Reconfigurable Intelligent Surface-Assisted Wireless Federated Learning With Imperfect Aggregation

This paper proposes a new Signal-to-interference-plus-noise ratio (SINR)-based Device selection, Power control, and Reconfigurable intelligent surface (RIS) configuration (SDPR) algorithm, which allows imperfect aggregation of wireless federated learning (FL) in RIS-assisted Non-Orthogonal Multiple Access (NOMA) systems. The SDPR algorithm selects the local models with SINRs within an acceptable range for global aggregations, benefiting FL from involving more local models with tolerable errors. The convergence of FL under the imperfect aggregation is analytically validated, where the influence of the local model quantization and modulation is captured through the translation of the SINR thresholds to the symbol error rates (SERs). Employing successive convex approximation and gradient descent, we jointly optimize the RIS configuration and the transmit powers of participating devices, thereby minimizing the convergence upper bound of FL under imperfect aggregation. Experimental results demonstrate that using SDPR, FL achieves superior convergence and accuracy by effectively utilizing model updates, even if they are received with errors. Moreover, more quantization bits do not necessarily offer better FL accuracy, and need to be tailored under specific SERs.

I. Introduction

Wireless Federated Learning (FL) enables distributed wireless devices to compute local model updates based on their individual data sets and only send their local model parameters to a designated parameter server, such as a base station (BS) [1], [2], [3]. The BS aggregates and updates the global model parameters, and then sends the global model parameters back to the participating devices. This repeats until convergence. However, limited communication resources in wireless networks can significantly affect the performance of FL, as the number of devices participating in FL can be heavily restricted [4]. To this end, it is important to jointly design device selection and communication resource allocation to facilitate the convergence of FL in resource-constrained wireless networks.