Detection and Mitigation of Label-Flipping Attacks in Federated Learning Systems with KPCA and K-Means
Dongcheng Li, W. Eric Wong, Wei Wang, Yao Yao, Matthew Chau
Abstract
Federated learning is a popular machine-learning technique that is often preferred due to its efficiency and data privacy. However, federated-learning systems face a serious threat of data poisoning that can cause the systems and predictions to fail if not treated in time. This study extends another study of data-poisoning attacks in federated-learning systems conducted by Tolpegin et al. We first investigate the effectiveness of the defense strategy suggested in Tolpegin’s study. Then we propose an improved defense strategy that emphasizes employing KPCA and K-mean clustering. It is demonstrated in this paper that our defense strategy, when combined with improved dimensionality-reduction algorithms, produces better results in defending against data-poisoning attacks in federated-learning systems.