Data sharing-based approach for Federated Learning tasks onEdge Devices

Abstract

Federated Learning (FL) enables edge devices to collaboratively train a global machine learning model. In this paradigm, the data is maintained on the devices themselves and a server is responsible for aggregating the parameters of the local models. However, the aggregated model may present convergence difficulties when the device data are non-independent and identically distributed (non-IID), that is, when they present a heterogeneous distribution. This work proposes an algorithm that extends data sharing-based solutions from the literature by considering privacy-flexible environment, where users agree to share a small percentage of their private, and privacy-sensitive environment, where it is assumed that the aggregator server has a set of public global data that is shared with users in the initial phase of the FL process. The proposed algorithm is evaluated in a distributed and centralized way considering a Human Activity Recognition (HAR) application. The results show that data sharing strategies indicate improved global model performance in non-IID scenarios.