Jin Zhuotong successfully defended his Master’s thesis titled „IdentityPreserving Iris Image Generation Using Generative Methods.“ This is a joint effort of h-da (Germany) and DTU (Denmark).
Generating synthetic Near-Infrared (NIR) iris images that preserve biometric identity under strict verification conditions remains a major challenge, especially given limited real data and privacy constraints. While generative models can produce visually realistic images, they often fail to maintain identity, limiting their utility for recognition-focused data augmentation.
This thesis proposes a reference-conditioned masked latent diffusion framework operating in the latent space of a KL-regularized iris autoencoder. A latent „keep mask“ preserves the identity-carrying iris region, while controlled noise is injected into periocular areas to introduce diversity without degrading iris microtexture. The method is benchmarked against two baselines: (i) StyleGAN2-ADA with identity-supervised SupCon-style training, and (ii) deterministic pupil dilation.
Evaluated on the ND-LG4000-LR dataset using aligned protocols and two open-source recognition systems (Worldcoin, TripletNN), results show that GAN-based methods, despite visual plausibility, suffer from identity drift even during image inversion and fail under low-FMR verification. In contrast, the proposed diffusion model maintains strong identity preservation across settings, requires only minor threshold adjustments, and achieves FID scores near the real-data empirical lower bound.