智能家居信息安全知识图谱构建研究OA
Research on Construction of Smart Home Information Security Knowledge Graph
随着智能家居行业的发展和信息安全需求的增加,设计时需考虑的信息安全因素逐渐变得复杂.标准文本中虽包含大量智能家居信息安全的设计要求,但文本之间缺乏结构关系,无法进行有效切割来提升大语言模型的知识抽取性能.为此,针对标准文本提出了一种大语言模型结合按标题切割策略-提示工程知识抽取方法.将长文档转化为Markdown格式,通过TF-IDF方法提取领域词,进行文本过滤,并依据Markdown半结构化特征进行文本分割,构建数据层;结合专家知识与智能家居信息安全领域多源数据,构建场景-设备-部件与危害-事故-控制链路本体,形成概念层;使用大语言模型、按标题切割的文本分割策略和"领域-模板-需求"的提示工程框架进行信息抽取,构建实例层.实验结果表明,经过领域词进行文本过滤后的大语言模型需要处理的文本减少了近60%,大大降低了调用API的成本,提出模型的最佳F1分数达到83.1%,通过消融实验,验证了提出的按标题切割的文本分割方法与提示工程框架可以极大地保留文本语义关联的合理性与有效性,再通过对比实验,证明了大语言模型比传统深度学习UIE模型明显更适配所提出的抽取方法,性能提升显著,说明大语言模型结合该方法在长文本的复杂结构语境下具有明显的优势.构建的知识图谱以场景-设备-部件为基础,连接了分散的危害与事故,并从标准中抽取出了相关的应对措施,提供了可视化查询和事故致因分析功能,有助于智能家居在设计时识别和评估潜在风险,预防信息安全事故.
With the advancement of smart home technology and growing demands for information security,the complexity of security-related design considerations has significantly increased.Although standard documents contain extensive design requirements for smart home information security,their unstructured nature hinders effective text segmentation to optimize knowledge extraction performance for large language models.This paper proposes a knowledge extraction methodology integrating large language models(LLMs)with a title-based segmentation strategy and prompt engineering.The implemen-tation involves three phases:(1)Data layer construction through format conversion to Markdown,domain-specific term extraction via TF-IDF for text filtering,and semi-structured text segmentation leveraging Markdown features;(2)Concept layer development by establishing dual-domain ontologies connecting scenarios-devices-components with hazards-accidents-control measures through expert knowledge integration;(3)Instance layer creation employing LLMs with a"domain-template-requirement"prompt engineering framework guided by title-based segmentation.Experimental results demonstrate that domain-specific text filtering reduces processing volume by approximately 60%,significantly lowering API(application programming interface)costs.The optimal F1-score of 83.1%validates the methodology's effectiveness,with ablation experi-ments confirming the semantic preservation capability of the title-based segmentation and prompt engineering frame-work.Comparative analysis reveals LLMs' superior performance over traditional deep learning model UIE(universal information extraction)in handling long-text contexts with complex structures,indicating substantial compatibility improvements.The constructed knowledge graph establishes connections between dispersed hazards and accidents through scenario-device-component relationships while extracting relevant countermeasures from standards.It provides visualized query capabilities and accident causation analysis,assisting smart home designers in identifying and evaluating potential risks to prevent information security incidents.
杨跃翔;郑怀城;刘学文;涂新雨
中国矿业大学(北京)管理学院,北京 100083中国矿业大学(北京)管理学院,北京 100083中国矿业大学(北京)管理学院,北京 100083中国矿业大学(北京)管理学院,北京 100083
信息技术与安全科学
智能家居知识图谱大语言模型知识抽取信息安全
smart homeknowledge graphlarge language modelknowledge extractioninformation security
《计算机科学与探索》 2026 (1)
169-181,13
国家重点研发计划(2022YFF0607100)中央基本科研业务经费项目(552023Y-10371).This work was supported by the National Key Research and Development Program of China(2022YFF0607100),and the Central Basic Business Research Funding Project of China(552023Y-10371).
评论