A Distributed Generative–Probabilistic Framework for Scalable Intrusion Detection in Smart Farming IoT Networks

Smart farming environments rely heavily on interconnected IoT–fog–cloud infrastructures, making them increasingly vulnerable to sophisticated cyber threats. The objective of this study is to develop a robust, scalable, and interpretable intrusion detection framework tailored for heterogeneous and resource-constrained agricultural IoT systems. The proposed method integrates a multi-stage generative–probabilistic pipeline: Principal Component Analysis (PCA) and Isolation Forest (iForest) operate at the fog layer for lightweight dimensionality reduction and early anomaly isolation, while the cloud layer employs an Intrusion Detection System Generative Adversarial Network (IDSGAN) model for advanced adversarial feature learning and Gaussian Mixture Models (GMM) for probabilistic attack-type clustering. Experimental analysis conducted on CIC IoT 2023 and CIC DIAD IoT 2024 datasets demonstrates strong detection performance, achieving Accuracy between 98.47–99.15%, F1-scores up to 0.996, and AUC values of 0.981–0.987, outperforming baseline IDS models. Clustering metrics including Silhouette (99.01%), ARI (98.97%), and NMI (98.91%) confirm highly coherent attack-grouping across major categories such as DDoS, Mirai, Spoofing, and Web-based intrusions. The novelty of this work lies in its distributed architecture combining edge-efficient processing with cloud-level generative learning, enabling low-latency, high-throughput, and interpretable detection suitable for real-world smart farming ecosystems. This framework thus offers a scalable and high-performing solution for securing agricultural IoT infrastructures.