Network Intrusion Detection Using Variational Autoencoders: A Comparative Evaluation with Generative Adversarial Approaches

Variational autoencoders have emerged as a promising approach for anomaly-based network intrusion detection, offering principled probabilistic anomaly scoring through their latent space distributions. This paper presents a comprehensive comparative evaluation of variational autoencoders against generative adversarial approaches for network intrusion detection, including VAE-GANs, adversarial autoencoders, and standalone GAN-based detectors. We evaluate each approach across multiple dimensions including detection accuracy, false positive rate, zero-day detection capability, adversarial robustne ss, training stability, and computational efficiency. Our experimental methodology employs three benchmark datasets and standardized evaluation protocols to ensure fair comparison. Results demonstrate that VAEs provide the most stable training dynamics and best computational efficiency, while VAE-GANs achieve the highest detection accuracy on known attacks at the cost of increased training complexity. Adversarial autoencoders demonstrate the best adversarial robustness. Based on our analysis, we propose a hybrid architecture that combines the complementary strengths of VAEs and adversarial approaches.