Machine Learning for Cross-Vulnerability Prediction in Smart Contracts

Abstract

Smart contracts are programs that are deployed on blockchain to automate the agreements among users. It is hard to fix a security vulnerability once a smart contract is deployed. Vulnerabilities in smart contracts have raised a lot of concerns. Many automated solutions are proposed to identify these vulnerabilities. Supervised machine learning algorithms require historical data to detect vulnerabilities. These solutions will not work if the historical data for a vulnerability in question is not available. To address this problem, we have proposed to use the historical data of another vulnerability. We identify similar features among the vulnerabilities and use them to build the models. The built models are used to test smart contracts with the type of vulnerability that does not have the historical data. We have conducted experiments on 4 datasets and found that worst case $F_{1}$-score is 73% and the best case is 93% and mostly F1-score was more than 80%.