Improving IP Geolocation With Target-Centric IP Graph (Student Abstract)

Kai Yang, Jiayang Li, Wenxin Tai, Zhenhui Li, Ting Zhong, Guangqiang Yin, Yong Wang

Introduction

IP geolocation—determining the physical location of a device based on its IP address—is a critical task for cybersecurity, content delivery, and targeted advertising. While modern methods have shifted from static database lookups to Graph Neural Networks (GNNs) that learn from network topology, they often struggle with data sparsity.

In traditional "router-centric" models, a target IP is typically linked only to the single router with the lowest latency. This approach frequently results in isolated nodes; in major datasets, nearly 10% of targets are completely isolated, and over 14% have fewer than 10 neighbors. Without sufficient landmark connections, GNNs cannot effectively aggregate the spatial information required for accurate prediction.

The TCGeo Framework

To address these limitations, this paper introduces TCGeo (Target-Centric Geolocation). The core innovation lies in a fundamental shift in graph construction.

TCGeo architecture overview: (a) network measurements, (b) target-centric graph construction, and (c) GNN-based location estimation.

Methodology

  1. Target-Centric (TC) IP Graph: Instead of connecting a target host to a single "best" router, TCGeo designates the target host as the central node and aggregates multiple last-hop routers. This creates a significantly denser network.
  2. Feature Engineering: Nodes are represented by 32-dimensional feature vectors. These include WHOIS identity data (6-dim), active network measurements like ping and traceroute (24-dim), and known coordinates for landmarks (2-dim).
  3. Attention-Based GNN: The model employs an attention mechanism to dynamically weight the relevance of different neighbors. A single GNN layer processes these representations, followed by a non-linear regression layer that outputs the estimated latitude and longitude.

Key Insights: Overcoming Sparsity

The authors conducted a "Sparsity Investigation" to understand the relationship between graph density and model performance.

Impact of neighbor counts on MAE (a) and comparison of isolated node counts between router-centric (RC) and target-centric (TC) approaches (b).
  • The "Power of 10": As shown in the figure above, Mean Absolute Error (MAE) drops sharply as the number of neighbors increases from 1 to 10, after which it plateaus. This indicates that a "critical mass" of approximately 10 neighbors is sufficient for high-accuracy geolocation.
  • Reducing Isolation: By allowing connections to multiple last-hop routers, TCGeo reduced the proportion of isolated nodes from 9.73% to 4.46% across the studied datasets.
  • Hybrid Strategy: To maintain computational efficiency, the authors use a hybrid approach: the target-centric method is prioritized for "lonely" IPs (those with few neighbors), while traditional methods are used for nodes that are already well-connected.

Experimental Results

TCGeo was evaluated against three large-scale datasets from New York, Los Angeles, and Shanghai. It consistently outperformed several baselines, including the previous state-of-the-art, TrustGeo.

Dataset Metric TCGeo (Ours) TrustGeo (Best Baseline)
New York MAE (km) 1.271 1.316
Los Angeles MAE (km) 2.658 2.793
Shanghai MAE (km) 5.189 5.457

The results demonstrate that TCGeo provides more precise, street-level estimates even in complex urban environments.

Conclusion and Future Work

TCGeo demonstrates that a simple yet strategic change in graph construction—shifting from router-centric to target-centric—can significantly mitigate the sparsity issues plaguing GNN-based IP geolocation.

Future research may focus on the computational overhead of gathering multi-router traceroute data and further refining the selection of "pertinent" topological neighbors to optimize performance in real-time production environments.

Summary updated

Abstract

Accurate IP geolocation is indispensable for location-aware applications. While recent advances based on router-centric IP graphs are considered cutting-edge, one challenge remain: the prevalence of sparse IP graphs (14.24% with fewer than 10 nodes, 9.73% isolated) limits graph learning. To mitigate this issue, we designate the target host as the central node and aggregate multiple last-hop routers to construct the target-centric IP graph, instead of relying solely on the router with the smallest last-hop latency as in previous works. Experiments on three real-world datasets show that our method significantly improves the geolocation accuracy compared to existing baselines.
BibTeX
@inproceedings{yang2024improving,
  author = {Kai Yang and Jiayang Li and Wenxin Tai and Zhenhui Li and Ting Zhong and Guangqiang Yin and Yong Wang},
  title = {{Improving IP Geolocation With Target-Centric IP Graph (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30529},
}
Figure