基于MHA与SDAE的Tor网站指纹识别模型  

Identifying Tor Website Fingerprinting Model Based on MHA and SDAE

作　　者：蒋首志 曹金璇[1] 殷浩展 芦天亮[1] JIANG Shouzhi;CAO Jinxuan;YIN Haozhan;LU Tianliang(School of Information Network Security,People's Public Security University of China,Beijing 10003&China)

机构地区：[1]中国人民公安大学信息网络安全学院,北京100038

出　　处：《信息网络安全》2022年第10期8-14,共7页Netinfo Security

基　　金：国家自然科学基金[61602489];中国人民公安大学基本科研业务费[2020JKF101]。

摘　　要：为解决Tor网站指纹识别技术在开放世界准确率低及概念漂移问题,文章提出一种基于MHA与SDAE的网站指纹识别模型一MHA-SDAE-GRU。首先将网站流量处理成序列格式;然后利用多头自注意力机制捕获输入数据的关键信息,并用堆叠降噪自编码器学习流量中的深层特征,增强模型的鲁棒性,通过GRU学习序列的前后关系;最后用Softmax函数输出结果。实验结果表明,MHA-SDAE-GRU模型在封闭世界的准确率高于CUMUL等算法,在开放世界的准确率和鲁棒性均优于CNN等算法,在概念漂移实验中对新数据的适应性优于CNN等算法。MHA-SDAE-GRU模型在网站指纹识别上具有优秀的表现。This paper aims at addressing the poor performance of identification technology in open world and the issue of concept drift by developing a new method to identify Tor website fingerprinting based on MHA and SDAE.First,this paper processed website traces into sequence form and extracts essential information of input data with muti-head attention,then the robustness was enhanced via learning deep features of traces with denoising autoencoder.The results were output by using Softmax after learning sequence relation with GRU.The results of experiments presents that accuracy of MHA-SDAE-GRU model in closed world is higher than CUMUL algorithm,accuracy and robustness in open world are better than other algorithms and adaptability to new data in concept drift experiments is better than the others.MHA-SDAE-GRU model plays an effective role in identifying tor website fingerprinting.

关 键 词：网站指纹 多头注意力 堆叠降噪自编码器 循环神经网络 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]