Machine Learning Assisted Traffic Classification Of User Activities At Programmable Data Plane

The increasing popularity of network terminals and the continuous emergence of new network applications have not only led to exponential growth in network traffic,but also increased the complexity of user activities traffic classification.Group encryption and dynamic port technology have brought new challenges to traffic classification.Researchers have attempted to deploy machine learning to the control plane of programmable networks to ensure classification accuracy,but there are problems such as high latency and high communication overhead.With the development of programmable data plane technology,network devices such as switches,routers,and network cards have gradually become programmable,providing the possibility of real-time user activities traffic classification and reducing communication costs.However,offloading machine learning technology to programmable data planes still faces many challenges,specifically manifested as:(i)limited storage resources for storing flow tables in network devices;(ii)The computing resources used for tasks other than parsing and forwarding data packets in network devices are limited;(iii)At present,the mainstream data plane programming languages have limitations in supporting complex operations such as floating point numbers and matrices.This thesis mainly studies how to achieve accurate classification of user activities traffic in programmable networks,aiming to utilize the limited resource conditions of network devices to construct suitable machine learning classification models and deploy them on programmable data planes.The main work content and innovation points of this thesis are as follows:Firstly,a framework for classifying user activities traffic in programmable networks is proposed,and machine learning classification models are unloaded onto programmable data planes based on different model transformation methods.This framework can be deployed in general network environments and highly dynamic network environments to achieve accurate classification of user activities traffic in different network environments.Secondly,a machine learning assisted traffic classification of user activities scheme is proposed.Design a storage structure based on hash tables and Sketches to record network traffic feature data to reduce memory footprint; Design a classification model based on clustering algorithm and decision tree algorithm to infer user activity label within a time window.The hard coding method is used to implement the deployment of machine learning classifiers in the programmable data plane.Experiments show that the proposed scheme can reduce the occupation of network device memory space while ensuring classification accuracy.Thirdly,A machine learning assisted traffic classification of user activities scheme in dynamic network environment is proposed.In response to the highly dynamic problem of traffic in dynamic network environments,the trained classification model is transformed into corresponding feature spaces,and feature tables and decision tables are generated.The machine learning algorithm is unloaded onto the programmable data plane using a matching action table approach.Experiments were conducted under dynamically changing Internet of Things traffic conditions,and the results showed that the scheme achieved good classification performance while detecting abnormal traffic,and possessed adaptability in dynamic network environments.In summary,this thesis investigates the problem of user activities traffic classification in programmable networks and explores the introduction of machine learning algorithms into the data plane,which has positive significance in improving traffic classification capabilities in different network environments.