Reproduce “One-for-All Does Not Work! Enhancing Vulnerability Detection by Mixture-of-Experts (MoE)” 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

Xu Yang, Shaowei Wang, Jiayuan Zhou, Wenhan Zhu

MoEVD, a deep learning approach for vulnerability detection, integrates a Mixture-of-Experts framework to overcome limitations of single-model designs, achieving a 12.8% F1-score improvement over state-of-the-art baselines and a 31.2% recall boost for rare vulnerability types. This method enhances detection across diverse vulnerability types, particularly those with long-tailed distributions.

Key insights

  • The Mixture-of-Experts framework, with specialized experts and dynamic routing, is more effective for vulnerability detection than traditional ensemble methods.
  • The accuracy of the router in directing inputs to the correct expert is critical, with correct assignments leading to a 133.7% performance difference in F1-score.
  • MoEVD effectively mitigates the long-tailed distribution problem in vulnerability datasets by enhancing detection for less common CWE types without sacrificing performance on frequent ones.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.