Identification of Attacks in WSN using AI & ML

Abstract

Wireless sensor networks (WSNs) have become indispensable in smart cities, healthcare surveillance, environmental perception, and industrial automation, but the nodes are extremely susceptible to denial-of-service attacks that can halt important processes due to limited resources Although recent developments have seen artificial intelligence and machine learning as intrusion detection systems, the existing methods have been associated with excessive computational overhead, lack of support to extreme class imbalance, lack of explainability, and lack of evaluation of energy consumption and inference latency, which are key considerations when implementing them in practical battery-powered WSN contexts The paper introduces a sparse, interpretable ensemble machine learning model, which combines Particle Swarm Optimization to tune the hyperparameters with a soft-voting ensemble of Random Forest, Decision Tree, and K-Nearest Neighbors classifiers with SMOTE-Tomek balancing and SHAP-driven interpretability The proposed model yields the highest evaluation performance on the WSN-DS dataset, with the highest accuracy of 99.85% and the highest F1-score of 99.84% with low energy consumption of 12.4 mJ and inference latency of 8.7 ms, which is much higher than the baseline and state-of-the-art models These findings illustrate a feasible and explainable remedy that can further the safe, energy-effective attack detection in community-based sensor networks.