基于文本⁃视觉多特征融合的非法网站识别研究  

Research on illegal website identification based on textual⁃visual multi⁃feature fusion

作　　者：熊凌龙 何月顺[1] 陈杰 杜萍[1] 韩鑫豪 XIONG Linglong;HE Yueshun;CHEN Jie;DU Ping;HAN Xinhao(School of Information Engineering,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330013

出　　处：《现代电子技术》2024年第9期97-103,共7页Modern Electronics Technique

基　　金：江西省重点研发计划项目(20224BBC41001);江西省网络空间安全智能感知重点实验室开放基金资助项目(JKLGIP202206)。

摘　　要：当前非法网站存在隐蔽性强、危害性高的特点,仅依赖单一特征的网站识别方法无法有效应对这种复杂性。针对上述问题,文中提出一种基于文本⁃视觉多特征融合的非法网站识别方法。首先构建基于ResNet⁃18的视觉特征提取模型和基于BERT⁃CNN的文本特征提取模型;然后通过设计的基于逻辑回归(LR)的融合算法对两种模型的分类结果进行融合;最后通过多轮次迭代训练得出最佳的非法网站判别模型。实验结果表明,文中构建的融合模型相较于依赖文本和视觉的单一特征模型的准确率分别高出4%和11%,能够更准确地识别非法网站。At present,illegal websites have the characteristics of strong concealment and high potential for harm,and the website identification methods relying only on a single feature fail to cope with the complexity effectively.In view of this,an illegal website identification method based on textual⁃visual multi⁃feature fusion is introduced.Initially,a visual feature extraction model based on ResNet⁃18 and a textual feature extraction model based on BERT⁃CNN are established.Subsequently,a fusion algorithm based on logistic regression(LR)is applied to integrate the classification results of the two models.The final illegal website identification model is refined by multiple iterations of training.The experimental results show that the accuracy of the fusion model constructed is 4%and 11%higher than that of the model relying only on textual feature or visual feature,respectively,so the proposed model can identify illegal websites more accurately.

关 键 词：非法网站识别 多特征融合 BERT ResNet CNN 深度学习 

分 类 号：TN915.08-34[电子电信—通信与信息系统] TP391.1[电子电信—信息与通信工程]