DWAMA: Dynamic weight-adjusted mahalanobis defense algorithm for mitigating poisoning attacks in federated learning

Abstract

Federated learning is a distributed machine learning approach that enables participants to train models without sharing raw data, thereby protecting data privacy and facilitating collective information extraction. However, the risk of malicious attacks during client communication in federated learning remains a concern. Model poisoning attacks, where attackers hijack and modify uploaded models, can severely degrade the accuracy of the global model. To address this issue, we propose DWAMA, a federated learning-based method that incorporates outlier detection and a robust aggregation strategy. We use the robust Mahalanobis distance as a metric to measure abnormality, capturing complex correlations between data features. We also dynamically adjust the aggregation weights of malicious clients to ensure a more stable model updating process. Moreover, we adaptively adjust the malicious detection threshold to adapt to the Non-IID scenarios. Through a series of experiments and comparisons, we verify our method’s effectiveness and performance advantages, offering a more robust defense against model poisoning attacks in federated learning scenarios.