IPGeoAI introduces a deep learning architecture that redefines IP geolocation as a sequential modeling task. It employs a Transformer Encoder to capture hierarchical dependencies in IP subnet structures.
The model integrates unstructured semantic context from Large Language Models (LLMs) via a Zero-Shot Feature Extraction pipeline. This transforms raw Autonomous System descriptions into structured metadata, such as distinguishing between 'University' and 'ISP'.
Semantic signals are fused into the network through a Multi-Head Cross-Attention module, bridging network topology and real-world semantics. Extensive offline evaluation on a proprietary dataset shows significant improvements in city-level accuracy and coverage.
In large-scale online tests, IPGeoAI achieved a statistically significant increase in downstream metrics, demonstrating its practical benefits for geolocation services.
Source: https://arxiv.org/abs/2609.04559
