面向电力智能终端的多模态程序安全检测方法  

A security detection method for multi-modal applications incorporating external knowledge

作　　者：李尼格 陈牧 卢子昂 潘鸣宇[3] LI Ni-ge;CHEN Mu;LU Zi-ang;PAN Ming-yu(State Grid Smart Grid Research Institute Co.,Ltd.,Nanjing 210000,China;Power Network Security Protection and Monitoring Technology Laboratory,Nanjing 210000,China;State Grid Beijing Electric Power Company,Beijing 100000,China)

机构地区：[1]国网智能电网研究院有限公司,南京210000 [2]电力网络安全防护与监测技术实验室,南京210000 [3]国网北京市电力公司,北京100000

出　　处：《信息技术》2025年第3期76-85,共10页Information Technology

基　　金：国家电网有限公司科技项目(5108-202218280A-2-405-XG)。

摘　　要：在新型电力系统发展形势下,终端本体层面的安全检测已无法满足实际需求,下沉至应用程序级别势在必行。文中提出一种面向电力智能终端的多模态程序安全检测方法,利用TF-IDF算法和多层次编码网络提取程序API调用序列的统计和语义特征,从知识图中提取与程序汇编文件注释内容相关的外部知识特征。采用基于transformer的多模态特征融合方法实现不同模态特征的有效融合并联合外部知识特征输入LSTM模型进行程序安全检测。实验结果表明,文中方法具有较好的检测性能,准确率最高可达96.5%。Under the new power system development situation,the security detection at the terminal ontology level can no longer meet the actual demand,and it is imperative to sink to the application level.This paper proposes a multi-modal programme security detection method for power intelligent terminals.TF-IDF algorithm and multilevel coding network are used to extract the statistical and semantic features of program API call sequences,and the external knowledge features related to the annotated contents of program assembly files are extracted from the knowledge graph.A transformer-based multi-modal feature fusion method is used to achieve effective fusion of different modal features and combine the external knowledge features into the LSTM model for program security detection.The experiment results show that the method in this paper has good detection performance,and the accuracy rate is up to 96.5%.

关 键 词：安全检测 外部知识 多模态 API调用序列 LSTM 

分 类 号：TN918.91[电子电信—通信与信息系统]