Adaptive personalized over-the-air federated learning with reflecting intelligent surfaces

Abstract

Over-the-air federated learning (OTA-FL) unifies communication and model aggregation by leveraging the superposition property of the wireless medium, enabling bandwidth-efficient learning via simultaneous transmission of model updates. In this paper, we consider a federated learning system over a heterogeneous edge network, where clients have heterogeneous compute resources and non-i.i.d. local datasets under a non-convex objective. We augment the network with Reconfigurable Intelligent Surfaces (RIS) to enhance the learning system. We propose a cross-layer framework that jointly designs communication, computation, and learning resources. Specifically, we adapt the number of local steps, power, and the RIS configuration. We consider channel noise and channel estimation errors in both uplink (model updates) and downlink (global model broadcast), employing dynamic power control for both. We provide the convergence analysis for the proposed algorithms and extend the framework to personalized learning. Experiments demonstrate that the proposed algorithms outperform the state-of-the-art joint communication and learning baselines.