Predicting Video Bitrate from Encrypted Streaming Traffic in SDN-based 5G Networks with ML

Video streaming applications currently dominate mobile network traffic. This predominance motivates network operators to optimize the network aiming at the quality of experience (QoE) of video streaming users. However, due to the widely adopted end-to-end traffic encryption, several key video-QoE indicators (KQI) that are useful to infer QoE are not readily available to the network operator. This work proposes a method to predict the video bitrate KQI from encrypted video streaming traffic. We compare the following machine learning (ML) algorithms for this task: Random Forest, Multilayer Perceptron, and Long Short-Term Memory networks. Since the only information available to the network operator are key performance indicators (KPIs), such as throughput, we use only information obtained from KPIs in our evaluations. We implement an open-source emulation setup for networking experiments to generate the data to train ML algorithms and a feature engineering procedure to obtain statistical features from the raw KPI data. Furthermore, we evaluate the proposed method in two ML learning cases: offline and incremental, and discuss issues regarding the generalization capability of the ML algorithms.