Enhancing Intrusion Detection in AIOT using Intelligent Feature Selection and Deep Learning Fusion

Abstract

As the industrial IoT and AIoT continue to rapidly advance, new security concerns have arisen as a result of the exponential growth in the volume of data transmitted via communication networks. When it comes to evolving cyberthreats, traditional safeguards like encryption and firewalls are often not enough. As a result, intrusion detection systems (IDS) are now essential for ensuring secure and reliable Internet of Things connectivity. This paper proposes a system that combines deep learning (DL) models with PCA and several feature selection procedures to enhance real-time intrusion detection. In order to improve classification performance and reduce dimensionality, five feature selection methods were evaluated and combined with principal component analysis (PCA): symmetrical uncertainty (SU) & Pearson analysis. Multiple classifiers were applied to the RT-IoT2022 dataset, including TabNet, DNNs, and ANNs. When compared to ANN (92.3614% accuracy) and TabNet  (94% accuracy), the combined performance of ANN, Pearson analysis, and PCA (98.6123% accuracy) was much better. Key attributes discovered were responsible for the performance increases. The results demonstrate that a powerful method for identifying threats in real-time AIoT environments may be achieved by integrating PCA with efficient feature selection, which in turn increases the accuracy and efficiency of IDS.