A Novel Method for Vulnerability Detection Based on Fusion and Hyperbolic Neural Network Graphs

Detecting vulnerabilities in source code is essential for maintaining cybersecurity in digital space. Recent research has highlighted the strong representational capabilities of Graph Neural Networks (GNNs) in modeling the structural features of source code. However, existing GNNs predominantly function in Euclidean space, which limits their ability to accurately represent features and adapt to complex structures, especially when dealing with source code that exhibits significant non-Euclidean properties. In this study, we introduce a novel approach that incorporates hyperbolic geometry into code vulnerability detection and present a feature learning framework situated in non-Euclidean space. The proposed method leverages the exponential growth property of hyperbolic space, which aligns well with the hierarchical features inherent in the tree-like structure of source code. By employing fusion graphs and hyperbolic GNNs (HGNNs), our approach enhances the logical and structural representation of source code, thereby improving the accuracy and comprehensiveness of vulnerability detection. Experimental results substantiate the effectiveness of the proposed method. The accuracy of the proposed vulnerability detection method has been significantly improved in comparisons with existing technologies, with enhancements ranging from 6% to 31%. Our technique achieves an accuracy of 87.34% on the Devign dataset utilizing a 3-layer HGNN with a curvature of c = 3.5 and an embedding dimension of 256, which is never reached by the existing methods. Given its superior accuracy in detecting vulnerabilities, this method is particularly beneficial for architectures based on smart contracts, blockchain-based Internet-of-Things systems, and blockchain-based cyber-physical systems.