Federated Learning with Local Fairness Constraints

In this paper, we study training a classifier satisfying local fairness constraints via federated learning. We compare two schemes: local fair training (LFT) combined with the ensemble method (LFT+Ensemble), and LFT combined with federated averaging (LFT+FedAvg). Our theoretical analysis shows that (i) LFT+FedAvg outperforms LFT+Ensemble in terms of fairness at the cost of frequent communication, and (ii) LFT+FedAvg may not reach the optimal fairness achievable through centralized training. The findings explain the success of recently proposed federated learning algorithms that virtually compute global fairness constraints with additional communication rounds. Moreover, we also present numerical experiments that support our findings in more general settings.