Enhanced Security for Wireless Sensor Networks through Lightweight Machine Learning-Based Anomaly Detection and Attack Classification
Wireless Sensor Networks (WSNs) have become indispensable components of modern cyber-physical systems, supporting healthcare, industrial automation, agriculture, environmental monitoring, and military surveillance. However, their limited computational resources, open wireless communication channels, and unattended long-term deployments expose them to diverse cyberattacks, whilst conventional security mechanisms remain inadequate for such constrained environments. This study aims to develop and evaluate an efficient Machine Learning-based framework for real-time anomaly detection and multi-class attack classification in WSNs. The objective is to enhance network security, maintain detection accuracy, and enable practical deployment on low-resource sensor hardware. Seven supervised and unsupervised learning models, including Random Forest, Support Vector Machine, XGBoost, LSTM, Isolation Forest, Autoencoder, and k-Nearest Neighbour, were assessed using three benchmark datasets: NSL-KDD, UNSW-NB15, and WSN-DS. A tailored feature engineering process selected the most informative network, temporal, and protocol attributes, followed by Bayesianoptimized ensemble learning. The proposed soft-voting ensemble outperformed individual models and achieved highly accurate intrusion detection with strong classification consistency and minimal false alarm rates. Cross-dataset evaluation confirmed robust generalization, while hardware profiling demonstrated low memory usage and fast inference, making the framework feasible for real-time embedded WSN deployment. This research provides a scalable and practical intelligent security solution for WSN environments. Its originality lies in combining high detection performance, lightweight deployment capability, and adversarial robustness analysis, offering significant value for securing future smart infrastructure and resource-constrained IoT systems.