eXplicability AI (XAI) for attack detection toward smart rural applications

This research evaluates the performance and computational efficiency of various AI models for intrusion detection in IoT environments, with the goal of enabling future deployment in Smart Rural scenarios. Leveraging the massive NF-UQ-NIDS-v2 dataset-comprising over 76 million labeled NetFlow records across 21 traffic classes-we benchmark five models, ranging from classical machine learning algorithms to deep learning architectures, across both high-performance and low-performance execution setups. The analysis covers standard classification metrics (accuracy, precision, recall, F1-score) and detailed resource usage indicators, including inference time, memory footprint, CPU cycles, and energy consumption per batch. Additionally, explainable AI techniques (SHAP and LIME) are employed to investigate model behavior and feature relevance under real-world constraints. Results show that classical models, particularly Random Forest and Decision Tree, achieve top-tier detection accuracy while maintaining minimal computational demands, making them strong candidates for constrained deployments. Deep learning models deliver comparable predictive performance but incur significantly higher resource consumption, requiring further optimization for practical use. Overall, this work provides a comprehensive evaluation framework and practical insights for selecting efficient and interpretable AI-based intrusion detection systems for rural and low-resource infrastructures.