Addressing Bias and Fairness using Fair Federated Learning: A Systematic Literature Review

1. Introduction

Due to the recent development of big data and LLM, issues of data, privacy protection, and security are being discussed more importantly than issues of data quantity. Representative examples include China's Cyber ​​Security Law, the Civil Code of the People's Republic of China, the European Union's General Data Protection Regulation (GDPR), the U.S. Health Insurance Portability and Accountability Act (HIPAA), and Korea's Personal Information Protection Act, which focus on privacy protection and data security.

Centralized data processing, which is mainly used for artificial intelligence, has limitations such as privacy protection, data security, data silos and accessibility, bandwidth and efficiency, scalability, and bias. Therefore, Federated Learning (FL), which is a distributed learning method, allows individual users in different spatial locations to collaborate with other users to learn machine learning models compared to centralized learning methods, and all personal data that may contain sensitive personal information can be stored on the device. By solving these core problems, FL not only provides a practical solution to the limitations of centralized learning, but also opens up possibilities for AI applications in sensitive or restricted environments such as healthcare, finance, and mobile services, and can enjoy the advantage of obtaining well-trained machine learning models without having to send privacy-sensitive personal data to a central server.

However, despite the growing interest in FL as a privacy-preserving machine learning framework, it still has significant limitations in addressing inherent bias and fairness issues. Currently, FL has problems with bias and fairness due to the distributed nature of data sources, which must be solved to effectively protect personal information, such as personal information, insufficient data volume, and statistical heterogeneity. Since FL models may perform well in the majority group but may not accurately predict the results of the minority group due to differences in data distribution between clients, it becomes complicated to detect and mitigate bias, which leads to fairness issues. Bias in FL can arise from uneven data distribution across devices, and models may not perform equally well across all users or scenarios due to skewed local data representations. Similarly, fairness in FL has not been sufficiently explored, particularly in terms of how to ensure fair model performance across diverse participant groups with varying data quality and quantity. Addressing these issues requires sophisticated fairness-aware strategies that not only improve model accuracy across diverse datasets, but also ensure that no group or demographic is systematically disadvantaged. Thus, pursuing fairness in FL requires a delicate balance between protecting user privacy, ensuring robust model performance, and promoting fair outcomes across all participating entities.

The need for Fair Federated Learning (FFL) provides a mechanism to reduce the various problems of FL, especially imbalances such as data heterogeneity, unfair resource distribution among clients, and the problem of discriminatory model training. FFL proposes a methodology that not only respects data privacy but also actively mitigates bias and promotes equity in model results. This includes designing novel algorithms that can handle heterogeneous data distributions and promote fairness, and potentially tailoring the learning process to give more weight to disadvantaged groups or develop models that can generalize across diverse data sources. FFL incorporates specific strategies such as reweighting training data across multiple nodes, modifying learning objectives to emphasize fairness, and using regularization techniques to penalize unfair outcomes. In addition, FFL's privacy-enhancing strategies such as data distribution, application of encryption techniques, differential privacy, access control and authentication mechanisms, and transparency of the FL protocol help enterprises and institutions develop fair and unbiased AI models while protecting their customers' data. By incorporating these strategies, FFL maintains data in local nodes, thereby enhancing privacy, and significantly improving the fairness of the resulting model. This dual focus on privacy and fairness makes FFL a powerful framework, especially for applications such as healthcare, finance, and public services, where fairness and non-discrimination are as important as data security. Therefore, leveraging FFL focuses on improving the fairness of FL and ensuring that all participants benefit fairly. FFL’s data privacy, model fairness, and non-discrimination effectively address the ethical issues inherent in machine learning systems by addressing privacy and fairness issues that may arise in centralized learning systems. This FFL approach can contribute to making FL a more practical and sustainable technology by addressing both technical and ethical issues. The evolution to FFL is essential for deploying FL in sectors such as healthcare and banking, where decision-making impacts are profound and have a direct impact on human lives.

This study focuses on systematically exploring how existing research addresses (or fails to address) the nuances of bias and fairness in the FL environment. We emphasize the need to integrate fairness into the FL paradigm and describe potential research directions for improving the reliability and applicability of FL models to ensure fairness and privacy. It is meaningful to study how technologies that address these bias and fairness issues can be further developed in the future. In addition, we systematically review recent research on privacy and data security and review, and examine the challenges and directions that should be derived in the future. The main contributions of this study are as follows: (1) We reviewed the development of FFL. (2) We explained the paper preprocessing process. (3) We reviewed the sources of bias and fairness in FL. (4) We reviewed the work of FFL in three aspects: data segmentation, privacy protection mechanisms, and methods to address heterogeneity. (5) We discussed the applications, challenges, and future research directions of FFL.

Figure 1. Overview of FL, FFL.

Figure 1. Overview of FL, FFL.

2. Methodology

This study followed the “Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)” guidelines[1].

2.1. Literature Review and Search Strategy

We performed an extensive search of articles published up to July 2024 in databases such as PubMed, Embase, Web of Science, IEEE Xplore Digital Library, ACM Digital Library, ELSEVIER, SpringerLink, Google Scholar, Semantic Scholar, Cornell University, Computing Research Repository, Database Systems and Logic Programming, and identified relevant English-language articles using the following search strings, slightly modified to fit the database: “Fair Federated Learning”, “Fairness-aware Federated Learning”, “Incentive Federated Learning”, “Efficient Federated Learning”, “Privacy-preservation Federated Learning”, “Recent Advanced Federated Learning”, “Fair Distributed computing methodology”. Similarly, we performed searches in Chinese databases CNKI and Wanfang. In addition, we manually searched various commentaries, editorials, and conference proceedings to ensure completeness.

2.2. Study Selection

Studies were included if they met the following criteria: 1) Bias of Federated Learning; 2) Problem of Federated Learning; 3) Future work of Federated Learning. Case reports, post hoc analyses, reviews, meta-analyses, and studies published in languages ​​other than English and Chinese were excluded.

Figure 2. Preferred Reporting Items for Systematic reviews and Meta-Analysis diagram.

Figure 2. Preferred Reporting Items for Systematic reviews and Meta-Analysis diagram.

2.3. Selection of Eligible Articles

The initial electronic search yielded 24,278 integrated studies from the selected databases, and after removing duplicates, 22,263 articles remained. Articles were further reviewed for eligibility based on titles and abstracts. After full-text review of these articles, 217 articles were deemed suitable for the final analysis. Figure 2 shows the study flow diagram according to the PRISMA guidelines. One Chinese article and 216 English articles were included in the final analysis. All studies were published between 2015 and 2024. Figure 3 shows the trends in FFL research over the past five years as investigated in this study. Figure 4 shows the default settings for each FFL item.

3. Causes of bias in FL and fairness in FFL

The systematic review covered a wide range of studies focusing on the causes of FL bias by category, as detailed in Table 1. The main findings of the study highlight bias and fairness in improving privacy in FL. The analysis was subdivided into several predefined criteria: conventional bias sources, sub-sampling, party selection, dropouts, heterogeneity, and method. Figure 1 shows the framework of conventional FL and FFL.

3.1. Causes of Bias in FL

The causes of bias in FL are generally as follows:

3.1.1. Conventional Bias Sources

Conventional Bias Sources in FL come from various causes that can occur just like in traditional machine learning environments. These traditional sources of bias are not unique to FL, but are connected to problems found in general data science and AI systems. Historical Bias, which may reflect the way data was collected or the real-world inequalities or stereotypes represented by the data, is a bias inherent in the data itself[16].

Measurement Bias occurs due to errors in the data collection process[12]. Sampling Bias occurs when the dataset does not adequately represent the entire population[17]. If data from a specific group is oversampled or omitted, this can bias the model’s ability to generalize. Aggregation Bias can occur during the process of aggregating models learned from multiple clients[18]. If the data distribution of each client is different, data from a specific client may have a greater impact than data from other clients during the central aggregation process. These traditional sources of bias should be considered in the design of FL systems, and if not properly managed through data collection, processing, and model learning methods, they can result in unfair predictions or decisions.

3.1.2. Sub-Sampling, Party Selection, and Dropouts

Among biases, subsampling, party selection, and dropout are particularly noteworthy due to their impact on the diversity and representativeness of data used to train models. The training purpose of subsampling refers to the practice of selecting a subset of data from a larger dataset. If the subset is not representative of the entire dataset, it can introduce bias and unintentionally emphasize or underestimate certain patterns or features in the data, which can distort the learning results.

Party selection in FL involves selecting clients (or parties) to contribute to the training of a shared model. The selection process can be a source of bias if certain types of clients are consistently selected over others. This can result in the model not performing well for all potential users, especially those whose data profiles are not well represented during the training process. Dropout refers to a situation where a client unexpectedly fails to complete a training task or fails to send updates back to the central server or aggregator of FL. This can lead to incomplete learning cycles, and the aggregated model updates can be distorted due to missing model updates from some clients. Frequent dropouts can lead to a type of selection bias, where the model overfits to the characteristics of clients with more stable or reliable engagement.

3.1.3. Data Heterogeneity

In FL, data heterogeneity refers to the variation in data distribution among different clients participating in the learning process. Heterogeneity is one of the major challenges of FL because it can significantly affect the performance and fairness of the resulting model. In FL, data heterogeneity occurs when data from different clients or devices differ in terms of distribution, volume, and features. This can occur due to different environments, user behaviors, and data collection protocols of different devices.

The impact of data heterogeneity on FL includes (1) Model performance: A model trained in a heterogeneous data environment may have difficulty generalizing well across all clients. A model that performs well on one client may not perform well on other clients due to differences in data characteristics. (2) Convergence speed: Data heterogeneity can slow down the convergence of the FL process. (3) Fairness issues: When data is heterogeneous, the federated model may be unintentionally biased toward data patterns of more dominant or more frequently participating clients.

3.1.4. Conventional Bias Sources

The impact of Fusion Methodologies on bias and fairness includes: (1) Variance reduction: A properly designed fusion methodology can reduce the variance of model performance across clients, enabling more stable and reliable predictions. (2) Fairness improvement: By adjusting the way local models are aggregated, biases that may arise from imbalanced or unrepresentative local datasets can be mitigated. (3) Generalization: Effective fusion can leverage the diversity of local datasets to improve the generalization of global models to unseen data and clients. Therefore, future directions for Fusion Methodologies focus on developing more adaptive, resilient, and fair methods for aggregating local models. This includes hybrid approaches that combine multiple methods and apply advanced machine learning techniques such as meta-learning and reinforcement learning to optimize the fusion process.

3.1.5. Systems Heterogeneity

System heterogeneity refers to the variability in hardware, network connectivity, computational power, and storage capacity of devices participating in the FL network. This type of heterogeneity is important because it directly affects the efficiency and effectiveness of the learning process across the distributed network. The effects of system heterogeneity on FL include: (1) Model learning efficiency: System heterogeneity can lead to significant differences in the time it takes for different devices to complete their respective share of the learning task. (2) Bias and fairness: If not properly managed, system heterogeneity can lead to bias in the training process.

3.2. Fairness and Recent Trends in FFL Research

FFL is a technique designed to provide a fair learning environment among various clients. This fairness is mainly classified into four types: Performance Fairness, Collaboration Fairness, Model Fairness, and Different client types Fairness. Each type plays an important role in the FFL environment and aims for balanced model development among clients.

Performance Fairness is fairness that ensures that all participating clients experience similar levels of model performance and is based on the principle that all participants should receive equal benefits. Since learning outcomes may vary depending on the amount or quality of data held by individual clients, performance fairness ensures that all participants contribute equally and receive equal value. Performance fairness increases motivation among participating clients and improves the reliability and satisfaction of the entire system.

Collaboration Fairness is fairness that ensures that all clients participate fairly in the learning process and that the costs and benefits incurred in the process are appropriately distributed. Collaborative fairness is important when multiple clients have different resources and helps all participants to be treated fairly and maintain participation. This promotes cooperation among participants and encourages long-term project participation, ensuring the performance and stability of the entire network.

Model Fairness refers to the fairness with which the trained model shows equal performance for various demographic groups. Model fairness is necessary to train a model without bias toward a specific group. A fairly designed model respects the diversity of users and becomes competitive in the market. Each of these types of fairness is essential to ensure that all participants benefit equally and contribute fairly in the complex environment of FFL. This enables the construction of a more sustainable and inclusive AI system.

Different client types Fairness considers the unique differences between clients, such as different computational capacities, dataset sizes, or data distributions. To fairly accommodate these differences, the FL process is customized to ensure that no client is disadvantaged due to its unique characteristics. The various fairness methods possessed by FFL are shown in Figure 5.

The purple mark on the map in Figure 6 is the distribution of FFL authors. Countries that received scores higher than 5 were the United States (80), China (78), Australia (11), Singapore (10), Republic of Korea (9), Canada (8), Great Britain (8), France (8), Israel (1), Sweden (1), Türkiye (1) Italy (1), and Denmark (1).

4. Bias for categorizations

In this section, we summarize the classification of FFLs in terms of five aspects: data partitioning, privacy mechanisms, applicable machine learning models, communication architectures, and heterogeneity resolution methods. The FFL metrics are listed in Table 2, and the fairness concept is illustrated in Figure 8.

4.1. Data Partition

FFL's data partitioning method is a differentiated approach that takes the fairness of FL as an important goal. As shown in Figure 7, according to the different distribution patterns of the data sample space and the feature space, FFL is divided into four categories: fair horizontal federated learning, fair vertical federated learning, fair federated transfer learning, and multimodal fair federated learning.

4.1.1. Fair Horizontal Federated Learning (FHFL)

Horizontal Federated Learning (HFL) is designed for scenarios where multiple participants (clients) share the same feature space but have different datasets with different samples. However, this form of partitioning inherently carries the risk of bias and unfairness. Different data such as data heterogeneity, imbalanced data, and non-IID data can lead to poor performance or bias due to distorted model training. FHFL is improved through various approaches that aim to mitigate bias and ensure more fair results for all participants, and ensure that the FL process is sensitive to data distribution and demographic differences among participants.

In FL, the bias is higher than that of centralized learning using models trained on the union of all data (1). For a wide range of fairness metrics, a global fairness training algorithm that can obtain the fairness of the global model using only summary statistics of local clients and directly minimizes the empirical loss was proposed[113]. An adversarial training method was proposed to train a FL model that satisfies individual fairness without accessing the underlying data distribution, thereby improving individual fairness and group fairness and protecting privacy[114].

4.1.2. Fair Vertical Federated Learning (FVFL)

Vertical Federated Learning (VFL) can cause bias in the learning process due to data imbalance, feature mismatch, privacy requirements, etc., which can reduce accuracy. Therefore, in order to solve the bias problem in FVFL, Secure Multi-Party Computation (SMPC), Fairness Algorithms, balanced data sampling, and weight adjustment can reduce bias while protecting privacy.

4.1.3. Conventional Bias Sources

Federated Transfer Learning (FTL) suffers from bias due to the lack of data homogeneity and limitations of pre-trained models when the model over-reflects the characteristics of the pre-trained data and fails to generalize to new data. Therefore, FFTL generalizes the model by using a multivariate approach, data compensation techniques, and adaptive learning processes to reduce bias and better adapt to new environments. Coston et al. proposed a new weighting method called Prevalence-Constrained Covariate Shifting (PCCS), which does not require protected attributes in the target domain, and Target-Processed Covariate Shifting (TFCS), which does not require protected attributes in the source domain[31].

4.1.4. Conventional Bias Sources

Multi-modal Federated Learning (MMFL) is a technology that learns by integrating various types of data modes (images, text, audio, etc.), but there is a risk of model bias due to imbalances between each data mode or overrepresentation of specific modes, which may reduce generalization ability. To address this bias, MMFFL adjusts weights through data balancing techniques, algorithm modifications, and diversification of model evaluations, and alleviates bias through independent evaluation metrics.

Figure 8. Overview of fairness in FFL.

Figure 8. Overview of fairness in FFL.

4.2. Privacy Preservation Mechanism

The main privacy protection mechanisms used in FFL include encryption methods, anonymization methods, perturbation methods, hardware-based protection, and Hybrid PPFL methods. These are similar to the mechanisms of basic FL, but FFL places greater emphasis on fairness and privacy. Figure 9 shows various types of privacy protection methods in FFL. Personal information protection indicators include pseudonymization technology and personal information protection technology.

Pseudonymization technology is used to maintain the usability of data while minimizing the possibility of personal identification. It is a method of lowering the possibility of identifying original data by removing or replacing direct identifiers from personal information. Table 3 presents the current landscape of personal information pseudonymization technologies.

4.2.1. Fair Cryptographic methods

4.2.1.1. Homomorphic Encryption, HE

In FFL, HE is used to ensure that each participant's data is processed fairly, and analysis can be performed without the risk of sensitive information leakage. Among the current HE methods, Fully Homomorphic Encryption is relatively more efficient and supports larger ciphertext sizes, but it is a slow process due to the large amount of computation required, while Partially Homomorphic Encryption is less efficient and supports smaller ciphertext sizes, but it is a fast process and requires less computational power. Jin et al. proposed FedML-HE, the first practical FL system with efficient HE-based security model aggregation, which significantly reduces computation and communication overhead during learning by selectively encrypting sensitive parameters, while providing customizable privacy features[38].

4.2.1.2. Secure Multi-party Computation, SMPC

In FFL, SMPC guarantees the data privacy of participants by performing necessary computation without directly sharing data, and optimizes computation among multiple participants through multi-party computation, which can effectively scale even in large-scale networks. PrivFairFL by Pentyala et al. proposed a method to train group fair ML models in privacy-preserving and cross-device FL without requiring clients to disclose sensitive attribute values ​​by combining FL with SMPC and DP[44]. SMPAI by Mugunthan et al. reduces privacy and data leakage risks based on SMPC and DP[45]. FL-IPFE by Chen et al. proposed a new SMPC algorithm to protect local gradients[46].

4.2.1.3. Secret Sharing, SS

As data security and privacy become increasingly important in FFL, secret sharing has become an essential tool for safely storing and managing access to critical data. This is especially necessary for distributing critical information among various managers or systems. FairFed by Ezzeldin et al. enabled group fairness using secret sharing techniques in a FL setting[50].

4.2.2. Anonymization methods

Data masking and shuffling are attempts to preserve data validity as much as possible while maintaining data confidentiality by transforming the actual value of the data or changing the order, respectively.

4.2.2.1. Data Masking

In FL, data masking is useful for enabling each client to collaborate with the central server to train a model without directly revealing its local data. Each client can participate in model training without exposing its personal information by masking its own data and sending it to the central server. The PrivMaskFL framework by Xiong et al. uses dynamic participant aggregate masking together with adaptive differential privacy to address bias and fairness issues and reduce communication overhead in FL environments[55].

4.2.2.2. Data Shuffling

In FL, data shuffling can resolve data imbalances between clients and prevent models from being overfitted to the data distribution of a specific client. When training a disease prediction model using medical data from various regions, the characteristics of the data collected from each region are different, so shuffling can ensure that the data characteristics of each region are evenly reflected in the model training process.

4.2.3. Perturbative methods

4.2.3.1. Differential Privacy, DP

DP in the existing FL is a privacy-preserving technology that adds small noise to the query response to prevent personal information from being inferred. In FFL, DP focuses on maintaining the accuracy of the overall model while preventing the exposure of each participant's data, and helps reduce discrimination that may arise from data-based decisions of individual participants, so that all participants are treated fairly. Ling et al.'s FedFDP used DP to adjust fairness while maintaining privacy and incorporated clipping applied to the loss value to maximize utility[67]. Salim et al. developed a DP-DLP mechanism to properly adjust the extended variance of Gaussian noise to meet the requirements of DP at a unique privacy level and hide the influence of a single data point on the learning stage[68].

4.2.3.2. Additive/Multiplicative Perturbation methods

The greatest advantage of additive perturbation is its ability to preserve the usability of data while enhancing privacy[117]. It protects privacy by adding controlled noise to data, while at the same time preserving the core statistical properties of data required for data analysis or machine learning algorithms. The greatest advantage of multiplicative perturbation is that it can preserve the privacy of data while preserving the original structure and relationships of data[118]. It effectively reduces the sensitivity of data while preserving the usability of data as much as possible. As a result, it can maintain high accuracy in data analysis and machine learning tasks.

PILE by Tang et al. developed a verifiable perturbation scheme that makes confidential local gradients verifiable for gradient verification[70]. Hu et al. proposed Fed-SMP, which mitigates the impact of privacy on model accuracy by using SMP where the local model is first sparse before being perturbed by Gaussian noise[71]. Li et al. proposed ANP, a stepwise noise perturbation operation, to better balance privacy and accuracy[72].

4.2.4. Hardware-based protection

Hardware-based protection methods in FL are approaches that use specially designed physical hardware devices to enhance the security and privacy of data. Hardware-based methods often provide stronger security than software-based methods, making them more resistant to a variety of cyber threats. Researchers at Imperial College London are exploring the intersection of systems, security, and machine learning, focusing on using hardware hardening mechanisms to ensure the confidentiality and integrity of FL models, helping to mitigate risks associated with bias and ensure fairness in the training and deployment of models[119].

4.2.5. Hybrid Privacy Preserving Federated Learning, Hybrid PPFL

Hybrid PPFL method in FL aims to protect data more effectively by utilizing the strengths of each privacy technology and complementing its weaknesses. For example, combining encryption and DP can improve data security and usability. It can complement vulnerabilities that may occur when using only one privacy protection mechanism and performance degradation issues by combining them in various ways, and can adjust protection techniques to suit various environments and requirements.

4.2.5.1. Model Aggregation

In the area of ​​FFL, model aggregation has been studied to mitigate bias and fairness issues, especially focusing on fair and privacy-preserving methods. In model aggregation, privacy issues are addressed while utilizing techniques such as SMPC for effective model aggregation. In FFL, additional weights or adjustments can be made to ensure fairness of each participant during the model aggregation process.

Guo et al. used encryption and hybrid models to enhance both privacy and fairness in FL systems[119]. They utilize confidential computing mode and secure encryption to isolate data and model learning processes from potential interference, thereby maintaining the fairness of model results. Truex et al. presented a hybrid model that combines various privacy mechanisms to ensure both data protection and fairness in model aggregation[120].

4.2.5.2. Sensitivity-based Weight Distribution

Sensitivity-based weight distribution is a hybrid approach that can help optimize model learning efficiency while maintaining data privacy. It ensures that all data sources contribute fairly to model learning, controls sensitive data from excessively influencing the model, and distributes weights according to importance to optimize learning efficiency.

Chatzikonstantinou et al. proposed a balanced and fair model by implementing a weight distribution technique that can effectively manage the difference in data contributions from various clients[121]. Dilley et al. modified the FedAvg method to introduce fairness constraints to ensure fair model training on diverse client datasets[122]. Chen et al. maintained local consistency while ensuring global convergence, reduced bias, and improved fairness in diverse data domains[123]. Zhao et al. effectively addressed bias by compensating for uneven data quality and distribution of clients[88].

4.3. Methods for solving heterogeneity

Methods for resolving heterogeneity in FFL are similar in basic structure to basic FL methods, but provide a differentiated approach, especially in terms of enhancing fairness and efficiency. These methods focus on asynchronous communication, sampling, fault tolerance mechanisms, and model heterogeneity. These methods effectively manage the diversity and heterogeneity of data through FFL and contribute to creating an environment in which all participants are treated fairly.

4.3.1. Asynchronous Communication

In the existing FL environment, not all clients may have the same network status or computational power. In FFL, asynchronous communication is further advanced to determine the priority of each client, ensuring more fair resource distribution and participation opportunities, and especially improving efficiency in environments where communication costs are high.

Chen et al. proposed an asynchronous online federated learning ASO-Fed framework in which edge devices perform online learning with continuous streaming local data and a central server aggregates model parameters of clients[124]. Baccarelli et al. proposed a novel asynchronous fair adaptation FL framework for stream-oriented IoT application environments, featuring time-varying operating conditions, heterogeneous resource constraints, non-IID local training data, and unreliable communication links[100]. Chen et al. proposed a novel easy-to-implement FL algorithm that controls the mismatch between the global model and the localized model using asynchronous settings and strategies, and adjusts the number of local epochs by estimating the aging, accelerating convergence, and preventing performance degradation due to stragglers[125]. Wang et al. proposed a novel asynchronous FL framework that adapts well to the heterogeneity of users, communication environments, and learning problems by considering both the potential delay in learning and uploading local models and the resulting staleness among receiving models that significantly impacts global model convergence[124].

4.3.2. Sampling

In FFL, sampling includes data imbalance resolution, stratified sampling, clustering-based sampling, and importance sampling. In FFL, sampling is adjusted to consider the data contribution and fairness of each client. It can prevent data bias or exclusion of minority opinions, and generate a more balanced model that reflects the opinions of all participants.

DivFL by Balakrishman et al. provides a thorough analysis of convergence in heterogeneous settings and shows several advantages of the approach by applying it to synthetic and real datasets, including improved learning efficiency, faster convergence, and more uniform performance across clients[106]. DELTA by Wang et al. characterizes the effects of client diversity and local dispersion and samples representative clients as information for global model updating[107]. Luo et al. designed an adaptive client sampling algorithm that addresses both systemic and statistical heterogeneity to minimize wall-clock convergence time[126].

4.3.3. Fault Tolerance

In FL, fault tolerance is used as distributed fault tolerance, redundancy techniques, and strong aggregation methods. Distributed fault tolerance creates a system that does not rely on a central server for all computations, reducing the risk of a single point of failure. Redundancy techniques implement redundant computation or data storage across nodes, allowing the system to recover from individual node failures without losing important information. The link between FL and fault tolerance is Redundancy Techniques, Checkpointing and Rollback mechanisms, and adaptive re-scheduling.

4.3.4. Model Heterogeneity

FFL provides a customized model for each client's data and computational capabilities by considering model heterogeneity. This allows all clients to optimally utilize their data to contribute to learning, and improves the fairness and performance of the overall model.

5. Fairness evaluation metrics

For the long-term viability of a fairness-aware FL system, it is critical to establish a set of performance evaluation measures. This allows us to comparatively evaluate the benefits of various proposed approaches. In general, various indicators are used in FL systems that recognize various fairness and are suitable for various scenarios and applications.

5.1. Evaluations of fairness

Fairness evaluation measures such as Average Variance, Pearson's Correlation Coefficient, and various distance measures are also used to evaluate the fairness of federated systems. If the variance of a particular system is less than the variance of another system, that system is considered fairer than the other system. Table 4 shows metrics and methods for evaluating fairness.

5.2. FFL model

Table 5 shows the state-of-the-art FFL model. It shows the latest FFL model in four categories: data-level, model level, server level, and future direction.

6. FFL Application in Wireless Communication

6.1. Improving Usability

FFL extends the application areas of traditional FL, especially providing an approach that enhances data fairness and privacy. Applications of FFL include usability improvements and unique use cases in wireless communications. First, in the application, FFL can improve the usability of mobile apps by learning user behavior patterns while protecting user privacy through mobile device optimization[202]. Second, in the service, it can be used in financial fraud detection, where banks and financial institutions can jointly train fraud detection models without sharing their data[203]. Third, it can be used in intelligent medical diagnosis systems, where patient data collected from various hospitals and medical institutions can be used to diagnose diseases and develop treatments[204]. By using FFL, each institution can jointly develop more accurate diagnosis models while maintaining the confidentiality of patient data. By integrating various data sources including patients’ genetic information, lifestyle habits, and health conditions, it can suggest treatment plans optimized for individuals.

6.2. Wireless Communication

FFL is a technology that can efficiently train machine learning models across various network devices while maintaining data privacy and security in wireless communications. It mainly focuses on optimizing model learning from distributed data sources and minimizing network latency. In intelligent transportation systems, data collected from vehicles is used to analyze real-time traffic conditions and optimize traffic flow. In smart grid management, smart meter data is used to adjust energy demand and supply in real time. Through FFL, each smart meter processes data individually, and a central aggregator integrates them to optimize energy distribution. FFL will further expand with the advancement of wireless communication technology in the future. In particular, the increase in IoT (Internet of Things) devices and the spread of 5G networks are expected to accelerate the development of this technology. First, personalized healthcare services will provide customized healthcare services by utilizing individual patient health data. Through FFL, data can be processed locally on each patient's device and integrated to suggest personalized health advice and treatment methods. Second, in disaster response systems, data collected from various sensors and cameras can be used to build a system that can respond quickly to natural disasters. FFL processes data from each sensor locally and provides real-time information for disaster response through central aggregation.

7. FFL Challenge and Future work

7.1. Privacy Preservation

FFL further enhances privacy protection, and includes mechanisms to ensure data fairness and protect data from overfitting in minority groups. Various research and development are underway as approaches to enhance privacy protection while improving model fairness. The future direction of FFL includes innovations in technical, ethical, and legal aspects. In technical innovation, advanced encryption techniques such as HE and SMPC are currently used in FFL systems. In the future, these technologies will further develop to improve processing speed and to efficiently process more complex data types and large-scale data. Here are some approaches for this: Extended HE: Currently, homomorphic encryption has limitations in computation, but it is necessary to expand the ability to process data in an encrypted state while reducing computational costs through more advanced forms. Utilization of large-scale language models such as GPT: We seek ways to derive personalized results through pre-trained models without directly processing user data. Establishment of international data privacy standards: We can improve the level of privacy protection in FFL by establishing privacy protection standards that can be commonly recognized by various countries and companies. Strengthening transparency and accountability: Provide transparency into the decision-making process of the FFL system and clarify the criteria for holding those responsible when privacy violations occur.

7.2. Communication Cost

In order to improve communication efficiency, FFL uses an optimized algorithm to minimize the amount of data transmitted by each participant and reduce communication costs. In particular, in large-scale distributed networks, where numerous participants share data with a central server, the high communication costs incurred in the process can reduce the efficiency of the entire system. Here are some approaches for this: Model Compression and Quantization: The size of the transmitted data can be reduced by compressing or quantizing model parameters. It reduces the amount of data required for model updates, thereby reducing the communication load. Sparse Model Training: Eliminate unimportant parameters of the model or minimize updates, and only exchange practically important parameters periodically. Asynchronous FL: Participants do not perform updates simultaneously, but communicate with the server asynchronously according to their own schedules, which helps reduce network latency and communication burden. Staleness-aware Aggregation: It reduces the need for communication by effectively utilizing even old updates, and finds the optimal balance between model freshness and efficiency. Edge Computing: It is effective in processing data close to the data generation point and exchanging only essential information with the central server, thereby reducing the amount of data to be transmitted and minimizing delay time. Network Topology Optimization: By optimizing the network topology, efficient data routing and resource usage can be promoted, and by shortening the data transmission path within the network, communication costs can be reduced. Resource Allocation: By efficiently allocating communicable resources and scheduling that takes into account the communication time of participants, network congestion and communication costs can be managed.

7.3. Systems Heterogeneity

In order for participants with various hardware, software, and network environments to efficiently train models and derive fair results, it is necessary to solve the heterogeneity problem. Here are some approaches for this. Model Compression: The size and complexity of the model are adjusted according to the computing ability of the participant, and each participant can train using a model suitable for their hardware. Layer-wise Adaptation: Only some layers of the model are learned locally by the participant, and important layers are learned centrally to manage heterogeneity. Asynchronous updates: Updates are performed asynchronously according to the network speed or availability of the participant, and each participant can transmit data at their own speed, which can improve the efficiency of the entire network. Bandwidth-aware Scheduling: The data transmission schedule can be optimized by considering the bandwidth of the participant, minimizing network congestion and reducing data transmission costs. Resource-aware Task Allocation: The learning task is allocated according to the computing ability of each participant, and all participants perform tasks suitable for their devices, which contributes to improving the overall learning efficiency. On-device Machine Learning: Use lightweight machine learning models to perform learning directly on each participant’s device, enabling learning even on devices with limited computing resources. Cross-platform Frameworks: Use cross-platform frameworks such as Lite or PyTorch Mobile to effectively run models on a variety of operating systems and hardware. Adaptive Code Deployment Develop tools that automatically optimize and deploy code to suit the characteristics of the platform.

7.4. Unreliable Model Upload

The problem of unstable model upload in FFL can reduce the learning efficiency and model accuracy of the entire system because participants can upload inaccurate or corrupted model updates to the central server. Here are some approaches: Pre-upload validation: Before participants upload models, a rigorous validation process is performed locally to evaluate the accuracy, consistency, and data quality of the model. Server-side validation: After receiving the uploaded model from the central server, additional validation is performed to check whether the model meets the set criteria. Quality-based incentives: Motivate participants to pay more attention to data quality and accuracy by providing greater rewards for participants who provide high-quality model updates. Penalty mechanism: Introduce penalties for participants who repeatedly submit inaccurate or corrupted updates to encourage them to provide only reliable data. Anomaly detection algorithm: Develop a machine learning-based algorithm to automatically detect outliers in uploaded model updates, so that the reliability of updates can be evaluated in an automated manner. Robust integration: When integrating updates from multiple participants, a robust integration mechanism that can tolerate outliers or errors is designed to enhance the stability of the overall model and minimize the impact of one or two inaccurate updates on the overall model. Each strategy can work complementarily to improve the overall reliability and efficiency of the system.

7.5. Multi-Center FL (MCFL)

MCFL is a method in which multiple data centers collaborate to build a common machine learning model. It is particularly difficult to secure communication, data homogeneity, and balanced data contributions between data centers. Therefore, the following approaches can be considered to effectively develop MCFL. Improved synchronization mechanism: Develop a more efficient data streaming and version management system for synchronization between data centers, and ensure that all participating centers have access to the latest data and model updates. Fair data contribution evaluation: Introduce a mechanism to fairly distribute rewards by evaluating the contribution of each data center, motivating all participating centers to contribute to the model, and improving the quality and quantity of data. Model Customization: Adopt a model structure that suits the characteristics of each data center, so that the characteristics of local data can be reflected as much as possible, which can improve the efficiency and accuracy of the overall model. Continuous model optimization: Continuously update and optimize the model through continuous feedback and performance monitoring, helping the model to quickly adapt to changing data conditions.

7.6. MMFFL Research Direction and Tasks

Data from different modalities are difficult to synchronize and can affect model learning, so it is necessary to develop a mechanism to effectively integrate features extracted from each modality. MMFFL can consider the following methods to maintain privacy while dealing with data imbalance and bias issues. Segmentation and appropriate feature extraction: Apply a feature extraction method optimized for each modality to extract the most useful information from each data type. Diversity promotion: Include data from participants with diverse backgrounds in each modality to increase the representativeness of the learning data, and ultimately promote fair model learning. Therefore, examples of future directions for MMFFL are as follows. Diagnostic systems in the medical field develop systems that provide more accurate and personalized diagnoses by integrating various forms of medical data (e.g., patient medical history information, X-ray images, and voice records). In this process, fairness of each data source is secured to provide equal treatment opportunities to all patients. Driver assistance systems in the automotive field integrate in-vehicle cameras, sensor data, and driver voice commands to learn driver behavior patterns and provide optimal driving support. This system ensures the same level of safety and convenience for all drivers by considering the fairness of each data source. It analyzes the user's text input, image upload, and interaction data through multimodal federated learning through the recommendation system of social media, and recommends personalized content. In this process, it ensures fair processing of user data, and aims to recommend unbiased content to all users. This direction will promote the development of multimodal federated learning and enable the transition to a more fair and inclusive technology. MMFFL offers the potential to develop more accurate and fair artificial intelligence models by utilizing rich and diverse data sources. However, multimodal research is still in its early stages, and future research and technological developments will further expand the possibilities of this approach.

7.7. Reliable Client Selection

FFL focuses on using only high-quality data by more thoroughly evaluating the reliability of participants, and significantly improving the fairness of data, privacy protection, and system reliability compared to existing FL. Selecting reliable clients is important for maintaining the quality of data and the accuracy and fairness of the model within the federated learning network. To solve the problems caused by the asymmetric data distribution and diverse computational capabilities among clients, the following approaches can be included. Behavior-based evaluation: Reliable clients can be identified by evaluating the client's previous participation history and the consistency of data provision. Introducing a point system: Motivate clients by awarding points whenever they continuously provide reliable data, and providing additional benefits when a certain number of points are accumulated. Feedback loop: Provide feedback on the client's data contribution so that the client source can improve the quality of data.

8. Conclusion

As the digital age exponentially increases the volume and variety of data, the need for ethical data utilization and strong privacy protection measures is growing. Fair Federated Learning (FFL) has emerged as an instrumental paradigm to address these rapidly growing needs. FFL has evolved to develop fair models that not only mitigate the risk of marginalizing individual participants or minority groups by integrating fairness into the core of its architecture, but also ensure fair benefits for all stakeholders. It also effectively addresses bias that can occur when relying on a single data source and prevents overfitting by learning from diverse data sources through bias reduction. Through this fairness enhancement and bias reduction, FFL enhances privacy by ensuring that important personal information is not transmitted over the network through local processing of data. This shift from traditional Federated Learning (FL) practices is toward correcting inherent bias and strengthening model integrity through a more comprehensive and balanced data representation.

Additionally, FFL excels at effectively deterring potential malicious interventions and data manipulation by strengthening the trustworthiness of shared models within federated networks. It advocates transparent and fair data contribution evaluation that resonates with ethical data handling practices, paving the way for safe and reliable model sharing. The evolution from FL to FFL exemplifies a holistic approach that reconciles the twin concerns of fairness and privacy, moving beyond simply optimizing model performance to embracing a governance framework that prioritizes equity and confidentiality. In essence, FFL not only advances FL methodologies, but also more closely aligns with the stringent regulatory and ethical standards that govern modern machine learning applications. This paper details the strategic implementation and challenges faced in realizing FFL, and highlights how this innovative approach can dramatically change the landscape of data-driven decision making by ensuring fairness and protecting participant privacy across a range of sectors.