| 摘要: |
| 为提高对抗性攻击在大规模图上的攻击效率,提出了基于子图采样的对抗样本生成方法. 该方法通过引入PageRank、余弦相似度及K跳子图等技术,提取与目标节点高度相关的子图,在大规模图上缓解了计算梯度效率较低的问题,在降低被攻击模型准确性的同时提升了攻击的隐蔽性. 实验结果表明： 所提出的对抗性攻击方法与基于梯度攻击的GradArgmax算法相比,在Cora数据集上提升了30.7%的攻击性能,且在Reddit大规模数据上能够计算GradArgmax算法无法计算的攻击扰动. |
| 关键词: 图神经网络 对抗性攻击 子图提取算法 |
| DOI：10.3969/J.ISSN.1000-5137.2024.02.004 |
| 分类号:TP181 |
| 基金项目:国家自然科学基金青年基金（62302308）,上海市青年科技英才扬帆计划（21YF1432900） |
|
| Subgraph sampling-based adversarial attack method for large-scale graphs |
|
GAO Xin1, AN Dongdong1, ZHANG Xiaofeng2
|
|
1.College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 201418, China;2.Shanghai Newtouch Software Co., Ltd., Shanghai 200120, China
|
| Abstract: |
| A subgraph sampling-based adversarial example generation method was proposed to enhance the efficiency of adversarial attacks on large-scale graphs. PageRank, cosine similarity, and K-hop subgraphs were employed to extract subgraphs highly relevant to the target node in this method, alleviating the issue of low gradient computation efficiency in large-scale graphs. The stealthiness of the attack was also increased while reducing the accuracy of the attacked model. Experimental results showed that attack performance was improved by 30.7% on the Cora dataset by this adversarial attack method compared to the GradArgmax algorithm, and attack perturbations on large-scale like Reddit dataset could be computed which the GradArgmax algorithm could not achieve. |
| Key words: graph neural network adversarial attack subgraph extraction algorithm |
Editorial Board