Accelerating and Securing Federated Learning via Intelligent Reflecting Surface

Abstract

Federated Learning (FL) provides a privacy-protected way to train Machine Learning (ML) models, where multiple edge devices upload the model parameters to the central server without sharing their local data. However, the stochastic nature of wireless channels and frequent communications involved in FL incur high latency while model parameters can be intercepted by eavesdroppers, which leads to privacy leakage. To address the above issues, we exploit the Intelligent Reflecting Surface (IRS) to reconfigure the wireless signal propagation, enabling secure transmission and fast convergence. In this paper, We aim to maximize the minimum secrecy rate at devices via jointly designing transmission power at devices and IRS phase shift. A Deep Reinforcement Learning (DRL) based algorithm is adopted to derive the optimal solution. Transmission power and IRS phase shift are treated as the action elements. These elements are dynamically adjusted through interactions between the agent and the environment, guided by a pre-defined optimal reward. Both the transmission power and IRS phase shift are obtained as outputs from the DRL neural network. Numerical results validate the efficiency of our proposed algorithm and demonstrate that the deployment of IRS can improve the transmission rate and secure the training of FL.