Test-Time training for domain adaption in IP geolocation

IP geolocation is a fundamental task in network management and cybersecurity, enabling services such as targeted advertising and content delivery optimization. However, the performance of geolocation models often deteriorates due to distribution shifts between training and testing data, such as changes in network topologies, new IP address allocations, or the emergence of previously unseen geographical regions. These challenges make it difficult for traditional geolocation models to generalize effectively in real-world scenarios. To tackle this issue, we introduce TTAGeo, a novel approach that applies test-time adaptation (TTA) techniques to dynamically adjust model parameters during inference. Our method integrates empirical risk minimization and self-supervised learning (SSL) to adapt geolocation models in a flexible and efficient manner. TTAGeo is particularly designed to handle the graph-structured nature of IP geolocation data, which involves complex feature correlations and topological relationships that are crucial for accurate predictions. Extensive experiments conducted on real-world benchmark datasets show that TTAGeo significantly enhances geolocation accuracy compared to baseline models. The results highlight its robustness and scalability for deployment in real-world network environments. TTAGeo represents a promising step toward more reliable and adaptive IP geolocation methods capable of meeting the challenges of evolving network landscapes.