Fairness-aware Federated Learning

Zhuozhuo Tu, zhiqiang xu, Tairan Huang, Dacheng Tao, Ping Li

Keywords: Learning Theory

Abstract: Federated Learning is a machine learning technique where a network of clients collaborates with a server to learn a centralized model while keeping data localized. In such a setting, naively minimizing an aggregate loss may introduce bias and disadvantage model performance on certain clients. To address this issue, we propose a new federated learning framework called FAFL in which the goal is to minimize the worst-case weighted client losses over an uncertainty set. By deriving a variational representation, we show that this framework is a fairness-aware objective and can be easily optimized by solving a joint minimization problem over the model parameters and a dual variable. We then propose an optimization algorithm to solve FAFL which can be efficiently implemented in a federated setting and provide convergence guarantees. We further prove generalization bounds for learning with this objective. Experiments on real-world datasets demonstrate the effectiveness of our framework in achieving both accuracy and fairness.

One-sentence Summary: We propose a new framework to address the fairness issues in federated learning and provide theoretical guarantees.