FedEquilibria: Towards Fair and Robust Federated Learning under Domain Skew

Hongjie Jia, Xing Li, Di Wu, Huan Wang, Jun Bai, Longxia Huang, Wei Xiang

Keywords: Federated Learning, Fairness, Domain Skew

Abstract: Federated Learning (FL) has been widely emerged as a promising paradigm for decentralized machine learning. However, its practical effectiveness is hindered by domain skew, a prevalent form of data heterogeneity where clients hold statistically different data distributions. Existing studies have revealed that this failure is rooted in two fundamental problems: (1) update conflicts, arising when clients’ learning objectives diverge, and (2) model aggregation bias, where conventional aggregation schemes neglect domain diversity, systematically favoring certain clients over others. To tackle these intertwined challenges, we propose FedEquilibria, a novel and fair federated aggregation framework. The core idea of FedEquilibria is to employ a server-side hybrid weighting mechanism consisting of two synergistic steps. First, it formulates aggregation as a multi-objective optimization problem, uniquely leveraging the Fisher Information Matrix (FIM) as a proxy for each client’s empirical objective. By computing conflict-aware weights on the server-side, FedEquilibria identifies a Pareto-optimal consensus of parameters that are structurally important across different domains. Second, to counteract aggregation bias, FedEquilibria calculates drift-aware weights based on the Euclidean norm of client updates, explicitly quantifying the fitting gap for each client and adaptively increasing the influence of underrepresented domains. Comprehensive experiments on benchmark datasets with pronounced domain shifts demonstrate that FedEquilibria significantly surpasses existing state-of-the-art methods. It achieves not only higher average accuracy but also substantially enhances fairness by improving the performance of the worst-performing clients, offering a principled solution for robust and equitable models in real-world FL systems.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 6956