Hierarchical Graph Feature Extraction Based on Multi-Information Contract Graph for Enhanced Smart Contract Vulnerability Detection

Abstract

With the development of deep learning, especially driven by advanced models such as Graph Neural Networks (GNN), smart contract vulnerability detection is gradually moving toward automation and intelligence. Although existing deep learning detection methods have improved the efficiency of vulnerability detection to some extent, they fail to fully explore and utilize the rich syntactic and semantic information in smart contracts and generally suffer from insufficient feature extraction. In this paper, we propose a new method for smart contract vulnerability detection that combines a Multi-Information Contract Graph (MIG) with a Hierarchical Graph Feature Extraction model (HGFE). MIG integrates key information such as control flow, data flow, and vulnerability feature flow within smart contracts, fully mining and utilizing the rich syntactic and semantic features of smart contracts, providing the model with comprehensive feature representation. HGFE applies a multilayer feature extraction strategy, combining global and local feature extraction, and comprehensively considers multiple dimensions of information within the contract graph, thereby fully extracting the features of the contract graph. The experimental results demonstrate that our method significantly enhances the ability to detect potential vulnerabilities in smart contracts, achieving a maximum accuracy and precision of 97.29% and 97.70%, respectively, outperforming other advanced methods.