GAN-driven iterative method for automated IoT intrusion detection architecture design and scalable training-set creation

In this study, we propose an innovative iterative methodology employing Generative Adversarial Networks (GANs) to achieve two primary objectives: (i) develop a comprehensive knowledge base termed TAG (Training-set for Architecture Generation) to facilitate the automated training of GANs for creating architecture configurations of Network Intrusion Detection Systems (NIDS) without manual intervention, and (ii) utilize the iterative framework with TAG to derive optimal deep learning-based intrusion classifier architecture. Our proposed method overcomes the limitations inherent in traditional Neural Architecture Search (NAS), which is often restricted by a predefined search space and incurs high computational overhead. To the best of our knowledge, this study represents the first effort to create a dedicated knowledge base for training GANs, which serves as a training dataset to produce NIDS architecture configurations. The proposed approach not only enhances the efficiency of generating NIDS configurations by reducing computational costs and accelerating convergence but also introduces a novel strategy for developing a structured knowledge base to direct GAN training. Experimental evaluations demonstrate that employing TAG to guide GAN training results in the generation of deep learning-based intrusion classifier architecture with a reported accuracy of 99%, an optimal False Alarm Rate (FAR) of 0.01, precision of 97%, and recall of 99%.