Smart Contract Vulnerability Detection Based on Residual Dilated Convolution with Multi-Head Attention

Abstract

In recent years, deep learning has been widely applied in smart contract vulnerability detection due to its automatic feature extraction and strong generalization capabilities. However, existing methods still face challenges such as redundant information in graph structures, insufficient utilization of data flow information, and single-scale feature extraction. To address these issues, we propose a function-level smart contract graph representation, namely the Multi-relational Semantic Graph (MSG), which employs various types of data flow edges to represent data dependency information within contracts. Subsequently, we introduce a detection model, REA_DCN, which combines a Residual Multi-scale Dilated Convolutional Network with a Multi-head Attention mechanism to capture syntactic and semantic features in the MSG. The model comprises two key modules: the Residual Multi-scale Dilated Convolutional Network (RE_DCN) can extract node features from three different dimensions, while the Multi-head Attention Network (MEA) is utilized for edge feature extraction. Experimental results on real-world datasets demonstrate that the highest score of REA_DCN in terms of accuracy, precision, recall and F1 score exceeds 97%, proving its effectiveness and feasibility.