IRS-Assisted Digital Over-the-Air Federated Learning
Yudi Pan, Zhibin Wang, Liantao Wu, Yong Zhou · 2023
For the purpose of training a machine learning model via exploiting data from multiple devices without compromising their privacy, federated learning (FL) has become a popular approach. Meanwhile, over-the-air computation (AirComp) enables concurrent model transmission to accelerate model aggregation in the context of FL. However, the performance of model aggregation is significantly hindered by adverse wireless channels. In this paper, we employ intelligent reflecting surface (IRS) to facilitate accurate model aggregation in AirComp-based FL. To ensure compatibility with existing communication standards, this paper adopts uniform quantization for both downlink model broadcast and uplink AirComp-based gradient aggregation. Furthermore, we quantitatively examine the impact of quantization errors on transmission accuracy and convergence bound. To mitigate signal distortion, we employ an alternating optimization algorithm that optimizes the beamforming vector at the base station, the transmit/receive scalars at the devices, and the phase shifts at the IRS. The simulation results provide compelling evidence for the effectiveness and robustness of our proposed method.