Meta-IDS-GAN: Utilizing Meta Learning to Enhance GAN-based Adversarial Traffic Generation

Abstract

In this thesis, we present Meta-IDS-GAN, a generic adversarial traffic generation framework that combines Model-Agnostic Meta-Learning (MAML) with Wasserstein Generative Adversarial Networks (WGAN) to enhance the transferability and authenticity of adversarial samples against black-box intrusion detection systems (IDS). By leveraging meta-learning, the generator adapts to multiple IDS models, producing adversarial traffic that effectively evades detection while preserving the statistical characteristics of real malicious traffic. Our experiments comparing Meta-IDS-GAN with IDS-GAN and AdvGAN demonstrate that Meta-IDS-GAN achieves the lowest detection rate across various IDS architectures, significantly improving attack generalization without sacrificing realism. The analysis based on Wasserstein distance and variance further confirms that Meta-IDS-GAN generates adversarial samples that balance evasiveness and authenticity. This study underscores the growing threat of meta-learned adversarial attacks in cyber security and highlights the need for IDS models to develop adaptive defense mechanisms against evolving adversarial strategies.