Attention-based Machine Learning Model for Smart Contract Vulnerability Detection

Abstract

Ethereum attracts extensive attention due to its distinctive function of smart contract and decentralized applications (Dapps). Since the number of contracts on blockchain has increased vigorously, various security vulnerabilities come up. Researchers rely on static symbolic analysis method at first, and it seems to perform well in the accuracy of vulnerability detection. However, this method requires manual analysis in advance and it needs to traverse all the possible execution paths to find out the vulnerable ones. The deeper the path goes, the more time it costs to detect the contracts. This paper proposes an approach to detect smart contracts vulnerability on blockchain by using machine learning(ML) methods. This approach aims to build a general benchmark for new vulnerability detection in order to reduce the demand of expert manpower. Moreover, the high-speed-performance ML algorithm makes quick detection comes true. As long as we adjust the threshold of the model, it can work as a fast prefilter for the traditional symbolic analysis tools in further improvement of accuracy.