燕山大学学报2025，Vol.49Issue(4)：309-318,10.DOI:10.3969/j.issn.1007-791X.2025.04.004
基于GNN因果推断的结构增强漏洞检测模型
Structure enhanced vulnerability detection model based on GNN causal inference
摘要
Abstract
To solve the problem that existing vulnerability detection methods extract graph structure features using simple graph neural network models and directly generalize information labels and graph structures out of distribution,leading to low detection efficiency.a structure-enhanced vulnerability detection model based on graph neural network causal inference,called SEVDM-GCI,is proposed.The model treats source code as a linear token sequence.First,it constructs a graph structure based on word co-occurrence relationships.Then,it divides the graph into a causal graph and a confounding graph through the residual connections of the graph neural network.Variables are strategically confounded to simulate the causal relationship between causal variables and labels.Finally,node embedding is performed on the causal and confounding graphs to enhance the graph's structural features.SEVDM-GCI is verified on a real benchmark dataset from CodeXGLUE,and the detection results show improvements of 3.15%,3.77%,and 2.57%in accuracy,precision,and F1 score,respectively,compared with the optimal baseline method.The performance of vulnerability detection is significantly improved.关键词
深度学习/图神经网络/因果推断/结构增强/漏洞检测Key words
deep learning/graph neural networks/causal inference/structural reinforcement/vulnerability detection分类
信息技术与安全科学引用本文复制引用
司文,赵富成,李硕,杨帅林,任家东..基于GNN因果推断的结构增强漏洞检测模型[J].燕山大学学报,2025,49(4):309-318,10.基金项目
国家自然科学基金资助项目(62376240) （62376240）
河北省自然科学基金资助项目(F2022203026,F2022203089) （F2022203026,F2022203089）
