Enhancing WSN Intrusion Detection: Two-Tier Feature Selection and Optuna- Optimized Ensemble Learning

The decentralized and resource-constrained nature of WSNs exposes them to multiple attacks. Intrusion detection serves as a vital protective measure for the security of wireless sensor networks (WSNs). This study introduces, a new Optuna-based stacking ensemble learning algorithm (OXCRF) was introduced for WSN intrusion detection. This method includes a two-stage feature selection method with the integration of SHapley Additive exPlanations (SHAP) and CatBoost, Mutual Information (MI), and cross-validated Recursive Feature Elimination (RFE) using a Random Forest classifier. This method efficiently reduces dimensionality and improves the model efficiency. The Synthetic Minority Oversampling Technique (SMOTE) handles data imbalance. The stacking ensemble utilizes XGBoost and CatBoost as base learners and a Random Forest as the meta-learner, whose hyperparameters are tuned using the Optuna framework. Model testing was performed using the NSL-KDD dataset for binary and multiclass intrusion predictions. The experimental results demonstrate that OXCRF achieves superior performance compared to individual classifiers and existing methods, with accuracies of 99.60% and 99.53% and misclassification rates of 0.0040 and 0.0047 for binary and multiclass classification, respectively. An ablation study proved the functionality of every component in the OXCRF framework for successful multiclass intrusion detection in WSNs, particularly with class overlap and data imbalance. These results validate the efficacy of the designed framework for improving intrusion detection in WSNs via effective feature selection, suitably balanced data processing, and ensemble learning optimization.