| 引用本文： |
-
庄园,徐宗一,李芊汐,孟少琛,张民,孙建国.基于大语言模型的智能合约漏洞检测与修复综述[J].信息安全学报,已采用 [点击复制]
- zhuangyuan,xuzongyi,liqianxi,mengshaochen,zhangmin,sunjianguo.A Review of Smart Contract Vulnerability Detection and Repair Methods Based on Large Language Modeling[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 随着去中心化应用的快速发展，智能合约的业务逻辑变得日益复杂，传统的基于程序分析的漏洞检测与修复方法面临着自动化程度低、检测准确率低、修复效果差的挑战。近年来，大语言模型（large language model, LLM）因其强大的代码语义理解能力与高质量代码生成能力，为解决现有挑战带来新思路。因此，对LLM在智能合约漏洞检测与修复领域的应用开展研究显得尤为重要。本文提出了基于LLM的智能合约漏洞检测与修复优化技术研究框架，分别从智能合约的漏洞检测和漏洞修复两方面进行梳理。在漏洞检测方面，将已有工作系统归纳为与传统程序分析方法结合的增强型策略和以大模型为核心的端到端策略。针对大模型在不同应用方法中的技术特点，将现有方法从应用场景、技术特点、漏洞覆盖度、性能表现等多维度进行对比分析。在漏洞修复方面，按照自动化程度分类并整理了智能合约漏洞修复工作，系统性对比了基于LLM的漏洞修复工作，探讨了在代码语义理解、漏洞定位、攻击路径分析以及补丁生成等关键技术方面的优势与局限性。此外，本文针对大模型在智能合约漏洞检测与修复领域中的常用数据集与评估指标进行了梳理分析，提出当前数据集构建与评估体系的主要特点以及存在的问题。最后，总结了现阶段LLM在智能合约漏洞检测与修复应用中面临的挑战，并给出未来展望。 |
| 关键词: 大语言模型 智能合约 漏洞检测 漏洞修复 信息安全 |
| DOI： |
| 投稿时间：2026-01-13修订日期：2026-06-22 |
| 基金项目:国家自然科学基金青年基金项目 |
|
| A Review of Smart Contract Vulnerability Detection and Repair Methods Based on Large Language Modeling |
|
|
|
| Abstract: |
| With the rapid development of decentralized applications, the business logic of smart contracts has become increasingly complex. Traditional vulnerability detection and remediation methods based on program analysis face challenges such as low automation, poor detection accuracy, and ineffective fixes. In recent years, large language model (LLM) have emerged as a promising solution to these challenges due to their robust code semantic understanding capabilities and high-quality code generation abilities. Therefore, researching the application of LLM in smart contract vulnerability detec-tion and repair has become particularly important. This paper proposes a research framework for optimizing smart con-tract vulnerability detection and repair using LLM, systematically organizing approaches from both vulnerability detec-tion and repair perspectives. For vulnerability detection, existing work is categorized into two strategies: enhanced ap-proaches combining traditional program analysis methods and end-to-end strategies centered on large models. Consider-ing the technical characteristics of large models across different application methods, existing approaches are compared and analyzed from multiple dimensions, including application scenarios, technical features, vulnerability coverage, and performance metrics. For vulnerability remediation, smart contract patching efforts are categorized and organized based on automation levels. LLM-based remediation work is systematically compared, exploring its advantages and limitations in key technologies such as code semantic understanding, vulnerability localization, attack path analysis, and patch gen-eration. Furthermore, this paper reviews commonly used datasets and evaluation metrics for large models in smart con-tract vulnerability detection and repair, identifying key characteristics and existing issues in current dataset construction and evaluation frameworks. Finally, it summarizes the challenges LLM applications face in smart contract vulnerability detection and repair at this stage and offers future directions. |
| Key words: large language model smart contract vulnerability detection vulnerability repair information security |
