CBS: A Deep Learning Approach for Encrypted Traffic Classification With Mixed Spatio-Temporal and Statistical Features

With the rapid advancement of the internet and online applications, traffic classification has become an increasingly significant topic in computer networks. Managing network resources, improving service quality, and enhancing cybersecurity are critical. Due to traffic encryption techniques, traditional traffic classification approaches have become ineffective and inaccurate. Therefore, the scientific community considers deep learning a high-performance approach for classifying encrypted traffic. This paper proposes an encrypted traffic classification approach, CBS, based on a deep learning technique. CBS can classify encrypted traffic at two levels using 1D-CNN, attention-based Bi-LSTM, and SAE deep network models. The proposed model classifies traffic types and applications based on a comprehensive set of session and packet-level features. CBS accurately distinguishes traffic classes using spatial, temporal, and statistical features extracted from packet content relationships, temporal relationships between packets in a session, and statistical characteristics of a work session. A traffic data augmentation technique based on a GAN network is employed to mitigate the impact of data imbalance on traffic classes. The proposed platform's performance is evaluated on the public ISCX VPN-Non VPN 2016 dataset. The results demonstrate that the platform accurately and efficiently identifies applications and classifies encrypted traffic. Compared to state-of-the-art methods, the proposed traffic classification model improves precision by 21.3%, accuracy by 13.1%, recall by 18.11%, and F1 score by 19.79%.