Sffl: Self-Aware Fairness Federated Learning Framework for Heterogeneous Data Distributions

Zhang, Jiale; Li, Ye; Wu, Di; Zhao, Yanchao; Palaiahnakote, Shivakumara

Abstract

Recent years have witnessed increasing development towards federated learning. However, federated learning has been proven to show biased predictions against certain demographic groups, such as sex or race, especially under heterogeneous data distributions. Training fair federated models under heterogeneous data distributions face the challenge of inherent unfair local training and bias propagation during aggregation and mismatch between local fairness and global fairness. Current fairness approaches for federated learning are struggling to balance fairness and privacy. More importantly, they neglect that the differences in update objectives between heterogeneous clients lead to difficulties in maintaining fair classification and learning among clients. To address these limitations, we propose a self-aware, fair federated learning framework, SFFL, which jointly improves fairness and performance under heterogeneous data distributions without the requirement for clients’ sensitive information. Firstly, we present the FairEM method, which considers the heterogeneous distributions as a combination of multiple underlying distributions and decomposes the clients’ training objects to the fair training objects on underlying distributions to alleviate the fairness and performance decrease caused by inconsistency update objects. Secondly, we introduce a self-aware aggregation method to mitigate the bias propagation across different component models without requiring sensitive statistics. Extensive evaluation results demonstrate the effectiveness of our proposed framework in achieving fairness and maintaining performance in heterogeneous data distributions.

Citation

Zhang, J., Li, Y., Wu, D., Zhao, Y., & Palaiahnakote, S. Sffl: Self-Aware Fairness Federated Learning Framework for Heterogeneous Data Distributions

• Working Paper Type: Working Paper
• Deposit Date: Nov 15, 2024
• Publicly Available Date: Dec 4, 2024
• Related Public URLs: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4885246