Reproduce “From Coverage to Causes: Data-Centric Fuzzing for JavaScript Engines” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kishan Kumar Ganguly, Tim Menzies
Researchers at North Carolina State University developed a data-centric fuzzing approach for JavaScript engines, leveraging Large Language Models to automatically derive static and dynamic features from historical vulnerabilities. This method identifies and prioritizes high-risk inputs, leading to the discovery of critical V8 bugs up to 18 times faster than existing fuzzers while achieving high prediction precision and improving execution rates.
Key insights
- An iterative, LLM-driven process can effectively automate the generation of both static and dynamic features from vulnerability data, overcoming the limitations and effort associated with manual heuristic design.
- The combined use of static and dynamic features significantly enhances the predictive performance of a vulnerability guidance model compared to using either feature type in isolation, demonstrating a strong synergistic effect.
- A minimal subset of features (top 25%) can achieve near-optimal predictive performance, allowing for substantial efficiency gains in feature collection and fuzzer execution without compromising detection capability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.