Research On User Actions Recognition Model Based On Mobile Network Traffic

In recent years,mobile intelligent devices(such as smart phones and tablets)have achieved more and more success in business,and have become an indispensable element in the daily life of billions of people around the world.Mobile devices are not only used for traditional communication activities,such as voice call and information communication,but also for a large number of multi-purpose applications,such as finance,gaming,video conference and online shopping.The daily online actions of mobile users not only bring the explosive growth of traffic,but also make the user actions complex and diverse.Therefore,in-depth analysis of mobile application traffic and identification of abnormal user action is of great significance to the security of mobile Internet.The main research contents of this thesis are as follows:1.In order to solve the interference problem of noisy traffic in the real mobile network environment,a mobile application user actions recognition model based on a nested pseudo-Siamese network is proposed.The model adopts the structure of a nested network.The inner pseudo-Siamese neural network is responsible for extracting the public traffic pattern of the specified user action from the pure network traffic without noise;the outer pseudo-Siamese neural network is responsible for comparing the real network traffic with the public traffic pattern,so as to quickly and accurately identify the specified user action.2.In order to reduce the workload of large-scale traffic collection and model training,a mobile application user action recognition framework based on transfer learning is proposed.Currently,the number of applications in active use reaches millions.It is not feasible to use traditional supervised learning methods to train a separate classifier for each application or the user action of each application.Therefore,this framework uses a network-based transfer leaning method to re-use part of the neural network structure and network parameters that have been trained,and use them for reference to another user action recognition classifier.Thus,the training time is greatly shortened,the accuracy of recognition is effectively improved,a large-scale of tag data collection is avoided,and the generalization performance of the mobile application user actions recognition model is enhanced.3.This thesis designs and implements a mobile application user action detection platform(UAMonitor),which can identify user operations on mobile devices by monitoring encrypted network traffic in a real mobile network environment,and detect abnormal user actions.The platform has a friendly interface,and realizes the functions of mobile terminal traffic collection,classification,and visual display,which can meet the functional requirements of real-time user actions detection.The research results of this thesis can be used for reference in the field of real-time detection of mobile network traffic and complex user action recognition.