Hybrid Reconfigurable Intelligent Surface Assisted Over-the-Air Federated Learning

Jiaqi Jin, Zhibin Wang, Liantao Wu, Yong Zhou

Abstract: By making full use of the superposition property, over-the-air computation (AirComp) enables low-latency model aggregation in wireless federated learning (FL). Meanwhile, reconfigurable intelligent surface can be adopted to mitigate the communication bottleneck of model aggregation in AirComp-based FL by introducing an additional reflective path. However, the double path loss attenuation in the reflective link limits the performance improvement delivered by passive RIS. To alleviate the detrimental effect of the double path loss attenuation, we propose to deploy a hybrid RIS with both active and passive elements to support over-the-air FL. We characterize the impact of gradient distortion on the convergence of FL and further formulate a gradient distortion minimization problem, while considering the modulus constraints of RIS. Furthermore, we develop an alternating minimization algorithm to implement joint design for the transmit scalars, RIS amplifying/reflecting coefficients, and receive beamforming. Simulation results show that our proposed hybrid RIS aided AirComp-based FL achieves superior performance in terms of test accuracy.