FedDCS: Dynamical Client Selection for Federated Learning in Mobile Scenarios with Label Distribution Shift

Federated learning (FL) is a distributed learning paradigm that enables clients to cooperatively train a global model without private data sharing. However, the label distribution of client local data usually deviates from the FL task requirements in mobile scenario due to the heterogeneity and dynamism of environment, termed label distribution shift, which will hamper FL convergence and decrease the performance of global model. In this paper, we propose a method to estimate the relevance of client label distribution to the FL task, leveraging the insight that label distribution divergence can be bounded by the similarity of corresponding models. Subsequently, we present FedDCS, a novel FL framework with an adaptive client selection scheme that dynamically select highly relevant clients to participate in FL to handle the label distribution shift of clients. We conduct complexity analysis to validate the efficiency of our method. To validate its effectiveness, we simulated various scenarios of client label distribution shift and conducted experiments on multiple real-world datasets. The extensive experiments show that FedDCS can significantly accelerate convergence of FL and achieve the best performance of global model than baselines.