Optimized Feature Reduction Techniques for Enhanced Network Threat Detection in Wireless Sensor Networks

Abstract

The security of Wireless Sensor Networks (WSNs) is currently seriously threatened by numerous threats. Consequently, a number of applications are offered to regulate data and information sharing along with the related security features that need to be maintained throughout data transfer. This study suggests an intelligent feature reduction methodology based on machine learning that uses Modified Principal Component Analysis (MPCA) to identify the properties most associated with the attacked classes that are being used. This could help with the machine learning model's complexity. The WSN-DS dataset was used to implement and test the suggested approach. This approach performs very well in intrusion detection for WSNs, attaining great accuracy and dependability. The proposed framework involves three key stages: (1) preprocessing the WSN-DS dataset, (2) applying MPCA to identify and retain the most critical features, and (3) implementing and testing multiple machine learning algorithms including Random forest (RF), Gradient Boosting (GB), Decision Tree (DT), Naive Bayes (NB), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Neural Networks (NN), on both the original and reduced feature sets. The experiment demonstrates that hybrid feature reduction techniques significantly enhance computational efficiency while maintaining or improving performance, particularly for robust algorithms like RF and GB. RF achieved near-perfect metrics across multiple attack types, with an F-measure of 99.92% for Flooding attacks and an increased recall of 99.70% for Blackhole attacks after reduction. These findings underscore the importance of algorithm selection and feature optimization tailored to specific attack scenarios, establishing hybrid feature reduction as a valuable approach to enhancing threat detection in WSNs.