An Exhaustive Investigation on Resource-aware Client Selection Mechanisms for Cross-device Federated Learning

Abstract

Federated learning (FL) is a distributed machine learning technique in which each client (FLClient) trains a model without revealing it’s local data to the server. Appropriate client selection is a crucial step towards ensuring the quality and robustness of the global model in a cross-device FL set-up. As such, various client selection mechanisms have been proposed, however, most of the mechanism makes the assumption of clients (devices) being mobile phones with uninterrupted power and compute resources supply. On the other hand, due to growing digitization in various industries, clients in a cross-device FL set-up can be resource-constrained IoT edge devices such as single board computers. To this end, there are a few resource-aware client selection mechanisms proposed in the literature. This paper provides a comprehensive, experimental comparative analysis of these mechanisms while resource-constrained IoT edge devices as clients. The effect of varying FL specific hyper-parameters on accuracy, convergence time and client retention is observed for all resource-aware client selection mechanisms so that a cognitive choice of the client selection mechanism can be made for a given application scenario.