XAI-DTBD: Explainable dynamic threshold-based backdoor detection in graph neural networks.

Backdoor detection in neural networks is an emerging area of research in which backdoor triggers are injected to alter the model's output. For example, injected triggers into the citation network by modifying the features and structures of certain nodes. Whenever a model faces such a trigger, it misclassifies the node. In this paper, we propose a novel method for detecting backdoors in Graph Neural Networks (GNNs), which utilises model activation clustering and dynamic thresholding. The proposed model, XAI-DTBD: Explainable Dynamic Threshold-based Backdoor Detection in Graph Neural Networks, begins by selecting the most important layer in the model for activation clustering. Utilizing the clustering metric to find the quality of a cluster and the sensitivity of the dataset. For dynamic thresholding, we use cluster measures, data sensitivity, and the percentage of data points poisoned. After finding the dynamic threshold for the specific data, compare it with other metrics for backdoor detection. Besides, explainable AI 2.0 methods such as SHAP and Grad-CAM are used to identify abnormal nodes and subgraphs, making the detection process more interpretable and trustworthy. Detailed experiments on benchmark datasets, such as Cora, CiteSeer, PubMed, and MUTAG, across various backdoor attack scenarios, show that XAI-DTBD achieves a high average detection rate with a low false positive rate compared to current state-of-the-art backdoor detection algorithms for GNNs.