BinShoo: Binary Code Similarity Detection Using Optimized Function Embedding and Siamese Neural Network

Binary code similarity detection (BCSD) is important for cybersecurity, software analysis, and security research, helping to identify malware, vulnerabilities, and duplicate code as well as analyze attack patterns. Traditional methods, which are often limited when dealing with obfuscation, are being outperformed by deep learning (DL)-based techniques that can learn and compare code features efficiently regardless of transformations. In this work, we introduce BinShoo, a DL-based BCSD framework that uses the advanced Proc2vec+ method derived from the original Proc2vec method. Specifically, BinShoo converts binary functions into vector form for DL analysis, allowing the model to detect complex similarities between functions despite obfuscation, optimization, and other transformations. Experimental results on the BinaryCorp dataset show that BinShoo outperforms Bindeep in similarity detection across the following scenarios: i) same architecture, different optimizations and compilers, ii) same optimization, different architectures and compilers, iii) different architectures, optimizations, and compilers.