关键词: 图数据, 图神经网络, 图对抗攻击, 对抗样本

Abstract:

Deep learning for graph data modeling has shown excellent performance in complex tasks, including node classification, link prediction, and graph classification. However, the subtle perturbation on the input is easy to cause deep neural networks false output. Graph neural networks also inherit this vulnerability, it has raised concerns for adapting graph neural networks in safety-critical areas such as finance and transportation. By investigating the mechanism and method of graph adversarial attacks, it can improve the understanding of vulnerability and robustness of graph neural networks, and promote the wider application of graph neural networks. Graph adversarial attack has become an urgent research topic for further development. Firstly, the related concepts of graph adversarial attack are introduced. Then, according to attack strategies, adversarial attack algorithms are classified into three categories, including topology attack, feature attack and hybrid attack. Moreover, each category is summarized, such as its core ideas and strategies. Furthermore, some typical attacks are compared, including specific implementation methods, advantages, and disadvantages. Through the analysis of the state of the art, the existing problems and development direction of graph adversarial attacks are summarized, which can provide help for further researching graph adversarial attacks.

Key words: graph data, graph neural network, graph adversarial attack, adversarial example