FedBEF : Federated Learning With Balance of Performance and Fairness

What the paper says

Federated learning (FL) faces significant challenges under Non‐IID data, primarily due to misaligned local gradients that point in conflicting directions across clients. This inconsistency creates a fundamental trade‐off between global model accuracy and fairness, as clients with large‐magnitude gradients often dominate the aggregation process, marginalising minority participants. Existing methods typically fail to explicitly resolve these directional conflicts, leading to suboptimal convergence or exacerbated bias. To address this, we propose FedBEF, a theoretically grounded aggregation framework that jointly optimises convergence speed and fairness. Unlike heuristic weighting schemes, FedBEF derives optimal client weights by minimising an upper bound of the global loss, formulated as a constrained quadratic program. The resulting weights explicitly penalise gradient conflicts (e.g., by discouraging large angular deviations) while promoting a coherent descent direction that aligns with the true global gradient. Moreover, to improve robustness against noisy or outlier updates—particularly pronounced under partial client participation and extreme heterogeneity, we introduce a Similar Neighbour Gradient (SNG) mechanism combined with adaptive momentum. By clustering clients based on gradient cosine similarity and smoothing within neighbourhoods, SNG effectively suppresses erratic updates without requiring additional communication. More importantly, we prove that FedBEF converges to a stationary point under non‐convex settings. Extensive experiments on CIFAR‐10 and CIFAR‐100 demonstrate its superiority over state‐of‐the‐art baselines such as FedProx and FedALA. Notably, under severe heterogeneity (CIFAR‐10, Dirichlet ), FedBEF achieves up to a 7.13% higher average accuracy. It also significantly enhances fairness, as evidenced by a substantial reduction in the coefficient of variation (standard deviation over mean) of per‐client accuracies.