FEDERATED LEARNING FOR INDUSTRIAL INTERNET OF THINGS

Enabling Trustworthy Federated Learning

in Industrial IoT: Bridging the Gap Between
Interpretability and Robustness
Senthil Kumar Jagatheesaperumal, Mohamed Rahouti, Ali Alfatemi, Nasir Ghani, Vu Khanh Quy, and Abdellah Chehri

Abstract manage data efficiently and securely, as highlight-

Federated Learning (FL) represents a paradigm ed in Nguyen’s study [3]. Particularly in IIoT net-
shift in machine learning, allowing collaborative work scenarios, which demand high throughput,
model training while keeping data localized. This low latency, and precise anomaly detection, FL’s
approach is particularly pertinent in the Industri- role becomes even more crucial. By incorporating
al Internet of Things (IIoT) context, where data deep reinforcement learning within an FL frame-
privacy, security, and efficient utilization of dis- work, FL-based anomaly detection techniques sig-
tributed resources are paramount. The essence nificantly reduce privacy risks, a concept validated
of FL in IIoT lies in its ability to learn from diverse, in Wang’s research [4].
distributed data sources without requiring central However, the challenge lies in integrating FL,
data storage, thus enhancing privacy and reduc- IIoT, and trustworthiness on a single platform, espe-
ing communication overheads. However, despite cially within the IIoT context, which is a relatively
its potential, several challenges impede the wide- unexplored area. The challenge is especially due
spread adoption of FL in IIoT, notably in ensuring to the tiny footprint of the IIoT devices to withhold
interpretability and robustness. This article focus- the FL model, and subsequently ensure security
es on enabling trustworthy FL in IIoT by bridging and privacy for the devices and model. Addressing
the gap between interpretability and robustness, this, the chameleon hash method with a config-
which is crucial for enhancing trust, improving urable trapdoor, as proposed in Wei’s study [5],
decision-making, and ensuring compliance with tackles vulnerabilities by employing FL for priva-
regulations. Moreover, the design strategies sum- cy-preserving data analysis in IIoT. This method has
marized in this article ensure that FL systems in been effectively implemented in a redactable med-
IIoT are transparent and reliable, vital in industrial ical blockchain, demonstrating enhanced accuracy
settings where decisions have significant safety and efficiency. Additionally, Li’s research introduc-
and economic impacts. The case studies in the es a multi-tentacle federated learning (MTFL) archi-
IIoT environment driven by trustworthy FL mod- tecture for the software-defined IIoT framework,
els are provided, wherein the practical insights comprising a stochastic tentacle data exchanging
of trustworthy communications between IIoT sys- (STDE) protocol and an efficient poisoning attack
tems and their end users are highlighted. detection algorithm [6]. These innovations collec-
tively mark significant advancements in the field.
Introduction Furthermore, federated cybersecurity, a decen-
Federated Learning (FL) has emerged as one of the tralized approach through FL, plays a pivotal role
promising platforms in collaboration with Artificial in identifying and mitigating security threats in the
Intelligence (AI), especially when applied to Indus- IIoT, as discussed in Ghimire’s study [7]. Federat-
trial Internet of Things (IIoT) networks [1]. With the edTrust, for instance, ensures reliable FL models
core elements of FL, it can be applied to intelligent by employing thirty metrics and six pillars to evalu-
IIoT systems by facilitating AI training at the net- ate trustworthiness in IoT and edge computing, as
work edge. Incorporating FL into IIoT networks noted by Sanchez [8]. This underscores its practi-
makes it possible to address scalability issues while cal application in IoT security, particularly in smart
also addressing concerns related to user privacy industries. Despite these developments, the exist-
and the confidentiality of industrial data [2]. ing literature on trustworthy FL, especially in IIoT,
The integration of FL across diverse indus- remains limited, lacking a forward-looking vision
tries such as smart cities, healthcare, and trans- for the potential, enabling technologies, challeng-
portation is revolutionizing data-sharing methods, es, and various design aspects of employing FL in
attack detection, and privacy preservation. These IIoT networks.
industries, closely linked with various Internet of The research questions tackled in this article
Things (IoT) services, benefit from FL’s ability to are outlined as follows:

Senthil Kumar Jagatheesaperumal is with Mepco Schlenk Engineering College, Sivakasi, Tamil Nadu, India; Mohamed Rahouti and Ali
Alfatemi are with Fordham University, USA; Nasir Ghani is with the University of South Florida, USA; Vu Khanh Quy is with Hung Yen Uni-
Digital Object Identifier: 10.1109/IOTM.001.2300274 versity of Technology and Education Hung Yen, Vietnam; Abdellah Chehri is with the Royal Military College of Canada (RMC), Canada.

38 2576-3180/24/$25.00 © 2024 IEEE IEEE Internet of Things Magazine • September 2024

Authorized licensed use limited to: Northeastern University. Downloaded on March 05,2025 at 19:46:26 UTC from IEEE Xplore. Restrictions apply.
 • How to achieve a balance between interpret- • Edge computing paradigms: By processing
Adversarial training
ability and robustness in FL systems within the data close to where it’s generated, edge
IIoT, ensuring both transparency for trust and computing reduces latency and bandwidth strengthens FL mod-
compliance, and resilience against diverse use, which is crucial for real-time applications els against malevo-
industrial challenges. in IIoT.
• What design strategies are effective for cre- Furthermore, to address the dual challenges of lent actors trying to
ating FL systems in IIoT that are both trans- interpretability and robustness in FL, several strate- impact model integ-
parent and reliable, considering the unique gies have been employed:
rity in the setting of
industrial requirements where decisions have
significant safety and economic impacts? Interpretability Enhancements IIoT, where security
• How can FL be adapted to meet the specific • Layer-wise relevance propagation (LRP): It is risks are common.
demands and characteristics of IIoT environ- one of the powerful XAI-based ML approach-
ments, including the aspects of data privacy, es that hold the capability to scale towards
security, and efficient utilization of distributed even the most complex DNNs. In IIoT, where
resources? decision-making processes must often be
These questions aim to address the complexities transparent and justifiable, LRP can be partic-
of integrating FL into IIoT with a focus on making ularly valuable. It allows for a detailed under-
these systems trustworthy and suitable for indus- standing of how input data influences the
trial applications. To address these questions, we model’s output, which is essential for trust in
present the following contributions: automated systems.
• We shed light on how to unlock the potential • Attention mechanisms: These mechanisms
of responsible and robust FL to transform IIoT can be specifically tuned to the data types
networks into a self-sustaining architecture. most commonly encountered in IIoT, like
• For the first time in the literature, we provide time-series sensor data, highlighting the criti-
a holistic overview of the design methodol- cal aspects for predictions or classifications.
ogies and challenging environments of IIoT.
Also, we envision the idea of integrating Robustness Enhancements
hybrid trustworthy FL design approaches for To strengthen the model’s robustness against pos-
IIoT networks. sible attacks, “adversarial training” entails supple-
• We sketch a road map with four different menting training data with adversarial instances
case studies along with the investigation for that are designed to trick the model. Adversarial
the successful realization of trustworthy FL in training strengthens FL models against malevolent
IIoT-empowered networks. actors trying to impact model integrity in the set-
The rest of this article is organized as follows. ting of IIoT, where security risks are common. On
The interpretability and robustness enhancements the other hand, the term “byzantine-resistant aggre-
in FL are covered. We present the design method- gation” describes methods used in the FL process
ologies of FL through deterministic, probabilistic, to reduce the impact of attacks. Byzantine-resistant
and adaptive approaches. This is followed by IIoT aggregation guarantees that malicious nodes do
case studies on manufacturing, energy, supply not disrupt the collaborative learning process in
chain, and environment monitoring. We then con- IIoT settings, where data integrity is crucial. This
clude the article and highlight future directions. improves the dependability and credibility of FL
models used in IIoT contexts. These methods are
Responsible and Robust Federated Learning essential for preserving FL models’ integrity in the
FL has emerged as a crucial technology, enabling face of possible hostile impacts and protecting IIoT
decentralized machine learning while address- systems’ efficacy and security.
ing key issues like communication efficiency, data • Adversarial training: In IIoT, where systems
privacy, and model accuracy. This is particularly can be targets for cyber-attacks, training
relevant in IIoT, where devices often operate in models with adversarial examples can signifi-
resource-constrained environments, and the integ- cantly improve their resilience, ensuring more
rity and confidentiality of data are paramount. reliable performance in hostile environments.
Recent advancements in FL for IIoT focus on opti- • Byzantine-resistant aggregation algorithms:
mizing communication efficiency, data privacy, Given the distributed nature of IIoT and the
and model accuracy [9]. potential for compromised nodes, these algo-
rithms are crucial for maintaining the integrity
FL Enhancement Techniques in IIoT Settings of the FL process, ensuring that malicious or
• Model compression: This technique reduces erroneous updates do not skew the model.
the size of the FL models, making them more The advancements in FL for IIoT aim to improve
suitable for transmission over networks with performance and prioritize the systems’ robustness
limited bandwidth, which is a common con- and transparency. These qualities are crucial for
straint in IIoT environments. FL’s acceptance and effective operation in indus-
• Differential privacy: This approach adds trial environments. The responsible and robust
noise to the data or model updates, thus approach to FL in the IIoT sets the stage for devel-
ensuring that individual data points cannot oping more secure, efficient, and reliable industrial
be reverse-engineered, an essential consider- automation and monitoring systems.
ation for maintaining data privacy in IIoT. Enhancements in FL for IIoT address particular
• Secure multi-Party computation: This meth- requirements such as managing heterogeneous
od enables multiple parties to collaboratively data, maintaining privacy, and supporting devices
compute a function over their inputs while with limited resources. Furthermore, adversarial
keeping those inputs private, which is vital in resilience against cyber-attacks should be given
collaborative IIoT settings. priority by FL approaches, as should the facilitation

IEEE Internet of Things Magazine • September 2024 39

Authorized licensed use limited to: Northeastern University. Downloaded on March 05,2025 at 19:46:26 UTC from IEEE Xplore. Restrictions apply.
 Coatings and seals
of real-time adaptation for prompt decision-mak-
ing. By guaranteeing strong model training, data
Deterministic Approaches
that can withstand privacy, and robustness against adversarial assaults Designing robust and efficient systems for chal-
harsh chemicals are in dynamic IIoT contexts, these improvements set lenging environments, particularly in industrial set-
apart IIoT-focused FL from other areas. tings, is a complex task that requires innovative
essential for pro- and resilient methodologies. This is particularly
tecting the devices’ Design Methodologies for Challenging true when implementing FL in the context of the
Industrial IoT and manufacturing processes.
internal and external Environments The reliability and performance of FL are put
components. to the test in high-temperature IIoT environments,
Guidelines for Equipping FL Architectures such as steel manufacturing or chemical process-
Consideration of different emerging technologies ing plants. According to Gaddam et al. [10], high
in association with FL architectures to handle temperatures can impact hardware performance
issues in particularly challenging contexts environ- and data transmission, which can disrupt learn-
ments, demands the following general principles ing processes. To address these challenges, it is
and guidelines for improving FL architectures in crucial for design methodologies to prioritize the
such settings: development of hardware and software that can
When equipping FL architectures, certain effectively withstand extreme temperatures. This
design considerations largely depend on optimiz- involves the utilization of heat-resistant materi-
ing for IIoT devices with limited resources and als in the construction of devices, as well as the
managing data heterogeneity using adaptive implementation of advanced cooling technolo-
learning rates and customized preprocessing tech- gies. On the software side, optimizing algorithms
niques. Incorporating adversarial robustness tech- for efficient data processing is crucial to minimize
niques like adversarial training and anomaly the duration the hardware is exposed to high tem-
detection, integrating robust privacy-preserving peratures. Furthermore, by implementing robust
mechanisms like FL with encryption or differ- data synchronization methods, it is possible to
ential privacy, and supporting real-time adapta- ensure consistent performance of FL even when
tion through online learning and adaptive FL there are thermal fluctuations.
algorithms are also necessary. Furthermore, the In FL for IIoT, decentralization improves
design demands lightweight model architectures dependability by lowering reliance on single
and edge computing solutions. FL architectures points of failure, encouraging redundancy, and
can successfully meet the special requirements of lowering data loss risks. Decentralized techniques
IIoT environments by addressing these issues and enhance system resilience, privacy, and security
design considerations. This allows for collabora- by decreasing vulnerability to centralized weak-
tive model training while protecting data privacy, nesses and aggregating insights locally, notwith-
guaranteeing model robustness, and promoting standing the constraints associated with gathering
prompt decision-making based on dynamic IIoT data from remote sources.
data streams. In environments where corrosive substances
Robust Communication and Processing: are prevalent, such as in certain types of chem-
Implement resilient and decentralized communi- ical manufacturing, traditional computing, and
cation protocols with an asynchronous approach, IoT devices can degrade quickly, hindering the
utilizing edge computing for local data processing effectiveness of FL [11]. To address this, design
to cope with unreliable connectivity. Additionally, methodologies must include the use of corro-
the incorporation of redundancy and fault-toler- sion-resistant materials in device fabrication. Coat-
ant methods as a proactive measure to effectively ings and seals that can withstand harsh chemicals
manage device failures in challenging scenarios is are essential for protecting the devices’ internal
recommended. and external components. From a software per-
Resource-Efficient and Adaptive Learning: spective, strategies such as redundant data paths
Develop FL algorithms that prioritize real-time and error-checking mechanisms can help maintain
processing and energy efficiency, considering the data integrity in these harsh conditions. Further-
constraints of demanding situations. Additionally, more, predictive maintenance algorithms integrat-
it is required to design adaptable models capable ed into the FL system can preemptively identify
of responding to dynamic environments, incor- and address potential issues caused by corrosive
porating reinforcement learning techniques for elements, thereby reducing downtime and main-
continuous improvements. taining steady performance.
Security and Privacy Preservation: Enhance In both high-temperature and corrosive envi-
data protection by integrating privacy-preserving ronments, the key is to develop FL systems that
technologies such as homomorphic encryption are resilient to physical challenges and capable of
and employing secure communication protocols. adapting to environmental changes. This involves
Concurrently, optimize data collection by lever- a continuous cycle of monitoring, learning, and
aging sensor technologies in collaboration with FL adapting, where the FL system evolves to main-
frameworks designed for robustness. tain optimal performance. Such deterministic
Continuous Monitoring and Collaboration: approaches in the design of FL systems are crucial
Establish continuous monitoring systems for FL for their successful deployment and operation in
model performance in challenging environments. challenging industrial environments.
Subsequently, it is recommended to foster inte-
gration among various sensor technologies, imple- Probabilistic Strategies
ment feedback loops, and evaluate FL structures In challenging IIoT environments, which may
based on real-world performance data from actu- include factors like varying network conditions,
al applications. fluctuating data quality, and unpredictable hard-

40 IEEE Internet of Things Magazine • September 2024

Authorized licensed use limited to: Northeastern University. Downloaded on March 05,2025 at 19:46:26 UTC from IEEE Xplore. Restrictions apply.
 IEEE INTERNET OF THINGS MAGAZINE 4

ware performance, probabilistic models can

enhance the resilience of FL systems. These models
account for uncertainties and variabilities in the
data and operational conditions [12]. By incor-
porating probabilistic reasoning, FL systems can
make more informed decisions, even in the face of
incomplete or noisy data. Techniques like Bayesian
inference can be used to update models dynami-
cally as new data becomes available, providing a
robust framework for learning under uncertainty.
To further adapt FL to challenging environ-
ments, techniques that focus on environmental
adaptability are essential. This includes develop-
ing adaptive algorithms that can adjust learning
parameters in response to changes in the environ-
ment. For instance, reinforcement learning strate-
gies can be employed, where the FL system learns
optimal actions through trial and error, adapting
to the environment’s dynamics [13]. Furthermore,
integrating reliable outlier detection and handling
mechanisms guarantees that the FL models are not
excessively impacted by abnormal data, which is
frequently encountered in unstable environments.
Overall, probabilistic strategies in FL offer a
promising avenue for designing systems resilient
to the uncertainties and variabilities of challeng-
ing IIoT environments. By leveraging these tech-
niques, FL systems can maintain performance and
reliability, even under adverse conditions.

AdAptIve hybrId ApproAches

The dynamic design fusion in FL involves com-
bining the predictability of deterministic models
with the flexibility of probabilistic approaches.
Deterministic models provide a solid foundation,
offering predictable outcomes under specific con-
ditions, which is vital for ensuring baseline perfor-
mance and reliability~\cite{boobalan2022fusion}.
On the other hand, probabilistic models excel in
handling uncertainties and variabilities inherent
in IIoT environments, like fluctuating network
conditions or variable data quality. By
Fig. integrating FIGURE
1. The Federated Learning 1. The Federated
Architecture Learning
in an Industrial Architecture
Environment in an Industrial
with IIoT-enabled devices andEnvironment with
XAI frameworks.
these approaches, FL systems can maintain stable IIoT-enabled devices and XAI frameworks.
operation while dynamically adapting to chang-
ing environmental factors. This fusionaccount for for
allows uncertainties
In anandIIoT
variabilities in the data
setting employing an and
adaptivemechanisms
hybrid guarantees that the FL models are not excessively
robust decision-making, balancing operational
the need for conditions [12]. Byto incorporating
approach FL integratedprobabilistic
with ExplainableimpactedArti-by abnormal data, which is frequently encountered
reasoning,
consistency with the ability to respond FL systems
to new ficialcan make more(XAI)
Intelligence informed decisions,theinarchitec-
frameworks, unstable environments.
and uncertain information. even in the face of incomplete
ture, shown or noisy
in Fig.data. Techniques like
1, dynamically combines Overall,
the probabilistic strategies in FL offer a promising
To further enhance the robustness Bayesian inference
of FL in can be usedofto both
benefits updatecentralized
models dynamically avenue for designing systems resilient to the uncertainties and
and decentralized
as new data becomesmachine
challenging IIoT environments, context-aware available,learning
providing a robustIIoT
models. framework
devices likevariabilities
sensors of challenging IIoT environments. By leveraging
methodologies are crucial. Theseforapproaches
learning under uncertainty.
and assembly machines process and analyze these techniques,
data FL systems can maintain performance and
involve designing FL systems that are not only locally, training AI models with data reliability,
specific to even under adverse conditions.
To further adapt FL to challenging environments, techniques
aware of their operational contextthat butfocuscan also their environments.
on environmental adaptability are This local processing
essential. This is key
adapt their learning and decision-making process- in preserving data privacy
includes developing adaptive algorithms that can adjust learn- and minimizingD. band- Hybrid Approaches
Adaptive
es based on this context [14]. This might include width usage. Here, the
ing parameters in response to changes in the environment.
virtualization layer makes
The dynamic design fusion in FL involves combining the
adjusting learning rates, modifying model param- resource management and allocation easier, guar-
For instance, reinforcement learning strategies can be em- predictability of deterministic models with the flexibility of
eters, or selectively weighting data based on its anteeing that IIoT devices are used effectively.
ployed, where the FL system learns optimal actions through probabilistic approaches. Deterministic models provide a solid
source and quality. Context-aware systems are Model aggregation is handled by FL aggrega-
trial and error, adapting to the environment’s dynamics [13]. foundation, offering predictable outcomes under specific con-
particularly adept at dealing with the heteroge- tion servers, and VMs in the cloud offer scalable
Furthermore, integrating reliable outlier detection and handling ditions, which is vital for ensuring baseline performance and
neous and dynamic nature of IIoT environments, computational resources. Integration takes place
where the operating conditions can vary wide- through APIs, which guarantee safe communica-
ly and unpredictably. By being responsive to the tion and cooperative model training, boosting FL’s
specific context in which they operate, FL systems credibility for IIoT.
can optimize their performance and maintain high In the adaptive hybrid model, these local
levels of accuracy and efficiency, even in the most devices periodically send model updates to a
challenging conditions. central server, which aggregates these updates to
These adaptive hybrid approaches in FL design refine a global model. However, unlike traditional
offer a pathway to creating reliable and adaptable FL, this approach can adaptively choose when to
systems, capable of thriving in the diverse and rely more on local models and incorporate more
often unpredictable landscapes of the IIoT. from the global model, depending on factors like

IEEE Internet of Things Magazine • September 2024 41

Authorized licensed use limited to: Northeastern University. Downloaded on March 05,2025 at 19:46:26 UTC from IEEE Xplore. Restrictions apply.
 IEEE INTERNET OF THINGS MAGAZINE 6

operational efficiency but also contributes to the

sustainability of energy systems.

supply chAIn mAnAgement

FL can significantly enhance supply chain man-
agement by providing insights into logistics with-
out compromising the data of individual suppliers.
For example, in a complex supply chain, FL can
be used to track and predict inventory levels,
ensuring timely restocking and reducing the risk
of supply shortages.

envIronmentAl monItorIng
In environmental monitoring, FL is used to pro-
cess data from various sensors deployed in dif-
ferent geographical locations. This data helps in
predicting environmental trends and potential
hazards, such as pollution levels or the likelihood
FIGURE 2. The
Fig. 2. The Dynamic Dynamic
Industrial Industrial
Management through Management
Personalized IIoT Framework driventhrough Personalized
by Federated Learning Models. IIoT of natural disasters. Such applications are crucial
Framework driven by Federated Learning Models. for proactive environmental protection and disas-
• XAI integration: More profound integration of XI tech- unique advantages for evaluating equipment health in low- ter management.
niques within FL models to enhance transparency and
trust.
temperature environments. This strategy preserves the confi-
dentiality of sensitive data and guarantees real-time insights,
These case studies demonstrate the versatility
• data diversity, model performance, and network
Advanced anomaly detection: Leveraging AI to detect which is essential for preserving equipment dependability and efficacy of FL in IIoT, highlighting its ability to
conditions. This adaptability ensures optimal learn-
and mitigate novel attacks and anomalies in FL networks.
Hybrid models: Combining the strengths of different
under challenging conditions.
offer robust solutions while adhering to responsible
Furthermore, FL is also used in mining operations to opti-
ing outcomes, balancing the need for customized data-handling practices. As the technology con-
•
learning paradigms (supervised, unsupervised, and rein- mize autonomous vehicles, track the condition of equipment,
local models with the benefits of a robust, gener-
forcement learning) within the FL framework for more
robust and interpretable models.
and improve extraction procedures in difficult terrain. Such tinues to evolve, it is expected to find even more
applications show the flexibility of FL in meeting certain
alized global model. industry requirements while maintaining data security and innovative applications across different sectors.
E. Practical Insights
The integration of XAI in this adaptive hybrid
privacy. The future of trustworthy FL in IIoT lies in
framework is crucial for maintaining transpar-
FL provides customized solutions for improving industrial
Enabling trustworthy FL in the context of IIoT presents a
unique set of challenges and opportunities, primarily centered developing holistic frameworks that integrate
ency
processes in challenging industrial ecosystems. and trust in the AI models. XAI tools pro-
By coopera- around balancing interpretability and robustness. FL, a decen- robustness and interpretability seamlessly. Poten-
vide insights into the decision-making processes
tively enhancing pressure control among dispersed devices,
FL can improve low-pressure processes in the chemical and
tralized approach to machine learning, allows for the training
of algorithms across multiple devices or servers while keeping tial directions include:
pharmaceutical sectors. FL is also useful for of both
evaluating local and global models, ensuring that
equip- the data localized. This method is particularly beneficial for • XAI integration: More profound integration of
AI decisions are interpretable and justifiable, a
ment health in low-temperature conditions and optimizing
temperature control in cryogenic industries, such as the man-
IIoT, where data security and privacy are paramount, and the
data is often generated in distributed environments [15].
XI techniques within FL models to enhance
ufacturing of liquefied natural gas. FL helps necessity
offshore oil in critical industrial applications. This
One key lesson learned is the importance of interpretability transparency and trust.
and gas operations improve overall safety inapproach
challenging and not only enhances operational efficien-
in FL models. In the industrial setting, stakeholders need • Advanced anomaly detection: Leveraging
remote environments and guarantee the quality of subsea to understand and trust the decisions made by AI models,
equipment. Extreme temperature and pressure cy changes
and decision-making
are in IIoT environments but
especially when these decisions can have significant safety and AI to detect and mitigate novel attacks and
addressed by FL in the aerospace manufacturingalso ensures
industry,
which also optimizes furnace conditions and forecasts equip-
that AI models remain understand-
financial implications. Interpretability in FL can be challenging
due to the distributed nature of the model training, but it’s
anomalies in FL networks.
able
ment failures to boost productivity. FL preserves dataand trustworthy to human operators, aligning
privacy, crucial for ensuring that the models are reliable and their • Hybrid models: Combining the strengths of
minimises communication overhead, and allows with the evolving needs and constraints of the
collaborative
model training directly on distributed IIoT devices, providing
decisions are justifiable. Techniques such as model-agnostic
methods, visualization tools, and simplified model architec-
different learning paradigms (supervised,
industrial sector. unsupervised, and reinforcement learning)
within the FL framework for more robust and
cAse studIes And prActIcAl InsIghts interpretable models.
The field of FL in IIoT has seen significant advance-
ments, with real-world applications demonstrating prActIcAl InsIghts
its potential for responsible and robust operations. FL provides customized solutions for improv-
Below are illustrative case studies and practical ing industrial processes in challenging industrial
insights from various industries. Figure 2 shows ecosystems. By cooperatively enhancing pres-
the dynamic personalized IIoT solutions through sure control among dispersed devices, FL can
secured and trustworthy FL models. improve low-pressure processes in the chemical
and pharmaceutical sectors. FL is also useful for
smArt mAnuFActurIng evaluating equipment health in low-temperature
In a smart manufacturing scenario, FL is utilized conditions and optimizing temperature control in
to enhance production processes by optimizing cryogenic industries, such as the manufacturing
them, all while prioritizing data privacy. FL can of liquefied natural gas. FL helps offshore oil and
identify patterns and inefficiencies by analyzing gas operations improve overall safety in challeng-
data across multiple manufacturing sites, all while ing and remote environments and guarantee the
ensuring that sensitive information is not shared. quality of subsea equipment. Extreme tempera-
For example, a multinational company can uti- ture and pressure changes are addressed by FL
lize FL to streamline its quality control procedures in the aerospace manufacturing industry, which
across various locations, resulting in improved effi- also optimizes furnace conditions and forecasts
ciency and decreased waste. equipment failures to boost productivity. FL pre-
serves data privacy, minimises communication
energy sector overhead, and allows collaborative model train-
In the energy sector, particularly in renewable ing directly on distributed IIoT devices, provid-
energy, FL is applied to predict maintenance ing unique advantages for evaluating equipment
needs and optimize energy distribution. By pro- health in low-temperature environments. This
cessing data from various wind farms or solar pan- strategy preserves the confidentiality of sensitive
els, FL can predict when maintenance is required, data and guarantees real-time insights, which is
thereby reducing downtime and maximizing ener- essential for preserving equipment dependability
gy output. This application not only improves under challenging conditions.

42 IEEE Internet of Things Magazine • September 2024

Authorized licensed use limited to: Northeastern University. Downloaded on March 05,2025 at 19:46:26 UTC from IEEE Xplore. Restrictions apply.
 Furthermore, FL is also used in mining operations and practical applications are geared towards devel-
to optimize autonomous vehicles, track the condi- oping methodologies and tools that enhance both
tion of equipment, and improve extraction proce- aspects, thereby making FL a viable and reliable
dures in difficult terrain. Such applications show the approach in the context of IIoT.
flexibility of FL in meeting certain industry require-
ments while maintaining data security and privacy. Conclusion
Enabling trustworthy FL in the context of IIoT While FL allows learning from diverse, distributed
presents a unique set of challenges and opportu- data sources without central data storage, enhanc-
nities, primarily centered around balancing inter- ing privacy and reducing communication needs,
pretability and robustness. FL, a decentralized its widespread adoption in IIoT is challenged by
approach to machine learning, allows for the train- issues ensuring model interpretability and robust-
ing of algorithms across multiple devices or servers ness. Trustworthy FL in the IIoT is a complex yet
while keeping the data localized. This method is promising domain. Bridging the gap between
particularly beneficial for IIoT, where data security interpretability and robustness requires a multifac-
and privacy are paramount, and the data is often eted approach involving advanced machine-learn-
generated in distributed environments [15]. ing techniques, robust security protocols, and an
One key lesson learned is the importance of overarching framework prioritizing transparency
interpretability in FL models. In the industrial set- and resilience. As technology evolves, the synergy
ting, stakeholders need to understand and trust of FL with IIoT will undoubtedly play a pivotal role
the decisions made by AI models, especially when in realizing the full potential of smart industries.
these decisions can have significant safety and
financial implications. Interpretability in FL can be References
[1] A. Vehabovic et al., “Ransomware Detection Using Feder-
challenging due to the distributed nature of the ated Learning with Imbalanced Datasets,” arXiv preprint
model training, but it’s crucial for ensuring that the arXiv:2311.07760, 2023.
models are reliable and their decisions are justifi- [2] D. C. Nguyen et al., “Federated Learning for Industrial Inter-
net of Things in Future Industries,” IEEE Wireless Commun.,
able. Techniques such as model-agnostic methods, vol. 28, no. 6, 2021, pp. 192–99.
visualization tools, and simplified model architec- [3] D. C. Nguyen et al., “Federated Learning for Internet of
tures have been explored to enhance interpretabili- Things: A Comprehensive Survey,” IEEE Communications
ty without compromising the model’s performance. Surveys & Tutorials, vol. 23, no. 3, 2021, pp. 1622–58.
[4] X. Wang et al., “Toward Accurate Anomaly Detection in
On the other hand, robustness in FL is essential Industrial Internet of Things Using Hierarchical Federated
for maintaining the integrity and performance of Learning,” IEEE Internet of Things J., vol. 9, no. 10, 2021, pp.
the models in the face of diverse and often noisy 7110–19.
industrial data. Robustness involves ensuring that [5] J. Wei et al., “A Redactable Blockchain Framework for Secure
Federated Learning in Industrial Internet of Things,” IEEE Inter-
the models are resistant to data anomalies, adver- net of Things J., vol. 9, no. 18, 2022, pp. 17,901–11.
sarial attacks, and other forms of disruptions that [6] G. Li et al., “Multitentacle Federated Learning Over Soft-
are common in IIoT environments. This requires ware-Defined Industrial Internet of Things Against Adaptive
advanced techniques like anomaly detection, data Poisoning Attacks,” IEEE Trans. Industrial Informatics, vol. 19,
no. 2, 2022, pp. 1260–69.
encryption, and secure aggregation protocols in the [7] B. Ghimire and D. B. Rawat, “Recent Advances on Federated
FL framework. The research in this area focuses on Learning for Cybersecurity and Cybersecurity for Federated
developing robust FL models that can withstand the Learning for Internet of Things,” IEEE Internet of Things J.,
challenges specific to industrial settings, such as vari- vol. 9, no. 11, 2022, pp. 8229–49.
[8] P. M. S. Sánchez et al., “Federatedtrust: A Solution for Trust-
able data quality and potential cybersecurity threats. worthy Federated Learning,” Future Generation Computer
An XAI model predicting machine failures Systems, vol. 152, 2024 , pp. 83–98.
through FL and IIoT frameworks in a manufactur- [9] P. Boobalan et al., “Fusion of Federated Learning and Indus-
ing plant could provide interpretable features such trial Internet of Things: A Survey,” Computer Networks, vol.
212, 2022, p. 109048.
as vibration levels, temperature fluctuations, and [10] A. Gaddam and V. Govender, “Detecting Sensor Faults and
health status to diagnose unexpected issues, aiding Outliers in Industrial Internet of Things,” Int’l. Conf. Sensing
in understanding the root causes of equipment fail- Technology, Springer, 2022, pp. 183–200.
ures and enabling proactive maintenance actions. [11] M. P. Cenci et al., “Eco-Friendly Electronics — A Compre-
hensive Review,” Advanced Materials Technologies, vol. 7,
Furthermore, robust XAI models ensure the consis- no. 2, 2022, p. 2001263.
tency of product quality by addressing quality vari- [12] V. Balasubramanian, M. Aloqaily, and M. Reisslein, “Fed-
ations in materials and environmental conditions. TSN: Joint Failure Probability Based Federated Learning for
The interpretability feature allows supply chain Fault-Tolerant Time-Sensitive Networks,” IEEE Trans. Network
and Service Management, 2023.
managers to understand better the factors affecting [13] L. U. Khan et al., “Federated Learning for Internet of Things:
performance and identify areas for improvement. Recent Advances, Taxonomy, and Open Challenges,” IEEE
Additionally, the robustness of the models helps Commun. Surveys & Tutorials, vol. 23, no. 3, 2021, pp.
manage supply chain disruptions, demand vola- 1759–99.
[14] R. Rentero-Trejo et al., “Using Federated Learning to
tility, and unforeseen market changes. In packag- Achieve Proactive Context-Aware IoT Environments,” J. Web
ing materials, interpretability allows inspectors to Engineering, 2022, pp. 53–74.
grasp the criteria used by XAI algorithms to detect [15] M. H. ur Rehman et al., “Trustfed: A Framework for Fair
defects and anomalies. To uphold product quality and Trustworthy Cross-Device Federated Learning in IIoT,”
IEEE Trans. Industrial Informatics, vol. 17, no. 12, 2021, pp.
and brand reputation, the robustness of the model 8485–94.
assists in estimating variations in surface textures,
defects, and packaging configurations. Biographies
Last, the integration of trustworthy FL in IIoT Senthil Kumar Jagatheesaperumal (senthilkumarj@mepcoeng.
ac.in) received his B.E. degree in Electronics and Communica-
hinges on achieving a balance between interpret- tion Engineering from Madurai Kamaraj University, Tamilnadu,
ability and robustness. While interpretability ensures India in 2003. He received his post-graduation degree in Com-
the transparency and trustworthiness of the models, munication Systems from Anna University, Chennai, India in
robustness safeguards them against the diverse chal- 2005. He completed his Ph.D. in Information and Communica-
tion Engineering from Anna University, Chennai, India in 2017.
lenges in industrial environments. Ongoing research He is currently working as an Associate Professor (Senior Grade)

IEEE Internet of Things Magazine • September 2024 43

Authorized licensed use limited to: Northeastern University. Downloaded on March 05,2025 at 19:46:26 UTC from IEEE Xplore. Restrictions apply.
 in the Department of Electronics and Communication Engineer- research, education, and outreach. Earlier, he was Associate Chair
ing, Mepco Schlenk Engineering College, Sivakasi, Tamilnadu, of the ECE Department at the University of New Mexico (UNM).
India. He received two funded research projects from National He has also held technical research and development positions at
Instruments, USA each worth USD 50,000 during the years several large corporations (including Nokia, IBM, and Motorola)
2015 and 2016. He also received another funded research proj- and some startups. His research focus areas include cybersecurity,
ect from IITM-RUTAG in 2017 worth Rs.3.97 Lakhs. His area of cyberinfrastructure design, disaster recovery, and online educa-
research includes Robotics, the Internet of Things, Embedded tion. His research has been supported by the NSF, DoD, DoE,
Systems, and Wireless Communication. He is a Life Member of Qatar Foundation, and several state and industry partners. He also
IETE and ISTE. received the NSF CAREER Award in 2005.

Mohamed Rahouti (mrahouti@fordham.edu) received the M.S. Vu Khanh Quy (quyvk@utehy.edu.vn) was born in Hai Duong,
and Ph.D. degrees from the University of South Florida in Math- Vietnam, in 1982. He received the M.Sc. and Ph.D. degrees
ematics Dept. and Electrical Engineering Dept., Tampa, FL, USA, from the Posts and Telecommunications Institute of Technology,
in 2016 and 2020, respectively. He is currently an Assistant Hanoi, Vietnam, in 2012 and 2021, respectively. He is currently
Professor, Department of Computer and Information Sciences, a Lecturer with the Hung Yen University of Technology and
Fordham University, Bronx, NY, USA. His current research focus- Education (UTEHY), Hung Yen, Vietnam. His research interests
es on computer networking, blockchain technology, Internet of include wireless communications, mobile computing, smart
Things (IoT), machine learning, and network security with appli- IoT systems, next-generation networks, Internet of Things, and
cations to smart cities. next-communication networks

Ali Alfatemi (aalfatemi@fordham.edu) is pursuing his Ph.D. in Abdellah Chehri [SM] (chehri@rmc.ca) is an Associate Profes-
Computer science at Fordham University, USA. He has exten- sor at the Royal Military College of Canada (RMC), Kingston,
sive industry experience as a Software Engineer. His research Canada. Before joining the RMC, he was an associate professor
interests encompass Artificial Intelligence and Machine Learning, at the University of Quebec (UQAC). He has an affiliate profes-
Network Intrusion Detection, and Bio-informatics. sor at the University of Quebec UQO, UQAC and an adjunct
professor at the University of Ottawa. He has served as guest/
Nasir Ghani (nghani@usf.edu) is a Professor of Electrical Engi- associate editor for several well-reputed journals. Dr. Chehri is a
neering and Program Director of the College of Engineering Senior Member of IEEE, a member of the IEEE Communication
MS in Cybersecurity. He is also Academic Research Director for Society (ComSoc), IEEE Vehicular Technology Society (VTS),
Cyber Florida, a state-funded cybersecurity center focusing on and IEEE Photonics Society.

44 IEEE Internet of Things Magazine • September 2024

Authorized licensed use limited to: Northeastern University. Downloaded on March 05,2025 at 19:46:26 UTC from IEEE Xplore. Restrictions apply.