基于大语言模型推理增强的函数级二进制与源代码相似性检测OA
Function-Level Binary-to-Source Code Similarity Detection Method Enhanced by Large Language Model Reasoning
二进制与源代码之间的相似性检测在软件成分分析、漏洞定位等领域应用广泛.然而,现有方法在复杂编译环境下难以对齐不同模态的语义,导致检索效果较差.为此,提出基于大语言模型推理增强的相似性检测框架BinThink.BinThink利用大语言模型构建二进制函数和源代码函数的统一语义表示,并结合语义引导机制将不同模态的函数映射到同一向量空间;同时,利用对比学习优化跨模态语义对齐能力,提升复杂编译环境下的函数检索效果.实验结果表明,在候选池大小为100时,BinThink的Recall@1相较于现有方法BinaryAI,在跨编译优化与跨指令架构的函数检索任务中分别提升7.8和8.5个百分点.
Binary-to-source code similarity detection is widely used in software component analysis and vulnerability localization.However,existing methods struggle to align semantics across different modalities in complex compilation environments,leading to poor retrieval performance.To address these challenges,a similarity detection framework enhanced by Large Language Model reasoning(Bin-Think)is proposed.A Large Language Model is utilized in BinThink to construct a unified semantic representation of binary and source code functions,and a semantic guidance mechanism is combined to map functions from different modalities to the same vector space.Simultaneously,contrastive learn-ing is leveraged to optimize cross-modal semantic alignment,improving function retrieval performance in complex compilation environments.Experimental results show that,when the candidate pool size is 100,BinThink achieves Recall@1 improvements of 7.8 and 8.5 percentage points compared with the existing method BinaryAI in function retrieval tasks across compilation optimization and instruction ar-chitectures,respectively.
常金为;孙佳佳;尹小康;张一筝;朱肖雅;李路凯
信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001
信息技术与安全科学
二进制与源代码相似性检测大语言模型函数级推理增强自适应推理标准化语义流
binary-to-source code similarity detectionLarge Language Modelfunction-level reason-ing-enhancementadaptive reasoningstandardized semantic flow
《信息工程大学学报》 2026 (4)
436-442,7
评论