Research On Network Traffic Classification Method Based On Transfer Learning

Accompanied by the high-speed progress of the era of big data,many network appliance products have occurred one after another,and the network structure has become increasingly complex.In the future,large-scale networks will present higher management requirements in terms of service quality and security.Understanding and discovering applications and their popular trends in the network can provide valuable information for network carriers to satisfy the demands of customers.Network traffic classification technology is recognized as one of the momentous technologies of network safety management and information skill safety.It improves the network configuration,reduces the risk of network security,and can according to user behavior data provide superior service for the corresponding network communication.In recent years,there have been plenty of solutions to the problem of network traffic classification,but there are still the following problems: The emergence of dynamic port and camouflage technology makes the classification of network traffic based on port number no longer applicable;Due to the limitations of laws and regulations and information security,most application protocols are not public,and the load based network traffic classification method is also limited;The network traffic classification algorithm on account of machine learning needs the training information and test information to satisfy the identical characteristic space and conditional probability distribution,but in the process of practical application,this prerequisite may not be met.At present,most network traffic classification methods are set for offline scenarios,which is not in line with the actual situation in which the network traffic data arrive in the form of“flow”.When the model is updated,the network traffic data needs to be retrained,which increases the waste of training time and related resources.To be directed against the weaknesses of traditional traffic classification methods,this paper introduces the transfer learning method to settle the complex problems of network traffic classification,and puts forward some innovative methods.The main innovations are as follows:(1)In order to settle the problem that the traditional network traffic classification method demands the consistent distribution of training and test information and sufficient training instance,which is difficult to achieve in natural utilization,a network traffic classification method based on similarity transfer is proposed.The case-based transfer learning theory is introduced to optimize the network traffic data from the two perspectives of characteristic attributes and similarity in the sample domain.The SAMME model is introduced to improve the weight update mechanism of Tr Ada Boost(boosting for transfer learning)algorithm to adapt it to multi-classification traffic tasks,and the fast convergence problem of source domain weight is solved by adding inhibition factors.The proposed network traffic classification method is checked on the data set Moore.The experimental results turn out that when the target sample size is insufficient,the proposed model can transfer selectively according to the similarity,effectively avoid the negative transfer problem,and has higher classification accuracy compared with the compared traffic classification model.(2)In order to settle the problem that traditional traffic classification methods can not realize online classification,an online transfer method of network traffic based on minimizing distribution difference is proposed.The method based on ensemble learning is introduced to train online classifiers in the source domain and target domain.In this model,the weight proportion is assigned to each classifier,and then the corresponding weight is dynamically adjusted according to the classification loss value;For the sake of decreasing distribution discrepancies between the source domain and the target domain online,Instead of each round of updating the transfer characteristic matrix of the target domain in actual time,and the time window is set to control the update rate of the matrix.The experimental results bear out that the suggested method can usefully diminish the spatial difference between realms online and dynamically adjust the matrix update frequency through the time window.Contrasted with the conventional online transfer learning technique,the mean average precision of network traffic classification can be promoted by 8.6%.