Biometric Leakage from Generative Models and Adversarial Iris Swapping for Spoofing Eye-based Authentication

Abstract

This thesis investigates the vulnerability of generative models trained on biometric data and explores digital spoofing attacks on iris-based authentication systems representative of AR/VR environments. We first explore how diffusion models trained on biometric data can memorize and leak iris images. Next, we evaluate the effectiveness of Cross-Attention GANs for iris-swapping attacks, demonstrating their ability to enable presentation attacks that spoof iris-recognition systems. Our experiments across several standard iris and VR datasets have an attack success rate of 100% within similar domains and generalize across domains with rates as high as 70%. Our findings highlight the need to consider vulnerabilities in biometric systems and strengthen defenses against digital presentation attacks produced by generative models.