| The Quick UDP Internet Connection(QUIC)protocol is a data transfer protocol proposed by Google to replace TCP.In order to have faster transmission speeds than TCP,QUIC introduces the new feature of 0-RTT,which improves the speed of establishing a connection by reducing the amount of unencrypted information exchanged during the handshake phase.In other words,the applicability of traffic classification methods based on headers or payloads is reduced,so how to classify QUIC traffic has become a hot issue in the current research.At present,in the field of network traffic classification,most of the existing classification models are single-task learning models,and for some tasks with a small number of labeled samples and a high characteristic dimension,the use of single-task learning models is prone to overfitting risks.Secondly,the single-task learning model often fails to achieve the best classification performance of all categories of traffic,and the recognition effect of one category may lead to a decrease in the recognition effect of the remaining categories.In addition,the complexity of the feature pattern varies from category to category,and the model needs to capture a feature pattern that can distinguish all categories,and its classification difficulty will increase significantly as the number of categories increases.In order to solve the above problems,this paper studies the traffic classification method based on multi-task deep learning,and the main research work is as follows:1.Traffic classification method based on soft parameter sharing multi-task learning.In order to further study the application of multi-task learning method in network traffic classification,this paper proposes a traffic classification method based on soft parameter sharing multi-task learning.This method determines the shared structure of the model through data-driven methods,designs auxiliary tasks and target tasks to learn in parallel,complements the learned domain knowledge,averages the noise difference of each,and can effectively reduce the risk of poor overfitting and generalization ability of a single task.After ten-fold crossvalidation,the experimental results show that the proposed method has achieved good performance on the public QUIC traffic dataset,and the average accuracy of the application identification,bandwidth demand prediction and duration prediction is better than that of the single-task learning method,and it is also better than the multi-task learning method based on hard parameter sharing.2.Traffic classification method based on hierarchical sharing multi-task learning.Aiming at the problem of improving the classification effect of all categories in a single traffic classification task at the same time,this paper proposes a multi-task learning method based on hierarchical sharing.This method divides the categories in a single task into multiple tasks,the categories with similar characteristic patterns are classified into the same task,and different tasks do not share the same hidden layers like other multi-task learning methods,but share different levels of network structure according to the characteristics of each task,forming a hierarchical structure,which is not only more targeted,but also reduces the classification difficulty of each task.In this paper,the experimental results show that the overall accuracy rate of the model not only reaches 98.49%,but also the precision,recall,and F1 value of most categories are improved compared with the single-task method. |
