GWS-Geo: A graph neural network based model for street-level IPv6 geolocation

Abstract

IP geolocation is a technique for inferring geolocation of a device based on characteristics of its IP addresses such as network measurements or queryable information. Traditional IPv4 geolocation methods are often inapplicable to IPv6 networks or produce unsatisfactory results due to the imperfection of network routing and the sparsity of IPv6 addresses. To improve the accuracy of IPv6 geolocation, we proposed GWS-Geo, a street-level IPv6 geolocation model based on Graph Neural Network(GNN). Our model includes preprocessing, pre-training, an improved GraphSAGE algorithm, and hierarchical classification. In the preprocessing step, we process node features and transform them into node embeddings, and transform landmarks’ latitude and longitude coordinates into area numbers. During pretraining, we assign weights to edges between IP addresses and input node information into the improved GraphSAGE algorithm. After pruning according to edge weights, the improved GraphSAGE executes graph convolution operations. Finally, we use hierarchical classification to divide the geolocation into finer granularity to obtain the location of the target IP address. Experimental results on three datasets (covering the Tokyo, New York, and Shanghai areas) show that the median error distance is within a range of 5.53km to 9.46 km, which is close to the level of street-level accuracy. Compared to popular geolocation algorithms such as SLG, IRLD, and MLP-Geo, our model reduces the median error distance by at least 15.99% and the average error distance by at least 16.36%.