Live-Stream Identification Based on Reasoning Network with Core Traffic Set.

The proportion of live-stream traffic in network traffic has been boosting rapidly in recent years, which brings severe challenges to network management. Particularly, some game lives transmit harmful information, which poses great potential harm to adolescents. Therefore, how to effectively identify live-stream traffic becomes an urgent issue to be solved. However, traditional works based on machine learning (ML) methods do not consider many important causalities between flows and video in a live-stream session, lacking interpretability, which inspires us to propose a novel reasoning method to identify different live-stream scenarios. Firstly, we first propose a new technical concept namely core traffic set to contain the most important and significant flows in a live-stream session, and then analyzed the relationship between video flow and other related flows to construct the core traffic set for each session. Then the features related to the live-stream content will be extracted, and a Live-Stream Reasoning Network (LSRN) is designed to infer the corresponding type of live-stream. To evaluate the effectiveness of the proposed approach, a set of experiments are conducted on the dataset collected from the three platforms. In addition, compared with state-of-the-art (SOTA) methods and ensemble classifiers, the results also show that LSRN can significantly contribute to identifying the live video.