Revealing QoE of Web Users from Encrypted Network Traffic

Internet Service Providers (ISPs) have a lot to gain from estimating the web browsing quality of their customers. However, unlike Content Providers who can easily access inbrowser computed application-level metrics to estimate web browsing quality, ISPs come short mainly because of traffic encryption. In this paper, we use exact methods and machine learning to estimate well-known application-level web browsing QoS metrics (such as SpeedIndex and Page Load Time) from raw encrypted streams of network traffic. Particularly, we present and open-source a unique dataset targeting a large set of popular pages (Alexa top-500), from probes from several ISPs networks, browsers software (Chrome, Firefox) and viewport combinations, for over 200,000 experiments. Our results show our models to be accurate, and we particularly focus on their ability to generalize to previously unseen conditions, giving guidance concerning their retraining.