Decentralized federated learning in wireless networks

Abstract

Current wireless applications such as autonomous driving, UAVs, IOT devices, etc. generate massive amounts of data that can be utilized to train machine learning models for decision making. Privacy, security and bandwidth constraints have led to the use of federated learning (FL) in wireless systems, where the problem of learning a centralized ML model is accomplished by training on client-specific local datasets in a server-based and decentralised setting. In this work, we propose the decentralized clustered wireless FL (CWFL), where a novel FL strategy is employed using clustering based on signal-to-noise ratio and channel strength of wireless links between the clients. We show that the proposed technique requires fewer channels as compared to the decentralized FL technique. Using the MNIST dataset, we demonstrate that CWFL outperforms the vanilla-FL and a power-allocation based improved FL strategy with respect to average accuracy and robustness across iterations