Asynchronous Hierarchical Federated Learning Based on Bandwidth Allocation and Client Scheduling

Abstract

Federated learning (FL) offers a promising solution in edge computing to overcome bandwidth limitations and privacy concerns associated with traditional cloud-based training. However, current FL methods often suffer from transmission delay and excessive communication resource usage. In this paper, we introduce an innovative asynchronous hierarchical FL approach based on bandwidth allocation and client scheduling. Specifically, we propose an efficient algorithm that dynamically assigns clients to edge servers based on client mobility during training and accelerates parameter uploading while taking into account the remaining bandwidth of the edge servers. Our experimental results demonstrate the effectiveness of our approach, particularly in scenarios with frequent client mobility. This research strongly supports the application of FL in edge computing and underscores the crucial role of resource allocation in addressing communication resource constraints and reducing the training time of FL.