Federated Learning at the Forefront of Fairness: A Multifaceted Perspective

Fairness in Federated Learning (FL) is emerging as a critical factor driven by heterogeneous clients’ constraints and balanced model performance across various scenarios. In this survey, we delineate a comprehensive classification of the state-of-the-art fairness-aware approaches from a multifaceted perspective, i.e., model performance-oriented and capability-oriented. Moreover, we provide a framework to categorize and address various fairness concerns and associated technical aspects, examining their effectiveness in balancing equity and performance within FL frameworks. We further examine several significant evaluation metrics leveraged to measure fairness quantitatively. Finally, we explore exciting open research directions and propose prospective solutions that could drive future advancements in this important area, laying a solid foundation for researchers working toward fairness in FL.

💡 Research Summary

Federated Learning (FL) has emerged as a powerful paradigm for training machine‑learning models across a multitude of edge devices while keeping raw data locally, thereby preserving privacy and reducing communication overhead. However, the inherent heterogeneity of participating clients—differences in data distribution, computational power, communication bandwidth, and availability—introduces significant fairness challenges that can undermine both the effectiveness of the global model and the willingness of clients to participate.

This survey provides the first comprehensive taxonomy that classifies fairness‑aware FL approaches along two orthogonal dimensions: model‑performance‑oriented and capability‑oriented methods. The authors first enumerate a rich set of fairness notions that have appeared in the literature, including client‑level fairness (equal participation), group fairness (equal performance across demographic groups), performance‑distribution fairness (uniform accuracy across clients), good‑intent fairness (minimizing worst‑case loss for protected groups), contribution‑based fairness (rewarding clients proportionally to their contribution), regret and expectation fairness (balancing incentive timing).

Model‑performance‑oriented approaches aim to ensure that the global model delivers comparable predictive quality to all clients regardless of data or resource disparities. The survey groups these methods into four categories:

1. Fairness‑as‑Optimization – Techniques such as FedISM and FedLF formulate single‑ or multi‑objective optimization problems that embed fairness constraints (e.g., sharpness‑aware loss, layer‑wise gradient alignment) to prevent any single client from dominating the model update.

2. Fairness‑aware Aggregation – Algorithms like FedHEAL and FairFed modify the server‑side aggregation step by weighting updates based on fairness metrics, discarding harmful updates, or aligning gradients to reduce domain bias.

3. Model Personalization for Fairness – Methods such as ShapFed‑W, DBE, and other personalization frameworks tailor model components to clients with similar data distributions, using tools like Shapley values or bidirectional knowledge transfer to mitigate representation bias.

4. Fairness‑aware Performance Evaluation – Beyond traditional statistical parity and equal‑opportunity measures, the survey highlights recent work (e.g., GLFOP) that decomposes unfairness into unique, redundant, and masked components, and even introduces adversarial scenarios (EAB‑FL) where malicious clients amplify group bias while preserving overall utility.

Capability‑oriented approaches focus on the structural inequities that arise from heterogeneous client resources. The authors discuss fair participant selection mechanisms that consider bandwidth, computation, and data volume when forming the initial client pool and when sampling per‑round participants. They also cover fair resource allocation and contribution scaling strategies that prevent high‑capacity devices from monopolizing the learning process, thereby preserving inclusivity for low‑resource participants.

The survey then surveys fairness evaluation metrics employed in FL research. In addition to classic metrics (Statistical Parity Difference, Equal Opportunity Difference), newer information‑theoretic metrics (Unique/Redundant/Masked Disparity) and multi‑objective trade‑off formulations are described, providing a richer quantitative toolkit for researchers.

Finally, the paper identifies several open challenges and future research directions:

• Unified frameworks that jointly optimize model‑performance fairness and capability fairness are still lacking.
• Dynamic adaptation to time‑varying client conditions (e.g., fluctuating bandwidth or battery) remains an open problem.
• Fairness‑privacy trade‑offs need systematic study, especially under differential privacy constraints.
• Standardized benchmarks covering diverse data modalities, device heterogeneity, and fairness scenarios are necessary for reproducible evaluation.
• Security‑fairness interplay (e.g., model‑poisoning attacks that target specific demographic groups) calls for defenses that simultaneously address robustness and equity.

In summary, this survey systematically maps the current landscape of fairness in federated learning, expands the discussion beyond client selection to include model performance disparities, and outlines a clear agenda for advancing equitable, robust, and scalable FL systems.