A comparative study of dimensionality reduction techniques for intrusion detection in IoT networks

The widespread adoption of Internet of Things (IoT) technology has driven significant advancements in fields such as agriculture, manufacturing, industry, and transportation. However, the highly interconnected and resource-constrained nature of IoT ecosystems makes them particularly vulnerable to cyberattacks. Although AI-based intrusion detection systems provide an effective protection, their deployment on IoT devices is hindered due to limited memory, processing power, and storage capacity. One strategy for addressing these limitations is dimensionality reduction, consisting of the removal of redundant or irrelevant features in order to reduce computational demands without compromising model accuracy. This work analyses the effectiveness of various dimensionality reduction approaches for the development of efficient and lightweight Random Forest models for anomaly detection in IoT environments. Among the considered methods, Permutation Feature Importance consistently produced the most balanced models, reducing inference time, model size, and RAM usage, while slightly enhancing predictive performance. Furthermore, the feasibility of model deployment in real-world environments was assessed through experiments on a resource-constrained Raspberry Pi device.