Transferable Graph Backdoor Attack

Graph Neural Networks (GNNs) have achieved tremendous success in many graph mining tasks, benefitting from the message passing strategy that fuses the local structure and node features for much better graph representation learning. Despite the excellent performance of GNNs, but similar to other type of deep neural networks, the robustness of GNNs is unsatisfactory. It have been disclosed by many works that GNNs are vulnerable to unnoticeable perturbations on both graph structure and node features. Many adversarial attacks have been proposed to disclose the fragility of GNNs under different perturbation strategies to create adversarial examples. However, less work has been done to show the vulnerability of GNNs under backdoor attack. To fill this gap, in this paper, we present GHAT, transferable GrapH bAckdoor aTtack. The core principle of GHAT is to poison training dataset with perturbation triggers that can lead to effective and transferable backdoor attack. The perturbation trigger for a graph is generated by performing the perturbation actions on the graph structure via a gradient based score matrix. Compared with the prior works, GHAT is different in several ways: it exploits a surrogate GCN model to generate perturbation trigger for black-box based backdoor attack; it generates sample-specific perturbation triggers which do not have fixed pattern; the attack of GHAT can be transferable to different GNN models when trained with the poisoned training dataset forged by GHAT. Through extensive evaluation on four real-world datasets, we demonstrate that GHAT shows much better attack effectiveness in regard to transferable backdoor attack on GNNs.