Distance-Statistical Based Byzantine-Robust Algorithms in Federated Learning

Abstract

New machine learning (ML) paradigms are being researched thanks to the current widespread adoption of AI-based services. Since it enables several users to cooperatively train a global model without disclosing their local training data, Federated Learning (FL) represents a new distributed methodology capable of attaining stronger privacy and security guarantees than current methodologies. In this paper, a study of the properties of FL is conducted, with an emphasis on security issues. In detail, a thorough investigation of currently known vulnerabilities and their corresponding countermeasures is conducted, focusing on aggregation algorithms that provide robustness against Byzantine failures. Following this direction, new aggregation algorithms are observed on a set of simulations that recreate realistic scenarios, in the absence and presence of Byzantine adversaries. These combine the Distance-based Krum approach with the Statistical based aggregation algorithm. Achieved results demonstrate the functionality of the proposed solutions in terms of accuracy and convergence rounds in comparison with well-known federated algorithms under a correct and incorrect estimation of the attackers number.