GNNHacker: Adaptive Subgraph Backdoor Attacking Method with Saliency Analysis and Joint Optimization

Xiaojie Wu, Yan Luo, Shujie Li, Xiping Hu, Qiang Liu, Chengming Li

Abstract: Subgraph backdoor attacks reveal critical vulnerabilities in graph neural networks (GNNs). They replace GNNs’ nodes and edges with elaborate triggers, causing the misclassification of graph-structured data examples into target categories specified by adversaries. To overcome the high computational overhead of existing attacking methods, we propose an adaptive subgraph backdoor attack method with saliency analysis and joint optimization called GNNHacker. Specifically, GNNHacker introduces a gradient-based saliency map to identify the nodes with high saliency scores as poisoning targets. Then, it adopts a joint optimization mechanism by simultaneously optimizing trigger generation and the training of backdoor GNNs. Experimental results over five public datasets show that the proposed method significantly outperforms baselines in terms of attacking capability and efficiency.

External IDs:dblp:conf/icic/WuLLHLL25