Scholarlink

<p>With the rapid advancement of blockchain technology, the security of smart contracts has become increasingly critical. Most existing automated tools focus on detecting vulnerabilities based on fixed patterns, making it difficult to identify complex business logic flaws—over 80% of which remain undetected. To address this gap, we propose DeepDesc, a novel framework for smart contract vulnerability detection that integrates Retrieval-Augmented Generation (RAG) with large language models (LLMs) and deep learning. DeepDesc employs GPT-3.5-turbo enhanced by RAG to generate high-quality code descriptions using 1,201 real-world vulnerability reports from Code4Rena. These descriptions, combined with the code, are used to train a deep learning classifier using Multimodal Low-rank Bilinear (MLB) pooling. Our method effectively handles both machine-auditable bugs (MABs) and machine-unauditable bugs (MUBs). Experimental results show that DeepDesc achieves an average F1 score of 1.00 for MABs and 0.90 for MUBs, significantly outperforming existing techniques.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DeepDesc: integrating retrieval-augmented generation with large language models for smart contract vulnerability detection

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Tao Tan,
  • Xiao Chen

摘要

With the rapid advancement of blockchain technology, the security of smart contracts has become increasingly critical. Most existing automated tools focus on detecting vulnerabilities based on fixed patterns, making it difficult to identify complex business logic flaws—over 80% of which remain undetected. To address this gap, we propose DeepDesc, a novel framework for smart contract vulnerability detection that integrates Retrieval-Augmented Generation (RAG) with large language models (LLMs) and deep learning. DeepDesc employs GPT-3.5-turbo enhanced by RAG to generate high-quality code descriptions using 1,201 real-world vulnerability reports from Code4Rena. These descriptions, combined with the code, are used to train a deep learning classifier using Multimodal Low-rank Bilinear (MLB) pooling. Our method effectively handles both machine-auditable bugs (MABs) and machine-unauditable bugs (MUBs). Experimental results show that DeepDesc achieves an average F1 score of 1.00 for MABs and 0.90 for MUBs, significantly outperforming existing techniques.