Balancing Complexity and Performance in Convolutional Neural Network Models for QUIC Traffic Classification

The upcoming deployment of sixth-generation (6G) wireless networks promises to significantly outperform 5G in terms of data rates, spectral efficiency, device densities, and, most importantly, latency and security. To cope with the increasingly complex network traffic, Network Traffic Classification (NTC) will be essential to ensure the high performance and security of a network, which is necessary for advanced applications. This is particularly relevant in the Internet of Things (IoT), where resource-constrained platforms at the edge must manage tasks like traffic analysis and threat detection. In this context, balancing classification accuracy with computational efficiency is key to enabling practical, real-world deployments. Traditional payload-based and packet inspection methods are based on the identification of relevant patterns and fields in the packet content. However, such methods are nowadays limited by the rise of encrypted communications. To this end, the research community has turned its attention to statistical analysis and Machine Learning (ML). In particular, Convolutional Neural Networks (CNNs) are gaining momentum in the research community for ML-based NTC leveraging statistical analysis of flow characteristics. Therefore, this paper addresses CNN-based NTC in the presence of encrypted communications generated by the rising Quick UDP Internet Connections (QUIC) protocol. Different models are presented, and their performance is assessed to show the trade-off between classification accuracy and CNN complexity. In particular, our results show that even simple and low-complexity CNN architectures can achieve almost 92% accuracy with a very low-complexity architecture when compared to baseline architectures documented in the existing literature.