<p>In this paper, the concepts of fog computing, machine learning and generative AI are utilized to construct an accurate and efficient intrusion detection system (IDS) for the Internet of Things. A hybrid approach in which the machine learning classifier for the IDS system is built using both real data and synthetic data is proposed. The real data used are collected at the edge of the network and deemed nonsensitive thereby preserving the security and privacy of the IoT environment. These real-time partial data are then used to train a generative AI model to produce synthetic samples that augment the minority class. We also implement principal component analysis (PCA) to minimize the system footprint of the developed models. To evaluate the proposal, we applied the approach to two datasets: an IoT traffic dataset (NSL-KDD) and an IoT telemetry dataset (ToN_IoT). The performance measures demonstrated a 75% reduction in the model loading time and a 58% reduction in the synthetic data generation time when PCA was applied to IoT traffic data. In addition, a classification accuracy of 86% was achieved for the IoT telemetry data using a random forest, while the precision, recall and F1 score attained for the minority-class labels improved, thus validating the effectiveness of the developed approach.</p>