AI Summary & Key Takeaways
AI InsightThe study contributes a transparent and reproducible benchmark framework that can support future domain-specific validation in smart agriculture cybersecurity.
📌 3 Key Takeaways
Abstract
Smart agriculture increasingly depends on Internet of Things (IoT) devices, networked sensors, cloud platforms, and data-driven decision systems. These technologies improve agricultural productivity and monitoring, but they also expand the attack surface of farm cyber-physical systems. This study proposes a benchmark-based explainable machine learning framework for intrusion detection in smart agriculture. The empirical evaluation focuses on the detection and explainability components of the framework, using a consolidated NSL-KDD dataset created by merging CSV-formatted versions corresponding to KDDTrain+.TXT and KDDTest+.TXT and then applying a stratified 80/20 train-test split. After binary target construction, the dataset contains 77,053 normal records and 71,463 attack records. Random Forest and XGBoost models were trained and evaluated for binary intrusion detection. Random Forest achieved an accuracy of 0.9958 and an AUC of 0.9998, while XGBoost achieved an accuracy of 0.9959 and an AUC of 0.9999 under the custom experimental design. SHAP and LIME were used to interpret global feature contributions and local prediction behavior. The strongest signals included traffic-volume variables, service indicators, login-related variables, and host-based traffic statistics. These high values are interpreted cautiously because the study uses a merged and resampled benchmark setting rather than canonical NSL-KDD validation or real agricultural IoT traffic. The generative AI communication component is framed as a future user-centered extension rather than an empirically evaluated contribution. The study contributes a transparent and reproducible benchmark framework that can support future domain-specific validation in smart agriculture cybersecurity.