Research on a Binary Code Similarity Detection Method Based on Jump-ModernBERT
Changxu Duan, Qiang Wei · 2025
Code similarity detection techniques play a pivotal role in various security domains. However, existing binary code similarity detection methods often suffer from high computational overhead, limited accuracy, and incomplete recognition of function-level semantic information. To address these challenges, this paper proposes a binary code similarity detection method based on the Jump-ModernBERT model. Jump-ModernBERT incorporates two major innovations: (1) an alternating attention mechanism that improves the efficiency of processing long sequences, and (2) a jump recognition mechanism that enables the model to learn the graph-structured information of functions, thereby capturing semantic information more comprehensively. Experimental results demonstrate that, compared with BERT and SBERT, the proposed model achieves higher accuracy in detecting the functional similarity of source code, with recognition rates improved by 64.83% and 56.28%, respectively.