Mitigating Demographic Bias of Federated Learning Models via Global Domain Smoothing

Huimin Zeng, Zhenrui Yue, Yang Zhang, Lanyu Shang, Ruohan Zong, Dong Wang

Keywords: Federated Learning, Demographic Fairness

Abstract:

Federated learning (FL) has shown impressive performance in training modern machine learning models from distributed data sources. However, the distributed training process of FL could suffer from a non-trivial bias issue, where the trained models are affected by the imbalanced distribution of the training data on local clients, and eventually lead to a severe bias of the aggregated global model. In
this paper, we propose a novel fairness-aware FL training framework Worst-Fair Domain Smoothing (WFDS) to address the bias issue of FL models from a domain-shifting perspective. Our framework consists of two concurrent components: 1) local worst-fair training, and 2) reference domain smoothing. The first module is designed to train fair local models and enforces the robustness of local fairness against domain shifts from local distribution to global distribution by performing worst-fair training. The second module simulates a reference data domain of the studied FL application for all clients, and implicitly reduces the domain discrepancy of training data among different clients. With reduced domain discrepancy, the fairness of each local model will be learned from similar training distributions despite on different clients. As such, improved global fairness can be achieved after aggregating the local models into the global model. Evaluation results on
multiple real-world datasets show that WFDS can achieve significant performance gains in demographic fairness compared to state-of-the-art baselines.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)