The designs of several fundamental challenges, including privacy, security, communication costs, heterogeneity, architecture and, differ depending on the domain and the use cases considered. Although several studies have focused on FL and IoT and discussed these methods in depth in terms of FL-IoT architecture, data services and applications, there are still more studies that have been conducted. The aim of this study is to provide an overview of the current state of research and summarize the most advanced methods that have been developed recently to solve these difficulties. As part of our research, we examine articles in related fields and conduct a thorough review of the latest survey studies in these areas. We use several categories to categorize the topics covered in FL survey articles. These categories include communication costs, heterogeneity, privacy/security as a primary challenge, FL architectures, implementation frameworks, communication protocols, and FL applications in different domains. The main Contributions of the manuscript are as follows:

1) The main focus of this manuscript is to thoroughly investigate and analyze existing FL-IoT reviews.

2) The manuscript classifies FL research into broad categories of communication protocol, implementation framework, FL architectures, challenges, and application areas.

3) It will address the various challenges of IoT applications and explore how FL can address these challenges.

4) Finally, the advantages of FL over traditional ML models will be discussed and a comparative analysis between FL (IID and Non-IID) and centralized ML algorithms will be conducted.

The following research questions (RQs) were formulated to accomplish the aim and objective of the review.

RQ1: What is the relationship between FL and IoT?

RQ2: What is the implementation framework for FL-IoT?

RQ3: What are various application areas of FL-IoT?

RQ4: What are various Challenges and limitations of FL-IoT?

RQ5: What is the need of FL over ML?

Therefore, the notable contribution of this research is as follows:

Federated learning uses the network of IoT devices to train machine learning models locally, which enables collaborative learning while protecting privacy and avoiding data aggregation at a central location. Further relationship between FL-IoT discussed in Section 4.3.

The paper’s structure is structured as follows: In Section 3, we provide an overview of the fundamentals of FL with IoT. Section 4 delves into the discussion of the FL-IoT framework, protocols, and various architectures of FL. Popular applications of FL, particularly utilizing IoT network data in domains such as smart healthcare, smart cities, smart industry, agriculture, transportation, and finance, are detailed in Section 5. In Section 6, we address the challenges that arise due to the decentralized nature of FL.

The literature search used the IEEE, ACM, arXiv, Elsevier, and Web of Science libraries. The general search terms used are: ‘federated learning’, ‘federated learning Internet of Things’, ‘Internet of Things applications’, or ‘federated learning architecture’. Following the initial literature investigation, each article’s title, keywords, and abstract were examined, and possibly relevant articles were obtained and tested for suitability using full-text articles. The PRISMA flow diagram provides a comprehensive overview of the research selection process. Fig. 1 depicts the entire procedure of searching and selecting literature. This procedure consisted of four stages: Identification, screening, eligibility, and inclusion. In the identification stage, 160 papers were gathered and 22 duplicate papers were removed. Furthermore, papers were screened with titles and keywords and 44 papers were excluded after screening. 104 papers were selected for full-text eligibility criteria, and out of them 12 were excluded because of no relevance and did not focus on quantitative evaluation. Therefore, 92 papers were chosen for full-text access.

The authors utilized five databases to conduct their electronic search, including IEEE, Elsevier, Arxiv, ACM, and WoS. The search was restricted to the English language. The AND function was utilized as a logical operator. A targeted search supplemented the computerized search. This consisted of a Google Scholar online search and a manual examination of the cited references of relevant publications using the search approach.

The search aimed to create a preliminary list of research that will be evaluated in more detail. The papers were then examined to determine whether they were appropriate and could be utilized to answer the research questions formulated, further we summarize some of the studies chosen based on the formulated research questions.

FL is an ML approach introduced by Google scientists that makes it possible to train models on decentralized devices with raw data without exchanging it. In an FL system, the model is trained jointly on the local data of each device, and only the model updates are shared with an aggregation server or between devices [28,29]. This privacy-friendly technique enables the development of ML models for distributed datasets while mitigating the privacy concerns associated with sharing sensitive information. It is particularly beneficial in scenarios where data is sensitive, located in different locations or cannot be centralized due to limitations such as network bandwidth. This approach is consistent with the principles of data privacy and security and is therefore suitable for applications in healthcare, finance and other areas where the protection of sensitive data is of paramount importance [30].

FL functions by conducting model training directly on individual devices, enabling the learning process to occur at the data sources themselves for both training and prediction [19]. Following the training phase, the models or model updates are transmitted back to a central server for aggregation. Later, the consolidated model is sent back to the devices utilizing principles derived from distributed computing, allowing efficient tracking and redistribution of models across diverse devices [31]. The process starts with the training of the local model on your client side. After training the model, each client sends its local model to the aggregation server, leaving the raw data on the client side. The aggregation server receives all local models from each client and starts training. After training the global model, the aggregation server sends this global model to all clients. This collaborative method maximizes training quality while minimizing the potential for data breaches. The local sneakers then retrieve the global update from the aggregator and continue this iterative process of computing subsequent local updates until the global training is complete [32].

Internet-based applications are in high demand today. Therefore, the IoT is the most important technology for the development of Internet-based applications. An IoT is a network in which different physical devices are connected to the Internet via different routers or network devices and exchange data. It is an intelligent technology that reduces human effort and makes physical devices easily accessible [33]. It enables remote control of devices and also has an autonomous control function that can be used to control any device without human interaction. It enables communication from person to person, person to thing and thing to thing. The IoT can be a thing in a world where everything can be connected and communicate with each other. There are numerous technologies such as big data, data analytics, AI, ML, wireless sensor networks (WSNs) and various sensor technologies, etc. In WSNs, multiple sensor nodes are deployed in different areas to monitor and control the relevant environmental conditions and collect the data [34,35]. IoT became popular after the invention of Radio Frequency Identification (RFID) chips.

The diverse types of communication in IoT systems are mentioned as follows [35]: i)

People to People (P2P): Communication/data transmission occurs between people to people through video calls, social calls, or social communication. It is also called “collaborative communication.”

IoT has a different vision to benefit various areas, such as the environment, industry, medicine, transportation, etc. Many researchers have explained the IoT in different ways in specific aspects and areas of interest [36]. The working model of IoT is structured in Fig. 2.

An IoT system consists of smart devices with embedded capabilities. These devices gather data through sensors and transmit it to an IoT gateway. The data is then either sent to the cloud/server for analysis or analyzed locally. These devices typically operate autonomously, with minimal human intervention. Users interact directly with the devices, configuring them, providing instructions, or accessing the data independently. IoT technologies have enhanced agricultural capabilities [37]. It can support farmers at any phase of their farming operations by offering the latest crop and weather data, enabling remote monitoring of their farms. Additionally, early detection of agricultural issues is possible, preventing the spread of diseases and safeguarding production. Agricultural IoT applications play a crucial role in boosting agricultural output and minimizing crop losses due to diseases [24].

The FL concept within IoT networks consists of two primary components: Firstly, the data clients, which are exemplified by IoT nodes, and secondly, an aggregation server as shown in Fig. 3. In this framework, IoT devices act as data clients, contributing their local data for model training. Meanwhile, an aggregation server strategically plays a crucial role in consolidating and processing the decentralized insights gathered from the various IoT devices.

This collaborative learning paradigm allows IoT devices to enhance their models collectively without the necessity of transmitting raw data to a centralized server. The aggregation server facilitates the synchronization of model updates, thereby promoting a more privacy-preserving and efficient approach to ML in the context of IoT networks. By decentralizing the learning process, FL concerns related to data privacy, network bandwidth, and latency while fostering continual improvement of models in a distributed and cooperative manner [38].

The FL-IoT process typically encompasses three main steps [28]:

1. Initialization of global model: The central server outlines the training task, including defining the intended application, generating an initial global model, setting hyperparameters, and determining the objective function j (θ, Dn) for each client n. The initialized global model is disseminated to local participants involved in the training.

The mathematical expression of the objective function is given by: (1)j(θ,Dn)=ΣnL(f(θ,x),y)

θ represents global model parameters, and Dn denotes the local dataset for the client. F (θ, x) predicts the output of the model for input x using θ, and y represents the true label. L denotes the loss function.

2. Local training of model: Each client participating in the FL network has its unique dataset. In every training round, a subset of clients or devices, denoted as Cn, is chosen, where n takes values from 1 to K, and K represents the number of selected clients. Each selected client Cn performs local model training using its local dataset Dn, optimizing local parameters θn to minimize the objective function J (θn, Dn). (2)θn(m1+1)=θn(m1)−η.∇j(θn(m1),Dn)

η is the learning rate, m_1, is for the iteration, and ∇J signifies the gradient function.

3. Local model aggregation: After the local model has been trained, the subsequent step is to collect updates from selected clients and then combine them to create an updated global model. The following are some of the aggregation methods that are utilized: Weighted averaging, simple averaging, and several others: (3)θ_g(m1+1)=N/1Σθn(m1+1)

θ_g signifies the updated global model parameters post-aggregation.

This method has substantial potential by achieving impressive learning accuracy, putting an emphasis on the preservation of privacy, and decreasing communication overhead, as proven in a comparative evaluation vs. traditional methodologies.

The communication protocol oversees the complete training process of FL, dealing with issues like security, unreliable device connections, and availability [39]. In the FL system, a cloud-based distributed service serves as the FL server, while end devices, such as phones, participate in FL operations. The protocol involves devices signaling readiness for FL operations, initiated by the server for specific FL populations identified by a global name. During each round, a subset of devices is selected to process a specific FL task, and devices remain connected to the server throughout the round. The server sends an FL checkpoint with global model parameters to participants, collects checkpoints from participants, and updates the global state for subsequent rounds. This iterative process ensures collaborative learning across distributed devices in the FL system. The FL communication protocol is structured into three phases during each training round: i.

Selection: Devices that fulfill certain suitability criteria regularly establish communication with the server via bidirectional streams. The server monitors the availability of clients by checking the vitality of these streams to ensure effective communication scheduling. In addition, an FL parameter server uses a client selection algorithm, such as FedCS [40], to choose active clients for training round participation. These selected clients then carry out a specified FL task.

In addition to the efficiency of communication, ensuring the security of communication during the transmission of local updates is another challenge that needs to be considered.

iv.

Secure aggregation: It refers to the use of cryptographic methods such as encryption, homomorphic encryption, secure multiparty computation (SMPC) and threshold cryptography. These methods are used to ensure the confidentiality of model updates transmitted from the participating devices to the central server. Encryption and homomorphic encryption are used to make individual updates unreadable during transmission so that the central server can summarize them without accessing the raw data [38].

To make the implementation of FLs easier, several frameworks and libraries have been developed. The building of FL models may be simplified with the help of these libraries, which include tools, application programming interfaces (APIs), and pre-built functionality.

Tensor Flow Federated (TFF): TFF is an open source framework for ML designed to perform computations on data distributed across multiple locations. It serves as the foundation for TensorFlow-based FL. Google is responsible for developing this framework for Python 3 for FL. In the real world, this paradigm has been put into practice for the first time. By using TFF, researchers can test new algorithms on their models and data. The interfaces of TFF consist of two basic levels. These layers are the FL API and the Federated Core API. When using TensorFlow, the FL application programming interface is an interface that simplifies the execution of actions associated with FL, model training and model evaluation. The three components that make up the interfaces provided by this layer are: Datasets, Federated Computing Builders and Models [43].

PySyft: PySyft makes it easy to identify and isolate private data from training models by using FL principles protected by a variety of privacy-enhancing methods. The only mode of operation PySyft is capable of is simulation mode. It must be combined with PyGrid and other projects of the OpenMined ecosystem to enable federated mode. Nevertheless, several projects, such as PySyft, are currently still under active development. The goal of PySyft is to make privacy-friendly ML techniques widely usable by providing Python connections and an interface reminiscent of traditional techniques. This enables the development and integration of new methods. One of PySyft’s main goals is to provide the highest possible level of protection to customers participating in the training. PySyft can provide robust privacy guarantees to data consumers by developing and deploying automated differential privacy protection. These guarantees are independent of the ML architecture used and the data itself. In this particular scenario, PySyft is primarily concerned with encryption and improving privacy security for clients through the use of homophobic encryption or encrypted computation. TensorFlow and Pytorch are the two types of libraries supported by PySyft [44].

Flower (FL over the World Wide Web): The impressive open source architecture called Flower, an acronym for “FL over the World Wide Web”, was developed to simplify FL operations across remote devices. The development of Flower was part of a research study conducted at Oxford University. Extending or rewriting a large number of components enables the creation of new, contemporary systems that can be stored for later use. Flower takes advantage of ML frameworks, even though the different frameworks for ML each have their own competencies. PyTorch, TensorFlow, Sci-kit-learn, TFLite and raw NumPy are some examples of well-known frameworks for ML. Flower not only has a large user base, but also a large community. With this in mind, this framework includes. This framework is characterised by its excellent documentation and tutorials that make implementation quick and easy [45].

IBM FL: Solutions for customer privacy protection, regulatory compliance in data integration and big data are the main focus of IBM’s FL, which is deployed in various locations. There is a wide range of deployment scenarios that can be configured with IBM FL. It is common for data centers and cloud instances from different vendors to participate in FL. IBM’s contributions to FL demonstrate IBM’s commitment to advancing innovative technologies with a strong focus on data protection. IBM’s FL integrates seamlessly with multiple ML libraries such as Keras, TensorFlow, PyTorch, SK Learn and RLLib, ensuring compatibility with its architecture. Providing APIs to create new FL algorithms increases its versatility and enables customization to different ML libraries and paradigms. In addition, fairness strategies within FL are emphasized to mitigate biases and reinforce the commitment to ethical and inclusive ML practices. Please note that the paraphrased content combines information from the indicated paragraphs for coherence and clarity [46]. Table 2 shows the comparison of supported features of some existing frameworks. The cell is left empty if the system does not support the corresponding feature.

Federated AI/(Federated AI Technology Enabler): It is an open source initiative with the aim of creating a safe and collaborative AI system. FATE-Flow is the name of the platform that hosts the FL pipeline. Inference processing, modeling, training, review and publishing are all applications of this technology. FATEBoard is a platform that allows users to view, explore and better understand specific ML models. It offers a variety of visualization options and displays the results in the form of tables or charts in different formats. The FATE network not only facilitates communication between devices, but also offers developers and scientists the opportunity to construct algorithms by using the Federation’s APIs with the help of various tools. The FATE framework is integrated into the KubeFATE element, which is the final component. By using cloud-based technologies such as Docker and Kubernetes, this aspect is used to deploy FATE. FATE not only offers comprehensive documentation, but is also easy to implement in practice [47].

In the next section we discussed the performance parameters utilized to evaluate FL-IoT framework along with implementation steps of FL and various vulnerabilities that were examined during FL-IoT implementation process.

Some papers, such as [4,9,15,51] explain FL architectures.

FL-IoT networks need optimization and convergence techniques to train models while addressing the challenges posed by decentralized, heterogeneous, and resource-constrained IoT devices [58,59]. Some key techniques are discussed as follows:

FL-IoT optimization algorithms: Various optimization algorithms such as Federated Averaging (FedAvg), Federated stochastic gradient descent (FedSGD) and FL with Adaptive Gradient Clipping (FLAG) are popular, which deal with decentralization and non-IID data distributions while incorporating FL with IoT devices. The local model updates from all IoT devices are collected by FedAvg and sent to the aggregation server. FedSGD is used to improve the convergence speed and efficiency of communication.

Model compression: IoT devices have limited computational resources and bandwidth. Model quantization and compression techniques such as knowledge distillation, weight pruning, and quantization are used. These approaches reduce the model size and minimize the communication cost during updates.

Data privacy-preserving techniques: Data privacy of individual clients is the main aim of FL. Secure aggregation, differential privacy, and homomorphic encryption approaches are used to ensure data privacy on the client side.

Learning rate: The AdaGrad, RMSpoprop, and Adam tools serve to offset the effects of non-stationary data distributions and accelerate convergence by dynamically adjusting the learning rate.

AI-based technologies are already widely used in the field of intelligent healthcare. One example of this is the use of intelligent imaging to identify diseases. When it comes to classic AI models, one of the most important issues is the question of data privacy, which arises from the transfer of data to the cloud for computer training. This is because, compared to other areas, the data stored in healthcare systems is extremely sensitive and subject to health regulations such as the United States Health Insurance Portability and Accountability Act (HIPPA) [60]. Traditional methods of AI are not sufficient to meet the privacy requirements of modern healthcare, emphasizing the need for IoT to protect privacy in healthcare ecosystems. FL operates in a decentralized manner, unlike traditional AI systems that depend on a central server for data analysis and involve the sharing of data. The use of this decentralized technology eliminates the need for data sharing, providing a solution that prioritizes privacy for healthcare applications [61].

The integration of smart devices and advanced infrastructures as well as integrated monitoring systems together with communication frameworks form a smart city. The goal of these ecosystems is to improve the quality of life of residents by facilitating the seamless supply of food, water and energy to end users. FL-IoT is a revolutionary technology that is revolutionizing smart cities, especially in the area of urban transportation. Table 5 shows the various performance metrics related to FL-IoT.

A variety of clients, such as vehicles, are involved in the process of cooperatively training globally shared AI models through the use of FL, which was recently developed to bring AI capabilities to the network edges to enable intelligent transportation. This eliminates the need for lengthy data transfer and protects user privacy.

FL can also provide realistic solutions to update the knowledge of industrial intelligent systems with different application areas such as automation and Industry 4.0 or Industry 5.0 without jeopardising data sharing or privacy. These solutions can be implemented without jeopardising the integrity of the data. For managing data exchange between robots for industrial activities, such as traffic routing, FL is an appealing technique. This is because it eliminates the possibility of unforeseen delays in network transmission [78,79].

FL-IoT can bring about a revolutionary change in the field of smart agriculture by improving a variety of elements of agricultural practices. In the field of agriculture, FL refers to the process of training ML models by utilizing data collected from a variety of decentralized sources, such as fields and sensors. This strategy facilitates the implementation of precision agriculture, which in turn improves crop yields, resource allocation and long-term sustainability, while maintaining the confidentiality of harvested data. Researchers like Antico et al. and Aggarwal et al. [28,83,84] classify maize and rice leaf crop diseases using FL-IoT by keeping the leaf image information at the farmer’s location [85].

The use of FL in finance improves data protection by enabling companies to work together on training models without having to share sensitive financial data. It not only ensures compliance with strict regulatory requirements, but also enables the construction of more accurate fraud detection models, risk assessment algorithms and customized financial services. However, the application of FL in banking requires customized enhancements to address real-world challenges. To effectively address these difficulties, user incentives need to be managed efficiently and access to personal data needs to be restricted. In reference [75], authors discuss all of the possible obstacles and problems that might arise while adopting FL in the banking sector, as well as the solutions to these problems. In reference [86] highlights issues in the bank loan approval process, where manual approval is often required due to insufficient data for automating decision-making through ML models. FL proves beneficial in overcoming data scarcity challenges by involving multiple financial institutions in global model training. In reference [87], the authors analyze the credit risk assessment process, proposing an FL-based model for predicting credit risks.

When it comes to dealing with the edge nodes of a network, there are a lot of challenges that can arise with FL from the resources. Various FL clients can pose particular challenges due to their limited performance or memory capacity, as well as their high energy consumption. A client with more memory may be able to perform more demanding calculations than a client with less memory. In addition, the energy budgets FL clients set during training may not be sufficient to meet the demands placed on the system. The combination of these characteristics leads to an increase in communication overhead and a decrease in system efficiency, both of which can lead to major difficulties when conducting model training. Utilizing EC for local computation on IoT devices proves to be a solution to reduce the resource burden and enable successful FL integration into various IoT ecosystems.

In FL environments, the inclusion of multiple devices offers performance benefits by training ML models on large datasets. However, there are also challenges, such as communication bottlenecks and increased computational costs, especially with a large number of clients. To address these issues, Hard et al. [91] focused on essential hardware prerequisites, required memory capacity, and computational capabilities for next-word prediction on a keyboard. Nishio et al. [40] addressed heterogeneous devices with varying computational power, aiming to optimize the computations on these devices, reducing processing times for lower-performance devices and minimizing upload times in poorer connection quality scenarios. Reference [92] shows a resource-aware FL architecture for mobile devices that addresses computational power constraints. The proposal includes a soft training technique to accelerate the training of straggler devices. These can partially train the model locally by hiding resource-intensive neurons that are later restored in the aggregation phase.

Network latency and convergence challenges in FL-IoT arise from delays in communication and model synchronization between distributed IoT devices. Limited computing capacity, energy constraints and bandwidth limitations contribute to these challenges. Managing network latency and ensuring timely convergence are critical to effective FL implementation in IoT, which includes strategies such as reducing device participation and optimizing communication rounds. Tackling communication overhead involves model compression techniques such as quantization and partitioning. Although this is effective, resource-constrained clients can face issues such as unsuccessful connections during training [93].

Wu et al. [94] presented a FL framework, FedKD, that focuses on adaptive mutual distillation of knowledge and dynamic compression gradient techniques. The authors conclude that the proposed framework reduces the communication cost by up to 94.8% and achieves better results by preserving the privacy of the data. An innovative approach is introduced in reference [95], Momentum FL by adding a momentum gradient framework at the aggregation server for accelerated convergence. Experimental results using the MNIST dataset are used to evaluate the convergence performance of MFL for a variety of machine-learning models.

Liu et al. [96] explore the approach applied to vehicular task offloading by discussing the various issues and challenges involved in edge servers with wireless networks.

Maintaining the confidentiality, integrity and privacy of data during model training on IoT devices is an important concern. However, due to the limited resources available, the privacy protection algorithms of the FL system may not work properly on these IoT devices [97]. Therefore, not only robust privacy safeguards need to be provided, but also new methods that are able to deal with fancy participants,that are efficient in terms of communication and that are frugal in terms of data processing. There are many studies that focus on the problems associated with peripheral devices that do not have sufficient computing capabilities. In order to build trust in FL-IoT applications, comply with privacy standards and develop effective communication and computing techniques for devices with limited power, it is essential to find solutions to these difficulties.

Zhang et al. [58] address security concerns in mobile and peripheral devices, proposing a Verifiable Privacy-preserving FL scheme to prevent gradient leakage and an online/offline method for lightweight gradient integrity verification. Reference [98] deploys privacy-aware Internet of Vehicles (IoV) services with FL in cloud-EC, utilizing distributed edge servers for model aggregation and homomorphic encryption for enhanced privacy.

Resource-constrained IoT devices, characterized by limited computing, storage and power resources, pose scalability challenges influenced by factors such as physical size and cost considerations. These devices, including drones and smartphones, are an essential part of applications such as environmental monitoring in smart cities and factories [99]. Their role in tasks like environmental monitoring and EC applications is significant [4]. Before integrating the FL into the IoT environment, it is important to close the gaps that currently exist and identify the latest hardware requirements for the IoT environment. Various hardware devices used in the FL-IoT EC scenario are listed in Table 6.

These above-mentioned hardware devices play a very important role in the FL-IoT computing environment and contribute to the distributive nature of FL.

In a heterogeneous environment, each device has different specifications, computing capacities, operating conditions, energy and storage capacities and data processing capabilities. The FL-IoT method can incorporate a wide range of heterogeneous devices in different ways. These may not only be alternative device performances, but also different platforms, different generations of devices or different amounts of data on which the local model is informed in this scenario. As a result, the duration of the training session can vary greatly from one customer to the next, and it would not be productive to assume that all participants have the same coverage [1]. The FL must be familiar with the various hardware configurations and be able to provide training on different platforms. In order to meet their system requirements, it is always necessary to select only the most trustworthy customers from among the connected customers. When it comes to successful FL collaboration, selective choice of participants that can be relied upon is absolutely essential. He et al. [100] explore training variation possibilities for heterogeneous devices in their paper. They present a formulation to maximize the efficiency of training, resource utilization, and heterogeneity. Feng et al. [101] examine the utilization of wireless power transfer and heterogeneous computing in FL. Their study suggests a framework for resource allocation and heterogeneous computing.

Effectively managing variations in heterogeneous hardware and non-IID data distribution is crucial for the success of FL systems, impacting the overall training procedure and global model accuracy. In 2022, Tahir et al. [102] offer a detailed analysis and overview of challenges associated with systematic and statistical heterogeneity. The study delves into various algorithms, including FedAvg, FedProx, FedPD, SCAFFOLD, and Fedmed, discussing their implications in addressing heterogeneity-related issues. Karimireddy et al. [103] focus on minimizing variance in local updates during data collection. Through experimental demonstrations, they find that SCAFFOLD exhibits a remarkable reduction in required communication rounds and proves resilient to data heterogeneity and client sampling issues, ultimately leading to faster convergence.

The scalability challenges of FL for IoT (FL-IoT) relate to difficulties in efficiently expanding the system. This is a major problem in terms of communication costs and aggregation of models on the central server with IoT devices. We need appropriate algorithms for aggregated models and communication protocols to overcome this scalability problem. In the FL environment, client selection is a very tedious task due to limited bandwidth and battery resources. Therefore, efficient client selection methods are essential to overcome these problems. Many researchers have explored these scalability issues with appropriate client selection mechanisms.

Zhang et al. [104] proposed a framework to solve the problem of scalability in order to minimize the communication costs between client and server. The authors presented a cooperative federated edge learning framework for high accuracy and low latency between mobile edge devices. Ye et al. [105] has addressed the problem of model aggregation and introduced an approach to model aggregation in vehicle clients, where a local model is selected and sent to the aggregation server while preserving the confidentiality of image quality and computational capability. FL-IoT thus needs to address scalability issues to successfully manage the growing network of IoT devices and maintain algorithms based on ML algorithms that are reliable and effective across different platforms.

ML methods have proven their effectiveness in various applications, but due to the centralized approach of traditional machine learning methods, they raise the problem of data privacy, have high communication costs and require more computational resources. The FL approach overcomes these problems by preserving the data on the client side and reducing communication costs as only the trained models are shared with the aggregation server. Many researchers have demonstrated the effectiveness of the FL approach over traditional ML algorithms. In Table 8 and Fig. 7, we have discussed some manuscripts that implement the FL approach and achieve better accuracy in training the model.

As it is clear from Fig. 7, there is a slight difference between the accuracies of ML and FL. In [69], the accuracy with the base ML approach is 96.8%, the accuracy with the FL approach is 91%, and in [28], the accuracy is 99% with both approaches. So, we conclude that with FL, we achieved almost the same results as with ML models by preserving data and reducing communication cost.

IID and non-IID mean the distribution of different data patterns to different devices. It is difficult to evaluate the actual non-IID performance of models because FL has a considerable number of hyperparameters, such as the number of clients, epochs, and the probability of a client failing. These hyperparameters change drastically from method to method, making it difficult to compare the performance of these algorithms. Dealing with non-IID data (device heterogeneity) is a major problem when implementing FL algorithms. In Table 9 and Fig. 8, we have discussed some manuscripts that implement the FL approach with IID and non-ID datasets and achieve better accuracy in training the non-ID data.

In Fig. 8, we compare the accuracy of the FL approach (IID and Non-ID) with the centralized approach and conclude that all three approaches have almost similar accuracy in terms of model accuracy, but with the FL approach we solve problems such as privacy, resource utilization and high communication costs. FL is an emerging distributed AI technique that has attracted great interest in realizing IoT services and applications that enhance privacy and are scalable.

Despite a thorough search across databases, some relevant publications may be missing because of the inclusion of only English-published studies, chosen search keywords, and database constraints.