Comparative Study of Feature Selection Methods for Deep Learning–based Intrusion Detection in IoT Networks

As the widespread proliferation of Internet of Things (IoT) networks, securing devices that are networked together is now a serious issue. Intrusion Detection Systems (IDS) based on deep learning have emerged as an intriguing solution due to their ability to learn complex patterns in network traffic. Yet, high-dimensional input data—since they are often redundant or irrelevant features—may hinder detection effectiveness and increase computational costs, particularly on resource-constrained IoT gateways. This paper is an extensive comparison of three distinct feature selection methods—Principal Component Analysis (PCA), Mutual Information–guided Multi-population Differential Evolution (MI-MPODE), and Statistical Moments ranking—implemented in the same 1D Convolutional Neural Network (CNN) architecture and compared on the CIC-IoT-2023 dataset. We provide complete theoretical derivations, mathematical expressions, and implementation details for each method. Experiments with 80/20 train/test split across 687,179 flows with labels for 34 classes of traffic prove that MI-MPODE has the highest detection accuracy (81.27%) with the least feature reduction of over 50%. PCA yields 75.38% with only 15 principal components, and Statistical Moments yield 48.99% with 30 features. We offer extensive per-class performance, computational overhead, and practical considerations for deploying IDS on IoT platforms.