VulTriNet: A software vulnerability detection method based on tri-channel network

Context:

Software vulnerabilities represent a critical concern in cybersecurity. As vulnerability patterns become increasingly complex, advanced detection methods are needed to fully analyze them. Recent studies have treated source codes as text using natural language processing (NLP) techniques. Subsequent advancements transformed programs into intermediate representations, utilizing graph neural network (GNN) for vulnerability learning. However, these approaches exhibit limitations in software vulnerability detection, as they fail to comprehensively analyze the features of source code.

Objective:

To solve this problem, we proposed a novel vulnerability detection method based on a tri-channel network (VulTriNet), which enables comprehensive analysis of source code and effective vulnerability detection.

Methods:

The Method integrates two graph-based and one textual code representation using three distinct methods to transform functions into multiple forms. Then, inspired by the RGB three-channel concept in the image domain, VulTriNet generates corresponding embedding vectors for these transformed representations, which are subsequently merged into a unified three-channel feature matrix. Finally, there is a CNN model integrated with attention mechanisms to improve the capability of detecting vulnerabilities.

Results:

Experimental results demonstrated that, compared to five state-of-the-art approaches, VulTriNet achieves, on average across different datasets: a 4.89% improvement in accuracy, a 3.41% increase in TNR, a 4.09% gain in TPR, and a 4.18% boost in F1-score.

Conclusion:

These results indicate that VulTriNet is more accurate and effective than previous studies. This hybrid analysis model strengthens vulnerability detection capabilities by simultaneously preserving contextual understanding of code and awareness of its structural relationships.