FedTCS: federated learning with time-based client selection to optimize edge resources

Client sampling in federated learning (FL) is a significant problem, especially in massive cross-device scenarios where communication with all devices is not possible. In this work, we study the client selection problem using a time-based back-off system in federated learning for a MEC-based network infrastructure. In the FL paradigm, where a group of nodes can jointly train a machine learning model with the help of a central server, client selection is expected to have a significant impact in FL applications deployed in future 6G networks, given the increasing number of connected devices. Our timer settings are based on an exponential distribution to obtain an expected number of clients for the FL process. Empirical results show that our technique is scalable and robust for a large number of clients and keeps data queues stable at the edge.