Quick expansion of network services diversity made network resource planning harder than ever. Different requirements of each service on the network, knowing network users' usage pattern and differentiated service to each traffic or preventing its propagation are fundamental reasons for traffic classification. On other hand extensive use of encryption and tunneling protocols on the networks made current methods unfunctional. Thus the network's quality of service management is affected by improper network traffic classification. Diversity of traffic and available encryptions for it requires a more complex modeling method which is capable of hidden pattern extraction. This research took advantage of Deep Learning methods to classify network traffic due to this level of uncertainty. The proposed method is trying to classify encrypted or tunneld network traffic by deploying Recurrent, Convulotional and Fully Connected Neural Networks with supervised training. Also to make the model as simple as possible, this research examined a range of preprocessing parameters which is involved in the training process like flow sample size, packet sample size, removal of source and destination address and ports and etc to make sure the optimal point for each one. Furthermore independence of classification from tunneling or encryption protocols is preserved by providing different encryption or tunneling on the same network service traffic with the same label in the training process. The results show that effects of preprocessing were previously unknown on network traffic classification accuracy. These preprocessing pipelines dramatically decreased required model parameters. Also this process helped the model’s input data to balance statistical features of packet header, packet payload and packets flow. After encryption of packet most of the statistical features in packets flow or packet headers preserve. So balancing these features to train Deep Neural Network improves final classification accuracy. This method unlike any other proposed methods can classify same traffic under different tunnels with the same label with an average up to 95% accuracy as well as having 30% less parameters to train than latest research. This accuracy and simplicity could lead to an effective role in today’s networks full of encryption and tunneling. Encrypted IP Traffic,Classificatio,Deep Learning Methods
