Neural decompilation candidate selection closes gap on obfuscated binaries
Original title: Rethinking Neural Decompilation: Candidate Selection as the Hidden Bottleneck
Generate-then-select pipeline with execution-guided selection reaches 88.8% pass@1 on obfuscated inputs.
LLM4Decompile architecture analysis reveals ground-truth tokens lie within top-five predictions in 95.6% of positions. A generate-then-select pipeline ranks 100 candidates per input using static and execution-based selectors. Negative perplexity emerges as strongest static selector while execution-guided selection nearly closes oracle gap on ExeBench after fine-tuning on newly released ExeBench-Obfuscated dataset.
- pass@100 of 93% on HumanEval-Decompile
- 95.6% ground-truth tokens in top-five predictions
- 88.8% pass@1 after fine-tuning
- Contradicts prior claims about LLM4Decompile
From the original article
1 min read at arxiv.orgReverse engineering obfuscated binaries is a core task in malware analysis and software security auditing.
By Felix Mächtle · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.