Enhancing the classification accuracy of IP geolocation

Location aware applications such as confinement of online transactions within acceptable pre-established locations, targeted online advertising, enforcement of digital content and territory rights and cloud auditing can benefit from a more accurate yet robust IP geolocation framework. Various approaches to IP geolocation have been well documented. The most recent approach casts IP geolocation as a machine learning classification problem. Casting IP Geolocation as a machine learning classification problem makes it possible to incorporate more geolocation information to the classifier hence improve accuracy. To enhance the classification accuracy of the existing classification framework, we expand it to include 6 different types of geolocation information (only 5 have been implemented in this thesis). This improves the accuracy in terms of error distance from 253.34 miles to 155.74 miles. To implement this classifier we come up with 4 major subsystems; (1) Measurement subsystem (collects and formats measurements from PlanetLab and US census website), (2) Training subsystem (reads measurements and then develops probability densities as likelihood), (3) Testing subsystem (uses the probability densities of the training set to fit them to the measurements of the testing set. The county with the highest probability density is the estimated county of the target.), (4) Decision subsystem (compares the estimated location to the true location of the target and calculates the error distance between them. Decides if to reclassify target or not).