da/sec scientific talk on Biometrics                            

Topic: AgeDiff: Latent Diffusion-based Face Age Editing with Dual Cross-Attention

by Marcel Grimmer
D19/2.03a, June 19, 2024 (Wednesday), 12.00 noon

Keywords — Face Recognition, Forensic analysis, Face Age Editing, Deep Learning, Diffusion Models

Abstract

The facial appearance changes over time due to natural ageing effects. In the context of face recognition (FR), comparing two mated samples with significant age differences can negatively impact recognition performance, as biometric systems presume the permanence of identity features. Typical application scenarios in which FR operators encounter large age differences include forensic investigations and migration management. To address the high demand for evaluating how facial ageing affects recognition performance, we propose AgeDiff, a Latent Diffusion-based Face Age Editing (FAE) model using a Dual Cross-Attention mechanism to ensure age transformation with high identity preservation and ageing accuracy. In this scientific talk, we introduce AgeDiff and demonstrate how FAE techniques can be used to evaluate the robustness of FR systems against facial ageing.