Determine the optimal Hidden Layers and Neurons in the Generative Adversarial Networks topology for the Intrusion Detection Systems
Ali Lamjid, Khairul Akram Zainol Ariffin, Mohd Juzaiddin Ab Aziz, Nor Samsiah Sani · 2022 International Conference on Cyber Resilience (ICCR) · 2022
Neural Networks (NNs) are a vast field of research, particularly in terms of adjusting hyper-parameters such as hidden layers (HLs) and hidden neurons (HNs). On the other hand, (NN)s presents the main component of GAN's two elements: (1) Generator (G) and (2) Discriminator (D). It improves GAN as NN optimizes its parts. The number of Hidden Layers (HLs) and Hidden Neurons (HNs) inside the neural networks remains an important research topic despite numerous rules trying to predict the values by some directives and calculations. However, the study of the deepness and wideness of both (G) and (D) is still under-explored, and exploring this topic in GAN was neglected. Hence, this research focuses on the (NN)s contained in the original architecture of GAN and targets to clarify the relation between the number of (HL)s and (HN)s in the Generator (G) and the Discriminator (D) inside GAN. Furthermore, the goal is to determine which (NN) should have the most significant number of (HL) and (HN) in order to improve the performance of the created samples. Therefore, experiments were carried out by (1) increasing the number of (HL)s equally or differently in the G and the D and (2) increasing the number of (HN)s in each (HL) either on G or D. Deep Neural Network (DNN) is chosen as an intrusion detection system (IDS) model and the benchmark dataset KDD99 was used for application. It has been observed that a large number of (HL)s and (HN)s in the (G) improved GAN training and convergence. It means that the (G) should have more critical (HL)s and (HN)s than the (D) to generate better samples. Thus, this study provides the researchers with a vital resource to optimize the GAN's (NN)s and to process the experiments in this field by standardizing the number of (HL)s and (HN)s. The experiments carried out revealed that better training and convergence are seen when the number of (HL)s is 10 and the number of (HN)s is 1024 in the (G). The highest accuracy completed is 0.9991 which is 10 (HL)s and 1024 (HN)s in the Generator, and 2 (HL)s and 64 (HN)s in the Discriminator.