基于注意力时间卷积网络的加密流量分类  

Encrypted Traffic Classification Based on Attention Temporal Convolutional Network

作　　者：金彦亮[1,2] 陈彦韬 高塬 周嘉豪 JIN Yanliang;CHEN Yantao;GAO Yuan;ZHOU Jiahao(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China;Shanghai Institute for Advanced Communication and Data Science,Shanghai University,Shanghai 200444,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444 [2]上海大学上海先进通信与数据科学研究院,上海200444

出　　处：《应用科学学报》2024年第4期659-672,共14页Journal of Applied Sciences

基　　金：上海市自然科学基金(No.22511103202);上海市产业项目(No.XTCX-KJ-2022-68)资助。

摘　　要：针对目前大多数加密流量分类方法忽略了流量的时序特性和所用模型的效率等问题,提出了一种基于注意力时间卷积网络(attention temporal convolutional network,ATCN)的高效分类方法。该方法首先将流量的内容信息与时序信息共同嵌入模型,增强加密流量的表征;然后利用时间卷积网络并行捕获有效特征以增加训练速度;最后引入注意力机制建立动态特征汇聚,实现模型参数的优化。实验结果表明,该方法在设定的两项分类任务上的性能都优于基准模型,其准确率分别为99.4%和99.8%,且模型参数量最多可降低至基准模型的15%,充分证明了本文方法的先进性。最后,本文在ATCN上引入了一种基于迁移学习的微调方式,为流量分类中零日流量的处理提供了一种新颖的思路。Aiming at the problem that most current encrypted traffic classification meth-ods ignore the timing characteristics in the traffic and the model efficiency,we propose an efficient classification method based on attention temporal convolutional network(ATCN).This method first embeds content information and timing information into the model to enhance the representation of encrypted traffic.Then it utilizes temporal convolutional network to capture effective features in parallel to increase training speed.Finally,we introduce attention mechanism to establish dynamic feature aggregation to optimize model parameters.Experimental results show the superior performance of our proposed method over the baseline in two classification tasks,achieving accuracy of 99.4%and 99.8%,respectively,while reducing the number of model parameters to a maximum of 15%of the baseline.Finally,a fine-tuning method based on transfer learning is introduced to the ATCN,which provides a novel approach for zero-day traffic processing in traffic classification.

关 键 词：加密流量分类 时间卷积网络 注意力机制 迁移学习 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]