Performance Evaluation of Distance-Statistical Based Byzantine-Robust Algorithms in Federated Learning

Due to its powerful learning capabilities and diverse applications, federated learning (FL) is becoming increasingly popular in the fields of wireless communications and machine learning (ML). Furthermore, since it enables several users to cooperatively train a global model without disclosing their local training data, FL represents a new methodology capable of attaining stronger privacy and security guarantees than current approaches. In this paper we propose new distance-statistical aggregation algorithms that provide robustness against Byzantine failures. In detail, a new class of aggregation algorithms is compared with the well-known federated algorithms on a set of simulations that recreate realistic scenarios (e.g. in the absence and presence of Byzantine adversaries). Achieved results demonstrate the functionality of the solutions in terms of accuracy and communication overhead, also under a correct and incorrect estimation of the attackers number.