Topology-Guided Adversarial Attack on Graph Neural Networks Based on Dynamic Node Selection
Abstract
Graph Neural Networks (GNNs) have been widely applied in numerous fields including the Internet of Things (IoT), yet they suffer from inherent security vulnerabilities, making them vulnerable to Graph Modification Attacks (GMA). Specifically, attackers can degrade the classification performance of GNNs by maliciously modifying the features or topological structures of a small number of nodes in the graph. However, attackers generally operate in black-box scenarios, with no access to the architecture, parameters, defense strategies and ground-truth labels of the victim model, nor to the graph topology and node features processed by defense mechanisms. This significantly limits the ability of attackers to identify critical nodes in the defender-revised real graph and craft effective adversarial samples. In response to the aforementioned problems, this paper proposes a topology-guided adversarial attack method based on dynamic node selection (DNS-TGA). Initially, we calculate the similarity between the victim model and the shadow model, which facilitates the construction of a graph representation that closely aligns with the victim network. Next, we employ the volcanic eruption algorithm (VEA) to identify critical nodes and their corresponding feature information from the generated graph representation. Finally, we establish a dual-discriminator that incorporates both the trained shadow model and the victim model. This discriminator enhances the attack model and produces highly effective perturbation samples. Experimental results indicate that, compared to other attack algorithms, DNS-TGA demonstrates superior attack performance across six representative GNN models, along with significant scalability and transferability.