Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion

David Geissbühler, Hatef Otroshi Shahreza, Sébastien Marcel

TL;DR: We introduce a new method inspired by the physical motion of soft particles subjected to stochastic Brownian forces, allowing us to sample identities distributions in a latent space under various constraints.

Abstract: Face recognition models are trained on large-scale datasets, which have privacy and ethical concerns. Lately, the use of synthetic data to complement or replace genuine data for the training of face recognition models has been proposed. While promising results have been obtained, it still remains unclear if generative models can yield diverse enough data for such tasks. In this work, we introduce a new method, inspired by the physical motion of soft particles subjected to stochastic Brownian forces, allowing us to sample identities distributions in a latent space under various constraints. We introduce three complementary algorithms, called Langevin, Dispersion, and DisCo, aimed at generating large synthetic face datasets. With this in hands, we generate several face datasets and benchmark them by training face recognition models, showing that data generated with our method exceeds the performance of previously GAN-based datasets and achieves competitive performance with state-of-the-art diffusion-based synthetic datasets. While diffusion models are shown to memorize training data, we prevent leakage in our new synthetic datasets, paving the way for more responsible synthetic datasets. Project page: https://www.idiap.ch/paper/synthetics-disco

Lay Summary: Existing real face recognition datasets are collected from the web, raising ethical and privacy concerns. This paper presents a new method for generating synthetic face recognition datasets, inspired by physics motion (i.e., Brownian motion) of tiny particles in fluids based on granular mechanics. Three algorithms are developed (called Langevin, Dispersion, and DisCo) that generate diverse synthetic faces by applying these physical processes to image generation. The generated datasets are used to train face recognition models, which are evaluated on a diverse set of benchmarks. Furthermore, this work promotes safe and ethical use of AI technologies in face recognition applications.

Link To Code: https://www.idiap.ch/paper/synthetics-disco

Primary Area: Social Aspects->Privacy

Keywords: Brownian Identity Diffusion, Face Recognition, Latent Space, Synthetic Dataset

Flagged For Ethics Review: true

Submission Number: 10820