Over-the-Air Hierarchical Personalized Federated Learning

Abstract

Data heterogeneity and communication bottleneck are two critical factors that limit the performance of federated learning (FL) over wireless networks. To address these challenges, this paper introduces a hierarchical personalized federated learning (HPFL) framework, which employs a three-tier network architecture to enable the simultaneous learning of a global model and multiple personalized local models. Meanwhile, over-the-air computation (AirComp) is leveraged to support communication-efficient device-to-edge and edge-to-cloud model aggregations. To provide useful guidance for enhancing learning performance, we derive the convergence bound of the proposed AirComp-assisted HPFL, taking into account the interference among different clusters as well as data heterogeneity across different devices. To minimize the impact of accumulated transmission distortion on learning performance, we formulate an optimization problem involving the beamforming design at both cloud and edge servers, followed by developing a successive convex approximation-based algorithm at the cloud server and an interference-aware algorithm at each edge server to perform the receive beamforming design. Simulation results demonstrate that our proposed framework outperforms other FL frameworks and transceiver design algorithms in terms of test accuracy.