YouTube QoE Estimation from Encrypted Traffic: Comparison of Test Methodologies and Machine Learning Based Models

Over the last few years, different client-side QoE monitoring apps have been developed that benchmark the performance of popular video streaming services. Such tools also provide the means for collecting ground truth data when developing models to estimate or classify QoE and various KPls from encrypted network traffic. We present a client-side YouTube QoE monitoring tool named ViQMon, which extracts YouTube performance data from the official app's Stats for Nerds window, and is applicable on various devices and platforms (Android, iOS). We compare ViQMon to approaches relying on YouTube's APls, and show relevant differences in buffering and application behavior in cases when videos are embedded and when videos are played in the official YouTube app. We further use ViQMon together with the collection of network measurements in both a laboratory and commercial mobile network to collect a large dataset of almost 500 YouTube videos streamed under different network conditions. The dataset is used to build machine learning based models for estimating QoE and various application-layer KPls solely from IP-level network traffic features. As such, the approach is applicable in the context of both TLS and QUIC traffic. The paper further compares and analyses the performance of the built models.