Enhanced Demographic Privacy in Face Recognition: From Images to Templates
Zohra Rezgui is a PhD student in the Department Datamanagement & Biometrics. (Co)Promotors are prof.dr.ir. R.N.J. Veldhuis and dr. N. Strisciuglio from the Faculty of Electrical Engineering, Mathematics and Computer Science.
This dissertation explores advanced methods to mitigate privacy risks associated with demographic profiling from both facial images and their numerical representations used for face recognition, known as templates. While templates are designed solely for identity verification and identification, they often reveal soft biometric attributes, such as gender and age, due to the overlap of identity and demographic features. This unintended leakage raises significant privacy concerns, particularly when demographic profiling is conducted without the user's consent.
The research employs a dual approach. First, at the image level, adversarial attacks are applied as proof of concept to demonstrate their potential to fool demographic classifiers. This approach shows that it is possible to reduce the accuracy of demographic inference without degrading image quality or compromising face recognition performance.
Second, at the template level, where identity and demographic features are deeply intertwined, adversarial techniques are extended to suppress demographic information from the face recognition templates while preserving recognition performance.
Beyond noise-based adversarial methods, the dissertation introduces a model-based approach. By incorporating constraints during the training process, this approach generates demographic-invariant features and achieves stronger privacy protections with minimal trade-offs in verification performance.
A final contribution is a method allowing dynamic control over the privacy-utility trade-off, which enables system operators to fine-tune the balance between privacy and performance without requiring extensive retraining.
These contributions significantly enhance privacy-preserving technologies in biometric systems. They pave the way for face recognition applications that are not only effective but also ethically and legally compliant. By reducing demographic profiling risks and maintaining system performance, this work supports the development of adaptable and privacy-conscious biometric solutions.