Feature Selection versus Feature Extraction Models for IoT Intrusion Detection System Using Convolutional Neural Network

Intrusion Detection Systems (IDS) are vital for securing Internet of Things (IoT) environments, which face escalating cyber threats due to the proliferation of interconnected devices. Traditional IDS methods that rely on signature-based detection fall short in addressing these evolving threats. This study aims to enhance IDS capabilities by evaluating feature selection and extraction techniques for multiclass label detection using Convolutional Neural Networks (CNNs). We preprocess the University of Nevada - Reno Intrusion Detection Dataset (UNR-IDD) through cleaning and normalization, followed by dimensionality reduction. Techniques such as Chi-Square, Lasso Regression, and Mutual Information were used for feature selection, while Principal Component Analysis (PCA), Factor Analysis, and Non-Negative Matrix Factorization (NMF) were used for feature extraction. Two CNN models, incorporating Conv1D and Conv2D layers, were trained and assessed. The findings reveal that Factor Analysis is the most effective feature extraction method, achieving accuracies of $86.1 \%$ with Conv1D and $87.6 \%$ with Conv2D CNNs, while Mutual Information excelled in feature selection, with accuracies of 78.5% for Conv1D and 79.2% for Conv2D CNNs. These results highlight the significant improvement in IDS performance when using advanced dimensionality reduction techniques. Future research should explore combining these techniques and applying them to diverse datasets to further enhance IDS accuracy and robustness across various IoT scenarios.