UncertaintyGeo: A Dirichlet Network Architecture for Evaluating IP Geolocation Uncertainty

Fine-grained IP geolocation plays a critical role in location-based services and cybersecurity. However, due to the impact of inherent noise on data features, most existing fine-grained IP geolocation methods fail to accurately predict the region of the target host. Although recent studies have focused on IP region prediction, they do not provide reliability assessments for the predictions, which limits the practical application of IP geolocation. To address these challenges, this paper proposes UncertaintyGeo, a fine-grained IP geolocation framework based on the Dirichlet network. We revisit IP geolocation from a classification perspective and introduce a “region-first, coordinate-second” paradigm rooted in the Dirichlet distribution. This framework first predicts the region of the target host, and uses the properties of the Dirichlet concentration parameters to evaluate the reliability of the predictions, then obtains coordinate predictions by weighting the centers of candidate regions based on prediction confidence. Experiments on real-world datasets from New York, Los Angeles, and Shanghai demonstrate that UncertaintyGeo significantly outperforms state-of-the-art methods in region-level prediction while achieving notable advantages in coordinate-level prediction and reliability estimation.