Mitigating Demographic Bias of Federated Learning Models via Robust-Fair Domain Smoothing: A Domain-Shifting Approach

Federated learning (FL) emerges as a promising solution to train machine learning (ML) models from distributed data sources. In FL, the heterogeneous and imbalanced data distribution of local clients could severely hurt the fairness of the aggregated global model. In this paper, we identify two key obstacles of developing fair FL models w.r.t. the global distribution: the domain shifts from local clients to the global data distribution and the fairness heterogeneity across local clients. Therefore, considering these two obstacles, we present a novel fairness-aware FL training framework Robust-Fair Domain Smoothing (RFDS) to address the bias issue of FL models from a unique domain-shifting perspective. In particular, we design two novel components to build RFDS: 1) local robust-fair training, and 2) reference domain smoothing. Local robust-fair training aims to train robust-fair local models whose fairness is robust against the domain shifts from local distributions to the global distribution. Reference domain smoothing reduces the heterogeneity of fairness across clients to improve the fairness of the aggregated global model. We further provide a theoretical analysis to show the connection between the domain discrepancy of local data distributions and the heterogeneity of fairness across clients. Empirical evaluation results on multiple real-world datasets show that RFDS achieves promising performance gains in improving demographic fairness compared to state-of-the-art baselines.