Cooperative D2D Partial Training for Wireless Federated Learning

Abstract

Federated learning (FL) is a promising distributed machine learning paradigm to train a machine learning model without the leakage of local data. However, as the sizes of models are increasing while Internet of Things (IoT) devices are heterogeneous and capability-limited, FL faces performance bottleneck. In this article, we propose a cooperative device-to-device (D2D)-based partial training scheme for wireless FL. Specifically, the IoT devices in each cluster extract and train the nonoverlapping submodels from the global model, and the trained submodels are transmitted to the cluster head (CH) to form a whole local model via D2D links. Then the CHs upload the local models to the server for global aggregation. We first conduct the convergence analysis for the proposed wireless FL scheme. Then a joint optimization problem is formulated to minimize the average delay by the optimization of model division, device selection, and bandwidth allocation. An efficient algorithm is proposed to solve this nonconvex problem. Comprehensive experiments verify the efficiency of the proposed scheme.