Impact of Adaptive Client Selection on Federated Learning for IoMT Ecosystem

This paper investigates the impact of two adaptive client selection mechanisms commonly employed in federated learning (FL) for dynamic networks, such as the Internet of Medical Things (IoMT), to enhance global model performance. Specifically, the study focuses on a performance-based client selection algorithm and a deep reinforcement learning (DRL)-based client selection (CS) algorithm. The performance-based approach utilizes a piecewise function to rank clients based on historical performance. In contrast, the DRL-based approach employs a neural network to approximate Q-values during the selection process. Simulation experiments using a medical security dataset demonstrate the superiority of the performance-based CS approach, accelerating global model convergence by 50% compared to the baseline FedAVG and 25% compared to the DRL-based CS approach.