HGL_GEO: Finer-grained IPv6 geolocation algorithm based on hypergraph learning

• Published in: Information processing & management Vol. 60; no. 6; p. 103518
• Main Authors: Ma, Zhaorui, Hu, Xinhao, Li, Na, Feng, Hao, Zhang, Shicheng, Li, Tianao, Liu, Fenlin, Zhou, Qinglei, Tian, Zhankui, Wang, Hongjian, Hu, Guangwu
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2023
• Subjects: HyperGraph Neural Network; IPv6 geolocation; Network mapping; IPv6 geolocation; HyperGraph Neural Network; Network mapping
• ISSN: 0306-4573; 1873-5371
• DOI: 10.1016/j.ipm.2023.103518

IP geolocation is necessary for applications such as location-aware ad recommendation, traceability, and fraud detection. Presently, graph neural network-based approaches focus on geolocation information transfer of neighboring nodes of the target IP node. However, terminal hosts in real physical proximity in tier networks are often not directly connected, graph learning tasks ignore geographic knowledge from more distant connecting relationships. To address the limitations of existing work, we propose a new HyperGraph Neural Network (hyperGNN)-based IPv6 geolocation framework HGL_GEO (HyperGraph Learning Geolocation), by modeling a set of IP host groups with strong geographic-location correlation using hyper-edges, and incorporating IP knowledge and connectivity relationships into the hypergraph. The target IP nodes then learn geolocation knowledge in the hypergraph convolution, thus enabling IPv6-based geolocation tasks. The final experimental results on a 10-month test data collection from three real regions (Shanghai, New York, and Tokyo) show that the median error distance ranges from 5.837 km to 8.52 km, and that the average error distance ranges from 7.012 km to 9.359 km, compared with those of TNN, MLP_Geo, and GWS_Geo IP geolocation algorithms and the GCN and GAT. The median error distance of HGL_GEO is at least 10.12% less and the average error distance is at least 3.3% less than those of deep learning models.