GFS: Gradient-Based Fairness-Aware Client Selection for Federated Learning
Zijian Wang, Danyang Xiao, Diying Yang, Weigang Wu · 2024
Statistical heterogeneity is one of the main causes of unfairness in federated learning (FL), which will make clients reluctant to participate in federated training. Recently, many client selection strategies have been proposed to handle client-wise data heterogeneity. However, most of these strategies neglect the cost of sampling data and the computing overhead of clients. In this paper, we study how to ensure fairness in scenarios that consider these costs. Firstly, we quantify the benefits for each client based on their cost. Then we define fairness by the Gini coefficient of the total benefits of clients. By our definition of fairness, we propose a gradient-based client selection strategy (GFS). GFS uses the gradients of clients to estimate the impact of selected clients on each client's benefit and fairness. GFS selects the subset of clients that maximizes a mixed function of average social welfare and the Gini coefficient to participate in federated training. Our experimental results show that compared to the baselines, GFS can improve the benefits of clients and ensure fairness better on FMNIST and CIFAR-10. Our code is available at https://github.com/SelectionStrategy/GFS_Strategy.