In this study, we apply the information-theoretic Privacy Funnel (PF) model to face recognition and develop a method for privacy-preserving representation learning within an end-to-end trainable framework. Our approach addresses the trade-off between utility and obfuscation of sensitive information under logarithmic loss. We study the integration of information-theoretic privacy principles with representation learning, with a particular focus on face recognition systems. We also highlight the compatibility of the proposed framework with modern face recognition networks such as AdaFace and ArcFace. In addition, we introduce the Generative Privacy Funnel (GenPF) model, which extends the traditional discriminative PF formulation, referred to here as the Discriminative Privacy Funnel (DisPF). The proposed GenPF model extends the privacy-funnel framework to generative formulations under information-theoretic and estimation-theoretic criteria. Complementing these developments, we present the deep variational PF (DVPF) model, which yields a tractable variational bound for measuring information leakage and enables optimization in deep representation-learning settings. The DVPF framework, associated with both the DisPF and GenPF models, also clarifies connections with generative models such as variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models. Finally, we validate the framework on modern face recognition systems and show that it provides a controllable privacy–utility trade-off while substantially reducing leakage about sensitive attributes. To support reproducibility, we also release a PyTorch implementation of the proposed framework.

The increasing reliance on large-scale datasets in machine learning poses significant privacy and ethical challenges, particularly in sensitive domains such as face recognition. Synthetic data generation offers a promising alternative; however, most existing methods depend heavily on external datasets or pre-trained models, increasing complexity and resource demands. In this paper, we introduce **AugGen**, a self-contained synthetic augmentation technique. AugGen strategically samples from a class-conditional generative model trained exclusively on the target FR dataset, eliminating the need for external resources. Evaluated across 8 FR benchmarks, including IJB-C and IJB-B, our method achieves **1–12% performance improvements**, outperforming models trained solely on real data and surpassing state-of-the-art synthetic data generation approaches, while using less real data. Notably, these gains often exceed those from architectural modifications, underscoring the value of synthetic augmentation in data-limited scenarios. Our findings demonstrate that carefully integrated synthetic data can both mitigate privacy constraints and substantially enhance discriminative performance in face recognition. Code and datasets will be made publicly available upon publication.

In this study, we harness the information-theoretic Privacy Funnel (PF) model to develop a method for privacy-preserving representation learning using an end-to-end training framework. We rigorously address the trade-off between obfuscation and utility. Both are quantified through the logarithmic loss, a measure also recognized as self-information loss. This exploration deepens the interplay between information-theoretic privacy and representation learning, offering substantive insights into data protection mechanisms for both discriminative and generative models. Importantly, we apply our model to state-of-the-art face recognition systems. The model demonstrates adaptability across diverse inputs, from raw facial images to both derived or refined embeddings, and is competent in tasks such as classification, reconstruction, and generation. For the source code visit: https://gitlab.idiap.ch/biometric/icassp2024.dvpf.

Recently, it has been exposed that some modern facial recognition systems could discriminate specific demographic groups and may lead to unfair attention with respect to various facial attributes such as gender and origin. The main reason are the biases inside datasets, unbalanced demographics, used to train theses models. Unfortunately, collecting a large-scale balanced dataset with respect to various demographics is impracticable. In this paper, we investigate as an alternative the generation of a balanced and possibly bias-free synthetic dataset that could be used to train, to regularize or to evaluate deep learning-based facial recognition models. We propose to use a simple method for modeling and sampling a disentangled projection of a StyleGAN latent space to generate any combination of demographic groups (e.g. hispanic–female). Our experiments show that we can synthesis any combination of demographic groups effectively and the identities are different from the original training dataset. We also released the source code1.1https://gitlab.idiap.ch/biometric/sg_latent_modeling