New Defense Mechanism for Federated Learning Introduced
/ 1 min read
🛡️✨ New defense mechanism enhances Federated Learning against model poisoning attacks. Researchers have introduced Kernel-based Trust Segmentation (KeTS), a novel approach designed to protect Federated Learning (FL) systems from model poisoning attacks, which compromise the accuracy of global models by injecting malicious updates. KeTS utilizes Kernel Density Estimation to effectively identify and segment malicious clients, even amidst benign outliers, outperforming existing defenses like Krum and Trim-Mean by over 24% on the MNIST dataset and 14% on Fashion-MNIST. The study demonstrates KeTS’s robust performance across various attack scenarios, highlighting its potential to significantly improve the resilience of FL systems in heterogeneous environments.
