Accurate Prediction of Streaming Video Traffic in TCP/IP Networks using DPI and Deep Learning

Video share of the Internet traffic is increasing day by day. This includes streaming on the go to/from mobile devices. These trends necessitate dynamic and robust resource allocation at Internet Exchange Point level to provide good quality of services to mobile video users. Any effective solution to this problem requires accurate predictions of the video traffic coming from or delivered to mobile devices. In this paper, we propose a framework to correctly identify and accurately predict the live streaming and video traffic. Deep packet inspection is used to identify the 23 most common live streaming and video traffic protocols. Subsequently Long Short-Term Memory neural network is used to predict the live streaming and video traffic over a prediction horizon of 6 hours with an average accuracy of 97.24% thus outperforming previous frameworks in both the accuracy and the prediction horizon. This technique can be used as a baseline towards a more effective application of traffic engineering techniques.