CrossDeep: A Hybrid Approach For Cross Version Binary Code Similarity Detection

Binary code similarity detection (BCSD) is utilized in various critical areas such as malware detection, vulnerability search, software version control, reverse engineering, digital forensics, and ensuring software integrity in distributed systems. Current approaches to the BCSD problem typically involve comparing specific features between binaries based on their control flow graphs and computing embedding vectors of binary functions, often employing deep learning algorithms for solution. Typically, current solutions involve comparing syntactic features captured from binary code, relying on domain knowledge. In this paper, a hybrid solution CrossDeep is proposed, employing multiple features, to address the cross-version BCSD problems. Hybrid method combines raw byte analysis with deep learning algorithm i.e. Convolutional Neural Networks (CNN), are proposed for enhancing binary code similarity detection across different versions. The method was implemented and evaluated on a custom dataset comprising approximately 312,142 samples. The results indicate good performance in the cross-version BCSD field, with a recall reaching 98.88%, better than the cutting-edge static solutions.