Network Anomaly Detection in the Internet of Things (IoT)

Abstract

This study comprehensively explores developing an efficient network anomaly detection system for the Internet of Things (IoT) through advanced machine learning techniques. The methodology encompasses data collection, preprocessing, feature engineering, and evaluating multiple machine-learning models. Random Forest emerges as the top-performing model, demonstrating impressive accuracy (98.11%), sensitivity (75.86%), specificity (98.71%), and G-Mean (86.53%). Decision Tree and K-Nearest Neighbors also exhibit commendable performances, highlighting the effectiveness of diverse machine-learning approaches in IoT anomaly detection. The proposed ensemble model, integrating Random Forest, XGBoost, and K-Nearest Neighbors, surpasses individual models with an accuracy of 98.14%, sensitivity of 78.75%, specificity of 98.62%, and a G-Mean of 88.12%. Leveraging a complex voting criterion and meticulous optimization through grid search enhances the model's predictive capabilities. Addressing class imbalance using the Synthetic Minority Over-sampling Technique (SMOTE) significantly improves sensitivity, specificity, and G-Mean. Sensitivity increases to 81.25%, specificity improves to 98.96%, and the G-Mean rises to 89.51%, enhancing overall model performance. Future research directions include exploring and optimising more sophisticated ensemble models, real-world deployment of the proposed model in diverse IoT scenarios, investigation of techniques for adapting to dynamic changes in IoT network behaviour, advanced hyperparameter tuning, and addressing potential vulnerabilities and security concerns. The study lays a solid foundation for effective IoT network anomaly detection, providing insights that can contribute to advancing anomaly detection techniques in the ever-evolving landscape of the Internet of Things.