摘要
针对大多数多模态加密流量分类方法使用特征级联的方式进行多模态融合,无法最优地利用不同模态的互补性信息问题,提出了一种基于交叉注意力机制的多模态加密流量分类方法。首先,通过试验分析,选择流量的有效载荷、数据包到达时间和长度序列及统计信息作为3种模态;其次,设计了3条路径使用神经网络学习3种模态特征;最后,将学习的高维特征使用交叉注意力机制进行融合,并使用ISCX VPN和ISCX nonVPN数据集对模型进行训练和测试。结果表明,模型的宏平均F1值分别达到96.95%和96.59%,与当前4种比较优秀的方法相比,均有明显提升;在相同数据集下,本方法的宏平均F1值较级联方式提升了2.49%,证实了交叉注意力机制在融合模态间互补信息方面的有效性。
To address the problem that most multimodal encrypted traffic classification methods use feature cascading for multimodal fusion,which cannot optimally utilize the complementary information of different modalities,a multimodal encrypted traffic classification method based on cross-attention mechanism is proposed.First,the payload,packet arrival time and length sequences,as well as the statistical information of the traffic are selected as the three modalities through experimental analysis.Then,three paths are designed to learn the three modal features using neural networks.Finally,the learned high-dimensional features are fused using the cross-attention mechanism,and the models are trained and tested using the ISCX VPN and ISCX nonVPN datasets.The results indicate that the macro-averaged F1 values of the model reach 96.95%and 96.59%,respectively,which are significantly improved compared with the four current better methods;under the same dataset,the macro-averaged F1 value of the proposed method is improved by 2.49%compared with the cascade method,which confirms the effectiveness of the cross-attention mechanism in fusing the complementary information between modalities.
作者
田鑫
丁要军
TIAN Xin;DING Yaojun(Gansu University of Political Science and Law,Lanzhou Gansu 730070,China)
出处
《通信技术》
2024年第7期739-747,共9页
Communications Technology
关键词
加密流量分类
多模态融合
特征级联
交叉注意力机制
encrypted traffic classification
multimodal fusion
feature cascading
cross-attention mechanism
