基于多粒度分层建模的恶意URL检测模型  

Malicious URL Detection Model Based on Multi-granularity Hierarchical Modeling

作　　者：肖军弼[1] 牟丹 XIAO Junbi;MU Dan(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao 266000)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,青岛266000

出　　处：《计算机与数字工程》2025年第2期505-509,共5页Computer & Digital Engineering

基　　金：中国高校产学研创新基金项目“SD-WAN网络中关键业务通信保障技术研究”(编号:2021FNA02007)资助。

摘　　要：恶意URL检测对于网络安全防护十分重要。针对传统机器学习中特征信息损失和现有深度学习方法上下文建模不充分的问题,论文提出了一种基于多粒度分层建模的恶意URL检测模型。该方法从字符和词汇两个特征粒度上进行建模,对于每个特征粒度先使用卷积神经网络建模局部上下文信息,引入注意力机制进一步建模上下文信息得到信息增强的特征表示,将特征多粒度建模和上下文分层建模相结合充分提取URL的特征表示来进行恶意检测。实验结果表明,论文模型的准确率达到98%,相较于现有方法在性能上有一定的提升。Malicious URL detection is very important for network security protection.To solve the problem of loss of feature in⁃formation in traditional machine learning and insufficient context modeling of existing deep learning methods,this paper presents a malicious URL detection model based on multi-granularity hierarchical modeling.This method models two feature granularities,character and vocabulary.For each feature granularity,firstly convolution neural network is used to model local context information,attention mechanism is introduced to further model the context information to get the enhanced feature representation,and the com⁃bination of feature multi-granularity modeling and context hierarchical modeling sufficiently extracts the feature representation of URL for malicious detection.The experimental results show that the accuracy of the model is 98%,which improves the performance of the existing methods.

关 键 词：恶意URL 分层建模 卷积神经网络 注意力机制 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]