Semantic-aware Graph Neural Network for Smart Contract Bytecode Vulnerability Detection

Abstract

In order to solve the problems of low detection accuracy and high false positive rate of traditional smart contract vulnerability detection methods and less consideration of bytecode level smart contract features in neural networks, a smart contract bytecode vulnerability detection method based on semantic perception graph neural network was proposed. First, in order to generate the control flow graph, the basic blocks divided by the smart contract bytecode were used as the nodes, and the call relationship between the basic blocks was extracted from the bytecode as the edges. Then, control flow graph is transmitted into the graph convolutional network for training to obtain the feature representation of the graph nodes; Afterwards, the contract bytecode instruction sequence is segmented, transformed into a word vector, embedded into a low-dimensional space and transmitted to a long short-term memory network for training. Then, the vector representation of bytecode semantic information was obtained. Finally, the generated node features and semantic features were spliced and transmitted to the full connection layer for dimensionality reduction. Combined with semantic information and node features, the vulnerability detection was carried out for smart contracts. The real smart contracts in public dataset were used for training and testing, and verified in two types of vulnerability classification datasets through traditional methods and artificial tags. The method proposed in this paper was compared with three traditional smart contract vulnerability detection tools and one smart contract vulnerability detection method based on neural network. The experimental results showed that the proposed network greatly improves the performance of network in terms of various indicators, and detects the contracts with vulnerabilities which are not detected by the other four methods. It shows that adding the bytecode semantic information to graph neural network can effectively improve the detection accuracy and reduce the false alarm rate.