关键词: 因果表示学习, 漏洞解释, 图神经网络, 伪相关性消除, 样本重加权, 独立性准则

Abstract: Vulnerability explanation is a critical task in software vulnerability mining,aiming to uncover the root causes of vulnerability formation and improve the quality of vulnerability remediation.Existing approaches for vulnerability explanation can be broadly categorized into factual reasoning and counterfactual reasoning methods.Factual reasoning focuses on identifying code subgraphs that are highly correlated with the model's prediction,while counterfactual reasoning seeks to discover subgraphs responsible for vulnerabilities by applying minimal perturbations that change the model's output.However,traditional methods lack explicit modeling of the causal relationships between code and vulnerabilities,making it difficult to distinguish spurious variables(spurious subgraphs)that are only statistically correlated with the vulnerability from the true causal variables(causal subgraphs)that directly lead to vulnerability formation.As a result,vulnerability explanation can be misled by spurious correlations,producing inaccurate conclusions.To address this issue,this paper proposes CausalVulGNN,a causal explanation framework for vulnerability analysis based on graph neural networks (GNNs).The proposed method firstly employs multi-layer GNN aggregation to capture structural and semantic features by integrating local node and edge information within the code graph.Then,it introduces a sample reweighting mechanism and a causal structure model(CSM)based on the Hilbert-Schmidt independence criterion(HSIC)to model causal relationships among high-level GNN representations.Unlike traditional correlation-based approaches,CSM simulates perturbations to high-level variables and quantifies their direct impact on model predictions,enabling the identification of subgraphs that are truly causal to vulnerabilities.This helps eliminate spurious subgraphs that are only statistically correlated with the output,thereby improving the fidelity of explanations.Experimental results demonstrate that CausalVulGNN achieves significant performance improvements across multiple vulnerability detection models and explanation methods,validating its effectiveness from a causal inference perspective.Systematic experiments are conducted on the real-world large-scale vulnerability dataset Big-Vul,covering popular detection models such as GCN,GGNN,GIN,and GraphConv.The results show that integrating CausalVulGNN leads to substantial improvements in standard explanation metrics(Accuracy,Precision,Recall,and F1-score)for widely used explainers including CFExplainer,GNNExplainer,and GNN-LRP.In particular,CFExplainer achieves an average 19.6% increase in Accuracy and over 28% improvement in Recall,confirming the effectiveness of CausalVulGNN for vulnerability explanation and analysis.

Key words: Causal representation learning, Vulnerability explanation, Graph neural networks, Spurious correlation elimination, Sample reweighting, Independence criterion