Key Research On Network Traffic Identification Based On Deep Learning Approach

With the development of network technology,especially Mobile Internet,social production and daily life had been increasingly dependent on the network.At the same time,the maintenance of cyberspace security and cyber malicious attacks had been in the process of mutual game.Network attacks such as Trojans,computer worms,denial of service have become more and more frequent,which seriously affecting people's normal use of the network.Network traffic identification,as the basis of network security,plays an important role in ensuring the rational operation of the network and maintaining information security.On the one hand,through the accurate identification of traffic,we can reduce the unnecessary network connection and avoid the risk of network attacks.On the other hand,the network manager can more effectively allocate network resources through traffic identification,thus provide better network services.Network traffic identification technology experienced from simple to complex development process accompanied by the improvement of people's network security awareness.In order to reduce unnecessary blocking from security devices such as firewalls,more and more network applications use port reuse technology,resulting in port based method basically failed.For pattern matching based DPI and flow statistics based DFI,there are difficulties in labeling massive samples and extract the features manually.In addition,it is difficult to achieve a good balance between the real-time ability and the identification accuracy when facing large-scale network data.So it is difficult to meet the demand of high-speed network traffic identification with a single traffic identification technology.In view of the above problems,we focus on the study of traffic identification using deep learning approach.The research is based on the traditional traffic identification technology and the achievements of computer vision in recent years.We propose the vision-meaningful image transformation of network traffic,the automatic feature extraction and recognition method based on unsupervised algorithm,and the implementation of traffic identification system for high speed network.Our main contributions are as follow:1.We propose a traffic data conversion method based on visual feature.After analyzing of the communication principle of network application and the characteristics of network application traffic,we focuses on the characteristics of network application traffic based on the same application layer protocol,and proposes a traffic data conversion method,including the effective data extraction and flow image transformation,so that the computer can follow the visual thinking to analyze and identify traffic.On this basis,we collect real traffic data in campus network gateway and create an open imageset of mobile traffic data(IMTD17).2.A method of traffic identification based on variational autoencdoer network(VAEN)is proposed.We studied the principle of unsupervised feature extraction of autoencdoer in deep learning method,and the error tolerance of variational autoencdoer algorithm based on probability distribution.On the basis of the automatic extraction of traffic features,we introduced the nonlinear fitting of multi-layer perceptron and multi-type regression classification,and network traffic identification is realized by a two stages learning,i.e.,unsupervised feature extraction and supervised classification.The ability of the model to extract the traffic features is illustrated by experiments,and the accuracy of traffic identification is verified.3.A method of traffic identification based on two-dimensional convolutional perception network(2D-CPN)is proposed.The convolutional autoencdoer algorithm based on convolution network is studied,and a convolution perception network model with visual feature extraction capability is designed.The 2D feature of the original image is transformed into a high-level feature representation by convolution network,while preserving the local correlation of the input image and realizing the global sharing of weights.The multi-layer perceptron is used to establish the mapping from the high-level convolution feature to the latent features of the encoder,so as to realize traffic identification.4.To maximize the real-time and accuracy requirements of traffic identification,a traffic identification system for high-speed network environment is designed and implemented.The overall scheme design and the overall realization method of the system are given.The function modules of the system are elaborated by using the idea of stratification and modularization.Then the performance of the system is verified and analyzed respectively for the network application traffic,mobile app traffic and malware traffic.The experimental results show that the system can meet the requirements of real-time detection at the high-speed network,and has good practicability.