DA-GNN: A smart contract vulnerability detection method based on Dual Attention Graph Neural Network

A smart contract is an automated computer program based on blockchain technology. In recent years, the security incidents of smart contracts have caused serious economic losses. However, existing smart contract vulnerability detection methods rely on fixed expert rules, resulting in reduced detection accuracy and scalability. Therefore, addressing the issues of low accuracy in traditional smart contract vulnerability detection methods and the insufficient feature extraction in neural network-based approaches for smart contracts, this paper introduces an intelligent contract vulnerability identification method, Dual Attention Graph Neural Network (DA-GNN). Firstly, DA-GNN transforms the operation code sequence of nodes in the smart contract Control Flow Graph (CFG) into a feature matrix of semantic features and relationships between nodes based on the five types of instructions we propose. Secondly, our proposed dual attention mechanism introduces node semantic features and relationship features between nodes into the GAT to achieve node embedding updates. The updated graph node information is fused through self-attention mechanism to obtain the graph features. Then, the classification and prediction of vulnerabilities are achieved through the classification module. Finally, we evaluated our method on 17,670 real smart contracts. The experimental results show that the precision in detecting integer overflow vulnerabilities, self-destruct vulnerabilities, and transaction sequence dependency vulnerabilities reaches 72.17%, 67.03%, and 73.66%, respectively.