FedAcc and FedAccSize: Aggregation Methods for Federated Learning Applications

Abstract

This paper presents the ability of the federated learning concept to create a collaboration between multiple devices using a shared global model, while still keeping data privacy to meet the General Data Protection Regulation (GDPR). In real-world application scenarios, this concept faces problems related to the defense of the global model from possible attacks and the compatibility with non-independent and identically distributed data (non-IID). This paper presents two aggregation algorithms compatible with non-IID data, which use a refined aggregation of the local model, based on their accuracy. Thus, the proposed algorithms can refine the confidence in each client, eliminate intruders and allow a safe aggregation of the global model. Testing scenarios performed for IID and non-IID data illustrate that the proposed algorithms are able to provide faster training and improved robustness against intruders, w.r.t. the well-known federated average algorithm.