DeinfoAttack: A heuristic graph adversarial attack algorithm leveraging graph topological information entropy

doi:10.1016/j.neucom.2025.131904

CSpace

	DeinfoAttack: A heuristic graph adversarial attack algorithm leveraging graph topological information entropy
	Wang, Yajing 1,3; Cao, Huawei 1,2; Song, Shuhan 1
	2026-01-14
发表期刊	NEUROCOMPUTING
ISSN	0925-2312
卷号	661 页码:12
摘要	Graph Neural Networks (GNNs) have been widely studied for their ability to capture node interdependencies and integrate node attributes, improving node classification accuracy. Recent research has shown that GNNs designed for heterophilic graphs can learn more information than traditional GNNs, bringing new challenges for developing attack and defense strategies across varying heterophily levels. However, existing adversarial attack algorithms have predominantly concentrated on homophilic graphs, often overlooking the heterophilic scenario where their effectiveness is compromised. We tackle this challenge and propose a novel universal heuristic attack algorithm DeinfoAttack, which stands for Decreasing informativeness while Attacking. For a deeper understanding, we theoretically analyze the more critical factor beyond homophily and heterophily: Graph Topological Information Entropy derived from Neighborhood Similarity Matrix S(A), to quantify the informativeness provided by the graph topology during GNN predictions. DeinfoAttack aims to maximize Graph Topological Information Entropy within a limited budget by preferentially targeting nodes that provide more graph topological information. Extensive experiments demonstrate that DeinfoAttack is efficient and effective, with the potential to reduce the GNNs' accuracy by up to 30 %. For example, the accuracy decreased from 74.2 % to 44.2 % after attacking Cora, and the accuracy decreased from 52.63 % to 22.59 % after attacking Chameleon. Furthermore, the overall performance is superior to existing attack algorithms in terms of attack effect, attack cost, and generality.
关键词	Graph adversarial attack Heterophilic GNNs Information entropy
DOI	10.1016/j.neucom.2025.131904
收录类别	SCI
语种	英语
WOS研究方向	Computer Science
WOS类目	Computer Science, Artificial Intelligence
WOS记录号	WOS:001614242300008
出版者	ELSEVIER
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/43098
专题	中国科学院计算技术研究所
通讯作者	Cao, Huawei
作者单位	1.Chinese Acad Sci, Inst Comp Technol, Beijing 100190, Peoples R China 2.Zhongguancun Lab, Beijing, Peoples R China 3.Univ Chinese Acad Sci, Nanjing 211135, Peoples R China
推荐引用方式 GB/T 7714	Wang, Yajing,Cao, Huawei,Song, Shuhan. DeinfoAttack: A heuristic graph adversarial attack algorithm leveraging graph topological information entropy[J]. NEUROCOMPUTING,2026,661:12.
APA	Wang, Yajing,Cao, Huawei,&Song, Shuhan.(2026).DeinfoAttack: A heuristic graph adversarial attack algorithm leveraging graph topological information entropy.NEUROCOMPUTING,661,12.
MLA	Wang, Yajing,et al."DeinfoAttack: A heuristic graph adversarial attack algorithm leveraging graph topological information entropy".NEUROCOMPUTING 661(2026):12.