Machine Learning-based BGP Traffic Prediction

Abstract

Accurate Internet traffic predictions can provide support to network operators for applications such as traffic engineering, bandwidth allocation, anomaly detection, etc. We apply and compare different forecasting techniques (traditional and machine learning-based techniques) on real BGP data that is collected from two well-known Internet exchange points (IXPs) to derive BGP future volume-based predictions. Our experimental evaluation shows that multivariate Bayesian Ridge outperforms all other forecasting techniques we consider. Through univariate LSTM, we are able to predict new BGP volume-based features. Furthermore, to study the impact of dataset size on BGP forecasting, we perform experiments on three BGP dataset sizes, i.e., Short (one-month), Medium (three-months), and Long (five-months) Periods. Our results show that the Short-Period BGP dataset seems to be sufficient for getting accurate predictions. We also present a use case study (forecast Google Leak anomaly) that supports our experimental evaluations. We provide our collected BGP datasets publically which will be helpful to perform further research experiments and analysis regarding BGP traffic predictions.