PRIMEVUL
收藏资源简介:
PRIMEVUL是由哥伦比亚大学、华盛顿大学及谷歌DeepMind等机构联合创建的一个新的安全漏洞检测数据集,旨在提升代码语言模型(Code LMs)在实际漏洞检测中的性能评估。该数据集包含6,968个易受攻击的函数和228,800个良性函数,覆盖了140种常见的缺陷枚举(CWEs)。PRIMEVUL采用了一套新颖的数据标注技术,实现了与人工验证基准相当的标注准确性,同时显著扩大了数据集规模。此外,PRIMEVUL还实施了严格的数据去重和按时间顺序划分数据的策略,以缓解数据泄露问题,并引入了更现实的评估指标和设置。此数据集不仅为研究人员提供了更具挑战性的评估平台,也揭示了现有模型在应对现实世界漏洞时的不足,凸显了在安全领域加强Code LMs训练的紧迫性。
PRIMEVUL is a novel security vulnerability detection dataset jointly created by institutions including Columbia University, the University of Washington, and Google DeepMind, aiming to improve the performance evaluation of code language models (Code LMs) in real-world vulnerability detection. This dataset contains 6,968 vulnerable functions and 228,800 benign functions, covering 140 common Common Weakness Enumerations (CWEs). PRIMEVUL adopts a novel set of data annotation techniques, achieving annotation accuracy comparable to human-validated benchmarks while significantly expanding the dataset scale. Additionally, PRIMEVUL implements strict data deduplication and chronological data partitioning strategies to mitigate data leakage issues, and introduces more realistic evaluation metrics and experimental settings. This dataset not only provides researchers with a more challenging evaluation platform, but also reveals the shortcomings of existing models when addressing real-world vulnerabilities, highlighting the urgency of strengthening the training of Code LMs in the field of cybersecurity.
数据集概述
数据集名称
- 名称: PrimeVul
- 目的: 用于训练和评估代码语言模型在现实世界漏洞检测场景中的表现。
数据集特点
- 数据多样性: 包含约7,000个漏洞函数和约229,000个良性函数,来源于真实的C/C++项目,覆盖超过140种常见弱点列举(CWEs)。
- 标签准确性: 采用创新的标签技术,达到人类级别的标签准确性,比现有自动标签方法准确度高出3倍。
- 数据纯净性: 通过彻底的数据去重和时间序列数据分割,最小化数据污染。
评估方法
- VD-Score: 衡量在不使开发者被误报淹没的情况下,遗漏安全缺陷的风险。
- 配对样本分析: 分析模型捕捉细微漏洞模式的能力,以及区分漏洞代码与其修复版本的能力。
数据集获取
- 获取方式: Google Drive链接




