Toward robust IoT intrusion detection: a novel approach using SMOTETomek and grid search optimization

The proliferation of Internet of Things (IoT) devices has created significant cybersecurity vulnerabilities, yet many existing intrusion detection systems (IDS) are too computationally intensive for resource-constrained edge nodes. This study aims to develop a lightweight, interpretable IDS pipeline that balances detection accuracy with low resource consumption for real-time IoT environments. Our methodology utilizes train-only SMOTETomek resampling to address extreme class imbalance, chi-square feature selection for dimensionality reduction, and a grid search-optimized K-nearest neighbors (KNN) classifier, with explicit benchmarking of inference latency and memory footprint. Evaluated on the CIC-IoT-2023 dataset, the proposed pipeline achieved 94.3% accuracy, 94.5% precision, 93.9% recall, and a 94.2% F1-score, while maintaining a minimal model size and low per-packet inference latency suitable for edge deployment. The proposed approach provides a practical trade-off between detection performance and efficiency, offering a transparent and deployable security solution for constrained IoT hawkware.