Stable Explanations for IoT Data Security: A Lightweight Sparse-Gated Intrusion Detection Model Robust Across Seeds and Perturbations
Journal of Al-Farabi for Engineering Sciences · 2026
DOI: 10.59746/beb4b354 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Al-Farabi for Engineering Sciences, 2026. Verfasst von Shihab Hamad Khaleefah.
Abstract (Originalsprache)
Intrusion detection is a core data-security control for Internet of Things (IoT) networks: it protects the confidentiality, integrity and availability of the data that devices exchange, and its alerts and explanations are themselves security evidence that analysts act on and auditors review. Explainable intrusion detection therefore helps a security analyst only if an alert’s explanation stays the same when the network flow changes slightly or the detector is retrained. Its propose StableGate-IDS, a compact neural IoT intrusion detector trained for explanation stability, which it treat as an integrity requirement on this evidence. A hard-concrete gate first selects a compact feature set and is then frozen; an attribution-consistency loss then aligns the gradient input attribution of each flow with that of a perturbed copy that keeps its prediction. The study evaluate it on Edge-IIoTset (15 classes) and CICIoT2023 (34 classes) against nine baselines, including XGBoost, LightGBM, FT-Transformer and TabM, over ten seeds with Holm-corrected paired tests. StableGate-IDS reaches the best macro-F1 on Edge-IIoTset (86.97%) and the second best on CICIoT2023 (88.30% vs. 89.39% for XGBoost). It uses 14–27 features and under 300 KB, 24–730 smaller than the tree ensembles. Under native explainers its one-pass explanations are the most stable of all models on CICIoT2023 and tied with LightGBM on Edge-IIoTset (top-5 Jaccard 0.932 and 0.944). On CICIoT2023 this holds at every noise level tested and under explanation-targeted perturbations. When every neural model is explained with the far costlier Expected Gradients, TabM becomes more stable. Even then, StableGate-IDS’s gradient input explanation matches or exceeds the Expected-Gradients explanations of all neural baselines at the training noise level. A consistency-weight sweep gives a smooth accuracy–stability trade-off, and ablations show that the loss drives local stability while the gate supports seed consistency. Also report where tree ensembles remain better: sparse perturbations and cross-seed ranking. Code and all run records are released.
Zitieren
APA
Khaleefah, S. H. (2026). Stable Explanations for IoT Data Security: A Lightweight Sparse-Gated Intrusion Detection Model Robust Across Seeds and Perturbations. Journal of Al-Farabi for Engineering Sciences. https://doi.org/10.59746/beb4b354
Vancouver
Khaleefah SH. Stable Explanations for IoT Data Security: A Lightweight Sparse-Gated Intrusion Detection Model Robust Across Seeds and Perturbations. Journal of Al-Farabi for Engineering Sciences. 2026. doi:10.59746/beb4b354
BibTeX
@article{Khaleefah2026,
author = {Khaleefah, Shihab Hamad},
title = {Stable Explanations for IoT Data Security: A Lightweight Sparse-Gated Intrusion Detection Model Robust Across Seeds and Perturbations},
journal = {Journal of Al-Farabi for Engineering Sciences},
year = {2026},
doi = {10.59746/beb4b354}
}
EUVIMED