As stalling is the worst Quality of Experience (QoE) degradation of HTTP adaptive video streaming (HAS), this work
presents a stream-based machine learning approach, ViCrypt, which analyzes stalling of YouTube streaming sessions in real-time from encrypted network traffic. The video streaming session is subdivided into a stream of short time slots of 1 s length, while considering two additional macro windows each for the current streaming trend and the whole ongoing streaming session. Constant memory features are extracted from the encrypted network traffic in these three windows in a stream-based fashion, and fed into a random forest model, which predicts whether the current time slot contains stalling or not. The presented system can predict stalling with a very high accuracy and the finest granularity to date (1 s), and thus, can be used in networks for real-time QoE analysis from encrypted YouTube video streaming traffic. The independent predictions for each consecutive slot of a streaming session can further be aggregated to obtain stalling
estimations for the whole session. Thereby, the proposed method allows to quantify the initial delay, as well as the overall number of stalling events and the stalling ratio, i.e., the ratio of total stalling time and total playback time.