Low Latency Federated Learning over Wireless Edge Networks via Efficient Bandwidth Allocation

Abstract

Federated learning has emerged as a solution to preserve data privacy while processing a large amount of data in parallel at the network edge, thus ensuring low latency and high data security. The involvement of a massive amount of edge devices in federated learning has raised various challenges due to device heterogeneity in computation and communication and limited resources. To reduce the latency of a resource-constrained network, optimal bandwidth sharing among edge devices participating in federated learning is necessary. This paper proposes an efficient bandwidth allocation method to reduce overall latency in federated learning rounds over resource-constrained edge networks. First, the optimization problem is formulated as the minimization of total latency in one round of federated learning. Then a fast converging iterative bandwidth allocation (FCI-BA) method is proposed to allocate the available bandwidth among participating edge devices efficiently. Based on experimental results, the proposed framework achieves optimal solutions regardless of the initial guess. Further experiments demonstrate the effectiveness of the approach for a large number of devices, and it is observed that the performance of the FCI-BA method is independent of the increase in the number of devices in the network. Therefore, the proposed method can be effectively applied to resource-constrained massive networks.