Research on Intelligent Vulnerability Detection Model Based on Multi-Modality

In response to the limitations of existing deep learning-based methods, which typically utilize either the source code of smart contracts or the graphs generated from them for vulnerability detection, often failing to effectively incorporate the rich syntactic, semantic, and structural information within smart contracts, this paper proposes a novel hybrid deep learning model CEGCN. This model is capable of integrating the source code of smart contracts with the semantic graphs generated from them, thereby enhancing the performance of vulnerability detection in smart contracts. Firstly, we construct Abstract Syntax Trees (AST) from smart contract code and optimize them to generate Semantic Graphs (SG) containing rich structural information. We then perform word2vec embedding on both SGs and the source code. Next, we utilize Convolutional Neural Networks (CNN) and Residual Graph Convolutional Networks (RGCN) along with an Edge Attention (EA) module to extract features from the embeddings of the source code and SGs, respectively. Finally, we concatenate the extracted features and employ a classifier to identify whether functions are vulnerable. Experimental results demonstrate that our proposed CEGCN model significantly improves the accuracy, precision, recall, and F1 score of smart contract vulnerability detection.