FL-ProtectX: Combating Label flipping in complex multi-agent environments

Federated Learning (FL) has emerged as a key enabler for secure and decentralized model training while preserving user data privacy. Nonetheless, FL is vulnerable to poisoning attacks; in which a malicious participant tries to corrupt the global model by sending manipulated updates. One common attack is label-flipping, where attackers intentionally mess up labels for training data on their end. Current defenses against these threats treat scenarios where a single adversarial group is cooperating in coordination. This misses the modern real-world scenario in which different adversarial groups have differing, conflicting objectives and act independently. This paper introduces FL-ProtectX, a defense mechanism designed to combat label-flipping attacks targeting heterogeneous multi-adversary environments. Our model also detects the adversary (one that pursues a set of attack goals) behind and can differentiate from other adversaries using their unique target attacks. FL-ProtectX: by utilizing last-layer gradient analysis and cosine similarity computation along with PCA for compression, provides accurate separation and ensures that malicious participants are detected. To facilitate global model aggregation, our method adaptively adjusts trust factors according to the calculated angular similarities between updates. We evaluate the effectiveness of FL-ProtectX on defending against adversary groups by experimentally studying label-flipping attacks on the CIFAR10 dataset and show that for most situations, it effectively neutralizes these attacks, reducing attack success rate from 17.12 to ${1 1. 0 3}$ while maintaining high accuracy in non-attack-related classes in multi-agent configurations.