Enhancing IoT Intrusion Detection with Stochastic LWTA Deep Networks for Real-World MQTT Traffic
Computer and Decision Making: An International Journal · 31.10.2026 · 3 (6S) · 876-894
DOI: 10.1922/comdem.v3i6s.18877 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 31.10.2026. Verfasst von Soumya und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Background
The rapid proliferation of Internet of Things (IoT) devices over MQTT-based infrastructures has significantly expanded the attack surface of cyber–physical systems. Reliable intrusion detection systems (IDS) are required to distinguish benign traffic from diverse attack patterns under realistic network conditions. Problem Statement/Research Gap: Most existing MQTT-based IDS studies rely on simulated datasets, limited attack taxonomies, or conventional machine learning models that are evaluated only under clean conditions. The robustness of deep architectures, particularly under noisy or perturbed inputs that approximate real deployments, remains insufficiently explored. Methodology: This work investigates IoT intrusion detection on the real-world MQTTEEB-D dataset, comprising 634,644 records and 6 traffic classes derived from MQTT/TCP flows. After SMOTE-based class balancing and Z-score standardization, the data are split into 70% for training, 15% for validation, and 15% for testing. The study benchmarks Gaussian Naive Bayes, a regularized Decision Tree, Random Forest, XGBoost, a ReLU-based multilayer perceptron (MLP), and a deep neural network with Stochastic Local Winner-Takes-All (S-LWTA) layers. In addition to standard test performance, robustness is assessed by injecting Gaussian noise with increasing standard deviation into the test features.
Results
On clean data, XGBoost achieves the best performance with test accuracy 0.8858 and Macro-F1 0.8806, closely followed by Random Forest (accuracy 0.8827, Macro-F1 0.8809). The proposed Deep S-LWTA network outperforms the plain MLP (accuracy 0.6080 vs. 0.5770; Macro-F1 0.6012 vs. 0.5807), but remains below the tree ensembles in absolute accuracy. Under strong noise (σ = 0.10), Deep S-LWTA becomes the most robust model, retaining accuracy 0.4661 and Macro-F1 0.4532, whereas XGBoost and Random Forest drop to accuracies of 0.2606 and 0.2955, respectively.
Conclusion
Real-world MQTT traffic analysis confirms that tree-based ensembles are the most accurate models for clean IoT intrusion detection, while the proposed S-LWTA architecture provides a competitive deep baseline with markedly better robustness to feature perturbations. These findings position S-LWTA as a suitable neural backbone for future work on adversarially robust IoT IDS across multiple datasets.
Zitieren
APA
Soumya, & Vincent, D. R. R. (2026). Enhancing IoT Intrusion Detection with Stochastic LWTA Deep Networks for Real-World MQTT Traffic. Computer and Decision Making: An International Journal, 3(6S), 876-894. https://doi.org/10.1922/comdem.v3i6s.18877
Vancouver
Soumya, Vincent DRR. Enhancing IoT Intrusion Detection with Stochastic LWTA Deep Networks for Real-World MQTT Traffic. Computer and Decision Making: An International Journal. 2026;3(6S):876-894. doi:10.1922/comdem.v3i6s.18877
BibTeX
@article{Soumya2026,
author = {Soumya and Vincent, Dr. Robin Rohit},
title = {Enhancing IoT Intrusion Detection with Stochastic LWTA Deep Networks for Real-World MQTT Traffic},
journal = {Computer and Decision Making: An International Journal},
year = {2026},
volume = {3},
number = {6S},
pages = {876--894},
doi = {10.1922/comdem.v3i6s.18877}
}
EUVIMED