Smart Contract Vulnerability Detection Based on Multi Graph Convolutional Neural Networks with Self-attention

Smart contracts can contain vulnerable program code, making their security a significant concern in the past few years. Traditional detection techniques rely on expert defined features and source code analysis, but both have scalability issues and high false alarm rates. This paper presents a method for the detection of vulnerabilities in smart contracts, employing a multi graph convolutional neural network and self-attention mechanism. The method models the key function information, flow of control, and flow of data information of the smart contract source code as a semantic graph to emphasize the relationship between the flow of data and the flow of control in program operation. It predicts learning edges between nodes in the graph using an edge prediction network and constructs a multi graph of the smart contract semantics. Additionally, it utilizes a self-attention mechanism to gather and extract features from many layers to achieve precise detection of smart contract vulnerabilities. Outcomes of an experiment demonstrate that the proposed vulnerability detection method has significant advantages in identifying reentrant vulnerabilities and timestamp dependency vulnerabilities, with accuracy rates of 92.5% and 92.67%.