Authors: Ankur Mishra
Addresses: Department of Computer Science and Engineering, Manonmaniam Sundaranar University, India
Abstract: Network traffic classification has been used for more than two decades for various applications, including QoS provisioning, anomaly detection, billing systems, etc. With the wide-spread adaptation of deep learning models in various fields, researchers also started adopting deep models for traffic classification task. However, the biggest challenge with deep models is that they need considerably larger training data in comparison with classical machine learning algorithms. In this paper, we propose a two-step training process that significantly reduces the number of labelled training samples. In the first step, we train a convolutional auto-encoder with an unlabelled and large public dataset. Then, in the second step, we transfer the models weights to a CNN model and we train the CNN model with only a few labelled samples. We compared our approach with two state-of-the-art methods and we showed that our approach outperforms.
Keywords: traffic classification; convolutional auto-encoders; convolutional neural networks; CNNs; deep learning; encrypted traffic identification.
International Journal of Information Systems and Management, 2020 Vol.2 No.2, pp.139 - 149
Received: 22 Nov 2019
Accepted: 18 Jun 2020
Published online: 07 Oct 2020 *