IP Geolocation with Two-tiered Neural Networks

The ability to accurately determine the geographic location of an arbitrary IP address has potential in many applications. Previous methods that are based on observing the relationship between network delay and physical distance are inaccurate. Methods based on delay similarity are more accurate but inefficient because they require information on a large number of landmark nodes near the destination to be collected and maintained. We propose a method that can overcome both problems. Our method maintains a stable collection of network observers and landmark nodes that covers the target area. Observers are network nodes from which we can issue measurement command such as ping and traceroute. Landmarks are IP addresses which are reachable from observers and for which the physical locations are well-known. With measurement results collected from these landmarks, we train a two-tier neural network that estimates the geolocation of arbitrary IP addresses. The first tier neural network locates a smaller region in which the target IP resides, and then the second tier estimates its location more accurately within that region. We intentionally limit the size of landmark collection to a controllable and reasonable number, and show that with only a limited number of landmarks, we are able to retain a high accuracy that is similar to previous methods. In our experiment, with 1547 landmarks across US territory, the median error of our estimation is 4.1 km on the entire test dataset. On half of the test regions which contain more than 100 landmarks, the median error can be decreased to 3.7 km.