Integration of federated learning with IoT for smart cities applications, challenges, and solutions

Introduction

Nobody wants unauthorized access to their data, whether a business or an individual. For the training of machine learning (ML) algorithms, a large quantity of high-quality data is required (Eyke, Green & Jensen, 2020). Federated learning (FL) is a method of training AI without needing a centralized server. Instead, it collaborates between a server and many devices. To make this function, no data must be transmitted back and forth (Ahmadi et al., 2022). It allows the storage of personal data locally, which reduces the risk of security breaches. When training the ML algorithms, traditional ML uses all the training data stored in the central server (Shen et al., 2022). It has a few issues; sometimes, connectivity between the device and the central server can be slow. Unauthorized individuals can access personal information. FL is a way to train models across multiple decentralized edge devices (Imteaj et al., 2021).

A global share model makes it possible for several devices to learn together. It will only send the collected information of the model, such as parameters and results, to the cloud, but it will use the data on devices to update the model (Jintasuttisak, Edirisinghe & Elbattay, 2022). For example, the keyboard model predicts the next word but also protects the confidentiality of text messages. It is a decentralized ML strategy that keeps data in its original location and decreases the amount of hardware infrastructure requirements (Bonawitz et al., 2021).

The IoT is a collection of interconnected physical objects that can exchange and integrate data with other devices and systems through sensor devices, applications, and other technologies (Noura, Atiquzzaman & Gaedke, 2019). The Internet Protocol (IP) is used to identify computers on the Internet and enable users to communicate with one another.

The IoT aims to create devices that can self-report information and data regularly and improve efficiency (Beg et al., 2022). Data security is an important point, and FL is a powerful strategy for protecting user privacy. In the domain of IoT, the sensors are used to gather data that may include privacy-sensitive personal details. Many IoT smart devices use FL, a distributed ML technique that protects IoT security and privacy (Pang et al., 2020). In Nguyen et al. (2021a) and Nguyen et al. (2021b), it has been proposed that FL can be used in various IoT applications, including intelligent healthcare, transportation, and Unmanned Aircraft Vehicles (UAVs).  FL has made it easier to provide intelligent healthcare services by allowing ML models without requiring medical institutions to share patient data (Aldahiri, Alrashed & Hussain, 2021). FL with IoT still has a lot of challenges.

To provide better patient care, experts in the healthcare industry require reliable technology. Extensive and diverse data sets are needed to train an algorithm for clinical purposes. FL plays an essential role in this kind of scenario. Companies use internal information sources to train the same algorithm (Lee & Shin, 2020). The FL can potentially bring significant positive changes in the healthcare profession. Thus, if FL is successfully applied, it could lead the way for universal precision medicine by generating models that are impartial in their predictions (Antunes et al., 2022). FL can also verify that the algorithm is running correctly without compromising the patient’s safety. The fourth industrial revolution, 4.0, completely alters how companies produce, improve, and distribute their products. The IoT, cloud computing, AI, and ML are new technologies that integrate business operations (Helo & Hao, 2021). FL allows computer-based algorithms to monitor and operate machinery, robots, and vehicles (Vijayakumar et al., 2019).

FL enables multiple participants to develop a global predictive model cooperatively. This strategy has received much attention in large-scale architectures and generic loT applications (Safri et al., 2022). Using digital technologies results in a higher level of automation, predictive maintenance, and self-optimization of improved efficiency (Naveed, Anwar & Haq, 2021). FinTech companies use a wide range of technology to manage their financial operations (Mehrban et al., 2020). Businesses that rely on FinTech face several challenges. These challenges are about obtaining approval and legal agreements, questions raised regarding the safety of the data, and the amount of time and expenses involved in collecting and transmitting data between networks. FL provides an easy solution for ML that is both encrypted and distributed. It enables users to train ML together on distributed data without the requirement for data transmission at any point during the process (Liu et al., 2022). It can solve problems and provide answers to FinTech. FL clears the way for FinTech to minimize possible risks (Kawase et al., 2019). It develops innovative and forward-thinking approaches for the benefit of both its customers and the businesses. It is reasonable for both parties to have trust in one another.

This study examines the obstacles and potential solutions to integrating FL and the Internet of Things (IoT). This statement underscores the significance of FL in protecting user privacy, enhancing the performance of models, facilitating adaptable scalability, and augmenting the quality of learning within the Internet of Things (IoT) networks. The study also discusses many obstacles faced by Federated Learning for the Internet of Things (FL-IoT). These challenges include resource management, Aggregation of updates, protection of privacy, security concerns, and issues related to learning and communication, standardized standards, and deploying machine learning capabilities on IoT sensors. This article identifies significant challenges and outlines potential future research lines. In ‘Literature Review’, we go through the most commonly used FL-IoT applications. ‘Methods and Techniques’ discusses the importance of FL and IoT and its challenges in detail, which is very important. ‘Results and Discussion’ represents the solutions and opportunities to FL-IoT Challenges. The conclusion and future work are described in ‘Conclusions and Future Work’. Figure 1 shows the article’s organization and Table 1 shows the list of abbreviations.

Literature Review

FL allows insurance companies to train the model with its data and then upload the trained model’s output to the federation server because this method is secure and encrypted (Wang, Dang & Zhou, 2019). The federation server will integrate the model results to produce a superior model and send feedback to each contributor. It provides efficient distribution of model training and reduces the amount of money spent on data transfer and storage. FL aims to train ML algorithms with different data sets (Wahab et al., 2021). A company could identify its customers’ behaviors without violating the data. The deployment of FL helps to prevent illegal activities. The algorithms could use the data to guide training and decision-making (Zheng et al., 2022). Natural language processing is a part of AI that helps computers understand, analyze, and change human language. It contributes to improved comprehension of the semantics of human languages. A massive amount of data is required to train extremely accurate language models. This information is readily available from portable electronic devices such as mobile phones and tablets. Because the textual data gathered from each edge device contains user information, privacy concerns challenge central language learning models in this context. In Liu et al. (2021a) and Liu et al. (2021b), the authors presented that it is possible to construct natural language processing models using an FL framework. The amount of data also significantly increases as technology improves. In modern IoT networks, including wearable technology, driverless cars, and smart homes, sensors collect and process data in real-time (Bianchi et al., 2019). FL techniques enable the development of models that quickly adapt to these systems. Various organizations are now using FL. Without exchanging data, organizations train their algorithms on a variety of datasets. FL seeks to protect the information gathered through several routes and keeps valuable data near to reach (Liu et al., 2020a; Liu et al., 2020b; Liu et al., 2020c). Personalization can be accomplished with FL, and the functionality of IoT applications can be improved. Figure 2 shows the application of FL for IoT.

The future generation of the Internet is the metaverse. It will provide a fully connected, comprehensive, and engaging online 3D virtual reality experience that can be accessed on personal computers and other devices using virtual and augmented reality (Lee et al., 2021). Users will have control of their avatars in the metaverse, which allows them to communicate with virtual objects. The information gathered through IoT devices and servers connects the virtual and real worlds. FL is an attractive potential solution that could facilitate collaboration between the edge and the server, improving global performance and enhancing the metaverse’s level of security and privacy. Wearable technologies from millions of people can gather motion-tracking data that can be trained on local machines and put together on an FL server (Wang et al., 2022). Due to its real-time predictive abilities, FL is an essential part of the technology behind autonomous vehicles. Road and traffic conditions can be updated in real-time, which can be reflected in the data for continuous improvement and quick decision-making (Zhang et al., 2021). One possible benefit is that it will make driving more relaxing and secure. There is a lot of potential for using federated ML in the automotive industry.

FL may reduce the time spent training autonomous vehicles and improve the accuracy with which wheel and steering angle predictions are made (Zhang et al., 2021). The exponential increase in IoT devices has had a major and very far impact on people’s regular routines. AI dramatically affects how IoT devices are trained to use ML (Abdellah, Mahmood & Koucheryavy, 2020). The term “federated learning” refers to a type of “distributed ML” that is implemented in a wide variety of “smart devices” that are connected to the IoT. FL is ideal for edge networks, where each IoT device at the edge acts as an autonomous client (Lim et al., 2020). Sensors connected to the IoT can immediately collect information about users, including private information. FL provides an excellent way to protect user privacy by ensuring that raw data collected for each IoT device is not shared with other entities (Dash, Sharma & Ali, 2022). Each IoT device learns models in an FL environment to perform its task. When training in a federated way, the raw data never leaves the machines, and only the model updates are sent to the centralized server. This significantly reduces the potential for leaking sensitive information (Rahman et al., 2020).

FL reduces the requirement to transfer IoT data to the server and helps minimize interaction delays caused by information offloading. It helps preserve system resources during the information phase, like frequency and transmission power (Tran et al., 2019). There is not enough information to build a high-quality model from a single IoT device. As a result of FL’s model-building infrastructure, all devices can work together to produce superior models (Liu et al., 2020a; Liu et al., 2020b; Liu et al., 2020c). FL is an effective method for enhancing the model’s overall performance in FL that allows flexible scalability and limited computation power available at several IoT devices across various geographic locations (Zhang et al., 2021). Increasing the network’s capacity is challenging because combining all data on a single server depletes computer resources at the edges (Ferrer, Marquès & Jorba, 2019). FL can increase IoT networks’ flexibility without adding more load to central servers. Instead, it enables more devices to join the framework (Bhati, Chugh & Bhati, 2022). FL decreases the cost of communications, which is helpful for IoT networks with low bandwidth (Victor et al., 2022). FL can potentially increase the whole training phase’s convergence speed and attain higher learning accuracy rates. FL is employed in various IoT applications, such as UAVs, smart healthcare, and smart transportation (Nguyen et al., 2021a; Nguyen et al., 2021b).

FL has facilitated ML modeling to simplify delivering smart healthcare services. Health data providers like hospitals are not required to share health information when using FL. They localize the training of the AI model and transfer the trained parameters to the aggregator, enabling it to be used in global computation (Fadlullah & Kato, 2021). FL encourages collaborative healthcare environments, involves rapid patient assessment and treatment, and protects individual privacy (Banabilah et al., 2022). The advanced vehicular services offered by FL, including autonomous driving, road safety prediction, and vehicle identification, enhance driver training efficiency (Javed et al., 2022).

Methods and Techniques

Results and Discussion

What FL and IoT challenges are faced during integration with organizations like smart businesses, cities, transportation, and healthcare?

FL is a relatively new field that has already made important contributions but also faces significant challenges. The following are key issues associated with FL (Li et al., 2020). These challenges prevent FL from being deployed on billions of IoT devices (Bouacida & Mohapatra, 2021). Figure 3 and Table 2 show the challenges in FL-IoT.

Table 2:

The challenges in FL-IoT.

SR. #	Challenges	The function of FL with IoT	Description	References
1	Resource Management in FL	Resource management is used in neural networks to use computational resources effectively.	Specific IoT devices have limited computational capabilities, and the required resources for FL training are not always achieved.	Tsukada, Kondo & Matsutani (2020)
2	Updates Aggregations	The server must quickly collect and update information for vehicles and IoT devices.	Blockchain technology solves all of these challenges because it has decentralized storage and continued maintenance of the FL model.	Ali, Karimipour & Tariq (2021)
3	Privacy protection and security issues in FL	IoT needs a secure FL system with the highest possible protection for gathering data.	The central server can corrupt after a few iterations; hackers can steal the data to generate the training gradients.	Ma et al. (2020)
4	Learning and communication issues in FL-IoT.	High-quality model training in IoT networks requires an FL-based algorithm.	Due to the challenges posed by diverse data and the resources provided by distributed IoT devices, traditional FL algorithms have a slow convergence speed.	Nilssom et al. (2018)
5	Standardized specifications for FL-IoT.	Need a new set of Wi-Fi protocols to handle the heavy traffic loads of future IoT applications.	The deployment of FL-IoT ecosystems depends on other significant computing services, such as cloud computing. Network standards and aspects play an essential role in this deployment with fast network speed.	López-Pérez et al. (2019)
6	AI Functions Deployed on IoT Sensors.	To increase the accuracy of hardware learning, it is necessary to use a memory-effective on-device sensor learning system.	Some IoT sensors cannot take part in training a full-scale AI model due to hardware, memory, and power limitations.	Dhar et al. (2019)
7	Time variation after the deployment of the model.	The requirement for on-device model changes creates a new issue for IoT devices, which often have limited functional capacities.	The necessity for on-device model updates introduces a new challenge for the IoT devices, which frequently have limited capabilities in terms of their functionality.	Liang et al. (2020)
8	Protect against the attacks of the adversary.	IoT devices are easy targets of phishing, identity fraud, and distributed denial of services (DDoS).	The IoT system must implement a simple security protocol to identify malicious devices.	Salim, Rathore & Park (2020)
9	Relation with accessible resources.	Depending on the characteristics of the feature, IoT configurations and data collected by various devices could differ significantly.	An IoT FL framework should be able to adapt data and computation load among multiple devices based on the resources available to those devices.	Saeed, Ozcelebi & Lukkien (2019)

DOI: 10.7717/peerjcs.1657/table-2

Resource management in FL

Federated learning has been investigated as a potentially useful option to train machine learning models at the network edge in a way that does not involve exchanging private user data. Because of the limited resources at the edge, new solutions need to be developed to maximize the use of the available software and hardware resources. It is necessary because the previously known solutions did not concentrate on resource management for the network edge, which is especially important for federated learning (Trindade, Bittencourt & Da Fonseca, 2022). To achieve updates on the server, all IoT devices need computing and storage resources. However, it is not always satisfied due to the limited computing power of IoT devices. A significant delay can occur at the server (Tsukada, Kondo & Matsutani, 2020).

Aggregation updates in FL-IoT

When moving data between networks, one of the most critical aspects of network security is ensuring that each device in the network has been properly authorized. Previously, the identification was based on public key infrastructure (PKI), which requires that each device in a system exchange its own private encrypted identity message to the local authentication center (LAC) (Akhter et al., 2021). The currently available privacy frameworks, such as the differential privacy framework (Caldas et al., 2018), are insufficient to resolve the users’ privacy.

Privacy protection and security issues in FL-IoT

FL can protect users’ privacy for the IoT systems that are distributed. FL still has several security and privacy flaws, and these challenges exist on both the client and server sides of the learning process (Lyu, Yu & Yang, 2020). Communication constraints, poisoning, and backdoor operations are currently the most specific security issues; inferential attacks are the most important to the privacy of FL. Future research must make FL adaptable in real-world situations (Mothukuri et al., 2021).

Learning and communication issues in FL-IoT

The major challenge in FL-IoT is the limited integration of learning and communication. Autonomous driving technology is making its way into regular vehicles with the advancement of vehicular IoT. A reliable self-driving system requires a real-time connection with a multi-access communication environment (Yarradoddi & Gadekallu, 2022). Spatial and temporal variations of the vehicular environment necessitate an intelligent solution that can adapt to the changing climate. The typical centralized-over-cloud solution requires the driving system to send a considerable quantity of raw data to the server, which could result in privacy leakage (Zhang et al., 2022).

Standardized specifications for FL-IoT

Future intelligent networks will include vertical FL-IoT use cases. Federated learning (FL) is a cutting-edge artificial intelligence approach. For this purpose, existing mobile network architecture must be amended (Shaheen et al., 2022). Network standards and elements like cloud analysis servers and edge-IoT protocols are essential, and these components can contribute to the completion of FL-IoT ecosystems (Ray, Dash & De, 2019). Unprocessed information is stored in advanced technology by a secret confidentiality service, which incorporates machine learning (ML) training while removing data connections (Alam & Gupta, 2022). There is a need for a developed system to improve the effectiveness of advanced learning systems.

AI functions deployed on IoT sensors

Internet of Things sensors are limited in training a comprehensive AI model due to the hardware, storage, and energy resources. Powerful machine learning algorithms typically require significant storage and power to train models. The main challenge is finding out the energy usage problem in FL-IoT systems. Improving the use of AI hardware on IoT sensors is an important task (Dhar et al., 2021).

Temporal changes after deployment

IoT sensing devices continuously collect updated information to upgrade the models for continuous learning. This data also improves the user experience (Tang et al., 2019). To continue offering services, it is also necessary to maintain training data on updated models, which presents a new issue for resource-constrained IoT devices. Most IoT devices have a limited amount of memory and memory resources. To overcome this problem, the requirement for a lightweight machine learning model can significantly reduce the memory needed for on-device training (Peltonen et al., 2020).

Protect against the attacks of the adversary

IoT devices are the targets for attackers, such as distributed denial of service (DDoS) (Parra et al., 2020). The traffic volume of IoT-based Attacks is increasing. Security updates can stop these attacks, but most IoT devices do not have sufficient processing power (Rizvi et al., 2018). The IoT system must be able to identify malicious devices that can destroy the training model. Implementing a lightweight security procedure and detecting malfunctioning devices in the IoT system is a solution (Jeon, Park & Jeong, 2020).

Relation to limited on-device resources

The existing machine learning models, particularly deep neural networks, demand significant computational power (Ren, Deng & Xie, 2022). Customized and specialized hardware is one way to speed up the training process for edge machine-learning applications. Edge devices have limited resources for memory storage, computational power, and information access (Qian et al., 2019). Recent neural network architectures are becoming increasingly complex. One of the most significant issues associated with neural network processing models is minimizing the number of memory accesses (Gobieski, Lucia & Beckmann, 2019). Table 2 shows an overview and summary of FL-IoT challenges in FL with IoT in Organizations.

What are the solutions and opportunities for these problems?

FL is a very emerging field with significant contributions but faces substantial challenges. New concepts and strategies are required to solve these challenges in the following subsections, the most popular solutions models that, when integrated with FL-IoT, may solve the challenges listed above. Integrating different solution models to FL-IoT can speed up the work and reduce the risks. The integration is presented in Fig. 4. Below, we have listed and explained all the important solution models in resolving issues in federated learning and IoT devices. Figure 4 and Table 3 show the integration of FL-IoT with solution models.

Table 3:

The integration of FL-IoT with solution models.

Challenges	Solution models	Models description	Role of models in IoT	References
Resource management in FL	Train neural networks using an FL architecture that considers the resources available on mobile platforms.	This architecture takes into consideration information regarding the available computational resources.	A soft training method is included to speed up IoT devices.	Xu et al. (2019)
	DNN architecture for mobile devices training.	This model can speed up the learning process.	It can speed up the running time and save memory resources.	Li, Wang & Kong (2018)
	Federated wireless edge network learning resource management.	Algorithms are provided in this model to maximize efficiency based on the relevance of the data and the effectiveness of the correlating processing and transmission resources.	Improve the training’s accuracy, consistency, and speed of convergence.	Balakrishnan et al. (2020)
Vehicles aggregation updates in FL-IoT	A solution to privacy problems in vehicles using encryption.	Information that has been most recently updated should be shared between global models and localized models.	Identifies the privacy concerns generated by the IoT.	(DOH and CHAE)
Privacy protection and security issues in FL	We are establishing composition theorems with complex mathematical formulations and perturbations methods.	Differential privacy techniques can be employed to protect training datasets from being hacked.	Prevents unauthorized access to the data in IoT devices.	Ma et al. (2020)
	A safe and secure aggregation method used in FL-based systems.	The goal is to provide the highest level of protection to unauthenticated network conditions.	Use encryptions in IoT devices for local changes.	Bonawitz et al. (2016)
Learning and communication issues in FL-IoT.	A fresh optimization approach for FL-based IoT networks.	This method can successfully train high-quality models, increasing communication effectiveness.	Errors accumulate in IoT devices.	Rothchild et al. (2020)
	Momentum-based FL.	Increase the rate at which FL algorithms arrive at solutions.	Minimize the loss of functions in IoT devices.	Liu et al. (2020a), Liu et al. (2020b) and Liu et al. (2020c)
Standardized specifications for FL-IoT.	Open Platform Communication Global Infrastructure.	Edge-IoT services are intended to be supported.	Implementation in edge-IoT settings.	Ray, Dash & De (2019)
	The latest iteration of Wi-Fi technology.	In the 5G and 6G eras, this technology will be capable of meeting the peak throughput needs that emerging applications for the IoT will establish.	Intelligent IoT services at the edge networks.	López-Pérez et al. (2019)
AI Functions Deployed on IoT Sensors.	A deep learning accelerator.	It is based on software that facilitates AI training on mobile devices.	Maximize using hardware resources for training data without impacting overall performance reliability.	Lane et al. (2016)
	A lightweight model training method.	This strategy makes use of the idea of model output exchange rather than the concept of parametric exchange.	It can solve communication latency issues in IoT devices.	Itahara et al. (2020)
Temporal changes after deployment	FedGKT method.	It is possible that using this strategy will reduce the number of trained memories that is necessary for efficient on-device learning.	Several small CNN models can transfer knowledge from IoT devices to a more prominent CNN stored on the cloud.	He, Annavaram & Avestimehr (2020)
Protect against the attacks of the adversary	DIoT solution.	Apply an FL approach to anomaly detection-based vulnerability scanning in the gateway.	Scan the gateways to IoT systems instead of storing data locally.	Nguyen et al. (2019)
Limited on-device Resources	Fed Mask solution.	A framework that is efficient in terms of processing and communication.	IoT devices can build a sparse model with decreased computing cost and memory footprint.	Li et al. (2021)

DOI: 10.7717/peerjcs.1657/table-3

DDN and federated averaging-based resource management solution for FL-IoT

Deep learning models like DNN require a considerable amount of CPU frequency to complete training tasks; it will be a good strategy to train the model directly on IoT devices (Jagmohan et al., 2018). In FL, the DNN models are distributed and taught on local machines. After that, the local parameters of the model are periodically aggregated in a centralized node to update the global model using an algorithm called federated averaging (FedAvg) (McMahan et al., 2017). Several strategies have been developed to simplify resource management in on-device FL training. To resource management problems, importance sampling and rank ordering-based algorithms were developed. These algorithms allow for the prioritizing of clients based on the resources they have accessible. The paper also presents algorithms for data-importance and computation communication-aware resources to optimize the efficiency of training convergence rate (Balakrishnan et al., 2020).

Blockchain-based aggregation updates solution for FL-IoT

FL solves privacy concerns to the maximum level. Blockchain technology is used in these scenarios to resolve the issues (Sharma, Park & Cho, 2020). In Unal et al. (2021), a secure architecture for FL in big data analytics services for IoT that uses blockchain technology was presented. The author provided an architecture that uses FL to train the model locally and then transmits the encrypted model to an edge-based artificial intelligence service. To provide accurate insights and predictions, the AI service compiles all of the regional models into one comprehensive model.

Perturbations methods for privacy and protection in FL-IoT

If a protection mechanism is not working, FL will act as a privacy and security firewall in intelligent IoT systems (Arisdakessian et al., 2022). By establishing composition theorems with complex mathematical formulations, perturbation methods such as differential privacy and dummies can be used to maintain the integrity of training datasets from being hacked. It is done to protect training data and hide personal information from external threats while guaranteeing convergence (Yin, Zhu & Hu, 2021). The findings of the study (Bonawitz et al., 2016) imply that there is a secure aggregation method for safe FL systems. It provides the best protection possible against server-mediated, unauthenticated network conditions. In Zhao et al. (2021), the FL scheme that protects users’ privacy for the extensive industrial data is represented. Differential privacy is applied via a Gaussian technique to shared parameters to ensure high user privacy protection.

Momentum federated learning-based solution in FL-IoT

The authors (Rothchild et al., 2020) suggested a new optimization approach for FL-based IoT networks; improving communication effectiveness standards can be trained with this model. In Liu et al. (2020a); Liu et al. (2020b); Liu et al. (2020c), an entirely new design for FL is presented: momentum federated learning. It aims to improve the FL algorithms’ ability to join solutions quickly. It minimizes the loss of functions in IoT devices.

The use latest iteration of WIFI in FL-IoT

Wi-Fi has become an integral part of our daily lives and is also one of the essential communication protocols for the Internet of Things. The IEEE 802.11 technical group has recently started discussing the possibility of issuing the IEEE 802.11 Wi-Fi protocol with a higher processing capacity, the next version of the Wi-Fi standard (Restuccia, 2021). In the 5G and 6G era, these standards provide service in deploying FL-IoT edge networks with smart IoT devices.

Deep learning accelerator solution for FL-IoT sensors

The research by Lane et al. (2016) suggests a deep learning accelerator based on software that facilitates AI and DL training on mobile devices. The central concept is a set of diverse processors, such as graphics processing units (GPUs), in which each computing unit uses various computational resources to process different inference phases of deep learning models. In Itahara et al. (2020), the authors outline a simple model training method that addresses the transmission cost connected with on-device FL training.

FedGKT-based solution for FL-IoT

Federated group knowledge transfer (FedGKT) is a method that can potentially lower the amount of learning memory needed for efficient on-device learning. IoT devices can transfer data from small CNN models to a more prominent CNN on a cloud server. This strategy will reduce the trained memories necessary for efficient on-device learning.

DIoT solution against the attacks of the adversary in FL-IoT

The Internet of Things system has to be able to recognize malicious connections that have the potential to corrupt the training model. The solution is implementing a simple security procedure and identifying devices in the IoT system that aren’t functioning properly. FL can provide an approach to IoT devices as a distributed system that can protect devices (Ghimire & Rawat, 2022). Defense Internet of Things (DIoT) is the IoT device gateway solution that employs an FL method for anomalous detection-based security vulnerabilities (Nguyen et al., 2019).

FedMask framework-based solution for limited on-device resources of FL-IoT

One strategy for speeding the training process for edge applications that use machine learning is customized and specialized hardware. Edge devices have limited resources for memory capacity, processing power, and access to information. Federated Mask (FedMask) is an efficient framework in terms of processing and communication. When the Fed-Mask algorithm is implemented, each node can learn a sparse binary mask that is heterogeneous and structured (Li et al., 2021). IoT devices can create a sparse model using this approach, resulting in lower computational costs and a smaller memory footprint. Table 3 shows the solution models to the above challenges in FL-IoT for implementation in organizations.

Conclusions and Future Work

ML is expanding and changing the technological world. FL applications are also facing challenges. Privacy is essential for everyone. Training an ML model using data stored on centralized servers is difficult. Different challenges and issues with FL-IoT integration are described in the manuscript mentioned above and improve businesses’ performance. We have shown the different challenges of federated ML with the IoT, which are described above. We have also shown solution models and methods which are solving these challenges. These solution models for FL-IoT are described in detail and mentioned in the above table and figure. There are still a lot of challenges in the field of FL with IoT. Despite the recent technological advancement, different models have made it possible to handle these challenges. Suppose we can find more solutions on time. In that case, it will be a tremendous driving factor for improving FL’s industrial and academic domains with the IoT.

Supplemental Information