Scholarlink

<p>The advancement of Web3.0 technology has brought about a urgent need for ensuring the safety and reliability of the software systems. Program analysis, a crucial aspect of software security research need a unified solution for various cross-language program. Moreover, the previous studies regard the necessity of capturing structured features from the ASTs, commonly holding a conception that plain text and compiled binary instructions are challenging to represent and identify. This paper proposes three methods for extracting structural information to demonstrate that the underlying principles of source code and instructions align with those of an abstract syntax tree. These methods include (1) embedding program instruction files directly into natural language, (2) embedding formatted source code into natural language, and (3) embedding misformatted, non-compilable source code into natural language. We train large-scale language models (LLMs) to identify defect and non-defect of WebAssembly. Experiment results demonstrate that program instruction analysis surpasses traditional techniques, achieving state-of-the-art accuracy exceeding 98.1%. Our study also suggests a practical approach of plain text embedding using a 7.65&#xa0;billion LLM. Interestingly, misformatted source codes are readable to humans but un-compilable, and the accuracy remains above 98.63%. This paper not only introduces novel instruction and plain text embedding approach for future program security analysis, but also provides new insights for subsequent research about the three program analysis forms of plain text, ASTs, and instructions.</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

LLM-based program analysis for source codes, abstract syntax trees and webassembly instructions

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Liangjun Deng,
  • Qi Zhong,
  • Yao Qiu,
  • Jingxue Chen,
  • Hang Lei,
  • Shunkun Yang,
  • Liming Zhou,
  • Hongyuan Cheng

摘要

The advancement of Web3.0 technology has brought about a urgent need for ensuring the safety and reliability of the software systems. Program analysis, a crucial aspect of software security research need a unified solution for various cross-language program. Moreover, the previous studies regard the necessity of capturing structured features from the ASTs, commonly holding a conception that plain text and compiled binary instructions are challenging to represent and identify. This paper proposes three methods for extracting structural information to demonstrate that the underlying principles of source code and instructions align with those of an abstract syntax tree. These methods include (1) embedding program instruction files directly into natural language, (2) embedding formatted source code into natural language, and (3) embedding misformatted, non-compilable source code into natural language. We train large-scale language models (LLMs) to identify defect and non-defect of WebAssembly. Experiment results demonstrate that program instruction analysis surpasses traditional techniques, achieving state-of-the-art accuracy exceeding 98.1%. Our study also suggests a practical approach of plain text embedding using a 7.65 billion LLM. Interestingly, misformatted source codes are readable to humans but un-compilable, and the accuracy remains above 98.63%. This paper not only introduces novel instruction and plain text embedding approach for future program security analysis, but also provides new insights for subsequent research about the three program analysis forms of plain text, ASTs, and instructions.