CLIENT SELECTION IN FEDERATED LEARNING

Abstract

In order to improve the effectiveness and efficiency of federated learning systems, I present Adaptive Weighted Client Selection for Federated Learning (AW-FedSelect), a novel client selection technique. AW-FedSelect weighs each client’s contribution according to local model quality and data distribution similarity, combining gradient-based selection and hierarchical clustering. We compare the performance of AW-FedSelect against traditional client selection techniques, such as random selection, power-law-based selection (POWD), clustering-based selection, and divergence-based federated learning (DivFL), by utilizing the FedAvg algorithm and a Convolutional Neural Network (CNN) model. We show through thorough experiments that AW-FedSelect outperforms other methods in terms of convergence speed, communication efficiency, and model accuracy. Our research highlights how AW-FedSelect may optimize client selection in federated learning systems, with promising results in a variety of fields.