VulnLens uses structured explanations for binary vuln detection
Original title: Structured Security Explanations for Learning-Based Vulnerability Detection
On a 526-function PrimeVul test set it reported 75.5% accuracy and 77.2% F1, ahead of DeepSeek V4 Pro and Gemini 2.5 Pro on those metrics.
VulnLens separates program-behavior analysis from the final vulnerability decision for binary C/C++ functions. A label-blind LLM explainer writes structured notes from seven fixed perspectives, and a supervised detector combines source code with those explanations through sequential cross-attention. The authors evaluate a PrimeVul-derived set covering ten CWE categories and test transfer to SecVulEval without retraining. They report gains over a code-only detector and over DeepSeek V4 Pro and Gemini 2.5 Pro on the published metrics.
Unconfirmed: Accuracy and F1 figures are author-reported and not independently confirmed.
- Matched explanations improved F1 by 7.3 points over a code-only detector.
- On SecVulEval, reported accuracy was 66.4% and F1 was 65.8%.
- The explainer does not receive vulnerability labels or CWE information.
- Transfer to SecVulEval was tested without retraining.
- Products
- VulnLens
- AI models
- DeepSeek V4 ProGemini 2.5 Pro
From the original article
1 min read at arxiv.orgSoftware vulnerability detection remains difficult because existing approaches rely on predefined analysis rules, learn vulnerability patterns implicitly from code and labels, or use large language models (LLMs) to make the final security judgment. It remains unclear whether structured descriptions of security-relevant program behavior can improve supervised detection.
By Fares Trad · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.