Over-the-Air Federated Learning Exploiting Channel Perturbation

Abstract

Federated learning (FL) is a promising technology which trains a machine learning model on edge devices in a distributed manner orchestrated by a parameter server (PS). To realize fast model aggregation, the uplink phase of FL could be carried out by over-the-air computation (OAC). On the one hand, engaging more devices in FL yields a model with higher prediction accuracy. On the other hand, the edge devices in OAC need to perform appropriate magnitude alignment to compensate for underlying channel coefficients. However, due to the limited power budget, this is not possible for devices experiencing deep fade. Consequently, these devices are excluded from the FL algorithm. In this paper, we propose a channel perturbation method so that no edge device is excluded due to experiencing deep fade. To this end, OAC is performed in multiple phases. In each phase, the radio frequency (RF) vicinity of PS’s antenna is intentionally perturbed by means of RF mirror structure coined in [1]. This yields independent realizations of channels between PS and devices in each phase. By using proper transmit scalars, all devices concurrently transmit their local model updates in each phase subject to a total power constraint. Then, the PS estimates the arithmetic sum of the local updates by properly combining the aggregated models obtained across all phases. The devices’ transmit scalars and PS’s de-noising factors can be efficiently found by solving a tractable optimization problem. Index Terms—Federated learning, over-the-air computation, edge machine learning, wireless communications.