Research Of Mobile Traffic Classification And Identification Based On Deep Learning Method

With the rapid development of mobile Internet and the increasing popularity of smart devices,network security issues involved in mobile traffic have gradually attracted people's attention.Traffic classification and identification is a major branch in the security field.Traffic classification and identification can help administrators analyze and manage the network more effectively.Besides,accurate application traffic identification will conduce to the prevention and response of user privacy leakage and network attacks.From the beginning of the Internet to the era of the Internet of Things,traffic classification and identification technology has developed to a more complex and more abstract stage.The proliferation of applications that randomly select port numbers and other issues have led to the inefficiency of port number-based classification methods.Currently widely used payload-based,statistical-based and behavior-based classification methods are also faced with the drawbacks of feature selection requiring a large amount of human involvement and low automation.Compared with traditional networks,mobile traffic has the characteristics of relatively single protocol,numerous applications,and wide use of third-party services and cloud services,which poses challenges for mobile traffic classification and identification research.Aiming at the above problems,this thesis studies the classification and identification of mobile traffic based on deep learning.It draws on the achievements in the field of computer vision and image identification in recent years and focuses on the image conversion method,automated feature extraction and classification methods.The main work and innovations of this paper include the following aspects:(1)According to the communication principle and traffic characteristics of mobile applications,the payload of the application layer protocol mainly used is studied,from which the valid information is extracted and analyzed.Then,a novel mobile traffic classification framework is designed based on CNN(Convolutional Neural Network).(2)The principle of generating traffic on Android mobile devices is studied.A dataset construction method based on user interface automatic test technology is proposed.This method can automatically simulate a series of common interactions between users and mobile devices and generate traffic for collection.On this basis,the real-world traffic generated by the mobile device is captured,and a mobile traffic data set of 100 applications is established.(3)According to the characteristics of the CNN and mobile traffic,two improved preprocessing methods are proposed.One is a method based on MCA(Multi-weight Conversion Algorithm),which draws on the results of computer vision and converts the traffic into data in the form of pseudo-image to classify and identify based on visual features.MCA fully considers the characteristics of traffic header,grading the conversion weights of different characters to enhance the features and reduce the confusion.The other method is called ETCC(Encrypted-Traffic-Compatible Conversion).The difference between this method and related research is that the encrypted traffic is taken into consideration.ETCC not only selects the message header but also uses some message entities to make the framework compatible with both encrypted and non-encrypted traffic while leveraging the efficiency of deep learning.(4)A CNN model suitable for mobile traffic is designed,and the parameters are tuned through comparison experiments using real-world data.Diverse experiments are conducted to evaluate the entire framework,and the results verify that it has good feature extraction ability and classification and ability for mobile traffic.