Intelligent Reflecting Surfaces Enhanced Federated Learning

Abstract

This paper investigates the problem of model aggregation for the federated learning aided by multiple intelligent reflecting surfaces (IRSs). Since the local parameters are transmitted over wireless channels, the undesirable propagation error will inevitably deteriorate the performance of global aggregation. The objective of this work is to reduce the signal distortion of over-the-air computation, thus the mean-square-error (MSE) is minimized by optimizing the transmit power at devices, controlling the receive scalar at the base station, and designing the phase shifts at IRSs, subject to the transmit power constraint for devices and the unit-modulus constraint for reflecting elements. To solve this intractable MSE minimization problem (P1), a closed-form expression for transmit power allocation is first derived, and the receive-scaling control subproblem (P2) is addressed by using the semidefinite relaxation. Then, the phase-shifting design subproblem (P3) is reduced to a feasibility-check problem, which is tackled by invoking the penalty method and successive convex approximation. After that, an alternating optimization algorithm is proposed to find a suboptimal solution for the non-linear and non-convex problem. Finally, compare to baselines, simulation results demonstrate that the designed algorithms for multi-IRS enhanced federated learning can converge faster and aggregate model more accurately.