Combining Graph Neural Networks with Expert Knowledge for Smart Contract Vulnerability Detection

Smart contract vulnerability detection draws extensive attention in recent years due to the substantial losses caused by hacker attacks. Existing efforts for security analysis heavily rely on rigid rules defined experts, which are labor-intensive and non-scalable. More importantly, expert-defined tend be error-prone suffer inherent risk of being cheated crafty attackers. Recent researches focus symbolic execution formal smart contracts detection, yet achieve a precise scalable solution. Although several methods have been proposed detect vulnerabilities contracts, there is still lack effort that considers combining patterns with deep neural networks. In this paper, we explore using graph networks expert knowledge detection. Specifically, cast rich control- data- flow semantics source code into graph. To highlight critical nodes graph, further design node elimination phase normalize Then, propose novel temporal message propagation network extract feature from normalized combine designed yield final system. Extensive experiments conducted all Ethereum VNT Chain platforms. Empirical results show significant accuracy improvements over state-of-the-art three types vulnerabilities, where our method reaches 89.15%, 89.02%, 83.21% reentrancy, timestamp dependence, infinite loop respectively.