This paper investigates machine learning driven cyberattack detection in Internet of Things networks. It tackles challenges posed by high-dimensional data and devices with limited resources. The study focuses on feature reduction methods to improve Edge AI efficiency. It compares feature selection techniques, such as Random Forest importance and Recursive Feature Elimination, with feature extraction methods, including Principal Component Analysis and Linear Discriminant Analysis. Several lightweight models are evaluated: Decision Tree, Random Forest, Logistic Regression, Multi-Layer Perceptron, and LightGBM. These models are tested using the CICIoMT2024 dataset for both binary and multi-label classification tasks. Performance is measured by accuracy, precision, recall, F1-score, and inference time on a workstation and a Raspberry Pi. The results reveal that feature selection outperforms feature extraction with appropriate frameworks. Decision Tree and Random Forest achieve the best result: 99.89% accuracy in binary classification and 99.61% in multi-label tasks when using Random Forest feature selection with five selected features. On the Raspberry Pi, Decision Tree stands out with inference times of 11.94 s for binary tasks and 25.73 s for multi-label tasks, making it suitable for edge computing. This research provides a practical guide for enhancing Internet of Things security across resource-constrained devices.