EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Network intrusion detection using a hybrid graph-based convolutional network and transformer architecture

Peter Appiahene, Samuel Opoku Berchie, Emmanuel Botchway iD, Michael Junior Ayitey, John Kwao Dawson, Henry Nii-Armah Mettle, Stephen Afrifa

PLOS One · 21.01.2026 · 21 (1) · e0340997

DOI: 10.1371/journal.pone.0340997 ↗

Kurz gefasst

Fachartikel in der Zeitschrift PLOS One, online erschienen am 21.01.2026. Verfasst von Peter Appiahene und 6 weiteren Autor:innen. Beteiligte Einrichtung: Department of Information Technology and Decision Sciences, University of Energy and Natural Resources, Sunyani, Ghana. Hauptthemen laut MEDLINE: Cloud Computing, Computer Security. Die Arbeit wurde bisher 15-mal zitiert (Crossref) und nennt 31 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Cloud computing continues to expand rapidly due to its ability to provide internet-hosted services, including servers, databases, and storage. However, this growth increases exposure to sophisticated intrusion attacks that can evade traditional security mechanisms such as firewalls. As a result, network intrusion detection systems (NIDS) enhanced with machine learning and deep learning have become increasingly important. Despite notable advancements, many AI-based intrusion detection models remain limited by their dependence on extensive, high-quality attack datasets and their insufficient capacity to capture complex, dynamic patterns in distributed cloud environments. This study presents a hybrid intrusion detection model that combines a graph convolutional layer and a transformer encoder layer to form deep neural network architecture. Using the CIC-IDS 2018 dataset, tabular network traffic data was transformed into computational graphs, enabling the model called “GConvTrans” to leverage both local structural information and global context through graph convolutional layers and multi-head self-attention mechanisms, respectively. Experimental evaluation shows that the proposed GConvTrans obtained 84.7%, 96.75% and 96.94% accuracy on the training, validation and testing set respectively. These findings demonstrate that combining graph learning techniques with standard deep learning methods can be robust for detecting complex network intrusion. Further research would explore other datasets, continue refining the proposed architecture and its hyperparameters. Another future research direction for this work is to analyze the architecture on other graph learning tasks such as link prediction.

Schlagwörter

MeSH-Schlagwörter (MEDLINE)

Cloud ComputingComputer SecurityGraph Neural NetworksConvolutional Neural NetworksAlgorithmsDeep LearningNeural Networks, ComputerMachine Learning

Autor:innen und Institutionen

  • Department of Information Technology and Decision Sciences, University of Energy and Natural Resources, Sunyani, Ghana.
  • Department of Computer Science, Sunyani Technical University, Sunyani, Ghana.

ORCID-Profile: Emmanuel Botchway

Zitieren

APA

Appiahene, P., Berchie, S. O., Botchway, E., Ayitey, M. J., Dawson, J. K., Mettle, H. N., & Afrifa, S. (2026). Network intrusion detection using a hybrid graph-based convolutional network and transformer architecture. PLOS One, 21(1), e0340997. https://doi.org/10.1371/journal.pone.0340997

Vancouver

Appiahene P, Berchie SO, Botchway E, Ayitey MJ, Dawson JK, Mettle HN, et al. Network intrusion detection using a hybrid graph-based convolutional network and transformer architecture. PLOS One. 2026;21(1):e0340997. doi:10.1371/journal.pone.0340997

BibTeX
@article{Appiahene2026,
  author = {Appiahene, Peter and Berchie, Samuel Opoku and Botchway, Emmanuel and Ayitey, Michael Junior and Dawson, John Kwao and Mettle, Henry Nii-Armah and Afrifa, Stephen},
  title = {Network intrusion detection using a hybrid graph-based convolutional network and transformer architecture},
  journal = {PLOS One},
  year = {2026},
  volume = {21},
  number = {1},
  pages = {e0340997},
  doi = {10.1371/journal.pone.0340997}
}
Kostenlos starten →