FairFL: A Fair Federated Learning Approach to Reducing Demographic Bias in Privacy-Sensitive Classification Models

The recent advance of the federated learning (FL) has brought new opportunities for privacy-aware distributed machine learning (ML) applications to train a powerful ML model without accessing the private training data of the participants. In this paper, we focus on addressing a novel fair classification problem in FL where the model trained by FL displays discriminatory bias towards particular demographic groups. Addressing the fairness issue in a FL framework posts three critical challenges: fairness and performance trade-offs, restricted information, and constrained coordination. To address these challenges, we develop FairFL, a fair federated learning framework dedicated to reducing the bias in privacy-sensitive ML applications. It consists of a principled deep multi-agent reinforcement learning framework and a secure information aggregation protocol that optimizes both the accuracy and the fairness of the learned model while respecting the strict privacy constraints of the clients. Evaluation results on real-world applications showed that FairFL can achieve significant performance gains in both fairness and accuracy of the learned model compared to state-of-the-art baselines.