Intrusion Classifier Architecture Generation with Zero Prior Knowledge Employing Generative Adversarial Networks

In this study, we introduce the IoT Intrusion Detector Architecture Generator (IoT-IDAG), a novel approach utilizing Generative Adversarial Networks (GANs) to generate optimal Deep Learning-based Network Intrusion Detection System (NIDS) architecture configurations without relying on prior knowledge. IoT-IDAG addresses the limitations of traditional Neural Architecture Search (NAS) methods, which are constrained by predefined search spaces, high computational costs, and produce a single architecture per training session. Our approach enhances the generation efficiency of NIDS architecture configurations by reducing computational cost and convergence time, while producing multiple NIDS architecture configurations in a single training session. The proposed model includes a generator that creates architecture descriptors and a discriminator that ensures these descriptors fit within a given distribution. Original methodologies for neuron mapping and hyperparameter computation have also been introduced to convert architecture descriptors into a feasible and implementable NIDS. Experimental results show that IoT-IDAG generates intrusion classifier architecture configurations that outperform hand-crafted DL-based NIDS in terms of detection accuracy, recall, and F-1 score.