针对图神经网络的单节点扰动攻击

Single Node Adversarial Attack Against Graph Neural Network

图神经网络(Graph Neural Networks,GNN)已经在各种图相关应用中显示出优异的性能。最近的研究表明,GNN模型容易受到精心构造的对抗性扰动攻击,导致模型性能下降。但先前的工作大多集中于修改图结构,这会改变图中重要的拓扑特征。对图数据上的间接对抗攻击进行了研究,提出了一种基于强化学习的单节点对抗攻击(Single Node Adversarial Attack,SNAA)来修改图中的节点特征。攻击设定在黑盒场景下,仅可以对测试数据进行有限次数的黑盒查询,并通过严格限制攻击预算以保证攻击不可察觉。通过在多个数据集上的实验表明,这种攻击对各种GNN模型都是有效的。

Graph neural networks(GNNs)have shown excellent performance in a variety of graph-related applications.Re-cent studies show that GNN models are vulnerable to carefully construct adversarial perturbations,resulting in degraded model per-formance.Most of the previous research on graph adversarial attacks focus on modifying the graph structure,which will change the important topology properties of the graph.Indirect adversarial attacks on graph data are studied,and a single node adversarial at-tack(SNAA)based on reinforcement learning to modify the node features in graphs is proposed.The attack is set in a black-box scenario,where only a limited number of black-box queries can be performed on the test data,and the attack budget is strictly limit-ed to ensure that the attack is imperceptible.Experiments on multiple datasets show that SNAA is effective against various GNN models.