Abstract:
Even though algorithms for adaptively setting video quality in online streaming are a hot topic in networking academia, little is known about how popular online streaming platforms do such adaptation. This creates obvious hurdles for research on streaming algorithms and their interactions with other network traffic and control loops like that of transport and traffic throttling. To address this gap, we pursue an ambitious goal: reconstruction of unknown proprietary video streaming algorithms. Instead of opaque reconstruction through, neural networks, we seek reconstructions that are easily understandable and open to inspection by domain experts. Such reconstruction, if successful, would also shed light on the risk of competitors copying painstakingly engineered algorithmic work simply by interacting with popular services. Our reconstruction approach uses logs of player and network state and observed player actions across varied network traces and videos, to learn decision trees across streaming-specific engineered features. We find that of 10 popular streaming platforms, we can produce easy-to-understand, and high-accuracy reconstructions for 7 using concise trees with no more than 20 rules. We also discuss the utility of such interpretable reconstruction through several examples.
