Agentic AI-Driven Risk-Aware Intrusion Detection: A Bayesian Spatiotemporal Graph Learning Framework for Secure 6G
IEEE Transactions on Cognitive Communications and Networking · 2026 · 12 · 10327-10342
DOI: 10.1109/tccn.2026.3711087 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Cognitive Communications and Networking, 2026. Verfasst von Yiting Yang und 4 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 44 Literaturangaben. Gefördert von: Natural Science Foundation on Frontier Leading Technology Basic Research Project of Jiangsu, National Natural Science Foundation of China, Fundamental Research Funds for the Central Universities.
Abstract (Originalsprache)
As sixth-generation networks evolve toward edge-native and agentic operation, network security functions are expected to support fast local perception, uncertainty assessment, and response decisions. Security agents deployed at the network edge therefore need uncertainty-aware predictions that indicate both the predicted attack category and the reliability of that prediction. However, conventional flow-based inspection often relies on per-flow state and deterministic predictors, which can increase edge-side overhead and lead to overconfident decisions under distribution shifts. This paper presents Bayesian Spatiotemporal Graph Convolutional Network (BSTGCN), a perception-to-belief module for agentic security pipelines. BSTGCN operates on flow-agnostic packet windows without maintaining per-flow state, constructs a dynamic $k$ -nearest-neighbor topology to capture behavioral similarity, and performs variational Bayesian spatiotemporal inference. The module outputs calibrated predictive distributions, quantifies epistemic and aleatoric uncertainty, and derives a posterior-predictive risk score for belief-driven alert triage. Experiments on the 5G-NIDD dataset show that BSTGCN achieves 96.74% accuracy and 91.65% macro-F1, while reducing Expected Calibration Error from 0.0084 for a spatiotemporal graph convolutional network with Monte Carlo dropout (STGCN+MC Dropout) baseline to 0.0041. In controlled out-of-distribution detection tests based on held-out attack types, BSTGCN attains an area under the receiver operating characteristic curve of $0.878\pm 0.04$ . The high-risk queue captures 86.5% of observed misclassifications in the top 20% ranked windows, indicating that the posterior-predictive risk score can prioritize likely errors under a fixed triage budget.
Förderung
- Natural Science Foundation on Frontier Leading Technology Basic Research Project of Jiangsu
- National Natural Science Foundation of China
- Fundamental Research Funds for the Central Universities
- National Key Research and Development Program of China
- Jiangsu Innovation Capacity Building Program
Zitieren
APA
Yang, Y., Zhan, H., He, S., Wang, J., & Huang, Y. (2026). Agentic AI-Driven Risk-Aware Intrusion Detection: A Bayesian Spatiotemporal Graph Learning Framework for Secure 6G. IEEE Transactions on Cognitive Communications and Networking, 12, 10327-10342. https://doi.org/10.1109/tccn.2026.3711087
Vancouver
Yang Y, Zhan H, He S, Wang J, Huang Y. Agentic AI-Driven Risk-Aware Intrusion Detection: A Bayesian Spatiotemporal Graph Learning Framework for Secure 6G. IEEE Transactions on Cognitive Communications and Networking. 2026;12:10327-10342. doi:10.1109/tccn.2026.3711087
BibTeX
@article{Yang2026,
author = {Yang, Yiting and Zhan, Hang and He, Shiwen and Wang, Jiaheng and Huang, Yongming},
title = {Agentic AI-Driven Risk-Aware Intrusion Detection: A Bayesian Spatiotemporal Graph Learning Framework for Secure 6G},
journal = {IEEE Transactions on Cognitive Communications and Networking},
year = {2026},
volume = {12},
pages = {10327--10342},
doi = {10.1109/tccn.2026.3711087}
}
EUVIMED