Traffic Classification System Based On Deep Learning

With the rapid development of the Internet,network traffic has experienced explosive growth in both quantity and variety.Therefore,network traffic classification has become an indispensable part of network management,and it has greatly helped the network intelligent operation,management,network quality of service(QoS)and network security.Moreover,as the demands of user privacy and data encryption continue to increase,more and more encrypted traffic appears on the network.Due to the diversity of encryption algorithms and encryption types,it brings great difficulties to network traffic classification.Traditional traffic classification methods can't meet the need of classification accuracy.Encrypted traffic classification has become a huge problem for operators and service providers.Deep learning has achieved great advantages in tasks such as classification and detection.Therefore,this paper proposes to integrate deep learning methods into encrypted network traffic classification.The main work of this paper is as follows:(1)This paper proposed a network traffic representation method suitable for deep learning model.Each traffic flow is represented as an N*M dimensional matrix according to the hierarchical structure of“flow-packet-byte",which is fed as input to our model.(2)Two deep learning models are proposed in this paper,Attention Based Bidirectional Long Short-Term Memory Networks(ABBiLSTM)and Hierarchical Attention Based Bidirectional Long Short-Term Memory Networks(HABBiLSTM).(3)Experiments on the ISCX VPN-NonVPN dataset demonstrated that when the dimension of the traffic flow is N=20 and M=1000,the classification accuracy is the highest.In addition,through the four task scenarios experiments,this paper found that ABBiLSTM and HABBiLSTM are better than traditional methods.Especially ABBiLSTM model achieved 93.6%F1 value on the most difficult 12 classification tasks,which is much higher than other methods.(4)According to the attention mechanism in the model,the classification weights of different categories are visualized and analyzed and the contribution of each data packet to the classification degree is obtained,which increases the interpretability of the model.