A Hybrid Deep Learning Model for Intrusion Detection in IoT Networks

The increasing complexity of network structures and the expanding number of network devices have rendered network security more vulnerable to sophisticated attacks, necessitating the implementation of an effective Intrusion Detection System to ensure data confidentiality, integrity, and availability. However, an accurate and robust intrusion detection system is scarce as it suffers from several challenges including class imbalance, high dimensionality, and inadequate performance. Therefore, this study proposes a deep learning-based hybrid model to address the identified challenges of Intrusion Detection Systems. Furthermore, the performance of the proposed model, alongside several deep learning models, was evaluated and compared for multi-class classification using a combined dataset titled CICIDS-Collection that compiles four distinct and current datasets. This dataset was optimized through a feature selection technique known as mutual information score which identifies the most relevant features. Additionally, sampling techniques such as undersampling and oversampling were employed to balance the data and a dimensionality reduction technique called Principal Component Analysis (PCA) was applied to reduce dimensionality while maintaining as much of the original information as feasible. Overall, the evaluation results demonstrated an accuracy of 99.24% for the hybrid model outperforming all utilized deep learning models in the study, thereby significantly impacting the application of Intrusion Detection Systems.