SFFL: Self-aware fairness federated learning framework for heterogeneous data distributions

Federated Learning (FL) has been proven to show biased predictions against certain demographic groups, such as sex or race. Recent advances in improving fairness in federated learning come with the price of leaking sensitive information and compromising the model performance, especially under heterogeneous data distributions, which is amore practical situation for federated learning. Therefore, how to ensure the clients obtain a fair model while preserving their information privacy is anew challenge. To address it, we propose SFFL, a self-aware fairness federated learning framework that jointly improves fairness and performance under heterogeneous data distributions without the requirement for clients' sensitive information. Specifically, SFFL first introduces a fair local training algorithm, FairEM, which decomposes the clients' training objectives into fair training objects based on underlying distributions to improve local fairness and model performance. Compared to existing methods, FairEM alleviates the decrease in fairness and performance caused by inconsistent update objectives. Moreover, we further propose a self-aware aggregation method to mitigate bias propagation during aggregation, which leverages a distance-based reweighting strategy to update the aggregation weights for each client. This method discards the requirement for clients' sensitive information and maintains high effectiveness. Extensive evaluation results demonstrate that our proposed framework can significantly improve the performance decrease under heterogeneous data distributions and enhance the privacy of clients. Numerically, our proposed SFFL improves the fairness for 0.3x to 2.1x in fairness and 2.36% to 9.87% in performance compared with the existing four SOTA works.