Comparison of Multiple Feature Selection Techniques for Machine Learning-Based Detection of IoT Attacks

Abstract

The Internet of Things (IoT) has become increasingly practical in applications such as smart homes, autonomous vehicles, and environmental monitoring. However, this rapid expansion has led to significant cybersecurity threats. Detecting these threats is critical, and while machine learning techniques are valuable, they struggle with high-dimensional data. Feature selection helps by reducing computational costs while maintaining model generalization. Selecting the most effective feature selection method is a crucial task. This research addresses this gap by testing five feature selection methods: Random Forest (RF), Recursive Feature Elimination (RFE), Logistic Regression (LR), XGBoost Regression (XGBoost), and Information Gain (IG) using the CIC-IoT 2023 dataset. It evaluates these methods when being used with five machine learning models: Decision Tree (DT), Random Forest (RF), k-Nearest Neighbors (k-NN), Gradient Boosting (GB), and Multi-layer Perceptron (MLP) using metrics like accuracy, precision, recall, and F1-score across three datasets. The results show that RFE, especially with the RF model, achieves the highest accuracy (99.57%) with 30 features. RF is the most stable, with accuracy from 83% to 99.56%. Additionally, the 5-feature scheme is best for implementing IDS on resource-limited IoT devices, with RFE paired with the k-NN model being the optimal combination.The Internet of Things (IoT) has become increasingly practical in applications such as smart homes, autonomous vehicles, and environmental monitoring. However, this rapid expansion has led to significant cybersecurity threats. Detecting these threats is critical, and while machine learning techniques are valuable, they struggle with high-dimensional data. Feature selection helps by reducing computational costs while maintaining model generalization. Selecting the most effective feature selection method is a crucial task. This research addresses this gap by testing five feature selection methods: Random Forest (RF), Recursive Feature Elimination (RFE), Logistic Regression (LR), XGBoost Regression (XGBoost), and Information Gain (IG) using the CIC-IoT 2023 dataset. It evaluates these methods when being used with five machine learning models: Decision Tree (DT), Random Forest (RF), k-Nearest Neighbors (k-NN), Gradient Boosting (GB), and Multi-layer Perceptron (MLP) using metrics like accuracy, precision, recall, and F1-score across three datasets. The results show that RFE, especially with the RF model, achieves the highest accuracy (99.57%) with 30 features. RF is the most stable, with accuracy from 83% to 99.56%. Additionally, the 5-feature scheme is best for implementing IDS on resource-limited IoT devices, with RFE paired with the k-NN model being the optimal combination.