AugGen: Generative Synthetic Augmentation Can Boost Face Recognition

Parsa Rahimi, Damien Teney, Sébastien Marcel

Keywords: Generative Model, Face Recognition, Synthetic Data

TL;DR: With our proposed sampling method, we demonstrate improved performance on a discriminative task when training on a mix of real and synthetic data, compared to training solely on the original real data.

Abstract:

As machine learning increasingly relies on large amounts of data, concerns about privacy and ethics have grown. Recently, methods for generating synthetic data to augment or replace real datasets have emerged to mitigate these concerns. In this paper, we demonstrate improved performance on a discriminative task when training on a mix of real and synthetic data, compared to training solely on the original real data. Our synthetic data is generated using a novel sampling method based on a conditional generative model and a discriminator, both trained exclusively on the original data, with no need for auxiliary data nor pre-trained foundation models. We consider the challenging task of face recognition, which is well known for its privacy and ethical issues. Using our augmented dataset, we demonstrate consistent improvements over the model trained on the original dataset, on various benchmarks including IJB-C and IJB-B by up to 5% while performing competitively with state-of-the-art synthetic data generation.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4882