Network Asset Detection Based on Weakly Supervised Learning

Guowei Zhang, Jie Zhao, Pengyuan Ma, Junjie WuInternational Journal of Computer Science and Information Technology 2025 · DOI

Abstract

Network asset detection technology is the basis for sorting out and counting Internet assets, timely managing vulnerable network assets, and anomaly detection. Due to various shortcomings in existing methods for obtaining device fingerprint information, the accuracy of evaluation results for network asset devices is low and the functionality is single. This paper proposes a network asset detection method based on weakly supervised learning, which aims to simulate real network data as much as possible to detect network assets and achieve more accurate results. The results indicate a significant improvement in the speed and accuracy of asset detection when using only 30% of labeled data.

Introduction

Network asset detection—identifying and cataloging devices, applications, and services on a network—is the foundational step for cybersecurity. Traditional detection methods typically fall into two categories: active scanning and passive traffic analysis. Active scanning (e.g., Nmap) can be disruptive and is often blocked by firewalls, while passive scanning frequently misses dormant devices. Machine learning (ML) offers a modern alternative by classifying device "fingerprints" from traffic behavior. However, traditional supervised ML requires vast amounts of perfectly labeled data, which is prohibitively expensive to maintain in dynamic, real-world networks. This research addresses the challenge of accurately identifying network assets when only a small fraction of network traffic data is labeled.

Methodology

The proposed approach frames the problem as a semi-supervised learning task, utilizing a pseudo-labeling algorithm to leverage unlabeled data.

Base Model Selection

The researchers evaluated several supervised models, including Decision Trees, Gradient Boosting, KNN, Naive Bayes, Random Forest, and SVM. The Decision Tree (DT) was selected as the backbone because it provided accuracy comparable to Random Forest while offering the faster response times necessary for real-time network monitoring.

Pseudo-Labeling Pipeline

The core of the method is an iterative pipeline designed to bootstrap model performance from limited labeled data:

  1. Initial Training: The DT is trained on a small subset of manually labeled traffic data.
  2. Label Generation: The model predicts labels for the unlabeled data pool.
  3. Selection: Predictions exceeding a specific confidence threshold (CFT) are accepted as "pseudo-labels."
  4. Retraining: The original labeled data and the new pseudo-labeled data are combined to retrain the model.
  5. Convergence: This process repeats until the model converges.
Algorithm 1: The proposed weakly supervised learning method using pseudo-labeling.

To prevent bias toward majority classes in early iterations, the algorithm enforces an equal number of samples for each pseudo-label class during the first 10 training cycles.

Data Preprocessing

Using a public network traffic dataset, the authors extracted 30 specific traffic features to create a robust device fingerprint. Notably, MAC addresses were removed from the feature set to ensure the model learns generalizable behavioral patterns rather than simply memorizing hardware identifiers.

Results and Discussion

The weakly supervised model demonstrated strong performance, particularly in low-data scenarios. When trained on just 30% of the available labeled data, the method achieved an accuracy of 0.72 and an F1 score of 0.74. This outperformed a standard Decision Tree trained on the same 30% split (0.69 accuracy, 0.70 F1) and even slightly surpassed a fully supervised Decision Tree trained on 100% of the labeled data (0.70 accuracy, 0.72 F1). This suggests that the pseudo-labeling process may have a regularizing effect that improves overall generalization.

A key insight from the study was the "Similar Device" heuristic. While the model initially struggled to distinguish between different equipment models from the same manufacturer, grouping these similar devices into single label classes improved accuracy to 88%. From a security perspective, this is often sufficient, as similar devices typically share vulnerabilities.

Conclusion

This research proves that data efficiency is achievable in network asset detection. The pseudo-labeling loop successfully bootstraps the model's understanding of the network environment using unlabeled traffic. Future work will explore more complex architectures, such as combining deep learning with transfer learning, and optimizing parameters for real-time data flow analysis in live environments.

Summary updated

Abstract

Network asset detection technology is the basis for sorting out and counting Internet assets, timely managing vulnerable network assets and anomaly detection. Due to various shortcomings in existing methods for obtaining device fingerprint information, the accuracy of evaluation results for network asset devices is low and the functionality is single. For this reason, proposes a network asset detection method based on weakly supervised learning, which aims to simulate the real network data in the network environment as much as possible to detect network assets, achieve more accurate results, so as to connect with the task of sorting out Internet assets and timely warning network asset vulnerabilities. The results indicate that there is a significant improvement in the speed and accuracy of asset detection when using only 30% of labeled data.
BibTeX
@article{zhang2025network,
  author = {Guowei Zhang and Jie Zhao and Pengyuan Ma and Junjie Wu},
  title = {{Network Asset Detection Based on Weakly Supervised Learning}},
  journal = {International Journal of Computer Science and Information Technology},
  year = {2025},
  doi = {10.62051/ijcsit.v5n1.13},
}
Figure