Robust Over-The-Air Federated Learning In Heterogeneous Networks

Abstract

The integration of Artificial Intelligence into wireless networks is rapidly advancing, with Federated Learning (FL) at the forefront as the preserving privacy strategy that enables learning on edge devices. Over-The-Air FL (OTA-FL) harnesses Multiple Access Channels (MACs) for efficient, low-latency global model aggregation. However, traditional OTA-FL methods are hindered by additive noise, fading, and heterogeneity across devices. This paper introduces NoROTA-FL, a robust OTA-FL framework designed to jointly address the occurrence of heterogeneity alongside wireless impediments such as noise and fading. The key aspect of NoROTA-FL is that we derive novel optimization problems to mitigate the effects of noise, fading, and heterogeneity, ensuring robust convergence in non-convex settings. The derived optimization problem can be solved locally, while only aggregation is required globally, hence achieving robustness in a scalable and efficient manner. Our experiments demonstrate superior stability and accuracy of NoROTA-FL, particularly in scenarios with stragglers and additive noise, outperforming state-of-the-art methods.