EVALUATING FEATURE DIMENSIONALITY REDUCTION TECHNIQUES FOR MACHINE LEARNING-BASED IoT INTRUSION DETECTION

Abstract

The rapid expansion of large-scale computer networks has led to a substantial increase in data traffic and system complexity, thereby heightening the risk of sophisticated cyberattacks. Because traditional intrusion detection methods often require considerable computational resources, the demand for lightweight yet effective detection mechanisms has become increasingly urgent. This study proposes and evaluates a flexible network attack detection framework that integrates dimensionality reduction techniques with high-performance classifiers. Two feature reduction methods Chi-square and Principal Component Analysis (PCA) are employed to optimize the feature set, followed by a comparative assessment using two classification models: a Deep Neural Network (DNN) and Extreme Gradient Boosting (XGBoost). The proposed system is trained and validated on publicly available benchmark datasets, demonstrating its ability to maintain high detection accuracy while significantly lowering computational costs. Experimental results further reveal that combining feature reduction with advanced machine learning models improves processing speed, reduces resource consumption, and enhances overall detection performance. These findings confirm the practicality and efficiency of the proposed framework for modern intrusion detection environments.