HFLGuard: clustering and autoencoder-based defense method against heterogeneous federated poisoning
Year: 2025
Type: article
Abstract: The distributed nature of federated learning made it vulnerable to attacks by malicious clients, leading to privacy leakage and model degradation. Additionally, the inherent heterogeneity and multi-source nature of real-world data distributions presented further challenges in mitigating these threats. To address these issues, a heterogeneous federated poisoning defense method HFLGuard was proposed. The method began by computing a weighted history gradient for each client to emphasize potential malicious “targets”. Subsequently, the K-means clustering algorithm provided an initial classification of gradients, followed by a Transformer-based autoencoder that detected malicious gradients and quantified their deviation. Finally, a dynamic trust score aggregation mechanism was then employed to penalize clients engaged in repeated poisoning attempts, assigning appropriate weights based on trust scores. Experiments conduct on benchmark datasets, including MNIST, FMNIST, CIFAR-10, and CIFAR-100, demonstrate that the proposed approach effectively mitigates both targeted and untargeted poisoning attacks.
Language: Chinese
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7117522603