Optimizing Feature Selection and Classification for Multi-Stage Cyber Attacks in IoT Networks

The rapid expansion of the Internet of Things (IoT) has led to an exponential increase in networkconnected devices, making IoT ecosystems highly vulnerable to cyber-attacks. Addressing the challenges of complex IoT traffic and multi-stage cyber threats, this research proposes a comprehensive machine learning framework that integrates feature engineering and ensemble modelling techniques. The framework employs random forest algorithms to enhance attribute selection, offering alternative insights compared to conventional techniques such as principal component analysis (PCA). A combination of base classifiers, including Support Vector Machine (SVM), AdaBoost, XGBoost, and Random Forest (RF), is integrated through a meta-model to capitalize on the strengths of diverse algorithms. The study further enhances attack prediction accuracy using hard and soft voting mechanisms, enabling probabilistic confidence in classifications. Robustness is validated through k-fold cross-validation, ensuring consistent model performance across varied datasets. Additionally, timing performance is assessed to identify the optimal configuration for real-time prediction in IoT systems. Experimental results demonstrate that the proposed framework effectively handles the intricacies of IoT traffic and provides reliable, efficient, and scalable solutions for predicting and mitigating cyber-attacks.