Migratability of Adversarial Node Attacks in Graph Neural Networks: A Comprehensive Study

Graph Neural Networks (GNNs) are able to discriminate and retain information about the intrinsic properties of nodes, their position in the graph structure, and their relationships to other nodes, thus enabling prediction of node classes. However, GNNs are vulnerable to attacks when performing node categorization tasks. It has been observed that adversarial attacks demonstrate remarkable migration capabilities, suggesting that such attacks are not constrained to a specific model but can achieve comparable attack outcomes on other models as well. This paper presents the inaugural comprehensive examination of the migratability of adversarial attacks within the domain of graph neural networks. In the course of our experiments, we subjected the source graph neural network model to five distinct escape attacks, encompassing a range of dimensions, including the optimization strategy, graph topology, gradient information, reinforcement learning, and agent model. We then migrate the adversarial samples generated by these attacks to the target graph neural network model and evaluate their impact. Our experiments validate the migratability of the adversarial attacks in three different scenarios, i.e., migration across datasets, migration across models, and migration across models and datasets. The experimental results show that the five adversarial attack methods selected in this paper for the node classification task exhibit good migratability in three different migration scenarios. The attacks have a considerable impact on the source graph neural network, and they also demonstrate robust attack effects after migrating to the target graph neural network.