On the Generation of Synthetic Identity Parades

In the prosecution of criminals, identity parades serve as forensic indications in court, allowing eyewitnesses to identify suspects from a set of candidate images with similar demographic attributes. The composition of identity parades is governed by strict candidate selection criteria and privacy protection policies to ensure the fair treatment of all individuals involved. Motivated by this challenge, we propose generating semantically consistent candidate images by exploiting the disentangled latent spaces of modern Generative Adversarial Networks (GANs). Specifically, we leverage recent GAN inversion techniques to map suspect images into the latent space of StyleGAN2, enabling the generation of synthetic identity parade candidates within a neighbouring margin region. These candidates exhibit semantically similar characteristics to the suspect while maintaining distinctiveness from one another. We conduct our evaluation on the UNCW-MORPH mugshot dataset to simulate realistic forensic scenarios, demonstrating strong attribute preservation rates for suspect images that align with the training dataset distribution of StyleGAN2.