sGuard+ : Machine Learning Guided Rule-based Automated Vulnerability Repair on Smart Contracts.
Cuifeng Gao,Wenzhang Yang,Jiaming Ye,Yinxing Xue,Jun Sun
DOI: https://doi.org/10.1145/3641846
IF: 3.685
2024-02-08
ACM Transactions on Software Engineering and Methodology
Abstract:Smart contracts are becoming appealing targets for hackers because of the vast amount of cryptocurrencies under their control. Asset loss due to the exploitation of smart contract codes has increased significantly in recent years. To guarantee that smart contracts are vulnerability-free, there are many works to detect the vulnerabilities of smart contracts, but only a few vulnerability repair works have been proposed. Repairing smart contract vulnerabilities at the source code level is attractive as it is transparent to users, whereas existing repair tools, such as SCRepair and sGuard , suffer from many limitations: (1) ignoring the code of vulnerability prevention; (2) possibly applying the repair to the wrong statements and changing the original business logic of smart contracts; (3) showing poor performance in terms of time and gas overhead. In this work, we propose machine learning guided rule-based automated vulnerability repair on smart contracts to improve the effectiveness and efficiency of sGuard . To address the limitations mentioned above, we design the features that characterize both the symptoms of vulnerabilities and the methods of vulnerability prevention to learn various vulnerability patterns and reduce false positives. Additionally, a fine-grained localization algorithm is designed by traversing the nodes of the abstract syntax tree, and we refine and extend the repair rules of sGuard to preserve the original business logic of smart contracts and support new vulnerability types. Our tool, named sGuard+ , reduces time overhead based on machine learning models, and reduces gas overhead by fewer code changes and precise patching. In our experiment, we collect a publicly available vulnerability dataset from CVE, SWC and SmartBugs Curated as a ground truth for evaluations. Overall, sGuard+ repairs more vulnerabilities with less time and gas overhead than state-of-the-art tools. Furthermore, we reproduce about 9,000 historical transactions for regression testing. It is shown that sGuard+ has no impact on the original business logic of smart contracts.
computer science, software engineering