoted by . . .). Now the agent
sends a Willing message to the sender (initiator) and f lag is set to false (line
17–19 Algorithm 1).

3 Formal Verification of the Distributed Algorithm

In this section, we present a formal verification process of the distributed algo-

rithm given in Sect. 2. The key concern of the algorithm for task execution in
the dynamic environment is the communication. Inter-agent communication is
required to ensure the proper coordination between agents/robots for efficient
mission execution in a dynamic environment. Linear temporal logic is used for
modeling and verifying the validity of the algorithm.
374 A. Nath and R. Niyogi

3.1 Formal Modeling and Verification of the Algorithm Using SPIN

Model Checker
We need to ascertain whether the distributed algorithm satisfies some properties;
for example, the initiator may send Conf irm messages only when it has received
a sufficient number of W illing messages; if an agent gets a Conf irm message in
the P romise state, then it goes to the Busy state; if an agent broadcasts request
message then, it must eventually receive either W illing or Engaged messages;
if an agent has sent W illing message, then it must eventually receive either
Conf irm or Not-Required message. We have used the SPIN model checker for
the verification of the algorithm.

3.1.1 LTL Model Checking

Model-checking is a technique for verifying that finite-state systems satisfy tem-
poral logic specifications [11,12]. The process of algorithm verification via model
checking has three steps: (i) Modeling of the system under consideration. (ii)
Specification of desirable properties of the system. (iii) Verifying whether the
system satisfies the specifications.
First, the system’s design is encoded in some formal modeling language (e.g.,
Promela) to obtain a model of the system. Second, the properties that the model
should satisfy must also be stated in an unambiguous, formal notation (e.g.,
LTL). Third, to verify the system, a tool (e.g., SPIN model checker) will take
as input the model and the specification (e.g., properties written in LTL) and
determines whether or not the system meets the specification.
Linear temporal logic (LTL)
For a set of atomic propositions (AP ), an LTL formula consists ofBoolean
operators ∧ (and), ¬ (negation), and temporal operators,  (next), (until).
Other temporal operators include  (always) - a property is satisfied now and
forever into the future and ♦ (eventually) - a property is satisfied at some time
in the future.
Syntax: LTL formulas over the set AP are formed  according to the following
grammar: ϕ :: = true | p | ϕ1 ∧ ϕ2 | ¬ϕ | ϕ | ϕ1 ϕ2 where:

• p is an atomic proposition (p ∈ AP).

• ϕ means ϕ is true in the next moment in time. 
• ♦ϕ means ϕ is true eventually in future moments, ♦ϕ = true ϕ
• ϕmeans ϕ is true in all future moments, ϕ = ¬♦¬ϕ
• ϕ1 ϕ2 means that ϕ1 is true until ϕ2 is true.

By combining the temporal modalities ♦ and , new temporal modalities are

obtained as follows: ♦ϕ means “infinitely often ϕ” and ♦ϕ means “eventually
forever ϕ”.
Semantics of LTL: The semantics of an LTL formula ϕ is defined by a sat-
isfaction relation, |=⊆ (N → 2p ) × N × Φ, where Φ is set of LTL formulas. For
σ : N → 2p , j ∈ N and ϕ ∈ Φ, we write (σ, j) |= ϕ to mean (σ, j, ϕ) ∈|=. The
 Formal Modeling, Verification, and Analysis of a Distributed Task 375

terminology is as follows: ϕ holds at position j of σ; σ satisfy ϕ at j; and σ

satisfies ϕ at σ(j). For more details we refer to [12].
LTL Model Checking: In order to check if an LTL formula ϕ holds in a tran-
sition system S, first a nondeterministic Büchi automata (NBA) is constructed
for the negation of ϕ . Now construct a product transition system of S and the
NBA, and then check the emptiness of the product automaton. For the details
of this construction, we refer to [12].

3.1.2 Modeling of the Algorithm in Promela

The modeling of the distributed algorithm is done in Promela. Promela programs
consist of processes, channels, and variables. In Promela mtype is used for defin-
ing symbolic names of numeric constants. As mentioned in the Algorithm 1, we
have abstracted five types of messages, declared as follows:
mtype = {REQUEST, WILLING, ENGAGED, CONFORM, NOTREQ};
which is functionally equivalent to the sequence of macro definitions:
#define REQUEST 5; #define WILLING 4; #define ENGAGED 3;
#define CONFORM 2; #define NOTREQ 1; Similarly the states of the agents are
defined:
#define IDLE 1; #define READY 2; #define PROMISE 3; #define BUSY 4;
To transfer the messages among active process Promela uses the channel.
The keyword chan is used to declare the channel. The channel declaration is
of two types: local and global. In the channel declaration, its capacity can be
initialized either by zero or some constant value. The messages are stored in first-
in-first-out (FIFO) manner. In this paper, agents (processes) communicate in an
asynchronous manner, and the channel capacity is defined by MESSAGE BUFFER.
This means that a channel, chan, can store at most MESSAGE BUFFER messages.
Each message is defined to have two fields of type mtype and int.
chan to agent [N+1] = [MESSAGE BUFFER] of {mtype, int};
where, #define MESSAGE BUFFER 4
The total number of agents in the environment is N and the number of other
agents k required to execute a joint action are defined as:
#define N 6; #define k 4
The processes in Promela are created by using keyword proctype and then
run statement is used to initialize the processes. It is, therefore, possible to create
multiple processes simultaneously. All processes run asynchronously, and each
process may take the finite number of parameters in SPIN. In this paper, the
to agent channel provides the communication means between different agents.
Each agent (process) is modeled in Promela as given below. A portion of the
Promela model (code) for the algorithms discussed in Sect. 2, is given below.
376 A. Nath and R. Niyogi

1 mtype = {REQUEST, WILLING, ENGAGED, CONFORM, NOTREQ, SELF1, SELF2};

2 /* TODO: Define states here */
3 #define IDLE 1; #define READY 2
4 #define PROMISE 3; #define BUSY 4
5 #define MESSAGE_BUFFER 4; #define N 6
6 #define K 4
7 chan to_agent [N+1] = [MESSAGE_BUFFER] of
8 {mtype, int};
9 hidden int idx; /* used in for loops */
10sent_will, sent_eng, sent_conf, sent_nreq;
11 bool received_req, received_will, received_eng, received_conf,
received_nreq;
12te_ready, state_promise, state_busy; /
13 bool at_Loc_Init;
14 #define sufficient_willing (c>=K)
15 int c;
16 ltl p0{[](sent_conf -> sufficient_willing)}
17 ltl p4{[](received_conf -> <> at_Loc_Init)} /*L3*/
18 proctype agent(int i) { /* to define the agent */
19 chan to_me = to_agent[i];
20 xr to_me;
21 byte state = IDLE;
22 int from, to;
23 bool is_willing[N + 1];
24 chan Q = [MESSAGE_BUFFER] of {int};
do
25 ::to_me?REQUEST(from)->received_req=true;
26 if
27 :: state == IDLE ->state = PROMISE; state_promise = true;

The do loop used here is an infinite loop in which a nondeterministic choice is

made between all instruction sequences starting with ::. Promela’s do ... od
provides a form of guarded commands in a CSP-like manner [13], where guards
restrict the set of sequences that can be chosen.
In the above code, there are 5 outer choices, since there are 5 different mes-
sages. We now explain the working of the first outer choice. When an (non-
initiator) agent i receives a Request message from an initiator agent (whose
identifier is f rom) via channel to me (the guard to me[i]?REQUEST(from)), if
the state is Idle, it changes the state to P romise and sends a W illing message
via channel to me to the initiator; if the state is not Idle it sends an Engaged
message via channel to me to the initiator.
 Formal Modeling, Verification, and Analysis of a Distributed Task 377

4 Experimental Results and Analysis

The verification results with identified safety and liveness properties are dis-
cussed here. Safety and liveness are two important kinds of properties provided
by all distributed systems. Safety property asserts that “nothing bad happens” in
the system. We identified some safety properties of the algorithm given in Algo-
rithm 1 and 2. The default safety property check of SPIN ensures the absence of
deadlock, and unreachable code. SPIN model checker checks a safety property
by trying to find a trace leading to the “bad” thing and if there is no such trace
found, it means that the property is satisfied. Liveness property asserts that
“something good will eventually happen”.

4.1 Safety Properties

We have identified the following safety properties:

Safety 1: Initiator may send Conf orm messages only when it has received
sufficient number of W illing messages. This property is expressed in Promela
(SPIN) as: [ ](sendConform -> c≥(k-1))

Safety 2: If an agent is in Promise state and it receives Conform message,

then its next state will be Busy.
[ ]((promise && receiveConform) -> X busy)
We have performed several experiments with the varying number of agents
N and the number of agents required k for different safety properties (Safety1,
Safety2). We have performed several experiments and the results of the runs
(like the one shown above) are summarized in Table 1 and Table 2 for k = 2 and
k = 3 respectively with varying number of agents N , for the default property.

Table 1. Results with varying N and k = 2 with default safety property of the SPIN

# agents (N) No. of states Memory (MB) Time (sec.) Verified?

3 22,127 71.422 0.073 Yes
4 310,312 184.313 1.31 Yes
5 3,217,309 2011.656 24.3 Yes
6 3,443,944 2011.266 26.6 Yes
7 2,999,766 2010.875 27.9 Yes
8 2,640,990 2010.875 29.6 Yes
9 2,381,396 2010.484 30.2 Yes
10 2,073,344 2011.266 30.3 No
11 1,893,540 2010.875 29.8 Yes
... ... ... ... ...
20 986,432 2010.484 27.9 Yes
378 A. Nath and R. Niyogi

Table 2. Results with varying N and k = 3 with default safety property of SPIN

# agents (N) No. of states Memory (MB) Time (sec.) Verified?

4 5,048,791 2011.266 18.6 Yes
5 4,213,500 2010.047 20.3 Yes
6 3,444,635 2011.656 26.9 Yes
7 2,999,164 2010.484 28.9 Yes
8 2,640,990 2010.875 29.4 Yes
9 2,382,352 2011.266 29.8 Yes
10 2,073,760 2011.656 29.8 Yes
11 1,893,160 2010.484 29.3 Yes
... ... ... ... ...
20 98,7027 2011.656 32.5 Yes

• The time, memory and the total number of states explored are low with the
small value of N and k.
• If the value of N is gradually increased and k is kept fixed (k = 2), the
time, memory and states explored also increases with a rapid rate up to
some extent, i.e. N = 6. After N = 6, we found that the value of time is
still increasing. Now, after N = 6, the states explored has started decreasing
slowly and memory used becomes almost constant.
• The value of time start decreasing after N = 10.

The evaluation of results indicates that the time, memory, and the number of
states explored during verification depend on the total number of agents present
in the environment. After some number of agents (N = 6) available in the
environment time, memory and states explored start decreasing and become
constant because of the high availability of idle agents in the environment, which
makes (c ≥ (k − 1)) easily satisfied, and no further work is required. For a fixed
value of N , when we vary k, the running time and states explored almost remain
the same.
The verification results corresponding to the safety properties (Safety1,
Safety2), given above, for k = 2 are given in Tables 3 and 4 respectively. We
have also performed several experiments by varying k for a fixed value of N for
these properties. The pattern is similar to that obtained for the default property.
 Formal Modeling, Verification, and Analysis of a Distributed Task 379

Table 3. Verification results for safety property (Safety 1) with k = 2

# agents (N) No. of states Memory (MB) Time (sec.) Verified?

3 22,127 71.179 0.031 Yes
4 310,312 182.898 0.437 Yes
5 2,073,075 1023.914 3.760 Yes
6 1,689,900 1023.914 4.070 Yes
7 1473750 1023.914 4.470 Yes
8 1306725 1023.914 4.530 Yes
9 1169175 1023.914 4.510 Yes
10 1021797 1023.914 4.370 Yes

Table 4. Verification results for safety property (Safety 2) k = 2

# agents (N) No. of states Memory (MB) Time (sec.) Verified?

3 22,127 71.179 0.031 Yes
4 310,312 182.898 0.421 Yes
5 2,073,075 1023.914 3.76 Yes
6 1,689,900 1023.914 4.41 Yes
7 1,473,750 1023.914 4.70 Yes
8 1,473,750 1023.944 4.56 Yes
9 1,169,175 1023.914 4.59 Yes
10 1,021,797 1023.914 4.38 Yes

4.2 Liveness Properties

We also have identified the following liveness properties:
Liveness 1: If an agent broadcasts Request message then, it must eventu-
ally receive either W illing or Engaged messages. This property is expressed
in Promela (SPIN) as:
[ ](sendRequest -> <> (receiveWilling || receiveEngaged))

Liveness 2: If an agent has sent W illing message then it must eventually receive
either Conf orm or Not-Required message. This property is expressed in Promela
(SPIN) as:
[ ](sendWilling -> <> (receiveConform || receiveNotrequired))
Liveness 3: If an agent has received a Conf orm message, then it will be even-
tually at the location of the initiator. This property is expressed in Promela
(SPIN) as:
[ ](receiveConform -> <> (at loc Init))

The verification results corresponding to the liveness properties (Liveness 1,

Liveness 2, Liveness 3), for k = 2 are given in Tables 5, 6 and 7 respectively.
380 A. Nath and R. Niyogi

Table 5. Verification results for liveness property (Liveness 1) with k = 2

# agents (N) No. of states Memory (MB) Time (sec.) Verified?

3 25,781 72.594 0.11 Yes
4 407,268 221.813 2.25 Yes
5 4,180,734 2010.875 29.60 Yes
6 3,424,694 2011.656 26.40 Yes
7 2,979,835 2010.875 28.50 Yes
8 2,641,520 2011.266 25.10 Yes
9 2,371,908 2010.875 22.00 Yes
10 2,062,962 2010.875 21.70 Yes

Table 6. Verification results for liveness property (Liveness 2) with k = 2

# agents (N) No. of states Memory (MB) Time (sec.) Verified?

3 23,361 71.813 0.078 Yes
4 352,623 200.719 1.58 Yes
5 4,180,736 2010.875 26.90 Yes
6 3,424,007 2011.266 26.00 Yes
7 2,979,834 2010.875 28.10 Yes
8 2,640,460 2010.484 29.40 Yes
9 2,371,432 2010.484 29.00 Yes
10 2,063,376 2011.266 27.90 Yes

Table 7. Verification results for liveness property (Liveness 3) with k = 2

# agents (N) No. of states Memory (MB) Time (sec.) Verified?

3 22,685 71.422 0.086 Yes
4 320,782 188.219 1.40 Yes
5 4,180,736 2010.875 25.20 Yes
6 3,424,008 2011.266 26.70 Yes
7 2,980,432 2011.266 29.60 Yes
8 2,640,990 2010.875 30.00 Yes
9 2,371,908 2010.875 29.80 Yes
10 2,062,962 2010.875 30.19 Yes

5 Conclusion
In this paper, we considered the problem of formal modeling, verification, and
analysis of a distributed algorithm for task execution in a dynamic multi-robot
environment. We modeled the distributed algorithm model in Promela, and sub-
 Formal Modeling, Verification, and Analysis of a Distributed Task 381

sequently, some important safety and liveness properties were identified and ver-
ified using SPIN model checker. The correctness of the algorithm is thus formally
verified. Extensive experiments with a varying number of agents are performed.
The results are quite encouraging, and it confirms the expected execution of the
algorithm.
The evaluation of results indicates that the time, memory, and the number of
states explored during verification depend on the total number of agents present
in the environment. After some number of agents (N = 6) available in the
environment time, memory and states explored start decreasing and become
constant because of the high availability of idle agents in the environment, which
makes (c ≥ (k − 1)) easily satisfied, and no further work is required.
The memory used and the number of states explored is almost constant
with all values of k and for fixed N = 12. The time taken has increased up to
k = 4, and after k = 4, it starts decreasing and becomes constant. The time,
memory, and the number of states explored depend on the number of agents
(N ) available in the environment, not k. Hence, the number of agents present in
the environment is the deciding factor for verification time, memory and states
explored.

Acknowledgment. The authors thank the anonymous referees for their valuable com-
ments that were helpful for improving the paper. The second author was in part sup-
ported by a research grant from Google.

References
1. Nath, A., Arun, A.R., Niyogi, R.: An approach for task execution in dynamic mul-
tirobot environment. In: Australasian Joint Conference on Artificial Intelligence,
pp. 71–76 (2018)
2. Holzmann, G.J.: The model checker SPIN. IEEE Trans. Softw. Eng. 23(5), 279–295
(1997)
3. Larsen, K.G., Pettersson, P., Yi, W.: UPPAAL in a nutshell. Int. J. Softw. Tools
Technol Transf. 1(1–2), 134–152 (1997)
4. Lomuscio, A., Qu, H., and Raimondi, F. MCMAS: a model checker for the veri-
fication of multi-agent systems. In: International Conference on Computer Aided
Verification, 2009, pp. 682–688. Springer, Heidelberg. https://doi.org/10.1007/978-
3-642-02658-4 55
5. Kwiatkowska, M., Norman, G., Parker, D.: PRISM: probabilistic symbolic model
checker. In: International Conference on Modelling Techniques and Tools for Com-
puter Performance Evaluation, 2002, pp. 200–204. Springer, Heidelberg. https://
doi.org/10.1007/3-540-46029-2 13
6. Tkachuk, O., Dwyer, M.B.: Environment generation for validating event-driven
software using model checking. IET Softw. 4(3), 194–209 (2002)
7. Abdallah, S., Lesser, V.: Organization-based cooperative coalition formation. In:
Proceedings IEEE/WIC/ACM International Conference on Intelligent Agent Tech-
nology, pp. 162–168 (2004)
8. Coviello, L., Franceschetti, M.: Distributed team formation in multi-agent systems:
stability and approximation. In: 51st IEEE Conference on Decision and Control
(CDC), pp. 2755–2760 (2012)
382 A. Nath and R. Niyogi

9. Tośić, P.T., Agha, G.A.: Maximal clique based distributed coalition formation for
task allocation in large-scale multi-agent systems. In: International Workshop on
Massively Multiagent Systems, pp. 104–120. Springer, Heidelberg (2004). https://
doi.org/10.1007/11512073 8
10. Wooldridge, M.: An automata-theoretic approach to multiagent planning. In: Pro-
ceedings of the First European Workshop on Multiagent Systems, Oxford Univer-
sity, pp. 1–15 (2003)
11. Clarke Jr, E.M., Grumberg, O., Kroening, D., Peled, D., Veith, H.: Model Checking.
MIT press, Cambridge (2018)
12. Baier, C., Katoen, J.P.: Principles of Model Checking. MIT press, Cambridge
(2008)
13. Hoare, C.A.R.: Communicating sequential processes. The origin of concurrent pro-
gramming, pp. 413–443 (1978)
 Deep Learning Library Performance Analysis
on Raspberry (IoT Device)

Luciana P. Oliveira1(B) , José Humberto da S. S. Santos2 , Edson L. V. de Almeida1 ,

José R. Barbosa1 , Arley W. N. da Silva3 , Lucas P. de Azevedo3 ,
and Maria V. L. da Silva3
1 IFPB Campus João Pessoa, Av. Primeiro de Maio, 720 - Jaguaribe, João Pessoa, PB, Brazil
[email protected], [email protected]
2 UFCG Campus Campina Grande, R. Aprı́gio Veloso, 882 - Universitário, Campina Grande,
PB 58428-830, Brazil
[email protected]
3 IFPB Campus Itaporanga, R. Projetada, SN - Vila Moco, Itaporanga, PB, Brazil

[email protected],
{lucas.azevedo,maria.silva.7}@academico.ifpb.edu.br

Abstract. The use of machine learning techniques in Internet of Things (IoT)

applications is increasing by the diversity of the IoT scenario. In addition,
machine learning libraries contribute to this growth by promoting stable and easy-
to-manipulate implementations, reducing the need to develop complex algorithms
and bringing researchers’ attention to other end-application activities. However,
library diversity can make it difficult to choose from, since factors such as com-
putational resource use are critical in an IoT scenario where such elements are
limited. This work analyzes the performance, energy consumption and resource
usage of two of the major machine learning libraries on the market, through stan-
dardized applications. It was found that the Pytorch 1.1.0 can be used with lower
consumption.

1 Introduction

The Machine Learning (ML) concept proposes to solve complex and repeating activ-
ities through the use of mathematical models and computational resources, making it
possible to develop computer programs capable of learning a certain task or predicting
a certain result [1].
Artificial Neural Network (ANN) is an ML algorithm widely used to solve many
problems, due to its ability to adapt in many situations and satisfactory results when
compared to other algorithms [2]. This algorithm is inspired by the functioning of the
human brain, using sets of neurons to formulate models capable of associating unknown
inputs to previously known classes (outputs). Despite the complexities, several IoT
applications reported in the literature present ANN as a viable alternative for solving

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 383–392, 2021.
https://doi.org/10.1007/978-3-030-75100-5_33
384 L. P. Oliveira et al.

various problems [3, 4]. However, in applications involving IoT, factors such as manip-
ulation of considerable data volumes, limited resource and low cost components are
considered [2, 5]. Such features require in-depth analysis of which modules to use in
IoT and ANN applications.
Today, several machine learning libraries offer optimized and standardized imple-
mentations of the algorithms [6–9]. Avoiding the need to redo the code and allowing
attention to be directed to the end applications. However, choosing a suitable library for
a given situation can be challenging when considering scenarios as limited resources,
as is often the case in IoT applications [14].
In this context, the main objective of this research is to compare two ML libraries in
Raspberry (IoT device) considering three training scenarios (one linear regression from
real estate data and two classifications with image processing: one in character recogni-
tion and one to identify model images). Therefore, this article presents the measurement
results using answers to the following questions:

• Do any of the libraries have the least resource use in all scenarios?
• What is the relationship between runtime and energy for each library?

In order to answer such questions, this article presents in Sect. 2 concepts of Artifi-
cial Intelligence and IoT to offer support for the understanding of the scenarios that will
be evaluated and the constraints of IoT devices. In Sect. 3, the methodology is described
in term of the components and phases to collect and analyze data that will be used to
compare libraries. In Sect. 4, the following items are presented: results of the processing
consumption, memory, runtime, and power consumption of each library. Finally, Sect. 5
presents conclusions.

2 Background
2.1 Artificial Intelligence, Machine Learning and Neural Networks

Artificial Intelligence (AI) is a concept approached in 1956 at the Dartmouth office,

considering artifacts might be able to learn any way to accomplish a specific task that
would be performed by humans. Despite the seeming fiction theme initially, the concept
received attention of the scientific community, and, in recent decades, with the increase
of computational resources such as processing, volatile memory and storage, it has been
possible to develop new approaches and applications using computers. Today, machine
learning is one from several concepts in Artificial Intelligence [10].
The concept of machine learning currently encompasses a large set of algorithms
and methods as illustrated in Fig. 1(a). The main techniques are based on statistical
models, conditioned repetition algorithms and natural systems modeling.
 Deep Learning Library Performance Analysis on Raspberry (IoT Device) 385

Fig. 1. AI.

Today the large volume of data, produced and made available by IoT scenarios,
drives the development of applications involving ML and Internet of the Things. In this
context, ANN algorithms have been gaining prominence due to the ability to adapt the
most diverse approaches and problems explored [2].
ANN are used in the functioning of the human brain in clusters of connected neurons
that form densely connected graphs, used by model generators by adjusting connections
during learning. Figure 1(b) illustrates an example of ANN structure and an artificial
neuron. The arrangement of layers in an ANN may vary by network type, but the most
common topology is an input layer, N intermediate layers, also calls hidden layers and
an output layer.
The artificial neuron (Fig. 1b) consists of a node of a graph, in which each edge
that connects the vertices (nodes) has a weight that is adjusted during learning. For
model adjustment, each i-neuronium sum the products between the weights (wni ) and
the inputs (xni ) plus an adjustment value (bi ), according to Eq. 1.
m
v j = ∑ wni xni + bi (1)
i=1

Finally, the signal is directed to a normally nonlinear function called the activa-
tion function (ϕ ), which limits the output value of the neuron (vi ) to a specific range,
resulting in a value of (yi ).
During the presentation of a training set for the multilayer perceptron (MLP), the
classification signal runs through the network from the input to the output generating the
result of the classification, then an error signal is generated in the reverse direction. Each
weight is initialized with a random value that is adjusted during the step training. In this
work, the weight adjustment is performed through the Backpropagation algorithm [12].
This algorithm performs the adjustment based on the error signals.
The Backpropagation algorithm uses the errors generated from the difference
between the expected output and the output obtained by the network to optimize the
MLP model [12]. For this, its applied the rule of the derivatives chain is applied to
perform the computation in the reverse direction. Finally, the values of the weights are
386 L. P. Oliveira et al.

updated with the result of the derivation plus a constant denominated rate of learning
(or bias).
Each weight is initialized with a random value that is adjusted during the step train-
ing. During the presentation of a training set for the MLP, the classification signal runs
through the network from the input to the output generating the result of the classifica-
tion. An error signal is generated from the difference between the predicted result and
the obtained result, this signal is used by the Backpropagation algorithm to adjust the
weights [12].
This algorithm and more complex variations such as Convolutional Neural Net-
works and deep learning algorithms are constantly used in cardiology and clinical anal-
ysis such as [13].

2.2 Internet of the Things - IoT

The Internet of Things (IoT) is constituted of smart objects that help automate services,
improve everyday life, and aid decision making. They are characterized by the ability
to collect and transmit data. Some of theses devices processing data to extract informa-
tion or knowledge before to transmit data and such functionalities can be executed by
artificial intelligence techniques (neural network, convulsion neural networks, decision
tree and others).
In addition, they have Internet connectivity but have low computing power and some
have a limited (non-rechargeable) battery life. For example, the ESP32 is a widely used
IoT microcontroller with low cost, WiFi and Bluetooth as transmission capabilities, a
coprocessor and advanced power modes.
On the other hand, Raspberry Pi is considered a minicomputer that has a modu-
lar structure. Its main features are portability and programmability, which provides its
use in different types of applications. This IoT device can operate on Linux operating
systems such as Raspibian or ubuntu. This minicomputer has a variety of interfaces
allowing the use of different modules as needed and also to execute machine learning
libraries to execute the processing of the data [14].
In this context, the challenge in an IoT environment is to execute artificial intel-
ligence mechanisms, considering the resource constraints of smart objects. This is
because the machine learning and other mechanisms implemented on smart objects
can influence the demand for processing, and hence the power consumption that will be
investigates in this study.

3 Methodology

This work is classified as its approach as a quantitative research, considering that its
results and its samples can be quantified numerically [15]. Regarding its nature, this
work is classified as applied research, since it aims to generate knowledge for practical
application.
During the experiments in Raspberry (in Fig. 2(c)), some codes with the python
language (sourceforge.net/projects/aina2020/files/deep learning/) were run to analyze
the deep learning libraries. These codes allow the automation to extract the execution
 Deep Learning Library Performance Analysis on Raspberry (IoT Device) 387

time (real), the memory utilization and energy consumption. The collect procedure was
executed ten times for each scenario and library combination, and then the runtime,
resident memory and energy consumption were averaged.
In terms of energy consumption, a Digital Minipa ET-2231 multimeter was used.
This multimeter is capable of measuring DC and AC voltage across buses and power
lines. In electronic circuits, it can measure current, frequency, duty cycle, capacitance,
test diodes, cable and wire continuity, and measure component resistance. A feature of
this multimeter is its connection via USB enabling communication with the computer,
facilitating the collection of information for the experiment. The multimeter used in the
experiment can be observed in Fig. 2(a).

Fig. 2. Energy measurement

Moreover, a plate in Fig. 2(b) was designed to maintain greater accuracy in mea-
surements. Such a plate makes it possible to interrupt the current, facilitating the mea-
surement process. The plate was made of fiberglass, using thermal transfer engraving,
through the corrosion process with iron perchloride. It consists of 4 connectors: Two-
way KRE post and two alligator clips.

3.1 Deep Learning Library

Several optimized implementations of deep learning algorithms can now be found,

reducing the need for development and allowing attention to be directed to the final
application. Among the most used in the literature is possible to highlight Tensorflow
and PyTorch, according to their official descriptions:

• Tensorflow: It is an open source machine learning platform with a comprehensive

and flexible set of machine learning tools, libraries and resources.
• PyTorch: It is a library optimized for deep learning using GPUs and CPUs.

Among the main criteria for choosing these libraries are the amount of materials
and quality of documentation available on the Internet, active community, stability of
implementations and diversity of applications reported in the literature. Both libraries
support application development involving features such as deep learning techniques,
several neural networks, benchmarking metrics, and so on. But details are presented in
Table 1.
388 L. P. Oliveira et al.

Table 1. Other features: tensorflow e pytorch

Feature Tensorflow PyTorch

Developed by Google Facebook
Creation date Feb. 2017 Oct. 2018
Open source Yes Yes
Main language for use Python Python
Multiple network topologies Yes Yes
Community Bigger Smaller
Graphic views Static Dynamic
Distributed processing Yes Yes
GPU processing Yes Yes
CPU processing Yes Yes

4 Experiment Description
Several applications are proposed in the literature involving concepts of IoT and
Machine Learning [2]. To reduce the scope, this work analyzed two types of applica-
tions: (i) image classification that is usually used in industrial automation processes,
allowing adjustments of mechanical arms, quality control of fruits, in addition to
applications such as identification of fire spots, analysis of animal behavior and etc.
[16, 17, 26]; and (ii) linear regression that is normally used when it is desired to predict
a future value from past and present entries, such as the next stock exchange variation
or the expectation of real estate values in some region.
These applications will be evaluated considering Pytorch 1.1.0, Pytorch 1.4.0 and
TensorFlow 2.3.1. They will be presented in more detail in the following sections.

4.1 Image Classification

Usually a Convolutional type network (CNN) is used to perform the image classifica-
tion. This type of network has the following information as input: a tensor formed by
the number of images x image width x image height x image depth. Then, a mathemat-
ical operation called convolution is applied, which performs the multiplication of the
matrix block (image) by a kernel mask, generating a map of shapes of equal or smaller
size that is directed to the next layer. Similar to a neural network explained previously,
each set of information tied to a pixel corresponds to an input neuron, so an image of
size 100 × 100 pixels generates 10,000 weights between two densely connected lay-
ers. However, through convolution, this value can be reduced according to the modified
kernel, so the same example can result in 400 pesos when using a 5 × 5 size kernel.
In this context, two scenarios were used: classification of clothing, with Fash-
ion Mnist dataset [23], with database available at https://github.com/zalandoresearch/
fashion-mnist, and classification of handwritten digits, with MNIST Dataset [24] avail-
able at http://yann.lecun.com/exdb/mnist. Each scenario has a dataset comprised of
 Deep Learning Library Performance Analysis on Raspberry (IoT Device) 389

60,000 small square 28 × 28 pixel grayscale images of items and the total number of
the test dataset was 10,000 images. In first scenario, the dataset has 10 types of clothing
(0: T-shirt/top, 1: Trouser, 2: Pullover, 3: Dress, 4: Coat, 5: Sandal, 6: Shirt, 7: Sneaker,
8: Bag, 9: Ankle boot). In second scenario, images of handwritten single digits between
0 and 9.

4.2 Linear Regression

The application for classification using numerical data can be performed with a con-
ventional implementation of a neural network, using each column of the database as
a network input. This type of application is one of the most common, as the data and
description of them can be obtained from several sources normally present in IoT sce-
narios. For application, Boston Housing Dataset [25], a database available at http://lib.
stat.cmu.edu/datasets/boston, was used, where it has 14 variables related to the prices
of each residence and the characteristics of its location, relative to the city of Boston,
Massachusetts, USA. Each sample is made up of statistical data on the number of res-
idences within 25000ft2, industrial areas, local crime rates, among others. The imple-
mented application adds the PRICE field, present in target, so that the algorithm can
work on the prices properly.

4.3 Scenarios IoT

A Raspeberry Pi 3B computer (Fig. 2(c)) was used to perform the experiment. This
choice was motivated due to the amount of memory and processing present in the
device, sufficient to execute a complex algorithm such as an ANN. This type of com-
puter is constantly used for IoT project design, home automation, media centers and
others [19], due to the low cost, ease of handling and increased computing resources
compared to commonly used microcontrollers [22].
The version has a Quad Core 1.2 GHz Broadcom BCM2837 64bit processor, 1
GB RAM, attached wirelles chip, 100 Gbits support Ethernet connection, 4 USB 2.0
ports, HDMI connection, micro SD memory card support, display ports Toutch screen.
Depending on an external power supply of 800 mA [20].
Raspbian Buster Lite 2019 operating system was used for installation, access and
execution of the evaluated libraries. This system offers the main tools present in a
conventional operating system, besides the possibility of installing other additional
resources that may be needed [21].

5 Results
The tests were performed with three applications (two image classification and one
linear regression), respectively involving image processing using two CNN network for
image recognition and a MLP network for price prediction using numerical data.
It is important to consider that the tests were performed only for the network train-
ing stage, as this is the stage that usually requires the most time and computational
resources.
390 L. P. Oliveira et al.

Table 2 shows the results of the CNN network execution, an application was used
using the dataset fashion, this is present in the Tensorflow and PyTorch/TorchVision
documentation, the networks in both libraries were configured as similar as possible,
being used 6 hidden layers of 32 neuron, when possible.

Table 2. Scenario 1: main fashion – classification with image processing and using CNN

Metric TensorFlow 2.3.1 Pytorch 1.1.0 Pytorch 1.4.0

Runtime (h:mm:ss) 31:31.75 36:46.98 01:37:35
CPU% 71.39654071 97.81109595 99.06338704
Memory% 72.14316126 23.22992269 22.14530339
Power (W) 4380.996956 5245.388272 12858.9161

Because it involves image processing, this type of network demands a high rate of
CPU consumption. It is possible to observe that both libraries had similar behavior, but
the Pytorch library performed the task with shorter memory consumption.
Table 3 presents the results of other CNN network with 6 hidden layers of 16 neu-
rons, using the dataset MNIST, composed to handwritten digits in image format.

Table 3. Scenario 2: handwritten digits - classification without image processing and using MLP

Metric TensorFlow 2.3.1 Pytorch 1.1.0 Pytorch 1.4.0

Runtime (h:mm:ss) 36:25:84 01:15:32 01:41:29
CPU% 79.84820607 64.09488372 93.0056865
Memory% 69.16402944 31.00126246 29.29963738
Power (W) 5398.766018 9295.956035 13872.23185

Results from the third scenario in Table 4 used fewer resources than scenario 1 and
scenario 2, for both libraries, because it work with numeric datas, more accessible to
processing in major processors, inclusive in Raspberry PI ARM. Also, for every sce-
nario, the tensor flow used the fewest resources.

Table 4. Scenario 3: linear regression using MLP

Metric TensorFlow 2.3.1 Pytorch 1.1.0 Pytorch 1.4.0

Runtime (h:mm:ss) 29:02.19 4:38.21 4:40.64
CPU% 36.12668588 25.4923913 25.46129032
Memory% 31.14899135 29.77681159 30.28853047
Power (W) 3002.291787 435.768093 440.269105
 Deep Learning Library Performance Analysis on Raspberry (IoT Device) 391

6 Conclusion
The presented results of this paper aimed at investigating the characteristics of
measurements in context of IoT, energy and artificial intelligence. This results has
helped to understand the resource consumption of the real experiments in raspberry.
After analyzing each measurement, the linear regression can be used with lower energy
consumption.
The results of this paper suggest that Pytorch 1.1.0 can be used with lower consump-
tion (memory and CPU) in all scenarios, lower energy consumption and quickly in some
scenarios. This articles encourages researches to continually execute real experiment in
new IoT devices, because experiments with real devices contain important information
to be used by energy consumption model and others.
In this direction, the results achieved in this study can be used as a guide for
researches to identify which parameters and metrics best fit for new articles can be
more broadly compared to existing publications.

Acknowledgments. The authors would like to thank the Federal Institute of Paraı́ba
(IFPB)/Campus João Pessoa for financially supporting the presentation of this research and, espe-
cially thank you, to the IFPB Interconnect Notice - No. 01/2020.

References
1. Angra, S., Ahuja, S.: Machine learning and its applications: a review. In: International Con-
ference on Big Data Analytics and Computational Intelligence (ICBDAC), pp. 57–60, March
2017
2. Zantalis, F., Koulouras, G., Karabetsos, S., Kandris, D.: A review of machine learning and
IoT in smart transportation. Future Internet 11(4), 94 (2019)
3. Niranjan, A., Mishra, G.: Etas - efficient traffic signal assistance system using deep learning.
In: ICCCNT, pp. 1–5, July 2018
4. Jan, T., Sajeev, A.S.M.: Boosted probabilistic neural network for IoT data classification. In:
DASC/PiCom/DataCom/CyberSciTech, pp. 408–411, August 2018
5. Ta-Shma, P., Akbar, A., Gerson-Golan, G., Hadash, G., Carrez, F., Moessner, K.: An inges-
tion and analytics architecture for IoT applied to smart city use cases. IEEE Internet Things
J. 5(2), 765–774 (2018)
6. Chollet, F., et al.: Keras (2015). https://keras.io
7. Abadi, M., et al.: TensorFlow: large-scale machine learning on heterogeneous systems 2015,
software available from tensorflow.org (2015). https://www.tensorflow.org/
8. Paszke, A., et al.: Automatic differentiation in PyTorch. In: NIPS Autodiff Workshop (2017)
9. Chen, T., et al.: MXNet: a flexible and efficient machine learning library for heterogeneous
distributed systems. In: Neural Information Processing Systems,Workshop on Machine
Learning Systems (2016). http://arxiv.org/abs/1512.01274
10. Kononenko, I.: Machine learning for medical diagnosis: history, state of the art and perspec-
tive. Artif. Intell. Med. 23(1), 89–109 (2001)
11. McCulloch, W.S., Pitts, W.: A logical calculus of the ideas immanent in nervous activity.
Bull. Math. Biophys. 5(4), 115–133 (1943)
12. Werbos, P.: Generalization of backpropagation with application to are current gas market
model. Neural Netw. 1, 339–356 (1988)
392 L. P. Oliveira et al.

13. Zhang, J., et al.: Automated detection and localization of myocardial infarction with staked
sparse autoencoder and treebagger. IEEE Access 7, 70 634–70 642 (2019)
14. Hentschel, K., Jacob, D., Singer, J., Chalmers, M.: Supersensors: Raspberry Pi devices for
smart campus infrastructure, pp. 58–62, August 2016
15. Gerhardt, T.E., Silveira, D.T.: Métodos de pesquisa. EditoraUFRGS (2009)
16. Valletta, J.J., Torney, C., Kings, M., Thornton, A., Madden, J.: Applications of machine
learning in animal behaviour studies. Anim. Behav. 124, 203 – 220 (2017). http://www.
sciencedirect.com/science/article/pii/S0003347216303360
17. Wang, G.: Machine learning for inferring animal behavior from location and movement
data. Ecol. Inform. 49, pp. 69–76 (2019). http://www.sciencedirect.com/science/article/pii/
S1574954118302036
18. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural
images with unsupervised feature learning (2011)
19. Salih, F., Omer, S.A.M.: Raspberry pi as a video server. In: International Conference on
Computer, Control, Electrical, and Electronics Engineering (ICCCEEE), August 2018
20. Sandeep, V., et al.: Globally accessible machine automation using raspberry Pi based on
internet of things. In: ICACCI, pp.1144–1147, August 2015
21. Fundation, R.P.: Raspberry Pi 3 model b (2016). https://www.raspberrypi.org/products/
raspberry-pi-3-model-b
22. Sandeep, V., et al.: Globally accessible machine automation using raspberry Pi based on
internet of things. In: ICACCI, pp. 1144–1147, August 2015
23. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-MNIST: a Novel Image Dataset for Benchmarking
Machine Learning Algorithms. arXiv:1708.07747, August 2017
24. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document
recognition. In: Proceedings of the IEEE, vol. 86(11), pp. 2278–2324, November 1998
25. Harrison, D., Rubinfeld, D.L.: Hedonic housing prices and the demand for clean air. J.
Environ. Econ. Manage. 5, 81–102 (1978). http://www.sciencedirect.com/science/article/pii/
0095069678900062
26. Menezes, A.H., Kelvin, R.D.O., Oliveira, L.P., Oliveira, P.J.D.S.: IoT Environment to Train
Service Dogs, S3C (2017). https://doi.org/10.1109/S3C.2017.8501386
 Network Topology-Traceable Fault
Recovery Framework with Reinforcement
Learning

Tatsuji Miyamoto1,2(B) , Genichi Mori1,2 , Yusuke Suzuki1,2 ,

and Tomohiro Otani1,2
1
KDDI Corporation, Tokyo, Japan
{tt-miyamoto,ge-mouri,uu-suzuki,tm-otani}@kddi.com
2
KDDI Research, Inc., Saitama, Japan
{tt-miyamoto,ge-mouri,uu-suzuki,otani}@kddi-research.jp

Abstract. Network Function Virtualization (NFV) is the focus of much

attention especially in the field of network operation automation because
of its efficient usage of resources and lower capital expenditures. How-
ever, NFV introduces additional complexity into network monitoring and
management due to the network function being independent of the hard-
ware. Moreover, maintaining the workflow still requires tremendous effort
in order to sustain the automation function. In fact, a vast amount of
workflows or program codes for automated deployment and the failure
recovery operation have to be created and maintained per type of service
and failure case. To address the above problems, previously we proposed
an artificial intelligence-assisted workflow management framework. This
work demonstrated that a scheme for fault-recovery operation automa-
tion. However, the conventional issue is to fix the dimensions of the input
vector for machine learning (ML) algorithms to prevent the relearning
repetition even if the network configuration and topology is changed. To
address the above issue, this paper proposes a reinforcement learning-
based fault recovery framework by applying the reinforcement learning
(RL) algorithm which can adapt to changes of network topology and con-
figuration. We demonstrate the effectiveness of the proposed framework
with testbed results.

1 Introduction

Network management in telecom networks is becoming ever more complex

because of 5G, and novel advances in software technology will become crucial
especially in procedures for network lifecycle management. At the same time,
demands for service quality such as in dealing with service outages and latency
will still remain or grow even higher. Automating network management is an
essential role of management, and thus ETSI ISG NFV has defined the manage-
ment and orchestration (MANO) [1]. However, implementing intelligence into
the orchestrator, and interworking with Operation Support Systems (OSSs) are
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 393–402, 2021.
https://doi.org/10.1007/978-3-030-75100-5_34
394 T. Miyamoto et al.

issues that still must be resolved. Linux foundation networking launched the
Open Network Automation Platform (ONAP) project aimed at achieving auto-
mated management especially for NFV [2]. ONAP provides data analytics for
troubleshooting and closed-loop functioning for interworking with design tools in
addition to providing the orchestration function. These activities provide oper-
ators with standard APIs that simplify the operational procedures and acceler-
ate the achievement of automation. However, from a maintenance perspective,
problems still remain. The types of software errors will likely increase in Vir-
tual Network Function (VNF) applications. To maintain automation, operators
need to update the automation mechanisms per type of error. In addition, the
implementation of automation programming (e.g. shell script) strongly depends
on individual specialist knowledge and skills. The maintainability of the scripts
consequently becomes complicated which leads to an increase in maintenance
costs despite the se of virtualization technology for OPEX saving. To address
this problem, a policy-based approach has been proposed that can simplify the
complex operational process for implementing the automation mechanism. The
policy-based approach presented in [3,4] essentially defines the design of actions
and operations to control resources based on monitored information through
sophisticated scripting (e.g., OpenStack Congress [5]). In [6], S. Calo et al. pro-
posed the Watson Policy Management Library (WPML) [7] as a generic policy
management library and authoring tool that can be used to define easy- to-
use policy definitions and templates for creating services and service assurance
activities. Although the policy-based approach can be used to provide automated
responses to defined actions and rules, it is difficult to apply when operators are
exposed to unexpected behavior.
Taking the above background into account, we have conducted two differ-
ent works. The first is designing a sustainable business process. By referring to
the specification for ETSI MANO defined by the Interfaces and Architecture
(IFA) working group in ETSI NFV as an enhanced Telecom Operation Map
(eTOM) process for fault recovery operation, we designed a standard for the
auto-healing process in TMForum [8]. However, maintaining the deepest level of
business processing is an issue that machine learning (ML) should address due to
vendor-specific implementation. We therefore proposed an Artificial Intelligence
(AI)-assisted closed-loop operation framework [9]. This work demonstrated a
scheme for fault-recovery operation cases. However, the conventional issue is to
fix the dimensions of the input vector for ML algorithms in order to prevent the
relearning repetition even if the network configuration or topology is changed.
In order to use ML techniques, network data must first be transformed so that
it can be utilized by ML techniques.
Various data structures exist to represent attributes of graphs such as the
matrix [10]. Graph classification is an important data mining task, and various
graph kernel methods have been proposed recently for this task. Once the kernel
matrix has been constructed, it is possible to classify the graphs with a Support
Vector Machine (SVM) [11,12], using the supplied kernel matrix. These methods
have proven to be effective, but they tend to have high computational overheads
 Network Topology-Traceable Fault Recovery Framework 395

especially for large graphs [13]. To prevent the computation complexity, an alter-
native approach to constructing a feature-vector for graph classification is pro-
posed [14]. However, the representation has shortcomings such as whether the
network may change over time in the case of failures when adding or removing the
nodes in an overlay network. In such cases, an adjacency matrix would have to
change its dimensions, namely, the input of the ML algorithm would no longer
be scale-free. To tackle these issues, we propose a network topology-traceable
fault recovery framework by applying RL. In the framework, we designed a data
representation procedure for preparing a set of input data for RL.
RL generally employs a simulator for training itself due to training which
requires huge number of trials. On the other hand, we should assume that a 5G
network on NFV is a complex structure which makes it difficult to simulate for all
patterns of network infrastructure behavior. Thus, we validate the effectiveness of
the proposed framework on a testbed by considering the following perspectives.
1. How many trials are required for convergence of learning?
2. How accurate is the network simulator required to be?
This paper addresses the above challenges and proposes a fault recovery
framework by applying RL. Section 2 describes the proposed workflow generation
framework. Section 3 describes the results of the experiment to validate the
effectiveness of the proposed method. Finally, Sect. 4 provides the conclusions.

2 Proposal of Fault Recovery Framework

To deal with the various patterns of network faults, we implemented a systems
of fault recovery automation as shown in Fig. 1. First, the NW Data Analyzer
collects via the monitoring protocol (e.g., SNMP) and information such as perfor-
mance monitoring (PM) data is collected from network infrastructure. Second,
the NW Data Analyzer estimates the root cause of the node and labels its per-
formance indexes. These data and vector values of the network topology graph
are transformed by a NW State Converter into understandable information for
machine learning. Finally, operation API of the workflow engine that suits the
network conditions is triggered by a reinforcement learning engine. Through rep-
etition of the above steps, the generation accuracy of the optimum workflow can
be improved dynamically.
The scope of the proposal has two parts: the topology representation method
of the NW State Convertor and Reinforcement Learning Engine which are
depicted as the red boxes in Fig. 1.

2.1 Data Representation

The operator makes a decision to select the fault recovery workflow based on
the status of the network such as topology, failure location and root cause. In
order to use this information as ML input, we represent a topology graph and
performance metric of the failure node as a multiple-layer matrix. Figure 2 shows
396 T. Miyamoto et al.

Fig. 1. Proposed fault recovery framework

the process of data representation. Regarding to network topology, it can be rep-

resented as a graph structure. On the other hand, when dealing with failures, it
is required to identify the location of each interface separately. Therefore in Step
1, we add interface connectivity to the graph structure and represent the graph
of the network and interface topology as a matrix. In this case, the topology can
be represented by using the adjacency matrix. In Step 2, an additional matrix
which indicates the failure location and root cause is lined up as the failure
matrix. When failure occurs, the element of the faulre matrix which corresponds
to the failure location and properties of the root cause are flagged as ‘1’. The
other elements are ‘0’ which means the normal state. Through the process of
representation, the NW state convertor outputs the multiple-layer matrix and
inputs to the RL engine.

2.2 Reinforcement Learning

Figure 3 shows the overview of the learning process for fault recovery operation.
Since we consider RL algorithms in this paper, determination of environment,
policy function and action value become crucial in this context. We consider
tasks in which the agent interacts with an environment through a sequence of
observations, actions and rewards. The goal of the agent is to select actions in
a way that maximizes cumulative future rewards. Thus, we use a deep convolu-
tional neural network to approximate the optimal action value function [15].
 
Q∗ (s, a) = max E rt + γrt+1 + γ 2 rt+2 + ...|st = s, at = a, π (1)
π

which is the maximum sum of rewards rt representing the change in network

infrastructure discounted by γ at each time step t. We define the optimal action-
 Network Topology-Traceable Fault Recovery Framework 397

Fig. 2. Process of network topology representation

value function Q∗ (s, a) as the maximum expected return achievable by follow-

ing any policy, after seeing some sequence s and then taking some action a,
π = P (a|s) is behavior policy mapping sequences to actions (or distributions
over actions). The action corresponds to executing recovery operation, and the
environment corresponds to a network state including network topology and fail-
ure matrix, respectively. The reward is calculated based on whether the action
affects the state positively or not. At each time-step the agent selects an action
at from the set of APIs of the workflow engine. The action is passed to the
network infrastructure and modifies its internal state and the failure matrix.
By using the proposed framework, it is assumed that we can obtain the
effect of autonomously improving the selection accuracy of tasks. In the next
chapter, we will validate this assumption and describe the results that verify the
effectiveness of the proposed method.

3 Experimental Results
We validate the effectiveness of the proposed framework on our testbed with the
following three verification cases.

• CASE I: CPU usage in a particular single process is 100%. Recover by restart-

ing the process.
• CASE II: Memory usage in two or more processes is 100%. Recover by restart-
ing the VM.
• CASE III: Similar to case II, CPU usage in two or more processes is 100%,
but it is not recovered by restarting the VM. Recover by rebuilding the VM.

In order to reproduce the trial of the above cases, we have developed a network
simulator using OpenAI gym [16]. In this simulator, the faulty node can be
fixed with Relay-B, and the CASE I and CASE II situations can be randomly
generated. By preventing the complexity of the network simulator perspective,
we define the scope of observation of the agent in RL as corresponding to 1 hop
398 T. Miyamoto et al.

Fig. 3. Reinforcement learning process for fault recovery operation

from the failure node. Regarding the action in RL, we implement 20 APIs of
fault recovery action as shown in Table 1. The reward is calculated as shown
in Listing 1. We consider that the attribute “time to recovery” and “high or
low risk” affect the operator’s decision making. Therefore, we implement the
following coefficients as a reward design.

1. Steps to recovery
2. Time to recovery
3. Risk level

Since 1) has been incorporated into the code shown in Listing 1, we implement
the coefficients 2) and 3) for each API as shown in Table 2.

Listing 1. Reward Design

1 Def step():
2 # define initial reward from −0.1
3 reward = −0.1
4
5 action = random or DNN selection
6
7 if NW state dont change:
8 reward += −1.0
9
10 if NW is recovered:
11 reward += +1.0
12
13 return reward

In the following subsection, we validate the effectiveness of the proposed

framework on a testbed from the same perspectives mentioned in Sect. 1.
 Network Topology-Traceable Fault Recovery Framework 399

Table 1. List of fault recovery actions

API# Node Action API# Node Action

0 LB Restart failed process 10 Relay-B Restart failed process
1 LB Restart failed server 11 Relay-B Restart failed server
2 LB Isolate failed server 12 Relay-B Isolate failed server
3 LB Instantiate standby server 13 Relay-B Instantiate standby server
4 LB Provisioning configuration 14 Relay-B Provisioning configuration
5 Relay-A Restart failed process 15 TG Restart failed process
6 Relay-A Restart failed server 16 TG Restart failed server
7 Relay-A Isolate failed server 17 TG Isolate failed server
8 Relay-A Instantiate standby server 18 TG Instantiate standby server
9 Relay-A Provisioning configuration 19 TG Provisioning configuration

Table 2. Validated reward design of recovery

API#0, API#1, API#2, API#3, API#4,

5,10,15 6,11,16 7,12,17 8,13,18 9,15,19
Time to recovery −0.1 −0.3 −0.2 −0.2 0
Risk level 0 −0.3 −0.5 0 0

3.1 How Many Trials Are Required for Convergence of Learning?

Figure 4 shows the F1-score of RL in the proposed fault recovery framework

when changing the learning rate. The result indicates the proposed framework
can retain its accuracy of selecting fault recovery action when the parameter of
the learning rate of RL is set lower than 0.01. Figure 5 shows the learning results
of RL by using a network simulator in CASE I & III, CASE II & III, respectively.
The average-loss value calculation result indicates that the learning process can
be converged. On the other hand, the average-q calculation satisfies the require-
ment of higher than 0.8 in about 4,000 steps (=total number of actions) for each
scenario. From the result the proposed framework requires a tremendous amount
of time for trials to prepare enough training data of RL. Thus, we should con-
sider that the proposed framework needs to prepare a network simulator from a
time efficiency perspective for application to a real network operation.

3.2 How Accurate Is the Network Simulator Required to Be?

In order to validate the requirement of the network simulator, we evaluate the

tolerance of the proposed framework against unexpected behavior of the network
when fault recovery action is executed. In this experiment, we randomly injected
a ‘1’ value into the element of the failure matrix which is generated after the
execution of fault recovery action. Figure 6 shows the F1-score when changing a
noise ratio of the failure matrix in the proposed framework. The result indicates
that the proposed framework can achieve higher performance until 13% of the
400 T. Miyamoto et al.

Fig. 4. F1-score of RL in proposed fault recovery framework when changing the learn-
ing rate

Fig. 5. Learning result of RL of proposed fault recovery framework

noise ratio. It shows that the behavior between the network simulator and real
network against the execution of fault recovery action should be coincident 87%
for applying the proposed framework. Actually, it depends on the acceptability
criterion of the assurance functionalities.
 Network Topology-Traceable Fault Recovery Framework 401

Fig. 6. F1-score when changing the noise ratio of the failure matrix in the proposed
framework

4 Conclusions

This paper proposes a reinforcement learning-based fault recovery framework

by applying an RL algorithm which can adapt to changes of network topology
and configuration. In the framework, we designed a data representation pro-
cedure for preparing a set of data for RL which is formed as a matrix of the
network-topology and failure state. We validated the effectiveness of the pro-
posed framework by implementing and simulating network infrastructure. The
result indicated that the learning process can be converged. On the other hand,
the proposed framework requires a tremendous amount of failure injection/recov-
ery operation trials for preparing enough training data of RL. Thus, we should
consider that the proposed framework needs to prepare a network simulator
from the time efficiency perspective for application to a real network operation.
In order to validate the network simulator requirement, we developed a simulator
by using OpenAI Gym and evaluated the tolerance of the proposed framework
against unexpected behavior of the network when fault recovery action is exe-
cuted. The result showed that the behavior between the network simulator and
implemented network infrastructure should be 87% coincident for application to
the proposed framework. The future work is how to shorten the time for trials
in the training process or the development of a network simulator by satisfying
the acceptability criterion of assurance functionalities.

Acknowledgements. This work was conducted as part of the project entitled

“Research and development for innovative AI network integrated infrastructure tech-
nologies” (JPMI00316) supported by the Ministry of Internal Affairs and Communica-
tions, Japan.
402 T. Miyamoto et al.

References
1. ETSI TR,GS NFV-MAN001,v1.1.1, December 2014
2. https://www.onap.org/ ONAP
3. Makaya, C., Freimuth, D., Wood, D., Calo, S.: Policy-based NFV management and
orchestration. In: IEEE Conference on NFV-SDN, pp. 128–134 (2015)
4. Romeikat, R., Bauer, B., Bandh, T., Carle, G., Schmelz, L., Sanneck, H.: Policy-
driven workflows for mobile network management automation. In: Proceedings of
the IWCMC 2010, pp. 1111–1115 (2010)
5. OpenStack Congress. https://wiki.openstack.org/wiki/Congress
6. Calo, S., Wood, D., Zerfos, P.: Technologies for federation and interoperation
of coalition networks. In: 12th International Conference on Information Fusion
(Fusion), pp. 1385–1392, July 2009
7. https://www.ibm.com/developerworks/community/groups/service/html/
8. https://inform.tmforum.org/features-and-analysis/2015/06/catalyst-recover-
first-resolve-next-closed-loop-control-for-managing-hybrid-networks/
9. Miyamoto, T., Kuroki, K., Miyazawa, M., Hayashi, M.: AI-assisted workflow man-
agement framework for automated closed-loop operation. In: Network Operations
and Management Symposium (NOMS), 2018 IEEE/IFIP, pp. 1–6, May 2018
10. Ramon, J., Gartner, T.: Expressivity versus efficiency of graphs kernels. In: 1st
International Workshop on Mining Graphs, Trees and Sequences (2003)
11. Scholkopf, B., Smola, A.J.: Learning with Kernels: Support Vector Machines, Reg-
ularization, Optimization, and Beyond. MIT Press, Cambridge, MA, USA (2001)
12. Vapnik, V.N.: The Nature of Statistical Learning Theory. Springer-Verlag, New
York (1995)
13. Gärtner, T., Flach, P., Wrobel, S.: On graph kernels: hardness results and efficient
alternatives. In: Schölkopf, B., Warmuth, M.K. (eds.) Learning Theory and Kernel
Machines, pp. 129–143. Springer, Berlin, Germany (2003)
14. Li, G., Semerci, M., Yener, B., Zaki, M.J.: Graph classification via topological and
label attributes. In: International Workshop on Mining and Learning with Graphs
(MLG), vol. 2, pp. 1–9 (2011)
15. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A., Veness, J., Bellemare, M., Graves,
A., Riedmiller, M., Fidjeland, A., Ostrovski, G., Petersen, S., Beattie, C., Sadik,
A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., Hassabis, D.:
Human-level control through deep reinforcement learning. Nature 518(7540), 529–
533 (2015)
16. https://gym.openai.com/
 Mobile Sensing Data Analysis for Road State
Decision in Snow Country

Yoshitaka Shibata1(B) , Akira Sakuraba1 , Yoshikazu Arai2 , Yoshiya Saito2 ,

and Jun Hakura2
1 Regional Coorporate Research Center, Iwate Prefectural University, Sugo, Takizawa 152-89,
Iwate, Japan
{shibata,a_saku}@iwate-pu.ac.jp
2 Faculty of Software and Information, Iwate Prefectural University, Sugo, Takizawa 152-52,
Iwate, Japan
{arai,y-saito,hakura}@iwate-pu.ac.jp

Abstract. In order to keep safety driving on the road in snow country, it is impor-
tant to know the road state and predict the dangerous locations in advance. In
this paper, the data from various in-vehicle sensor state of a snow country using
various in-vehicle sensors are analyzed to identify the road state using AI tech-
nology. The various sensor data and the decided road state are organized as road
state GIS platform system to widely show the road states on map. The accuracy
of decision of road state is evaluated based on the prototype of the proposed road
state information platform.

1 Introduction

Mobility and transportation is the most important means to maintain safe and reliable
daily life and economic activity in the world. Autonomous driving is limitedly starting
in some developed countries on highway and exclusive roads where the driving lanes
are clear, the driving direction is the same and opposite driving car lanes are completely
separated [1]. On the other hand, autonomous driving does not become popular in snow
countries because of bad road surface conditions in winter. Even people are getting old
and driving ability are getting worse, they have to drive their cars to continue their daily
lives. In fact, more than 90% of car accidents in snow area in winter in Japan is due to
slipping car on snowy or iced road [2]. More advanced road state sensing and decision
system to identify the dangerous locations are indispensable for driving in winter.
In order to resolve those problems, we introduce a new generation wide area road
surface state information platform based on crowd sensing. In road sensing, the sensor
data from various environmental sensors including 9 axis dynamic sensor, far infrared
road surface temperature sensor, air humid/temperature sensor, laser sensor, and GPS
attached on vehicle are integrated to precisely determine the various road states and
identify the dangerous locations. This road information is transmitted to the neighbor
vehicles and road side servers using V2X communication network. The received road
information is displayed on the monitor while indicating the dangerous location and

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 403–413, 2021.
https://doi.org/10.1007/978-3-030-75100-5_35
404 Y. Shibata et al.

alerting by voice annotation. This road information is also collected in wide area as big-
data to the cloud computing server through the roadside servers and analyzed to predict
the future road states and unobserved local roads by combining with open weather data
and 3D terrain data along roads. Thus, wide area road state GIS information platform can
be attained and safer and more reliable driving even in winter season in snow countries
and can extend to autonomous driving.
In the following, the related works with road state information by sensor and V2X
technologies are explained in section two. A general system and architecture of the
proposed road surface state information platform are explained in section three. Next,
crowed sensing system to collect various sensor data and to identify road states is intro-
duced in section four. Preliminary experiment by prototyped platform to evaluate our
proposed system is precisely explained in section five. In final, conclusion and future
works are summarized in section seven.

2 Related Works
There are several related works with the road state sensing method using environmental
sensors. Particularly road surface temperature is essentially important to know snowy
or icy road state in winter season by correctly observing whether the road surface
temperature is under minus 4 °C or over.
In the paper [3], the road surface temperature model by taking account of the effects
of surrounding road environment to facilitate proper snow and ice control operations is
introduced. In this research, the fixed sensor system along road is used to observe the
precise temperature using the monitoring system with long-wave radiation. They build
the road surface temperature model based on heat balance method.
In the paper [4–6], cost effective and simple road surface temperature prediction
method in wide area while maintaining the prediction accuracy is developed. Using the
relation between the air temperature and the meshed road surface temperature, statistical
thermal map data are calculated to improve the accuracy of the road surface temperature
model. Although the predicted accuracy is high, the difference between the ice and snow
states was not clearly resolved.
In the paper [7], a road state data collection system of roughness of urban area roads
is introduced. In this system, mobile profilometer using the conventional accelerometers
to measure realtime roughness and road state GIS is introduced. This system provides
general and wide area road state monitoring facility in urban area, but snow and icy
states are note considered.
In the paper [8], a measuring method of road surface longitudinal profile using build-
in accelerometer and GPS of smartphone is introduced to easily calculate road flatness
and International Road Index (IRI) in offline mode. Although this method provides easy
installation and quantitative calculation results of road flatness for dry or wet states, it
does not consider the snow or icy road states.
In the paper [9], a statistical model for estimating road surface state based on the
values related to the slide friction coefficient is introduced. Based on the estimated the
slide friction coefficient calculated from vehicle motion data and meteorological data is
predicted for several hours in advance. However, this system does not consider the other
factors such as road surface temperature and humidity.
 Mobile Sensing Data Analysis for Road State Decision in Snow Country 405

In the paper [10], road surface temperature forecasting model based on heat balance,
so called SAFF model is introduced to forecast the surface temperature distribution on
dry road. Using the SAFF model, the calculation time is very short and its accuracy is
higher than the conventional forecasting method. However, the cases of snow and icy
road in winter are not considered.
In the paper [11], blizzard state in winter road is detected using on-board camera
and AI technology. The consecutive ten images captured by commercial based camera
with 1280 × 720 are averaged by pre-filtering function and then decided whether the
averaged images are blizzard or not by Convolutional Neural Network. The accuracy of
precision and F-score are high because the video images of objective road are captured
only in the daytime and the contrast of those images are almost stable. It is required to
test this method in all the time.
In the paper [12], road surface state analysis method is introduced for automobile tire
sensing by using quasi-electric field technology. Using this quasi-electric field sensor,
the changes of road state are precisely observed as the change of the electrostatic voltage
between the tire and earth. In this experiment, dry and wet states can be identified.
In the paper [13], a road surface state decision system is introduced based on near
infrared (NIR) sensor. In this system, three different wavelengths of NIR laser sensors
is used to determine the qualitative paved road states such as dry, wet, icy, and snowy
states as well as the quantitative friction coefficient. Although this system can provides
realtime decision capability among those road states, decision between wet and icy states
sometimes makes mistakes due to only use of NIR laser wavelength.
With all of the systems mentioned above, since only single sensor is used, the number
of the road states are limited and cannot be shared with other vehicles in realtime. For
those reasons, construction of communication infrastructure is essential to work out at
challenged network environment in at inter-mountain areas. In the followings, a new
road state information platform is proposed to overcome those problems.

3 Road Surface State Information Platform

In order to resolve those problems in previous session, we introduce a new generation

wide area road surface state information platform based on crowd sensing and V2X
technologies as shown in Fig. 1. The wide area road surface state information platform
is organized mainly by mobile wireless nodes, so called Smart Mobile Box (SMB) and
roadside wireless nodes, so called Smart Relay Shelters (SRS). Each SMB is furthermore
organized by a sensor server unit and wireless communication unit, and installed in a
vehicle. The sensor server unit includes various sensor devices such as semi-electrostatic
field sensor, an acceleration sensor, gyro sensor, temperature sensor, humidity sensor,
infrared sensor and sensor server. Using those sensor devices, various road surface states
such as dry, rough, wet, snowy and icy roads can be qualitatively and quantitatively
decided.
On the other hand, the wireless communication unit in SMB and SRS includes mul-
tiple wireless network devices with different N-wavelength (different frequency bands)
wireless networks such as IEEE802.11b/g/n (2.4 GHz), 11ac (5.6 GHz), 11ad (28 GHz),
ah (920 MHz) and organizes a cognitive wireless node. The network node can selects
406 Y. Shibata et al.

the best link among the cognitive wireless network depending on the observed network
quality, such as RSSI, bitrate, error rate by Software Defined Network (SDN). If none
of link connection is possible, those sensor data are locally and temporally stored in
the database unit as internal storage until the vehicle approaches to the region where a
link connection to another mobile node or roadside node is possible. When the vehicle
enters the region, it starts to connect a link to other vehicle or roadside node and transmit
sensor data by DTN Protocol. Thus, data communication can be attained even though
the network infrastructure is not existed in challenged network environment such as
mountain areas or just after large scale disaster areas.
Thus, in our system, SRS and SMB organize a large scale information infrastructure
without conventional wired network such as Internet or wireless network such as cellular
network. The collected sensor data and road state information on a SMB are directly
exchanged to near SMB by V2V communication. Those received data from near SMB
are displayed on the digital map of the viewer system while alerting the dangerous
locations such as icy road location. Thus, the vehicle can know and pay attention to the
ahead road state before passing through.
On the other hand, when the SMB on own vehicle approaches to the SRS, the
SMB make a connection to the SRS, delivers the sensor data and the decided road state
information in the other SMBs through the N-wavelength wireless network. At the same
time, those collected sensor data are sent to the cloud computing in data center through
the cloudlet in SRS. By combining those sensor data form SRSs on various locations
with public mesh weather data such as temperature, pressure, amount of rain and snow
as open data, a wide area road states can be geographically and temporally predicted
and provided as road state Geographic Information System (GIS) information service
through public cellular network not only for current running drivers but for even ordinal
users.

Fig. 1. Road surface state information platform

Thus, the proposed system can provide safer and more reliable driving environment
even though winter season in snow countries. This network not only performs various
 Mobile Sensing Data Analysis for Road State Decision in Snow Country 407

road sensor data collection and transmission functions, but also provide Internet access
network functions to transmit the various data, such as sightseeing information, disaster
prevention information and shopping and so on as ordinal public wide area network for
residents. Therefore, many applications and services can be realized.

4 Road State Sensing System

In order to detect the precise road surface conditions, such as dry, wet, dump, showy,
frozen roads, various sensing devices including 9 axis dynamic sensor, far infrared
road surface temperature sensor, air humid/temperature sensor, laser sensor, and GPS
are integrated to precisely and quantitatively detect the various road surface states and
determine the dangerous locations on GIS in sensor server as shown in Fig. 2.

Fig. 2. Road state sensing system

4.1 9-Axis Dynamic Sensor

9-axis dynamic sensor including 3-axis accelerator, 3-axis gyro sensor and 3-axis elec-
tromagnetic sensor is used to measure vertical amplitude of roughness and horizontal
along the road as well as horizontal drift due to snow ruts. The signal from 9-axis
dynamic sensor is sub-sampled and converted by A/D convertor with several mseconds,
and then transformed from time series to frequency spectra density. By reducing high
frequency noise by low pass filter, the residual essential signal is input to the decision
maker to qualitatively determine the road surface state. The quantitative values of the
9-axis dynamic sensor can be also used to control the velocity and steering angle of
autonomous driving car (Fig. 3).

4.2 Far-Infrared Temperature Sensor

The far-infrared temperature sensor observes the road surface temperature to determine
whether the road surface is snowy or icy without touching the road surface as shown in
Fig. 4. The dynamic range of the observed temperature is very wide from −40–500 °C.
Since its response time and sampling rate are 0.1 and 10 Hz, realtime decision of road
surface state can be realized. However, since the emissivity of infrared amount depends
on the objective material of road, calibration due to the difference of materials of road
must be made to maintain the measurement accuracy.
408 Y. Shibata et al.

Fig. 3. Axis dynamic sensor

Fig. 4. Far-infrared temperature sensor

4.3 Air Temperature/Humidity Sensor

The air temperature sensor is used to measure the air temperature outside of vehicle
and know the relation to the air and road surface temperatures. If the relation between
both temperatures can be clearly understood, the road surface temperature can estimate
by this relation. Then, since the road surface temperature is related to the road state,
the future road state can be predicted using the future road surface temperature if other
whether condition is the same. Since the element to detect air temperature part is platinum
resistance thermometer sensor, the measurement accuracy is high within plus minus 0.3°
although the response time of air temperature is long around 1–2 min.
On the other hand, humidity sensor is used to know the air is dry or wet and snowy
or icy. If air is wet, the probability of snowy road state is high, whereas, if air is dry,
the probability of icy road state is high in winter. The element of humidity sensor part
is made by polymer capacity typed sensor, the measurement accuracy is low with plus
minus 3%.

Fig. 5. Air temperature/humidity sensor

 Mobile Sensing Data Analysis for Road State Decision in Snow Country 409

Thus, more detailed road states in both winter and summer seasons can be determined
by combination of those sensors among possible road states (Fig. 5).

4.4 Near Infrared Laser Sensor

The far-infrared laser sensor precisely measures the friction rate of snowy and icy road
states. Using three laser beams with different wave lengths (λ1, λ2, λ3) and considering
their reflection and absorption rates on the road surface, the road states such as dry, wet,
snowy and icy states in addition to the friction rate, can be determined (Fig. 6).

Fig. 6. Near-infrared laser sensor

4.5 Global Positioning System

GPS is required to obtain the exact current time and the locations of the vehicle. The
all of the sampled sensor data are linked with the sampled time and the location by
longitude and latitude indexes and managed as database such as the relation database in
sensor server. When the vehicle goes through tunnel, the longitude and latitude indexes
cannot be detected. In this case, the locations of the vehicle can be calculated using 9-axis
dynamic sensor. As results, all of the time and location data can be always extracted.

5 Prototype and Field Experiment

In order to evaluate the sensing function and accuracy of sensing road surface states,
both sensor and communication servers, and various sensors are set to the vehicle as
shown in Fig. 7. The experimental vehicle running trials was made in the Morioka city
which is a snow country in Northern Jan during January to February at both daytime
and night about 4 h to evaluate each sensor data and decision accuracy in realtime on the
winter road with various road states such as dry, wet, snowy, damp and icy states shown
in Fig. 8. In order to evaluate the road surface decision function, the video camera is
also used to compere the decision state and the actual road surface state.
410 Y. Shibata et al.

Fig. 7. Crowd road surface sensing system

Fig. 8. Experimental field at daytime and night

Fig. 9. Result 1: road state distribution for daytime and night

5.1 Evaluation and Analysis of Various Sensor Data

Figure 9 shows the road state distribution on both daytime in dry and damp states and
night in snowy and icy road states. The gray, green, yellow and red mean dry, damp,
snowy and icy, respectively. At daytime, since it is cloudy, the road surface state is dry
 Mobile Sensing Data Analysis for Road State Decision in Snow Country 411

and damp. Whereas at night, the road surface sate is snowy and icy. By evaluating and
comparing those idenfied road states by the sensors and the video camera, our road
surface state sensing system can identify the real winter states more than about 80%
accuracy.

Fig. 10. Result 2: road surface temp. & friction coefficient

Fig. 11. Result 3: air, road surface temps and humidity

Fig. 12. Result: pearson correlation

Figure 10 shows the road surface temperature and the friction coefficient at both
the same daytime and night. At daytime, the road surface temperature is higher and the
412 Y. Shibata et al.

friction coefficient is larger in average, whereas at night, the surface temperature is lower
and almost constant and the friction coefficient is lower in average. It is concluded that
night cold road surface is more dangerous than daytime and the dangerous locations can
be precisely identified on the road map.
Figure 11 shows air temperature, road surface temperature and humidity at both the
same daytime and night. Since at daytime, both the air temperature and humidity vary
as elapsed time and then the road surface temperature vary, whereas at night, all the air
temperature, humidity and road surface are lower and almost constant. It is concluded
that the night time is easy to be snowy or icy due to lower temperature and humidity
compared with daytime.
Figure 12 shows the Pearson correlation among road temperature, air temperature,
humidity and friction coefficient. It is clear that the correlation between the road temper-
ature and air temperature, and the correlation between the friction and road temperature
are is high at both daytime and night. It is concluded that by observing correctly air
temperature and road surface temperature, the average road state in the same area can
be detected whereas by measuring road temperature and the friction coefficient, precise
dangerous locations, such as snowy or icy state can be identified.

6 Conclusions

In this paper, we introduced road state information platform to realize safe and reliable
driving in winter in snow countries based on various in-vehicle environmental sensors.
The sensor data are analyzed to identify the road state using AI technology. The various
sensor data and the decided road state are organized as road state GIS platform system to
widely show the road states on the road map. In order to evaluate the proposed system, a
prototype is designed and developed. The accuracy of decision of road state is evaluated
on the actual public road. From this field experiment, our proposed system could pro-
vide reasonable accuracy to identify the correct road states although more experimental
measurement data are required to improve the accuracy of road state identification. As
future work, we will deploy the EV based connected autonomous driving based on our
road state information platform in challenged road and network environment.

Acknowledgement. The research was supported by JSPS KAKENHI Grant Numbers JP

20K11773, Strategic Information and Communications R&D Promotion Program Grant Num-
ber 181502003 by Ministry of Affairs and Communication and Strategic Research Project Grant
by Iwate Prefectural University in 2020.

References
1. SAE International: Taxonomy and Definitions for Terms Related to Driving Automation
Systems, for On-Road Motor Vehicles. J3016_201806, June 2018
2. Police department in Hokkaido: The Actual State of Winter Typed Traffic Accidents,
November 2018. https://www.police.pref.hokkaido.lg.jp/info/koutuu/fuyumichi/blizzard.pdf
 Mobile Sensing Data Analysis for Road State Decision in Snow Country 413

3. Takahashi, N., Tokunaga, R.A., Sato, T., Ishikawa, N.: Road surface temperature model
accounting for the effects of surrounding environment. J. Japan. Soc. Snow Ice 72(6), 377–390
(2010)
4. Fujimoto, A., Nakajima, T., Sato, K., Tokunaga, R., Takahashi, N., Ishida, T.: Route-based
forecasting of road surface temperature by using meshed air temperature data. In: JSSI&JSSE
Joint Conference, pp. 1–34, September 2018
5. Saida, A., Sato, K., Nakajima, T., Tokunaga, R., Sato: A study of route based forecast of road
surface condition by melting and feezing mass estamation method using weather mesh data.
In: JSSI&JSSE Joint Conference, pp. 2–57, September 2018
6. Hoshi, T., Saida, A., Nakajima, T., Tokunaga, R., Sato, M., Sato, K.: Basic consideration of
wide-scale road surface snowy and icy conditions using weather mesh data. Monthly report
of Civil Engineering Research Institute for Cold Region, No. 800, pp. 28–34, January 2020
7. Fujita, S., Tomiyama, K., Abliz, N., Kawamura, A.: Development of a roughness data collec-
tion system for urban roads by use of a mobile profilometer and GIS. J. Japan Soc. Civ. Eng.
69(2), I_90–I_97 (2013)
8. Yagi, K.: A measuring method of road surface longitudinal profile from sprung acceleration
and verification with road profiler. J. Japan Soc. Civ. Eng. 69(3), I_1–I_7 (2013)
9. Mizuno, H., Nakatsuji, T., Shirakawa, T., Kawamura, A.: A statistical model for estimating
road surface conditions in winter. In: Proceedings of Infrastructure Planning (CD-ROM). The
Society of Civil Engineers, December 2006
10. Saida, A., Fujimoto, A., Fukuhara, T.: Forecasting model of road surface temperature along
a road network by heat balance method. J. Civ. Eng. Japan 69(1), 1–11 (2013)
11. Okubo, K., Takahashi, J., Takechi, H., Sakurai, T., Kokubu, T.: Possibility of blizzard detec-
tion by on-board camera and AI technology. Monthly report of Civil Engineering Research
Institute for Cold Region, No. 798, pp. 32–37, November 2019
12. Takiguchi, K., Suda, Y., Kono, K., Mizuno, S., Yamabe, S., Masaki, N., Hayashi, T.: Trial
of quasi-electrical field technology to automobile tire sensing. In: Annual Conference on
Automobile Technology Association, 417–20145406, May 2014
13. Casselgren, J., Rosendahl, S., Eliasson, J.: Road surface information system. In: Proceedings
of the 16th SIRWEC Conference, Helsinki, 23–25 May 2012. https://sirwec.org/wp-content/
uploads/Papers/2012-Helsinki/66.pdf
14. Ito, K., Shibata, Y.: V2X communication system for sharing road alert information using
cognitive network. In: Proceedings of 8th International Conference on Awareness Science
and Technology, Taichung, Taiwan, pp. 533–538, September 2017
15. Ito, K., Shibata, Y.: Experimentation of V2X communication in real environment for road alert
information sharing system. In: Proceedings on IEEE International Conference on Advanced
Information Networking and Applications (AINA 2015), Gwangju, Korea, pp. 711–716,
March 2015
16. Ito, K., Shibata, Y.: Estimation of communication range using Wi-Fi for V2X communication
environment. In: The 10th International Conference on Complex, Intelligent, and Software
Intensive Systems (CISIS 2016), Fukuoka, Japan, pp. 278–283, July 2016
 Towards Industrial 5G: An LTE+ Case
Study Report

Alexander Gogolev(B) and Markus Aleksy

ABB Corporate Research Center, Ladenburg, Germany

{Alexander.Gogolev,Markus.Aleksy}@de.abb.com

Abstract. This manuscript evaluates the robustness of the small LTE+

campus setup prototype for industrial automation, addressing archi-
tecture bottlenecks and the communication latencies. It shows that a
channel overload or a simple Denial-of-Service attack can have a crucial
impact on the communication latencies in such setups. The findings sug-
gest that future LTE+ small campus setups could benefit from certain
enhancements in their wired segment.

1 Introduction and Motivation

The ongoing progress in the development of wireless communication technologies,
such as Long-Term Evolution (LTE), Narrowband Internet of Things (NB-IoT),
and 5G enable new types of industrial applications. Traditional industrial sys-
tems usually rely on wired communication technologies to connect sensors and
actuators. However, these type of communication needs to be enhanced by wire-
less technologies in future to address new developments, such as mixed reality
applications, automated guided vehicles, moving robots and drones or achieving
higher flexibility required by increasing demand for highly customized products
and adaptable production facilities.
Additionally, the future of industrial automation promises to flatten the hier-
archy in automation systems, unify the hardware and software, and bring the
Industrial Internet of Things (IIoT) to the field level [1].
Enabling devices with IIoT on the application level is often suggested with
Open Platform Communications and Unified Architecture (OPC UA) [2], and its
early evaluations already show promising results [3–7]. 5G is another technology
that could foster IIoT development for industrial automation on the commu-
nication plane. With its ultra-reliable-low-latency communication (URLLC) 5G
vendors promise to unlock the use cases previously inaccessible due to limita-
tions of the wired communication. Already there is evidence that indeed a low-
latency communication can be achieved with 5G [8,9]. But there are still aspects
that need careful assessment. For instance, Industry 4.0 concept suggests that
new automation technologies should offer open standard and non-vendor bind-
ing interfaces and tools (as does, for instance, another next-gen communication
platform, IEEE TSN [10]), and the 5G is currently developing such capabilities
in 5G ACIA Alliance [11]. Another aspect is the communication reliability and
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 414–423, 2021.
https://doi.org/10.1007/978-3-030-75100-5_36
 Towards Industrial 5G: An LTE+ Case Study Report 415

robustness—it is known that mobile networks can handle high throughput, but
in automation it is more important to deliver “little data on time”, rather than
“a lot of data delivered a bit later”. Deterministic wireless communication stud-
ies show promising approaches [12–16], but if 5G aims to compete with wired
industrial communication in terms of determinism, its robustness needs to be
revisited.
Below we extend the assessment of the new communication technologies for
industrial automation [6,7,17,18] with an LTE+ campus network. The test setup
is rated as “LTE+” since at the moment of the experiment there is no finalized
standard to directly attribute the capabilities of the studied campus cell to the
5G. This paper reflects on the findings from the perspective of communication
robustness, since it is one of the most vital aspects for industrial systems.
The paper is organized as follows. Next section details the 5G application in
the scope of the industrial automation. Section 3 explains the test system setup.
Then, Sect. 4 presents the use cases and assumptions that guided our research.
Sections 5 and 6 present and discuss test results. Finally, Sect. 7 concludes our
work.

2 Industrial 5G: Campus Networking

The today’s business of 5G vendors is creation and operation of the end-user
mobile networks and their cooperation. In such systems network management is
typically handled by the network operator, or a service provider–a single entity
that exclusively controls throughput allocation and connection guarantees. Sup-
porting multiple simultaneous subscriptions which can be streaming video feeds
at high data rates, entails handling sporadic data bursts, often combined with
high user mobility, both in terms of distance and speed. Such networks need to
ensure data exchange with high throughput, but not-so-high-reliability, due to
low critical data exchange.
In industrial automation (IA), whether it is process automation, or mobile
robots, the end systems is not typically moving at speeds of up to 300 km/h,
for distances on hundreds of kilometers (as could a typical high-speed train pas-
senger). Industrial end systems then could be relatively slow and their mobility
range could be confined within a limited space of a factory territory, i.e., a
few square kilometers. Industrial automation, however, requires uninterrupted
service for its end systems, because of, e.g., safety reasons. To guarantee such
uninterrupted service IA often chooses to control all of the automation pyramid
vertically: from the cloud to the factory floor. This way an automation archi-
tecture can be traced, verified and validated to ensure the necessary safety and
reliability. Or, an architecture could be modular, allowing division into segments,
for maintenance, redundancy or other purposes.
To meet requirements from different IA use cases and still maintain some
generics, 5G campus networks seem to be a good candidate for pilot evaluations.
5G campus is a cluster of 5G network devices bounded by licensed frequencies
and deployed within a limited territory. Such an approach has little challenges
when it comes to reducing of the users’ speed and possible mobility range.
416 A. Gogolev and M. Aleksy

However, vertical/modular integration would require such a campus network

to expose some interfaces to the existing systems (typically, running fieldbuses),
or the systems with upcoming technologies (e.g., running IEEE TSN), and local
engineers and their tools. This would require a significant re-work of the cur-
rent mobile network centralized and proprietary management. 5G Alliance for
Connected Industries and Automation [11] established a task work group to
define a standard exposure interface for such network management, to foster the
adoption of 5G in IA.
But the largest and the most important hitch is the reliability of the data
exchange which can affect the safety and execution continuity of the automated
process. To a certain extent the safe execution of the process control can be
ensured by the higher-level control software that could, for instance, shut down
the system in a fail-safe manner to guarantee safely operation in case of inter-
rupted communication. But, this still leaves the process continuity at question,
which can lead to disruptions in the functioning of the factory.
The aspect of communication reliability and robustness in 5G campus net-
work needs assessment in at least two of its segments:
• Wireless communication
• Overall system architecture
The robustness of wireless segment has been addressed with physical
enhancements such as multiple-input and multiple-output (MIMO), orthogo-
nal frequency-division multiplexing (OFDM), etc., and we leave it outside this
paper. However the latter can be evaluated with a network engineering tool-set
on the given example of the small cost-effective LTE+ campus network.

2.1 Industry Performance Requirements

Communication performance and robustness are crucial for industrial automa-
tion systems, and are typically assessed with timing parameters of data delivery,
such as latency and jitter. Industrial automation spans timing requirements from
motion control that require those in µs-range to, for instance, process automa-
tion, that often deals with millisecond-granular control cycles. 5G claims to
match the latter latencies thus paving its way into such industries.
Robustness of wireless communication is, due to limitations of shared com-
munication media, typically worse than that of wired communication. Hence
URLLC for Industrial IoT systems is often motivated with the assumption that
it will not hinder the communication robustness. Rather it claims to add flexibil-
ity, opening up new use cases that were previously inaccessible. Since the purely
wireless aspects are omitted in this paper, the main focus is on the campus
network architecture.

2.2 5G and Latency Accumulation

The discussions of 5G URLLC for industrial applications place the latency in
data exchange in a millisecond, or even sub-millisecond range [8,9]. Such claims,
 Towards Industrial 5G: An LTE+ Case Study Report 417

however, do not always clearly specify the jitter, packet loss rate, and overall
determinism.
As mentioned above, IA spans over multiple fields with rather varying
requirements. For instance, motion control’s µs-granular control loops are clearly
unachievable with 5G URLLC. It could satisfy, on the other hand, the latency
requirements for process automation or mobile robots, given that communication
is stable and robust. We skip the assessment of stability and robustness in the
wireless segment, but rather focus on the architecture of the 5G campus network
that needs to reflect these requirements (which differ from the requirements of
the “regular” user mobile networks).
Therefore, the open questions are:
1. Can the suggested architecture of the 5G campus network cope with high-load
traffic?
2. How would the channel overload, or a denial-of-service (DoS) attack affect
the end-to-end latency?
3. What is the quantifiable impact of such overload?
We try to reflect on this below, using a real-life LTE+ setup.

3 Test Setup Description

The test setup is shown in Fig. 1 and consists of:
1. 1 LTE+ Base Station
2. 2 LTE User Equipment devices, attached to the Linux machines
3. 1 management PC (Linux)
4. 1 non-managed Gigabit Ethernet bridge
5. 4 Linux PCs, connected to the bridge and management PC
We refer to the setup as LTE+, since at the moment of the testing there is no
finalized standard that could directly attribute the capabilities of the setup to
“Industrial 5G”. This means that the HW and SW is operational but cannot be
strictly speaking referred to as “5G” because the feature sets first need to be
defined and rated in the standard.

3.1 User Equipment

User Equipment is a term used in mobile networks to designate the end user
connection point on the system. It is, typically, a mobile phone or a modem.
The test setup uses two PCs with two commercial-off-the-shelf (COTS) LTE
UEs (User Equipment – a modem), communicating to the base station (BS).
We have observed that for such UEs (available as COTS and thus often used
for testing), the communication latency does not always correlate with data
rate/speed. In our setup the goal is to minimize the former, since automation
devices rarely produce much traffic, but do require low latency communication.
For this reason, they need to be configured in specific modes of operations, which
418 A. Gogolev and M. Aleksy

in some cases requires some low-level access and change of firmware to provide
lower latency instead of higher speed. These modes are commonly referred as
“stick” and “router” modes, and the latter generally offers higher throughput
(and is typically set as default mode). The “stick” mode may reduce the effec-
tive throughput but does show better latency. Thus, in our test the UEs are
configured to run in the “stick” mode.

3.2 5G Enablement
Base station is then connected to a management PC(P Cm ) over a Gigabit Eth-
ernet cable and a switch, as shown in Fig. 1.
P C3 and P C4 are connected to P Cm , and exchange data with the P C1 and
P C2 . They use independent physical network interfaces on the P Cm to avoid
traffic interference. All machines run Ubuntu 18.04 LTS, except P Cm – it runs a
vendor-chosen Linux version with hardware binding (required to secure the BS
management).

P C1 / U E1 P C3

BS 1 Bridge 2 P Cm P C6

P C2 / U E2 P C5 P C4

Fig. 1. Network setup with two LTE (U E) devices, and an LTE+ base station (BS).
Solid gray line represents the Gigabit Ethernet connections, dashed lines – wired traffic,
dotted lines – traffic over wireless link. Blue dotted line and the green dashed one –
ICMP traffic between UEs and test PCs, loose dotting and dashing designates the
wireless link. Red dash-dotted line designates disturbance traffic.

The network topology is schematically represented in Fig. 1, where bridge

Bridge connects base station BS to the base-station management PC – P Cm .
The physical network topology is organized in a way that all of the traffic to/from
U Es is going through the ports 1 and 2 on the base station and P Cm , respec-
tively. Four independent physical interfaces of P Cm are connected to the Bridge,
P C3 , P C4 and P C6 , respectively.
It can be seen that in the test setup P C1 and P C2 (to the left of the BS)
communicate to their counterparts (to the right of the BS) – P C3 and P C4 using
both wireless and wired links. They use physically independent UE devices (i.e.,
modems) with different QoS configuration: P C1 has a wireless bearer configured
in Qci 8, where P C2 uses Qci 9 (higher priority). Qci could be explained as
the broadband equivalent of the Ethernet QoS priority distinction, though not
exactly right so, due to fundamental differences in defining the QoS in mobile
 Towards Industrial 5G: An LTE+ Case Study Report 419

networks [19]. What is important here is that P C2 communicates on a “more

important link” than P C1 . Once the data from P C1 and P C2 crosses over to
the BS it is forwarded to the bridge and further towards P Cm . From the P Cm
it is forwarded to P C3 and P C4 using physically separate interfaces. Thus, the
data between P C1 and P C3 and P C2 and P C4 only shares the wired connection
between the BS and the P Cm .
In a wireless segment (between UEs and BS) the Qci distinction ensures that
in case of saturation (around 1 Gbps of combined throughput), the link with
higher Qci receives better service. However, once the traffic leaves the BS towards
P Cm and further, it should be served using Ethernet QoS mechanisms (e.g., out
of TSN toolbox). TSN combines several such mechanisms of which some pre-date
the TSN itself and therefore are available since a while. In the direction UEs →
PC, even if the wireless media is saturated with high-density data-exchange, BS
would ensure the delivery of the data from the high-importance link towards the
“trunk” port on the P Cm . However, if the communication is bi-directional (and
it typically is) the reverse direction (UEs ← PCs) is also important. Here the
determinism of the data exchange also needs to be ensured, but with Ethernet-
based toolset, as mentioned above. But, even if the respective tool-set is in place
and properly configured, it would be beneficial to have similar support on the BS
side of the wired link. Moreover, the starvation of the P Cm hardware resources
still poses a challenge: if P Cm is not the most capable one, processing gigabits
of traffic can exhaust its computing capabilities.
To mitigate some of the most obvious bottlenecks the test setup connects both
P C3 and P C4 with separate physical network interfaces. Respective forwarding
rules are then added to the forwarding tables of the P Cm .
Generating high-throughput disturbance traffic incurs additional CPU load
on the PCs, which may affect the performance measurements. To avoid this,
the PCs that generate disturbance and the ones performing the actual commu-
nication are physically independent machines. P C6 generates the disturbance
traffic towards the P C5 that has to go out of the same physical interface of the
P Cm as the traffic towards the BS. Thus, the setup generates “non-important”
disturbing traffic and “important” Internet Control Message Protocol (ICMP)
traffic on physically independent machines, which then share the same physical
interface on the P Cm . The latter, as one can notice, represents the single point
of failure in this architecture. However, the software bundle on the P Cm is pre-
pared by the broadband vendor, and implements a network configuration of a
certain complexity. This software implements certain features, and utilizes the
available hardware to ensure the necessary determinism and guarantees for the
traffic on the broadband side of the test network.

4 Methodology and Performance Reference

To evaluate the data exchange latency and robustness we use the Linux ping
command (IMCP echo request/reply) with an interval of 20 ms (empirically
derived). The round trip times (RTT, referred further as request latency) are
registered on the client side and averaged over 100000 cycles.
420 A. Gogolev and M. Aleksy

The unidirectional disturbance traffic of 800 Mbps is generated using an

Iperf tool [20]. It can be shown that for instance, Ostinato packet generator [21]
provides a steadier flow of packets/frames and therefore can be better suited
for such tests [17]. However, the steady data flow is just a specific use case of a
saturated link, while a typical data flow is rather sporadic. In any case the Iperf
has already provided the saturation sufficient to distinguish a serious latency
growth, as shown in the following section.
Let us briefly reference previous findings with regard to latency in new indus-
trial network technologies. In [6,17], the latency of the ICMP echo request
is studied in a multi-hop TSN network showing that RTTs can reach sub-ms
latency with Time-Aware Shaping [22] or Frame Preemption [23]. It was also
shown that achieved latencies stay stable even with disturbance traffic up to
100% of the channel capacity.

5 Latency Measurements
To evaluate the impact of traffic saturation, the testing is performed using ICMP
packets of 64 bytes. The disturbance is consequently tested with a different
throughput (∈[400, 500, 800] Mbps), using 1518B packet size.
The Table 1 presents the baseline latencies achieved in a network with and
without disturbance.

Table 1. Average RTTs over LTE+ setup

LTE+ Session Disturbance

None 400 Mbps 600 Mbps 800 Mbps
High priority (Qci 9) 20 ms 20 ms 22 ms 35 ms
Low priority (Qci 8) 20 ms 20 ms 22 ms 35 ms

The test did not exceed the 800 Mbps of the disturbance since it would
saturate the link even further, and suffocating Ethernet links is out of scope of
this paper as it has been studied previously.
It shows that e2e communication latency noticeably grows with 800 Mbps of
induced traffic, or with effectively 80% of the channel utilization. It also shows
that both sessions (with Qci = 9 and Qci = 8) are equally disturbed, despite
different levels of QoS guarantees. It is known that in LTE+ networks QoS is
ensured by means of Qci bearer classes (now also known as “5ci” in 5G ACIA [11],
with similar function as QoS priority levels of Ethernet) which shape traffic
sessions in the wireless segment: between BS and UEs. However, it is evident
that Qci distinction in the wireless segment is not sufficient to ensure full e2e
determinism: there are guarantees missing elsewhere in the system. Since the
system is built of wired and wireless segments, and wireless segment is presumed
to be handled with Qci mechanisms, it leaves the wired segment to blame. The
test sessions cross over both segments in reverse directions:
 Towards Industrial 5G: An LTE+ Case Study Report 421

• U E1 ↔ P C3
• U E2 ↔ P C4
One could argue that in the direction U E1 , U E2 → P C3 , P C4 the BS and
Qci-based QoS will ensure that high important traffic is not disturbed. However,
this argument does not hold for direction P C3 , P C4 → U E1 , U E2 the traffic
needs guarantees before it gets to the BS where Qci mechanisms can take over.
So, if the communication channel is congested in the direction P C3 , P C4 →
U E1 , U E2 the communication session latency may worsen. This explains the
observations above.

6 Analysis
In the real-life mobile operator setups, the network management functions are
not likely to run on a single PC that has a single point of failure, such as in
the described setup, but rather a HW-based routers or multi-port-and-cpu-core
server blades that are difficult to saturate with a simple DoS attack. Still, since a
BS has just one port to which such router or a server is likely to be connected, the
same observations hold. Moreover, “campus network” is marketed as an entry
concept to industrial market, which must be reasonably cost-competitive, and
therefore is likely to use the simplest setup available, similar to the described one.
In such a setup the observations above are even more just. One important aspect
of the adoption of the 5G is the compatibility of the QoS, or service guarantee
levels, which can be difficult due to the different media capabilities and essential
difference in requirements. Therefore, matching the 5G’s QCIs (5CIs) to, e.g.,
Ethernet QoS needs to be organized in either flexible, or well-motivated manner.
Furthermore, Qci guarantees must be then translated into the wired part of the
setup, to ensure the QoS throughout the cell, campus and the whole mobile
network.
In addition to this, integration with existing and future automation systems
requires a thoroughly weighted choice of interfaces that is, perhaps, aligned with
existing solutions for other “yet to come” communication technologies, such as
IEEE TSN. Such solutions and their motivation can be found in the state of the
art discussions in, e.g., FLC group [24].
Next, performance difference with wired technologies such as TSN or other
deterministic Ethernet fieldbuses, must be accounted for in 5G and reflected
properly. Statements such as “5G campus can be perceived from outside as a
single TSN bridge”, can be misleading and should be avoided, since TSN bridges
guarantee 1, 000 times lower latencies than current 5G.

7 Summary
This paper extends earlier work [6,7,17] on new industrial communication tech-
nologies assessing the LTE+ campus network from a point of communication
reliability and robustness.
422 A. Gogolev and M. Aleksy

We show that the studied campus network has a bottleneck in its design
that can severely hinder the communication determinism. We illustrate this by
channel overloading that simulates a simple denial-of-service attack that signifi-
cantly increases the e2e communication latency. This indicates that studied 5G
campus network concept is yet not mature enough for robustness testing, nor
for industrial deployment. However, it can be used as a testbed for early tests in
performance and integration. Last, but not least—the Ethernet QoS mechanisms
can be suggested for enhancing the wired segment of the 5G campus network,
where possible.
Further research can, therefore, address issue of increasing the efficiency and
determinism on the 5G side, for instance, adding software and hardware on those
bottleneck interfaces (if overall architecture stays unchanged).

Acknowledegements. This research was supported by the German Federal Ministry

of Education and Research (BMBF) under grant number 16KIS0721. The responsibility
for this publication lies with the authors.

References
1. Unlock the value of Industry 4.0. https://www.ericsson.com/en/internet-of-
things/industry4-0
2. OPC Foundation. Open platform communications unified architecture. https://
opcfoundation.org/about/opc-technologies/opc-ua/
3. Veichtlbauer, A., Ortmayer, M., Heistracher, T.: OPC UA integration for field
devices. In: 2017 IEEE 15th International Conference on Industrial Informatics
(INDIN), pp. 419–424, July 2017
4. Pfrommer, J., Ebner, A., Ravikumar, S., Karunakaran, B.: Open source OPC
UA PubSub Over TSN for realtime industrial communication. In: 2018 IEEE
23rd International Conference on Emerging Technologies and Factory Automation
(ETFA), vol. 1, pp. 1087–1090 (2018)
5. OPC UA TSN a new solution for industrial communication. https://www.br-
automation.com/smc/953ce46647cb909f0cce603249fb229e29f0a30a.pdf
6. Gogolev, A., Braun, R., Bauer, P.: TSN traffic shaping for OPC UA field devices.
In: 2019 IEEE 17th International Conference on Industrial Informatics (INDIN),
vol. 1, pp. 951–956, July 2019
7. Gogolev, A., Mendoza, F., Braun, R.: TSN-Enabled OPC UA in field devices. In:
2018 IEEE 23rd International Conference on Emerging Technologies and Factory
Automation (ETFA), vol. 1, pp. 297–303 (2018)
8. Pilz, J., Mehlhose, M., Wirth, T., Wieruch, D., Holfeld, B., Haustein, T.: A tactile
internet demonstration: 1ms ultra low delay for wireless communications towards
5g. In: IEEE Conference on Computer Communications Workshops (INFOCOM
WKSHPS), vol. 2016, pp. 862–863 (2016)
9. Soldani, D., Guo, Y.J., Barani, B., Mogensen, P., Chih-Lin, I., Das, S.K.: 5G for
ultra-reliable low-latency communications. IEEE Netw. 32(2), 6–7 (2018)
10. Institute of Electrical and Electronics Engineers. Time-sensitive networking task
group. http://www.ieee802.org/1/pages/tsn.html
11. 5G Alliance for Connected Industries and Automation. https://www.5g-acia.org/
about-5g-acia/
 Towards Industrial 5G: An LTE+ Case Study Report 423

12. Zhang, S., Zhang, H.: A review of wireless sensor networks and its applications.
In: IEEE International Conference on Automation and Logistics, vol. 2012, pp.
386–389 (2012)
13. Zheng, L., Lu, N., Cai, L.: Reliable wireless communication networks for demand
response control. IEEE Trans. Smart Grid 4(1), 133–140 (2013)
14. Jonsson, M., Kunert, K.: Towards reliable wireless industrial communication with
real-time guarantees. IEEE Trans. Ind. Inform. 5(4), 429–442 (2009)
15. Fan, B., Krishnan, H.: Reliability analysis of DSRC wireless communication for
vehicle safety applications. In: 2006 IEEE Intelligent Transportation Systems Con-
ference, pp. 355–362 (2006)
16. Berger, A., Pichler, M., Haselmayr, H., et al.: Energy-efficient and reliable wireless
sensor networks-an extension to IEEE 802.15.4e. JEURASIP J. Wireless Commun.
Netw. 126, 1–2 (2014)
17. Gogolev, A., Bauer, P.: A simpler TSN: traffic preemption vs. scheduling. In: 2020
IEEE Conference on Emerging Technologies and Factory Automation (ETFA),
vol. 1, September 2020
18. Abukwaik, H., Gogolev, A., Groß, C., Aleksy, M.: Opc ua realization for simplified
commissioning of adaptive sensing applications for the 5g iiot. Internet Things 11,
100221 (2020)
19. The 3rd Generation Partnership Project (3GPP). Policy and charging con-
trol architecture. https://portal.3gpp.org/desktopmodules/Specifications/
SpecificationDetails.aspx?specificationId=810
20. iPerf – a tool for active measurements of the maximum achievable bandwidth on
IP networks. https://iperf.fr/
21. Ostinato – packet generator. https://ostinato.org/
22. Institute of Electrical and Electronics Engineers. 802.1Qbv – Enhancements for
Scheduled Traffic. http://www.ieee802.org/1/pages/802.1bv.html
23. Institute of Electrical and Electronics Engineers. 802.1Qbu – Frame Preemption.
https://www.ieee802.org/1/pages/802.1bu.html
24. OPC Foundation. Field Level Communications (FLC) Initiative. https://
opcfoundation.org/flc/
 Thermal Environment Simulator for Office
Building with Multi-agent-Based Simulation
Technique

Kodai Murakami(B) , Toshihiro Mega, and Noriyuki Kushiro

Faculty of Computer Science and Systems Engineering, Kyusyu Institute of Technology,

Kitakyushu, Japan

Abstract. There is a growing interest to introduce split-ductless air conditioning

systems into an office building both for realizing energy saving and comfort to
office workers. Building energy management systems (BEMS) are expected to
establish sophisticated controlling methods for the split-ductless air conditioning
system. For creating the methods, simulators for thermal environment in a build-
ing are required. However, existing simulators necessitate a building information
model (BIM) and huge amount of data for parameter tuning. Moreover, the exist-
ing simulators have difficulties to express spatial and temporal dynamics in each
space in a building. The spatial and temporal dynamics in a space are indispens-
able for creating methods for controlling split-ductless air conditioning systems,
because the systems independently manage devices dispersed in a room. A thermal
environment simulator with multi-agent-based simulation (MAS) technique was
proposed in the paper. The simulator can work without BIM and visualize spa-
tial and temporal dynamics with limited numbers of parameters, which are easily
obtained from BEMS. The simulator with MAS was implemented and applied to a
field test in winter and summer seasons. As a result, high precision for reproducing
spatial and temporal thermal dynamics was confirmed through the field test.

1 Introduction

Building Energy Management Systems (BEMS) has been introduced to reduce green-
house gas emissions into office buildings [1]. Electricity energy consumption for air
conditioning reaches above 30% in office buildings [2, 3]. The air conditioning systems
are regarded as a major target of energy reduction to the BEMS. Energy consumption for
air conditioning systems is affected by various of factors: outdoor temperature, humidity,
density of office workers in a space. The BEMS ought to identify these factors and decide
how to control the air conditioning systems efficiently. However, the operation for energy
reduction in air conditioning systems often causes deterioration of comfortabilities for
office workers, due to trade-off between saving energy and realizing comfort.
Various kinds of research have been conducted to relax incompatible features among
saving energy and realizing comfort by the BEMS. For examples, [4] balanced energy
efficiency and office workers’ comfortability with machine learning techniques. [5] pro-
posed algorithms by utilizing information elicited from WEB to improve comfortability.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 424–435, 2021.
https://doi.org/10.1007/978-3-030-75100-5_37
 Thermal Environment Simulator for Office Building 425

Energy prediction based on huge historical data with machine learning techniques [6] is
one of promising solutions for relaxing the trade-off, because the results of prediction
help the BEMS plan optimal ways and control schedule for the air conditioning systems.
[7] provided a prediction method, which combined accuracy and readability, with het-
erogeneous mixture modeling. Both accuracy and readability of the model contribute to
create an optimal control and schedule method to BEMS.
Recently, split-ductless air conditioning systems are introduced into an office build-
ing. The split-ductless air conditioning systems realize energy saving and comfort to
office workers by managing independent devices dispersed in a room. However, the
existing simulators necessitate a building information model (BIM) and huge amount
of data for parameter tuning. The existing simulators have difficulties to express spatial
and temporal dynamics in each space in a building. The spatial and temporal dynam-
ics are indispensable information for creating methods for controlling split-ductless air
conditioning systems.
In this paper, a thermal environment simulator with multi-agent-based simulation
(MAS) technique was proposed. The simulator works without BIM and visualize spatial
and temporal dynamics with limited numbers of parameters, which are easily obtained
from BEMS. The simulator was implemented and applied to a field test. As a result,
high accuracy was confirmed through the experiment.
The rest of paper is organized as follows: Sect. 2 summarizes modeling approaches
used in simulator for BEMS. Section 3 proposes concept of thermal environment simula-
tor for office building with MAS technique, and Sect. 3.1 explains method for identifying
parameters required for the proposed simulator. Section 4 summarizes the field test of
the simulator and results, and Sect. 5 concludes the paper.

2 Thermal Environment Simulator for Office Building

This section surveys existing thermal simulation methods used in BEMS. Thermal
environment simulator with MAS technique is proposed base on the surveys.

2.1 Categories of Modeling Approaches for Simulation

Modeling approaches in BEMS are classified into the following three categories [8, 9].
Features of each approach are summarized in Table 1.

2.1.1 Category1: Forward Approach

Forward approach is regarded as a white box approach. It contains descriptions of math-
ematical models on physics [9]. The mathematical models utilize the results of build-
ing information modeling (BIM) as a basis of simulation and requires many physical
constants for calculation.
The forward approach is classified into the following three methods: Computational
Fluid Dynamics (CFD), zone, and multi-zone approaches [10]. The CFD approach
describes flow of field on a micro perspective. In this approach, heat transfer in a building
is expressed as fluid in three dimensions [11]. [12] proposed volume element model for
426 K. Murakami et al.

Table 1. Features for each modelling approach

Categories Features
Forward approach Pros: High accuracy of simulation base on physics
Cons: Requiring huge number of parameters for simulation and
necessity of complex mathematical models for simulation
Data driven approach Pros: Simple mathematical or statistical model
Cons: Less parameters for simulation and huge training data
Grey box approach Pros: Limited mathematical model
Cons: Identification unknown parameters

3D dynamic building thermal simulation, and applies CFD approach in order to exam-
ine local thermal effect. Zone approach is a simplified version of the CFD approach, in
which a room is divided into multiple zones and the physical equations are solved for
each cell in each zone. [13] presents a zonal model to predict air temperature distribution
in buildings with calculating mass air flows between different zones. Multi-zone app-
roach is a method of modeling airflow and contaminant transfer between different rooms
or between outside or wall-to-room by means of node-to-node transfer, and assumes that
temperature and contaminant concentration are uniform within a single node. [14] eval-
uates assumption that air momentum effects can be neglected in a multizone airflow
network models.
Several researches with forward approach for improving accuracies were challenged
by solving complex mathematical equations explaining thermal environment in building
through long period of tuning time [18, 19]. These challenges increase necessities for
more minute BIM and more granular parameters, which are not easy to obtain from the
existing BEMS.

2.1.2 Category2: Data-Driven Approach

The data-driven approach, known as a black box approach, constructs prediction model
with huge training data with machine learning techniques [15]. The data-driven approach
constructs the prediction model without any mathematical equations on physical, it
requires just data for learning, and statistical models on machine learning techniques.
For example, [16] predicted room temperature and relative humidity by autoregressive
linear and nonlinear neural network models accurately.
Data driven approach has a merit to predict objective variables at specific time and at
specific position in a building, however it has clear restrictions to explain temporal and
spatial thermal dynamics in a space, due to nonexistence of physical rules. Therefore, it
has difficulty to apply for determining control methods for the distributed air conditioning
systems.

2.1.3 Category3: Grey Box Approach

Gray box approach is an intermediate modeling approach between the white box (for-
ward) and the black box (data-driven) approaches. The gray box approach uses restricted
 Thermal Environment Simulator for Office Building 427

number of mathematical equations on physics and identified parameters required for the
equations on black box approach. In [17], plural of gray box models was estimated
their accuracies for predicting thermal dynamics in multi-zones of office building. As a
result, the gray model was confirmed that they realized high accuracies in comparison
with the white box and black box approaches. However, the gray-box approach remains
difficulties to select an appropriate mathematical model from several kinds of governing
equations, and to identify parameters including the selected mathematical model with
data elicited by BEMS.

2.2 Solutions for Issues in Existing Simulation Model

The following are requirements for thermal simulator in this study:

1. The simulator should predict spatial and temporal dynamics in each space in a
building for controlling and scheduling split-ductless air conditioning systems.
2. The simulator does not require BIM for constructing prediction model, because the
simulator should be applied to any scales of building, which has difficulty to obtain
BIM.
3. The simulator should identify all the parameters required for simulation by data
easily obtained from BEMS.

2.2.1 Concept of Thermal Environment Simulator for Office Building with Multi-
agent-Based Simulation Technique
To solve the requirements, we propose a new model to describe thermal dynamics with
fewer parameters. Multi-agent-based simulation modeling (MAS) is introduced as the
modeling technique for thermal dynamics. MAS is a bottom-up approach to express
an entire system through interactions among autonomous, rule-based actors called as
agents, and simulate complex behaviors in the system by defining simple interaction
rules among agents [20].

Fig. 1. Conceptual diagram of the proposed simulation model.

428 K. Murakami et al.

By introducing the MAS to thermal dynamics simulator, we are succeeded in reduc-

ing mathematical equations and parameters required for simulation. The simulator with
MAS technique requires three basic heat conditions and three parameters. Figure 1 shows
image of heat transfer in the room and overall diagram of the proposed simulation model.
Heat transfer is determined by the following three types of heat transfer mechanism.
At the first, the heat transfer in which a space transfers heat obtained from a heat
source to an adjacent space was defined as a heat exchange model as follows:

1. Heat exchange model - If the temperatures of the adjacent spaces are Tspace and

Tspace , respectively, the amount of heat transferred in unit time, Qspace , is
 
  
Qspace = α Tspace − Tspace  (1)

At the second, the effect of convective heat from the air that the air conditioning
produces was modeled by assuming that particles called heat charge are emitted from
the air conditioning. The heat charge is periodically emitted from the air conditioner and
transfers heat to the surrounding space as it moves, as shown in Fig. 1. Heat transfer into
space by heat charge was defined as follows:

2. Heat convection model - Assuming that the temperature of the emitted heat charge
is Theat , the temperature of the surrounding space is Tspace , and the distance between
the heat charge and the space is r, the amount of heat transferred from the heat charge
to the space in unit time, Qconvection , is
 
Theat − Tspace 
Qconvection = β (2)
r2
The detailed conceptual diagram of the heat charge model is shown in Fig. 2. The
heat charge is defined as properties: coordinates, energy, temperature, velocity, radius,
and direction of movement.

Fig. 2. Properties and conceptual diagram of the heat charge model.

 Thermal Environment Simulator for Office Building 429

Finally, radiant heat model is defined as follows, assuming that radiant heat is gen-
erated from people, equipment such as PCs, floors, ceilings, windows, and walls as heat
sources, and that heat is transferred to adjacent spaces.

3. Radiation heat model - If the temperature of the heat source is Tbarrier and the
temperature of the adjacent space is Tspace , the amount of heat transferred Qradiation
is
 
Qradiation = γ Tbarrier − Tspace  (3)

Equations (1), (2) and (3) follow Fourier’s law, which is a fundamental principle of
heat transfer engineering.
Effect of forced convection from air conditioning ordinally needs to solve complex
and numerous equations on fluid mechanics by using the CFD. In this study, the effects of
pseudo forced convection come down to the Fourier’s law, assuming that the heat charge
interacts with surrounding spaces through agent simulation. This makes the simple. In
addition, heat sources such as people and walls were also considered as agents to affect
spatial and temporal dynamics of heat.
Equations (1), (2), and (3) are basic physical models that consolidate the constants
necessary for fluid mechanism analysis, e.g., heat transfer coefficient, thermal conduc-
tivity, and viscous stress, respectively. The proportionality constants α, β, and γ are
parameters, which represent amount of heat transfer at a unit temperature difference in
unit time. Since the parameter are regarded as aggregation of many constants including
various physical quantities, the proportionality constants α, β, and γ need to be identified
anew. The parameters were determined experimentally in this study. Method to identify
the parameters was also developed and details of the method are explained in Sect. 3.

3 Implementation of Thermal Environment Simulator

with Multi-agent Modeling Technique
3.1 Implementation of Simulator with Multi-agent Modeling Technique
The simulator was developed by the MAS software “artisoc” [22] on the model proposed
in Sect. 2. Figure 3 shows the overview of the simulator and office layout used for an
example of simulation. Lists of functions of the simulator is shown in Table 2.
The simulator starts only with suction temperature of each air conditioners, which
can be easily obtained by BEMS, and works on predetermined control schedule of the air
conditioners. Users set conditions for simulations (date, time, outdoor temperature etc.)
on the panel and directly observe spatial and temporal thermal dynamics as animation.
Measurement device for identifying the parameters required for the simulator was
also developed in the study. The parameters are determined experimentally through the
examination in an actual building prior to starting simulation. Various studies have been
conducted to determine thermal properties in existing buildings. [21] proposed accurate
mathematical algorithm for evaluating heat transfer by using infrared thermography. An
overview of the device for identifying the parameters with multi-thermal cameras and
the results of are presented in the next section.
430 K. Murakami et al.

Table 2. Functional requirements for the developed simulator

Functions Details
Input (csv file) BEMS data: initialization data
control data: Air conditioning control file in minutes
Output (csv file) Air conditioning information and observed temperature
Heat map Red when the temperature is high, white when the temperature is low
Control panel Setting of observation temperature position, air conditioning and
input/output data
Graph visualization Power consumption, observation temperature, indoor temperature

Fig. 3. The simulator we developed and the layout we used as a reference

3.2 Parameter Tuning in Lab and Methods for Tuning

As shown in the left image of Fig. 4, paper cylinders were hung at equal intervals from the
ceiling in a space, and each paper cylinders were photographed with 6 thermal cameras
(LIR-AX8) at every 10 s. The parameters were determined through calculating heat
balance based on changes of temperature in paper cylinders for unit time. The image on
right of Fig. 3 shows temperature of each paper cylinder.
 Thermal Environment Simulator for Office Building 431

Temperature

Fig. 4. Scene for the experiment with device (left), thermal image obtained from device (right).

Three proposed models (Eq. (1), (2) and (3)) are applied to describe respective heat
exchange mechanism among agents. In the heat exchange model, heat is exchanged
between spatial agents. In the convection model, heat is exchanged between heat charge
and space. In the radiant heat model, heat is exchanged between objects such as floors
and walls and space. Since the temperature of the space cannot be measured directly
with thermal cameras, the parameters are determined by replacing paper cylinders as
spatial agents. The corresponding positional relationships for each model are shown in
Fig. 5.

Fig. 5. Parameter identification method using a paper cylinder

Multi values should be determined in the radiation model, because effects from walls,
windows, ceilings and floors are different. In this study, we assumed that walls, floors,
and ceilings have almost the same thermal properties and give the same values. Each
value of parameter was tuned through experiments on the simulator prior to field test
and the results of parameter identification are shown in Table 3.
432 K. Murakami et al.

Table 3. Parameter identification results

Parameter value Measured value [j/s · ◦ C]

Heat exchange model (α) 0.520
Convection heat model (β) 0.610
Radiation heat model: window (γ) 0.039
Radiation heat model: wall, ceiling, floor, human (γ) 0.033

4 Field Test and Estimation of Simulation

4.1 Basic Setups for Field Test
Field test was conducted for evaluating accuracy of prediction with the simulator.
The simulator was installed into an office building (8000 m3 ) in Tokyo and evaluates
by comparing the results of simulations with data obtained from the BEMS. Table 4
shows the details of the field test.

Table 4. Details of field test

Items Details
Location 8-story office buildings (8000 m3 ) in Tokyo
Data acquisition period (summer) 8:00 am–7:30 pm on 11/Aug.–17/Aug., 20.Aug.–23/Aug.
in 2020
Data acquisition period (winter) 8:00 am–7:30 pm on 14/Dec.–20/Dec. in 2020

The same layout for the experimental site was reproduced on the simulator (Fig. 3),
and the accuracy of the simulator was evaluated by comparing the inlet temperature
obtained from the BEMS with the temperature of the space agent, which were occupying
the same position, calculated by the simulator. The green squares allotted to 10–13, 18–
21, and 27 in Fig. 3 depict arrangement for devices of split-ductless air conditioning
systems in the room. The simulator started with conditions of the floor at 8:00 am,
which obtained from the BEMS, and simulates spatial and temporal dynamics in the
floor 7:30 pm. Few minutes were required for simulate on notebook computer. The
performance of the simulator seemed to be practical for use.

4.2 Parameter Tuning in Lab and Methods for Tuning

Figure 6 and 7 show comparison results from 8:00 am to 7:30 pm for each space depicted
as 10–13 and 18–22 on 12th and 13th Aug., respectively. Differences between measured
value and simulated value were small at every space and every time. We confirmed that
the simulator can reproduce spatial and temporal thermal dynamics in the floor through
the field test.
 Thermal Environment Simulator for Office Building 433

Fig. 6. Temperature comparison results for August 12

Fig. 7. Temperature comparison results for August 13

5 Conclusion

Thermal environment simulator for office building with multi-agent-based simulation

(MAS) was proposed in this paper so as to create sophisticated controlling methods for
the split-ductless air conditioning systems in BEMS.
The simulator satisfies the following requirements:

1. The simulator should predict spatial and temporal dynamics in each space in a
building for controlling and scheduling split-ductless air conditioning systems.
434 K. Murakami et al.

2. The simulator does not require BIM for constructing prediction model, because the
simulator should be applied to any scales of building, which has difficulty to obtain
BIM.
3. The simulator should identify all the parameters required for simulation by data
easily obtained from BEMS.

To solve the requirements, we propose a new model to describe thermal dynam-

ics with fewer parameters. MAS is introduced as the modeling technique for thermal
dynamics. By introducing the MAS technique to thermal dynamics simulator, we are
succeeded in reducing mathematical equations and parameters required for simulation.
The simulator works without BIM and visualize spatial and temporal dynamics with
limited numbers of parameters, which are easily obtained from BEMS.
The simulator was implemented and applied to a field test. The simulator was
installed into an office building (8000 m3 ) in Tokyo area and evaluated its accuracies by
comparing the results of simulation with the data obtained by the BEMS. As a result,
we confirmed that the simulator can reproduce spatial and temporal dynamic thermal
environment accurately.

Acknowledgements. We would like to express our special thanks to Mr. Fuyuki Sato in Mitsubishi
Electric Building Techno Service. Discussions with him help progress of the study.

References
1. U.S. Energy Information Administration (EIA): Annual Energy Outlook 2020 (2020)
2. International Organization Standardization: ISO50001:2018 Energy management (2018)
3. The Energy Conservation Center, Japan (ECCJ): Energy Conservation for Office Buildings
2010 (2010)
4. Carreira, P., Costa, A., Mansur, V., Arsénio, A.: Can HVAC really learn from users? A
simulation-based study on the effectiveness of voting for comfort and energy use optimization.
Sustain. Cities Soc. 41, 275–285 (2018). https://doi.org/10.1016/j.scs.2018.05.043
5. Papantoniou, S., Kolokotsa, D., Kalaizakis, K.: Building optimization and control algorithms
implemented in existing BEMS using a web based energy management and control system.
Energy Build. 98, 45–55 (2015)
6. Yang, C., Létourneau, S., Guo, H.: Developing data-driven models to predict BEMS energy
consumption for demand response systems. In: Ali, M., Pan, J.S., Chen, S.-M., Horng, M.
(eds.) Modern Advances in Applied Intelligence. Lecture Notes in Computer Science (Lecture
Notes in Artificial Intelligence), vol. 8481, pp. 188–197. Springer, Cham (2014)
7. Kushiro, N., Fukuda, A., Kawatsu, M., Mega, T.: Predict electric power demand with extended
goal graph and heterogeneous mixture modeling. Information 10(4), 134 (2019)
8. Coakley, D., Raftery, P., Keane, M.: A review of methods to match building energy simulation
models to measured data. Renew. Sustain. Energy Rev. 37, 123–141 (2014)
9. Kumar, A.: A review on modeling and simulation of building energy systems. Renew. Sustain.
Energy Rev. 56, 1272–1292 (2016)
10. Foucquier, A., Robert, S., Suard, F., Stephan, L., Jay, A.: State of the art in building modeling
and energy performances prediction: a review. Renew. Sustain. Energy Rev. 23, 272–288
(2013)
11. Wendt, John F. (ed.): Computational Fluid Dynamics, 3rd edn. Springer, Heidelberg (2009)
 Thermal Environment Simulator for Office Building 435

12. Yang, S., Pilet, T.J., Ordonez, J.C.: Volume element model for 3D dynamic building thermal
modeling and simulation. Energy 148, 642–661 (2018)
13. Inard, C., Bouia, H., Dalicieux, P.: Prediction of air temperature distribution in buildings with
a zonal model. Energy Build. 24(2), 125–132 (1996)
14. Wang, L., Chen, Q.: Evaluation of some assumptions used in multizone airflow network
models. Build. Environ. 43(10), 1671–1677 (2008)
15. Amasyali, K., El-Gohary, N.M.: A review of data-driven building energy consumption
prediction studies. Renew. Sustain. Energy Rev. 81(1), 1192–1205 (2018)
16. Mustafaraj, G., Lowry, G., Chen, J.: Prediction of room temperature and relative humidity by
autoregressive linear and nonlinear neural network models for an open office. Energy Build
43, 1452–1460 (2011)
17. Berthou, T., Stabat, P., Salvazet, R., Marchio, D.: Development and validation of a gray box
model to predict thermal behavior of occupied office buildings. Energy Build. 74, 91–100
(2014)
18. Khaddaj, M., Srour, I.: Using BIM to retrofit existing buildings. Procedia Eng. 145, 1526–1533
(2016)
19. Volk, R., Stengel, J., Schultmann, F.: Building information modeling (BIM) for existing
buildings – literature review and future needs. Autom. Constr. 38, 109–127 (2014)
20. Macal, C.M., North, M.J.: Tutorial on agent-based modelling and simulation. J. Simul. 4,
151–162 (2010)
21. Baldinelli, G., Bianchi, F., Rotili, A., Costarelli, D., Seracini, M., Vinti, G., Asdrubali, F.,
Evangelisti, L.: A model for the improvement of thermal bridges quantitative assessment by
infrared thermography. Appl. Energy 211, 854–864 (2018)
22. artisoc4.0: Specialized platform for MAS. https://mas.kke.co.jp/en/artisoc4/
 Message Ferry Routing Based
on Nomadic Lévy Walk in Wireless
Sensor Networks

Koichiro Sugihara and Naohiro Hayashibara(B)

Kyoto Sangyo University, Kyoto, Japan

{i1986089,naohaya}@cc.kyoto-su.ac.jp

Abstract. Message ferry is a way of communication in Intermittently

Connected Delay-Tolerant Wireless Sensor Networks. It is for collecting
data from stationary sensor nodes. The efficiency of message delivery
using message ferries depends on the routing scheme. Nomadic Lévy
Walk that is a variant of Lévy Walk is an eligible candidate for a mes-
sage ferry routing scheme. It includes homing behavior in addition to the
behavior similar to Lévy walk with strategic moving the home position.
In this paper, we conducted several simulation runs on a WSN to mea-
sure the message delivery ratio of message ferries with several routing
schemes. We also evaluate that Nomadic Lévy Walk is suitable for mes-
sage ferry routing. Our simulation result shows Nomadic Lévy Walk is
practically useful as a message ferry routing for collecting data in WSNs.

1 Introduction

Various messaging protocols in Intermittently Connected Delay-Tolerant Wire-

less Sensor Networks (ICDT-WSNs) have been proposed so far [10]. This is
because WSNs are attracted attention in various research fields. In this type of
network, stationary sensor nodes which have wireless communication capability
with limited communication range are located in a field, and messages and sensor
data sent by a node are delivered to the destination in a Store-Carry-Forward
manner because end-to-end paths do not always exist. It means that each mes-
sage or data is carried to the destination by some mobile entities which repeat
message passing to each other.
Now, we focus on the Message Ferrying based approach for data collection
in ICDT-WSNs. Message ferries (or MF for short) are mobile nodes that move
around the field to collect sensor data from wireless sensor nodes. Each sensor
node sends data to an MF with short-range communication capability (e.g.,
Bluetooth) when it gets close.
Therefore, the efficiency of data collection depends on the routing scheme
of MFs. Most of the message ferry routing schemes assume a fixed route; for
instance, MFs move on a circular path [9,21,22]. However, the fixed route based
routing scheme does not work if a part of the route is not available.

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 436–447, 2021.
https://doi.org/10.1007/978-3-030-75100-5_38
 Message Ferry Routing Based on Nomadic Lévy Walk 437

We focus on a random walk based routing scheme for message ferry routing.
In particular, Lévy walk has recently attracted attention due to the optimal
search in animal foraging [7,20] and the statistical similarity of human mobil-
ity [11]. It is a mathematical fractal that is characterized by long segments
followed by shorter hops in random directions. Lévy walk has also been used to
improve the spatial coverage of mobile sensor networks [18], to analyze the prop-
erties of an evolving network topology of a mobile network [5], and to enhance
grayscale images combined with other bio-inspired algorithms (i.e., bat and fire-
fly algorithms) [6]. It is also considered to be particularly useful for data col-
lection and finding outliers of sensor values to detect incidents in a large-scale
delay-tolerant sensor network.
Nomadic Lévy Walk (NLW) is a variant of Lévy walk, which has been pro-
posed by Sugihara and Hayashibara [16]. An agent starts from a sink node and
returns to the sink node with the given probability α. Each sink node changes
its position according to the given strategy with the probability γ.
We suppose to collect sensor data by using battery-driven autonomous elec-
tric vehicles as MFs. Moreover, they have an on-board camera to drive along
the road and avoid obstacles. Each of them basically moves with a mobile power
source vehicle. Then, it departs from the sink to explore sensor nodes for col-
lecting data. It requires to go back to the sink to charge its battery.
In this paper, we apply NLW as a message ferry routing and measure the
cover ratio in a WSN. In this context, agents and sink nodes are considered
to be MFs and mobile power source vehicles. MFs are require to collect data
from sensors efficiently. So, we measure the cover ratio that indicates the ratio
of sensor nodes from which MFs collect data in our simulation. First, we analyze
the simulation result with the parameter α and γ of NLW to clarify the impact
of these parameters on the cover ratio. Then, we compare two message ferry
routings, NLW and Random Way Point (RWP), regarding the cover ratio.

2 Related Work

We now introduce several research works related to message ferry in WSNs and
WSNs. Then, we explain existing movement models that can be used for message
ferry routing.

2.1 Message Ferry in WSNs and DTNs

Tariq et al. proposed Optimized Way-points (OPWP) ferry routing method to

facilitate connectivity on sparse ad-hoc networks [4]. According to the simulation
result, OPWP-based MFs outperforms other ferry routing method based on
Random way point (RWP).
Shin et al. apply the Lévy walk movement pattern to the routing of message
ferries in DTN [12]. They demonstrated message diffusion using message ferries
based on the various configuration of Lévy walk. According to the simulation
438 K. Sugihara and N. Hayashibara

result, the ballistic movement of message ferries (i.e., smaller scaling parameter
of Lévy walk) is efficient regarding the message delay.
Basagni et al. proposed the notion of a mobile sink node and its heuris-
tic routing scheme called Greedy Maximum Residual Energy (GMRE) [3]. The
motivation of the work is to prolong wireless sensor nodes that are deployed in
a large field and send data to the sink periodically. The mobile sink as a data
collection and processing point moves around the field to save the energy con-
sumption of sensor nodes. The simulation result showed that the mobile sink
with the proposed routing scheme improved the lifetime of sensor nodes.
Alnuaimi et al. proposed a ferry-based approach to collect sensor data in
WSNs [1,2]. It divides a field into virtual grids and calculates an optimal path
for a mobile ferry to collect data with a minimum round trip time. It utilizes
a genetic algorithm and the node ranking clustering algorithm for determining
the path.

2.2 Movement Models

Birand et al. proposed the Truncated Levy Walk (TLW) model based on real
human traces [5]. The model gives heavy-tailed characteristics of human motion.
Authors analyzed the properties of the graph evolution under the TLW mobility
model.
Valler et al. analyzed the impact of mobility models including Lévy walk on
epidemic spreading in MANET [19]. They adopted the scaling parameter λ = 2.0
in the Lévy walk mobility model. From the simulation result, they found that the
impact of velocity of mobile nodes does not affect the spread of virus infection.
Thejaswini et al. proposed the sampling algorithm for mobile phone sensing
based on Lévy walk mobility model [18]. Authors showed that proposed algo-
rithm gives significantly better performance compared to the existing method in
terms of energy consumption and spatial coverage.
Fujihara et al. proposed a variant of Lévy flight which is called Homesick
Lévy Walk (HLW) [8]. In this mobility model, agents return to the starting
point with a homesick probability after arriving at the destination determined
by the power-law step length. As their result, the frequency of agent encounter
obeys the power-law distribution though random walks and Lévy walk do not
obey it.
Most of the works related to Lévy walk assume a continuous plane and hardly
any results on graphs are available. Shinki et al. defined the algorithm of Lévy
walk on unit disk graphs [14]. They also found that the search capability of Lévy
walk emphasizes according to increasing the distance between the target and the
initial position of the searcher. It is also efficient if the average degree of a graph
is small [13].
Sugihara et al. proposed the novel mobility model called Nomadic Lévy
Walk (NLW) as a variant of HLW and the sink relocation strategy based on
a hierarchical clustering method [16,17]. They conducted simulations to mea-
sure the cover ratio of unit disk graphs. The simulation result showed that the
mobility model cover a wide area with preserving homing behavior.
 Message Ferry Routing Based on Nomadic Lévy Walk 439

3 System Model
We assume a wireless sensor network so that each stationary sensor node is
deployed in a field. So, we model the field by a unit disk graph U DG = (V, E)
with a constant radius r. Each node v ∈ V is located in the Euclidean plane, and
an undirected edge (link) {vi , vj } ∈ E between two nodes vi and vj exists if and
only if the distance of these nodes is less than 2r. Nodes in a graph represent
wireless sensor nodes. Note that r is the Euclidean distance in the plane. We
assume that any pair of nodes in the graph has a path (i.e., connected graph).
It does not mean network connectivity. It just means that there is a geometric
route to any sensor node in the field. UDGs are often used for modeling road
networks. We assume the field restricts the movement of mobile entities by links.
We also assume a mobile entity called a message ferry (MF), which collects
messages and data from sensor nodes, and forwards them to the destination.
This assumption is similar to the existing work [1,2,12,22]. Practically speaking,
both MFs and mobile nodes have short-range communication capability such
as Bluetooth, ad hoc mode of IEEE 802.11, Near Field Communication (NFC),
infrared transmission, and so on. Each MF starts moving from its home position,
called sink, and it goes back to the sink position occasionally.
There are two types of ferry interaction.
• Ferry-to-Ferry interaction. Message ferries exchange messages each other if
they are at the same node (i.e., they are in their communication range).
• Ferry-to-Node interaction. Each sensor node can send data piggybacked on a
message to a MF when it is physically close to the node (i.e., it is at the node
in the graph).
We suppose that MFs are autonomic electric vehicles and sinks are mobile
power source vehicles in practice. Thus, MFs are required to come back to their
sink to charge their battery.
We also assume that each MF identifies its position (e.g., obtained by GPS),
which is accessible from an MF, and it has a compass to get the direction of a
walk. Each MF has a set of neighbor nodes, and such information (i.e., positions
of neighbors) is also accessible. Moreover, an MF has an on-board camera to
drive along the road and avoid obstacles. On the other hand, it has no prior
knowledge of the environment.

4 Nomadic Lévy Walk

We proposed a variant of Lévy walk called Nomadic Lévy Walk in our previous
work [16,17] to improve the ability of the broad area search while preserving
the homing behavior. NLW is an extension of HLW [8], and holds the following
properties in addition to HLW.

Nomadicity: Each sink moves its position with the given probability γ.
Sink relocation strategy: The next position of the sink is decided by a particular
strategy.
440 K. Sugihara and N. Hayashibara

The movement of each MF obeys HLW. Thus, their trajectory is radially

from their sink. Moreover, they move their sink at the probability γ in Nomadic
Lévy walk. In fact, the fixed sink restricts the area that each MF explores. The
property of Nomadicity is expected to improve coverage by each MF.
We now explain the detail of the algorithm of Nomadic Lévy walk on unit
disk graphs in Algorithm 4.

1: Initialize:
s ← the position of the sink
c ← the current position
o←0  orientation for a walk.
P N (c) ← the possible neighbors to move.
2: if Probability: α then
3: d ← the distance to s
4: o ← the orientation of s
5: else
6: d is determined by the power-law distribution
7: o is randomly chosen from [0, 2π)
8: end if
9: if Probability: γ then
10: s ← P (c)  update the position of the sink according to the given strategy.
11: end if
12: while d > 0 do
13: P N (c) ← {x|abs(θox ) < δ, x ∈ N (c)}
14: if P N (c) = ∅ then
15: d←d−1
16: move to v ∈ P N (c) where v has the minimum abs(θov )
17: c←v
18: else
19: break  no possible node to move.
20: end if
21: end while

At the begging of the algorithm, each MF holds the initial position as the
sink position s. In every walk, each MF determines the step length d by the
power-law distribution and selects the orientation o of a walk randomly from
[0, 2π). It can obtain a set of neighbors N (c) and a set of possible neighbors
P N (c) ⊆ N (c), to which MFs can move, from the current node c. In other
words, a node x ∈ P N (c) has a link with c that the angle θox between o and the
link is smaller than π2 .
In unit disk graphs, it is not always true that there exist links to the desig-
nated orientation. We introduce δ that is a permissible error against the orien-
tation. In this paper, we set δ = 90. It means that MFs can select links to move
in the range ±90 with the orientation o as a center.
In a given probability α, the MF goes back to the sink node (line 2 in Algo-
rithm 4). In this case, it sets the orientation to the sink node as o and the
distance to the sink as d.
 Message Ferry Routing Based on Nomadic Lévy Walk 441

Each MF changes its sink position with the given probability γ (line 10 in
Algorithm 4) and then starts a journey from the sink.

4.1 Sink Relocation Strategy

Each sink replaces its position according to the given strategy in NLW. Obvi-
ously, the sink position is one of the important factors for covering a wide area
of a graph. Each sink manages the history H of the sink positions. We have
proposed several sink relocation strategies for NLW [16,17]. We use a strategy
for preventing biased positions by using hierarchical clustering [17].
In this strategy, each sink assumes an ordered set of the history of the sink
positions Bhist with the size h to store the coordinates of the locations at which
the sink has been located in the past. It calculates the reverse orientation of each
position in Bhist .
The sink positions could be biased at a particular area of the graph while relo-
cating. As a result, the cover ratio could not be improved with MFs. For this rea-
son, we use the unweighted pair group method with arithmetic mean (UPGMA)
clustering method [15] to detect a bias of sink positions. UPGMA method is an
agglomerative hierarchical clustering method based on the average distance of
members in two clusters. It calculates a pairwise distance pdist(A, B) between
two clusters A and B. pdist(A, B) is computed as follows.
1 
pdist(A, B) = d(a, b) (1)
|A| · |B|
a∈A b∈B

d(a, b) is the distance between the element a and b. In each step, a cluster A
merges another cluster B so that pdist(A, B) is the minimum value. Finally, all
elements are in one cluster, and the merging process can be represented as a
binary tree.
First, each sink constructs a matrix of pdist({x}, {y}), where x, y ∈ H.
pdist({x}, {y}) is calculated by the Eq. 1. Note that the distance d(x, y)
is Euclidean distance. Then, two clusters {x} and {y} merge into one if
pdist({x}, {y}) is the minimum value and pdist({x}, {y}) ≤ T . This merging
phase is repeated until there is no candidate to merge. As a result, the sink
determines the past sink positions are biased if the size of the biggest cluster is
more significant than  |H|
2 .
In the case of detecting a bias of sink positions, the next sink s will be
relocated with the following condition, where C ⊆ H is the biggest cluster.

pdist(C, {s}) > T

5 Performance Evaluation
We measure the average cover ratio, which indicates the ratio of sensor nodes
from which MFs have collected data. It means that the efficiency of data collec-
tion by MFs. First of all, we show the cover ratio by MFs using Nomadic Lévy
442 K. Sugihara and N. Hayashibara

Walk (NLW) as a message ferry routing with the parameter α and γ. Then, we
compare the cover ratio with two message ferry routings; NLW and Random way
point (RWP).

5.1 Environment
In our simulation, we distribute 1,000 nodes at random in the 1,000 × 1,000
Euclidean plane. Each node has a fixed-sized circle r as its communication radius.
A link between two nodes exists if and only if the Euclidean distance of these
nodes is less than 2r. Thus, the parameter r is a crucial factor to form a unit
disk graph. We automatically generated unit disk graphs such that two nodes
have an undirected link if these circles based on r have an intersection.

5.1.1 Radius r of Unit Disk Graphs

We set r ∈ {35, 50} as a parameter of the environments in our simulations. r
is a compatible parameter with the diameter of a unit disk graph. Since nodes
obtain more links to other nodes that are not connected by gaining r, it tends
to become small. The average degrees of nodes deg(U DG) are 14.6 with r = 35
and 28.9 with r = 50, respectively. The diameters of unit disk graphs are 24
with r = 35 and 16 with r = 50.

5.2 Parameters
The scaling parameter λ and the homesick probability α are common to NLW
and HLW. We set λ = 1.2 for all simulations because it is known as the parameter
that realizes the efficient search on unit disk graphs [13].
The number of MFs is set as k ∈ {2, 5, 8}. Each MF has a particular sink
node (e.g., mobile power source vehicle).

5.3 Performance Criteria

We use the following criteria for measuring the performance of MFs.
Cover ratio: This is a ratio of V , and a vertex covers V  ⊆ V where ∀v ∈ V 

is visited by MFs at least once. Thus, the ratio is computed by VV . In this
paper, we calculate the average cover ratio of ten simulation results for each
configuration.
The cover ratio indicates how much MFs cover the field in terms of the ratio
of sensor nodes from which MFs have collected data.

5.4 Simulation Result

We show the simulation results of the cover ratio of NLW based MFs on unit disk
graphs. Then we compare the message delivery ratio with two routing schemes;
NLW and RWP.
 Message Ferry Routing Based on Nomadic Lévy Walk 443

5.4.1 On the Parameter α of Nomadic Lévy Walk

Figure 1, 2 and 3 show the cover ratio of k MFs using NLW as a message ferry
routing on unit disk graphs (UDGs) with r = 50 where k = 2, 5, 8, respectively.
The parameter α is a probability with which each MF goes back to the sink. On
the simulations, we configured the parameter of α in {0.2, 0.5, 0.8} and γ = 0.5
to observe the impact of the parameter α on the cover ratio.
According to the results, the average cover ratio is inverse proportional to
α. The cover ratio of α = 0.2 with k = 2 improves 23.5% at 2000 simulation
steps compared to that of α = 0.8. The gap between them is getting smaller by
increasing k. Then, it is only 3.8% with k = 8.
It means that the cover ratio improves with a lower probability that each MF
goes back to the sink. Moreover, the impact of α on the cover ratio is emphasized
with a small number of MFs. On the other hand, the impact could be negligible
by increasing the number of MFs k.

r = 50 r = 50 r = 50
100 100 100

80 80 80

60 60 60
Cover ratio

Cover ratio

40 40 Cover ratio 40

20 20 20

NLW (0.2, 0.5) NLW (0.2, 0.5) NLW (0.2, 0.5)

NLW (0.4, 0.5) NLW (0.4, 0.5) NLW (0.4, 0.5)
NLW (0.6, 0.5) NLW (0.6, 0.5) NLW (0.6, 0.5)
NLW (0.8, 0.5) NLW (0.8, 0.5) NLW (0.8, 0.5)
0 0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Simulation steps Simulation steps Simulation steps

Fig. 1. Cover ratio of Fig. 2. Cover ratio of Fig. 3. Cover ratio of

NLW with λ = 1.2, NLW with λ = 1.2, NLW with λ = 1.2, α ∈
α ∈ {0.2, 0.4, 0.6, 0.8} α ∈ {0.2, 0.4, 0.6, 0.8} {0.2, 0.4, 0.6, 0.8} and γ =
and γ = 0.5 on UDG and γ = 0.5 on UDG of 0.5 on UDG of r = 50
of r = 50 with k = 2. r = 50 with k = 5. with k = 8.

Figure 4, 5 and 6 show the cover ratio of an MF with NLW on unit disk
graphs with r = 35.

r = 35 r = 35 r = 35
100 100 100

80 80 80

60 60 60
Cover ratio

Cover ratio

Cover ratio

40 40 40

20 20 20

NLW (0.2, 0.5) NLW (0.2, 0.5) NLW (0.2, 0.5)

NLW (0.4, 0.5) NLW (0.4, 0.5) NLW (0.4, 0.5)
NLW (0.6, 0.5) NLW (0.6, 0.5) NLW (0.6, 0.5)
NLW (0.8, 0.5) NLW (0.8, 0.5) NLW (0.8, 0.5)
0 0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Simulation steps Simulation steps Simulation steps

Fig. 4. Cover ratio of Fig. 5. Cover ratio of Fig. 6. Cover ratio of

NLW with λ = 1.2, NLW with λ = 1.2, NLW with λ = 1.2, α ∈
α ∈ {0.2, 0.4, 0.6, 0.8} α ∈ {0.2, 0.4, 0.6, 0.8} {0.2, 0.4, 0.6, 0.8} and γ =
and γ = 0.5 on UDG and γ = 0.5 on UDG of 0.5 on UDG of r = 35
of r = 35 with k = 2. r = 35 with k = 5. with k = 8.
444 K. Sugihara and N. Hayashibara

The results have a similar tendency to the ones with r = 50. The average
cover ratio of α = 0.2 with k = 2 improves 20.3% at 2000 simulation steps
compared to that of α = 0.8. The gap between them is also getting smaller by
increasing k.

5.4.2 On the Parameter γ of Nomadic Lévy Walk

Figure 7, 8 and 9 show the cover ratio of k MFs using NLW as a message ferry
routing on unit disk graphs (UDGs) with r = 50 where k = 2, 5, 8, respectively.
On their simulations, we configured the parameter of γ in {0.2, 0.5, 0.8} and
α = 0.2 to observe the impact of the parameter γ on the cover ratio.
According to the results, The influence of the parameter γ is negligible on
UDGs of r = 50. The difference between them is at most 3% regarding the cover
ratio.

r = 50 r = 50 r = 50
100 100 100

80 80 80

60 60 60
Cover ratio

Cover ratio

40 40 Cover ratio 40

20 20 20

NLW (0.2, 0.2) NLW (0.2, 0.2) NLW (0.2, 0.2)

NLW (0.2, 0.5) NLW (0.2, 0.5) NLW (0.2, 0.5)
NLW (0.2, 0.8) NLW (0.2, 0.8) NLW (0.2, 0.8)
0 0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Simulation steps Simulation steps Simulation steps

Fig. 7. Cover ratio of Fig. 8. Cover ratio of Fig. 9. Cover ratio of

NLW with λ = 1.2, NLW with λ = 1.2, NWL with λ = 1.2, α =
α = 0.2 and γ ∈ α = 0.2 and γ ∈ 0.2 and γ ∈ {0.2, 0.5, 0.8}
{0.2, 0.5, 0.8} on UDG {0.2, 0.5, 0.8} on UDG on UDG of r = 50 with
of r = 50 with k = 2. of r = 50 with k = 5. k = 8.

Figure 10, 11 and 12 show the cover ratio of an MF with NLW on unit disk
graphs with r = 35. On UDGs of r = 35, the gap between the configurations of
γ widens compared to the one on UDGs of r = 50. However, the gap is still at
most 6% regarding the cover ratio.

r = 35 r = 35 r = 35
100 100 100

80 80 80

60 60 60
Cover ratio

Cover ratio

Cover ratio

40 40 40

20 20 20

NLW (0.2, 0.2) NLW (0.2, 0.2) NLW (0.2, 0.2)

NLW (0.2, 0.5) NLW (0.2, 0.5) NLW (0.2, 0.5)
NLW (0.2, 0.8) NLW (0.2, 0.8) NLW (0.2, 0.8)
0 0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Simulation steps Simulation steps Simulation steps

Fig. 10. Cover ratio of Fig. 11. Cover ratio of Fig. 12. Cover ratio of
NLW with λ = 1.2, NLW with λ = 1.2, α = NLW with λ = 1.2, α =
α = 0.2 and γ ∈ 0.4 and γ ∈ {0.2, 0.8} 0.6 and γ ∈ {0.2, 0.8} on
{0.2, 0.8} on UDG of on UDG of r = 35 with UDG of r = 35 with k =
r = 35 with k = 2. k = 5. 8.
 Message Ferry Routing Based on Nomadic Lévy Walk 445

5.4.3 Comparison of Nomadic Lévy Walk and Random Way Point

We compare the two message ferry routing schemes, NLW and RWP, regarding
the average cover ratio.
MFs with RWP are exploring a UDG heading to way-points randomly located
on the graph. On the other hand, MFs with NLW have to go back to the sink
with a given probability α.
We clarify the impact of this homing behavior on the cover ratio through the
comparison.
Figure 13, 14 and 15 show the average cover ratio of NLW and RWP on unit
disk graphs with r = 50 and Fig. 16, 17 and 18 show that with r = 35.
On the parameter of UDG r, the difference is negligible regarding the average
cover ratio.
RWP improves 12.6% and 15.4% with k = 2 at 1000 simulation steps regard-
ing the average cover ratio compared to NLW in r = 50 and 35, respectively.
However, the gap between them is less than 5% with k = 5, and it is almost the
same with k = 8.

r = 50 r = 50 r = 50
100 100 100

80 80 80

60 60 60
Cover ratio

Cover ratio

Cover ratio

40 40 40

20 20 20

NLW (0.2, 0.5) NLW (0.2, 0.5) NLW (0.2, 0.5)

NLW (0.2, 0.8) NLW (0.2, 0.8) NLW (0.2, 0.8)
RWP RWP RWP
0 0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Simulation steps Simulation steps Simulation steps

Fig. 13. The compari- Fig. 14. The compari- Fig. 15. The comparison
son of NWL and RWP son of NWL and RWP of NWL and RWP on
on UDG with k = 2. on UDG with k = 5. UDG with k = 8.

r = 35 r = 35 r = 35
100 100 100

80 80 80

60 60 60
Cover ratio

Cover ratio

Cover ratio

40 40 40

20 20 20

NLW (0.2, 0.5) NLW (0.2, 0.5) NLW (0.2, 0.5)

NLW (0.2, 0.8) NLW (0.2, 0.8) NLW (0.2, 0.8)
RWP RWP RWP
0 0 0
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Simulation steps Simulation steps Simulation steps

Fig. 16. The compari- Fig. 17. The compari- Fig. 18. The comparison
son of NWL and RWP son of NWL and RWP of NWL and RWP on
on UDG with k = 2. on UDG with k = 5. UDG with k = 8.

The cover ratio of NLW is worse than that of RWP with a small number of
MFs. The gap could be negligible if we deploy a few additional MFs. According
to the results, MFs with NLW have to go back to their sink; nevertheless, the
cover ratio with NLW could be almost the same as that with RWP by adding a
few MFs.
446 K. Sugihara and N. Hayashibara

6 Conclusion
In this paper, we evaluated the cover ratio of NLW-based MFs in intermittently
connected delay-tolerant wireless sensor networks (ICDT-WSNs). We also com-
pared NLW-based MFs with RWP-based MFs regarding the cover ratio.
NLW-based MFs improve the cover ratio with a lower probability α that each
MF goes back to its sink. On the other hand, the impact of γ on the cover ratio
is negligible.
On the comparison of different routing schemes, RWP-based MFs is better
than NLW-based MFs regarding the cover ratio, especially with a small number
of MFs (i.e., k = 2). However, the difference between them is negligible by
increasing k.
In this paper, we assume that each MF has a limited battery, and it should be
charged periodically. MFs using NLW have to go back to their sink; nevertheless,
the cover ratio with NLW could be almost the same as that using RWP by
adding a few MFs. NLW-based MFs are practically useful in this usage scenario
in ICDT-WSNs.

References
1. Alnuaimi, M., Shuaib, K., Abdel-Hafez, K.A.M.: Data gathering in delay tolerant
wireless sensor networks using a ferry. Sensors 15(10), 25809–25830 (2015)
2. Alnuaimi, M., Shuaib, K., Abdel-Hafez, K.A.M.: Ferry-based data gathering in
wireless sensor networks with path selection. Procedia Comput. Sci. 52, 286–293
(2015)
3. Basagni, S., Carosi, A., Melachrinoudis, E., Petrioli, C., Wang, Z.M.: Controlled
sink mobility for prolonging wireless sensor networks lifetime. Wireless Netw.
14(6), 831–858 (2008)
4. Bin Tariq, M.M., Ammar, M., Zegura, E.: Message ferry route design for sparse
ad hoc networks with mobile nodes. In: Proceedings of the 7th ACM International
Symposium on Mobile Ad Hoc Networking and Computing, MobiHoc 2006, pp.
37–48. Association for Computing Machinery (2006)
5. Birand, B., Zafer, M., Zussman, G., Lee, K.W.: Dynamic graph properties of mobile
networks under levy walk mobility. In: Proceedings of the 2011 IEEE Eighth Inter-
national Conference on Mobile Ad-Hoc and Sensor Systems, MASS 2011, pp. 292–
301. IEEE Computer Society (2011)
6. Dhal, K.G., Quraishi, M.I., Das, S.: A chaotic lévy flight approach in bat and firefly
algorithm for gray level image. Int. J. Image Graph. Sign. Process. 7, 69–76 (2015)
7. Edwards, A.M., Phillips, R.A., Watkins, N.W., Freeman, M.P., Murphy, E.J.,
Afanasyev, V., Buldyrev, S.V., da Luz, M.G.E., Raposo, E.P., Stanley, H.E.,
Viswanathan, G.M.: Revisiting lévy flight search patterns of wandering albatrosses,
bumblebees and deer. Nature 449, 1044–1048 (2007)
8. Fujihara, A., Miwa, H.: Homesick lévy walk and optimal forwarding criterion of
utility-based routing under sequential encounters. In: Proceedings of the Internet
of Things and Inter-cooperative Computational Technologies for Collective Intel-
ligence 2013, pp. 207–231 (2013)
 Message Ferry Routing Based on Nomadic Lévy Walk 447

9. Kavitha, V., Altman, E.: Analysis and design of message ferry routes in sensor
networks using polling models. In: 8th International Symposium on Modeling and
Optimization in Mobile, Ad Hoc, and Wireless Networks, pp. 247–255 (2010)
10. Li, Y., Bartos, R.: A survey of protocols for intermittently connected delay-tolerant
wireless sensor networks. J. Netw. Comput. Appl. 41, 411–423 (2014)
11. Rhee, I., Shin, M., Hong, S., Lee, K., Kim, S.J., Chong, S.: On the levy-walk nature
of human mobility. IEEE/ACM Trans. Netw. 19(3), 630–643 (2011)
12. Shin, M., Hong, S., Rhee, I.: DTN routing strategies using optimal search patterns.
In: Proceedings of the Third ACM Workshop on Challenged Networks, CHANTS
2008, pp. 27–32. ACM (2008)
13. Shinki, K., Hayashibara, N.: Resource exploration using lévy walk on unit disk
graphs. In: The 32nd IEEE International Conference on Advanced Information
Networking and Applications (AINA-2018) (2018)
14. Shinki, K., Nishida, M., Hayashibara, N.: Message dissemination using lévy flight
on unit disk graphs. In: The 31st IEEE International Conference on Advanced
Information Networking and Applications (AINA 2017) (2017)
15. Sokal, R.R., Michener, C.D.: A statistical method for evaluating systematic rela-
tionships. Univ. Kansas Sci. Bull. 38, 1409–1438 (1958)
16. Sugihara, K., Hayashibara, N.: Message dissemination using nomadic lévy walk on
unit disk graphs. In: Barolli, L., Hussain, F.K., Ikeda, M. (eds.) Proceedings of
CISIS 2020, pp. 136–147. Springer (2020)
17. Sugihara, K., Hayashibara, N.: Performance evaluation of nomadic lévy walk on
unit disk graphs using hierarchical clustering. In: Proceedings of the 34th Int’l
Conferences on Advanced Information Networking and Applications (AINA-2020),
pp. 512–522. Springer (2020)
18. Thejaswini, M., Rajalakshmi, P., Desai, U.B.: Novel sampling algorithm for human
mobility-based mobile phone sensing. IEEE Internet Things J. 2(3), 210–220 (2015)
19. Valler, N.C., Prakash, B.A., Tong, H., Faloutsos, M., Faloutsos, C.: Epidemic
spread in mobile ad hoc networks: determining the tipping point. In: Proceed-
ings of the 10th International IFIP TC 6 Conference on Networking - Volume Part
I, NETWORKING 2011, pp. 266–280. Springer-Verlag (2011)
20. Viswanathan, G.M., Afanasyev, V., Buldyrev, S.V., Murphy, E.J., Prince, P.A.,
Stanley, H.E.: Lévy flight search patterns of wandering albatrosses. Nature 381,
413–415 (1996)
21. Zhao, W., Ammar, M., Zegura, E.: A message ferrying approach for data delivery
in sparse mobile ad hoc networks. In: Proceedings of the 5th ACM International
Symposium on Mobile Ad Hoc Networking and Computing, pp. 187–198. Associ-
ation for Computing Machinery (2004)
22. Zhao, W., Ammar, M., Zegura, E.: Controlling the mobility of multiple data trans-
port ferries in a delay-tolerant network. In: IEEE INFOCOM 2005. Maiami, FL,
USA (2005)
 Image Similarity Based Data Reduction
Technique in Wireless Video Sensor
Networks for Smart Agriculture

Christian Salim(B) and Nathalie Mitton

Inria, Villeneuve d’Ascq, France

{christian.salim,nathalie.mitton}@inria.fr

Abstract. Nowadays, to improve animal well being in livestock farm-

ing or beekeeping application, a wireless video sensor network (WVSN)
can be deployed to early detect injury or Asiatic hornets attacks. WVSN
represents a low-cost monitoring solution compared to other technolo-
gies such as the closed circuit television technology (CCTV). WVSNs are
composed of low-power resource-constrained video sensor nodes (motes).
These nodes capture frames from videos at a given frequency (frame
rate) and wirelessly send them to the sink. The big amount of data
transferred from the nodes to the sink consumes a lot of energy on the
sensor node, which represents a major challenge for energy-limited nodes.
In this paper, we introduce two complementary mechanisms to reduce
the overall number of frames sent to the sink. First, the Transmission
Data Reduction algorithm (TDR) run on the sensor node leverages the
similarity degree of consecutive images. Second, the Inter-Nodes Simi-
larity algorithm (INS) exploits the spatio-temporal correlation between
neighbouring nodes in order reduce the number of captured frames. The
results show a 95% data reduction, surpassing other techniques in the
literature by 30% at least.

Keywords: Wireless video sensor networks · Data reduction ·

Geometric correlation · Inter-nodes similarity

1 Introduction
In recent years, the need for smart agriculture has increased [14]. Surveillance
systems are deployed to support it, especially in remote sites in order for instance
to monitor the animals behaviour, understand and protect the livestock and
vegetables farming. To achieve it, a wireless video sensor network (WVSN) can
serve as an energy-efficient and low cost solution. In WVSN, the system normally
operates periodically, sending all the captured frames to the sink [12]. However,
the system can be an event driven system sending only the frames that show
a change in the area of interest. The wireless video sensor nodes are energy-
limited nodes that monitor an area of interest according to their Field of Views
(FOVs) and send the videos/frames to the sink. Shooting a video and capturing
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 448–459, 2021.
https://doi.org/10.1007/978-3-030-75100-5_39
 Image Similarity Based Data Reduction in WVSN for SA 449

its frames consume a lot of energy but sending this great number of frames to the
sink is even worse. In order to cover all the area of interest and prevent black
holes, in general, WVSN are such that several FoVs overlap with each other.
To reduce the energy consumption at the sensor node level, we aim to reduce
the amount of data sent to the sink leveraging the overlap between FoVs and
by applying a node embedded lightweight algorithm in order to cope with the
nodes limited processing and energy resources.
In this paper, we propose a two-level data reduction approach. The first one,
Transmission Data Reduction algorithm (TDR), runs on each node and selects
only the critical frames to be sent to the sink. The second one, the Inter-Nodes
Similarity algorithm (INS), leverages the spatio-temporal correlation between
neighbour nodes to reduce the number of sent and captured frames. The TDR
algorithm compares consecutive captured frames on the node based on a norm
simple euclidean distance similarity method [11]. According to the degree of sim-
ilarity with the last frame sent to the sink, the node decides whether to send
this frame or not [12]. If the decision is to send the image, the node creates a
new image imgdif f that only includes the difference between the two compared
images. imgdif f is sent to the sink instead of the whole image. In this case, data
reduction is attained when several consecutive frames represent a high similarity
percentage (stable situation). To furthermore increase the data reduction, the
INS algorithm computes the correlated degree of neighbouring nodes and com-
pares their frames using the same norm simple euclidean distance metric. If the
result detects a similarity for n consecutive frames, the node with the highest
energy level sends the frames to the sink, the other one goes to idle mode and
its frame rate is set to the minimum F R = 1. We compare this new method to
another algorithm from the literature based on the colour and edge similarities
and on the number of consecutive frames sent per period [10]. The results show
that our approach reduces the amount of energy by 95% outperforming the other
approach, while capturing all the events.
The remainder of this paper is organised as follows. Section 2, browses the
related work. Section 3 describes the transmission data reduction algorithm while
Sect. 4 details the INS algorithm The experimental results and the comparison
with other methods are given in Sect. 5. Finally, Sect. 6 concludes and provides
perspectives and future work.

2 Related Work

In this section, we discuss some previous works regarding data reduction in

WSN and in WVSN in particular. A lot of techniques based on scheduling [3,6],
intrusion location and clustering [4,13], compressive sensing [16], data similar-
ity [10,11], and frame rate adaptation [7,15] exist for data reduction and for
energy-efficiency. The authors in [13] used the kinematics functions to predict
the next location of the intrusion in the area of interest in order to increase the
frame rate adaptation of the targeted nodes. This approach reduced the energy
dissipation of the overall network, since only the targeted nodes are capturing
450 C. Salim and N. Mitton

and not all the nodes in the zone of interest. This approach can come as a
complementary solution to our method which focuses on the nodes detecting
the intrusion and not on the projected positions of the intrusion. In [11] and
[15], the authors worked on data reduction by adapting the frame rate of the
nodes. This type of approaches is helpful and could be studied later on as a
complementary work of the approach presented in this paper. The geometric
correlation between several sensor nodes has been exploited by several authors
such as [1,8,9]. In [1], the authors transform the coordinates of every point from
3D to a 2D coordinate system according to some reference points. Computing
the disparity and the lengths of the reference vectors can determine whether
for example two cameras are geometrically correlated according to a predefined
threshold. In [10], the authors proposed a similarity function to detect the per-
centage of similarity between consecutive frames to attain data reduction by
sending the critical frames to the sink. They merge it with a correlation over-
lapping technique at the coordinator level. The proposed colour-edge similarity
function is time and energy consuming especially in the colour part where every
channel is taken into consideration. They compared the frames on the coordi-
nator level to further reduce the amount of frames. In this paper, the proposed
method reduces the number of frames sent from each node to the coordinator by
using the L2-norm distance to measure the similarity. It is a way less heavy to be
implemented at the sensor node unlike other techniques that use the colour-edge
and pixel similarities as in [7,10,15]. As explained in Sect. 3 it is a less complex
technique regarding the execution time comparison with other approaches for
the same purpose of data reduction. Unlike other approaches that use a 3D to
2D transformation to implement a similarity algorithm as in [1], we complement
the approach by proposing a 2-D spatial correlation between neighbour nodes
based on some geometrical conditions using simple equations. If 2 nodes are said
“correlated”, the Inter-Nodes Similarity algorithm (INS) runs on both nodes,
the one with the most residual energy sends the data, the other goes into idle
transmission mode and lowers its frame rate to its minimum F R = 1.

3 Data Reduction

In WSN, sensor-nodes operate periodically, and send a huge amount of data to

the coordinator simultaneously. This scenario consumes a lot of energy on each
sensor node representing the most important challenge in WSN in general. In
WVSN, packets sent are multimedia, which emphasises even more the challenge
of energy consumption. To overcome this challenge and to lower its importance
on the feasibility of the network deployment, we proposed a data reduction
method at the sensor node to reduce the number of raw images sent from each
node to the sink. This technique combines two mechanisms presented by the
Transmission Data Reduction algorithm (TDR) and the Inter-Nodes Similarity
algorithm (INS).
 Image Similarity Based Data Reduction in WVSN for SA 451

3.1 Transmission Data Reduction

The Transmission Data Reduction algorithm (TDR) is described in Algorithm 1.

TDR algorithm leverages the consecutive frames similarity as in [11], using the
norm simple euclidean distance similarity method to compare images. If two
consecutive sensed frames are estimated as similar (no new information is rep-
resented in the second frame), the node does not send the second frame to the
coordinator. The node counts the number of consecutive similar frames, if it
surpasses nb (the required number of consecutive similar frames), the state of
the area of interest monitored by the node is considered as stable (situation =
1). This technique uses the norm L2 relative error function tested in C++ in
OpenCV and compared to other techniques in [11] where the time to execute
this technique is way less important than other techniques (edge, colour, ...) for
the same purpose of data reduction. To apply this function, each image is trans-
formed into a matrix. If the similarity normsim = dist(f ri , f rlast ) between those
images does not surpass a certain threshold of similarity (normsim < thsim ), it
means the images are not similar. In this case, the sensor-node creates and sends
to the sink a new image called imgdif f by adopting the absolute images difference
function as shown in the equation below:

imgdif f = absdif f (M AT A, M AT B) (1)

This function computes the absolute value of the difference for each pixel, for
each channel of the two matrices. The imgdif f is an image that represents the
difference between both images using less bits. In this case, the size in Bytes of
the transmitted packets decreases because only the difference between images is
sent rather than the full new image.
An example of the imgdif f is shown in Fig. 2c representing the difference
image imgdif f of the frames presented in Figs. 2a and 2b taken from the dataset
in [5]1 . As we can see, the shadow of the moving person and the person are
the only differences taken into account. The last sent image to the coordinator
is stored on the sensor node. Every new frame is compared to this last frame
sent to the sink, it can replace this last frame if it represents a critical event or
a change in the monitored zone as shown in Algorithm 1. In the remainder of
this paper, the sent image imgdif f is called critical frame. We called it critical,
because the node sends only an image to the coordinator if an event or a change
is happening in the area of interest. This whole process is explained in Fig. 1.
Note that, the reconstruction algorithm at the sink level is out of the scope of
this paper.

1
http://changedetection.net.
452 C. Salim and N. Mitton

Fig. 1. Video sensor node and sink behaviours

(a) The First Frame (b) The Second Frame (c) The Difference Frame

Fig. 2. The difference scenario

Algorithm 1. Transmission Data Reduction Algorithm T DR on Sensor Node

Su
Set f rlast as the last sent frame to the sink
2: Set Situationu = 0
Set nbsim = 0 the number of consecutive similar frames
4: while Energy > 0 do
for each frame i do
6: if normsim < thsim then
nbsim = 0; f rlast = f ri
8: Sends imgdif f to the sink
else
10: nbsim = nbsim + 1
if nbsim > nb then
12: Situationu = 1 # Stable Situation
end if
14: end if
end for
16: end while
 Image Similarity Based Data Reduction in WVSN for SA 453

(a) Field of View (b) Two points Method (c) Angle Condition

Fig. 3. Overlapping model

4 Inter-Nodes Correlation: The Overlapping Method

In the above sections, we were interested in reducing the amount of sent frames
from each node to the sink by applying a similarity function at the sensor node
itself. To furthermore reduce the data transmission, we now focus on the spatial
correlation between neighbouring sensor nodes. In our approach, two nodes are
considered as neighbours if they are in the communication range of each other.
Neighbour nodes exchange their vector of metrics defined in the video sensing
model below to check whether their FoVs overlap enough to apply the learning
process and the similarity function. We apply the same norm distance similarity
function on frames captured by two overlapping neighbour sensor nodes. This
technique reduces the number of the overall captured and sent frames from the
nodes to the sink, by only sending one out of the two frames, depending on the
residual energy of the two nodes as shown in Algorithm 2. Before describing the
Inter-Nodes Correlation (INS) approach, we introduce the video sensing model,
the characteristics of every video sensor node and the geometric condition to
determine what nodes overlap and can thus proceed with the INS algorithm.

4.1 Video Sensing Model

We consider a 2-D model of a video sensor node where z = 0 (XOY plane)

and all the captured frames are compared as 2-D images not taking account of
the third dimension. A video sensor node S is represented by the Field of View
(FoV) of its camera. Unlike a wireless sensor which range is omnidirectional, a
FoV covers only a part of the surrounding area of a video sensor as shown in
−
→
Fig. 3a. A FoV is a vector of 4-tuple S(P, Rs , V , α) where P is the position of
→
−
S, Rs is its sensing range, V is the vector representing the line of sight of the
camera’s FoV e.g. its sensing direction, and α is the offset angle of the FoV.
In [10] the authors presented the FoV with 4 points a,b,c and the gravity
centre g as shown in Fig. 3b. The overlapping area is detected based on those
points as explained in the upcoming subsections. A point P1 is said to lay in
the FoV of a video sensor node S at position P if and only if the two following
conditions are satisfied:
454 C. Salim and N. Mitton

1. d(P, P1 ) ≤ Rs , where d(P, P1 ) is the Euclidean distance between P and P1 :

−−→
P P1  ≤ Rs
−−→ →
− −−→ −→ −−→ →
−
2. The angle between P P1 and V is within [−α, +α] : P P1 . V ≥ P P1 × V ×
cos(α).

In the remainder of this paper, we consider that all video nodes are identical,
with the same sensing range Rs and the same offset angle α. In the next section,
we discuss the spatial correlation of overlapping neighbour nodes based on some
geometrical conditions, and how we use it in order to increase the data reduction
at the sensor node level.

4.2 Overlapping Sensor Nodes

In this part, we describe the geometric condition which determines if two neigh-
bours are overlapping nodes. This condition is a combination of the angle con-
dition between the FoV’s of the nodes and the ratio of the overlapping area
between them.

4.2.1 The Angle Condition

The angle between two neighbouring sensor nodes is defined as the angle between
the vectors of their FoVs. If the angle between the FoVs of two neighbouring
nodes is wide, these two nodes can not take part in the similarity comparison
function to decide which one among them sends the data to the sink. Indeed,
that means that they are not sensing the same area of interest. A frame from two
different perspectives can be widely different. To be able to define two neighbour
sensor nodes as overlapping sensor nodes, the angle between their FoVs must
not surpass a certain angle threshold. In order to determine the angle between
vectors (V and V ) of the sensor nodes S and S  respectively as shown in Fig. 3c,
we use the scalar product between those sensor nodes. Since all sensor nodes are
identical, both vectors V and V have the same length l. The scalar product
can be calculated in two formats. The first one according to their coordinates
(x and y) where V = (XV , YV ) and V = (XV  , YV  ) :

V.V = XV × XV  + YV × YV  (2)
The second format is given according to the length of each vector and to the
angle between both, as follows:

V.V = l2 × cos(V, V ) (3)

−
→ →
−
Where l = V   =  V .
The below equation defines and computes the angle θ between the two vectors
according to both formats of the scalar product:

θ = arccos((XV × XV  + YV × YV  )/l2 ) (4)

 Image Similarity Based Data Reduction in WVSN for SA 455

If an angle threshold thangle is defined as 30◦ , the angle between V and V

must remain less than 30◦ so the two sensor nodes S and S  can proceed to the
next step (the two points strategy), to be able at the end to take part in the
similarity function process.

4.2.2 The Two Points Condition

Inspired from [10] we present below the two points condition as the second part
of the geometric condition. A node S  satisfies the two points condition with
another node S if g (the gravity centre of abc) belongs to the FOV of S alongside
any other point between a, b and c from S  s FOV as shown in Fig. 3b. In this
figure, S1 , S2 and S3 satisfy this condition separately with S. In this scenario,
each sensor node can be a candidate alongside S to apply the learning process
and the similarity function between them on the one having the most residual
energy. However, only S1 and S2 satisfy at the same time the angle and the two
points conditions. For this purpose, S1 and S2 will be the nodes selected as two
overlapping sensor nodes to test the inter nodes similarity algorithm IN S as
explained in the next section.

4.3 Inter-Nodes Similarity

As mentioned earlier, the comparison between the frames captured by overlap-
ping sensor nodes is done at the start of each stable situation in the zone of
interest. We used the same norm simple euclidean distance function to measure
the similarity. If the similarity computed ratio INsim is greater than a prede-
fined threshold INth for the first frame as shown in Algorithm 2, it means that
the two correlated nodes present the same information when shooting the zone
of interest. In this case, only one out of the two sensor nodes is selected to send
the data, depending on the residual energy RN of both nodes. If Su has more
energy left than Sv (RNSu > RNSv ), Su continues to send the data while Sv
goes to transmission idle mode and its frame rate is reduced to F R = 1 frame
per second on Sv . Once an event occurs in the FoV of Sv the frame rate is back
to its maximum of 30 frames per second. Figure 4 shows the behaviour of two
neighbour nodes while applying this overlapping technique and INS algorithm.

Algorithm 2. IN S run on Su
Set nsim = 0 the number of similar frames
Get situationv and f rv0 from Sv
3: if situationu = 1 AND situationv = 1 AND INsim > INth then
if RNSu > RNSv then
Su continues sending the frames running TDR
6: else
F Ru = 1 (and stop sending)
end if
9: end if
456 C. Salim and N. Mitton

Fig. 4. Two overlapping sensor nodes behaviour running INS

5 Experimental Results

Several simulations have been conducted to validate our approach at the sensor
node, aiming to minimise the number of frames sent from the node to the sink.
We compare our approach to MASRA and SSA algorithms in [10] and PPSS
algorithm in [2]. We have used a Matlab simulator. A scenario is introduced in
Fig. 3c where 2 video-sensor nodes S1, S2 are deployed as overlapped neighbour
nodes to monitor the same area of interest from different perspectives as shown
in Fig. 3c. The angle threshold is set to 15◦ , however in this case α = 10 and
S1 covers the gravity centre and one other point of the FoV of S2 , which makes
them two overlapping neighbour nodes in order to apply the INS algorithm on
both. The main purpose in our work is to send to the sink the frames that
represent the critical situations while maintaining the information by applying
TDR and INS algorithms. We have used 2 Microsoft LifeCam VX-800 cameras
to shoot a short video of 120 s (3600 frames captured per node), each camera is
connected to a laptop to do the processing via a Matlab simulator. Each camera
captures 30 frames/s. In our scenario, an intrusion is detected at the following
time-intervals:
S1: 40 s from 15 to 55 ; S2: 40 s from 20 to 60.
In the rest of the time, both sensor nodes had a stable situation. After conducting
several simulations, the number of high similarities needed to detect a stable
situation is set to n = 60 which means 2 s of stability. The similarity threshold
thsim is set to 85% in TDR algorithm and 75% in INS algorithm. Note that a
full compressed image size is equal to 30KB and the difference image imgdif f
 Image Similarity Based Data Reduction in WVSN for SA 457

Table 1. TDR algorithm over 120 s Table 2. Overall reduction

Node Situation Captured frames Sent frames Total frames Type frames Size frames
S1 Stable 2300 2 7200 Captured all 216000 KB
S1 Event 1200 240 504 Critical all 15120 KB
S2 Stable 2300 2 504 imgdif f 10080 KB
S2 Event 1200 260
ALL All 7200 504

size is 20 KB. The thresholds and all the parameters can be adapted according
to the application and the QoS required.

5.1 Transmission Data Reduction

Table 1 shows the number of frames sent by each node to the sink over 120 s
using the TDR algorithm. This algorithm allows the reduction of the number
and size of frames transmitted by the sensor-node as shown in Tables 1 and 2.
A transmission data reduction of more than 90% is reached by sending only
the critical frames. However the size reduction reached 95% due to the image
difference imgdif f applied on every selected/critical frame. Once in a stable
situation, the first frame is sent and due to similarity all the other frames are
not transmitted to the sink.

5.2 Inter-Nodes Similarity

The INS algorithm runs on both nodes at the same time with TDR algorithm.
In this section we consider that S1 has more residual energy than S2 , which puts
S2 into the idle mode and reduces its frame rate to F R2 = 1 frame per second.
In this case the data reduction is emphasised also on the sensing process for S2
as shown in Table 3. In this case, the number of frames sent to the sink remains
the same as in TDR alone, but the number of captured frames reduces by more
than 50% on S2 and 33% overall.

Table 3. TDR and INS-120 s Table 4. S1 and S2 data reduction

Node Situation Captured frames Sent frames Technique Captured Sent Size in KB
S1 Stable 2300 2 All data 7200 7200 216000
S1 Event 1200 240 MASRA+ SSA [10] 2540 1310 39300
S2 Stable 133 2 PPSS [2] 3600 3600 108000
S2 Event 1200 260 TDR + INS 4833 504 10080
ALL All 4833 504
458 C. Salim and N. Mitton

5.3 Comparison

We compare to the scheduling algorithm P P SS [2] and to the M ASRA algo-

rithm [10] where the colour and edge similarities were both present to detect the
difference between images. The difference between those approaches in terms of
data reduction is shown in Table 4. As we can notice from Table 4, TDR and
INS algorithms outperform MASRA algorithm [10] and PPSS algorithm [2] in
terms of transmission data reduction by 70% and 90% respectively and in terms
of the size of the data transmitted. This technique reduced the transmitted data
by 95% and the sensed data by 33% of the overall data captured and trans-
mitted. However, the sensing data reduction is better on MASRA algorithm by
45% which implements the frame rate adaptation of each node according to the
scene but computes the geometric correlation on the sink not on the node, in
this purpose they transmit 3 times more than our current approach. The overall
gain remains positive by more than 30%.

6 Conclusion

In this paper, we introduced a distributed data reduction approach for wireless

video sensor network. This approach is composed of two complementary algo-
rithms, the transmission reduction algorithm (TDR) and the inter-nodes similar-
ity (INS) algorithm. TDR detects the similarities between consecutive captured
frames and decides to send only the difference image of each critical frame to the
sink. The results show that the proposed algorithms did not miss any event in
the recorded video sequence. The second part of the approach consists in a geo-
metrical condition computed by neighbour nodes to detect whether they overlap
and reduce accordingly their frame rate or transmission rate. Comparing our
approach with MASRA and PPSS algorithms in terms of data reduction shows
that our algorithm outperforms both algorithms by 30% and 90% respectively.
As future works, first of all we aim to compare the relevance of our mathemat-
ical model by comparing it to other approaches. In order to increase the data
reduction, the frame rate adaptation technique at the sensor node level can be
taken into account.

Acknowledgements. This work was partially supported by a grant from CPER

DATA, by LIRIMA Agrinet project and the Inria/CEFIPRA Associate team DC4SCM.

References
1. Dai, R., Akyildiz, I.F.: A spatial correlation model for visual information in wireless
multimedia sensor networks. IEEE Trans. Multimedia 11(6), 1148–1159 (2009)
2. Elson, J., Girod, L., Estrin, D.: Fine-grained network time synchronization using
reference broadcasts. SIGOPS Oper. Syst. Rev. 36, 147–163 (2002)
3. Feng, J., Zhao, H.: Energy-balanced multisensory scheduling for target tracking in
wireless sensor networks. Sensors 18(10), 3585 (2018)
 Image Similarity Based Data Reduction in WVSN for SA 459

4. Ghosal, A., Halder, S.: A survey on energy efficient intrusion detection in wireless
sensor networks. J. Ambient Intell. Smart Environ. 9, 239–261 (2017)
5. Goyette, N., Jodoin, P.M., Porikli, F., Konrad, J., Ishwar, P.: Changedetection.net:
a new change detection benchmark dataset. In: IEEE Workshop on Change Detec-
tion (CDW-2012) (2012)
6. Jiang, B., Ravindran, B., Cho, H.: Probability-based prediction and sleep schedul-
ing for energy-efficient target tracking in sensor networks. IEEE Trans. Mob. Com-
put. 12(4), 735–747 (2013)
7. Monika, R., Hemalatha, R., Radha, S.: Energy efficient surveillance system using
WVSN with reweighted sampling in modified fast haar wavelet transform domain.
Multimed. Tools Appl. 77(23), 30187–30203 (2018)
8. Mowafi, M.Y., Awad, F.H., Aljoby, W.A.: A novel approach for extracting spa-
tial correlation of visual information in heterogeneous wireless multimedia sensor
networks. Comput. Netw. 71, 31–47 (2014)
9. Nu, T.T., Fujihashi, T., Watanabe, T.: Power-efficient video uploading for crowd-
sourced multi-view video streaming. In: IEEE Global Communications Conference
(GLOBECOM) (2018)
10. Salim, C., Makhoul, A., Couturier, R.: Similarity based image selection with frame
rate adaptation and local event detection in wireless video sensor networks. Mul-
timed. Tools Appl. 78, 5941–5967 (2019)
11. Salim, C., Makhoul, A., Couturier, R.: Energy-efficient secured data reduction tech-
nique using image difference function in wireless video sensor networks. Multimed.
Tools Appl. 79, 1801–1819 (2020)
12. Salim, C., Makhoul, A., Darazi, R., Couturier, R.: Combining frame rate adapta-
tion and similarity detection for video sensor nodes in wireless multimedia sensor
networks. In: IWCMC (2016)
13. Salim, C., Makhoul, A., Darazi, R., Couturier, R.: Kinematics based approach
for data reduction in wireless video sensor networks. In: International Conference
on Wireless and Mobile Computing, Networking and Communications (WiMob)
(2018)
14. Salim, C., Mitton, N.: Machine learning based data reduction in WSN for smart
agriculture. In: International Conference on Advanced Information Networking and
Applications, pp. 127–138. Springer (2020)
15. Tayeh, G.B., Makhoul, A., Demerjian, J., Laiymani, D.: A new autonomous data
transmission reduction method for wireless sensors networks. In: IEEE Middle East
and North Africa Communications Conference (MENACOMM) (2018)
16. Veeraputhiran, A., Sankararajan, R.: Feature based fall detection system for elders
using compressed sensing in WVSN. Wireless Netw. 25, 287–301 (2019)
 Using Clusters for Approximate
Continuous k-Nearest Neighbour Queries

Wendy Osborn(B)

Department of Mathematics and Computer Science, University of Lethbridge,

Lethbridge, AB T1K 3M4, Canada
[email protected]

Abstract. This paper proposes several strategies for continuous k-

nearest neighbour query processing in location-based services. All uti-
lize a clustered dataset to create safe regions. The only previously pro-
posed strategy to do so has limited validation of the safe region on the
client, which leads to only approximate k-nearest neighbour results being
obtained. This work improves upon both the creation of the safe region
and the validation of the safe region on the client. An evaluation of our
strategy, including a comparison versus the existing strategy, show sig-
nificant improvements in accuracy, and up to over 90% accuracy. When
compared to repeated k-nearest neighbour search, our strategy is compu-
tationally significantly faster for larger datasets. Obtaining significantly
high accuracy at low computational costs is significant for our strategies.

Keywords: Continuous query processing · k-nearest neighbour ·

Approximate · Clustering · Location based services

1 Introduction
Location Based Services (LBS) [16] obtain and display information for a user on
their mobile device (e.g., smartphone) from a remote server, given the type of
information being sought and their current location. An example of such a query
is the locations of the nearest k restaurants, given the user’s current location. If
the user stays in the same location, a one-time result containing the nearest k
restaurants can be obtained from the server and transmitted to them. However,
if the moving by foot or by car, the initial result will become invalid. Therefore,
the location of the nearest k restaurants must be repeatedly updated for the
user who is on the move. A continuous k-nearest neighbour query [12] takes a
sequence of a user’s locations on their trajectory and transmits them to a remove
server to retrieve and/or update the actual k nearest points of interest (POIs)
for to the user. However, an approximate continuous k-nearest neighbour query
considers the trade-off between accuracy and efficiency by retrieving k POIs
that are not guaranteed to be the closest to the user, but are considered “close
enough” to meet their needs. Referring back to the example, the exact k nearest
restaurants may be sought by the user as they are moving. However, the user

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 460–471, 2021.
https://doi.org/10.1007/978-3-030-75100-5_40
 Using Clusters for Approximate Continuous k-Nearest Neighbour Queries 461

may be satisfied with restaurants which are not be the actual k closest to them
but still within walking or driving distance to them.
Strategies for handling continuous k-nearest neighbour queries have been
proposed previously [1,3–14]. These can be classified as one of the following:
– Repeated application of a standalone k-nearest neighbour strategy when the
location of the user changes.
– Storing a subset of data on the mobile device, so that calls to the server can
be reduced or minimized.
– Storing a safe region (and corresponding set of POIs that it contains) on the
mobile device, so that calls to the server can be reduced or minimized. In a
safe region, the set of points remains a valid result (e.g., a set of k-nearest
neighbours) while the query remains inside of it.
– A hybrid approach that combines safe region and caching.
In addition, previously proposed continuous k-nearest neighbour strategies have
one or more of the following limitations:
– Searching data or data structures (e.g., spatial access methods) repeatedly,
– Caching too much data on the user’s mobile device,
– Complex strategies required to generate new safe regions when needed,
– Upfront knowledge of the user’s travel trajectory, and
– Lack of use of a clustered dataset, which can help improve performance.
Specifically, most existing strategies do not consider the last limitation, which
is the use of clustering in order to improve the performance of continuous query
processing. A clustered dataset can assist in improving the performance of cre-
ating a safe region that resides on the user’s mobile device. The only previously
proposed strategy that considers the use of a clustered dataset for processing
continuous k-nearest neighbour queries [14] is an approximate strategy. It only
selects the clusters that overlap the query point, and only considers whether the
query point is still inside the selected clusters. This alone does not guarantee an
accurate k-nearest neighbour result, as the distance from the query point to the
k th nearest neighbour may be greater than the distance to one or more sides of
the safe region [14]. Therefore, the accuracy was found to be approximately 60%
at worst, with at worst 80% of queries producing inaccurate results [14].
Therefore, in this paper several new strategies for continuous k-nearest neigh-
bour query processing are proposed that improve upon the previously proposed
strategy in [14]. Two aspects are the focus of this paper: the creation of the
safe region using the clustered point set on the server, and additional validity
testing of a safe region on the mobile device (i.e., client). An experimental evalu-
ation show some significant improvements in accuracy over the strategy in [14] -
up to almost 100% in some cases. In addition, our strategies are computationally,
significantly faster for larger datasets over repeated searching.

2 Background
In this section, some background is summarized that is utilized for the work to be
proposed here. First, the concepts clusterMBRs and superMBR are presented.
462 W. Osborn

118.32 , 4.83 , 141.42 , 77.61 , C0 118.32 , 77.61 , 118.32 , 77.61 , 1

55.62 , 218.54 , 86.18 , 228.24 , C1 123.34 , 69.35 , 123.34 , 69.35 , 2
1.75 , 38.69 , 56.01 , 85.18 , C2 127.79 , 60.78 , 127.79 , 60.78 , 3
166.11 , 115.14 , 204.77 , 166.11 , C3 131.64 , 51.95 , 131.64 , 51.95 , 4
212.15 , 55.62 , 228.24 , 100.89 , C4 134.87 , 42.81 , 134.87 , 42.81 , 5
60.77 , 85.50 , 112.75 , 127.78 , C4 137.48 , 33.51 , 137.48 , 33.51 , 6
100.89 , 177.06 , 154.39 , 212.15 , C5 139.45 , 24.05 , 139.45 , 24.05 , 7
4.83 , 131.63 , 51.91 , 141.42 , C6 140.76 , 14.48 , 140.76 , 14.48 , 8

Fig. 1. Cluster MBRs Fig. 2. Point set of cluster C0

Then, the strategy proposed in [14] is summarized, as this is the strategy used for
comparison purposes in the evaluation (Sect. 4). Finally, the clustering strategies
which are used in this work are summarized.

2.1 Safe Regions, ClusterMBRs and the SuperMBR

A safe region is a subregion of space (and the points or objects that reside in it)
in which the result of the query remains valid while the query continues residing
within the region. This approach is utilized by many strategies that process
continuous spatial queries in order to minimize the number of requests to the
server that must be made. One such type of safe region is a superMBR, which
is formed by utilizing one or more clusterMBRs. A clusterMBR is a minimum
bounding rectangle (MBR) that contains a set of points that were formed by a
clustering strategy. Therefore, a clusterMBR is created for each cluster in a set,
and will represent that cluster. First proposed in [15] for continuous region query
processing, the superMBR is also utilized for approximate continuous k-nearest
neighbour query processing in [14].
Figures 1 and 2 depict a set of clusterMBRs, and the points that are contained
by cluster 0, The lower left-hand corner of a clusterMBR (lx,ly) is represented
by the first two values, while the upper right-hand corner (hx,hy) is represented
by the next two values. The last value in each record is the identifier for each
cluster or point. For each record in Fig. 2, the first and third value are the same,
as are the second and fourth values, since each is representing a point.

2.2 Continuous k-Nearest Neighbour Using the SuperMBR

The approximate continuous k-nearest neighbour strategy proposed in [14] uti-
lizes a clustered dataset on the server to create a superMBR. The strategy takes
a query point q and transmits it to a remote server, which creates an initial
superMBR and corresponding point set to send back to the client. The client
displays the k nearest neighbours from this set. This superMBR is considered
valid until q moves outside of it. Whenever this happens, another request for a
new superMBR and point set is made to the server.
On the server, one or more clusterMBRs are selected for creating the
superMBR. In [14], two scenarios are considered: 1) q overlaps one or more
 Using Clusters for Approximate Continuous k-Nearest Neighbour Queries 463

(a) Valid

(b) Invalid

Fig. 3. SuperMBR creation Fig. 4. Valid and invalid cases

clusterMBRs, and 2) q overlaps no clusterMBRs. Figure 3 (from [14]) depicts

both scenarios. As we can see, query Q1 overlaps clusterMBRs C4 and C5. Both
are selected for creating the superMBR, which is represented in the figure by
the orange rectangle that encloses both clusterMBRs. Query Q2, however, does
not overlap any clusterMBR. In this scenario, the closest clusterMBR - C6 - is
chosen, and the superMBR encloses both C6 and Q2. This is done in an attempt
to reduce the number of requests made to the server.
One significant limitation of this strategy is with the validation approach for
keeping an existing superMBR on the client, which is a simple test to ensure
that q is still within the bounds of the superMBR. Figures 4a and 4b (from
[14]) depicts some scenarios that can arise from this simple verification. In Fig.
4a we see the scenario in which the k th nearest neighbour is safely inside the
superMBR. However, in Fig. 4b, we see that the distance between q and the k th
nearest neighbour is greater than the distance between q and one side of the
superMBR. As we can see from the circle formed from the k th on the outer edge,
some of the circle resides outside of the superMBR. It is possible that one or
more points in this area are closer to q, which makes the result in the superMBR
invalid. In addition, Fig. 3 depicts the scenario in which a clusterMBR - C3 -
overlaps the superMBR created for Q1. However, since C3 did not overlap Q1,
it was not chosen to be part of the superMBR. The problem here is that C3
contains points that are part of a valid result, but will not be transmitted to the
server.
464 W. Osborn

2.3 Clustering Strategies

The work to be proposed in the next section can utilize a set of points that
is clustered with any clustering strategy. For the evaluation, the following clus-
tering strategies are utilized to see how they affect the strategies [2]: KMeans,
Expectation-Maximization (EM), Farthest-First (FF), Density and Canopy. The
implementation of these strategies in the WEKA1 data mining software was uti-
lized. WEKA was chosen due to the author’s familiarity with it, and also due to
its stable implementation.
KMeans uses partitioning to split a set of points into k clusters. First, k
seed points are chosen randomly from the set of points, before all other points
are assigned to the closest seed point. Then, a centroid point is calculated for
each new cluster. Then, each point is re-assigned to a different cluster if it is
closer to a different centroid. This is repeated while points continue to move
between clusters. The partitioning-based method EM performs almost identically
to KMeans. However, instead of using a distance function for assigning points
to a cluster, EM uses a probability distribution function. FF also performs very
similarly to KMeans. The main difference is that only the first seed point is
chosen randomly. The other k − 1 seeds are the furthest away from each other
and the first point seed. For the Density strategy, a base clustering algorithm
(such as KMeans) is utilized to fit the data to a specified distribution. Finally, the
Canopy strategy partitions the data into proximity regions, using two distances
– one that dictates the distance a point can be from the centre of a cluster, and
another that dictates the number of clusters formed.

3 Strategies for SuperMBR Creation and Validation

In this section, four strategies for approximate k-nearest neighbour query pro-
cessing are proposed. All utilize a clustered set of points on the server to create
a superMBR and corresponding point set for the client. In addition to utilizing
the strategy for superMBR creation from [14], the strategies utilize an additional
superMBR creation strategy, and a new client-side validation strategy. Both are
proposed below.

3.1 Creation of SuperMBR on Server

The new strategy for creating a superMBR attempts to address the issue of
missing points due to overlap with the superMBR but not with the query point.
In addition, this strategy attempts to reduce the number of requests to the server
that the client must make by incorporating additional clusterMBRs into the
superMBR (i.e., which also leads to storing extra points on the client). However,
a trade-off does exist since memory is limited on most mobile devices. To control
this, the value β specifies the number of clusterMBRs that the superMBR can
be composed of altogether.
1
http://www.cs.waikato.ac.nz/ml/weka.
 Using Clusters for Approximate Continuous k-Nearest Neighbour Queries 465

Fig. 5. Overlap Fig. 6. No overlap

Both scenarios—initial overlap, and no initial overlap—are addressed in

this strategy. First, the query q is tested for overlap against the set of clus-
terMBRs. If one or more clusterMBRs overlap q, then a superMBR is created
that encompasses the initial chosen clusterMBRs. Then (if necessary), addi-
tional clusterMBRs are incorporated into the superMBR. This next step is
performed by testing the remaining clusterMBRs for overlap with the exist-
ing superMBR. If any clusterMBRs are found to overlap the superMBR, the
superMBR is expanded to enclose these clusterMBRs. This process repeats until
β clusterMBRs have been incorporated. Figure 3 depicts an example for β = 3.
Earlier, we saw that C3 overlapped the initial superMBR. It is now incorporated
into the superMBR.
If no clusterMBRs are found to overlap q, then the β closest clusterMBRs to
q are chosen for creating the superMBR. Figure 6 depicts an example for β = 3.
The clusterMBRs C6, C7 and C3 are the closest to Q2.

3.2 Validity Test of SuperMBR on Client

The strategy proposed for testing the validity of a superMBR extends upon the
strategy from [14] by also considering the boundaries of the superMBR with
respect to q. As mentioned earlier, the distance of the k th nearest neighbour
from q (i.e., dist(q, k th nn)) may be greater than the distance from q to one or
more sides of the superMBR (i.e., dist(q, side), where side = lx, hx, ly, or hy).
This may lead to points just outside of the boundary actually being closer than
the k th nearest neighbour. However, it may be possible that the k th nearest
neighbour since the superMBR is valid, and no points outside of the boundary
exist. To allow for some flexibility in this situation, the value α is proposed that
specifies the difference between dist(q, k th nn) and dist(q, side) that is permitted:

α <= dist(q, k th nn) − dist(q, side), for all side = lx, ly, hx, hy (1)
466 W. Osborn

#Point #Queries
500 20
1,000 28
5,000 63
10,000 89
50,000 198
100,000 283

Fig. 7. Validation strategy Fig. 8. #Queries per data size

Figure 7 depicts this for α <= dist(q, k th nn) − dist(q, hy). For no flexibility,
α = 0 (i.e., the k th nearest neighbour must be properly within the boundaries
of the superMBR. Setting α to anything greater than zero will permit some
flexibility. Our strategies incorporate both the strict and the flexible use of α.

3.3 Four Strategies

The strategies above for superMBR creation and validity testing, and the strat-
egy for superMBR creation from [14] are applied in various way in the following
four strategies:

– Strategy 1: Validity Test (α = 0), superMBR creation strategy from [14],

– Strategy 2: (α = 1), superMBR creation strategy from [14],
– Strategy 3: (α = 0), new superMBR creation strategy,
– Strategy 4: (α = 1), new superMBR creation strategy.

4 Evaluation

In this section, all four cluster-based continuous k-ne arest neighbour query
processing strategies that are proposed above are evaluated. All strategies are
evaluated for accuracy. In addition, all strategies are compared again the cluster-
based approach proposed in [14] that is summarized above in Sect. 2. Finally,
some observations concerning the running times of the strategies is presented.
All experiments utilize a simulated client-server application that was imple-
mented on an IntelR CoreTM2 Duo CPU E8757 with 2.66 GHz CPU and 1835
MB RAM, running the CentOs Linux 7 OS.
Sets of n points were √ all experiments. Each set of n points is drawn
√ utilized for
from a region of size n ∗ 10 x n ∗ 10 units. Altogether, twelve sets of points
were generated, with six sets assuming a random distribution and containing 500,
1,000, 5,000, 10,000, 50,000, and 100,000 points respectively, and the remaining
six sets assuming an exponential (i.e., skewed) distribution and containing the
same number of points as the uniform data sets.
 Using Clusters for Approximate Continuous k-Nearest Neighbour Queries 467

For clustering our sets of points, we utilized the KMeans, Density (with
a KMeans core), EM, FF, and Canopy implementations from WEKA. Other
clustering algorithms existed in WEKA. However, they either did not work for
larger sets of points, or they did not have the option to state the number of
clusters, which was required for one set of tests below.
For all tests, a trajectory of query points that proceeds diagonally through
the same space occupied by the points is utilized to evaluate all strategies. The
number of query points for each trajectory depends on the number of points in
the dataset, which is depicted in Fig. 8. The same trajectory is utilized for both
the random and exponential sets of points of the same size.
The four tests that are conducted vary:

– The number and distribution of points,

– The number of nearest neighbours,
– The number of clusters formed in the set of points, and
– The clustering algorithm utilized to create the clusters.

In addition, the following percentage accuracies are calculated:

– Perfect: The percentage of completely accurate results (i.e. all k-nearest

neighbours were correct),
– Overall: The percentage of overall accurate results, which includes both the
completely accurate results and the partially correct results for queries that
did not produce a set of completely accurate k nearest neighbours, and
– Imperfect: The percentage of accurate results in queries that did not result
in completely true k nearest neighbours.

Due to space limitations, only the first two percentages are reported for a
subset of the results that were generated during the evaluation. For each subset of
results presented, there are five pairs of accuracy values. The first pair belongs to
the strategy originally proposed in [14] while pairs 2 to 5 belong to the strategies
Strategy 1 to Strategy 4 (see end of Sect. 3), respectively.

4.1 Results-Varying Data Size and Distribution

For these tests, the size of the point set varied between 500 and 100000 points.
The number of nearest neighbours was set at 3, and the number of clusters set
at 50 and created using KMeans clustering. Figures 9 and 10 depict the results
of the 10000, 50000 and 100000 point sets for both uniform and exponential
distributions, respectively. Overall, most strategies are achieving at least 80%
accuracy, with some achieving 90+% accuracy.
For the larger data sets, results show modest to significant improvements
for most of the proposed strategies over the original strategy. The accuracy for
perfect results in the uniform point sets (Fig. 9) is approximately 11–12% higher
than the original strategy for 10000 and 100000 points, while it is significantly
higher for 50000 points, at almost 55% higher in most cases. The only exception
– Strategy 2 for 100000 points—is significantly lower. As Strategy 2 permits the
468 W. Osborn

Fig. 9. Number of objects vs accuracy (uniform)

Fig. 10. Number of objects vs accuracy (exponential)

“sphere” of the k th result to extend outside of the superMBR, valid points may
exist in this small region of space that are not being retrieved when needed.
With a stricter requirement of the “sphere” being inside the superMBR (as with
Strategy 1), these situations would be avoided. In addition, overall we see similar
trends with the overall percentage of accuracy values.
For the exponential point sets (Fig. 10), we see more significant improve-
ments across all of the strategies for both perfect and overall accuracies. The
50000 point set sees the most modest improvement at 17%, while the 10000 and
100000 point sets achieving more significant improvements at between 27–35%.
For the tests on point set below 10000 points, the improvements are negligible for
the most part—the only exception being the exponential 5000 point set, which
achieves approximately a 50% improvement across all strategies. Therefore, given
the results the strategies benefit the larger point set sizes the most.

4.2 Results-Varying Number of Clusters

For this set of tests, the number of clusters varied between 10 and 100. The
uniformly distributed 10000-point set was used, as well as 3 near neighbours
 Using Clusters for Approximate Continuous k-Nearest Neighbour Queries 469

Fig. 11. Number of clusters vs. accuracy

Fig. 12. Number of nearest neighbours vs. accuracy

and KMeans clustering. Figure 11 depicts the results that used 10, 30, 50, 70
and 90 clusters. From the results, we see that the more significant improvements
happen in the smaller and larger number of clusters. For 10 and 30 clusters,
improvements of between 17–21% are achieved, while those for 90 clusters are
between 20–30%. Notice again that the one instance of poor performance occurs
with Strategy 2. This also occurs in the results for some of the other clusters not
shown here. Again, this is probably due to the nature of the data, and having
some too close to other clusters that were fetched while the cluster those points
resided in was overlooked due to the non-negative α value.

4.3 Results-Varying Number of Nearest Neighbours

For this set of tests, the number of nearest neighbours varied between1 and 5.
The uniformly distributed 10000-point set was used, as well as 50 clusters created
using KMeans clustering. Figure 12 depicts the results. It appears that the value
of k does not significantly affect the improvement in accuracy. Although there is
improvement as the value of k increases, the improvement is approximately 8%
at the lower end and at most 16% at the upper end of k. However, the accuracies
are 90% in many cases.
470 W. Osborn

4.4 Results-Varying Clustering Strategy

Finally, for this last set of tests varied the strategies used for creating the clusters.
Again, the uniformly distributed 10000-point set was used, as well as 50 clusters
and 3 nearest neighbours. Results (not shown due to space limitations) show
modest improvements of between 5–10% across all strategies, except for Fastest
First (FF), which achieved more significant improvements—between 26–33%.
FF was found to be the poorest performing clustering strategy with respect to
the original strategy [14], so this may explain the improvement. The issues that
exist in an FF clustering of data may have been addressed by the new strategies.
Finally, overall, most accuracies were at least 80%, with several over 90%.

4.5 Running Time

Data concerning the running time of our strategies was not recorded. However,
some observations were noted during the evaluation of the above strategies. First,
even though the superMBR creation and validity test strategies proposed here
involve more computation that that proposed in [14], all strategies appeared to
execute in a similar amount of time and did not appear to take more time than the
original proposed strategy. This is significant since the strategy in [14] was found
to run significantly faster than repeated k-nearest neighbour searching, especially
for the larger point sets. Repeated searching took up to 5–6 h to execute, where the
original strategy, and the new ones proposed here, took a matter of a few seconds
to execute. Therefore, even though the new strategies require more work, they are
not increasing the running time in a noticeable way at all.

5 Conclusion

In this paper, several new strategies for approximate continuous k-nearest neigh-
bour query processing are proposed. All utilize a clustered set of points on the
server to create a superMBR and corresponding point set for the client. In addi-
tion to utilizing the original strategy for superMBR creation from [14], the strate-
gies utilize an additional superMBR creation strategy, and a new client-side vali-
dation strategy. Both new superMBR strategies attempt to overcome limitations
of the original strategy by creating a larger save superMBR, as well as additional
and flexible validity testing of the superMBR on the client.
The new strategies were evaluated and compared again the original strategy.
Results show that: 1) in many instances, at least 80% (and in some case, over
90%) accuracy in both the number of perfect results, as well as the overall
accuracy, was achieved, 2) up to 50% improvement over the original strategy
was achieved, and 3) the new strategies work very well for larger point sets. One
limitation that was discovered was for Strategy 2, which achieved most of the
worst performance among the strategies.
Some new research directions emerged from this work. The first is to explore
additional extensions to superMBR creation so that results can be 100% accurate
 Using Clusters for Approximate Continuous k-Nearest Neighbour Queries 471

more often than what is being accomplished now. The second is to explore the
use of other types of clustering strategies, as all of those we applied here are
partitioning methods. In particular, it would be interesting to see how density-
based clustering strategies would apply to our work. Finally, further evaluation
of our work versus other safe-region based approaches is required. In addition,
further evaluation is required to determine the running time, the amount of
data transmission to and from the server, how many times a new safe region is
required, etc., of all strategies.

References
1. Cheng, R., Lam, K.Y., Prabhakar, S., Liang, B.: An efficient location update mech-
anism for continuous queries over moving objects. Inf. Syst. 32(4), 593–620 (2007)
2. Frank, E., Hall, M., Witten, I.: The weka workbench (2016). https://www.cs.
waikato.ac.nz/ml/weka/Witten et al 2016 appendix.pdf. Accessed 14 Feb 2021
3. Gao, Y., Zheng, B.: Continuous obstructed nearest neighbor queries in spatial
databases. In: Proceedings of the 2009 ACM SIGMOD International Conference
on Management of Data, pp. 577–590 (2009)
4. Gupta, M., Tu, M., Khan, L., Bastani, F., Yen, I.L.: A study of the model and
algorithms for handling location-dependent continuous queries. Knowl. Info. Syst.
8(4), 414–437 (2005)
5. Huang, Y.K., Chen, C.C., Lee, C.: Continuous k-nearest neighbor query for moving
objects with uncertain velocity. GeoInformatica 13(1), 1–25 (2009)
6. Huang, Y.K., Chen, Z.W., Lee, C.: Continuous k-nearest neighbor query over mov-
ing objects in road networks. In: Proceedings of the International Conference on
APWeb/WAIM, pp. 27–38 (2009)
7. Ilarri, S., Bobed, C., Mena, E.: An approach to process continuous location-
dependent queries on moving objects with support for location granules. Jnl. Syst.
Soft. 84(8), 1327–1350 (2011)
8. Iwerks, G.S., Samet, H., Smith, K.P.: Maintenance of k-nn and spatial join queries
on continuously moving points. ACM Trans. Database Syst. 31(2), 485–536 (2006)
9. Ku, W.S., Zimmermann, R., Wang, H.: Location-based spatial query processing
with data sharing in wireless broadcast environments. IEEE Trans. Mob. Comput.
7(6), 778–791 (2008)
10. Lam, K.Y., Ulusoy, Ö.: Adaptive schemes for location update generation in execu-
tion location-dependent continuous queries. Jnl. Syst. Soft. 79(4), 441–453 (2006)
11. Liu, F., Hua, K.A.: Moving query monitoring in spatial network environments.
Mob. Netw. Appl. 17(2), 234–254 (2012)
12. Mouratidis, K., Papadias, D.: Continuous nearest neighbor queries over sliding
windows. IEEE Trans. Knowl. Data Eng. 19(6), 789–803 (2007)
13. Osborn, W.: Continuous k-nearest neighbour strategies using the MQR-tree. In:
Proceedings of the 21st International Conference on Network-Based Information
Systems (2018)
14. Osborn, W., Anderson, C.: Approximate continuous nearest neighbour query pro-
cessing in clustered point sets. In: Proceedings of the 11th Annual IEEE IEMCON
Conference (2020)
15. Osborn, W., Keykavoos, F.: Continuous region query processing in clustered point
sets. In: Proceedings of the 16th International Conference on Mobile Systems and
Pervasive Computing, pp. 282–288 (2019)
16. Schiller, J.H., Voisard, A. (eds.): Location-Based Services. Morgan Kaufmann
(2004)
 Enhancement of Medical Information
Security Through Lightweight
Cryptography and Steganography
in the Platform of Cloud-IoT Ecosystem

Humayra Binte Arfan, Zinia Anzum Tonni(B) , Saria Jahin Taluckder,

and Rashed Mazumder

Jahangirnagar University, Dhaka, Bangladesh

[email protected]

Abstract. In the 21st century, an innovative method is the online medi-

cal integrated system. It has significant applications, such as reducing the
expense of the healthcare system, enhancing the efficacy of the hospital
management system, and improving the effectiveness of human capital.
The online medical system can undoubtedly play a critical role in the
COVID-19 pandemic scenario, and it does as well. A road map of the
secured acquisition of medical data from COVID-19 patients, encoding,
and distribution to the cloud will be seen in this article. The major
aim of this paper is to design a framework for protecting cloud med-
ical data where resources are very limited. Blowfish and Elliptic-curve
hybrid cryptographic techniques have been used for encryption. Using
the 2D Discrete Wavelet Transform (DWT) method at the fog layer of
the cloud, the encrypted data is then embedded into a stego-image for-
mat. For a stego-image, the distortion of the cover image must be left
unnoticeable. The PSNR, MSE, and SSIM values are most important for
maintaining this hiding capacity. The PSNR value differed from 53.08 to
60.36, while the MSE value varied from 0.06 to 0.31 with different data
sizes with respect to different formats of the grayscale cover image. For
all experimental conditions, the Structural Similarity Index Test (SSIM)
values were 1. The proposed model with its stego image PSNR, MSE,
and SSIM values has demonstrated that this proposed model can be a
suitable substitute to secure medical data in cloud.

1 Introduction
Medical information security is an essential method for securing the sensitive
data but most often computerized image and relevant data of patient are trans-
mitted across public networks [1]. There is another tool of Internet of Things
(IoT) that has extensive usage in healthcare system, such as it can provide
benefits like simultaneous reporting, end-to-end connectivity, affordability, and
remote medical assistance services [2]. Cloud-based Medical Care has various

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 472–482, 2021.
https://doi.org/10.1007/978-3-030-75100-5_41
 Medical Information Security in Cloud-IoT 473

issues related to its security such as illegal access policy issues, and data protec-
tion. In addition, the exchange of medical information over public networks is
vulnerable as there are different types of security threats. When data or health
information is in public network (non-trusted communication channel), there is
a high chance to leak that information. The COVID-19 epidemic poses an enor-
mous challenges to the global health system where maintaining physical distance
is mandatory. If the system is based on offline, we do not need to worry about
the privacy issues. However, the offline health database management system is
not suitable in this 21st century. So, the virtual medication and privacy issues
both are important and mandatory nowadays. Therefore, a secure framework is
needed for transmitting, storing, and processing of medical information.

2 Background Studies

Hamid et al. proposed a security model for privacy preservation of healthcare

system using pair based cryptography for authenticating key exchange proce-
dure [3]. Though there is a double security level, fog nodes may be an entry
point of the attacker to steal the important data [4]. Mohammad Elhoseny et al.
proposed a model combination of hybrid encryption scheme and steganography
technique [5]. It gives the higher security in terms of PNSR, MSE and SSIM
value. It can be unsuitable for resource constraint areas as AES and RSA both
need large storage capacity [6]. Rajesh et al. proposed a novel tiny symmet-
ric encryption algorithm (NTSA) providing enhanced security for the transfer
of text files through the IoT network by introducing additional key confusions
dynamically for each round of encryption [7]. Tan et al. proposed a method to
secure the image using reversible watermarking technique having tamper detec-
tion ability [8]. It assures the integrity and authenticity of the image and offers
very less run time with good quality of image.
Usman et al. proposed a new image steganography approach for securing
medical data using Swapped Huffman tree coding to apply lossless compres-
sion and manifold encryption to the payload, before embedding into the cover
image [9]. Baby et al. proposed data securing technique using DWT that used
for hiding multiple color images into a single color image. Secret images are
embedded into R,G and B color planes [10]. For combined method of cryptog-
raphy and steganogarpy, some papers are remarkable. Sharma et al. proposed
a method using blowfish and LSB steganography was applied to secure images
from attackers [11]. Haddad et al. proposed a new joint watermarking JPEG-LS
scheme that extracts the message from the compressed image bit stream [12].
It is based on the full lossless JPEG-LS en-coding and does not decompress the
compressed image bit stream to access the message. Al Haj et al. proposed a
system that combines encryption standards with watermarking techniques [13].
Chinnasamy et al. proposed a design of secure cloud storage for health-
care data by using a Hybrid encryption of user information using Blowfish
and enhanced RSA algorithms [14]. To avoid false requests, the information is
474 H. B. Arfan et al.

encrypted through a symmetric algorithm and keys are encrypted using an asym-
metric algorithm. Ardiansyah et al. proposed a combination of two Steganog-
raphy domains LSB and DWT coupled with Cryptography 3-DES method [15].
This model provides higher PSNR and less MSE value. Tseng et al. proposed
a novel algorithm named Swapped Huffman code Table (SHT algorithm) join-
ing compression and encryption based on the Huffman Coding [16]. This model
shows that the security could be achieved without sacrificing the compression
ratio and performance.

3 System Design

The proposed model composes of three continuous processes:

1. The confidential patient’s data is compressed using Huffman encrypted

compression technique to pass through the transmission line. After being
encrypted from the IoT device, medical data will be transmitted to the fog
layer.
2. Then, the medical data will be encrypted using a hybrid encryption scheme
that is Blowfish and ECC encryption algorithms in fog layer.
3. Finally, the encrypted data is being concealed in a random medical cover
image using 2D-DWT steganography technique producing a stego-image and
uploaded to the cloud.

Figure 1 is describing our proposed model where we have used compression,

cryptography, and steganography method.

Fig. 1. System model

3.1 Terminal Devices

Data Compression Technique: Medical information contains very personal

and vital details of individuals. For this reason, losing very little data can lead
to many problems. Based on this, we have used Huffman coding for compressing
and securing the medical information through the transmission line as it is a
lossless compression technique.
 Medical Information Security in Cloud-IoT 475

3.2 Encryption in Fog Node

3.2.1 Incorporation Process into Fog Node

We have integrated the encryption technique into the fog node as IoT devices
have such limited capacity. It is assumed that this fog node is situated very close
to the cloud. So that with less time, output can be accomplished.

Fig. 2. Sequence diagram of data in fog node platform

3.2.2 Hybrid Cryptography and Steganography

Hybrid cryptography using Blowfish and ECC are considered to be used here
for encryption. To generate ciphertext, Blowfish is used to encrypt data. Using
ECC asymmetrical cryptography, Blowfish keys are distributed. Cipher text is
embedded in a cover image using 2D-DWT steganography. So that the undesired
individual may presume that the stego-image is the original data. If the intruder
can assume the existence of stego-image authentication without understanding
the key, he has to crack the blowfish encryption again, which has unpredictable
key length and is difficult to unlock.

Fig. 3. Visualization of encryption process using cryptography and steganography.

476 H. B. Arfan et al.

3.2.3 Cryptography Using Blowfish

There are several cipher modes that can be used in Blowfish. We selected the
Cipher Feedback (CFB) method for this project. Since it can be used for any size
of data. Inputs in the Blowfish algorithm are sender-defined data and key. The
key is encrypted by calling the cipher method, and the cipher text is eventually
accessed using the CFB mode.

3.2.4 Key Sharing Using Elliptic Curve Cryptography

Symmetrical cryptographic algorithms exchange the sender-recipient key. How-
ever, at the time of sharing, a third party can eavesdrop on the transmission
lines to access data. To prevent this, symmetric keys may be distributed using
asymmetric cryptography, i.e. public cryptography. The public and private keys
can be produced by using the Elliptic Curve Cryptography (ECC) algorithm.
Blowfish’s key can be transmitted from sender to recipient in cryptographic for-
mat with the assistance of the private and public recipient keys. As ECC is in
our consideration to distribute the key between sender and recipient, it will be
implemented further.

Fig. 4. Hybrid cryptography utilization

3.2.5 Steganography Scheme

We have used “Haar” filters in the 2-level 2D DWT algorithm in our project.
In 2-Level 2D DWT, the sub band LL1 is split into a further four sub bands is
seen in Fig. 5. Our encrypted text has been hidden in HH2 and HL2 sub bands.

Fig. 5. Original image to 2-level 2D DWT transformation [5]

 Medical Information Security in Cloud-IoT 477

In the DWT algorithm, the cover image and encrypted text have been
inserted in the input section. Finally we get the embedded image as output.
Besides using the HH2 of LL1 sub band, HL2 can also be used if the information
is too large to hide in single area.

4 Result Analysis

The implementation of proposed model is carried out using Python 3 in Jupyter

Notebook and Matlab R2017a running on a personal computer with a 2.27 GHz
Intel (R) Core (TM) I3 CPU, 16 GB RAM and Windows 10 as the operating
system.
As a cover image, grayscale images from various formats are utilized here to
legitimize our proposed technique. The text size in 15, 30, 45, 60, 100, 128, 256
bytes have been picked as plain content. The details of the experimented images
of DICOM dataset [17,18] that represents the Image(1), Image(2), Image(3) and
DME dataset [19] that represents the Image(4) are specified in Table 1.

Table 1. Details of experimented images

No Name Resolution Format

Image (1) MRIBrain 400 × 400 PNG
Image (2) Handskel 450 × 600 BMP
Image (3) Kidney 630 × 429 JPG
Image (4) Dieye3 768 × 576 Tiff

4.1 Visual Comparison

Figure 6 to Fig. 9 displays the embedded images where the encrypted text of 256
bytes has been used. If we analyze these images, it is almost impossible to make
a distinction between the original and the embedded image through naked eye.
At that point, we can say these embedded pictures are same as original one.

4.2 Compression Ratio

The data compression ratio is also expressed as the compression power, which
determines the reduction in the size of the data. The values are expressed in
Table 2 having the compressed and uncompressed size of different text size.
478 H. B. Arfan et al.

Fig. 6. Original and embedded image Fig. 7. Original and embedded image
of image (1) of image (2)

Fig. 8. Original and embedded image Fig. 9. Original and embedded image
of image (3) of image (4)

Table 2. Original size and compressed size after Huffman coding

Original size Compressed size Compression ratio

15 byte 8byte 1.875
30 byte 19 byte 1.579
45 byte 29 byte 1.552
60 byte 38 byte 1.579
100 byte 64 byte 1.563
128 byte 82 byte 1.561
256 byte 164 byte 1.561

4.3 Statistical Result

Table 3 shows the after effect of three measurable parameters PSNR, MSE,
SSIM. The outcome is acquired from embedded grayscale pictures with differ-
ent inserted text size in different image formats. From this table, it tends to
be assessed that PSNR esteem practically reachable to 60 to 53. The MSE var-
ied from .06 to .32. SSIM is consistently 1 which demonstrates no differences
between original and embedded picture.
 Medical Information Security in Cloud-IoT 479

Table 3. Result of the statistical parameters for gray scale image with different text
size.

Image (1) Text size (byte) PSNR MSE SSIM

15 58.49 0.09 1
30 58.42 0.09 1
45 58.38 0.09 1
60 58.27 0.10 1
128 58.04 0.10 1
256 57.57 0.1138 1
Image (2) 15 53.31 0.30 1
30 53.28 0.30 1
45 53.28 0.30 1
60 53.24 0.31 1
128 53.12 0.31 1
256 53.08 0.32 1
Image (3) 15 60.36 0.06 1
30 60.23 0.06 1
45 60.17 0.06 1
60 59.98 0.07 1
128 59.59 0.07 1
256 58.91 0.08 1
Image (4) 15 57.57 0.11 1
30 57.56 0.11 1
45 57.53 0.11 1
60 57.54 0.11 1
128 57.51 0.12 1
256 57.44 0.12 1

4.4 Histogram Analysis

Figure 10 illustrates the histogram of both original and embedded image.

Encrypted text of 256 byte data has been used in embedded image to have
the histogram.
480 H. B. Arfan et al.

Fig. 10. Original and embedded histogram of Image (1), Image (2), Image (3), Image
(4)

4.5 Lightweight Justification

Another thought is, how much size has been expanded of the embedded image
contrasted with the original image. As we have chipped away at the asset require-
ment zone, it is additionally a major worry to direct the consumed space of the
embedded picture. Table 4 shows that embedded image is not too large, so less
space from memory will be required in the cloud.

Table 4. Size details of embedded image

Image name Original image size (kb) Embedded image size (kb)
Image (1) 96.2 110
Image (2) 29.2 242
Image (3) 23.3 117
Image (4) 78.6 200

Table 5 shows the memory usage and run session of proposed technique which
have been performed creating a virtual environment for this project. This out-
come demonstrated that proposed model takes less assets to execute its activity.
 Medical Information Security in Cloud-IoT 481

Table 5. Time and memory utilization

Text size (byte) Encryption and Decryption time (ms) Memory (MB)
100 14 100.2
128 17 100.3
256 24 101.7

4.6 Summary
Our proposed framework performs very well for JPG image having 58.91 PSNR
worth and MSE esteem is .08 for embedding 256-byte text data . It likewise
gives a decent outcomes to other picture format like tiff, png and bmp. This
framework produces 56.75 PSNR esteem in average, SSIM is consistently 1 and
MSE is .16 considering the tested picture formats.

5 Conclusion
The proposed architecture has developed using 2D DWT steganography tech-
nique and hybrid cryptographic algorithm, where Blowfish and ECC are consid-
ered to use. This architecture is suitable to process and transfer the medical infor-
mation in a secure manner through the cloud storage using resource-constrained
devices. It performs very well for jpg image format. However, the result of our
proposed model is not satisfactory in bmp image format comparing to other
formats. This result is obtained from different gray scale image format those
having different text size. Though ECC is in our system design, we will try to
implement this key exchanging technique in future. In our research model, we
use only text as medical information. In near future, we will try to consider the
medical images as input for secured transmission.

References
1. Avudaiappan, T., Balasubramanian, R., Pandiyan, S.S., Saravanan, M., Laksh-
manaprabu, S.K., Shankar, K.: Medical image security using dual encryption with
oppositional based optimization algorithm. J. Med. Syst. 42(11), 1–11 (2018)
2. Patel, N.: Internet of things in healthcare: applications, benefits, and chal-
lenges. Internet: https://www.peerbits.com/blog/internet-of-things-healthcare-
applications-benefits-andchallenges. html. Accessed 21 March 2019
3. Al Hamid, H.A., Rahman, S.M.M., Hossain, M.S., Almogren, A., Alamri, A.: A
security model for preserving the privacy of medical big data in a healthcare cloud
using a fog computing facility with pairing-based cryptography. IEEE Access 5,
22313–22328 (2017)
4. Atlam, H.F., Walters, R.J., Wills, G.B.: Fog computing and the internet of things:
a review. Big Data Cogn. Comput. 2(2), 10 (2018)
5. Elhoseny, M., Ramı́rez-González, G., Abu-Elnasr, O.M., Shawkat, S.A., Arunku-
mar, N., Farouk, A.: Secure medical data transmission model for IoT-based health-
care systems. IEEE Access 6, 20596–20608 (2018)
482 H. B. Arfan et al.

6. Bhardwaj, I., Kumar, A., Bansal, M.: A review on lightweight cryptography algo-
rithms for data security and authentication in IoTs. In: 2017 4th International
Conference on Signal Processing, Computing and Control (ISPCC), pp. 504–509.
IEEE, September 2017
7. Rajesh, S., Paul, V., Menon, V.G., Khosravi, M.R.: A secure and efficient
lightweight symmetric encryption scheme for transfer of text files between embed-
ded IoT devices. Symmetry 11(2), 293 (2019)
8. Tan, C.K., Ng, J.C., Xu, X., Poh, C.L., Guan, Y.L., Sheah, K.: Security protection
of DICOM medical images using dual-layer reversible watermarking with tamper
detection capability. J. Digit Imaging 24(3), 528–540 (2011)
9. Usman, M.A., Usman, M.R.: Using image steganography for providing enhanced
medical data security. In: 2018 15th IEEE Annual Consumer Communications &
Networking Conference (CCNC), pp. 1–4. IEEE, January 2018
10. Baby, D., Thomas, J., Augustine, G., George, E., Michael, N.R.: A novel DWT
based image securing method using steganography. Procedia Comput. Sci. 46,
612–618 (2015)
11. Sharma, M.H., MithleshArya, M., Goyal, M.D.: Secure image hiding algorithm
using cryptography and steganography. IOSR J. Comput. Eng. (IOSR-JCE)
(2013). e-ISSN, 2278-0661
12. Haddad, S., Coatrieux, G., Cozic, M., Bouslimi, D.: Joint watermarking and lossless
JPEG-LS compression for medical image security. Irbm 38(4), 198–206 (2017)
13. Al-Haj, A., Abdel-Nabi, H.: Digital image security based on data hiding and
cryptography. In 2017 3rd International Conference on Information Management
(ICIM), pp. 437-440. IEEE, April 2017
14. Chinnasamy, P., Deepalakshmi, P.: Design of secure storage for health-care cloud
using hybrid cryptography. In: 2018 Second International Conference on Inventive
Communication and Computational Technologies (ICICCT), pp. 1717–1720. IEEE,
April 2018
15. Ardiansyah, G., Sari, C.A., Rachmawanto, E.H.: Hybrid method using 3-DES,
DWT and LSB for secure image steganography algorithm. In: 2017 2nd Interna-
tional conferences on Information Technology, Information Systems and Electrical
Engineering (ICITISEE), pp. 249-254. IEEE, November 2017
16. Tseng, K.K., Jiang, J.M., Pan, J.S., Tang, L.L., Hsu, C.Y., Chen, C.C.: Enhanced
Huffman coding with encryption for wireless data broadcasting system. In: 2012
International Symposium on Computer, Consumer and Control, pp. 622–625.
IEEE, June 2012
17. McEvoy, F.J., Svalastoga, E.: Security of patient and study data associated with
DICOM images when transferred using compact disc media. J. Digit Imaging
22(1), 65–70 (2009)
18. Hacklander, T., Martin, J., Kleber, K.: An open source framework for modification
and communication of DICOM objects. Radiographics 25(6), 1709–1721 (2005)
19. Rabbani, H., Allingham, M.J., Mettu, P.S., Cousins, S.W., Farsiu, S.: Fully auto-
matic segmentation of fluorescein leakage in subjects with diabetic macular edema.
Invest. Ophthalmol. Vis. Sci. 56(3), 1482–1492 (2015)
 Blockchain-Based Trust Management Approach
for IoV

Amal Hbaieb1,2,3(B) , Samiha Ayed2 , and Lamia Chaari3

1 National School of Electronics and Telecommunications of Sfax, Sfax, Tunisia
2 LIST3N Laboratory, University of Technology of Troyes, Troyes, France
[email protected]
3 CRNS-SM@RTS (Laboratory of Signals, systeMs, aRtificial Intelligence and neTworkS),

Sfax, Tunisia

Abstract. Internet of Vehicles (IoV) brings new specific vehicular use cases fur-
ther from Vehicular Ad-hoc Networks (VANETs) and offers sophisticated applica-
tions for more efficient Intelligent Transport Systems (ITS) in which autonomous
driving technology is the quantum leap in smart connected vehicles. However,
this innovation brings a security concern and remains cautious and suspicious as
vehicular networks become more vulnerable. Traditional security techniques may
not tackle all the security issues and provide the comprehensive protection in the
current IoV system scenarios. It is therefore of paramount importance to orga-
nize trust in such environment and distinguish dishonest entity from trustworthy.
Accordingly, trust management-based approaches are further needed to assist for
effective vehicular networks deployment. In this paper, a trust management-based
model is proposed to secure the vehicular communication system. The introduced
trust approach executes a serie of security activities to ensure the trustworthiness
of the exchanged data. In our scheme, vehicles also validate the trustworthiness
of received data relying on location-trust metric. In this way, our approach will be
efficient in calculating and maintaining trust scores in vehicular networks while
accommodating related imperatives as Quality of Service (QoS) satisfaction.

1 Introduction

Internet of Vehicles (IoV) networks implement several autonomous vehicles capable of

communicating with each other. This type of network is essentially characterized by the
high mobility and the sensitive nature of transmitted data. Therefore, the field of use
of IoV is confronted with very restrictive requirements in terms of Quality of Service
(QoS) and security. Recently, there are several related works that have been interested
in the study of IoV by looking at the network technical performance aspects. However,
these works have not take into consideration the security aspect. From a security point
of view, several works have focused on the study of security mechanisms in IoV without
considering the impact of the proposed solutions on the performance of such network.
Some of these solutions are based on cryptographic techniques. These techniques need
to meet more stringent QoS requirements (e.g., low time complexity). Besides, their use
requires and consumes a lot of memory and energy, which makes their consideration in

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 483–493, 2021.
https://doi.org/10.1007/978-3-030-75100-5_42
484 A. Hbaieb et al.

the IoV unprofitable regarding network performances. In this paper, we aim to propose
a secure management of IoV that allows later to maintain its performance. Our solution
is based on the trust management concept. Trust management consists of a set of steps
where a node assigns a trust value to another node during their interaction, in order to
mitigate the maleffects of malicious and selfish nodes actions. The major of existing
approaches for the trust management in IoV do not take into account the impact of the
use of trust management techniques on the main QoS parameters. Thus, we propose
to design a trust management scheme based on the Blockchain technology that will
enable to deal with security and essential QoS requirements regarding its features. Our
introduced trust approach executes a serie of security activities to ensure the trustwor-
thiness of the exchanged data. In our scheme, vehicles validate the trustworthiness of
received data relying on the reputation and the location metrics. In this way, our approach
will be efficient in calculating and maintaining trust scores in IoV while accommodat-
ing related imperatives as QoS satisfaction. Our paper is organized into five sections
as follows: Sect. 2 is dedicated to the related works and recent researches focused on
Blockchain-based trust management in IoV. Section 3 provides our proposed system
communication model where we identify the components of the envisaged communica-
tion system architecture, as well as our network model. Section 4 validates our proposal
through a use case simulation. Finally, Sect. 5 concludes the paper.

2 Related Works
Blockchain technology [1] is having broad potential for the establishment of trust man-
agement in vehicular networks context. Indeed, the merits of Blockchain can be exploited
to deal with the faced key issues related to centralization, security, and privacy, when
maintaining, managing, tracking, and sharing data related to vehicles, traffic condi-
tions, traffic violations, surveillance, and weather information. We refer here some ideas
that were elaborated based on Blockchain to mainly address security [2–6], and pri-
vacy preservation aspect [7, 8] for vehicular networks. In [2], the authors proposed a
Blockchain-based trust management scheme where the main process consists of trust
rates generation and uploading step, trust value offsets calculation step (based on rep-
utation and knowledge metrics), and miner election as well as consensus mechanism
application for new blocks generation. The receiver entity relies on Bayesian inference
rule to assess message-credibility and later generate trust scores for these received con-
tents. Next, the computed trust values are uploaded to the RSUs nodes. Trust values
offsets of vehicles nodes are obtained using weighted aggregation. The scheme applies
Proof-of-work and proof-of-stake consensus. The Blockchain-based trust scheme in [7]
is combined with a privacy preservation approach that is managed by identity-based
group signatures for an efficient conditional privacy. The work uses Proof-of-work and
Practical Byzantine Fault Tolerates for security reinforcement. Similarly, the work [8]
exploited Blockchain features to build an anonymous reputation management scheme
that addresses trust and privacy issues in VANET. The reputation management is based
on experience and recommendations metrics. The scheme uses further reward and pun-
ishment models to improve the trust establishment. We refer also that the Named Data
Networking (NDN) technology was proposed to assist in providing trustworthy commu-
nications within Blockchain-based networks models, such in [9] where authors explored
 Blockchain-Based Trust Management Approach for IoV 485

this combination to enhance the trust in VANET. We note that the major of these works
do not consider effectively the impact of their used schemes on the QoS parameters (e.g.,
by using less intensive computationally Blockchain). Therefore, we aim to tackle this
limitation in our proposal.

3 Proposed System Communication Model

3.1 Architecture Overview
In this subsection, we introduce our proposed Blockchain-based trust management app-
roach for IoV network. Figure 1 illustrates the overall architecture of the envisioned
communication system. By keeping in mind the major QoS issues when deploying
trust management solution in vehicular networks, we propose the elaboration of two-
layered trust database by means of Blockchain: (1) a general Blockchain which will
serve to exhibit the global trustworthiness of the communication system, and (2) a local
Blockchain which will be assigned for the different zone areas of the network to pro-
ceed a local trust management. The main use of local Blockchains is to meet lower
delays while processing time-sensitive data and handling their security. The core of our
proposed trust management model consists of ordinary,lightweight nodes, edge nodes,
RSUs nodes, and trusted authority nodes. The ordinary nodes represent the main users
of the communication system, which refer generally to the smart vehicles (as we will
settle on V2V and V2I communications for our later simulation). The ordinary nodes act
as smart vehicles and they are resource constraint nodes (e.g., storage, computation, and
battery life). They consist of sensing, communication, and basic local processing layer.
Each vehicle is installed with sensors, actuators, embedded gateway [10], and on-board
units which has wireless communication capability to permit the vehicle to communi-
cate with other nodes to share messages. The trusted authority nodes are assumed to be
trusted (i.e., deployed for privacy issue dealing). For example, trusted authority node
undertakes the work to register (further revoke) the new nodes that want to participate
in the communication system. Trusted authority nodes also maintain the database that
contains correlation between nodes’ public address and real identities. The edge nodes
handle trust-service requests/responses (e.g., trust values computation and trust values
exhibition). We consider that the edge nodes are semi-trusted nodes. The edge nodes
have more computational power than ordinary nodes, and have further cache storage.
They host the local Blockchains which will be transferred later to RSUs; along with
smart contracts execution and mining consensus performing. As usual, the RSUs nodes
act as static nodes to provide required services (e.g., traffic information, road event, etc.)
and undertake the computational work to add the local Blockchains into the global chain
as we aforementioned.
Hence, the scalability and the availability of the network can be achieved as edge
and RSUs nodes are placed in a distributed way. Regarding privacy, a cryptography-
based mechanism can be integrated to ensure more reliable communication. The trusted
authority nodes are deployed to help in privacy-preservation (e.g., certificates issuance
for participating nodes). The proposal use Interplanetary File System (IPFS) along with
the Blockchain technology to allow the storage of the large amount of generated data
(particularly within an autonomous driving context) in a decentralized and distributed
486 A. Hbaieb et al.

Fig. 1. Proposed network model: overall architecture.

way. The smart contracts can also be implemented to allow the trustworthy transaction
and automate the network processes. IPFS hashes will be used in order to link and track
the data in the Blockchain smart contracts. Moreover, to optimize the execution time of
nodes queries, cache memory is introduced (e.g., to store frequently asked data). The
procedure of nodes requests requires a large execution time that involves propagation
delay and transmission delay. As IoV is a time-sensitive system, cache memory is applied
at the edge nodes to meet this time complexity enhancement.

3.2 Main Procedures

The main phases of our Blockchain-based approach consist of (1) trust values computa-
tion by edge nodes, (2) local mining mechanism and blocks generation, (3) trust values
uploading to the local Blockchains, (4) global mining mechanism and blocks generation,
and (5) trust values uploading to the global Blockchain. The edge vehicles are respon-
sible, as aforementioned, for determining the local trustworthiness of the IoV network
and for hosting the internal Blockchains. In this subsection, we highlight the procedure
of local trust management. The process of local trustworthiness assessment is preceded
by an authentication procedure, and it stands also to apply an incentive algorithm to give
reward or penalty to engaged nodes (i.e., messages senders, queried nodes). We term V
= [V 1 , V 2 , …, V i ] the set of i lightweight vehicles in the IoV, E = [E 1 , E 2 ,…, E k ] the
representation of k edge vehicles; 1 ≤ k and TA = [TA1 , TA2 …, TAl ] the set of l trusted
authority nodes; 1 ≤ l which are responsible to validate the identities of the nodes in
the IoV communication system. The basic idea to gather the local trust information is to
rely on the reputation of the sender node and to suggest to the recipient to evaluate the
credibility of the received message. That is to say, the trust score refers to the combi-
nation of defined reputation value and message-credibility value. Basically, each sender
will have an initialized reputation value. Figure 2 depicts the process applied in this local
trust management phase. We assume that every lightweight vehicle can quickly query
the trust score of the sender node (i.e., inquires about the trustworthiness of a node from
 Blockchain-Based Trust Management Approach for IoV 487

which it received messages) from the nearest edge vehicle E k . The trust values that k
edge vehicles keep for j sender nodes are denoted as R = [R1 , R2 , …, Rj ]. In fact, each
queried E k calculates the sender trust value based on reputation and location metrics. We
refer that the trust values are maintained in the E k corresponding IPFS, while the most
frequent ones are stored in the E k respective cache. For example, once a lightweight
vehicle obtains a message from another vehicle, it identifies the later’ trust, describing
how to react on the received content. Regarding authentication, a vehicle node V i (V i ∈
V) that wants to join the IoV network uses its assigned pair of public-private keys (termed
as (Pum i , Pvm i ); 1 ≤ m, where m is the cardinality of generated keys) and receives a
digital certificate from a trusted authority node TAu .

Fig. 2. Process of local trust management.

Obviously, the vehicle V i generates the pair of keys (Pum i , Pvm i ) to reach a classical
encrypted communications, and the V i associated certificate issued by the trusted author-
ity node TAu is used to correlate V i public key (Pum i ) and real identity. The delivered
certificate by TAu to V i contains the V i public key, the generated signature of the entire
certificate (Sigu ), the information about the certificate issuer TAu (IF u ), the expiration
date of the given certificate (Expi ), as well as the V i trust value (Ri ) (i.e., V i initial
reputation value, or V i trust value given by an edge vehicle E k ):
m
Certi,u = Puim , Sisgu , IFu , Expi , Ri (1)

Step 1: Calculation of Reputation Value

The reputation value of a sender node will be calculated by an edge vehicle based on
the knowledge metric and available recommendations provided from other surrounding
edge vehicles. The knowledge metric refers either to the sender’ initialized reputation,
or its trust established by the previous queried edge vehicle and retrieved from the IPFS
(from the cache memory of the edge vehicle; if available). The number of queried recom-
menders as the related number of hops will be defined taken into account their impact on
the calculation accuracy. The edge vehicle E k can consider further the recommendation
available from the nearest RSU to establish the reputation value. Moreover, we define
three weightiness levels for exchanged events: (1) high level (concerning crucial and
severe events, e.g., accident or information about road work), (2) ordinary level (e.g.
beacon messages comprising driving status), and (3) soft level (concerning less impor-
tant events, e.g., normal traffic information). For instance, let the vehicle V i requests the
488 A. Hbaieb et al.

nearest edge vehicle E k to exhibit the trust value of another vehicle V j from which it
receives a message about an event Evh (Evh ∈ [Ev1 , Ev2 , …, Evh ]; 1 ≤ h), and let the
event Evh an accident event. Suppose that the edge vehicle E k communicates with its
direct neighbor E q in order to proceed the required value calculation. The queried edge
vehicle E k computes the V j reputation value as follows:
t

n
Rtj = αRt−1
j +β γq Recq,j + δ RSUa,j (2)
q=1

where Rj t denotes the reputation value of the sender V j given by the edge vehicle E k
at the current time t, nt is the recommenders number (in the current scenario, we have
only one recommender; E q ), Recq,j refers to the recommendation of the edge vehicle E q
about V j , RSU a,j represents the recommendation about V j by the nearest placed RSU a
(RSU a ∈ [RSU 1 , RSU 2 , …, RSU b ]), γ q stands for E q recommendation credibility, and
the variables α, β, and δ are the fixed weights of knowledge factor (Rj t−1 ), edge nodes
recommendations, and RSU a recommendation, respectively. We point that α, β, δ ∈ [0,
n
1], α + β + δ = 1, and γ q = 1. Next, E k will pass to judge V j message-credibility.
q=1
The edge nodes are considered as semi-trusted nodes. Therefore, after making the V j
message acceptance decision, a reward or a punishment for involved recommenders will
be carried out with respect to their associated weights (regardless Evh weightiness level):
e.g., there are two kinds of actions that will make recommender weight (γ q ) increasing.
First, recommender gave a good recommendation and V j message is admitted. Second,
recommender did not support V j when V j message is dropped. Also, recommender
weight will be decreased if it supported V j whereas V j message is denied, or if it gave
bad recommendation whereas V j message is accepted.
Step 2: Calculation of Message-Credibility Value
As indicated above, the second phase in the process of local trust establishment consists
of evaluating the E k degree of belief on the V j message-credibility about the event Evh .
This task is relied on the location metric, as the location proximity is an important
attribute that can assess the data credibility (i.e., data are likely more accurate as the
sender node V j is close to the location of the event Evh ). Let C m = [C m1 t , C m2 t , …,
C mj t ] the set of messages-credibility values estimated by k edge vehicles. A C mj t value
corresponds to the credibility of Evh sent from V j to V i and defined by E k . The calculation
of the location proximity of V j to Evh allows E k to determine the C mj t value.
 
minimum locVj , locEvh
Cmj =
t   (3)
maximum locVj , locEvh

Where |locVj – locEvh | < rholoc , locVj and locEvh are the locations of the sender node V j
and the event Evh , respectively. The rholoc indicates the admissible threshold difference.
If the marked condition is not satisfied, we assign value 0.1 to the credibility value C mj t .
Now the reputation score and the message-credibility value of the sender node V j are
available, we can get the related final trust value using the Bayes rule (as we aim to predict
 Blockchain-Based Trust Management Approach for IoV 489

the final trust value of the sender node by taking into account its message-credibility
value that is derived after its reputation score establishment):
   
  P Cmj t /Rt × P Rt
j j
P Rtj /Cmj
t
=   (4)
P Cmj t

where P(Rj t ) is the reputation of the sender node V j , and P(C mj t ) is the V j associated
message-credibility. When P(Rj t /C mj t ) exceeds the preset threshold for the trustworthi-
ness value (Thrstr ), the edge node E k regards V j message as true. The edge node E k
reports the ultimate trust score of V j to the vehicle V i which will execute the acceptance
task of Evh . Once the E k decision has been made on the validity of the event Evh , it
performs a payoff-based algorithm for the sender node V j in order to guide the update of
its trustworthiness. This incentive algorithm counts on the weightiness level of the event
E k to adjust the old trust of the sender (see the following algorithm). The sender V j is
rewarded with a positive reinforcement of its old trust value if the event Evh is confirmed,
otherwise, the V j is penalized by diminishing its trustworthiness. For example, in the
above scenario the event Evh corresponds to an accident, hence the old trust of V j will
be decreased in half once the accident is denied.

Furthermore, we can suppose that many sender nodes have send, quasi simultane-
ously, the same message about a particular event Evh to vehicle V i . The later node will
query Evh validity as usual from the nearest edge vehicle E k . E k will execute the calcu-
lation steps of senders’ reputation values and message-credibility values. The vehicle V i
accepts the event Evh from the sender V j of which E k has confirmed its reliability first.
Finally, all senders will be rewarded in case of Evh confirmation. We can also extend
to the scenario where the node V j sends to V i , quasi-simultaneously, many messages
related to different events. In this case, E k has to determine the coherent trust value
of the sender V j . Assume that V i receives two different messages (Evh and Evg ) from
V j . The queried edge vehicle E k starts to estimate its belief on them (using the above
corresponding equations). Thereafter, E k compares the generated results (P(Rj t /C mj t )
according to Evh and P(Rj t /C mj t ) according to Evg ) with the preset threshold Thrstr ,
to verify the validity of both Evh and Evg . E k will get conflicting about V j trust (e.g.,
Evh is validated and Evg is decayed, and note that V j has now two associated trust val-
ues). At this end, E k must estimate a new proper trust value for V j to preserve the trust
dataset consistency. This value can be deduced by calculating the geometric average of
490 A. Hbaieb et al.

the obtained trust results (e.g., P(Rj t /C mj t )-Evh and P(Rj t /C mj t )-Evg ), with the following
manner:

  nbj  
P Rtj /Cmjt
= nbj P Rtj /Cmjt (5)
j=1

nbj
where nbj is the number of sent messages from V j , and P(Rj t /C mj t ) represents the
j=1
product over the generated trust values of V j . Then, E k compares one more time this
ultimate value with the trust threshold Thrstr to run the payoff algorithm for V j . So, to
recap the local trust management, each new node that want to participate the network
has to register within the trusted authority node. Next, this participating entity generates
private/public keys, and submits to trusted authority node its initial public key to prove its
legal identity. The trusted authority node will issue an initial certificate to this node (will
be used for tracking the real identity of node in case of disputes). Whenever receiving a
message, the node can request the trust value of the sender from the nearest edge node.
The edge node checks first requester identity. Then, it undertakes the work of required
trust value computation. This procedure includes (1) reputation value calculation step
and (2) message-credibility value calculation step. So, the edge node verifies its cache
memory. If the old trust value of the sender is present at the cache, it will be used directly
for the calculation, otherwise, the edge node gets access to the IPFS to retrieve this trust
data; if available. Next, the edge node exhibits the new determined trust value to the
requester node which will make the decision of sender message acceptance. Edge nodes
begin the mining for adding the new calculated trust value into the local Blockchain.
Also, the miner can be rewarded (e.g., valid transaction number-based). Edge nodes
store the frequent required trust values in their caches. Upon being created, edge nodes
upload the local Blockchain into the RSUs, which will check in turn the validity of this
Blockchain and add it into the global chain. Likewise, the miner can be rewarded after
a valid transaction. Similarly, the following steps can be executed, for example, while
an event witness overall scenario. Whenever an event is occurred and ordinary node is
a witness, it sends a report to the nearest edge node. The later will check the sender
identity. Then, the edge node verifies the report validity, and accordingly it can update
the trust value of the sender. Mining consensus process is triggered once the sender trust
data is not on the local Blockchain. As aforementioned, the metrics used for trust value
computation are reputation-based (for sender reliability assessment), and location-based
(for message-credibility assessment). However, we can also take into account the time of
the received event when estimating message-credibility as this factor helps in verifying
message freshness. The problem of selfishness or low node participation can be tackled
by the involvement of an incentive mechanism. For example, whenever an ordinary
node produces a valid event report, it gets incentive (in term of trust adjusting, or may
certificate revocation). Besides, an announcement protocol, such in work [7], may be
involved to assist in nodes authentication (within an event witness context).
 Blockchain-Based Trust Management Approach for IoV 491

4 Simulation Results
In this section, we conduct the performance evaluation of our trust management scheme
to validate its effectiveness. We implement the approach in the Python environment,
where we apply the Ethereum Blockchain. To this end, we define the parameters’ values
used in our evaluation. We set the type of vehicles nodes before the beginning of the
evaluation. We suppose that these nodes are authenticated (i.e., each node has already
generated its pair of public-private keys, and has received a certificate from a trusted
authority node). As aforementioned, the vehicles nodes in our model are grouped into
areas, based on the location factor. Hence, in each area, we select 2%, 3%, 10% and
85% as maximum percentages of the members of RSUs, trusted authority, edge, and
lightweight nodes, respectively. Also, we consider 40% as the maximum percentage of
malicious lightweight nodes that provide false messages. The major nodes of IoV are
assumed to be likely honest. Therefore, the selected proportions seem to be realistic. Our
scenario is considered for a highway road, where the speed of vehicles varies between
90 km/h and 150 km/h, and the maximum nodes density is 150 nodes. In order to begin
the calculation of the trust value, we initialize the reputation value for each sender at
0.4, also, we define the threshold of the trust at 0.7. Regarding the values of associated
weights, we set 0.5, 0.3, 0.2 to α, β, and δ, respectively. We consider that the knowledge
factor (where α is assigned to), is the most important metric that will be taken into
account in the calculation of the trust value. Finally, the maximum number of hops
between queried recommenders is defined by 2 hops. We evaluate our proposal in V2V
and V2I environment with up to three different traffic density scenarios: 10, 50, and 150
vehicle nodes, where 2%, 5%, 10%, 20%, 25%, 30%, and 40% of lightweight vehicles
are malicious, and where we suppose that the received events are critical (accident event).
We assume that the recommenders are trusted. Under these parameters, we estimate our
communication system trustworthiness. The proposal proves its effectiveness in term
of detection rate of malicious nodes. The sender vehicles that reported invalid events
have yielded a weak value of trust. The scheme proves also that it supports the scalability
requirement. Figure 3 illustrates the outcomes of the detection rate of malicious vehicles.
For instance compared to the approach proposed in [7], our scheme presents better
performance concerning detection rate. Nevertheless, it is notable that the higher the
percentage of dishonest recommenders in the system is, the easier for them to damage
the communication system. We assess also the time complexity factor. Figure 4 depicts
that the proposal reaches a good tradeoff between latency and trust management process
accomplishment. The latency corresponds to the total spending time for establishing the
trust in the communication system (i.e., trust values calculation, and scores upload and
update in local and global Blockchains). Finally, Fig. 5 illustrates the communication
overhead ratio obtained from carrying out our test. In this paper, the communication
overhead is stated as the number of valid packets that are transferred between nodes. We
evaluate this parameter over the percentage of malicious vehicles. As shown in Fig. 5,
the ascending percentage of malicious vehicles in the communication system reduces the
communication overhead. Thence, the communication overhead requirement is satisfied
as the number of detected malicious nodes has a significant impact on such criteria.
492 A. Hbaieb et al.

Fig. 3. Detection rate of Fig. 4. Total spending time for Fig. 5. Communication
malicious vehicles in the establishing trust in the overhead over malicious
communication system. communication system vehicles in the
communication system

5 Conclusion
In this paper we present a Blockchain-based trust management approach for IoV com-
munication system. By using two layers of Blockchain, each node is able to asses the
trustworthiness of the entities that communicate with. The trust estimation is based on
reputation and location metrics. By determining a few parameters, each participating
node can perceive a reasonably accurate view of the IoV network. The implementation
results demonstrate the efficiency of our work. The use of Blockchain solution allowed
us to find the good trade off between security and QoS in IoV environment. The proposed
network model shows that is can be dynamic, scalable, light and reliable. We plan to
further take into account the privacy preserving and define a detailed adversary model in
the future. Besides, the unavailable discussion regarding the implementation with real
simulation tools in this paper will be the focus of our future work.

Acknowledgments. This work was supported by the Ministère de L’Enseignement supérieur, de

la Recherche et de l’Innovation–France, and the FEDER; TBAN project.

References
1. Puri, V., Kumar, R., Van Le, C., Sharma, R., Priyadarshini, I.: A vital role of blockchain
technology toward internet of vehicles. In: Handbook of Research on Blockchain Technology,
pp. 407–416. Elsevier (2020)
2. Yang, Z., Yang, K., Lei, L., Zheng, K., Leung, V.: Blockchain-based decentralized trust
management in vehicular networks. IEEE IoT J. 6(2), 1495–1505 (2019)
3. Yang, Y.T., Chou, L.D., Tseng, C.W., Tseng, F.H., Liu, C.C.: Blockchain-based traffic event
validation and trust verification for VANETs. IEEE Access 7, 30868–30877 (2019)
4. Javaid, U., Aman, M.N., Sikdar, B.: DrivMan: driving trust management and data sharing
in VANETS with blockchain and smart contracts. In: 2019 IEEE 89th Vehicular Technology
Conference (VTC2019-Spring), pp. 1–5. IEEE (April 2019)
5. Dong, W., Li, Y., Hou, R., Lv, X., Li, H., Sun, B.: A blockchain-based hierarchical repu-
tation management scheme in vehicular network. In: 2019 IEEE Global Communications
Conference (GLOBECOM), pp. 1–6. IEEE (December 2019)
6. Chai, H., Leng, S., Zhang, K., Mao, S.: Proof-of-reputation based-consortium blockchain for
trust resource sharing in internet of vehicles. IEEE Access 7, 175744–175757 (2019)
 Blockchain-Based Trust Management Approach for IoV 493

7. Liu, X., Huang, H., Xiao, F., Ma, Z.: A blockchain-based trust management with conditional
privacy-preserving announcement scheme for VANETs. IEEE IoT J. 7(5), 4101–4112 (2019)
8. Lu, Z., Wang, Q., Qu, G., Liu, Z.: Bars: a blockchain-based anonymous reputation system for
trust management in vanets. In 2018 17th IEEE International Conference on Trust, Security
and Privacy in Computing and Communications/12th IEEE International Conference on Big
Data Science and Engineering (TrustCom/BigDataSE), pp. 98–103. IEEE (August 2018)
9. Khelifi, H., Luo, S., Nour, B., Moungla, H., Ahmed, S.H.: Reputation-based blockchain for
secure NDN caching in vehicular networks. In: 2018 IEEE Conference on Standards for
Communications and Networking (CSCN), pp. 1–6. IEEE (October 2018)
10. Hbaieb, A., Rhaiem, O.B., Chaari, L.: In-car gateway architecture for intra and inter-
vehicular networks. In 2018 14th International Wireless Communications Mobile Computing
Conference (IWCMC), pp. 1489–1494. IEEE (June 2018)
 Improved Visual Analytic Process Under
Cognitive Aspects

Samar Bouazizi1(B) and Hela Ltifi2

1 Faculty of Sciences and Techniques of Sidi Bouzid, Computer Science and Mathematics
Department, University of Kairouan, Kairouan, Tunisia
2 REGIM-Lab.: Research Groups in Intelligent Machines, LR11ES48, National Engineering
School of Sfax, University of Sfax, 3038 Sfax, Tunisia
[email protected]

Abstract. The Visual Analytics (VA) systems aids to ameliorate the decision-
making activity, since these systems relates visual to automatic analysis to extract
knowledge and explore data. In this paper, we present an adaptation of a cognitive
Visual Analytics with a well-known Visual Analytics process under cognitive
aspects: intelligent data collection, fusion and synchronization and their analysis
using the adapted cognitive data mining technique. To prove the effectiveness of
our approach, we have applied it during a formative assessment. We have used the
Eye Tracking as well as the Electroencephalogram (EEG) as an intelligent data
collection source and the Fuzzy Logic as data mining technique. We performed
an evaluation for proving the utility and usability of this developed system.

Keywords: Cognitive Visual Analytics · Eye-tracking · EEG · Fuzzy logic ·

Formative assessment

1 Introduction
The importance of user interfaces in the decision support systems research field has
been emphasized [1] as the main purpose is to more integrate the decision-maker into
the decision-making process. Consequently, a new interactive system combining the
capacity of automatic processing and the capacity of human reasoning has taken place.
This system is called Visual Analytics (VA) systems [2].
Actually, visual analysis combines interactive visualization with automated methods
of data analysis to provide interactive and scalable decision support. Given the evolution
of the field of VA, researchers introduced several interactive and automatic methods and
measures for sense-making and insight generation [3]. However, the data collection pro-
cess is generally done using traditional tools and methods such as database management
systems. In this context, we propose to integrate automatic cognitive data collection in
the VA process applied in education field, specifically in a formative assessment. The
collected data must be computationally analyzed. The uncertain nature of the collected
data requires the use of specific technique: The Fuzzy Logic that is based on “degrees
of truth” allowing representing the human behavior with the system [4].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 494–506, 2021.
https://doi.org/10.1007/978-3-030-75100-5_43
 Improved Visual Analytic Process Under Cognitive Aspects 495

The rest of this paper is presented as follows: in Sect. 2, we provide a short study
background concerning decision support systems and VA. Subsequently, we formally
define our cognitive VA framework in Sect. 3. Within Sect. 4, we present our proposal
application. Finally, Sect. 5 summarizes our work and paves the way for future research.

2 Background of the Study

To manage large quantities of complex data, to generate knowledge and to make appro-
priate decisions, Decision Support Systems (DSS) are intelligent and visual [1, 19].
Such DSS are named VA systems. The term “VA” related Visual Analysis to Automatic
Analysis [2]. It is a multidisciplinary field which mainly integrates: (1) data mining, (2)
information visualization and (3) perception and cognition. The VA couples the auto-
mated analysis techniques with interactive visualizations to understand, interpret and
effectively decide from massive and complex data sets [2]. In the literature, several VA
researches works proposed pipelines for designing VA systems. We begin with Wjik
process [9].

Fig. 1. The Wijk Visual Analytics process

As shown in Fig. 1, the analyst gains new ideas by perceiving the image. These ideas
will be translated into knowledge, hence the emergence of new hypotheses. Finally,
after exploration and analysis, the analyst provides suitable specifications which will
be considered in the visualization to improve his/ her knowledge. Later, Keim et al. [2]
proposed VA process that is based on the combination of the human and the automatic
data processing (cf. Fig. 2).
In the sub-process of automatic data analysis based on the data mining methods (cf.
Fig. 2), the central phase makes it possible to automatically and intelligently extract
the precious information called models or “patterns” from data by means of automatic
analysis algorithms. Visualization is important in the sub-process of visual analysis of
data and patterns. On the one hand, it provides interactive visual representations of the
data and patterns generated by the mining method. On the other hand, it ensures efficient
Analyst-Image interaction. Hence the improvement in the quality of the knowledge
extracted by perceiving the visualization adequate to the decision-making situation.
Sacha et al. [10], basing on the Keim et al. VA process [2], proposed an adaptation of
the Keim et al.’s process.
496 S. Bouazizi and H. Ltifi

Fig. 2. The Keim et al. Visual Analytics process

Fig. 3. The Sacha et al. Visual Analytics process

As presented in Fig. 3, the first part (computer part) is the same as proposed by Keim
et al. [2]. However, they detailed the activities to be carried out to obtain knowledge
in the second part. So, they specified the phases necessary to reach knowledge which
are grouped into three loops of exploration, verification and knowledge generation. The
first is used to explore, allowing to choose the appropriate actions on the visualization
according to the results. The second loop consists of verifying the information making
it possible to automatically formulate hypotheses on the data analysis from the results.
The last loop aims to generate knowledge.
The review of these processes shows that the VA is still an area of research that needs
further improvements. Our work takes place in this context.

3 Proposed Cognitive Visual Process

In recent years, a remarkable research development in the field of VA, the main purpose
of which is to help users easily explore, analyze and interpret data. This visual analysis
approach transforms information into knowledge that represents the science of analytical
reasoning facilitated by interactive visual interfaces. Our contribution concerns the two
parts of the visual analysis process of Keim et al. [2], which is illustrated in Fig. 4.
As presented by Fig. 4, we have adapted the process of Keim et al. [2] (visible in
Fig. 4A) by integrating cognitive aspects (visible in Fig. 4B). Our adaptation is marked
in green in Fig. 4B.
 Improved Visual Analytic Process Under Cognitive Aspects 497

Fig. 4. The proposed cognitive Visual Analytics process

The process begins by a cognitive data collection. Contrary to traditional data col-
lection, we propose to use artificial intelligence to mimic the way the human mind reads
data. This method learns to automatically recognize a lot of data rather than traditional
configuration by experts. It is more adapted to collect data from visual interfaces. A set
of artificial intelligence tools can be applied such as the eye tracking [11], etc. The need
for this method comes from the interest presented by different fields of application such
as those of monitoring and assessments. In the case where different data collection tools
have been used, it is a question of merging and synchronizing them. Otherwise, an only
transformation is made on the collected data to prepare them for data mining. The data
fusion and synchronization consist of searching connection between the collected data
from the used cognitive tools for finally creating a single dataset before automatic and
visual analysis.
The treated data will be used in the data mining algorithm and mapped for creating
visualization at the same time. Concerning the automatic analysis, the applied data
mining algorithm must be adapted to the cognitive collected and synchronized data. For
498 S. Bouazizi and H. Ltifi

this reason, we named it cognitive data mining in the Fig. 4B. Concerning the visual
analysis, we have specified the cognitive visualization process of the user/decision-
maker/analyst. It consists of three main steps: perception, recognition and reasoning.

a) Perception. Visual perception is the means by which users interpret visual interfaces
on a computer screen. it is an activity that usually relies on information delivered by
one’s senses. It is the process of collecting and processing sensory information and the
resulting awareness.
b) Recognition. It is the ability of the human brain to identify what has been previously
perceived. Recognition is a cognitive skill that allows users to retrieve information stored
in memory and compare it to the represented information. Recognition is, in fact, a type
of recovery. It consists of having access to information from the past stored in the human
memory.
c) Reasoning. It is a cognitive process that makes it possible to pose a problem in a
thoughtful way with a view to obtaining one or more results. The objective of reasoning
is to better define/understand what has been perceived and then recognized and to verify
its reality, using alternately different “laws” and experiences.

Combining visual and automatic analysis allows users to better generate insights
and knowledge for making appropriate decisions. This proposal, as they will be demon-
strated throughout this article, aims to add interesting potential for VA to support human
decision-making. For this, we present its application on the Education field.

4 Application for Formative Assessment

A. Context
Visual Analytics in Education field proposes methods and platforms for delivering
education information through the Internet [13], trying to reach the improvement
of the learning performance of the students (groups and individuals) and forecast
their academic performance level in each subject. Many conditions are considering
like the students’ environment, social competences and emotional maturities in our
context, we propose to design and develop a Visual Analytics system, based on
the proposed process, which allows formative assessments in university courses.
Following we present the process application.
B. Process Application

Cognitive Data Collection

This research came from the difficulties that teachers face in recognizing whether their
students are benefiting from formative assessment, or not. For this, we combined the
eye tracking data and EEG signals taken from students while editing, compiling and
executing a JAVA program. The data collection was performed while making the students
in front of the computer screen doing their tasks. Hence, 55 students in the same class
were equipped with: The Electroencephalogram (EEG) device and Eye tracker:
 Improved Visual Analytic Process Under Cognitive Aspects 499

1. The Electroencephalogram (EEG). Which is a test made to the brain for measuring
the electrical activity. The EEG signals are voltage signals emerged in the neural activ-
ities. These signals can be used either to classify the state of mind or find a disorder
in neural activity [13]. The EEG also allows evaluating the mental workload in many
situations (mental fatigue, memory load while performing tasks) because the signals are
sensitized to variations in mental workload.
The EEG device used is Biosemi Active amplifiers at rate of 512 Hz. The data
was recorded from 72 channels. This device sets new standards for multi-channel, high
resolution bio potential measurement systems for research applications. The EEG signals
was band-pass filtered from 0.1 to 100 Hz. Each participant is instrumented with biosemi
active amplifiers.

2. Eye tracker. Eye Tracking is a powerful tool that has been implemented for analyzing
the user attention while during information processing tasks [11]. It is considered as an
attention index to evaluate cognitive conflicts in human-machine interaction [16]. Typi-
cally, the recorded data have temporal information on the area of interest that participants
are interested in or the fixation of the eyes on the screen.
An Eyelink 1000 tracker is used to collect the eye movement data. This device has
high speed camera. It has been designed to interchangeably fit into different mounting
options and assures the unprecedented low noise, fast sampling rates and stability [15].
Every participant putted this device, while the formative assessment in progress, and
keep doing its tasks as usual.
Data Fusion and Synchronization
After recording data from the eye tracker and EEG, we have combined these data with
EEGLab MATLAB toolbox. The original raw EEG data obtained was preprocessed in
EEGLab. The parameters of each recorded technique (eye tracking and EEG signals)
were set as input in the fuzzy logic technique with MAMDANI fuzzy inference system
and the output was the students’ concentration level. We have eight input parameters
which are: area of interest, heat map, fixation, saccade, artifact and frequency, bright-
ness, noise. The first four parameters are the parameters of the eye tracking technique,
the following two parameters are the parameters of the EEG signals and the last two
parameters made to improve the student’s concentration level since the brightness degree
as well as the noise level strongly influences on the concentration of the student.
Automatic Analysis
The cognitive data mining in our context was done using the fuzzy logic technique. The
objective is to perform an automatic analysis for the prepared data in order to predict
whether a participant is distracted or concentrated.
The fuzzy logic is a computing approach based on “degrees of truth” rather than the
usual “true or false” Boolean logic (1 or 0). It deals with the ‘0’ and the ‘1’ as extreme
cases of the truth, but also the various states of this truth located between the two. It
seems closer to the functioning of our brain.
Motivation for Using the Fuzzy Logic. In this work we deal with human being which
means that we deal with the uncertainty because, in our case, a student could have eyes
fixed on screen and high EEG signal frequency but in reality, he/she is distracted.
500 S. Bouazizi and H. Ltifi

The choice of the fuzzy logic is based on the following reasons: the fuzzy logic
technique supports ambiguity and uncertainty that can be issued from the EEG and eye
tracker devices. Also, it offers flexibility as well as simplicity of use. Fuzzy algorithms
are often robust, in the sense that they are not very sensitive to changing environments
and erroneous or forgotten rules. In addition, the fuzzy methods usually have a shorter
development time than the conventional methods.
To apply the fuzzy logic, we have associated linguistic variables to the parameters
of the prepared data as presented in Table 1. To treat uncertainty in recorded data is to
express them in fuzzy numbers form related to the linguistic variables of the membership
functions. A fuzzy number is equal to a quantity whose value is imprecise. In Table 2,
are visible the fuzzy numbers of the linguistic terms of the fuzzy logic output variable
named Concentration_level. The result will imitate more truthfully the behavior of the
student.

Table 1. Linguistic variables of the input parameters.

Parameters Linguistic variables Description

Area of interest Poor focus, average focus, high focus The area that the participant focuses
in the most
Heat map Poor n_o_g_f, average n_o_g_f, high Indicate the rate of participants’
n_o_g_f number of gaze fixation
Fixation Low fixation, average fixation, high Whether the participant fixed his
fixation eyes on screen for longtime
Saccade Short, average, long The rate of participants’ saccade
Artifact Low, medium, high The rate of signals that recorded on
the EEG and not generated by the
brain
Frequency Low, average, high The frequency of the EEG signal
Brightness Low, medium, high The brightness degree on the screen
in front of the student
Noise Low, medium, high The noise level in the class

The fuzzy logic application is supposed to treat the recorded results and reasons
using linguistic variables. It applies a set of rules and takes place in three steps:
1. Fuzzification. The numeric variables are converted to linguistic variables with a
membership function defined in the interval [0, 1].

2. Motor Inference. To apply the fuzzy logic rules in order to tolerate imprecision
and uncertainty. We have developed 52 rules. A fuzzy rule is a linguistic IF–THEN
construction,
 Improved Visual Analytic Process Under Cognitive Aspects 501

Table 2. Linguistic terms of the output variable ‘concentration_level’.

Linguistic terms Fuzzy numbers: From–To Mean

Distracted 0–10 5
Motivated not 5–15 10
concentrated
Not motivated and 10–20 15
concentrated
Motivated and 15–25 20
concentrated

We used the Mamdani’s fuzzy inference method [16]. Mamdani’s work was founded
on fuzzy algorithms for complex systems and decision processes [17].
3. Defuzzification. To produce single output and to transform this linguistic output
variable into numerical data.
Visual Analysis
Visual Mapping. The generated visualization is presented in Fig. 5. It is subdivided
into three main parts: Fig. 5A for students’ information, Fig. 5B for visual Analysis and
Fig. 5C for automatic analysis.
The goal of this visualization is to interactively analyze the data of one participant to
decide whether this participant is benefiting of this formative assessment. It allows teach-
ers to evaluate the concentration of their students (participants). The interface presents
the formative assessment of the student colored in red (cf. Fig. 5A).
The representation of the collected data is presented in Fig. 5B, which is subdivided
into two parts: the eye tracker part that indicates the data collected by the EyeLink
(participants’ fixation and saccade) and the part of EEG signal that indicates the curve
generated by the EEGLab with the indication of the participants’ fixation and saccades
time during the record phase. The legend in the right helps to specify the fixation (in
blue) and saccade (in green). The third part is the fuzzy logic (cf. Fig. 5C) contains three
parts: the rule (generated in the Mamdani fuzzy inference system) that correspond to
the current student data, the parameters (the inputs) values of the recorded data from
the current student and finally the result which indicate the concentration level of this
student.
Cognitive Activities of the Visualization
The visual analysis of the developed interface involves three main cognitive activities,
as presented in Sect. 3.

1. Perception. The visual interface integrates several visual elements (i.e. Histograms,
Heat map, Angular histogram, Main Sequence and curves) related to the eye tracker and
502 S. Bouazizi and H. Ltifi

Fig. 5. The proposed cognitive Visual Analytics process

EEG data/parameters (cf. Fig. 5B). They are graphic representation that structure and
explore causal relationships between data parts. The interactivity mechanisms (zoom in,
zoom out and Legend) support visualization flexibility. In fact, the teacher can click over
nodes (cf. Fig. 5A) to display monitoring information of other. In addition, the teacher
can visualize computerized results generated by Fuzzy Logic technique.
2. Recognition. The teacher interacts with the visual representations to gain knowledge
associated to the students’ concentration level, to retrieve data associated to the displayed
information from the dataset using the interaction techniques, and to visually interpret the
differences and/or similarities between the displayed data. To understand the generated
result, the teacher can apply the interaction mechanisms by zooming or filtering, which
can help him /her to concentrate and pay attention. He/she enthusiastically interplays
between global and local views of the displayed data. Teacher can take different viewing
positions while representations can reveal other important information.
3. Reasoning. The teacher looking at the interactive visual histograms, heatmaps, main
sequence and curves can determine why the students has such concentration level and
then analyzes this information using his/her visual reasoning skills.

After acquiring knowledge on the monitored students’ concentration level generated

by the system and his/her tacit knowledge, the teacher (i.e., decision-maker) can react
to improve the quality of its formative course.
 Improved Visual Analytic Process Under Cognitive Aspects 503

After acquiring knowledge on the monitored students’ concentration level generated

by the system and his/her tacit knowledge, the teacher (i.e., decision-maker) can react
to improve the quality of its formative course.

5 Evaluation

A. Utility Evaluation
To evaluate the utility of our adapted VA system, we have applied the confusion matrix
[18] (Fig. 6) to assess the prediction ability of our fuzzy logic. Figure 6 presents our
confusion matrix concerned by a dataset composed of 55 students where all instances
are classified according to the output linguistic variables.

Fig. 6. The confusion matrix

Figure 7 presents the predictive measures generated by the confusion matrix.

These measures are calculated by output linguistic variable (Distracted, Moti-
vated_not_concentrated, Concentrated_not_motivated, and Concentrated_and_ moti-
vated). For example, for the “Concentrated_ and_motivated” class the True Positive (TP)
Rate (TPR), which corresponds to cases where the instances are correctly classified, is
equal to 1. The False Predictive (FP) Rate, corresponding to the instances classified as
“Concentrated_and_motivated” but actually belong to another class, is equal to 0. The
accuracy rates (precision) are interesting, the lowest of which is 85,7%.

Fig. 7. Prediction measures

504 S. Bouazizi and H. Ltifi

B. Usability Evaluation
The aim is to assess our visual mapping usability; we have made a controlled experiment
where 30 persons have participated. In this study, we have designated 3 categories of
academic participants: novice, knowledge-intermittent and expert. Table 3 presents the
user profiles that have been involved for the evaluation task.

Table 3. Participants profile.

Types of participants Number Age Skill

Novice 10 students 18–20 Modest experience in
using computer
Knowledge-intermittent 10 scientific students 21–34 Good experience in using
computer: 15–17 years
Expert 10 computer science teachers 35–45 Expert experience in using
computer

Evaluation Protocol. We divided the participants randomly into groups of five. We

asked them each individually to remember strategy implementation problems that they
had already experienced. We collected that individual experience and then let the groups
work for 45 min. Half of the groups using non-visual user interface support and half of the
groups using our he visual system. After that, we issued a post experiment survey, where
we collected subjective metrics such a satisfaction with the outcome and satisfaction
with the process. We also captured all the results. 45 min later, after a distraction task,
we issued a recall test to see how much these participants remember from the meeting
they had.

Evaluation Results. We found that the results are very intriguing. Not only we did find
that the productivity of the teams using our visual tool was significantly higher but also
the quality of their work was higher even more. So, on individual basis the recall and
knowledge gains that we tested were much higher in the visual tool support groups than
in the non-visual groups. Through this controlled experiment, we have shown that our
tool is very relevant for knowledge acquiring and decision-making.

6 Conclusion

The proposed process integrates cognitive aspects from data collection [21] to the auto-
matic and visual analysis. It allows studying and analyzing cognitively collected data to
gain related knowledge [20]. The proposed process was applied to support the formative
assessments. We have developed several visualization techniques to visually analyze
cognitive data sets issued from Eye Tracker and EEG devices. These data are displayed
using interactive graph representations (histogram, heatmap, etc.) to be analyzed and
 Improved Visual Analytic Process Under Cognitive Aspects 505

visually interpreted (perceived, recognized, and reasoned) for generating related knowl-
edge. We have developed the Fuzzy Logic technique as a data mining technique able
to deal with and analyze the collected cognitive data. We have evaluated the system
utility by verifying the prediction ability of the fuzzy logic using the matrix confusion,
which shows interesting results. In addition, we have conducted a controlled experiment-
based usability evaluation for this visual tool. It showed that the knowledge gains were
significant. Such evaluation reflected the feasibility of the suggested process.
Future outlook of this work concerns the application of the proposed process for
acquiring knowledge other kinds of patterns such as the association rules and proba-
bilities. We plan also to propose a cognitive VA process that generates decisions from
knowledge.

References
1. Ltifi, H., Kolski, C., Ben, A.: Combination of cognitive and HCI modeling for the design of
KDD-based DSS used in dynamic situations. Decis. Support Syst. 78, 51–64 (2015)
2. Keim, D., Andrienko, G., Fekete, J.-D., Görg, C., Kohlhammer, J., Melançon, G.: Visual
analytics: definition, process, and challenges. In: Kerren, A., Stasko, J.T., Fekete, J.-D., North,
C. (eds.) Information Visualization. Lecture Notes in Computer Science, vol. 4950, pp. 154–
175. Springer, Heidelberg (2008)
3. Blascheck, T., John, M., Kurzhals, K., Koch, S., Ertl, T.: VA2: a visual analytics approach
for evaluating visual analytics applications. IEEE Trans. Vis. Comput. Graph. 22(1), 61–70
(2016)
4. Ltifi, H., Amri, S., Ben Ayed, M.: Fuzzy logic–based evaluation of visualizations generated
by intelligent decision support systems. Inf. Vis. 17(1), 3–21 (2016)
5. Borra, S., Dey, N., Bhattacharyya, S., et al. (eds.): Intelligent Decision Support Systems.
Applications in Signal Processing. De Gruyter, Berlin, Boston (2019)
6. Li, G., Kou, G., Peng, Y.: A group decision making model for integrating heterogeneous
information. IEEE Trans. SMC Syst. 48(6), 982–992 (2018)
7. Fayyad, U.M., Djorgovski, S.G., Weir, N.: Automating the analysis and cataloging of sky sur-
veys. In: Fayyad, U.M., Piatetsky-Shapiro, G., Smyth, P., et al. (eds.) Advances in Knowledge
Discovery and Data Mining, pp. 471–494. The MIT Press, Menlo Park, CA (1996)
8. Ltifi, H., Benmohamed, E., Kolski, C., Ben Ayed, M.: Enhanced visual data mining process
for dynamic decision-making. Knowl. Based Syst. 112, 166–181 (2016)
9. Wijk, V.J.J.: The value of visualization. In: IEEE Visualization, pp. 79–86 (2005)
10. Sacha, D., Stoffel, A., Stoffel, F., Kwon, B.C., GKeim, D.A.E.: Knowledge generation model
for visual analytics. IEEE Trans. Vis. Comput. Graph. 20(12), 1604–1613 (2014)
11. Rayner, K.: Eye Movements and Visual Cognition: Scene Perception and Reading. Springer,
New York (2012)
12. Wang, S.-L., Hwang, G.-J.: The role of collective efficacy, cognitive quality, and task cohesion
in computer-supported collaborative learning (CSCL). Comput. Educ. 58(2), 679–687 (2012)
13. Pradhan, C., Shariar Rahaman, M., Sheikh, A., Kole, A., Maity, T.: EEG signal analysis using
different clustering techniques. In: Abraham, A., Dutta, P., Mandal, J.K., Bhattacharya, A.,
Dutta, S. (eds.) Emerging Technologies in Data Mining and Information Security: Proceedings
of IEMIS 2018, Volume 2, pp. 99–105. Springer, Singapore (2019)
14. Dehais, F., Causse, M., Vachon, F., Tremblay, Sébastien.: Cognitive conflict in human–
automation interactions: a psychophysiological study. Appl. Ergon. 43(3), 588–595 (2012)
15. Research SR: Eyelink 1000 users’ manual, version 1.5. 2 (2010)
506 S. Bouazizi and H. Ltifi

16. Mamdani, E.H., Assilian, S.: An experiment in linguistic synthesis with a fuzzy logic
controller. Int. J. Man Mach. Stud. 7, 1–13 (1975)
17. Zadah, L.A.: Outline of a new approach to the analysis of complex systems and decision
processes. IEEE Trans. Syst. Man Cybern. 3(1), 28–44 (1973)
18. Mohana Shankar, P.: Pedagogy of Bayes’ rule, confusion matrix, transition matrix, and
receiver operating characteristics. Comput. Appl. Eng. Educ. 27(2), 510–518 (2019)
19. Jemmaa, A., Ltifi, H., Ayed, M.: Multi-agent architecture for visual intelligent remote health-
care monitoring system. In: Abraham, A., Han, S.Y., Al-Sharhan, S.A., Liu, H. (eds.) Hybrid
Intelligent Systems. Advances in Intelligent Systems and Computing, vol. 420, pp. 211–221.
Springer, Cham (2016)
20. Ogiela, L.: Syntactic approach to cognitive interpretation of medical patterns. In: Xiong,
C., Huang, Y., Xiong, Y., Liu, H. (eds.) Intelligent Robotics and Applications, pp. 456–462.
Springer, Heidelberg (2008). https://doi.org/10.1007/978-3-540-88513-9_49
21. Ogiela, L.: Cognitive informatics in automatic pattern understanding and cognitive informa-
tion systems. In: Yingxu Wang, D., Zhang, W.K. (eds.) Advances in Cognitive Informatics
and Cognitive Computing, pp. 209–226. Springer, Heidelberg (2010)
 Secure Distribution of Factor Analysis of Mixed
Data (FAMD) and Its Application
to Personalized Medicine of Transplanted
Patients

Sirine Sayadi1,2(B) , Estelle Geffard2 , Mario Südholt1 , Nicolas Vince2 ,

and Pierre-Antoine Gourraud2
1STACK Team, IMT Atlantique, Inria, LS2N, Nantes, France
{Sirine.Sayadi,Mario.Sudholt}@imt-atlantique.fr
2 University of Nantes, Nantes University Hospital, INSERM, Research Center in
Transplantation and Immunology, UMR 1064, ATIP-Avenir, Nantes, France
[email protected],
{Nicolas.Vince,Pierre-Antoine.Gourraud}@univ-nantes.fr

Abstract. Factor analysis of mixed data (FAMD) is an important statistical tech-

nique that not only enables the visualization of large data but also helps to select
subgroups of relevant information for a given patient. While such analyses are
well-known in the medical domain, they have to satisfy new data governance
constraints if reference data is distributed, notably in the context of large consor-
tia developing the coming generation of personalised medicine analyses.
In this paper we motivate the use of distributed implementations for FAMD
analyses in the context of the development of a personalised medicine application
called K I TAPP. We present a new distribution method for FAMD and evaluate its
implementation in a multi-site setting based on real data. Finally we study how
individual reference data is used to substantiate decision making, while enforcing
a high level of usage control and data privacy for patients.

1 Introduction

Big data analysis techniques are increasingly popular to extract new information from
massive amounts of data to improve decision making, notably in the medical sector.
A major challenge for clinicians consists in safely making correct treatment decisions
based on ever growing amounts of patient data. This problem calls for new analysis
techniques and algorithms, in particular, for precision medicine.
Precision Medicine, that is using genetic or molecular profiling for optimizing care
of small patient groups, will probably become the standard of care in the next decades.
It represents a deep revolution in health care also because not only patients will be cov-
ered but also healthy individuals. For this (r)evolution of medicine to become reality,
it is necessary to deliver the evidence of the efficiency and cost-effectiveness of preci-
sion medicine. This, in turn, requires decisions concerning patients to be substantiated

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 507–518, 2021.
https://doi.org/10.1007/978-3-030-75100-5_44
508 S. Sayadi et al.

by an analysis of large-scale reference data that is relevant for their personal situation
compared to others [2].
Dimension reduction is a major technique for transforming large multi-dimensional
data spaces into a lower dimensional subspaces. This is while preserving significant
characteristics of the original data. Among dimension reduction methods, the most
common method is Principal Component Analysis (PCA) [3], which enables dimen-
sion reduction for quantitative data variables. Other methods are Factor Analysis of
Mixed Data (FAMD) [5], which performs dimension reduction for mixed (quantitative
and qualitative) data variables, and dictionary learning (DL) [4] one of the most power-
ful methods of extracting features from data.
FAMD analysis provides simplified representations of multi-dimensional data
spaces in the form of a point cloud within a vector subspace of principal components.
If two points are close to each other in this cloud, a strong global similarity exists
between them with respect to the selected principal components. In the biomedical field,
this kind of analysis is frequently used to present patients groups in a simplified and
visual way for a large range of complex clinical data encompassing quantitative data
(for example obtained from biological exams) and qualitative data (for example gen-
der information). The result are actionable representations of each patient’s individual
characteristics compared to those of others.
In the context of a French public-private partnership KTD-innov and a H2020 EU
project EU-train, FAMD has been used for dimension reduction as part of the clinical
decision support system K I TAPP (the kidney transplant application) [6]. This precision
medicine web-application computes predictive scores and represents distributions of
patients’ variables in a subgroup of reference patients after kidney transplantation. The
application is conceived to relay the intuition and experience of clinicians by means
of on-demand computations and graphical representations. One of K I TAPP’s key func-
tionalities consists in the “contextualization” of patients relative of a population of ref-
erence (POR). To this end, it first uses FAMD for dimension reduction, then applies a
percentile statistical modelling [21] algorithm, and visualizes the relations of patients
to the POR.
Medical studies often involve large national or international collaborations (such as
our KTD-innov and EU-Train projects). Simple centralization schemes for the place-
ment of data and computations are frequently not applicable in this context because data
and computations may not be shared due to legal reasons, security/privacy concerns
and performance issues). To deploy this kind of medical services in larger contexts dis-
tributed systems and algorithms for precision medicine have to be provided. One of the
main lines of research around the KTD-innov and EU-TRAIN projects is the imple-
mentation of a reference database integrated into a distributed computing infrastructure
allowing a secure access to data while respecting the European GPDR data protection
regulation [12]. However, very few distributed algorithms have been developed for and
applied to the domain of precision medicine.
Data sharing and analysis placement are generally difficult due to governance, reg-
ulatory, scientific and technical reasons. Analyses are often only possible “on premise.”
Furthermore, researchers and institutions may be averse to lose control over both data
usage. In addition, huge volumes of data are intrinsically difficult to share or transfer,
 Secure Distribution of Factor Analysis of Mixed Data (FAMD) 509

notably because of cost of the use of computational, storage and network resources.
On the other hand distributed architectures enable more flexible data governance
strategies and analysis processes by freeing them from centralization constraints [9].
Decentralized databases enable performing local calculations on patient data, without
any individual data circulating outside the clinical centers generating the data. To this
end, one may strive for distributed statistical calculations. Fully-distributed analyses
have been proposed, see Sect. 2, for contextualizing the state of a patient relative to
POR data stored in a distributed database. Such algorithms have to meet requirements
of scalability, security and confidentiality [10], as well as availability properties and
right to privacy properties [11]. These criteria are difficult to satisfy, however, because
the statistical significance and accuracy of analyses often directly depend on the number
of cases or individuals included in the database.
A solution to these problems can be based on harnessing distributed analyses that
manipulate sensitive data on the premises of their respective owners and harness dis-
tributed computations if non sensitive, aggregated, summarized or anonymized data is
involved.
In this paper, we present two main contributions:

– We motivate and define requirements for distributed algorithms for dimension reduc-
tion in the context of the K I TAPP project.
– We present a novel distributed FAMD algorithm for dimension reduction in the pres-
ence of sensitive data in precision medicine and apply it to a contextualization prob-
lem.
The rest of this article is organized as follows. Section 2 presents related works.
Section 3 presents the kidney transplantation application (K I TAPP) and its use of FAMD
dimension reduction. Section 4 presents our distributed algorithm FAMD and a corre-
sponding implementation. Section 5 provides experimental results, notably concerning
privacy requirements, and a performance evaluation. Finally, Sect. 6 summarizes our
findings and proposes some future work.

2 Related Work
Parallel versions of PCA dimension reduction algorithms have already been proposed.
Liang et al. [14] propose a client-server system and send singular vectors and singu-
lar values U Σ V from the client to the server. Feldman et al. [15] have shown how to
compute PCAs by sending smaller matrices U Σ instead of sending all matrices of a sin-
gular value decomposition, thus improving on the communication cost. Wu et al. [16]
have introduced an algorithm that improves on the storage and data processing require-
ments and harnesses Cloud computing for PCA dimension reduction. These proposals
send matrices of synthesized data of the original data and not real data. This is very
interesting for biomedical analyses in order to ensure privacy of patient data. Imtiaz
et al. [17] have improved Feldman et al.s’ proposal by adding privacy guarantees using
differential privacy.
510 S. Sayadi et al.

To the best of our knowledge, no distributed FAMD algorithm has already been
proposed. In this paper, we propose a distributed FAMD on the basis of a distributed
PCA algorithm. Our algorithm is structured into two parts (similar to Pagès [5]):

1. Transform qualitative data into quantitative one using complete disjunctive

tables [13], thus transforming the original FAMD dimension reduction problem into
a PCA one.
2. Perform the dimension reduction of the distributed PCA based on Feldman et al. s’
proposal [15].

3 The Kidney Transplantation Application (K I TAPP)

Chronic kidney failure affects approximately 10% of the world population and can pro-
gressively lead to end-stage kidney disease requiring replacement therapy (dialysis or
transplantation). Kidney transplantation is the best treatment for end-stage kidney dis-
ease [19].

3.1 K I TAPP Overview

Data from approximately 1500 renal transplantation, including clinical and immuno-
logical items, were collected since 2008 as part of a French national project.
K I TAPP enables personalized contextualization algorithm to be harnessed to com-
pare data trajectories of a given patient (POI) to a sub-population with similar charac-
teristics (POR) selected by filters or distance measures. The information relative to a
graft is selected from similar cases at the time of the graft. With the help of clinicians
and knowledge of the existing body of research, we defined a set of variables to select
the sub-population of reference.
We propose three population contextualization algorithms: compare a given
patient’s data to PORs with

1. similar characteristics selected by filters or approaches based on statistical analysis;

2. the nearest neighbor method or
3. the cluster method.

With our filter approach (1), the POR is defined according to selected filters made avail-
able to the clinician, such as age, gender and Body Mass Index (BMI). Methods (2) and
(3) are based on the results of an FAMD. Following this analysis, we can then select
a POR by close neighbor method (2): by selecting the N individuals most similar to a
POI or we can select a POR by clustering (3): by selecting the individuals in the same
cluster as our POI.
The visualization of contextualized information is done by comparing a POI’s bio-
logical data (creatinemia) and its evolution over time post-transplantation (clinical vis-
its) to a POR that is represented by their median and percentile values.
 Secure Distribution of Factor Analysis of Mixed Data (FAMD) 511

3.2 Motivation for Distributed Analyses for K I TAPP

We intend to harness the K I TAPP application as part of large-scale cooperations with

many (national and international) partners. To meet the challenge of harnessing medical
data while keeping sensitive data on premise or ensure strong data protection if data
is moved, computations are often performed today over distributed databases that are
linked to a computation integrator that enables a center to interact with and access
some data from remote sites. Each clinical center collects, stores and controls their own
patients’ data. The founding principle of the architecture is that no data of individuals
circulates outside the centers. However, this sharing paradigm is very restrictive and
inhibits a large range of potential analyses to be performed - either because sensitive
data cannot be appropriately protected or the analysis cannot be performed sufficiently
efficient.
The need for local storage and distribution of reference data is motivated by the
actionable value and publication value. It provides the possibility of controlling locally
who has accessed to data, what are the usages of the local data and how to limit then
should it be needed. The use of distributed infrastructure is a central element of multi-
stakeholders data governance.
We are therefore working on more general distributed analysis architectures and
implementations that ease collaboration as part of multi-centric research projects, where
each center can control and account for their own patients’ data usage even if located
remotely. Contextualization then has to be performed with respect to large-scale dis-
tributed medical databases that are maintained at different sites.

4 Distributed FAMD

In the following we first provide an overview over the architecture and properties of our
algorithm before defining it in detail.œ

4.1 Overall Architecture and Properties

Factor analysis of mixed data (FAMD) [5] is a method of dimension reduction of vari-
ables including mixed quantitative and qualitative data into fewer components for infor-
mation synthesis reasons. This analysis can defined, for instance using matrix opera-
tions, as follows:
FAMD = PCA + MCA (1)
where PCA is a principal component analysis dimension reduction for quantitative vari-
ables and MCA is a multiple correspondence analysis dimension reduction for qualita-
tive variables.
Overall, our algorithm works as follows. As a first step, we transform the qualitative
variables into quantitative ones using complete disjunctive coding (CDC) [20] that is
performed locally on each site. In a second step we perform dimension reduction by
means of a distributed PCA in order to obtain a secure and distributed FAMD algorithm.
512 S. Sayadi et al.

Fig. 1. Collaboration architecture

We harness the distributed cooperation architecture shown in Fig. 1. Data transfor-

mation and dimension reduction analysis are performed locally at multiple sites sepa-
rately. The coordination between the sites is done by an aggregator site, which receives
synthetic data and performs the overall dimension reduction. Imtiaz et al. [17] have pro-
posed a secure and distributed algorithm for PCA dimension reduction. This algorithm
uses differential privacy as a security technique and synthetic data for communication
between nodes.
The resulting algorithm has two important properties :
• Low communication cost: The communication cost of parallel and distributed
FAMD algorithms essentially depends on the size of the matrices transferred
between sites. Many dimension reduction algorithms require matrices of size D × D
to be sent, where D is the number of data items to be analyzed (that is, transplanta-
tion data in our case). In contrast, Imtiaz et al.’ [17] algorithm requires matrices to
be sent of type D × R where R is the number of variables and (typically) R  D.
• Security/privacy awareness: Our algorithm satisfies two interesting characteristics:
(1) differential privacy is used for data protection and (2) communication between
the sites and the aggregator involves only synthesized data PS and not the original
data, which minimizes possibilities of data theft and also supports data protection.

We harness the same principle and properties while providing two new contribu-
tions: (1) a transformation of qualitative variables into quantitative variables to obtain
a secure and distributed algorithm FAMD dimension reduction and (2) a scalable dis-
tributed implementation and evaluated it by analyzing real-world biomedical data on a
realistic grid environment.

4.2 Algorithm Definition

Algorithm 1 presents our secure and distributed FAMD algorithm. Lines 1–16 imple-
ment the first step, the transformation of a full FAMD problem into a (qualitative)
PCA problem. Each site begins by calculating, for each quantitative variable, the cor-
responding mean μk , standard deviation σk and Centering and Reduction Function
Xi,k = σ1 (xi,k − μk ). For each qualitative variable, Complete Disjunctive Coding using
k
 Secure Distribution of Factor Analysis of Mixed Data (FAMD) 513

Algorithm 1: Distributed FAMD Algorithm

Input : Data matrix Xs ∈ RD×Ns for s ∈ [S] of N elements and P variables, with C
quantitative variables and M qualitative variables; ε , δ : privacy parameters; j :
reduced dimension;
Output: V j : Matrix of eigenvectors on top j
1 foreach site s ∈ S do
2 foreach element i ∈ N do
3 foreach element k ∈ C do
4 Compute the mean of the variable μk ;
5 Compute the standard deviation of the variable σk ;
6 Compute the Centering and Reduction Function Xi,k = σ1k (xi,k − μk );
7 return Xi,k ;
8 end
9 foreach element k ∈ M do
10 Apply the Complete Disjunctive Coding using (ade4 package on R);
11 Compute the effective of the modality Nk ;
12 Compute the proportion pk = Nk /N;
x
13 Compute the Indicator Weighting Function Xi,k = √ i,k p,k ;
14 return Xi,k ;
15 end
16 end
17 Compute As = N1s Xs XsT ;
18 Generate D × D symmetric Matrix E where Ei, j : i ∈ [D], j ≤ i drawn i.i.d. from

N(0, Δε2,δ ) where Δε ,δ = Ns1 ε 2log( 1.25
δ ), Ei, j = E j,i ;
19 Compute As = As + E;
20 Perform SVD(As ) = U Σ U T ;
21 Compute Ps = U Σ 1/2 ;
22 Send Ps to the aggregator;
23 end
24 Compute A = 1s ∑ss=1 Ps PsT ;
25 Perform SV D(A) = V Λ V T ;
26 Send V j to all sites ;
27 return V j ;

the ade4 package from the R language is then applied in order to transform the qualita-
tive variables into a quantitative variable, followed by the computation of the modalities
Nk and proportions pk = Nk /N to compute the indicator weighting function Xi,k .
The second step, the dimension reduction proper, is implemented on lines 17–23.
Each site calculates the (second moment) matrix As = NS 1
Xs XsT . The application of the
scheme of differential privacy (following Dwork et al.’ proposal [18]) is performed by
generating the noise matrix E of size D×D and the estimated differential privacy matrix
As = As + E on line 19. Each site then performs the Singular Value Decomposition
SVD(As )) of matrix As to compute the matrix (Ps = U Σ 1/2 ) and broadcast it to the
aggregator.
514 S. Sayadi et al.

At the aggregator site, the server computes, see lines 24–26, the matrix A =
1
s ∑ss=1 Ps PsT
of all sites. It performs next the global Singular Value Decomposition
SVD(A) = V Λ V T .

5 Experimentation
In this section we report on experiments involving analyses over real medical data that
we have carried out on a real heterogeneous large-scale grid infrastructure. We report
on our setup, and evaluate our implementation w.r.t. three criteria.

5.1 Setup
Our experiments have been carried out on renal transplantation data available in the
European database Divat [22]. In order to compare with results from the K I TAPP
project, we have applied our distributed algorithm to its analyses on 11,163 transplanta-
tion data. We started by divided the data file before transfer and analysis on the different
sites.
We have implemented our distributed algorithm and executed it in a grid-based envi-
ronment featuring different distributed architectures, ranging from placing all clients on
different (geo-distributed) machines to placing them as one cluster on only one machine.
This distributed environment constitutes a realistic architecture of a medical collabora-
tion involving the research and clinical centers, the partners of the K I TAPP project.
We have implemented our distributed algorithm using the Python and R programming
language using 860 lines of code. The whole distributed system can be deployed and
executed on an arbitrary number of sites of the Grid’5000 infrastructure using a small
script of only eight commands.
The Grid’5000 platform is a platform, built from eight clusters in two European
countries for research in the field of large-scale distributed systems and high perfor-
mance computing. For our experiment, we have reserved a machine as a server (aggre-
gator) executing a Python program to manage the analysis, client interactions and gen-
eration of the final result. To create a number of client sites we have reserved machines
distributed over five different sites in France.

5.2 Results and Performance Evaluation

The K I TAPP-motivated FAMD dimension reduction analysis we employed as a test
case has been executed on the basis of 11,163 transplantation operations characterized
using 27 qualitative and quantitative variables distributed over five sites. We have set
the dimension reduction parameter j on the server to two. Figure 2 shows the resulting
two-dimensional subspace after application of our distributed FAMD analysis.
In order to distribute the POR selection for POI contextualization between sites, we
have applied the K-means unsupervised clustering technique to the result of the dis-
tributed FAMD dimension reduction analysis. FAMD and clustering enables grouping
of patient data according to their similarity and proximity relative to principal compo-
nents. Figure 3 presents the result of k-means clustering, three independent data clusters
 Secure Distribution of Factor Analysis of Mixed Data (FAMD) 515

Fig. 2. Distributed FAMD for 5 sites. Fig. 3. Clusterning of distributed FAMD.

that correspond exactly to the result of the (centralized) sequential algorithm that is used
as part of the K I TAPP project.
Each cluster is characterized by specific variables combinations. The green cluster
corresponds to living donors. The yellow cluster corresponds to deceased donors and
the purple one to deceased donors with expanded criteria, such as aged we > 50 years or
subject to hypertension or creatinine levels ≥ 133 μ mol/L. Note that we always obtain
the same clusters independent from the number of sites that participate in the distributed
FAMD analysis if we operate it with the same data, which shows a strong scalability
potential of our proposed algorithm.
In the following we evaluate three properties of our implementation: (i) the quality
of the reduction technique in the presence of noise introduced by the differential pri-
vacy technique using a notion of captured energy, (2) execution time and (3) commu-
nication cost. For evaluation purposes we consider three architectures: our distributed
FAMD reduction technique (denoted“DPdis” below), a more centralized version where
W where all the second moment matrix As of each clients are aggregated at the server
(denoted “fulldis”), and a fully-centralized FAMD version (denoted “pooled”).
Captured Energy/Utility. Following Imtiaz et al. the captured energy q is used to
evaluate the quality of V j principal directions based on the difference in information
utility between the case where all data is centralized qpooled , all second moment matrix
As of each site are distributed qfulldis and secure distributed proposed FAMD algorithm
qDPdis by data size and number of sites.
The captured energy is defined as the matrix multiplication q = tr(V j (A)T AV j (A)
measuring the amount of optimal eigenvalues captured in the subspace FAMD. For any
other sub-optimal sub-spaces, the value would be less than the optimal value.

– Energy per site. We have varied the number of S sites that participate in this analysis
by keeping the total number of samples N = 11163 (i.e. we decreased the size Ns
of each site). Figure 4 shows a deterioration in the performance of qfulldis and qDPdis
for an increasing number of sites. This decrease in performance is explained by the
decrease in the number of elements per site. In addition, the presence of high vari-
ance noise degrades the number of eigen-directions stronger than the noise which
516 S. Sayadi et al.

are detected by the PCA instead of capturing all the j directions which present the
data.
– Energy by data size. Figure 5 shows an increase in performance of captured energy
q as a function of the elements number per site. qfulldis and qDPdis have almost the
same performance in captured energy for the two variation of sites number and data
size, qpooled always keeps better performance.

Fig. 4. Captured energy Fig. 5. Captured energy (q) Fig. 6. Execution time by sites
(q) by sites number. by data size. number.

Execution Time. We have varied the number of S sites by keeping the global number
of samples N = 11163 (i.e. we decreased the size Ns of each site) and we have measured
the execution time. Figure 6 shows that execution time decreases with increasing sites
number. Our proposed approach d pdis always keeps the least execution time. This is
due to the Lower communication cost explained in 4.1 section.
Communication Cost/Data Sharing: The lower cost of communication of our pro-
posed algorithm introduced in the previous section allows for a minimum sharing of
data (matrices Ps are shared and not As ). In the case of our experiments with distri-
bution on 5 sites with global samples equal to 11, 163 and 27 features, the quantity of
data shared by all clients is equal to 11, 163 × 27 = 301.401 values instead of sending
sqr(11, 163) = 124.612.569 values.

6 Conclusion and Future Work

FAMD dimension reduction is an important tool for transforming complex data into
lower-dimensional sub-spaces while preserving important characteristics of the orig-
inal data. This technique is generally useful to reduce complexity and support deci-
sion making. In this paper, we have motivated the use of dimension reduction for geo-
distributed biomedical collaborations that require distributed models and implementa-
tions of biomedical algorithms with distributed implementation. Its evaluation on a real
geo-distributed grid infrastructure using real data has valided its efficiency, scalability,
privacy protection properties.
As future work, we will focus on extensions of federated learning as a more general
method for the definition of secure and distributed biomedical analyses.
 Secure Distribution of Factor Analysis of Mixed Data (FAMD) 517

References
1. Hood, L.: Systems biology and p4 medicine: past, present, and future. Rambam Maimonides
Med. J. (2013)
2. Gourraud, P., Henry, R., et al.: Precision medicine in chronic disease management: the mul-
tiple sclerosis bioscreen. Ann. Neurol. 76(5), 633–642 (2014)
3. Jolliffe, I.T.: Principal Component Analysis. Springer Series in Stat. Springer (2002)
4. Shakeri, Z., Sarwate, A.D., Bajwa, W.U.: Sample complexity bounds for dictionary learning
from vector and tensor valued data. In: Rodrigues, M., Eldar, Y. (eds.) Information Theoretic
Methods in Data Science, Chapter 5. Cambridge University Press (2019)
5. Pagès, J.: Multiple Factor Analysis by Example using R. Chapter 3 (2014)
6. Herve, C., Vince, N., et al.: P218 The kidney transplantation application (KITAPP): a visual-
ization and contextualization tool in a kidney graft patients’ cohort. Hum. Immunol. 78, 216
(2017)
7. KTD-Innov. www.ktdinnov.fr. Accessed 07 Jan 2021
8. EU-TRAIN. eu-train-project.eu . Accessed 07 Jan 2021
9. Brous, P., Janssen, M., et al.: Coordinating decision-making in data management activities: a
systematic review of data governance principles. In: International Conference on Electronic
Government. Springer (2016)
10. Boujdad, F., Gaignard, A., et al.: On Distributed Collaboration for Biomedical Analyses WS
CCGrid-Life (2019)
11. Scheel, H., Dathe, H., Franke, T., Scharfe, T., Rottmann, T.: A privacy preserving approach to
feasibility analyses on distributed data sources in biomedical research. Stud. Health Technol.
Inform. 267, 254–261 (2019)
12. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016
on the protection of natural persons with regard to the processing of personal data and on
the free movement of such data, and repealing Directive 95/46/EC (General Data Protection
Regulation). http://data.europa.eu/eli/reg/2016/679/oj. Accessed 07 Jan 2021
13. Greenacre, M., Blasius, J.: Multiple Correspondence Analysis and Related Methods. Chap-
man and Hall/CRC Press, London (2006)
14. Liang, Y., Balcan, M., Kanchanapally, Y.: Distributed PCA and k-Means Clustering (2013)
15. Feldman D., Schmidt, M., Sohler, C.: Turning big data into tiny data: constant-size coresets
for K-means, PCA and projective clustering. In: Proceedings of the Twenty-fourth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2013, pp. 1434–1453 (2013)
16. Wu, Z., Member, Li, Y., Plaza, A., Li, J., Xiao, F., Wei, Z.: Parallel and distributed dimen-
sionality reduction of hyperspectral data on cloud computing architectures. IEEE J. Select.
Top. Appl. Earth Obser. Remote Sens. 9, 1–9 (2016)
17. Imtiaz, H., Sarwate, A.: Differentially private distributed principal component analysis. In:
2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
pp. 2206–2210 (2018)
18. Dwork, C., McSherry, F., Nissim, K., Smith, A.: Calibrating noise to sensitivity in private
data analysis. In: Halevi, S., Rabin, T. (eds) Theory of Cryptography. TCC 2006. Lecture
Notes in Computer Science, vol. 3876. Springer, Berlin, Heidelberg (2006)
19. Hill, N., Fatoba, S., et al.: Global prevalence of chronic kidney disease - a systematic review
and meta-analysis. PLoS One (2016)
20. Mellinger, M.: Correspondence analysis in the study of lithogeochemical data: general strat-
egy and the usefulness of various data-coding schemes. J. Geochem. Explor. 21(1–3), 455–
469 (1984). ISSN 0375-6742
21. Sayadi, S., Geffard, E., Südholt, M., Vince, N., Gourraud, P.: Distributed contextualization
of biomedical data: a case study in precision medicine. In: AICCSA 2020 - 17th IEEE/ACS
International Conference on Computer Systems and Applications, pp. 1–6, November 2020
518 S. Sayadi et al.

22. Divatfrance. www.divat.fr. Accessed 07 Jan 2021

23. Balouek, D., Carpen Amarie, A., Charrier, G., et al.: Adding virtualization capabilities to the
grid’5000 testbed. In: Cloud Computing and Services Science. Springer (2013)
 Comparative Study of Traditional Techniques
for Unsupervised Autonomous Intruder
Detection

Anik Alvi(B) , Tarem Ahmed, and Mohammad Faisal Uddin

Independent University, Bangladesh (IUB), Dhaka, Bangladesh

{alvi,tarem,faisal}@iub.edu.bd

Abstract. In this paper we investigate five traditional techniques to extract fea-

tures within a face image, and we evaluate them by applying the Kernel-based
Online Anomaly Detection (KOAD) algorithm. The main objective of this work
is to explore the various fundamental feature extraction techniques that can be
used to identify whether a person’s face is covered by a mask or not. Although
face covering or wearing a mask is recommended during this global COVID-19
pandemic, deliberate face occlusion is considered to be a suspicious activity in
a normal scenario. Even during this pandemic, it may be considered suspicious
if someone is covering his/her face inside an ATM booth or an apartment com-
plex during odd hours, for instance. Our proposed framework detects such intru-
sion activity by combining a traditional face detection algorithm with KOAD.
Comparative analysis is performed for each filter used and we show that our
proposed system achieves high detection accuracy with low computational com-
plexity, while also providing the added benefits of being adaptive, portable, and
involving low infrastructural costs.

1 Introduction

In the modern world, visual surveillance is present in almost every place. From malls
to banks, offices to homes, every place is monitored with many cameras for security
reasons. It becomes tedious and monotonous for a security officer to stare at multi-
ple screens for a long periods of time. This leads to many suspecting activities going
unnoticed. To avoid such incidence, there is a need for automated visual surveillance
algorithms which can detect any intrusion activity removing the burden of human secu-
rity officers to constantly monitor. Before the COVID-19 pandemic, wearing a mask
was an unusual activity and was regarded as suspicious, especially in sensitive places
like ATM booths and offices. Now, not wearing a mask in these places can be consid-
ered an anomalous activity. For both scenarios, detecting and informing such activity
to security officers has become an important issue. We hope that the present pandemic
scenario is a temporary one, upon the end of which camera operators would go back to
signaling the wearing of a mask as the anomalous activity to be alerted to.
We investigate a framework consisting of Face Detection, Feature Extraction and
Unsupervised Pattern Matching steps to develop a system of autonomously detecting
anomalous images from an image sequence. The objective is to develop a framework
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 519–530, 2021.
https://doi.org/10.1007/978-3-030-75100-5_45
520 A. Alvi et al.

which will signal the absence of a mask in the present pandemic state-of-affairs, and
conversely the presence of any face covering in the old normal future. To this end, we
incorporate the legacy Viola-Jones Face Detector [2], perform a comparative study of
the fundamental feature extraction techniques of Wavelet Decomposition [3], the Gabor
Filter [4], and Canny [5], Laplacian [6], and Sobel Edge Detectors [7], and the Kernel-
based Online Anomaly Detection (KOAD) [8, 9] algorithm to learn in unsupervised
manner.
Whilst a plethora of commercial tools for mask detection have recently sprung up
during the pandemic [10], a vast majority of these software remain proprietary. Their
algorithmic foundations are not available in the public domain, which make it difficult
to customize the software or to further build on it. We believe that developing image
processing algorithms based on established theoretical foundations will lead to robust
schemes which stand the test of time and offer a wider scope of applications.

1.1 Related Work

Face mask detection has been a new area of research due to the pandemic. With the gov-
ernment wanting to monitor whether the citizens are properly wearing masks or not,
many software companies [10] are linking existing face detection and occluded face
detection algorithms to detect whether a person is wearing a mask or not. Recently, a
deep learning model like Inception V3 [11] was trained in a masked face data set and
achieved almost full accuracy. For the case of visual intruder detection, research is being
performed to recognize posture [12] and detect intruders. Bouchrika et al. [13] used gait
analysis to identify and track a person. Even occluded face detection can be considered
as intruder detection from the assumption that the person is concealing his/her face.
Algorithms like Face Attention Network [14] have used anchor-level attention to high-
light the features from an occluded face to detect it. One of the disadvantages of all
these algorithms is that they are not hardware independent, and so require expensive
sophisticated hardware to be installed.

1.2 Our Contribution

In this paper, we investigate traditional algorithms that can be used to detect whether a
person is wearing a mask or not. Our proposed method use a face detector [2] to extract
faces from images, use various filters to extract features from the face, and uses a pattern
matching algorithm [15–17] to identify whether a person may be flagged as an intruder
or not. Our paper is limited to detecting whether a face is covered by mask or not. In
the real-world environment, different objects may appear which is outside the scope of
this paper. The following primary contributions are presented, with each experimentally
demonstrated:
1. To the best of our knowledge, this work constitutes the first thorough exploration of
the legacy techniques to the problem at hand.
2. Results of our proposed approach are compared for two different environment set-
tings. While the first setting is appropriate for public areas (e.g. ATM booths, Gov-
ernmental facilities, Airports, etc.) under surveillance, the second setting is more
 Study of Traditional Techniques for Unsupervised Autonomous Intruder Detection 521

appropriate for home or small office settings where more familiar faces appear reg-
ularly. Accuracy comparisons between the filters is presented for two different envi-
ronmental settings with two different data sets.
3. To make our algorithm more efficient in comparison to other algorithms in terms
of space and time, we have successfully utilized a compression algorithm with the
system.

1.3 Outline of Paper

The rest of the paper is organized as follows. Section 2 describes our proposed frame-
work of the traditional techniques. Section 3 presents our experiments and compares
performance within each traditional technique. We conclude in Sect. 4 with a layout of
our future work.

2 Proposed Framework
2.1 Face Detector
We used the fundamental Viola Jones face detection algorithm [2] to extracting faces
from an image. It uses Haar-like features which consists of dark regions and light
regions. To apply these Haar-like features in face images, it first computes an inte-
gral image which is an intermediate representation of an image which basically adds
side-by-side pixels. So, when Haar-like features are applied and a darker region needs
to be subtracted from a lighter region, it can easily be done using integral image.
The AdaBoost algorithm is then used to find a strong classifier which is a linear
combination of weak classifiers. To find the weak classifier, at each iteration, the algo-
rithm finds the error rate of each feature and selects the feature with the lowest error
rate for that iteration.
A cascade classifier is then used which is made up of multi-stage classifiers. Each
stage is a strong classifier which is produced by AdaBoost algorithm. If an input image
receives a negative result at a specific stage of the classifier, then it is immediately
discarded; the image needs to pass through all the stages with positive result to be
considered a face, and hence proved to be very quick in detecting a face from an image.
Although, it is not effective to detect masked faces since it cannot find the difference
in pixels that is required by Haar-like features. So, instead we extract the whole image
for masked faces to get a larger variation within the images.

2.2 Feature Extraction

2.2.1 Wavelet Decomposition
After extracting images from the face detector, the images are converted to grayscale
and standard two-dimensional Haar wavelet decompositions [3] are performed. The
reason for doing so is to work with the images in the frequency domain rather than
space domain. This is because in the space domain, with minor camera movements,
differences between specific pixels in different images may arise, which is not desirable.
Moreover, it is better to consider and compare each image as a whole. The wavelet
decomposition represents an image in a way that reflects variation in neighboring pixel
522 A. Alvi et al.

intensities, and to perform image compression. This makes it suitable to be fed into
algorithms and to look for patterns between higher order statistics of the pixels.

2.2.2 Gabor Filter

Gabor filters have been previously used in face recognition [4]. It also enhances step
edges in an image but does not filter out information such as illumination and hue from
the original image.

2.2.3 Canny Edge Detector

An elegant way of eliminating noise is provided by the Canny edge detector [18]. The
Canny edge detector filters out information such as illumination and hue from the orig-
inal image, while preserving the important structural properties [5]. This results in an
“edge” image where the step edges are enhanced.

2.2.4 Laplacian Edge Detector

The Laplacian edge detector is based on the Laplace operator [6]. It is similar to the
Canny edge detector but is more sensitive to noise within the image. The step edges
enhancement in Laplacian is also less compared to Canny step edges.

2.2.5 Sobel Edge Detector

The Sobel edge detector uses the Sobel operator, which is a gradient based method on
the first order derivatives [7]. Since Sobel uses two kernels, one on the horizontal axis
and one on vertical axis, so better step edge enhancement is obtained than Canny or
Laplacian edge detectors.

2.3 Kernel-Based Online Anomaly Detection Algorithm

The Kernel-based Online Anomaly Detection (KOAD) algorithm is a learning algo-
rithm based on the so-called “kernel trick” which uses a kernel function to map the
input image onto a feature space of much higher dimension [19]. The idea is to use
a suitable kernel function that computes the inner product of the input vectors in the
feature space without explicit knowledge of the features vectors themselves. The kernel
function used is:
k(xi , x j ) = φ (xi ), φ (x j ) (1)
where xi , x j denote the input vectors and φ represents mapping onto feature space.
Considering a set of multivariate measurements {xt }t= T . When these measurements
1
are associated with mapping φ , the feature vectors that depict normal behaviour are
expected to cluster. It is then able to explain the region of normality using a relatively
small dictionary of approximately linearly independent elements {φ (x˜j )}mj=1 [20].
Comparing this with approximation threshold ν , the equation of projection error is:
m
δt = mina || ∑ a j φ (x˜j ) − φ (xt ) ||2 < ν (2)
j=1

where a = {a j }mj=1 is the optimal coefficient vector.

 Study of Traditional Techniques for Unsupervised Autonomous Intruder Detection 523

Fig. 1. Sample images from data set 1.

Fig. 2. Sample images from data set 2.

This projection error is calculated for each timestep in KOAD. This error measure
δt is then compared with two thresholds ν1 and ν2 , where ν1 < ν2 , and if δt < ν1 ,
KOAD infers that xt is sufficiently linearly dependent on the dictionary and represents
normal behaviour, otherwise if δt > ν2 , it concludes that xt is far away from the realm
of normality and immediately raise a “Red1” alarm to immediately signal an anomaly.
In the case if ν1 < δt < ν 2, KOAD infers that xt is sufficiently linearly independent
from the dictionary to be considered an unusual event. It may indeed be an anomaly, or
it may represent an expansion or migration of the space of normality itself. So, KOAD
raises an “Orange” alarm, keeps track of the contribution of the relevant input vector xt
in explaining subsequent arrivals for l timesteps, and then takes a firm decision on it.

3 Experiments
3.1 Data Sets
Our experiments were performed on two data different sets. The first data set has been
created by us, where we have collected 100 different people images from the internet.
Within this data set, five intruder images were inserted at random time intervals. Sample
images from Data Set 1 are presented in Fig. 1. For the second data set, we use a Georgia
Institute of Technology face database [21]. This data set contains 750 images of 50
different people with each person having 15 images of different pose, expression and
illumination. From this data set [21], we have taken 190 images and included 10 masked
images at random time steps. Figure 2 represents the sample images of this data set.
From these two data sets, we have considered them as representative of two different
environmental scenarios. Data Set 1 is relevant for ATM surveillance and Data Set 2 is
appropriate for small office or home surveillance.
524 A. Alvi et al.

Fig. 3. Gabor edge images corresponding to sample raw images from data set 1.

Fig. 4. Gabor edge images corresponding to sample raw images from data set 2.

Fig. 5. Canny edge images corresponding to sample raw images from data set 1.

Fig. 6. Canny edge images corresponding to sample raw images from data set 2.

Face detector was not required for Data Set 2 since the images were cropped to 150
× 150 pixels with background also removed. The images from both these data sets were
pre-processed using the filters mentioned earlier. Each filter represented the same input
image in different ways. The results obtained are present using the figures that follow.
Study of Traditional Techniques for Unsupervised Autonomous Intruder Detection 525

Fig. 7. Laplacian edge images corresponding to sample raw images from data set 1.

Fig. 8. Laplacian edge images corresponding to sample raw images from data set 2.

Fig. 9. Sobel edge images corresponding to sample raw images from data set 1.

Fig. 10. Sobel edge images corresponding to sample raw images from data set 2.
526 A. Alvi et al.

3.2 Results
Data Set 1 contained five intruders and Data Set 2 contained 10 intruders, inserted at
random time intervals. Figures 3, 4, 5, 6, 7, 8, 9, 10 shows the different feature vectors
for each technique. So, for each method, we ran KOAD over a range of values for the
thresholds ν1 and ν2 and found suitable thresholds that gave a high detection rate for
both data sets. In order to check the detection rate, we needed to compare them with the
time steps in which anomalies were identified in the data set. For Data Set 1, anomalous
time steps were: t = 27, 50, 66, 70, 80 and for Data Set 2, anomalous timesteps were:
t = 21, 32, 53, 74, 104, 116, 127, 148, 169, 185.

Fig. 11. Progression of the KOAD detection statistic δt with time when the Gabor Filter used
as the feature extractor. Each time step represents each image. KOAD efficiently detects both
correct and falsely claimed anomalies. The left sub-figure corresponds to Data Set 1 and the right
sub-figure corresponds to Data Set 2.

Table 1 summarizes the results found from each method.

Gabor and wavelet decomposition yield similar results for same threshold parame-
ters. Sobel also has similar threshold values and achieves a higher detection rate with
lowest false alarm rate for Dataset 2. As a result, Sobel with KOAD performs well if
the same images of different orientations exist within a data set. Laplacian and Canny
require further investigation for better results.
Figures 11, 12, 13, 14 present the progression of the KOAD detection statistic δt
versus timestep t for each filter and the values found in each of the two datasets. The
green dotted line in each figure denotes the lower threshold ν1 and the red dotted line
denotes the upper threshold ν2 .
In Fig. 11, KOAD detects four out of the five intruders in Data Set 1 and nine out of
the 10 intruders in Dataset 2. Although the false alarm rate increases slightly from Data
Set 1 to Data Set 2, high accuracy is nonetheless achieved.
In Fig. 12, KOAD also detects four out of the five intruders in Data Set 1 and nine
out of the 10 intruders in Data Set 2 for almost the same parameters for KOAD. False
alarm rate also rises slightly from Data Set 1 to Data Set 2 due to increase in images,
but high accuracy is still achieved.
Just like for the previous methods, KOAD in Fig. 13 also detects four out of the five
intruders in Data Set 1 and nine out of the 10 intruders in Data Set 2 for almost the same
parameters for KOAD. False alarm rate decreases significantly from Data Set 1 to Data
Set 2, showing that it performs much better in data sets that contain repeated images of
different orientations.
 Study of Traditional Techniques for Unsupervised Autonomous Intruder Detection 527

Table 1. Detection rate and false alarm rate of different methods for two data sets.

Method Threshold, Threshold, Data set 1: Data set 1: Data set 2: Data set 2:
for ν1 ν2 Detection False alarm Detection False alarm
feature rate rate rate rate
extraction
Gabor 0.3 0.6 80 12.632 90 14.737
Wavelet 0.3 0.6 80 21.053 90 36.842
decompo-
sition
Sobel 0.4 0.6 80 34.737 90 3.684
Laplacian 0.2 0.4 80 56.842 50 27.895
Canny 0.3 0.6 Invalid 100 Invalid 100

Fig. 12. Progression of the KOAD detection statistic δt with time when wavelet decomposition
used as the feature extractor. Each time step corresponds to each image. KOAD also efficiently
detects both correct and falsely claimed anomalies. Left sub-figure corresponds to Data Set 1
while right sub-figure corresponds to Data Set 2.

Fig. 13. Progression of the KOAD detection statistic δt with time with Sobel edge detector used
as the feature extractor. Each time step corresponds to each image. KOAD also efficiently detects
both correct and falsely claimed anomalies. Left sub-figure corresponds to Data Set 1 and right
sub-figure corresponds to Data Set 2.
528 A. Alvi et al.

In Fig. 14, KOAD achieves moderate performance. In Data Set 1, it detects four out
of the five intruder images, with high false alarms. In Data Set 2, it detected five out
the 10 intruders with a lower false alarm. Further research is needed to achieve high
accuracy with using the Laplacian edge detector.
The representation from the Canny edge detector, when applied to KOAD, could
not differentiate between normal face image and intruder images. Further research is
needed here. The results of using the Canny edge detector are thus not presented here
in the interest of space.
Since every filter gives a different feature representation of the same image, the
parameters of KOAD are changed for better detection rate with each filter. Optimum
settings were identified by trial and error. However, the KOAD algorithm parameters
were kept constant for each filter across the two different data sets, to conclude that the
results are not sensitive to parameter settings after an initial training period. Detailed
complexity and parameter sensitivity analysis is not presented here in the interest of
space.

Fig. 14. Progression of the KOAD detection statistic δt with time with the Laplacian edge detec-
tor used as the feature extractor. Each time step corresponds to each image. KOAD moderately
detects both correct and falsely claimed anomalies. Left sub-figure corresponds to Data Set 1 and
right sub-figure corresponds to Data Set 2.

4 Conclusion and Future Work

In this paper, we investigated traditional methods that can extract features from a face.
Five traditional techniques have been used to extract features within a face image and
we evaluated them by applying Kernel-based Online Anomaly Detection (KOAD) algo-
rithm. We created two frontal face image datasets with two different environment set-
tings. We also included a traditional face detector to this application to extract faces.
KOAD processed each face and signalled an anomaly if the face was covered, suspect-
ing it to be an intruder. Being a learning algorithm, the system became recursive and
easily signalled if any intruder was present. Its computational complexity is indepen-
dent of time. KOAD can achieve almost accurate detection with low false alarm rates
and have proven to be adaptive to changes in the environment. Furthermore, feature
extraction techniques like Gabor, Sobel and Haar wavelet decomposition also proved to
be effective methods for KOAD to use.
 Study of Traditional Techniques for Unsupervised Autonomous Intruder Detection 529

The future work will focus on analysing the parameter sensitivity for this applica-
tion. We would like to compare this work with benchmark work performed by other
people similar to this application. We would also like to search for different image
enhancement techniques that can be used in Laplacian and Canny edge detectors to
achieve better detection performance.

References
1. Valera, M., Velastin, S.: A review of the state-of-the-art in distributed surveillance systems.
In: Velastin, S., Remagnino, P. (eds.) Intell. Distrib. Video Surveill. Syst., pp. 1–30. Institu-
tion of Electrical Engineers, London, UK (2008)
2. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In:
Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pat-
tern Recognition. CVPR 2001, Kauai, HI, USA, pp. I (2001). https://doi.org/10.1109/CVPR.
2001.990517
3. Zhang, D.: Wavelet ransform. In: Fundamentals of Image Data Mining. Texts in Computer
Science. Springer, Cham (2019)
4. Bui, L., Tran, D., Huang, X., Chetty, G.: Face recognition based on Gabor features. In: 3rd
European Workshop on Visual Information Processing, Paris. France, pp. 264–268 (2011).
https://doi.org/10.1109/EuVIP.2011.6045542
5. Canny, J.: A computational approach to edge detection. In: IEEE Trans Pattern Anal. Mach.
Intell. PAMI-8(6), 679–698 (1986). https://doi.org/10.1109/TPAMI.1986.4767851.
6. Wang, X.: Laplacian operator-based edge detectors. IEEE Trans Pattern Analy. Mach. Intell.
29(5), 886–890 (2007). https://doi.org/10.1109/TPAMI.2007.1027
7. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter
using the Sobel operator. IEEE J Solid-State Circ. 23(2), 358–367 (1988). https://doi.org/10.
1109/4.996
8. Ahmed, T., Coates, M., Lakhina, A.: Multivariate online anomaly detection using kernel
recursive least squares. In: Proceedings IEEE INFOCOM, Anchorage, AK, USA (2007)
9. Ahmed, T., Oreshkin, B., Coates, M.: Machine learning approaches to network anomaly
detection. In: Proceedings of ACM/USENIX Workshop on Tackling Computer Systems
Problems with Machine Learning Techniques (SysML). Boston MA, USA (2007)
10. Datakalab. Company Webpage (2016).https://www.datakalab.com/
11. Chowdary, G.J., Punn, N.S., Sonbhadra, S.K., Agarwal, S.: Face mask detection using trans-
fer learning of inceptionV3. In: Bellatreche, L., Goyal, V., Fujita, H., Mondal, A., Reddy, P.K.
(eds) Big Data Analytics. BDA, : Lecture Notes in Computer Science, vol. 12581. Springer,
Cham (2020)
12. Nar, R., Singal, A., Kumar, P.: Abnormal activity detection for bank ATM surveillance. In:
2016 International Conference on Advances in Computing, Communications, and Informat-
ics (ICACCI), Jaipur, India, (2016), pp. 2042–2046.https://doi.org/10.1109/ICACCI.2016.
7732351.
13. Bouchrika, I., Carter, J.N., Nixon, M.S.: Towards automated visual surveillance using gait
for identity recognition and tracking across multiple non-intersecting cameras. Multimed.
Tools Appl. 75, 1201–1221 (2016). https://doi.org/10.1007/s11042-014-2364-9
14. Wang, J., Yuan, Y., Gang, Y.: Face Attention Network: An Effective Face Detector for the
Occluded Faces. ArXiv abs/1711.07246: n. pag (2017)
15. Anika, A., Karim, K.L.D., Muntaha, R., Shahrear, F., Ahmed, S., Ahmed, T.: Multi image
retrieval for kernel-based automated intruder detection. In: IEEE Region 10 Symposium
(TENSYMP). Cochin, India, pp. 1–5 (2017)
530 A. Alvi et al.

16. Ahmed, T., Pathan, A.K., Ahmed, S: Adaptive algorithms for automated intruder detection in
surveillance networks. In: 2014 International Conference on Advances in Computing, Com-
munications and Informatics (ICACCI), New Delhi, India, pp. 2775–2780 (2014). https://
doi.org/10.1109/ICACCI.2014.6968617
17. Ahmed, T., Ahmed, S., Ahmed, S., Motiwala, M.: Real-time intruder detection in surveil-
lance networks using adaptive kernel methods. In: 2010 IEEE International Conference on
Communications, Cape Town, South Africa, pp. 1–5 (2010)
18. Au, C., Skaff, S., Clark, J.: Anomaly detection for video surveillance applications. In: Pro-
ceedings of IEEE ICPR, Hong Kong, China (2006)
19. Scholkopf, B., Smola, A.: Learning with Kernels. MIT Press, Cambridge, MA, USA (2001)
20. Engel, Y., Mannor, S., Meir, R.: The kernel recursive least squares algorithm. IEEE Trans.
Sign. Proc. 52(8), 2275–2285 (2004)
21. Georgia Tech Face Database. ftp://ftp.ee.gatech. edu/pub/users/hayes/facedb/
 A Novel Dynamic Authentication Method
for Sybil Attack Detection and Prevention
in Wireless Sensor Networks

Wadii Jlassi1(B) , Rim Haddad2 , Ridha Bouallegue3 , and Raed Shubair4

1 Innov’COM Lab/Sup’Com, National Engineering School of Tunis, University Tunis El
Manar, Tunis, Tunisia
2 Laval University, Quebec, Canada
[email protected]
3 Innov’COM Lab/Sup’Com, University of Carthage, Tunis, Tunisia
[email protected]
4 New York University, Abu Dhabi, UAE

[email protected]

Abstract. One of the most concerns in wireless sensor networks is security. Wire-
less sensor networks are vulnerable to attacks in every layer of the network. They
are extremely necessitous for securing network protection. Sybil attack is one of
those attacks in which a legal node is converted into a Sybil one which is a replica
node with a different personality using a similar ID. This kind of identity theft
attack can be classified as Sybil attack. This attack mainly influences routing pro-
tocols and operations such as voting, data aggregation and reputation evaluation.
This paper proposes a dynamic and accurate method for Sybil attack detection
and prevention. The proposed method generates a routing table that holds infor-
mation about deployed nodes. The base station verifies the keys of nodes that will
communicate and transmit data between them. Data transmission occurs between
nodes once they get the signal from the base station. The message authentication
method will progress data transmission in the network and enhance the throughput
compared to RPC.

Keywords: Wireless sensor networks · Sybil attack · Random password

comparison · Message authentication method

1 Introduction
Wireless Sensor Networks (WSNs) consist of small sensor nodes co-working to monitor
and obtain information about the area of interest [1]. They are termed ad hoc networks
as they need no infrastructure. Therefore, WSNs are used in many areas such as environ-
mental conservation, domestic applications, Military surveillances, and so on. The basic
task of sensor networks is to sense the events, gather efficient data and send it to their
requested destination [2]. The parameters of the WSN such as the sensing data, total
nodes, the life span of sensors, the energy consumption, geographical location of the
sensor placement, the environment, and the context are critical in the design of WSN.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 531–541, 2021.
https://doi.org/10.1007/978-3-030-75100-5_46
532 W. Jlassi et al.

Clustering is used to divide the network into several sections called clusters with the
objective to prolong the lifetime of the wireless sensor network. Each cluster has one
node which is called cluster head. This node collects data from the other member nodes
of the cluster and then sends it to base station. Now the main focus area of research in
clustered sensor networks is energy and security. Security of such networks is a challenge
especially when they are employed in critical applications [3]. The important goal of
security implementation is to reduce energy consumption and data security.
Wireless sensor networks are susceptible to attack starting from the physical layer
and going all the way up the stack to the application layer. Based on the WSN layers,
various types of attacks are classified as follows. Malicious node, Desynchronization
belongs to the Application layer. Desynchronization, Flooding belongs to Transport
Layer. Sybil, Selective forwarding, Sinkhole, Wormhole belongs to Network layer. Colli-
sion, Exhaustion belongs to the Datalink layer. Jamming, Interference, Node Replication
Attack belongs to the physical layer.
Sybil attack is a type of security threat. A threat is a set of circumstances that has the
potential to cause loss or harm [4]. It can be easily launched in WSN due to its nature
of unstructured and distributed topology and wireless communication. It is one of the
primary attacks that would facilitate the onset of many different attacks in the network.
The Sybil attacker also called the adversary captures a legal node or inserts an illegal
node in the network. This malicious node forges fake identifiers or duplicates existing
nodes identifiers in different areas of the network. This attack decreases the network
lifetime, creates suspicion in the genuineness of the data, reduces the effectiveness of
fault-tolerant schemes and poses a threat to geographic routing protocols. Also, malicious
node attracts a heavy and considerably disrupts routing protocols. It disturbs operations
such as data aggregation, voting and reputation evaluation. A malicious node which
enters the network with multiple IDs is shown in Fig. 1.

Fig. 1. Wireless sensor network with Sybil node

Various algorithms are proposed to detect and confront this attack in WSN, used dif-
ferent techniques such as radio resource testing technique [5], Neighboring Information
[6], node location based on RSSI [7], random key pre-distribution, Random password
 A Novel Dynamic Authentication Method for Sybil Attack 533

comparison [8], RSSI of cluster head, Leach routing protocol [9], Neighboring Informa-
tion and broadcast two-hop packets [10], observer nodes and neighboring information
[11]. These techniques require heavy computational complexity which leads to reduc-
tion in energy, since energy consumption in WSN is limited, and have a limitation in
their detection range.
This paper concentrates on the security in wireless sensor networks and proposes
a dynamic and accurate algorithm to detect and prevent Sybil attack. The rest of the
paper is organized as follows: Sect. 2 reviews related works (existing works). Section 3
describes the proposed algorithms. Results of simulations and discussions are presented
in Sect. 4. Finally, Sect. 5 concludes the paper and reference papers are included.

2 Related Works

Wireless sensor nodes perform the operation of transmitting the data from the source to
the destination which should be made in an efficient way so that the data transmission
between the sender and the receiver will be in an effective manner. The sensor nodes
have limitations in storage, power, latency, constraint bandwidth and reduced corporal
size [12]. Since WSN nodes have limitations, safekeeping is the most important role
needed to detect and prevent malicious activities in the network [13].
Sybil attack was introduced for the first time by Douceur [14] in peer-to-peer net-
works. It is a matter of critical importance and consternation in network security leading
to many fake identities that can cause disruption in the network [15]. The detection
techniques of sybil attack are classified as localized, distributed and centralized. In other
words, they could be classified based on various characteristics such as key distribution,
broadcast, multicast, location, random or group deployment.
In [16], a solution based on a centralized base station (BS) has been proposed. Each
node in WSN sends the list of neighbors and their locations to BS. The same node ID in
two lists with inconsistent locations results in a clone detection. BS has the mission to
remove the clones.
In [17], a position-based Sybil attack detection approach has been proposed verifying
the physical position of each node. Sybil nodes can be detected using this approach
because they will appear to be at exactly the same position as the malicious node that
generates them.
An RSSI-based locating scheme has been introduced in [18]; it uses proportion of
RSSIs from different receivers in order to calculate nodes locations in the network. This
is sufficient to detect Sybil nodes since all of them are positioned in nearby locations.
A protocol has been proposed in [19] which uses identifier-based encryption. This is
used to avoid Sybil attacks by not allowing malicious nodes to acquire various identifiers.
In [20], an algorithm has been proposed to find Sybil nodes in WSNs. It consists of
three phases. In the first phase, they send ID and power value to the head node, the head
node checks for nodes with power value under the threshold value. In the second phase,
if the nodes are very close then the distance between the receiver and the sender is zero
so the node suffers from Sybil attack. In the third phase, the routing procedure in the
cluster is verified if there was a hop between the Sybil identities.
534 W. Jlassi et al.

Like [21], has provided a method to avoid Sybil attacks. The information on routes
is collected by an intelligence algorithm during network activity. The Sybil nodes are
detected according to their energy changes when the network is active. Another algorithm
[22] has been provided to avoid Sybil attacks in wireless sensor networks, based on
clients’ puzzles and learning Automata (LA). A Sybil attack detection based on a trust-
based model [23] has been provided which incorporates the concept of fuzzy expert
system and neural network.
A lightweight trust system [24] has been proposed which uses energy as a metric
parameter for a hierarchical WSN for dealing with Sybil attack in WSNs. This system
can reduce the communication overhead in the network. The evaluation of this system
views efficiency and scalability for detecting Sybil attacks in terms of true and false
positive detection in a heterogeneous WSN.
A mobile agent-based Sybil attack detection method has been proposed in [25]
which uses three parameters: random key pre-distribution, random password and thresh-
old value. The network performance evaluation before and after the detection of Sybil
nodes based on: the throughput and the packet delivery ratio. Another algorithm called
Sybil attack Detection Algorithm (SDA) [26] based on mobile agent has been proposed
to detect and prevent the sybil attack. It uses three parameters namely random key
pre-distribution, random password generation and threshold value. The suspect node is
informed to the BS and the BS alerts the network nodes.
Like [27], has provided a lightweight, dynamic algorithm for detecting Sybil nodes
in mobile wireless sensor networks. This algorithm uses watchdog nodes to monitor
the traffic passively and assign bit wise tags to mobile nodes based on their movement
behaviors, and then detects Sybil nodes according to the labels and node id, during
detection phase.
Another algorithm [11] has been proposed which uses some observer nodes to detect
Sybil attacks in mobile wireless sensor networks. Two phases proposed in this algorithm:
monitoring phase which observer nodes record the number of meeting occurrences of
other nodes in a vector called history, for R time periods. Then, detection phase in
which observer nodes cooperate and identify Sybil nodes based on the content of history
vectors.
A novel Sybil attack detection protocol (NoSad) [28] is proposed. This protocol is
a localized method using intra-cluster communication and RSSI value to identify and
isolate the Sybil attack in WSN.
In general, the main disadvantages of the existing algorithms for detection Sybil
attack are: high rate of false detection, high cost (need for nodes locations), no scalability
and complexity of detection algorithms.

3 Proposed Work
The main objective of this paper is to propose a method for detecting and preventing
Sybil attacks in wireless sensor networks. The network G is composed of N number of
nodes deployed randomly and a base station BS.
The proposed algorithm generates a routing table (rtable RPC) which holds the
information of each node’s id, time and password. The intermediate node in the route
 A Novel Dynamic Authentication Method for Sybil Attack 535

is identified between source and destination. The intermediate node’s information is

compared with the RPC database during the communication in the middle of nodes.
Based on the comparison results, it decides whether it is a Sybil or a normal node. A
Sybil node will not be able to submit the dynamic password which is allowed to all the
nodes in the network. The system model goes through the following steps:

• Step 1: The network G is composed of N number of nodes deployed randomly.

• Step 2: Generate table ti for ni, ti = {c1, c2, c3} where ci = NID, time, pwd.
• Step 3: Discover the route among the elected nodes S, D are source and destintion
nodes.
• Step 4: Check the route and verify the Sybil activity then pass the data.

Random password creates a new password every few second for each node and sends
it to all nodes in the network. After communication between source and destination, the
destination node’s id, time delay and the random password corresponding to the time
delay will be compared with RPC database. If the information matches, the node is
considered to be a normal node and the source node will send data otherwise the node is
considered as a Sybil node. The nearest neighbor is found by calculating the minimum
distance from the source node as given in (1).

dij = (xj − xi)2 + (yj − yi)2 ∀i, j denotes node index (1)

The Random password comparison method algorithm is given below:

• Network G, which has n number of nodes, G = {n1, n2, n3….ni, nn}

• Table T = {t1, t2, t3….ti,….tn} where ti is table for ith node.
• Table T is commonly available
• in BS where it can be authenticate all nodes in G.
• c1, c2, c3 are the three columns in the table ti where c1 is the node id, c2 is the time,
c3 is the password generated for node
• Elect any two nodes S, D as Source and Destination Node in G.
• Node D sends a data request message to node S.
• A route discovered between S and D using the sub-procedure DISROUT(S, D) is
obtained.
• Now S sends data in the same route verifying for Sybil activity by CHKROUT (S, D)
in all the nodes found in the route. Then node S sends a data to D.

The algorithm gives the values c1, c2 and c3 for each node and they are stored in the
rtable. After discovering the route between source and destination, every nearest node
should submit their information (id, time password) which is compared with the rtable
values. If the information matches, the node is chosen as the nearest neighbor and added
to the route. Otherwise, it means that the current node is discarded and another node is
chosen for the same process. This process stopped if the destination node is discovered.
RPC generates the route by adding the genuine node in its path from source to des-
tination node using several sub procedures DISROUT(S, D) and CHKROUT(S, D).The
536 W. Jlassi et al.

sub procedures detect the route from source to destination before transmitting the data.
The Sybil node is detected during the data transmission.
Figure 2 represents respectively sub procedures DISROUT (S, D) and CHKROUT
(S, D).

Fig. 2. A flowchart of sub procedures DISROUT AND CHKROUT

Even though the detection ratio is high, this technique has its own drawback such as
high energy consumption on the nodes near the BS or sink due to forwarding packets.
Application of RPC needs a lot of time for road research and distance calculation.
We propose to use RPC method with message authentication method to improve the
data transmission between nodes. In the network G, a node Ni sends a request messages
to node Nj with its key, as msg(Ni).Node Nj submits its key message with msg(Nj),and
later, both keys are verified by the base station and an ok signal produced for sending
the data. Data transmission occurs between Ni and Nj once they get the signal from the
base station.
 A Novel Dynamic Authentication Method for Sybil Attack 537

The message authentication and passing method are given in Fig. 3.

Fig. 3. A flowchart of the message authentication method

4 Simulation Results
In this section, to prove the performance of our proposed method, NS-2 software is used
for simulation. Table 1 shows the initial parameters of Wireless Sensor Network for
simulating 100 nodes with a network size of 100 * 100.
Figure 4 presents the network deployment with 100 nodes and one Sybil node detec-
tion. All nodes are constructed under a single BS. The number of created nodes is 100
and 0 is the base station. We analyze the performance by calculating the average delay
of data packet transfer and the throughput.
As shown in Fig. 5, our proposed method’s average delay of the data packet transfer
of the network is very less than RPC. The Average delay extends when the number of
nodes increase respectively for the different methods. For message authentication, the
delay is less compared to RPC due to the time of detecting of Sybil node by the sub
procedures DISROUT and CHECKROUT.
Figure 6 shows a Comparison of Throughput between the existing method RPC with
the proposed method. In the case of message authentication method, the throughput will
538 W. Jlassi et al.

Table 1. Simulation parameters

Parameter Level
Area 100 m × 100 m
Nodes 100
Packet size 50 bytes
Transmission protocol UDP
Simulation time 100 s
Propagation Two ray ground
Placement Random
Application Traffic CBR
Routing Protocol AODV
Initial energy 100 J

Fig. 4. Network deployment

reach 98% while in RPC, the throughput will be 78%. The proposed method improves
security and increases the communication in the wireless sensor networks compared to
RPC.
 A Novel Dynamic Authentication Method for Sybil Attack 539

RPC Proposed method

4.5
4
Average delay (%)

3.5
3
2.5
2
1.5
1
0.5
0
10 20 30 40 50 60 70 80 90 100
Number of nodes

Fig. 5. Comparison of average delay of the data packet transfer between RPC with proposed
method.

RPC Proposed method

Packet delivred successfully (%)

100
90
80
70
60
50
40
30
20
10
0
30 60 90 120 150 180 210 240 270 300
Time(ms)

Fig. 6. Comparaison of Throughput between existing method RPC with proposed method.
540 W. Jlassi et al.

5 Conclusion
In this paper, we present a new method for detecting and preventing Sybil attacks in wire-
less sensor networks. Simulation results proved the efficiency of the proposed method
compared to the existing method RPC. The proposed method presents a method to
validate the route by the authentication of each node. Besides, we compared the data
packet transfer using our method and RPC. For message authentication, the delay is less
compared to RPC due to the time of detecting of Sybil node by the sub procedures DIS-
ROUT and CHECKROUT. We made also another comparison of Throughput between
RPC method and the proposed method. We noticed that our method improves security
and increases communication in the wireless sensor networks. As future work, we will
study the use of a novel method for detecting Sybil attacks with a guarantee of energy
consumption of the network.

Acknowledgments. We are grateful for the support of the Department of Electrical and Computer
Engineering at the New York University of Abu Dhabi (NYU).

References
1. Rawat, P., Singh, K.D., Chaouchi, H., Bonnin, J.M.: Wireless sensor networks: a survey on
recent developments and potential synergies. J. Supercomput. 68, 1–48 (2014)
2. Manjunatha, T.N, Sushma, M.D., Shivakumar, K.M.: Security concepts and sybil attack
detection in wireless sensor networks (2016)
3. Akyildiz, I.F., Kasimoglu, I.H.: Wireless sensor and actor networks: research challenges. Ad
Hoc Networks (2014)
4. Dhamodharan, U.S.R.K., Vayanaperumal, R.: Detecting and preventing sybil attacks in
wireless sensor networks using message authentication and passing method (2015)
5. Newsome, J., Shi, E., Song, D., Perrig, A.: Te sybil attack in sensor networks: analysis and
defenses. In: International Symposium on Information Processing in Sensor Networks (2004)
6. Ssu, K.F., Wang, W.T., Chang, W.C.: Detecting sybil attacks in wireless sensor networks using
neighboring information. Proc. Comput. Networks 53, 3042–3056 (2009)
7. Demirbas, M., Song, Y.: An RSSI-based scheme for Sybil attack detection in wireless sensor
networks. In: IEEE Computer Society International Symposium on World of Wireless, Mobile
and Multimedia Networks (2006)
8. Amuthavalli, R., Bhuvaneswaran, R.S.: Detection and prevention of sybil attack in wireless
sensor network employing random password comparison method. J. Theoricaland Appl. Inf.
Technol. 67(1), 236–246 (2014)
9. Jangra, A., Priyanka, S.: Securing LEACH protocol from Sybil attack using jakes channel
scheme (JCS). In: International Conferences on Advances in ICT for Emerging Regions
(2011)
10. Rafeh, R., Khodadadi, M.: Detecting Sybil nodes in wireless sensor networks using two-hop
messages. Indian J. Sci. Technol. 7(9), 1359–1368 (2014)
11. Jamshidi, M., Ranjbari, M., Esnaashari, M., Qader, N.N., Meybodi, M.R.: Sybil node detection
in mobile wireless sensor networks using observer nodes. Int. J. Inf. Visualization 2(3),
159–165 (2018)
12. Birami, K.S.: Sybil attack in Wireless Sensor Network (2016)
 A Novel Dynamic Authentication Method for Sybil Attack 541

13. Sharmila, S., Umamaheswari, G.: Detection of Sybil attack in Mobile Wireless Sensor
networks. Int. J. Eng. Sci. Adv. Technol. 2(2), 256–262 (2012)
14. Douceur, J.R.: The sybil attack. In: International Workshop on Peer-to-Peer Systems, pp. 251–
260. Springer, Heidelberg (2002)
15. Sharmila, S., Umamaheswari, G.: Detection of sybil attack in mobile wireless sensor networks.
Int. J. Eng. Sci. Adv. Technol., 256–262 (2017)
16. Eschenauer, L., Gligor, V.D.: A key-management scheme for distributed sensor networks.
In: Proceedings of the 9th ACM Conference on Computer and Communications Security,
pp. 41–47 (2002)
17. Manjunatha, T.N., Sushma, M.D., Shivakumar, K.M.: Sybil Attack Detection Through on
Demand Distance Vector Based Algorithm in Wireless Sensor Networks (2015)
18. Triki, B., Rekhis, S., Boudriga, N.: An RFID based system for the detection of sybil attack
in military wireless sensor networks. In: Computer Applications and Information Systems
(wccais), World Congress on, pp. 1–2. IEEE (2014)
19. Sujatha, V., Anita, E.M.: Detection of sybil attack in wireless sensor network (2015)
20. Sharmila, S., Umamaheswari, G.: Detection of Sybil Attack in Mobile Wireless Sensor
Networks (2016)
21. Muraleedharan, R., Ye, X., Osadciw, L.A.: Prediction of Sybil Attack on WSN using Bayesian
Network and Swarm Intelligence. Wireless Sensing and Processing (2008)
22. Jamshidi, M., Esnaashari, M., Meybodi, M.R.: An algorithm for defending sybil attacks
based on client puzzles and learning automata for wireless sensor networks. In: 18th National
Conference of Computer Society of Iran (2017)
23. Sarigiannidis, P., Karapistoli, E., Economides, A.A.: Detecting Sybil attacks in wireless sen-
sor networks using UWB ranging-based information. Expert Systems with Applications,
pp. 7560–7572 (2015)
24. de Sales, T.B.M., Perkusich, A., de Sales, L.M., de Almeida, H.O., Soares, G., de Sales, M.:
ASAP-V: a privacy-preserving authentication and sybil detection protocol for VANETs. Inf.
Sci. 23(4), 208–224 (2016)
25. Pecori, R.: S-Kademlia: A trust and reputation method to mitigate a Sybil attack in Kademlia.
Comput. Netw. 94, 205–218 (2016)
26. Prasad, K., Mallick, C.: Mobile agent based sybil attack detection algorithm for wireless sensor
network. In: International Conference on Emergent Trends in Computing and Communication
(2015)
27. Jamshidi, M., Zangeneh, E., Esnaashari, M., Meybodi, M.R.: A lightweight algorithm for
detecting mobile Sybil nodes in mobile wireless sensor networks. Comput. Electr. Eng. 64,
220–232 (2017)
28. Angappan, A., Saravanabava, T.P.: Novel Sybil attack detection using RSSI and neighbour
information to ensure secure communication in WSN. J. Ambient Intell. Humanized Comput.
(2020)
 An Effective Location-Based Alert Messages
Dissemination Scheme for Software Defined
Vehicular Networks

Raoua Chakroun(B) , Slim Abdellatif, and Thierry Villemur

LAAS-CNRS, 7 Avenue du Colonel Roche, Toulouse, France

{rchakrou,slim,villemur}@laas.fr

Abstract. Software Defined Networking (SDN) is considered as a key paradigm

for future vehicular networks and its adoption may pave the way to novel
approaches to support emerging and legacy ITS services, such as Alert Message
(AM) dissemination. Historically, to meet the low dissemination delay require-
ments and full vehicle coverage, Vehicular Ad-hoc NeTworks (VANETs) were
considered as the underlying assumption of most proposed AM dissemination
schemes. With SDN, vehicular-to-infrastructure (V2I) links are a viable alterna-
tive that adds to Vehicle-to-vehicle (V2V) links. Moreover, a global view of the
network at a centralized controller complemented with information on vehicles
and their route helps to adjust the dissemination procedure to improve its perfor-
mance and efficiency. In this paper, we propose a location-based AMs dissemi-
nation scheme that combines V2I broadcasts with V2V rebroadcasts in order to
provide a low delivery delay, very limited collisions, high information coverage
with insignificant signaling and network overhead. The originality of our scheme
stands in the selection process of V2V rebroadcasts which is based on vehicles’
location with respect to predefined rebroadcast points selected by the controller.
Our proposal is evaluated and compared to legacy techniques. It shows significant
improvements in delivery delays and network resource utilization.

1 Introduction
As an obvious evolution of legacy ITS (Intelligent Transportation Systems), Coopera-
tive ITS (C-ITS) have been pushed in the front of the stage with the idea of enabling
communications between vehicles (V2V: Vehicle-to-Vehicle), vehicles and infrastruc-
ture (V2I: Vehicle-to-Infrastructure), and more generally vehicles and all their sur-
roundings (V2X: Vehicles-to-everything). The rationale is that thanks to these new com-
munication alternatives, the perception of vehicles on their surrounding and their envi-
ronment is enriched, its accuracy is improved and the perception horizon is enlarged. As
a consequence, novel C-ITS services that further improve road safety, traffic efficiency,
comfort and convenience to drivers and passengers are envisioned, e.g. cooperative
maneuver, bird’s eye view, cooperative awareness (e.g. Emergency vehicle warning,
Alert for an accident), etc. [11, 15].
One of the challenges facing the development of C-ITS is the provision of network
connectivity services that meet the various Quality of Service (QoS) expectations of C-
ITS services. One direction that has adhesion from academia and Industry is to consider
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 542–552, 2021.
https://doi.org/10.1007/978-3-030-75100-5_47
 An Effective Location-Based Alert Messages Dissemination Scheme For SDVN 543

a hybrid multi-access vehicular network that combines, amongst: DSRC (Dedicated

Short Range Communication) which is foreseen for rapid dissemination and delivery of
critical safety messages, and LTE (Long Term Evolution) which is foreseen for heavier
and less stringent message exchanges [15, 20]. One promising approach towards this
hybridization is to apply the Software Defined Networking (SDN) paradigm as lever-
age to unify the network control of the above-cited technologies with lots of promising
opportunities and outcomes in terms of improved network resource usage and perfor-
mance [7, 9, 16]. For these reasons, this work considers SDN-based vehicular networks.
Point-to-multi-point communications and more specifically Geo-Broadcasting are
at the heart of many C-ITS services. Alert Message AM dissemination is such a service,
which upon an emergency or a risky situation (accident or vehicle breakdown), alert
messages are generated for some time and spread over a geographical area of interest as
fast as possible to allow other vehicles to react properly very quickly [3]. Historically,
message alert dissemination was carried out by a DSRC based VANET (Vehicular Ad-
hoc NeTworks) operating on a dedicated communication channel by inviting vehicles
to blindly rebroadcast the alerts. This method, known as simple flooding, causes a huge
amount of unnecessary re-transmissions and collisions (known as the broadcast storm
problem), thus wasting bandwidth, increasing dissemination delay, and lowering packet
delivery ratio [17]. Lots of proposals from the literature tried to mitigate the broadcast
storm problem by controlling the vehicles that are allowed to rebroadcast (using some
form of random selection or clustering techniques to group vehicles) or reducing the
contention. But, most were designed for VANETs and neither assume the presence of
V2I links nor take advantage of the benefits that a global view of the network at a
centralized controller can bring to the dissemination scheme.
In this paper, we propose an alert messages dissemination scheme that primarily
exploits V2I transmissions and complements with V2V transmissions in order to reach
vehicles located in areas that are not covered (white zones) or poorly covered (grey
zones) by transmissions from the infrastructure. The originality of our scheme stands in
the selection process of relay vehicles which is based on a local decision at each vehicle
based on its location with respect to predefined rebroadcast points that are computed by
the controller based on its knowledge of the radio performance and coverage as well as
road traffic conditions. In addition, our proposal induces marginal extra signaling and
network overhead.
The rest of the paper is organized as follows: Sect. 2 presents an overview of existing
work on AM dissemination. Section 3 describes the system model as well as the pro-
posed location-based dissemination algorithm. Section 4 evaluates and discusses our
simulation results. Finally, Sect. 5 concludes the paper.

2 Related Work
Alert message dissemination has been widely studied in a VANET/V2V context. All
aim at mitigating the effect of the ”broadcast storm” while ensuring high information
coverage. These works can be broadly classified as follows.
Lots of the proposed schemes try to control the flooding procedure of AMs either by
reducing the number of rebroadcasters or by reducing the contention between transmit-
ting vehicles. In the former case, some vehicles are selected to relay the AMs [2, 14, 19].
544 R. Chakroun et al.

This selection is based on vehicles’ characteristics, distance from the sender, local den-
sity, interests, transmit power, etc., and helps to limit re-transmissions and contention
at the cost of reduced information coverage. In the latter case, MAC level protocol
parameters are adjusted on a per vehicle basis in order to statistically assign different
back-off periods to vehicles [6]. By so doing, collisions are reduced with no impact on
information coverage. However, useless transmissions are not avoided.
Another class of proposals set aside the flooding logic and try to guide AMs dis-
semination by organizing vehicles in clusters (groups of vehicles) and defining how AM
are propagated between and within clusters [8, 10, 18]. Many criteria are used for clus-
ter formation as well as cluster-head selection, e.g. neighborhood, direction/destination,
relative velocity, etc. A last class of proposals is those based on routing protocols that
proactively or reactively compute point-to-multi-point routes to all known vehicles [4].
They clearly exhibit the most predictable information coverage and efficient resource
usage, but this comes at the cost of increased network overhead and complexity embed-
ded in vehicles.

3 Location-Based Alert Messages Dissemination

3.1 System Model

An SDN controller is assigned to each region, it manages all the RSUs that fall inside.
Each controller identifies whites and grey zones in its region and selects K rebroad-
cast zones where vehicles can rebroadcast an AM sent by an RSU in order to enhance
the overall coverage and reach all the vehicles located in grey or white zones. These
rebroadcast zones are defined by a rebroadcast point Pi (xi , yi ), i ∈ [1, K], with xi and yi
are the GPS coordinates of Pi , and a radius dmax in the order of a few meters from the
rebroadcast point. The contribution of this paper concerns the dissemination procedure,
as a consequence, a basic broadcast point placement is adopted as explained hereafter.
Real experiments in [20] show that the packet loss rate and delay increase when the
distance between vehicles and RSU is greater than a threshold Rth which also depends
on mobility, due to the wireless propagation channels. According to [15], an RSU will
be deployed every 2 km in Europe with a theoretical coverage defined by R = 850 m.

Fig. 1. Rebroadcast zones

 An Effective Location-Based Alert Messages Dissemination Scheme For SDVN 545

We assume that the controller is able to define a threshold distance Rth for each RSU
from which the quality degrades and the probability of reception decreases. From this
threshold radius (Fig. 1), it can build a regular polygon with K, K ∈ [5, 17] equal sides
of 2Rmax (±50 m) (where Rmax is the maximum radius coverage of a vehicle), to avoid

Fig. 2. Location-based alert messages dissemination

Algorithm 1: Alert message dissemination

Input:
AM : Alert Message
Pi , i ∈ [1, k], dmax : rebroadcast zones
V j : current node, Pos(V j ) : its position, Pos(V j ) : its direction
S : Sender
1 AM.Reception() /* AM received */
2 if S = RSU then
3 Pi ← F(Pos(V j ), dir(V j )) /* find the nearest rebroadcast zone */
4 if d(Pi ,V j ) ≤ dmax then
5 V j ← CR /* V j is a Candidate Relay */
6 T ← T1 /* Calculate time to wait T */
7 else if dmax < d(Pi ,V j ) ≤ Rmax then
8 V j ← CRB /* V j is a Candidate Relay Backup */
9 T ← T2

10 else
11 if (AlreadyReceive (AM) = True) and (V j = CR or V j = CRB) then
12 if d(Pi , S) < d(Pi ,V j ) then
13 Stop (t) and Discard()

14 else if ( AlreadyReceive (AM) = True ) and (V j = CRE) then

15 Stop (t) and Discard()
16 else if V j ∈
/ {RSU} and T T L = 0 then
17 V j ← CRE /* V j is a Candidate Relay Exceptional */
18 T ← T3
19 else
20 Discard ()

21 LaunchTimer (T) /* start timer */

546 R. Chakroun et al.

collisions between two vehicles in two adjacent rebroadcast zones, whose polygon ver-
tices represent the rebroadcast points. In this case, a rebroadcast zone i is represented
by the vertex Pi with a margin dmax of a few meters. If a rebroadcast zone does not serve
any road area, it will be omitted.

3.2 Dissemination Procedure

The proposed AM dissemination scheme is described in Algorithm 1. At handover, on
the re-association request with a new RSU, a vehicle V j receives the list of the rebroad-
cast zones that are under the coverage of the RSU. An AM is broadcasted with the
position of the sender and a predefined Time To Live T T L (used to control the rebroad-
casts of vehicles located in white zones as shown in Fig. 2). Vehicles that receive an AM
check their location: inside, around, or away from the rebroadcast zones. The vehicle
closest to the broadcast point is the best candidate for the rebroadcast. This is achieved
by assigning, on an AM reception, a personalized arbitration waiting time to each vehi-
cle. The closer a vehicle is from the rebroadcast point, the lower is its waiting time.
Let Pi = F(pos(V j ), dir(V j )) a function which takes the GPS position of vehicle
pos(V j ) = (x j , y j ) and its direction as parameters, and returns the closest rebroadcast
zone in its direction.
Referring to Algorithm 1, if a vehicle V j receives an AM, first, it checks the sender:
Step 1 [Lines 2–3] If the sender is an RSU, the node finds the closest rebroadcast point Pi
to its location, and it checks its position according to the rebroadcast zone, otherwise,
the message is received from a relay vehicle, go to step 4
Step 2 [Lines 4–6] If V j is in the rebroadcast zone, the node is declared as Candi-
date Relay CR and the vehicle in the rebroadcast zone closest to the rebroadcast
point has the highest priority to rebroadcast, so it has the minimum waiting time T1 .
Another factor that we consider is the distance between the RSU (sender) and V j
(receiver) to avoid the simultaneous transmissions of multiple vehicles sitting at the
same distance from the rebroadcast point. Hence, a relaying metric G1 that is used
to calculate the time to wait T1 , is given by :

d(V j , S) d(V j , Pi )
G1 = (1 − α ) + α (1 − ) (1)
R dmax
where : d(V j , S) indicates the distance between vehicle V j and RSU, R is maximum
transmission coverage of the RSU, d(V j , Pi ) is the distance between vehicle V j and
rebroadcast point Pi , dmax is the radius of a rebroadcast zone. α determines the influ-
ence of each component (distance from the rebroadcast point and distance from the
RSU) to compute G1 . Hence, if we consider a DSRC enabled vehicle, the waiting
time T1 is derived as follows:

T1 = (CWmin − G1CWmin )Ts (2)

where CWmin and Ts are respectively the minimum contention window and the time
slot duration of DSRC’s medium access technique, namely CSMA/CA (Carrier
Sense Multiple Access/Collision Avoidance) [1]. Go to step 7, otherwise, go to step 3
 An Effective Location-Based Alert Messages Dissemination Scheme For SDVN 547

Step 3 [Lines 7–9] If V j is in the radius coverage Rmax of rebroadcast point, then, the
node is declared as Candidate Relay Backup CRB and the relaying metric G2 is
given by :
d(V j , S) d(V j , Pi )
G2 = (1 − α ) + α (1 − ) (3)
R Rmax
Hence for a DSRC enabled vehicle, the waiting time T2 is as follows:

T2 = [CWmin + [(CWmax −CWmin ) − G2 (CWmax −CWmin )]]Ts (4)

where CWmax is the maximum contention window related to unicast transmissions.

T2 is set between CWmin and CWmax to avoid collisions with nearby CR vehicles. This
is particularly useful when the road traffic is low or sparse since, in case of no vehicle
in a broadcast zone, a close-by vehicle is given the opportunity to rebroadcast the
AM. Go to step 8, Otherwise discard the message
Step 4 [Lines 10–11] If V j receives the same message another time, and V j is a CR or
a CRB go to step 5, otherwise, go to step 6
Step 5 [Lines 12–13] If V j and the sender node are in the same rebroadcast zone, i.e.
the distance between the rebroadcast point and the sender is less than the distance
between the rebroadcast point and V j , stop the timer, the procedure exits, otherwise,
go to Step 9
This is to allow only the closest vehicle to the rebroadcast point to rebroadcast (if
both vehicles are in the same rebroadcast zone). It also avoids that a CR or CRB
related to a rebroadcast point to be prevented from rebroadcasting when it receives
a rebroadcast AM related to another rebroadcast point.
Step 6 [Lines 14–15] If V j receives the same message another time, and V j is a Can-
didate Relay Exceptional CRE, stop the timer, the procedure exits, otherwise, go to
Step 7
Step 7 [Lines 16–18] If V j is not attached to any RSU and the TTL is not expired, V j
is declared as CRE. In this case, the farthest vehicle from the sender, but still in its
transmission range, has the highest priority to rebroadcast, so the relaying metric G3
is given by:
d(V j , S)
G3 = (5)
Rmax
Hence the waiting time T3 is calculated as follows:

T3 = (CWmin − G3CWmin )Ts (6)

Otherwise, Discard message

Step 8 [Line 21] Start back-off Timer T
Step 9 Having waited for the time T , V j decrements T T L and forwards the message.

4 Performance Evaluation

The objective is to assess the performance of our method: (1) to reach all the vehicles
sitting in the area of interest in a very short time, (2) to avoid collisions, and (3) to
548 R. Chakroun et al.

effectively use network resources. Also, our proposal is compared to the flooding tech-
nique [5] and V2I broadcasts (with no V2V rebroadcasts). The following performance
metrics are used:
Collision Ratio: The collision ratio is the percentage of MAC collisions divided by the
number of sent packets.
Information Coverage: computed as the total number of vehicles that successfully
receive the AM divided by the number of all the vehicles.
Packet Delivery Ratio: is the ratio between the number of successfully received pack-
ets at the vehicles of the area of interest and all transmitted packets
Dissemination Delay: The dissemination delay is the total time required to deliver the
AM to all the vehicles in the area of interest.

4.1 Simulation Setup

In the experiments, an event based network simulator Netsim was used. In order to make
the most realistic simulations, SUMO (Simulation of Urban Mobility) bidirectionally
coupled with Netsim as described in [12] was used. Netsim includes implementations
of IEEE 1609.4 and IEEE 802.11p communication standards. It further includes Basic
Safety Message (BSM) handling and beaconing for cooperative awareness messages
(CAMs). In order to create realistic road traffic scenarios on a real map with SUMO, we
used OpenStreetMap as the geographic data source and selected an area of 2 km × 2 km
located in the city center of Toulouse, France. The traffic of vehicles was generated
randomly. We inserted vehicles into the network topology at a constant rate (30, 75,
100, 200, 300, 400, and 500 vehicles). The speed of vehicles was varied from 0 to
20 m/s, i.e. vehicles had different accelerations at different timestamps. The maximum
transmission range of each vehicle was set to Rmax = 250 m. In all simulated scenarios,
there was two RSUs separated by a distance of 2 km. Each one was responsible for the
transmission of alert messages to vehicles located in its transmission range. The other
simulation parameters are shown in Table 1.

4.2 Rebroadcast Zone Selection

For the performance evaluation of our dissemination procedure, we considered a basic
and empirical rebroadcast points/zones selection method described hereafter. The goal
is to compute a threshold distance Rth associated to the two considered RSUs and build
our polygon from which we derive the locations of rebroadcast points. To that end, each
RSU broadcasts a control message every 100 ms for 500 s. After each packet, we logged
the distance of each vehicle from the RSUs, the vehicles that received the message, and
the packet status (success or error). From the performance analysis, we observed that the
packet loss rate dramatically increases when the distance between the vehicle and RSU
is strictly greater than Rth  600 m. From this threshold, we built a regular octagon for
each RSU with 8 sides of 459 m(0.765Rth [13]) length and 8 vertices which represent
our rebroadcast points. We did not place all the rebroadcast points as the RSUs are in
the corners; we had a total of 9 rebroadcasts zones shown in Fig. 3.
 An Effective Location-Based Alert Messages Dissemination Scheme For SDVN 549

Fig. 3. Toulouse’s city scenario and rebroadcast zones selection

Table 1. Simulation configuration parameters

Parameter Value
Simulation time 500 s
Alert start 10 s
Packet generation rate 10 packets/s
AM packet size 1460 bytes
Propagation model Nakagami m = 3
Slot time 16 μ s
CWmin 15
CWmax 1023
dmax 16 m
α 0.8

4.3 Performance Analysis

Collision Ratio. Figure 4 shows the collision ratio as a function of vehicle density. Even
for high traffic density, we observe that, compared to flooding, our method significantly
reduces packet collisions (around 0.001%). This is due to the fact that only vehicles
around rebroadcast zones are allowed to relay the message with different waiting times
for each vehicle. The occurrence of some packet collisions is due to the behavior of
the CSMA/CA mechanism. If the channel is busy, relay nodes must wait for a random
back-off time before re-transmission. So when two vehicles are very close to each other,
550 R. Chakroun et al.

they have almost a similar waiting time, thus, they may rebroadcast at the same time if
the channel is busy.
Information Coverage. Figure 5 shows the percentage of information coverage versus
vehicle density. For low traffic density, the percentage of information coverage of the
proposed technique is close to other techniques, but it is still better. This is due to
the fact that, in such a scenario, the presence of vehicles around all rebroadcast zones
at each AM transmission is not guaranteed. Indeed, vehicles can join and leave the
network quickly according to the random mobility of SUMO without receiving any
message. When increasing traffic density, flooding is a bad option as shown in the figure.
Our method outperforms flooding and V2I approaches, and in contrast to these latter,
the information coverage strictly increases as more vehicles are covering the area of
interest. This means that rebroadcasts are taking place without causing collisions.
Dissemination Delay. Figure 6 shows the dissemination delay versus vehicle density.
The results show that the dissemination delay of our technique is significantly lower
than those of the other techniques, especially in high traffic density. Indeed, at each
message period, there is a relay vehicle in or around the rebroadcast zone to transfer
quickly the message to other vehicles outside the RSUs coverage. For an urban area,
the frequent collisions inherent to the flooding technique cause an additional delay, and
the mono-broadcast V2I technique takes a longer time to reach the maximum number
of vehicles in the area, as the vehicle must be in the coverage of an RSU to receive the
message. In low density, fewer vehicles are available in the network, so the presence
of vehicles around each rebroadcast zone from the first AM message is not guaranteed,
thus, our technique takes more time to reach all vehicles, e.g. for vehicles density of 30
vehicles, the total time required is 518.8 ms. Indeed, the average speed of vehicles is
between 7 and 13 m/s, the vehicle can only move approximately 1 m in 100 ms, there-
fore vehicles take more time to draw near the rebroadcast zones. It is worth noting that,
in this low density scenario, more than 75% of vehicles have the message delivered
within 100ms.

Fig. 4. Collision ratio vs. vehicle density Fig. 5. Information coverage vs. vehicle
density
 An Effective Location-Based Alert Messages Dissemination Scheme For SDVN 551

Fig. 6. Dissemination delay vs. vehicle den- Fig. 7. Packet delivery ratio vs. vehicle den-
sity sity

Packet Delivery Ratio. Figure 7 shows the packet delivery ratio versus vehicle density.
The results show that the packet delivery ratio of our proposal is slightly better than
the V2I technique. This is due to the fact that relay vehicles rebroadcast only reach
their neighbors. Because of the blind re-transmission which may cause a collision or a
useless transmission, flooding suffers from poor packet delivery ratios.

5 Conclusion and Future Works

In this paper, we propose an alert message dissemination scheme based on vehicle loca-
tions to provide high coverage with fast delivery. The main peculiarities of our solution
are: First, the combination of V2I transmissions, where RSUs broadcast alert messages
and V2V re-transmissions, where some selected relay vehicles rebroadcast the mes-
sages after a personalized waiting time, Second, the definition of broadcast zones by
the SDN controller where relay vehicles are allowed to rebroadcast the message to
reach all vehicles in the area of interest and notably those located in white or poorly
covered areas. In this paper, we have proposed a static and simple method to define
rebroadcast zones. The results obtained from our study prove the effectiveness of our
scheme by maintaining its high delivery ratio, avoiding packet collisions, and ensuring
rapid dissemination and high information coverage.
The main perspective of this work is to propose a machine-learning-based dynamic
placement method of rebroadcast zones, which adapts to traffic conditions, weather,
signal quality, etc.

Acknowledgements. This work is funded by Continental Digital Service France (CDSF) in the
framework of the eHorizon project.
552 R. Chakroun et al.

References
1. Ahizoune, A.: A protocol for broadcasting messages in vehicular networks. thesis manuscript
(2011)
2. Liu, B., Jia, D., Wang, J., Lu, K.: Cloud-assisted safety message dissemination in VANET-
cellular heterogeneous wireless network. IEEE Syst. J. 11, 128–139 (2017)
3. Salim, F.D., Loke, S.W., Rakotonirainy, A., Srinivasan, B., Krishnaswamy, S.: Collision pat-
tern modeling and real-time collision detection at road intersections. In: Intelligent Trans-
portation Systems Conference (ITSC), pp.161–166 (2007)
4. Aldabbagh, G., Rehan, M., Hasbullah, H., Rehan, W., Chughtai, O.: A driver safety informa-
tion broadcast protocol for VANET. Appl. Math. Inf. Sci. 10, 451–468 (2016)
5. Ciccarese, G., Blasi, M.D., Marra, P.: On the use of control packets for intelligent flooding in
VANETs. In: IEEE Wireless Communications and Networking Conference (WCNC 2009),
Budapest, pp. 1–6 (2009)
6. Virdaus, K., Kang, M., Shin, S., Lee, C.G., Pyun, J.: A counting-based broadcast model of
emergency message dissemination in VANETs. In: International Conference on Ubiquitous
and Future Networks (ICUFN), Milan, pp. 927–930 (2017)
7. Zheng, K., Kan, L., HOU, H., Lu, Q., Meng, N., Hanlin, L.: Soft-defined heterogeneous
vehicular network: Architecture and challenges. IEEE Network 30, 72–80 (2016)
8. Liu, L., Chen, C., Qiu, T., Zhang, M., Li, S., Zhou, B.: A data dissemination scheme based
on clustering and probabilistic broadcasting in VANETs. Veh. Commun. 13, 78–88 (2018)
9. Salahuddin, M.A., Al-Fuqaha, A., Guizani, M.: Software-defined networking for rsu clouds
in support of the internet of vehicles. IEEE Int. Things J. 2, 133–144 (2014)
10. Ali, M., Malik, A.W., Rahman, A.U., Iqbal, S., Hamayun, M.M.: Position-based emergency
message dissemination for Internet of vehicles. Distrib. Sens. Netw. 15, 1550147719861585
(2019)
11. Boban, M., Kousaridac, A., Manlakis, K., Eichinger, J., Xu, W.: Use Cases, Requirements,
and Design Considerations for 5G V2X, IEEE Vehicular Technology Magazine (2017)
12. Netsim simulator. https://www.tetcos.com/vanets.html
13. Polygon construction. https://debart.pagesperso-orange.fr/geoplan/polygone-regulier.mobi
le.html
14. Oliveira, R., Montez, C., Boukerche, A., Wangham, M.S.: Reliable data dissemination pro-
tocol for VANET traffic safety applications. Ad Hoc Netw. 63, 30–44 (2017)
15. Report of the working group: communication technologies for cooperative ITS. https://www.
ecologie.gouv.fr/sites/default/files/Rapport
16. Toufga, S., Abdellatif, S., Owezarski, P., Villemur, T., Relizani, D.: Effective prediction of
V2I link lifetime and vehicle’s next cell for software defined vehicular networks: a machine
learning approach. In: IEEE Vehicular Networking Conference (VNC), Los Angeles, CA,
United States (2019)
17. Tseng, Y.C., Ni, S.Y., Chen, Y.S., Sheu, J.P.: The broadcast storm problem in a mobile ad hoc
network. In: The Fifth Annual ACM/IEEE International Conference on Mobile Computing
and Networking, pp. 152–162 (1999)
18. Velmurugan, V., Manickan, J.M.L.: A efficient and reliable communication to reduce broad-
cast storms in VANET protocol. Cluster Computing (2019)
19. Wu, L., Nie, L., Fan, J., He, Y., Liu, Q., Wu, D.: An efficient multi-hop broadcast protocol
for emergency messages dissemination in VANETs. Chinese J. Electron. 26, 614-623 (2017)
20. Xu, Z., Li, X., Zhao, X., Zhang, M.H., Wang, Z.: DSRC versus 4G-LTE for connected vehicle
applications: a study on field experiments of vehicular communication performance. J. Adv.
Transp. 2017, (2017)
 Energy-Efficient Fuzzy Geocast Routing
Protocol for Opportunistic Networks

Khuram Khalid1 , Isaac Woungang1(B) , Sanjay Kumar Dhurandher2 ,

and Jagdeep Singh3
1
Department of Computer Science, Ryerson University, Toronto M5B 2K3, Canada
{khuram.khalid,iwoungan}@ryerson.ca
2
Department of Information Technology, Netaji Subhas University of Technology,
New Delhi, India
3
Department of Computer Science and Engineering, Sant Longowal Institute of
Engineering and Technology, Longowal, Punjab, India
[email protected]

Abstract. This paper proposes an energy-efficient version of our

recently proposed Fuzzy-based Check-and-Spray Geocast (FCSG) rout-
ing protocol (called EFCSG) for opportunistic networks (OppNets). The
proposed EFCSG protocol utilizes the Mamdani fuzzy econtroller along
with fuzzy attributes such as node’s residual energy, buffer space, and
movement to determine the relay nodes that are eligible to participate
in the routing process. Through simulations using real mobility traces,
EFCSG is shown to outperform the Energy Efficient Check-and-Spray
Geocasting (EECSG), the Fuzzy Geocast Routing Protocol for OppNets
(FCSG), and the Fuzzy Geocasting mechanism in OppNets (F-GSAF),
in terms of delivery ratio, average latency, number of hop count, and
overhead ratio, under varying Time-to-Live (TTL) and buffer size.

1 Introduction
OppNets are a type of challenged networks where high node mobility and inter-
mittent connectivity are frequent [1]. In this context, there is no guarantee of
an end-to-end path between the sender and receiver of a message like in tra-
ditional network to perform the message routing. Instead, a store-carry-and-
forward mechanism [2] is utilized, where every node receiving the data packets
is expected to first store the message in its local buffer (if space permits), then
eventually pass the message opportunistically to an encounter node in its neigh-
bourhood (assuming that this node is eligible to carry the data packets), and so,
until the message eventually gets to its intended destination.
Several routing protocols for OppNets implementing this paradigm have been
discussed in the literature [3], each of which have used its own opportunistic
route decision selection mechanism. These mechanisms involve the use of various
different parameters including the node’s attributes or a combination of them,
the hope being that the selected nodes become the best possible forwarders
of the message toward its destination. Examples of such mechanisms include
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 553–565, 2021.
https://doi.org/10.1007/978-3-030-75100-5_48
554 K. Khalid et al.

geocasting techniques [4], fuzzy logic-based techniques [5], a combination of these

techniques, and much more [3].
In a geocast technique [4], the message is first delivered from the source
node to the destination region (also called destination cast) instead of a selected
node, then it is disseminated within that destination cast using some kind of
control flooding mechanism. Doing so requires the consideration of temporal
and spatial domains, along with some constraints imposed in the availability
and accessibility of the nodes’ resources [4]. In such technique, every node that
resides within the destination cast is a potential receiver of the message if the
message destination address corresponds with its current location. A node that
does not reside to the destination cast cannot receive the message. It should
also be noted that frequently visited locations (or Point of Interests) are those
locations that are often visited during the control flooding process. From an
implementation standpoint, a cast is defined as a point (pair of two dimensional
coordinates) of polygon; and to identify the position of a point outside or inside
the polygon, some methods such as the Angle Summation test, Crossing test,
and Triangle test algorithms can be utilized [4,6].
Another way to achieve an effective message routing in OppNets is by using
fuzzy-logic [7] in the design of the utility function that decides on the selection
of suitable forwarders of a message from source to its destination. This is moti-
vated by the fact that one of the unique features of fuzzy logic is that diverse
(and sometimes conflicting parameters) of a given problem can be integrated and
evaluated as a single parameter to be used for the purpose of decision making.
In a nutshell, fuzzy logic [7] is a process that involves performing an analysis of
some information using the so-called fuzzy sets, where each of these sets is used
to describe the output and input variables, noting that every fuzzy set is charac-
terized by a membership function. The role of this function is to map each input
value to a membership value that belongs in the range [0,1]. It should be noted
that a membership value closer to 1 reveals that the corresponding input belongs
to the fuzzy set with a high degree. On the other hand, a small membership value
mean that the considered fuzzy set is not suitable for that input. Indeed, the
relationship between the inputs and corresponding outputs are governed by the
so-called fuzzy rules, which ultimate goal is to dynamically characterize the over-
all behavior of the system. Therefore, a fuzzy system is typically composed of
fuzzification and rule base, inference engine and de-fuzzification components.
The fuzzification component considers a crisp numeric value as input and
uses the rule base activation to map it to a fuzzy output (known as a suitable
ensemble of fuzzy sets), then assigns it a degree of membership to any of the fuzzy
sets. The inference engine component is the one that determines in which way the
obtained fuzzy sets can be aggregated to form the expected fuzzy solution space.
Considering the fuzzy output as entry, the de-fuzzification component converts
it into a crisp value that can be used for decision making or action control. This
paper integrates fuzzy and geocast techniques in designing a routing protocol
for OppNets (called Energy-efficient Fuzzy Check-and-Spray Geocast (EFCSG)).
Two phases are involved in the design, namely: (1) Message forwarding to the
destination cast using a fuzzy logic technique; and (2) Intelligent control flooding
 Energy-Efficient Fuzzy Geocast Routing Protocol 555

of the message to all the nodes within the geocast region using the so-called
Check-and-Spray mechanism.
The paper is organized as follows. Section 2 discusses some representative
work on fuzzy geocast routing protocols for OppNets. Section 3 describes the
proposed EFCSG protocol. Section 4 presents the simulation results on the per-
formance of the EFCSG protocol. Section 5 concludes the paper.

2 Related Work
The use of fuzzy logic for designing routing protocols for challenged networks
such as OppNets is recent. Few representative works in this direction are
described in [5–13].
Dhurandher et al. [5] proposed a geocast routing protocol for OppNets, in
which some fuzzy attributes are used in a fuzzy controller to decide on the best
forwarders nodes to carry the message toward the destination node. In their
proposed approach, a multi-copying spray mechanism is utilized for relaying the
message up to the destination cast. Next, within this cast, the message is flooded
in an intelligent manner, in such a way the undelivered copies of the messages
that are already delivered to the destination are removed from the buffers of
nodes.
Rahimi and Jamali [10] proposed a DTN routing protocol for vehicular net-
works based on fuzzy logic. In their scheme, a fuzzy controller is constructed
using parameters like node’s velocity, number of nodes, node’s direction, and
node’s distance from destination, to determine the suitability of a node to be
selected as best forwarder to carry it toward the destination. This suitability
is quantified by the calculation of the so-called node’ chance value; the highest
that value, the better chance the node has to be selected.
Banerjee and Dutta [11] presented an adaptive geocasting scheme for mobile
ad hoc networks, in which a node that belongs to the destination cast is expected
to receive a notification as being member of a geocast group and is intimated
to stay within a prescribed logically bounded network region for a limited time
period, until it is ready to accept the multicast data packet to be forwarded
toward the destination. In this process, the location information is used to reduce
the multicast delivery overhead.
Jain et al. [12] proposed a routing scheme for DTNs, which is also based
on fuzzy logic. In their scheme, the message size, the number of message repli-
cas, and message remaining battery, are used as inputs to a fuzzy controller to
make the decision on choosing the messages to be scheduled when a node is
encountered. This involves the design of a mechanism to determine the number
of message replicas in the network.
Mottaghinia and Ghaffari [13] proposed a distance and energy-aware routing
protocol for delay tolerant mobile sensor networks, in which a fuzzy controller
uses the energy of sensor nodes and the distance from the sink as parameters
to determine the next hop selection of a source node. Another fuzzy controller
is designed for buffer management purpose, which uses the number of message
replicas, the survival time, and message size as input parameters.
556 K. Khalid et al.

Recently, Khuram et al. [8] proposed a fuzzy-based geocast routing protocol

for OppNets (called FCSG), in which the output of the considered fuzzy con-
troller is used as pattern to determine the direction of the node’s movement and
the node’s speed when a source node meets an encounter. Based on this, the
direction movement of any relay node selected to forward the message toward
the destination is obtained as the angle between the line that joins the sender
node to the center of the destination cast and the line along which the node is
currently moving. As such, the smaller this angle is, the better is the chance of
successful message delivery to destination. This paper proposes an energy effi-
cient version of the FCSG scheme (called EFCSG) and evaluate its performance
against that of the FCSG [8], EECSG [9] and F-GSAF [5] schemes, in terms of
delivery ratio, average latency, number of hop count, and overhead ratio, under
varying Time-to-Live (TTL) and buffer size.

Fig. 1. Fuzzy controllers used in phase 1 of EFCSG design [8].

3 EFCSG Design
Assuming that a source node, say SN , wants to send a message m to a destina-
tion node, say D, the design of the proposed EFCSG routing scheme consists of
two phases as shown in Algorithm 1.

Notations used in Algorithm 1:

• C: number of remaining copies of a message
• VH: very high
• H: high
• M: medium
• L: low
• VL: very low
 Energy-Efficient Fuzzy Geocast Routing Protocol 557

Table 1. Movement controller [8].

Movement controller
Direction Speed 2 Movement
VL L L
VL M L
VL H L
L L L
L M L
L H M
M L L
M M M
M H H
H L M
H M H
H H H

• SN: sender node

• RN: receiver node
• m: message
• L(): likelihood function

• Phase 1: Forwarding the Message Towards the Destination Cast

In this phase, we apply a multi-copying spray technique involving the use
of two fuzzy controllers. More precisely, the message m generated by SN is
assigned a payload data, a cast definition (set of two dimensional points form-
ing the geographic cast), a pair of epoch times that define the lifetime of the
message, and an integer C representing the maximum number of remaining
copies of m that is allowed to generate and send forward to any potential
encountered node during the forwarding process of m.
When node SN opportunistically encounters another node, say RN, the like-
lihood of RN is calculated by means of a sequence of fuzzy controllers (as
depicted in Fig. 1).
In Fig. 1, the so-called Movement Controller uses as inputs the RN speed
and direction (angle between the line that joins SN to the center of the geo-
cast region and the line on which RN is moving at the moment), along with
parameters given in Table 1 to determine the movement of RN toward the
destination cast. It should be noted that the variable representing the direc-
tion varies from 0 to 180 (in degree) whereas the variable representing the
speed varies from 0 to 15 milliseconds (ms).
Next, the so-called Likelihood Controller considers the RN movement
obtained from the first controller, along with the buffer space (in bytes)
558 K. Khalid et al.

Table 2. Likelihood controller [8].

Likelihood controller
Energy Buffer 2 Movement Likelihood
L L L VL
L L M L
L L H M
L M L VL
L M M L
L M H M
L H L VL
L H M M
L H H H
M L L L
M L M M
M L H H
M M L L
M M M M
M M H H
M H L L
M H M M
M H H H
H L L L
H L M M
H L H M
H M L L
H M M H
H M H VH
H H L L
H H M VH
H H H VH

and remaining energy of RN , as inputs to determine the likelihood of RN

(denoted L(RN, m)) to receive m, based on parameters that are given in
Table 2. If L(RN, m) ≥ L(SN, m), then m is forwarded to RN and its C
value is decreased accordingly based on the current likelihood value of RN
as per Table 2; otherwise, RN is ignored, its C value is kept unchanged, and
a new search for another suitable RN is launched. In this process, if at some
point in time, the value C of RN reach 0 while message m has not yet reach
its destination D, then no additional message copies will be generated and
eventually, message m will be loss in the network.
 Energy-Efficient Fuzzy Geocast Routing Protocol 559

Algorithm 1. EFCSG Algorithm

Initialization during message generation:
m is created
Sender node and destination geocast region are defined
C is initialized

Phase 1: Moving a message towards its destination cast

1: for each SN that meets the next RN do
2: drop expired messages from buffer
3: for each m in the SN buffer do
4: if m already exists in RN then
5: skip this m and go to the next m in the for loop
6: end if
7: if C > 0 then
8: if RN is located in geocast area then
9: forward message to RN
10: C=0
11: // In case RN is not located within the
12: // geocast region, use Tables I and II to
13: // calculate the likelihood of RN to receive
14: // the message.
15: else if L(RN,m) > L(SN,m) then
16: forward a copy of message to RN
17: switch L(RN,m) do
18: case VH: C = C-5, break;
19: case H: C = C-4, break;
20: case M : C = C-3, break;
21: case L : C = C-2, break;
22: case VL : C = C-1, break;
23: end if
Phase 2: Distributing message copy to all nodes inside the geocast
24: else if C = 0 and m is in geocast region then
25: if RN is inside the destination geocast then
26: if SN energy is greater than a preset threshold value then
27: if RN does not have message m already then
28: Forward m to RN
29: end if
30: end if
31: end if
32: end if
33: end for
34: end for
560 K. Khalid et al.

• Phase 2: Forwarding the Copies of Message m to all Nodes Within

the Geocast Region
In this phase, an energy-efficient controlled flooding technique is implemented,
which consists of two steps: (1) Checking the energy level of node RN (inside the
geocast region) against an energy level threshold T ; and checking the presence
of the already existing message m in the RN cast as well as the current loca-
tion of RN ; and (2) Flooding the message m to only those nodes RN for which
the checking in (1) was successful. Afterwards, RN becomes SN and Phase 1 is
invoked again. This process is continued until its termination. It should be noted
that for the calculation of T , the residual energy of the source/intermediate
node (CurrentEnergy(node)) and the residual energy of the neighbouring
node of SN (neighbourEnergy(node)) are calculated as follows:

CurrentEnergy(node) = getCurrentEnergy(srce/inter, node)

neighbourEnergy(node) = getCurrentEnergy(neighbour)
where getCurrentEnergy() is a function returning the residual energy of the
corresponding node. Based on these values, we get

T = neighbourEnergy(node)/count

where count is the total number of neighbouring nodes of the current node.

Fig. 2. Delivery ratio versus buffer size.

4 Performance Evaluation
The performance of the proposed EFCSG routing protocol is evaluated using the
ONE simulator [14], then compared against the performances of the FCSG [8],
EECSG [9] and F-GSAF [5] schemes, in terms of delivery ratio, average latency,
number of hop count, and overhead ratio, under varying Time-to-Live (TTL)
and buffer size, using the INFOCOM 2006 real mobility traces [15].
 Energy-Efficient Fuzzy Geocast Routing Protocol 561

For these simulations, the delivery ratio is determined as the ratio of messages
successfully delivered to the destination at the end of simulations vs. the total
number of messages generated in the network. Besides, the overhead ratio is
considered a measure of the bandwidth efficiency, calculated as:
Number of relayed messages − Number of delivered messages
(1)
Number of delivered messages

4.1 Simulation Results

In Fig. 2, it is observed that the delivery ratio increases when the buffer size
is increased. Also, in terms of delivery ratio performance, E-FCSG is about
1.73% better than FCSG, 21.76% better than EECSG, and 31.65% better than
F-GSAF. This is due to the fact that when the buffer size is increased, more
messages are stored in the node’s buffer, therefore the higher the chances of
message delivery.

Fig. 3. Average latency vs. buffer size

Fig. 4. Hop count vs. buffer size

562 K. Khalid et al.

In Fig. 3, it is observed that the average latency increases as the buffer size is
increased. The reason is that when the size of the buffer increases, more messages
stay in the node’s buffer and less messages are dropped during buffer manage-
ment, causing the average latency to increase. In terms of latency performance,
it is observed that E-FCSG is about 5.22% better than FCSG, 17.31% better
than EECSG, and 22.37% better than F-GSAF.
In Fig. 4, it is found that the hop count decreases as the buffer size is
increased. This is due to the fact that, the more space the buffer size has, the
more time the messages can reside in a relay node and in this case, the hop count
decreases. In terms of hop count performance, it is E-FCSG is about 2.65% better
than FCSG, 21.24% better than EECSG, and 27.47% better than F-GSAF.

Fig. 5. Remaining energy vs. buffer size

In Fig. 5, it is observed that the remaining energy increases as the buffer size
is increased. This is due to the fact that, the energy threshold check within the
geocast area saves the nodes from further spending energy in scanning, receiving,
and sending packets. From the result, it is found that E-FCSG is about 21.78%
better than EECSG in terms of remaining energy.
In Fig. 6, it is observed that the delivery ratio increases as the TTL is
increased. In terms of delivery ratio performance, it is found that E-FCSG is
about 1.09% better than FCSG, 11.92% better than EECSG, and 31.05% better
than F-GSAF. This is attributed to the fact that in the EECSG design, a node
gets more time to find a suitable relay node to forward the message. Similar
type of results are observed for the average latency, hop count, and average ratio
metrics as depicted in Fig. 7, Fig. 8, Fig. 9, and Fig. 10, respectively.
 Energy-Efficient Fuzzy Geocast Routing Protocol 563

Fig. 6. Delivery ratio vs. TTL

Fig. 7. Average latency vs. TTL

Fig. 8. Hop count vs. TTL

564 K. Khalid et al.

Fig. 9. Overhead ratio vs. TTL

Fig. 10. Remaining energy vs. TTL

5 Conclusion
In this paper, we have proposed an energy-efficient version of the fuzzy-based
check-and-spray geocast routing scheme (called EFCSG) for OppNets. Using the
INFOCOM 2006 real mobility traces under, we have shown by simulations that
the proposed EFCSG outperforms the FCSG, EECSG, and F-GSAF routing
schemes with respect to delivery ratio, average latency, number of hop count,
and overhead ratio, under varying Time-to-Live (TTL) and buffer size. As future
work, we plan to investigate the message security in the propose EFCSG scheme,
by designing a security-aware EFCSG protocol to defend against network attacks
such as blackhole attacks.

Acknowledgements. This work is partially sponsored by a grant held by the 2nd

author, from the National Science and Engineering Research Council of Canada
(NSERC), Reference number: RGPIN-2017-04423.
 Energy-Efficient Fuzzy Geocast Routing Protocol 565

References
1. Lilien, L., Kamal, Z.H., Bhuse, V., Gupta, A.: The concept of opportunistic net-
works and their research challenges in privacy and security. In: Mobile and Wireless
Network Security and Privacy, pp. 85–117. Springer, Boston (2007)
2. Huang, C.M., Lan, K.C., Tsai, C.Z.: A survey of opportunistic networks. In: 22nd
International Conference on Advanced Information Networking and Applications-
Workshops (AINA Workshops 2008), pp. 1672–1677. IEEE (2008)
3. Kuppusamy, V., Thanthrige, U.M., Udugama, A., Förster, A.: Evaluating forward-
ing protocols in opportunistic networks: trends, advances challenges and best prac-
tices. Fut. Internet 11(5), 113 (2019)
4. Rajaei, A., Chalmers, D., Wakeman, I., Parisis, G.: GSAF: efficient and flexible
geocasting for opportunistic networks. In: 2016 IEEE 17th International Sympo-
sium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM), pp.
1–9. IEEE (2016)
5. Dhurandher, S.K., Singh, J., Woungang, I., Takizawa, M., Gupta, G., Kumar,
R.: Fuzzy geocasting in opportunistic networks. In: International Conference on
Broadband and Wireless Computing, Communication and Applications, pp. 279–
292. Springer, Cham (November 2019)
6. Ma, Y., Jamalipour, A.: Opportunistic geocast in disruption-tolerant networks. In:
2011 IEEE Global Telecommunications Conference-GLOBECOM 2011, pp. 1–5.
IEEE (December 2011)
7. Zadeh, L.A.: Soft computing and fuzzy logic. In: Fuzzy Sets, Fuzzy Logic, and
Fuzzy Systems: Selected Papers by Lotfi a Zadeh, pp. 796–804 (1996)
8. Khalid, K., Woungang, I., Dhurandher, S.K., Singh, J.: A fuzzy-based check-and-
spray geocast routing scheme for opportunistic networks. IEEE Syst. J. (2020,
submitted). ManuscriptID ISJ-RE-20-10831
9. Khalid, K., Woungang, I., Dhurandher, S.K., Singh, J., Rodrigues, J.J.P.C.:
Energy-efficient check-and-spray geocast routing protocol for opportunistic net-
works. Information 11(11), 504 (2020)
10. Rahimi, S., Jamali, M.A.J.: A hybrid geographic-DTN routing protocol based on
fuzzy logic in vehicular ad hoc networks. Peer-to-Peer Netw. Appl 12(1), 88–101
(2019)
11. Banerjee, A., Dutta, P.: Fuzzy controlled adaptive geocasting in mobile ad hoc
networks. Int. J. Inf. Technol. 14(2), 109–131 (2008)
12. Jain, S., Chawla, M., Soares, V.N., Rodrigues, J.J.: Enhanced fuzzy logic-based
spray and wait routing protocol for delay tolerant networks. Int. J. Commun. Syst.
29(12), 1820–1843 (2016)
13. Mottaghinia, Z., Ghaffari, A.: Fuzzy logic based distance and energy-aware routing
protocol in delay-tolerant mobile sensor networks. Wireless Pers. Commun. 100(3),
957–976 (2018)
14. Keränen, A., Ott, J., Kärkkäinen, T.: The ONE simulator for DTN protocol eval-
uation. In: Proceedings of the 2nd International Conference on Simulation Tools
and Techniques, pp. 1–10 (March 2009)
15. Scott, J., Gass, R., Crowcroft, J., Hui, P., Diot, C., Chaintreau, A.: CRAWDAD
dataset cambridge/haggle (v. 2009-05-29). CRAWDAD Wireless Network Data
Archive (2009)
 Time Synchronization for Automotive
Prototyping and Testing Platforms

Esraa Elelimy(B) and Andrew Azmy

Valeo, Cairo, Egypt

{esraa.elelimy,andrew.azmy}@valeo.com

Abstract. In typical automotive prototyping and testing platforms,

data streams between several heterogeneous components. Each of these
components will usually have its own internal clock. However, these
clocks will differ in their resolution, drift rate, and other properties.
Furthermore, some components will provide unstamped data, especially
when integrating automotive and non-automotive components such as
digital applications and simulation tools. As a result, having a time syn-
chronization and data stamping mechanism is indispensable to these
platforms to accurately report when an event has occurred or when a
sensor or an application has sent their data. In this paper, we present an
abstracted time synchronization approach that facilitates synchronizing
and time stamping data from diverse automotive and non-automotive
elements, regardless of the component’s underlying communication pro-
tocol. Moreover, this approach can achieve high synchronization accuracy
even when individual components are distributed over multiple networks.

1 Introduction
Prototyping is an integral part of developing complex products. In the automo-
tive industry, prototyping platforms are regularly used in the early development
stages to test and validate software applications to ensure their safety and func-
tionality before the production phase. Moreover, prototyping platforms speed
up the process of exploring new innovative ideas and novel concepts [4]. For
such platforms, the ability to test a wide range of applications is dependent
on the components integrated within the platform. Consequently, a prototyping
platform usually consists of various elements such as Electronic Control Units
(ECUs), Front Cameras, Radars, and many other sensors and controllers.
In some cases, building a physical prototyping platform might be infeasible;
one approach in such situations is to develop virtual prototyping platforms.
However, these platforms do not eliminate the need for physical ones [1,11].
Other approaches involve a combination of physical and virtual components.
For example, in cases where a specific sensor is not available, a model could
be developed using simulation tools such as CarMaker [8], SCANeR studio [21]
or unity [22] and data from that model could be passed to the other platform
components.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 566–576, 2021.
https://doi.org/10.1007/978-3-030-75100-5_49
 Time Synchronization for Automotive Prototyping and Testing Platforms 567

Whether physical, virtual, or a combination of both, these platforms require

a robust time synchronization mechanism as many automotive applications are
time-sensitive. One example of such applications is Advanced Driving Assis-
tance Systems (ADAS), which include several Cameras, radar, and ECUs; all of
these components need to have a standard reference time and be synchronized
together. Another example would be data logging from different sensors, which
also requires precise time stamping.
In addition to the previous examples, and with the recent growing appeals
for connecting the vehicles with smart-gadgets and IoT applications [5], the
prototyping platforms need to encompass all these new tools and applications.
This integration adds more diversity to the communication protocols and the
data types flowing in the vehicle. Subsequently, exposing several limitations to
the previous synchronization methods used in prototyping platforms.
This paper presents a time synchronization mechanism that facilitates syn-
chronization between automotive and non-automotive components, either digital
applications or simulation tools, as it is built upon an abstracted platform for
interoperability called Let’s DO [3] that allows integrating heterogeneous compo-
nents in a seamless manner. Additionally, the presented approach has the ability
to synchronize such components even when they are distributed across different
networks.

2 Background and Related Work

A Vehicle network contains various physical components; each uses a different
underlying communication protocol. The standard protocols found in a vehicle
network are FlexRay [15], Ethernet [2], and Controller Area Network(CAN).
However, each of these protocols has different properties and capabilities, mak-
ing time synchronization a complicated task [9]. For example, FlexRay runs on
a predefined time pattern. However, CAN supports priority-based transmission,
meaning that when multiple messages are transmitted, the one with higher prior-
ity dominates the bus. Ethernet has inherent latencies, so when using Ethernet-
based communication for time-sensitive tasks there will be a need for timing
information for the transmitted data [9] that is why various standards are avail-
able for time synchronization. As mentioned before, connecting the vehicle with
smart-gadgets adds more protocols to the vehicle network and presents new
challenges as most of these gadgets and applications do not provide stamped
data.
A plentiful of literature have discussed time synchronization, and several
methods and platforms were developed to address this problem. However, these
platforms were mostly addressing specific sub-problems in the automotive indus-
try and did not generalize or perform well when the problem is expanded. In
addition to that, the ability to communicate and synchronize with either digital
applications or simulation tools was not discussed enough in the literature.
In the following sections, we will mention several platforms and methods
that were meant to solve the time synchronization problem, we will discuss their
568 E. Elelimy and A. Azmy

shortcomings and limitations that led to the development of the synchronization

mechanism which will be introduced in this paper.

2.1 Time Sensitive Networking

AUTOSAR 4.2.2 and IEEE 802.1AS, also known as Time-Sensitive Networking

(TSN), are widely used in the automotive industry for time synchronization.
AUTOSAR harmonizes the time synchronization between FlexRay, CAN, and
Ethernet using a modified version of the IEEE 802.1AS standard [9]. However,
this approach still has some limitations. Mainly, it requires a statically predefined
network topology, and the synchronization mechanism strictly depends on that
topology. Besides, integrating new components to such topology can not be easily
configured, especially when they have different characteristics, such as digital
gadgets.

2.2 Synchronization in RTmaps

RTmaps [18] is a platform for prototyping automotive applications. It provides

an internal synchronization module to adjust the clocks across multiple dis-
tributed machines [13]. However, it does not provide a smooth way to inte-
grate and synchronize with non-automotive digital components, which provide
unstamped data or simulation tools that usually use its own reference time.

2.3 Synchronization in ROS

While Robot Operating System (ROS) [20] allows integration with heteroge-
neous components and applications, the framework itself does not support time-
synchronization across multiple machines [16], making it not ideal for automotive
prototyping.
To integrate components that do not have reference time in ROS, the devel-
opers will have to implement their synchronization mechanism as done by Fregin
et al. in [7] where they implement Precision Time Protocol (PTP) for synchroniz-
ing four cameras for an ADAS system; as ROS does not provide such capabilities.
The lack of time synchronization in ROS adds more complexity and overhead
for the developers when implementing time-sensitive applications.

2.4 Precision Time Protocol

IEEE 1588 standard, also known as Precision Time Protocol (PTP), [17] is
broadly used for time synchronization. PTP was developed to overcome the
limitations in previous protocols such as the Network Time Protocol (NTP) [19]
and Global Positioning System (GPS), as PTP provides higher accuracy than
NTP and allows for hardware time stamping. PTP is also suitable for indoor
applications as it does not depend on satellite signals, which makes it more
robust than GPS in such situations [14].
 Time Synchronization for Automotive Prototyping and Testing Platforms 569

According to the IEEE 1588 standard, clocks are arranged in a master-slave

hierarchy where each slave synchronizes with its master based on information
from several messages exchanged between them [6]. At the base of this hierarchy
is the Grand-Master clock, which is usually selected by the best master clock
algorithm and is considered as the time source to the network. This clock’s
choice depends on its properties and accuracy relative to the other clocks in the
network. The master clock sends cyclic multicast sync messages to all the other
clocks (slave clocks) in the network, while the slaves respond using a unicast
message to the master.
The messages exchanged between the master and each slave starts with a
sync message fired from the master, followed by a follow-up message containing
the timestamp at which the sync message was sent. The sync message arrives
at the slave end at time t2, and shortly after that, the follow-up message will
arrive carrying the timestamp t1. After that, the round-trip communication time
needs to be calculated to accurately get the time offset between the master and
the slave clock. As a result, the slave clock will send a delay request message
at time t3, reaching the master at time t4, and the master will respond with a
delay response message containing the actual time it received the request [14].
Figure 2 illustrates the sequence of messages exchanged between the master and
each slave. Using t1, t2, t3, t4, the slave clock can now calculate the time offset
and update its internal clock according to the following equations (Fig. 1):

D1 = t2 − t1 (1)
D2 = t4 − t3 (2)
D1 − D2
Of f set = (3)
2
D1 + D2
M eanP athDelay = (4)
2

Fig. 1. PTP timing diagram

3 Time Synchronization in Let’s DO

Previous platforms and synchronization methods had several pitfalls and limita-
tions. This paper targets two of these limitations. Firstly, the lack of the ability
570 E. Elelimy and A. Azmy

to synchronize with digital applications and IoT gadgets. Secondly, the complex-
ities involved when synchronizing hardware components and applications that
use different communication protocols.
In the following section, we will introduce Let’s DO [3], a platform for inter-
operability between different software components. After that, we will present
an abstracted synchronization mechanism built on Let’s DO and addresses the
previously mentioned shortcomings and is the main contribution of this paper.
Finally, we will then present the results obtained using that module in different
configurations.

3.1 Let’s DO

Let’s DO, introduced in [3], is a platform for interoperability and data exchange
between heterogeneous components. It provides standardized and abstract com-
munication interfaces allowing different elements and applications to communi-
cate with each other, including digital applications, simulation tools, and other
automotive and non-automotive components.
In Let’s DO, each computing element is called a node. Nodes communi-
cate with each other using the message-oriented-middleware libraries that have
abstracted and standardized APIs for communication. These standardized APIs
allow nodes to send and receive messages to each other over the Let’s DO net-
work regardless of their original underlying communication protocols, as Let’s
DO offer a variety of pre-defined messaging adapters and utility libraries.
Let’s DO has two messaging standards, Message Queuing Telemetry Trans-
port (MQTT) and Data Distribution Service (DDS). While these standards
allow interoperability and efficient communication among non-homogenous com-
ponents, they lack having synchronization capabilities.

3.2 Method

The time synchronization mechanism introduced in this paper is built as a utility

library within the Let’s DO framework [3], employing its messaging libraries.
It is also based on the IEEE 1588 standard [17] with the addition of several
modifications to be suitable for prototyping platforms.
Firstly, we created a timestamp message that could be easily imported and
used in all Let’s DO messages. We also added utilities to fill the timestamp for
any node whether it supports software or hardware time stamping.
The PTP standard is originally built upon IP and User Datagram Protocol
(UDP) [14], which complicates synchronizing components that use other pro-
tocols. To address this issue, the synchronization mechanism presented here is
built upon the standardized and abstracted Let’s DO APIs for sending and
receiving messages, making it independent of the underlying protocols used for
communication by each node.
In a Let’s DO network, each distributed node will either be a slave node or a
master node. Using Let’s DO standard APIs for publishing and subscribing, the
master node will publish a sync message and a follow-up message periodically
 Time Synchronization for Automotive Prototyping and Testing Platforms 571

every 10 ms. These messages will be published to a pre-specified topic to which

all the other nodes, slave nodes, are subscribing. Each slave node will publish
an independent delay request to the master node, in addition to the timing
information in the delay request, we added an additional ID attribute to identify
the correspondence between each slave and its message request. The master will
then publish a delay response to the node with the same ID as in the delay
request. Based on these messages exchanged, each slave node can now calculate
its offset and delay according to Eqs. (1) to (4).
Our module supports having a dynamically changing network topology,
where new nodes can enter an existing Let’s DO network at any time and still
able to synchronize with the master. This feature was not available in the pre-
vious synchronization platforms used in automotive applications. Besides, this
synchronization module can synchronize Let’s DO nodes running either on the
same network or different networks.

3.3 Experimental Results

We performed several experiments to characterize the accuracy of the introduced
synchronization module in Let’s DO. As previously mentioned, Let’s DO sup-
ports both MQTT and DDS as messaging standards. The developer can use the
desired standard during run-time through simple configurations. The nodes used
in the following experiments were configured to use MQTT.
In the first experiment, all the nodes were connected to the same WiFi
network with WPA2 security and a link speed of 65 Mb/s. Figure 2 shows the
network architecture for this experiment. In the second experiment, each node
existed in a different network and they were communicating through a common
broker on a server as indicated in Fig. 3.

Fig. 2. Distributed nodes communicating over a local network

572 E. Elelimy and A. Azmy

Fig. 3. Distributed nodes communicating over multiple networks

Both experiments were repeated several times to ensure the reproducibility

of the results. In each trial, the offset between the nodes and the master node
was recorded.
Figure 4 shows the average offset between the nodes and the master when all
the nodes were in the same network. As indicated, the average offset was 3.7 ms
with a standard deviation of 7.2 ms, and Root Mean Square Error (RMSE)
of 8.4 ms. For automotive prototyping, having a synchronization accuracy of
approximately four milliseconds is enough to test various algorithms and appli-
cations as most of the sensors used will have a refresh rate in sub-milliseconds.
For the second set of experiments, the average offset was six milliseconds with
a standard deviation of 10.6 ms, and Root Mean Square Error (RMSE) 12.2 ms
as indicated in Fig. 5. In the second case, the synchronization accuracy dropped
compared to the first experiments which were expected due to the lower quality
of service and long delays in the messages transmitted when the computing nodes
do not exist on the same network.
In [12], Mani et al. tested the synchronization accuracy over MQTT when no
additional mechanisms were employed. It was shown that the synchronization
RMSE was 162.5 ms, which is significantly higher than what was presented in
our results.
 Time Synchronization for Automotive Prototyping and Testing Platforms 573

Fig. 4. Synchronization accuracy on a Wifi Network

Fig. 5. Synchronization accuracy over different networks

4 Applications

Digitization is currently transforming every aspect of day-to-day life. Every

industry is racing to evolve to meet digital lifestyle demands and expectations.
The automotive industry is not an exception. Manufacturers are modifying their
product development strategies to meet consumer-driven requirements, which
has started by integrating digital services into vehicles’ infotainment systems
and are targeting novel solutions to further enable mobility as a service.
The need for such transformation goes deeper, reaching the development
cycle of almost all vehicle systems. This need has been greatly emphasized with
the increasing demand of remote access to all business aspects since the start of
the recent pandemic (COVID-19) [10].
574 E. Elelimy and A. Azmy

Manufacturers directed their efforts to make all physical systems needed by

their engineering teams remotely and virtually accessible, enabling them to con-
tinue being productive while working remotely.

4.1 Testing ADAS System

One of the significant activities in the development cycle of advanced driver-
assistance systems (ADAS) is in-vehicle testing. It is considered as one of the
conclusive phases of testing that should cover every possible use case and sce-
nario, taking the system under test to its limits before approving its release.
Converting such a step to a virtual activity avails numerous advantages and
savings, but simultaneously presents various challenges and required adapta-
tions. Firstly, a simulation software that can closely mimic the whole environ-
ment surrounding the vehicle is needed. Sensor models within such environment
would send relevant information to another computing node running the soft-
ware under test. This can be a PC running the software, or the actual hardware
control unit with the software flashed on it. The simulation node also sends out
reference (ground-truth) information to another node that responsible for judg-
ing the performance of the system under test based on a set of predefined key
performance indicators (KPIs). The diagram in Fig. 6 shows a sample for such
distributed testing environment.

Fig. 6. Sample of a distributed testing environment

For such a loop, temporal synchronization of the exchanged data is crucial

for formulating a useful judgement of the system performance. Synchronization
issues in the network connecting all the aforementioned nodes might hinder all
testing results useless.
 Time Synchronization for Automotive Prototyping and Testing Platforms 575

5 Conclusion
In this paper, we presented the importance of having a synchronization mech-
anism for automotive prototyping platforms. We also discussed the limitations
of existing platforms for prototyping as there is either the limitation of connect-
ing heterogeneous components or the lack of synchronization support. We then
introduced the synchronization module in the Let’s Do framework addressing
the previously mentioned gaps in the literature; having a mechanism embedded
in an abstracted framework like Let’s DO solves synchronizing heterogeneous
components and digital applications since it is independent of any particular
protocol. It was also shown that the presented module achieves a synchroniza-
tion accuracy that is sufficient for time-sensitive applications in the automotive
industry.

References
1. Cioroaica, E., et al.: Prototyping automotive smart ecosystems. In: 2018 48th
Annual IEEE/IFIP International Conference on Dependable Systems and Net-
works Workshops (DSN-W), pp. 255–262. IEEE (2018). https://doi.org/10.1109/
DSN-W.2018.00072
2. Kozierok, C.M., Correa, C., Boatright, R.B., Quesnelle, J.: Automotive Ethernet:
The defnitive guide (2005)
3. ElHakim, R., et al.: Let’s DO - Automotive Platform for Interoperability
4. Elverum, C.W., Welo, T.: The Role of early prototypes in concept development:
insights from the automotive industry. Procedia CIRP 21, 491–496 (2014). https://
doi.org/10.1016/j.procir.2014.03.127
5. Fehling, C., et al.: Your coffee shop uses cloud computing. IEEE Internet Comput.
18(5), 52–59 (2014). https://doi.org/10.1109/MIC.2014.101
6. Flammini, A., Ferrari, P.: Clock synchronization of distributed, real-time, indus-
trial data acquisition systems. In: Sciyo, MV (ed.) Data Acquisition (2010).
https://doi.org/10.5772/10458
7. Andreas, F.: ROSCon 2017: building a computer vision research vehicle with
ROS. https://roscon.ros.org/2017/presentations/ROSCon%202017%20Computer
%20Vision%20Research%20Vehicle.pdf
8. GmbH, IPG Automotive. CarMaker. Accessed 25 Aug 2020, https://ipg-
automotive.com/products-services/simulation-software/carmaker/
9. Jesse, B.: Time Synchronization in Automotive Ethernet Networks Balancing
Act Between AUTOSAR, IEEE, and TSN. Vector (2016). https://assets.vector.
com/cms/content/know-how/ technical-articles/Ethernet Timesync Automobil-
Elektronik 201608 PressArticle EN.pdf
10. Kude,T.: Agile Software Development during COVID19. http://knowledge.essec.
edu/en/innovation/agile-software-development-during-after-COVID19.html
11. Kulkarni, A., et al.: Virtual prototyping used as validation tool in automo-
tive design. Semant. Scholar (2011). https://doi.org/10.36334/modsim.2011.a5.
kulkarni
12. Mani, S.K., et al.: A System for Clock Synchronization in an Internet of Things
(2018). ArXiv:1806.02474 [Cs]
576 E. Elelimy and A. Azmy

13. Nashashibi, F.: /Sup RT/M@ps: a framework for prototyping automotive multi-
sensor applications. In: Proceedings of the IEEE Intelligent Vehicles Symposium
2000 (Cat. No. 00TH8511), pp. 99–103. IEEE Xplore (2000). https://doi.org/10.
1109/IVS.2000.898325
14. Neagoe, T., et al.: NTP versus PTP in computer networks clock synchronization.
In: 2006 IEEE International Symposium on Industrial Electronics, vol. 1, pp. 317–
362. IEEE Xplore (2006). https://doi.org/10.1109/ISIE.2006.295613
15. Corporation, N.I.: Flexray automotive communication bus overview (2015)
16. Clock - ROS Wiki. http://wiki.ros.org/Clock
17. IEEE Standard for a Precision Clock Synchronization Protocol for Networked Mea-
surement and Control Systems. In: IEEE Std 1588–2008 (Revision of IEEE Std
1588-2002), pp.1–300 (2008). https://doi.org/10.1109/IEEESTD.2008.4579760
18. Intempora RTMaps. https://intempora.com/products/rtmaps/
19. Ntp.Org: Home of the Network Time Protocol. http://www.ntp.org/
20. Robot Operating Sytem (ROS). https://www.ros.org
21. SCANeR Studio. https://www.avsimulation.com/scanerstudio/#
22. Unity. https://unity.com/products/unity-simulation
 Localization of Wireless Sensor Networks Using
Negative Knowledge Bilateration Techniques

Saroja Kanchi(B)

Department of Computer Science, Kettering University, Flint, MI 48504, USA

[email protected]

Abstract. In this paper, we propose an anchor based localization technique for

localization of nodes in wireless sensor networks given range measurements.
While much of the research has focused on gathering location information based
on existing range measurements, this paper focuses on obtaining location using
the knowledge of lack of range information between two nodes termed as neg-
ative knowledge. Three variations of the negative knowledge technique are used
to develop localization algorithm which include negative knowledge bilateration,
joint negative knowledge bilateration and double negative knowledge bilatera-
tion. The simulation results show that the proposed algorithm localizes signifi-
cantly more nodes than traditional trilateration. This improvement holds in cases
of sparse and dense random networks, connected random networks, and, H-shaped
random networks. The number of nodes localized by the proposed algorithm is
significantly higher for varying percentage of anchors and varying network radii
for each of these types of networks.

1 Introduction

With the explosion of mobile devices incorporated with sensors and the rapid growth
of IoT technology, algorithms and protocols for wireless sensor networks have received
considerable attention. Sensors have limited capability in the realms of computation,
energy, storage and mobility. When deployed, they transmit sensed data to a centralized
server or to nearest sink node, and, the sensed data is used for various applications such
as surveillance, monitoring etc. Often, when deployed in an environment that is subject
to mobility such as underwater missions, vehicles involved in combat etc. their locations
change over time. Since equipping sensors with GPS consumes tremendous energy, the
locations of deployed sensors becomes unknown over time. Localization of sensors is
defined as the problem of finding geo-location of sensor nodes under different models
of communication and different deployment architectures. Localization algorithms have
received tremendous attention in last three decades as described in the survey papers
[3–5] in the literature.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 577–587, 2021.
https://doi.org/10.1007/978-3-030-75100-5_50
578 S. Kanchi

Techniques for localization of nodes in wireless sensor networks can be categorized

using various assumptions about available information for localization: whether sensor
nodes remain stationary or mobile during execution; whether anchor nodes (anchor
based) present or no anchor nodes (anchor free) are present; whether distance information
between the nodes is available (range-based) or is not available (range-free). In addition,
3D localization, indoor localization, underwater localization and localization under noisy
information require vastly different techniques [5].
Radio range of sensor node is the radius within which a node can transmit and receive
signals to and from other sensor nodes. The range information between nodes can be
obtained using RSSI (Received Signal Strength Indicator), ToA (time of arrival of a
signal) or TDoA (Time difference between arrival of signals), AoA (Angle of Arrival
of signal). RSSI is the cheapest of the four, however, it is more error prone. A range
based distributed localization algorithm using Multi Dimensional Scaling (MDS) was
proposed in [6]. The researchers in [9] present a range based source localization in which
distance measurement of the path loss model uses ratio approach and also least squares
approach. A received signal strength indicator (RSSI) based localization technique called
cognitive maximum likelihood (C-ML) was developed in [8] to localize nodes with the
positioning scheme that relies on the statistical path loss model. Oğuz-Ekim et al. propose
a range based maximum-likelihood (ML) algorithm in [7]. Bayesian approach was used
in [14] in which position is achieved by using iterative least square (ILS) rather than
lateration. In [10], the authors provide an anchor free localization technique using the
range measurements and triangulation. An extended range and redundant association of
intermediate nodes (TERRAIN), is presented in [10]. TERRAIN uses (assumption based
coordinates) ABC and transmits data to the other nodes to construct final mapping. Moore
et al. [12] propose a range measurement based anchor free localization algorithm, using
trilateration technique in ‘Robust Distributed network Localization with noisy range
measurements’ (RODL). It adopts a cluster based localization, which uses quadrilaterals
to avoid flip ambiguities.
Range-free localization with beacons using centroid approach was presented in [1], in
which localization is accomplished by each node computing the centroid of the locations
of the anchors from which it gets information. The algorithm reduces the accumulation of
error due to low anchor density by finding the location using approximate point in triangle
test (APIT). In the range-free localization presented in [13], connectivity information is
used by finding number of hops between nodes. DV-Hop technique combines the radio
range information with the proximity to anchors to deal with the issue of low anchor
density within the proximity of the node. MDS (Multidimensional Scaling) [2] is an
anchor free technique used in both range based and range free localization and hop
distance is used in the absence of range measurements to compute distances between
nodes.
It is well known that a node’s location can be uniquely computed in 2D if the
Euclidean distances to three other nodes whose locations are known are available. This
technique of location determination is termed as trilateration. Multilateration is the exten-
sion of trilateration technique used to consolidate the distances received from several
nodes. When range measurements and anchors are available, multilateration is used to
propagate location information from one node to another thus increasing the number of
 Localization of Wireless Sensor Networks 579

anchors over time in the network. Collaborative multilateration is presented in [15]. In

[11], the limitation of trilateration-based approach is removed by proposing localization
of wheels. Applying the rigidity theory, researchers in [16] propose the concept of ver-
ifiable edges and derive the conditions for an edge to be verifiable. On this basis, they
design a localization approach with outlier detection, which explicitly eliminates ranges
with large errors before location computation. In [17], authors propose a distributed
localization protocol, GROLO, that is able to perform efficient distributed localization
through an adaptive global rigidity mechanism especially designed for the resource lim-
ited IoT nodes. In order to increase the positioning accuracy of the multi-channel location
algorithm in a multipath environment, [18] proposes an optimal multi-channel trilatera-
tion positioning algorithm (OMCT) by establishing a novel multi-objective evolutionary
model. The authors in [19] provide a localization algorithm based on shadow edges as
a necessary and sufficient localizability condition.
In this paper, we propose a localization algorithm based on negative knowledge
bilateration. The algorithm is compared to trilateration technique and we show by sim-
ulation that when localization includes negative knowledge bilateration techniques, the
algorithm localizes significantly more nodes in a variety of network or varying density
and with various number of anchors.
Section 2 provides the background of the WSN models as a network graph, the
model used and the trilateration technique. Section 3 describes the various negative
knowledge bilateration techniques used in the algorithm. Section 4 presents the algorithm
and discusses implementation details. Section 5 presents the results of simulation and
Sect. 6 provides conclusion.

2 Background
The 2D localization problem of wireless sensor network considered here is the problem
of finding mapping of the sensors in a WSN on to the 2D space such that given range
measurements between pairs of sensor nodes equals with the Euclidean distance between
the sensor nodes specified by the mapping. If we represent the underlying network graph
G = (V,E) with vertices representing sensor nodes and edge weights e(u,v) representing
the distance between vertices u and v when range measurement between u and v is
given, the problem of localization is to determine the unique mapping m() of vertices
V to points in 2D, such that dist(m(u),m(v)) = e(u,v) of the graph G for every pair of
vertices u,v eV where dist denotes the Euclidean distance between corresponding points
on the plane. It is known that the underlying graph of a network has to be globally rigid
[20] for the nodes in the network to be localizable. Here we consider networks that are
random networks, connected networks and networks of H-shape and find large number
of localizable nodes in all these types of networks.
Given a WSN, the range information between sensor nodes is available if and only
if they are within the radio range of each other. It is assumed all nodes in the network
have the same radio range and it is called the radius of the network. It is assumed that
all nodes remain stationary for the duration of the algorithm. There are some special
nodes marked anchors which are equipped with GPS and they know their locations a-
priori. The localization problem therefore is to find the geo-locations of the unlocalized
(non-anchor) nodes.
580 S. Kanchi

When a node has distance to three nodes whose locations are known in 2D, the node’s
location can be uniquely found using trilateration. The technique of using trilateration
is described in Sect. 3. The process of iterative trilateration begins by identifying an
unlocalized node in the network and localizing it if it has available, the range measure-
ments to three neighbors whose locations are known. Inversely, nodes that are adjacent
to three anchors can be localized initially as it is done in the algorithm in this paper.
These localized nodes now act as anchors to help localize other unlocalized nodes. The
process of iterative trilateration cannot progress when unlocalized nodes do not have
distances to three neighbors that are localized. The success of trilateration is therefore
sensitive to the location of the anchors within the network and order of localization of
the unlocalized nodes. Trilateration stops progressing, even if there are several nodes
with known locations in the network. Our algorithm progresses beyond trilateration by
using negative knowledge bilateration techniques.

Trilateration: Here, we first describe how an unlocalized node can find its unique loca-
tion using the geometric technique of trilateration. That is, given the distances d 1 , d 2
and d 3 from the unlocalized node A to three localized nodes P1 , P2 , and P3 respectively,
the location of the node A can be found using geometric properties on 2D plane. See
Fig. 1.

Given, P1 = (x1 , y1 ), P2 = (x2 , y2 ), and P3 = (x3 , y3 ), then the coordinates of

A = (x, y) can be found by solving the following three equations:
⎫
(x − x1 )2 + (y − y1 )2 = d12 ⎬
(x − x2 )2 + (y − y2 )2 = d22 (1)
⎭
(x − x3 )2 + (y − y3 )2 = d32

Fig. 1. Trilateration

which essentially is a solution to the equation

2MX = b (2)
 Localization of Wireless Sensor Networks 581

where
 
2(x3− x1 ) 2(y3− y1 )
M = (3)
2(x3− x2 ) 2(y3− y2 )

with
 
d12 − d32 − x12 − y12 + x32 + y32
b= (4)
d12 − d22 − x22 − y22 + x32 + y32

and
 
x
X = (5)
y

3 Negative Knowledge Bilateration Techniques

Bilateration refers to a node having distance information to two neighbor nodes as
opposed to three that is required for trilateration. this information provides two possible
positions for the unlocalized node. clearly, additional information is needed in order use
bilateration for determining the location of the unlocalized node. Given two possible
positions of a node, lack of distance information to a localized node is used determine
which of the two positions of the unlocalized node is indeed the correct geo-location.
In the next section, we describe the details of three supplemental information used with
bilaterable nodes, namely, negative knowledge bilateration, joint negative knowledge
bilateration, and, double negative knowledge bilateration. To maximize the number of
localized nodes, we run the trilateration technique after each of the negative knowledge
technique that are applied.
Negative Knowledge Bilateration: Without having distances to three localized nodes,
if an unlocalized node has distances available to two localized nodes, we call the node
bilaterable. As a first step, the novel technique of negative knowledge, that is, lack of
distance to a localized node that is within the sensor radius of one of the two possible
locations reveals that one of locations would not be correct location, due to the missing
distance information. Thus the negative knowledge of missing distance is used to confirm
the location of the unlocalized node A.
Assume that the node A is adjacent to only two other localized nodes in the graph,
say P1 , and P2 , with distances d 1 and d 2 , then clearly, A must be at the intersection of
two circles with centers P1 , and P2 and radii d 1 and d 2 respectively. Let us call the two
possible locations of node A, PA and PA’ . See Fig. 2. We propose a novel approach where
missing distance information from a localized node is used to determine the location of
A as either PA or PA’ . Let us assume that there is a localized node B within the sensor
radius s of the possible location PA of node A, and since there is no edge between B and
A then the location of A would be at PA’ . Similarly, if there is a node B, within the sensor
radius s of the possible location PA’ of node A, and since there is no edge betweenB and
A then the location of A would be at PA . This is proved in Theorem 1 below:
582 S. Kanchi

Fig. 2. Negative knowledge bilateration

Theorem 1: Assume that the possibility of trilateration is examined for all unlocalized
nodes in a network graph. Given distances two localized nodes P1 and P2 from A, and
a localized node B within sensor radius of one of the two possible locations, then the
node A is localizable.

Proof: We examine the possibility of the point A being at location PA and PA’ . Assume
that there a sensor node that located within the sensor radius around the center of PA .
Assume that there is a localized node B within this radius. Since there is no edge between
the node A and the node B, clearly node A cannot be at location PA. The reason why there
is no edge between A and B is because if there was an edge, we would have localized
the node A using trilateration technique as it is also adjacent to P1 and P2 . A must be
located at location PA’ . Thus the absence of edge between a localized node B and the
node A tells us the location of node A.

To use the negative knowledge bilateration, we identify an unlocalized node A with

two edges to localized nodes and find the two intersection points of the circles formed
by centers P1 and P2 and radii d 1 and d 2 as shown below.

(x − x1 )2 + (y − y1 )2 = d12
(6)
(x − x2 )2 + (y − y2 )2 = d22

However, it is possible that there is no localized node within the sensor radius of PA
in which case, we try to find a localized node around radius s of PA’ . If neither circles
have a localized node, then we move to localize other nodes in the network and return to
localizing A using Joint Bilateration or Double Negative Bilateration as described below.
 Localization of Wireless Sensor Networks 583

Joint Bilateration: Assume that the two possible positions PA and PA’ of bilaterable
node A. Also assume that Negative Knowledge Bilateration above has been attempted
to resolve the position of A and the position of A was not resolved. For each pair of
bilaterables A and B with PA ,PA’ as possible positions for A, and PB ,PB’ for B in the
WSN, where e(A,B) is given, we compute dist(PA ,PB ), dist(PA’ ,PB ), dist(PA ,PB’ ), and
dist(PA’ ,PB’ ). One of these four must equal d(A,B) since the all possible positions of
bilaterables A and B have been determined. We can thus jointly localize both A and B to
the corresponding locations that match e(A,B). When e(A,B) is not available as edge in
the graph, we then attempt to localize using double negative knowledge bilateration as
described below.

Double Negative Knowledge Bilateration: We now explore the case where there is no
localized node B within the sensor radius of two possible locations of the unlocalized
node PA and PA’ . In this case, we would not be able to confirm the correct location of
A. Assume that there is a localized node B that is within the radius s (adjacent to P1 or
P2 ) of the graph but is not adjacent to A. If the Euclidean distance dist(A,B) < s, then
the correct position of A is PA’, since graph would have an edge between A and B and
clearly A has only two edges to localized nodes P1 and P2. Conversely, if dist(A ,B) <
s, the A is localized to position of PA (Fig. 3).

Fig. 3. Double negative knowledge bilateration

4 Algorithm

The vertices of the WSN graph that are not anchors are initially set to unlocalizd and
all anchor nodes are set to localized. The goal of the localization algorithm is to find
location of unlocalized nodes using techniques described in Sect. 2 and Sect. 3. Once
the location is determined, the unlocalized node becomes localized.
Negative Knowledge Bilateration Algorithm:
584 S. Kanchi

1. Mark each anchor in the WSN as a localized node.

2. Repeat:
a. Repeat:
Perform trilateration by finding
trilaterables among all adjacencies of
localized nodes.
Until (no nodes can be marked localized)
b. Mark a node “bilaterable” if it is adjacent to
precisely two localized nodes during the process
above.
c. For each node that is bilaterable, apply negative
knowledge localization described in Section 3.
d. Perform joint bilateration as described in Section
3 for each pair of bilaterables which are adjacent
to each other in the WSN.
e. For each remaining bilaterable, use double
negative knowledge bilateration, if applicable

3. Until (no additional localized node is found)

In the centralized version of the code above, each node sends its neighborhood
information i,e its adjacencies to the centralized server and the server computes the
location of each node.

5 Simulation Results

The simulation was performed using by placing nodes in a 100 × 100 square for ran-
dom and connected graphs. Each data point is collected over average of 50 networks of
the given parameters. Node density is the number of nodes that is within the radio
range of nodes. For node density of nd, the number of nodes generated would be
(nd*10,000/π*s2 ), where s is the network radius. The node density of 8 for random
networks and node density of 10 was used to generate connected graphs and these
represent highly sparse graphs.
Given a specified radius, nodes were placed using uniform distribution with real value
co-ordinates in the 100 × 100 2 . For connected networks, 1-connectivity was checked
using breadth first search on the underlying graph. The percentage of anchors used
comparing the performance was 15%, which is the lowest we have seen in the literature.
The anchors were uniformly distributed in the network. The edges of the network graph
connect nodes whose Euclidean distance is less than or equal to the radius of the network.
Figures 4 and 5 show random network of radius 10 and connected network of radius
10 respectively as a sample graph. Figure 6 shows the results of localization on connected
networks of various percentage of anchors on a sparse graph of radius 12. Note that
the number of nodes localized using the negative knowledge localization techniques is
significantly higher for all anchor percentages. Similar results can be seen for random
networks in Fig. 8. Figure 7 shows the number of nodes localized for various radii from
sparse to dense graph. Again it is noted that the new algorithm performs significantly
better for various radii. The percentage of anchors is set to 15% in this case. Similar
 Localization of Wireless Sensor Networks 585

results can be found for random networks as shown in Fig. 9. Figure 10 shows a H-shaped
graph of radius 10 and Fig. 11 shows the performance of the algorithm for H-shaped
graph.

Fig. 4. Random network Fig. 5. Connected network

Fig. 6. Connected graphs: anchors varied Fig. 7. Connected graphs: radius varied

Fig. 8. Random graphs: anchors varied Fig. 9. Random graphs: radius varied
586 S. Kanchi

Fig. 10. H-Graph with radius 10 Fig. 11. H-Graphs: radii varied

6 Conclusion
In this paper, we have presented an improved localization algorithm which used lack
of distance information in three different ways to localize more nodes than traditional
trilateration. Simulation results show that significantly more nodes are localized using
proposed algorithm for random sparse and dense networks, connected networks and
H-shaped networks.

References
1. Bulusu, N., Heidemann, J., Estrin, D.: GPS-less low-cost outdoor localization for very small
devices. IEEE Pers. Commun. 7(5), 28–34 (2000)
2. Shang, Y., Ruml, W., Zhang, Y., Fromherz, M.P.: Localization from mere connectivity. In:
Proceedings of the 4th ACM International Symposium on Mobile Ad Hoc Networking &
Computing, pp. 201–212, June 2003
3. Mao, G., Fidan, B., Anderson, B.D.: Wireless sensor network localization techniques.
Comput. Networks 51(10), 2529–2553 (2007)
4. Amundson, I., Koutsoukos, X.D.: A survey on localization for mobile wireless sensor net-
works. In: International Workshop on Mobile Entity Localization and Tracking in GPS-less
Environments, pp. 235–254. Springer, Heidelberg, September 2009
5. Chowdhury, T.J., Elkin, C., Devabhaktuni, V., Rawat, D.B., Oluoch, J.: Advances on local-
ization techniques for wireless sensor networks: a survey. Comput. Networks 110, 284–305
(2016)
6. Ji, X., Zha, H.: Sensor positioning in wireless ad-hoc sensor networks using multidimensional
scaling. In: IEEE INFOCOM 2004, vol. 4, pp. 2652–2661. IEEE, March 2004
7. Oğuz-Ekim, P., Gomes, J.P., Xavier, J., Stošić, M., Oliveira, P.: An angular approach for range-
based approximate maximum likelihood source localization through convex relaxation. IEEE
Transactions on Wireless Communications 13(7), 3951–3964 (2014)
8. Bandiera, F., Coluccia, A., Ricci, G.: A cognitive algorithm for received signal strength based
localization. IEEE Trans. Signal Process. 63(7), 1726–1736 (2015)
9. Xu, Y., Zhou, J., Zhang, P.: RSS-based source localization when path-loss model parameters
are unknown. IEEE Commun. Lett. 18(6), 1055–1058 (2014)
10. Savarese, C., Rabaey, J.M., Beutel, J.: Location in distributed ad-hoc wireless sensor net-
works. In: 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing.
Proceedings (Cat. No. 01CH37221), vol. 4, pp. 2037–2040. IEEE, May 2001
 Localization of Wireless Sensor Networks 587

11. Yang, Z., Liu, Y., Li, X.Y.: Beyond trilateration: On the localizability of wireless ad-hoc
networks. In: IEEE INFOCOM 2009, pp. 2392–2400. IEEE, April 2009
12. Moore, D., Leonard, J., Rus, D., Teller, S.: Robust distributed network localization with noisy
range measurements. In: Proceedings of the 2nd International Conference on Embedded
Networked Sensor Systems, pp. 50–61, November 2004
13. Singh, S.P., Sharma, S.C.: Range free localization techniques in wireless sensor networks: a
review. Procedia Comput. Sci. 57(7–16), (2015). 3rd.
14. Coluccia, A., Ricciato, F.: RSS-based localization via Bayesian ranging and iterative least
squares positioning. IEEE Commun. Lett. 18(5), 873–876 (2014)
15. Savvides, A., Park, H., Srivastava, M.B.: The bits and flops of the n-hop multilateration prim-
itive for node localization problems. In: Proceedings of the 1st ACM International Workshop
on Wireless Sensor Networks and Applications, pp. 112–121, September 2002
16. Yang, Z., Wu, C., Chen, T., Zhao, Y., Gong, W., Liu, Y.: Detecting outlier measurements based
on graph rigidity for wireless sensor network localization. IEEE Trans. Vehicular Technol.
62(1), 374–383 (2012)
17. Wu, H., Ding, Z., Cao, J.: GROLO: realistic range-based localization for mobile IoTs through
global rigidity. IEEE Internet Things J. 6(3), 5048–5057 (2019)
18. Fang, X., Chen, L.: An optimal multi-channel trilateration localization algorithm by radio-
multipath multi-objective evolution in RSS-ranging-based wireless sensor networks. Sensors
20(6), 1798 (2020)
19. Oliva, G., Panzieri, S., Pascucci, F., Setola, R.: Sensor networks localization: extending
trilateration via shadow edges. IEEE Trans. Automat. Contr. 60(10), 2752–2755 (2015)
20. Eren, T., Goldenberg, O.K., Whiteley, W., Yang, Y.R., Morse, A.S., Anderson, B.D., Bel-
humeur, P.N.: Rigidity, computation, and randomization in network localization. In: IEEE
INFOCOM 2004, vol. 4, pp. 2673–2684. IEEE, March 2004
 UAVs Route Planning in Sea Emergencies

Nicholas Formica1 , Leonardo Mostarda2 , and Alfredo Navarra1(B)

1
Mathematics and Computer Science Department,
University of Perugia, Perugia, Italy
[email protected], [email protected]
2
Computer Science Division, University of Camerino, Camerino, Italy
[email protected]

Abstract. We consider the First Boat Rescue (FBR) problem, a new

challenging variant of the well-known Electric Vehicle Routing problem.
It comes from practical scenarios of rescue activities in the sea. In partic-
ular, usually boats relatively close to the coast that require some medical
intervention are served by rescue lifeboats, even though most of the times
the intervention is unnecessary. As alternative and cheaper approach,
here we propose the usage of drones, equipped with basic medical tools
that might suffice to solve the request. The main issue is that drones
are battery powered and their lifespan is rather short. Hence, recharg-
ing activities are required in order to accomplish their missions. To this
respect, we exploit a recent development about a charging station com-
posed of a buoy that recharges itself by means of the sea movement and
provides a charging pad for a drone. It comes out that by suitably dispose
a set of such buoys in the area of interest to be monitored, it is possible
to serve many boats alarms by means of drones, without wasting fuel
and time for rescue lifeboats. We present a resolution algorithm for FBR
along with various simulations concerning random and ad-hoc scenarios.

1 Introduction

Boating is the leisurely activity of travelling by boat, or the enjoyment useof

a boat, such as powerboats, sailboats, or man-powered vessels, for travelling as
well as sports activities, such as fishing or water-skiing. It is a popular activity,
and there are millions of boaters worldwide. The statistics that are reported in
2013 by the Italian Ministry of Infrastructure and Transport show that boating
activity is usually performed by lots of boats with small and medium size. More
precisely, 41% of boats have an overall length that is less than 10 m, 58.8% have
a length between 10 and 24 m while only 0.3% of the boats have a length that
exceeds 24 m.

The work has been supported in part by the Italian National Group for Scientific
Computation GNCS-INdAM, project “Travelling Salesman Problem 2.0” – Progetti di
Ricerca 2020, and by the Italian project “HALYomorpha halys IDentification” (HALY-
ID), funded by ICT AGRIFOOD MIPAAF, CUP J99C20000820003.

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 588–599, 2021.
https://doi.org/10.1007/978-3-030-75100-5_51
 UAVs Route Planning in Sea Emergencies 589

Every year, boating activities require to deal with a high number of rescuing
calls. In Italy in 2013, 6166 boats called for help. During the summer period (35
days) about 1152 people called for aid in Marche region. The help procedure
requires a rescue lifeboat to reach the boat asking for help and assist it. The
main problem is that most of the help calls are false alarms. More precisely, only
2.4% of the calls requires rescuing, thus lifeboat assistance. False alarms cause
waste of resources (like fuel) but most importantly they prevent the lifeboat to
quickly assist other boats that are really in need. In order to solve this problem
we propose the use of telemedicine via Unmanned Aerial Vehicles (UAVs). A
UAV that is equipped with audio/video devices and few medical sensors (e.g.,
temperature, pulse oximeter, heartbeat and glucometer) is sent to the boat.
A doctor from a remote base station can connect to the UAV, evaluate the
emergency level and decide whether or not a lifeboat is needed. Our solution
also requires the use of an energy sea buoy [12] that can be used to recharge the
UAV when long distance needs to be covered.

1.1 Related Work

Our problem is closely related to the Electric Vehicle Routing Problem (eVRP)
[4,7] which extends the classical Vehicle Routing Problem (VRP) in order to con-
sider the limited autonomy range of electric vehicles. This limitation is solved
by planning battery charging/replacing stations. eVRP models are usually based
on the following assumptions: (i) the vehicles can fully or partially charge their
batteries at every charging station, and (ii) the battery charging is linear. The
author in [8] presents a survey on the electric vehicle routing problem. They
focus on battery electric vehicles for logistic purposes by considering two main
problems that are limited driving range and the need for an additional recharging
infrastructure. They incorporate the maximum-route distance into the vehicle-
scheduling model and consider battery swapping or full-fast charging. They
model the single-depot vehicle-scheduling problem and propose a heuristic. The
work presented in [2] is quite close to ours. The routing problem is related to
electric vehicles that can use different recharge technologies and can perform
partial recharges. A set of feasible routes is a feasible solution if all customers
are visited once and no more than a fixed number of vehicles is used. The goal
is to minimise the total recharging cost. This is estimated by considering the
number of recharge cycles after which a battery is replaced. In [10], the authors
consider partial nonlinear charging. They propose a hybrid metaheuristic that is
applied to a set of realistic instances. A set of customers are served by using an
unlimited and homogeneous fleet of electric vehicles and charging stations can
handle an unlimited number of vehicles simultaneously. Each customer is visited
exactly once; each route starts and ends at the depot and the objective function
minimises the total route time. In [13], the authors study a Multi-Depot Green
Vehicle Routing problem where vehicles start from different depots, serve cus-
tomers, and, at the end, return to the original depots. The limited capacity of
vehicles forces them to visit fuel stations. The objective is to minimise the total
carbon emissions. A Two-stage Ant Colony System is proposed to find a feasible
590 N. Formica et al.

and acceptable solution for this NP-hard problem. In [1], the authors present a
green vehicle routing problem. They assume a graph that is composed of ver-
tices representing customer locations, alternative fuelling stations (AFSs), and a
depot. A solution is a set of vehicle tours with minimum distance. Each vehicle
starts and returns from a depot, visits a set of customers without exceeding the
autonomy driving range. Each tour may include a stop at one or more AFSs to
allow the vehicle to refuel en route. In [9], the authors present a general eVRP
where electric vehicles can recharge at a charging station during their delivery
operations. Their formulation consider the costs associated with time and elec-
tricity consumption. All vehicles start and end at the same depot. The proposed
eVRP finds the minimum travel time cost and energy cost. Our problem, and in
general the eVRP, are closely related to the classical VRP [5,11] and to the Multi
Depot Vehicular Routing Problem (MVR) [6] which in turn generalises the Trav-
elling Salesman Problem. Such problems are known to be NP-hard. They study
how goods can be delivered from various depots by using a set of home vehicles
that can serve a set of customers via a given road network. MVR finds a route for
each vehicle that starts and finishes at its depot and the global transportation
cost is minimised. In [3], the authors present the Steiner Travelling Salesman
Problem which is closely related to ours. In fact, each salesman is assigned to
visit a subset of nodes, called terminals, possibly passing through some further
nodes not required to be visited. The goal is to find minimum-weight closed
walks.

1.2 The Marche Region Case Study

We have considered the Italian Marche region case study. In particular, the area
of the cost selected in the sea of Marche concerns 175 km × 40 km mission
region. A possible configuration of base stations and charging stations is shown
in Fig. 1A. Sea buoys act as charging stations for drones. The charging stations
can charge drones with “contact based” charging systems1 (see Fig. 2). The pad
provides up to 500W loss-free charging system for the drones. Oil platforms can
also be used to host charging stations for charging drones. Figure 1B shows the
deployment of oil platforms in the sea of Marche region.
We considered a consumer grade drone2 which can be seen in Fig. 2a. It has
4280 mAh battery which can allow the drone a 27 min flight. The drone can
fly at maximum speed of 94 km/h. The battery can be charged in about 27 min
with linear quick charging. Audio, video and some sensors (such as oxygen,
temperature) will be used during the rescuing. These are powered by a different
battery thus rescuing do not affect the drone energy consumption. The rescue
time when serving a boat is 15 min in average.

1
https://skycharge.de/charging-pad-outdoor.
2
The DJI Inspire 2.
 UAVs Route Planning in Sea Emergencies 591

A B

Fig. 1. A) Charging buoy grid, B) Oil platforms

(a) The drone (b) Pad connection

Fig. 2. The contact-based charging pad for the drones.

Our problem can be seen as a novel variation of the eVRP. A set of drones
need to save a set of boats by going via one or more charging buoys. Each
buoy can host at most on drone at the time. The goal is to minimise the time
needed to rescue all the boats. This may require a drone to partially charge its
battery. In our problem, drones are not required to return at the departing base
station. The organisation of buoys into grid can be a peculiarity that can be
further analysed. The Marche region scenario suggests similar case studies such
as drone patrolling and woods monitoring.

2 Problem Formalisation
In this section, we provide the formalisation of our problem. We consider a
fixed infrastructure of buoys given by a rectangular grid G of n rows and m
columns. Each vertex of the grid represents a buoy, i.e. a recharging station.
The bottom row of the grid represents base stations from where drones can
start their journeys. The number of drones that can be exploited for a mission
is denoted by k, whereas the number of boats that may require intervention is
denoted by B. Boats can appear in any point within the area delimited by G.
592 N. Formica et al.

First Boat Rescue (FBR)

Input: A rectangular grid G embedded in the plane of size n × m buoys where
the bottom row represents the possible starting base stations; k drones;
B boats. Each buoy is associated with a charging power eb . Each drone
is associated with a battery of power ed ; Each boat is represented by a
point in the area covered by G. Furthermore, the time required to fully
or partially recharge a drone on a buoy, as well as the speed of flying,
depend on the used technology
Solution: For each drone, a trajectory that starts and ends in a base station (not
necessarily the same one) that can be traced by the drone without
draining its battery
Goal : Minimise the total time required to serve all the B boats

Clearly, the resolution of the problem highly depends on the chosen technol-
ogy. In particular, the capabilities of the drones in terms of flying autonomy and
speed are crucial.
From the point of view of pure combinatorial optimisation, the problem
becomes easily unfeasible. In fact, one may consider an instance where plenty of
boats surround a single buoy so that the charging energy available is not enough
to guarantee the drone to fly over all the boats, whereas other buoys cannot
be exploited because too far apart. On the other hand, the depicted scenario is
rather unrealistic. This suggests to consider “reasonable” instances dictated by
practical observations. Moreover, depending on the traffic, i.e., the number of
boats one may expect to serve and the number of available drones, one parame-
ter that can be suitably tuned to increase the percentage of successful missions
concerns the density of the underlying grid. Our investigation is then intended
to provide a suitable heuristic for the resolution of the problem in practical sce-
narios. Moreover, our study might be useful also to understand what should be a
satisfying size for the underlying grid in order to be able to serve all the expected
boats.

3 Resolution Algorithm

In this section, we describe our heuristic for the resolution of the FBR problem.
Algorithm 3 shows a portion of the corresponding pseudo-code which concerns
the computation of the trajectories that the selected drones must trace.
In order to provide the correct input to Algorithm 3, we need to solve an
instance of the Multi-Depot Vehicle Routing (MVR) problem [4,6,7]. As observed
in Sect. 1.1, this is a well-known variant of the standard VPR where there are
many depots (one for each vehicle) from where vehicles can start and end their
journeys. In practice, by ideally considering the drones with infinite battery, the
resolution of MVR provides a preliminary solution for FBR. The solution of MVR
is provided by a set of trajectories, one for each vehicle (drone) involved so as
to minimise the time required to serve the requests (boats). Each trajectory τ
 UAVs Route Planning in Sea Emergencies 593

Fig. 3. Algorithm for FBR starting from a solution of MVR.

is represented by a sequence (v0 , v1 , . . . , vc ) of c + 1 points, for some c > 0,

that must be reached by a vehicle. Moreover, each trajectory τ is associated
with a Costτ which represents the time needed by a vehicle to trace τ . Clearly,
MVR does not take into account the battery constraint for the vehicles, hence
no intermediate stops (buoys) are necessary. It turns out that an optimal solu-
tion to MVR represents a lower bound to our FBR. Again, as MVR is NP-hard
[4,6,7], of course the same holds for FBR. In our implementation, we refer to
the code provided in https://github.com/brandhaug/mdvrp-ga to solve MVR,
which makes use of genetic techniques.
The obtained solution is then processed by means of Algorithm 3 in order to
check (see Line 3) whether the computed trajectories can be traced as they are,
or (Lines 4–6) they need some recharging activities according to the capabilities
of the drones we consider. If a buoy vb , with charging power eb and maximum
charging power maxb , is inserted between two consecutive points vi−1 and vi of a
computed trajectory τ (see Lines 7–10), then Costτ must be updated of quantity
t(vi+1 , vb ) + t(vb , v1 ) − t(vi+1 , vi ) (Line 11), where function t(x, y) returns the
time needed by the drone to move from x to y, whereas x and y range among
the set of buoys and boats.
The algorithm also applies some backtrack method (up to four hops back-
ward, see Line 13) when the current trajectory cannot be traced. This might
happen for instance if the buoys around the current tested position have been
already assigned to other drones and the remaining charge is not sufficient to
reach the desired target. To this respect, it is also worth noting that the order
in which the various trajectories returned by solving MVR are considered might
imply different solutions or even lead to impossibility for FBR. For this reason, in
our implementation we consider all (or a subset of) the permutations to increase
the chances to find a (good) solution.
594 N. Formica et al.

Our heuristic also optimises the times required for charging activities. In fact,
drones are assumed to start with the full charge but, when buoys are exploited,
only the percentage of battery required to reach the subsequent buoy is con-
sidered as charging requirement. This allows the use of the same buoy to serve
different drones, with the only constraint that only one drone per time lands on
such a buoy.

4 Experiments
In this section, we first remind all the settings under which our simulations have
been conducted. In particular, the chosen parameters directly come from the
practical scenario described in Introduction. Then, we show the results we have
obtained for ad-hoc and random instances.

4.1 Settings
Based on the practical scenario described in Introduction, as field of action we
consider a rectangular grid G of size 3 × 8. As shown in Fig. 1A, this allows the
coverage of an area of 175 × 40 km2 . In fact, according to the characteristics of
the buoys and those of the drones, the chosen setting allows a good compromise
in terms of density and feasibility in real scenarios.

• Subsequent columns are distanced 25 km one from the other; subsequent rows
are distanced 20 km one from the other.
• The bottom row of the grid represents the base stations from which drones
can start to fly; we assume at most one drone per base station;
• Initially, every buoy b (and every drone d, resp.) is assumed to be fully charged
with a power of maxb (maxd , resp.) corresponding to 27 min of charging
(flying, resp.);
• As observed, the considered drone can fly for 27 min at a maximum speed of
94 km/h; We assume a constant average speed of 80 km/h;
• Once a boat is reached, no battery consumption is assumed as the drone
provides further instrumentation self-supplied;
• The time required for serving a boat is assumed to be of 15 min on average;
• A drained battery of the drone requires 27 min to get fully charged;
• The time required by a drone d to accomplish all its planned trajectory τd is
given by the difference between the arrival time and the starting time;
• The time required by a drone d to accomplish its mission (i.e., for serving
all the boats it is assigned to) is given by the difference between the time in
which it finishes to serve the last boat of its trajectory and the starting time.

Boats are randomly distributed in the area covered by G. We vary the number
of boats from 5 to 20. Concerning the number of drones, it is assumed that up to
8 drones (one per base station) can be used. However, according to the input
instance, it is not always necessary nor recommended to make use of the whole
swarm.
 UAVs Route Planning in Sea Emergencies 595

Fig. 4. Solution for B = 10.

We also conduct some experiments with specific locations for the boats with
the intent to stress the system in what we intuitively consider extreme cases.
These would come out with very low probability from the uniform distribution.

4.2 Simulations

The first experiments reported in Figs. 4, 5 and 6 concern the uniform distribu-
tion of boats from 5 to 20. The case with 5 boats is not reported because is very
similar to the case of 10. When the density of the boats is up to 10, the resolution
of the problem is rather natural and straightforward from the solutions obtained
for MVR.
In Fig. 4, we find a solution for FBR when 10 boats are considered. Num-
bers associated to buoys and boats are just meant as ids to univocally refer to
such elements. The obtained solution makes use of 3 drones since the usage of
further drones would increase the total time of intervention. The solution has
been obtained as a natural extension of the one obtained for MVR by using the
considered genetic algorithm. The first trajectory from the left requires a time
mission of 1 h and 42.25 min which includes 45 min required to serve three boats
and 22 min for the charging activity at buoy 19. The total time to trace the
whole trajectory is instead of 2 h and 43 min. This is given by the mission time
plus the charging activity at buoy 18 (of 24 min) and the time required to trace
the segment from 18 to 1. It is worth mentioning that the time required to trace
the whole trajectory and to serve the three boats, produced by the resolution of
the underlying MVR problem, was of 1 h and 42 min.
The longest trajectory is the third one. It requires a total of 3 h and 14 min,
against 2 h and 15 min obtained for MVR. The second trajectory shows a case
where the drone starts and ends its journey in two different base stations.
In Fig. 5, we find a solution for FBR when 15 boats are considered. The
longest trajectory is the second one which requires a total of 4 h and 54.75 min,
against the 3 h and 39.75 obtained for MVR. Actually the length of this trajectory
seems to suggest that it would be better to split it between two drones. However,
the usage of another drone may possibly make the whole solution more costly,
596 N. Formica et al.

Fig. 5. Solution for B = 15.

Fig. 6. Solution for B = 20.

because it could require to trace longer trajectories in order to ensure the two
drones can find enough energy on the buoys for recharging activities. To this
respect, it is interesting to see how the buoy 33 has been used for allowing the
drone to keep on tracing its trajectory.
In Fig. 5, we find a solution for FBR when 20 boats must be served. Consid-
ering that the field of intervention contains 24 buoys, then 20 boats constitute a
good threshold above which it might happen that no solution can be found, at
least within ‘reasonable time’, i.e. without waiting for buoys to get back their
charge. In fact, we have to keep in mind that we are dealing with rescue activi-
ties. By our experiments, we observed that B = 20 is a good limit that, under the
studied scenario and with the uniform distribution, guarantees to find effective
solutions. Of course in a context where more boats are expected, the density of
the buoys must be reconsidered.
The longest trajectory is the first one which requires a total of 3 h and
53.25 min, against the 2 h and 31.5 min obtained for MVR. As it can be seen,
just a few of the available buoys remain unused and also the number of drones
involved is 6 out of the 8 that are available.
Concerning ad-hoc examples, in Fig. 7, we report two cases where 10 boats
are thicken in a very restricted area. In Fig. 7a, by means of two drones it has
been possible to find a solution. Contrarily, with the density shown in Fig. 7b,
 UAVs Route Planning in Sea Emergencies 597

(a) Dense solvable case. (b) Dense unsolvable case.

Fig. 7. Dense cases with B = 10.

Fig. 8. Solution for B = 20 in the case where boats are faraway from the depots.

no reasonable solution exists. Clearly, in both cases it would be better to use

a rescue lifeboat rather than drones as some non-false alarms might be easily
faced. Another ad-hoc positioning of the boats is shown in Fig. 8. We consider
20 boats disposed faraway from the depots. The obtained solution is indeed
rather satisfactory. It makes use of just three drones and the longest trajectory
requires 5 h and 4 min against the 3 h and 23.25 min obtained from the resolution
of the underlying MVR problem. Moreover, the time of mission, is just of 3 h and
56.5 min, i.e., the drone requires 1 h and 7.5 min to come back to the depot.

5 Conclusions and Future Works

We introduced the First Boat Rescue (FBR) problem, a new challenging variant
of the well-known Electrical Vehicular Routing Problem. Starting from the offi-
cial data provided by the Italian Ministry of Infrastructure and Transport, we
have envisioned a possible solution that exploits advanced technologies based on
drones. Hence, we presented a resolution algorithm along with an extensive case
study in order to show various peculiarities.
Apart for its practical interest, FBR represents an interesting combinatorial
problem that deserves further investigation. We know it is NP-hard in general
and that sometimes even feasibility cannot be reached, provided that it is not
598 N. Formica et al.

reasonable to wait for the recharging time of the buoys. However, it would be
challenging to characterise the problem with respect to the topology of the under-
lying network, to the relation between drones and buoys and to their capabilities.
In particular, many parameters might be considered that may completely change
the resolution strategy. For instance, is the rectangular grid the best way to dis-
pose the buoys? It is reasonable to spread the buoys according to the expected
traffic of boats. So far, we have considered a uniform distribution but perhaps
different data might suggest different infrastructures. Another consideration may
refer to the weather. Trajectories might take into consideration the current wind
or the possibility of thunderstorms.
Furthermore, the practical environment suggests that an online version of
FBR should be considered. How trajectories might be modified on-the-fly if boats
are assumed to appear in real time and not known in advance? Dynamicity cer-
tainly represents a great challenge. To this respect, another interesting scenario
considers buoys with different charges and drones suitably positioned in vari-
ous buoys, not necessarily on the depots. Actually, one question might concern
what would be the most suitable placement for the drones in order to better serve
possible requests and whether drones should move to re-arrange themselves once
one or more drones started a mission. In doing so, drone might become better
prepared to further requests. Finally, as we have shown, all the technology for
real experiments is available. Hence, it would be very interesting to verify how
the proposed strategy may behave in practice. This may provide suggestions for
modifications in case the parameters mentioned above, and not considered so far
in our solutions, reveal to be decisive to accomplish required missions.

References
1. Erdoğan, S., Miller-Hooks, E.: A green vehicle routing problem. Transp. Res. Part
E: Logistics Transp. Rev. 48(1), 100–114 (2012). https://doi.org/10.1016/j.tre.
2011.08.001
2. Felipe, Á., Ortuño, M.T., Righini, G., Tirado, G.: A heuristic approach for the
green vehicle routing problem with multiple technologies and partial recharges.
Transp. Res. Part E: Logistics Transp. Rev. 71, 111–128 (2014). https://doi.org/
10.1016/j.tre.2014.09.003
3. Gabrel, V., Mahjoub, A.R., Taktak, R., Uchoa, E.: The multiple Steiner TSP with
order constraints: complexity and optimization algorithms. Soft Comput. 24(23),
17957–17968 (2020). https://doi.org/10.1007/s00500-020-05043-y
4. Ghorbani, E., Alinaghian, M., Gharehpetian, G.B., Mohammadi, S., Perboli, G.:
A survey on environmentally friendly vehicle routing problem and a proposal
of its classification. Sustainability 12(21), 9079 (2020). https://doi.org/10.3390/
su12219079
5. Han, M., Wang, Y.: A survey for vehicle routing problems and its derivatives. In:
IOP Conference Series: Materials Science and Engineering 452, 042,024 (2018).
https://doi.org/10.1088/1757-899x/452/4/042024
6. Karakatic, S., Podgorelec, V.: A survey of genetic algorithms for solving multi
depot vehicle routing problem. Appl. Soft Comput. 27, 519–532 (2015). https://
doi.org/10.1016/j.asoc.2014.11.005
 UAVs Route Planning in Sea Emergencies 599

7. Lalla-Ruiz, E., Erdelić, T., Carić, T.: A survey on the electric vehicle routing
problem: Variants and solution approaches. J. Adv. Trans. 2019, 5075,671 (2019).
https://doi.org/10.1155/2019/5075671
8. Li, J.Q.: Transit bus scheduling with limited energy. Transp. Sci. 48(4), 521–539
(2014). https://doi.org/10.1287/trsc.2013.0468
9. Lin, J., Zhou, W., Wolfson, O.: Electric vehicle routing problem. Transp. Res.
Procedia 12, 508–521 (2016). In: Tenth International Conference on City Logis-
tics, 17—19 June 2015, Tenerife, Spain. https://doi.org/10.1016/j.trpro.2016.02.
007. http://www.sciencedirect.com/science/article/pii/S2352146516000089
10. Montoya, A., Guéret, C., Mendoza, J.E., Villegas, J.G.: The electric vehicle routing
problem with nonlinear charging function. Transp. Res. Part B: Methodol. 103,
87–110 (2017). https://doi.org/10.1016/j.trb.2017.02.004
11. Mor, A., Speranza, M.G.: Vehicle routing problems over time: a survey. 4OR 18(2),
129–149 (2020). https://doi.org/10.1007/s10288-020-00433-2
12. Viet, N.V., Wu, N., Wang, Q.: A review on energy harvesting from ocean waves
by piezoelectric technology. J. Model. Mech. Mater. 1(2) (2017). https://doi.org/
10.1515/jmmm-2016-0161
13. Zhang, W., Gajpal, Y., Appadoo, S.S., Wei, Q.: Multi-depot green vehicle routing
problem to minimize carbon emissions. Sustainability 12(8), 3500 (2020). https://
doi.org/10.3390/su12083500
 SQUAB: A Virtualized Infrastructure for
Experiments on BGP and its Extensions

Naoki Umeda1(B) , Naoto Yanai1 , Tatsuya Takemura1 , Masayuki Okada2 ,

Jason Paul Cruz1 , and Shingo Okamura3
1
Osaka University, Suita, Japan
[email protected]
2
University of Nagasaki, Sasebo, Japan
3
National Institute of Technology, Nara College, Nara, Japan

Abstract. The border gateway protocol (BGP), which is known as a

backbone protocol of the Internet, is constantly the target of many hijack
attacks. To combat such attacks, many extensions of BGP have been
developed to make BGP more secure. However, to perform experiments
to evaluate their performance, most BGP extensions require the utiliza-
tion of platforms, such as testbeds, with high operating costs. In this
paper, we propose SQUAB (Scalable QUagga-based Automated Configu-
ration on Bgp), a lightweight evaluation tool for protocols under devel-
opment and for protocols that will be developed by a user with actual
devices locally. SQUAB can configure BGP networks automatically, and
thus it can significantly reduce the overhead of experiments on BGP
and its extensions. Unlike conventional testbeds, SQUAB can set up
BGP networks locally and its execution requires only a computational
resource of a typical laptop computer. We used SQUAB in experiments
to check the validity of functions based on network topologies in the real
world. Our results show that SQUAB can configure a network composed
of 50 routers within 52.9 s and consumes only 354.7 MB of memory. Fur-
thermore, as a case study, we confirm that SQUAB can also provide
experimental networks, even including other protocols, e.g., BGPsec.

1 Introduction
Border gateway protocol (BGP) [22] is a protocol designed to exchange route
information among networks called autonomous systems (ASes) by assigning
unique numbers called AS numbers. BGP is utilized as a backbone protocol of
the Internet, and many vulnerabilities have been reported [20]. In fact, BGP
was found to be targeted by an average of 4.82 hijacks per day [28], and all
hijacks inflicted serious damage. Therefore, guaranteeing the validity of route
information is an important and urgent issue in Internet security.
Many countermeasures against BGP hijacking have been proposed in the
past years [20]. However, evaluations of the effectiveness and risks of these
countermeasures are insufficient. This lack of evaluation is mainly due to BGP
itself, i.e., it cannot lend itself well to supporting experiments [24].
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 600–613, 2021.
https://doi.org/10.1007/978-3-030-75100-5_52
 SQUAB: A Virtualized Infrastructure for Experiments on BGP 601

Therefore, many efforts have been devoted to developing tools that can effec-
tively evaluate BGP and its extensions. In general, an evaluation tool should be
able to support network configurations and manage interconnections between
BGP routers in an experiment [24]. However, existing evaluation tools require
users to configure networks manually, which is a challenge even for experts, when
designing BGP configurations that operate as intended [12,16]. This means that
the existing tools have limitations in scaling experiments. Notably, when users
want to evaluate their designed protocol locally, they need to manually set up
a network environment for their experiments. Such manual setup, therefore,
requires additional effort from users during their experiments.
In this paper, we present SQUAB (Scalable QUagga-based Automated Config-
uration on Bgp), a novel tool that configures BGP environments automatically
for local experiments. It is a user-friendly tool for networks of BGP routers, i.e.,
BGP networks. SQUAB can configure BGP networks automatically based on
only an easy interface written by a user. SQUAB makes use of virtual contain-
ers, making it lightweight, i.e., it can be executed by a typical laptop computer.
SQUAB enables users to set up their experimental environment locally, and
thus users will not need to register and pay for a service, which are necessary in
conventional testbeds.
As a preferable use case, SQUAB can be used as the first step in the eval-
uation of protocols under development or that will be developed by a user as
a proof-of-concept. Evaluating the design of a new protocol in a testbed often
requires extra overhead due to additional processes, e.g., user registration and
setting up a server, as described above. In contrast, after developing a software
router of their designed protocol, users can use SQUAB to evaluate their pro-
tocol for any network configuration scale by introducing a container image of
the router. We demonstrate that SQUAB can configure networks mixing BGP
and BGPsec [15], which provides the validity of route information by using dig-
ital signatures for BGP, through a reference implementation for BGPsec. We
also discuss further insights about use cases and the limitations of SQUAB. We
released SQUAB and the reference implementation described above publicly1 .
In summary, the contributions of this paper are as follows:
• Design and implementation of SQUAB, a tool that allows users to perform
experiments on BGP easily.
• Public release of SQUAB and BGPsec reference implementation.

2 Research Motivation

In this section, we first describe BGP and define the requirements of SQUAB.
We then describe the technical difficulty in satisfying the requirements.

1
https://github.com/han9umeda/SQUAB.
602 N. Umeda et al.

2.1 Border Gateway Protocol

In BGP, ASes exchange the route information, which includes IP prefixes and
routes to a destination, i.e., AS PATH, with their neighbors. BGP routers then
decide the best routes to each IP prefix in accordance with the received route
information and the static policy defined locally by the operators if there are
multiple routes for the IP prefix. In doing so, the ASes append their own AS
numbers to AS PATH and advertise the IP prefix and the AS PATH to the
neighbors as the best route. As a result, chains of AS PATH to reach each IP
prefix are configured. BGP requires operators to prepare various configuration
files, e.g., information on the neighbors, policies in transmission, and reception
of route information between ASes. Such a complicated setting may lead even
experts to misconfigure BGP settings. Meanwhile, operators of each AS just
prepare for its configuration file, and thus BGP has no function to manage the
whole Internet from a higher perspective. BGP cannot guarantee the validity
of the route information nor detect incorrect route advertisement, i.e., miscon-
figuration or route hijacking. Consequently, despite the development of many
BGP extensions, it remains unclear if these extensions are operable on network
topologies in the real world.

2.2 Goal

Our goal is to develop a tool that allows users to flexibly evaluate the validity
of their protocol design. Such a tool should have the following properties:

• Automated configuration: Users only need to use one command to set up an

experimental environment for BGP. A BGP configuration is difficult to set
up, and thus an automated configuration is useful for all users.
• Lightweight: Execution time is short enough, and memory requirement is low
enough for a common laptop computer.
• Local deployment: Experimental environments are deployed locally. In doing
so, a tool should be stand-alone, and any other resources, such as a server,
for experiments, are unnecessary. Consequently, users will not need to register
and pay for a service.

When designing an experimental tool, its safety and extendability should be

taken into account. Hereafter, safety means never disrupting Internet borders,
and extendability means that a new protocol can be quickly introduced into the
tool as long as an implementation of the protocol is available.
Meanwhile, we assume the following conditions for designing a tool that
satisfies the three properties described above. First, routers are required to be
software routers that can be deployed on the tool to be developed. In particular,
we focus on Quagga Routing Suite [7] and its extensions in the current speci-
fication (see Sect. 3.3 for details). Second, software routers are implemented in
advance, i.e., the development of the software routers themselves is out of the
scope of this paper.
 SQUAB: A Virtualized Infrastructure for Experiments on BGP 603

2.3 Technical Difficulty

Achieving the goals described in the previous subsection is non-trivial. In general,

a configuration file in the conventional BGP defines a setting for some ASes. To
configure experimental environments at the network-level, a user often needs to
prepare all configuration files to make them consistent. This is different from the
technology of designing a single router of observation point [24]. As described in
Sect. 1, operating BGP precisely is often challenging even for experts [12,16].
To avoid the challenge above, existing works have utilized testbeds that
can support network configurations in their experiments. However, the use of
testbeds inherits additional overhead. Experimental environments may be set up
locally using numerous virtual machines, but such an approach often requires a
user to have knowledge about virtual machines and network topologies. Thus,
the use of testbeds does not achieve the lightweight and local deployment
properties.

3 Design of SQUAB

In this section, we present SQUAB (Scalable QUagga-based Automated Con-

figuration on Bgp). We first describe the design concept of SQUAB and then
explain the modules used as building blocks of SQUAB. Finally, we present the
interface of SQUAB, which corresponds to our central idea, and the workflow.

3.1 Design Concept

Our design concept is to virtually configure experimental networks consisting of

virtualized containers for BGP software routers through a user-friendly interface.
Loosely speaking, a user specifies a configuration from three points of view:
the existence of an AS, the function for each BGP router, and the connections
between ASes. After a user specifies the three information through the user-
friendly interface, SQUAB will automatically create networks including virtual
containers of routers. SQUAB automatically completes settings, e.g., such that
BGP routers connect with each other, and thus the user can begin experiments
immediately. Therefore, SQUAB achieves the automated configuration property.
Moreover, the user can execute the automated configuration described above on
a computer locally, and thus SQUAB achieves the local deployment property.
SQUAB achieves the lightweight property by combining virtual containers to
construct an experimental environment. In a conventional experimental environ-
ment, virtual machines that virtualize not only applications but also hardware,
such as CPU and memory, are utilized. However, a virtual machine also virtu-
alizes functions that are not necessary for the experiment, thus increasing the
load. On the other hand, by utilizing a virtual container, hardware virtualization
becomes unnecessary, and then only the application level, i.e., the routers are
virtualized. Consequently, SQUAB drastically reduces the overhead caused by
the experimental environment.
604 N. Umeda et al.

3.2 Modules and Architecture

SQUAB utilizes virtual containers and software routers as elemental technolo-
gies. A virtual container is a virtual environment that packages the application
execution environment and isolates resources, such as processes and networks,
from the host OS. As mentioned above, virtual containers virtualize only at the
application layer in comparison with a conventional virtual machine which virtu-
alizes from the hardware layer. Virtual containers make applications run faster,
and they use fewer resources than virtual machines. Moreover, by virtualizing
the application itself, it is possible to provide an execution environment without
depending on the host environment. Docker2 is a known platform for virtual
containers [19].
A software router is a software that can provide the functions of a router on
the host OS by packaging the network router. Quagga and BIRD [1] are known
open-source BGP software routers used in network operations. There are many
extensions of these software routers that support BGP extension protocol, e.g.,
BGP-SRx [5] for BGPsec.

Fig. 1. The architecture of SQUAB. SQUAB executes Docker to create a network

consisting of Docker containers, which include Quagga routers.

Figure 1 shows the architecture of SQUAB and its position. SQUAB receives
configuration files and commands, and executions Docker. SQUAB runs each
module in a coordinated manner.
We provide the behavior of each module in detail in the next section.

3.3 Methodology
We describe the design of SQUAB in detail. In the following, we describe
the methodology using Quagga as a software router and Docker as a virtual
container.
2
https://www.docker.com/.
 SQUAB: A Virtualized Infrastructure for Experiments on BGP 605

Fig. 2. Example of a network topology to be evaluated and the SQUAB configuration

file: The left-side shows the topology to be evaluated and the right-side shows an
example of the SQUAB configuration file to set up the topology.

3.3.1 Setup of Network Configuration

Figure 2 shows the topology and SQUAB configuration file which the user gives
as input to SQUAB. The users manually create the file corresponding to the
topology they want to configure. The SQUAB configuration file is written in the
YAML format and can be rewritten to change the experimental patterns. In the
hash nest of the AS Setting key, the user inputs the AS numbers and docker
container image names each AS uses as a router, e.g., Quagga. In the hash nest
of the Peer info key, the user inputs the connection information for each AS.

Fig. 3. The workflow of SQUAB for creating experimental environments: The workflow
consists of three processes. By writing the SQUAB configuration file shown in Fig. 2,
SQUAB will create a corresponding docker-compose file and its resulting networks on
Docker.

3.3.2 Workflow
Figure 3 shows the SQUAB environment creation workflow. First, SQUAB reads
the SQUAB configuration file provided as input by a user, and then it generates
the corresponding docker-compose file. The docker-compose file is described so
that the docker engines can interpret the contents similar to the input files. As the
first step of generating a docker-compose file, SQUAB reads all AS information
from the user-input SQUAB configuration file. SQUAB assigns a router-specific
IP address to each AS to avoid duplication, and then it writes to the docker-
compose file to generate the corresponding docker network. Next, SQUAB reads
606 N. Umeda et al.

the peer information from the user-input SQUAB configuration file. SQUAB
creates the docker networks and AS routers for the AS-to-AS connection, and
then it assigns a network IP address to each network to avoid duplication. After
that, SQUAB runs the docker-compose to set up all containers and networks,
and then it creates BGP configuration files by giving arguments to the shell
script which was pre-installed in the container of a router. Finally, SQUAB runs
the routing daemons for each container.
The various experimental data are collected by calling standard UNIX com-
mands or the shell scripts pre-installed in the container of a router. We uti-
lized Python 3.8.5, Quagga 0.99.22.4, and Docker Desktop 3.0.4 to implement
SQUAB.

4 Experiments
We developed SQUAB according to the design described above and used it in
experiments to validate its functions and evaluate its performance. As a case
study, we used SQUAB to configure networks mixing BGP routers and BGPsec
routers.

4.1 Purpose Experiments

We conducted the following three experiments.
Function Validation: We confirm that SQUAB sets up the topology as
defined in the SQUAB configuration file, and then confirm that routers correctly
exchange the route information in the experimental environment.
Performance Evaluation: We measure the running time of SQUAB to set
up and remove the experimental environment. We also measure the memory
consumption of the containers by using SQUAB. Hence, we confirm that SQUAB
is a lightweight tool that requires only machine resources, e.g., a typical laptop
computer.
Case Study: We configure networks such that BGP and BGPsec [15], which is
a security extension of BGP, are mixed, and then check if their routers correctly
exchange route information. Hence, we confirm that it is easy to embed a software
router that supports a BGP extension protocol into SQUAB.

4.2 Experimental Settings

We describe the experimental settings. The experiments were conducted on Mac-
Book Pro 2017(macOS 11.1) with 2.3 GHz Dual-Core Intel Core i5, 8 GB memory.
Function Validation: We validate the functions of SQUAB using an actual
topology centered on NAITWAYS(AS57119). Concretely, we obtain the topology
information from BGPlay3 , an Internet topology visualization tool. Based on the
3
https://stat.ripe.net/widget/bgplay.
 SQUAB: A Virtualized Infrastructure for Experiments on BGP 607

information, we manually prepare the SQUAB configuration file and then input
it into SQUAB. After that, we observe the behavior of SQUAB.
Performance Evaluation: We measure the running time for setting up or
removing the experimental environment using SQUAB by changing the number
of routers placed in a linear topology from 2 to 50. Then, we measure the memory
consumption of all the containers by using the docker stats command. We
measured the memory consumption three times for each pattern and then take
their average.
Case Study: We modify SQUAB to handle BGP-SRx, a software router running
BGPsec. At that time, we checked how much work the modification required. In
addition, we configure networks mixing BGP routers and BGPsec routers, and
then confirm that the routers correctly exchange the route information and that
the function is working correctly as BGPsec.

Fig. 4. The running time of SQUAB: Fig. 5. The memory consumption of

The blue line represents the time for set- SQUAB: The memory consumption of all
ting up an environment and the orange containers to set up the environment is
line represents the time for removing the measured.
environment.

4.3 Results
We show the results of the three experiments describe above.
Functional Validation: We confirm that SQUAB correctly sets up the topol-
ogy centered on NAITWAYS(AS57119) and then records all the details on a
video4 as a demonstration. At the beginning of the video, the topology cen-
tered on NAITWAYS output by BGPlay is displayed. Next, the video shows
the SQUAB configuration file corresponding to the topology at 0:05. SQUAB in
Python script format is run at 0:11. SQUAB then reads the SQUAB configura-
tion file and creates a network and containers of the router. After waiting for all
the routers on the topology exchange route information, the script that collects
the setting information and routing tables of all the routers is run at 2:00, and
4
http://www-infosec.ist.osaka-u.ac.jp/media/aina2021 squab demo.mp4.
608 N. Umeda et al.

then its results are displayed from 2:10 to 3:50. The results show the IP prefixes
and the corresponding AS PATHs, which are consistent with the results of the
exchange of route information in the topology displayed at the beginning of the
video. Thus, we confirm that SQUAB can set up the environment correctly. In
addition, the script that removes the experimental environments is run at 4:00
and the results confirm that SQUAB can remove the environment.
Performance Evaluation: Figure 4 shows the the running time of SQUAB for
setting up and removing each topology. The setting up took 11.9 s for 10 routers
and 52.9 s for 50 routers, showing that the time tends to increase by about one
second per router. On the other hand, the removing took 13.1 s for 10 routes
and 21.4 s for 50 routes, showing that the time tends to increase by about 0.12 s
per router.
Figure 5 shows the memory consumption of all the containers to set up the
environment. The memory requirement of SQUAB is 71.0 MB for 10 routers and
354.7 MB for 50 routers, showing that each router tends to consume 7 MB.
Case Study: We checked the workload for enabling SQUAB to handle BGP-
SRx. First, for embedding in the function to set up the experimental environ-
ment, we needed to add 44 lines of code to the original 326 lines of code, i.e., 370
lines of code totally. We also created a Dockerfile that builds a container image
of a BGP-SRx router and a script to generate both a key and a certificate for
signing and verification. Besides, we modified the script for Quagga, which writes
the BGP configuration file in a router, to correspond with the BGPsec config-
uration file of BGP-SRx. As the script to install BGP-SRx into the image for
building the container image, we adopt the original code of BPG-SRx5 . Mean-
while, certificates into the rpkirtr svr, the RPKI emulator by the BGP-SRx
suite, cannot be registered due to a bug. Thus, we need to operate the container
of rpkirtr svr and then register the certificate manually after setting up the
experimental environment.
We write the SQUAB configuration file corresponding to the topology of
NAITWAYS utilized in the experiments on the functional validation and then
run SQUAB. Figure 6 shows the setting information and routing tables of the
BGP-SRx router. As shown in the figure, the routers correctly exchange the
route information. In addition, we confirm that routers correctly conduct the
verification by BGPsec through the existence of v(v,-), which indicates the
issuance of a certificate.

5
https://github.com/usnistgov/NIST-BGP-SRx.
 SQUAB: A Virtualized Infrastructure for Experiments on BGP 609

Fig. 6. The setting information and routing tables of the BGP-SRx router: From the
routing table of the BGP-SRx router, it can be seen that there is the character of
v(v,-) in the line of the router that issued the certificate.

5 Discussions
In this section, we discuss the performance of SQUAB based on the experimental
results, use cases, and limitations of the current specification.

5.1 Performance

We first discuss the running time of SQUAB. The running time is different
between setting up and removing an environment according to Fig. 4. For setting
up an environment, each container with a router needs to generate the BGP
configuration file and run the daemon. Therefore, the running time for setting
up an environment is longer than that for removing an environment.
610 N. Umeda et al.

Next, the memory consumption of SQUAB is 7 MB per router according to

Fig. 5. This result indicates that, for example, the memory consumption is about
700 MB even when a network topology consisting of 100 routers with 50 inter-
AS connections is set up. We thus believe that SQUAB enables a typical laptop
computer to run experiments even with medium-scale topologies.

5.2 Use Cases

SQUAB is applicable to the existing protocols for BGP extensions. We describe

D-BGP [23] and APVAS [14] as use cases of SQUAB.
First, we describe the application to D-BGP, a general-purpose protocol for
BGP extensions that supports them on the same network. D-BGP has a func-
tion to tunnel routers that have no support for the extensions. Although the
function of D-BGP has not been evaluated on the network composed of only D-
BGP routers, SQUAB can provide networks composed of various routers, which
support only BGP extensions except for D-BGP. A prototype implementation
of D-BGP is publicly available6 . It is implemented as an extension of Quagga,
and hence, by making a container image composed of the prototype of D-BGP
and existing configure file generation script, SQUAB can deal with D-BGP.
Next, we describe the application to APVAS, a protocol for BGPsec that
reduces the memory consumption of routers more than state-of-the-art proto-
cols by introducing aggregate signatures [10], which can aggregate individually
generated signatures of BGPsec into a single short signature. Although the per-
formance of APVAS has not been evaluated on the actual topology, SQUAB
can potentially provide the evaluation. The latest APVAS prototype is imple-
mented as an extension of BIRD, and a user can use SQUAB to evaluate the
performance of APVAS by modifying the syntax specification of the BGP con-
figuration file from Quagga to BIRD. In this way, SQUAB can be applicable for
software routers other than Quagga.

5.3 Limitations

We describe several limitations of the current specification of SQUAB. First, the

current specification does not support configuration of AS policy, e.g., the local-
preference [13] to override the AS path length metric of the base BGP operation
and the multi exit discriminator (MED) attributes [18] to provides a mechanism
for BGP speakers to convey to an adjacent AS the optimal entry point into the
local AS.
Second, network topologies generated by SQUAB are limited. Concretely,
the networks are composed of only AS border routers. In other words, SQUAB
cannot represent components of an intra-AS to configure and maintain routing
tables within a single AS, e.g., the network latency due to complicated networks
or random errors inside the AS. To introduce the functions described above,

6
https://github.com/RS1999ent/quagga evolvabilty.
 SQUAB: A Virtualized Infrastructure for Experiments on BGP 611

implementation to simulate intra-AS is necessary. Likewise, the latency on inter-

AS communication should be considered to represent networks in the real world
further.
Finally, if users want to conduct an experimental evaluation of new protocols,
they need to develop their software routers in advance. Moreover, the scripts
of SQUAB to set up routers may need updates. However, the automation of
developing software routers and updating the scripts is out of the scope of the
current specification. These topics are reserved for future work.

6 Related Works
In this section, we briefly review related works in terms of experimental tools
for BGP, testbed developments, and BGP security issues in the past decades.
Experimental Tools for BGP: Although there are many tools for measuring
empirical values on BGP [11,17,21,26], they cannot interact with the existing
ecosystem. This means that evaluation, including the current policy and connec-
tivity of actual ASes, is limited. In contrast, PEERING [24] is the first tool that
takes into account ecosystems in the real world. However, PEERING special-
izes in experiments concerning borders and is thus unsuitable for experiments
to measure each AS’s connectivity on the Internet.
Testbed Developments: Many existing testbed environments, such as Cloud-
Lab [2], XSEDE [9], and Emulab [4], provide hardware as a computer resource
in contrast to SQUAB. More specifically, Emulab [4] provides the testbed envi-
ronments specialized for network experiments, and also there are many Emulab-
based tools by customizing to each application such as EPIC [25] and DETER [3].
Emulab provides not only computer resources but also network connections rep-
resenting network delays as experimental environments. However, the use of
Emulab is restricted to a responsible project leader, e.g., a professor or a person
who has permission from the leader, and this approval may unexpectedly con-
sume time. PlanetLab [6] and WIDE [8] are projects where participants bring
machine resources and then configure networks to share an experimental environ-
ment. However, the participants need to meet several conditions, such as provid-
ing two Linux machines with global IP addresses. Compared with the testbeds
mentioned above, SQUAB is based on containers whose cost is lightweight and
thus it allows experiments on only a typical laptop computer without any third-
party resources. Therefore, anyone can use SQUAB to quickly start experiments.
The closest tool to SQUAB is DOCKEMU [27], which uses Docker to set up
experimental network environments. However, DOCKEMU is based on NS-3, a
network simulation tool, while SQUAB provides network emulation by virtue of
Docker networks.

7 Conclusion
In this paper, we proposed SQUAB, a lightweight evaluation tool for experi-
ments on BGP and its extensions. We publicly released our implementation of
612 N. Umeda et al.

SQUAB, including its reference implementation for BGPsec. In the design of a

new protocol, the effectiveness of the protocol should be evaluated through vari-
ous experiments, and SQUAB makes such evaluation easy. We thus believe that
SQUAB will assist researchers in developing future BGP extensions and their
social implementation. We are in the process of introducing AS policy support
such as the local-preference attributes and the multi exit discriminator (MED)
attributes.

Acknowledgments. This research was supported in part by the Japan Society for
the Promotion of Science KAKENHI Numbers 18K18049, and Innovation Platform for
Society 5.0 at MEXT.

References
1. The bird internet routing daemon. https://bird.network.cz/
2. Cloudlab https://www.cloudlab.us/
3. The deter project. https://deter-project.org/
4. emulab. https://www.emulab.net/portal/frontpage.php
5. Nist bgp secure routing extension (bgp / srx) prototype https://www.nist.gov/
services-resources/software/bgp-secure-routing-extension-bgp-srx-prototype
6. Planetlab https://planetlab.cs.princeton.edu/
7. Quagga routing suite. https://www.nongnu.org/quagga/
8. Wide. https://www.wide.ad.jp/
9. Xsede. https://www.xsede.org/
10. Boneh, D., Gentry, C., Lynn, B., Shacham, H.: Aggregate and verifiably encrypted
signatures from bilinear maps. In: Proceedings of EUROCRYPT 2003, pp. 416–432
(2003)
11. Chen, K., Choffnes, D.R., Potharaju, R., Chen, Y., Bustamante, F.E., Pei, D.,
Zhao, Y.: Where the sidewalk ends: extending the internet as graph using tracer-
outes from p2p users. In: Proceedings of the 5th International Conference on
Emerging Networking Experiments and Technologies, pp. 217–228 (2009)
12. Fogel, A., Fung, S., Pedrosa, L., Walraed-Sullivan, M., Govindan, R., Mahajan, R.,
Millstein, T.: A general approach to network configuration analysis. In: Proceedings
of NSDI 2015, pp. 469–483. USENIX Association (2015)
13. Griffin, T., Huston, G.: Bgp wedgies. RFC 4264 (2005). https://rfc-editor.org/rfc/
rfc4264.txt, published: RFC 4264
14. Junjie, O., Yanai, N., Takemura, T., Okada, M., Okamura, S., Cruz, J.P.: APVAS:
reducing memory size of as path validation by using aggregate signatures (2020).
CoRR abs/2008.13346, https://arxiv.org/abs/2008.13346
15. Lepinski, M., Sriram, K.: BGPsec Protocol Specification (2017). https://doi.org/
10.17487/RFC8205, https://rfc-editor.org/rfc/rfc8205.txt, published: RFC 8205
16. Mahajan, R., Wetherall, D., Anderson, T.: Understanding BGP misconfiguration.
In: Proceedings of SIGCOMM 2002, pp. 3–16. ACM (2002)
17. Mao, Z.M., Rexford, J., Wang, J., Katz, R.H.: Towards an accurate as-level tracer-
oute tool. In: Proceedings of the 2003 conference on Applications, Technologies,
Architectures, and Protocols for Computer Communications, pp. 365–378 (2003)
18. McPherson, D.R., Gill, V.: BGP MULTI EXIT DISC (MED) Considerations. RFC
4451 (2006). https://www.rfc-editor.org/rfc/rfc4451.txt, published: RFC 4451
 SQUAB: A Virtualized Infrastructure for Experiments on BGP 613

19. Merkel, D.: Docker: lightweight linux containers for consistent development and
deployment. Linux J. 2014(239), 2 (2014)
20. Mitseva, A., Panchenko, A., Engel, T.: The state of affairs in BGP security: a
survey of attacks and defenses. Comput. Commun. 124(2018), 45–60 (2018)
21. Peterson, L., Bavier, A., Fiuczynski, M.E., Muir, S.: Experiences building plan-
etlab. In: Proceedings of the 7th Symposium on Operating Systems Design and
Implementation, pp. 351–366 (2006)
22. Rekhter, Y., Hares, S., Li, T.: A Border Gateway Protocol 4 (BGP-4)
(2006). https://doi.org/10.17487/RFC4271, https://rfc-editor.org/rfc/rfc4271.txt,
published: RFC 4271
23. Sambasivan, R.R., Tran-Lam, D., Akella, A., Steenkiste, P.: Bootstrapping evolv-
ability for inter-domain routing with d-bgp. In: Proceedings of SIGCOMM 2017,
pp. 474–487. ACM (2017)
24. Schlinker, B., Arnold, T., Cunha, I., Katz-Bassett, E.: Peering: virtualizing BGP
at the edge for research. In: Proceedings of CoNEXT 2019, pp. 51–67. ACM (2019)
25. Siaterlis, C., Genge, B., Hohenadel, M.: Epic: a testbed for scientifically rigor-
ous cyber-physical security experimentation. IEEE Trans. Emerg. Topics Comput.
1(2), 319–330 (2013)
26. Staff, R.N.: Ripe atlas: a global internet measurement network. Internet Protoc.
J. 18(3), 1–4 (2015)
27. To, M.A., Cano, M., Biba, P.: Dockemu–a network emulation tool. In: 2015 IEEE
29th International Conference on Advanced Information Networking and Applica-
tions Workshops, pp. 593–598. IEEE (2015)
28. Vervier, P., Thonnard, O., Dacier, M.: Mind your blocks: on the steal thiness of
malicious BGP hijacks. In: Proceedings of NDSS 2015. Internet Society (2015)
 Scalability Support with Future Internet
in Mobile Crowdsourcing Systems

Peron Sousa and Antonio A. de A. Rocha(B)

Institute of Computing, Fluminense Federal University, Rio de Janeiro, Brazil

peron [email protected], [email protected]

Abstract. The popularization of mobile devices with multiple

resources, the real-time location, and the availability of “human com-
puting”, provided by a global crowd of users, motivated the emergence
of Mobile Crowdsourcing Systems. This new paradigm presents several
requirements such as assertiveness in task assignment, security, mobility
support and scalability. In this paper, scalability is the ability to meet
increasing demand without performance loss, a point ignored in many
proposals that make use of the traditional model (client-server). The
solution of this problem is also the main focus of this work. To address
the issue we review the state of the art and analyze where each proposal
failed. We designed a new architecture and worked on its implementation
to carry out a proof of concept that confirmed its feasibility. Experiments
evaluated the performance of the components that make up the solution
and the results showed that the architecture offers a reduction in latency
between 26.31% and 32.03%. Another interesting point is that the new
architecture, unlike the state of the art, has no overhead and no need to
manage dozens of machines around the world. It was also the first with
mobility support for Mobile Crowdsourcing Systems.

1 Introduction
Mobile Crowdsourcing System (MCS) is a new paradigm that seeks to take
advantage of an already installed base of devices with multiple features to locate
workers and assign them tasks that have restrictions of time and space. Moreover,
these tasks require human cognition to be performed quickly, cheaply, and with
greater efficiency.
An example of an MCS request is the location of parking spaces. People can
easily identify available spots and report with images and/or text messages. If
we were to use an autonomous system based on ultrasound to find it, not only
would we need special hardware, but also sophisticated algorithms to ensure
data reliability [1].
New challenges have arisen with the advent of these systems and scalability
is one of them, should MCS intend to support a multitude of users scattered
around the globe, who are in constant movement. In our proposal, scalability is
the ability to meet growing demand without loss of performance; a point ignored
in many proposals which make use of the traditional model (client-server). In

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 614–625, 2021.
https://doi.org/10.1007/978-3-030-75100-5_53
 Scalability Support with Future Internet in Mobile Crowdsourcing Systems 615

this model, the backend is responsible for registering users, distributing tasks,
receiving responses and, in some cases, consolidating information. However, this
type of system does not scale, losing performance, and may stop when a large
number of users and tasks arrive in a short time. In the state-of-the-art, we
find solutions involving load balancing with active replication of servers, proxy,
Software Defined Networking (SDN), elastic computing, edge computing, and
Device to Device (D2D). However, they have limitations [2].
Our goal is to demonstrate how our architecture, called Snow Flurry (inspired
by the eponymous mobiles of sculptor Alexander Calder), solves the issue of
scalability more efficiently than state-of-the-art and with mobility support. In
this work, the efficiency is measured by latency and proved by experiment com-
parative with another solution. To achieve the objectives, we used the Global
Name Service (GNS) of the MobilityFirst architecture. This GNS (also known as
Auspice) enables a Globally Unique Identifier (GUID), mobility, and contextual
search [3,4]. Because it is able to address the issue of mobility and most of the
scalability issues involved in MCS architectures.
With the realization of the proof of concept (PoC) we verified the viability
of our proposal. With performance tests, volume, and infrastructure analyzes
we show: that the architecture offers a reduction in latency between 26.31% and
32.03% when compared to the state-of-the-art; dispenses the monitoring and
maintenance of a complex infrastructure with approximately 200 hosts. These
contributions occur along with meeting the requirements for a scalable architec-
ture. In addition, it is also the first, to the best of our knowledge, to support
mobility.

2 Snow Flurry’s Proposal and Implementation

In this section we will introduce the proposed Snow Flurry architecture through
each of its elements; this project aims to scale and support mobility for crowd-
sourcing systems. Then, we will introduce the implementation and explain how
the Snow Flurry architecture works. We will also show the role of each compo-
nent throughout the process and discuss commonalities with other MCS.

2.1 Proposed Architecture

Figure 1 presents an overview of the Snow Flurry architecture. In it, the optional
elements are represented by dashed edges surrounding the Content Distribution
Network (CDN). Therefore, the mandatory components are: the notification ser-
vice, the backend and the mobile device.
Notification Service: We designed Snow Flurry to work with a notification
service that uses the push method. These services allow you to send messages
to the user without the need to take the initiative to check for new messages.
Also, the push method increases the number of tasks completed and reduces
overhead [5].
616 P. Sousa and A. A. A. Rocha

Fig. 1. Snow flurry architecture

Backend: The backend is the component responsible for keeping the workers’
information, from the register (with name, email, geolocation, etc.) to the result
of their work in some scenarios. This item becomes critical when we think of a
global reach crowdsourcing system that should support millions of users around
the world. Using a client/server architecture provided by a supercomputer would
result in a solution with a critical point of failure and varying levels of quality of
service, as users’ perceived latency would vary with each other’s geographic loca-
tion, that is, the further away from the server, the worse would be the application
response time. One solution to these points is to distribute servers geographically.
Geographically Distributed Servers solves two issues: (1) prevents the system
from having a critical point of failure and (2) reduces latency, improving the
quality of service offered to the user. Consensus Algorithm is needed to be used
for maintaining replicated data consistent. Whenever a change occurs on one
server, it is necessary to write that change on others. Yet, simultaneous changes
 Scalability Support with Future Internet in Mobile Crowdsourcing Systems 617

can occur on different servers and a decision is needed to define which change
will be considered.
Distributed Database Search algorithm is necessary to locate data across mul-
tiple servers. It has some challenges. In our specific cases, one of them is join
operations between distributed tables. An alternative is the use of document-
oriented databases (a type of NoSQL). NoSQL Servers (with document-oriented
databases) solve the problem of interactions by avoiding them through denor-
malization, yet they are possible by using key-value, column family, and graphs.
It means a more dynamic structure (such as JSon) that can be used for context
definition, such as saying that a crowdsourcing worker is a taxi driver, photogra-
pher, or even both. This brings us to another need for Snow Flurry architecture,
Context Support, which jointly with NoSQL databases can provide a new struc-
ture aligned with system needs whenever changes prove necessary. Although
most modern NoSQL databases already offers, specific Geolocation Algorithms
still need to be produced for each emerging scenario.
Finally, Mobility Support here is the ability of the device to change identifiers
without losing established connections, i.e., being able to change IP and, conse-
quently, network without having to restart an operation (such as a download in
progress). Therefore, we understand that this support must allow both worker
and requester to be in transit while performing tasks and transmitting responses
in mobile crowdsourcing systems.
Mobile Device: While in crowd sensing systems smartphones are more
requested due to low user interaction and the multiple features offered by these
devices (such as GPS, accelerometer and cameras), in mobile crowdsourcing sys-
tems we may consider other types of devices, (such as, tablets and notebooks)
because the greater interest is in the cognitive power of the user. But, Resources
for Attendance of Tasks may also be useful. Ideally, the worker’s device has GPS,
accelerometer, cameras, among other features to perform tasks (which directs,
but does not limit the system for smartphone users).
In addition to enabling campaign launching, the in-app requester interface
also brings with it a strategy that enables scalability in mobile crowdsourcing sys-
tems, which is the distribution of the task assignment algorithm. By embedding
this algorithm in the app, we remove from the backend the burden of centraliz-
ing the task assignment process and all the computational effort involved in this
event, diluting this burden between requesters’ devices. The app needs to have
a lightweight local relational database (SQL local DB ) to enable the tracking of
tasks, both from the point of view of workers and the management of requesters.
This information only exists on the end devices, i.e., it will not be kept at the
backend of the system.
From a security perspective, one of the safest communication models that
exists is the combined use of symmetric with asymmetric cryptography algo-
rithms. Thus, the security module in the app must enable the creation of asym-
metric cryptographic keys and user registration at the backend with the con-
sequent submission of the public key created for it. Afterwards, updates to the
base must occur through commands signed by some symmetric cryptography
618 P. Sousa and A. A. A. Rocha

algorithm, this signature must be encrypted with the private key generated by
the asymmetric cryptographic algorithm before sending with the command to
the backend.
Finally, at bottom left part of Fig. 1, we can see represented the Direct Com-
munication between Devices (DCD), when the worker’s mobile device communi-
cates directly with the requester’s mobile device, so the arrows exit and return to
the mobile device component. And, at the left side of Fig. 1, the dotted arrows
connecting the mobile device to the content distribution network component
represent a direct communication between a requester using a web interface and
a worker using the app.

2.2 Implementation

The Snow Flurry implementation consists of 3 main components: an Android

App, GNS Auspice and Firebase Cloud Messaging. The interface with MCS can
be made exclusively by the App we developed for Android devices, i.e., serve
both workers and requesters. We will call workers the users who respond to the
tasks of the requesters. Therefore, the requesters are the users who create the
tasks in the MCS. We will call campaign the creation of a task (or set of related
tasks) and its distribution to more than one worker. Optionally, the architecture
can rely on an Internet page composed of HTML/JavaScript and distributed by
a global CDN service.
As mentioned, the second element is a GNS. For this work we chose the
Auspice [4], because its project considered the popularization of mobile devices
and prepared for such event. It is able to keep billions of records (user accounts)
and support millions of queries/updates per day, sent from all over the world [3].
The Android application enables the registration in Auspice and the data update.
Among these data we highlight the geolocation, the context information and the
token of the notification service. This communication is illustrated by the number
one in Fig. 2. Unlike Peng et al. [6], in which the backend is updated every 2 min,
our proposal only forwards an update to the GNS if a change occurs.
In Fig. 2, the number two illustrates the search for workers by the requester.
At this point, we send the geolocation of the point of interest (PoI) within a
radius of 1 km (by default). The PoI is obtained when the map is opened and
changes as we move it, see an example in Fig. 3. The frontend receives the data
and presents them, redoing this process every 5 s (by default).
Figure 3 illustrates the user interfaces of the worker. It allows to draw vector
shapes on the map. On the screen we see five workers, three with the blue
background, one green and one red. Each one defines its own radius of action and
informs this to Auspice. The requester views the derived perimeter by clicking
on the marker. In the example of Fig. 3, compared to the green worker, the blue
collaborator serves a larger area.
 Scalability Support with Future Internet in Mobile Crowdsourcing Systems 619

Fig. 2. Snow Flurry operation, from task assignment to results reception

In the process, once the location has been chosen, it is defined the task with
a given context and deadline. In this example we indicate the context (color)
and add the necessary information to the task. In the sequence, the frontend
identifies the workers who serve the context (color) and assigns them the task
through Firebase Cloud Message, circle number three from Fig. 2. Then, the
event indicated by circle number four of Fig. 2 occurs. In it, Firebase delivers
the notification to the device. This delivery uses the token that the machine
received from Firebase. This token is persisted in the GNS by the app and
collected by the requester, circle one and two of Fig. 2.
The GNS can store tokens from other notification services, such as the Apple
Push Notification Service (APNS) and the Microsoft Push Notification Service
(MPNS). We opted for Firebase, due to its integration with the Android Studio
development environment, but there are other ways the worker can access the
tasks. Some solutions use the pull method, in which the worker searches for
their existence. Firebase works with push, in this method the task is sent to
the worker. According to Kandappu et al. [5], push increases the number of
completed tasks and reduces the overhead.
Tasks are routed preferably by notification services. In these services, when
the worker is disconnected, the tasks/notifications are stored for a period. After
this wait, without successful delivery, the notification is discarded. If the worker
informed the token that identifies him in these services to Auspice, the appli-
cation will send notifications through them. That way, if the user leaves their
device unplugged for a long time (vacation, for example), they will not receive
old notifications that have missed the requester’s deadline. We can also notify
workers through the Android app. In Fig. 3, we see how the solution looks in the
application during the worker search step and view of their radius, as seen on
page to the Internet. Similarly, we can view the opening of a task through the
application.
620 P. Sousa and A. A. A. Rocha

Fig. 3. User interfaces of the worker

Circle number one in Fig. 2 also represents the completion of the task, at
this time the worker sends the data and/or files to the Auspice. In its structure,
Auspice uses a document-oriented database (MongoDB). Due to the way it was
implemented, we can use the same function used to create a context for storing
data in JSon format. After the deadline, the requester collects the responses, as
shown by circle number two in Fig. 2. In this way, when using the Auspice, we
are (1) able to focus on the development of the functionalities of the MCS, (2)
save lines of code, (3) reduce environment settings, and (4) we do not need to
monitor and maintain replicas. Our proposal uses the Spark Plan (free), its has
no limits for Cloud Messaging.
For proof of concept we have developed the two types of features that con-
sume GNS services, the web page and the app for Android devices. Requesters
can use both the web page and the application, while workers only the applica-
tion. The Android application allows registration in Auspice and data update,
among which we highlight the geolocation, context information and notification
service token. This communication is illustrated by the number one circle in
Fig. 2. Unlike Peng et al. [6], which updates the rear every 2 min, Snow Flurry
only forwards an update to GNS if a change occurs.
 Scalability Support with Future Internet in Mobile Crowdsourcing Systems 621

The DCD was not implemented in PoC. In direct communication, due to the
mobility of the devices involved, there is the possibility that one of the tips or
both of them change network, exchanging IP. To avoid loss of communication we
can use MSocket. It is a Java library (analogous to the conventional Socket) that
allows the mobility of both ends without loss of connection. This solution is part
of the research on the MobilityFirst architecture and works together with Aus-
pice, using cryptographic keys in communication. Apart from some additional
code adaptations, we can simply exchange Socket and ServerSocket for MSocket
and MServerSocket, respectively. This way, it is possible to provide mobility
to an application that works with a TCP communication without relying on a
DNS [7]. We don’t use this library in implementing Snow Flurry, because it still
needs some adaptations to run perfectly on the Android environment. Mobility
support was provided by an Auspice API that is consumed by HTTP requests.

3 Comparing Snow Flurry to the State-of-the-Art

In this section we briefly present the state-of-the-art in scalability for crowdsourc-

ing and compare it to Snow Flurry proposal. In the proposal of Yan et al. [8],
mCrowd is presented, which is a mobile crowdsourcing platform acting as a proxy
between users (requesters and workers) and the platforms Amazon Mechanical
Turk (MTurk) and ChaCha. By analyzing the proposal, we can conclude that
mCrowd must face a similar problem to that found in the client-server model.
Even if partner platforms are able to ensure scalability, mCrowd may represent
a bottleneck between them and users. To avoid this problem, Hara et al. [9] pro-
posed a decentralized system, where special servers, known as Server Access Net-
work Entities (SANE), are distributed geographically to dilute the load. Auspice
can perform this entire operation for our proposal in a similar way, having several
servers and positioning the information geographically close to those interested.
SANE architecture uses Distributed Hash Table (DHT) to organize its com-
ponents and the work presents an implementation with a relational database,
MySQL. As the architecture has a very specific use, it acts as a proxy between
crowdsourcing clients and crowdfunding servers, the relational database model is
sufficient. In addition to not being tied to a specific use, the Snow Flurry architec-
ture also has algorithms for the proper positioning of information. Auspice uses
an algorithm called Gigapaxos [10] that has the ability to quickly manage and
reconfigure a very large number of replica groups. Li et al. [11] conceptualized
a decentralized blockchain-based framework for crowdsourcing called CrowdBC.
The idea is not to depend on any trusted third party, to guarantee user privacy
and low transaction fees.
In the survey of Xiao et al. [2], we find a study on scalability for mobile
crowd sensing (application similar to crowdsourcing, but with little or no human
interaction, and it focuses on the collection of sensor data). One of those systems
is proposed by Marjanović et al. [12] and uses edge computing for crowd sensing
application. Basically, this work advocates a cloud architecture with intermediate
servers (edge) closer to users to reduce latency. Our proposal already does this,
622 P. Sousa and A. A. A. Rocha

since Auspice has servers spread around the world that organize the data in a
way that leaves them geographically close to users. Marjanović’s et al. research
was the one that focused on scalability the most, but the work concentrated
efforts on finding out how many intermediate servers would be needed to make
their proposal viable and confined themselves to saying that edge computing can
provide ultra-low latency (less than 1 ms), without presenting evaluations.
Medusa, by Ra et al. [13], is another a crowd sensing solution that offers ser-
vices with a high level of abstraction in specifying the steps required to complete
a detection task and which employs a distributed run-time system to coordinate
the execution of these tasks between smartphones and a cluster in the cloud. It
works similarly to mCrowd of Yan et al. [8], as it also intermediates with MTurk.
However, Ra et al. [13] states that Medusa needs not to be restricted to a single
machine and that it can work with as many logically independent simultaneous
servers as needed. The work presents an implementation with MySQL without
giving further details on how the application handles database growth.
Table 1 summarises a comparison among state of the art architectures, includ-
ing the Snow Flurry proposal. From the table, we can note that Snow Flurry
covers most of the aspects that other architectures do not, except for the item
“Anonymity” which will be the subject of future work.

Table 1. Comparison between architectures

Structure (NoSQL)

Application and/or
Elastic Computing
Centered Element

Mobility Support
Task Assignment

Context Search
Geographically

Database Full
Cloud and/or
Push Method

Source Code
Distribution

Distribution
Geolocation
Information
Notification
Distributed

Distributed

Distributed
Replication

Encryption
Anonymity
Algorithm

Available
Database
Search in

Dynamic
Network
Content

Method
Servers

Geocrowd [14] ✓ - ✗ ✗ ✗ - - - - ✓ - - - - - ✗
gMission [15] ✓ - ✗ ✗ ✗ - - - - ✓ - - - - - ✗
CrowdBC [11] ✗ - ✗ - ✗ ✓ ✓ ✓ - - ✗ - - ✓ ✓ ✗
mCrowd [8] ✓ - ✗ - ✓ ✗ - - - - - - - - - ✗
SANE [9] - - - - - ✓ ✓ ✗ - - - - - - - ✗
Edge [12] ✗ ✗ - - ✓ ✓ ✓ ✗ - ✓ ✗ - - - - ✗
Medusa [13] ✓ ✗ ✗ - ✓ - ✓ - - ✓ ✗ - ✗ - - ✓
Snow flurry ✗ ✓ ✓ ✓ ✗ ✓ ✗ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✗ ✓

4 Experiment

In this section we will describe the scenario, the metric used in the comparison
and how the values were obtained. We will present the results obtained in the
tests, discuss implementation points that affect the performance and to make
considerations about the complexity and costs involved in a global infrastructure.
 Scalability Support with Future Internet in Mobile Crowdsourcing Systems 623

Due to the unavailability of MCS, we will present a comparison of Snow Flurry

implementation against the Medusa by Ra et al. [13], a mobile crowd sensing
system.
To evaluate the performance of a single instance of GNS, we used a Linode
machine, consisting of 1 CPU core with 64-bit, running Ubuntu 16.04 operat-
ing system, 2 GB RAM, 30 GB SSD storage, 2 TB transfer, 40 Gbps input,
1000 Mbps output and located in Newark, NJ, USA. We installed Java version
1.8.0 151, the GNS and we carried a load of 5000 workers in the system.
To evaluate the performance of our solution we will use as latency metric
on each type of communication. To rule out the delay due to the network, the
requests were triggered from the machine itself where the implementation was.
We performed 30 repetitions in each case to obtain the boxplots shown in Fig. 4.

Fig. 4. Performance of requests without network delay.

We can see that actual overhead costs are only 34.47 ms on average for
Medusa, which allows for a yield of 1740 instances of tasks per minute on a single
server. The creators of Medusa claim that this component is highly parallelizable
since each task instance is relatively independent and can be assigned a separate
Task Controller. Thus, if the task transfer rate becomes a problem, it would be
possible to leverage the cloud computing elastic resources to scale the system [13].
Following the same principle, we can see that our solution is able to serve
more requests per machine, since the average times of 25.4 and 23.43 ms, enable,
respectively, the attendance of 2362 and 2560 requests per minute. Values rep-
resenting a capacity increase between 35.75% and 47.13%, with a decrease in
latency between 26.31% and 32.03%. If our proposal worked with elastic com-
puting, these data would already indicate a possible savings compared to the
reduction in the provisioning of resources.
624 P. Sousa and A. A. A. Rocha

5 Conclusion
In this paper we discuss the problem of scalability in mobile crowdsourcing sys-
tems and the techniques used in state of the art. We presented a new architecture,
called Snow Flurry, which use a global name service and cloud services for worker
notification. We established a comparative table with the other solutions, which
was possible to realize that the only point of attention of Snow Flurry is the
absence of anonymity. We conducted a proof of concept with the Snow Flurry
implementation and performed comparative experiments with a mobile crowd
sensing system, due to unavailability of mobile crowdsourcing system solutions.
The results presented a reduction in latency between 26.31% and 32.03%.

References
1. Ganti, R.K., Ye, F., Lei, H.: Mobile crowdsensing: current state and future chal-
lenges. IEEE Commun. Mag. 49(11), 32–39 (2011)
2. Xiao, Y., Simoens, P., Pillai, P., Ha, K., Satyanarayanan, M.: Lowering the barriers
to large-scale mobile crowdsensing. In: Proceedings of the 14th Workshop on Mobile
Computing Systems and Applications, p. 9 (2013)
3. Sharma, A., Tie, X., Uppal, H., Venkataramani, A., Westbrook, D., Yadav, A.: A
global name service for a highly mobile internetwork. ACM SIGCOMM Comput.
Commun. Rev. 44(4), 247–258 (2015)
4. Raychaudhuri, D., Nagaraja, K., Venkataramani, A.: Mobilityfirst: a robust and
trustworthy mobility-centric architecture for the future internet. ACM SIGMO-
BILE Mob. Comput. Commun. Rev. 16(3), 2–13 (2012)
5. Kandappu, T., Misra, A., Cheng, S.-F., Jaiman, N., Tandriansyah, R., Chen,
C., Lau, H.C., Chander, D., Dasgupta, K.: Campus-Scale Mobile crowd-tasking:
deployment & behavioral insights. In: Proceedings of the 19th ACM Conference on
Computer-Supported Cooperative Work & Social Computing, pp. 800–812. New
York (2016)
6. Peng, X., Gu, J., Tan, T.H., Sun, J., Yu, Y., Nuseibeh, B., Zhao, W.: Crowdservice:
serving the individuals through mobile crowdsourcing and service composition. In:
31st IEEE/ACM International Conference on Automated Software Engineering,
pp. 214–219 (2016)
7. Bronzino, F., Nagaraja, K., Seskar, I., Raychaudhuri, D.: Network service abstrac-
tions for a mobility-centric future internet architecture. In: Proceedings of the
Eighth ACM International Workshop on Mobility in the Evolving Internet Archi-
tecture, pp. 5–10. ACM (2013)
8. Yan, T., Marzilli, M., Holmes, R., Ganesan, D., Corner, M.: mCrowd: a platform for
mobile crowdsourcing. In: Proceedings of the 7th ACM Conference on Embedded
Networked Sensor Systems, pp. 347–348 (2009)
9. Hara, T., Springer, T., Bombach, G., Schill, A.: Decentralised approach for a
reusable crowdsourcing platform utilising standard web servers. In: Proceedings
of the ACM Conference on Pervasive and Ubiquitous Computing Adjunct Publi-
cation, pp. 1063–1074 (2013)
10. Venkataramni, A., Gao, Z., Gu, T., Anantharamu, K.: Scalable fine-grained recon-
figurable replica coordination. University of Massachusetts (UMass), Technical
Report (2018)
 Scalability Support with Future Internet in Mobile Crowdsourcing Systems 625

11. Li, M., Weng, J., Yang, A., Lu, W., Zhang, Y., Hou, L., Liu, J., Xiang, Y., Deng,
R.H.: Crowdbc: a blockchain-based decentralized framework for crowdsourcing.
IEEE Trans. Parallel Distrib. Syst. 30(6), 1251–1266 (2019)
12. Marjanović, M., Antonić, A., Žarko, I.P.: Edge computing architecture for mobile
crowdsensing. IEEE Access 6, 10:662–10:674 (2018)
13. Ra, M.R., Liu, B., La Porta, T.F., Govindan, R.: Medusa: A programming frame-
work for crowd-sensing applications. In: Proceedings of the 10th International Con-
ference on Mobile Systems, Applications, and Services, pp. 337–350 (2012)
14. Kazemi, L., Shahabi, C.: Geocrowd: enabling query answering with spatial crowd-
sourcing. In: Proceedings of the 20th International Conference on Advances in
Geographic Information Systems, pp. 189–198 (2012)
15. Chen, Z., Fu, R., Zhao, Z., Liu, Z., Xia, L., Chen, L., Cheng, P., Cao, C.C., Tong,
Y., Zhang, C.J.: gmission: a general spatial crowdsourcing platform. Proc. VLDB
Endowment 7(13), 1629–1632 (2014)
 HyDiLLEch: A WSN-Based Distributed
Leak Detection and Localisation in Crude
Oil Pipelines

Safuriyawu Ahmed1(B) , Frédéric Le Mouël1 , Nicolas Stouls1 ,

and Gislain Lipeme Kouyi2
1
Univ Lyon, INSA Lyon, Inria, CITI, EA3720, 69621 Villeurbanne, France
{Safuriyawu.Ahmed,frederic.le-mouel,nicolas.stouls}@insa-lyon.fr
2
Univ Lyon, INSA Lyon, DEEP, EA7429, 69621 Villeurbanne, France
[email protected]

Abstract. One of the major failures attributed to pipeline transporta-

tion of crude oil is oil leakages and spills. Hence, it is monitored via sev-
eral classical leak detection techniques (LDTs), which are more recently
implemented on centralised wireless sensor networks (WSN)-based leak
detection and monitoring systems (LDMS). However, the LDTs are
sometimes prone to high false alarms, and the LDMS are sensitive to
single points of failure. Thus, we propose HyDiLLEch, a distributed
leakage detection and localisation technique based on a fusion of sev-
eral LDTs. In this work, we implemented HyDiLLEch and compared it
to the individual LDTs in terms of communication efficiency and leak-
age detection and localisation accuracy. With HyDiLLEch, the number
of nodes detecting and localising leakages increases by a maximum of
four to six times, thereby eliminating single points of failures. In addi-
tion, we improve the accuracy of localisation in nodes physically-close
to the leak and maintain an average of 96% accuracy with little to no
communication overhead.

1 Introduction
Oil leakages and spills (OLS) in pipelines occur due to ageing infrastructures,
corrosion, and most commonly third-party interference. OLS has both economic
and environmental effects. Some of them include the annual loss of up to 10
billion USD in the United States [1], pollution of water and land resources, fatal
accidents caused by resulting explosions, etc. Therefore, several measures are put
in place to help monitor, detect and predict such failures in a timely manner.
These measures can be broadly categorised into human-based and non-
human-based leakage detection and monitoring systems (LDMS). The human-
based LDMS comprises the use of community-based surveillance, security
personnel, specialised helicopters, etc., for detecting and localising leakages.
Although they work, their drawbacks include long delays, high cost, inefficiency,
etc. This is ascertained with the recorded loss of crude oil from the pipeline net-
work of Shell Nigeria to the tune of 11000 barrels per day in 2018 due to failures,

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 626–637, 2021.
https://doi.org/10.1007/978-3-030-75100-5_54
 HyDiLLEch: A WSN-based Distributed Leak Detection and Localisation 627

an increment of nearly 550% compared to the previous year [2]. The non-human-
based (sensorised) LDMS, on the other hand, detect leakages via the supervisory
control and data acquisition (SCADA), fibre optic or copper installed along the
length of the pipeline. More advanced sensorised LDMS include the use of mul-
tisensory network, i.e. Wireless Sensor Networks (WSN), which employ various
detection techniques and are considered to be more efficient than others [3,4].
However, existing works such as [5] proposed WSN-based LDMS that are
centralised, making them susceptible to Single Point of Failures (SPOF). This is
in addition to other constraints (energy consumption, robustness, etc.) associated
with WSNs accompanied by sensitivity, accuracy, high false alarm rates, etc. [6]
related to the used leak detection techniques (LDTs). Together, these challenges
might explain the limited adoption of sensorised LDMS for pipelines, as shown
in [7].
Thus, in this article, we present HyDiLLEch: a fusion of several detection
techniques suitable for resource-constrained networks. HyDiLLEch is aimed at
improving the resilience to failures (communication, node, third party inter-
ference, etc.) by removing the SPOF associated with centralised systems in a
distributed manner. We examine its efficiency in terms of accuracy of localisa-
tion, energy consumption, and communication overhead compared to the clas-
sical LDTs. The remainder of this article is structured as follows. In Sect. 2,
related work is discussed. In Sect. 3, we detail some required hydraulics back-
ground. Our contribution is presented in Sect. 4 and discussed in Sect. 5. We
conclude and present future works in Sect. 6.

2 Related Work
Sela et al. [8] worked on a robust placement of sensors in a pipeline network
using robust mixed integer optimisation (RMIO) and robust greedy approxi-
mation (RGA) as enhancement of the nominal versions. In most cases, RMIO
and MIO outperformed the other versions in the conducted test. [9] worked on
fault detection in water pipelines using an approximate solution of the minimum
set cover problem by adopting the minimum test cover approach. Both works
are based on the assumption that a single sensor can detect failures in multiple
pipelines, making them robust to node failures.
Rashid et al. [6] used machine learning for leakage detection and classification
in pipelines. They compared several machine learning algorithms like the sup-
port vector machine (SVM), K-nearest neighbour (KNN) and Gaussian mixture
model (GMM), and SVM outperformed the rest in terms of sensitivity, speci-
ficity, and accuracy for leak size estimation. They, however, did not consider
leakage localisation. Another work based on artificial neural networks (ANN) is
that of Roy [10]. He proposed the use of ANN, consisting of input, hidden and
output layers as an optimisation tool for accurate leakage detection in pipeline
networks. While results from conducted experiments showed improved accuracy
in leak detections, ANN may not be suitable for resource-constrained networks
like WSNs.
628 S. Ahmed et al.

Beushausen et al. [11] worked on detecting transient leaks using a statistical

pipeline leak detection method through the analysis of the pressure, flow, and
modified volume balance of crude oil. Although the system successfully detected
some of the transient leaks in the pipeline, localisation error was up to 10 km,
amongst other challenges. Authors in [12] proposed a non-intrusive leak detec-
tion method for fluid pipelines. They estimated the effects of air bubbles in the
proposed system’s efficiency and evaluated its accuracy and suitability in detec-
tion. Also demonstrated in [5] are two leak detection techniques, i.e. the negative
pressure wave method and the gradient method. Experimental results showed
that both methods could be used to detect and locate leaks. The gradient-based
is more energy efficient due to its low sample rate but has slightly lower accuracy
compared to the other method.
Whereas these works have considered various aspects of LDMS, the indus-
trial implementation in the midstream is still in the early stages [7]. Research [13]
shows that majority of failures in pipelines are caused by vandalism and third-
party interference. We have therefore taken into account such situations to pro-
pose a decentralised detection and localisation of leakages to address these chal-
lenges. To the best of our knowledge, no work has considered distributed leakage
detection based on their causal factors.

3 Leakages in Pipelines

Software LDTs such as the pressure point analysis (PPA), the gradient-based
method (GM) and the negative pressure wave method (NPWM), amongst oth-
ers, have proven to be more efficient than others. They are non-invasive LDTs
and easy to implement and deploy on existing infrastructure [5]. Therefore,
HyDiLLEch is a combination of these techniques. In this section, we give a
short reminder about leakages in pipelines and how these techniques work.
Due to frictional resistance, in a transmission pipeline with a steady state, i.e.
presenting no leakage, the pressure decreases with distance. Figure 1 illustrates
the Bernoulli’s principle, which shows the pressure gradient changes of the crude
oil as it travels along a pipeline. P0 and PL are the pressures at the inlet and the
outlet of the pipeline, respectively in a steady state. P0 decreases with distance
along the pipeline resulting in a relatively constant pressure gradient (PG) for
every measurement points. In a leak case, a negative pressure wave (NPW) -
travelling in opposite directions from the point of leak - is generated by the
occurrence of a leak [5,14]. This results in different pressures P0leak at the inlet
and PLleak at the outlet of the pipeline compared to P0 and PL .
Methods to detect and localise a leakage include PPA, GM and NPWM,
amongst others. PPA is a detection technique used in determining dissipation of
flow in pipeline fluid implemented by measuring the pressure at various points
along the pipeline using the well known Bernoulli equation (Eq. 1).

Pa V2 Pb V2
za + + a = zb + + b + Eab (1)
ρg 2g ρg 2g
 HyDiLLEch: A WSN-based Distributed Leak Detection and Localisation 629

Fig. 1. The pressure gradient distribution in a pipeline before and after leakage

where Eab is the energy head loss, V = velocity, d = Pipeline’s inner diameter,
L = the distance from point a to point b,
Pa = pressure at point a, ρ = fluid mass density , g = gravitational force
and za = the elevation at point a.
The GM makes use of the two steady states, i.e. the states before a leak
(P Gleak leak
0−Q ) and after a leak (P GQ−L ). Given the difference in these pressure gra-
dients as shown in Fig. 1, leakage can be localised with Eq. (2).

L × dP Gleak
Q−L + (dp0 − dpL )
Q= (2)
Q−L − dP G0−Q
dP Gleak leak

where Q = the estimated leak location, L = the pipeline length, dp0 = average
increment in the pipeline’s initial cross-section, dpL = average increment in
pipeline final cross-section, dP Gleak leak
0−Q /dP GQ−L are the average increments in
the pressure gradient before/after the leak point.
The NPWM, on the other hand, makes use of the arrival of the NPW front
at the sensor nodes. The NPW speed c can be calculated with (Eq. 3). But,
the amplitude of the wave attenuates with distance. Thus, (Eq. 4) gives us the
maximum detectable distance of the wave front from the point of leakage.
1
c=  (3)
1 d
ρ( K + Y.w )

Ab = Aa ∗ e−αD (4)
where ρ = Fluid density, K = Fluid’s modulus of elasticity, d = Pipeline’s
inside diameter, Y = The Young’s modulus and w = Pipeline’s wall thickness,
Ab = the amplitude at sensing point b, Aa = the initial amplitude i.e. at sensing
point a, α = attenuation coefficient and D = the distance between two sensing
points.
630 S. Ahmed et al.

Each of the techniques mentioned above has some limitations. For example,
the accuracy of the GM is tightly coupled with the accuracy of the sensor nodes.
The NPWM is dependent on the ability to accurately detect the wave front,
which reduces the detecting distance. The NPWM is also less power-efficient
than the GM due to its higher sampling rate [5]. Using PPA to detect leakages
in transient states and localise leaks is impossible without combining it with
other changes resulting from the leak [15].
Therefore, we propose combining these LDTs (PPA, GM and NPWM) by
taking advantage of their various strengths to improve the accuracy and sen-
sitivity of leakage detection and localisation while minimising their individual
weaknesses.

4 Contributions

We propose a distributed architecture based on several sensors deployed on

the pipeline with a mesh connection that enables pre-processing of informa-
tion among geographically close sensors. This allows distributed detection and
localisation of leakages at the sensor level. For the communication aspect, we
propose the use of a LPWAN in short and long-range communication and cellu-
lar network as a backhaul as they are more widely deployed.

Fig. 2. Detection and localisation of leakages

Based on this architecture, we introduce HyDiLLEch, an LDT based on

the fusion of PPA, GM, and NPWM. It aims to minimise false positives and
improve accuracy through a distributed algorithm by defining appropriate dis-
tance between nodes and interconnecting them until 2-hops. This is achieved
through the spatial correlation of data shared among geographically close sensor
nodes. The data containing pressure and the NPW information is shared with
neighbouring sensors in a single-hop or double-hop manner. Detection and local-
isation are implemented locally, i.e. on the sensors using this partial information
shared amongst the nodes.
 HyDiLLEch: A WSN-based Distributed Leak Detection and Localisation 631

Propagation-Based Node Placement: Some of the factors that affect the

performance and efficiency of an LDMS is the placement of sensors on a pipeline.
Several approaches exist based on the maximum communication range [16], the
definition of the shortest distance between the event and a sensor [8,9], or the
placement of nodes only at the key junctions [1]. By our side, we assume that
leakage events are stochastic in nature, cannot be pre-determined and may occur
at any point in the pipeline. Also, researches [2,13] shows that over 90% of
leakages are caused by third party interference and vandalism, where the mean
value of the failure rate is 0.351 per km-year. Hence, we propose the deployment
of several sensors along a pipeline segment with two constraints :

• The distance between sensors must be less than half the sensor’s maximum
communication range to ensure a two-hop interconnection.
• The distance between sensors and event source must be sufficiently small to
make small leakage events detectable by at least three nodes considering the
amplitude of the NPW (Eq. 4). This is to ensure detection and localisation
in the event of node failure.

In this work, we consider a linear and uniform deployment of sensors on the

pipeline. We limit the number of collaborations among the sensors to a maxi-
mum of 2-hops. This allows us to minimise the various interferences resulting
from such collaborations and optimising the energy consumption for all the sen-
sors. Therefore, 2-hops data sharing represents additional information from two
upstream nodes and two downstream nodes (Fig. 2).
Leakage Detection: In HyDiLLEch, leakage detection is implemented by mak-
ing use of PPA to pre-estimate the expected pressure at every sensor node
location. According to Eq. (1), we can find the pressure gradient P G0−L for
a pipeline. Thus, the pressure gradient in a steady state at every point for a
horizontal pipeline can be estimated as follows:
P0 PL 1
P G0−L = ( − ) (5)
ρg ρg L
A threshold is set to accommodate the difference between the real value and
possible calibration errors from the sensor readings. The first step in detect-
ing leakage is by comparing the sensed pressure and the pre-estimated pressure
threshold. A difference between these values represents a possible leakage. Then
we check the pressure values to determine the pressure gradient on both side
of the current node i ( P G(i−1)−(i) , P G(i)−(i−1) ) and finally, we check for the
arrival of the NPW front. Thus, we can correctly detect the presence of leakages
and reduce false positives.
Leakage Localisation: Each node detecting a leakage does its localisation
estimation by two methods. The first one is GM, using pressure gradients and
Eq. (2), while the second is considering the time of arrival of the NPW front [17].
It can be calculated by:

Q = node s of f set + c × δt (6)

632 S. Ahmed et al.

where c is the negative wave speed estimated by Eq. (3), δt the difference
in arrival time of the signal at the upstream and downstream nodes and node’s
offset is the distance between the nodes multiplied by the detecting node’s index.

Algorithm 1. HyDiLLEch (Double-Hop)

1: {Initialisation while no leakage}
2: Form Neighbourhood
3: Get PG in steady state
4: Set upper, and lower threshold values for PG
5: for ever do
6: Get pressure data from neighbours
7: Calculate local gradients P G(i−1)−(i) , P G(i−2)−(i) , P G(i)−(i+1) and P G(i)−(i+2)

8: Detect NPW
9: if local PG is outside the threshold and NPW detected, share NPW data then
10: Try to localise using GM data (Eq. 2)
11: Try to localise using NPW data (Eq. 6)
12: else
13: No leak detected
14: end if
15: end for

Finally, HyDiLLEch is defined with a neighbourhood of 1-hop (2 nodes) or

2-hops (4 nodes) to deal with a possible node failure - respectively referred to as
HyDiLLEch-1 and HyDiLLEch-2. Algorithm 1 shows the main structure of the
2-hops version.
Both versions result in an increment in the number of sensor node detecting
and localising leakages (NDL) with relatively high localisation accuracy. The
results of the simulation are discussed next in the following section.

5 Simulations and Results

Metrics used to determine the efficiency of LDTs include accuracy of leakage

localisation, sensitivity to leaks, amongst others. In this work, we focus on these
two metrics to determine the efficiency of HyDiLLEch compared to the classical
LDTs, including the energy consumption and communication overhead due to
the distributed approach of our LDT.
Using NS3, we simulate crude oil propagation and leakages on a horizontal
transmission pipeline for a single-phase laminar flow within industrially-defined
operational criteria. The pipeline properties, such as the length, material, etc.,
are those of existing long haul transmission pipelines sourced from the Depart-
ment of Petroleum Resources (DPR) of Nigeria. Tables 1 and 2 outlines the
simulation parameters.
 HyDiLLEch: A WSN-based Distributed Leak Detection and Localisation 633

Table 1. Pipeline and oil characteristics Table 2. Network simulation parameters

Material Carbon steel Number of sensors 21

Pipeline Length (L) 20 km Number of gateways 1
Wall thickness (w) 0.323 m PHY/MAC model 802.11ax Ad hoc
Inside diameter (d) 0.61 m Transmit power 80 dBm
Height/elevation (z) 0m Transmit distance 20 km
Oil kinetic viscosity 2.90 mm2 /s Error model YANS
Temperature 50 ◦ C Propagation Loss Log-distance
Oil density (ρ) 837 kg/m3 Path Loss (L0 ) 46.67 dB
Inlet pressure (P0 ) 1000 psi Reference distance (d0 ) 1m
Reynolds no (Re ) 1950 Path-Loss Exponent (σ) 3.0
Velocity (V ) 2 m/s Packet size 32 bytes
Molecular Mass (m) 229 Data rate 1 Kbps
Oil elasticity (K) 1.85 × 105 psi Distance between sensors 1 Km
Carbon steel elasticity (Y ) 3 × 106 psi
Gravitational force (g) 9.81 m/s2
Constant (e) 2.718
Coefficient of friction (λ) 0.033
Wave speed (c) 14.1 m/s

These parameters enable us to calculate the values of the NPW speed using
Eq. (3). Expected pressure (PL ) at the pipeline’s outlet in a steady state is firstly
estimated by the Bernoulli equation defined in Eq. (1). The resulting value is
then used to determine the pressure gradient using Eq. (5). As a preliminary
work, all tests are conducted in ideal conditions, i.e. no communication or node
failures and results are discussed in the following subsections.

5.1 Localisation Accuracy

In our first simulations, we implemented two LDTs (NPWM and GM) in a cen-
tralised way [18]. While they both localise the leakage at an average of 99% and
98%, respectively, with different variance, the centralised detection and localisa-
tion make them susceptible to SPOF and less robust to other types of failures, i.e.
communication failure and others outlined in Sect. 3. In this section, we explain
how HyDiLLEch performs relative to this drawback.
Both single-hop (HyDiLLEch-1) and double-hop (HyDiLLEch-2) versions
of HyDiLLEch were simulated, and the results are represented in Fig. 3 and
4 respectively. The number of NDL are four for HyDiLLEch-1 and six for
HyDiLLEch-2, a significant increment compared to one gateway in the cen-
tralised version. These sensors, which are represented as N1−6 in Fig. 3 and
Fig. 4, make use of the spatially correlated data (pressure gradient and negative
pressure wave arrival) received from neighbouring sensors.
634 S. Ahmed et al.

Fig. 3. Single-hop HyDiLLEch Fig. 4. Double-hop HyDiLLEch

Thus, in HyDiLLEch-1, the average localisation accuracy on all four nodes

is above 96%, with nodes N2 and N3 maintaining the highest accuracy of
about 99%. HyDiLLEch-2 has similar results with two additional NDLs. Nodes
physically-close to the leak -N2 , N3 , N4 , N5 - are also maintaining the highest
accuracy of about 99%.
Note that from Figs. 3 and 4, we have several nodes localising using different
principal information (dpGM or dpNPW). The hybrid method is kept loosely
coupled to allow robustness of detection and to improve fault tolerance in cases
where one or more performance influencing factors such as noise, communication
failure, node failures, etc. is particularly detrimental to one of the localisation
techniques. The choice of the most accurate estimated leak location will be made
through a consensus algorithm to be implemented in the data and service layer
in our future work.

5.2 Communication Efficiency

To compare the communication efficiency of the LDTs, we consider the cost of

communication by the number of packets used by each LDT. We also evalu-
ate their energy consumption based on the sampling rate and the radio energy
consumption of the net devices. Obtained results are discussed next.
Communication Overhead: Figure 5 shows the number of packets exchanged
between the sensors and the gateway (for NPWM and GM-based on the total
number of packets needed to localise) and amongst the sensors for HyDiLLEch-1
and HyDiLLEch-2. The GM has the lowest number of exchanged packets com-
pared to NPWM and the two versions of HyDiLLEch. As expected, data cor-
relation among neighbours increased the overhead in the number of packets in
HyDiLLEch. Thus, HyDiLLEch-1 relative to GM has an overhead of about 50%
with a corresponding 200% increment in the number of NDL. HyDiLLEch-2 has
approximately 60% overhead with a 400% increment in the number of NDL as
well. On the contrary, both versions of HyDiLLEch requires a lower number of
packets when compared to NPWM packets with an even higher increment of
400% and 600% in the number of NDL respectively.
 HyDiLLEch: A WSN-based Distributed Leak Detection and Localisation 635

Fig. 5. Communication overhead by Fig. 6. Sampling energy consumption of

number of packets the sensors

Fig. 7. Radio energy consumption Fig. 8. Cumulative energy consumption

Energy Consumption: We evaluated the energy consumption of each LDT

based on the sampling rate of the sensors and radio energy consumption by the
network duty cycle. Results shown in Fig. 6 reveals the energy consumed per
second by the sensors. The high sampling rate of NPWM results in more energy
consumption compared to both GM and HyDiLLEch. Combining both tech-
niques in HyDiLLEch allows first determining the offset region before enhancing
the location accuracy by the expected time of arrival of the NPW. Thus, we need
less than 2% of NPWM sampling rate, allowing us to take advantage of the low
energy consumption of GM while maintaining the high localisation accuracy of
NPWM.
In the radio energy consumption case, we analyse the rate of change of the
energy consumed by the netdevice for every duty cycle. From Fig. 7, the most
significant change can be seen during the first and second duty cycles. This is due
to the difference in connections required among the participating nodes by each
LDT. Once connections are established and as the cycle increases, the energy
consumption converges and becomes stable for each of the methods. Figure 8, on
the other hand, shows the cumulative energy consumption of all the methods.
Energy consumption compared to NPWM is reduced by approximately 86% and
83% for HyDiLLEch version 1 and 2, respectively, in the first duty cycle, and
increase by 6–7% compared to GM per NDL. Both versions show a more linear
rate of change in consumption as the cycle increases, which is very similar to
that of GM.
636 S. Ahmed et al.

6 Conclusion and Future Works

In this paper, we proposed a new LDT –HyDiLLEch– based on a fusion of several
detection techniques using a node placement strategy that allows distributed
detection and localisation of small to big-sized leakages.
Simulation results show that HyDiLLEch in both the single and double-hops
versions increased the NDLs by four to six times, respectively, thereby eliminat-
ing the SPOF problem related to the classical LDTs. In terms of communication
efficiency, both versions of HyDiLLEch performed comparatively well with little
to no communication overhead. We also showed that the global energy consump-
tion of HyDiLLEch is relatively low compared to NPWM and similar to GM. In
our future work, we will examine the fault tolerance and resilience of HyDiLLEch
by introducing communication, node, etc. failures to the system. The choice of
the most accurate estimated leak location will be made through an asynchronous
consensus to be implemented in the data and service layers. Once the data and
service layers are implemented, experiments will be carried out on real pipeline
networks.

Acknowledgment. This work was supported by the Petroleum Technology Develop-

ment Fund.

References
1. Slaughter, A., Bean, G., Mittal, A.: Connected barrels: Transforming oil and gas
strategies with the Internet of Things. Technical Report Deloitte Center for Energy
Solutions (2015)
2. SPDCN, “Security, theft, sabotage and spills,” Shell Petroleum Development Com-
pany of Nigeria Limited, Technical Report 2017
3. Henry, N.F., Henry, O.N.: Wireless sensor networks based pipeline vandalisation
and oil spillage monitoring and detection: Main benefits for Nigeria oil and gas sec-
tors. In: The SIJ Transactions on Computer Science Engineering & its Applications
(CSEA), vol. 3, no. 1 (January 2015)
4. Yu, H., Guo, M.: An efficient oil and gas pipeline monitoring systems based on
wireless sensor networks. In: International Conference on Information Security and
Intelligent Control, pp. 178–181 (August 2012)
5. Ostapkowicz, P.: Leak detection in liquid transmission pipelines using simplified
pressure analysis techniques employing a minimum of standard and non-standard
measuring devices. Eng. Struct. 113, 194–205 (2016)
6. Rashid, S., Akram, U., Qaisar, S., Khan, S.A., Felemban, E.: Wireless sensor net-
work for distributed event detection based on machine learning. In: 2014 IEEE
International Conference on Internet of Things (iThings), and IEEE Green Com-
puting and Communications (GreenCom) and IEEE Cyber, Physical and Social
Computing (CPSCom), pp. 540–545 (September 2014)
7. Slaughter, A., Mittal, A., Bansal, V.: Bringing the digital revolution to midstream
oil and gas. Technical Report (2018)
8. Sela, L., Amin, S.: Robust sensor placement for pipeline monitoring: mixed integer
and greedy optimization. Adv. Eng. Inf. 36, 55–63 (2018)
 HyDiLLEch: A WSN-based Distributed Leak Detection and Localisation 637

9. Perelman, L.S., Abbas, W., Koutsoukos, X., Amin, S.: Sensor placement for fault
location identification in water networks: a minimum test approach. Automatica
72, 166–176 (2016)
10. Roy, U.: Leak detection in pipe networks using hybrid ANN method. Water Con-
serv. Sci. Eng. 2(4), 145–152 (2017)
11. Beushausen, R., Tornow, S., Borchers, H., Murphy, K., Zhang, J.: Transient leak
detection in crude oil pipelines. In: International Pipeline Conference (2004)
12. Santos, A., Younis, M.: A sensor network for non-intrusive and efficient leak detec-
tion in long pipelines. In: 2011 IFIP Wireless Days (WD), pp. 1–6 (October 2011)
13. Ambituuni, A., Hopkins, P., Amezaga, J.M., Werner, D., Wood, J.M.: Risk Assess-
ment Of A Petroleum Product Pipeline In Nigeria: The Realities Of Managing
Problems Of Theft/sabotage (2015)
14. Sowinski, J., Dziubinski, M.: Analysis of the impact of pump system control on
pressure gradients during emergency leaks in pipelines. In: E3S Web of Conferences,
vol. 44, p. 00166 (2018)
15. Sheltami, T.R., Bala, A., Shakshuki, E.M.: Wireless sensor networks for leak detec-
tion in pipelines: a survey. J. Ambient Intell. Humanised Comput. 7, 347–356
(2016)
16. Guo, Y., Kong, F., Zhu, D., Tosun, A.S., Deng, Q.: Sensor placement for lifetime
maximization in monitoring oil pipelines. In: Proceedings of the 1st ACM/IEEE
International Conference on Cyber-Physical Systems, ICCPS 2010 (2010)
17. Wan, J., Yu, Y., Wu, Y., Feng, R., Yu, N.: Hierarchical leak detection and localiza-
tion method in natural gas pipeline monitoring sensor networks. Sensors (Basel)
12(1), 189–214 (2012)
18. Ahmed, S., Mouël, F.L., Stouls, N.: Resilient IoT-based monitoring system for
crude oil Pipelines. In: Proceedings of the 7th International Conference on Internet
of Things: Systems, Management and Security (IOTSMS), IEEE (2020)
 Tactful Opportunistic Forwarding: What
Human Routines and Cooperation
Can Improve?

Rafael Lima Costa1,2,3(B) , Aline Carneiro Viana3 , Artur Ziviani4 ,

and Leobino Nascimento Sampaio2
1
École Polytechnique/IPP, Palaiseau, France
2
Federal University of Bahia (UFBA), Salvador, Brazil
{rlimacosta,leobino}@ufba.br
3
Inria, Palaiseau, France
[email protected]
4
National Laboratory for Scientific Computing (LNCC), Petrópolis, Brazil
[email protected]

Abstract. Opportunistic D2D forwarding algorithms have leveraged

human mobility characteristics to improve cost-effective content deliv-
ery. Most previous proposals focused on traditional or simplistic human-
centered metrics to improve performance in scenarios such as cellular
data offloading. Still, there is a need to approximate algorithm’s met-
rics to inherent in-depth aspects of human mobility hidden into real
datasets while leveraging more realistic scenarios interesting to mobile
operators. This work proposes TOOTS, a novel human-aware opportunis-
tic D2D forwarding strategy for cost-effective content delivery on cellu-
lar networks. TOOTS features a dissemination policy and a forwarding
algorithm that leverages wireless encounter patterns, temporal, spatial,
geographic, and direction awareness to improve cost-effectiveness deliv-
ery. These characteristics are extracted from NCCU and GRM datasets.
We compare TOOTS with the most popular state-of-art social-aware
algorithm, Bubble Rap, combined with three dissemination policies. The
TOOTS evaluation results show increased performance in terms of deliv-
ery rate, delivery latency, and overhead.

Keywords: Human-centered computing · Social computing · Network

architectures · Device-to-Device (D2D) forwarding

1 Introduction

The advent of 5G wireless networks will drive new business models in nearly
every vertical industry. Device-to-Device (D2D) communication and opportunis-
tic forwarding are considered enablers of 5G [1] and 6G [19] network communica-
tion and, consequently, emerging applications. They represent a new paradigm

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 638–652, 2021.
https://doi.org/10.1007/978-3-030-75100-5_55
 Tactful Opport. Forwarding: What Human Routines and Coop. 639

to offload traffic from wireless communication networks, which have attracted

multiple research initiatives involving user participation [13]. D2D opportunistic
forwarding algorithms tackled cost-effective and timely delivery of data [5,9]:
that is, delivering as much content as possible with less overhead and delay. In
these scenarios, contents (or messages) are forwarded user-to-user, from source to
destination in an opportunistic fashion (i.e., relying on user devices’ intermittent
connectivity).
Previous research discussed opportunistic network deployment to deal with
networking performance challenges, such as traffic growth (e.g., through data
offloading). Most initiatives focused on proposing algorithms and evaluating their
metrics. However, they still lack more evaluations with real traces to show the
benefits of using opportunistic communication in real-world applications (i.e., in
scenarios with more realistic settings). These algorithms typically consider user
encounters and individual mobility [5,9,10], points of interest (PoIs) [9], and
time-evolving social ties between node pairs [8]. Apart from that, not much was
done to approximate the evaluation metrics to broader inherent aspects of human
mobility while targeting QoE and QoS [2]. There is thus a lack of initiatives
beyond traditional techniques or limited human-mobility features to identify
routines (spatiotemporal patterns), related consequences (e.g., wireless encoun-
ters), and movement decisions (e.g., motion direction) with more granularity and
precision. This work focuses on metrics able to reflect human mobility features
during periods of the day. Still, each population has particular habits that can
change wireless contact dynamics [4]. Therefore, such metrics must be evaluated
throughout different populations, while spatiotemporal aspects must fit each
population’s habits. In this work, we combine features extracted from a real-
world and a synthetic dataset and apply them into a novel Tactful Opportunistic
cOmmunicaTion Strategy (TOOTS). This strategy features a dissemination pol-
icy and a forwarding algorithm. In the former, we leverage users’ spatiotempo-
ral properties and induced wireless encounters to choose content-disseminating
nodes that have shown previous encounter routines with destination nodes. The
forwarding algorithm relies on nodes’ popularity, direction and displacement,
network-cell (as PoIs) visiting, and proximity.
In [2], we presented and detailed the features of TOOTS, including mathe-
matical formulations and results from metrics characterizations using MACACO
(a private European Dataset) [17]. In [2], we discussed mobility datasets han-
dling and characterization. The same methodology is applied here with different
datasets. Conversely, the word tactful, which is part of the strategy’s name,
means having or showing skill and sensitivity in handling with people. In [3], we
discuss the Tactful Networking paradigm, whose goal is to add perceptive senses
to wireless networks by assigning it with human-like capabilities of observation,
interpretation, and reaction to daily-life features and associated entities. Our
contributions with TOOTS are thus the following:

– The TOOTS design combining human-aware metrics in a more granular and

precise way than state-of-art to improve networking and system performance;
640 R. L. Costa et al.

– The comparison of the TOOTS effectiveness in each stage using a real-world

and a synthetic dataset with state-of-art enhanced Epidemic and Store-wait-
forward, and Bubble Rap alternatives;
– TOOTS reaches 100% delivery rate with respectively 28%, and 73% less deliv-
ery latency, and with 16%, and 27% less overhead in the real-world and syn-
thetic datasets.
– Sharing the notion of going beyond traditional techniques when dealing with
human-aware metrics to reach a superior system and networking performance.
We also emphasize the need to evaluate such opportunistic strategies through
more realistic scenarios and real datasets, incentivizing mobile network oper-
ators to deploy opportunistic networks more widely.
The results mentioned above cope with [3], which talks about the need to
have a more in-depth look into human characteristics to build tactful network-
ing solutions. The remainder of this work is organized as follows. Section 2
discusses related work. Section 3 presents the mobility models (datasets) used
and motivates the strategy design. Section 4 discusses TOOTS, including its
architecture, features, and algorithms for content-dissemination and forward-
ing. Section 5 presents the experimentation results, analysis, and discussions.
Section 6 concludes this paper and points out future work.

2 Related Work
Over the years, different works extracted user mobility metrics to assist routing
algorithms and opportunistic forwarding strategies. Due to the dynamic human
mobility, choosing the nodes to forward the content is challenging.
In [5], the authors introduced Bubble Rap, a distributed forwarding algo-
rithm for delay-tolerant networks. The algorithm exploits centrality and com-
munity detection, characteristics remarked as less volatile than user mobility.
Bubble Rap outperformed other state-of-art algorithms and is used as a bench-
mark with recently published research [9,14]. Despite its broad contributions,
the major shortcomings relate to small scenarios (available traces had less than
100 nodes) evaluations. Further, community detection algorithms are computa-
tionally expensive and require parameters calibration, which is not feasible in
real-time D2D scenarios.
In [8], the authors propose dLife, a forwarding algorithm that relies on con-
tact duration and degree during daily hour periods. They capture the social
behavior and interactions dynamism, resulting in improved delivery rate, cost
(i.e., overhead), and latency. Like Bubble Rap, dLife was simulated through small
traces. The evaluation setup and scenario also had shortcomings (e.g., content
size ranging from 1 to 100 kb).
GROUPS-NET, an opportunistic algorithm that relies on social group meet-
ings, is introduced in [10]. This algorithm requires a centralized computation
of group-to-group paths whenever a given device wishes to send content to a
destination. This requirement is challenging in larger populations. Furthermore,
GROUPS-NET is validated within two campus datasets, lacking other kinds of
 Tactful Opport. Forwarding: What Human Routines and Coop. 641

environment evaluations. In large-scale scenarios, they had a similar delivery

ratio as Bubble Rap but with reduced overhead.
In [9], the algorithm SAMPLER combines individual mobility, PoIs, and
social-awareness for opportunistic routing. Despite showing improved delivery
ratio, reduced overhead, and delivery latency, this work has a few shortcomings.
The algorithm requires static relay points deployment and social community
computations, which require parameters calibration. Second, there is a lack of
details about the simulation scenario and parameters to turn possible repro-
ducibility for comparison reasons. Finally, the evaluations need more realistic
settings. For example, the authors considered any contact as enough to forward
a message. Other parameters, such as message time-to-live, node buffer size, and
more realistic content-size, must be evaluated.
TOOTS Positioning: From the related initiatives described herein, we con-
sider that understanding human-behavior and context from mobility is impor-
tant for networking solutions. Different from such studies, we use metrics with a
more granular link with time. We focused on identifying different human activi-
ties during periods of the day. Through analyzing the datasets and our metrics,
we found that their coefficients vary according to the period [2]. Considered
human-mobility features are routines (spatiotemporal patterns), related conse-
quences (e.g., wireless encounters), and movement decisions (e.g., motion direc-
tion). Metrics and insights based on generic aspects might not apply to all pop-
ulations, so datasets of different kinds (i.e., different environments) and more
users need to be evaluated. Facets such as node density, proximity, transporta-
tion mode, and even cultural factors can change contact dynamics. Therefore, it
is necessary to choose the temporal aspects and other characteristics accordingly.

3 Rationale
Datasets Description: We use the real-world NCCU trace [16], which features
the displacements (given by GPS coordinates) of 115 users on a campus (3764 m
× 3420 m) throughout two weeks. Additionally, we adopted the synthetic GRM
dataset, which captures social regularity by human mobility [11]. The GRM
dataset has 1000 users displacements in a 1500 m x 1500 m area throughout two
weeks and applies here to simulate a larger network (i.e., with more nodes).
Overhead and Latency Evaluation: In Fig. 1, we evaluate the Epidemic
forwarding with a consumer set size equal to 1% of each dataset amount of nodes.
Note that 1% represents a limited number of consumers, which puts the Epidemic
strategy in a challenging scenario. Depending on datasets’ characteristics and the
strategy’s kind, this can result in lower bound results for dissemination strategies,
which will also be used in the TOOTS evaluation hereafter. Although mostly
evaluated in the literature, we illustrate the performance constraints related
to the Epidemic forwarding algorithm using the two mentioned datasets. The
Epidemic algorithm always tries forwarding a content to an encountered node
that does not have it. Epidemic forwarding is known as having smaller delays
and a higher delivery ratio; however, it brings a heavy network overhead burden.
642 R. L. Costa et al.

Fig. 1. Epidemic forwarding average delivery latency performance evaluation.

Source and destination nodes were chosen randomly with a content generation
per hour during one full day. The experiments were carried 30 times, and so the
confidence intervals are presented.
The Epidemic algorithm delivers the contents with an average latency of 2.6
and 3.6 h, respectively, for NCCU and GRM. The delivery-rate was 100% on
those experiments, but with 98.44% and 99.62% average network “infection”
(overhead). The latter means that most network nodes became forwarders. For
reaching such a delivery rate and small delivery latency, the network was flooded,
which is not feasible in the real-world and is highly costly. Furthermore, one
of the Epidemic Forwarding characteristics is a higher average of hops (i.e.,
intermediary nodes) for delivering the message. For example, there are 6.62 and
8.85 average hops in these experiments for content delivery on NCCU and GRM.
This characteristic brought an insight and motivation for our strategy: taking
the content closer to the destination could reduce hops, delay, and overhead by
combining rational dissemination with intelligent forwarding decisions. In the
next section, we introduce TOOTS.

4 TOOTS Architecture and Strategy Description

TOOTS Architecture: Figure 2 describes the architecture for TOOTS. It fea-
tures a mobile network owned by a cellular operator with an infrastructure of
mobile edge computing (MEC) [7]. In this MEC site, the operator stores con-
tents commonly requested by its subscribers. The operator’s goal is to offload
content through opportunistic D2D communication, assuming the users are will-
ing to cooperate through incentives (e.g., data plan savings). We assume that
the operator controls the offloading process while the forwarding decisions hap-
pen in the user devices based on locally calculated metrics. TOOTS learns from
users’ mobility for seven days. Upon having a set of user requests for a given
content, the operator runs our Tactful Dissemination Policy (TDP) (Sect. 4.1)
to choose a set of disseminator nodes. In Fig. 2, we illustrate the offloading of two
 Tactful Opport. Forwarding: What Human Routines and Coop. 643

different contents through the chosen disseminators. Subsequently, an algorithm

(Sect. 4.2) runs locally (in the node) for taking a forwarding decision.

Fig. 2. TOOTS architecture.

TOOTS Metrics: TOOTS combines a novel temporal approach described

below with other insights into a Tactful Dissemination Policy (TDP)(Sect. 4.1);
the other metrics herein described are combined into a forwarding algorithm
(Sect. 4.2). Following, we resume these metrics, which are detailed (includ-
ing mathematical formulations) and evaluated in [2]. We discuss the intuitions
behind the use of each metric in the descriptions of the algorithms.

– Temporal Approach: We divided the day into six non-uniform periods of

different durations. This decision is justified by associating the periods to
the instants the population travels and has longer confinements (e.g., the
period from 06:00 to 09:59 accounts for most individuals’ home-to-work dis-
placement, while from 10:00 to 13:59 there is more confinement at work and
smaller displacements on lunchtime). With this temporal division, the inten-
tion is to extract decision factors increasingly accurate and closer to the real
human routines. In [2], we found metrics coefficients heterogeneity per period
in MACACO Dataset. The same phenomenon was observed here with NCCU
and GRM, which reinforces our intuition.
– Social Awareness Centrality Degree (CDp (u)): Measures the social
bonds of a user u in a period of the day p, that is, his number of encounters
given a specific communication range. Someone with a higher degree is more
“popular” (i.e., has further encounters).
– Coverage Area Radius of Gyration (RGp (u)): Quantifies in meters a
user (u) individual mobility related to a center of mass, calculated from his
movements in a period (p).
– Sojourn Time (STpc (u)): Quantifies in hours the stay of a user u in a network
cell c in a period p. We divided [2] the limited geographic space from each
dataset into cells (original local operators’ cells were too large) and calculated
users’ stay.
644 R. L. Costa et al.

– Destination Proximity as Geographical Awareness (M Ppc (u)): We cal-

culate the maximum proximity a node u reached towards a cell c in a period
p through a geodesic formula.
– Geographic Direction Awareness: Analyzes the last 30 min of node mobil-
ity to determine if its displacement is towards a given cell.

4.1 1st TOOTS Phase: TDP Policy

Herein, we detail our Tactful-Based Dissemination Policy (TDP). TDP (Algo-

rithm 1) chooses nodes based on their social behavior according to each time
window. The intuition behind this is to start the offloading process with the
content closer to their consumers. Following, we show how we apply the pre-
dictive regularity of user routines for choosing content disseminators. From the
dynamic contact graph Gt = (V, Ep ), where V is the set of users (mobile nodes),
and Ep is the set of edges found during the period p of the week k. We assume
that there is a logically centralized operator entity that, at the end of a week k,
receives the following information from each one of its nodes u ∈ V :

– (a): σ(u)p - the set of users v encountered by u in each period p.

– (b): ΔCLIDp (u) - the average local improved centrality degree metric [15] for
a user u in a period p. This metric accounts for the number of contacts (i.e.,
encounters), considering their duration and earliness. With that said, higher
coefficients belong to nodes with more contacts that had longer durations and
happened earlier in a time window (i.e., in a period p).

Algorithm 1: SelectDisseminators
input : Gt , C, c, p
output: D(c)
1 begin
2 D(c) ←− ∅
3 while C = ∅ do
|σp (u) ∩ C|
4 let u ∈ V maximizing | + ΔCLIDp (u)|
|C|
5 D(c) ←− u
6 C ←− C − σ(u)
7 end
8 Return D(c)
9 end

During the week k +1, upon having a set of consumers C(c) ∈ V interested in
a content c, Algorithm 1 runs for choosing disseminator nodes. We justify the use
of σ(u)p obtained at the week k by the fact that the policy tries to select nodes
that had direct contact with the consumers, and due to their routines, they are
 Tactful Opport. Forwarding: What Human Routines and Coop. 645

most likely to repeat those interactions [12]. We prioritize the direct contacts,
i.e., higher proximity, as they are most likely to repeat, possibly minimizing
the spending of network and user device resources. Further, the CLID measures
neighborhood coverage capabilities through contacts and is used to identify the
nodes’ popularity in the network, giving importance to their contacts’ duration
and earliness. We use this metric because, in an opportunistic communication
scenario, a contact that can transmit a message (i.e., the encounter lasts enough
to send the data) earlier in the time window might decrease the total delivery
latency. Furthermore, depending on the content size, choosing short contacts can
waste node resources without proper message transmission.
Algorithm 1 selects a set of disseminators D(c) for offloading a content c
upon period p. D(c) starts empty as the algorithm runs when there is a set of
Consumers C requisitions for a content c. The policy selects (line 4) the user
with the higher coefficient for the direct contacts with Consumers (σp (u) ∩ C),
normalized by C’s size, summed with the ΔCLIDp (u) (u’s average CLID in p).
The algorithm then removes the Consumers covered by u (line 6). In this way, we
guarantee a next user u with a similar coefficient for direct contacts do not get
selected. Algorithm 1 loops (line 3) until the consumer set is not fully covered
(C = ∅). The policy assumes that there are no isolated users in the traces. In
Sect. 5, we evaluate the performance of the proposed dissemination policy with
different sizes of consumer set (1%, 5%, and 10% of each dataset amount of
nodes).

4.2 2nd TOOTS Phase: Human-Aware Forwarding

After choosing the disseminator set D(c) for a content c, the operator sends the
content to each node u ∈ D(c), for storing in its local buffer. From this moment,
upon an encounter, any node carrying c runs Algorithm 2 to decide on the next
forwarding neighbor. The input is: the destination node d (that is, d ∈ C), the
content c, the period p, the encountered node v, and l as the coordinates of d’s
cell. Each node stores the metrics coefficients locally by period p during the whole
simulation, and a node u transmits c to v only if a certain algorithm condition
is satisfied. First, if v already has c (line 2), u waits for the next encounter.
Second, if v = d (line 4), that is, v is the destination, c is transmitted. The
content transmission is successful if the edge e between u and v is persistent
enough (there is no topology change causing a disconnection) until the time t
necessary to transfer c.
As explained in Sect. 4, the geographic space was divided into cells. If v is
at l (line 6), the nodes avg. Centrality degree in the period p are compared. If v
has a higher avg. CD, it means v met more nodes during p, so c is forwarded to
v, given his higher dissemination capacity inside that cell. In the contrary case
(line 9), if ΔRGp (v) > ΔRGp (u) and ΔSTpcell (v) > ΔSTpcell (u), v is less “popular”
(lower ΔCDp ) than u. Nevertheless, v can potentially cover a larger area inside
the cell l (given his higher RG) and routinely stays longer in l (given his higher
avg. ST). Hence, c is then forwarded to v (line 10). If v is not in l, his recent
mobility towards l is checked. If the direction test is true (line 12), i.e., v moves
646 R. L. Costa et al.

Algorithm 2: HumanAwareForwarding
input : d, c, p, v, l
output: A forwarding Decision
1 begin
2 if c ∈ v then
3 exit(0)
4 else if v = d then
5 v ←− c
6 else if v.cell = l then
7 if ΔCDp (v) > ΔCDp (u) then
8 v ←− c
9 else if ΔRGp (v) > ΔRGp (u) and ΔSTpl (v) > ΔSTpl (u) then
10 v ←− c
11 else
12 if testDirection(v, l) then
13 v ←− c
14 else if ΔM Ppl (v) < ΔM Ppl (u) then
15 v ←− c
16 end
17 end

towards the aim and has the potential to reach or to get closer to l, c is sent
to v. Finally, if the direction test is false, the Alg. checks the avg. Destination
proximity. If ΔM Ppl (v) < ΔM Ppl (u) , c is forwarded to v, meaning that node v got
closer (or visited) l during p and it is more likely to repeat this behavior.

5 Experimentation Results and Analysis

As stated previously, the operator’s goal is to offload content through oppor-

tunistic communication. The contents consist of 60 s advertising videos with
the average size of a YouTube 720P HD 30fps video, ranging from 11–14 MB.
TOOTS learns from users mobility for seven days; one day is for content gen-
eration, with an offloading task started by the operator every hour, and three
days are the delivery deadline [15]. The consumers of each content are chosen
randomly, with different consumer set sizes. The operator injects the contents
on the network through their subscribers, who run a forwarding algorithm upon
each contact. Every node has an 802.11/11 Mbps network interface. The commu-
nication range is 30 m (avg. for WiFi Direct). This scenario simulation is through
the Opportunistic Network Environment (ONE) Simulator [6], with NCCU and
GRM imported as mobility models. Each experiment was carried 30 times. The
confidence intervals appear when necessary.
 Tactful Opport. Forwarding: What Human Routines and Coop. 647

5.1 TDP Results and Analysis

(a) Average Delivery Rate (b) Average Delivery Latency

Fig. 3. Tactful dissemination policy average delivery rate and average delivery latency
performance evaluation.

This section evaluates the TDP policy only. For this, we modify two literature
forwarding strategies, named store-wait-forward (a.k.a., direct delivery) and epi-
demic forwarding. This modification consists of using the dissemination nodes
given by the TDP policy as the starting content sources, executing then each of
the literature strategies. Figure 3 plots the performance of the TDP-enhanced
store-wait-forward and epidemic forwarding. The store-wait-forward transmits
the message only if the encountered node is the destination (consumer). In both
traditional and TDP-enhanced versions, the store-wait-forward has zero over-
head. The traditional epidemic had 98.44%, and 99.62% overhead, as shown in
Sect. 3. On the other hand, the TDP-enhanced epidemic might get a compara-
ble delivery rate and decreased delay. TDP-enhanced store-wait-forwarding has
a potentially higher delivery rate and decreased delivery latency, as its success,
though still depending on direct encounters, is improved by the TDP’s initial
content dissemination. Improvements might still take time (or never happen)
depending on the source and destination social bonds.
Figure 3(a) reveals that combined with store-wait-forward, TDP induces an
average delivery rate of 82% on NCCU and 75% on GRM, regardless of the num-
ber of consumers evaluated (1%, 5%, and 10%). TDP enhancement assures choos-
ing nodes with past direct encounters with the consumers, which favors the deliv-
ery rate increase. Compared with the traditional store-wait-forward, the TDP-
enhanced average delivery rate with 1% of consumers was 18% higher on NCCU
and 60% higher on GRM. In Fig. 3(a), as expected, the TDP-enhanced epidemic
forwarding reaches a 100% average delivery rate in all scenarios. Besides, the
TDP enhancement favors the decrease of the average hop count of the tradi-
tional Epidemic forwarding.
648 R. L. Costa et al.

In Fig. 3(b), we evaluate the average delivery latency. We see that regardless
of the size of consumers set, the contents take on average 13h–15h to be deliv-
ered with TDP-enhanced store-wait-forward on NCCU and 18h–19h on GRM,
respectively. Thanks to the TDP policy, most of the delay-tolerant content is
forwarded in an acceptable time [15]. Compared with the traditional store-wait-
forward with 1% of consumers, TDP-enhanced reduced the average delivery
latency by 23% on NCCU and 48% on GRM. As expected, when the forwarding
is the enhanced epidemic, the average delay is smaller (18.6% less on NCCU
and 64.5% less on GRM with 1% of consumers), but with comparable network
infection (overhead), which is unfeasible or very costly in real scenarios. From
these findings above, we look for a forwarding algorithm able to increase the
delivery rate with as lower as possible overhead and delivery latency.

5.2 TOOTS Forwarding Results and Analysis

(a) NCCU - Delivery Rate (b) GRM - Delivery Rate

Fig. 4. Delivery rate performance comparison of TOOTS, bubble rap-CDP, bubble

rap-TDP, and bubble rap-RDP on NCCU and GRM datasets.

Finally, TOOTS (i.e., the full strategy) is evaluated in terms of delivery rate,
delivery latency, and overhead within NCCU and GRM datasets. The chosen
scenario is with random consumer sets with a size equal to 1% of dataset nodes
(as used for the epidemic performance evaluation). TOOTS’ performance is com-
pared with Bubble Rap, the most popular social-based forwarding algorithm for
delay-tolerant content. In Bubble Rap, each node gets a GlobalRank, and a
LocalRank. The first measures the popularity in the whole network, while the
second stores the popularity inside a community. Both centrality ranks are cal-
culated with the C-Window technique, which measures nodes’ popularity as the
average number of unique nodes encountered throughout a time window of fixed
length from the last 24 h. Furthermore, Bubble Rap identifies social communities,
 Tactful Opport. Forwarding: What Human Routines and Coop. 649

and each node must belong to at least one. Given these metrics, the forwarding
strategy transmits a content c carried by u to v when: v is the destination node
(d); v has higher GlobakRank or belongs to d’s community; v belongs to d’s
community and has higher LocalRank.
We combine Bubble Rap with three dissemination policies responsible for
injecting the contents into the network upon the start of the offloading process
by the operator:
– Random-Based Dissemination Policy (RDP): selects origin nodes ran-
domly (i.e., working like the traditional Bubble Rap);
– Centrality-Based Dissemination Policy (CDP): selects the higher cen-
trality node as the origin;
– Tactful-Based Dissemination Policy (TDP): proposed in Sect. 4.1.
Selects the nodes with more direct encounters and higher CLID .
In Fig. 4(a), we evaluate the delivery rate for both TOOTS and Bubble Rap
on NCCU trace. TOOTS achieved a 100% delivery ratio and had the fastest
delivery (90% of the contents delivered in up to 9 h). Bubble Rap-CDP failed
in delivering all contents, even in the NCCU’s smaller population, which has
overall high centrality degree nodes. Furthermore, using a CDP in the real-world
is not feasible, as it creates a bottleneck and tends to drain the most “popular”
users’ devices resources. Bubble Rap-RDP reaches a 100% delivery ratio but is
expected to show higher overhead (analysis to follow). Finally, Bubble Rap-TDP
reaches a 100% delivery ratio but takes slightly more time to forward all contents
successfully than TOOTS.
In the GRM trace (Fig. 4(b)), TOOTS delivers 100% of the contents, with
an average of 97% in up to 12 h. Bubble Rap-CDP shows an even worst delivery
performance than in NCCU. That happens due to (i) communication bottle-
necks, as there is a higher node density, and each node interface can transmit
only one content at a time, and (ii) the fact that the GRM dataset has overall
nodes with much lower centrality degrees. Bubble Rap-RDP delivers 100% but
is also expected to show higher overhead (analysis to follow). Bubble Rap-TDP
reaches close to 94% delivery ratio. This result can also be explained by GRM’s
lower centrality degrees (explanation to follow), calling for more time to reach
100% delivery ratio in Bubble Rap-TDP.
In Fig. 5(a), we evaluate the delivery latency. TOOTS has the lower average
delivery latency on both datasets, followed by Bubble Rap-RDP on NCCU and
Bubble Rap-TDP on GRM. We remind that on GRM, only TOOTS and Bubble
Rap-RDP reached 100% delivery rate in up to 72h. On GRM, Bubble Rap-
CDP shows a much higher average latency, justified by the bottlenecks on the
fewer higher centrality nodes and by a dataset characteristic: from the content-
generation time till the end of the simulation, over 90% of the nodes have low
centrality degrees (up till 0.4). These dataset characterization results were not
plotted as we focus on the full strategy proposal and its results.
Finally, in Fig. 5(b), we find the Bubble Rap-CDP with the smaller over-
head on both datasets. On NCCU, this happens as the higher centrality degree
node can find the destination directly for most of the contents generated. On
650 R. L. Costa et al.

(a) Avg. Delivery Latency (b) Avg. Infection % (Overhead)

Fig. 5. Delivery latency and overhead performance comparison of TOOTS, bubble

Rap-CDP, bubble Rap-TDP, and bubble Rap-RDP on NCCU and GRM datasets.

the other hand, still, a bottleneck arises. Despite the lower overhead, Bubble
Rap-CDP has a shallow delivery rate compared with the other proposals. The
time windows used by Bubble Rap here also might be a reason for the slower
delivery performance. Bubble Rap-RDP presents the worst-case in terms of over-
head. Compared with Bubble Rap-TDP, there is respectively 53% and 65% less
overhead on NCCU and GRM. On the other hand, we must remember this com-
bination does not deliver 100% of the contents on GRM when the deadline is
up to 72h. Using other real datasets with larger populations would be interesting
to evaluate this strategy. TOOTS is the fastest strategy able to deliver 100% of
the contents on both datasets. Its overhead is respectively 10% and 17% smaller
than Bubble Rap-RDP.

6 Conclusion and Future Work

Despite remarked as an enabler of future networking generations since 4G/LTE,
opportunistic communications have not been widely deployed by cellular opera-
tors. Recent research discusses the D2D-enabled 5G and 6G [18,19]. Novel initia-
tives in opportunistic communication will arise through a deeper understanding
of the human behavior hidden into the datasets and the use of AI [19].
This work discussed a more realistic application scenario and introduced a
human-aware opportunistic communication strategy named TOOTS. Compared
to Bubble RAP and in both evaluated datasets (i.e., NCCU and GRM), TOOTS
shows increased delivery performance and reduced delivery delay and overhead.
Besides, TOOTS does not rely on complex communities calculation requiring
parameter calibration. These characteristics make Bubble Rap challenging to be
implemented in real-world scenarios.
Aiming for even better performance, future work related to TOOTS must
evaluate the involved decision factors individually. In particular, in the real
 Tactful Opport. Forwarding: What Human Routines and Coop. 651

dataset NCCU, TOOTS showed a better correlation with the leveraged human
characteristics. The time-window proposal also made it possible to calculate
more precise metrics reflecting the routine-based human mobility behavior at
different daily periods.
There is still a lack of datasets with larger populations to be evaluated for
real-world data offloading applications. We found that a dissemination strategy
to inject the contents into the network is valuable in a scenario such as the
one presented. Further, research must also evaluate the presented algorithm and
other similar proposals regarding different contact-range, available bandwidth,
variable content-size, node buffer size, and node energy constraints.
Furthermore, when it comes to human-aware metrics in the big data era,
there are also many possibilities. Thus, future research shall invest efforts in
proposing different metrics to identify human routines and social aspects from
mobility data, but from other sources (e.g., personality traits [3].
Finally, it is necessary to discuss different application scenarios to have a
clearer idea of what opportunistic D2D communication will be able to do or
enhance. Showing opportunistic gains in more realistic scenarios and innovative
applications can motivate cellular network operators to invest more efforts in
the context of human-centered networks.

Acknowledgements. This work was supported in part by the CAPES

(88881.135682/2016-01; 88887.468235/2019-00), CNPq (465.560/2014-8; 310.201/2019-
5; 432064/2018-4), FAPERJ(E-26/203.046/2017), FAPESB (TIC0004/2015), INCT-
CiD, EMBRACE associate team Inria project, and ANR MITIK project (PRC
AAPG2019).

References
1. Akyildiz, I.F., Nie, S., Lin, S.C., Chandrasekaran, M.: 5g roadmap: 10 key enabling
technologies. Comput. Netw. 106 17–48 (2016) Supplement C
2. Costa, R.L., Carneiro Viana, A., Ziviani, A., Sampaio, L.N.: Towards human-aware
d2d communication. In: IEEE DCOSS, pp. 173–180 (2020)
3. Costa, R.L., Viana, A.C., Ziviani, A., Sampaio, L.N.: Tactful networking: humans
in the communication loop. IEEE TETCI, pp. 1–16 (2020)
4. Cuttone, A., Lehmann, S., González, M.C.: Understanding predictability and
exploration in human mobility. EPJ Data Sci. 7(1), 2 (2018)
5. Hui, P., Crowcroft, J., Yoneki, E.: Bubble rap: Social-based forwarding in delay-
tolerant networks. IEEE TMC 10(11), 1576–1589 (2011)
6. Keränen, A., Ott, J., Kärkkäinen, T.: The one simulator for dtn protocol evalua-
tion. In: Proc. ICST, Simutools (2009)
7. Mao, Y., You, C., Zhang, J., Huang, K., Letaief, K.B.: A survey on mobile edge
computing: The communication perspective. IEEE Commun. Surv. Tutor 19(4),
2322–2358 (2017)
8. Moreira, W., Mendes, P., Sargento, S.: Opportunistic routing based on daily rou-
tines. In: IEEE WoWMoM, pp. 1–6 (2012)
9. Nunes, I.O., Celes, C., Nunes, I., Vaz de Melo, P.O.S., Loureiro, A.A.F.: Combining
spatial and social awareness in d2d opportunistic routing. IEEE COMMAG 56(1),
128–135 (2018)
652 R. L. Costa et al.

10. Nunes, I.O., de Melo, P.O.S.V., Loureiro, A.A.F.: Leveraging d2d multihop com-
munication through social group meeting awareness. IEEE Wirel. Commun. 23(4),
12–19 (2016)
11. Nunes, I.O., Celes, C., Silva, M.D., Vaz de Melo, P.O., Loureiro, A.A.: GRM: group
regularity mobility model. In: Proceedings ACM MSWiM, pp. 85–89 (2017)
12. Oliveira, E.M.R., Viana, A.C., Sarraute, C., Brea, J., Alvarez-Hamelin, I.: On the
regularity of human mobility. Pervasive Mob. Comput. 33, 73–90 (2016) (Supple-
ment C)
13. Rebecchi, F., Dias de Amorim, M., Conan, V., Passarella, A., Bruno, R., Conti, M.:
Data offloading techniques in cellular networks: a survey. IEEE Commun. Surv.
Tutorials 17(2), 580–603 (2015)
14. Tao, J., Wu, H., Shi, S., Hu, J., Gao, Y.: Contacts-aware opportunistic forwarding
in mobile social networks: a community perspective. In: IEEE WCNC, pp. 1–6
(2018)
15. Thilakarathna, K., Viana, A.C., Seneviratne, A., Petander, H.: Design and analysis
of an efficient friend-to-friend content dissemination system. IEEE TMC 16(3),
702–715 (2017)
16. Tsai, T., Chan, H.: Nccu trace: social-network-aware mobility trace. IEEE COM-
MAG 53(10), 144–149 (2015)
17. Viana, A., Jaffres-Runser, K., Musolesi, M., Giordano, S., Fiore, M., Olmo, P.:
Mobile context-Adaptive CAching for COntent-centric networking (2013). http://
macaco.inria.fr Accessed Jan 2021
18. Waqas, M., Niu, Y., Li, Y., Ahmed, M., Jin, D., Chen, S., Han, Z.: A comprehensive
survey on mobility-aware d2d communications: principles, practice and challenges.
IEEE Commun. Surv. Tutorials 22(3), 1863–1886 (2020)
19. Zhang, S., Liu, J., Guo, H., Qi, M., Kato, N.: Envisioning device-to-device com-
munications in 6g. IEEE Netw. 34(3), 86–91 (2020)
 Cooperative and Coordinated Mobile Femtocells
Technology in High-Speed Vehicular
Environments
Mobility and Interference Management

Rand Raheem(B) , Aboubaker Lasebae, Miltos Petridis, and Ali Raheem

Faculty of Science and Technology, Middlesex University, London NW4 4BT, UK

{R.H.Raheem,A.Lasebae,M.Petridis,A.Raheem}@mdx.ac.uk

Abstract. In future networks, most users who will be accessing wireless broad-
band will be vehicular. Serving those users cost-effectively and improving their
signal quality has been the main concern of many studies. Thus, the deployment of
Mobile Femtocell (Mobile-Femto) technology on public transportation is seen to
be one of the promising solutions. Mobile-Femto comes with its mobility and inter-
ference challenges. Therefore, eliminating the Vehicular Penetration Loss (VPL)
and interference while improving signal quality and mobility for train passengers
is the main concern of this paper. The initial system-level evaluation showed that
the dedicated Mobile-Femto deployment has great potential in improving users’
experience inside public transportation. The Downlink (DL) results of the Pro-
posed Interference Management Scheme (PIMS) showed significant improvement
in Mobile-Femto User Equipment (UE) gains (up to 50%) without impacting the
performance of macro UEs. In contrast, the Uplink (UL) results showed noticeable
gains for both macro UEs and Mobile-Femto UEs.

1 Introduction
There are always new challenges on future mobile networks especially with the traffic
explosion in today’s networks. Previous studies predicted that from 2020 the number
of connected devices to the internet will exceed 50 billion which in return will require
a tremendous increase in the network capacity. Future networks will enable users to
be connected to the internet from anywhere at any time and with anything. This is
without neglecting that future transportation systems may play an integral role in wireless
networks by providing additional communication capabilities and becoming part of the
internet communication infrastructure. Thus, to achieve that sort of connectivity, it is
highly desirable to reduce the environmental impact of mobile communication systems
by intelligently deploying new wireless nodes to enhance UEs wireless connectivity.
Mobile-Femto technology is seen to be a promising solution for future networks to
improve passengers’ wireless connectivity inside public transportation [1, 2].
Nowadays, one of the most used transportation services that save time and effort are
High-Speed Trains (HSTs). HSTs are used all over the world e.g. the TGV Eurostar in
Europe, the Shinkansen train in Japan and the HST in China with a speed exceeds the
350 km/h [3]. Due to the well-shield carriages, the HSTs are more exposed to Doppler

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 653–666, 2021.
https://doi.org/10.1007/978-3-030-75100-5_56
654 R. Raheem et al.

frequency shift, high VPL (up to 40 dB), low Handover (HO) success rate, and the
UL/DL interference issues [3]. The European “Shift2Rail” project [4] has made it clear
that there is a persistent need for seamless wireless connectivity and high quality of
services to be provided for passengers in future rail development. Although plenty of
research has discussed GSM-R networks [5] and 5G networks [6]; not much researches
dealt with the communications network regarding future railway services. On the other
hand, Mobile Communications Enablers for Twenty-twenty (2020) Information Society
(METIS) is one of the few projects that are concerned with the possibility of improving
the wireless network connectivity by simply using low power small Base stations (BSs).
This project has also introduced the proposal of placing mobile BSs in vehicles, such as
cars, trucks, buses, and trains to improve vehicular UE performance.
Therefore, the main contribution of this paper is improving the performance and
signal quality of train passengers via installing small-cells technology inside train car-
riages. That is without neglecting the importance of discussing the main two challenges
here; mobility and interference that accompany the deployment of such technology.

2 Related Work
The European 5G project group (METIS) predicted that from 2020, most users who
require reliable connection to the internet will be train passengers where the number
will exceed 300 users’ devices per train. Whereas, the E-train projects from the Interna-
tional Union of Railways (UIC) confirmed that the railway services will spread widely
in the next few years and dominate people’s lives. In addition, there will be more railway
services that are focused on realizing the objective of smart, green, and integrated trans-
port systems [7]. That has created significant motivation to improve the performance of
passengers in public transportations, reduce signal outage, and mitigate interference in
future networks. Providing indoor coverage on HSTs with outdoor BSs is not the most
convenient solution due to high VPLs caused by the Faraday cage characteristics of
railcars [8]. This fact has led to poor signal quality inside train carriages were offering
broadband services is not always possible. Although, broadband access on trains could be
achieved by installing BSs close to the railways; however, this solution is not convenient
to train operators due to the high investment needed to deploy such equipment. Besides,
such solutions increase the number of unnecessary HOs; as a result, this has focused the
research community on offering solutions that take advantage of the existing wireless
infrastructure to propose efficient methods to manage mobility in a seamless way for the
end UEs. Hence, a cooperative moving Relay Node (RN) system deployed on HSTs has
been introduced to enhance the coverage area for HSTs passengers and to overcome the
VPL issues [9, 10]. In contrast, a novel HO scheme that utilises the Coordinated Mul-
tipoint Transmission (CoMP) in high-speed Femtocell networks has been proposed to
ensure seamless, deployable, and efficient HO procedure for the train passengers when
the train switches between different coverage areas [11, 12]. Thus, installing small cells
inside HSTs to serve train passengers is the future of next-generation networks. How-
ever, this solution comes with its deployment challenges e.g. interference and mobility
issues especially in the HST environment. Therefore, this paper pays more attention to
the previous two challenges and provides optimal solutions to improve signal quality
inside HST environments.
 Cooperative and Coordinated Mobile Femtocells Technology 655

3 Mobile Femto Scenario in High Speed Train Environment

Mobility management is an essential component in mobile cellular communication sys-
tems because it offers clear benefits to the end UE in the HST environments i.e. low delay
services to voice and real-time video connections. Thus, coming up with an effective
solution to overcome the above challenges was the main motivation behind proposing the
Cooperative and Coordinated Mobile-Femtos (CCMFs) system in HST environments.
The CCMFs would be implemented on trains or other large-spatial-dimensions vehi-
cles where they are connected to one another through the CrX2 coordination interface as
shown in Fig. 1. This allows high optimisation for the group HO procedure and Backhaul
(BH) link connection.

Fig. 1. CCMFs scenario in HST environments

However, one of the main challenges in deploying the CCMF technology is the exces-
sive interference caused by the different access and BH links. That has been classified by
the following DL and UL interference scenarios. The possible DL interference scenarios
in HST CCMF are illustrated in Fig. 2. Interference (1) is caused by neighboring eNBs’
signals to the macro UEs while the BH links interference (2) occurs because of the high
Transmission Power (TP) of neighboring eNBs to neighboring Mobile-Femtos. As in
interference (2); interference (3) is caused by the high TP of the serving eNB to the
Mobile-Femto UEs within its coverage. In contrast, interference (4) occurs when the
Mobile-Femto TP is very high to the extent that it causes interference to the macro UEs
within the same eNB coverage. Finally, interference (5) is the interference between the
Mobile-Femtos themselves.
On the other hand, the possible UL interference scenarios in HST CCMF are
demonstrated in Fig 3.
The UL interference (1) is caused by Macro UEs to other neighboring eNBs, while
interference (2) is caused by the BH UL transmission of the Mobile-Femto to neighboring
eNBs. Interference (3) occurs when the Mobile-Femto UE causes interference to the
Macrocell signals while interference (4) is caused by the macro UE to the access links.
As for DL, interference (5) occurs between the access links of the many deployed CCMFs
within the same Macrocell coverage.
656 R. Raheem et al.

Fig. 2. DL interference scenarios

Fig. 3. UL interference scenarios

However, interference is not the only challenge in the HST environments because
mobility plays another significant role in this environment. That is because; the for-
warding of the data from the source eNB (SeNB) to the target eNB (TeNB) is the main
goal of offering seamless HOs in future networks. Data forwarding typically takes place
over the X2 interface between eNBs. In the presence of Mobile-Femto deployments,
handing-over the Mobile-Femto itself between Macrocells and forwarding the data is
a major challenge in today’s networks. That HO procedure takes a place whenever the
SINR between the Mobile-Femto and the source eNB gets weaker. In that case, the
Mobile-Femto initiates a HO procedure to maintain its connectivity with the hosting
network. The choice of target eNBs depends on the hosting train path and direction as
discussed in our previous work [13]. However, the mobility of the installed CCMF in
HSTs makes it even more challenging for the cellular operator to control the generated
interference of these small cells. Therefore, there are several techniques that could be
 Cooperative and Coordinated Mobile Femtocells Technology 657

used by service providers to mitigate the interference between eNBs and CCMF and
between the CCMFs themselves in future networks.

3.1 Coverage Optimisation

An important challenge for the CCMFs is optimising their radio coverage area dynam-
ically. This has helped in achieving the desired level of performance for mobile trans-
mission, avoiding undesired interference, and reducing power consumption [14, 15].
Consequently, optimising the coverage plan procedure is important to have good radio
conditions everywhere within the vehicular environment. This can be achieved by miti-
gating the impact of path loss (PL) which in turn can be achieved by installing Mobile-
Femtos in appropriate locations with appropriate TP [16]. The NLOS PL equation of
the BH links is given by

PL(L) = 34.53 + 38 log10 (L) (1)

Where L is the distance in meters and no less than 20 m.

On the other hand, the LOS PL equation of the access links is given by

PL(L) = 30.18 + 26 log10 (L) (2)

The NLOS PL model is for the BH link between the Mobile-Femto and the eNB. In
contrast, a constant PL has been used for the LOS access link between Mobile-Femtos
and their served UEs within the same train carriage.

3.2 Transmission Power Control

Controlling the TP of eNBs and Mobile-Femtos is another efficient method to miti-

gate interference in current and future networks. However, insufficient TP with poor
Femtocell deployment leaves the network with dead zones (holes) [17]. Whereas, if
the Femtocell TP is high with dense Femtocells deployment, this can create excessive
interference between the different parties within the network. In the CCMFs scenario,
the channel between the Femtocell and its UEs is considered to be less affected by the
interfered signals from other BSs. This is due to the short distance between the transmit-
ter and the receiver. In addition, the high penetration loss of the train environment (well
shielded carriages) prevents outside signals from passing through to train passengers.
This makes it clear that train carriage chassis can work as a barrier to isolate train pas-
sengers from outside signals. Thus, train passengers can be served only by the installed
Femtocell within their carriage; as a result, the interference can be mitigated.
Hence, the previous two possible solutions have been used for years to reduce the
impact of interference in current and future networks. However, in HST environment
there is always worst case scenario due to the high mobility challenge. Therefore, to
overcome such scenarios, the following interference mitigation scheme is presented to
demonstrate its positive impact on future networks and vehicular UEs performance.
658 R. Raheem et al.

4 Proposed Interference Management Scheme

The frequency scheme shown in Fig. 4 divides the Macrocell coverage into center and
edge zones.
F0 sub-band is allocated for center zones, whereas F1 , F2 and F3 are assigned to
edge zones of Macrocell1 , Macrocell2 and Macrocell3 respectively. Thus, applying the
frequency reuse scheme between the CCMF in HST environment has positive impact in
mitigating the generated interference between transmission signals. That is because the
Mobile-Femto inside the train carriage allocates frequency sub-bands that have not been
used by Macrocells or other neighboring Mobile-Femtos. It can only reuse frequencies of
further Mobile-Femtos that would not cause interference to its served UEs. However, in
real life scenarios the frequency reuse scheme works by receiving Macrocell dimensions,
no. of Mobile-Femtos and locations. More network characteristics are required like BS
TPs, received power from serving and interfering cells and the white Gaussian noise.
Based on the received Macrocell characteristics and used TPs, the inner cell radius is
calculated to differentiate between centre and cell edge zones as illustrated in Fig. 5.
The following algorithm summaries the frequency reuse scheme discussed earlier.

Fig. 4. The PIMS

 Cooperative and Coordinated Mobile Femtocells Technology 659

Fig. 5. Inner cell radius

Algorithm 1. PIMS for Macrocell/CCMF

Create Network topology

for inner cell radius equal to 0: R
for frequency band equal to 0: Total-subcarriers
Allocate frequency band for macrocells
if inner cell radius >= distance mfemto
Mobile-Femto belongs to center zone
Allocate frequency band for mfemto
else
if inner cell radius < distance mfemto
Mobile-Femto belongs to edge zone
Allocate frequency band for mfemto;
endif
endif
for all UEs equal to 1 : U calculate
Signal to noise ratio
Capacity
Throughput
endfor
endfor
endfor

5 System Level
In order to study the impact of the PIMS on UEs; the SINR, throughput and capacity
will be calculated. The Macro and Mobile-Femto UEs can be interfered by the DL/UL
signals of neighboring eNBs and Mobile-Femtos. Therefore, on subcarrier n the received
Macro UE SINR is given by
SINRm(D),n =
 2
 
PeNB
x G1,eNB,n2  PL(x)ε (3)
neNB eNB  2 nMFemto MFemto
G  |G2,MFemto,n |2 +Pnoise
eNB =1 Px 1,eNB ,n + MFemto=1 Px
660 R. Raheem et al.

The serving eNB and the neighbouring eNB’ TPs are presented by PxeNB and PeNB
x
respectively. The channel gain between the Macro UE m and serving eNB is presented
by G1,eNB,n while the channel gain from neighbouring eNBs is denoted by G1,eNB ,n . In
contrast, the TP of neighbouring Mobile-Femtos is presented by PxMFemto whereas the
channel gain between the Macro UE m and the neighbouring Mobile-Femto is presented
by G2,MFemto,n . The PL model has already been calculated and it is expressed by PL
where x is the distance between the Macro UE and its serving eNB. The VPL and
the white noise power are presented by ε and Pnoise respectively. On the other hand,
the Mobile-Femto UEs can be interfered by eNB signals or any other neighbouring
Mobile-Femto signals, thus, on subcarrier n the received Mobile-Femto UE SINR is
given by

SINRmf(A),n =
PMFemto
x Closs (4)
neNB nMFemto MFemto  2
|G1,eNB,n | 2
+ MFemto =1 Px G 
2,MFemto ,n +Pnoise
eNB
eNB=1 Px

In Eq. (4), there is only a constant system loss since there is no channel gain over
the LOS access link with the served Mobile-Femto where the only existing channel
gain (G2,MFemto ,n ) is between the Mobile-Femto UEs and other neighbouring Mobile-
Femtos. In contrast, on subcarrier n, the Macro UEs capacity can be calculated as
 
Cm,n = BW. log2 1 + αSINRm(D),n (5)
where the divided bandwidth for subcarrier n is represented by BW and α is a constant
coding margin for target Bit Error Rate (BER) which is given by α = −1.5/ln(5BER).
As a result, the overall throughput of the serving eNB M is given by

ThroughputM = βm(D),n Cm(D),n (6)
m n

β m,n represents the assigned subcarrier for Macro UEs. When β m(D),n = 1, the subcarrier
n is assigned to Macro UE m, otherwise β m(D),n = 0. Thus, every Macro UE in every time
 m
slot is allocating different subcarrier within the same Macrocell. The Nm=1 βm(D),n = 1
for ∀ k as Nm is the no. of Macro UEs and k is the available Physical Resource Blocks
N
(PRBs). On the other hand, the expected mfmf=1 βmf (A),n = 3 for k ∈ FMobile-Femto where
the no. of Mobile-Femto UEs is given by Nmf and FMobile-Femto represents the available
sub-bands allocated to Mobile-Femtos. This makes it clear that the PIMS reuses the full
frequency bands three times in every Macrocell. As a consequence, the capacity of the
Mobile-Femto UE on subcarrier n is given by
 
Cmf,n = BW. log2 1 + αSINRmf(A),n (7)
On the other hand, the overall throughput of the serving Mobile-Femto can be given
by
 
ThroughputMFemto = βmf(A),n Cmf(A),n (8)
mf n

Hence, the subsequent results will show that the PIMS is greatly capable of preventing
the interference among the Macro and CCMFs UEs.
 Cooperative and Coordinated Mobile Femtocells Technology 661

6 Results and Discussion

The system-level simulator written in Matlab [18] has been used to evaluate the network
performance before and after implementing the PIMS in the HST environment. Figure 6
illustrates network topology and train path where the CCMFs will be installed inside
train carriages. In our scenario the x-position of the tracks does not change and the train
moves in a straight line throughout the cell edges of Macrocells at a constant speed of
100 km/h.

Fig. 6. Network layout

As it is shown above, the eNBs are uniformly deployed and have a 600 m cell radius
thus, the inter-site distance is 1200 m. Two loads have been considered 10 and 25 mobile
macro UEs. A full eNB buffer is considered where there are always buffered data ready
for transmission for each node. The urban NLOS Microcell model has been considered
which is more appropriate to describe the communication links and to eliminate the
interference issue as much as it is possible in such High speed environment. The back-
haul links use omni-directional antenna, instead of 7 dBi directional antenna. The access
link log normal fading Standard Deviation (STD) is 0 dB. The reason for those changes
is because the MFemto and the vehicular UE are inside. Table 1 represents some of the
considered parameters for simulating the required scenarios while more detailed tables
can be found in our previous work [19]. In our previous work we have discussed the
situations where choosing different parameters have an impact on the achieved perfor-
mance with an in depth comparison between different scenarios of being served by a
Macrocell, fixed or mobile femtocells.

6.1 Downlink Scenario

Figure 7(a) illustrates the gained throughput of cell edge UEs who are being served by
the CCMF when the PIMS is applied for the DL signals. The achieved results show that
the gain is greater for small MFemto TPs and small VPLs compared to the case when
there is no use of the PIMS. In that case, the interference from the Macro network is at
its highest as the low VPL does not isolate the indoor UEs from outdoor interference
signals. However, high VPL provides natural protection against outside interference.
662 R. Raheem et al.

Table 1. Simulation parameters

Parameter Value Parameter Value

System bandwidth 10 MHz No. of BH links 5
Traffic module Full buffer No. of Mobile-Femto UEs 1–5
Mobile-Femto type In-band, full duplex Train position Cell edge
No. of cells 19 VPL 10–30 dB
No. of Macro UEs 10–25 Mobile-Femto TP 0–20 dBm (DL)
No. of carriages 5 Cell edge UE PL PL > 120 dB

Thus, when the VPL and the TP are both high, the gain from the PIMS is low as the
SINR is high in this case. Whereas Fig. 7(b) clearly shows that using the PIMS does not
drop Macro UE’s performance significantly as the loss is estimated to be about 3% at
its most based on the used parameters. Please note that the different colors in the two
charts represent the achieved gain for different TPs (0, 10 and 20 dBm) and VPLs (10,
20 and 30 dB). The changes in each parameter were needed for experimental purposes
to study the impact of each on the achieve UE gain.

(a) (b)
Fig. 7. (a) Vehicular UE’s gain. (b) Macro UE’s gain

6.2 Uplink Scenario

Figure 8 illustrates the UL Macro UE’s performance in HST environment with and
without the used PIMS. It also shows the performance of the network without the Macro
UE Transmit Power Control (TPC). Consequently as expected, TPC improves the cell-
edge Macro UE performance, and the PIMS improves this even further.
On the other hand, Fig. 9 illustrates the performance for train passengers (Mobile-
Femto UEs) with and without the use of the PIMS.
 Cooperative and Coordinated Mobile Femtocells Technology 663

Fig. 8. UL performance of Macro UEs with the PIMS and 10 Macro UEs

Fig. 9. Vehicular UL performance with the PIMS and 10 Macro UEs

Hence, it is obvious that the achieved results reflect the improvement in both Macro
and Mobile-Femto UE performance after applying the PIMS. However, with the PIMS,
the 95% throughput of macro UEs has slightly deteriorated. This is due to the fact that
cell-edge UEs are given priority in the PRBs allocation while leaving fewer options for
the Macro UE. This is in fact a characteristic of most Fractional Frequency Reuse (FFR)
methods, which are more concerned with improving the cell-edge UEs performance.
However, the simulation was repeated by increasing the no. of UEs in the cell to 25 UEs
per cell. The results are shown in Fig. 10 and Fig. 11 for Macro UEs and train passengers
respectively.
When the 95% performance of Macro UEs and Mobile-Femto UEs are being com-
pared, it has clearly been noticed a large difference in the achieved performance between
the two. This is due to the fact that the BH capacity is divided between the Mobile-Femto
UEs inside the train carriages where each user gets an equal share. Furthermore, the PIMS
does not degrade the 95% TP compared to the case without the PIMS. The reason is that
the PIMS improves the BH performance; since the BH links are always being treated as
cell edge UEs, giving them an advantage.
664 R. Raheem et al.

Fig. 10. UL performance (Macro UEs) with the PIMS and 25 Macro UEs

Fig. 11. UL performance (vehicular UEs) with the PIMS and 25 Macro UEs

7 Conclusion

The future of the railway industry is expected to rely upon smart transportation systems
that provide train passengers with new services. These smart railways are more into
improving users experience when they are using public transportation. The CCMFs
technologies were the optimal solution to mitigate the impact of mobility, VPL, PL
while boost users performance and internet services inside train carriages. However,
every technology comes with its challenges and the main challenge in HST environment
is the interference. Therefore, a frequency reuses scheme was introduced to mitigate the
impact of the generated interference in HSTs environment and maintain signal services
inside train carriages.

Acknowledgments. The authors would like to thank all those who contributed to the completion
and success of this work. Also, the Faculty of Science and Technology at Middlesex University
played significant role in backing this work at all stages of the study.
 Cooperative and Coordinated Mobile Femtocells Technology 665

References
1. Raheem, R., Lasebae, A., Aiash, M., Loo, J., Colson, R.H.: Mobile Femtocells utilisation
in LTE vehicular Environments: vehicular penetration loss elimination and performance
enhancement. J. Veh. Commun. 9, 31–42 (2017)
2. Raheem, R., Lasebae, A., Aiash, M., Loo, J.: Mobility management for vehicular user
equipment in LTE/Mobile Femtocell Networks. Int. J. Inf. Syst. Serv. Sect. 9(4), 60–87 (2017)
3. Ai, B., He, R., Zhong, Z., Guan, K., Chen, B., Liu, P., Li, Y.: Radio wave propagation scene
partitioning for high-speed rails. Int. J. Antennas Propag. 2012, 7 (2012)
4. Shift2Rail: Rails Proejcts. https://www.shift2rail.org. Accessed 03 Jan 2021
5. Zhong, Z.D.: Key Issues for GSM-R and LTE-R. In: Dedicated Mobile Communications for
High-speed Railway. Advances in High-speed Rail Technology, pp. 19–55. Springer, Germany
(2018). https://doi.org/10.1007/978-3-662-54860-8_2
6. METIS project, Deliv. D1.1: Scenarios, Requirements and KPIs for 5G Mobile and Wireless
System, Technical report. https://cordis.europa.eu/docs/projects/cnect/9/317669/080/delive
rables/001-METISD11v1pdf.pdf. Accessed Apr 2014
7. HORIZON 2020 Work Programme 2018–2020. Smart, Green and Integrated Trans-
port Revised, EC Decision C(2020)6320. https://ec.europa.eu/programmes/horizon2020/en/
h2020-section/smart-green-and-integrated-transport. Accessed Sept 2020
8. Vivas, P.N., Campo, C., Rubio, C.G., Carrion, A.R.: Communication technology for vehicles.
In: 4th International Workshop, Nets4Cars/Nets4Trains, Vilnius, pp. 174–185 (2012). https://
doi.org/10.1007/978-3-642-29667-3
9. Scott, S., Leinonen, J., Pirinen, P., Vihriala, J., Phan, V.V., Latva-aho, M.: A cooperative mov-
ing relay node system deployment in a high speed train. In: IEEE 77th Vehicular Technology
Conference (VTC2013-Spring), Germany (2013). https://doi.org/10.1109/VTCSpring.2013.
6691818
10. Small Cells and infrastructure Blog: Small Cells on the Train – A 2 hop solution. https://sma
llcells.3g4g.co.uk/2015/07/small-cells-on-train-2-hop-solution.html. Accessed 05 Jan 2021
11. Chae, S., Nguyen, T., Jang, M.Y.: A novel handover scheme in moving vehicular femto-
cell networks. In: 2013 5th International Conference on Ubiquitous and Future Networks
(ICUFN), Vietnam (2013). https://doi.org/10.1109/ICUFN.2013.6614800
12. Pan, Meng-Shiuan., Lin, Tzu-Ming., Chen, Wen-Tsuen.: An enhanced handover scheme for
mobile relays in LTE-a high-speed rail networks. IEEE Trans. Veh. Technol. 64(2), 743–756
(2015)
13. Raheem, R., Lasebae, A., Loo, J.: Mobility management in LTE/Mobile Femtocell Networks:
outage probability and drop/block calls probability. In: IEEE 3rd International Conference
on Emerging Research Paradigms in Business and Social Sciences, ERPBSS-2015, Dubai,
United Arab Emirates (2015)
14. Ma, T., Pietzuch, P.: Femtocell Coverage Optimisation Using Statistical Verification. https://
lsds.doc.ic.ac.uk. Accessed 20 Jan 2021
15. Lu, Z., Bansal, T., Sinha, P.: Achieving user-level fairness in open-access femtocell-based
architecture. IEEE Trans. Mob. Comput. 12(10), 1943–1954 (2013). https://doi.org/10.1109/
TMC.2012.157
16. Alqudah, Y.A.: On the performance of Cost 231 Walfish Ikegami model in deployed 3.5
GHz network. In: Proceedings of the International Conference on Technological Advances in
Electrical, Electronics and Computer Engineering (TAEECE), Konya, Turkey (2013). https://
doi.org/10.1109/TAEECE.2013.6557329
17. Raheem R., Lasebae A., Aiash M., Loo J., Colson R. H.: Interference management and system
optimisation for femtocells technology in LTE and future 4G/5G networks. In: Advances in
Communications: Reviews’ Book Series, vol. 1. IFSA Publishing, Barcelona, Spain (2017)
666 R. Raheem et al.

18. Nt.tuwien.ac.at. Vienna LTE-A Simulators – nt.tuwien.ac.at (2021). https://www.nt.tuwien.

ac.at/research/mobile-communications/vccs/vienna-lte-a-simulators/. Accessed 13 Feb 2021
19. Raheem, R.: Interference management and system optimisation for femtocells technology in
LTE and future 4G/5G networks. Ph.D. Thesis submitted to Middesex Univerity. https://epr
ints.mdx.ac.uk/21255/1/RRaheemThesis.pdf. Accessed 14 Feb 2021
 Resource Allocation for Millimeter Wave
D2D Communications in Presence of
Static Obstacles

Rathindra Nath Dutta(B) and Sasthi C. Ghosh

Advanced Computing and Microelectronics Unit, Indian Statistical Institute,

203 B. T. Road, Kolkata 700108, India
{rathin r,sasthi}@isical.ac.in

Abstract. Device-to-device communications using millimeter-wave fre-

quencies has already been well acclaimed due to its improved spectral
reuse, higher data rate and lower delay. Whereas, such high frequency
communications with smaller wavelengths suffers from high propagation
and penetration losses, thus requires short distance obstacle-free line-of-
sight (LOS) communications. We allow relay aided communications in
cases where direct LOS communication path is blocked by some obsta-
cle. In this work we proposed an algorithm for efficient frequency channel
allocation for the direct, as well as relay aided, blockage free LOS com-
munications for real time application scenario. We also control the trans-
mission powers allocated to the devices for better channel allocation and
links activation, which in turn improves the overall system throughput.
Simulation results show that our proposed algorithm produces better
solutions in comparison to an existing work in terms of link activation,
system throughput, fairness and energy efficiency.

1 Introduction

Modern smart user equiptments (UEs) with their bandwidth intensive applica-
tions like video calls, video streaming services and mobile gaming have already
saturated the capacities of traditional cellular networks. To cope up with it,
several cooperative communication strategies have been proposed among which
communication in the millimeter-wave (mmWave) frequency has gained con-
siderable interest for enabling device-to-device (D2D) communication in next
generation cellular networks to satisfy such high data rate demand [10,13].
In D2D communications, two nearby devices are allowed to directly com-
municate with each other with limited or no involvement of the base station.
The D2D communication in mmWave frequency can dramatically improve the
spectral reuse, data rate, delay and energy efficiency. However, these mmWave
signals suffer from extremely high propagation and penetration losses making it
ideal for short distance obstacle-free line-of-sight (LOS) communications [9–11].
Using directional antennas and beamforming technique, one can establish LOS
communication links between two devices with almost zero interference to other
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 667–680, 2021.
https://doi.org/10.1007/978-3-030-75100-5_57
668 R. N. Dutta and S. C. Ghosh

devices outside the beamwidth [1,10]. Thus using beamforming and mmWave
one can have high speed pseudo-wire like communications within the beamwidth,
when LOS exists between the corresponding transmitter-receiver pair.
In cases where a direct communication suffers from blockage due to obsta-
cles, relay assisted communications are being used. Optimal relay selection and
resource allocation problems have been studied in [4,6]. However, these works
do not consider presence of obstacles in the communication paths. In case of
static obstacles, prior knowledge of obstacle locations can be made available by
satellite imaging or they can be learned [11]. The relay selection problem in the
presence of static obstacles have also been studied by several authors [5,11]. In
[5] authors consider a 3D model for the obstacles and find a relay-assisted path
giving maximum throughput among all candidate paths for backhaul communi-
cations. In [11] authors learn the obstacles without resorting to satellite imaging
and provide relay assisted LOS path when direct communication is blocked.
Be it direct or relay assisted, all active links must be assigned with frequen-
cies, making frequency allocation a challenging problem in D2D communications.
With limited number of available frequency channels, reuse of frequencies is a com-
mon practice. In this setting, frequency allocation to the active links while keep-
ing the interference level as low as possible, is a challenging task, in fact proven to
be NP-complete. The problem is more complex for relay aided real-time applica-
tion scenario, where data received at a relay device in a time slot must be sent out
immediately in the next time slot. The frequency allocation problem for D2D com-
munications has been explored in [4,6,7,13,15]. The work which is closest to ours,
is the resource allocation strategy investigated in [7]. Here the authors have pro-
posed a framework for efficient frequency channel allocation as well as controlling
the transmission power of the devices. They begin by a preliminary grouping of
the devices with respect to assigned frequencies and later reform these groups by
exchanging and adding new links. However, they neither considered the presence
of the obstacles nor allowed real time relay aided communications.
To the best of our knowledge none of the works including the ones mentioned
above have considered the frequency assignment problem with relay selection in
the presence of static obstacles under multiple frequency reuse scenario supporting
real time relay aided application scenario. To this end, we developed a framework
having the knowledge of static obstacle locations for a D2D overlay communica-
tion network in which we aim to optimally allocate channel resources and max-
imize the number of links activated to improve overall system throughput. We
propose an algorithm which allocates frequency channels to the requesting D2D
pairs having direct LOS communication paths. In cases where direct LOS path
is blocked due to obstacles we allow one-hop relay assisted LOS communications
whenever possible. We are considering a real-time application scenario where data
received at a relay device in a time slot must be sent out immediately in the next
time slot. Therefore our algorithm solves the problem for two consecutive time
slots because a subset of links activated in the first slot must also be activated in
the second slot. Moreover our algorithm incorporates power control mechanisms
for efficient channel allocation by maximizing the number of links activated while
maintaining energy efficiency and fairness of the resource allocation.
 Resource Allocation for D2D Communications 669

The rest of the paper is arranged as follows. The system model is given in
Sect. 2. In Sect. 3, we mathematically formulate the problem under considera-
tion. Our proposed algorithm is described in Sect. 4 and its simulation results is
presented in Sect. 5. Finally we conclude this work with Sect. 6.

2 System Model
We consider a single cell D2D overlay communications scenario in the presence
of static obstacles where the UEs can be classified into D2D pairs and relays
depending upon their intention to communicate. There are M D2D pairs who
are willing to communicate and N relays which are providing relaying services
to them. Half duplex communication is assumed and hence a UE can not com-
municate and relay simultaneously. We assume time is discretized into time slots
t0 , t1 , · · · with a small Δt time span for each slot. We denote a pair of successive
time-slots as a superslot. We also discretize the entire region as a grid of small
squared cells of size a × a.
Mobility Consideration: We consider all D2D devices as pseudo-stationary,
that is they do not change their position for the duration of a superslot. For a partic-
ular superslot, position of an UE can be determined with great accuracy [8]. Thus
we can solve the problem independently for each superslot. Position of a device in
some unit grid cell is approximated by the center of the small squared cell.
Obstacle Modelling: Each obstacle is approx-
imated by its minimum enclosing rectangle. We
assume that sizes and locations of these rectan- R B
gles corresponding to respective static obstacles
are known a priori through satellite imagery. The
grid cells containing some portion of an obstacle
A
are marked as blocked. In Fig. 1 all the grid cells
containing the obstacle, denoted by a black rect-
angle, has been marked in grey as blocked.
Fig. 1. Relay aided communi-
Direct and Relay Aided Communication: We cation in the presence of an
call an LOS link of a D2D pair to be blocked if obstacle
there is a blocked cell on the LOS path between
them. For a D2D pair if there is an obstacle-free LOS path they can commu-
nicate directly. Whereas, if their LOS path is blocked due to some obstacle in
between, the D2D pair may communicate via an intermediate relay providing
LOS path between them. We are only considering one hop relay assisted D2D
communications in such cases. We further assume that an idle UE can serve as
a relay for only one D2D pair. We denote D to be the set of D2D pairs having
LOS links and DR as the set of D2D pairs requiring a relay. As shown in Fig. 1
direct LOS path is blocked for the A − B D2D pair, which may communicate
via a relay R.
Candidate Relays: For a D2D pair, all idle UEs having LOS links within its
vicinity form the set of candidate relays for that pair. An idle UE can be present
670 R. N. Dutta and S. C. Ghosh

in more than one candidate sets, but can only be selected for a single D2D pair
for relaying.We denote Rk to be the set of candidate relays for the D2D pair k
and R = Rk .
k∈D R

Communication Channel: We assume mmWave directional antennas are

being used for the D2D communications. For this we use flat-top model of
beamforming with directional antennas [1,10] where mainlobe antenna gain
Gm = 2π/θ and sidelobe gain Gs = 0. Here θ is the beamwidth. We assume
the directional antennas are automatically aligned to each other based on the
device positions and we ignore the alignment overhead being negligible com-
pared to actual data transmission [2]. Following [1,12], the received power for
the transmitted power P can be computed as RTx ,Rx = kφGTx GRx P d−α Tx ,Rx ,
where dTx ,Rx is the distance between the transmitter (Tx ) and the receiver (Rx ),
λ
α is the pathloss exponent, k = ( 4πd 0
)2 with d0 and λ being the reference dis-
tance for the antenna farfield and the wavelength respectively, φ is the shadowing
random variable, GTx and GRx are the antenna gains at the transmitter and the
receiver respectively. If the receiver falls within beamwidth of the transmitter,
we set GTx = Gm , and GTx = Gs otherwise. Similarly, if the transmitter falls
within beamwidth of the receiver, we set GRx = Gm , and GRx = Gs otherwise.
Interference Consideration: Consider a D2D pair k = (k1 , k2 ), k ∈ DR where
device k1 needs to send some data to device k2 via relay r, r ∈ Rk . In the first
time slot of a superslot, k1 plays the role of a transmitter, r becomes the receiver
and k2 does not take part. In the second slot, r becomes the transmitter, k2 plays
the role of a receiver and k1 sits idle. While for a D2D pair i = (i1 , i2 ), i ∈ D
with device i1 directly transmitting to device i2 , i1 plays the role of a transmitter
while i2 receives. Such a D2D pair can be opportunistically scheduled in either
of the two time slots of a superslot. For a specific time slot, where a transmitter-
receiver pair (Tx , Rx ) is scheduled for transmission, TX may cause interference
to all active receivers other than Rx receiving in the same frequency channel as
of Rx only if those receivers fall inside the critical region of (Tx , Rx ) as defined
below.
Critical Region: Consider a transmitter Tx
transmitting with beamwidth θ to a receiver Rx
where the line joining Tx and Rx makes an angle
ψ with respect to the positive X axis. We define
critical region CRTx ,Rx of (Tx , Rx ) as the ψ ±θ/2
region as depicted in Fig. 2.

Fig. 2. Critical region

 Resource Allocation for D2D Communications 671

3 Problem Formulation
We define the following binary variables ∀ i ∈ D, k ∈ DR , r ∈ Rk , f ∈ F , s ∈
{1, 2}.

1 if i-th D2D pair communicates directly using frequency f in slot s
Xi,f,s =
0 otherwise

1 if k-th D2D pair communicates via relay r using frequency f in slot s
Yk,r,f,s =
0 otherwise

For a D2D pair i ∈ D transmitting using frequency f ∈ F , the interferences

during the first slot and the second slot of a superslot are given in (1) and (2)
respectively. Here Ra,b denotes the received signal power at device b transmitted
from device a.
  
Ii,f,1 = Rj1 ,i2 Xj,f,1 + Rk1 ,i2 Yk,r,f,1 (1)
j∈D ,i=j,i2 ∈CRj1 ,j2 k∈D R ,i2 ∈CRk1 ,r r∈R k

  
Ii,f,2 = Rj1 ,i2 Xj,f,2 + Rr,i2 Yk,r,f,2 (2)
j∈D ,i=j,i2 ∈CRj1 ,j2 k∈D R r∈R k ,i2 ∈CRr,k2

The corresponding SINR value is given in (3), where η0 denotes thermal noise.
Ri1 ,i2
SIN Ri,f,s = , ∀i ∈ D, f ∈ F , s ∈ {1, 2} (3)
η0 + Ii,f,s
Similarly, for a D2D pair k ∈ DR transmitting via relay r ∈ Rk using frequency
f ∈ F , the interferences during the first and the second slots of a superslot are
given in (4) and (5) respectively.
  
Iˆk,r,f,1 = Ri1 ,r Xi,f,1 + Rk1 ,r Yk ,r ,f,1 (4)
i∈D ,r∈CRi1 ,i2 k ∈D R ,k =k,r∈CRk ,r r  ∈R k
1

  
Iˆk,r,f,2 = Ri1 ,k2 Xi,f,2 + Rk1 ,k2 Yk ,r ,f,2
i∈D ,k2 ∈CRi1 ,i2 k ∈D R ,k =k r  ∈R k ,k2 ∈CRr ,k
2
(5)
For each k ∈ DR , r ∈ Rk , f ∈ F the SINR values for the two slots are given in
(6).

Rk,r,f,1 = Rk1 ,r Rk,r,f,2 = Rr,k2

SIN , SIN (6)
η0 + Iˆk,r,f,1 η0 + Iˆk,r,f,2
In order to satisfy the required data rate, the SINR value of each active link i
must be above some threshold Ti . This is formalized as the following constraints.

SIN Ri,f,s ≥ Ti Xi,f,s ∀i ∈ D, f ∈ F , s ∈ {1, 2} (7)

Rk,r,f,s ≥ Tk Yk,r,f,s ∀k ∈ DR , r ∈ Rk , f ∈ F , s ∈ {1, 2}
SIN (8)
672 R. N. Dutta and S. C. Ghosh

Furthermore, we must also have the following constraints on the binary vari-
ables.

(Xi,f,1 + Xi,f,2 ) ≤ 1, ∀i ∈ D (9)
f ∈F
 
Yk,r,f,s ≤ 1, ∀k ∈ DR , s ∈ {1, 2} (10)
r∈R k f ∈F
 
Yk,r,f,s ≤ 1, ∀r ∈ R, s ∈ {1, 2} (11)
k∈D R ,r∈R k f ∈F
 
Yk,r,f,1 − Yk,r,f,2 = 0, ∀k ∈ DR , r ∈ Rk (12)
f ∈F f ∈F

Constraint (9) ensures that a direct link can be scheduled at either of first or
second slot. Constraints (10) and (11) ensure that for relay aided communication,
a pair can choose only one relay and a particular relay can be chosen by at most
one pair, respectively. Constraint (12) ensures that if a source to relay link is
scheduled in the first slot, the relay to destination link must be scheduled in the
second slot.
For each k ∈ DR , r ∈ Rk the effective SINR of the two slots is computed as:

  
Rk,r = min(
SIN Rk,r,f,1 Yk,r,f,1 ,
SIN Rk,r,f,2 Yk,r,f,2 )
SIN (13)
f ∈F f ∈F

We aim to maximize the total number of links activated as expressed bellow.

⎧ ⎫
⎨      ⎬
max Xi,f,s + Yk,r,f,1 (14)
⎩ ⎭
i∈D f ∈F s∈{1,2} k∈D R r∈R k f ∈F

Throughput obtained from the active links can be computed using Shannon’s
capacity formula data rate = B log(1 + SIN R), where B is the channel band-
width.

4 Scheduling and Resource Allocation Algorithm

We begin by preprocessing the blocked cells to compute a visibility graph which
we later use to check for an obstacle-free LOS path between any two cells. With
the candidate relay sets computed for each relay aided D2D pairs we proceed
for the resource allocation. We first perform an initial frequency and power
assignment for the direct D2D pairs which is later refined with exchanges and
addition of new direct, as well as relay aided, D2D links. We also optimize
the power allocation during the frequency assignment process itself for efficient
resource utilization. Finally we compute the system throughput as per the final
frequency and power assignment.
 Resource Allocation for D2D Communications 673

Precomputing Visibility Graph

Consider two grid cells with indices (i1 , j1 ) and (i2 , j2 ) respectively and a line l
joining the centers of these two cells. We say two cells are visible to each other if
and only if the LOS path from any point in one cell to any point in the other cell
is not obstructed by any obstacle. Considering each cell is of size a × a, the two
cells are visible to each other if and only if there is no blocked cells within the
2a strip centered around the line l. We reduce the search space by considering
only the blocked cells in the 2D range [(i1 , i2 ), (j1 , j2 )]. The blocked cells in this
range can efficiently be obtained by storing all blocked cells in a 2D range tree
that uses fractional cascading [14].
For each cell we check visibility of all other cells within a circle of radius d.
Here d is the transmission range, i.e., the distance up to which signal strength
does not fall below the required threshold using the Friis free space propagation
model [3]. We store this data in a binary array visible of size O(hvd2 ), where h
and v are the number of horizontal and vertical grid cells respectively. Algorithm
1 formalizes this procedure.

Algorithm 1: Compute visibility graph

1 Store the blocked cells in a 2D range tree TR
2 Create the visible array
3 for i1 = 1 to h do
4 for j1 = 1 to v do
5 for i2 = i1 − d to i1 + d do
6 for j2 = j1 to j1 + d do
7 Run rectangular range query [(i1 , i2 ), (j1 , j2 )] on TR and store
the resultant cells in B
8 Let l be the line joining the center of grid cells (i1 , j1 ) and
(i2 , j2 )
9 if ∃b ∈ B such that b ∈ 2a wide strip around l then
10 visible[(i1 , j1 ), (i2 , j2 )] = visible[(i2 , j2 ), (i1 , j1 )] = f alse
11 else
12 visible[(i1 , j1 ), (i2 , j2 )] = visible[(i2 , j2 ), (i1 , j1 )] = true

Selecting Candidate Relays

An idle device r may act as a potential relay for a D2D pair k = (k1 , k2 ) ∈ DR
if and only if r is within the transmission range of k1 and k2 is also in the
transmission range of r. Therefore, in order to act as a potential relay for this
D2D pair, r must reside in the intersection of the two d radius circles centered at
the cells of k1 and k2 . To reduce the search space for finding the candidate relay
devices, we consider the minimum enclosing rectangle containing the intersection
of these two circles and perform a 2D range search with respect to this rectangle.
Figure 3 depicts this idea. Such an axis parallel rectangle can uniquely be
identified by any of its two opposite corners. In this figure P and Q are two
674 R. N. Dutta and S. C. Ghosh

such corners such that P.i < Q.i and P.j < Q.j, where A.i and A.j denote row
and column indices of the grid cell containing some point A. Let C1 and C2
be the points where the two circles intersect with each other. We can compute
P.i, P.j, Q.i and Q.j as follows.

max(k1 .i, k2 .i) − d if min(C1 .j, C2 .j) < k1 .j < max(C1 .j, C2 .j)
P.i =
Q min(C1 .i, C2 .i) otherwise

max(k1 .j, k2 .j) − d if min(C1 .i, C2 .i) < k1 .i < max(C1 .i, C2 .i)
P.j =
min(C1 .j, C2 .j) otherwise

min(k1 .i, k2 .i) + d if min(C1 .j, C2 .j) < k1 .j < max(C1 .j, C2 .j)
Q.i =
max(C1 .i, C2 .i) otherwise
P 
min(k1 .j, k2 .j) + d if min(C1 .i, C2 .i) < k1 .i < max(C1 .i, C2 .i)
Q.j =
max(C1 .j, C2 .j) otherwise

Fig. 3. Relays
The candidate relay sets Rk , ∀k ∈ DR can be obtained using the Algorithm
2, where dist(k1 , r) is the distance between k1 and r.

Algorithm 2: Candidate relay selection

1 Store all idle devices as potential relay nodes in a 2D range tree TR
2 foreach k ∈ DR do
3 Compute the two intersection points C1 and C2
4 Compute locations of P and Q as stated in Figure 3
5 Run a rectangular range query [(P.i, Q.i), (P.j, Q.j)] on TR and store the
results in Rk
6 foreach r ∈ Rk do
7 if dist(k1 , r) > d or dist(k2 , r) > d then
8 Rk = Rk \ r

The Scheduling and Resource Allocation Algorithm

With F number of orthogonal frequency channels available, we denote F1 =
{1, 2, . . . , F } and F2 = {F + 1, F + 2, . . . , 2F } as the set of frequencies avail-
able for the first and the second time slot respectively. We define F = F1 ∪ F2 as
the total frequency set for the two slots. We construct an interference graph G in
which vertices are the D2D pairs in D and we put an edge from vertex i to vertex
j with i = j if and only if there exists an obstacle-free LOS path from i1 to j2 and
j2 ∈ CRi1 ,i2 . Let F be the set of available colors. We apply a simple greedy color-
ing algorithm on G to color as many vertices as possible with the given set of col-
ors in F . With this greedy coloring, we color the vertices of the graph by ordering
the vertices by their degrees and allocate a conflict free color to each vertex when-
ever possible. More often than not it might be the case that the whole graph can-
not be colored with only |F | colors. This coloring represents an initial frequency
 Resource Allocation for D2D Communications 675

assignment. For each frequency f ∈ F we define a group Gf containing all those

D2D pairs whose corresponding vertices are colored with color f . Since two D2D
pairs belonging to two different frequency groups can not interfere with each other,
we handle each of these frequency groups independently of each other. The D2D
pairs having no frequency assigned to them are put into a set C . We now allo-
cate minimum possible power satisfying the required SINR threshold Ti to each
D2D pair i in Gf by using the following linear programming (LP) based power
control algorithm: minimize Pi subject to constraints Pmin ≤ Pi ≤ Pmax
i∈Gf
and Ri1 ,i2 ≥ Ti (η0 + Rj1 ,i2 ). We should note that the initial
j∈Gf ,i=j,i2 ∈CRj1 ,j2
frequency assignment obtained by the graph coloring has already taken care of
the interference, i.e., no device in a particular group Gf can interfere any other
device in the same group. But at later stages of our algorithm we possibly add
more D2D pairs into a particular group which might introduce interference in
that group. This justifies the presence of the interference term (the term) in
the second constraint in our power control LP. This initial frequency and power
allocation process is formalized as Algorithm 3. Next we revise each frequency
group Gf , f ∈ F by exchanging a D2D pair in Gf by a D2D pair in C such that
maximum power allocated to a D2D pair in Gf reduces. This in turn improves
the chance of adding more D2D pairs to Gf . We repeat this process until no such
exchange is possible. At this point we run Algorithm 5 to add more direct D2D
pairs from C into Gf . When there are no direct D2D pairs which can be added in
this way we move our attention to relay aided D2D pairs and invoke Algorithm 6.
This refinement process is formalized as Algorithm 4.

Algorithm 3: Initial frequency and power assignment algorithm

1 Create vertices corresponding to each pair i ∈ D for the interference graph G
2 foreach i ∈ D do
3 foreach j ∈ D do
4 if i = j and j2 ∈ CRi1 ,i2 then
5 Add an edge between vertices i and j into G

6 Apply greedy graph coloring on G

7 Set Gf = {i| vertex i gets color f } ∀f ∈ F
8 Set C = {i| vertex i does not get any color f ∈ F }
9 foreach f ∈ F do
10 Run linear programming based power control algorithm on Gf to find Pi for
i ∈ Gf

In Algorithm 5, to add more direct D2D pairs into a frequency group Gf

we adopt a simple greedy strategy. We find the D2D pair î from C causing
minimum interference into Gf while transmitting at Pmax . We then apply the
power control algorithm on Gf ∪ {î}. If there exists a feasible power allocation
676 R. N. Dutta and S. C. Ghosh

we permanently move î from C into Gf otherwise we discard î for this group.

We repeat this process with rest of the D2D pairs in C until no new pair can be
added into Gf . Here RTmax
X ,RX
denotes the received power at RX from transmitter
TX transmitting at Pmax .

Algorithm 4: Group refinement algorithm

1 foreach f ∈ F do
2 set Ctmp = C
3 while Ctmp = Φ do
4 let i be a D2D pair in Ctmp
5 set î = argmax {Pi }
i∈Gf

6 set Gtmp = Gf ∪ {i} \ {î}

7 run power control algorithm on Gtmp
8 if there is a feasible power allocation for Gtmp and max Pi < Pî then
i∈Gtmp
9 set Gf = Gtmp and C = C ∪ {î} \ {i}
10 set Ctmp = Ctmp \ {i}
11 run Algorithm 5 for Gf
12 run Algorithm 6

Algorithm 5: Adding direct D2D pairs

1 set Ctmp = C
2 while Ctmp = Φ do

3 set î = argmin Rimax
1 ,j2
i∈C tmp j∈Gf ,j2 ∈CRi1 ,i2

4 set Gtmp = Gf ∪ {î}

5 run power control algorithm on Gtmp
6 if there is a feasible power allocation for Gtmp then
7 set Gf = Gtmp
8 set Ctmp = Ctmp \ {î}
9 set C = C \ Gf

Algorithm 6 deals with relay aided D2D pairs from DR . To establish com-
munication of a D2D pair k ∈ DR via some relay r ∈ Rk we need to assign some
frequency f1 ∈ F1 for the D2D pair k  = (k1 , r) in slot 1 and some frequency
f2 ∈ F2 for the pair k  = (r, k2 ) in slot 2. This is equivalent to adding the D2D
pair k  into group Gf1 and k  into group Gf2 . We again adopt a greedy selection
strategy to schedule the relay aided D2D pairs. We select a D2D pair k̂ ∈ DR
with relay r̂ ∈ Rk̂ along with two frequencies fˆ1 ∈ F1 and fˆ2 ∈ F2 , such that
k  and k  cause minimum interferences to the corresponding groups among all
 Resource Allocation for D2D Communications 677

other possible choices. We again run the power control algorithm for the groups
Gfˆ1 ∪ {k  } and Gfˆ2 ∪ {k  }. If there exists feasible power allocations for both
cases, we permanently add k  and k  to the respective groups and mark r̂ as
used. We mark k̂ as done and do not consider it for further iterations. We repeat
this process until no more choice remains.
Based on the final frequency and power assignment we obtain the system
throughput as per the Shannon capacity formula. The worst-case time complex-
ity of the proposed scheduling and resource allocation algorithm is O(M F β)
where M is the number of requesting D2D pairs, F is the number of available
frequencies and β is the time complexity for solving LP with at most M number
of variables and 2M constraints.

Algorithm 6: Adding relay aided D2D pairs

1 set Ctmp = D R and Rused = Φ
2 while Ctmp = Φ do
3 k̂, r̂, fˆ1 , fˆ2 =
⎧ ⎫
⎨ ⎬
max max
argmin Rk 1 ,i2
+ Rr,j
k∈Ctmp ,r∈R k \Rused ,f1 ∈F 1 ,f2 ∈F 2 ⎩i∈G 2 ⎭
f1 ,i2 ∈CRk1 ,r j∈Gf ,j2 ∈CRr,k
2 2
4 set k = D2D pair (k̂1 , r̂) and k = D2D pair (r̂, k̂2 )
5 run power control algorithm on both Gfˆ ∪ {k } and Gfˆ ∪ {k } separately
1 2
6 if there is a feasible power allocation for both Gfˆ ∪ {k } and Gfˆ ∪ {k } then
1 2
7 set Gfˆ = Gfˆ ∪ {k } and Gfˆ = Gfˆ ∪ {k }
1 1 2 2
8 assign relay r̂ to D2D pair k̂
9 set Rused = Rused ∪ {r̂}
10 set Ctmp = Ctmp \ {k̂}

5 Simulation Results
We consider a simulation environment similar to [7]. We have taken a cell of size
500 m × 500 m with total 200 UEs among which requesting D2D pairs varying
from 30 to 95 and while the rest act as relays. The distance between a D2D pair
varies from 1 m to 20 m. For the mmWave D2D communications in 60 GHz fre-
quency we adapt the channel parameters from [10,12]. More specifically, channel
bandwidth is 20 MHz, beamwidth is 45◦ , the maximum transmission power is
24 dBm, SINR threshold is 15 dB, thermal noise is –174 dBm/Hz, pathloss expo-
nent is 1.88, and the shadowing random variable follows log normal distribution
with zero mean and 1.2 dB standard deviation. Here λ = fc where c is the speed
of light and f = 60 GHz. For the visibility computation we consider the grid
cells are of size 5 m × 5 m.
678 R. N. Dutta and S. C. Ghosh

We compare our proposed algorithm with the D2D resource allocation and
power control (DRAPC) algorithm presented in [7] in terms of number of D2D
pair links successfully activated, overall system throughput achieved, energy effi-
ciency and fairness of resource allocation. For the energy efficiency we consider
number of bits transmitted per unit power measured in bit/mJoule, while the
fairness of the throughput of the requesting pairs is computed following Jain’s
fairness index [7].
Since DRAPC does not consider presence of any obstacles, for a fair compar-
ison we first consider a simulation environment with no obstacles. As evident in
Fig. 4 our proposed algorithm outperforms the DRAPC algorithm in all of the
four criteria. This improvement can be attributed to the fact that, given a chan-
nel assignment, DRAPC uses a naive greedy strategy for power optimization
while our proposed algorithm optimally solves an LP for the same. By minimiz-
ing the power assignment to a frequency group our algorithm widens the scope
for packing more D2D pairs in the same group while satisfying the required
SINR.
Furthermore, we also consider two more simulation environments, one with
15 obstacles and another with 25 obstacles each of size 10 m × 10 m placed in
the area uniformly at random. In Fig. 5 we again observe the improvement of
the results as the number of obstacles are increased from zero to 15 and 25. This
improvement comes from the fact that, while the DRAPC algorithm can only
activate D2D pairs having direct LOS communication paths, our algorithm can
significantly improve the solution by considering relay aided communications.
Energy Efficiency

50
Active Links

40 0.8
30 0.7
0.6
20
40 60 80 100 40 60 80 100
Load Load
0.8
5
Throughput (Gbps)

Fairness

0.7
0.6
4
0.5
0.4
3 40 60 80 100
Load
2
40 60 80 100 Proposed (without obstacles)
Load DRAPC

Fig. 4. Experimental results with no obstacles

 Resource Allocation for D2D Communications 679

50

Energy Efficiency
Active Links
40 0.8
30 0.7
0.6
20
40 60 80 100 40 60 80 100
Load Load
0.8
0.7

Fairness
0.6
5 0.5
Throughput (Gbps)

0.4
4 40 60 80 100
Load
3 Proposed with 15 obstacles
DRAPC with 15 obstacles
2
40 60 80 100 Proposed with 25 obstacles
Load DRAPC with 25 obstacles

Fig. 5. Experimental results with varying number of obstacles

6 Conclusion
We have addressed the channel and power allocation problem in context of
mmWave D2D overlay communications in the presence of static obstacles. To
this end, we have proposed an algorithm which by extensive simulations is shown
to outperform the DRAPC algorithm in terms of number of links successfully
activated, system throughput achieved, energy efficiency and fairness of resource
allocation.

References
1. Cai, L.X., Cai, L., Shen, X., Mark, J.W.: Rex: a randomized exclusive region
based scheduling scheme for mmwave wpans with directional antenna. IEEE Trans.
Wireless Commun. 9(1), 113–121 (2010)
2. Congiu, R., Ghadikolaei, H.S., Fischione, C., Santucci, F.: On the relay-fallback
tradeoff in millimeter wave wireless system. In: IEEE Conference on Computer
Communications, INFOCOM Workshops, San Francisco, CA, USA, April 10-14,
pp. 622–627 (2016)
3. Friis, H.T.: A note on a simple transmission formula. Proc. IRE 34(5), 254–256
(1946)
4. Hoang, T.D., Le, L.B., Le-Ngoc, T.: Joint mode selection and resource allocation
for relay-based d2d communications. IEEE Commun. Lett. 21(2), 398–401 (2017)
5. Hu, Q., Blough, D.M.: Relay selection and scheduling for millimeter wave backhaul
in urban environments. In: IEEE 14th International Conference on Mobile Ad Hoc
and Sensor Systems (MASS), pp. 206–214 (2017)
6. Kim, T., Dong, M.: An iterative Hungarian method to joint relay selection and
resource allocation for d2d communications. IEEE Wireless Commun. Lett. 3(6),
625–628 (2014)
680 R. N. Dutta and S. C. Ghosh

7. Lai, W., Wang, Y., Lin, H., Li, J.: Efficient resource allocation and power control
for lte-a d2d communication with pure d2d model. IEEE Trans. Veh. Technol.
69(3), 3202–3216 (2020)
8. Moore, S.K.: Superaccurate GPS chips coming to smartphones in 2018. https://
spectrum.ieee.org (2017)
9. Pi, Z., Khan, F.: An introduction to millimeter-wave mobile broadband systems.
IEEE Commun. Mag. 49(6), 101–107 (2011)
10. Qiao, J., Shen, X.S., Mark, J.W., Shen, Q., He, Y., Lei, L.: Enabling device-to-
device communications in millimeter-wave 5g cellular networks. IEEE Commun.
Mag. 53(1), 209–215 (2015)
11. Sarkar, S., Ghosh, S.C.: Relay selection in millimeter wave d2d communications
through obstacle learning. In: 2020 International Conference on Communication
Systems Networks (COMSNETS), pp. 468–475 (2020)
12. Singh, D., Ghosh, S.C.: Mobility-aware relay selection in 5g d2d communication
using stochastic model. IEEE Trans. Veh. Technol. 68(3), 2837–2849 (2019)
13. Tehrani, M.N., Uysal, M., Yanikomeroglu, H.: Device-to-device communication in
5g cellular networks: challenges, solutions, and future directions. IEEE Commun.
Mag. 52(5), 86–92 (2014)
14. Willard, D.E.: New data structures for orthogonal range queries. SIAM J. Comput.
14(1), 232–253 (1985)
15. Yu, G., Xu, L., Feng, D., Yin, R., Li, G.Y., Jiang, Y.: Joint mode selection and
resource allocation for device-to-device communications. IEEE Trans. Commun.
62(11), 3814–3824 (2014)
 Bandwidth Slicing with Reservation Capability
and Application Priority Awareness for Future
Vehicular Networks

Abdullah A. Al-khatib1(B) , Abdelmajid Khelil1 , and Mohammed Balfaqih2,3

1 Institute for Data and Process Science, Department of Computer Science,
Landshut University of Applied Sciences, Landshut, Germany
{s-aalkha,khelil}@haw-landshut.de
2 Department of Computer and Network Engineering,
University of Jeddah, Jeddah, Saudi Arabia
3 Department of Automobile Transportation, South Ural State University, Chelyabinsk, Russia

Abstract. This paper addresses the bandwidth slicing for different types of con-
nected vehicle applications with diverse demands in heterogeneous networks. We
develop a joint network and device levels slicing scheme based on bandwidth
reservation and allocation optimization. The scheme enables vehicles to reserve
bandwidth in advance for their critical applications. It categorizes and prioritises
vehicular applications based on their degree of criticality. The reserved bandwidth
is first allocated. Then, the sub-optimal allocation of the remaining bandwidth is
determined and allocated according to degree of criticality of vehicular appli-
cations to achieve maximal resource utilization. Moreover, we employ machine
learning to accurately predict the time for bandwidth re-slicing, whereas exceeding
this time causes high cost of latency and consequently degrades the application per-
formance. Our proposed scheme outperforms the conventional slicing allocation
schemes in terms of efficiency and reliability of bandwidth utilization.

1 Introduction
The emergence of the fifth generation (5G) and beyond networks are expected to pro-
vide Vehicle-to-everything (V2X) communication and facilitate various new vehicular
applications such as cooperative, assisted, or autonomous driving. These applications
have great potential to address traffic related issues, including reducing traffic accidents
and congestion, and improving the comfort of passengers and the energy efficiency dur-
ing travel. Vehicular applications are considered as a real time traffic due to that, both
high network connectivity and high bandwidth are required. Besides, ultra-low latency
is needed for safety applications such as obstacle detection and lane change [1]. These
requirements complicate the problem of providing robust network resources on-time
and fulfilling the desired Quality of Service (QoS) [2].
To meet such crucial requirements, the 5G and beyond networks are intended to
adopt many emerging technologies while exploiting heterogeneous networks (HetNets).
The exploration of the unlicensed bands and underutilized spectrum through different
wireless access technologies has been proposed [3]. For instance, Long-Term Evolution

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 681–691, 2021.
https://doi.org/10.1007/978-3-030-75100-5_58
682 A. A. Al-khatib et al.

(LTE) devices can utilize WiFi unlicensed bands to support bandwidth critical applica-
tions. However, vendors and service providers in a HetNet must coordinate for resource
management, which is a very difficult process hindering innovations in network archi-
tecture. Thus, the concept of network slicing has been identified, in the first release of
5G normative specifications [4], to facilitate multi-tenancy, multi-service support, and
provide on-demand network services for different application scenarios. Network slicing
refers to creating multiple end-to-end virtual networks over a common infrastructure.
Each network slice is logically isolated to serve different types of applications with
diverse requirements.
Software Defined Network (SDN) and Network Function Virtualization (NFV)
paradigms are seen as promising enabler for network slicing, in which NFV enables
implementing network functions that are realized in costly hardware platforms as soft-
ware appliances placed on low-cost commodity hardwares or running in the cloud com-
puting environment [5]. This architecture separates control plane and data plane to realize
the flexible control of network traffic. Since the centralised SDN controller has a global
network view, it is deployed to dynamically make routing decisions that satisfy diverse
service demands with maximum cost-efficiency. Within a slice, the required bandwidth
and computation resources can be allocated on-demand depending on a set of QoS
parameters and other indicators such as latency, reliability, and energy efficiency.
Bandwidth slicing solutions are classified according to slicing level into network-
level and device-level solutions. The former refers to allocating the aggregated band-
width of a HetNet among the BSs. The latter is the process of allocating BS resources
to its associated vehicles which might occur based on reactive or proactive approaches.
The reactive allocation occurs after application performance degradation or significant
demand for more bandwidth. In addition, it is accomplished upon vehicle request and
negotiation with SDN controller resulting in high cost of signaling and latency. The
proactive allocation, on the other hand, is challenging in terms of the required vehicles
movement tracking and future position forecasting. It is necessary to ensure that the over-
head of employing proactive method is reasonable compared to its benefit. In addition,
resource allocation must be timely and accurate especially for delay-critical applications.
The early allocation could lead to a waste of resources while the late allocation might
lead to service interruption, and consequently degrading the user experienced service
quality. This paper is extents to our prior work (reservation-based slicing [6]), with focus
on a joint network and device level slicing scheme based on reservation and bandwidth
allocation optimization in HetNets. The solution enables vehicles to reserve bandwidth
in advance for their critical applications. Moreover, it categorizes and prioritises vehic-
ular applications based on their degree of criticality. With the softwarization function,
SDN controller aggregates the total bandwidth from HetNet. The SDN controller spec-
ifies how the bandwidth is allocated among the BSs. The network-level slicing rate for
each BS is determined according to calculated device-level slicing rates. The bandwidth
ratio of each BS is determined proactively in two steps: (i) The reserved bandwidth is
assigned to vehicles that reserved resources for their delay-critical applications, and (ii)
the sub-optimal allocation of the remaining bandwidth is determined and allocated based
on criticality degree of vehicular applications to achieve maximal resource utilization.
 Bandwidth Slicing with Reservation Capability and Application Priority Awareness 683

Moreover, a machine learning model is utilized to predict the accurate time of bandwidth
re-slicing, which consequently reduces the latency.
The remainder of the paper is organized as follows. We present network slicing
concept and types as well as the requirements of different vehicular applications. Next,
the related work is discussed and compared. Then, we describe our proposed priority
and reservation-based slicing scheme including the prediction model of re-slicing time.
Consequently, we present the performance evaluation of the proposed scheme. Finally,
we conclude the paper and give an overview of our future work.

2 Network Slicing Concept and Types

The network slicing utilizes available physical or virtual resource of wireless networks to
enable service providers to efficiently manage and customize their virtual networks, and
thereby guarantee the demanded QoS [7]. The network slicing can be classified based on
the type of resource and network function sharing into Core Network (CN) slicing and
Radio Access Network (RAN) slicing [8]. The CN refers to dividing networks into more
fine-granular network functions [9]. The functions are related to control plane and user
plane functionalities, such as mobility management, session management, and authenti-
cation. Each function becomes flexible, scalable, programmable, and auto configurable
and can be managed by SDN and NFV [10]. On the other hand, the RAN refers to divid-
ing the physical resources (bandwidth and power) of various Radio Access Technologies
(RAT) between multiple virtual slices. The bandwidth slicing is more challenging due
to the different nature and parameters of wireless resources, such as time/frequency
resources, frame size, and Hybrid Automatic Repeat Request (HARQ) options.
With respect to slicing level, the bandwidth slicing solutions are categorized into
network-level and device-level solutions. In the former, the bandwidth resources of a
HetNet are allocated among the BSs. In a process known as slicing, BS resources are
allocated to its associated users; this process may be based on either one of reactive or
proactive approaches. In reactive approach, the user first submits its demand for band-
width; only then does the process of negotiation between BS and SDN controller begin.
This process may result in significant latency, leading to service degradation during
the negotiation process. The latency period is dependent on the time taken during the
period of request and initiation of response. In contrast, in proactive approach, predic-
tion or reservation are used prior to the allocation of user demand of BS resources. The
latency is reduced, in proactive approach, by assigning the bandwidth resource proac-
tively to vehicles application based on requirements criteria to ensure optimized network
performance.
Implementing network slicing in vehicular networks enhances network agility and
reconfigurability. It enables a network operator to guarantee diverse QoS requirements
of different vehicular applications by composing slices flexibly. Based on the mini-
mum level required of reliability and latency, the applications can be categorized and
prioritized accordingly into three classes:
684 A. A. Al-khatib et al.

(i) Internet and infotainment applications (priority-III):

This class aims to facilitate more infotainment experience such as web browsing,
and video streaming. It is expected that such applications will be more on-demand upon
the widespread of automobile driving, where drivers need media consumption. The
requirement of low latency is not critical for these applications and the reliability is not
a concern [8]. The required data rate is 0.5 Mb/s for web browsing, and up to 15 Mb/s
for HD video [11].

(ii) Traffic safety and efficiency applications (priority-II):

The main objective of traffic safety and efficiency applications is to reduce the traffic
accidents and improve traffic efficiency. For instance, to notify vehicles in surrounding
area about emergency stop event to trigger safer behavior, a vehicle sends a message about
this event to Roadside Unit (RSU), which in turn relays the message to its surrounding
vehicles. Such applications require low latency (i.e., 100 ms maximum latency) and high
reliability (i.e., about 99%). The required data rate is about 1 Mb/s.

(iii) Autonomous driving related applications (priority-I):

They enable a vehicle to operate with little to no human assistance. For example, to
create an image of their surroundings and navigate traffic, bird’s eye view application is
used. They stream data from sensors such as cameras or radar at intersections to identify
eventual pedestrians or free places. Compared to traffic safety applications, autonomous
driving applications require ultra-low latency (i.e., 10 ms maximum latency), ultra-high
reliability (i.e., at least 99.999%), and high data rate (i.e., at least 10 Mb/s).

3 Related Work
Several efforts have been dedicated for bandwidth slicing in different use cases including
vehicular networks. Allocating the network bandwidth statically between operators was
the main slicing. Due to the variation of users and load distribution on idea of the
earliest bandwidth slicing solutions for network-level each cell, dynamic bandwidth
allocation solutions have been proposed based on the allocated device-level slices which
are determined reactively or proactively.
A reactive bandwidth allocation scheme for two-tier HetNet Networks was proposed
in [12]. Optimal bandwidth allocation ratios of each BS were determined based on the
service type of their end users to maximize the network utility. Although two types of
applications were considered, there is no prioritization of critical applications. Moreover,
the scenario of having multi-association was not considered. An architecture of Mobile
Edge Computing (MEC) and SDN was proposed in [13] for autonomous vehicles to
address the issue of high data traffic volume and high number of computing/storing
tasks. A dynamic multi-resource management scheme by migrating computing task and
radio bandwidth slicing was introduced. An algorithm of dynamic bandwidth allocation
among multiple virtual wireless operators and service providers was presented in [14].
 Bandwidth Slicing with Reservation Capability and Application Priority Awareness 685

A weighted proportionally fairness/protection criteria was developed and considered to

achieve desirable fair bandwidth allocation. Pareto-optimality of user association to base
stations was also considered. The fairness was considered in terms of utility gains and
capacity savings across different operators and service providers and their associated
users. The main limitation of the proposed algorithm is the high complexity.
Several network slicing solutions have been proposed to allocate bandwidth reac-
tively or proactively upon forecasting the network load and bandwidth usage in the next
time window. However, the main limitation of adopting forecasting models is the high
cost and complexity due to the need of vehicle movement tracking and predicting their
future locations. A new metric was defined based on the available bandwidth that can be
allocated to a single very active bearer. In [15], an optimization problem with the objec-
tive of minimizing network overheads while satisfying QoS requirements of each virtual
network was formulated. A regressive moving average prediction method was developed
to obtain the optimal allocation ratios with the consideration of load forecasting.
The authors in [15] proposed a mechanism for 5G Network Slice Broker that involves
network slicing traffic forecasting and admission control. The slice forecasting module
was developed to analyse the network slices traffic patterns and provide forecasting
information to the admission control module. Based on this information, admission
control policies are applied to take a decision on tenants requests for the next time
window. Another solution for 5G Network Slice Broker was proposed in [16] to measure,
predict usage, and provision slice resources by deriving the throughput of an individual
wireless link. A short-term prediction model was developed to predict resource usage
by a slice based on historical collected data of the network.

4 Bandwidth Slicing Scheme with Reservation Capability

and Application Priority Awareness

The vehicles in the proposed scheme can reserve the needed bandwidth proactively. The
reservation process must occur within a suitable time before the actual need. At the
Fig. 1, a slice manager in charge of designing a proper bandwidth slice for the vehicles
and interacts with the SDN controller to deploy the bandwidth slicing.
Vehicle vi (i = 1, 2,…, n) issues a request to slice manager to reserve bandwidth
slicing for a certain period at a certain time t and location in which vi will be located
at t. We categorize, and prioritise vehicular applications based on the minimum level
required of reliability and latency into three classes. Each vi has a positive constant
weightω, where ω indicates that the corresponding vehicular application has a first,
second or third priority class. The bandwidth slicing scheme is initiated by aggregating
the total network bandwidth from all BSs. The network-level slicing is determined upon
obtaining the total reserved bandwidth βrr by vehicles and number of active associated
vehicles at each BS. First, the slice manager obtains the reserved bandwidth rate at each
BS. If the total reserved bandwidth did not exceed the total bandwidth of HetNet, the
slice manager then determines the sub-optimal allocation rate of the remaining bandwidth
β rmg at each BS based on the applications of the associated vehicles and their priorities.
To obtain the sub-optimal allocation of β rmg , an optimization problem is formulated.
The optimization objective is the utility maximization (see Fig. 2) of β rmg for various
686 A. A. Al-khatib et al.

Virtual Resources

Slice Manager Application Priority III

Logical link Application Priority II

Wired Backhaul link
SDN
Application Priority I
Controller

Obstacle Distance
(b) (a)

Fig. 1. Bandwidth slicing in two-tier HetNet

vehicle applications with different priorities, which is the summation of utilities achieved
by all vehicles with various priorities. The network utility maximization depends on
number of associated vehicles at each BS, bandwidth reservation ratio, and bandwidth
device-level slicing ratio. The input of the formulated problem includes:

• β t : The total aggregated network bandwidth from all BSs.

• βrr: The ratio of total reserved bandwidth.
• ω: The weight of vehicular application representing their prioritization level.
• X BS : The association pattern (X BS ), where X BS equal one if vi is associated with the
BS and X BS equal zero if not.
• fr : The sub-optimal device-level slicing ratio received from the BS for vi , which is
determined according to their X BS , the priority level of their application, and the total
number of associated vehicles with the BS.

The following constraints are considered in the problem formulation to maximize

the utilization of β rmg :

• Each vehicle can only associate with one BS during each bandwidth allocation period.
It is assumed that there is no vehicle with multi association patterns.
• The association pattern must be binary in which a vehicle is either associated to a BS
or not.
• The allocated bandwidth for each slice is the lowest acceptable bandwidth, which is
not less than periodic packets arrival rate.
• The allocated bandwidth guarantees that the achievable rate is not less than the
effective bandwidth.
• The summation of ratio βrr and ratio β rmg is equal to one.
 Bandwidth Slicing with Reservation Capability and Application Priority Awareness 687

Fig. 2. Formulation components of the remaining bandwidth β rmg optimization.

A case study of two-tier HetNet is shown in Fig. 1. The network consists of multiple
Macro BSs and RSUs with an overlap coverage. The Macro BSs (Bm ) are placed at the
first tier. In the second tier, small RSUs (Rs ) cells are placed. The Bm and Rs are directly
connected to edge routers of the core network via wired backhaul links.
Alternative Concave Search (ACS) algorithm [17] is utilized to solve the optimization
problem. It is a maximization method based on block-relaxation, where only the variables
of an active block are optimized while the variables of the other blocks are fixed. In this
case study, the bandwidth slicing ratios β Bm and β Rs are firstly fixed to obtain sub-
optimal association patterns. Then, the sub-optimal bandwidth slicing ratios β Bm and
β Rs (where β Bm and β Rs are β rmg for Bm and Rs respectively) are obtained by fixing the
association patterns. The ACS algorithm iteratively solves for sub-optimal bandwidth
slicing ratios and association patterns.

5 Prediction of Re-slicing Time

To re-allocate the total bandwidth between BSs, the re-slicing process must be conducted
in a certain time window. The βrr will be assigned to vehicles based on their reservation
time and location. However, the re-slicing of the β rmg must be timely accurate especially
in the scenario of delay-critical applications. The early re-slicing of β rmg could lead to
a waste of resources while the late re-slicing might lead to service interruption, and
consequently degrading the user experienced service quality [10].
An example of such scenario is shown in Fig. 1. Vehicle vi acquires the allocated
bandwidth slicing that satisfies QoS requirement of its current applications. For example,
the slice manager during the slice reservation allocated 10 Mbps to Vehicle at position
(a). Vehicle was predicted to transmit an image of unknown obstacle ahead with 500 ms
maximum latency requirements to its Edge computing to identify the obstacle. However,
Vehicle at the position (b) decided to change lane and will use lane changing warning
688 A. A. Al-khatib et al.

service with 100 ms maximum latency requirement. When Vehicle starts another appli-
cation with lower latency requirement (i.e., higher QoS requirements), the slice manager
in such scenario must be aware of maximum allowable latency of vehicle applications
and increase the allocated bandwidth slice.
It is significant for the slice manager to predict the accurate time of bandwidth re-
slicing. Exceeding this time causes higher latency and consequently degrades the QoS
of applications. We employ a machine learning model for predicting the re-slicing time.
We use a traffic simulation-based dataset as an input flow to train the machine learning
model. Here, we apply our re-slicing time prediction model in the scenario of obstacle
detection application. First step starts with the generation of the simulation dataset that
consists of one million rows. Each row consists of five fields, Vehicle-ID, Section-ID
for each road network link, a velocity, a distance, and the critical moment which is
used in traffic management decision-making process. The critical moment is the time
that delay-critical application performance will start degrading if the re-slicing process
is not performed. The value of the critical moment depends on the type of vehicular
application. The dataset considers Vehicle-ID, Section-ID for each road network link,
velocities within the range (20–220) (km/h) and distances to an obstacle within the range
(20–220) (meters). The dataset is divided into training dataset and testing dataset.
The training dataset is used in the second step to train different machine learning mod-
els and determine the critical moment of obstacle detection application. The examined
models are Naive Bayes (NB), Decision Tree (DT), K-Nearest Neighbor (KNN), Sup-
port Vector Machine (SVM), and Random Forest (RF). These machine learning models
learn to predict the suitable re-slicing time without resource wasting or degrading QoS
of applications. The accuracy of these predictions depends on how well the model can
be generalized from the training data. The first and second steps are time-consuming
processes and thus they are performed offline in advance. They are performed once,
then the trained machine learning models are saved and reused later. During the training
phase, the KNN model outperformed the other model, thus it is adopted in our solution.
In the third step, the re-slicing solution loads the trained machine learning model and
decides the time of bandwidth re-slicing based on the data given in the first arrival flow.

6 Performance Evaluation

The case study presented in Fig. 1 is utilized to demonstrate the effectiveness of the
proposed joint network and device levels bandwidth slicing scheme over two-tier HetNet.
The transmit power is set to 40 dBm for each Bm and 30 dBm for each Rs . The distance
between each Bm and its under-laid RSU is set as 400 m, as recommended by 3GPP [4].
The locations are fixed for Bm and Rs . The total number of vehicles is varied from 50
to 300 vehicles and uniformly distributed between Bm and Rs according to a Poisson
Point Process (PPP). The total reserved bandwidth rate is set to 10%. The random data
arrivals and current data queue backlogs of the Macro BS is 20 packet/s, and the average
rate packet arrivals at each queue of either the Macro BS or an RSU is 5 packet/s. The
default parameters values are obtained from 3GPP [4].
Figure 3 studies the effect of different reservation rates in each Rs . The number of
vehicles in each Rs is increased gradually by 25 vehicles. The results show that the
 Bandwidth Slicing with Reservation Capability and Application Priority Awareness 689

Fig. 3. Sub-optimal slicing ratio of β Rs with Fig. 4. Sub-optimal slicing ratio of β Rs

different βrr ratio achieved by different slicing solution with
variations of number of vehicles with priority
I application.

Fig. 5. Network utility of different bandwidth slicing solutions.

sub-optimal bandwidth ratio of the remaining bandwidth β Rs for Rs increases gradually

till the number of vehicles reaches the peak value (i.e., the bandwidth is completely
allocated). After the peak value, the sub-optimal ratio is decreased gradually because
the added vehicles share the slices with the existing vehicles in the network. The peak
value is reached with a smaller number of vehicles when the reserved slicing rate βrr is
higher. The peak values are 50, 75, 125, and 250 vehicles when the reserved slicing rates
βrr are 10%, 30%, 50%, and 70%, respectively. The number of vehicles that reserve the
bandwidth for their applications is a key decision to obtain a good system performance.
Different scenarios have different suitable reserved slicing rate depending on the number
of the vehicles and network parameters.
690 A. A. Al-khatib et al.

Figure 4 compares the sub-optimal slicing ratio of the remaining bandwidth β Rs in

our proposed scheme, dynamic slicing, and static slicing. The number of vehicles with
priority I application in each RSU is gradually increased in 25 vehicle increments. In
static slicing, fixed rate from the aggregated bandwidth is allocated to the RSUs. Thus,
increasing the number of vehicles with priority I application has no effect on the sub-
optimal slicing ratio. The bandwidth is totally allocated when the number of vehicles with
priority-I reaches 250 vehicles in the proposed scheme and 150 vehicles in the dynamic
slicing scheme. This is because our scheme allocates more bandwidth to vehicles with
priority-I application instead of other applications, which consequently decreases the
chance for applications with low priority to be associated with their preferred network.
The bandwidth resource might not be fully utilized because of the mismatch between
the demand and actual bandwidth allocation. Hence, the achieved network utility of the
macro BS by different slicing solution is evaluated with variations of number of vehicles
as shown in Fig. 5. The network utility, and thus the network throughput of the bandwidth
using the proposed scheme is significantly improved compared to static slicing. This can
be explained as the static slicing allocates the bandwidth to BSs with a fixed rate without
consideration of the demands of its associated vehicles applications. The network utility
of the proposed scheme is higher if the reserved bandwidth is fully utilized because the
vehicles applications requirements are considered. However, the network utility will be
lower if the reserved bandwidth is not utilized or partially utilized.

7 Conclusion
In this article, we developed a joint network and device level slicing scheme based on
reservation and bandwidth allocation optimization. The scheme categorizes and priori-
tises vehicular applications based on their degree of criticality. Accordingly, the vehicles
reserve bandwidth in advance for their critical applications. First, the slice manager allo-
cates the reserved bandwidth, then obtains and allocates the sub-optimal of the remaining
bandwidth according to degree of criticality of vehicle applications to achieve maximal
resource utilization. Moreover, a machine learning model is employed to predict the
accurate time of bandwidth re-slicing. Exceeding this time causes higher latency and
consequently degrades the fidelity of applications. Finally, the numerical results con-
firmed that the proposed scheme allocates the demanded bandwidth for delay-critical
applications more efficiently than the current slicing solutions. Our future work will
investigate the reservation pricing possibility and the scenario of reservation cancelation.

References
1. Chebaane, A., Khelil, A., Suri, N.: Time-critical fog computing for vehicular networks.
Chapter in Wiley Book. Fog Comput. Theory Pract. (2020). ISBN: 978-1-119-55169-0.
https://onlinelibrary.wiley.com/doi/abs/10.1002/9781119551713.ch17
2. Palattella, M.R., Soua, R., Khelil, A., Engel, T.: Fog computing as the key for seamless
connectivity handover in future vehicular networks. In: Proceedings of the 34th ACM/SIGAPP
Symposium On Applied Computing (SAC) (2019)
 Bandwidth Slicing with Reservation Capability and Application Priority Awareness 691

3. Huang, J., Qian, F., Gerber, A., Mao, Z.M., Sen, S., Spatscheck, O.: A close examination
of performance and power characteristics of 4G LTE networks. In: Proceedings of the 10th
International Conference on Mobile Systems, Applications, and Services (MobiSys), pp. 225–
238 (2012)
4. Study on Management and Orchestration of Network Slicing for Next Generation Network
(Release 15): Document TS 28.801, 3GPP (2018)
5. Kreutz, D., Ramos, F.M., Verissimo, P.E., Rothenberg, C.E., Azodolmolky, S., Uhlig, S.:
Software-defined networking: a comprehensive survey. In: Proceedings of the IEEE, vol.
103(1), pp. 14–76 (2014)
6. Al-Khatib, A.A., Khelil, A.: Priority-and reservation-based slicing for future vehicular net-
works. In: Proceedings of the 3rd International Workshop on Advances in Slicing for
Softwarized Infrastructures (S4SI@NetSoft), pp. 36–42 (2020)
7. Mei, J., Wang, X., Zheng, K.: Intelligent network slicing for V2X services toward 5G. IEEE
Netw. 33(6), 196–204 (2019)
8. Zhou, G., Zhao, L., Liang, K., Zheng, G., Hanzo, L.: Utility analysis of radio access network
slicing. IEEE Trans. Veh. Technol. 69(1), 1163–1167 (2019)
9. Ksentini, A., Frangoudis, P.A.: Toward slicing-enabled multi-access edge computing in 5G.
IEEE Netw. 34(2), 99–10 (2020)
10. Wang, G., Feng, G., Quek, T.Q., Qin, S., Wen, R., Tan, W.: Reconfiguration in network
slicing—optimizing the profit and performance. IEEE Trans. Netw. Serv. Manage. 16(2),
591–605 (2019)
11. Study on LTE-based V2X Services (Release 14): 3GPP TR 36.885 Std. (2016)
12. Ye, Q., Zhuang, W., Zhang, S., Jin, A.L., Shen, X., Li, X.: Dynamic radio resource slicing for
a two-tier heterogeneous wireless network. IEEE Trans. Veh. Technol. 67(10), 9896–9910
(2018)
13. Peng, H., Ye, Q., Shen, X.S.: SDN-based resource management for autonomous vehicular
networks: a multi-access edge computing approach. IEEE Wirel. Commun. 26(4), 156–162
(2019)
14. Caballero, P., Banchs, A., De Veciana, G., Costa-Pérez, X.: Multi-tenant radio access network
slicing: statistical multiplexing of spatial loads. IEEE/ACM Trans. Netw. 25(5), 3044–3058
(2017)
15. Sciancalepore, V., Samdanis, K., Costa-Perez, X., Bega, D., Gramaglia, M., Banchs, A.:
Mobile traffic forecasting for maximizing 5G network slicing resource utilization. In: Pro-
ceedings of the IEEE Conference on Computer Communications (INFOCOM), pp. 1–9
(2017)
16. Gutterman, C., Grinshpun, E., Sharma, S., Zussman, G.: RAN resource usage prediction for a
5G slice broker. In: Proceedings of the Twentieth ACM International Symposium on Mobile
Ad Hoc Networking and Computing, pp. 231–240 (2019)
17. Gorski, J., Pfeuffer, F., Klamroth, K.: Biconvex sets and optimization with biconvex functions:
a survey and extensions. Math. Meth. Oper. Res. 66(3), 373–407 (2007)
 Proposal of a Disaster Response AI System

Tomoyuki Ishida(B) and Tenma Hirose

Fukuoka Institute of Technology, Fukuoka, Fukuoka 811-0295, Japan

[email protected], [email protected]

Abstract. In the event of a large-scale natural disaster, the Disaster Countermea-

sures Headquarters refers to past disaster information and disaster response that
are not stored in a database. It usually takes a lot of time to refer to this infor-
mation. Therefore, there is a possibility that disaster response will be delayed.
In this study, we have developed a method that automatically proposes disaster
response such as shelter operation and transportation of relief supplies using AI
technology from information such as past disaster response records and regional
disaster prevention plans.

1 Introduction
In the event of a disaster, the disaster countermeasures headquarters is an organization
established in the national, prefectural, and municipal governments to smoothly collect
and transmit disaster information and make decisions on disaster response actions. The
disaster countermeasures headquarters must function effectively to respond promptly
and appropriately. When a disaster occurs, the disaster countermeasures headquarters
will be busy with various disaster responses, such as evacuation shelter management,
management and provision of relief supplies, and public relations activities. However,
the Cabinet Office (Japan) has been developing an “Integrated Disaster Management
Information System [1]” for sharing disaster prevention information among disaster
prevention related organizations since 2005. It is not easy to transfer technology to
municipalities since this system is a large-scale system. Besides, the Ministry of Inter-
nal Affairs and Communications started the operation of “L-alert (Public Information
Commons) [2]” in 2011, which quickly and accurately conveys disaster information to
residents. This system does not have a viewpoint of disaster response at the disaster
site level. Currently, there are few approaches to support using information technology
based on the disaster countermeasures headquarters of each municipality.

2 Previous Study
Takahagi et al. [3–5] developed comprehensive disaster prevention and mitigation infor-
mation system consisting of a disaster information registration system, a disaster infor-
mation sharing system, and a disaster information transmission common platform. The
disaster information registration system digitizes various disaster information handled

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 692–700, 2021.
https://doi.org/10.1007/978-3-030-75100-5_59
 Proposal of a Disaster Response AI System 693

by the disaster countermeasures headquarters. The disaster information sharing system

aggregates and analyzes the digitized disaster information. Besides, the disaster infor-
mation transmission common platform quickly distributes disaster information to res-
idents via various media and information transmission tools. However, Hirohara et al.
[6] developed a decision-making support cloud system for disaster countermeasures
using an interactive large-scale ultra-high-definition display. This system supports the
organization of enormous amounts of information and decision-making by displaying
digitized disaster information for each content on a large display installed at the disaster
countermeasures headquarters. However, both studies did not consider the utilization of
regional disaster prevention plans or past disaster response records. Thus, we implement
a disaster response AI system that proposes a disaster response at the disaster counter-
measures headquarters by AI using the disaster response job data of Chiba Prefecture in
Typhoon Faxai and Hagibis in 2019.

3 System Architecture

Figure 1 shows the system’s architecture. The system’s architecture consists of the appli-
cation server group and database server group. The application server group consists of a
disaster response operation management server, disaster response operation evaluation
server, and disaster response operation suggestion server. The database server group
consists of disaster information, disaster response records, local government staff, and
relief supply information storage.

Fig. 1. System architecture of the disaster response AI system.

694 T. Ishida and T. Hirose

4 Prototype System
Figure 2 shows the menu screen of the disaster response AI system. The user can reg-
ister disaster information, relief supply information, and staff information for disas-
ter response by selecting the “Register Disaster Information,” “Register Relief Supply
Information,” and “Register Staff Information” menus at the bottom of the screen. When
the user selects “Register Disaster Information” from the menu screen of the disaster
response AI system, the screen transitions to the disaster information list screen (Fig. 3).
The user can register new disaster information by selecting the “Add” button on the dis-
aster information list screen. Figure 4 shows the new disaster information registration
screen. When the user edits the registered disaster information record, the disaster infor-
mation edit screen is displayed by selecting the “Edit” button on the disaster information
list screen. Furthermore, when the user deletes the registered disaster information record,
the disaster information deletion screen is displayed by selecting the “Delete” button on
the disaster information list screen.
When the user selects “Register Relief Supply Information” from the menu screen of
the disaster response AI system, the screen transitions to the relief supply management
screen (Fig. 5). The user can register new relief supply information by selecting the “Add”
button on the relief supply management screen. Figure 6 shows the new relief supply
information registration screen. On the relief supply information registration screen, the
relief supplies are managed by registering the number of relief supplies stockpiled by
the municipality.

Fig. 2. Menu screen of the disaster response AI system.

 Proposal of a Disaster Response AI System 695

Fig. 3. Disaster information list screen for registering disaster information.

Fig. 4. New disaster information registration screen.

Fig. 5. Relief supply management screen.

696 T. Ishida and T. Hirose

Fig. 6. Relief supply information registration screen.

When the user selects “Register Staff Information” from the menu screen of the dis-
aster response AI system, the screen transitions to the staff management screen (Fig. 7).
The user can register new staff information by selecting the “Add” button on the staff
management screen. Figure 8 shows the new staff information registration screen.
 Proposal of a Disaster Response AI System 697

Fig. 7. Staff management screen.

Fig. 8. Staff information registration screen

On the disaster response proposal screen (Fig. 9) of the disaster response AI system,
the user can search the disaster response proposal derived from the past disaster response
by AI. Based on the past disaster response evaluation, the number of relief supplies to
be dealt with in the future is visualized as a disaster response proposal on this screen.
698 T. Ishida and T. Hirose

Fig. 9. Disaster response proposal screen.

Figure 10 shows the input/output sequence of the disaster response AI system. When
the user registers disaster information in the disaster response AI system, the disaster
information is stored in the database through the application server. Besides, when
the user re-registers the registered disaster information, the disaster information in the
database is updated through the application server. Furthermore, when the user deletes the
registered disaster information, the disaster information stored in the database through
the application server is deleted. The same flow applies to registration, editing, and
deleting relief supply management information and staff management information.
 Proposal of a Disaster Response AI System 699

Fig. 10. Input/output sequence of the disaster response AI system.

700 T. Ishida and T. Hirose

5 Conclusion
In this study, we implemented the disaster response AI system to improve the efficiency
of disaster response and prompt information sharing at the disaster countermeasures
headquarters. This disaster response AI system accumulates regional disaster prevention
plans, past disaster response records, relief supplies information, and disaster response
staff information. Thus, this system supports the decision-making of disaster counter-
measures headquarters by visualizing appropriate disaster response personnel allocation
and relief supply information in a disaster event. In the future, we plan to implement a
function for external applications and expand the stored data.

Acknowledgments. This study was supported by JSPS KAKENHI Grant Number JP19K04972
and the research grand of The Telecommunications Advancement Foundation (TAF) of Japan.

References
1. Cabinet Office (Japan), Integrated Disaster Management Information System Maintenance,
Accessed January 2021
2. Foundation for Multi Media Communications, L-alert (Public Information Commons), https://
www.fmmc.or.jp/commons/, Accessed January 2021.
3. Takahagi, K., Ishida, T., Sakuraba, A., Sugita, K., Uchida, N., Shibata, Y.: Construction of
a mega disaster crisis management system. J. Internet Serv. Inf. Secur. (JISIS) 5(4), 20–40
(2015)
4. Takahagi, K., Ishida, T., Sakuraba, A., Sugita, K., Uchida, N., Shibata, Y.: Proposal of the
disaster information transmission common infrastructure system intended to rapid sharing of
information in a time of mega disaster. In: Proceedings of the 18th International Conference
on Network-Based Information Systems, pp. 505–510 (2015)
5. Takahagi, K., Ishida, T., Uchida, N., Shibata, Y.: Proposal of the common infrastructure sys-
tem for real-time disaster information transmission. In: Proceedings of the 30th International
Conference on Advanced Information Networking and Applications Workshops, pp. 673–676
(2016)
6. Hirohara, Y„ Ishida, T., Uchida, N., Shibata, Y.: Proposal of a disaster information cloud system
for disaster prevention and reduction. In: Proceedings of the 31st International Conference on
Advanced Information Networking and Applications Workshops, pp. 664–667 (2017)
 An Efficient Image Codec Using
Spline-Based Directional Lifting Wavelet
Transform and an Optimized
EZW Algorithm

Rania Boujelbene(B) and Yousra Ben Jemaa

L3S Laboratory, ENIT, University of Tunis-El Manar, Tunis, Tunisia

{rania.boujelbene,yousra.benjemaa}@enis.tn

Abstract. This paper proposes a new image compression system with

an optimal spline-wavelet transform based on adaptive directional lifting
and an improved Embedded zero-tree wavelet (EZW) coding algorithm.
The performance of the proposed scheme is evaluated using objective and
subjective criteria. Experimental results have shown promising perfor-
mance compared with the results obtained by the JPEG2000 standard.

1 Introduction
Technology advancement has enabled people not only to capture a large amount
of data but also to exchange these data. It has also greatly changed the speed and
access to the internet, thus allowing a widespread use of Web applications such
as games, bank services, streaming video platform (Netflix) and social networks
which have been changing dynamically.
Thus, one should be able to compress these data in very efficient formats in
order to store them or transmit them over given bandwidth channels without
degrading the quality [1,2].
As a result of this, image compression techniques have been developed with
a lot of new algorithms and also with variations of the already existing ones.
In addition, it is always necessary to change these techniques in order to
adapt them to the evolution of Web applications and media transmission or
storage.
At present, the most advanced image compression systems are based on three
processing steps namely decorrelation, quantization and entropy coding specially
the arithmetic coding which is more commonly used. The first step consists in
changing the representation space of the image. Indeed, a transform approach
generally consists in applying a linear transformation to the intensity values
representing the image in order to extract the main components. Among the most
successful approaches in this field, wavelet decompositions consist in representing
the image by a sum of small waves of varying intensity, size and position. A study
on existing orthogonal wavelets [3] has been done and has proved that these

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 701–712, 2021.
https://doi.org/10.1007/978-3-030-75100-5_60
702 R. Boujelbene and Y. B. Jemaa

types of wavelets are not symmetrical and introduce distortions during the base
change. To fill this gap, other approaches are proposed and prove the need to
study the bi-orthogonal wavelets [4] and more precisely the family of splines [5].
However, these approaches do not faithfully represent the detailed information
of the image.
In the second step, in order to keep the progressive side of the wavelet trans-
form, several methods of organization and quantization are required [6]. These
include EZW, SPIHT, SPECK and EBCOT. These encoders are applied to the
coefficients obtained by the wavelet transform [7].
Among the existing wavelet image compression techniques, EZW [8] is one
of the first and powerful algorithms. It is shown effective, fast in execution and
computationally simple method for the stream bit. Based on the conclusion
which was presented in [9] and which states that the EZW is more promising than
other existing coding schemes and can bring more improvements, many attempts
to enhance its features and to reduce its limitations have been suggested in the
literature [9].
Unlike existing approaches [10–13], we dealt with the two steps of image
compression in order to construct a new efficient wavelet image compression
system. Our contribution is emphasized as follows:
• Enhance the first step of decorrelation by employing a new spline wavelet
transform based on directional lifting, which allows to further reduce the
magnitude of the high-frequency wavelet coefficients.
• Improve the second step of quantification by using an enhanced EZW coding
algorithm which aims at optimizing the binary coding by decreasing the total
number of bits contained in the symbols.
The remaining sections of this paper are organized in the following way. In
Sect. 2, we present the block diagram for the proposed image compression codec.
We describe the principle of the spline-based directional lifting wavelet trans-
form and the different steps of the proposed NE-EZW coding algorithm. The
experimental results are presented in Sect. 3. We illustrate the robustness of the
proposed compression system through a comparative study with the JPE2000
standard. Finally, Sect. 4 concludes the paper and presents the future work.

2 Proposed Codec Scheme for Wavelet Image

Compression
We present here the block diagram of our optimal wavelet image compression
scheme. Indeed, we are the first to present this scheme which is composed of
three connected blocs as shown in Fig. 1.
A Spline wavelet transform based on adaptive directional lifting (ADL-SWT)
represents the transform which combines the spline filter with the adaptive direc-
tional lifting (ADL).
After the wavelet transform step, we try to get a way to code the wavelet
coefficients into an effective result by taking into account the storage space and
 An Efficient Image Codec Using Spline-Based Directional Lifting 703

Fig. 1. Block diagram for optimal codec.

the redundancy. An enhanced EZW image coding (NE-EZW) is the best way
which ensures a triple trade-off between the total number of symbols/the size
of the compressed image/the quality of the reconstructed image. This algorithm
consists in combining two adaptive approaches, which makes it possible to have
fewer redundant symbols with fewer bits.
In addition, in order to further improve the performance of the proposed
NE-EZW, we use arithmetic coding as an entropy coding.
Once the input image has been coded, it is saved or sent through the commu-
nication channel to the receiver who needs to use this code in order to reconstruct
the input image. This is the decoding process which consists of the arithmetic
decoding, the NE-EZW decoding and the inverse ADL-SWT transform.

2.1 ADL-SWT Transform

Instead of alternately using the lifting-based prediction in the horizontal or ver-
tical direction, the ADL performs the prediction in windows of high pixel cor-
relation. For lossy image compression, unlike conventional methods that use the
ADL with the biorthogonal 9/7 filter [14], this technique mixes ADL with a
spline wavelet filter. In fact, we have concentrated on the polynomial spline for
the calculation of the filter taps. Lately, it was shown in [5] that the polynomial
spline wavelet filter of fifth order provides the best performance as compared to
the most efficient existing filters such as the biorthogonal 9/7.
Hence, to construct this performed scheme, the best spline filter of fifth order
is combined with the ADL by incorporating the coefficients calculated by this
filter into the ADL. Thus, the proposed ADL-SWT is employed as the representa-
tion of our image compression system. The proposed 2-D ADL-SWT involves two
separable transforms. The schematic representation of this transform is shown
in Fig. 2.
Let X[m,n] be a 2-D signal, where m and n represent the row and column
indices, respectively. Firstly, carry out 1-D ADL-SWT on each image column,
producing a vertical low-pass subband (L) and a vertical high-pass subband (H).
Secondly, carry out 1-D ADL-SWT on each row of L and H.
After one-level decomposition, one low-pass subband (LL) and three high-
pass subbands (LH,HL and HH) are generated. In other words, the subband
decomposition structure of 2-D ADL-SWT is the same that of 2-D DWT.
For ADL-SWT, unlike DWT which does transform along the fixed direction,
the selected filtering need to be encoded as side information. So, to reduce the
704 R. Boujelbene and Y. B. Jemaa

Fig. 2. 2-D Adaptive directional lifting spline wavelet transform.

overhead bits for the direction information, the image is divided into regions of
approximately uniform edge orientations using quadtree segmentation. All the
pixels in the local region are predicted and updated along the uniform direction
which is chosen in a rate–distortion optimal sense. The main steps of the ADL-
SWT transform are illustrated by Algorithm 1.

2.2 NE-EZW Coding Algorithm

To increase the compression performance, the NE-EZW is used which can be

viewed as an extension of the EZW coding technique. Unlike the previous
improved versions of EZW [15,16], the new algorithm ensures a triple trade-
off between the total number of symbols/the size of the compressed image/the
quality of the reconstructed image.
This algorithm suggests the use of two new added symbols in its dominant pass
as compared to the original EZW. This increases the number of coefficients not to
be encoded and decreases the total number of symbols by adding two steps namely
remove the last T and de-bit encoding that aim at eliminating unnecessary sym-
bols. In addition, the NE-EZW greatly reduces the amount of information by the
use of the compressor cell operator in the binary coding step which aims at coding
a consecutive number of symbols by a small number of bits.
The NE-EZW has six steps known as initialisation, dominant pass (or signif-
icance mapping pass), remove the last T, de-bit encoding, binary encoding and
subordinate pass. The process of its coding algorithm is illustrated by Algorithm 2.

3 Experimental Results

3.1 Data and Evaluation Criteria

To assess the efficiency of the proposed approach, a number of numerical exper-

iments have been performed on a number of natural and medical images. For
this purpose, the standard Waterloo 8-bit greyscale image set [17] containing 12
images of different sizes (Lena, Barbara, Boat, Mandrill, Zelda, Goldhill, Pep-
pers, House, Washsat, France, Montage, Library) and a two medical images [18]
have been considered to perform the required tests.
 An Efficient Image Codec Using Spline-Based Directional Lifting 705

Algorithm 1. ADL-SWT transform.

1: The image is adaptively divided into blocks of variable sizes based on hierarchical quadtree
segmentation.
2: Each block of 16 × 16 can be divided into three modes (16 × 16, 8 × 8 and 4 × 4).
3: Perform the ADL of spline wavelet filter on these modes
4: if The enegry metric (sum of absolute coefficients in high frequency subband) related to
one of these block modes is smaller than the other then
5: Select this mode
6: end if
7: The output is divided to variable-size regions based on hierarchical quadtree segmentation.
8: Each block of 16 × 16 can be divided into three modes (16 × 16, 8 × 8 and 4 × 4).
9: Perform the ADL of spline wavelet filter on these modes
10: if The enegry metric related to one of these block modes is smaller than the other then
11: Select this mode
12: end if
13: The inverse transform is performed to reconstruct the image by giving the optimal lifting
direction as side information

For a fair comparison, the proposed codec has been compared with the
JPEG2000 [19]. Peak Signal to Noise Ratio (PSNR), Structural Similarity Index
Measure (SSIM) and Bjøntegaard Delta PSNR (BD-PSNR) were used in our
experiments as a set of performance criteria. The PSNR is defined as follows:

(P eak)2
P SN R(dB) = 10log10 [ ], (1)
M SE
where P eak is equal to 255 for the images in 8 bits per pixel (bpp) and

 M
N 
1
M SE = (f (i, j) − f(i, j))2 , (2)
M ×N 1 1

where f , f and M × N represent the original image, the reconstructed image

and the total number of pixels in the image, respectively.
Besides PSNR, to further evaluate the quality and the distortion for the
reconstructed image using the presented techniques, we use SSIM which is con-
sidered to be correlated with the quality perception of the human visual system
(HVS). The SSIM is a decimal value between 0 (zero correlation with the original
image) and 1 (exact same image). It is defined as follows:

SSIM = [l(f, f)]α [c(f, f)]β [s(f, f)]σ (3)

where: (f, f), l(f, f), c(f, f) and s(f, f) represent respectively two images, lumi-
nance comparison, contrast comparison and structural comparison between two
images.
α > 0, β > 0 and σ > 0 are used to adjust the importance of the three
parameters.
To calculate the coding efficiency between different codecs based on PSNR
measurements, a Bjøntegaard model was proposed by Gisle Bjntegaard. It is
706 R. Boujelbene and Y. B. Jemaa

Algorithm 2. NE-EZW.
1: Th0 = 2(log2 (max(abs(coef f s))) /* Initialization */
2: k=0
3: Dominant List=All coefficients
4: Character Encoding List=[ ]
5: Subordinate List=[ ]
6: for each coefficient x in the Dominant List /* Dominant Pass */ do
7: if (abs(x) ≥ Thk ) then
8: if at least one descendant ≥ Thk then
9: if x>0 then
10: Put P on Character Encoding List
11: else
12: Put N on Character Encoding List
13: end if
14: else
15: if x>0 then
16: Put Pt on Character Encoding List
17: Don’t code its descendants
18: else
19: Put Nt on Character Encoding List
20: Don’t code its descendants
21: end if
22: end if
23: else
24: if x is non-root of a zerotree (it is a descendant from a Zerotree Root) then
25: Don’t code this coefficient
26: end if
27: if x is a zerotree root then
28: Put T on Character Encoding List
29: else
30: Put Z on Character Encoding List
31: end if
32: end if
33: end for
34: for each symbol(P, Pt , N and Nt ) in the Character Encoding List do
35: Put its magnitude on the Subordinate List
36: Remove it from the Dominant List
37: end for
38: Scan the Character Encoding List in inversed order/* Remove last T */
39: Delete all the T after (P, Pt , N or Nt ) while encounter the first (P, Pt , N or Nt )
40: Encode the Character Encoding obtained from the previous step by making the use of
decreasing bits coding/* De-Bit encoding Algorithm */
41: Transform the output of the previous step into binary coding/* Binary encoding Algorithm */
42: for each coefficient in the Subordinate List/* Subordinate Pass */ do
43: if magnitude of coefficient is in the Bottom Half of [T hk , 2T hk ] then
44: output ’0’
45: else
46: output ’1’
47: end if
48: end for
49: T hk+1 =T hk /2 /* Halve threshold */
50: k=k+1
51: Go to the Dominant Pass: Repeat until the reconstructed image is reached or until the
number of transferable bits required is exceeded

used to calculate the average PSNR and bit rate differences between two rate-
distortion (R-D) curves obtained from the PSNR measurement when encoding a
content at different bit rates. The Bjøntegaard delta PSNR (BD-PSNR), which
corresponds to the average PSNR difference in dB for the same bit rate is used
in our experiments.
 An Efficient Image Codec Using Spline-Based Directional Lifting 707

3.2 Performance Analysis of the Proposed Codec

We presented in this section the performance of our system in terms of recon-

structed image quality.
The variations of the PSNR versus the bitrate for all test images is given by
the curves represented in Figs. 3 and 4.

41 38
38

37
40
36
36
39
35
34
38
PSNR(dB)

PSNR(dB)

PSNR(dB)
34

37 32
33
Proposed codec
Proposed codec
Proposed codec JPEG2000
36 JPEG2000 32
JPEG2000 30

35 31

28
30
34
0,25 0,50 0,75 1,00 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 0,25 0,50 0,75 1,00
Bitrate(bpp) Bitrate(bpp) Bitrate(bpp)

(a) Lena (b) Barbara (c) Boat

42 37

36
28

35
40

34
PSNR(dB)

PSNR(dB)

26
PSNR(dB)
Proposed codec
33 JPEG2000
Proposed codec Proposed codec
JPEG2000 38 JPEG2000 32

24
31

30
36
0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 0,25 0,50 0,75 1,00
Bitrate(bpp) Bitrate(bpp) Bitrate(bpp)

(d) Mandrill (e) Zelda (f) Goldhill

30

38 38

28
PSNR(dB)

PSNR(dB)

PSNR(dB)

36
26 36

Proposed codec Proposed codec Proposed codec

JPEG2000 JPEG2000 JPEG2000
34 24

34

22
0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Bitrate(bpp) Bitrate(bpp) Bitrate(bpp)

(g) Peppers (h) House (i) Washsat

32 42

40 28

30
38
26
36
PSNR(dB)

PSNR(dB)

PSNR(dB)

28

34 24
26
Proposed codec Proposed codec Proposed codec
JPEG2000 JPEG2000 JPEG2000
32
22
24 30

20
0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Bitrate(bpp) Bitrate(bpp) Bitrate(bpp)

(j) France (k) Montage (l) Library

Fig. 3. PSNR (in dB) versus the bitrate (bpp) of the proposed and JPEG2000 codecs
for all grayscale test images.
708 R. Boujelbene and Y. B. Jemaa

40

42
38
PSNR(dB)

PSNR(dB)
36

Proposed codec 40 Proposed codec

JPEG2000 JPEG2000
34

32
38
0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
Bitrate(bpp) Bitrate(bpp)

(a) MRI (b) MR3DBrain100001

Fig. 4. PSNR (in dB) versus the bitrate (bpp) of the proposed and JPEG2000 codecs
for all medical test images.

By comparing the different values of PSNR, we see clearly the effectiveness

of our proposed codec as compared to the most-well-known codec JPEG2000 for
all test images.
The coding efficiency was also evaluated using the average PSNR gain (i.e.
BD-PSNR) measure. According to the simulation results shown in Tables 1 and
2, we conclude that the proposed codec outperforms the JPEG2000 standard for
different test images.
Indeed, the average PSNR gain of the grayscale and medical test images
reaches 0.48 dB and 0.72 dB, respectively. The performance, reported in Tables 3
and 4, are computed in terms of SSIM. It can be observed that for all test images
the proposed codec produces high SSIM values when compared to the JPEG2000
for all bit-rates.

Table 1. BD-PSNR gain of the proposed algorithm against JPEG2000 for all grayscale
test images

Image Lena Barbara Boat Mandrill Zelda Goldhill Peppers House Washsat France Montage Library
BD- 0.328 0.55 0.945 0.488 0.637 0.57 0.303 0.323 0.21 0.263 0.838 0.157
PSNR(dB)
Average 0.477

Table 2. BD-PSNR gain of the proposed algorithm against JPEG2000 for all medical
test images

Image BD-PSNR(dB)
MRI 0.948
MR3DBrain100001 0.483
Average 0.716
 An Efficient Image Codec Using Spline-Based Directional Lifting 709

Table 3. SSIM versus the bitrate (bpp) of the proposed and JPEG2000 codecs for all
medical test images.

Image Bitrate (bpp) Methods

Proposed codec JPEG2000
MRI 0.25 0.9231 0.9187
0.5 0.9412 0.938
1 0.984 0.9832
MR3DBrain100001 0.25 0.9452 0.945
0.5 0.9712 0.9636
1 0.9834 0.9801

Figure 5 shows the reconstructed MRI medical image for the proposed and
JPEG2000 codecs at 0.25 bpp. In this subjective test, we compare the perceptual
quality between the two codecs. It can be observed that the visual quality of the
image in Fig. 5(b) is better than the other.

(a) JPEG2000 (b) Proposed codec

Fig. 5. Subjective assessment of the different codec for the MRI medical image at 0.25
bpp (a) PSNR = 32.1 dB (b) PSNR = 34.76 dB
710 R. Boujelbene and Y. B. Jemaa

Table 4. SSIM versus the bitrate (bpp) of the proposed and JPEG2000 codecs for all
grayscale test images.

Image Bitrate (bpp) Methods

Proposed codec JPEG2000
Lena 0.25 0.9771 0.9712
0.5 0.9918 0.9859
1 0.9981 0.9913
Barbara 0.25 0.9321 0.9268
0.5 0.9759 0.9646
1 0.9974 0.9968
Boat 0.25 0.955 0.9413
0.5 0.9721 0.9696
1 0.9987 0.9862
Mandrill 0.25 0.7932 0.7919
0.5 0.885 0.8788
1 0.976 0.97
Zelda 0.25 0.9611 0.9691
0.5 0.9799 0.9775
1 0.9941 0.9937
Goldhill 0.25 0.8315 0.8219
0.5 0.9391 0.937
1 0.9725 0.97
Peppers 0.25 0.9678 0.9602
0.5 0.9744 0.973
1 0.9891 0.9807
House 0.25 0.8123 0.8044
0.5 0.9134 0.9091
1 0.9628 0.961
Washsat 0.25 0.8966 0.8953
0.5 0.9316 0.9307
1 0.9867 0.9899
France 0.25 0.7188 0.7097
0.5 0.84 0.833
1 0.9356 0.9303
Montage 0.25 0.878 0.8737
0.5 0.9536 0.94
1 0.9897 0.9859
Library 0.25 0.5987 0.5981
0.5 0.7482 0.748
1 0.8746 0.8741
 An Efficient Image Codec Using Spline-Based Directional Lifting 711

4 Conclusion
In this paper, we presented a novel wavelet image compression system which
is constructed with a new spline wavelet transform based on adaptive direc-
tional lifting and an optimized coding algorithm based on the notion of zerotree.
Experimental results have shown the superiority of the proposed system over
the most known algorithm JPEG2000 in terms of PSNR, BD-PSNR and SSIM
for different test images.
In the future, we plan to use our proposed image compression technique in the
Wireless Multimedia Sensor Networks (WMSN) in order to check its efficiency
in this field.

References
1. Doutsi, E.: Retina-inspired image and video coding. PhD thesis, Université Côte
d’Azur (2017)
2. Bovik, A.C.: Handbook of Image and Video Processing. Academic press,
Cambridge (2010)
3. Rasool, U., Mairaj, S., Nazeer, T., Ahmed, S.: Wavelet based image compression
techniques: comparative analysis and performance evaluation. Int. J. Emerg. Tech-
nol. Eng. Res. (IJETER) 5(9), 9–13 (2017)
4. Prasad, P.M.K., Umamadhuri, G.: Biorthogonal wavelet-based image compression.
In: Artificial Intelligence and Evolutionary Computations in Engineering Systems,
pp. 391–404. Springer (2018)
5. Boujelbene, R., Jemaa, Y.B., Zribi, M.: An efficient codec for image compression
based on spline wavelet transform and improved SPIHT algorithm. In: 2017 Inter-
national Conference on High Performance Computing & Simulation, HPCS 2017,
17–21 July 2017, Genoa, Italy, pp. 819–825. IEEE (2017)
6. Singh, A., Kirar, K.G.: Review of image compression techniques. In: 2017 Inter-
national Conference on Recent Innovations in Signal Processing and Embedded
Systems (RISE) (2017)
7. Hussain, A.J., Al-Fayadh, A., Radi, N.: Image compression techniques: a survey in
lossless and lossy algorithms. Neurocomputing, 300, 44–69 (2018)
8. Shapiro, J.M.: Embedded image coding using zerotrees of wavelet coefficients. IEEE
Trans. Sign. Process. 41(12), 3445–3462 (1993)
9. Boujelbene, R., Jemaa, Y.B., Zribi, M.: A comparative study of recent improve-
ments in wavelet-based image coding schemes. Multim. Tools Appl. 78(2), 1649–
1683 (2019)
10. Boujelbene, R., Jemaa, Y.B.: A new structure of spline wavelet transform based on
adaptive directional lifting for efficient image coding. Sign. Image Video Process.
14(7), 1451–1459 (2020)
11. Boujelbene, R., Boubchir, L., Jemaa, Y.B.: Enhanced embedded zerotree wavelet
algorithm for lossy image coding. IET Image Process. 13(8), 1364–1374 (2019)
12. Jiang, X., Song, B., Zhuang, X.: An enhanced wavelet image codec: SLCCA
plus. In: 2018 International Conference on Audio, Language and Image Processing
(ICALIP), pp. 163–167. IEEE (2018)
13. Mander, K., Jindal, H.: An improved image compression-decompression technique
using block truncation and wavelets. Int. J. Image Graph. Sign. Process. 9(8), 17
(2017)
712 R. Boujelbene and Y. B. Jemaa

14. Antonini, M., Barlaud, M., Mathieu, P., Daubechies, I.: Image coding using wavelet
transform. IEEE Trans. Image Process. 1(2), 205–220 (1992)
15. Che, S., Che, Z., Wang, H., Huang, Q.: Image compression algorithm based on
decreasing bits coding. In: 2009 Fifth International Conference on Information
Assurance and Security, vol. 1, pp. 217–220. IEEE (2009)
16. Brahimi, T., Laouir, F., Boubchir, L., Ali-Chérif, A.: An improved wavelet-based
image coder for embedded greyscale and colour image compression. AEU-Int. J.
Electron. Commun. 73, 183–192 (2017)
17. Greyscale and color databases for the evaluation of image and signal processing
algorithms. http://links.uwaterloo.ca/Repository.html
18. Medical database for the evaluation of image and signal processing algorithms.
http://www.international-education-biometrics.net/medeisa/
19. Xiaolin, W.: Lossless compression of continuous-tone images via context selection,
quantization, and modeling. IEEE Trans. Image Process. 6, 656–664 (1997)
 An Innovative Monocular Mobile Object
Self-localization Approach Based
on Ceiling Vision

Alfredo Cuzzocrea1,2(B) , Luca Camilotti3 , and Enzo Mumolo3

1
IDEA Lab, University of Calabria, Rende, Italy
[email protected]
2
LORIA, Nancy, France
3
University of Trieste, Trieste, Italy
{camilotti,mumolo}@units.it

Abstract. This study deals with the estimation of the position of a

mobile object using ceiling landmarks images acquired by a low resolu-
tion camera placed on a mobile object. The mobile object is moving in an
indoor environment where light is given by electric lamps with circular
holders. The images of the circular holders are projected on the image
plane of the camera and are processed by means of computer vision algo-
rithms. The pixels of the images of the light holders on the ceiling are
mapped to the pixels of the images of the light holders on the image plane
of the camera by means of a two dimensional dynamic programming algo-
rithm (2D-DPA). The projection distortions are thus compensated and
this reduces the estimation errors. The algorithm described in this paper
estimates the distance from the camera lens to the center of the land-
marks using only ceiling vision. Localization can be easily obtain from
such distance estimations. The projections are geometrically described
and the distance estimation is based on the pixels mapping information
obtained by 2D-DPA.

Keywords: Localization · Ceiling landmarks · Mobile object · Two

dimensional dynamic programming

1 Introduction
Self-localization of mobile objects is a fundamental requirement for autonomy.
Mobile objects can be for example a mobile service robot, a motorized wheelchair,
a mobile cart for transporting tasks or similar. Self-localization represents as well
a necessary feature to develop systems able to perform autonomous movements
such as navigation tasks. Self-localization is based upon reliable information
coming from sensor devices situated on the mobile objects. There are many
This research has been made in the context of Excellence Chair in Computer Engi-
neering – Big Data Management and Analytics at LORIA, Nancy, France.

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 713–730, 2021.
https://doi.org/10.1007/978-3-030-75100-5_61
714 A. Cuzzocrea et al.

sensors available for that purpose. The early devices for positioning are rotary
encoders. If the encoders are connected to wheels or legs movement actuators, rel-
ative movements of the mobile object during its path [1] can be measured. Then,
mobile object positioning can be obtained with dead-reckoning approaches. Dead
reckoning [1] is still widely used for mobile robot positioning estimation. It is also
true that dead-reckoning is quite unreliable for long navigation tasks, because of
accumulated error problems. Other popular sensor devices for self-localization
are laser or sonar based range finder devices and inertial measurement devices.
In outside scenarios the most popular approaches are based on Global Position-
ing System (GPS). Due to the importance of self-localization, many other solu-
tions for indoor environment have been proposed so far with different cost and
accuracy characteristics. For example the Ultra Wide Band radio signal indoor
localization systems [2], or the Bluetooth-based angle of arrival radio devices
[3], or a combination of them. However these systems have serious limitations in
cost and reliability, respectively. Another important type of sensors which may
be used for cost effective self-localization are the CCD cameras, which require
computer vision algorithms for localization such as for example visual odometry,
[23]. Mobile objects vision self-localization is currently an open research field [5]
and an increasing number of new methods are continuously proposed. As a mat-
ter of fact we have to consider that self-localization of mobile objects requires
centimeter-level accuracy and Computer Vision is one of the most cost-effective
techniques able to reach that accuracy. Consequently, some surveys of Computer
Vision based self-localization techniques appeared recently in the literature, [6].
In this paper we describes a novel Computer Vision algorithm for estimat-
ing the distance from the camera lens to the center of ceiling landmarks with
circular shape using a monocular low cost webcam. From the distance, mobile
object localization approaches can be easily developed and a simple example is
provided in this paper. The images of the ceiling landmarks are projected on
the image plane of the camera. The projection is analytical described, but the
projections distortions, which may arise especially when low cost devices are
used, may affect the results. To take into account the projection distortions in
order to obtain a better precision of the results, we use an approximation of
the two-dimensional dynamic programming (2D-DPA) algorithm [4] which finds
a sub-optimal mapping between the image pixels of the ceiling landmarks and
the image plane pixels of the projected landmarks. Since optimum 2D-DPA is
NP-complete, in fact, many approximations have been developed. For example,
the 2D-DPA technique described by Levin and Pieraccini in [25] has an expo-
nential complexity in the image size, while Uchida and Sakoe describe in [24]
a Dynamic Planar Warping technique with a complexity equal to O(N 3 9N ).
Lei and Govindaraju propose in [26] a Dynamic Planar Warping approximation
with a complexity of O(N 6 ). However each approximation has some limitation in
terms of continuity of the mapping. In this paper we use a approximation of the
optimum 2D-DPA with a complexity of O(N 4 ) [22] which is implemented on a
GPU to obtain real-time performance. When the landmark is far from the cam-
era or if the environments has low lighting conditions, an high quantization noise
 An Innovative Monocular Mobile Object Self-localization 715

may arise in acquired images. However the algorithm we describe in this paper
is particularly robust against noise due especially to the use of two-dimension
DPA.
This paper is organized as follows. Section 2 reports on related work. Section 3
the localization problem is described, and in Sect. 4 the projection distortion is
geometrically described, while in Sect. 5 the two-dimensional Dynamic Program-
ming approximation is described. In Sect. 6 the proposed algorithm is sketched
and in Sect. 7 the computer vision algorithms for the detection of landmarks on
the image plane are reported. Section 8 sketches a possible global localization
approach of the mobile object. In Sect. 9 we report some experimental compari-
son of the proposed algorithm with state of the art algorithm. Finally, Sect. 10
concludes the paper with concluding remarks and a suggestion of future works.

2 Related Work
Many papers on vision-based mobile robot self-localization appeared recently in
the literature. For example Avgeris et al. describe in [18] a self-localization algo-
rithm for mobile robots that uses cylindrical landmarks resting on the floor and
a single pivotal camera with an horizontal angle of view of 30-degree. Each cylin-
drical landmark has a different color in order to be easily detected by the robot.
However, frontal vision could be occluded by objects or people. Such interference
can be avoided by placing the landmarks on the ceiling, so that the camera is
tilted toward the ceiling. Ceiling vision has been used by many authors to per-
form mobile robot localization. One of the early proposals is described in [20] and
is based upon a digital mark pattern and a CCD camera. The camera is tilted, so
the horizontal distance from the ceiling mark pattern is obtained measuring the
ratio between the length and the width of the pattern picture. Kim and Park,
[7], acquire ceiling images in a small area with a fish-eye lens camera. Ceiling
outlines are detected by means of adaptive binarization and segmentation. Robot
pose is obtained after identification of the ceiling region and the determination
of the center and the momentum of the region. Lan et al. describe in [8] a mobile
robot positioning algorithm based on artificial passive landmarks placed on the
ceiling and infrared sensors. The landmarks are made of reflective film 2D struc-
tures containing dots assigned to unique ID’s. The infrared sensors consist of an
infrared camera and an infrared LED array. A similar approach is described in
[9] where artificial passive reflective landmarks are placed on the ceiling and an
infrared camera plus an infrared LED source are used to capture the reflection
the IR light on the landmark for estimating the robot pose. Wang et al. describe
in [21] a vision control system which capture ceiling RGB images with a camera
placed on the robot, convert the image to HSV color space and use V channel
images to reduce the effect of illumination lamps. The common objects and the
straight lines on the ceiling are detected by template matching and used to esti-
mate the robot orientation. Other Computer Vision based approaches are based
on the Free Space Density concept. For example A. Ribacki et al. use an upward
facing camera to extract the ceiling boundaries for estimating the ceiling space
716 A. Cuzzocrea et al.

density from the current image [11]. Other authors, for example [12,13] use the
ceiling depth images for robot localization. In these approaches self-localization
is obtained from Principal Component Analysis of ceiling depth images. Ceil-
ing vision is used by many other authors to perform self-localization of mobile
robots. For example Qing Lin et al. describe in [17] a visual odometry algorithm
based on a monocular camera which points to the ceiling. The algorithm uses
several local features detectors for matching the features between two sequential
frames of the ceiling.
In addition, it should considered the emerging integration of these topics with
the innovative big data trend (e.g., [14–16]). Here, the main research perspective
is that to take into account the well-known 3V model of big data, including
volume, velocity and variety.

3 Problem Description
We show in Fig. 1 a mobile object in an indoor environment. The movable object
is equipped with a camera set tilted towards the ceiling at an angle ϕ. We call h
the distance between the camera and the ceiling. Moreover in Fig. 2 the horizon-
tal and vertical angles of view of the camera, called θx , and θy respectively, are
highlighted. The direction towards which the camera is oriented is shown with
the ‘Camera Direction’ arrow. The ceiling landmark is shown in Fig. 1 with a
segment with a greater thickness and the image plane of the camera is shown
with a segment orthogonal to the camera direction. The ceiling landmark is pro-
jected to the landmark on the image plane. The visual landmarks positioned on
the ceiling used in this approach are the lighting holders shown as that shown in
Fig. 3. We choose landmarks with isotropic shapes on the plane because in this
way the distortion components due to image rotation can be eliminated. The
simpler isotropic shape is the circle. As shown in Fig. 3, the lines of pixels on the
image plane are all parallel to the reference abscissa on the ceiling plane regard-
less of the angle of the camera with respect to the landmark. It is important
to remark that each landmark must be distinguishable from the others and its

ceiling camera
ceiling direction
landmark

h = camera lens
to h landmark in the
ceiling distance image plane

= camera tilt angle

floor

Fig. 1. A mobile object with a camera on it, tilted toward the ceiling.
 An Innovative Monocular Mobile Object Self-localization 717

Fig. 2. The horizontal and vertical angles of view of the camera.

Fig. 3. An example of the circular lamp holder used in this paper

coordinate in the global reference system must be known. A schematic represen-

tation of a mobile object and some landmarks with the orthonormal reference
system centered on the camera lens is shown in Fig. 4. The reference abscissa
changes dynamically in relation to the direction of the focal axis. The reference
abscissa, in fact, is always normal to the focal axis and at the same time it is
parallel to the horizon.
The landmarks must be distinguishable from each other. There are many
possible solutions for making the landmarks distinct. A simple possibility is
to paint each holder with a different color. More recently, the characteristic
frequency of fluorescent lights has been used, for instance in [19]. In this paper
we used the simplest solution, namely we painted adjacent lamp holders with
different colors. For this reason the landmarks in Fig. 4 are represented with
different colors, where for simplicity the three circular landmarks positioned on
the ceiling are colored in red, blue and green. Figure 4 shows that the landmarks
which fall within the visual field of the camera are projected onto the image
plane of the camera. Of course we know in advance the physical position of
each landmark in the global reference system. On the other hand the landmark
colours can be detected using well known computer vision techniques.
718 A. Cuzzocrea et al.

blue landmark
ceiling

red landmark green landmark

image plane

z
camera visual field

camera x
mobile object

Fig. 4. Schematic representation of orthonormal reference system, landmarks and

image plane.

4 Projective Transformations

The projective transformation is the linear transformation of coordinates

reported in (1).
p = T p (1)
where p represents a generic point in space expressed in homogeneous coordi-
nates, relative to the orthonormal reference system S described by the quadruple
(O, î, ĵ, k̂). The projected point p is expressed in coordinates relative to the ref-
erence system S  described by the quadruple (O , iˆ , jˆ , kˆ ), where iˆ = î, jˆ has
the direction of the segment M Q and k̂  has the direction of the normal to the
segment M Q.
Since p is expressed with the three components (xp , yp , zp ) and p has the
three components (xp , yp , zp ), Eq. (1) can be also written as follows
⎛ ⎞ ⎛ ⎞
xp xp
⎝ yp ⎠ = T ⎝ yp ⎠ (2)
zp zp

Such a transformation maintains the properties of collinearity, that is, the points
which in S belong to a line, are aligned in a line also in S’. However, projective
transformation may not be defined for every point of S, in the sense that some
points could be mapped in S’ at infinity.
Let us view Fig. 4 from the left side, that is the y −z plane of the orthonormal
reference system which has its origin coinciding with the center of the camera
lens. This plane is highlighted in Fig. 5, where the ceiling is at z = h, and the
field of view of the camera is shown with points M and E. Let us assume that a
 An Innovative Monocular Mobile Object Self-localization 719

landmark falls within the vertical angle of view. Then, the center of the landmark
is the point C. On the other hand, if we view Fig. 4 from the front side, that
is the x − z plane, we obtain the system shown in Fig. 6. Of course the camera
image plane, which is the plane normal to the focal axis in Fig. 4, is shown with
the segment M − Q in Fig. 5 and segment G − I in Fig. 6.

Fig. 5. Plane y − z in orthonormal reference system

Suppose we fix a point P on the ceiling. If the point falls within the field of
view of the camera it is shown as P in Fig. 5. Let (px , py , pz ), with pz = h, be the
coordinates of P . The point P is projected to the image plane of the camera to
the point P  , which has coordinates (xp , yp , zp ). Also the center of the landmark
in Fig. 5 is projected to the point C  and the segment M − E is projected to
the segment M − Q in the image plane. In this model, the focal distance of the
device or other characteristic parameters are not taken into account. It is in
fact a purely ideal model, which has the only purpose of deriving the relations
that define the projective transformation from the orthonormal system whose
origin coincides with the center of the camera lens to the image plane system.
The latter is chosen independently of the characteristics of the camera. With
reference to the Figs. 5 and 6, we introduce the following geometric variables
characteristic of the problem.
–
θy π
Φ=ϕ+ − (3)
2 2
– The distance a from the origin to the barycenter of the landmark projected
on the image plane:
h 1
a = OC  = − h(tan(ϕ) + ) cos(ϕ) (4)
sin(ϕ) tan(ϕ)
720 A. Cuzzocrea et al.

Fig. 6. Plane x − z in orthonormal reference system

– The abscissa of the point P’ on the image plane:

b 1
= M C  = C  Q = h(tan(ϕ) + sin(ϕ)) (5)
2 tan(ϕ)
Equations (4) and (5) are developed in Appendix A. Moreover, we define the
following two variables:
G = −h (tan Φ tan ϕ + 1) (6)
and
F = h (tan ϕ − h tan Φ) (7)
We remark that the following considerations are based on three coordinate sys-
tems, namely an orthonormal reference system centered on the camera lens,
shown in Fig. 4, an orthonormal reference system on the image plane and a sys-
tem on the ceiling plane which is simply translated by h with respect to that
centered on the camera lens. In general, points on the systems centered on the
camera lens and on the ceiling are denoted with a capital letter, such as P, while
that on the image plane of the camera are denoted with a capital letter plus an
apex such as P’. In this case, P’ is the P point projected on the image plane. If
we look at the landmark seen from the orthonormal reference system centered
on the camera lens, its barycenter is located at (xc , yc ). A generic point on the
ceiling has coordinate (x, y) and the same point projected on the image plane is
(x , y  ). The coordinates of a generic point on the landmark is given relative to
its barycenter: (x = xc + xr ) and (y = yc + yr ). According to Figs. 6 and 5 the
offsets xr , yr are projected to the image plane in x , y  .
Assume now we have an optimum mapping between images. In other words,
assume that, having two images A and B, A = {a(i, j)|i, j = 1, ..., N } and
B = {b(u, v)|u, v = 1, ..., M }, we can estimate the mapping function
   
u x(i, j)
F (i, j) = = (8)
v y(i, j)
 An Innovative Monocular Mobile Object Self-localization 721

which maps each pixel (i, j) of one image to the pixel (u, v) of the other image
such that the difference between the two images is minimized, as shown in (9).

min a(i, i) − b(u, v) (9)

where u = x(i, j) and v = y(i, j). Such mapping is performed through a two
dimensional Dynamic Programming operation [24]. 2D-DPA is the base of image
matching algorithms called Elastic Image Matching. Unfortunately, the Elastic
Image Matching operation is NP-complete [29]. For this reason we devise an
approximation which reduces the 2D-DPA operation complexity to O(N 4 ), as
described below.
The barycenter of the landmarks, (xc , yc ), are estimated using the following
Proposition.
Proposition 1. By measuring the abscissa and ordinate (x , y  ) of a generic
point on the landmark projected on the image plane we can estimate the coordi-
nate (xc , yc ) of the ceiling landmark using the following equations:

h cos(ϕ − γy )(x − g)
xc = + g − xr (10)
a sin(γy )

aG + ayr tan(ϕ) − (y  − 2b )(yr + F )

yc = b
(11)
y − 2 − a tan(ϕ)

Proof. In Appendix B we give a sketch of the derivations.

A different estimation of the coordinates of the landmark barycenter is

obtained for all the points P inside the landmarks. A sequence of barycenter
coordinates xc , yc are thus obtained, of which we compute the expected value.
The algorithm is thus sequentially divided into two parts: estimation of E(xc )
and E(yc ) by measuring the dimension y  and x of the distorted image on the
image plane.
The distance from the camera lens and the landmark in the ceiling reference
system is thus the following:

d= E(xc )2 + E(yc )2 ) (12)

with reference to Figs. 6 and 5, where C = (xc , yc , zc ) is the barycenter of the

landmark in the reference system (O, i, j, k). We obtain he sub-optimal cor-
respondence, pixel by pixel, between a reference image and a distorted image
by means of approximated two dimensional dynamic programming, . Our algo-
rithm therefore uses the deformation of the image to derive the distance of the
landmark, i.e. it is intended to determine how the perspective has distorted the
image.
The coordinates of the barycenter of the ceiling landmarks are obtained using
the coordinate x measured on the image plane and xr using the mapping func-
tion, and in terms of y  and yr . Clearly (xr , yr ) and (x , y  ) are both known
722 A. Cuzzocrea et al.

because they are derived from the coordinates of the pixels in the pattern and
in the test images respectively. What associates the two pixels is the mapping
relationship described in (8) obtained by 2D-DPA.
The characteristic that differentiates the algorithms present in the literature
from the one developed in this paper is the statistical character of the obtained
estimate. The algorithm based on dynamic programming is able to calculate
a position estimate for each single pair of associated pixels from the mapping.
The advantage is that a large number of points are used, which contribute to the
calculation of the average distance value. This makes the estimate more truthful,
especially when the landmark is very distant, which results in a smaller image
and a greater quantization error.

5 2D Dynamic Programming Based Image Mapping

Technique (2D-DPA)
For the sake of coherence with what we write below, we repeat now the mapping
considerations summarized above about images A and B using instead images X
and Y . Given the two images, X = {x(i, j)} and Y = {y(u, v)}, the mapping
of one image to the other is represented by the operation


N 
N
D(X, Y ) = min x(i, j) − y(u, v)
i=1 j =1

where u = x(i, j), v = y(i, j) is the mapping function between the pixels
of X and Y . The quantity D(X, Y ) gives a distance between the image X and
the optimally deformed Y , the optimal warping function x(i, j), y(i, j) gives
an interpretation of the image X according to the generation model Y.
Given the i−th row of the X image and the j−th row of the Y images,
namely Yj = (yj,1 , yj,2 , . . . , yj,N ), Xi = (xi,1 , xi,2 , . . . , xi,N ) respectively,
the distance between the two rows is obtained by applying a 1D-DPA [27] for
finding a warping among the two rows as described in (13). Here the map M 
is, say, over (n, m) coordinates, so that Ml = ((il , nl ), (jl , ml )).

M M
min
 l=1 d(Ml ) min
 l=1 xi l ,n l − yj l ,m l 
M M
d(Xi , Yj ) = = (13)
M 2N
Finally, the distance between the two images is obtained by (14). In this case
the map M  is between all the rows of X and Y. As before, |M  | is the length
of the path.
min k d(M  k )

D(X, Y ) = M
|M  |
|M |
min l =1 d(M l )
M 
min k d(Xi k , Y jk ) min k 2N
M M
= =
|M  | 2N
 An Innovative Monocular Mobile Object Self-localization 723

|M  |
min{ k min l=1 xi,n l − yj,m l }
M  M
= (14)
4N 2
Let us assume that the images are of equal size, that is N × N pixels. Then
the length of the optimum path between the two images is equal to 2N . The
local distances in each point of this path is obtained with other 1D-DPA with
paths of length 2N . The total length is the sum of 2N along the 2N long path,
giving 4N 2 at the denominator. The complexity of the described operation is
O(N 2 N 2 ) = O(N 4 where N is the image dimension.

6 Proposed Algorithm
The algorithm described in this paper is summarized in the following Algorithm.
The inputs of the algorithm are the two gray-scale images imgA and imgB
which are the landmark on the image plane and on the ceiling respectively. We
perform the 2D-DPA algorithm on these two images to obtain the mapping
function as result. the mapping function is represented with a linked list where
each node is the map related to the two pixels. The function get() give as result
the value of the pixel on the image indicated as input and is used to get the
values of the two pixels linked by the map on the two landmark images. To
decide if the pixel is a landmark pixel or not, we consider their gray levels. The
landmarks have a lower values with respect to the environment and thus if the
pixel values is less then a thereshold, the pixel is a landmark pixel.

Input:imgA , imgB
Output: distance
img=Detect(imgA );  get the landmark in the image plane
id=identify(img);  identify the landmark
head=2D-DPA(imgA , imgB );
ptr=head;  head is the list of mapping function
repeat
pixA = get(imgA , ptr);  pixel of imgA
pixB = get(imgB , ptr);  pixel of imgB
if (pixA ≤ L)&&(pixB ≤ L) then  the pixels are in the landmark
Compute xc , yc with (10) and (11)
sum y + = yc ;
sum x + = xc ;
counter++;
ptr = ptr → next;
until ptr == N U LL
yc = ys um/counter;
xc = xs um/counter;

distance = x2c + yc2 ;
return distance
724 A. Cuzzocrea et al.

7 Computer Vision Approach for Extracting Landmark

Images

We briefly summarize in this Section the computer vision operations we did on

the image acquired from the ceiling. The problem is to detect from the image
plane the isotropic images which represent the landmark. Another operation,
which is not reported here, is the identification of the landmark. The simplest
way is to draw the landmarks with different colors, since the computer vision
operations to identify the colors are very simple. There are however many other
ways which can be used for the identification, typically based on some type
of code drawn inside the landmark. Of course the computer vision operations
are slightly more complex than using different colors. More importantly, the
computer vision operations to decode drawn codes could need greater camera
resolution.
We report in Fig. 7 the Computer Vision algorithms we applied on the original
image for extraction of isotropic images. The algorithms are described as follows:

– The acquired image is first transformed in gray-scale, and then its edges are
obtained via the Canny’s operator, obtaining the Edge image.
– From the Edge image, its contours are extracted, obtaining the Contour1
image.
– The Contour1 image is processed via morphological analysis. More precisely
the opening operation with circular structuring element, is applied to Con-
tour1 image in order to eliminate the little Side Dishes. The edges are then
extracted again with the Canny operator, and then the contours are extracted
again, finally obtaining the Contour2 image.
– Ellipse fitting is applied to Contours2 image. Based on the position and size
of the found ellipses, square portions are cut out from original image. Most
likely, the landmarks are contained in one of the extracted portions.

The results are shown in Fig. 8. These results refer to the input image shown
in Fig. 3.

Fig. 7. Block diagram of the computer vision algorithms.

 An Innovative Monocular Mobile Object Self-localization 725

Fig. 8. Processed results, with reference to Fig. 3

8 Localization

The localization of the mobile object is an issue we leave open as starting from
distance estimation several possible solution can be developed. However, just to
point out a possible simple idea based on trilateration, we report Fig. 9. This
figure shows a global reference system which is related to the indoor environ-
ment is shown. Another reference system which is rotated and translated with
respect to the first one. The origin of second reference system is centered on the
camera lens of the mobile object. Note that the x − y planes shown in Fig. 9
correspond to the ceiling plane. The mobile object identify the landmarks and
knows in advance their location coordinate in the global reference system. Our
algorithm estimates the distance from the mobile object and the detected land-
marks. Therefore, we can think to draw a circle with center on the landmark

angle
of view

ca
m
er
a
re
fe
re
nc
e
sy
environment reference system st
em

Fig. 9. Sketch of a possible localization by trilateration

726 A. Cuzzocrea et al.

and radius equal to the estimated distance. If at least three different landmarks
are detected, the mobile object can be localized in global reference system.

9 Experimental Results

The experiments has been made using an Intel I7 CPU with 8cores running at
3.07 GHz and a memory of 24 GB. Then, the two dimensional DPA algorithm
has been written in the CUDA framework and executed on a NVidia Kepler TM
GK110 device. A low cost 640 × 480 webcam is used for image acquisition. In
Fig. 10 we report the average error of the estimation distance from the camera
lens and the barycenter of the landmarks. As a general consideration regarding
these results, if the camera tilt-angle is high (i.e. if the inclination of the optical
axis is close to the perpendicular to the ceiling) the error is quite small, but the
field of view turns out to be very limited. To take advantage of wider fields of
view, higher tilt-angles must be used. In this case, however, the error is higher.
Furthermore, if the light in the environment become worse, the average error
increases. Our algorithm, however, is quite robust against noise. The curve drawn
in Fig. 10 with solid line is obtained by the algorithm described in this paper.
The curve in the middle is related to the approach developed in 2019 by Avgeris
et al. and described in [18]. Finally, the higher curve is related to the work
proposed by Ogawa et al. in [20]. Despite being quite old we include this result
because its setting is very similar to this paper (the camera is directed towards
the ceiling with a tilt angle equal to 30◦ ). The errors are in any cases well above
that obtained by all the other algorithms.

Average error
15%
proposed, tilt = /3
Avgeris, tilt = /3
Ogawa, tilt= /2

10%
Error

5%

0%
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Camera - Landmark Distance [m]

Fig. 10. Average errors of the estimated distance

 An Innovative Monocular Mobile Object Self-localization 727

10 Final Remarks and Future Work

In this paper we present an algorithm to measure the distance of a mobile object
to the lightning lamps used as ceiling landmarks in indoor environment. The
algorithm has many attractive features, mainly the accuracy, which is better than
many other visual-based algorithms. Also, the distance measurements algorithm
is robust against noise. Quantization noise can be high in low lighting condition
of the environment and if the distance from landmarks and camera is high. The
negative outcome of the algorithm is the high complexity of 2D-DPA which, even
if polynomial, can lead to high computational times. In [22], however, we show
how the 2D-DPA when implemented on a NVidia Kepler TM GK110 device
leads to computation time less then 100 ms, for image size of 100 × 100 pixels.
This paper naturally opens to the development of localization algorithms
based on our distance estimation algorithm. The global localization is in fact
under development. Another open important issue is the landmark placement.
Finally, the estimation of the orientation of the mobile object is another funda-
mental problem not addressed in this paper. The use of the characteristic fre-
quencies of fluorescent lamp is an interesting method to identify the landmarks.
In this case, adaptive and artificial intelligence metaphors, perhaps inherited by
different scientific context (e.g., [30,31]), may be considered. Future works will
be focused on these open points.

Acknowledgements. This research has been partially supported by the French PIA
project “Lorraine Université d’Excellence”, reference ANR-15-IDEX-04-LUE.

Appendix A
Referring to Fig. 5, we derive below the geometric variables reported in Sect. 4.
Consider first Eq. (4). a = OC  = OC − C  C. From the right trian-
gle OCCy we have OC = sin(ϕ) h
. Moreover, from M C  C we have
C  C = M C cos ϕ =(W C − W M ) cos ϕ = ( tan h
ϕ
+ h tan Φ) cos ϕ. There-
fore
h 1
a= −h + tan Φ cos ϕ
sin(ϕ) tan ϕ
. 
1
Considering Eq.(5), we have 2b = M C sin ϕ = h tan ϕ
+ tan Φ sin ϕ.

Appendix B

We now report a sketch of the derivation of the two propositions reported in

Sect. 4.
Let us start with 7. Regarding Fig. 5, the angle formed by segments OR and
OP is equal to (Φ − γy ), so tan (Φ − γy ) = R P
OR
= [y tan ϕ −h (tan Φ tan ϕ +1)] cos ϕ
h (tan ϕ +y −h tan Φ ) cos ϕ
.
728 A. Cuzzocrea et al.

In addition to simplifying the cos ϕ, we use the definition of F and G reported

above.
G = −h (tan Φ tan ϕ + 1)
and
F = h (tan ϕ − h tan Φ) .
y tan ϕ+G
Then we have: tan (Φ − γy ) = y +F
. We conclude that y  = b
2
+
a y tan ϕ+G
y +F
By setting y = yr + yc we obtain the landmark coordinate yc
reported in (11).
Going now back to 7, let us consider Fig. 6. For lack of space we only state
that, according to considerations very similar to that just described, we can
conclude that
a sin γy (x − g)
x = g + (15)
h cos ϕ − γy
where g = GD = DI = a tan θ2x . As we did previously, we substitute
x = xr + xc in 15 and thus we can obtain xc , described in (10).

References
1. Campbell, S., O’ Mahony N., Carvalho, A., Krpalkova, L., Riordan, D., Walsh, J.:
Where am I? localization techniques for mobile robots a Review. In: 6th Interna-
tional Conference on Mechatronics and Robotics Engineering, ICMRE, Barcelona,
Spain, 12-15 February, pp. 43–47 (2020)
2. He, C., Xia, Y., Yu, C., Jiang, C.: A multi-hop distributed indoor localization
algorithm for ultra-wide-band sensor network. In: 16th International Conference on
Control, Automation, Robotics and Vision (ICARCV), 13–15 December, Shenzhen,
China, pp.1335–1340 (2020)
3. Kumar, G., Gupta, V., Tank, R.: Phase-based Angle estimation approach in indoor
localization system using bluetooth low energy. In: International Conference on
Smart Electronics and Communication (ICOSEC), Trichy, India, pp. 904–912, 10–
12 September 2020
4. Glasbey, C.A.: Two-dimensional generalisations of dynamic programming for image
analysis. Stat. Comput. 19(49), 49–56 (2009)
5. Wei, A.: Shang, survey of mobile robot vision self-localization. J. Autom. Control
Eng. 7(2), 98–101 (2019)
6. Morar, A., Moldoveanu, A., Mocanu, I., Moldoveanu, F., Radoi, I.E., Asavei, V.,
Gradinaru, A., Butean, A.: A comprehensive survey of indoor localization methods
based on computer vision. Sensors 20(9), 1–36 (2020)
7. Kim, Y.G., Park, T.H.: Localization of mobile robots from full detection of ceiling
outlines. In: Proceedings of the IEEE International Conference on Information and
Automation, Ningbo, China, pp. 1515–1520, August 2016
8. Lan, G., Wang, J., Chen, W.: An improved indoor localization system for mobile
robots based on landmarks on the ceiling. In: IEEE International Conference on
Robotics and Biomimetics (ROBIO), 3–7 December, Qingdao, China, pp. 1395–
1399 (2016)
 An Innovative Monocular Mobile Object Self-localization 729

9. Vidal, J., Lin, C-Y.: Simple and robust localization system using ceiling landmarks
and infrared light. In: IEEE International Conference on Control and Automation,
Kathmandu, Nepal, pp. 583–587, 1–3 June (2016)
10. Wang, W., Luo, Z., Song, P., Sheng, S., Rao, Y., Soo, Y.G., Yeong, C.F., Duan, F.:
A ceiling feature-based vision control system for a service robot. In: Proceedings
of the 36th Chinese Control Conference, Dalian, China, pp. 6614–6619, 26–28 July
2017
11. Ribacki, A., Jorge, V.A.M., Mantelli, M., Maffei, R., Prestes, E.: Vision-based
global localization using ceiling space density. In: IEEE International Conference
on Robotics and Automation (ICRA), Brisbane, Australia, pp. 3502–3507, 21–25
May (2018)
12. Carreira, F., Calado, J., Cardeira, C., Oliveira, P.: Navigation system for mobile
robots using PCA-based localization from ceiling depth images: experimental val-
idation. In: 13th APCA International Conference on Automatic Control and Soft
Computing, Ponta Delgada, Azores, Portugal, pp. 159–164, 4–6 June 2018
13. Carreira, F., Calado, J.M.F., Cardeira, C.B., Oliveira, P.J.C.R.: Enhanced PCA-
based localization using depth maps with missing data. J. Intell. Robot. Syst.
February 2015
14. Chatzimilioudis, G., Cuzzocrea, A., Gunopulos, D., Mamoulis, N.: A novel dis-
tributed framework for optimizing query routing trees in wireless sensor networks
via optimal operator placement. J. Comput. Syst. Sci. 79(3) (2013)
15. Cuzzocrea, A., Mansmann, S.: OLAP visualization: models, issues, and techniques,
2nd edn. In: Encyclopedia of Data Warehousing and Mining (2009)
16. Cuzzocrea, A., De Maio, C., Fenza, G., Loia, V., Parente, M.: OLAP analysis of
multidimensional tweet streams for supporting advanced analytics. In: Proceedings
of the 31st Annual ACM Symposium on Applied Computing (2016)
17. Lin, Q., Liu, X., Zhang, Z.: Mobile robot self-localizationusing visual odometry
based on ceiling vision. In: Symposium Series on Computational Intelligence, Xia-
men, China, pp. 1435–1439, 6–9 December 2019
18. Avgeris, M., Spatharakis, D., Athanasopoulos, N., Dechouniotis, D., Papavassiliou,
S.: Single vision-based self-localization for autonomous robotic agents. In: 7th Inter-
national Conference on Future Internet of Things and Cloud Workshops, Istanbul,
Turkey, pp.123–129, 26–28 August 2019
19. Zhang, C., Zhang, X.: Visible light localization using conventional light fixtures
and smartphones. IEEE Trans. Mob. Comput. 18, 2968–2983 (2018)
20. Ogawa, Y., Lee, J-H., Mori, S., Takagi, A., Kasuga, C., Hashimoto, H.: The posi-
tioning system using the digital mark pattern-the method of measurement of a
horizontal distance. In: IEEE SMC 1999 Conference Proceedings, Tokyo, Japan,
pp. 731–741 (1999)
21. Wang, W., Luo, Z., Song , P., Sheng, S., Rao, Y., Soo, Y.G., Yeong, C.F., Duan, F.:
A ceiling feature-based vision control system for a service robot. In: 136th Chinese
Control Conference (CCC), Dalian, pp. 6614–6619 (2017)
22. Cuzzocrea, A., Mumolo, E., Pirro, D., Vercelli, G.: An efficient CUDA-based
approximate two-dimensional dynamic programming algorithm for advanced com-
puter vision applications. In: IEEE International Conference on Systems, Man, and
Cybernetics, Budapest, Hungary, pp.4251–4258, 9–12 October 2016
23. He, M., Zhu, C., Huang, Q., Ren, B., Liu, J.: A review of monocular visual odom-
etry. Vis. Comput. 36(5), 1053–1065 (2020)
24. Uchida, S., Sakoe, H.: A monotonic and continuous two-dimensional warping based
on dynamic programming. Fourteenth Int. Conf. Pattern Recogn. 16–20(August),
521–524 (1998)
730 A. Cuzzocrea et al.

25. Levin, E., Pieraccini, R.: Dynamic planar warping for optical character recogni-
tion. In: 1992 IEEE International Conference on Acoustics, Speech, and Signal
Processing, San Francisco, California, USA, pp. 149–152, 23–26 March 1992
26. Lei, H., Govindaraju, V.: Direct image matching by dynamic warping. In: IEEE
Conference on Computer Vision and Pattern Recognition Workshops, Washington,
DC, USA, 27 June-2 July (2004)
27. Vinotha, K.: Bellman equation in dynamic programming. Int. J. Comput. Algo-
rithm, 05(1), 35–37 (2016)
28. Becker, C., Joaquin, S., Tokusei, K., Latombe, J.-C.: Reliable navigation using
landmarks. IEEE Conf. Robot. Autom. 2, 401–406 (1995)
29. Keysers, D., Unger, W.: Elastic image matching is NP-complete. Pattern Recognit.
Lett. 24(1–3), 445–453 (2003)
30. Cannataro, M., Cuzzocrea, A., Pugliese, A.: XAHM: an adaptive hypermedia
model based on XML. In: Proceedings of the 14th SEKE International Confer-
ence (2002)
31. Cuzzocrea, A.: Combining multidimensional user models and knowledge represen-
tation and management techniques for making web services knowledge-aware. Web
Intell. Agent Syst. Int. J. 4(3), 289–312 (2006)
 Analysis of Visualized Bioindicators Related
to Activities of Daily Living

Tomokazu Matsui1(B) , Kosei Onishi1 , Shinya Misaki1 , Hirohiko Suwa2 ,

Manato Fujimoto1 , Teruhiro Mizumoto3 , Wataru Sasaki4 , Aki Kimura4 ,
Kiyoyasu Maruyama4 , and Keiichi Yasumoto1
1 Nara Institute of Science and Technology, Ikoma, Japan
[email protected]
2 Nara Institute of Science and Technology/RIKEN, Ikoma, Japan
[email protected]
3 Osaka University, Suita, Japan

[email protected]
4 Mitsubishi Electric Corporation, Tokyo, Japan

[email protected]

Abstract. Estimating the stress associated with activities of daily living in the
home is important for understanding and improving residents’ quality of life.
However, conventional stress estimation methods are mostly used in specific sit-
uations, and stress has not been estimated for long-term activities of daily living.
In this paper, we visualize changes in stress caused by activities such as house-
hold chores by measuring bioindicators associated with daily living activities in
the home, and to obtain knowledge for establishing a method for estimating long-
term daily quality of life. To collect lifestyle and stress data, two subjects com-
pleted a lifestyle and stress measurement questionnaire daily in a smart home for
approximately two weeks. We visualized the obtained data and confirmed that the
results suggest several relationships between daily living activities and bioindi-
cators of stress level, for instance, cleaning the room and dishwashing resulted in
more stress compared to smartphone use and sleep.

1 Introduction
One of the recent research challenges in the field of information science is to improve
peoples’ quality of life (QoL) using the Internet of Things (IoT) and information and
communication technology (ICT). In particular, it is desirable to have people work on
improving their own QoL in their home, which is a private space. Therefore, there has
been a great deal of research on in-home living to improve residents’ QoL [1–4].
A promising approach to improve QoL is to reduce stress in daily life. However,
although it is possible to measure stress in daily life with simple stress checkers and
questionnaires, it is difficult to identify the causes of stress. To reduce stress in daily
life, we need to understand the stressors in our lives.
Performing household chores is an essential activity that residents must perform in
their dwellings. Neglecting household chores has a negative impact on the resident’s
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 731–744, 2021.
https://doi.org/10.1007/978-3-030-75100-5_62
732 T. Matsui et al.

hygiene and environment [5]. However, excessive household chores can cause a lot of
stress to the occupants. If we record the stresses that residents feel during household
chores, we can use this as a basis for providing an in-home service that reduces stress
and improves QoL, with a focus on household chores, which takes up most of the time
in daily life. For example, if someone has to get up early the next morning after work
and go out, he or she could perform only light household chores to prepare for the next
day’s work while still being able to complete them. It also provides an opportunity to
rethink the way we approach household chores. Specifically, to reduce the burden of
household chores, residents can improve their household chore methods, introduce new
appliances and household chore services, and review the division of household chores
between family members and housemates. However, not enough research has been done
to measure the stress (burden) of household chores.
In this paper, we propose a method for estimating QoL by measuring bioindicators
of people performing chores in the home and using these indicators to estimate stress.
We measured the life activity data of two subjects for two weeks in a smart home at the
Nara Institute of Science and Technology, and evaluated their subjective impressions
using a questionnaire. The results indicate that several relationships exist between daily
activities and bioindicators, such as the tendency of increased stress while performing
various household chores and the tendency of relaxation during rest and entertainment.

2 Related Work

Here, we describe existing studies related to QoL improvement and stress estimation.
Quality of life is a measure of satisfaction with one’s living conditions. The study of
QoL [6] started originally as a concept to discuss a patient’s QoL after medical treat-
ment. Nowadays, it is used not only in the medical field, but also as a concept related to
general QoL, such as work–life balance and happiness. Work–life balance describes the
harmony (or disharmony) between work and daily life when not working, and the idea
is to improve and enhance QoL by balancing work and home life and avoiding mental
stress from work and depression from physical fatigue.
In particular, QoL that is directly related to a person’s health is termed health-related
QoL (HRQoL), and is evaluated across various domains such as physical state, psy-
chological state, social interaction, economic and occupational state, and religious and
spiritual state [7]. The World Health Organization (WHO) is trying to quantitatively
evaluate HRQoL with their own metric, called the WHOQOL [8, 9] and short form
WHOQOL (SF) [10]. These indices are evaluated using paper questionnaires, but the
WHOQOL-100 form [8] requires answers for 100 items in six domains, and the SF-36
form [10] requires answers for 36 items in eight domains. Because of the heavy burden
of completing these long questionnaires, routine assessment of the QoL is a difficult
task.
Amenomori et al. [11] proposed a method to continuously measure HRQoL with
a smaller questionnaire by exploiting tracking and biometric data from smartphones
and smartwatches. In addition, Asma et al. [12] used the Hamilton Depression Rating
Scale to identify negative psychological states; Garcia et al. [3] proposed the Oldenburg
Burnout Inventory to grasp working stress; and Natasha et al. [4] proposed a method to
 Analysis of Visualized Bioindicators Related to Activities of Daily Living 733

quantitatively measure and estimate the next day’s health, stress, and happiness using a
smart device. These studies used smart devices to improve QoL by measuring and esti-
mating the user’s state to provide feedback to the user. However, the causes of decreased
QoL in daily life have yet to be determined.
One possible reason for decreased QoL is the stress of daily life. There are many
ways to estimate the stress a person is feeling [13, 14]. Here, we describe a method for
estimating stress from biometric information acquired by sensors.
Heart rate variability can be an indicator of various physiological functions, includ-
ing autonomic functions [15, 16]. Heart rate variability is measured non-invasively using
small sensors and thus can be used to estimate stress without significant inconvenience.
However, without information related to the activities of daily life, identifying the activ-
ities that cause stress is impossible, making it difficult to make adjustments that could
reduce stress levels.
In this study, we visualized changes in stress caused by household chores using
bioindicators associated with activities of daily living in the home. Our purpose is to
obtain knowledge for establishing a method for estimating QoL in long-term daily life.
In the experiment, two participants completed an initial questionnaire and a pre-bedtime
questionnaire during the experiment, and data collection was conducted in a smart home
testbed of the Nara Institute of Science and Technology.

3 Proposed Method

To correlate life activities with stress, we measured stress-related bioindicators, col-

lected subjective assessments, and tracked life activities. We describe how they were
measured in the following sections.

3.1 Stress-Related Bioindicators

The heartbeat has periodic variation called heart rate variability. Analysis of this vari-
ability can be used to estimate the stress state [15, 16].
Various methods have been proposed for heart rate variability analysis. Frequency-
domain analysis uses the low-frequency/high-frequency (LF/HF) ratio. LF is the power
spectrum in the low-frequency range of 0.04 to 0.15 Hz and represents the activity of
sympathetic and parasympathetic nerve activity associated with blood pressure varia-
tion. HF is the power spectrum in the high-frequency range of 0.15 to 0.4 Hz and repre-
sents the activity of the sympathetic and parasympathetic nerve activity associated with
respiratory variation. Thus, the LF/HF ratio indicates the activity of the sympathetic
nervous system. In a relaxed state, both LF and HF signals appear, but the HF compo-
nent is relatively high and the LF/HF ratio is low. In a stressful state, the LF/HF ratio
is higher because the HF component is less prominent. Therefore, if an individual has
a high LF/HF ratio during a particular activity, it is presumed that he or she is feeling
stress. Thus, heart rate variability may be an important indicator in stress estimation.
The heart rate interval between electrocardiogram R peaks, or the RR interval (RRI),
and the standard deviation (SD) of the RRI are also important indices for estimating
734 T. Matsui et al.

stress [17, 18]. Toyohuku et al. [19] proposed a simple method for estimating parasym-
pathetic nerve activity using a Lorenz plot (also called a Poincaré plot) of the RRI. In
this method, the n-th value of the RRI is plotted on a graph with the n-th value on the
horizontal axis and the n + i-th value on the vertical axis. Specifically, all the plot points
are projected onto the y = x and y = −x axes, the SD of the distance from the origin
(0, 0) of the y = x axis is given as σx , the SD of the distance from the origin (0, 0) of
the y = −x axis is given as σ−x , and the ellipse output from the following equation is
the area:
S = π × σx × σ−x , (1)
where the area of the output ellipse represents the magnitude of the RRI fluctuation.
This method has been used to judge whether the parasympathetic nervous system has
been activated [20–22]. In this study, the LF/HF ratio and RRI were measured, visual-
ized, and analyzed for stress estimation.
This was accomplished with a chest-mounted WHS-3 heart rate sensor [23], which
can measure the LF/HF ratio and transmit the data with Bluetooth communication.
Then, the LF/HF ratio can be displayed and recorded on a tablet device.

3.2 Subjective Assessments Related to Stress

Questionnaires were administered before the experiment and before bedtime each day
to collect the subjects’ subjective evaluations. Subjects reported their preferences for
performing household chores and their level of fatigue. Table 1 lists the questions about
the preferences and strengths of each household chore performed before the experiment,
where each item was rated on a 5-point scale. Table 2 shows the pre-bedtime question-
naire (7-point scale except for the first three questions) about the stress and fatigue of
each household chore performed during the day.
The pre-bedtime questionnaire in Table 2 is based on the NASA Task Load Index
(NASA-TLX) [24], which was developed to visualize the factors of workload, and
designed as a questionnaire about all the chores performed throughout the day. There are
six NASA-TLX rating scales: Mental Demand, Physical Demand, Temporal Demand,

Table 1. Content of pre-experiment questionnaire.

Item Description
Washing preferences (up to 5) Psychological factors such as likes and dislikes
Cleaning preferences (up to 5)
Cooking preferences (up to 5)
Dishwashing preferences (up to 5)
Washing skill (up to 5) A high degree of skill, such as being familiar
Cleaning skill (up to 5)
Cooking skill (up to 5)
Dishwashing skill (up to 5)
Frequency of doing household chores (description) How to do your daily chores, and how to order them?
Housework needs (up to 5)
 Analysis of Visualized Bioindicators Related to Activities of Daily Living 735

Table 2. Content of pre-bedtime questionnaire.

Item Description
Current physical condition (up to 5) Current status grasping
Current fatigue Level (up to 5)
Satisfaction or sense of accomplishment in accomplishing
household chores (up to 5)
The most stressful chores Psychological factors due to the stress of
household tasks (one of “laundry, cleaning,
Laundry stress (up to 7)
cooking, and washing dishes”)
Cleaning stress (up to 7)
Cooking Stress (up to 7)
Dishwashing stress (up to 7)
Household chores that caused the most fatigue (physically) Physical elements of household tasks due to fatigue
(one of “laundry, cleaning, cooking, and
Laundry fatigue (up to 7)
dishwashing”)
Cleaning fatigue (up to 7)
Cooking fatigue (up to 7)
Dishwashing fatigue (up to 7)
“Mental Demand: How much mental and perceptual activity Workload factor identification
was needed? (up to 7)
Criterion: choose the one with a larger load if you feel that the
task demands more or the content is more complex”
“Physical Demand: How much physical activity did you need?
(up to 7) Criteria: if you feel that the amount of work required
by the task is too much, choose the one with a larger load”
“Temporal Demand: How much time pressure did you feel for
the progress of the task, the rate of progress, and the timing of
the task occurring? (up to 7) Criterion: if you are pressed for
time and in a hurry to do the work, choose the one with the
greater load”
“Performance: How well do you think you are doing in setting
tasks in accomplishing your goals? (up to 7) Criterion: in this
case, if I felt I could perform the household tasks well, I would
choose satisfaction”
“Effort: How hard did you work on the required task? (up to 7)
Criterion: if you are forced to use a lot of ingenuity in your own
way to do the work, choose the one with a larger load”
“Frustration: How did you feel frustrated, stressed, satisfied, or
relaxed while performing the task? (up to 7) Criterion: choose
the one with a larger load if you were feeling frustrated when
performing the task”

Performance, Effort, and Frustration. Although NASA-TLX is mainly used to identify

workload factors, it was used to measure the various demands of the overall housework
activity in this study. The questionnaire was administered at the end of each day, and
each item is rated on a 7-point scale.

3.3 Daily Activities

Participants recorded the start and end times of each daily activity using an activity
recording application on a tablet device. Figure 1 shows the operation screen of the
736 T. Matsui et al.

Fig. 1. Activity annotation application.

Activity Recording Application for Everyday Life. Each activity has an icon that can
be tapped to switch from one activity to another. The current activities are indicated by
the icon turning blue; icons representing all other activities remain gray.

4 Experiment

This section describes the experiment we conducted to collect activity data and bioindi-
cator data on daily life, and visualize the data to clarify the relationship between activ-
ity and bioindicators. This experiment was approved by the Ethics Review Board of
the Nara Institute of Science and Technology (Approval Number 2019-I-9). The RRI
and LF/HF ratio values in the data were preprocessed using an interquartile range to
remove outliers. Note that the data on bathing activity were measured during the time
just before bathing to the removal of the measurement device and from the time of wear-
ing the device again after bathing to the operation of the application to select another
activity.

4.1 Experimental Methods

Two participants (students) lived in the one-bedroom smart home testbed at the Nara
Institute of Science and Technology during the data collection experiment, and data
were collected for 13 days from December 27, 2019, to January 8, 2020. The experiment
was conducted for the first 7 d for subject ID01 and for the later 6 days for subject ID02.
Table 3 lists definitions of the target activities and definitions of the target activities’
start and end. We asked the participants to fill out a questionnaire before the experiment
and to complete a bedtime questionnaire every day during the experiment term. The
participants were instructed to wear the WHS-3 at all times during the experiment,
except when bathing. We also instructed them to record the start time and end time
of each activity of daily living by using the application, which was developed for this
experiment for mobile devices. These were the only restrictions on the participants.
 Analysis of Visualized Bioindicators Related to Activities of Daily Living 737

Table 3. Target activities.

Activity Definition of activity Start End

StayingAtHome While at home Going in through the Stepping out of the doorway
front door
PersonalHygiene Hygiene activities Entering the bathroom Leaving the bathroom
such as hand washing
and tooth brushing
Cooking Cooking ingredients Entering the kitchen Cooking finish
in the kitchen
Eating Eating a food Start eating the food Finish eating the food
WashingDishes Washing dishes Start washing the Dishwashing is over
dishes
UsingToilet Going to the Stepping out of the Going in through the front door
bathroom doorway
Bathing Washing the body in Entering the bathroom Getting out of the bathroom
the bathroom and
soaking in the bathtub
CleaningRoom Vacuuming the room Turn on the vacuum Turn off the vacuum cleaner
cleaner
CleaningBathroom Cleaning the Entering the bathroom Getting out of the bathroom
bathroom
Laundry Putting the laundry in Finish putting in the Finish folding the laundry
the washing machine laundry, start folding
the laundry
WorkingOnPC Working with a PC Open PC Close the PC
WatchingTelevision Watching TV Turn on the TV Turning off the TV
PlayingGames Playing video games Turn on the game Turning off the game console
console
ReadingBook Reading books Open the book Close the book
UsingSmartphones Using smartphones Operating with a Put your smartphone
smart phone
Sleeping Sleeping Getting into bed Get out of bed
Dressing Preparing for an Start preparing Ready to go
outing, such as
makeup and changing
clothes

4.2 Analysis Results

In this section, we first present the results of the analysis of the overall tendency by
participant, followed by the results of analysis by activity and Lorenz plots.

4.2.1 Results of Overall Tendency Analysis

Tables 4 and 5 show results of the pre-bedtime questionnaire and the pre-experiment
questionnaire, respectively. For the pre-bedtime questionnaire, Table 4 gives the average
738 T. Matsui et al.

Table 4. Results for pre-bedtime questionnaire.

Items ID01 ID02

Current physical condition (up to 5) 3.71 2.83
Current fatigue level (up to 5) 3.14 3.00
Satisfaction or sense of accomplishment in 3.00 3.17
accomplishing household chores. (up to 5)
Chores that were the most stressful Cooking Cooking
Laundry stress (up to 7) 2.14 3.50
Cleaning stress (up to 7) 1.71 2.83
Cooking stress (up to 7) 3.29 3.4
Dishwashing stress (up to 7) 2.57 2.67
Housework that was the most exhausting. (physically Cooking Cooking
and mentally)
Laundry fatigue (up to 7) 2.00 3.20
Cleaning fatigue (up to 7) 2.00 3.20
Cooking fatigue (up to 7) 3.57 3.17
Dishwashing fatigue (up to 7) 2.86 2.33
Mental demand (up to 7) 4.14 4.67
Physical demand (up to 7) 3.57 2.17
Temporal demand (up to 7) 2.57 2.67
Performance (up to 7) 4.71 2.17
Effort (up to 7) 3.43 3.33
Frustration (up to 7) 3.14 3.17

responses to the questions, most of which were rated on a scale of 1 to 7, and the ques-
tions requiring a textual response (to select household chores) show the most frequent
responses. The RRI histograms for each participant obtained throughout the experiment

Table 5. Results for pre-experiment questionnaires. (up to 5)

Items ID01 ID02

Laundry preferences 5 3
Cleaning preferences 2 5
Cooking preferences 3 4
Dishwashing preferences 4 2
Laundry skill 5 2
Cleaning skill 2 4
Cooking skill 3 2
Dishwashing skill 4 4
Frequency of doing Often do housework after Perform chores when needed
household chores accumulating
The need for household 5 4
chores
 Analysis of Visualized Bioindicators Related to Activities of Daily Living 739

are shown in Fig. 2, and the LF/HF ratio histograms for each participant are shown in
Fig. 3.
Figure 2 shows that the shape and location of the frequency distribution of RRI
differed between participants. The histogram of participant ID01 has a multimodal dis-
tribution, suggesting that the variation of RRI between activities is relatively large, and
many RRI values are overall in a low range. In contrast, the histogram of ID02 is a nor-
mal distribution, and the RRI variation between activities is relatively small. From the
results of the bedtime questionnaire, ID01 rated him/herself less stressed and fatigued
from housework than did ID02. In contrast, ID01 is evaluated as being overloaded, as
indicated by the NASA-TLX-referenced questions, such as various requirements and
work generation, which may affect the RRI.
Figure 3 shows that the LF/HF ratio histograms do not show as clear a distribu-
tional difference as the RRI histograms. The ID01 distribution has more frequencies in
the LF/HF ratio region than does the ID02 distribution. In other words, ID01 is more
susceptible to stress due to daily activities than ID02. Regarding ID02, the frequency
decrease is larger in the region around the LF/HF ratio of 3, compared with ID01. Thus,
the figures indicate that activities can be divided into those with larger LF/HF and those
with lower LF/HF.

4.2.2 Analysis Results by Activity

Table 6 shows the total duration of each activity for each participant, as well as the
corresponding means and SDs of the RRI and LF/HF ratio for each activity. In this
experiment, participants were encouraged to perform natural activities as usual, and
therefore, the duration of each activity varied from subject to subject. Both participants
tended to spend less time cleaning the room and dishwashing compared to sleeping
and PC work. Also, a large difference appears between participants in the durations of
activities such as cooking and TV watching. Thus, the participants engaging in stressful
activities attained a higher index of stress during the experimental period, which may
affect the shape of the distribution of bioindicator histograms. Some activities may be
performed in parallel, such as watching TV during household chores. In this case, the
collected biometric indices are used to calculate the activities’ RRIs, LF/HF, and Lorenz
plots as though they affect all activities taking place at that time.
The standard deviation of the RRI (hereafter, SDNN) is lower for both participants
for all household activities (room cleaning, cooking, laundry, and dishwashing) com-
pared to the other activities. The SDNN is known to be lower in a stressed state and
higher in a relaxed state, indicating higher stress in the overall activity. The results of
the pre-experiment questionnaire revealed that participants ID01 and ID02 do not pre-
fer cleaning and dishwashing, respectively, as their household chores. Each participant’s
SDNN is relatively low for each activity, suggesting an association between household
chore preferences and biometric indices.
RRI boxplots for each participant’s activities are shown in Figs. 4a and 4b, and the
corresponding LF/HF ratio boxplots are shown in Figs. 5a and 5b. From Figs. 4a and 4b,
the data for household chores activities show relatively low RRI variability, indicating
a stressed state. In contrast, data for sleep and smartphone use show relatively high RRI
variability, indicating a relaxed state. Figures 5a and 5b show that watching television
740 T. Matsui et al.

Fig. 2. RRI histograms.

Fig. 3. LF/HF ratio histograms.

Table 6. Statistics for each activity.

Activity ID01 ID02
Time [min] RRI mean [ms] RRI SD [ms] LF/HF mean LF/HF SD Time [min] RRI mean [ms] RRI SD [ms] LF/HF mean LF/HF SD
Bathing 140.1 554.9 202.7 1.015 0.993 109.2 643.2 112.9 3.883 2.070
CleaningBathroom 15.6 460.2 52.1 1.758 1.337 0.0 0.0 0.0 0.0 0.0
CleaningRoom 25.4 455.5 29.3 1.237 0.959 29.6 724.6 72.6 3.835 1.773
Cooking 100.6 563.2 37.9 3.816 2.450 222.5 734.6 87.2 4.448 2.243
Dressing 15.0 578.0 137.3 2.492 1.714 26.6 658.7 72.3 4.053 2.161
Eating 163.0 555.2 54.2 4.175 2.291 162.1 748.8 118.2 4.431 2.201
Laundry 41.8 560.6 48.2 3.259 2.119 6.4 756.0 76.0 2.695 1.618
PersonalHygiene 23.1 614.8 84.0 5.086 2.059 22.5 766.4 87.4 3.284 1.650
ReadingBook 143.2 652.9 52.4 3.388 1.998 291.4 883.2 103.8 3.422 2.056
Sleeping 4180.6 846.8 95.6 2.284 2.112 3276.9 1106.6 199.4 1.764 1.710
StayingAtHome 8634.9 737.4 133.8 2.965 2.235 6024.6 979.8 210.4 2.458 2.013
UsingSmartphone 1016.1 745.5 100.3 4.092 2.268 199.4 936.7 162.6 3.120 2.078
UsingToilet 135.0 672.8 113.3 2.995 2.171 91.5 728.4 85.8 2.379 2.369
WashingDishes 43.3 586.1 52.9 3.832 2.664 40.2 654.3 67.2 4.591 1.970
WatchingTelevision 2073.5 617.2 70.2 3.976 2.211 177.9 868.0 122.7 2.337 1.647
Working on PC 3013.0 655.1 62.1 3.557 2.096 1226.2 919.0 109.6 2.699 1.827

and personal hygiene have different stress tendencies (high for ID01 and low for ID02)
between participants, showing the effects of individual differences, the type of content
watched on television, and different physical conditions of the participants.
According to the results of the pre-bedtime questionnaire, the most frequent stress-
ful activity of each participant is cooking; however, according to each boxplot, the RRI
tended to be relatively low and the LF/HF relatively high during this activity, reflecting
subjective evaluations. Comparison with the results of the pre-experiment questionnaire
shows that the LF/HF values of ID02’s cooking and dishwashing activities are higher
 Analysis of Visualized Bioindicators Related to Activities of Daily Living 741

Fig. 4. RRI boxplots.

Fig. 5. LF/HF ratio boxplots.

when he or she had a high preference for the activity. In contrast, the self-reported
preferences and strengths for activities may not always correspond to the LF/HF ratio,
as seen in the washing and cleaning activities of ID01. This result is affected by the
influence of the devices, each bioindicator, and other mental or physical factors not
considered. In particular, bioindicators such as LF/HF ratio and RRI are affected by var-
ious mental and physical factors, such as psychological stress, individual differences,
and chronic stress from the environment. Therefore, while the results obtained by this
experiment can be regarded as a guide for estimating QoL for general living activities
in the home, they are not absolute.

4.2.3 Analysis Results with Lorenz Plots

The Lorenz plot for each participant is shown in Figs. 6a and 6b. The Lorenz plot of
ID02 showed that the data tended to be more scattered in the high RRI region and
denser in the low RRI region. This indicates that fluctuations in the RRI are large, and
the parasympathetic nervous system is activated during sleep and relaxation. In contrast,
the Lorenz plot of ID01 shows that the data spread in the high RRI region is slightly
742 T. Matsui et al.

Fig. 6. Lorenz plots for all activities.

Fig. 7. Areas of lorenz plots.

wider than that in the low RRI region, and that the variability of data in the high RRI
region is smaller. In other words, the fluctuation of the heart rate interval is small during
sleep and relaxation, suggesting the possibility that the participants were not able to
rest sufficiently. There is no significant difference in the shape of the Lorenz plots for
each activity. Overall, the activities with a small area on the Lorenz plots show a wide
distribution only in the y = x direction with small differences in shape and clustering
in the low RRI region. For activities with large areas, the distribution is wide not only
in the y = x direction but also in the y = −x direction, and extends into the high RRI
region.
Figures 7a and 7b show the area of the Lorenz plot for each participant’s activity.
The figure shows that, for both participants, sleeping, smartphone use, and watching TV
have a large Lorenz plot area and activated the parasympathetic nervous system. The
area is small for various household tasks, and the parasympathetic nervous system is
inactive, indicating that they are under stress. The area of Lorenz plots corresponding to
 Analysis of Visualized Bioindicators Related to Activities of Daily Living 743

each participant’s activity is also small, although participants ID01 and ID02 indicated
cleaning and dishwashing as their least preferred activities in the pre-experiment ques-
tionnaire. These results reveal that household chore preferences and loads can be esti-
mated from the area information of Lorenz plots calculated from the RRI. The Lorenz
plot area of ID01’s bathing activity is considered to be abnormally high. This is the data
obtained within the range while sensor attachment and detachment happened.

5 Conclusion
In this paper, we visualized the subjective evaluation of stress and bioindicators for each
activity for estimating the QoL variability in daily life activities. The results showed a
tendency of increased stress during various household chores and a tendency of relax-
ation during rest and recreation, although the subjective evaluations did not necessarily
coincide with the stress indicated by the bioindicators. A more generalizable evalu-
ation requires a longer-term experiment with more participants. As future work, we
will explore the possibility of providing services, such as optimal timing estimation for
housekeeping and relaxation activity recommendations based on daily life activities, by
accounting for the chronic stress to which each subject is subjected and personal charac-
teristics, such as the ability to recover from stress. Generally, parasympathetic activity
varies greatly among individuals, and real-time stress assessment by Lorenz plot area
is difficult because of the need to normalize the area for each person. However, if we
generate Lorenz plots in advance and calculate the area at rest or during each activity,
real-time stress assessment for each individual is possible by calculating the difference
between the usual area and the area at a specific time.

Acknowledgements. This study was supported in part by the Japan Society for the Promotion
of Science, Grants-in-Aid for Scientific Research numbers JP20H04177 and JP19K11924.

References
1. Al-Surimi, K., Al-Harbi, I., El-Metwally, A., Badri, M.: Quality of life among home health-
care patients in Saudi Arabia: household-based survey. Health Qual. Life Outcomes 17(1),
1–9 (2019)
2. Imanishi, M., Tomohisa, H., Higaki, K.: Impact of continuous in-home rehabilitation on
quality of life and activities of daily living in elderly clients over 1 year. Geriatr. Gerontol.
Int. 17(11), 1866–1872 (2017)
3. Garcia-Ceja, E., Osmani, V., Mayora, O.: Automatic stress detection in working environ-
ments from smartphones’ accelerometer data: a first step. IEEE J. Biomed. Health Inform.
20(4), 1053–1060 (2015)
4. Jaques, N., Taylor, S., Nosakhare, E., Sano, A., Picard, R.: Multi-task learning for predicting
health, stress, and happiness. In: NIPS Workshop on Machine Learning for Healthcare (2016)
5. Cerrato, J., Cifre, E.: Gender inequality in household chores and work-family conflict. Front.
Psychol. 9, 1330 (2018)
6. Group, W., et al.: The development of the world health organization quality of life assessment
instrument (the whoqol). In: Quality of Life Assessment: International Perspectives, pp. 41–
57. Springer, Heidelberg (1994)
744 T. Matsui et al.

7. McKenna, S.: Quality of life and pharmacoeconomics in clinical trials (1997)

8. Group, W.: The world health organization quality of life assessment (whoqol): development
and general psychometric properties. Soc. Sci. Med. 46(12), 1569–1585 (1998)
9. Group, W., et al.: Development of the world health organization whoqol-bref quality of life
assessment. Psychol. Med. 28(3), 551–558 (1998)
10. Ware Jr., J.E.: Sf-36 health survey update. Spine 25(24), 3130–3139 (2000)
11. Amenomori, C., Mizumoto, T., Suwa, H., Arakawa, Y., Yasumoto, K.: A method for simpli-
fied hrqol measurement by smart devices. In: International Conference on Wireless Mobile
Communication and Healthcare, pp. 91–98. Springer (2017)
12. Ghandeharioun, A., Fedor, S., Sangermano, L., Ionescu, D., Alpert, J., Dale, C., Sontag,
D., Picard, R. : Objective assessment of depressive symptoms with machine learning and
wearable sensors data. In: 2017 Seventh International Conference on Affective Computing
and Intelligent Interaction (ACII) (2017)
13. Lovibond, P.F., Lovibond, S.H.: The structure of negative emotional states: comparison of the
depression anxiety stress scales (DASS) with the beck depression and anxiety inventories.
Behav. Res. Ther. 33(3), 335–343 (1995)
14. Tennant, R., Hiller, L., Fishwick, R., Platt, S., Joseph, S., Weich, S., Parkinson, J., Secker, J.,
Stewart-Brown, S.: The Warwick-Edinburgh mental well-being scale (WEMWBS): devel-
opment and UK validation. Health Qual. Outcomes 5(1), 63 (2007)
15. Kim, H.G., Cheon, E.J., Bai, D.S., Lee, Y.H., Koo, B.H.: Stress and heart rate variability: a
meta-analysis and review of the literature. Psychiatry Invest. 15(3), 235 (2018)
16. Sztajzel, J., et al.: Heart rate variability: a noninvasive electrocardiographic method to mea-
sure the autonomic nervous system. Swiss Med. Wkly. 134(35–36), 514–522 (2004)
17. Baek, H.J., Lee, H.B., Kim, J.S., Choi, J.M., Kim K.K., Park, K.S.: Nonintrusive biological
signal monitoring in a car to evaluate a driver’s stress and health state. Telemedicine e-Health,
15(2), 182–189 (2009)
18. Yamamoto, K., Toyoda, K., Ohtsuki, T.: Spectrogram-based non-contact RRI estimation by
accurate peak detection algorithm. IEEE Access 6, 60369–60379 (2018)
19. Toyofuku, F., Yamaguchi, K., Hagiwara, H.: Simplified method for estimating parasympa-
thetic nerves activity by Lorenz plot of ECG RR intervals. Jpn. J. Ergon. 42(Supplement),
512–515 (2006)
20. Brennan, M., Palaniswami, M., Kamen, P.: Poincare plot interpretation using a physiological
model of HRV based on a network of oscillators. Am. J. Physiol.-Heart Circulatory Physiol.
283(5), H1873–H1886 (2002)
21. Kamen, P.W., Krum, H., Tonkin, A.M.: Poincare plot of heart rate variability allows quan-
titative display of parasympathetic nervous activity in humans. Clin. Sci. 91(2), 201–208
(1996)
22. Mourot, L., Bouhaddi, M., Perrey, S., Rouillon, J.-D., Regnard, J.: Quantitative poincare plot
analysis of heart rate variability: effect of endurance training. Euro. J. Appl. Physiol. 91(1),
79–87 (2004)
23. Union Tool Co. Sensor products. https://www.uniontool.co.jp/en/product/sensor/. Accessed
31 Jan 2021
24. Hart, S.G.: Nasa task load index (TLX). volume 1.0; computerized version (1986)
 Non-contact Person Identification
by Piezoelectric-Based Gait Vibration Sensing

Keisuke Umakoshi1 , Tomokazu Matsui1 , Makoto Yoshida1,2 , Hyuckjin Choi1 ,

Manato Fujimoto1(B) , Hirohiko Suwa1,3 , and Keiichi Yasumoto1
1 Nara Institute of Science and Technology, Ikoma, Japan
{manato,h-suwa}@is.naist.jp
2 Onkyo Corporation, Higashiosaka, Japan

[email protected]
3 RIKEN, Wako, Japan

Abstract. Floor vibrations caused by walking person (hereinafter, gait vibra-

tions) have recently been explored as a way to determine their locations and iden-
tities, and the technologies for measuring such gaits may enable low-cost elderly
monitoring services and crime prevention systems. In this paper, we report on
the development of a system that can accurately capture both high- and low-level
signal gait vibrations using a piezoelectric sensor, and then propose a new sys-
tem that can identify a walking person based on a small number of footsteps. Our
proposed system uses two key approaches to accurately obtain such gait vibra-
tions. The first uses a combination of a source follower circuit in parallel and a
piezoelectric sensor. The second involves widening the dynamic range by the use
of a dual power supply drive. We then show how we can increase the accuracy of
our system by combining multiple footsteps rather than using a single footstep,
and thus achieve a more robust system regardless of the distance between the
sensor and the target. In experiments comparing five different machine learning
(ML) models conducted with six test participants to evaluate our system, person
identification results obtained using only a single footstep showed accuracy lev-
els up to 70.8% of average F-measure when using the Light Gradient Boosting
Machine (LightGBM) classifier, while for other methods, the average F-measures
were 63.1%, 75.9%, and 87.1% in cases of using the first, first and second, and
from first to third footsteps from each back-and-forth walk test, respectively.

1 Introduction
Amidst the ongoing improvements to sensing technologies, various companies and
research institutes have been working on estimating human activities and their contexts
as a means to providing more sophisticated services. When attempting to grasp the
human context, it is essential to answering a number of questions (Where is it happen-
ing? When did it happen? Who is involved? What are they doing and why?) to under-
stand a person’s activities in detail. Most existing studies related to person identification
methods seek ways to determine such information, even though it is often difficult to
identify. However, the research continues for effective methods because inexpensive

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 745–757, 2021.
https://doi.org/10.1007/978-3-030-75100-5_63
746 K. Umakoshi et al.

and accurate person identification procedures have the potential to make excellent con-
tributions to elderly monitoring services and crime prevention systems.
Many previous methods have been developed to help identify people [1–8]. Of
these, biometrics [1, 2] are among the best-known technologies. However, face recog-
nition [1] and fingerprint recognition [2] systems are often burdensome because they
require the user to perform certain predetermined tasks, such as looking at the cam-
era for authentication. Other methods involve extracting personal movements for iden-
tification with sensors such as cameras, Wi-Fi Channel State Information (CSI), and
backscatter communication. However, it is difficult for camera-based methods [3] to
distinguish identities if the subject is not clearly visible in the camera view field due to
bright lighting, obstacles, etc. Also, Wi-Fi-based walking identification methods [4, 9]
require observations of a large number of footsteps, while backscatter communication-
based methods [5] require very strong radio power, which limits its availability. Cur-
rently, the person identification technology using gait data (gait vibrations) [6–8] has
been attracting attention as a way to resolve these problems. However, while the iden-
tification accuracy achieved with this technology is high, there are still unsolved prob-
lems. For example, the method is not always usable in narrow environments and it is
often expensive to install.
In response to these issues, we herein propose a new system that efficiently detects
low-level signal gait vibrations with a single piezoelectric sensor and can distinguish
individuals with high accuracy levels using data from, e.g., three footsteps or fewer.
The major contributions of this paper are summarized as follows.
First, we designed a circuit to detect both high- and low-level gait vibration signals
by (1) widening the output signal level of a piezoelectric sensor using a field-effect tran-
sistor (FET), (2) designing a low-noise circuit using a circuit simulator, and (3) widen-
ing the dynamic range of the newly designed circuit by adopting a dual supply battery
drive for voltage amplification. The resulting system can obtain accurate gait vibration
data even when the gait vibration signal level is low. Second, we carried out evaluation
experiments utilizing the proposed system with the designed low-noise circuit. Specifi-
cally, we evaluated the personal identification accuracy utilizing five different machine
learning models (SVM, Random Forest, Xgboost, LightGBM, DNN). As a result, the
average F-measure for the person identification using only a single footstep was up to
70.8% when using the LightGBM classifier. Finally, to identify individuals with just
a few footsteps, we evaluated our system in three ways: using the first footstep, the
first and second footsteps, and from the first to third footsteps, of each back-and-forth
walking test, which will be described in detail later. The obtained results showed that
the average F-measures when using the LightGBM classifier were 63.1%, 75.9%, and
87.1%, respectively.

2 Related Work

Recently, various studies utilizing gait vibrations generated in buildings have been con-
ducted [10]. In this section, we describe the existing research related to our work.
 Non-contact Person Identification by Piezoelectric-Based Gait Vibration Sensing 747

2.1 Geophones/Seismic Sensor-Based System

Zhang et al. [11] proposed a method for estimating the number of people in a building.
To accomplish this, they installed geophones at five locations under floors in a building
and counted people if their gait vibrations exceeded a preset threshold. When evalu-
ating the effectiveness of their proposed method, they compared the predicted number
of people with the log data of the room access control system and obtained a correla-
tion coefficient of 0.836. Separately, Mirshekari et al. [12] detected events such as door
opening, closing, and walking persons by acquiring floor vibrations and using machine
learning (ML) to analyze those events. That work showed that adapting the training
model generated at a specific location to different locations improved the event detec-
tion rate by up to 10 times compared to other methods. However, these studies focused
on estimating the number of people or detecting events without considering the issue of
person identification.
Meanwhile, Pan et al. [6] proposed a person identification system called FootprintID
that combines supervised learning and iterative transductive learning models in which
they applied ML to gait data acquired in the case of an average walking speed. They
then annotate that gait data with various walking speed data using iterative transduc-
tive learning to generate accurate training models. This method reduced the amount
of labeled data and achieved an accuracy rate of up to 96% for 10 participants. How-
ever, this system did not use low-level signal gait vibrations that are sensitive to noise,
requires seven footsteps of data to achieve sufficiently high accuracy, and could only
identify individuals in wide open space. In other words, this system can only be used
under specific conditions, such as within large rooms.
Clemente et al. [7] constructed a highly accurate person identification system by
adding new features based on [13]. In their system, four seismic sensors are installed
at the four corners of a 2.5 m × 3.0 m area and used to identify individuals using a
simple majority voting method. In that system, F-measure of 63% was obtained for
each individual sensor, while F-measure of 71.9% was obtained using a voting method
based on the weighting of each sensor. However, this system is expensive to install and
requires at least four sensors to identify an individual with one footstep.

2.2 Piezoelectric Sensor-Based System

To resolve the problems described in Sect. 2.1, we decided to base our system on piezo-
electric sensors, which have been attracting attention due to their low cost and ease of
installation in furniture and floors. Another attractive characteristic is that piezoelectric
sensors are mostly insensitive to sounds from speakers and human voices, and are thus
relatively resistant to environmental noises, which is a difference from a normal micro-
phone. Hence, even with obstacles in the vicinity, vibrations transmitted to the floor
or furniture are less affected by the noise than wireless communication-based systems
such as Wi-Fi.
Kashimoto et al. [14] proposed a system that estimated a user’s location by using
various vibrations generated by the user’s activity. However, this study focused solely
on location estimation rather than individual identification. In addition, this system is
748 K. Umakoshi et al.

based on a circuit with insufficient voltage amplification for obtaining fine-granularity

floor vibration data.
Separately, Akiyama et al. [8] proposed a method of estimating a user’s walking
direction using two horizontally installed piezoelectric sensors. Their system generated
differential data from the acquired gait vibrations and used linear discriminant analy-
sis to estimate the user’s walking direction. Our system uses parallel source-follower
circuits. Source-follower circuit is a circuit that lowers the output impedance while
keeping the voltage constant. This significantly increases the current amplification, and
thus allows detailed walking vibrations to be obtained. From the acquired results, it
was possible to classify the walking direction of a person with about 83% accuracy.
However, this system only estimates the walking direction without identifying the indi-
vidual. Additionally, despite the use of a single power supply consisting of two AA
batteries, the dynamic range is relatively narrow and low-level signal gait vibrations
cannot be correctly acquired due to noise. Furthermore, when the amplifier circuit gain
is adjusted to obtain low-level signal gait vibrations, the original dynamic range is nar-
row and the high-level signal gait vibration becomes saturated (unusable) in the sensor
vicinity.

2.3 Our Approach

To resolve the problems highlighted in Sects. 2.1 and 2.2, we constructed a new system
that can efficiently detect low-level signal gait vibrations using a single piezoelectric
sensor, and thus identify individuals with high levels of accuracy using as few footsteps
as possible. In tests conducted to evaluate our system, we begin by assuming that a
single user is walking within the detection range of our piezoelectric sensor.

3 System Requirements
To realize person identification using gait vibrations, we set the following four require-
ments for our proposed system:

Req. 1: The system should be able to accurately acquire low-level signal gait vibra-
tions.
Req. 2: The system should be unaffected by internal and natural noise.
Req. 3: The system should be capable of detecting a wide range of gait vibrations.
Req. 4: The system should be able to identify the individual in the fewest footsteps
possible.

Suppose the output impedance of the piezoelectric sensor is higher than the input
impedance of the amplifier circuit. In that case, an output voltage drop occurs, which
makes it difficult to detect vibrations generated at a distance from the sensor, which in
turn means that it is necessary to devise a way to reduce the output impedance of the
piezoelectric sensor. In the circuit shown in Fig. 1, the output impedance is reduced by
introducing a source-follower circuit equipped with a field-effect transistor (FET), thus
fulfilling Req. 1.
 Non-contact Person Identification by Piezoelectric-Based Gait Vibration Sensing 749

Since walking gait vibration characteristics are mainly expressed in terms of their
low-frequency component, to accurately capture those characteristics to the greatest
extent possible, it is vital to design a circuit that is not susceptible to low-frequency
noise. Hence, to fulfill Req. 2, we used an LTspice1 circuit simulator to produce a design
that is unaffected by such noise.
For Req. 3, since it is important to widen the dynamic range to detect a wide range
of gait vibrations, we obtained increased voltage by adopting a dual power supply drive.
Here, the dynamic range refers to the ratio of the maximum and minimum signals. As a
result, in addition to low-level signals, unsaturated high-level signal gait vibrations can
be acquired without saturation, as well as low-level signal, thus fulfilling Req. 4.

4 Proposed System

In this section, we describe our proposed person identification system, which is based
on a piezoelectric sensor, as well as an effective method for using this system.

Fig. 1. Circuit diagram.

4.1 Circuit Structure

The circuit in Fig. 1 consists of a piezoelectric sensor, the FET, and an operational
amplifier. The 7BB-41-2L0 piezoelectric sensor developed by Murata Manufacturing
Co., Ltd. is used in our system. This circuit uses a FET source follower circuit to prevent
the voltage drop caused by a case of the output impedance of the piezoelectric sensor
from being higher than the input impedance of the amplifier circuit. As shown in Fig. 1,
the current amplification ratio was increased by parallel source-follower circuit, which
made it possible to acquire unsaturated low-level gait signal vibrations.
The source-follower circuit only amplifies the voltage current, which is essential to
obtaining accurate low-level signal gait vibration data. To accomplish this, we amplify
1 https://www.analog.com/en/design-center/design-tools-and-calculators/ltspice-simulator.
html.
750 K. Umakoshi et al.

the signal by using a non-inverting amplifier circuit with an amplification gain of 10x.
To minimize the noise effects, circuit elements were selected using the LTspice circuit
simulator.
However, there is still a problem because high-level signal vibrations become sat-
urated in the vicinity of the piezoelectric sensor, which means they cannot be acquired
even if a gain adjustment is applied to acquire the low-level signal vibrations. In this
research, we address this issue by supplying a dual power supply to both the source-
follower and non-inverting amplifier circuits, thereby achieving a much wider dynamic
range than that would be possible with conventional systems.
Finally, since humming noise is a serious problem in gait vibration feature extrac-
tion, we use AA batteries (1.2 V × 4 batteries) to remove that effect.

Fig. 2. System flow.

4.2 Person Identification Method

In this section, we describe the person identification method designed to take advantage
of our piezoelectric sensor-based system. Figure 2 shows the system flow. The objective
of our method is to use ML to identify individuals via the five steps of 1) training data
collection, 2) gait event detection, 3) normalization and denoising, 4) feature extraction,
and 5) person identification model construction. Each step is described below.

4.2.1 Training Data Collection

In Step 1, we collected gait vibrations to be used as training data. As shown in Fig. 2, we
asked participants to walk normally along a section of flooring to obtain gait vibration
data. Since private homes are the target environment of this study, participants were
asked to walk without shoes but while wearing socks. Our system collected data using
a data logger capable of sampling at 10 kHz.

4.2.2 Gait Event Detection

In Step 2, after collecting the training data, we detected single footstep gait events
from the measured vibration waveforms. In this study, we use a change point detec-
tion algorithm that uses background noise variance as a threshold. To detect gait events,
we obtained gait vibration variance values separated by two millisecond intervals and
 Non-contact Person Identification by Piezoelectric-Based Gait Vibration Sensing 751

shifted by one millisecond. The starting point of the gait event is when the variance
value exceeds 10 times the threshold, and the end point of the gait event is when it falls
below 1.5 times the threshold.

4.2.3 Normalization and Denoising

In Step 3, we normalized the gait events and removed the background noise. Since
differences in signal levels due to the distance relationship between the piezoelectric
sensor location and the gait vibration location could cause a decrease in identification
accuracy, we resolved this problem by dividing the gait signal by the sum of the squares
of that signal. This process normalizes the gait event and prevents the deterioration of
identification accuracy. Additionally, since it is possible that gait vibration characteris-
tics could be affected by building vibrations, we use the wavelet transform to remove
such noise.

Table 1. Extracted features.

Domain Feature names

Time domain Event duration, Standard deviation, Entropy,
First five peak values, Maximum peak value,
Maximum peak location,
Five values before the maximum peak,
Five values after the maximum peak
Frequency domain Spectra, Centroid, First five peak values,
First five peak locations, Peak number

4.2.4 Feature Extraction

In Step 4, after normalizing the gait events and removing background noise, we
extracted the features needed to identify individuals from their gait events. Table 1
shows the list of features. This system utilizes 13 features that have been shown to
be highly significant for identifying individuals, as reported in [7].

4.2.5 Person Identification Model Construction

Finally, we used the features described in the previous section to identify individuals.
To construct our person identification model, we used five ML models: support-vector
machine (SVM), Random Forest, XGboost, Light Gradient Boosting Machine (Light-
GBM), and deep neural network (DNN), which are considered to be representative
classification models. In this study, gait data were collected under the condition that
no vibrations other than the walking vibration occurred. For each training model, the
hyper- parameters were adjusted using the grid search method.
752 K. Umakoshi et al.

5 Experiment
In this section, we describe evaluation experiments conducted using our proposed sys-
tem. First, the objective and overview of the experiments are described, followed by the
experimental conditions and evaluation methods.

Fig. 3. Experimental environment.

5.1 Experimental Objective and Overview

To evaluate the effectiveness of our piezoelectric sensor-based person identification sys-
tem, we conducted an evaluation experiment in a smart home environment of the Nara
Institute of Science and Technology (NAIST), shown in Fig. 3. Initially, we asked our
test participants to walk along a specific path and collected his/her gait vibrations by
using our system. Figure 4 shows the measurement devices used in this experiment. As
can be seen in the figure, one piezoelectric sensor is placed on the floor at the mid-
point of the walking path 10 cm away from the wall and 40 cm from the walking path.
A 215 g brass weight is placed on top of that piezoelectric sensor to make it easier to
measure gait vibrations. The sampling rate is 10 kHz, and the data are logged on the lap-
top personal computer (PC) using an Analog Discovery 2 measurement device. Since
this experiment tested a scenario in which a person enters and leaves in a room, we
instructed the participants to enter the room, walk naturally on the path, turn around,
and then exit at their own pace. All of the experiments in this study were conducted with
the approval of the NAIST Ethics Review Committee (Approval Number: 2020-I-16).

5.2 Experimental Conditions

At the initial stage of our experiment, which was conducted with the cooperation of six
participants (five males and one female). Each participant was asked to walk in their
normal manner along the path shown in Fig. 3 while wearing socks. More specifically,
each participant was instructed to walk forward 5 m from the starting point, turn around,
and then return to the original position. In this experiment, we collected data of three
 Non-contact Person Identification by Piezoelectric-Based Gait Vibration Sensing 753

Fig. 4. Measurement devices.

sessions, each of which consisted of 10 consecutive round-trips, from each participant.

To avoid collecting data consecutively from the same participant, we changed partici-
pants or inserted breaks between sessions.

5.3 Evaluation Method

Four evaluation methods were used in these experiments. First of all, we generated an
identification model for each participant using just a single footstep. For each train-
ing data set, we constructed a person identification model using the five ML mod-
els described in Sect. 4.2.5, and then evaluated the resulting identification accuracy by
leave-one-session-out cross-validation. The most accurate person identification model
acquired was then used for the three other evaluation methods. More specifically, to
determine whether individuals could be identified via a small number of footsteps, we
evaluated the individual identification accuracy using the first, first and second, and first
to third footsteps of each back-and-forth walking trip. In this experiment, precision,
recall, and F-measure (indicated by F1) were used to evaluate the proposed system.

5.3.1 Person Identification Using a Single Footstep

For person identification using a single footstep, each ML model returns a confidence
level that is the probability of belonging to each class. The class with the highest confi-
dence value is considered the identification result. In our evaluation, one round-trip con-
sists of approximately 18 footsteps and each participant performs 10 round-trip walks,
which means that one session generates approximately 180 footsteps of data of a person.

5.3.2 Person Identification Using Multiple Footsteps

In cases where the identification result is decided by majority vote for each footstep,
split votes may result and the person may not be correctly identified. For example, if
the first footstep is “label A,” the second footstep is “label B,” and the third footstep is
“label C,” it would be impossible to determine the person by a simple majority vote.
754 K. Umakoshi et al.

Table 2. Person identification results using a single footstep.

SVM Random forest XGBoost LightGBM DNN

Recall Precision F1 Recall Precision F1 Recall Precision F1 Recall Precision F1 Recall Precision F1
A 58.8% 54.3% 56.5% 62.8% 62.4% 62.4% 61.3% 60.5% 60.8% 63.4% 64.9% 64.1% 69.7% 57.2% 62.8%
B 72.8% 67.9% 70.3% 81.6% 60.4% 69.4% 79.9% 69.2% 74.0% 82.0% 69.5% 75.1% 76.3% 72.5% 74.3%
C 75.4% 66.8% 70.8% 65.5% 77.6% 70.3% 69.5% 76.6% 72.1% 72.0% 77.1% 73.8% 66.8% 76.4% 70.8%
D 54.8% 70.1% 61.4% 63.2% 63.8% 63.5% 67.3% 68.6% 67.8% 69.0% 71.5% 70.0% 62.9% 71.0% 66.7%
E 69.3% 72.2% 70.7% 60.2% 82.2% 69.5% 71.0% 78.0% 74.1% 74.1% 80.8% 77.1% 70.1% 76.5% 72.6%
F 59.1% 61.9% 60.4% 61.3% 58.5% 59.9% 63.9% 63.2% 63.5% 65.3% 64.5% 64.8% 67.5% 67.4% 67.4%
Ave. 65.0% 65.5% 65.0% 65.7% 67.4% 65.8% 68.8% 69.3% 68.7% 70.9% 71.3% 70.8% 68.9% 70.1% 67.4%
Acc. 64.8% 65.5% 68.4% 70.5% 68.8%

Accordingly, this study uses the confidence value and sums up this value for each class.
As a result, the class with the highest confidence value becomes the identification result
for all of the footsteps. As mentioned earlier, each session generates approximately 180
footsteps of data per person. From these data, we constructed test data in three differ-
ent ways. Since the first method is for an evaluation using only the first footstep, we
extracted all the first footsteps from each back-and-forth walk of each participant. Con-
sequently, this test data contain data of 20 first footsteps. Similarly, we construct evalu-
ation test data using the first two footsteps, and then the first three footsteps. Hence, for
the first two and first three footsteps, we collect data of 40 and 60 footsteps, respectively.
During all experiments, if some gait events are omitted due to the low signal level, we
consider the following measurement detected by the sensor as the next footstep’s data.

6 Results
In this section, we describe the results of our experiments. First, we show the results
of person identification using only a single footstep. Then, we show the results of the
person identification from each back-and-forth walk in three ways: the first footstep,
the first and second footsteps, and the first to third footsteps.

6.1 Person Identification Using a Single Footstep

Table 2 shows the results of person identification using a single footstep. These results
show that the accuracy of LightGBM is the highest, with an average F-measure of
70.8%. Moreover, the other training models identified the participants with an average
F-measure of 60% or better accuracy. Participant E, who shows the highest accuracy,
has a mean F-measure of 77.1%, and Participant A, who has the lowest accuracy, has a
mean F-measure of 64.1%, thereby indicating that our proposed system is robust and can
identify individuals with relatively high accuracy. However, this result does not guaran-
tee good identification accuracy in cases involving low-level signal gait vibrations, such
as situations where gait vibrations are generated at an excessive distance from the sen-
sor. Therefore, we conducted experiments using LightGBM, which showed the highest
recognition accuracy, for three cases of identification data: the first footstep, the first
and second footsteps, and the first to third footsteps of each back-and-forth walk.
 Non-contact Person Identification by Piezoelectric-Based Gait Vibration Sensing 755

Table 3. Results using multiple footsteps.

Recall Precision F1 Accuracy

First footstep only 62.9% 69.9% 63.1% 63.0%
First and second footsteps 76.2% 78.5% 75.9% 76.1%
First to third footsteps 87.2% 88.1% 87.1% 87.3%

Fig. 5. Confusion matrix (using the first footstep only).

6.2 Person Identification Using Multiple Footsteps

Table 3 shows the person identification results for the first and multiple footsteps from
each back-and-forth walk, while the confusion matrices for these results are shown in
Figs. 5 to 7. The average F-measure for the first footstep detected by the system from
the walking start was 63.1%, which indicates that our system is able to acquire gait
vibrations generated at an expected distance from the sensor. The average F-measure
using the first and second footsteps is 75.9%, thus achieving a 12.8% accuracy improve-
ment compared to the case of using only the first footstep. The reason for this is that the
confidence of each class is higher because of the observation of two footsteps of gait

Fig. 6. Confusion matrix (using the first and Fig. 7. Confusion matrix (using the first to
second footsteps). third footsteps).
756 K. Umakoshi et al.

vibration. Furthermore, the mean F-measure of the person identification results when
using the first to third footsteps of the walking test was 87.1%, which was 24% more
accurate than using only the first footstep, and 11.2% more accurate than using the first
and second footsteps. These results show that the proposed system can identify individ-
uals with an average F-measure of more than 80% by observing gait vibrations from
their first three footsteps.

7 Conclusion
In this paper, we proposed a new system that efficiently detects low-level signal gait
vibrations with a single piezoelectric sensor and a ML-based method that can identify
an individual with high accuracy from a small number of footsteps. To show the effec-
tiveness of our proposed system, we conducted a person identification experiment in
a smart home at the NAIST. As a result, we obtained two findings. First, the obtained
accuracy using five different ML models showed that the average F-measure for per-
son identification using only a single footstep was up to 70.8% when using LightGBM.
Second, in order to investigate whether individuals could be identified with just a few
footsteps, we evaluated the accuracy of individual identification in three ways of using
the first footstep, the first and second footsteps, and the first to third step footsteps, from
each back-and-forth walk. The mean F-measures of those cases were 63.1%, 75.9%, and
87.1%, respectively. These results showed that the proposed system achieves a recog-
nition accuracy comparable to existing systems that can identify an individual with an
average of 83% with five footsteps [13] and an average accuracy of 96% with seven
footsteps [6]. In the future, we plan to improve the circuit to acquire gait vibrations
at lower signal levels. In addition, we plan to select additional features to improve the
system accuracy and to make it possible to more accurately identify individuals with
fewer footsteps.

Acknowledgement. This study was supported in part by the Japan Society for the Promotion of
Science, Grants-in-Aid for Scientific Research number JP20H04177.

References
1. Sun, Y., Chen, Y., Wang X., Xiaoou, T.: Deep learning face representation by joint
identification-verification. Adv. Neural Inf. Process. Syst. 1988–1996 (2014)
2. Hrechak, A.K., McHugh, J.A.: Automated fingerprint recognition using structural matching.
Pattern Recogn. 23(8), 893–904 (1990)
3. Jiewen, Z., Ruize, H., Yiyang, G., Liang, W., Wei, F., Song, W.: Human identification and
interaction detection in cross-view multi-person videos with wearable cameras. In: Proceed-
ings of the 28th ACM International Conference on Multimedia, pp. 2608–2616 (2020)
4. Korany, B., Karanam, C.R., Cai, H., Mostofi, Y.: Xmodal-id: Using wifi for through-wall
person identification from candidate video footage. In: The 25th Annual International Con-
ference on Mobile Computing and Networking, pp. 1–15 (2019)
5. Higashino, T., Uchiyama, A., Saruwatari, S., Yamaguchi, H., Watanabe, T.: Context recogni-
tion of humans and objects by distributed zero-energy IoT devices. In: 2019 IEEE 39th Inter-
national Conference on Distributed Computing Systems (ICDCS), pp. 1787–1796. IEEE
(2019)
 Non-contact Person Identification by Piezoelectric-Based Gait Vibration Sensing 757

6. Pan, S., Yu, T., Mirshekari, M., Fagert, J., Bonde, A., Mengshoel, O.J., Noh, H.Y., Zhang,
P.: Footprintid: indoor pedestrian identification through ambient structural vibration sensing.
Proc. ACM Interact. Mob. Wear. Ubiquitous Technol. 1(3), 1–31 (2017)
7. Clemente, J., Song, W.Z., Valero, M., Li, F., Liy, X.: Indoor person identification and fall
detection through non-intrusive floor seismic sensing. In: 2019 IEEE International Confer-
ence on Smart Computing (SMARTCOMP), pp. 417–424. IEEE (2019)
8. Akiyama, S., Yoshida, M., Moriyama, Y., Suwa, H., Yasumoto, K.: Estimation of walk-
ing direction with vibration sensor based on piezoelectric device. In: 2020 IEEE Inter-
national Conference on Pervasive Computing and Communications Workshops (PerCom
Workshops), pp. 1–6. IEEE (2020)
9. Zeng, Y., Pathak, P., Mohapatra, P.: Wiwho: wifi-based person identification in smart spaces.
In: 2016 15th ACM/IEEE International Conference on Information Processing in Sensor
Networks (IPSN), pp. 1–12 (2016)
10. Wan, C., Wang, L., Phoha, V.V.: A survey on gait recognition. ACM Comput. Surv. (CSUR)
51(5), 1–35 (2018)
11. Zhang, Y., Pan, S., Fagert, J., Mirshekari, M., Noh, H.Y., Zhang, P., Zhang, L.: Occupant
activity level estimation using floor vibration. In: Proceedings of the 2018 ACM Interna-
tional Joint Conference and 2018 International Symposium on Pervasive and Ubiquitous
Computing and Wearable Computers, pp. 1355–1363 (2018)
12. Mirshekari, M., Fagert, J., Bonde, A., Zhang, P., Noh, H.Y.: Human gait monitoring using
footstep-induced floor vibrations across different structures. In: Proceedings of the 2018
ACM International Joint Conference and 2018 International Symposium on Pervasive and
Ubiquitous Computing and Wearable Computers, pp. 1382–1391 (2018)
13. Pan, S., Wang, N., Qian, Y., Velibeyoglu, I., Noh, H.Y., Zhang, P.: Indoor person identifica-
tion through footstep induced structural vibration. In: Proceedings of the 16th International
Workshop on Mobile Computing Systems and Applications, pp. 81–86 (2015)
14. Kashimoto, Y., Fujimoto, M., Suwa, H., Arakawa, Y., Yasumoto, K.: Floor vibration type
estimation with piezo sensor toward indoor positioning system. In: 2016 International Con-
ference on Indoor Positioning and Indoor Navigation (IPIN), pp. 1–6. IEEE (2016)
 Design of Healthcare Information Sharing
Methods Using Range-Based Information
Disclosure Incentives

Miku Honda1(B) , Joji Toshima2 , Takuo Suganuma3 , and Akiko Takahashi1,2

1 National Institute of Technology, Sendai College, 4-16-1 Ayashi-Chuo,
Aoba-ku, Sendai 989-3128, Japan
[email protected], [email protected]
2 Graduate School of Information Sciences, Tohoku University, 6-3-09 Aoba, Aramaki-aza
Aoba-ku, Sendai 980-8579, Japan
[email protected]
3 Cyberscience Center, Tohoku University, 6-3 Aoba, Aramaki-aza Aoba-ku,
Sendai 980-8578, Japan
[email protected]

Abstract. The means by which healthcare information is collected and utilized

are progressing. Collecting and analyzing such information on a large scale
can result in the extraction of useful information and further service develop-
ment. However, the provision of healthcare information has privacy considera-
tions because healthcare information contains personal information. Therefore, in
this paper, we propose a healthcare information sharing method that promotes the
sharing of useful healthcare information with ample privacy safeguards using auto-
negotiation considerations and user/system intentions. In addition, we evaluate the
effectiveness of the proposed method in simulation experiments.

1 Introduction
With the widespread use of wearable devices that continuously collect sensor data,
the collection and utilization of user health data (e.g., sleep time and walking time) is
increasing [1, 2]. However, although collecting and analyzing such information on a
large scale may yield useful information, such information is generally managed and
utilized by companies that provide devices. Here, the services the user can access with a
device (information provider) while providing information are limited to those provided
by the company. Thus, there are methods to appropriately collect information based on
the consent of the information provider [3]. These methods are based on the notion that
information owned by an individual belongs to that individual, and they should have the
right to use it freely.
The goal is to widely use and distribute the information to business operators (infor-
mation collectors), e.g., companies, and not limit the spread of information to compa-
nies that provide the devices. In such an information sharing context, public healthcare
information provision imposes a burden on users due to the personal nature of healthcare
information. As a result, users may be reluctant to provide healthcare information, which

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 758–769, 2021.
https://doi.org/10.1007/978-3-030-75100-5_64
 Design of Healthcare Information Sharing Methods 759

may limit the amount of data gathered for analysis. Thus, it is difficult to collect large
volumes of high-quality healthcare information. Therefore, a new framework is needed
to promote the use of healthcare information among a wide range of information collec-
tors. The new framework will prioritize the collection of large amounts of high-quality
healthcare information while considering privacy. To provide healthcare information on
an information provider’s own initiative, it is necessary to have a motivation to consider
the risks incurred by the information provider and assume that the information provider
may provide the information.
For this problem, we proposed a method to promote information provision by giv-
ing incentives while considering the protection of information. Using auto-negotiation
methods, we proposed a way for determining the required level of incentives and infor-
mation protection level. It is necessary to gain user buy-in while considering the sensitive
nature of healthcare information, objective information provider/information collector
evaluations, and a combination of the two.
We propose a healthcare information sharing method that promotes the provision
of useful healthcare information with appropriate privacy safeguards without burdening
information providers and collectors. In addition, we apply the proposed healthcare
information sharing method to a healthcare information sharing system and evaluate its
effectiveness through simulation experiments that assume different types of information
providers.

2 Related Work and Proposed Method

2.1 Related Work

As the populations in Japan, Northern Europe, and the United States of America continue
to rapidly age, the medical care and welfare burden continues to increase. As a result,
it is becoming increasingly difficult to respond to conventional medical and insurance
systems [4]. There are some methods to sharing information that promote adherence
to a healthy diet and individualized physical activity [5]. Sharing healthcare informa-
tion is effective relative to improving both health and lifestyle and sharing healthcare
information is required to create personalized healthcare services [6].
To promote the provision of healthcare information, it is necessary to have a motiva-
tion to provide healthcare information. The conditions under which personal information
may be provided are described as follows. The information provider receives economic
benefits, and the provision of information improves the convenience of services. Incen-
tives are an effective way to influence human behavior [7], and money is the most
effective incentive, but it can be replaced with a voucher. It is effective to give a small
incentive to promote changes in individual behavior; however, given privacy considera-
tions, information providers are often reluctant to make their own personal information
available.
Healthcare information contains personal information, and the handling risk of such
information is high. Therefore, it is important to utilize such information with care.
760 M. Honda et al.

However, the utilization of healthcare information has not progressed [8]. Regarding
utilization of personal information on the Internet, e.g., browsing history and location
information, various privacy infringement risk indicators and personal information pro-
tection methods have been proposed. However, it is possible that certain individual
privacies will be abused as a result of access to such information because it is impossible
to completely quantify privacy risk using the existing method [9]. Thus, it is important
to develop a sense of satisfaction from the individual to promote information provision.
One effective way to increase survey cooperation rates is through financial incentives
[10, 11]; however, the extent to which incentive requirements vary is related to each
individual’s degree of privacy awareness due to considering the psychological burden
of the information provider [12]. In addition, personal information may be abused, and
disclosure of information that users wish to keep private is becoming problematic.
Therefore, it is becoming increasingly important to control the information disclosed
by the information providers [13]. Such information providers must disclose personal
information by adjusting the types of information provided based on data collectors’
intentions. However, it is difficult for general users to determine appropriate personal
information to provide.

2.2 Technical Problems and the Proposal

As discussed in Sect. 2.1, methods for sharing personal health information are available;
however, providing healthcare information is a concern for users owing to privacy prob-
lems. To address this issue, some methods have previously been proposed to promote
information provision through incentives. However, the amount of incentive required to
promote healthcare information provision differs for each diffident individual. In addi-
tion, it is difficult for general users to determine appropriate healthcare information to
provide. Thus, it is a burden for both the information provider and information collector
to negotiate a benefit for both parties while considering such factors. There are techni-
cal problems involved in promoting healthcare information sharing through incentives,
particularly when considering the information protection and usage intent.

(P1) It is difficult to provide users sufficient incentives to motivate them to provide

healthcare information.

Different people require different amounts of incentives to provide such sensitive

information. In addition, it is difficult to determine the requisite level of incentive because
user-provided healthcare information is used in different ways.

(P2) It is difficult to reasonably protect sensitive user-provided information.

This difficulty exists because sharing sensitive information is mutually beneficial,

but it is possible that personal information may be abused, which could lead to leakage
of information the user wishes to keep private.
 Design of Healthcare Information Sharing Methods 761

Therefore, it is necessary to determine the information to be disclosed based on the

user’s unique intentions. Therefore, this study proposes the following to solve the above
technical issues.

(S1) Healthcare information should be shared using an auto-negotiation method

between software agents.

We propose the promotion of useful healthcare information sharing via consensus

building using auto-negotiation to consider the intentions of both information providers
and information collectors. Given that agents reflect intentions between information
providers and collectors, the level of incentives and information protection level can be
determined by automated negotiation between agents. Thus, this method is expected to
encourage users to provide confidential healthcare information.

3 Proposed Healthcare Information Sharing Method

3.1 Immediate Evaluation of Healthcare Information

Fig. 1. Example of healthcare information

The information-provider agent obtains healthcare information data that can be acquired
on a daily basis from smart devices, including wearable devices, equipped with a wide
variety of sensors. Healthcare information is sent to information-collector agents owned
by people seeking data such as companies, developers, and organizations using network
technology. The information-collector agent immediately evaluates the user-provided
healthcare information according to several factors, e.g., quality and sensitivity. In this
study, among smart devices, wearable devices are treated as testbeds. The evaluation
point is data that can be obtained from a wearable device worn by an information
provider. In this evaluation, only numerical data, e.g., blood pressure and sleep time, are
considered because it is easy to uniquely determine the evaluation compared with verbal
data. Healthcare information is evaluated quantitatively using the Mahalanobis distance
762 M. Honda et al.

measure [14]. The Mahalanobis distance is calculated using a group of high-quality

healthcare information established by an information-collector agent, which relies on
immediate evaluation of the healthcare data. We consider that the preference information
of healthcare information differs depending on the given information collector. There-
fore, good healthcare information should be defined individually by each information
collector. Here, we categorize healthcare information into seven classes of numerical
data (Fig. 1) that can be acquired from wearable devices.
Considering the above, the Mahalanobis distance used for immediate evaluation is
defined as follows:

−1
d = (x − μ)
 (x − μ)
 (1)

where d denotes the Mahalanobis distance, x denotes the data provided by information
provider, and μ denotes the mean vector calculated from the set of high-quality healthcare
information established by the information-collector agent.
The closer the data are to a set of high-quality healthcare information, the smaller
the Mahalanobis distance. Thus, it is possible to perform high-level evaluation of high-
quality healthcare information. Here, the Mahalanobis distance is normalized to [0, 1]
and used to immediately evaluate the healthcare information.

3.2 Utility Functions

Auto-negotiation is performed between the information-collector agent and information-
provider agent. Here, the information-provider agent reflects the will of the information
provider, and the information-collector agent reflects the will of the information col-
lector. The level of incentives and amount of information disclosed is determined via
auto-negotiation between these agents. This auto-negotiation is performed based on
the “alternating offers” negotiation protocol [15], which is frequently used in bilateral
negotiations. In the alternating offers negotiation protocol, each agent selects one of
the three actions (Accept, Offer, or End Negotiation). Each agent will select an “Offer”
until “Accept” or “End negotiation” is selected. This behavior alternates between the
information-collector agent and information-provider agent. The negotiation algorithm
is shown in Algorithm 1. When selecting an “offer, each agent offers a proposal in the
range of [0, 1], with their utility function considered as a probability distribution.

(a) Accept.
The agent accepts the bid presented by the other agent. The bid’s incentive level
and amount of information disclosed are determined as a result of negotiation, and
the negotiation is completed.
 Design of Healthcare Information Sharing Methods 763

return End Negotiation .

(b) Offer.
The agent rejects the bid presented by the other agent. The agent presents a new
bid to the other agent.

(c) End Negotiation.

The agent rejects the bid presented by the other agent. The agent does not present
a new bid and ends the negotiation.

Each agent determines its conduct according to each utility function. The utility
function quantifies the degree of satisfaction, i.e., the utility, a person feels as a result
of hypothetical profit or loss. Therefore, it is necessary to properly express how utilities
differ based on the intentions of the individuals who experience the profits and losses.
The utility value is defined in the range [0, 1]. Here, if the utility is 0, the person is
very dissatisfied with the profit or loss, and, if the utility is 1, the person is completely
satisfied with the profit or loss.
Considering the above, the utility function is defined as follows:
2
U = ui × wi (2)
i=1
where U denotes the utility function, u1 denotes the utility of information protection,
w1 denotes the intentional weight of information protection, u2 denotes the utility of the
incentive, and w2 denotes the intentional weight of the incentive.
Here, intentional weights w1 and w2 are given in the range [0, 1]. When summed, w1
and w2 should equal 1. Parameters w1 and w2 express the difference between the user’s
and system’s intentions relative to incentives and information protection.
764 M. Honda et al.

Utility ui is expressed by a different equation between the system agents and user
agents. Here, utility uu1 , uu2 is a value that expresses the degree of user satisfaction, and
the value of utility us1 , us2 expresses the system’s degree of satisfaction. uu1 , uu2, and
us1 ,us2 are expressed as follows:
 r
Uui = x − Ip + 1 (i = 1, 2), (3)

Usi = Ie − (x(1 − Ie ))r (i = 1, 2), (4)

where Ip denotes the parameter of the information provider’s sensitive evaluation of

healthcare information. Ip is defined in the parameter [0, 1]. Ie denotes the immediate
evaluation of the healthcare information (Sect. 3.1). In addition, parameter r denotes the
degree of risk preference.
The utility function is outlined in Fig. 2(a) and Fig. 2(b). Note that the information
provider’s utility function is a monotonically increasing function, and the information
collector’s utility function is a monotonically decreasing type function. Here, utility
function U takes the form of a risk-averse function if the risk preference degree r
becomes greater than 1 and becomes a risk-loving function if r is less than 1.

Fig. 2. (a) Utility function of information collector (b) Utility function of information provider

3.3 Design of Healthcare Information Sharing Model

This model comprises an information-provider side agent that operates on behalf of
the information provider, an information-providing agent (PA), an information-collector
agent that operates on behalf of the information collector, an information-collector agent
(CA), and a healthcare information sharing system (HIS) that manages the collected
healthcare information. The model is shown in Fig. 3. The processing flow of this model
is summarized as follows. Note that each step outlines the equivalent process shown in
Fig. 3.
 Design of Healthcare Information Sharing Methods 765

Fig. 3. Healthcare information-sharing model

Step 1. The PA obtains the information provider’s healthcare information.

Step 2. The PA obtains the information provider’s sensitive evaluation of the healthcare
information.
Step 3. The CA evaluates the healthcare information based on the quality and sensitivity
of the information.
Step 4. The PA selects the CA with the highest evaluation of the value of the information
and performs an automated negotiation. The agent determines the information to be
disclosed and the level of incentive.
Step 5. The CA gives the PA the incentive established based on the valuation (Step 4).
Step 6. The CA sends the anonymized healthcare information to the HIS.

Table 1. Types of information providers

This method promotes information sharing by providing an incentive to the user

according to the quality of the shared information in consideration of the extent to which
the information is protected.
766 M. Honda et al.

Table 2. Types of methods

4 Experiment and Discussion

4.1 Evaluation Experiment of Healthcare Information Sharing Method

To confirm whether the proposed method can collect high-quality healthcare information
while reducing incentives and the degree of information protection, we performed a
simulation experiment to compare the proposed method to two simple methods.
Here, the four types of information providers considered in this experiment are
detailed in Table 1. These information providers assuming 500 people each, who have
different qualities and sensitive evaluation of healthcare information. In this experiment,
the utility function parameters (Sect. 3.2) were set to w1 = 0.5, w2 = 0.5, and r =
2. The behaviors of the information providers and information collectors followed the
processing flow presented in Sect. 3.3.
Here, each information provider provided healthcare information every day for seven
days. We assumed three information collectors. The information-collector agent gen-
erated and evaluated random numbers from the range of the immediate evaluation of
healthcare information (shown in Table 1). The information-provider agent selected
the information-collector agent who highly evaluated the healthcare information, and
they automatically negotiated with the collector who highly evaluated the healthcare
information. In the auto-negotiation, the upper limit of the negotiation was 30 times,
and, if the upper limit was exceeded, the utility, amount of incentive obtained, and
degree of information protection was 0. If the auto-negotiation succeeded, the informa-
tion provider provided healthcare information using the utility calculated by the utility
function as a probability. The experimental results were taken as the average results of
auto-negotiation between an information provider’s agent and the system’s agent relative
to the requisite amount of incentive, the information protection levels, the quality of the
collected healthcare information, the number of collected healthcare information, and
the utility. The amount of incentive, information protection levels, utility, and quality of
the collected healthcare information were calculated as the average value of the experi-
mental results for the three types of information providers. The proposed method yields
excellent results for various types of information providers. Therefore, we confirm that
the proposed method is excellent for any type of information provider. As a result, it
is desirable to reduce incentives and the information protection level and increase the
utility, the quality of the collected information, and the number of collected healthcare
information. We confirmed the effectiveness of the proposed method by comparing the
results of these experiments to two simple methods (Table 2). The experimental results
 Design of Healthcare Information Sharing Methods 767

are shown in Fig. 4. Regarding incentive, the proposed method scored −62% com-
pared to the require simple method and +47% compared to the quality simple method.
Here, regarding the degree of information protection, the same tendency for the amount
of incentive was observed. For the information protection level, the proposed method
scored −63% compared to the require simple method and +56% compared to the quality
simple method. For utility, the proposed method scored −32% compared to the require
simple method and +24% compared to the quality simple method. For the amount of
healthcare information, the proposed method scored −15% compared to the require
simple method and +19% compared to the quality simple method. In addition, for infor-
mation quality, the proposed method scored +9% compared to the require simple method
and +9% compared to the quality simple method.

Fig. 4. Experimental results

4.2 Discussion
The experimental results demonstrate that the proposed method reduced the amount
of incentive and information protection compared to the require simple method but
increased these factors compared to the quality simple method. In the quality simple
method, the amount of incentive and information protection level presented by the infor-
mation collector is the experimental results. Information collectors seek to reduce incen-
tives and the information protection level as much as possible; therefore, we considered
that the incentive and information protection level was the lowest compared to any other
method. We found that the amount of incentives and degree of information protection
level were minimized. In addition, utility was the lowest among all methods, and the
amount of healthcare information provided was also the smallest because the quality
simple method only prioritizes the intent of the information collector. Thus, the quality
simple method does not provide healthcare information as the utility felt by the infor-
mation provider decreases. Since the require simple method accepts all information
provider requests, the utility felt was high, and the probability of providing health care
information increased. Therefore, the require simple method can collect the most health
768 M. Honda et al.

care information. However, the amount of incentive and information protection level was
the highest among all methods, and the sustainability of services is considered in the
follow. The proposed method collected higher quality health care information compared
to the simple methods because healthcare information was collected by considering
the intentions of both the information provider and information collector through auto-
negotiations. Therefore, the information quality obtained by the proposed method was
greater than that of the simple methods; thus, it is possible to collect a large amount of
useful healthcare information using the proposed method.
Thus, we consider that the experimental results confirmed the effectiveness of the
proposed healthcare information collection method.

5 Conclusion
In this paper, we have proposed a method to promote information provision by giving
incentives while considering the protection of healthcare information. We proposed pro-
moting the sharing of useful healthcare information by consensus building using the
auto-negotiation considering the intentions of information provider/information collec-
tor. In the proposed method, agents reflect the intentions of information providers and
information collectors, and the amount of incentive and items to provide information
are determined automatically via auto-negotiation between the agents. In addition, to
evaluate the effectiveness of the proposed healthcare information sharing method, we
performed simulation experiments. The experimental results confirmed that the pro-
posed method outperforms the simple method and that it enables dynamic evaluation of
healthcare information.
In future, we plan to evaluate the effectiveness of the proposed method by considering
various other conditions, e.g., the probability of a user providing healthcare informa-
tion and continuation rate. We also plan to conduct an experiment in which healthcare
information is acquired automatically from smart devices, e.g., wearable devices.

Acknowledgments. This work was supported by JSPS KAKENHI Grant Number 19K04895.

References
1. Itao, K., Komazawa, M., Katada, Y., Itao, K., Kobayashi, H., Luo, Z.W.: Age-related change
of the activity of autonomic nervous system measured by wearable heart rate sensor for long
period of time. In: Cipresso, P., Matic, A., Lopez, G. (eds.): Pervasive Computing Paradigms
for Mental Health, Springer International Publishing, pp. 33–38 (2014)
2. Swan, M.: Sensor mania! the internet of things, wearable computing, objective metrics, and
the quantified self 2.0. J. Sens. Actuator Netw. 1(3) 217–253 (2012)
3. Mitchell, A.: Right Side Up: Building Brands in the Age of the Organized Consumer. Harper
Collins Business, US (2001)
4. Dall, T.M., et al.: An aging population and growing disease burden will require alarge and
specialized health care workforce by 2025. Health Aff. 32, 2013–2020 (2013)
5. DeVahl, J., King, R., Williamson, J.W.: Academic incentives for students can increase par-
ticipation in and effectiveness of a physical activity program. J. Am. Coll. Health, 53(6),
295–298 (2005)
 Design of Healthcare Information Sharing Methods 769

6. Pickard, K., Swan, M.: Big desire to share big health data: a shift in consumer attitudes toward
personal health information (2014)
7. Farooqui, M.A., Tan, Y.T., Bilger, M., Finkelstein, E.A.: Effects of financial incentives on
motivating physical activity among older adults: results from a discrete choice experiment.
BMC Public Health 14(1), 1–9 (2014)
8. Perera, G., Holbrook, A., Thabane, L., Foster, G., Willison, D.J.: Views on health information
sharing and privacy from primary care practices using electronic medical records. Int. J. Med.
Inform. 80(2), 94–101 (2011)
9. Machanavajjhala, A., et al.: -diversity: privacy beyond k-anonymity. ACM Trans. Knowl.
Discov. Data (TKDD) 1(1) (2007)
10. Hann, I.H., Hui, K.L., Lee, T.S., Png, I.P.: The value of online information privacy: an
empirical investigation. Ind. Organ. 0304001, University Library of Munich, Germany (2003)
11. Danezis, G., Lewis, S., Anderson, R.: How much is location privacy worth? In: Proceedings
of WEIS (2005)
12. Sasank, R., Deborah, E., Mark, H., Mani, S.: Examining micro-payments for participa-
tory sensing data collections. In Proceedings of the 12th ACM International Conference
on Ubiquitous Computing (2010)
13. Westin, A.F.: Privacy and freedom. Wash. Lee Law Rev. 25(1),166 (1968)
14. Qunigob, M.: An introduction to mahalanobis distance for MTS methods. Qual. Eng. 9(1),
13–21 (2001)
15. Rubinstein, A.: Perfect equilibrium in a bargaining model. Econometrica, 97–109 (1982)
Author Index

A Boukhalfa, Fouzi, 25
Abassi, Ryma, 113 Bouras, Christos, 13
Abdellatif, Slim, 542 Bourgeois, Julien, 332
Abou Jaoude, Chady, 39 Bylykbashi, Kevin, 63
Adhinugraha, Kiki, 272
Agha, Khaldoun Al, 195 C
Ahmed, Safuriyawu, 626 Camilotti, Luca, 713
Ahmed, Tarem, 519 Cao, Jiannong, 102
Aleksy, Markus, 414 Chaari, Lamia, 483
Al-khatib, Abdullah A., 681 Chakroun, Raoua, 542
Almansor, Ebtesam Hussain, 163 Choi, Hyuckjin, 745
Alvi, Anik, 519 Costa, Rafael Lima, 638
Aly, Alhassan, 1 Cruz, Jason Paul, 600
Cuzzocrea, Alfredo, 713
Ampririt, Phudit, 63, 82
Arai, Yoshikazu, 403
Arfan, Humayra Binte, 472 D
Ayed, Samiha, 483 da Silva, Arley W. N., 383
da Silva, Maria V. L., 383
Azmy, Andrew, 566
de Almeida, Edson L. V., 383
Azuma, Masaya, 92
de Azevedo, Lucas P., 383
de Castro, Ariel G., 181
Dhurandher, Sanjay Kumar, 553
B Diallo, El-hacen, 195
Babbar-Sebens, Meghna, 208 DIB, Omar, 195
Balfaqih, Mohammed, 681 Duan, Chen, 307
Barbosa, José R., 383 Duolikun, Dilawaer, 172
Barolli, Admir, 82 Durresi, Arjan, 208
Barolli, Leonard, 63, 73, 82, 92 Durresi, Mimoza, 208
Ben Halima, Mohamed, 285 Dutta, Rathindra Nath, 667
Bennani, Samir, 149
Bouallegue, Ridha, 531 E
Bouazizi, Samar, 494 El Fatmi, Sihem Guemara, 113
Boujelbene, Rania, 701 Elelimy, Esraa, 566

© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2021
L. Barolli et al. (Eds.): AINA 2021, LNNS 225, pp. 771–773, 2021.
https://doi.org/10.1007/978-3-030-75100-5
772 Author Index

Elias Mallouhy, Roxane, 39 L

Enokido, Tomoya, 125, 172 Lasebae, Aboubaker, 653
Le Mouël, Frédéric, 626
F Lee, Ming-Chang, 136
Ferworn, Alex, 220 Li, Bo, 51
Formica, Nicholas, 588 Li, Li-feng, 51
Fujimoto, Manato, 731, 745 Lin, Jia-Chun, 136
Lipeme Kouyi, Gislain, 626
G Lorenzon, Arthur F., 181
Geffard, Estelle, 507 Ltifi, Hela, 494
Ghosh, Sasthi C., 667 Luizelli, Marcelo C., 181
Gogolev, Alexander, 414 Lv, Gaofeng, 320
Gourraud, Pierre-Antoine, 507
Gran, Ernst Gunnar, 136 M
Guyeux, Christophe, 39 Makhoul, Abdallah, 39, 332
Maruyama, Kiyoyasu, 731
H Matsui, Tomokazu, 731, 745
Haddad, Rim, 531 Matsuo, Keita, 63
Hadded, Mohamed, 25 Mazumder, Rashed, 472
Hadj Sassi, Mohamed Saifeddine, 285 Mega, Toshihiro, 424
Hakura, Jun, 403 Misaki, Shinya, 731
Hayashibara, Naohiro, 436 Mitton, Nathalie, 448
Hbaieb, Amal, 483 Miyamoto, Tatsuji, 393
Hirose, Tenma, 692 Mizera–Pietraszko, Jolanta, 297
Honda, Miku, 758 Mizumoto, Teruhiro, 731
Hooper, Katrina, 220 Monteiro, Maxwell E., 259
Hussain, Farookh Khadeer, 163 Mori, Genichi, 393
Hussain, Fatima, 220 Mostarda, Leonardo, 588
Moussa, Mohamad, 332
I Muhlethaler, Paul, 25
Ikeda, Makoto, 63, 92 Mumolo, Enzo, 713
Imamura, Mitsuyoshi, 233, 245 Murakami, Kodai, 424
Ishida, Tomoyuki, 692
Iwanow, Maksymilian, 345 N
Nakamura, Shigenari, 125
J Nath, Amar, 370
Jemaa, Yousra Ben, 701 Navarra, Alfredo, 588
Jlassi, Wadii, 531 Niyogi, Rajdeep, 370
Jóźwiak, Alicja M., 297
Jóźwiak, Piotr P., 297 O
Juan, Angel A., 1 Okada, Masayuki, 600
Okamoto, Shusuke, 73
K Okamura, Shingo, 600
Kalogeropoulos, Rafail, 13 Oliveira, Luciana P., 383
Kanchi, Saroja, 577 Omote, Kazumasa, 233, 245
Kataoka, Kotaro, 355 Onishi, Kosei, 731
Kaur, Davinder, 208 Osborn, Wendy, 460
Kchaou, Amira, 113 Otani, Tomohiro, 393
K˛edziora, Michał, 297 Ouali, Imene, 285
Khalid, Khuram, 553
Khalidi Idrissi, Mohammed, 149 P
Khelil, Abdelmajid, 681 Peng, Wei, 307, 320
Kimura, Aki, 731 Petridis, Miltos, 653
Kushiro, Noriyuki, 424 Piranda, Benoit, 332
Author Index 773

Q Takemura, Tatsuya, 600

Qafzezi, Ermioni, 63 Takizawa, Makoto, 82, 125, 172
Taluckder, Saria Jahin, 472
R Taniar, David, 272
Rahayu, Wenny, 272 Tonni, Zinia Anzum, 472
Raheem, Ali, 653 Toshima, Joji, 758
Raheem, Rand, 653 Touati, Haifa, 25
Raynal, Michel, 102
Rebei, Abir, 25 U
Retbi, Asmaâ, 149 Uchimura, Shota, 92
Rivera, Samuel J., 208 Uddin, Mohammad Faisal, 519
Rocha, Antonio A. de A., 614 Umakoshi, Keisuke, 745
Rossi, Fábio D., 181 Umeda, Naoki, 600
Uslu, Suleyman, 208
S
Saito, Yoshiya, 403 V
Sakamoto, Shinji, 73, 82 Viana, Aline Carneiro, 638
Sakuraba, Akira, 403 Viegas, Pablo B., 181
Salim, Christian, 448
Villemur, Thierry, 542
Sampaio, Leobino Nascimento, 638
Vince, Nicolas, 507
Santos, José Humberto da S. S., 383
Sasaki, Wataru, 731
W
Sayadi, Sirine, 507
Wali, Ali, 285
Shibata, Yoshitaka, 403
Wang, Baosheng, 307
Shobha, G., 355
Wang, Pengcheng, 320
Shubair, Raed, 531
Woungang, Isaac, 553
Singh, Jagdeep, 553
Wróbel, Tobiasz, 297
Sneha, M., 355
Souabi, Sonia, 149 Wrzuszczak-Noga, Jolanta, 345
Sousa, Peron, 614
Stouls, Nicolas, 626 X
Südholt, Mario, 507 Xhafa, Fatos, 1
Suganuma, Takuo, 758
Sugihara, Koichiro, 436 Y
Suwa, Hirohiko, 731, 745 Yanai, Naoto, 600
Suzuki, Yusuke, 393 Yasumoto, Keiichi, 731, 745
Szczepanik, Michał, 297 Yoshida, Makoto, 745

T Z
Tada, Yoshiki, 92 Zhang, Wenxin, 320
Takahashi, Akiko, 758 Ziviani, Artur, 638