A Statistical Framework to Forecast Duration and Volume of Internet Usage Based on Pervasive Monitoring of NetFlow Logs

In this paper, we address an important and practical problem - namely to forecast the duration and volume of Internet usage of a subject based on pervasive and unobtrusive of past history. Unfortunately though, profiling users can have privacy ramifications. In this paper, we present a statistical framework to forecast duration and volume of Internet usage of subjects via processing NetFlow logs from routers. Briefly, NetFlow logs are network level information of IP packets as they traverse a router, but they do not contain the packet payload. In our experimental study, Internet traffic logs of octets and durations of 66 subjects in a college campus were collected (via privacy-preserving NetFlow records) in a pervasive and unobtrusive manner for a month. By applying times series forecasting techniques, we demonstrate that predictions on duration and volume of usage at future times can be made based on past usage, with very good precision. Furthermore, our results also show that with more historical data, prediction accuracy improves further. We believe that our problem in this paper has not been addressed in the literature. We also believe that our contributions in this paper have important consequences in enabling privacy preserving techniques to manage network resources for administrators, cyber security via behavioral based authentication, and smarter advertising.