Large-Scale IP Usage Identification via Deep Ensemble Learning (Student Abstract)

Zhiyuan Wang, Fan Zhou, Kunpeng Zhang, Yong Wang

Abstract

Identifying the usage scenarios of IP addresses—such as home broadband, data centers, or enterprise networks—is a critical task for cybersecurity and digital marketing. This paper introduces a novel deep learning framework that combines differentiable decision trees with Neural Ordinary Differential Equations (ODEs) to classify IP blocks. By treating feature transformations as a continuous flow, the proposed model achieves superior accuracy compared to traditional machine learning and state-of-the-art tabular models across diverse geographical regions.

Introduction

While IP geolocation is a mature field, IP usage scenario identification (IPusi) remains a significant challenge due to the volatility and noise of individual network behaviors. Distinguishing between a home user, a corporate network, or a server farm is vital for filtering bot traffic and flagging suspicious activities. This research addresses the limitations of traditional classification methods by proposing a two-pronged approach: smarter data grouping and a sophisticated neural architecture designed for tabular network data.

Methodology

The proposed framework leverages three key innovations to overcome the noise inherent in network data:

  1. IP Blocking: Instead of analyzing individual IPs, the system segments them into blocks based on WHOIS and BGP routing information. This aggregation provides stable statistical features that smooth out the "jitter" of single-user behavior.
  2. Differentiable Oblivious Decision Trees (ODT): The model utilizes stochastic routing to make decision trees differentiable, allowing them to be trained via backpropagation within a deep learning pipeline. This bridges the gap between traditional tabular methods and modern neural networks.
  3. Neural ODE Integration: Tree layers are stacked using Neural Ordinary Differential Equations, treating the data transformation as a continuous flow rather than discrete steps. This enables the learning of complex, high-order feature interactions.

Experimental Setup

The model was evaluated using large-scale real-world datasets from four regions: Sichuan, Shandong, Chongqing (China), and Illinois (USA). These datasets categorized millions of IPs into four scenarios: Home Broadband (HB), Dedicated Enterprise Network (DEN), Cellular Network (CN), and Data Center (DC).

Table 1: Statistical overview of the regional datasets used in the study, highlighting the scale of IP addresses and blocks analyzed.

Results and Discussion

The proposed model was benchmarked against several baselines, including traditional machine learning (SVM, LDA), boosting frameworks (CatBoost, XGBoost), and specialized tabular deep learning models (TabNet, AutoInt). The primary metric for evaluation was the Area Under the ROC Curve (AUC).

Table 2: Performance comparison (AUC) of the proposed model against baseline methods across three regions.

As shown in Table 2, the proposed model consistently outperformed all baselines. While the improvement over XGBoost is incremental in some regions, the deep ensemble approach provides a more robust framework for capturing expressive information flow. Furthermore, the model demonstrated strong regional generalizability, suggesting that network usage patterns share universal characteristics across different geographies.

Conclusion

This research demonstrates that bridging the gap between decision trees and deep learning through differentiable routing and Neural ODEs significantly improves IP usage identification. The move from individual IP analysis to block-level analysis serves as a force multiplier for classification accuracy. Future work will focus on reducing computational overhead and exploring the temporal dynamics of IP usage in the context of dynamic allocation and VPNs.

Summary updated

Abstract

Understanding users' behavior via IP addresses is essential towards numerous practical IP-based applications such as online content delivery, fraud prevention, and many others. Among which profiling IP address has been extensively studied, such as IP geolocation and anomaly detection. However, less is known about the scenario of an IP address, e.g., dedicated enterprise network or home broadband. In this work, we initiate the first attempt to address a large-scale IP scenario prediction problem. Specifically, we collect IP scenario data from four regions and propose a novel deep ensemble learning-based model to learn IP assignment rules and complex feature interactions. Extensive experiments support that our method can make accurate IP scenario identification and generalize from data in one region to another.
BibTeX
@inproceedings{wang2022large,
  author = {Zhiyuan Wang and Fan Zhou and Kunpeng Zhang and Yong Wang},
  title = {{Large-Scale IP Usage Identification via Deep Ensemble Learning (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21675},
}
Figure