Client Selection in Hierarchical Federated Learning
Silvana Trindade, Nelson L. S. da Fonseca
Abstract
Federated Learning is a promising technique for providing distributed learning without clients disclosing their private data. In Hierarchical Federated Learning, edge servers partially aggregate the parameters of their connected clients’ models, improving scalability and reducing computational overhead on the central server. To speed up the convergence of the global model, only those clients with potential contributions to the model performance will participate in model training. This paper introduces a two-step client selection approach for hierarchical federated learning and three novel algorithms, which consider a large set of features in this selection and the client’s contributions to the model performance. Compared to selected baseline algorithms, the proposed client selection algorithms reduce CPU utilization by more than 50%, memory usage by 80%, and energy consumption by 50%.