基于预训练与新型时序图神经网络的智能合约漏洞检测方法

Smart contract vulnerability detection method based on pretraining and novel timing graph neural network

针对现有深度学习漏洞检测方法对合约字节码特征挖掘不足、漏洞语义表征不精准,且传统图神经网络模型对合约语句的时序信息学习能力不足,提出一种基于预训练与时序图神经网络的智能合约漏洞检测方法。首先,通过预训练模型将智能合约字节码建模为漏洞语义感知的合约图结构。其次,结合自注意力机制,设计了一种新颖的基于事件驱动的时序图神经网络模型,实现对合约执行中时序信息的有效抽取。最后,聚焦于可重入漏洞、时间戳依赖漏洞以及Tx.origin身份认证漏洞,通过120932份真实合约数据集进行大量的评估实验,结果表明所提方法的检测效果显著优于现有方法。

To address the limitations of current deep learning-based methods in extracting contract bytecode features and representing vulnerability semantics,as well as the shortcomings of the traditional graph neural networks in learning temporal information from contract statements,a method for detecting vulnerabilities in contracts was proposed based on pre-trained and temporal graph neural network.Firstly,the pre-trained model was used to transform smart contract bytecode into a vulnerability semantics-aware contract graph structure.Then,combined with a self-attention mechanism,the event-driven temporal graph neural network was designed to extract temporal information during contract execution.Finally,focusing on reentrant vulnerabilities,timestamp dependency vulnerabilities,and Tx.origin authentication vulnerabilities,extensive experiments were conducted on a dataset of 120932 actual contracts.The results show that the proposed method significantly outperforms existing approaches.