City-level IPv6 geolocation algorithm based on self-attention-driven router sequence learning

IP geolocation uses IP addresses to determine the geolocation of network devices, supporting location-based services (LBS) such as online advertising, content access control, and cybercrime tracking. Current city-level IP geolocation algorithms are designed for IPv4 networks, their key assumption of relationships between geolocation and network measurement data, such as static city boundary routers, may not be suitable in rapid-growing IPv6 networks. Therefore, transferring current geolocation algorithms to IPv6 networks often lead to much lower accuracy. To offer reliable city-level geolocation services for IPv6 networks, this paper introduces a city-level IPv6 geolocation algorithm based on self-attention-driven router sequence learning (SRSL-Geo). Instead of finding static mapping rules, SRSL-Geo models dynamic mapping patterns in router sequences of IPv6 targets. SRSL-Geo leverages self-attention-driven deep neural networks to place more importance on routers which are related to city-level geolocation. The experiments in real-world IPv6 networks covering 22 cities of China, Germany, and Japan show SRSL-Geo improves the accuracy of city-level geolocation accuracy averagely by 46.38%, compared with mainstream algorithms.