VulTriage: Triple-Path Context Augmentation for LLM-Based Vulnerability Detection
Researchers have introduced VulTriage, a novel framework designed to enhance automated vulnerability detection in software security using Large Language Models (LLMs). While LLMs demonstrate strong code understanding, direct prompting often results in missed vulnerabilities or false alarms due to subtle semantic differences. VulTriage addresses this by employing a triple-path context augmentation strategy. The Control Path extracts and verbalizes structural data like AST, CFG, and DFG to reveal dependencies. The Knowledge Path retrieves relevant CWE-derived vulnerability patterns via hybrid retrieval. The Semantic Path summarizes functional behavior before final judgment. These contexts are integrated into a unified instruction to guide more reliable reasoning. Experiments on the PrimeVul pair test set indicate that VulTriage achieves state-of-the-art performance, surpassing existing deep learning and LLM baselines. Ablation studies confirm the efficacy of each path, while tests on the Kotlin dataset demonstrate robust generalization in low-resource and class-imbalanced settings. The source code is publicly available, marking a significant advancement in applying AI to cybersecurity tasks.
Wire timeline
VulTriage: Triple-Path Context Augmentation for LLM-Based Vulnerability Detection
Researchers have introduced VulTriage, a novel framework designed to enhance automated vulnerability detection in software security using Large Language Models (LLMs). While LLMs demonstrate strong code understanding, direct prompting often results in missed vulnerabilities or false alarms due to subtle semantic differences. VulTriage addresses this by employing a triple-path context augmentation strategy. The Control Path extracts and verbalizes structural data like AST, CFG, and DFG to reveal dependencies. The Knowledge Path retrieves relevant CWE-derived vulnerability patterns via hybrid retrieval. The Semantic Path summarizes functional behavior before final judgment. These contexts are integrated into a unified instruction to guide more reliable reasoning. Experiments on the PrimeVul pair test set indicate that VulTriage achieves state-of-the-art performance, surpassing existing deep learning and LLM baselines. Ablation studies confirm the efficacy of each path, while tests on the Kotlin dataset demonstrate robust generalization in low-resource and class-imbalanced settings. The source code is publicly available, marking a significant advancement in applying AI to cybersecurity tasks.
cs.AI updates on arXiv.org