A defense mechanism for federated learning in AIoT through critical gradient dimension extraction

Abstract

Leveraging the distributed nature of the Internet of Things (IoT), Federated Learning (FL) facilitates knowledge transfer among heterogeneous IoT devices, enhancing the capabilities of Artificial Intelligence of Things (AIoT) while preserving data privacy. However, FL is susceptible to poisoning attacks such as label flipping, Gaussian, and backdoor attacks. Most existing defense strategies rely on robust aggregation algorithms that use the statistical properties of gradient vectors to counteract poisoning attacks, however, they often overlook the non-independent and identically distributed (non-iid) nature of client data, limiting their effectiveness in the IoT. We propose a method that combines cross-node Top-k gradient vector compression and Principal Component Analysis (PCA) dimensionality reduction to extract critical gradient dimensions. By clustering these essential dimensions and performing filtering, our approach effectively distinguishes malicious from benign clients in non-iid data scenarios. Additionally, we introduce a client trust-score assessment mechanism that continuously monitors client behavior and applies secondary filtering, further improving the identification of malicious clients. Experimental results on the CIFAR-10, MNIST, DomainNet, and Flowers102 datasets demonstrate that our method achieves higher model accuracy and robustness in non-iid data settings compared to existing defense strategies.