Seq2Path: a sequence-to-path-based flow feature fusion approach for encrypted traffic classification

Chengxi Jiang, Shijie Xu, Guanggang Geng, Jian Weng, Xinchang Zhang

Abstract: With the increasing awareness of user privacy protection and communication security, encrypted traffic has increased dramatically. Usually utilizing the flow information of the traffic, flow statistics-based methods are able to classify encrypted traffic. However, these methods require a large number of packets and manual selection of statistical features. In this paper, we propose a novel encrypted traffic classification method (Seq2Path), which fuses flow features by using path signature theory to translate feature sequences into a traffic path. Then, the statistical features of the traffic path are generated by computing its signature; and finally, these features are used to train a machine learning classifier. Our experiments on four datasets containing three types of traffic (HTTPS, VPN and Tor) show that Seq2Path achieves stable performance and generally outperforms state-of-the-art methods.