Dyff-VD: Binary Code Vulnerability Detection Based on Dynamic Feature Fusion

Vulnerabilities lurking within binary code constitute a major threat to software security and need to be effectively detected. In recent years, under the influence of the wave of artificial intelligence, considerable advancements have been achieved in the research domain of binary code vulnerability detection based on deep learning. Nevertheless, constrained by the analytical complexity inherent in binary code, existing methods still encounter challenges in learning its syntax and semantics. In this paper, we propose a dynamic feature fusion BERT learning mechanism and, based on this, propose a binary code vulnerability detection method, Dyff-VD, which simultaneously learns the fused features on assembly code and pseudocode. We evaluate Dyff-VD on the NVD dataset, and the results indicate that Dyff-VD surpasses the state-of-the-art baselines by 3.41%, 5.91%, and 5.85% in terms of Fl-score, precision, and accuracy, respectively. Visualization further substantiates the superiority of Dyff-VD in vulnerability detection.