An attack detection mechanism in smart contracts based on deep learning and feature fusion

Abstract

The rapid growth of Ethereum has spurred widespread adoption of smart contracts, enabling substantial financial transactions. Once deployed on the blockchain, smart contracts are immutable, rendering them unmodifiable even if vulnerabilities are present. In recent years, numerous attacks exploiting these vulnerabilities have caused significant financial losses. Although prior research has improved vulnerability detection in source code or bytecode before deployment, identifying attacks that exploit vulnerabilities during the execution phase after deployment remains a significant challenge. These challenges arise from the limited adaptability of predefined detection rules and an overreliance on opcode sequence names, which often neglects a comprehensive analysis of opcode sequence properties. In this study, we propose an advanced multidimensional feature fusion technique designed to detect attacks during the execution phase of smart contracts. By leveraging deep learning, our approach enhances detection accuracy through a comprehensive analysis of attack behaviors across four dimensions: operation objects, action behaviors, functional categories, and gas consumption. Extensive experiments demonstrate that our method achieves a detection accuracy of 97.21% and a weighted F1-score of 97.21%, confirming its effectiveness in identifying attacks.