Enhancing Code Vulnerability Detection via Vulnerability-Preserving Data Augmentation
summary
Claims a fine-grained detector that trains five binary classifiers for different CWE types and combines them by majority vote, built on an edge-aware GGNN and a dataset mined from GitHub containing 5 CWEs (counts: 44, 8351, 2672, 362—as given). Introduces a vulnerability-preserving augmentation with five mutation operators driven by forward/backward slicing to expand scarce classes. Reports higher F1 and lower false positives than static tools and several deep-learning baselines.
AI-written summary · about
chapters (11)
- 0:00Vulnerability detection enhancement
- 1:08Mixed vulnerability types problem
- 2:13FGWorld framework overview
- 3:20Vulnerability-preserving augmentation
- 4:43Five mutation operators
- 5:27Edge-aware GNN architecture
- 6:31Baseline comparison
- 7:59Mutation operators effectiveness
- 8:44Edge-aware GNN vs traditional GNNs
- 9:05Conclusion and future work
- 9:26Q&A
related talks
30mBoosting Vulnerability Intelligence: How Accurate CWE Mappings Transform ML Model Performance
21mState of Attack Surface Elimination in a World Plagued by Vulnerable Software
56mCVE Is The Worst Vulnerability Framework (Except For All The Others)
28mHow Do We Leverage CVE Root Cause Mapping and CWE Data to Prevent New Vulnerabilities?
31mValidating Vulnerability Analysis with Statistical Analysis of Metadata
57mVulnerability Root Cause Mapping with CWE
30mData Driven Security
01mEvolution of CVEs and Vulnerability Management