Reproduce “The Seeds of the FUTURE Sprout from History: Fuzzing for Unveiling Vulnerabilities in Prospective Deep-Learning Libraries” 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
Zhiyuan Li, Jingzheng Wu, Xiang Ling, Tianyue Luo, Zhiqing Rui, Yanjun Wu
The FUTURE framework introduces a fuzzing methodology to uncover vulnerabilities in newly developed deep-learning libraries, utilizing historical bug patterns from mature libraries and fine-tuning large language models for code conversion. It identified 148 bugs, including 10 CVEs, in Apple MLX, Huawei MindSpore, and OneFlow, while also exposing 7 previously undetected flaws in PyTorch.
Key insights
- Historical bug patterns and complex API sequences from mature DL libraries can be effectively converted and leveraged to discover vulnerabilities in new, nascent frameworks.
- Targeted fine-tuning of Large Language Models (LLMs) with minimal documentation and code examples from new libraries successfully overcomes the inherent knowledge lag, enabling high-quality code generation and conversion.
- A bidirectional "history to future and back" cycle allows vulnerabilities found in new libraries to provide feedback and uncover previously undetected bugs in older, established source libraries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.