A Gradient Boosting Framework with Phi-K Correlation-Based Feature Selection for Intrusion Detection in IoT Networks
Internet of Things (IoT) devices generate vast network traffic but can hardly run heavyweight intrusion detection systems due to resource constraints. Moreover, the signature-based detectors often fail against zero-day attacks. This study proposes a lightweight detection framework that combines Phi-K correlation-based feature selection with gradient boosting. By using the RT-IoT2022 dataset of over 120,000 real IoT traffic flows, I reduced 82 candidate features to 34 (a 58.5% reduction) and compared nine models spanning logistic regression, decision trees, bagging, random forests, and XGBoost. The tree-based ensembles clustered at the top of the comparison, with no meaningful difference between them; the tuned XGBoost model achieved 99.87% accuracy, 99.92% recall, 99.93% precision, and an F1-score of 0.9993 on held-out test data, with a false-negative rate of 0.08% and a false-positive rate of 0.64%. The feature-importance analysis shows that timing features are dominant: maximum flow inter-arrival time (40.9%) ranks first, and maximum forward inter-arrival time (27.7%) ranks second, which suggests that a small set of timing features can effectively support accurate, resource-efficient detection at IoT gateways. Limitations include a single-dataset evaluation and bias from random splitting, and I also address the outline of extensions for multi-class classification and strategies to improve robustness.
Authors
- Emily Ma
Publication Details
- Journal
- American Journal for Young Scientists
- Published
- 2026-11-13
- DOI
- https://doi.org/10.67194/ajfys.v3i2.001
- Primary Topic
- Network Security and Intrusion Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00