Dynamic gradient filtering in federated learning with Byzantine failure robustness

Federated Learning (FL) introduces a novel methodology with the potential to achieve enhanced privacy and security assurances compared to existing methods. This is achieved by allowing multiple users to collaboratively train a global model without revealing their individual training data. Moreover, the popularity of federated learning is on the rise in fields like wireless communications and machine learning (ML), thanks to its robust learning capabilities and versatile applications. In this paper we propose a novel Dynamic Gradient Filter (DGF) aggregation algorithms that provide robustness against Byzantine failures. Specifically, the received updates are first processed to generate clients' distance based scores, which are compared with the average value of these scores in order to remove malicious updates before final aggregation. DGF undergoes comparison with established federated algorithms through a series of simulations that replicate realistic scenarios, including scenarios both with and without the presence of Byzantine adversaries. We also demonstrate that DGF can be readily tailored to enhance the efficiency of the FL process without compromising its distinctive qualities. The attained results showcase the effectiveness of the solutions in terms of convergence rate, efficiency and resilience.