A slice-level software vulnerability detection method based on hyperbolic gated graph neural network
Yu Liu , Bin Liu , Shihai Wang , Tengfei Shi , Haoran Li , Shudi Guo
As technological progress drives software complexity, traditional vulnerability detection can’t meet rising security demands. AI-driven detection is now a trend. Recent research indicates deep learning can automate code semantic analysis for proactive risk identification and better detection efficiency. In deep learning-based vulnerability detection, graph-based source code representation prevails, capturing code structures better than sequence-based methods. But it fails to model tree-like dependency relationships well, resulting in poor feature extraction. To solve this, we propose a Slice-based Hyperbolic Gated graph neural network for Vulnerability Detection, named SHGvd. It first extracts key semantic features through normalization, graph slicing and embedding. Then it employs a novel Hyperbolic Gated Graph Neural Network that utilizes the negative curvature of hyperbolic space to enhance hierarchical code relation capture.Moreover, with a trainable curvature strategy, it can dynamically adjust the hyperbolic space curvature according to data. Finally, it employs convolutional, pooling layers and multi-layer perceptrons for classification. Evaluated on Big-Vul, SHGvd outperforms existing models, achieving 80.55% accuracy and 81.37% F1-score, showing superior feature extraction.
Vulnerability detection / Source code representation / Graph neural network / Hyperbolic space
Higher Education Press
/
| 〈 |
|
〉 |
PDF
(4779KB)
