Wireless Sensor Networks (WSNs) are widely deployed across many critical infrastructures because of their low cost, flexibility, and efficiency, besides their broad applicability across various domains. A normal WSN is basically composed of small, battery-powered nodes that monitor environmental conditions and endorse real-time applications; thus, WSNs are very prone to security threats, especially Denial of Service (DoS) attacks that disrupt network functionality. The traditional security mechanisms tend to fall short, as WSNs lack the computational, memory, and energy resources for complex classification algorithms. In this regard, we aim to develop a resilient Intrusion Detection System (IDS) optimized for WSNs, capable of real-time detection and mitigation against security breaches. Our IDS leverages an online ensemble learning approach based on the Adaptive Random Forest with Hoeffding Adaptive Tree (ARF+HAT) algorithm, alongside three dimensionality reduction (DR) techniques: Gaussian Random Projection (GRP), Random Bernoulli Projection (RBP), and Hashing Trick (HT). The IDS operates in three online phases: pre-processing, dimensionality reduction, and attack detection, where DR selectively retains the most relevant features to reduce computational demands. Extensive simulations validate the IDS’s effectiveness, with HT combined with the ARF+HAT achieving high accuracy, precision, recall, and F1-score, while offering superior memory and runtime efficiency. HT’s low false alarm rate and enhanced energy efficiency underscore its suitability for real-time intrusion detection in resource-constrained WSN environments, positioning it as a practical and robust solution for improving WSN security.