Enhancing the Classification Accuracy of IP Geolocation

The ability to localize Internet hosts is appealing for a range of applications from online advertising to localizing cyber attacks. Recently, measurement-based approaches have been proposed to accurately identify the location of Internet hosts. These approaches typically produce erroneous results due to measurement errors. In this paper, we propose an Enhanced Learning Classifier approach for estimating the geolocation of Internet hosts with increased accuracy. Our approach extends an existing machine learning based approach by extracting six features from network measurements and implementing a new landmark selection policy. These enhancements allow us to mitigate problems with measurement errors and reduces average error distance in estimating location of Internet hosts. To demonstrate the accuracy of our approach, we evaluate the performance on network routers using ping measurements from PlanetLab nodes with known geographic placement. Our results demonstrate that our approach improves average accuracy by geolocating internet hosts 100 miles closer to the true geographic location versus prior measurement-based approaches. Published in the Proceedings of the IEEE MILCOM 2012, 29 Oct-1 Nov 2012, Orlando, FL.