Efficient IoT Intrusion Detection Using Mutual Information And Pca-Based Feature Reduction With Ensemble Machine Learning

The rapid proliferation of Internet of Things (IoT) devices has significantly expanded the attack surface for cyber threats, making robust Intrusion Detection Systems (IDS) indispensable. This paper presents an efficient machine learning-based IDS for IoT networks leveraging Mutual Information (MI) for feature selection and Principal Component Analysis (PCA) for dimensionality reduction, applied to the IoT-23 dataset. MI identifies six discriminative features from eight preprocessed attributes, which PCA further transforms into four principal components preserving 99% of the total data variance. Four ensemble classifiers—Random Forest (RF), XGBoost, Bagging-SVM (BG-SVM), and AdaBoost—are evaluated under Bayesian hyperparameter optimization with 5-fold cross-validation. All models achieve accuracy exceeding 99.7%, with RF attaining peak performance at 99.834% accuracy, 99.589% precision, 99.753% recall, and 99.671% F1-score. Compared with the notable reference deep learning ensemble proposed in [14] on the identical dataset, the proposed method surpasses in accuracy (+0.584%), recall (+4.753%), and F1-score (+1.671%) while requiring significantly lower computational resources. The results confirm the effectiveness of MI-PCA feature engineering for lightweight, high-accuracy intrusion detection in resource-constrained IoT environments.