Wireless Sensor Network Intrusion Detection: A Review of Statistical, Machine Learning, and Federated Methods

Abstract

Wireless Sensor Networks have become essential infrastructure across environmental monitoring, healthcare, smart agriculture, industrial automation, and smart city applications. However, their distributed, resource-constrained, and physically accessible design creates serious security risks. This paper surveys cybersecurity challenges and intrusion detection in WSNs, tracing the field from statistical methods to machine learning, deep learning, explainable AI, and federated learning. Our analysis shows that while signature-based and anomaly-based detection provide foundational security, they face significant limitations in dynamic, resource-limited WSN environments. Machine learning and deep learning achieve strong results, Random Forest reaching 98–99.5% accuracy and deep learning architectures exceeding 99% in some studies, but these figures depend heavily on dataset, preprocessing, and evaluation choices. Explainable AI techniques such as SHAP and LIME address the "black-box" problem by enabling transparent security decisions, while federated learning enables privacy-preserving, communication-efficient collaborative detection. Significant gaps remain: integrated frameworks that detect both data theft and eavesdropping are lacking, passive attacks receive limited attention, developing-region contexts are under-addressed, and energy-aware adaptive security remains aspirational. We identify future directions including lightweight deep models, energy-aware adaptive security, hybrid architectures, and robust, poisoning-resilient federated mechanisms.