FedDCS: A distributed client selection framework for cross device federated learning

Abstract

In cross-device Federated Learning (FL) the existing client selection approaches are centralised, making the assumption that the participatory clients are resource-full and restrict their type to mobile phones. These assumptions limit the applications of FL and make it difficult for the global server to monitor the client (s) resources; as a consequence of which a high client drop-out rate is observed. Thus, a distributed cross-device FL framework with the provision of resource & time-aware client selection is required. This paper proposes a distributed cross-device FL client selection framework (FedDCS) which employs a novel data-type, resource, and time-aware protocol for FL client selection. FedDCS is designed to offload the client selection and their resource monitoring processes to the local edge servers with the objective to minimize client(s) drop-outs, model convergence time, number of message exchanges, and time to achieve a certain level of accuracy. Extensive experiments on two benchmark data sets: MNIST and FashionMNIST, are performed on the PyTorch-based FL platform to evaluate the performance of FedDCS on such metrics. The results show 100% client retention in FedDCS during FL model training for both MNIST and FashionMNIST datasets using IID and Non-IID settings. For the MNIST dataset, FedDCS shows 82% and 36% improvement in the time taken to reach the desired testing accuracy, and 18% and 10% improvement in model convergence time, for IID and Non-IID setting respectively as compared to the best performing state-of-the-art, i.e., FedMCCS. For the FashionMNIST dataset, FedDCS shows 17% and 6% improvement in the time taken to reach the desired testing accuracy, and 4% and 1% improvement in model convergence time, for IID and Non-IID setting respectively as compared to the best performing state-of-the-art, i.e., FedMCCS. In addition, FedDCS also significantly reduces the number of message exchanges during FL.