TACSan: Enhancing Vulnerability Detection with Graph Neural Network

With the increasing scale and complexity of software, the advantages of using neural networks for static vulnerability detection are becoming increasingly prominent. Before inputting into a neural network, the source code needs to undergo word embedding, transforming discrete high-dimensional text data into low-dimensional continuous vectors suitable for training in neural networks. However, analysis has revealed that different implementation ideas by code writers for the same functionality can lead to varied code implementation methods. Embedding different code texts into vectors results in distinctions that can reduce the robustness of a model. To address this issue, this paper explores the impact of converting source code into different forms on word embedding and finds that a TAC (Three-Address Code) can significantly eliminate noise caused by different code implementation approaches. Given the excellent capability of a GNN (Graph Neural Network) in handling non-Euclidean space data and complex features, this paper subsequently employs a GNN to learn and classify vulnerabilities by capturing the implicit syntactic structure information in a TAC. Based on this, this paper introduces TACSan, a novel static vulnerability detection system based on a GNN designed to detect vulnerabilities in C/C++ programs. TACSan transforms the preprocessed source code into a TAC representation, adds control and data edges to create a graph structure, and then inputs it into the GNN for training. Comparative testing and evaluation of TACSan against other renowned static analysis tools, such as VulDeePecker and Devign, demonstrate that TACSan’s detection capabilities not only exceed those methods but also achieve substantial enhancements in accuracy and F1 score.