Multi-Task Over-the-Air Federated Learning in Cell-Free Massive MIMO Systems

Abstract

Wireless devices are expected to provide a wide range of AI services in 6G networks. The increasing computing capabilities of wireless devices and the surge of wireless data motivate the use of privacy-preserving federated learning (FL). In contrast to centralized learning that requires sending large amounts of raw data during uplink transmission, only local model parameters are uploaded in FL. Meanwhile, over-the-air (OtA) computation is considered as a communication-efficient solution for fast FL model aggregation by exploiting the superposition properties of wireless multi-access channels. The required communication resources in OtA FL do not scale with the number of FL devices. However, OtA FL is significantly affected by the uneven signal attenuation experienced by different FL devices. Moreover, the coexistence of multiple FL groups with different FL tasks brings about inter-group interference. These challenges cannot be well addressed by conventional cellular network architectures. Recently, Cell-free Massive MIMO (mMIMO) has emerged as a promising technology to provide uniform coverage and high rates via joint coherent transmission. In this paper, we investigate multi-task OtA FL in Cell-free mMIMO systems. We propose optimal designs of transmit coefficients and receive combining at different levels of cooperation among the access points, aiming to minimize the sum of OtA model aggregation errors across all FL groups. Numerical results demonstrate that Cell-free mMIMO significantly outperforms conventional Cellular mMIMO in term of the FL convergence performance by operating at appropriate cooperation levels.