Evaluating Machine Learning Efficacy for DoS Intrusion Detection in Wireless Sensor Networks

Abstract

Wireless Sensor Networks (WSNs) are integral to mission-critical applications, including environmental monitoring, smart infrastructure, and healthcare. However, they are particularly vulnerable to denial-of-service (DoS) attacks, which can deplete the node's energy and disrupt communication. This study examines the effectiveness of various machine learning algorithms in enhancing intrusion detection within WSNs, focusing on balancing detection accuracy and computational efficiency. Utilising the Network Simulator-2 (NS-2) generated WSN-DS dataset, seven algorithms—K-Nearest Neighbours (KNN), Decision Tree (DT), Random Forest (RF), Naïve Bayes (NB), Stacking Classifier, AdaBoost, and Artificial Neural Network (ANN)—were implemented and evaluated. The experimental results indicate that AdaBoost achieved the highest overall performance, with an accuracy of 99.7%, ROC-AUC of 0.996, and detection speed of 1.6 min, underscoring its suitability for real-time intrusion detection. Stacking and Random Forest also demonstrated high accuracy (99.7% and 99.6%, respectively) but required slightly longer detection times of 7.07 and 7.33 min, respectively. In contrast, KNN exhibited the longest detection time (86.2 min) due to its high computational overhead, whereas Naïve Bayes was the fastest (0.02 min) but had lower precision (0.757) and F1 score (0.771). AdaBoost demonstrated superior detection accuracy, efficiency, and adaptability under constrained WSN conditions, outperforming all other algorithms across multiple performance metrics. These findings offer a practical benchmark for developing lightweight, high-performance intrusion detection systems in resource-limited wireless sensor environments, thereby enhancing the resilience and reliability of next-generation WSN infrastructures.