David-Egea/CWE-20-CFA

CWE-20-CFA是一个平衡的C/C++函数数据集，标记为良性或漏洞，针对CWE-20类别（不当输入验证）。每个记录包括原始源函数、代码属性图、图输入、标签、标识符和对抗性样本标志。数据集通过最小编辑的对抗性增强对来翻转漏洞标签，减少人为关联，并支持基于归因的可解释性。适用于代码漏洞检测、鲁棒性评估、解释性研究和源代码的图学习任务。

CWE-20-CFA is a balanced dataset of C/C++ functions labeled benign vs vulnerable under the CWE-20 category. Each record includes the original source function, Code Property Graph, graph inputs for GNN models, label, identifiers, and a flag for counterfactual provenance. The dataset was created by pairing minimally edited counterfactuals that flip vulnerability labels, reducing spurious correlations and enabling attribution-based explainability. It is suitable for code vulnerability detection, robustness evaluation, explainability research, and graph learning on source code.