EUVIMEDEUVIMED entdecken
Crossref

CFMD-i: Communication-Efficient Clustered Federated Multidomain Learning for Robust Network Intrusion Anomaly Detection

Chunjiong Zhang, Yunchun Su, Zengmin Xu, Wenjing Guo, Zhenbing Liu

IEEE Internet of Things Journal · 15.07.2026 · 13 (14) · 30958-30973

DOI: 10.1109/jiot.2026.3685960 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Internet of Things Journal, erschienen am 15.07.2026. Verfasst von Chunjiong Zhang und 4 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 50 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Guangxi Natural Science Foundation.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In highly heterogeneous Internet of Things (IoT) environments, multisource network intrusion data exhibit significant nonindependent and identically distributed characteristics, accompanied by severe class imbalance, privacy constraints, and insufficient terminal computing power. These factors make it challenging for traditional centralized or federated anomaly detection models to achieve stable performance. To address these challenges, this article proposes an efficient and adaptive dynamic clustering-based federated multidomain learning framework with initial dynamic selection (CFMD-i). First, a federated multidomain optimization objective is constructed. By sharing cross-domain knowledge and introducing adversarial domain weights, the model’s robustness across multiple attack domains is enhanced. Subsequently, a dynamic hierarchical clustering mechanism is designed. Nodes are adaptively partitioned based on cosine similarity of node model parameters, Kullback–Leibler (KL) divergence, and intracluster compactness, enabling simultaneous acquisition of overall generalization capability and domain-level personalized inference capability. Concurrently, we introduce communication optimization mechanisms, including low-rank singular value decomposition (SVD) compression, differential transmission of model parameters, leapfrog communication strategy, and quantization-encoded residual compensation, significantly reducing bandwidth consumption while enhancing stability in weak network environments. Experiments on KDDcup’99, AWID, and CICIDS2017 datasets validate the method’s effectiveness. Compared to FedAvg, Dis-DAGMM, DIOT, and the latest clustering-based federated learning (FL) method ZeKoC, CFMD-i achieves 4%–12% $F1$ -score improvements across multiple attack domains. It maintains stable convergence while reducing communication volume by over 95% and demonstrates reliable performance even under 12% packet loss conditions.

Förderung

  • National Natural Science Foundation of China
  • Guangxi Natural Science Foundation

Zitieren

APA

Zhang, C., Su, Y., Xu, Z., Guo, W., & Liu, Z. (2026). CFMD-i: Communication-Efficient Clustered Federated Multidomain Learning for Robust Network Intrusion Anomaly Detection. IEEE Internet of Things Journal, 13(14), 30958-30973. https://doi.org/10.1109/jiot.2026.3685960

Vancouver

Zhang C, Su Y, Xu Z, Guo W, Liu Z. CFMD-i: Communication-Efficient Clustered Federated Multidomain Learning for Robust Network Intrusion Anomaly Detection. IEEE Internet of Things Journal. 2026;13(14):30958-30973. doi:10.1109/jiot.2026.3685960

BibTeX
@article{Zhang2026,
  author = {Zhang, Chunjiong and Su, Yunchun and Xu, Zengmin and Guo, Wenjing and Liu, Zhenbing},
  title = {CFMD-i: Communication-Efficient Clustered Federated Multidomain Learning for Robust Network Intrusion Anomaly Detection},
  journal = {IEEE Internet of Things Journal},
  year = {2026},
  volume = {13},
  number = {14},
  pages = {30958--30973},
  doi = {10.1109/jiot.2026.3685960}
}
Kostenlos starten →