Byzantine-Robust Federated Learning Based on Dynamic Gradient Filtering

Federated Learning (FL) introduces an innovative approach with the potential to enhance privacy and security compared to current methods. This is accomplished by enabling multiple users to collectively train a global model without disclosing their individual training data. Despite its advantages, this paradigm is vulnerable to poisoning attacks, wherein an adversary manipulates local model updates and injects them into the aggregating process. In this paper we propose a novel Robust and Dynamic Gradient Filter (RDGF) aggregation algorithms that provide resilience against model poisoning and backdoor attacks. In detail, the gradients received undergo initial processing to produce distance-based scores for clients, which are then compared to the average value of these scores. This comparison serves to eliminate malicious gradients before the final aggregation. We evaluate RDGF with well-known federated algorithms in a series of simulations that replicate realistic scenarios. These scenarios encompass situations with various percentages of Byzantine adversaries. The attained results showcase the effectiveness of the solutions in terms of both robustness and accuracy.