Reproduce “Generalization-Enhanced Code Vulnerability Detection via Multi-Task Instruction Fine-Tuning” 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
Xiaohu Du, Ming Wen, Jiahao Zhu, Zifan Xie, Bin Ji, Huijun Liu, Xuanhua Shi, Hai Jin
A novel framework, VulLLM, applies instruction-tuned Large Language Models in a multi-task learning paradigm for code vulnerability detection, achieving enhanced generalization and robustness against adversarial attacks compared to prior state-of-the-art models.
Key insights
- Multi-task instruction fine-tuning compels Large Language Models to learn deep-seated, generalized vulnerability features rather than superficial correlations.
- Leveraging generative LLMs like GPT-4 with a Chain-of-Thought with Self-Verification mechanism can produce high-quality natural language interpretations of vulnerabilities, enriching training signals.
- Auxiliary tasks such as vulnerability localization and interpretation provide crucial, diverse learning signals that significantly improve overall detection generalization and robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.