Computation–Communication Resource Allocation for Federated Learning System with Intelligent Reflecting Surfaces

Federal learning (FL) is a distributed learning algorithm, which enables user devices (UDs) to collaboratively train model through wireless networks. In order to enhance the performance of FL systems, an intelligent reflector surfaces (IRS) is deployed in the cell to assist the communication between UD and BS. Specifically, the IRS is composed of an array of passive reflective elements that is used for improving the communication link by varying the phase shift of the intended signal. In this paper, the computation and communication resource allocation of the FL system with IRS is designed to minimize the training latency, which is difficult to solve due to non-convex and the coupling of variables. In order to solve this problem, block coordinate descent (BCD) technique is first used to block the variables for alternately optimizing. Then, the user selection, the UDs’ CPU frequency and transmit power allocation are optimized by invoking the Majorize - Minimization(MM) algorithm, and IRS phase shift is obtained by using semidefinite relaxation (SDR) and Gaussian randomization. Numerical results show that the proposed algorithm effectively reduces the training time of the FL system.