@Article{cmc.2026.085248,
AUTHOR = {Adi Wibowo, Deris Stiawan, Chonlameth Arpnikanondt, Ahmad Heryanto, Didi Rosiyadi, Mohd Yazid Idris, Fandi Kurniawan, Tuul Triyason, Rahmat Budiarto},
TITLE = {A Robustness Aware Intrusion Detection Framework Using Ensemble Learning and Imbalance Aware Preprocessing for Smart Home IoT Networks},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/28362},
ISSN = {1546-2226},
ABSTRACT = {The proliferation of smart home Internet of Things (IoT) devices has introduced critical security vulnerabilities that demand robust and efficient intrusion detection mechanisms. Existing machine learning-based Intrusion Detection Systems (IDS) often suffer from class imbalance bias, limited cross domain generalization, and insufficient robustness evaluation under adversarial conditions. This study proposes a robustness-aware intrusion detection framework that systematically integrates imbalance-aware preprocessing and ensemble machine learning into a unified pipeline specifically designed for resource-constrained smart home IoT environments. The framework employs Min Max normalization and Synthetic Minority Over sampling Technique (SMOTE) with strict train-test separation to prevent data leakage, operating directly on the seven packet-level traffic features available in the capture rather than on a reduced subset. Three machine learning models Decision Tree (DT), Random Forest (RF), and XGBoost are systematically compared against deep learning baselines (CNN and LSTM) across three benchmark IoT datasets: COMNETS Smart Home, IoT-23, and TON IoT. Under a conventional random-split evaluation, XGBoost achieves the highest detection performance (97.8% accuracy, 97.0% F1) at substantially lower computational cost than the deep-learning alternatives; a leakage-controlled evaluation, however, shows that much of this reflects evaluation artifacts, with the honest macro-F1 falling to 0.558 and reliable detection of the scarce benign (minority) class remaining unverified. Cross-dataset experiments are reported as a preliminary analysis; because the feature spaces could only be approximately harmonized, these results are indicative rather than conclusive evidence of cross-domain generalization. Robustness analysis under Gaussian noise injection (5%–15%) and random feature perturbation (not gradient-based adversarial attacks) reveals that ensemble models exhibit significantly lower performance degradation compared to single-tree approaches. Statistical significance testing (paired <i>t</i>-test, <i>p</i> &lt; 0.05) validates that the observed improvements are not attributable to random variation. These results show that, once evaluation artifacts are controlled for, the honest performance of these lightweight models is modest, underscoring the importance of leakage-aware evaluation for smart-home IoT intrusion detection.},
DOI = {10.32604/cmc.2026.085248}
}