Balancing privacy and fairness: Client selection in differential privacy-based federated learning

Federated Learning (FL) tackles the data island problem by enabling collaborative model training across clients while safeguarding privacy. However, most existing work only cared about the parameter privacy protection, while ignoring the model training efficay. To this end, in this paper, we propose AdaDPCS-FL, an Adaptive Budget Allocation and Client Selection method tailored for Federated Learning. This approach consists of two steps, where in the step, an adaptive privacy budget allocation strategy based on model similarity and a reversion mechanism are designed to speed up training convergence while still keeping privacy preservation in FL. In the second step, addressing the fairness of client selection in the FL process, it proposes a contribution-based online client selection mechanism is further proposed with the consideration of the fairness of client selection, in which, a multi-armed bandit scheme is tailored to optimize the client selection. Theoretically, the proposed method satisfies the properties of differential privacy, convergence guarantee, and a constant upper bound on cumulative regret $O\left(K\sqrt{lnRlnR}\right)$. Experiments on real datasets demonstrate superior performance over baselines like FedProx and FedAvg. Moreover, with the privacy guaranteed, the test accuracy by our proposed method can be improved approximately 4% compared to DP-FedAvg and FedBDP in heterogeneous settings.