Fast and Accurate Flow Priority Detection in Encrypted Networks using Reinforcement Learning

In this paper, we tackle the network flow category and priority detection problem, exacerbated by the widespread adoption of encrypted protocols like HTTPS and QUIC. We present a novel approach that leverages real-time flow priority on a 5G User Plane Function (UPF) node. By predicting flow priority with the first packet, our method enables early priority treatment for all flows, significantly improving the overall throughput of priority flows. We employ a hybrid approach combining Reinforcement Learning (RL) with Deep Packet Inspection (DPI), and Machine Learning (ML) solutions for fast and accurate flow priority detection. RL is utilized for fast and early prediction, while ML is employed for accurate flow priority prediction with additional context. Our experimental results on captured and open-source datasets demonstrate a prediction time of less than 2 microseconds and an accuracy rate of up to 91% for the Multilayer Perceptron (MLP) algorithm. Our RL algorithm exhibits good cumulative rewards and accurate flow priority predictions over time. Our use case has the potential to enable intelligent flow handling in 5G UPF nodes, resulting in higher throughput and efficient network resource allocation.