tCLD-Net: A Transfer Learning Internet Encrypted Traffic Classification Scheme Based on Convolution Neural Network and Long Short-Term Memory Network

Abstract

The Internet is about to enter the era of full encryption. Traditional traffic classification methods only work well in non-encrypted environments. How to identify the specific types of network encrypted traffic in an encrypted environment without decryption is one of the foundations for maintaining cyberspace security. Traffic classification based on machine learning relies heavily on the prior knowledge of experts to construct feature sets. Although traffic classification based on deep learning can reduce human intervention, it requires a large amount of labeled data for parameter determination. This paper proposes a tCLD-Net model that combines transfer learning and deep learning. It can be trained on a small amount of labeled data to distinguish network encrypted traffic with a high accuracy. It pre-trains a CLD-Net model in the source domain data set, and fixes the parameters of the convolutional neural network module in it, and trains and tests it in the target domain data set. In order to verify the effectiveness of the tCLD-Net model, we use the ISCX public data set to conduct experiments. The results show that our proposed model can complete 100 epoches training in 208 seconds when the training set only occupies 20% of the target domain. And achieve a classification accuracy rate about 86%. This is 4% higher than the model without pre-training, and the training time is only one third of the model without pre-training.