Robust Over-the-Air Federated Learning

Abstract

Interest continues to grow in using federated learning (FL) for a variety of signal processing and communications applications. This paper focuses on a robust design for FL to mitigate the effects of noise and fading channels. To enhance the efficiency of FL in bandwidth-limited environments, over-the-air (OTA) computation has been proposed based on the superposition property of a wireless multiple-access channel (MAC). However, OTA FL inherently faces challenges with channel noise and wireless channel fading in the wireless MAC, which could degrade optimization procedure and significantly reduce the accuracy of the trained model. To tackle this challenge, we introduce a novel approach using a Kalman filter (KF)-based OTA FL algorithm in this paper.