Reproduce “Multi-View Decompilation for LLM-Based Malware Classification” 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

Bercan Turkmen, Vyas Raina

This work introduces a multi-view decompilation strategy to improve the accuracy of Large Language Model (LLM)-based malware classification by feeding an LLM pseudo-C code generated by two different decompilers. The approach achieved F1-score improvements of up to 13.9 percentage points for some LLMs, primarily by increasing recall for malicious samples, and demonstrated that combining views significantly benefits smaller, lower-cost models.

Key insights

  • Decompilers often produce distinct outputs for the same binary, leading to different LLM predictions, which indicates their outputs offer complementary information.
  • Presenting multiple decompiler views to an LLM helps it identify more malicious behaviors that might be obscured or unclear in a single decompiler's output.
  • The multi-view approach significantly closes the performance gap between smaller, lower-cost LLMs and larger, more expensive models, making automated malware triage more practical and cost-effective.

Summary by alphaXiv.

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