Decentralized Federated Learning Over Random Access Channel

In this letter, a Federated Learning (FL) system where a server does not exist is investigated. In the absence of the server, entire learning process including exchange of model updates is conducted in a distributed manner. Hence, communication protocol is also required to be decentralized. When large number of devices communicate distributively, heavy congestion of communication is inevitable, which leads to huge amount of time for decentralized FL. This letter proposes a novel method to enhance communication efficiency when the decentralized FL system exploits random access protocol. By leveraging the learning characteristics of updates provided by decentralized FL, devices decide on transmission based on their size of dataset, achieving rapid model convergence with low communication overhead. In addition to that, adapting transmission probability is also proposed. Through extensive experiments, we validate our proposed scheme which outperforms existing studies in both case of homogeneous and heterogeneous data distribution.