Energy-Efficient Federated Learning With Intelligent Reflecting Surface

Abstract

Federated learning is a new paradigm to support resource-intensive and privacy-aware learning applications. It enables the Internet-of-Things (IoT) devices to collaboratively train a global model to accomplish a machine learning task without sharing private data. In practice, the IoT devices powered by batteries finish the local training and interact with the central server via wireless links. However, the repeated interaction between IoT devices and the central server would consume considerable resources. Motivated by the emerging technology of intelligent reflecting surface (IRS), we propose to leverage the IRS to reconfigure the wireless propagation environment to maximize the utilization of the available resources. Specifically, we consider the critical energy efficiency issue in the reconfigurable wireless communication network. We formulate an energy consumption minimization problem in an IRS-assisted federated learning system subject to the completion training time constraint. An iterative resource allocation algorithm is proposed to jointly configure the parameters with proven fast convergence. Simulation results validate that the proposed algorithm converges fast and can achieve significant energy savings, especially when the number of reflecting elements is large and when the IRS is properly configured.