Over-the-Air Federated Learning Under Time-Varying Wireless Channels Using OTFS

Abstract

Due to increased scalability and enhanced data privacy, federated learning (FL) techniques become popular among the research community. Specifically, FL over wireless networks finds a broad spectrum of applications. However, they are often difficult to realize in practice primarily due to the challenges posed by wireless channels. In this paper, we propose an over-the-air (OTA) computation-based FL scheme suitable for time-varying wireless channels. An FL over time-varying wireless channel is challenging, therefore, the proposed scheme is designed based on orthogonal time frequency space (OTFS) modulation to make it resilient to time-varying wireless channels. To perform OTA computation, we develop a novel precoding technique under an optimal transmit power allocation for each user. Since precoding requires channel state information (CSI) of all the devices involved in FL, frequent exchange of CSI may increase control overhead significantly and eventually affect the performance of the FL scheme. Therefore, we develop an extended Kalman filter based CSI estimation scheme to reduce control overhead due to CSI exchange. Numerical simulations reveal that the proposed scheme has a better convergence rate and prediction accuracy than the orthogonal frequency-division multiplexing based learning schemes.