Source note
VulWeaver: Weaving Broken Semantics for Grounded Vulnerability Detection
Summary
VulWeaver is an LLM-based vulnerability detector that tries to fix broken static-analysis semantics before asking the model to reason about code. It combines an enhanced dependency graph, broader context extraction, and guideline-driven prompting to improve accuracy and reduce brittle pattern matching.
Problem
- Source-code vulnerability detection misses real semantics when static analysis builds incomplete or noisy call, control-flow, and data-dependency graphs, especially around reflection, polymorphism, and global definitions.
- Many learning and LLM systems analyze only local or one-directional slices, so they miss inter-procedural and implicit context such as sanitizer definitions or global configuration.
- Plain chain-of-thought prompting can key on variable names and other lexical hints instead of program behavior, which hurts reliability on patched code and adversarial renaming.
Approach
- VulWeaver builds a unified dependency graph from Joern outputs by merging call, control-flow, and data-dependency relations at statement level.
- It then enhances that graph with a neuro-symbolic method: deterministic rules plus LLM inference to repair missing or spurious semantics, with explicit handling for Java polymorphic and reflective calls and added global nodes parsed with Tree-sitter.
- From the enhanced graph, it extracts a holistic vulnerability context for each sensitive API call by combining explicit slices with implicit context such as usage, definition, and declaration information that normal slicing misses.
- It maps sensitive APIs to vulnerability types and uses meta-prompting with expert-written, CWE-specific reasoning guidelines so the LLM follows a fixed analysis procedure over the extracted context.
- It runs multiple LLM queries and uses majority voting to stabilize predictions; the knowledge base covers 102 CWE types.
Results
- On PrimeVul4J, VulWeaver reports precision 0.81, recall 0.70, and F1 0.75.
- On the same dataset, it beats the best learning-based, LLM-based, and agent-based baselines by 23%, 15%, and 60% in F1, respectively.
- Its VP-S score is 0.58, which is 164% higher than the best baseline, indicating stronger separation between vulnerable code and patched counterparts.
- On C/C++ PrimeVul with minimal adaptation, it reaches F1 0.78 and still outperforms all baselines.
- In a robustness study with adversarial identifier renaming, standard CoT prompting lost 11% to 39% F1 and 44% to 1,050% VP-S across five LLMs, while meta-prompting kept F1 drops to 0% to 6% and VP-S reductions within 16%.
- In real-world use, VulWeaver found 26 true vulnerabilities across 9 Java projects, with 15 confirmed by developers and 5 assigned CVEs; an industrial deployment found 40 confirmed vulnerabilities.