Smart Contract Vulnerability Detection Based on Deep and Cross Network

Abstract

Smart contract is the key core part of the blockchain’s operation, which has the feature that it cannot be changed after deploying, resulting in the fact that the vulnerabilities associated with smart contracts cannot be fixed. Currently, deep learning has the best performance for detecting smart contract vulnerabilities. However, the common deep learning method treats the smart contract source code as a text sequence, ignoring the running logic and data flow of the program, and can not extract the constructed features. In this paper, a modular vulnerability detection model Link-DC is proposed. the model takes the contract graph constructed by smart contracts and the pattern features defined by expert knowledge as the original data input to capture richer feature information through stitching. Then, the deep and cross network is adopted to construct low-dimensional and sparse features into high-dimensional nonlinear features, thus improving the performance and efficiency of detection. Extensive experiments are conducted on smart contracts with vulnerabilities on the Ether and VNTChain platforms. The experiment results show that the high detection accuracy performance of the Link-DC model is obtained, where the accuracy of detection of re-entry vulnerability, timestamp-dependent vulnerability, and infinite loop vulnerability reaches 94.37%, 92.11%, and 85.29%, respectively.