Machine Learning

Quantum
Computing
Smart
Manufacturing
History

DECEMBER 2024 www.computer.org

 PUBLISH WITH THE
IEEE COMPUTER SOCIETY

Break
Free.
You Have
Choices.
It’s Author’s Choice:
IEEE Computer Society provides
all publishing models; open-
access, hybrid, and traditional
options to accommodate the
unique needs of all researchers.

www.computer.org/cfp
 IEEE COMPUTER SOCIETY computer.org

STAFF
Editor Periodicals Portfolio Senior Managers Director, Periodicals and Special Projects
Lucy Holden Carrie Clark and Kimberly Sperka Robin Baldwin

Production & Design Artist Periodicals Operations Project Specialists Senior Advertising Coordinator
Carmen Flores-Garvey Priscilla An and Christine Shaughnessy Debbie Sims

Circulation: ComputingEdge (ISSN 2469-7087) is published monthly by the IEEE Computer Society. IEEE Headquarters, Three Park Avenue, 17th
Floor, New York, NY 10016-5997; IEEE Computer Society Publications Office, 10662 Los Vaqueros Circle, Los Alamitos, CA 90720; voice +1 714 821 8380;
fax +1 714 821 4010; IEEE Computer Society Headquarters, 2001 L Street NW, Suite 700, Washington, DC 20036.
Postmaster: Send address changes to ComputingEdge-IEEE Membership Processing Dept., 445 Hoes Lane, Piscataway, NJ 08855. Periodicals Postage
Paid at New York, New York, and at additional mailing offices. Printed in USA.
Editorial: Unless otherwise stated, bylined articles, as well as product and service descriptions, reflect the author’s or firm’s opinion. Inclusion in
ComputingEdge does not necessarily constitute endorsement by the IEEE or the Computer Society. All submissions are subject to editing for style,
clarity, and space.
Reuse Rights and Reprint Permissions: Educational or personal use of this material is permitted without fee, provided such use: 1) is not made for
profit; 2) includes this notice and a full citation to the original work on the first page of the copy; and 3) does not imply IEEE endorsement of any third-
party products or services. Authors and their companies are permitted to post the accepted version of IEEE-copyrighted material on their own Web
servers without permission, provided that the IEEE copyright notice and a full citation to the original work appear on the first screen of the posted copy.
An accepted manuscript is a version which has been revised by the author to incorporate review suggestions, but not the published version with copy-
editing, proofreading, and formatting added by IEEE. For more information, please go to: http://www.ieee.org/publications_standards/publications
/rights/paperversionpolicy.html. Permission to reprint/republish this material for commercial, advertising, or promotional purposes or for creating new
collective works for resale or redistribution must be obtained from IEEE by writing to the IEEE Intellectual Property Rights Office, 445 Hoes Lane,
Piscataway, NJ 08854-4141 or [email protected]. Copyright © 2024 IEEE. All rights reserved.
Abstracting and Library Use: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy for private use of patrons,
provided the per-copy fee indicated in the code at the bottom of the first page is paid through the Copyright Clearance Center, 222 Rosewood Drive,
Danvers, MA 01923.
Unsubscribe: If you no longer wish to receive this ComputingEdge mailing, please email IEEE Computer Society Customer Service at help@
computer.org and type “unsubscribe ComputingEdge” in your subject line.
IEEE prohibits discrimination, harassment, and bullying. For more information, visit www.ieee.org/web/aboutus/whatis/policies/p9-26.html.

IEEE Computer Society Magazine Editors in Chief

Computer IEEE Intelligent Systems IEEE Security & Privacy

Jeff Voas, NIST San Murugesan, Western Sean Peisert, Lawrence
Sydney University Berkeley National Laboratory
Computing in Science and University of California,
& Engineering IEEE Internet Computing Davis
İlkay Altintaş, University Weisong Shi, University of
of California, San Diego Delaware IEEE Software
(Interim EIC) Sigrid Eldh, Ericsson,
IEEE Micro Mälardalen University,
IEEE Annals of the History Hsien-Hsin Sean Lee, Sweden; Carleton
of Computing Intel Corporation University, Canada
Troy Astarte,
Swansea University IEEE MultiMedia IT Professional
Balakrishnan Prabhakaran, Charalampos Z.
IEEE Computer Graphics University of Texas at Dallas Patrikakis, University of
and Applications West Attica
Pak Chung Wong, Trovares IEEE Pervasive Computing
and Bill & Melinda Gates Fahim Kawsar, Nokia Bell Labs
Foundation (Interim EIC) and University of Glasgow

2469-7087/24 © 2024 IEEE Published by the IEEE Computer Society December 2024 1
DECEMBER 2024 • VOLUME 10 • NUMBER 12

30
The Quantum
34
“Sensoring”
40
AI for Water
Cybersecurity the Farm
Threat May
Arrive Sooner
Than You Think
 Machine Learning
8 Orchestrating Networked Machine Learning
Applications Using Autosteer
ZHENYU WEN, HAOZHEN HU, RENYU YANG, BIN QIAN, RINGO W. H. SHAM,
RUI SUN, JIE XU, PANKESH PATEL, OMER RANA, SCHAHRAM DUSTDAR, AND
RAJIV RANJAN

16 Feature Interactions on Steroids: On the Composition

of ML Models
SVEN APEL, CHRISTIAN KÄSTNER, AND EUNSUK KANG

Quantum Computing
22 Distributed Quantum Machine Learning: Federated
and Model-Parallel Approaches
JINDI WU, TIANJIE HU, AND QUN LI

30 The Quantum Cybersecurity Threat May Arrive Sooner

Than You Think
PETE FORD

Smart Manufacturing
34 “Sensoring” the Farm
JOANNA F. DeFRANCO, NIR KSHETRI, AND JEFFREY VOAS

40 AI for Water
FERAS A. BATARSEH AND AJAY KULKARNI

History
46 Computer Networking Initiatives in One of the World’s
Remote Cities
T. ALEX REID

56 Lessons From the Father of Software Engineering

RICARDO VALERDI

Departments
4 Magazine Roundup
7 Editor’s Note: Keeping Up With the Rise of Machine Learning
60 Conference Calendar

Subscribe to ComputingEdge for free at

www.computer.org/computingedge
 Magazine Roundup

T he IEEE Computer Society’s lineup of 12 peer-reviewed technical magazines covers cutting-edge topics
ranging from software design and computer graphics to Internet computing and security, from scientific
applications and machine intelligence to visualization and microchip design. Here are highlights from recent issues.

often at other institutions. An HPC

Certifiability Analysis of software ecosystem approach is
Machine Learning Systems required to support dependencies Integrating GPT as an
for Low-Risk Automotive on third-party scientific software. Assistant for Low-Cost
Applications This January–March 2024 Com- Virtual Reality Escape-
puting in Science & Engineering Room Games
In this article, featured in the Sep- article describes the U.S. Exascale
tember 2024 issue of Computer, the Computing Project (ECP) contribu- In this article, featured in the
authors analyze the existing safety tions to HPC software ecosystem July/August 2024 issue of IEEE
standard, ISO 26262, for automo- challenges. Computer Graphics and Appli-
tive applications, determining the cations, the authors explore the
certifiability of machine learning integration of generative pre-
(ML) approaches used in low-risk trained transformer (GPT), an AI
automotive applications. This can Making Innovation in the language model developed by
help with assuring the security and Mexican Silicon Valley: OpenAI, as an assistant in low-
safety of ML-based autonomous The Early Years of El cost virtual escape games. Their
driving systems, gaining the trust Centro de Tecnología de study focuses on the synergy
of regulators, certification agen- Semiconductores (1981–2001) between virtual reality (VR) and
cies, and stakeholders. GPT, aiming to evaluate its per-
This article, featured in the April– formance in helping solve logi-
June 2024 issue of IEEE Annals of cal challenges within a specific
the History of Computing, tells context in the virtual environ-
Providing a Flexible and the early story of El Centro de Tec- ment while acting as a person-
Comprehensive Software nología de Semiconductores (CTS) alized assistant through voice
Stack Via Spack, an Extreme- as a site of innovation. It argues interaction.
Scale Scientific Software that, along with economic and sci-
Stack, and Software entific development goals, CTS
Development Kits furthered political and geopolitical
change agendas for IBM and Mex- Effective Adversarial
To manage the complex demands ico. These included reorganizing Examples Identification of
of modern high-performance com- labor around global supply chains Credit Card Transactions
puting (HPC), software applica- and maintaining specific power
tions increasingly depend on soft- dynamics between the Global Credit cards are a prevalent
ware developed by other teams, North and South. method of transactions but are

4 December 2024 Published by the IEEE Computer Society  2469-7087/24 © 2024 IEEE
 susceptible to forgery, lead- critical national infrastructure partition-based reference frame
ing to numerous cases of fraud. (CNI). They propose a generic DT- recompression (RFRC) scheme
As a result, much research driven deception-based solu- which is capable of compress-
has been focused on employ- tion called securing cyberphys- ing a variable-size to a fixed bit
ing artificial intelligence (AI) to ical systems through DT-driven of access unit. Compared with
achieve high detection perfor- deception (INCEPTION), which the conventional schemes, which
mance. However, the accuracy can serve as a research-based DT- compress fixed-size partition
of these AI-based methods may driven deception platform for CNI. to variable bits, this work can
be challenged by attack tech- increase the compression ratio by
niques using adversarial exam- greatly reducing the redundancy
ples. To address this issue, this of the partitions’ boundaries.
July/August 2024 IEEE Intelligent Mosaic Pages: Big TLB Reach
Systems article utilizes neu- With Small Pages
ron activation status distribu-
tion and deep neural networks In this article, featured in the July/
as detection tools. Furthermore, August 2024 issue of IEEE Micro, Energy Communities:
the experiments employ three the authors introduce mosaic Pervasive Technologies and
methods to generate adversarial pages, which increase transla- Collective Futures
examples, showcasing the effec- tion lookaside buffer (TLB) reach
tiveness of the proposed detec- by compressing multiple, discrete Energy communities are emer-
tion approach. translations into one TLB entry. gent sociotechnical constella-
Their results show that Mosa- tions, where local actors col-
ic’s constraints on memory map- lectively organize green energy
pings do not harm performance, initiatives. In this article, fea-
Digital-Twin-Driven and there are no conflicts before tured in the April–June 2024 issue
Deception Platform: memory is 98% full—at which of IEEE Pervasive Computing,
Vision and Way Forward point a traditional design would the authors draw on emerging
also likely swap. research and five situated design
Digital twin (DT) technology cases to illustrate the many
provides new opportunities to nuances involved in envisioning
enhance the robustness and and developing technology for
resilience of critical infrastruc- Adaptive Detachable energy communities. Reflections
ture. In this July/August 2024 Partition-Based Reference from these cases suggest a need
IEEE Internet Computing article, Frame Recompression for for design approaches that bet-
the authors discuss the poten- Video Coding ter account for the diverse social
tial of DTs as a proactive secu- lifeworlds into which these tech-
rity enabler and present a vision In this April–June 2024 IEEE Mul- nologies are meant to foster
for using DTs as a deception plat- tiMedia article, the authors pres- local engagement toward sus-
form to thwart cyberattacks on ent an adaptive detachable tainable futures.

www.computer.org/computingedge 5
MAGAZINE ROUNDUP

transformations across multi-

ple domains, including software
A Viewpoint: Safer Heaps development. In this May/June
with Practical Architectural 2024 IT Professional article, the
Security Primitives authors discuss ChatGPT’s role in
software engineering, including
In this article, featured in the July/ an investigation of implications
August 2024 issue of IEEE Security and applications highlighting
& Privacy, the authors argue that ChatGPT’s code-assistance capa-
architectural security primitives bilities. Through a series of analy-
are a promising basis for fast and ses, they discuss the real impact
secure program heaps. They dis- of ChatGPT on open source soft-
cuss MPKAlloc, a recent research ware development.
effort demonstrating the con-
crete benefits of this approach
using Intel MPK to harden a pro-
Stay connected by
duction allocator.
discovering the valuable
tools and resources of IEEE:
Create a personalized
experience
Integrating Static Quality
Get geo and
Assurance in CI Chatbot
interest-based
Development Workflows recommendations
Schedule, manage,
To fill a gap in proposals to inte- or join meetups
grate automated quality assur- virtually
ance mechanisms into the chat- Read and download
bot development workflow, the your IEEE magazines
authors of this article from the Stay up-to-date
September/October 2024 issue
Join the IEEE with the latest news
of IEEE Software present a con-
tinuous integration workflow Computer Locate IEEE members
by location, interests,
for chatbot development, imple- Society and affiliations
mented as GitHub actions, and computer.org/join
show its usefulness by its appli-
cation to open source chatbots.

ChatGPT for Software

Development: Opportunities
and Challenges

Rapid natural language pro-

cessing advances, such as Ope-
nAI’s ChatGPT, promise profound

6 ComputingEdge December 2024

 Editor’s Note

Keeping Up With the Rise

of Machine Learning

T he applications of machine
learning (ML) have
rocketed in recent years. ML is
sky-
present AUTOSTEER, a software
platform for deploying ML applica-
tions across hardware, cloud, and
improve farming and water reg-
ulation. Computer article “‘Sen-
soring’ the Farm” analyzes how
now being used in a vast range edge devices. The authors of IEEE the Internet of Things (IoT) can
of industries including transpor- Software article, “Feature Inter- be used to increase farming effi-
tation, healthcare, marketing, actions on Steroids: On the Com- ciency and addresses the chal-
cybersecurity, customer service, position of ML Models,” propose lenges of agricultural computing.
data analysis, social media, and rethinking ML model composition The authors of the Computer arti-
even space exploration. The pos- to create stronger and more accu- cle “AI for Water” show how AI can
sibilities of ML seem endless. And rate specifications. improve water access, treatment,
yet—are software developers Despite the exciting poten- and management.
keeping up? This issue of Com- tials of quantum computing, Present-day engineers can
putingEdge examines the appli- there are also risks and obsta- look back on their predecessors’
cations of ML, where it needs cles to consider. The authors of contributions to draw valuable
improvement, and how to improve “Distributed Quantum Machine lessons about computing that are
it, as well as the intersection of Learning: Federated and Model- still applicable today while avoid-
ML and quantum computing. The Parallel Approaches,” from IEEE ing the same mistakes. The article
articles go on to discuss cyber Internet Computing, explore the “Computer Networking Initiatives
threats in quantum computing as challenges of implementing two in One of the World’s Remote Cit-
well as history lessons about soft- types of quantum ML methodol- ies,” from IEEE Annals of the History
ware engineering and computer ogies and suggest potential solu- of Computing, tells the story of the
networking. tions. “The Quantum Cyberse- risky—but rewarding—purchase
Developers are coming curity Threat May Arrive Sooner of the first time-sharing computer
up with new ways to apply ML Than You Think,” from Computer, in Australia. In “Lessons From the
with improved specifications. reveals how quantum computers Father of Software Engineering,”
In “Orchestrating Networked will eventually have the ability to from Computer, the author reflects
Machine Learning Applications break current encryption. on the life of Barry Boehm and how
Using Autosteer,” from IEEE Inter- Automation is being employed he impacted the design and man-
net Computing, the authors for agricultural purposes—to agement of software.

2469-7087/24 © 2024 IEEE Published by the IEEE Computer Society December 2024 7
 EDITOR: Ewa
EDITOR: Ewa Deelman,
Deelman, [email protected]
[email protected]
This article originally
appeared in
DEPARTMENT:
DEPARTMENT: VIEW FROM THE
THE CLOUD
CLOUD
vol. 26, no. 6, 2022

Orchestrating Networked
Orchestrating Networked Machine Learning
Machine Learning Applications
Applications Using Autosteer
Using Autosteer
Zhenyu Wen and Haozhen Hu , Zhejiang University of Technology, Hangzhou, 310023, China
Zhenyu Wen and
Renyu Yang Haozhenof
, University Hu , Zhejiang
Leeds, University
LS2 9JT, of Technology, Hangzhou, 310023, China
Leeds, U.K.
Renyu Yang, Ringo
Bin Qian , University of Leeds,
W. H. Sham, LS2Sun
and Rui 9JT, Leeds, U.K. University, NE1 7RU, Newcastle, U.K.
, Newcastle
Bin
Jie Qian , Ringo W.
Xu, University H. Sham,
of Leeds, andLS2
Leeds, Rui 9JT,
Sun U.K.
, Newcastle University, NE1 7RU, Newcastle, U.K.
Jie Xu, University
Pankesh Patel , of Leeds, Leeds,
University LS2Carolina,
of South 9JT, U.K.Columbia, SC, 29208, USA
Pankesh Patel, Cardiff
Omer Rana , University of South
University, CF24Carolina, Columbia,
3AA, Cardiff, U.K. SC, 29208, USA
Omer RanaDustdar
Schahram , Cardiff, University, CF24Vienna,
TU Wien, 1040, 3AA, Cardiff,
AustriaU.K.
Schahram Dustdar
Rajiv Ranjan , TU Wien,
, Newcastle 1040, Vienna,
University, NE1 7RU,Austria
Newcastle, U.K.
Rajiv Ranjan , Newcastle University, NE1 7RU, Newcastle, U.K.

A platform for orchestrating networked machine learning (ML) applications

Aover distributed
platform environmentsnetworked
for orchestrating is described. ML applications
machine learning (ML)are transformed
applicationsinto
over distributed
environments is described. ML applications are transformed intoand
automated pipelines that manage the whole application lifecycle production-
automated pipelines
grade
that implementations
manage are automatically
the whole application constructed.
lifecycle We present AUTOSTEER,
and production-grade a
implementations are
software platform that can deploy ML applications on various hardware resources—
automatically constructed. We present AUTOSTEER, a software platform that can deploy
ML applicationsusing
interconnected on various hardwarenetwork
heterogeneous resources—interconnected
resources—across cloud using heterogeneous
and edge
network
devices. resources—across cloud and edge
Device placement optimization devices.
and model Device placement
adaptation optimization
are used as control
and model
actions adaptation
to support are usedrequirements
application as control actions to support
and maximize the application
performancerequirements
of ML
and
modelmaximize theover
execution performance of MLcomputing
heterogeneous model execution overThe
resources. heterogeneous
performancecomputing
of
resources. The performance of deployed applications is continually monitored
deployed applications is continually monitored at runtime to overcome performance at runtime
to overcome performance degradation due to incorrect application parameter
degradation due to incorrect application parameter settings or model decay. Three settings
or model decay. Three real-world applications are used to demonstrate how AUTOSTEER
real-world applications are used to demonstrate how AUTOSTEER can support
can support application deployment and runtime performance guarantees.
application deployment and runtime performance guarantees.

M
achine learning (ML) systems and applica- ML applications executing over a networked platform
tions are intrinsically nondeterministic and are arguably complex systems, which have to be continu-
need to operate in an environment that is ously updated and maintained. ML applications need to
constantly evolving, and contains ever-changing data. be transformed into automated pipelines that manage
Typically, a networked ML application consists of a the whole application lifecycle and build production-
variety of components for data collection, device con- grade ML implementations. A pipeline workflow, typically
trol, model inference (e.g., speech recognition, object in the form of a graph representing the component inter-
detection), which are deployed and managed at differ- connections in an ML application, can comprise: data
ent locations, i.e., either on locally managed servers or management, model learning (model selection, training,
remotely in cloud data centers or edge environments. and hyperparameter selection), model testing, and valida-
tion and model deployment. Thereafter, runtime manage-
ment is responsible for ensuring performance guarantee,
i.e., end-to-end model performance optimization and
model update,1 so that the deployed ML applications can
be dynamically modified to runtime environment.
1089-7801  2022 IEEE
Digital Object Identifier 10.1109/MIC.2022.3180907 Doing so manually is generally unrealistic and not
Date of current version 23 December 2022. scalable, particularly when thousands of ML applications

8 December2022
November/December 2024 Published
Published by the IEEEby the IEEE Society
Computer Computer SocietyIEEE
 51
2469-7087/24 © 2024 IEEE
Internet Computing
VIEW FROM THE CLOUD VIEW FROM THE CLOUD

FIGURE 1. Conceptual workflow.

are submitted and maintained in edge and cloud plat- degradation caused by obsolete application parameters
form that may be composed of hundreds of devices with setting or model decay. Finally, we use three real-world
heterogeneous hardware and software specifications. applications that are executed upon AUTOSTEER to
Continuous and automatic orchestration plays a pivotal showcase how the mechanisms are engaged in the appli-
role in deploying, managing, and synchronizing models of cation deployment and runtime maintenance.
the ML applications across multiple tiers in a distributed
computing environment. For instance, the trained mod- MOTIVATION
els will be published and delivered to specific cloud serv-
ers or edge devices to run inference. Some specific Motivating Examples
applications, e.g., federated learning tasks require on- We primarily categorize the networked ML applications
device training, indicating more complex device place- into a) centralized off-site ML applications that can be
ment and model synchronization. Moreover, model trained offline or offsite, and b) distributed on-site/feder-
decay arising from changes in data, would inevitably ated ML applications that must build their models using
diminish model accuracy over time. Hence, an orchestra- local dataset on individual device and, in some cases,
tor calls for observation of the performance deviation share and aggregate models with other peer devices.
and redeployment of the updated models. Centralized off-site learning applications: A smart
Deploying such networked ML systems, particularly home application allows users to observe the occu-
in an IoT and edge environment can be challenging due pancy of their house, remotely control the smart devi-
to the difficulty in managing the complexity of heteroge- ces (e.g., LEDs, air conditioner) via smartphone and
neous network and hardware resources. A variety of devi- even automatically control the smart devices. For
ces are used for data exchange, model training, and data example, a smart home application can automatically
analysis encompassing edge devices (such as IoT gate- adjust the temperature of air conditioners based on
ways and base stations) and servers (such as GPU, CPU, the occupancy, weather, and so on.
and TPU-based devices). Existing ML model development Distributed on-site/federated learning applications:
can be computationally expensive and resource inten- A high-quality brain tumor detection application relies
sive, which impede the effective deployment of applica- on a huge amount of magnetic resonance imaging data
tions, particularly those with strict latency requirements that is only locally available and managed within a spe-
to resource-constrained devices. cific institution domain due to GDPR and other privacy
In this article, we propose a platform solution to regulations. A shared model is typically distributed to
deployment and runtime management for the pipelines different data owners and trained locally. Locally trained
of networked ML applications. We devise AUTOSTEER, a models will be combined into a consensus model.
management system that can automatically deploy net-
worked ML applications over heterogeneous network Research Scope and Overview
and hardware resources while ensuring their perfor- In general, the pipeline for such an application can be
mance through deployment plan optimization and model depicted as the workflow in Figure 1. The pipeline starts
adaptation. At runtime, AUTOSTEER continually monitors with and augments an initial model that has been
the performance of deployed applications and automati- trained offline along with a reference to metadata and
cally performs model update to mitigate performance the associated data sources on which the model has

52
www.computer.org/computingedge
IEEE Internet Computing November/December 2022 9
VIEW FROM THE CLOUD VIEW FROM THE CLOUD

been trained. Thereafter, the workflow management solution under numerous constraints. The manifestation
platform typically addresses two fundamental prob- of heterogeneity intrinsically stems from the static
lems: planning for device placement and model adapta- attributes of the hardware, such as CPU, GPU, memory,
tion in the deployment phase and model execution SSD, and network bandwidth, and of the software includ-
performance guarantee in the runtime phase. ing operating system version, clock speed, and particu-
Determining the placement of ML components on larly software libraries. The compatibility of a given
available resources remains a key challenge—espe- hardware or library version even becomes a hard con-
cially due to heterogeneity of resources. In addition, straint, for any violations of such requirements would
models have to be converted, for example through completely fail the deployment. For example, some com-
model pruning,2 posttraining quantization,3 and identi- ponents are compiled for ARM Mali cannot be executed
fying a “focus” for the associated model through distil- on Nvidia GPU. The network constraints, such as band-
lation techniques. This enables the generated models width sharing among colocated components or network
to best fit the target device, balancing the model size latency specified by each individual component, will fur-
with accuracy of prediction. Significant recent efforts ther exacerbate the planning complexity.
in this area include TinyML and EdgeML. 2) Model adaptation: The advancement of deep
Once the plan of deployment comes into effect, run- models, such as recurrent neural network and convolu-
time management ensures that the model performance tional neural network leads to the substantially
can be monitored and overcomes model staleness. In increased parameter number and the resultant compu-
the automated and continuous pipeline, triggers can be tational cost, which hinders the real-world model
used to update application parameters or retrain the deployment into embedded and edge devices. Hence,
stale model with fresh data when performance observ- model pruning and compression can be used to reduce
ably degrades due to dynamic environment changes, model size, remove redundant weights, such that pre-
such as network speed drop, workload bursting, model trained models can better adapt to portable devices
drift, or lack of generalization. For applications of feder- with limited resources (e.g., memory, CPU, power, and
ated learning and distributed training, the platform run- bandwidth) and be applied into real-time applications.
time also needs to enforce efficient on-device training. 3) Enabling dependent components within a pipe-
A key focus of this work is to devise an orchestration line: Each individual ML model has its own specification
system for supporting multiple ML model development and format of input and output data. Dependencies are
and performance optimization. In addition, the system referred to as the interactions, such as the data flows
needs to scale to support both application size and and remote callings, among interconnected compo-
resource heterogeneity. To underpin precise perfor- nents. This would be problematic and challenging par-
mance monitoring and anomaly detection while measur- ticularly when components deployed on various devices
ing platform health and resource utilization, we also need are interconnected via different network types and pro-
to track and inspect (distributed) system fingerprints— tocols. Hence, it is imperative to design an effective
consisting of various performance indicators and appli- data messaging system to orchestrate the data flow
cation metrics, such as drift and prediction scores. and manage the network traffic across different models
while considering the particular specification and data
format.
CHALLENGES Optimized runtime management: Improper applica-
We elaborate on these specific challenges facing the tion parameter setting or model decay could result in
ML workflow platform in the following notable aspects. poor performance of an ML application and even fail-
Complexity of device placement and model adap- ures. The first task of runtime management is to per-
tation: Planning for a pipeline of a given ML applica- form end-to-end and intraapplication optimization.
tion indicates a mapping procedure between awaiting Application parameters (e.g., model accuracy, task off-
models and available computing resources on the loading rate) need to be adjusted at runtime to ensure
devices. To accommodate the specific demands of the allocated resource can guarantee the expected per-
diverse distributed or federated learning applications, formance level. To do so, the orchestration system
infrastructure resources have become increasingly should be capable of automatically detecting any per-
heterogeneous, making the planning a far more intri- formance degradation of the deployed applications and
cate task. then dynamically work out the optimal configuration to
1) Device placement: Successfully deploying sizeable rescue the abnormal performance. Second, in the face
components of the ML applications served in the plat- of any model failures, the orchestration system should
form requires stringent capacity check and optimization automatically perform local on-device training while

10 ComputingEdge
November/December 2022 IEEE Internet Computing 53
December 2024
VIEW FROM THE CLOUD VIEW FROM THE CLOUD

FIGURE 2. Architecture of AUTOSTEER.

synchronize and aggregate the up-to-date global mod- such as model accuracy and end-to-end latency. To
els on the fly. achieve an automatic deployment, we need to trans-
Low-cost platform monitoring and troubleshooting: late this knowledge to machine-understandable lan-
Monitoring is one of the primary issues in maintaining guage. We use a UML-based visual domain specific
ML applications and systems; outline or anomaly detec- language4 that can easily represent the component
tion is important to find out unexpected model predic- dependencies within an application and specify the for-
tion or any system-wide issues in the early stage. mat and source of input and output of each individual
However, anomaly detection and trouble-shooting component. As a result, the interactions between com-
could be challenging as high-quality labeled data are ponents, such as data flows and service calls, are
sparse and difficult to obtain and hence only semisuper- loosely coupled through interfaces and agnostic about
vised or unsupervised approaches could be applied. The any model updates. Apart from the application specifi-
overhead is another non-negligible consideration when cation, standardized resource specification is the key
designing application instrumentation and metric col- to automatic and efficient deployment. We exploit5 for
lection. This usually indicates a tradeoff between the specifying the available underlying computing resour-
accuracy and granularity of the measured data. Hence, ces and the hardware and software requirements of
the platform solution of infrastructure monitor should each application.
have an overall co-design of metric sampling, storage Planning optimization for device placement: To navi-
and real-time analysis. gate the algorithmic complexity, the orchestrator in
AUTOSTEER adopts two optimization techniques: gradi-
ent-based optimization6 and reinforcement learning
SYSTEM DESIGN
(RL).7 Gradient-based approaches work upon a realistic
In response to the aforementioned challenges, we
model to formalize an optimization problem and usually
develop AUTOSTEER, an orchestration platform for
have relatively low time complexity without the need of
application deployment and runtime management. In
a priori knowledge or experience, which are therefore
this section we mainly highlight a set of key techniques
suitable for new applications. In contrast, RL-based
used for implementing the orchestration mechanism.
methods can learn the optimal planning from the experi-
Figure 2 describes the architecture of AUTOSTEER.
ences and can better support the uncertainties com-
pared the gradient-based solutions.
Automatic Application Deployment We also construct an efficient data messaging sub-
Application and resource specification: The user sub- system where two types of dependencies are
mits an ML application with execution logic, pretrained defined—data flow and service call. Since the orches-
models and specifies the pertaining requirements, tration system needs to deliver a large volume of data

54
www.computer.org/computingedge
IEEE Internet Computing November/December 2022 11
VIEW FROM THE CLOUD VIEW FROM THE CLOUD

in distributed environments, high system throughput We specify model parameters based on extensive
becomes a critical system objective. We employ the benchmarking experiments and transform the prob-
publish/subscribe paradigm implemented in Apache lem of finding the “best” setting of parameters into an
Kafka to underpin the data flows. The service call, on optimization problem using techniques, such as con-
the other hand, is implemented through RESTful APIs, vex optimization, evolution- and gradient-based meth-
as the precise command delivery is the primary goal. ods. RL is an alternative approach that uses statistical
Both the AUTOSTEER publish/subscribe and RESTful or deep learning model where the application parame-
paradigms can be implemented upon a vast majority ters are the actions of the agent, and the available
of network types and protocols, hence capable of sup- computing resources represent the environment. The
porting most networked ML applications. system performance is represented by the reaction of
the environment to the actions. As opposed to the
optimization-based approaches that have better
Model Adaptation
interpretability but need extra hand-crafted modeling
Computation optimization aims to improve the execu-
process, the RL-based approaches have better repre-
tion efficiency of different computation units associated
sentation capabilities and can learn to set optimal
with the model (e.g., vector–vector, vector–matrix, and
application parameters from experience.
matrix–matrix operations) on various hardware. Optimiz-
ing the execution pipeline of the computation graph of a
neural network can further improve model performance.
Model Update
We use TensorRT along with the adjustment of weights
Coping with the drift: During the lifecycle of an ML appli-
and numerical precision associated with the activation
cation, the relationship between the input variables and
function (e.g., INT8 and FP16). Model architecture optimi-
the performance of the targeting prediction inevitably
zation improves the efficiency of on-device computation
experiences constant change and drift over time. The
through well-designed models, such as MobileNetV2,
model drift usually originates from the following
ShuffleNet etc.,—part of the TensorFlow-Lite toolkit). We
aspects. 1) Invalid measurement indicator: the replace-
use YOLOv38 to strike a balance between computation
ment of data collection devices may give rise to differ-
efficiency and model accuracy.
ent value spaces and a broken device could always
In addition, more advanced and customizable
deliver nil reading. 2) Concept drift: data distribution or
approaches, such as neural architecture search (NAS)9
statistical characteristics, which is uncertain and fre-
and model compression can be implemented in AUTO-
quently varying over time, may lead to concept drift. 3)
STEER further. NAS automates the search of an optimal
Data drift: the model effectiveness is also prone to
network structure with the aid of RL or genetic algo-
inherent changes, such as the seasonal temperature
rithm-based approaches. However, it is computation-
rise and fall. Drifts can be roughly categorized into sev-
intensive and tends to be problematic given the portable
eral classes: sudden drift (sudden change of the data
devices with limited resources. Model compression is
pattern), gradual/incremental drift (new pattern that
thus extensively studied in three notable aspects: model
replaces the old ones within a period of time), and reoc-
pruning that removes the redundant parameters within
curring drift (old patterns repop up later).
the networks; quantization that reduces the weights pre-
It is imperative to detect such drifts, understand the
cision, and knowledge distillation10 that trains a new
degree of drift and intervene the model for adapting to
small model based on a larger model. Quantization is the
changing environments. There are three representative
most straightforward approach at the risk of precision
classes of drift detection. 1) Error rate-based approaches
degradation and model pruning is the most well-estab-
focus on the online detection of errors or sudden
lished approach but requires extra calibration process.
changes for triggering the model update. 2) Data distribu-
Integrating mixed techniques in the platform is already
tion-based approaches mainly measure the statistical
underway for building more adaptive and robust models.
similarities between the original data and the new data
and check if the difference is sufficient for model update.
End-to-End Application Optimization 3) Hypothesis test-based approaches, built upon the pre-
In a networked ML system, computational and net- vious two methods, apply various hypothesis tests to
work resources are dynamically available at different quantify further the severity of model drift. Based on
levels. Application parameters, such as input rate and these approaches, our solution can determine when to
the targeted accuracy need to be adjusted, in intervene according to the starting and ending points of
response to the ever-changing traffic congestion, to the drift, where to intervene, i.e., localizing the concept/
assure the end-to-end latency or system throughput. data drift in the feature space, and how to intervene, in

12 ComputingEdge
November/December 2022 IEEE Internet Computing 55
December 2024
VIEW FROM THE CLOUD VIEW FROM THE CLOUD

FIGURE 3. Edge-cloud video analysis application and an early performance comparison. (a) Illustration of an edge-cloud video
analysis system. (b) Performance of workload optimizer in different system working conditions.

the light of the type and degree of the drift, by adaptively The metric tracking and tracing system of our orches-
choose model update strategies. The most straightfor- tration infrastructure collects system logs, model
ward approach is the model retraining and updating. For metrics (task execution status, prediction statistics,
concept drift, we ensemble several base classifiers or uti- and evaluation metrics as baselines), system metrics
lize knowledge transfer learning for the emerging new tar- (request latency, error rates, network status, etc.), and
get variables. resource metrics (CPU utilization, memory utilization,
System implementation: The amount of data GPU usage, etc.) in real time, and ships them to a
engaged in the model update has an impact on the centralized analytic platform. We adopt the random
training effectiveness and the system overhead: less sampling mechanism on each agent that is deployed
data can reduce computation and storage cost but only on each physical node, for reducing the overhead of
reflect the latest data distribution; more data are benefi- data collection. More advanced technologies, such as
cial for reshaping models with higher precision, along sketch12 can be further added. Anomaly Detector com-
with increased overhead. We employ an adaptive win- prises real-time event-based processing units, used for
dow-based solution to select the optimal data amount identifying per-application performance degradation
used for on-device training and/or global model syn- while Root-cause Analyzer is implemented to trouble-
chronization via ADWIN11 algorithm: instead of using a shoot the causes of performance degradation based
fixed time window, the algorithm calculates the drift on the collected performance indicators.
rate from all possible windows and selects the best cut
that reveals the optimal drift level. We modularize and
implement the drift detection and alarming system in CASE STUDY: EDGE-BASED REAL-
AUTOSTEER. The detection module is responsible for TIME VIDEO ANALYTICS
data retrieval and extraction of data statistical proper- In this section, we showcase a real-world application
ties, and we then leverage hypothesis tests to evaluate backed up by the deployment and runtime manage-
the drift degree. Once the alarming system confirms the ment mechanisms in AUTOSTEER.
existence of the model drift, we employ techniques in As shown in Figure 3(a), we develop a video analyti-
Section Model Adaptation for efficient on-device train- cal application following the edge-cloud paradigm. A
ing. For federated learning applications, once local set of video generating devices (e.g., traffic surveil-
model has been updated, we also trigger gradient aggre- lance cameras, drones, mobile phones) produce live
gation to keep the global model up-to-date. video streams, which are then processed either on
low-power edge devices (e.g., Raspberry pi, Jetson
Infrastructure Monitor and Nano, computing chips), or GPU cluster in cloud data-
Maintenance centers. We prototype the video analytic application
To learn how the applications perform, we either collect via object detection models yolo3 and the wide area
general-purpose telemetry metrics in a black-box man- network communication between edge devices and
ner or instrument, as an integral part of the models, the data center is implemented by using the real-time
subsystems or system services, in a white-box manner. video stream transmission protocol.

56
www.computer.org/computingedge
IEEE Internet Computing November/December 2022 13
VIEW FROM THE CLOUD VIEW FROM THE CLOUD

The heterogeneity of edge nodes and the interplay Institute Post-Doctoral Enrichment Award Pro-
among the edge and cloud introduce uncertainties gramme, and in part by U.K. Alan Turing Pilot Project.
regarding network latency, hardware slowdown, or fail-
ures. As discussed in the “End-to-End Application Opti-
mization,’’ section the collected fingerprints and system
status are mathematically modeled with a hierarchy
REFERENCES
1. B. Qian et al., “Orchestrating the development lifecycle
queuing model that reveals the relationships between
of machine learning-based IoT applications: A
the workload offloading rate (between the edge and
taxonomy and survey,” ACM Comput. Surv., vol. 53,
cloud) and the system latency and throughput. We then
no. 4, pp. 1–47, 2020.
formulate a min-latency optimization problem bounded
2. M. Zhu and S. Gupta, “To prune, or not to prune:
by a minimal throughput threshold. For model optimiza-
Exploring the efficacy of pruning for model
tion, we implement two gradient-based optimization
compression,” 2017, arXiv:1710.01878.
algorithms (i.e., PGD-VAO and PGS-VAO) to ascertain a
3. B. Jacob et al., “Quantization and training of neural
solution to minimizing the overall latency. All compo-
networks for efficient integer-arithmetic-only
nents are containerized and deployed at both the edge
inference,” in Proc. IEEE Conf. Comput. Vis. Pattern
and the cloud side via AUTOSTEER.
Recognit., 2018, pp. 2704–2713.
Figure 3(b) shows the performance of our pro-
4. T. Eterovic, E. Kaljic, D. Donko, A. Salihbegovic, and
posed algorithms under empty, normal, and busy sys-
S. Ribic, “An Internet of Things visual domain
tem workloads. Specifically, we insert video chunks
specific modeling language based on UML,” in Proc.
into system buffering queues to simulate different
25th Int. Conf. Inf., Commun. Automat. Technol.,
workloads. Then, we test our algorithms against the
2015, pp. 1–5.
other state-of-the-art task-offloading approaches, i.e.,
5. T. Binz, U. Breitenbu
€ cher, O. Kopp, and F. Leymann,
DeepDecision and FastVA. We can see that with the
“TOSCA: Portable automated deployment and
increase of the workload, the system latency is
management of cloud applications,” in Advanced
increasing as well. It is also clear that our modeling-
Web Services. New York, NY, USA: Springer, 2014,
based algorithms (e.g., PGS-VAO, PGD-VAO, FastVA)
pp. 527–549.
perform better than nonmodeling-based algorithms.
6. D. Maclaurin, D. Duvenaud, and R. Adams, “Gradient-
based hyperparameter optimization through
CONCLUSION reversible learning,” in Proc. Int. Conf. Mach. Learn.,
Most prior work related to ML applications focuses on 2015, pp. 2113–2122.
algorithm design and optimization for better training 7. R. S. Sutton and A. G. Barto, Reinforcement Learning:
ML models. An introduction. Cambridge, MA, USA: MIT Press, 2018.
Although such work is essential for specific applica- 8. J. Redmon and A. Farhadi, “Yolov3: An incremental
tions, there are few studies on the holistic orchestration improvement,” 2018, arXiv:1804.02767.
solution to maintaining the lifecycle of networked ML 9. T. Elsken, J. H. Metzen, and F. Hutter, “Neural
applications. In this article, we first highlight several key architecture search: A survey,” J. Mach. Learn. Res.,
challenges facing the orchestration systems. We then vol. 20, no. 1, pp. 1997–2017, 2019.
present a set of techniques to deploy ML applications 10. J. H. Cho and B. Hariharan, “On the efficacy of
onto resources across cloud and edge devices and knowledge distillation,” in Proc. IEEE/CVF Int. Conf.
assure their runtime performance, making models being Comput. Vis., 2019, pp. 4793–4801.
served free from model decay and performance degra- 11. A. Bifet and R. Gavalda, “Learning from time-changing
dation due to inappropriate parameter setting. These data with adaptive windowing,” in Proc. 2007 SIAM Int.
assist in finding effective pathways to automating the Conf. Data Mining, 2007, pp. 443–448.
management of networked ML applications at produc- 12. T. Yang, Y. Zhou, H. Jin, S. Chen, and X. Li, “Pyramid
tion level, although, admittedly, it still calls for significant sketch: A sketch framework for frequency estimation
effort in large-scale engineering practices and integra- of data streams,” in Proc. VLDB Endowment, 2017,
tion with wider domain-specific scenarios. pp. 1442–1453.

ACKNOWLEDGMENTS ZHENYU WEN is a professor with the Institute of Cyberspace

This work was supported in part by U.K. EPSRC under Security, Zhejiang University of Technology, Hangzhou,
Grant EP/T01461X/1, in part by U.K. Alan Turing 310023, China. Contact him at [email protected].

14 ComputingEdge
November/December 2022 IEEE Internet Computing 57
December 2024
VIEW FROM THE CLOUD VIEW FROM THE CLOUD

HAOZHEN HU is currently working toward the master’s JIE XU is a chair professor with the School of Computing, the
degree with the College of Information, Zhejiang University of University of Leeds, LS2 9JT, Leeds, U.K., and chief scientist of
Technology, Hangzhou, 310023, China. Contact him at BDBC, Beihang University, Beijing, China. Contact him at
[email protected]. [email protected].

RENYU YANG is a research fellow with the University of Leeds, PANKESH PATEL is a researcher with AI Institute, University

LS2 9JT, Leeds, U.K. Contact him at [email protected]. of South Carolina, Columbia, SC, 29208, USA. Contact him at
[email protected].

BIN QIAN is a postgraduate research student with the school

OMER RANA is a full professor with the School of Computer Sci-
of computing, Newcastle University, NE1 7RU, Newcastle,
ence and Informatics, Cardiff University, CF24 3AA, Cardiff, U.K.
U.K. Contact him at [email protected].
Contact him at [email protected].

RINGO W. H. SHAM is a research technician with the school of SCHAHRAM DUSTDAR is a full professor of computer sci-
computing, Newcastle University, NE1 7RU, Newcastle, U.K. ence with TU Wien, 1040, Vienna, Austria. Contact him at
Contact him at [email protected] [email protected].

RUI SUN is a postgraduate research student in the school of RAJIV RANJAN is a chair and professor with Newcastle Univer-
computing, Newcastle University, NE1 7RU, Newcastle, U.K. sity, NE1 7RU, Newcastle, U.K., and with the China University of
Contact him at [email protected]. Geosciences, Wuhan, China. Contact him at [email protected].

ADVERTISER INFORMATION

Advertising Coordinator Central US, Northwest US, Southeast US, Asia/Pacific:

Eric Kincaid
Debbie Sims Email: [email protected]
Email: [email protected] Phone: +1 214-553-8513 | Fax: +1 888-886-8599
Phone: +1 714-816-2138 | Fax: +1 714-821-4010 Cell: +1 214-673-3742

Midwest US:
Advertising Sales Contacts Dave Jones
Email: [email protected]
Mid-Atlantic US, Northeast, Europe, the Middle East Phone: +1 708-442-5633 | Fax: +1 888-886-8599
and Africa: Cell: +1 708-624-9901
Dawn Scoda
Email: [email protected]
Jobs Board (West Coast and Asia), Classified Line Ads
Phone: +1 732-772-0160
Cell: +1 732-685-6068 | Fax: +1 732-772-0164 Heather Buonadies
Email: [email protected]
Southwest US, California:
Phone: +1 623-233-6575
Mike Hughes
Email: [email protected]
Cell: +1 805-208-5882 Jobs Board (East Coast and Europe), SE Radio Podcast

Marie Thompson
Email: [email protected]
Phone: +1 714-813-5094

58
www.computer.org/computingedge
IEEE Internet Computing November/December 2022 15
 EDITOR: Tim Menzies, North Carolina State University, [email protected]
This article originally
appeared in
DEPARTMENT: SE FOR AI
vol. 39, no. 3, 2022

Feature Interactions on
Steroids: On the Composition of
ML Models
Sven Apel, Christian Kästner, and Eunsuk Kang

FROM THE EDITOR

Is machine learning (ML) different? What lessons can we learn from software engineering (SE) that would
help ML applications? Here, it is argued that “feature interaction” (a well-studied problem in ML) is a nat-
ural framework within which to discuss developing ML apps.
This column publishes commentaries on the growing field of SE for AI. Submissions are welcomed
and encouraged (1,000–2,400 words, each figure and table counts as 250 words, try to use fewer than
12 references, and keep the discussion practitioner focused). Please submit your ideas to me at timm
@ieee.org.—Tim Menzies

O
ne of the key differences between tradi- CHALLENGES IN COMPOSING
tional software engineering and machine ML MODELS
learning (ML) is the lack of specifications Many systems do not use just one ML model but
for ML models. Traditionally, specifications provide a compose multiple models to solve complex prob-
cornerstone for compositional reasoning and for the lems. To automatically generate captions for images,
divide-and-conquer strategy of how we build large and one could try to learn a model that directly takes an
complex systems from components, but these are hard image and produces a caption, but a state-of-the-art
to come by for machine learned components. While solution decomposes the problem into three steps
the lack of specification seems like a fundamental with different models.1 First, a visual detector (a
new problem at first sight, in fact, software engineers convolutional neural network) predicts which of
routinely deal with iffy specifications in practice. We 1,000 common objects is visible in the image, then a
face weak specifications, wrong specifications, and language model (a maximum-entropy model) takes
unanticipated interactions among specifications. ML these objects and generates 500 plausible sen-
may push us further, but the problems are not funda- tences with them, and finally a caption ranker (a
mentally new. Rethinking ML model composition from deep multimodal similarity model) takes both the
the perspective of the feature-interaction problem original image and the generated sentences and
highlights the importance of software design. scores the combination to pick the best sentence as
the caption.
At a first glance, this looks like a great divide-
and-conquer story. We break down the problems
Digital Object Identifier 10.1109/MS.2021.3134386 into steps. Each model can be developed and
Date of current version: 18 April 2022 tested independently, using different modeling and

16 December 2024 Published by the IEEE Computer Society  2469-7087/24 © 2024 IEEE
 SE FOR AI

Visual Language Caption

implementation techniques, pos- Detector Model Reranker
sibly reusing models or training 1. teddy 0.92 1. A teddy 1. A blender
data from other domains. 2. on 0.92 bear. sitting on top
3. cake 0.9 2. A stuffed of a cake.
At a second glance though,
4. bear 0.87 bear. 2. A teddy
it seems we cannot really reason 5. stuffed 0.85 bear in front
modularly about problems. For . . . . . .
of a birthday
15. blender 0.57 108. A
example, as shown by Nushi et al.,2 cake.
blender
all three models somehow contrib- sitting on top 3. A cake
of a cake. sitting on top
ute to the poor caption “A blender
of a blender.
sitting on top of a cake” for the
image in Figure 1. The visual detec- FIGURE 1. Problems in each component leading to the poor caption. (Copyright ©
tor detects a blender where there 2017, Association for the Advancement of Artificial Intelligence. All rights reserved.)
is none, the language exhibits low
common-sense awareness with
“a blender sitting,” and the ranking model picks that For ML components, we do not have any meaning-
sentence, even though it includes a word with a low ful specifications in the traditional sense. For many
object-detection score. So, no component is behav- problems, we already have a hard time capturing what
ing perfectly or compensating for problems in oth- it means for a single prediction of a model to be “cor-
ers. There are no clear responsibilities or boundaries rect” (humans may not agree). Even if we had a strict
between the components that could be used to assign
blame. So, if we already have a problem with compos-
ing three models, how do we expect to build reliable WHILE THE LACK OF SPECIFICATION
systems with, say, the 18 models in Baidu’s self-driving SEEMS LIKE A FUNDAMENTAL NEW
car system?3 PROBLEM AT FIRST SIGHT, IN FACT,
SOFTWARE ENGINEERS ROUTINELY
THE ROOT OF THE PROBLEM: DEAL WITH IFFY SPECIFICATIONS IN
A LACK OF SPECIFICATIONS PRACTICE.
The core of the problem is that we do not have clear
specifications for what each model is supposed to
do. For traditional software components, a specifica- binary notion of correctness for any single prediction,
tion tells us whether a component’s output is either we would not expect a model to make correct predic-
correct or wrong for a given input—not “pretty good” tions for every possible input. In line with the aphorism
or “95% accurate.” We also would not accept correct “all models are wrong, but some are useful,” we do not
answers for 98% of all inputs, but would instead con- evaluate the correctness of a model but whether the
sider the component to have a bug if it produced a model fits our problem, usually approximated through
wrong result. In systems with multiple components, accuracy measures. These do not compose nicely.
we can assign blame by checking which component
produced a wrong output according to its specifica- THE LIMITS OF DECOMPOSITION:
tion (even if our specifications in practice are often FEATURE INTERACTIONS
weak and textual). While it may seem that modular decomposition in tra-
Specifications enable modular reasoning based ditional software systems is obvious and clean, unfor-
on logic: we can understand the consequences from tunately, that is not always the case either. It is not
combining two modules in terms of the composed even just a question of whether we write formal or
specifications. This is a cornerstone for a sound informal, strong or weak specifications, but a ques-
divide-and-conquer strategy and modularity, and it tion of whether this would be possible in the first
enables the vast reuse of libraries in building modern place. The study of feature interactions has made this
and complex software systems. amply clear.

www.computer.org/computingedge 17
SE FOR AI

There are many examples of feature interactions, we decompose them anyway and, necessarily,
and they typically follow a common pattern. We make simplifying assumptions that may not actu-
decompose the system into components, then design, ally hold. Humans simply cannot deal with com-
develop, and test these components in separation, but plexity beyond a certain scale—we do not have
finally observe unexpected behavior when compos- the cognitive capacity. We need to abstract and
ing them. The canonical example is a call forwarding decompose. We may make simplifications and
feature in a phone system that competes with an inde- create specifications that are weak or even wrong
pendently developed call waiting feature on how to for specific cases. We may do this intentionally for
respond to the same call on a busy line. We can often good reasons, hoping that we can resolve issues at
blame interaction problems on weak specifications composition time.
that did not anticipate the interaction, but in general The good news is that, most of the time, a
the problem is much bigger. divide-and-conquer approach pays off and imper-
fect decompositions work. For example, most
home-automation components work well together as
LIKELY THE MOST POWERFUL a rule, and control mechanisms can often dynamically
STRATEGY IN ADDRESSING THE adjust for unanticipated interactions. The problems
FEATURE-INTERACTION PROBLEM IS that reach the surface are the remaining unantici-
TO PREPARE FOR THE POSSIBILITY pated interactions that surprise us.
OF INTERACTIONS AS PART OF THE
SYSTEM DESIGN. COPING WITH
FEATURE INTERACTIONS
Feature interactions have been actively studied since
In complex systems, it is often not possible to at least the 1990s. The community has learned how
cleanly separate behavior and divide a problem into to build systems that work reasonably well despite
subproblems. Behavior is often antimodular. This is weak or even partially wrong component specifica-
particularly apparent when the real world is involved. tions. Dedicated analysis, design, and control tech-
Consider a home-automation system with a heating niques can anticipate and compensate for some
component, a ceiling fan, and a component to open modularity violations.
windows. We could specify the behavior of each While not a silver bullet, it is likely that there are
controller and reason about how each component insights from a long tradition of coping with feature
interacts with the environment, but the actions of interactions that may help us to better understand
components may influence each other through physi- and build systems composed of both multiple ML
cal processes in the environment (such as heated and traditional components. In both worlds, we
air moving through the ceiling fan to the open win- explicitly deal with modularity and composition
dow). Components may also compete for the same problems stemming from weak or missing specifica-
resources, such as electricity or human attention. tions. Let’s look at three strategies to manage fea-
To truly understand how the components behave in ture interactions.
concert, we would need to fully understand the envi-
ronment in addition to the actual components. Even Detection Through Testing
if we could model the environment (such as the room It is widely recognized that unit testing or compo-
layouts and thermodynamics), we could not reason nent verification is clearly not sufficient, even in tra-
about components individually but only about the ditional software systems; integration testing and
system as a whole. As feature-interaction pioneer system testing are important too. This observation
Michael Jackson framed it: “the physical world has also holds for systems with ML components, which
no compositionality.” is just another reminder to evaluate the entire sys-
A key insight from the study on feature inter- tem (often in production) and not just the prediction
actions is that, even when systems are complex, accuracy of individual models.

18 ComputingEdge December 2024

 SE FOR AI

Detection Through system design will make it easier for developers or end
Better Specifications users to resolve interactions when found. If we antici-
A long history of research on feature interactions pate that interactions may happen, clever system
has shown that better requirements engineer- designs can automatically select default resolutions,
ing can help to anticipate interactions, for example that is, design the system to handle unknown interac-
through systematic inspection of potential interac- tions gracefully.
tion points or through model checking of combined In several domains, automatic domain-specific
specifications. Even weak specifications can be use- default-resolution mechanisms have been success-
ful to detect problems through inspection, such as ful. For example, in self-driving cars, multiple compo-
goal models and resource models, for example, to nents (such as cruise control, emergency braking, and
analyze whether multiple home-automation compo- maps) may provide possibly conflicting suggestions
nents might compete for electricity or human atten- for the target speed, and, anticipating such conflicts,
tion. Nhlabatsi et al. provide a concise overview of the system can be designed to resolve them by always
common kinds of conflicts and different kinds of picking the lowest (safest) suggested speed. 5 In
interactions that can help to guide an inspection.4 Android, the system is designed to ask the user which
In an ML setting, we may be able to reason to some of multiple apps should open a link if a conflict is
degree about goals, resources, and maybe even weak detected at runtime. It is important to note that these
specifications to detect certain kinds of interac- strategies usually need to be designed for a specific
tions, especially if models interact through the envi- problem, which is their strength and weakness at the
ronment and shared resources. However, given how same time: the system can resort to resolutions that
hard it is to provide even weak specifications of ML leverage domain knowledge, but it is difficult to trans-
components, leveraging system design might be a fer this type of solution to other domains.
more promising solution. Systems with multiple ML components usually
already naturally isolate the models and communicate
Design for Interactions through messages where resolution can be focused.
Likely the most powerful strategy in addressing the Data fusion can be considered as a resolution strategy
feature-interaction problem is to prepare for the pos- that is already common in many system architectures.
sibility of interactions as part of the system design, The fusion strategy can either be defined manually
designing the system to 1) prevent certain interactions (like picking the lowest speed) or learned with another
by isolating components and 2) prepare for resolving ML model. In a sense, the ranking component of our
interactions when they eventually occur. initial image-captioning scenario can be seen as a
Isolation is a common design strategy to shield domain-specific fusion mechanism that combines
components from each other; an example is Android outputs of object detection and language models,
apps that cannot access each other’s internal state or trained on task-specific data. The ranking component
files. However, full isolation is rarely desired as com- is carrying out coordination logic to resolve interac-
ponents should often work together to achieve a goal. tions that has been learned itself! We suspect that
Designs will therefore typically allow specific kinds of there are many opportunities to think very deliber-
interactions but often require the use of permitted ately about communication channels and formats to
and possibly controlled communication channels; the restrict the kind of data exposed from models and,
system can intercept, modify, or block messages at more importantly, data fusion steps to define or even
runtime if it serves the system specification (for exam- learn default resolutions for interactions.
ple, not making phone calls without the corresponding
app permissions).
Once an interaction has been detected, it can be
resolved with additional coordination logic, such as
E ven though the lack of proper specifications may
make ML models appear special compared to
traditional software systems, there are many parallels
one component overwriting the behavior of another. around how to design a system to anticipate interac-
Anticipating the need for resolution as part of the tions with a healthy dose of system thinking. That

www.computer.org/computingedge 19
SE FOR AI

is, we need to focus on system design, not just the interaction: The security threat from within software
design of ML model architectures. systems,” Prog. Informat., no. 5, pp. 75–89, 2008, doi:
The key point is to realize that decomposition 10.2201/NiiPi.2008.5.8. https://www.nii.ac.jp/pi/n5/5
without perfect modularity is okay. Decomposition _75.html
is a best-effort approach, but we need to anticipate 5. C. Bocovich and J. M. Atlee, “Variable-specific resolu-
interactions, prepare for them as part of the sys- tions for feature interactions,” in Proc. 22nd ACM SIG-
tem design and development process, and make SOFT Int. Symp. Found. Softw. Eng., Hong Kong, China,
feature interactions a first-class concern to reason Nov. 2014, pp. 553–563, doi: 10.1145/2635868.2635927.
about them when they occur. We need to embrace 6. M. Mitchell et al., “Model cards for model reporting,”
design methods of managing interactions, and in Proc. Conf. Fairness, Accountability, Transparency,
ML itself may provide a powerful tool for learning Atlanta, GA, USA, Jan. 2019, pp. 220–229.
feature-interaction-resolution strategies. 7. T. Gebru et al., “Datasheets for datasets,” Mar. 23, 2018,
At the same time, it is worth exploring what kind of arXiv:1803.09010v3.
specifications, however partial, we can provide for ML 8. C. Kästner, E. Kang, and S. Apel, “Feature interactions
models. Describing goals of models or assumptions on steroids: On the composition of ML models,” May 13,
made in training data selection or modeling can help 2021, arXiv:210506449K.
us to reason at least partially, about compositions.
The recent adoption of more structured documenta-
tion, such as model cards6 and data sheets,7 can pro- SVEN APEL is the chair of software engineer-
vide inspiration for providing structured, and possibly ing at Saarland University and Saarland Infor-
even machine-readable, information about models. matics Campus, Saarbrücken, 66123, Germany.
An extended version of this article can be found in Contact him at [email protected].
Kästner et al. 2021.8
CHRISTIAN KÄSTNER is an associate profes-
ACKNOWLEDGMENT sor in the School of Computer Science at
Sven Apel’s work has been funded by the German Carnegie Mellon University, Pittsburgh, Penn-
Research Foundation, grant 389792660, as part of sylvania, 15213, USA. Contact him at kaestner
the Transregional Collaborative Research Center @cs.cmu.edu.
248–Center for Perspicuous Computing; see https://
perspicuous-computing.science/. EUNSUK KANG is an assistant professor in
the School of Computer Science at Carnegie
REFERENCES Mellon University, Pittsburgh, Pennsylvania,
1. H. Fang et al., “From captions to visual concepts and 15213, USA. Contact him at eunsukk@andrew
back,” in Proc. IEEE Conf. Comput. Vision Pattern .cmu.edu.
Recognit., 2015, pp. 1473–1482, doi: 10.1109/CVPR.2015
.7298754.
2. B. Nushi, E. Kamar, E. Horvitz, and D. Kossmann, “On
human intellect and machine failures: Troubleshooting
integrative machine learning systems,” in Proc. 31st
AAAI Conf. Artif. Intell., vol. 31, Feb. 2017, pp. 1017–1025.
3. Z. Peng, J. Yang, T.-H. (Peter) Chen, and L. Ma, “A first
look at the integration of machine learning models in
complex autonomous driving systems: A case study
on Apollo,” in Proc. 28th ACM Joint Meeting Eur. Softw.
Eng. Conf. Symp. Found. Softw. Eng., Nov. 2020, pp.
1240–1250, doi: 10.1145/3368089.3417063. WWW.COMPUTER.ORG/COMPUTINGEDGE
4. A. Nhlabatsi, R. Laney, and B. Nuseibeh, “Feature

20 ComputingEdge December 2024

 IEEE CS EXECUTIVE COMMITTEE
President: Jyotika Athavale
President-Elect: Hironori Washizaki
www.computer.org Past President: Nita Patel
First VP: Grace A. Lewis
Second VP: Nils Aschenbruck
PURPOSE: Engaging professionals from all areas of computing,
the IEEE Computer Society sets the standard for education Secretary: Mrinal Karvir
and engagement that fuels global technological advancement. Treasurer: Darren Galpin
Through conferences, publications, and programs, IEEE CS VP, Member & Geographic Activities: Kwabena Boateng
empowers, guides, and shapes the future of its members, and VP, Professional & Educational Activities: Cyril Onwubiko
the greater industry, enabling new opportunities to better serve VP, Publications: Jaideep Vaidya
our world. VP, Standards Activities: Edward Au
OMBUDSMAN: Contact [email protected]. VP, Technical & Conference Activities: Terry Benzel
CHAPTERS: Regular and student chapters worldwide provide 2023–2024 IEEE Division VIII Director: Leila De Floriani
the opportunity to interact with colleagues, hear technical 2024–2025 IEEE Division V Director: Christina M. Schober
experts, and serve the local professional community. 2024 IEEE Division V Director-Elect: Thomas M. Conte

PUBLICATIONS AND ACTIVITIES IEEE CS BOARD OF GOVERNORS

Computer: The flagship publication of the IEEE Computer Term Expiring 2024:
Society, Computer, publishes peer-reviewed technical content that Saurabh Bagchi, Charles (Chuck) Hansen, Carlos E. Jimenez-
covers all aspects of computer science, computer engineering, Gomez, Daniel S. Katz, Shixia Liu, Cyril Onwubiko
technology, and applications. Term Expiring 2025:
Periodicals: The society publishes 12 magazines, 18 journals İlkay Altintaş, Mike Hinchey, Joaquim Jorge, Rick Kazman,
Conference Proceedings & Books: Conference Publishing Carolyn McGregor, Andrew Seely
Services publishes more than 275 titles every year. Term Expiring 2026:
Standards Working Groups: More than 150 groups produce Megha Ben, Terry Benzel, Mrinal Karvir, Andreas Reinhardt,
IEEE standards used throughout the world. Deborah Silver, Yoshiko Yasuda
Technical Communities: TCs provide professional interaction
in more than 30 technical areas and directly influence computer IEEE EXECUTIVE STAFF
engineering conferences and publications. Executive Director and COO: Sophia Muirhead
Conferences/Education: The society holds more than 215 Interim General Counsel and Chief Compliance Officer:
conferences each year and sponsors many educational activities, Ahsaki Benion
including computing science accreditation. Chief Human Resources Officer: Cheri N. Collins Wideman
Certifications: The society offers three software developer Managing Director, IEEE-USA: Russell Harrison
credentials. Chief Marketing Officer: Karen L. Hawkins
Managing Director, Publications: Steven Heffner
AVAILABLE INFORMATION
Staff Executive, Corporate Activities: Donna Hourican
To check membership status, report an address change, or obtain Managing Director, Member and Geographic Activities:
information, contact [email protected]. Cecelia Jankowski
Chief of Staff to the Executive Director: Kelly Lorne
IEEE COMPUTER SOCIETY OFFICES Managing Director, Educational Activities: Jamie Moesch
IEEE Standards Association Managing Director: Alpesh Shah
WASHINGTON, D.C.: LOS ALAMITOS:
2001 L St., Ste. 700, 10662 Los Vaqueros Cir., Chief Financial Officer: Thomas Siegert
Washington, D.C. 20036-4928 Los Alamitos, CA 90720 Chief Information Digital Officer: Jeff Strohschein
Phone: +1 202 371 0101 Phone: +1 714 821 8380 Managing Director, Conferences, Events, and Experiences:
Marie Hunter
Fax: +1 202 728 9614 Email: [email protected]
Managing Director, Technical Activities: Mojdeh Bahar
Email: [email protected]
IEEE OFFICERS
IEEE CS EXECUTIVE STAFF
President & CEO: Thomas M. Coughlin
Executive Director: Melissa Russell
President-Elect: Kathleen Kramer
Director, Governance & Associate Executive Director:
Past President: Saifur Rahman
Anne Marie Kelly
Director & Secretary: Forrest D. Wright
Director, Conference Operations: Silvia Ceballos
Director & Treasurer: Gerardo Barbosa
Director, Information Technology & Services: Sumit Kacker
Director & VP, Publication Services & Products: Sergio
Director, Marketing & Sales: Michelle Tubb
Benedetto
Director, Membership Development: Eric Berkowitz
Director & VP, Educational Activities: Rabab Kreidieh Ward
Director, Periodicals & Special Projects: Robin Baldwin
Director & VP, Membership and Geographic Activities:
Deepak Mathur
Director & President, Standards Association:
James E. Matthews III
Director & VP, Technical Activities: Manfred J. Schindler
Director & President, IEEE-USA: Keith A. Moore
revised 29 October 2024
 EDITOR: Qun
EDITOR: Qun Li,
Li, [email protected]
[email protected]
This article originally
appeared in
DEPARTMENT: EMERGING INTERNET TECHNOLOGIES
DEPARTMENT:
EMERGING INTERNET TECHNOLOGIES vol. 28, no. 2, 2024

Distributed Quantum Machine Learning:

Distributed Quantum Machine
Federated and Model-Parallel Approaches
Learning: Federated and Model-
Jindi Wu , Tianjie Hu , and Qun Li , William & Mary, Williamsburg, VA, 23187, USA
Parallel Approaches
Jindi Wu , Tianjie Hu , and Qun Li , William & Mary, Williamsburg, VA, 23187, USA
In this article, we explore two types of distributed quantum machine learning (DQML)
methodologies: quantum federated learning and quantum model-parallel learning. We
discuss
In the challenges
this article, encountered
we explore two types in of DQML, propose
distributed potential
quantum solutions,
machine and highlight
learning (DQML)
methodologies: quantum federated
future research directions learning
in this rapidly andfield.
evolving quantum model-parallel
Additionally, learning.
we implement twoWe discuss
the challenges
solutions encountered
tailored to the two in DQML,
types propose
of DQML, potential
aiming solutions,
to enhance and highlight
the reliability of thefuture research
directions
computinginprocess.
this rapidly evolving
Our results field.
show theAdditionally,
potential ofwe implement
DQML two solutions
in the current Noisy tailored
to the two types of DQML,
Intermediate-Scale Quantum era.aiming to enhance the reliability of the computing process. Our
results show the potential of DQML in the current Noisy Intermediate-Scale Quantum era.

D
istributed quantum machine learning (DQML) image processing, which require the efficient handling of
is an emerging field that combines quantum large datasets while preserving local privacy. The current
machine learning (QML) with distributed com- cost of training even a modest QML model is quite high.
puting. QML utilizes distinctive quantum mechanics For instance, for a QML circuit of a quantum convolu-
properties, like superposition and entanglement, to tional neural network (CNN) that utilizes eight qubits and
potentially enhance traditional machine learning algo- includes approximately 150 trainable parameters on a
rithms. However, the current stage of quantum com- training set of 500 training instances, it may take approxi-
puting technology, often referred to as the Noisy mately $20,000 to train a model on current quantum
Intermediate-Scale Quantum (NISQ) era, imposes limi- computers. QML and DQML are generally not practical
tations on the size and complexity of QML models at present, but in the future, we believe that they have
implemented with variational quantum circuits (VQCs). the potential to perform better than classical computers.
These constraints can restrict the performance and The DQML approaches, while promising, confront
applicability of QML methods. To address these chal- unique and significant challenges distinct from those
lenges, integrating QML with distributed computing has in classical distributed machine learning. In this article,
emerged as a strategic and forward-looking approach. we focus on two specific areas within the realm of
This hybrid approach aims to overcome the individual DQML: quantum federated learning (QFL) and quan-
limitations of quantum devices by harnessing distributed tum model-parallel learning.
quantum computing’s power to manage and process QFL is a case of distributed learning with data par-
complex tasks across multiple quantum computing allelism.1 In this approach, multiple quantum comput-
nodes, thus amplifying the capabilities of QML models. ing nodes, each with its own local dataset, collaborate
DQML holds promise for a diverse array of appli- to train a shared QML model. Each node processes its
cations, especially those demanding complex com- own data independently, ensuring privacy and security
putations that can utilize the distinct advantages of by not transferring raw data between nodes. Instead,
quantum computing. For instance, DQML can signifi- only model updates are communicated across the net-
cantly enhance molecular simulation and drug discov- work. The updates generated by each computing node
ery by enabling more efficient modeling of molecular are collectively aggregated, effectively synthesizing the
interactions. Furthermore, DQML can be highly benefi- insights learned from distinct local datasets. These
cial in fields such as financial modeling and medical consolidated updates are then redistributed to each
node. This approach effectively combines computa-
1089-7801 © 2024 IEEE
tional power and data from diverse sources, enhancing
Digital Object Identifier 10.1109/MIC.2024.3361288 the learning process while maintaining data confidential-
Date of current version 16 April 2024. ity, a key aspect in scenarios where data privacy is crucial.

22 March/AprilDecember
2024 2024 Published by thePublished by the Society
IEEE Computer IEEE Computer Society  Internet Computing
IEEE 65
2469-7087/24 © 2024 IEEE
EMERGING INTERNET TECHNOLOGIES EMERGING INTERNET TECHNOLOGIES

Quantum model-parallel learning is a method of dis- Qubit

tributed learning characterized by the paradigm of A qubit, which carries quantum information, is the fun-
model parallelism. The model parallelism in quantum damental building block of quantum computing. Due
model-parallel learning is particularly advantageous for to its unique property of superposition, it can exist in
handling large-scale QML models that exceed the com- multiple states simultaneously. The state of a single
putational and memory capacities of individual quan- qubit can be represented as jwi ¼ aj0i þ bj1i with
tum nodes.2 In this framework, the QML model is a, b 2 C. Here, jaj2 and jbj2 represent the probabilities
partitioned into submodels, distributed across multiple of measuring the qubit as j0i and j1i, respectively, i.e.,
quantum computing nodes, with each node handling a jaj2 þ jbj2 ¼ 1: Superposition empowers quantum com-
distinct submodel of the entire model. The computing puters with increased computational power, allowing
nodes process their assigned partition of data and them to explore and process numerous potential solu-
compute intermediate results in parallel, which are then tions to a problem in parallel. Furthermore, entangle-
communicated to other nodes or a central coordinator ment allows two or more qubits to establish
for generating the final outcome. The unique aspect of correlations in which changes in the state of one
quantum model-parallel learning is its ability to leverage entangled qubit instantaneously impact the state of
the individual computational strength of each node while the other, irrespective of the spatial separation
jointly contributing to the construction of a comprehen- between them.
sive model. Through the distribution of computational
workload and the facilitation of parallel processing, Quantum Gates
quantum model-parallel learning opens up new possibil- Quantum computation involves manipulating qubit
ities for addressing more intricate QML tasks. states with quantum gates, represented by unitary
QFL and quantum model-parallel learning each pre- matrices denoted as U (satisfying the conditions
sent unique benefits for specific QML tasks, yet they U † U ¼ UU † ¼ I). The fundamental quantum gates
also face several challenges. In our study, we conduct include the Pauli-X, Pauli-Y, Pauli-Z, Hadamard (H), and
a thorough examination of these challenges and pro- controlled-X (CNOT) gates. These gates serve as foun-
pose potential solutions to mitigate these issues. dational building blocks for assembling more intricate
These challenges include quantum errors, scalability, quantum algorithms. Among these gates, the CNOT
communication, and hardware diversity. Moreover, we gate is a two-qubit gate that establishes correlations
implement two specific solutions to enhance the reli- between qubits. In addition, quantum rotation gates,
such as RX(h), RY(h), and RZ(h), provide precise manipu-
ability of the two types of DQML approaches. In partic-
lation of quantum states through an angle h.
ular, a key challenge in QFL arises from the global
model being susceptible to a wide array of local errors.
Quantum Errors
Because the local models are trained under device-
Quantum computing is error-prone due to the inherent
specific errors, the aggregation of these error-impacted
instability of the quantum system and the immature
local models can significantly compromise the efficacy
manufacturing of quantum computers. A quantum cir-
of the overall QML model. To address this issue, we
cuit (program) comprises a set of gates that manipulate
minimize the impact of errors on local models as much
quantum data, and the processed data are acquired
as possible. For quantum model-parallel learning, devis-
through quantum measurement operations. Each opera-
ing a strategy to partition a large-scale QML model
tion in this process can introduce errors into the quan-
without compromising its functionality and reliability
tum system, potentially resulting in inaccuracies in the
presents a considerable challenge. We address these
quantum circuit’s outcomes. In particular, errors in quan-
issues by carefully designing the submodels of the QML
tum computing arise from various sources, including
model, taking into account both the architecture and
quantum state preparation, quantum gates, measure-
the reliability of the available quantum computing
ment, and crosstalk. Moreover, the state of qubits can be
nodes. Our results demonstrate the significant poten-
influenced by errors due to decoherence and dephasing.
tial of DQML in the current NISQ era.
QML
BACKGROUND VQCs are a widely adopted approach for constructing
Quantum Computing QML models. These models are tailored for executing
Quantum computing is a revolutionary paradigm of particular tasks, including optimization and classification,
computation that leverages the principles of quantum by harnessing the capabilities of quantum circuits and
mechanics to perform complex calculations. integrating them with classical optimization methods.

66
www.computer.org/computingedge
IEEE Internet Computing March/April 2024 23
EMERGING INTERNET TECHNOLOGIES EMERGING INTERNET TECHNOLOGIES

Ansatz 2) Local training: On the i-th quantum computing

A VQC, implementing a QML model, is composed of node, the client trains its individual model U ðhi Þ
three parts: 1) the encoding unit is responsible for con- using the private dataset with several update
verting classical data into quantum data, which can be steps.
processed by the quantum circuit. Popular encoding 3) Local model submission: Each participating
techniques include angle encoding and amplitude client submits their updated local parameters h
encoding. 2) The variational block constitutes the core to the central server.
of the VQC, featuring a sequence of parameterized 4) Model aggregation: The central server aggre-
quantum rotation gates. The rotation gates equipped gates the received local parameters from cli-
with trainable parameters function analogously to neu- ents and integrates them into a global model
P
rons in classical neural networks. The entire variational represented as h0 ¼ N i¼1 ni hi , where N is the
block is organized into layers, and a block consisting of number of clients involved in this procedure,
L layers can be represented as and ni denotes the corresponding weight of the
i-th client.
U ðhÞ ¼ UL ðhL ÞUL�1 ðhL�1 Þ::: U1 ðh1 Þ (1)
5) Global model distribution: The updated global
(3) The measurement unit involves the measure- parameters h0 are distributed to clients for the
ment of one or more qubits to obtain the outcome cor- subsequent round of training.
responding to the input data with observable O: 6) Iteration: Repeat steps 2–5 for multiple rounds
until convergence or the desired model perfor-
Optimization mance is achieved.
QML employs a hybrid quantum–classical approach to
train the model. Once a model on a quantum computer
processes specific input data, the classical computer Quantum Model-Parallel Learning
takes over to optimize the model’s parameters, guided Quantum model parallelism is a distributed approach
by a predefined cost function in QML, especially for large-scale QML models. In this
C ¼ h0jU † ðhÞOUðhÞj0i: (2) method, a complex quantum model is partitioned into
submodels, which are then distributed across multiple
This optimization can be executed through various
quantum computing nodes.
methods, including gradient-based approaches like the
In such a large-scale QML model, the overall unitary
stochastic gradient descent algorithm or non-gradient-
operation U ðhÞ is partitioned into K submodels
based techniques, such as the parameter-shift algo-
fU1 ðh1 Þ, U2 ðh2 Þ, :::, Uk ðhk Þg: Each submodel, denoted as
rithm. The entire training procedure in QML involves
Ui ðhi Þ, is allocated to a distinct quantum computing
repeating the optimization step until the parameters of
node, with i ranging from 1 to K: Similarly, the input
the model converge.
data D are divided into subsets fD1 , D2 , :::, Dn g, cor-
responding to the submodels fU1 ðh1 Þ, U2 ðh2 Þ, :::,
QFL
Un ðhn Þg within the input layer of the QML model, where
QFL is a sophisticated process that blends the princi-
n � K: Starting with the submodels in the input layer,
ples of quantum computing with federated learning’s
each submodel within the same layer operates inde-
distributed model training approach. The QFL system
pendently. The output generated by a submodel is
consists of a central server and multiple quantum com-
then transmitted to the submodel in the subsequent
puting nodes, each allocated to different clients. In this
layer, using either classical or quantum communication
configuration, clients maintain their data locally on
channels. The submodel in the final layer of the QML
their assigned quantum computing nodes. The objec-
model generates the ultimate output that corresponds
tive is to collaboratively train a QML model that bene-
to the specific input data. This process guarantees the
fits from the aggregated data across all nodes, while
seamless flow of information and computation across
ensuring that each client’s private information remains
the distributed quantum system.
unshared and secure.3 Formally, the QFL procedure for
a VQC-based QML model can be represented as
follows: RELATED WORK
The design of the QFL model can be varied to align
1) Initialization: Each client initiates a local model with the specifications of the available quantum devi-
U ðhÞ with the same ansatz and parameters on ces and data. For instance, a hybrid quantum–classical
their respective local devices. transfer learning model is developed and deployed on

24 ComputingEdge
March/April 2024 IEEE Internet Computing 67
December 2024
EMERGING INTERNET TECHNOLOGIES EMERGING INTERNET TECHNOLOGIES

each computing node for federated learning.4 This originate from multiple computing nodes, which can
model utilizes a pretrained classical model to compress lead to suboptimal performance. Additionally, quantum
classical input data into a small size, followed by model-parallel learning produces final outcomes based
leveraging a VQC-based QML model for decision mak- on noisy intermediate results. As the complexity of the
ing. In addition, a QFL approach based on quantum model increases, errors accumulate, consequently dimin-
data is proposed.5 Furthermore, beyond a VQC, the ishing the fidelity of the outcomes.
QFL model can be constructed using various methods. Error-mitigation strategies such as circuit optimiza-
For example, the model may consist of several layers tion and result postprocessing can be utilized to effec-
that incorporate a differing number of qubits.6,7 tively improve the reliability of each computing node.
A scalable QML is an instance of quantum model- For instance, the solution presented in the subsequent
parallel learning.2 The approach divides a large-scale section aims to alleviate the influence of various errors
QML model into two distinct layers. The first layer com- in the overall model. It achieves this by specifically
prises individual subcircuits, each designed to learn reducing the error impact on each computing node
from segments of a training instance. The second layer through the application of circuit-optimization techni-
then aggregates these intermediate results, enabling ques. However, a comprehensive solution for error miti-
further exploration of the correlations between data gation has not yet been fully realized. In addition,
segments. In addition, the approach can be imple- quantum errors continuously change in unpredictable
mented with a single quantum computing node due to ways. Therefore, a model trained to mitigate errors dur-
the independence between the submodels of the large ing the training phase may not be effective against
QML model. Similarly, a quanvolutional neural network errors encountered during testing. A potential solution
achieves scalability by constructing quantum convolu- to address fluctuating quantum errors involves contin-
tional kernels that emulate the functionality of the uously updating the model to accommodate current
classical convolutional kernel used in classical CNNs.8 quantum error conditions, but this method may lead to
This quantum kernel slides over the input data to considerable overhead. Alternatively, the model can be
extract abstract features, mirroring the process in clas- trained while accounting for shifted errors.9
sical CNNs.
Scalability
CHALLENGES Although distributed approaches can enhance the
Although QFL and quantum model-parallel learning scalability of QML tasks to a certain extent, scalability
offer benefits by integrating distributed computing remains a significant challenge in the field. For QFL,
with QML, such as enhanced privacy protection and the scale of the QML model deployed on an individual
improved model scalability, there are still several chal- quantum computer is constrained by the limited quan-
lenges that need to be addressed. tum computing resources. The circuit width, for instance,
is confined by the number of qubits available on the
Quantum Errors quantum hardware. These qubits, serving as the register
Quantum errors present a significant challenge in dis- for data processing, limit the data’s dimensionality that
tributed quantum computing systems as the errors the model can process. To mitigate this, a common strat-
accumulate through noisy operations and vary over egy is to compress data to fit the available qubit capacity.
time in an unpredictable manner. Additionally, various encoding methods have been pro-
During the training process of a QML model, posed to represent data within a limited number of
parameter updates serve a dual purpose: they learn qubits. However, these methods can potentially diminish
from the training dataset and simultaneously capture the utility of data for specific tasks or introduce signifi-
the error pattern to mitigate the impact of quantum cant overhead and errors, posing a tradeoff between
errors. Nevertheless, QML models cannot completely data representation efficiency and the fidelity of the
eliminate the influence of these errors. The error pat- information processed. Moreover, the depth of the circuit
tern captured by a QML model is influenced by both is constrained due to the accumulation of errors and the
the ansatz of the model and the specific quantum instability of the quantum system. Therefore, for QFL
device. When the level of error is sufficiently high, the tasks, the scale and complexity of local models remain
reliability of the model can be significantly compro- limited. This limitation can subsequently constrain the
mised. Furthermore, error patterns vary across different performance and effectiveness of these local models.
quantum devices. Consequently, in QFL approaches, Within quantum model-parallel learning, the scale
the aggregated model suffers from varying errors that of the QML model is dictated by the specifications of

68
www.computer.org/computingedge
IEEE Internet Computing March/April 2024 25
EMERGING INTERNET TECHNOLOGIES EMERGING INTERNET TECHNOLOGIES

available quantum devices and the requirements of cir- in distributed QML. Different quantum systems use
cuit partitioning. The capacity of each individual quan- different physical implementations for qubits, such
tum computing node plays a critical role in determining as trapped ions, superconducting circuits, or topolog-
how the QML model should be partitioned. Excessive ical qubits, each with unique characteristics and
partitioning of the QML model into numerous small seg- limitations. This diversity poses a challenge when
ments can compromise its integrity. Hence, achieving an attempting to create a standardized QML model
optimal balance in the partitioning process is crucial for capable of running on various quantum hardware
the effective functioning of a large-scale QML model. platforms. Furthermore, the interoperability between
Furthermore, as the scale of the QML model diverse quantum systems becomes complex due to
increases, the challenge of the barren plateau emerges.10 differences in their control and measurement proto-
This phenomenon, encountered in the training of large- cols. This necessitates the development of adaptable
scale quantum neural networks, is characterized by QML algorithms that can be efficiently transpiled
the gradient of the cost function becoming extremely and optimized for various hardware architectures.
small. As a result, it becomes inefficient to train the In essence, hardware diversity in distributed QML
QML model. To tackle the barren plateau problem, rec- demands a robust and flexible approach to algorithm
ommended strategies include careful parameter initial- design and system optimization to ensure effective
ization and the adoption of problem-specific ansatzes. and reliable performance across heterogeneous quan-
Nevertheless, there remains a need for more advanced tum computing platforms.
solutions to effectively address this issue.
APPROACHES
Communication In this section, we introduce two strategies designed
The communication channels, both classical and quan- to enhance the reliability of QFL and quantum model-
tum, between quantum computing nodes in a dis- parallel learning by applying an error-mitigation tech-
tributed QML system present challenges. In quantum nique. These techniques primarily involve minimizing
model-parallel learning, the outputs of subcircuits, gate errors by optimizing quantum circuit design. In
obtained through measurements, are transmitted to the particular, we design the QML models to minimize the
central node via classical communication channels. necessity for SWAP gates according to the qubit topol-
However, these measurement results may not entirely ogy on the machines. These gates are inserted into the
represent all the information of the intermediate states circuit to enable the application of multiqubit gates to
as measurements are typically made on a single basis, nonadjacent qubits. Moreover, we prioritize choosing
which can result in the loss of significant information. qubits and gates with lower error rates to further
One potential solution to this issue is the classical reduce the error rate of the QML model.
shadow technique, which involves performing a series of Based on the distinctive characteristics of QFL
measurements on a quantum state and using the results and quantum model-parallel learning, various strate-
to create a “shadow” or a classical approximation of the gies can be employed to design the QML model. In
state. However, the overhead will rise exponentially as the context of QFL, all participating nodes train the
the number of qubits increases. An effective strategy for QML model using the same logical ansatz. Neverthe-
reducing measurement overhead is to reconstruct the less, the transpiled circuits of the model vary and
complete state of a single qubit and fully utilize the infor- capture distinct error patterns, adversely affecting
mation of this single qubit to scale up the size of the the performance of the aggregated model. In contrast,
problem that the model can solve.11 Alternatively, the in the quantum model-parallel learning approach, a
processed quantum states at the nodes could be trans- large-scale QML model is partitioned into several
mitted through quantum communication channels. Yet, submodels, each potentially with a different ansatz.
the development of stable and reliable quantum com- The overall performance of the entire model depends
munication channels, essential for a functional quantum on the reliability of each submodel. Therefore, our
network, is still in its nascent stages. Overcoming these approach to QFL centers around designing the QML
communication hurdles is crucial for the efficient opera- model while taking into account the qubit topology
tion of distributed QML systems. and quality of all the participating computing nodes.
On the contrary, for quantum model-parallel learning,
Hardware Diversity we design submodels of the entire QML model by
Hardware diversity, particularly the various techniques separately considering the qubit topology and quality
used to implement qubits, poses a significant challenge of each node involved.

26 ComputingEdge
March/April 2024 IEEE Internet Computing 69
December 2024
EMERGING INTERNET TECHNOLOGIES EMERGING INTERNET TECHNOLOGIES

QFL ansatz ensures that no noisy SWAP gates are required

In QFL, our strategy begins with a careful selection pro- during the circuit transpilation, thereby preserving
cess, where we carefully select a subset of qubits from the reliability of the circuit. This QML model is then
each participating node. These chosen subsets are deployed on nodes, selecting the highest-quality qubit
characterized by similar qubit connectivity and rela- set that meets this topology requirement. This strategy
tively lower error rates. Subsequently, we construct the not only enhances the fidelity of the outcomes but also
QML model to adapt the qubit connectivity of these minimizes the impact of errors on the trained model.
selected subsets. This method guarantees that the We assess the effectiveness of our proposed approach
transpiled QML models require the minimum number by conducting the training phase on a simulator that
of SWAP gate insertions across all nodes. By doing so, incorporates the noise models of specific quantum
we aim to minimize the impact of errors on each local devices. Subsequently, we test the trained model on
model and, consequently, reduce the disturbance in actual quantum devices. The trained model attained
the aggregated model caused by differing local error accuracies of 93%, 96%, and 96% on the three respec-
patterns. This method provides a more unified and tive nodes involved, as illustrated in Figure 1(d). For
error-resilient approach to QFL, optimizing both indi- a baseline comparison, the accuracies of a model
vidual and aggregated model performance. trained without accounting for the specific architec-
For concreteness, given a QML task focused on a ture of the quantum devices across nodes are 90%,
binary classification task that categorizes images com- 86%, and 88%. The observations indicate that the
posed of four pixels [as shown in Figure 1(a)], we con- model with circuit optimization using our approach
struct a QML model that incorporates four qubits. outperforms the baseline model without circuit optimi-
There are several open source tools available for imple- zation in terms of accuracy across all computing
menting QML. For instance, Qiskit by IBM, PennyLane nodes, highlighting the efficacy of our method. More-
by Xanadu, and TensorFlow Quantum by Google. These over, the lower accuracy achieved by the baseline indi-
tools provide support for QML, but currently, there cates that although the trained QML model can offset
are no tools specifically designed for DQML. In our errors to a certain degree, its effectiveness is still nota-
research, we created a simulated federated learning bly diminished by these errors. This is mainly because
environment utilizing Qiskit on the IBM Quantum the transpiled circuit of the model incurs a large vol-
platform. For this setup, three quantum computers, ume of errors due to the extensive insertion of SWAP
ibm_lagos, ibm_perth, and ibm_nairobi, were selected gates, which are intricately eliminated in our proposed
as individual quantum computing nodes to train local approach.
QML models. These three computers have identical
qubit topologies, as depicted in Figure 1(b). From each Quantum Model-Parallel Learning
computer, we select a subset of qubits within which Typically, the width of a QML model is determined by
three qubits are adjacent to a central qubit, and these the size of the data to be processed. For instance, to
are chosen for their lower error rates, as illustrated in encode a data instance consisting of eight compo-
the green box in Figure 1(b). Then we construct the nents, a QML model would require eight qubits using
QML model as depicted in Figure 1(c). This specific the angle-encoding method. In general, a large-scale

FIGURE 1. QFL with error mitigation. (a) The dataset comprises two classes, with each image containing four pixels. (b) The qubit
topology of quantum computing nodes involved in the QFL environment. (c) The specifically designed QML ansatz, tailored with
the qubit topology to eliminate the need for inserting SWAP gates. (d) The test accuracy of the model on actual quantum devices.

70
www.computer.org/computingedge
IEEE Internet Computing March/April 2024 27
EMERGING INTERNET TECHNOLOGIES EMERGING INTERNET TECHNOLOGIES

FIGURE 2. Quantum model-parallel learning with error mitigation. (a) A QML model is designed with two subcircuits, each con-
structed with consideration of the architecture and reliability of the involved nodes. (b) A QML model is constructed and then par-
titioned based on the number of qubits available on the nodes. (c) The test accuracy of the models implemented with different
approaches.

QML model is first built and then divided into submo- (seven qubits) of one node to process seven compo-
dels according to the capacities of the available quan- nents of the data instance. Meanwhile, the second sub-
tum computing nodes. circuit utilizes two qubits to process the remaining
This approach frequently necessitates inserting component and to integrate the intermediate results,
additional gates and can compromise the correlation as shown in Figure 2(b).
within data instances, leading to a significant reduction In Figure 2(c), we depict the accuracy of two QML
in the model’s reliability. Conversely, our approach models implemented using distinct methods: one
entails designing the submodels while considering with circuit optimization achieving 96% accuracy and
both the qubit topology and the quality of resources another without optimization achieving 69% accuracy.
available on the quantum computing nodes. These It’s clear that the optimized model exhibits significantly
submodels are then integrated into a complete QML superior accuracy compared to the baseline model
model. This strategy aims to preserve the fidelity of the without circuit optimization. There are three main rea-
output of each submodel, thereby enhancing the over- sons for the observed performance difference. First,
all performance of the model. our proposed method for designing subcircuits using
For example, we consider a QML-based classifica- circuit-optimization techniques eliminates the need for
tion task focused on handwritten digits 0 and 1, where noisy SWAP gates, thereby reducing the overall error
images from the Modified National Institute of Stand- rate. Second, data instances are partitioned evenly in
ards and Technology dataset are downscaled to eight our approach. This method preserves coherence within
components using the principal component analysis each data segment and ensures that the final result is
method. The quantum system in this scenario com- evenly influenced by both parts. In contrast, the base-
prises two quantum computing nodes: ibm_lagos and line method partitions the data unevenly, leading to a
ibm_perth, with their qubit topology illustrated in biased final result. Third, although the baseline method
Figure 2. Based on the characteristics of these two can also be optimized to minimize additional SWAP
quantum computing nodes, we design a QML model gates, the first submodel in this approach still includes
consisting of two submodels: the first subcircuit, con- many gates necessary for the model’s functioning. Given
sisting of four qubits, processes the first half of the that the fidelity of a quantum circuit exponentially
image, while the second subcircuit, encompassing five decreases with the increasing number of gates, the
qubits, handles the second half of the image and inte- larger size of the subcircuit in the baseline method likely
grates the processed results from the first half. The results in significantly lower fidelity. Therefore, care-
complete QML model is depicted in Figure 2(a). As a fully partitioning the model across available computing
baseline for comparison, another method implemented nodes is advantageous for maintaining higher fidelity.
for this task involves partitioning the QML model based
solely on the resource capacity of the computing CONCLUSION
nodes. For processing data instances with eight values, In this article, we explore two methodologies within
the first subcircuit is constructed using all the qubits DQML: QFL and quantum model-parallel learning.

28 ComputingEdge
March/April 2024 IEEE Internet Computing 71
December 2024
EMERGING INTERNET TECHNOLOGIES EMERGING INTERNET TECHNOLOGIES

A significant challenge in distributed QML is managing 17th Int. Conf. Mobility, Sens. Netw. (MSN),
quantum errors. To address this, we introduce an error- Piscataway, NJ, USA: IEEE Press, 2021, pp. 145–152,
resilient approach to QFL and quantum model-parallel doi: 10.1109/MSN53354.2021.00035.
learning, employing an ansatz construction based 8. M. Henderson, S. Shakya, S. Pradhan, and T. Cook,
on qubit topology. Empirical evaluations highlight the “Quanvolutional neural networks: Powering image
effectiveness of our solutions, underscoring the poten- recognition with quantum circuits,” Quantum Mach.
tial of distributed QML methodologies. Intell., vol. 2, no. 1, p. 2, 2020, doi: 10.1007/s42484-020-
00012-y.
9. Z. He, B. Peng, Y. Alexeev, and Z. Zhang,
REFERENCES
“Distributionally robust variational quantum
1. C. Ren et al., “Towards quantum federated learning,”
algorithms with shifted noise,” 2023, arXiv:2308.14935.
2023, arXiv:2306.09912.
10. J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush,
2. J. Wu, Z. Tao, and Q. Li, “wpScalable quantum neural
and H. Neven, “Barren plateaus in quantum neural
networks for classification,” in Proc. IEEE Int. Conf.
network training landscapes,” Nature Commun.,
Quantum Comput. Eng. (QCE), Piscataway, NJ, USA:
vol. 9, no. 1, 2018, Art. no. 4812, doi: 10.1038/s41467-018-
IEEE Press, 2022, pp. 38–48, doi: 10.1109/QCE53715.
07090-4.
2022.00022.
11. J. Wu, T. Hu, and Q. Li, “MORE: Measurement and
3. B. McMahan, E. Moore, D. Ramage, S. Hampson, and
correlation based variational quantum circuit for
B. A. Arcas, “Communication-efficient learning of
multi-classification,” in Proc. IEEE Int. Conf. Quantum
deep networks from decentralized data,” in Proc.
Comput. Eng. (QCE), Piscataway, NJ, USA: IEEE Press,
20th Int. Conf. Artif. Intell. Statist., PMLR, 2017,
2023, pp. 208–218, doi: 10.1109/QCE57702.2023.00031.
pp. 1273–1282.
4. S. Y.-C. Chen and S. Yoo, “Federated quantum
machine learning,” Entropy, vol. 23, no. 4, pp. 460–473, JINDI WU is a Ph.D. student focusing on quantum error miti-
2021, doi: 10.3390/e23040460. gation and quantum machine learning in the Department of
5. M. Chehimi and W. Saad, “Quantum federated Computer Science, William & Mary, Williamsburg, VA, 23187,
learning with quantum data,” in Proc. IEEE Int. Conf. USA. Contact her at [email protected].
Acoust., Speech Signal Process. (ICASSP), Piscataway,
NJ, USA: IEEE Press, 2022, pp. 8617–8621, doi: 10.1109/
TIANJIE HU is a Ph.D. student focusing on quantum error
ICASSP43922.2022.9746622.
correction in the Department of Computer Science, William &
6. Q. Xia and Q. Li, “QuantumFed: A federated learning
Mary, Williamsburg, VA, 23187, USA. Contact him at thu04@
framework for collaborative quantum training,” in
wm.edu.
Proc. IEEE Global Commun. Conf. (GLOBECOM),
Piscataway, NJ, USA: IEEE Press, 2021, pp. 1–6,
doi: 10.1109/GLOBECOM46510.2021.9685012. QUN LI is a professor with the Department of Computer Sci-
7. Q. Xia, Z. Tao, and Q. Li, “Defending against byzantine ence, William & Mary, Williamsburg, VA, 23187, USA. Contact
attacks in quantum federated learning,” in Proc. him at [email protected].

72
www.computer.org/computingedge
IEEE Internet Computing March/April 2024 29
 EDITOR: Joseph Williams, Infosys Consulting, [email protected]

This article originally

appeared in
DEPARTMENT: COMPUTING ARCHITECTURES
vol. 56, no. 2, 2023

The Quantum Cybersecurity

Threat May Arrive Sooner Than
You Think
Pete Ford , QuSecure

Recent articles predict that the quantum computing market will expand 500% by 2028.
Billions of dollars are currently pouring into the quantum computer industry to build the
first fault-tolerant quantum computers.

F
ault-tolerant quantum computers (FTQCs) Recall Y2K, when we had to upgrade many of the
will be a gold mine for humanity and business. computing systems in the world to prevent potential
Recent articles predict that the quantum computer errors related to the formatting and storage
computing market will expand 500% by 2028, so we of calendar data for dates in and after the year 2000.
know there’s a quantum race building where quantum Now we have Q-Day or Y2Q, where the entire world
computers could solve some of our most pressing and needs to change the encryption so we can safely use
important problems. As a result, billions of dollars are the Internet.
currently pouring into the quantum computer industry Many ask, “When will a CRQC be available?” Like
to build the first FTQC. However, there is a dark side to searching for gold in a mine, the answer 1 is more
this wonderful invention as quantum computers have opaque than many admit, with views clouded by tech-
been mathematically proven to have the ability to nical jargon and very complex descriptions.
break the cybersecurity that most of the world uses. If you have not heard of Q-Day, you are not alone; it
In this article, I discuss a few ways quantum comput- is the time when FTQCs will be available and can break
ers may be used for theft or disruption sooner than we current asymmetric-key encryption (the encryption
might think. the entire world uses) and become CRQCs. Y2K and
Q-Day have several things in common. They are/were
CRYPTOGRAPHICALLY RELEVANT both global computing efforts, cost a lot of money,
QUANTUM COMPUTERS and have a time constraint. Although we knew about
When a quantum computer is powerful enough to Y2K seven years beforehand, our awareness of Q-Day’s
break our current encryption, most people have no timing is opaque, yet we do know that billions are being
idea how the balance of power in the world could shift. spent to build an FTQC.2
Think about what could happen if an adversarial nation The pending upgrade to postquantum-resilient
state had access to national secrets, financial sys- cybersecurity will be the largest upgrade in informa-
tems, and data, and treasure troves of personal data. tion technology history. Current World Economic
A quantum computer with this power is sometimes Forum predictions 3 indicate that more than 20 bil-
called a cryptographically relevant quantum computer lion devices will need to be upgraded or replaced to
(CRQC), and in an ever-increasing digital world, we can operate safely and successfully in a postquantum
all understand the gravity of such an event. world to keep data safe. Q-Day will arrive, but we
don’t know the exact timing like we did for Y2K. The
first strategic crisis of the 21st century is already
Digital Object Identifier 10.1109/MC.2022.3227657 underway, and thankfully, we are aware of it and can
Date of current version: 8 February 2023 act now.

30 December 2024 Published by the IEEE Computer Society  2469-7087/24 © 2024 IEEE
 The first well-known encryption-breaking proof NISQ problem. Although the time until Q-Day is still
came from Peter Shor in 1994. His algorithm4 showed unknown, these advances should alert us to be ready
that an FTQC with 4,099 coherent qubits could break for a strategic surprise.
the public-key encryption standards we use today. Other quantum scientists9 are exploring tradeoffs
This algorithm and coherent qubit counting have between circuit depth-efficient and qubit-efficient
been a focal point of Q-Day timing. Deep technical models as a mixed approach to encoding schemes that
debates on FTQC gains, daily physics breakthroughs, are easier to deploy on a NISQ machine. Other scien-
and technology funding are riddled with scientific tists are reducing errors in NISQ machines by making
details. Certainly, an FTQC with 4,099 qubits will be them more visible10 and easier to work around. These
one of the big treasures buried in the quantum mines. new techniques and algorithms offer faster paths to
National governments are the main sources paying for solve the nonpolynomial hard math problems protect-
this quantum breakthrough as national security and ing our encryption. This is another threat to breaking
strategic surprise are at stake. India, Russia, Japan, asymmetric encryption, even before an FTQC is ready.
the European Union, and Australia have substantial Studying these developments is important as the race
programs. However, the two main countries invest- to Q-Day speeds up, and the threat could become a
ing in this quantum race5 are China and the United reality sooner than we realize.
States. China is leading by far, spending an estimated
US$13 billion since 2015. U.S. spending during the same
period is estimated at US$2.1 billion. THE PENDING UPGRADE TO
POSTQUANTUM-RESILIENT
NOISY INTERMEDIATE SCALE CYBERSECURITY WILL BE THE
QUANTUM COMPUTERS LARGEST UPGRADE IN INFORMATION
However, there are other ways to get to a CRQC TECHNOLOGY HISTORY.
without needing 4,099 qubits. Noisy intermedi-
ate scale quantum computers (NISQs) and hybrid
classic-quantum computing6 are quietly making These new options using NISQ’s and hybrid com-
strong strides toward solving nonpolynomial hard pilers to break encrypted data shed light on other
math in the near term. As these quantum techniques possible nation-state plans. As competitive nation
advance, we may find quantum-approximate optimiza- states spend many billions of dollars on quantum
tion algorithms7 or new hybrid classic quantum solu- computing, not all the funding in these nation states
tions that can break current encryption before a large will be spent on developing FTQC’s and quantum
FTQC or CRQC is online. As an example, Zapata Com- hardware. Obviously, some funds will be spent to
puting used a heuristic algorithm called variational rapidly decrypt data that have already been stolen.
quantum factoring 8 to show that a quantum hybrid When a well-funded group11 of smart scientists and
solution has promise. This technique of blending clas- programmers pull in the same direction, they can
sical and quantum computing allowed better trad- accomplish surprising results. We should be aware
eoffs between available quantum resources (primarily of others developing these or similar techniques to
quantum circuit depth and solution accuracy). These accomplish devastating results. It is critical for us
smart tradeoffs leveraging readily available quantum to prepare now before a dramatic strategic surprise
computing resources could be a workaround of the makes it too late.

www.computer.org/computingedge 31
COMPUTING ARCHITECTURES

Commercial and government experts are con- Symp. Found. Comput. Sci., 1994, pp. 124–134, doi:
stantly reviewing the traction gained toward a CRQC 10.1109/SFCS.1994.365700.
using current scientific methods. For reluctant 5. S. Rollo, “The quantum tech arms race is on,” Asia Times,
bystanders, the future will be troublesome as it is vola- Mar. 2022. Accessed: Oct. 3, 2022. [Online]. Available:
tile, uncertain, complex, and ambiguous if we rely on https://asiatimes.com/2022/03/the-quantum-tech
rigid certainty for encryption standards. -arms-race-is-on/
There are solutions to the CRQC problem, which 6. M. Swayne, “Quantum AI may need only minimal
can now be implemented. Postquantum cybersecurity data—Proof takes step toward quantum advantage,”
(PQC) refers to a new form of cybersecurity that is Quantum Insider, pp. 2–3, Aug. 2022. [Online]. Available:
resistant to quantum computing attacks from CRQC https://thequantuminsider.com/2022/08/24/quantum
and the nearer-term approaches. Unlike existing -ai-may-need-only-minimal-data-proof-takes-step
encryption that will be broken by powerful quantum -toward-quantum-advantage
computers, PQC uses specific quantum-resilient algo- 7. A. H. Karamlou, W. A. Simon, A. Katabarwa, T. L. Scholten,
rithms12 in combination with other features to ensure B. Peropadre, and Y. Cao, “Analyzing the performance
that enterprise and government organizations will have of variational quantum factoring on a superconducting
protections in place to withstand these attacks, even quantum processor,” 2020, arXiv:2012.07825v1.
before CRQCs arrive. For PQC to be deployed quickly 8. A. Katabarwa and Y. Cao. “Analyzing the performance
and effectively, building in backward-compatible PQC of variational quantum factoring on a superconducting
software solutions that protect our encryption on quantum processor.” Zapata. Accessed: Oct. 8, 2022.
existing hardware allows us to react quickly as quan- [Online]. Available: https://www.zapatacomputing.com
tum computing advances happen. /publications/analyzing-the-performance-of-variational
-quantum-factoring-on-a-superconducting-quantum

D esigning future systems with this resilience, and

not just ruggedness for success, should be our
standard. This will help in the opaque nature of the
-processor/
9. A. Glos, A. Krawiec, and Z. Zimborás, “Space-efficient
binary optimization for variational quantum comput-
race upon us. Time is never on our side as defenders. ing,” npj Quantum Inf., vol. 8, Apr. 2022, Art. no. 39, doi:
FTQCs or nearer-term alternative solutions may come 10.1038/s41534-022-00546-y.
online well before we expect. With a resilient offensive 10. “Researchers propose new quantum error correction
design and mindset, we should embrace crypto agility technique.” HPCwire. Accessed: Oct. 7, 2022. [Online].
as our postquantum information goals are not only Available: https://www.hpcwire.com/off-the-wire
achievable but are a race we can win if we put the right /researchers-propose-new-quantum-error-correction
measures in place. -technique/
11. M. Masiowski, N. Mohr, H. Soller, and M. Zesko, “Quantum
REFERENCES computing funding remains strong, but talent gap raises
1. “QRP3–Quantum race: Sprint, marathon, false start?” concern,” McKinsey, pp. 4–6, Jun. 15, 2022. [Online]. Avail-
Project Q Sydney. Accessed: Oct. 5, 2022. [Online]. able: https://www.mckinsey.com/capabilities/mckinsey
Available: https://projectqsydney.com/qrp3-quantum -digital/our-insights/quantum-computing-funding
-race-sprint-marathon-false-start/ -remains-strong-but-talent-gap-raises-concern
2. “Global quantum computing market.” SkyQuestt. Accessed: 12. “eSecurity planet reports: Best encryption software for
Sep. 28, 2022. [Online]. Available: https://skyquestt.com 2022,” QuSecure, San Mateo, CA, USA, 2022. [Online].
/sample-request/quantum-computing-market Available: https://www.qusecure.com/esecurity
3. “Transitioning to a quantum-secure economy,” World Eco- -planet-reports-best-encryption-software-for-2022/
nomic Forum, Geneva, Switzerland, 2022. [Online]. Available:
https://www3.weforum.org/docs/WEF_Transitioning
20%to_a_Quantum_Secure_Economy_2022.pdf PETE FORD is senior vice president of federal operations at
4. P. W. Shor, “Algorithms for quantum computation: QuSecure, San Mateo, CA 94403 USA. Contact him at pete
Discrete logarithms and factoring,” in Proc. 35th Annu. @qusecure.com.

32 ComputingEdge December 2024

Unlock
Your
Potential
WORLD-CLASS CONFERENCES — Stay
ahead of the curve by attending one of our
195+ globally recognized conferences.

DIGITAL LIBRARY — Easily access over 900k

articles covering world-class peer-reviewed
content in the IEEE Computer Society
Digital Library.

CALLS FOR PAPERS — Discover

opportunities to write and present your
ground-breaking accomplishments.

EDUCATION — Strengthen your resume

with the IEEE Computer Society Course
Catalog and its range of offerings.

ADVANCE YOUR CAREER — Search the

new positions posted in the IEEE Computer
Society Jobs Board.

NETWORK — Make connections that count

by participating in local Region, Section,
and Chapter activities.

Explore membership
today at the IEEE
Computer Society
www.computer.org
 EDITOR: Joanna F. DeFranco, The Pennsylvania State University; [email protected]

This article originally

appeared in
DEPARTMENT: INTERNET OF THINGS
vol. 56, no. 10, 2023

“Sensoring” the Farm

Joanna F. DeFranco , The Pennsylvania State University
Nir Kshetri , University of North Carolina at Greensboro
Jeffrey Voas , IEEE Fellow

Smart farming is gaining attention. But research needs to continue to address the challenges
that come with farming efficiency.

S
eventy-five years ago, The Farmer’s Almanac Agriculture is now a US$1.2 trillion dollar industry.2
was a key predictor for when to plant and reap. In the mid-20th century during the arrival of comput-
Today, it is still used for longer-range planning ers, farming applications began to emerge. Today, the
and forecasting, however, today’s newest weapon for industry continues to advance in this area, where the
“when to do what” on a farm during a planting season Internet of Things (IoT) is the main technology that
is not a book, it’s sensors embedded into the land and facilitates smart and precision farming. Given the
air that provide real-time data such that plans for how labor intensiveness of this industry, it makes sense
to make a crop more successful emerge before and that farmers use computing power to increase preci-
after planting. sion and efficiency.3
Agricultural tools have evolved for thousands of
years. The first plows were made out of forked sticks AGRICULTURAL COMPUTING
to pull through the soil. That was the predecessor to Agricultural computing innovations are assisting the
plows integrating machinery to reduce soil resistance, industry with improving efficiency, increasing crop
and now, modern-day plows utilize GPS systems, yield, and preserving natural resources. According
allowing for precise operations. In addition to innova- to the statistics portal Statista, the IoT was the most
tive machinery, farmers today have access to many influential agricultural technology (AgTech) innova-
computing developments that assist with increased tion of 2022 and had a higher impact on the agricul-
efficiency and improvement of agricultural productiv- tural market than any other AgTechs.4 The primary
ity, precision, and predictability. computing technologies related to the IoT are under
In the mid-19th century, the agriculture industry the umbrella of smart farming, which includes pre-
began educating farmers as the United States needed cision farming technologies. Recently, agricultural
to develop reliable food supplies for the betterment researchers have created systems and frameworks to
of society. Accordingly, baccalaureate degrees in help farmers efficiently connect the numerous sensor
agriculture were launched. The first American school nodes and devices that monitor soil, water, and envi-
to offer baccalaureate degrees in agriculture was ronmental data as well as control the flow of water and
The Farmer’s High School of Pennsylvania, now fertilizer to crops.5,6 Other researchers have created
known as The Pennsylvania State University. Its first
graduating class in 1861 had 13 students. In 2020,
37,721 agricultural degrees were awarded across the
United States.1 DISCLAIMER
The authors are completely responsible for the
content in this message. The opinions expressed
Digital Object Identifier 10.1109/MC.2023.3296762
here are their own.
Date of current version: 20 September 2023

34 December 2024 Published by the IEEE Computer Society  2469-7087/24 © 2024 IEEE
 TABLE 1. Agricultural IoT technologies.

Technology Examples
Livestock monitoring Livestock feeding is optimized using sensors to track behavior, health, and activity.
Soil monitoring Moisture, nutrient, and temperature data are used to determine fertilization and watering needs.
Irrigation management and automation Water usage is optimized.

Agricultural robots for crop harvesting Picking and harvesting tasks are automated; aerial crop assessments are provided.

Crop monitoring (yield and health) Using sensors, images, and artificial intelligence, these systems can detect signs of disease,
pests, and harvest to optimize and predict yield. Sensors are used to collect in real time and
use historical data to predict the quality of soil and crop health.

Equipment monitoring and maintenance Sensors are used to monitor machine health, performance, fuel, and maintenance.

Security Intrusion and theft monitoring/alert services are provided.

systems to optimize the use of natural resources such

as water and solar power.7 Some of the key examples THE PRIMARY COMPUTING
of the IoT’s uses in agriculture are listed in Table 1. TECHNOLOGIES RELATED TO THE
IOT ARE UNDER THE UMBRELLA OF
LIVESTOCK MONITORING SMART FARMING, WHICH INCLUDES
South African multinational mobile telecommunica- PRECISION FARMING TECHNOLOGIES.
tions company MTN, which operates in many African
and Asian countries, has partnered with Aotoso Tech-
nology to provide IoT-based connected collars for cattle that the soil is dry. The solutions are cost efficient and
in the Sudanese market. Farmers use Subscriber Iden- thus accessible for smallholder farmers.10
tity Module cards on the collar and on their cell phones
to get vital information about the cattle. The informa- IRRIGATION MANAGEMENT
tion helps them develop feeding and breeding strate- AND AUTOMATION
gies. The device also monitors reproduction and lacta- IoT-based irrigation systems are helping improve the
tion, which can help farmers increase their income.8 availability of water for farmers. For example, Nai-
robi, Kenya-based solar irrigation company SunCul-
SOIL MONITORING ture provides smallholder farmers irrigation and solar
The IoT can help maintain appropriate soil moisture pumping solutions.11 Using SunCulture’s off-grid tech-
conditions and nutrient availability, which can help nology, farmers can extract up to 3,000 L of water
increase the effect of fertilizers. As an example, Kenya’s per hour from wells up to 70-m deep.12 The system
IoT-based smart irrigation system, Illuminum Green- also has built-in algorithms, which study the weather
houses, are powered by solar panels and sensors, to optimize performance. Based on the predicted
which work together to create an optimal environment weather patterns, SunCulture’s smart machines send
for crops.9 Farmers can use their mobile phones to phone and text alerts to farmers about the appropri-
control key indicators such as temperature, humidity, ate irrigation timing.13 If too much water is pumped
and soil moisture. An automated watering system sup- out of the ground, pressure in the aquifer gradually
plies a precise amount of water if the sensors detect decreases. To ensure that the water supply does not

www.computer.org/computingedge 35
INTERNET OF THINGS

dry out, other IoT solutions can be used to evaluate was being used by 7,200 farmers in the United States,
the optimal amount of water to be pumped. As noted South Africa, Australia, and New Zealand.17
previously, farmers, for instance, can use sensors
to detect soil moisture and pump water only when EQUIPMENT MONITORING
needed. IoT-based irrigation systems can thus mini- AND MAINTENANCE
mize water consumption. In March 2023, Indian multinational conglomerate
Mahindra’s AgTech arm Krish-e launched IoT-based
AGRICULTURAL ROBOTS Krish-e Smart Kit (KSK), which provides equipment
To assist with helping with labor shortages and owners with detailed insights about their farm equip-
increasing population, robotics has been integrated ment, such as tractors harvesters and rice transplant-
into the agricultural domain in many areas to reduce ers. The KSK uses GPS to track and remotely monitor
some of the heavy lifting, such as picking and gath- various parameters using a smartphone or desktop.18
ering fruit, harvesting vegetables, and dealing with The kit can be installed on any brand of equipment,
weeds. In addition, drones, considered flying robots, which can help improve fleet performance, reduce
provide aerial images that assist farmers with quick equipment downtime, prevent unauthorized usage,
assessments of crop health. Other uses of robots in and reduce maintenance costs.19
farming are greenhouses (mentioned earlier), which
SECURITY
IoT-based solutions can also help prevent theft of live-
AUTONOMOUS ROBOTS HAVE A 98% stock and agricultural produce. The U.K.-based IoT
ACCURACY WHEN PICKING FRUIT BUT
solutions provider Smarter Technologies Group’s cat-
DO REQUIRE THE SUPERVISION OF
tle collars send out a signal to a livestock management
ONE INDIVIDUAL.
dashboard every 15 min. This enables farmers to gain
remote visibility of their livestock. Cattle owners can
can provide more vegetables to urban areas.14 also set up specific geofences and alerts, which notify
Tortuga AgTech is a farming robot system. In addi- them when a livestock moves outside a designated area.
tion to the labor-intensive tasks, it provides a way to The company’s smart fence gate alarms can inform cat-
address challenges such as damage caused by human tle owners when there are unauthorized cattle move-
hands during the harvesting process. Autonomous ments.20 Likewise, MTN’s IoT-based connected collars,
robots have a 98% accuracy when picking fruit but discussed previously, can also help prevent cattle theft.
do require the supervision of one individual.15 iRo-
bot, another farming system, has also entered the AGRICULTURE TECHNOLOGY AS A
harvest-automation arena. In addition to picking and SERVICE
harvesting, this technology can also continue running The agriculture technology-as-a-service (ATaaS) mar-
in extremely hot weather.16 ket, valued at more than US$3.4 billion, is driven by
small-scale farmers and their growing demand for pre-
CROP MONITORING cision agriculture.21 Smaller-scale farmers can use
Vodacom’s MyFarmWeb has tools that collect and ana- ATaaS to enhance productivity and efficiency. The
lyze data from multiple IoT sensors across a farm, which industry is predicted to reach US$4.93 billion by 2028.22
can help to improve productivity and optimize farm- The IoT’s potential to transform farming practices has
ing practices. For instance, its ITESTLeaf feature can been one of the key drivers of the ATaaS market.23
be used to view and compare leaf tissue based on his- ATaaS can
torical data. Its precision pest-monitoring feature pro-
vides a historical analysis of pests and trends. Likewise, › streamline operations and increase efficiency
the MyYield feature can help analyze areas with vari- using crop monitoring, livestock monitoring, soil
ous levels of yields and identify the underlying causes analysis, irrigation, crop yield, and autonomous
of areas with low yields. As of early 2022, MyFarmWeb farming machinery

36 ComputingEdge December 2024

 INTERNET OF THINGS

› provide real-time information about crops to public to source some of their food. For example, farm-
make informed decisions about planting, fertil- ers in India are also using IoT systems to increase food
izing, and pest control. production in urban areas.28 Greenhouse technology
is used as well as innovative solutions such as vertical
To achieve these goals, ATaaS providers integrate farms (growing crops on top of each other), hydropon-
advanced sensors and other technologies such as ics (growing plants without the use of soil), aeroponics
remote sensing and drones in their solutions. Farm- (growing plants with only water and nutrients), and
ers have access to a wide range of indicators that the aquaponics (growing plants and fish together in the
sensors collect and measure, which can be used for same environment), are being utilized along with sen-
accurate crop monitoring, and take proactive actions sors, data collection, and a device to monitor.
that can help them generate the highest revenue. For As the world’s population increases, so too does
instance, farmers can detect potential problems such the need for food. However, tillable land is being turned
as disease, pest infestations, and soil degradation, into data centers and wind farms, while other tillable
which can help with timely interventions.24 land is becoming untillable due to climate change
and political conflicts. AgTech is an exciting and
AGRICULTURE COMPUTING ever-expanding science with the potential to address
CHALLENGES these challenges. Computing plays a large role, not
In general, smart farming is environmentally friendly only from sensors and the IoT, but from computing’s
given the improvements discussed earlier. But soci- role in creating new fertilizers, pesticides, and seeds.
ety needs to consider the challenges that come along Many of the systems described in this article are
with changes. For example, some research states that used by farmers to make informed decisions to optimize
agricultural efficiency reduced expenses of water and the use of things like fertilizer and pesticide use and
pesticides, but some claim that the positive outcomes natural resources (water/fuel) to minimize environmen-
of more efficient use of machinery and irrigation sys- tal impact. For example, water delivery can be adjusted
tems may result in higher use of pesticides and fertil- based on real-time weather sensor data. Controlling
izers, which can also be a result of climate change.25 It water usage not only preserves a natural resource but
is a fact that climate change has increased pests that also reduces costs and improves crop yields.
destroy crops. The point is, the entire system needs to Data collection (from different sensors) also can
be considered, even when improvements are made. assist in predicting crop yield, reducing equipment
To grasp this statement, one can use the analysis downtime (watching equipment health), and, overall,
approach of systems thinking. Specifically, one of the improves efficiency in many aspects of farming.
systems thinking laws by Peter Senge is, “Today’s prob- Therefore, smarter farming is just smarter, and this
lems come from yesterday’s solution.” For example, topic will be gaining more attention in the immediate
Brazil is the largest coffee producer but also the largest years.
pesticide consumer.26 The pesticides used are effective
for coffee production, but not good for the ecosystem REFERENCES
and health of the system in which it resides (this includes 1. “Agriculture.” DATAUSA. Accessed: Jul. 14, 2023. [Online].
people). The pesticides used have contaminated some Available: https://datausa.io/profile/cip/agriculture
water supplies and been linked to respiratory problems, 2. “Ag and food sectors and the economy,” U.S. Dept. Agri-
high blood pressure, cancer, and cardiovascular dis- culture, Washington, DC, USA, Jan. 26, 2023. Accessed:
ease. Sustainable alternatives could be considered. Jul. 14, 2023. [Online]. Available: https://www.ers.usda.gov
Agroforestry, the practice of integrating trees, shrubs, /data-products/ag-and-food-statistics-charting-the
and livestock, has been known to create benefits.27 -essentials/ag-and-food-sectors-and-the-economy
/#:~:text=Agriculture%2C%20food%2C%20and%20

L ike the victory gardens encouraged in the United

States after World War I and World War II to meet
population demands, urban planners are looking to the
related%20industries%20contributed%20roughly
%20%241.264,of%20this%20sum%E2%80%94
about%200.7%20percent%20of%20U.S.%20GDP

www.computer.org/computingedge 37
INTERNET OF THINGS

3. “Precision farming vs. digital vs. smart farming: What’s 13. E. Mungai, “AI is a game changer for small companies,”
the difference?” DTN, Burnsville, MN, USA, Mar. 2021. Africa Sustainability Matters, pp. 1–8, Feb. 2021.
Accessed: Jul. 14, 2023. [Online]. Available: https:// Accessed: Jul. 14, 2023. [Online]. Available: https://
www.dtn.com/precision-farming-vs-digital-farming-vs africasustainabilitymatters.com/ai-is-a-game-changer
-smart-farming-whats-the-difference/ -for-small-companies
4. “Smart agriculture 2022.” Statista. Accessed: Jul. 14, 14. S. Gossett, “16 agricultural robots and farm robots you
2023. [Online]. Available: https://baba-blog.com/smart should know,” Builtin, Mar. 7, 2023. Accessed: Jul. 14,
-agriculture/ 2023. [Online]. Available: https://builtin.com/robotics
5. A. K. Nalendra, D. Wahvudi, M. Mujiono, M. N. Fu’ad, /farming-agricultural-robots
and N. Kholila, “IoT-Agri: IoT-based environment control 15. “Tortuga AgTech.” Builtin. Accessed: Jul. 14, 2023. [Online].
and monitoring system for agriculture,” in Proc. 7th Int. Available: https://builtin.com/company/tortuga-agtech
Conf. Informat. Comput. (ICIC), Denpasar, Indonesia, 16. “Harvest automation.” Builtin. Accessed: Jul. 14, 2023.
2022, pp. 1–6, doi: 10.1109/ICIC56845.2022.10006964. [Online]. Available: https://builtin.com/company
6. S. K. Sah Tyagi, A. Mukherjee, S. R. Pokhrel, and K. K. /harvest-automation
Hiran, “An intelligent and optimal resource allocation 17. S. Brooks, “Vodafone brings digital transformation
approach in sensor networks for smart Agri-IoT,” IEEE to farming in Europe,” Enterprise Times, Feb. 2022.
Sensors J., vol. 21, no. 16, pp. 17,439–17,446, Aug. 2021, Accessed: Jul. 14, 2023. [Online]. Available: https://
doi: 10.1109/JSEN.2020.3020889. www.enterprisetimes.co.uk/2022/02/25/vodafone
7. V. V. Reddy S, B. Jaison, A. Balaji, D. Indumathy, S. -brings-digital-transformation-to-farming-in-europe/
Vanaja, and J. J. Jeya Sheela, “Agri-IoT: A farm monitor- 18. “Krish-e launches IoT-based Smart Kit for farm
ing and automation system using Internet of Things,” in equipment monitoring,” Agriculture Post, Mar. 2023.
Proc. 2nd Int. Conf. Electron. Renewable Syst. (ICEARS), Accessed: Jul. 14, 2023. [Online]. Available: https://
Tuticorin, India, 2023, pp. 639–642, doi: 10.1109 agriculturepost.com/agritech/krish-e-launches
/ICEARS56392.2023.10085235. -iot-based-smart-kit-for-farm-equipment-monitoring/
8. “Leading digital solutions for Africa’s progress,” 19. “Krish-e launches IoT based Smart Kit for farm equip-
MTN Group Ltd., Fairland, South Africa, Dec. 2020. ment,” Mahindra, Mar. 2023. Accessed: Jul. 14, 2023.
Accessed: Jul. 14, 2023. [Online]. Available: https:// [Online]. Available: https://www.mahindra.com
www.mtn.com/wp-content/uploads/2023/03/2020 /news-room/press-release/en/krishe-launches
-UN-Global-Compact-Report.pdf -iot-based-smart-kit-for-farm-equipment
9. M. Mutiga, “Kenya’s smart greenhouse texts when 20. B. Wingrave, “The smart way to combat livestock
your tomatoes need watering,” The Guardian, Jan. 2016. theft.” Smarter Technologies. Accessed: Jul. 14, 2023.
Accessed: Jul. 14, 2023. [Online]. Available: https://www [Online]. Available: https://smartertechnologies.com
.theguardian.com/global-development/2016/jan/05 /blog/livestock-theft/
/kenya-smart-greenhouse-tomatoes-watering-farming 21. “Agriculture technology-as-a-service global market
10. J. Stewart, “Challenges surrounding IoT deployment in report 2023: Decreasing agriculture workforce drives
Africa,” Compare the Cloud, 2019. Accessed: Jul. 14, 2023. demand,” GlobeNewswire, Mar. 2023. Accessed: Jul. 14,
[Online]. Available: https://www.comparethecloud.net 2023. [Online]. Available: https://www.globenewswire
/articles/challenges-surrounding-iot-deployment-in-africa/ .com/en/news-release/2023/03/28/2635448/28124/en
11. “SunCulture wins project of the year.” EEP Africa. /Agriculture-Technology-as-a-Service-Global-Market
Accessed: Jul. 14, 2023. [Online]. Available: https: -Report-2023-Decreasing-Agriculture-Workforce
//eepafrica.org/sunculture-wins-project-of-the-year -Drives-Demand.html
/#:~:text=SunCulture%20is%20based%20in%20Nairobi 22. Emergen Research, “Agriculture technology as a
12. “The 10 most innovative European, Middle Eastern, and service market to witness explosive growth by 2028 |
African companies of 2021,” Fast Company, Mar. 2021. IBM Corporation, Accenture plc, Trimble,” News Wires,
Accessed: Jul. 14, 2023. [Online]. Available: https:// Apr. 2022. Accessed: Jul. 14, 2023. [Online]. Available:
www.fastcompany.com/90600369/europe-middle https://www.einnews.com/pr_news/566974142
-east-africa-most-innovative-companies-2021 /agriculture-technology-as-a-service-market-to

38 ComputingEdge December 2024

 INTERNET OF THINGS

-witness-explosive-growth-by-2028-ibm-corporation 2017. Accessed: Jul. 14, 2023. [Online]. Available:

-accenture-plc-trimble https://www.fs.usda.gov/nac/assets/documents
23. “Global agriculture technology-as-a-service market.” /agroforestrynotes/an35g09.pdf
Spherical Insights. Accessed: Jul. 14, 2023. [Online]. 28. B. Anuradha, R. Pradeep, E. Ahino, A. Dhanabal, R. J.
Available: https://www.sphericalinsights.com/reports Gokul, and S. Lingeshwaran, “Vertical farming algorithm
/agriculture-technology-as-a-service-market using hydroponics for smart agriculture,” in Proc.
24. “Agriculture technology-as-a-service market - A global Int. Conf. Intell. Syst. Commun., IoT Secur. (ICISCoIS),
and regional analysis: Focus on product, application, Coimbatore, India, 2023, pp. 432–437, doi: 10.1109/
and country analysis - Analysis and forecast, 2022-2027,” ICISCoIS56541.2023.10100527.
BIS Res., 2023. Accessed: Jul. 14, 2023. [Online]. Avail-
able: https://bisresearch.com/industry-report
/agriculture-technology-service-market-report.html JOANNA F. DeFRANCO is an associate professor of software
25. “What is smart farming, and what are its advantages engineering, associate director of the D.Eng. in Engineering
and disadvantages?” Hurley Farms. Accessed: Jul. 14, program at The Pennsylvania State University, Malvern, PA
2023. [Online]. Available: https://www.visithurleyfarms 19355 USA, and an associate editor in chief of Computer.
.com/what-is-smart-farming-and-what-are-its Contact her at [email protected].
-advantages-and-disadvantages/
26. University of Copenhagen, “Unsustainable coffee NIR KSHETRI is a professor of management in the Bryan
production is making more and more people sick, says School of Business and Economics, University of North
study,” Phys.Org, Jun. 2023. Accessed: Jul. 14, 2023. Carolina at Greensboro, Greensboro, NC 27412 USA, and the
[Online]. Available: https://phys.org/news/2023-06 “Computing’s Economics” column editor for Computer. Con-
-unsustainable-coffee-production-people-sick.html tact him at [email protected]
27. M. Vaughan, N. Adamson, and K. MacFarland, “Using
agroforestry practices to reduce pesticide risks to JEFFREY VOAS, Gaithersburg, MD 20899 USA, is the editor
pollinators and other agriculturally beneficial insects,” in chief of Computer. He is a Fellow of IEEE. Contact him at
U.S. Dept. Agriculture, Washington, DC, USA, Jun. [email protected].

IEEE COMPUTER SOCIETY

Call for Papers
Enhance the credibility
and prestige of your research
by publishing with
a globally recognized and
respected organization.

GET PUBLISHED
www.computer.org/cfp

www.computer.org/computingedge 39
general-cfp-hHoriz-Sept24.indd 3 9/30/24 6:22 PM
 EDITOR: Hsiao-Ying Lin, IEEE Member, [email protected]

This article originally

appeared in
DEPARTMENT: ARTIFICIAL INTELLIGENCE/
MACHINE LEARNING vol. 56, no. 3, 2023

AI for Water
Feras A. Batarseh , Virginia Tech
Ajay Kulkarni, Commonwealth Cyber Initiative, Virginia Tech

Water availability is a prerequisite for flourishing economies, protecting public health,

and national prosperity. An overview of the potentials of artificial intelligence for water
is presented.

I
n the United States, there are about 153,000 public STORIES OF INTEREST
drinking water systems1 and more than 16,000 Critical infrastructures, such as smart grids, nuclear
publicly owned wastewater treatment plants.2 plants, medical monitoring systems, smart farms,
While systems control and data acquisition (SCADA) and intelligent water systems (IWSs), are deemed
has become a standard in large water treatment and obsolete (and dangerous to human life) without com-
distribution plants,3 a very small (unknown) percent- prehensive measures to protect them and secure
age of those plants have cyber defenses in place.4 their outcomes. 9 The workings of these systems
are usually governed by laws and policies as well
MOTIVATION as domain-specific best practices.10 However, the
SCADA infrastructures, as evidence shows, 5 are very challenge of securing those cyberphysical systems
prone and vulnerable to cyberthreats. Additionally, is exacerbated by the dependency of their cyber,
there has been a rising number of attacks, especially human, biological, and physical components on each
on water plants, in the United States and around the other. For instance, a programmable logic controller
world. The traditional cybersecurity approach to uti- controlling the pH levels of water at a treatment plant
lizing firewalls and double authentication is bene- is an example of a cyber and bio interdependency;
ficial and can mitigate multiple forms of cyberat- smart sensors reading water flow in municipal water
tacks; however, more sophisticated attacks, such as pipes is an example of physical and cyber dependen-
data poisoning, data manipulation, minimum pertur- cies; and so on.11 Wastewater treatment plants dump
bations, concealed attacks,6 info stealer, botnets, treated water (effluent) into rivers all over the coun-
and ransomware, require algorithms that can detect try, while U.S. Environmental Protection Agency pol-
unusual activities (that is, outlier events)7; classify icies dictate acceptable amounts of phosphorous
the source of adversarial actions; and perform attack and nitrogen. For instance, an unwarranted modifica-
mitigation activities. The current state of the art pro- tion of that value as being measured by sensors at a
vides evidence that artificial intelligence (AI) is the wastewater treatment facility can cause severe envi-
leading approach to such defenses 8 due to its ability ronmental damage to rivers and lakes due to excess
to adequately identify unwarranted pattern shifts in amounts of such chemicals—that is not hypothet-
networks and datasets, a feature that is not achiev- ical; recent motivation examples of applying AI for
able using traditional approaches. water are presented.

Data poisoning and water poisoning

Digital Object Identifier 10.1109/MC.2022.3231142 In the last two decades, U.S. water systems have
Date of current version: 8 March 2023 been exposed to different cyberthreats.12 In 2015, the

40 December 2024 Published by the IEEE Computer Society  2469-7087/24 © 2024 IEEE
 U.S. Department of Homeland Security responded to droughts) have affected the range of decisions start-
25 cyber incidents related to the water sector, repre- ing at the farm and reaching the household consump-
senting a 78.6% increase in the number of reported tion of certain goods.
cases since 20142 —one of the highest rate increases
among all sectors. Cyberattacks have continued to Water Security
grow exponentially relative to the attention given We suggest that the status quo definition of water
to the sector by governments, operators, practitio- security ought to expand from merely covering water
ners, and academics—deeming the sector unsafe.13 “availability” to including cyber-, physical, and biose-
In February 2021, a plant operator in Oldsmar, FL, curity aspects.17 Conventional definitions don’t sat-
saw his cursor being moved around on a computer isfy the rising need to cover all areas of how water
screen, starting various software functions that can be secure. The United Nations provides a work-
control the water being treated. While the opera- ing definition for water security as follows11: “Water
tor assumed that it was another employee access- security is defined here as the capacity of a popu-
ing the system remotely, the intruder boosted (that lation to safeguard sustainable access to adequate
is, data poisoning) the value of one data point— quantities of acceptable quality water for sustain-
the sodium hydroxide (lye) amount—by 100 times. ing livelihoods.” Based on the three-pillared chal-
Sodium hydroxide, the main ingredient in clean- lenge involved in the sector (cyber, bio, and phys-
ing liquids, is added to treat the acidity of water and ical), we propose the following new definition for
remove metals from water for purposes of human water security encompassing emerging trends and
consumption (that is, drinking). Increasing its conventional challenges related to water availabil-
amount to higher levels can cause poisoning, burns, ity and quality:
and multiple other health risks. 4
“Water security is the capacity of nations to safe-
Water, farmers, and irrigation guard the quality and availability of water for all
The Brookings Institution reports an average of 30 desired purposes of society, which includes ensuring
farmers committing suicide per day in India due to measures related to securing access; valid treatment
water-related events.14 Albeit this is a fluctuating processes; cyber hygiene; and mitigating risks associ-
rate, the cause of farmers’ uncertainty is increasingly ated with environmental factors, data collection,
due to water access, extreme weather events, rain- biological threats, and emerging technologies.”
fall, the lack of water, inaccurate forecasts of water
levels in rivers and lakes,15 and unwarranted chem- While we understand that this definition is debat-
icals in water used for agricultural irrigation. Such able, the goal of presenting such a definition is to urge
events affect agricultural yield, the quality of crops, scientists and practitioners in the water sector to
crop disease, and even the prices of agricultural com- consider the novel aspects.
modities around the world.16 As is known, nimble agri-
culture and food security is vital to human survival. AREAS OF APPLICATION
In the past three years, global agriculture has been Water has an obvious effect on economies besides
negatively affected by many water-related shocks. farming and drinking water. Societies flourish next
Unprecedented uncertainties (such as floods and to water bodies18; sanitation and health19 are not

www.computer.org/computingedge 41
ARTIFICIAL INTELLIGENCE/MACHINE LEARNING

manageable or possible without access to water; and design of monitoring and control networks, and the
manufacturing is heavily dependent on reliable water management of tunnels and pipes during extreme
sources. It is rather difficult to capture all potential AI weather events. 24
applications in the water sector; here, a brief review Optimization through AI is performed via mul-
of three examples is elaborated. tiple approaches, but the most common ones include
genetic algorithms (GAs), deep learning (DL), and
Water treatment and management reinforcement learning (RL). For instance, if a water
As AI becomes more developed and deployed fur- treatment plant aims to minimize nitrogen in the efflu-
ther across critical infrastructure, operators will be ent, then a DL or GA optimization approach could be
blindsided if they rely only on their past experience utilized. RL algorithms can, for instance, reinforce
or expertise when making decisions (such as deciding practices that lead to better water quality or increase
on the number of pumps to operate during a storm, water volume (gallons) processing per day.
assessing the performance of the adsorption pro-
cess, and evaluating water quality). Future leaders Agricultural irrigation and farming
Essential crops (such as wheat, corn, and soybeans),
livestock, fruit, and all agricultural commodities
“WATER SECURITY IS THE CAPACITY require water to survive. However, access to water is
OF NATIONS TO SAFEGUARD THE not always guaranteed; for instance, in drought-prone
QUALITY AND AVAILABILITY OF areas, smart irrigation is a critical strategy due to
WATER FOR ALL DESIRED PURPOSES the scarcity of water.25 Urban and vertical farming
OF SOCIETY, WHICH INCLUDES
are similar; in those scenarios, precision irrigation is
ENSURING MEASURES RELATED TO
important to maintain farm finances and create profit
SECURING ACCESS; VALID TREATMENT
for the farmers.16 Such AI-driven decision-making
PROCESSES; CYBER HYGIENE; AND
processes (for example, crop yield prediction, live-
MITIGATING RISKS ASSOCIATED WITH
stock price forecasting, and optimizing the right mix
ENVIRONMENTAL FACTORS, DATA
of biodegradable pesticides) are heavily dependent
COLLECTION, BIOLOGICAL THREATS,
on big data. Data, however, are prone to poisoning,
AND EMERGING TECHNOLOGIES.”
validation issues, and incorrectness. Poisoned data
can change farming recommendations; manipulate
need to possess a fundamental knowledge of AI to smart-irrigation systems’ outcomes; and compromise
better lead and protect their institutions. 20 A mod- water meters, humidity sensors, water pumps, and
ern water treatment plant, referred to as an IWS, has other agricultural control devices.
hundreds of sensors and actuators—for pipes, tanks,
reservoirs, and pumps. Such equipment has inter- Water economics and policy
and intradependencies that increase the complex- Water and the environment are difficult spaces to
ity of detecting a breach.21 Accordingly, AI can be regulate, mainly due to the shared nature of their
used at wastewater treatment plants for multiple use resources (that is, the tragedy of the commons). Poli-
cases in decision making. 22 Decision-making support cymakers need to refer to experts to understand the
involves optimization techniques to allocate an opti- domain and create reasonable policies that can gov-
mal combination of factors that maximize/minimize ern the space fairly. Park et al.,26 for instance, utilized
a numerical objective function (that is, the factor AI, Shapley Additive exPlanations (SHAP), and partial
affected by the decision). From the onset of hydraulic dependence plots for identifying important variables
modeling, optimization techniques have been critical that affect algal bloom in rivers: generally, a challeng-
to water distribution networks. 23 Areas of applica- ing (and potentially subjective) aspect to measure.
tion that can benefit from AI optimization algorithms One of the theories in computational quantification is
include: optimizing energy consumption, the number referred to as value loading.27 It presents matters with
of pumps used at a certain point in time, the optimal a subjective concept, such as policy evaluation—of

42 ComputingEdge December 2024

 ARTIFICIAL INTELLIGENCE/MACHINE LEARNING

the Clean Water Act 33 U.S.C. §1251 et seq. (1972), for the water industry to investigate and further apply
instance—that could be defined mathematically and AI for decision making while considering assurances
measured in a more empirical manner. (such as security) for deriving actionable intelligence.
However, we argue the following for the application Besides cybersecurity, explainability, and correct-
of AI for public policy. Science used as a foundation for ness, other challenges involved with AI’s deployability
statutes is ever changing, and in many cases, different to the water sector include adoption by operators; the
contexts could lead to different results. Accordingly, black box nature of AI models; and data privacy issues—
data-driven lawmaking has to be one of the major all of which are applicable to most sectors. Ultimately,
ways of constructing and evaluating the success of however, these issues ought to be addressed, espe-
statutes—a direction that is becoming progressively cially because water has no substitute.
inevitable and is also increasingly further backed up by
the public.28 One of the ongoing debates in water and REFERENCES
environmental law is the Water of the United States 1. R. Clark, S. Panguluri, T. Nelson, and R. Wyman,
issue—that is, which authority (state versus federal) “Protecting drinking water utilities from cyberthreats,”
prevails and who has control over water bodies of the J. Amer. Water Works Assoc., vol. 109, no. 2, pp. 50–58,
United States. Feb. 2017, doi: 10.5942/jawwa.2017.109.0021.
Multiple administrations have tackled this issue, 2. “Water and wastewater systems sector-specific plan,”
but what is very interesting is the involvement of dif- U.S. Dept. Homeland Secur., Washington, DC, USA,
ferent scientific assumptions as a basis for criteria 2015. [Online]. Available: https://www.cisa.gov/sites
that define answers to the debate. For instance, the /default/files/publications/nipp-ssp-water-2015-508.pdf
Obama administration used the significant nexus test, 3. J. Hubert, Y. Wang, E. Alonso, and R. Minguez, Using
while the Trump administration used the narrower Artificial Intelligence for Smart Water Management
definition derived from the Supreme Court’s Rapanos Systems. Mandaluyong, Philippines: Asian Develop-
versus USA case, 547 U.S. 715 (2006). Both are deemed ment Bank, 2020.
reasonable, but one would ask: Which one should be 4. A. Hassanzadeh et al., “A review of cybersecurity
followed? And without sufficient data, how does one incidents in the water sector,” J. Environmental Eng.,
measure the outcome? The answer lies in empirical vol. 146, no. 5, May 2020, Art. no. 03120003, doi: 10.1061
evaluations and scientific experimentation (not expert /(ASCE)EE.1943-7870.0001686.
opinions or political partisanship)—multiple research 5. M. Hameed et al., “Application of artificial intelligence
institutes produce research that generates results and (AI) techniques in water quality index prediction:
directions that ought to support a certain direction, A case study in tropical region, Malaysia,” Neural
although contradicting in some cases, but referring to Comput. Appl., vol. 28, no. 1, pp. 893–905, Dec. 2017, doi:
such scientific debates usually leads to a consensus 10.1007/s00521-016-2404-7.
that is backed up by different dimensions. 6. J. Chen, X. Zhang, R. Zhang, C. Wang, and L. Liu,
In some cases, U.S. law cannot be defined in isola- “De-Pois: An attack-agnostic defense against data
tion; therefore, the national well-being also has to be poisoning attacks,” IEEE Trans. Inf. Forensics Security,
in the balance when it comes to ratifying statutes that vol. 16, pp. 3412–3425, May 2021, doi: 10.1109/TIFS.2021
contradict or confirm international treaties. For such .3080522.
difficult goals to be balanced in the same legal realm, 7. J. Zheng, P. Chan, H. Chi, and Z. He, “A concealed poi-
AI can be one of the main referees in determining soning attack to reduce deep neural networks robust-
the best version of a statute as it is being crafted or ness against adversarial samples,” Inf. Sci., vol. 615, pp.
amended. 758–773, Nov. 2022, doi: 10.1016/j.ins.2022.09.060.
8. H. Mehmood, D. Liao, and K. Mahadeo, “A review of arti-

A I has been helping cure disease, create art, drive

cars, perform surgery, and identify crime; the
water sector is no different. In this article, the use
ficial intelligence applications to achieve water-related
sustainable development goals,” in Proc. IEEE/ITU Int.
Conf. Artif. Intell. Good (AI4G), 2020, pp. 135–141, doi:
cases, AI methods, and applications aim to encourage 10.1109/AI4G50087.2020.9311018.

www.computer.org/computingedge 43
ARTIFICIAL INTELLIGENCE/MACHINE LEARNING

9. L. Corominas, M. Garrido-Baserba, K. Villez, G. Olsson, U. 20. E. E. Kim et al., “Water economy with smart water
Cortés, and M. Poch, “Transforming data into knowledge system in the City of Carouge,” in Proc. IEEE Int. Conf.
for improved wastewater treatment operation: A critical Omni-Layer Intell. Syst. (COINS), 2022, pp. 1–6, doi:
review of techniques,” Environmental Modelling Softw., 10.1109/COINS54846.2022.9854990.
vol. 106, pp. 89–103, Aug. 2018, doi: 10.1016/j.envsoft 21. J. Ktari, T. Frikha, M. Hamdi, H. Elmannai, and H. Hmam,
.2017.11.023. “Lightweight AI framework for industry 4.0 case study:
10. J. Keck and J. Lee, “Embracing analytics in the water Water meter recognition,” Big Data Cogn. Comput., vol.
industry,” J. Water Resour. Planning Manage., vol. 147, 6, no. 3, p. 72, Jul. 2022, doi: 10.3390/bdcc6030072.
no. 5, May 2021, Art. no. 02521002, doi: 10.1061/(ASCE) 22. L. Hao, J. Gao, and J.-a Wang, “Simulating the effects of
WR.1943-5452.0001375. water reuse on alleviating water shortage,” in Proc. 4th
11. “Water security and the global water agenda: A Int. Conf. Bioinf. Biomed. Eng., 2010, pp. 1–4, doi: 10.1109
UN-Water Analytical Brief,” UN Water, Hamilton, ON, /ICBBE.2010.5515327.
Canada, 2013. [Online]. Available: https://collections. 23. G. Sebestyen, A. Hangan, and Z. Czako, “Anomaly
unu.edueserv/UNU:2651/Water-Security-and-the detection in water supply infrastructure systems,”
-Global-Water-Agenda.pdf in Proc. 23rd Int. Conf. Control Syst. Comput. Sci.
12. N. K. Velayudhan, P. Pradeep, S. N. Rao, A. R. Devidas, (CSCS), 2021, pp. 349–355, doi: 10.1109/CSCS52396
and M. V. Ramesh, “IoT-enabled water distribution .2021.00064.
systemsA comparative technological review,” IEEE 24. W. Y. Mao et al., “Trustworthy AI solutions for cyber-
Access, vol. 10, pp. 101,042–101,070, Sep. 2022, doi: biosecurity challenges in water supply systems,” Int.
10.1109/ACCESS.2022.3208142. FLAIRS Conf. Proc., vol. 35, May 2022, Art. no. 130664,
13. A. Harsha Vardhan, B. Subramanyam, M. Lakshmi doi: 10.32473/flairs.v35i.130664.
Reddy, G. Gowtham Reddy, and A. Ramesh, “Anomaly 25. V. Radhakrishnan and W. Wu, “IoT technology for smart
detection in water distribution systems,” in Proc. 4th water system,” in Proc. IEEE 20th Int. Conf. High
Smart Cities Symp. (SCS), 2021, pp. 219–222, doi: 10.1049 Perform. Comput. Commun.; IEEE 16th Int. Conf.
/icp.2022.0344. Smart City; IEEE 4th Int. Conf. Data Sci. Syst. (HPCC/
14. R. Shamika, A Reality Check on Suicides in India (Brook- SmartCity/DSS), 2018, pp. 1491–1496, doi: 10.1109/HPCC
ings India IMPACT Series). New Delhi, India: Brookings /SmartCity/DSS.2018.00246.
Institution India Center, 2015. 26. J. Park, W. H. Lee, K. T. Kim, C. Y. Park, S. Lee, and T.-Y.
15. M. Lunani, “Artificial intelligence for water and Heo, “Interpretation of ensemble learning to predict
wastewater: Friend or foe?” Opflow, vol. 44, no. 6, pp. water quality using explainable artificial intelligence,”
6–7, Jun. 2018, doi: 10.1002/opfl.1017. Sci. Total Environ., vol. 832, Aug. 2022, Art. no. 155070,
16. F. Batarseh, M. Gopinath, A. Monken, and Z. Gu, “Public doi: 10.1016/j.scitotenv.2022.155070.
policymaking for international agricultural trade using 27. N. Bostrom, Superintelligence: Paths, Dangers, Strate-
association rules and ensemble machine learning,” gies. Oxford, U.K.: Oxford Univ. Press, 2015.
Mach. Learn. Applicat., vol. 5, Sep. 2021, Art. no. 28. F. Batarseh and R. Yang, Federal Data Science: Trans-
100046, doi: 10.1016/j.mlwa.2021.100046. forming Government and Agricultural Policy Using
17. A. Chastain-Howley. “How big is big data among water Artificial Intelligence. London, U.K.: Elsevier, 2017.
utilities?” Water Online. Accessed: Oct. 29, 2022.
[Online]. Available: https://www.wateronline.com/doc
/how-big-is-big-data-among-water-utilities-0001 FERAS A. BATARSEH is an associate professor with the
18. G. Abramowitz, “Towards a benchmark for land surface Department of Biological Systems Engineering, Virginia
models,” Geophys. Res. Lett., vol. 32, no. 22, Nov. 2005, Tech, Arlington, VA 22203 USA. He is a Senior Member of
Art. no. L22702, doi: 10.1029/2005GL024419. IEEE. Contact him at [email protected].
19. N. Hellen and G. Marvin, “Explainable AI for safe water
evaluation for public health in urban settings,” in Proc. AJAY KULKARNI is a postdoctoral associate with the Com-
Int. Conf. Innov. Sci., Eng. Technol. (ICISET), 2022, pp. monwealth Cyber Initiative, Virginia Tech, Arlington, VA
1–6, doi: 10.1109/ICISET54810.2022.9775912. 22203 USA. Contact him at [email protected]

44 ComputingEdge December 2024

 IEEE COMPUTER SOCIETY

Publications Seek
2026 Editors in Chief
Application Deadline: 1 March 2025

IEEE Computer Society seeks applicants for editor in chief for the following publications:

• Computer magazine • IEEE Transactions on Mobile Computing

• IEEE Computer Graphics and Applications • IEEE Transactions on Services Computing
• IEEE Security & Privacy • IEEE Transactions on Software Engineering
• IEEE Transactions on Emerging Technologies • IEEE Transactions on Visualization and
in Computing Computer Graphics

Our publications are the cornerstone of professional activities for our members and the
community we serve. We seek candidates who are IEEE members in good standing, have strong
familiarity with our publications, and possess an excellent understanding of the field as it relates 1260 IEEE TRANSACTIONS ON BIG DATA, VOL. 9, NO. 5, OCTOBER 2023

to academic, industry, and governmental areas. Applicants must have successful experience
Epidemic Spread Modeling for COVID-19 Using
Cross-Fertilization of Mobility Data 09.23

Computer
Anna Schmedding , Riccardo Pinciroli , Lishan Yang , Member, IEEE, and Evgenia Smirni , Fellow, IEEE P U T T I N G A I I N T O P R A C T I C E

developing a diverse team of individuals to serve key editorial board roles. Demonstrated
Abstract—We present an individual-centric model for COVID- In this article, we use the data provided by the Korean Center
19 spread in an urban setting. We first analyze patient and route for Disease Control (KCDC) and local governments during the VOLUME 38, NUMBER 2 MARCH/APRIL 2023
STATE OF TELEMEDICINE—PART

SEPTEMBER 2023
data of infected patients from January 20, 2020, to May 31, 2020,
first wave of the disease in South Korea. In contrast to the
collected by the Korean Center for Disease Control & Prevention
(KCDC) and discover how infection clusters develop as a function Google data, the KCDC data focus on individual patients and
of time. This analysis offers a statistical characterization of mobility allow the development of an individual-centric model of the
habits and patterns of individuals at the beginning of the pandemic. COVID-19 epidemic. Infected individuals are monitored3 and
While the KCDC data offer a wealth of information, they are also their movements are logged using CCTV, cellphones, and credit

managerial skills are also required to ensure content and issue development, and timely
by their nature limited. To compensate for their limitations, we
card transactions. The KCDC records patient movements in
use detailed mobility data from Berlin, Germany after observing
that mobility of individuals is surprisingly similar in both Berlin plain text (i.e., natural language) without any unified rule. These
and Seoul. Using information from the Berlin mobility data, we logs are parsed through automated code and rule-based methods
cross-fertilize the KCDC Seoul data set and use it to parameterize to extract keywords that are then used with web mapping service
an agent-based simulation that models the spread of the disease in APIs (e.g., Google Maps) to extract geographical coordinates
an urban environment. After validating the simulation predictions
(i.e., latitude and longitude) and other data. The parsed logs are
with ground truth infection spread in Seoul, we study the impor-
tance of each input parameter on the prediction accuracy, compare made publicly available [1] and being collected by KCDC are

processing of submissions. Terms begin 1 January 2026.

the performance of our model to state-of-the-art approaches, and deemed trustworthy.
show how to use the proposed model to evaluate different what-if To the best of our knowledge, the KCDC logs are the only
counter-measure scenarios. publicly available data that contain patient-centric information
Index Terms—Data analysis, simulation models, individual- in great detail: they report on the patient mobility, i.e., traveled
centric models, COVID-19, disease spread modeling, cross- distance and the sequence of locations visited on a daily basis, the

STATE OF TELEMEDICINE—PART I
fertilization. date of the onset of symptoms, whether and when the patient got
in contact with other patients that are also diagnosed. This leads
I. INTRODUCTION to our first research question, RQ1: What statistical information

O N MARCH 11, 2020, the WHO1 declared COVID-19 can be extracted by the KCDC mobility data to parameterize an
the first pandemic caused by a coronavirus. Since then, agent-based simulation that models the spread of the disease?
a tremendous amount of data has been collected to help pub- The KCDC logs are a valuable resource for studying the spread

 The last version of the KCDC data set contains data col-
lic policy decisions that limit the spread of COVID-19. For of COVID-19, yet they have limitations:

For complete information on how to apply, please go to

example, Google2 provides time-series data of infections at a
coarse granularity (i.e., as a function of the area’s population, no lected up to May 31, 2020 (the KCDC data set has not been
information is provided at the granularity of single individuals). updated since then). By that date, approximately 11,500
Epidemiological simulation and mathematical models have been COVID-19 cases were confirmed in South Korea [2], but

 Some locations visited by patients (e.g., locations where

used to predict the spread of the disease. Typically, model only 35% of them have been logged into the data set.
effectiveness is tied to its input parameterization.
people live) are not recorded due to privacy concerns.
Manuscript received 8 July 2022; revised 16 January 2023; accepted 14 Consequently, patient infection information and route data

www.computer.org/press-room/seeking-2026-editors-in-chief
February 2023. Date of publication 24 February 2023; date of current version 1 do not always coincide. For example, there are patients that

AI’s 10 to Watch
September 2023. This work was supported in part by the National Science Foun- infect each other even if their routes do not cross. This may

 Patient and route data may be incomplete (i.e., some

dation under Grants IIS-2130681 and IIS-1838022, and in part by the Ministry of
University and Research under the PRIN project SEDUCE 2017TWRCNB and happen when patients belong to the same household.

Deep Anomaly Analytics

the PNRR project VITALITY (ECS00000041) Spoke 2 Astra. Recommended
for acceptance by S. Mohanty. (Corresponding author: Riccardo Pinciroli.) attributes are occasionally missing, such as the type of
Anna Schmedding and Evgenia Smirni are with the Computer Science De-
partment, William and Mary, Williamsburg, VA 23185 USA (e-mail: akschmed- locations visited by some patients) and require manual

 There is route data information for only a portion of the

[email protected]; [email protected]). completion before analyzing the data set.
Riccardo Pinciroli is with the Computer Science Department, Gran Sasso
Science Institute, 67100 L’Aquila, Italy (e-mail: [email protected]).
Lishan Yang is with the Computer Science Department, George Mason patients. Patient movement has been logged only for the
University, Fairfax, VA 22030 USA (e-mail: [email protected]). 15% of all confirmed cases by May 31.

Volume 56 Number 9
Digital Object Identifier 10.1109/TBDATA.2023.3248650
1 https://bit.ly/3izwIdL
2 https://bit.ly/3H5YU1V 3 https://bit.ly/3VMQvVm

2332-7790 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information. w w w .computer.org /i ntel l i g ent
vol. 56 no.9 www.computer.org/com

9128 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 23, NO. 10, OCTOBER 2024
2316 IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 16, NO. 4, JULY

A Blockchain-Based Copyright Protection

IIEEE COMPUTER GRAPHICS AND APPLICATIONS
IEEE INTELLIGENT SYSTEMS

03.23 Scheme With Proactive Defense

Computer

Xiaohu Chen, Anjia Yang , Member, IEEE, Jian Weng , Member, IEEE, Yao Tong,
Cheng Huang , Member, IEEE, and Tao Li

BETTER LIVING VOLUME 44, NUMBER 4 JULY/AUGUST 2024

Abstract—Copyright protection, including copyright registra- that the proposed scheme has good security and effic
MARCH 2023

VOLUME 38, NUMBER 4 JULY/AUGUST 2023 tion, copyright transfer and infringement penalty, plays a critical can be applied for the copyright protection.

THROUGH CHALLENGES
role in preventing illegal usage of original works. The mainstream

Apply Today!
traditional copyright protection schemes need an authority online Index Terms—Blockchain, copyright protecti
all the time to handle copyright issues and face some problems authentication-preventing signature.
July/August 2023

such as intricate copyright transfer, single point of failure and so

on. To alleviate the burden of the authority, a few blockchain-based I. INTRODUCTION

I
copyright protection schemes are proposed. However, most of them
do not consider copyright transfer, and their infringement penalty
NTERNET technology has greatly promoted
may only happen after copyright owners discover the infringement nation of data, but also makes the illegal repro
behavior (i.e., “ex-post penalty”). In this article, we propose a new unauthorized usage of the data extremely rampant
security strategy, called “Proactive Defense” in copyright protec- data from being illegally used, and to track and puni
Fig. 2. Layout of experimental fields and sites (the red dots are the location of RPs). (a) Layout diagram. (b) Experimental site. tion which can prevent infringement before it occurs. With our violators, it is important that the data owner can prov
proposed proactive defense strategy, we design a secure copyright
protection scheme which provides advantages of compact copyright
ership (i.e., copyright) over the contents. For exam
transfer and prior infringement penalty. More concrete, both copy- that a musician composed a piece of music, with th
JULY/AUGUST

right registration and transfer are regarded as transactions and copyright, he was free to play and disseminate it wi
recorded to the blockchain. Based on the double-authentication- being taken away as creator by others. In addition, t
prevention signature and non-interactive zero-knowledge proof can be transferred, which can raise incomes of the m
techniques, illegal copyright transfer can be detected and the in-
timely and full enforcement of the rights of creators
BETTER LIVING THROUGH CHALLENGES

fringement penalty can be done automatically with a tailored smart

2024

contract before the completion of the transfer. Our security analysis inators can stimulate the vitality of innovation and
shows that the proposed scheme can achieve all desirable security the whole society. As a consequence, copyright can
properties. Moreover, we implement our scheme in Java and eval- prosperity and development of literature, art and sc
Text Classification and Sentiment Analysis

uate the performance experimentally. Experimental results show Acting as a kind of digital assets, copyright also
problems. For example, It is easy to duplicate th
across borders illegally, but hard to detect and pu
Manuscript received 18 September 2022; revised 20 January 2023; accepted fringement. Therefore, a set of effective methods a
13 February 2023. Date of publication 23 February 2023; date of current protect the copyright. Copyright protection, includi
Multisensory Experiences in eXtended Reality

version 8 August 2023. The work of Xiaohu Chen was supported in part by the
Key-Area Research and Development Program of Guangdong Province under registration, copyright transfer and infringement pen
Grants 2020B0101090004 and 2020B0101360001. The work of Anjia Yang was critical role in preventing illegal usage of original wo
supported in part by the National Key Research and Development Program of requirements need to be satisfied associated with t
China under Grant 2021ZD0112802, in part by the National Natural Science
Foundation of China under Grant 62072215. The work of Jian Weng was sup- protection: generating a unique label for the copy
ported in part by the National Natural Science Foundation of China under Grants registration phase, ensuring the fairness in the transf
61825203 and U22B2028, in part by the Major Program of Guangdong Basic and punishing the malicious party and protecting the ho

Text Classification and

Fig. 3. Schematic illustration and APP interface of localization software system designed based on the proposed algorithm. (a) Localization system schematic
diagram. (b) App interface. Multisensory Experiences in Applied Research Project under Grant 2019B030302008, in part by the National
Key Research and Development Plan of China under Grant 2020YFB1005600,
in part by Guangdong Provincial Science and Technology Project under Grant
the infringement penalty phase.
The mainstream traditional copyright protection

eXtended Reality
2021A0505030033, in part by the Science and Technology Major Project of based on watermarking and digital fingerprinting t
Sentiment Analysis
Tibetan Autonomous Region of China under Grant XZ202201ZD0006G, and
shadowing models are shown in Table III. Moreover, to better 4.26 dBm, which is 21.40%, 15.98%, 13.77%, 14.29%, and in part by National Joint Engineering Research Center of Network Security [1], [2], [3], [4], [5], [6]. Most of these schemes
visualize the differences among the models, we illustrate the 9.36%, respectively, compared with the algorithms lognormal Detection and Protection Technology, and Guangdong Key Laboratory of Data copyright protection by adding a watermarking or d
Security and Privacy Preserving under Grant 2017B030301004. Recommended printing to the file to be protected. By checking the
cumulative distribution function (CDF) of the RSS estimation shadowing, covSE, covMater, covRQ, and covMater+RQ. This for acceptance by Y. Zhang. (Corresponding author: Anjia Yang.)
errors using different methods in Fig. 4. The four APs shown in is because the IRPLS model has good adaptability; it can capture Xiaohu Chen, Anjia Yang, Jian Weng, and Tao Li are with the Col- in the files, data owners can detect illegal data usa
Fig. 5 reveal the lower and upper bounds of the RSS estimation the RSS changes of spatial information in a sparse state and lege of Cyber Security, Jinan University, Guangzhou 510632, China (e-mail: protect their rights through other methods, such as
[email protected]; [email protected]; [email protected]; the authority. Some studies consider secure copyr
along the central axis based on the IRPLS method and the RSS appropriately expand RSS data. [email protected]).
that can resist malicious copyright owner (seller) o
Volume 44

estimation of several other algorithms. The radio maps of four The CDF curve is an important evaluation standard that intu- Yao Tong is with the Fongwell Data Limited Company, Guangzhou 510623,
APs generated by the IRPLS prediction model are shown in itively reflects the performance of the algorithm. Its horizontal China (e-mail: [email protected]). [8]. However, these traditional copyright protecti
Volume 38, Number 4

Cheng Huang is with the Department of Electrical and Computer Engi- need an authority online all the time to handle copy
Volume 56 Number 3

Fig. 6 axis represents the RSS estimation error, and its vertical axis rep- neering, University of Waterloo, Waterloo, ON N2L 3G1, Canada (e-mail:
As shown in Table III, the proposed model achieves better resents the cumulative distribution probability within a specific [email protected]). which poses several problems such as intricate copyr
Number 4

performance in the RSS estimation errors for different algo- error. As shown in Fig. 4, IRPLS obtains the best performance, Digital Object Identifier 10.1109/TSC.2023.3246476 procedure, single point of failure and so on [9].
rithms. In the case of sparse sampling (0.046 RPS per square followed by covMater+RQ, while RSS estimation of classical www.computer.org/cga
www.computer.org/intelligen t 1939-1374 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
meter), the RSS estimation error of the IRPLS model reaches lognormal shadowing exhibits the worst performance. See https://www.ieee.org/publications/rights/index.html for more information.
vol. 56 no.3
no. 1 www.computer.org/computer

6924 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 30, NO. 10, OCTOBER 2024
EDITOR: Fei-Yue Wang, [email protected]
IEEE SECURITY & PRIVACY

COLUMN: AI EXPERT
02.23
08.23
Simulation Driven AI: From Artificial
Computer

Computer

2022 Volume 3 to Actual and Vice Versa

TRUSTWORTHY AI—PART I Li Li , Tsinghua University, Beijing, 100084, China
FEBRUARY 2023

Yilun Lin , Shanghai AI Laboratory, Shanghai, 200232, China

AUGUST 2023

Yutong Wang and Fei-Yue Wang , Chinese Academy of Sciences, Beijing, 100190, China
INCLUSIVE PRIVACY AND SECURITY

In this perspective, we discuss the important role of simulations in building state-of-

the-art artificial intelligence (AI) systems. We first explain why simulations become
vital in building complex AI systems. Then, we study some challenges and candidate
solutions related to simulation-based AI systems. Finally, we discuss future
research directions in this field.

A
rtificial Intelligence (AI) research is embracing decades. For example, reinforcement learning assumes
a new age of renaissance due to neural net- that AI systems consist of single or multiple agents that
works.1 However, we still do not have a com- learn to achieve a goal in a complex, potentially uncer-
monly accepted definition for AI while researchers tain environment via a process of trial-and-error.5 Build-
around the world are aggressively addressing various ing such an intelligent system can be viewed as a series
aspects of AI from different viewpoints. of tightly correlated trial-and-error actions. The designer
TRUSTWORTHY AI—PART I

Reflecting and summarizing the idea of universal tries an initial setup first, then improves their building
COMPETITIONS

language proposed by Leibniz, the concept of comput- strategy by observing agent behavior in a test environ-
VOLUME 22

ability and intelligent machinery proposed by Turing,2 ment until the system is successfully built.
as well as the framework of Cybernetics proposed by To demonstrate how simulation plays a more criti-
Wiener,3 we arrive at three basic functions of an AI cal role in building AI systems, we will first introduce
system :4 how to use simulations to build three functions of
NUMBER 5

intelligent systems, respectively, and then discuss sev-

› The Representation of AI. First, an AI system eral important issues related to simulation-based AI.
should be able to translate a special task into a
certain computational problem, by using a uni-
SIMULATIONS FOR AI SYSTEMS
SEPTEMBER/OCTOBER 2024

versal language.
› The Design of AI. Second, an AI system should Simulation for AI Representation
be able to find an algorithm to solve the compu- Simulation usually refers to the imitation in a virtual
tational problem. world for the operation of a real-world process/system.
› The Test of AI. Third, an AI system should be able It is a powerful tool to help us understand how a com-
to find an algorithm to validate the answer to the plex process/system really works.6 In many situations,
computational problem. we do not have an overall analytical formula to describe
the collective mechanism of a complex process/system.
Reviewing the developing history of AI, we find that Instead, we can develop some models to represent the
Fig. 15. Comparison with state-of-the-art neural dynamic clothing deformation methods. Rows from top to bottom : ground truth, SNUG [14], Neural Cloth
WWW.COMPUTER.ORG/SECURITY

various ideas and methods have been proposed to key characteristics, behaviors, and functions of the Simulation (NCS) [13], HOOD [18], and our prediction. Columns represent deformation with various garments and motions: (a)-(b) moving forward and hanging
implement these three functions during the last eight individual components within this process/process. leg raise, (c)-(e) palm striking and kicking, (f)-(h) layup.
This allows us to use the simulation to represent the
operation of the process/system over time.7
CS Election Candidates, p. 106
Volume 56 Number 8

When the simulation model is determined, we can As observed in the results, for motions with low dynamics: our proposed spectral strategies. In scenarios where movements
Volume 56 Number 2

1541-1672  2023 IEEE collect lots of simulated data that can then be viewed forward movement (a) and hanging leg lifts (b), prior methods
can yield roughly reasonable deformations, but some finer de-
involve actions such as the character’s palm strikes and body
twisting in cases (c) and (d), we can observe a continuous swing-
IEEE Division VIII Delegate/Director
Candidates, p. 122
Digital Object Identifier 10.1109/MIS.2023.3235676 as the sampled input–output pairs of the complex pro-
Date of current version 13 February 2023. cess/system. We can then use other tools (e.g., logical
September/October 2024 tails are lacking in HOOD’s prediction. In contrast, our method ing motion of the dress hem from left to right. However, SNUG
Vol. 22, No. 5 demonstrates its capability to predict a substantial portion of does not adequately capture this dynamic behavior, particularly
folds and dynamic trends, primarily due to the effectiveness of in the fold direction of the dress. This deficiency in capturing vol. 56 no.8 www.computer.org/com
vol. 56 no.2
no. 1 www.computer.org/computer January/February 2023 Published by the IEEE Computer Society IEEE Intelligent Systems 3

2128 IEEE TRANSACTIONS ON SOFTWARE ENGINEERING, VOL. 50, NO. 8,

IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 11, NO. 3, JULY-SEPTEMBER 2023 2217

EDITOR RON VETTER

University of North Carolina Wilmington;
A Flexible Heuristic to Schedule Distributed SPOTLIGHT ON TRANSACTIONS [email protected]

Analytic Applications in Compute Clusters 05.23

Computer

Francesco Pace , Daniele Venzano, Damiano Carra , and Pietro Michiardi

2022 Volume 3

Trustworthy AI—Part II
Abstract—This work addresses the problem of scheduling user-defined analytic applications, which we define as high-level
compositions of frameworks, their components, and the logic necessary to carry out work. The key idea in our application definition, is
to distinguish classes of components, including core and elastic types: the first being required for an application to make progress, the Fig. 3. Simplified execution steps based on trace in Fig. 2. Notation: gray columns: RTC steps, white columns: stable configurations, boldface
dispatched event. (Trivial RTC steps where CEs are discarded or events are forwarded to the doActivity are omitted.)
MAY 2023

latter contributing to reduced execution times. We show that the problem of scheduling such applications poses new challenges, which
existing approaches address inefficiently. Thus, we present the design and evaluation of a novel, flexible heuristic to schedule analytic
applications, that aims at high system responsiveness, by allocating resources efficiently. Our algorithm is evaluated using trace-driven
simulations and with large-scale real system traces: our flexible scheduler outperforms current alternatives across a variety of metrics,
including application turnaround times, and resource allocation efficiency. We also present the design and evaluation of a full-fledged
system, which we have called Zoe, that incorporates the ideas presented in this paper, and report concrete improvements in terms of
efficiency and performance, with respect to prior generations of our system.

Contributions of
Index Terms—Scheduling, distributed applications, distributed systems

Quantum Factoring
1 INTRODUCTION

T HE last decade has witnessed the proliferation of numer-

ous distributed frameworks to address a variety of
large-scale data analytics and processing tasks. First, MapRe-
request the deployment of the framework they need. The
general underlying idea is that of sharing cluster resources
among a heterogeneous set of frameworks, as a response to

on Quantum
duce [1] has been introduced to facilitate the processing of static partitioning, which has been dismissed for it entails
bulk data. Subsequently, more flexible tools, such as Dryad low resource allocation [8], [9], [10]. Existing systems divide
[2], Spark [3], Flink [4] and Naiad [5], to name a few, have the resources at different levels. Some of them, e.g., Mesos
been conceived to address the limitations and rigidity of the and YARN, target low-level orchestration of distributed com-

Research
TRUSTWORTHY AI—PART II

MapReduce programming model. Similarly, specialized puting frameworks: to this aim, they require non-trivial
libraries such as MLLib [6] and systems like TensorFlow [7] modifications of such frameworks to operate correctly.
have seen the light to cope with large-scale machine learning Others, e.g., Kubernetes [22] and Docker Swarm [23], focus
problems. In addition to a fast growing ecosystem, individ- on provisioning and deployment of containers, and are thus
ual frameworks are driven by a fast-pace development oblivious to the characteristics of the frameworks running in
model, with new releases every few months, introducing such containers. To the best of our knowledge, no existing Clayton Ferner, University of North Carolina Wilmington concept of using quantum mechan-
substantial performance improvements. Since each frame- tool currently addresses the problem of scheduling analytic ics as a basis for computing was first
work addresses specific needs, users are left with a wide applications as a whole, leveraging the intrinsic properties of
introduced around 1980 by Paul Be-
choice of tools and combination thereof, to address the vari- the frameworks such applications use. This installment of Computer’s series highlights nioff.1 Since that time, researchers
ous stages of their data analytics projects. The endeavor of this paper is to fill the gap that exists in
The context depicted above has driven a lot of research current approaches, and raise the level of abstraction at which work published in IEEE Symposium on have been working on developing ac- Fig. 4. Overview of the semantic classes defined in fUML and PSSM for state machines and doActivities. Notation: rectangles: UML (meta)
[8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], scheduling works. We introduce a general and flexible defi- tual quantum computers as well as al- vertices: implicit concepts in PSSM; solid lines: associations; dashed lines: dependencies.
[21] (see Section 7 for a detailed discussion) in the area of nition of applications, how they are composed, and how to Foundations of Computer Science. Although we gorithms using quantum computing.
resource allocation and scheduling, both from academia and execute them. For example, a user application addressing Algorithms have been developed that
the industry. These efforts materialize in cluster manage- the training of a statistical model involves: a user-defined typically highlight articles from work published demonstrate quantum supremacy, that
test cases, executing the test cases in Cameo Simulation Toolkit, A. Overview of the Operational Semantics in PSS
and examining the source code and debugger executions of the
ment systems that offer simple mechanisms for users to program implementing a learning algorithm, a framework
is, quantum computers can solve prob- The fUML specification defines an operational
(e.g., Spark) to execute such a program together with infor- in IEEE Computer Society journals, this month’s lems with a super polynomial speedup
Papyrus Moka4 [31] reference implementation. for activities and actions, which is extended for
mation about its resource requirements, the location for We highlight the parts that necessitate careful considera-
 F. Pace, D. Venzano, and P. Michiardi are with the Data Science Department, chines in PSSM. These specifications define an execu
Eurecom, Sophia Antipolis 06904, France. E-mail: {francesco.pace, daniele. input and output data and possibly parameters exposed as exception is due to the prevalence and over classical computers. tion when developing state machine models (e.g., concurrent for a subset of the UML language. The execution
venzano, pietro.michiardi}@eurecom.fr. application arguments. Users should be able to express, in a Until recently, quantum comput- behaviors or non-deterministic choices in priorities), and incon- cludes an abstract execution engine, classes for eve
 D. Carra is with the Computer Science Department, University of Verona,
Verona 37129, Italy. E-mail: [email protected].
simple way, how such an application must be packaged and significance of quantum computing research ing has remained on the fringes of sistencies in the specification artifacts that could cause misun-
 EDITOR: Alexander
EDITOR: Alexander B. Magoun,
Magoun, [email protected]
[email protected]
This article originally
appeared in
DEPARTMENT:
DEPARTMENT: ANECDOTES
ANECDOTES
vol. 45, no. 4, 2023

Computer
Computer Networking Networking Initiatives in One of
Initiatives
the World’s inRemote
One of the Cities World’s
Remote Cities
T. Alex Reid , The University of Western Australia, Crawley, WA, 6009, Australia

T. Alex Reid , The University of Western Australia, Crawley, WA, 6009, Australia

This article describes some computing initiatives made by members of the

University of Western Australia, located in arguably the most isolated capital city in
This article describes some computing initiatives made by members of the University of
the world.
Western These initiatives
Australia, located incenter around
arguably theonline and networking
most isolated capital capabilities,
city in the world. These
predominantly arising from the installation, in 1965,
initiatives center around online and networking capabilities, of the first time-sharing arising
predominantly
computer
from in Australia.inThis
the installation, far-sighted,
1965, of the firstiftime-sharing
risky, purchase set the university
computer on This
in Australia. a far-
course that led to many more initiatives, encompassing significant computer
sighted, if risky, purchase set the university on a course that led to many more initiatives,
resource sharing,
encompassing a ground-breaking
significant online library
computer resource system,
sharing, early online education
a ground-breaking online library
system, early
programs, online
and education
an early programs,
multihost and an early
packet-switched multihost
network. Mostpacket-switched
research network.
Most
concerning isolation and innovation suggests that isolation operates as a break on as
research concerning isolation and innovation suggests that isolation operates
ainnovation,
break on innovation, but the
but the Western Western Australian
Australian experience experience belies that conclusion.
belies that conclusion.

WESTERN AUSTRALIA MADE A line of computers incorporating the same instruction

LATE START IN COMPUTING set and designed for all users large and small [7], [14].

P
erth, the capital city of the State of Western
Australia, is often considered to be the most
isolated capital city in the world. Regarded by TIME-SHARING COMES TO
many at the time as something of a “quiet backwater,” AUSTRALIA
so it was not surprising that it was not until 1962 that However, the university’s Computing Centre Director,
Perth (and hence the whole of the 2.6 million square Dennis Moore, was familiar with MIT’s multiuser, time-
kilometres of Western Australia) took delivery of its shared computing Project MAC, and believed that this
first computer, a Bendix G15, bought by the Main was the way of the future [8], [11], [14]. There were only
Roads Department to help in the design of freeway two companies worldwide at the time offering to sell
extensions and interchanges [28]. It was followed that such computers, though none had yet been sold. These
same year by an IBM 1620 bought by the University of were the Canadian division of the British firm Ferranti
Western Australia (UWA) [19], [28]. This was well after Limited, and the Digital Equipment Corporation (DEC)
most other Australian capitals (and universities) had in Massachusetts, USA [1]. IBM did not support time-
acquired computers [17], [18], [19]. sharing on its System/360 range until 1970 [7].
Given this slow start, it is perhaps rather surprising The funding of public universities in Australia at
that Perth did not long remain quiet in computing. By the time was split equally between the Federal and
1964, the capacity of the university’s IBM 1620 was State governments, with triennial grants made avail-
overwhelmed by demand from university and external able for capital items such as computers. The univer-
users, and a search began for a much larger replace- sity obtained such a capital grant and issued tenders.
ment. IBM were convinced that the university would There were four contenders—-IBM, ICT (later ICL),
just buy it from their stable, especially since they had DEC, and Ferranti. Each made presentations to the
just announced their System/360 range, being the first university’s Computer Users Group [13]. The Ferranti-
Packard salesman was very persuasive about the ben-
efits of time-sharing and “almost had the Computer
Users Group cheering” for its 6000 model, recounted
1058-6180  2023 IEEE
Moore [14]. He was in fact selling the DEC computer
Digital Object Identifier 10.1109/MAHC.2023.3325497 [11], [13]. Designed originally to share computing with
Date of current version 23 November 2023. equipment operating in laboratories, the PDP-6 was a

46
72 December 2024of the History of Computing
IEEE Annals Published by the IEEEby
Published Computer Society  Society
the IEEE Computer 2469-7087/24 © 2024
October-December 2023 IEEE
 ANECDOTES
ANECDOTES

full time-sharing machine incorporating the hardware Australia for many years to come, paving the way for the
to do this efficiently [1]. creation of the Western Australian Regional Computing
Despite efforts by IBM to reverse the decision, the Centre in 1972 (see below).
University resolved to buy a PDP-6. This was an This acquisition of a revolutionary and architectur-
extraordinary decision. Digital Equipment was a small, ally pioneering computer system allowed a wide range
almost unknown, company; time-sharing was a great of novel applications to be developed, in particular
laboratory experiment, but there were no systems running online experiments in laboratories around the
commercially available and in use in the field; and university. This was exemplified by the Department of
Perth was the furthest place in the world, the almost Psychology, one of whose fields of research was in
exact antipodes, from Digital’s factory in Maynard, visual perception. Being able to use the computer in
Massachusetts. When the computer was delivered in real time to control images that human subjects were
May 1965, it became the first time-shared computer to allowed to see, and how they perceived them, enabled
be delivered to Australia [2], and among the first in the significant advances in understanding the human per-
world. The foresight and bravery of Moore and Birkett ception system [30]. This led subsequently to the
Clews and the university was quite remarkable. invention of the Betagraph, a visual display system
Along with the computer came the DEC systems that relied on the human eye and brain’s ability to fill
engineer responsible for developing the time-shared in missing information, as when a moving car is per-
operating system, who proceeded to spend much ceived through a picket fence [30]. Controlling such
time refining the software—he sometimes used a experiments by computer is widespread today, but it
camp stretcher so he could work on the computer was a new experience for researchers in the 1960s.
overnight [14]. Thus, time-sharing was introduced to Similar advances were made in the fields of crystallog-
Australia, incidentally bolstering Digital Equipment’s raphy, physics, and physiology, among others.
market position there.
LATIN INSCRIPTIONS
SHARED COMPUTING Not all uses of the PDP-6 were based on its online
capability, though that was often exploited at different
Numerous terminals were connected to the PDP-6
stages of the project. An example was Professor E.
over telephone lines. These were originally teletypes,
John Jory’s creation in the early 1970s of an index to
adapted from telex machines, operating at 110 or 300
all the inscriptions found on ancient monuments
bits per second (approximately 10 and 30 characters
around Rome. These had been accumulating since
per second), but the PDP-6 promised more than just
1862 into Volume VI of the Corpus Inscriptionum Lat-
connecting terminals; it was acquired in part to enable
inarum (CIL). The solution was to create a key word in
a range of laboratory experiments to connect and be
context (KWIC) index, but much of the data entry and
directly controlled by the computer. These included a
program development was undertaken online. The
diffractometer and a spectrometer in the Physics
data entry took well over two years, the main problem
Department, an analogue computer and field recorder
being that the inscriptions were two to three thousand
in engineering, a flying spot scanner in the Crystallog-
years old, and every inscription entered produced
raphy Laboratory, a rat race and a perception labora-
something new to be programmed for, because there
tory operated by Psychology, and a flying spot scanner
was no standard format. Resolving these issues with
in the Pathology Department [14]. In the early days,
Jory able to access the data online from his office
these devices were interfaced by an electronics engi-
shortened the project considerably.
neer, Ian Nicholls, employed by the Centre, who under-
When the index’s 7315 pages were published in
took some ground-breaking work [14].
1975 by the Academy of Sciences of the German Dem-
Other agencies around Perth, including CSIRO and
ocratic Republic, Jory received international recogni-
the technical divisions of various government depart-
tion. Since the appearance of Jory’s KWIC index,
ments, had drawn on the capabilities of UWA’s IBM
computer applications and databases have had a
1620. With the advent of the much more powerful PDP-6
major influence on epigraphic studies [8], [10].
in 1965, along with its time-sharing and remote connec-
tivity capabilities, this usage burgeoned, encouraged by
the university who saw this sharing almost as an obliga- UNIVERSITY OF WA FORTRAN
tion given the remoteness of Perth. This laid the ground- TRANSLATOR, UNIWAFT
work for the high degree of collaboration and sharing Another project which did not altogether rely on the
that characterized public sector computing in Western online nature of the PDP-6 was the development of a

www.computer.org/computingedge
October-December 2023 IEEE Annals of the History of Computing 7347
 ANECDOTES
ANECDOTES

Fortran compiler. The PDP-6’s standard Fortran com- government departments were able to connect to it
piler produced correct, fast, and reliable code, but it remotely. UWA encouraged this sharing and estab-
was unsuitable for teaching purposes: it was slow to lished charges for computer use in order to ensure
compile, produced arcane error messages, and equitable access. Many of these external organiza-
stopped analyzing (“parsing”) the program at the first tions in any case had close ties with the engineering
error encountered. Of course, the online nature of the or scientific disciplines within the university, who, like
PDP-6 meant that many users could develop and test the Computer Society, ran regular professional devel-
their Fortran programs online. However, the volume of opment seminars for their alumni. Accordingly, consid-
students wishing to do so outstripped the computer’s erable use was made of the PDP-6 by organizations
capacity to support them simultaneously. The goal like CSIRO, and the technical divisions of the Main
was therefore to develop a fast Fortran compiler that Roads Department, the WA Water Authority, and the
would have rich error messages and which undertook State Electricity Commission. In addition, a number of
as much parsing as possible at each attempt. The private engineering consultancy firms also became
result, released for the start of the 1971 academic customers.
year, was a system called Uniwaft—University of Inevitably, it was not long before the capacity of
Western Australia Fortran Translator. It was itself writ- the PDP-6 was sorely tested. Initially, various upgrades
ten in Fortran, which made coding the system that were made (e.g., additional memory, magnetic tapes,
much faster—and eased its transport to subsequent and disk drives for storage and program “swapping”).
computers [22], [24]. But by 1970, it became clear that a major upgrade or
The system was extended in 1972 to WA State high replacement would be required [8], [14]. As indicated
schools, called Miniwaft, using specially prepared pre- above, major purchases of this kind were funded on a
printed punched cards. Chads on these were pushed triennial basis by the Australian Universities Commis-
out using a paper clip, based on a similar arrangement sion (AUC), an agency of the Australian federal gov-
developed at Monash University [31]. ernment. For the triennium 1970–1972, the AUC had
ruled that it would only support grants to universities
which were prepared to share computing resources, in
ONLINE EDUCATION
order to take advantage of the economies of scale
The potential for online teaching presented by the
available at that time for large computers. Discussions
PDP-6 was not missed, even though systems like Uni-
ensued between the universities and colleges in all the
waft and Miniwaft were batch processing systems.
Australian capitals to attempt to reach agreement.
Before 1977, there was no department of computer
However, only in the case of Western Australia was
science at UWA, but Centre staff were active in the
agreement reached: A heritage of the sharing that had
Australian Computer Society, with the WA Branch
been happening there for several years. No university
having been founded by the then Director, Dennis
in any other city received a computer grant in that tri-
Moore [19]. Staff was keen to promote this new com-
ennium [19].
puter architecture and its implications for the future
In order to set this arrangement for sharing on a
of computing. A series of professional development
more formal basis, it was agreed to set up the Western
seminars was launched, with some using the PDP-6
Australian Regional Computing Centre (WARCC),
[23], [26]. In particular, a simple database package,
housed and managed by the University of Western
Data Management Package (DAMP), that imple-
Australia, under the direction of a Board of Manage-
mented the CODASYL-linked database architecture,
ment made up of representatives of the major users
was used to teach database architecture. In one of
(see [8, Section V]). The primary users would consist of
the first examples of online education in Australia, stu-
the UWA, other tertiary education institutions, state
dents (in this case, Computer Society members) cre-
government departments, and statutory authorities,
ated an online database and interacted online with
and the CSIRO, serving in particular their scientific
their data [26].
and engineering computing needs. The computer
would be expected to support local batch processing,
THE FOUNDATION FOR THE WA remote batch processing, and remote interactive
REGIONAL COMPUTING CENTRE, processing (the role of direct computer-controlled lab-
WARCC oratory experiments by that time had largely been
One of the consequences of having a time-shared taken over by minicomputers). The Centre would oper-
computer, featuring online connections on campus ate along commercial lines, charging all users (includ-
and off, was that the technical divisions of many ing the university) according to the computer time

48
74 ComputingEdge
IEEE Annals of the History of Computing December
October-December 20232024
 ANECDOTES
ANECDOTES

FIGURE 1. Reconstructed IBM 1620 console [photo credit

FIGURE 3. Front entrance to WARCC, UWA [photo credit
Alex Reid].
UWA].

used, and all expenses (including staff salaries) would

be met from this revenue. The university would pro- some time, the decision was made to purchase a
vide accounting and some administrative services, Cyber 72 (a later model of the 6400) from CDC, which
along with power, cleaning, space, etc., and would bill was delivered to the Centre, by then housed in a new
the Centre accordingly. Computer charges were to be extension to the university’s Physics Building, in
set at a level that would cover all expenses, and also August 1972 (see Figure 3).
to allow the accumulation of a small reserve for pur- The Cyber 72 was the most powerful computer yet
chasing additional equipment. installed in WA, and was among the most powerful in
WARCC officially came into existence on 1 January the country (see [8, Section V-E]). It had 60-bit words
1972, after over two years of discussion and negotia- (compared with the PDP-6’s and PDP-10’s 36-bit
tion [8], [13], [14]. With $A470,000 available from the words), and all peripheral input/output activity
1970–1972 Triennium, and the prospect of similar fund- (e.g., disk drives, tape drives, card reader, printer,
ing available from the 1973–1975 Triennium, proposals remote communications) was handled by multiple
were received from Digital Equipment Corporation for peripheral processors, so that the main central proc-
a PDP-10 (a later version of the PDP-6) and Control essing unit was dedicated to calculating, at which it
Data Corporation for a CDC 6400. In the end, and after excelled. The 60-bit words gave it very high precision
in arithmetic calculations, which greatly pleased the
scientists and engineers among its users. This Cyber
72 was later upgraded to a Cyber 73, and subsequently
to a Cyber 720. Curiously, and presumably for market-
ing purposes, the Cyber 72 was in fact identical to the
Cyber 73, but incorporated a circuit designed to slow
it down; upgrading it to a Cyber 73 involved simply
removing that circuit (and the payment of additional
funds). The computer ran the SCOPE (later NOS/BE)
operating system and provided a wide range of scien-
tific and other software. The Uniwaft Fortran compiler
was transported to the Cyber with little difficulty.
Soon, many terminals and a good number of
remote job entry systems were connected to the
Cyber from across Perth. The PDP-6 had been trans-
FIGURE 2. The PDP-6 after delivery to UWA, L to R: DEC Aus- ferred to WARCC in 1972, and its time-sharing feature
tralian sales manager Ron Smart, UWA deputy vice-chancel- was still very much favored for certain applications. In
lor professor John Birkett Clews, UWA Computing Centre due course, therefore, it was decided to replace it with
director Dennis Moore [photo credit UWA].
a compatible, but much faster PDP-10 with its KA10
CPU. Use of these systems continued to grow, and

www.computer.org/computingedge
October-December 2023 IEEE Annals of the History of Computing 49
75
ANECDOTES
ANECDOTES

FIGURE 6. WARCC revenue growth 1972–1989 ([27, p. 1]).

activities ensuring its continued economical operation

FIGURE 4. The main console of the cyber72 [photo credit [32]. By then, WARCC had been operating successfully
for 20 years, with gross revenue growing at an annual
UWA].
compound rate of 25% (see Figure 6), a testament to
the networking among public-sector computer users
WARCC with it, with regular upgrades of the main
in WA, seeded by what the PDP-6 had initiated.
computing equipment funded by the growth in usage
and thence revenue (the PDP-10 was replaced by a
DECsystem-10 with a KL10 CPU in March 1980). At the
same time, the WARCC diversified its services into LIBRARY CIRCULATION SYSTEM,
applications software programming, networking, mini- LOANLY
computer support, and subsequently microcomputer Inspired by the power of online systems, David Noel,
sales and support, and facilities management (e.g., the systems librarian in the University Library, had a
housing and operating the State Health Department’s vision for a loan system where library users recorded
computers). their loans themselves using online terminals [15].
By 1990, WARCC had a staff of over 100, an annual Some public libraries in the U.K. had embarked on this
turnover of $A10 million, and was highly regarded course, but no university or college libraries worldwide
throughout the State (see Figures 5 and 6 [27]). Indeed, could be found to have attempted this. The volumes
the Public Accounts and Expenditure Review Commit- in the UWA library had already each been equipped
tee of the WA State Legislative Assembly strongly rec- with an 80-column punched card, which provided the
ommended the WARCC model of computing for the title, author, and basic identification data; students
rest of the WA public sector [29]. Ultimately, however, and staff of the university had identification cards
mainframe computers became uneconomical in this that resembled abbreviated, plastic, punched cards
environment, and by 1992 those at WARCC were containing their unique staff or student number.
decommissioned, with the Centre’s diversified

FIGURE 5. WARCC sources of revenue 1989 ([27, p. 1]). FIGURE 7. LOANLY users at work [photo credit UWA].

50
76 ComputingEdge
IEEE Annals of the History of Computing December
October-December 2024
2023
 ANECDOTES
ANECDOTES

Consequently the library commissioned WARCC to for use ON-campus, but was unable to get them
develop this system, with custom-built terminals for approved for use off-campus [8], [14].
reading book and borrower cards from a U.K. com- The problems encountered in licensing data com-
pany. A DEC PDP-11/40 was selected, and program- munications equipment was one reason why hardware
ming began at the start of 1973. Despite some initiatives gave way to the use of software to enhance
setbacks, a working system was implemented by mid- data communications capabilities. The main reason
1975 [15], [25]. This integrated the records of books was however the flexibility that software afforded, and
that had been borrowed with records of all the books it was clearly going to be a much more fruitful
in the library’s collection. This enabled remote catalog approach in the long run. Several networking projects
enquiries: if sought items were on loan, they could be were initiated, starting with the clustering of remote
recalled by the enquirers themselves online. Thus was terminals. Some computer manufacturers already had
implemented a radical, self-charging, university circu- “terminal cluster” systems, but they were very expen-
lation system and described at the time as “trend-set- sive and not particularly flexible. It was decided
ting” [4], [16]. Its name, LOANLY, was a homophonic instead to buy small minicomputers (e.g., the DEC
nod to the character in the very popular, if bleak, TV PDP-11/10) and program them to emulate these termi-
spy series Callan. A caricature of Lonely was used in nal clusters. This relatively simple adaptation
promoting the system. increased staff competence in communications soft-
Some valuable lessons about the nature of online ware that opened up many more possibilities [21].
systems were learnt through this project and con- At that time (1973) WARCC operated a Cyber 73; it
veyed to the Australian computing community also still ran the PDP-6 then the PDP-10. It had
through a lecture tour conducted by Alex Reid, become obvious quite early on that these two archi-
together with a lead article in the Australian Computer tectures complemented each other rather well—one
Bulletin [20]. being a powerful batch processor, with the other ideal
Of course, circulation control is just one of many for remote multitasking. Neither firm’s machines could
applications of automation in libraries. Since it is per- undertake both modes well together.
haps where library staff time could save the most A PDP-11/10-based terminal cluster system was
time, it has generally been the first function to be installed at the Western Australian Institute of Tech-
automated. Other applications followed, including nology (later, Curtin University) to connect the insti-
acquisitions, catalog maintenance, serials manage- tute to the WARCC PDP-10. It was also found
ment. Initially developed independently, commercial beneficial to replace the terminal handling system on
providers overtook in-house applications in the mid- the PDP-10 with a similar minicomputer, emulating the
1980s with comprehensive software systems like DEC-supplied interface. Beyond the significant sav-
URICA and LIBERTAS [4], [9]. ings, this move more importantly opened up the
opportunity to adapt both sets of emulators to accom-
modate traffic destined for the Cyber as well as the
PACKET-SWITCHED NETWORKING PDP-10. The Cyber relied heavily on remote job entry
Based on its early start in data communications, con- (RJE) systems, which provided a card reader and
necting ON- and OFF-campus remote terminals to a cen- printer, and a communications link to the Cyber: This
tral computer over telephone lines, WARCC soon enabled remote users to submit their jobs to the
realized that data communications networks repre- Cyber using equipment on their own premises. These
sented significant potential. Initially, the focus had RJE systems were also expensive, and replacing them
been upon hardware developments, to facilitate con- by much cheaper emulators had the added benefit
necting terminals. These included the design and con- that these could be programmed to perform other
struction of modems (to enable reliable transmission functions. One of these was to turn them into conver-
over long distances), and multiplexors—-to enable sational RJEs (CRJEs), whereby remote users could
sharing of communications links by several terminals develop their programs interactively, submit them to
[13]. In the late 1960s and early 1970s, the cost of tele- the Cyber for processing, and interact online with the
phone lines in Australia was relatively prohibitive—in results [13], [21].
part due to its long distances, but also due to the Another challenge which proved to be simple given
monopoly enjoyed by the Postmaster-General’s the background of the WARCC Network team, was
Department (later Telecom), which was not only the interfacing the Cyber and DECsystem-10 directly with
sole provider but was also the regulator. For example, each other, so that data files could be moved between
the Computing Centre staff had developed modems the two systems. It had become clear that the ideal

www.computer.org/computingedge
October-December 2023 IEEE Annals of the History of Computing 7751
 ANECDOTES
ANECDOTES

way to manage data traffic between computers and

end-user devices was to employ “packet switching.”
As early as 1974 the Centre’s staff began researching
the technique, but trod very cautiously in view of
others’ failure in this field. By 1977 packet switching
had become fairly well-established overseas, e.g., with
ARPANET (precursor to the Internet [6], [12]) connect-
ing some universities and research laboratories in the
USA, but it was not extended to Australia. The monop-
oly data communications provider in Australia, Tele-
com, had plans for a common packet-switching
service, but was slow to implement them [3], [19].
Meanwhile, talk of a Western Australian Regional
FIGURE 8. WA regional network in 1978 (from [13]).
Computer Network started to gain momentum, again
with some caution, lest some agencies felt that this
was a political move to grab control of public sector software and protocols. In the end, the same approach
computing [8], [13]. But no one could deny the logic was taken as with the Cyber, that is making the PDP-
that, in a state as large as WA, and with such a scat- 11s look like DEC-10s to the network; the first such link
tered population, it made sense for public sector enti- was installed early in 1978 on the PDP-11/34 at the
ties to share communication lines to remote centers. Nedlands College of Advanced Education over a 2400-
While the political investigations continued, with the bps line. The network expanded rapidly, with a CRJE
State government, for instance, collecting data on being connected, as well as the Cyber 172, at Main
traffic volumes and projections, WARCC continued to Roads Department, along with various other minicom-
investigate the technology. puters and terminal clusters. Toward the end of 1980,
Overall, the goal was to devise a networking envi- when CSIRONET became operational, a two-way gate-
ronment, which would rationalize the diversity of com- way based on a PDP-11/40 was installed. Work contin-
munications links required, while providing a vehicle ued on connecting other public-sector computers,
for both terminal and file traffic between multiple such as the Interdata 8/32 at Murdoch University, and
hosts and other networks. various IBM systems that were installed across the
Computer suppliers were canvassed, but few rose public service, as well as gateways to other national
to the challenge. Ultimately, only DEC’s offering, DEC- and international networks, such as ARPANET [21].
NET, seemed promising, and WARCC launched a Thus, what appears to be the first packet-switched
project to experiment with it. It was found to be sur- multihost network in Australia came into operation by
prisingly straightforward to connect the DEC-10 at early 1978 (see Figure 8). It was recognized that there
WARCC with one that by then had been installed at were several overheads in employing packet-switching
WAIT. By the spring of 1977, a packet-switched multi- (as opposed to use of devices like PABX’s), but it was
host network had been installed between WARCC and clear that this was the path of the future in view of its
WAIT that employed a 4800-bps synchronous line flexibility and versatility, and this has indeed been
rented from Telecom, using PDP-11/40-based DN87 borne out by history. Interestingly, as with ARPANET,
interfaces in front of each DEC-10 [21]. It performed one of the chief benefits of this regional network was
remarkably well. The team then interfaced WARCC’s the widespread uptake in the use of email [21].
Cyber to this network through emulation—making the
Cyber look like a DEC-10 to the network and the net-
work look like a standard terminal cluster to the EPILOGUE
Cyber. No changes were made to the operating sys- Research findings from a variety of sources (e.g., [5])
tems of either the DEC-10 or the Cyber—an important suggest strongly that isolation impedes innovation.
requirement to minimize future support. However, the evidence from the projects described in
By the beginning of 1978 the Cyber had joined this this article, originating in arguably the world’s most
fledgling, packet-switched, multihost network. The isolated city, would seem to belie that formulation.
next phase of expansion was to connect PDP-11 com- The cited research is based on patents and start-ups,
puters at various agencies across Perth. This proved but those are only one measure of innovation. No pat-
surprisingly difficult, as it was discovered that there ents were sought from any of UWA’s initiatives, and
were two incompatible versions of the DECNET there were no notable start-ups, with one significant

52
78 ComputingEdge
IEEE Annals of the History of Computing December
October-December 20232024
 ANECDOTES
ANECDOTES

exception—Monte Sala, whose work is described else- [8] K. Falloon, “Cyberhistory,” M.S. thesis, Univ. Western
where in this issue. Australia, Crawley, WA, Australia, 2001. Accessed: Sep.
The simplest explanation of why so many notable 12, 2023. [Online]. Available: https://research-
initiatives arose at UWA is the early start, which it repository.uwa.edu.au/en/publications/cyberhistory
gained in online systems and networking through its [9] H. Groenewegen, “Four decades of library automation:
bold purchase of Australia’s first time-shared com- Recollections and reflections,” Australian Library J.,
puter. All the initiatives described have at their core vol. 53, no. 1, pp. 39–53, Feb. 2004.
the online paradigm and demonstrate a progressive [10] E. J. Jory, “Problems and prospects for the
realization of computer networking. The networking production of computer compiled indices to
went beyond physical data communications to foster epigraphic works,” Antiquities Africaines, vol. 9,
the close-knit networking of computer users through- no. 1, pp. 15–22, 1975.
out the city of Perth. At the time of these develop- [11] J. A. N. Lee, R. M. Fano, A. L. Scherr, F. J. Corbato, and
ments, Perth was a city of half a million people, but has V. A. Vyssotsky, “Project MAC (time-sharing
since quadrupled in population. It has lost the charac- computing project),” IEEE Ann. Hist. Comput., vol. 14,
teristic which led Dennis Moore to observe that it was no. 2, pp. 9–13, 1992.
small enough for people to be able to talk to each [12] B. Leiner et al., “Brief history of the internet,” Internet
other, but large enough to support technical initiatives Soc., 1997. [Online]. Available: https://www.
[13], [14]. Now that further developments of networking, internetsociety.org/internet/history-internet/brief-
in particular the Internet, have shrunk distances world- history-internet/
wide, Perth can no longer be considered a “backwa- [13] D. Moore, “Computers, communications and
ter”—if indeed it deserved that epithet 60 years ago. cooperation,” Annual WA Comput. Soc. Conf. Bunbury,
1978. Accessed: Sep. 12, 2023. [Online]. Available:
https://alex-reid.com/History/Moore-Bunbury-ACS-
ACKNOWLEDGMENTS 1978.pdf
The author would like to gratefully acknowledge the [14] D. Moore, interview with Penny Collings for the National
very helpful advice given by the reviewers and editors. Library’s, “History of ICT in Australia oral history project,”
Any remaining errors are entirely the author’s recorded Feb. 7, 2015. Accessed: Sep. 12, 2023. [Online].
responsibility. Available: http://catalogue.nla.gov.au/Record/6807543?
lookfor¼ORAL%20TRC%20moore%20%23[format:Audio]
&offset¼4&max¼128, https://alex-reid.com/History/
BIBLIOGRAPHY Dennis-Moore-7Feb15-complete-transcript.htm
[1] G. Bell, C. Mudge, and J. McNamara, Computer [15] D. G. Noel and T. A. Reid, “System and file design in the
Engineering: A DEC View of Hardware Systems Design. LOANLY real-time circulation system,” in Proc. 17th
Bedford, MA, USA: Digital Press, 1978. Biennial Conf., 1973, pp. 479–490.
[2] P. Budne, “DEC PDP-6 serial numbers,” Apr. 2022. [16] D. G. Peake, “Library automation in Australia: The state
Accessed: Sep. 12, 2023. [Online]. Available: https:// of the art,” Prog., Electron. Library Inf. Syst., vol. 15,
www.ultimate.com/phil/pdp10/pdp6-serials.html no. 1, pp. 11–23, 1981.
[3] R. Clarke, “Origins and nature of the internet in [17] Pearcey Foundation, “CSIRAC among the first
Australia,” 2004. Accessed: Sep. 12, 2023. [Online]. electronic stored program computers,” 2021.
Available: https://www.rogerclarke.com/II/OzI04.html Accessed: Sep. 12, 2023. [Online]. Available: https://
[4] L. Clyde and M. Middleton, “Library automation in www.pearcey.org.au/initiatives/csirac/csirac-among-
Western Australia,” Library Autom. Syst. Inf. Exchange, the-first-electronic-stored-program-computers/
vol. 21, no. 1, pp. 5–15, Jul./Aug. 1990. [18] T. Pearcey, A History of Australian Computing.
[5] A. Contigiani and M. Testoni, “Geographic isolation, Dandenong, Australia: Chisholm Inst. Technol., 1988.
trade secrecy, and innovation,” Res. Policy, vol. 52, [19] G. Philipson, “A vision splendid: The history of
no. 8, Oct. 2023, Art. no. 104825, doi: 10.1016/j. Australian computing,” Australian Comput. Soc., 2017.
respol.2023.104825. Accessed: Sep. 12, 2023. [Online]. Available: https://
[6] S. D. Crocker, “Arpanet and its evolution–A report www.acs.org.au/content/dam/acs/acs-publications/
card,” IEEE Commun. Mag., vol. 59, no. 12, pp. 118–124, ACS-ebook-2017_A-Vision-Splendid_The-History-of-
Dec. 2021. Australian-Computing.pdf
[7] B. O. Evans, “System/360: A retrospective view,” IEEE [20] A. Reid, “The trials and tribulations of an on-line
Ann. Hist. Comput., vol. 8, no. 2, pp. 155–179, Apr.–Jun. computer project,” Australian Comput. Bull., vol. 2,
1986, doi: 10.1109/85.150016. no. 1, pp. 6–14, Feb. 1978.

www.computer.org/computingedge
October-December 2023 IEEE Annals of the History of Computing 7953
 ANECDOTES
ANECDOTES

[21] A. Reid, “Network developments at WARCC 1965- [29] E. S. Ripper and G. I. Gallop, “Report on computing in
1980,” Australian Comput. J., 1981. Accessed: Sep. 12, government,” Public Accounts and Expenditure
2023. [Online]. Available: https://alex-reid.com/History/ Review Committee, WA Legislative Assembly, Western
Network-Developments-WARCC-1981.pdf Australia, Australia, 1990.
[22] T. A. Reid, “A guide to Fortran programming and [30] J. Ross, “A new type of display relying on vision’s
UNIWAFT,” UWA Computing Centre, Perth, Western sensitivity to motion,” J. Physiol., vol. 271, no. 2,
Australia, 1971. pp. 2P–3P, Oct. 1977.
[23] T. A. Reid, “Programming for on-line data [31] L. G. Whitehouse, “MINITRAN Monash University
management,” Australian Computer Society student fortran,” Monash Univ. Computer Centre,
Professional Development Seminar, 1971. Melbourne, Australia, 1969.
[24] T. A. Reid, “UNIWAFT— The University of WA Fortran [32] B. Whitford, “Mainframes and micros: Western
translator,” in Proc. 5th Australian Comput. Conf., 1972. Australian regional computing centre,” in Computer
[25] T. A. Reid, D. G. Noel, and C. P. R. Greaves, “LOANLY—- Excellence: Computer Companies in Australia are
A real-time library circulation system,” in Proc. 6th Seeking and Attaining Excellence, B. Whitford, Ed.
Australian Comput. Conf., 1974, p. 45. Perth, Australia: Beaumont Publishing House, 1991,
[26] T. A. Reid and D. G. Moore, “Computer programming and pp. 76–81.
data management for on-line systems,” Univ. Western
Australia Extension Service, 25-week course, 1971. T. ALEX REID is an honorary professorial fellow at the University
[27] T. A. Reid, “Director’s annual report,” WA Regional of Western Australia, Crawley, WA 6009, Australia. Contact him
Computing Centre, Perth, Western Australia, 1990.
at [email protected].
[28] T. A. Reid, “Computing,” in Historical Encyclopedia of
Western Australia, J. Gregory and J. Gothard, Eds.
Perth, Western Australia: UWA Press, 2009,
pp. 223–225.

Get Published in the

New IEEE Transactions
on Privacy
This fully open access journal is now soliciting papers
for review.
IEEE Transactions on Privacy serves as a rapid publication
forum for groundbreaking articles in the realm of privacy and
data protection. Be one of the first to submit a paper and
benefit from publishing with the IEEE Computer Society! With
over 5 million unique monthly visitors to the IEEE Xplore®
and Computer Society digital libraries, your research can
benefit from broad distribution to readers in your field.

Submit a Paper Today!

Visit computer.org/tp to learn more.

54
80 ComputingEdge
IEEE Annals of the History of Computing December
October-December 20232024
IEEE COMPUTER SOCIETY D&I FUND

Drive Diversity
& Inclusion in
Computing

Supporting projects
and programs that
positively impact
diversity, equity, and
inclusion throughout
the computing DONATE TODAY!
community.
 EDITOR: Phil Laplante, The Pennsylvania State University, [email protected]

This article originally

DEPARTMENT: SOFTWARE ENGINEERING appeared in

Lessons From the Father of vol. 56, no. 1, 2023

Software Engineering
Ricardo Valerdi , University of Arizona

The recent passing of Barry Boehm motivated the opportunity to synthesize

lessons from his contributions in software engineering. This article summarizes
six lessons that impacted how software is designed, built, and managed.

F
ormer Brooklyn Dodgers baseball player Jackie the project is at risk of failing. This expanded focus
Robinson once said, “A life is not important captured the many views of how we define success for
except in the impact it has on other lives.” Barry software developments. Rather than emphasizing just
Boehm, known as the “father of software engineering,” performance, Barry valued happy developers, satisfied
was a perfect example of this. His contributions liter- superiors, and pleased users equally.
ally changed how software is builtbuilt—from Beijing The characterization of a manager as a negotiator
to London and everywhere in between. More impor- was innovative at the time because it shifted the empha-
tantly, he influenced academics and practitioners in sis from an organizer of tasks to a “packager of solutions.”
the field to think about how to embed more scientific It also placed a focus on the developer and the user as
rigor and discipline into software engineering. equal participants. Barry understood that people are
It would be impossible to summarize all of Barry’s motivated by good will and self-interest. Accordingly, it
contributions in one article. Instead, it would be more was important to let their value preferences surface to
beneficial to identify some of the lessons learned understand what features were most critical. This idea
that every software developer and software manager applies to any project that has multiple collaborators.
should know. In fact, I strongly believe that Barry’s Whether the objective is to build a sports arena or to
contributions apply beyond the context of software write a new corporate policy, Theory W reminds us that
development. Their relevance to product development the most important thing is to ensure which success
and technology management is driven by the broad critical stakeholders need to be satisfied.
applicability of his work. Here are six lessons I learned
from having Barry as my mentor during my doctoral LESSON 2: RESOLVE
studies and subsequently as a collaborator for the MODEL CLASHES
past 20 years. (See also “Barry on the Tennis Court.”) Sometimes it is impossible to make everyone happy.
Barry developed a simple way to think about such
LESSON 1: STRIVE FOR WIN-WIN “sticky” situations through the model-based systems
Barry’s objective function was to make everyone a architecting and software engineering approach.2
winner. He referred to this as win-win, or Theory W. The Disagreements between parties were reframed into
idea was that software project managers will be fully “clashes” between them and represented as differ-
successful if and only if they make winners of all of the ent types of models: product models, process mod-
other participants in the software process: superiors, els, property models, and success models. Through
subordinates, customers, users, maintainers, and so this reframing of preferences, it became possible for
on.1 If any success critical stakeholder is left out, then success-critical stakeholders to visualize or reason
about the prospective system and its likely effects to
better deal with them. These models captured product
Digital Object Identifier 10.1109/MC.2022.3219527 attributes; the process for developing the system; prop-
Date of current version: 9 January 2023 erties such as cost, performance, or dependability; what
it means for the system to be successful; and the means

56 December 2024 Published by the IEEE Computer Society  2469-7087/24 © 2024 IEEE
 to resolve the many conflicts that occur as the product
BARRY UNDERSTOOD THAT PEOPLE
is architected, designed, developed, tested, and fielded.
ARE MOTIVATED BY GOOD WILL AND
An example of a model clash is when a product model
SELF-INTEREST.
aims for high reliability and high performance, while a
property model aims for low cost and a short develop-
ment schedule. Barry’s approach gave stakeholders a where the end result is not well defined and would ben-
way to surface their preferences, perform tradeoffs, and efit from incremental steps that provide information and
come to the negotiating table with specific needs. This feedback along the way. Each iteration, or spiral, allows
facilitated negotiations about things that mattered the stakeholders to identify risks, experiment with solutions,
most by requiring collaborators to prioritize and reach identify corrective actions, and reevaluate the results,
consensus on what was important. The simplicity of this which would ultimately lead to better products.
approach facilitated its use throughout the software
community. But model clashes occur in areas outside of LESSON 4: A SOLUTION
software too. As a result, the approach was put to work THAT IS TOO EXPENSIVE IS NOT
across systems, hardware, and software communities A SOLUTION
to resolve conflicts and reach decisions as to what was Economic analysis techniques were relatively new to
best for all of those concerned. software engineering in the 1980s. Limited computing
power and memory increased the importance of cost-
LESSON 3: ITERATE TO GAIN benefit analyses of software product features. Barry
KNOWLEDGE AND UNDERSTAND knew that “it is often worth paying for information
STAKEHOLDER PREFERENCES because it helps us make better decisions.”4 To make
Another of Barry’s fundamental contributions was the this possible, he developed the constructive cost
idea that software development should be an iterative model (COCOMO). Besides helping people understand
process, rather than a specification-driven process. This the cost consequences of their decisions, the model
approach, appropriately called the spiral model, was cre- allowed its users to conduct tradeoffs and come up
ated to overcome the limitations of the waterfall model, with a plan that helped them realize the goals set by
which recommended that software be developed in suc- collaborators. Barry elevated the importance of cost
cessive stages.3 Such top-down structured approaches as a critical decision criterion for projects and democ-
had other difficulties, such as assuming uniform pro- ratized the ability for people to generate their own
gression of the system’s evolution and the inability to cost estimates based on a set of parameters that were
accommodate software reuse, among other things. known to influence the cost of software.
The innovation of the spiral model is that it reframed COCOMO became the standard measuring stick
software development as an incremental process with for industry and government to better understand
multiple milestones along the way. Ultimately, the goal how to keep software development affordable. It also
was to identify risks early in the development process allowed decision makers to quantify the effects of differ-
and resolve them as early as possible. The spiral model ent parameters on cost, such as software reliability, pro-
eventually became the universal software development grammer capability, and project schedule. The broad
process. Barry once said that he knew he had made it adoption of COCOMO gave rise to the development
when he read about the spiral model in a Dilbert cartoon. of algorithmic models that could be used to forecast
Such an iterative approach can be useful in situations the cost and schedule of projects in their early stages.

www.computer.org/computingedge 57
SOFTWARE ENGINEERING

BARRY ON THE TENNIS COURT

B arry Boehm was known to many as a humble intel-

lectual—a gentle giant in our field. But if you ever
faced him on the tennis court, you would witness an en-
way to hit a forehand was with a long backswing and a long
follow-through. One-handed backhands were the norm, and
(as Barry demonstrated) the sooner you came up to the net
tirely different side of him. With a racquet in his hand, the to hit a volley, the more likely you were to win the point.
man was a beast. His style of play was aggressive, and Barry’s game wasn’t his only secret weapon. He had
his stamina on the court was on par with a Wimbledon home court advantage. This came from knowing where
champion because he ran three miles every morning. all of the cracks were on the tennis court. You see, Barry’s
When I first met Barry in the early 2000s as a new house was near the Santa Monica fault line, which meant
graduate student in his lab at USC, he was wearing a that the ground moved over time. This led to sporadic
Wimbledon sweater. I didn’t think much of it. I thought he cracks on one side of the tennis court. They were small
was simply a tennis fan. Barry walked slowly, spoke qui- cracks, but if a tennis ball bounced on one, there was no
etly, and took veeeeeery long pauses between thoughts. way of predicting where it would go. Barry seemed to al-
I asked him if he played tennis, and he confidently said, ways aim for the cracks and would occasionally hit them.
“I do!” and invited me to play. I thought we would meet When he did, it was impossible to return the ball. It was
at the courts on campus or maybe at a park in Santa an unfair match. An opponent in great physical condi-
Monica. Instead, he invited me to his house because he tion with an aggressive approach and the home court
had his own tennis court in his backyard. That’s when I advantage was a difficult one to beat.
realized my professor was serious about tennis. During the five years of being Barry’s grad student, I
Going to someone’s house to play tennis is not an did not beat him once. As soon as I defended my doctoral
everyday occurrence, especially in urban Los Angeles. dissertation, I decided it was time for his streak to end.
Nevertheless, I was confident in my tennis game because (By this time I had learned where all the cracks were.) But,
I had played competitive tennis since I was young. When I more importantly, I had learned to never hit it to his fore-
showed up at his house, Barry was dressed in old, beat-up hand because it was his aggressive side. I loved playing
tennis attire from the 1980s. His wardrobe featured a tennis with Barry because he loved to compete. Initially, I
partially ripped shirt, shoes with holes in them, and glasses was embarrassed to lose to someone 40 years older than
that fit crookedly on his face. I soon learned that this was me, but he was the best player of his generation I had
simply a facade. He was about to rain down aces on me and ever played against. When I am his age, I hope to be half
approach the net on every single point. Every point. Who as talented, half as humble, and half as competitive. I will
does that? Barry did. Because Barry learned to play tennis dearly miss my professor and tennis partner, but I recog-
in the 1950s when the racquets were wooden and the only nize that the ball is in my court to carry on his legacy.

LESSON 5: BE PROBLEM DRIVEN

IF TWO DESIGNS HAVE THE SAME
FUNCTIONALITY BUT DIFFER IN
INSTEAD OF METHOD DRIVEN
As a researcher, Barry was not limited to a fixed set
COSTS, A MODEL LIKE COCOMO CAN
of tools or methods. Instead, he was motivated by
HELP BREAK THE TIE.
real-world problems that required robust solutions and
an open mindset. The best example is the COCOMO
This capability is also helpful when comparing pos- suite of models, which was motivated by questions
sible design alternatives and conducting a wide range that the basic COCOMO model could not answer. The
of tradeoffs. If two designs have the same function- COCOMO model answered two questions: How many
ality but differ in costs, a model like COCOMO can software developers do I need to build this system,
help break the tie. Alternatively, if a project exceeds and how long will they take to complete the project?
a budget, then either its forecasted cost or schedule As the sophistication of software products evolved
should be adjusted or it should be declared infeasible. over time, new development paradigms emerged, the

58 ComputingEdge December 2024

 SOFTWARE ENGINEERING

effects of which could not be covered by COCOMO.

One of these paradigms was the increased use of THE UNDERLYING PREMISE OF THE
commercial off-the-shelf (COTS) software to acceler-
CONE OF UNCERTAINTY IS THAT
RISK NEEDS TO BE QUANTIFIED
ate development schedules and reduce labor costs.
AND TRADEOFFS MUST BE MADE TO
Another was the deployment of agile methods. To
PERFORM THIS QUANTIFICATION.
address these new challenges, Barry developed new
calibrations or other models, like COCOTS.5
This important lesson from Barry’s work also
highlights his versatility as a researcher. By being schedule, cost versus reliability, schedule versus person-
problem driven, Barry focused on identifying a prob- nel availability, and so on). Such tradeoffs are important
lem to be solved and then worked diligently to solve in light of current development frameworks because
it. This resulted in an impactful body of work that has they allow decision makers to look at the big picture
left a lasting legacy in software engineering. Such an rather than just at the economics of the situation.
approach can apply anywhere there are challenging
problems to solve. Being method driven might still
be helpful in some situations, but one could become
trapped by “having a hammer and only being interested
J ust as Babe Ruth was known as the greatest base-
ball player of all time, Barry is considered to be the
best software engineer of all time. Barry was a Fellow
in nails.” Barry has demonstrated that being problem of multiple professional societies, including IEEE, and
driven can be much more gratifying in the long run. he had the power to convene thought leaders from
around the globe to move the field forward. Barry was
LESSON 6: an intellectual giant: every one of his contributions
EMBRACE UNCERTAINTY was a home run. Most important of all, Barry was the
Barry’s approach to being risk driven in a world with Most Valuable Person.
model clashes was centered on the idea that life is
full of uncertainties. To illustrate this notion, he pop- REFERENCES
ularized the cone of uncertainty, which is the idea 1. B. W. Boehm and R. Ross, “Theory-W software project man-
that uncertainty decreases at an increasing rate as a agement principles and examples,” IEEE Trans. Softw. Eng.,
project progresses.4 Cost estimates are subject to a vol. 15, no. 7, pp. 902–916, Jul. 1989, doi: 10.1109/32.29489.
high degree of uncertainty at the beginning of a proj- 2. B. W. Boehm, “Escaping the software tar pit: Model clashes
ect when information is limited. As more information and how to avoid them,” SIGSOFT Softw. Eng. Notes, vol. 24,
is obtained from prototypes, user feedback, and itera- no. 1, pp. 36–48, Jan. 1999, doi: 10.1145/308769.308775.
tive development, the uncertainty tends to decrease. 3. B. W. Boehm, “A spiral model of software development
The applications of the cone of uncertainty extend and enhancement,” ACM SIGSOFT Softw. Eng. Notes, vol.
to scenarios where uncertainty must be accounted for 11, no. 4, pp. 14–24, Aug. 1986, doi: 10.1145/12944.12948.
in a forecast. This notion is used in hurricane forecast- 4. B. W. Boehm, Software Engineering Economics.
ing, where the projected path of a hurricane is bounded Englewood Cliffs, NJ, USA: Prentice-Hall, 1981.
by an error cone. Instead of uncertainty decreasing as 5. B. W. Boehm et al., Software Cost Estimation With
in software development, uncertainty increases to COCOMO II. Englewood Cliffs, NJ, USA: Prentice-Hall,
reflect the error associated with the possible paths 2000.
of a storm. In both cases, uncertainty is incorporated
into a forecast, which leads us to embrace uncertainty
by modeling it rather than being intimidated by it. RICARDO VALERDI is a professor and head of the Depart-
The underlying premise of the cone of uncertainty ment of Systems & Industrial Engineering at the University
is that risk needs to be quantified and tradeoffs must of Arizona, Tucson, AZ 85721 USA. He completed his Ph.D. at
be made to perform this quantification. Such tradeoffs the University of Southern California under Barry Boehm’s
allow “value” to be determined as different parameters mentorship and played many tennis matches against him.
are evaluated and tradeoffs are made (cost versus Contact him at [email protected].

www.computer.org/computingedge 59
 Conference Calendar

I EEE Computer Society conferences are valuable forums for learning on broad and dynamically shifting top-
ics from within the computing profession. With over 200 conferences featuring leading experts and thought
leaders, we have an event that is right for you. Questions? Contact [email protected].

JANUARY MARCH and Trust wor thy Machine

4 January 1 March Learning), Copenhagen,
• VLSID (Int’l Conf. on VLSI • HPCA (IEEE Int’l Symposium Denmark
Design and Int’ l Conf. on on High Performance Com- 16 April
Embedded Systems), Banga- puter Architecture), Las Vegas, • COOL CHIPS (IEEE Sympo-
lore, India USA sium on Low-Power and High-
14 January 4 March Speed Chips and Systems),
• ICOIN (Int’l Conf. on Infor- • SANER (IEEE Int’l Conf. on Tokyo, Japan
mation Networking), Chiang Software Analysis, Evolution 22 April
Mai, Thailand and Reengineering), Montreal, • PacificVis (IEEE Pacific Visu-
27 January Canada alization Conf.), Taipei City,
• AIxVR (IEEE Int’ l Conf. on 8 March Taiwan
Ar tif icial Intelligence and • VR (IEEE Conf. Virtual Reality 26 April
eXtended and Virtual Reality), and 3D User Interfaces), Saint • ICSE (IEEE/ACM Int’l Conf.
Lisbon, Portugal Malo, France on Software Eng.), Ottawa,
17 March Canada
FEBRUARY • PerCom (IEEE Int’l Conf. on Per- 28 April
9 February vasive Computing and Com- • VTS (IEEE VLSI Test Sympo-
• BigComp (IEEE Int’l Conf. on munications), Washington, DC, sium), Tempe, USA
Big Data and Smart Comput- USA
ing), Kota Kinabalu, Malaysia 31 March MAY
17 February • DATE (Design, Automation & 4 May
• ICNC (Int’l Conf. on Computing, Test in Europe Conf.), Lyon, • ARITH (IEEE Symposium on
Networking and Communica- France Computer Arithmetic), El Paso,
tions), Honolulu, Hawaii, USA • ICSA (IEEE Int’l Conf. on Soft- USA
26 February ware Architecture), Odense, • FCCM (IEEE Annual Int’l Sym-
• VISIGRAPP (Int’l Joint Conf. Denmark posium on Field-Program-
on Computer Vision, Imaging • ICST (IEEE Conf. on Software m a b l e C u s t o m C o m p u t-
and Computer Graphics The- Testing, Verification and Vali- ing Machines), Fayetteville,
ory and Applications), Porto, dation), Napoli, Italy USA
Portugal • MOST (IEEE Int’l Conf. on
• WACV (IEEE/CVF Winter Conf. APRIL Mobility, Operations, Services
on Applications of Computer 9 April and Technologies), Newark,
Vision), Tucson, USA • SaTML (IEEE Conf. on Secure USA

60 December 2024 Published by the IEEE Computer Society  2469-7087/24 © 2024 IEEE
5 May Computer Vision and Pattern 7 July
• HOST (IEEE Int’l Symposium Recognition), Nashville, USA • IOLTS (IEEE Int’l Symposium
on Hardware Oriented Secu- 18 June on On-Line Testing and Robust
rity and Trust), San Jose, USA • CBMS (IEEE Int’l Symposium System Design), Ischia, Italy
11 May on Computer-Based Medical • SERVICES (IEEE World Con-
• ISPASS (IEEE Int’l Symposium Systems), Madrid, Spain gress on Services), Helsinki,
on Performance Analysis of • ICHI (IEEE Int’l Conf. on Health- Finland
Systems and Soft ware), Ghent, care Informatic s), Rende, 8 July
Belgium Italy • COMPSAC (IEEE Annual Com-
12 May 21 June puters, Soft ware, and Applica-
• SP (IEEE Symposium on Secu- • ISCA (ACM/IEEE Annual Int’l tions Conf.), Toronto, Canada
rity and Privacy), San Fran- S y mp osium on C omputer 15 July
cisco, USA Architecture), Tokyo, Japan • ICALT (IEEE Int’ l Conf. on
19 May 23 June Advanced Learning Technolo-
• CCGrid (IEEE Int’l Symposium • CSF (IEEE Computer Secu- gies), Changhua, Taiwan
on Cluster, Cloud and Internet rity Foundations Symposium), 21 July
Computing), Tromsø, Norway Santa Cruz, USA • ICCP (IEEE Int’l Conf. on Com-
• ICDE (IEEE Int’l Conf. on Data • DSN (Annual IEEE/IFIP Int’l p u t at i o na l Ph o t o g r a p hy),
Eng.), Hong Kong Conf. on Dependable Systems Toronto, Canada
26 May and Networks), Naples, Italy • ICDCS (IEEE Int’l Conf. on Dis-
• FG (IEEE Int’l Conf. on Auto- • SVCC (Silicon Valley Cyberse- tributed Computing Systems),
matic Face and Gesture Recog- curity Conf.), San Francisco, Glasgow, United Kingdom
nition), Tampa/Clearwater, USA USA
26 June
JUNE • IEEE Cloud Summit, Washing-
2 June ton, DC, USA
• MDM (IEEE Int’ l Conf. on 30 June
Mobile Data Management), • EuroS&P (IEEE European Sym-
Irvine, USA posium on Security and Pri-
3 June vacy), Venice, Italy
• IPDPS (IEEE Int’l Parallel and • ICME (IEEE Int’l Conf. on Mul- Learn more
Distributed Processing Sym- timedia and Expo), Nantes,
about IEEE
posium), Milano, Italy France
4 June Computer
• ISMVL (IEEE Int’l Symposium JULY Society
on Multiple-Valued Logic),
Montreal, Canada
6 July
• ISVLSI (IEEE Computer Society
conferences
10 June Annual Symposium on VLSI), computer.org/conferences
• CVPR (IEEE/CVF Conf. on Kalamata, Greece
 IEEE COMPUTER SOCIETY CAREER CENTER

Career Accelerating
Opportunities
Explore new options—upload your resume today

careers.computer.org

Changes in the marketplace shift demands for vital skills and talent.
The IEEE Computer Society Career Center is a valuable resource tool
to keep job seekers up to date on the dynamic career opportunities
offered by employers.
Take advantage of these special resources for job seekers:

JOB ALERTS TEMPLATES WEBINARS

CAREER RESUMES VIEWED

ADVICE BY TOP EMPLOYERS

No matter what your career level, the IEEE Computer

Society Career Center keeps you connected to
workplace trends and exciting career prospects.