Deep Reinforcement Learning-Based Asymmetric Convolutional Autoencoder for Intrusion Detection

Abstract

In recent years, intrusion detection systems (IDSs) have become a critical component of network security, due to the growing number and complexity of cyber-attacks. Traditional IDS methods, including signature-based and anomaly-based detection, often struggle with the high-dimensional and imbalanced nature of network traffic, leading to suboptimal performance. Moreover, many existing models fail to efficiently handle the diverse and complex attack types. In response to these challenges, we propose a novel deep learning-based IDS framework that leverages a deep asymmetric convolutional autoencoder (DACA) architecture. Our model combines advanced techniques for feature extraction, dimensionality reduction, and anomaly detection into a single cohesive framework. The DACA model is designed to effectively capture complex patterns and subtle anomalies in network traffic while significantly reducing computational complexity. By employing this architecture, we achieve superior detection accuracy across various types of attacks even in imbalanced datasets. Experimental results demonstrate that our approach surpasses several state-of-the-art methods, including HCM-SVM, D1-IDDS, and GNN -IDS, achieving high accuracy, precision, recall, and F1-score on benchmark datasets such as NSL-KDD and UNSW-NB15. The results emphasize how effectively our model identifies complex and varied attack patterns. In conclusion, the proposed IDS model offers a promising solution to the limitations of current detection systems, with significant improvements in performance and efficiency. This approach contributes to advancing the development of robust and scalable network security solutions.