Optimized Power Control for Over-the-Air Federated Averaging With Data Privacy Guarantee

Abstract

This paper investigates the privacy-preserving power control problem for <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">over-the-air federated edge learning</i> (Air-FEEL), which features “one-shot” aggregation and significantly reduced communication latency. In Air-FEEL, the inevitable random perturbation encountered in the aggregation process due to the corruption by channel fading and noise poses a fundamental trade-off between the privacy and accuracy, as larger perturbation may harm the accuracy but benefit the privacy, and vice versa. Therefore, power control in Air-FEEL, as the main approach to regulate the said random perturbation, needs to be judiciously designed for balancing the accuracy-privacy trade-off. This, however, is a largely uncharted area. To bridge the research gap, we first analyze the convergence behavior (in terms of the optimality gap) and the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">differential privacy</i> (DP) performance of Air-FEEL with respect to the power control policy at different iterations. Then, to achieve the maximized training accuracy under given DP guarantee requirement, we minimize the optimality gap by jointly optimizing the power control at edge devices and the denoising factors at the edge server, subject to a set of power constraints at individual edge devices and the DP constraint. Experimental results validate that the proposed power control policy for Air-FEEL achieves a better privacy-accuracy trade-off compared with the benchmarks.