数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (6): 111-123.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.06.011

doi: 10.11871/jfdc.issn.2096-742X.2025.06.011

• 技术与应用 • 上一篇    下一篇

基于LoRA微调大模型在网络安全等级保护测评报告质量检测中的应用研究

吴建华1,2,*(),刘振宇1,2,曾睿1,2,王文瑄1,2,易勇1,2,王世轶1,2   

  1. 1.上海市计算机软件评测重点实验室，上海 201112
    2.上海计算机软件技术开发中心，上海 201112
  • 收稿日期:2025-02-14 出版日期:2025-12-20 发布日期:2025-12-17
  • 通讯作者: 吴建华
  • 作者简介:吴建华，上海计算机软件技术开发中心网络安全技术研究所所长，硕士，高级工程师，主要研究方向为网络安全，测试技术研究。
    本文负责确立论文选题、制定论文框架与修改论文。

Research on the Application of Fine-Tunned Large Language Models Based on LoRA in Quality Evaluation of the Security Level Protection Assessment Reports

WU Jianhua1,2,*(),LIU Zhenyu1,2,ZENG Rui1,2,WANG Wenxuan1,2,YI Yong1,2,WANG Shiyi1,2   

  1. 1. Shanghai Key Laboratory of Computer Software Testing Evaluating, Shanghai 201112, China
    2. Shanghai Development Center of Computer Software Technology, Shanghai 201112, China
  • Received:2025-02-14 Online:2025-12-20 Published:2025-12-17
  • Contact: WU Jianhua
  • About author:WU Jianhua is the director of the Cybersecurity Research Institute at the Shanghai Computer Software Technology Development Center. He holds a master's degree and is a senior engineer. His main research areas include cybersecurity and testing technology.
    In this paper, he is responsible for determining the research topic, formulating the paper framework, and revising the paper.
    E-mail: wjh@sscenter.sh.cn

RichHTML

3

PDF

0

摘要：

【目的/意义】网络安全等级保护测评对提升网络系统的安全性、满足合规要求以及促进持续改进具有非常重要的意义。但是，测评记录的错误、矛盾等情况在报告中时有发生，严重的甚至会影响到报告的结论。【方法】因此，本文创新性地提出将基于LoRA微调的大语言模型应用于等保报告审核。本文先将检测需求转换为一个基于逻辑推理和自然语言理解的多分类任务，然后据此构建微调数据集，并基于该数据集对glm-4-9b-1m-chat、chatglm3-6b与deepseek-r1-distill-qwen-7b大模型进行微调。【结果】实验结果显示，微调后的大模型与基座模型相比，其Bleu-4和Rouge-1值有大幅提升，并且多分类的准确率达到了87%。【结论】经过微调，大模型的输出质量更高，能较好地满足等级保护测评报告质量检测这一垂直领域的要求。本文提出的LoRA微调大模型为等级保护测评报告等类型的文档质量检测提供了新思路。

关键词: 大语言模型, LoRA微调, 网络安全, 文档质量检测, 等保测评

Abstract:

[Purpose/Significance] The network security level protection assessment is of great importance in improving the security of network systems, meeting compliance requirements, and promoting continuous improvement. However, errors and contradictions in evaluation records often occur in the reports, which can seriously affect the conclusions of the reports. [Methods] This paper innovatively proposes a method applying a fine-tunned large language model based on LoRA to the review of level protection reports. First, the detection needs are converted into a multi-classification task based on logical reasoning and natural language understanding. Then, a fine-tuning dataset is constructed based on this, and the glm-4-9b-1m-chat, chatglm3-6b and deepseek-r1-distill-qwen-7b large models are fine-tuned based on the dataset. [Results and Analysis] The experimental results show that the fine-tuned large model has a significant improvement in Bleu-4 and Rouge-1 values compared to the base model, and the accuracy of the multi-classification reaches 87%. [Conclusion] After fine-tuning, the output quality of the large model is improved and can better meet the requirements of quality inspection for graded protection reports in this vertical domain. The LoRA fine-tuning of the large model proposed in this paper provides a new perspective for quality evaluation of the level protection report documents and other types of documents.

Key words: large language models, LoRA fine-tuning, network security, document quality detection, level protection evaluation