Features that Matter: Feature Selection for On-line Stalling Prediction in Encrypted Video Streaming

Abstract

Despite the vast literature and major developments in HTTP adaptive video streaming (HAS) technology, stalling events due to re-buffering still are by far the worst Quality of Experience (QoE) degradation, and thus, represent a major issue for ISPs. In this paper, we address the problem of real-time QoE monitoring of HAS, focusing on the detection of re-buffering events and QoE-relevant metrics, for the particular case of YouTube. Given the wide adoption of end-to-end encryption, we resort to machine learning models to predict these metrics directly from the analysis of the encrypted traffic. The salient feature of our approach ViCrypt is its ability to perform QoE predictions in real-time, during the course of an ongoing YouTube streaming session, relying on constant memory, stream-like inputs continuously extracted from the encrypted stream of packets. We show through empirical evaluations that ViCrypt can predict the occurrence of re-buffering events with a time granularity as small as one second with very high accuracy. By aggregating independent predictions, ViCrypt is able to accurately estimate per-video-session QoE-relevant metrics such as initial playback delay, number of re-buffering events and stalling ratio. In this situation, we investigate if a decent prediction performance can also be reached by selecting reduced feature sets based on the relevance of the features. Moreover, we explore the potential of including recurrent features, namely, stalling predictions from past time slots, to improve the prediction performance.