Network-Based Video Quality Assessment for Encrypted HTTP Adaptive Streaming

Abstract

End-to-end encryption brings new challenges for mobile network operators to obtain key quality indicators (KQIs) of the HTTP adaptive streaming (HAS), which are vital for quality of experience (QoE) of the end users. Targeting on this issue, this paper proposes a network-based solution to extract KQI-related features from encrypted traffic; then, the video quality is assessed through machine-learning-based methods. First, content-independent features are extracted resorting to traffic characteristics and a typical HAS player model. Then, back-propagation neural network and random forest are applied to evaluate stalling and estimate initial buffering delay, which are known as the most important KQIs for QoE. In addition, to evaluate the performance of our solution, we construct a data acquisition platform, and 4733 encrypted video sessions are collected from YouTube by crowdsourcing. Experiment results show that the proposed solution outperforms the previous methods, where initial buffering delay is effectively estimated with an absolute error less than 1 s for up to 80% videos, and the accuracy of stalling detection reaches 88.5%.