Enhanced Intrusion Detection Systems Dataset Synthesis Using Conditional Generative Adversarial Networks with the Adaptive Whale Optimization Algorithm

Abstract

Modern networks, especially in the IoT era, require intrusion detection systems (IDSs) to ensure security and integrity. The growing diversity and data-driven nature of cyberattacks necessitate the generation of high-quality synthetic attack data for training robust detection models. In this domain, Generative Adversarial Networks (GANs) have proven to be a versatile and powerful tool. However, GAN training is unstable, as it often fails to converge and produces suboptimal outputs. To address these challenges, this study proposes a Conditional GAN (CGAN) enhanced with the Adaptive Whale Optimization Algorithm (AWOA). Relax-Alpha is a feedback mechanism that adjusts the generator’s updates based on trends in discriminator accuracy. It aims to provide dynamic training control, stabilizing training, and increasing the realism and diversity of synthetic samples. Several adaptation strategies were evaluated, including discrete and continuous alpha adjustments. The model was tested on the CIC IoT 2023 dataset through 5-fold cross-validation. The experiments show strong classification and generation quality; CGAN-AWOA achieved 98.59% accuracy.