RFLDefense: robust federated learning defense against targeted attacks

Abstract

Abstract Federated learning (FL) is a distributed machine learning framework that protects data privacy by allowing multiple participants to jointly train models without sharing local data. However, the distributed nature of FL systems makes them susceptible to backdoor attacks, where attackers inject malicious updates to disrupt the performance of the global model, causing it to fail in specific tasks or inputs. The stealthiness and complexity of these attacks, especially when participant data distributions are non-independent and identically distributed (non-IID), limit the effectiveness of existing defense methods that primarily rely on anomaly detection. Additionally, some defense mechanisms rely on external validation datasets, which may be difficult to obtain in real-world scenarios or could be exploited by attackers. To address these issues, we propose a novel defense method called RFLDefense, comprising three core modules: hybrid similarity calculation, sign consistency verification, and robust layer aggregation. RFLDefense effectively identifies and suppresses malicious updates by analyzing the direction of updates and the distribution of parameter signs, while ensuring the robustness and performance consistency of the global model during the aggregation process. Experimental results on three widely used datasets demonstrate that RFLDefense can effectively defend against various backdoor attacks, outperforming existing defense algorithms while having minimal impact on model utility.