[论文] When LLM Decompilers Recompile More and Preserve Less

论文概要

研究领域: ML 作者: Chang Liu, Edward Raff, Kristopher Micinski 发布时间: 2026-09-04 arXiv: 2609.05370

中文摘要

反编译从编译后的机器码恢复高级源码，是漏洞检测和恶意软件分析等安全任务的基础。传统反编译器如Ghidra和Hex-Rays将无法解析的内容暴露为可见占位符，常输出不可编译或执行的伪代码；基于LLM的反编译器生成干净惯用C代码，现在几乎完全通过可重编译性和可重执行性评判：输出是否构建并通过附带的输入/输出测试。本文表明这些指标可能奖励错误路径：函数可能重编译并通过所有附带测试，但在其他合法输入上偏离；披露的漏洞可能从反编译代码中消失且无崩溃可见痕迹。现有测试套件均无法捕获这些失败。为解决此差距，提出Decompile-Diverge行为比较oracle，不依赖固定或手工测试：为每个函数合成驱动程序，从参考实现增长fuzzing语料库，并在相同输入上重新运行反编译代码以检测函数行为变化。在九个配置的八个系统上，通过所有附带测试的候选仍在我们输入语料库上与原始实现偏离：总体4.9%，单系统最高13%。在300个真实GitHub库函数和287个CVE相关函数上，可重编译性与行为一致性可能分离：最强精炼LLM将Ghidra构建率从75%提升至90%，而匹配率从74%降至62%；披露漏洞中多达十分之一在其输出中表现出崩溃缺失。源码级分析将偏离归因于引入的字段、类型、被调用者和守卫，替代了传统工具留下的可见未知。

原文摘要

Decompilation recovers high-level source from compiled machine code and serves as a foundation for security tasks such as vulnerability detection and malware analysis. Traditional decompilers like Ghidra and Hex-Rays expose whatever they cannot resolve as visible placeholders and often emit pseudocode that will not compile or execute; LLM-based decompilers produce clean, idiomatic C and are now judged almost entirely by recompilability and re-executability: whether the output builds and passes its shipped input/output tests. We show that these metrics can reward the wrong path: a function may recompile and pass every shipped test yet diverge on other legitimate inputs, and a disclosed vulnerability may disappear from the recompiled code with no visible trace of the crash. Neither failure i...

#论文 #arXiv #ML #小凯