Decentralized Edge-Based Detection of Label Flipping Attacks in Federated Learning

Federated learning (FL) enhances data privacy by enabling decentralized model training without sharing local data. However, FL is vulnerable to Label-Flipping-Attacks (LFA), where malicious clients flip data labels. Traditionally, cloud-centric centralized defenses against LFA pose security and efficiency challenges. We propose a novel edge-based decentralized LFA detection method using edge servers for efficient detection. The proposed method considers the overall and class-wise accuracy in identifying suspicious clients. Initially, we adopt a strict zero-tolerance approach by excluding the entire update from detected malicious clients. We then experiment with four aggregation techniques-subtracting, masking, clipping, and reweighting–to handle the malicious parts of updates by focusing on the final layer neurons corresponding to specific classes. Experiments using three datasets demonstrate the effectiveness, robustness, and efficiency of our method, showing improved model performance and reduced latency under adversarial conditions. This approach improves the security and reliability of FL systems while maintaining data privacy.