Accelerating Federated Learning Under Client Dropout via Joint Bandwidth Allocation and Prototype Fine-Tuning in Mobile Edge Computing Networks

Abstract

Deploying federated learning (FL) in mobile edge computing (MEC) networks enables collaborative model training while preserving the privacy of raw data. However, due to system heterogeneity, statistical heterogeneity of mobile clients (MCs), and client dropout, optimizing bandwidth allocation and fine-tuning the global model is essential for accelerating FL. We formulate a problem to reduce the sum of waiting time of the MCs and edge server (ES), which can indirectly accelerate the FL process, and decompose it into two subproblems. We then propose a joint bandwidth allocation and prototype fine-tuning (BAPFT) framework to address them. In each round, BAPFT dynamically allocates bandwidth to the participating MCs, determines a deadline, thereby reducing the waiting time of the MCs and ES caused by system heterogeneity under client dropout. Moreover, BAPFT adopts mixup prototypes maintained at the ES to fine-tune the global model, which mitigates the negative impact of missing updates from the dropped MCs and statistical heterogeneity. Extensive experiments on three datasets demonstrate that BAPFT significantly reduces model training latency by an average of 16.78% and rounds by an average of 5.72% compared to baseline frameworks.