Joint Mode Selection and Resource Allocation for D2D-Assisted Wireless Federated Learning

Abstract

Straggling link is a well-known bottleneck in wireless federated learning (FL), which would cause a significant decrease on the learning performance and increase the learning latency. Distinguishing from existing approaches, a device-to-device (D2D)-assisted wireless FL framework is proposed in this letter to address this challenge. The stragglers can successfully upload the local models to base station (BS) via neighbors in the D2D network. Moreover, to further improve the learning efficiency, an optimization problem is formulated to minimize the learning latency per iteration. To effectively solve this problem, three sub-problems are decomposed and a joint mode selection and resource allocation algorithm is developed to achieve the approximate optimal solutions. In the end, the effectiveness of the proposed algorithm is demonstrated by comprehensive experiments. Compared against the baselines, our proposal can obtain the better learning performance and lower learning latency.