Cross-Architecture Binary Code Similarity Detection Based on Hybrid Neural Networks

Binary code similarity detection is a technique that analyzes the semantic and structural features of binary code, compares it with known vulnerable code for similarity, and determines whether the tested code contains vulnerabilities. With the increasing complexity of firmware vulnerabilities in IoT devices, binary code similarity detection has become a core task in firmware security analysis. However, existing methods face challenges in accurately modeling binary code semantics across architectures(ARM/x86/MIPS) and compilers(GCC/Clang), resulting in low precision and accuracy. This paper proposes a novel hybrid model, CABech, which integrates a natural language processing module(DeBERTa-V3) with a Multi-Track Graph Convolutional Network(MTGCN) for semantic feature vector extraction. The model also incorporates Low-Rank Adaptation(LoRA) technology into the NLP module to reduce training parameters and computational costs. DeBERTa- V3 enhances semantic modeling of binary code through its decoupled attention mechanism.MTGCN mitigates the over-smoothing issue in traditional GCN by separating structural features of binary code via multi-track message passing. The low-rank adaptive matrix fine-tunes the parameters of DeBERTa- V3,making the model suitable for resource-constrained firmware analysis terminals. Experiments demonstrate that our model outperforms state-of-the-art methods in cross-architecture and cross-compiler tasks, achieving improvements of 3.4% in accuracy, 4.0% in precision, and 1.9% in F1-score.