Reproduce “Attention Distance: A Novel Metric for Directed Fuzzing with Large Language Models” 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
Wang Bin, Ao Yang, Kedan Li, Aofan Liu, Hui Li, Guibo Luo, Weixiang Huang, Yan Zhuang
Researchers from Peking University and collaborators introduce Attention Distance, a metric that integrates Large Language Model-extracted semantic insights with structural proximity to guide directed grey-box fuzzing (DGF). This approach improves vulnerability reproduction time by up to 3.43x on average over baseline AFLGo and enhances other state-of-the-art fuzzers.
Key insights
- Integrating LLM-derived semantic understanding directly into the core guidance mechanism of fuzzing significantly enhances the efficiency and direction of vulnerability discovery.
- Attention Distance creates a more 'layered and informative' distribution of path priorities, effectively alleviating the clustering of uniform distance values inherent in purely structural metrics.
- The LLM model can effectively recognize and prioritize vulnerability-associated code at function, basic-block, and line levels, providing fine-grained, semantically informed guidance for fuzzers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.