MTLH: Video QoE Monitoring for Encrypted Traffic via Multi-Task Learning with Hierarchy

With the continuous growth of video traffic on the whole Internet, it is of vital importance for content providers to provide high-quality service. Monitoring the QoE (Quality of Experience) of video can help them to improve their service. However, most of the video traffic is encrypted traffic which has unreadable payload and less information in the transport layer that makes monitoring QoE a challenging task. QoE is measured from multiple metrics and many methods have been proposed to infer it. Traditional machine learning based methods have two main shortcomings. 1) These methods ignore the relevance of these metrics which can do a favor to the inference of QoE. 2) They need to extract the transport layer features and statistic features manually. In this paper, we propose multi-task learning with hierarchy (MTLH) model for inferring video QoE metrics. We use shared layer to extract deep features automatically from traffic tendency pictures which reveal the variation tendency of video streaming, that characterizes video QoE. In addition, we utilize the hierarchy to promote fine-grained start-up delay tasks. Finally, we use multitask learning architecture to learn the correlation between each task and use their correlation to refer QoE metrics. We evaluate our model for YouTube adaptive video streams on two datasets. The proposed approach achieved an average 9.0% higher of precision score and an average 11.5% higher of recall score than the state-of-the-art method in two datasets.