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

Wireless Sensor Networks (WSNs) are key for ubiquitous computing. Despite advantages, they face securitychallenges due to decentralized nature and threats. Intrusion detection helps protect WSNs from securitythreats. This study proposes an Optuna-implemented stacking technique (OXCRF) the method combinesSHapley Additive exPlanations, CatBoost, Mutual Information, and cross-validated Recursive FeatureElimination with Random Forest for feature selection, while SMOTE handles data imbalance. The stackingensemble, XGBoost, CatBoost and Random Forest are used as the base learners, with hyperparametersbeing optimized using Optuna. Experiments on the NSL-KDD and UNSW-NB15 datasets show that OXCRFachieves higher accuracy (99.60% for binary and 99.53% for multiclass on NSL-KDD; 98.62% for binaryand 83.67% for multiclass on UNSW-NB15) and lower misclassification rates (0.0040 and 0.0047 on NSLKDD; 0.0138 and 0.1633 on UNSW-NB15) compared to baseline models. Running an ablation studyshowed that OXCRF components worked as expected for multiclass intrusion detection in WSNs withoverlapping classes and imbalanced data. The framework is efficient through feature selection, balanceddata distribution and improved ensemble learning.