Structural Sensitivity of Graph Neural Networks for Intrusion Detection: An Intervention Study
IEEE Access · 2026 · 14 · 91372-91384
DOI: 10.1109/access.2026.3703442 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ki On Chan und 3 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: U.S. Department of Energy, National Science Foundation.
Abstract (Originalsprache)
The rapid increase in Internet of Things (IoT) devices in critical infrastructure and enterprise networks has increased cybersecurity risk and motivated the development of graph-based intrusion detection methods. This work presents an intervention-based analysis of structural sensitivity in GNN-based intrusion detection. Using the ToN-IoT and CIC-IDS-2017 datasets, we represent the network as a flow-based directed multigraph for message passing and a simplified undirected graph for structural analysis. We compare weighted cross-entropy (wCE), focal loss, and a combined loss, and our results suggest that wCE is generally competitive across models and datasets. We also conduct ablation studies, suggesting that both node and edge features influence performance, with sensitivity patterns that vary across architectures and datasets. Finally, we evaluate graph interventions including triangle removal, bridge removal, and degree-preserving rewiring. The results suggest that structural sensitivity depends on both the model and the dataset: bridge removal has limited measurable effect across settings, while degree-preserving rewiring tends to produce relatively larger distributional shifts for some model–dataset pairs. Overall, the findings suggest that graph-based intrusion detection is influenced by the interaction between architecture, feature representation, and graph structure.
Förderung
- U.S. Department of Energy
- National Science Foundation
Zitieren
APA
Chan, K. O., Kim, Y., Jo, J., & Arifuzzaman, S. (2026). Structural Sensitivity of Graph Neural Networks for Intrusion Detection: An Intervention Study. IEEE Access, 14, 91372-91384. https://doi.org/10.1109/access.2026.3703442
Vancouver
Chan KO, Kim Y, Jo J, Arifuzzaman S. Structural Sensitivity of Graph Neural Networks for Intrusion Detection: An Intervention Study. IEEE Access. 2026;14:91372-91384. doi:10.1109/access.2026.3703442
BibTeX
@article{Chan2026,
author = {Chan, Ki On and Kim, Yoohwan and Jo, Ju-Yeon and Arifuzzaman, Shaikh},
title = {Structural Sensitivity of Graph Neural Networks for Intrusion Detection: An Intervention Study},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {91372--91384},
doi = {10.1109/access.2026.3703442}
}
EUVIMED