Defending Against Backdoors in Federated Learning by Self Supervised Learning and Anomaly Client Detection

Federated Learning (FL) is a distributed learning framework that enables multiple parties to train models without disclosing potentially sensitive personal data, breaking down data silos and ensuring strong privacy protection. However, its decentralized nature means that individual participants can be compromised, potentially introducing backdoors into the model through data or gradient poisoning. Recent approaches have either focused on detecting backdoors during the aggregation process to reduce contamination or on erasing the backdoors post-training, using techniques such as model pruning. However, these methods often lead to a degradation in model performance and rely on several idealized conditions, such as having a clean dataset on the server side for detection, which is unrealistic in real-world scenarios. To address these challenges, we propose a novel method that serves as an effective defense against DBA(DISTRIBUTED BACKDOOR ATTACKS) in FL. Our method is designed to resist two types of attack settings and performs well even under client-side data heterogeneity. Specifically, our approach employs Self supervised learning to train a cleaner feature extractor and integrates a two stage malicious detection method to identify suspicious local updates, effectively mitigating backdoor attacks. We conduct comprehensive experimental evaluations across different datasets and attack scenarios, showing that our method outperforms state-of-the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\operatorname{art}($</tex> SOTA) defenses against both single-round and continuous backdoor attacks. Notably, our approach provide a significant reduction in the backdoor effect(e.g., attack success rate is less than 14.5%) while maintaining nearly unchanged performance on the main task.