FedBayes: An Aggregation Method for Federated Learning that uses Bayesian Regression

Over the years, the concept of Federated Learning (FL) has gained popularity based on compliance with the General Data Protection Regulation (GDPR). This framework allows multiple participants to collaborate on training a shared global model, without exchanging training samples. In real-world applications, the data of different participants are usually nonindependently and identically distributed (non-IID); in these cases, the clients are vulnerable to producing biased local models, for which FL-based aggregation could be improper. Several problems may also arise when some clients work with an insufficient amount of data or are exposed to potential attacks. To address these main issues, this paper proposes a new algorithm that evaluates the accuracy of models derived from participating clients and uses Bayesian regression to aggregate the local models. Through this strategy, the number of clients who can negatively affect the global model can be reduced, whether they have been attacked or have obtained poor performance for other reasons (e.g., an insufficient amount of data). The experiments are performed using a previously proposed algorithm, FedAccSize, together with this new extension, FedBayes. In the considered scenarios, both independent and identically distributed data (IID) and non-IID data are used to demonstrate the ability of the proposed algorithms to obtain an aggregation of the global model which can be done safely. Based on the comparison with the established algorithms FedAvg and FedAvgM, it results that assessing the quality of the local models is crucial for aggregation.