CEGT: Smart contract vulnerability detection via Connectivity-Enhanced GCN-Transformer

Abstract

The deployment of smart contracts on blockchains is rising rapidly. Accurate detection of security vulnerabilities in smart contracts can significantly minimize property losses. However, most existing machine learning (ML)-based models for smart contract vulnerability detection models overlook the contract graph structures and sequence information, reducing detection effectiveness. This study presents a Connectivity-Enhanced GCN-Transformer (CEGT), a method for detecting smart contract vulnerability detection that integrates graph and sequence models to enhance vulnerability detection accuracy. We improve node connectivity by identifying additional paths between nodes in the graph and augment the representation capability of node features through an additional orthogonal transformation layer, which performs an orthogonal transformation on the weight matrix. Moreover, we designed a novel attention mechanism, termed the dynamic attention mechanism, based on the sequence model and inspired by the concept of dynamic routing in capsule networks. Such a dynamic attention mechanism within the sequence model is introduced to integrate structural and sequential information of smart contracts, thereby enhancing vulnerability in detection accuracy. Our experiments demonstrate that CEGT surpasses state-of-the-art methods in detecting Reentrancy, Timestamp dependence, and Integer overflow vulnerabilities, achieving F1 scores of 93.47%, 89.33%, and 91.27%, respectively. This enables us to achieve greater accuracy in detecting smart contract vulnerabilities, helping to identify potential risks, reduce security threats, and ensure the reliability and safety of blockchain applications.

Editor’s note: Open Science material was validated by the Journal of Systems and Software Open Science Board.