Encrypted traffic classification: the QUIC case

Abstract

The QUIC protocol is a new reliable and secure transport protocol that is an alternative to TLS over TCP. However, compared to TLS, QUIC obfuscates the connection hand-shake and the server name indication domain, making a simple inspection more challenging. The classification of QUIC traffic has also received less attention than that of TLS. In this work, we present a comprehensive study aiming to explore the challenges of QUIC traffic classification. We selected three models: 1) multi-modal CNN, 2) LighGBM, and 3) IP-based classifier, and evaluated their properties using a large one-month CESNET-QUIC22 dataset with 102 web service labels. The developed classifiers reached up to 88% accuracy and set the new baseline in fine-grained QUIC service classification. Moreover, the real nature of the dataset and its long time span allowed us to collect a number of insights and measure the classifiers' performance in the presence of data drift.