da/sec scientific talk on Biometrics

Topic: Identity‐Preserving Synthetic Iris Generation via Latent Diffusion

by Fabian Stockhardt
D19/2.03a, June 18, 2025 (Wednesday), 12.00 noon

Keywords — Iris Recognition, Deep-learning, Image Generation, Identity Preservation, Latent Diffusion

Abstract

„Iris‐based biometrics remain one of the most reliable modalities for identity verification, but strict privacy regulations and the scarcity of large‐scale public datasets limit progress in deep-learning, cross‐sensor and “iris on the move” applications. In this work, we leverage latent diffusion models (LDMs) to generate synthetic near‐infrared (NIR) iris and periocular images that preserve subject identity. We conduct two main phases of synthesis:

Identity Reproduction: To increase intra‐subject variability, we fine‐tune LDMs on an existing real‐iris dataset to learn robust identity embeddings. These embeddings enable interpolation across session patterns, thereby extending each subject’s data diversity without sacrificing identity consistency.

New Subject Generation: By structuring a stable semantic space of learned identities, we try to sample “unseen” embeddings to produce entirely novel, privacy‐compliant iris identities.“