Adversarial feature generation for ML-based intrusion detection in the petrochemical industry

Machine learning-based intrusion detection systems (IDS) have been adopted widely to secure industrial control systems (ICS), including the petrochemical industry, against evolving cyber threats. However, recent studies show that these IDS are often ineffective against adversarial attacks. Therefore, in this study, we examine latent-space vulnerabilities in variational autoencoder (VAE)-based IDS by proposing an adversarial feature generation method, which manipulates the reconstructions of VAE and misclassifies intrusions as normal activity. By perturbing inputs to match latent representations with benign distributions, we demonstrate that VAEs are susceptible to evasion rates of 96% for Decision Tree (DT) and 100% for SVM classifiers, outperforming GAN-based attacks by 15% on a simulated fluid-catalytic cracking (FCC) dataset. Statistical validation using Wilcoxon test ($p<0.001$) confirms that normalization processes linearly degrade reconstruction fidelity, exposing predictable adversarial dynamics. These findings emphasize the urgent need for adversarial training and compact latent-space designs to enhance the resilience of ML-based IDS in mission-critical petrochemical ICS.