Graduation Semester and Year
Spring 2026
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Computer Science
Department
Computer Science and Engineering
First Advisor
Shirin Nilizadeh
Abstract
Face obfuscation systems aim to protect individuals in shared photos and videos against unauthorized face recognition while preserving useful visual information for legitimate viewers and downstream applications. In practice, however, obfuscation methods face three persistent challenges. First, privacy is often evaluated under limited threat models and may not hold against stronger or mismatched attackers. Second, utility degradation can be subtle, such as changes in expression, pose, or fine facial details, yet still important for human perception and downstream analysis. Third, privacy protection and utility preservation may differ across demographic subgroups, creating unequal privacy guarantees.
This dissertation advances the robustness and fairness of face obfuscation systems through three research goals. The first goal studies StyleGAN-based face de-identification as a utility-preserving obfuscation method. We evaluate whether StyleGAN style mixing can change identity while preserving useful facial attributes, and compare its privacy and utility against existing generative obfuscation methods such as CIAGAN and DeepPrivacy. Through machine-learning-based evaluation, human studies, and statistical analysis, we show that StyleGAN can provide a strong privacy–utility trade-off and can preserve key attributes such as gender and expression better than several existing alternatives.
The second goal investigates whether face obfuscation methods provide equitable privacy protection across demographic groups. To support this analysis, we develop FairDeFace, a unified framework for auditing visual privacy protection systems. FairDeFace integrates benchmark datasets, face detection and recognition models, obfuscation methods, attacker settings, utility metrics, image-quality metrics, and fairness metrics. Across hundreds of experiments, FairDeFace reveals that privacy protection is not uniform across demographic groups and that demographic bias can arise from multiple components of the evaluation pipeline, including the obfuscation method, the detector, and the recognizer.
The third goal proposes a demographic-preserving face obfuscation method based on constrained StyleGAN2 latent editing. Motivated by the demographic disparities identified by FairDeFace, this method treats selected protected attributes, such as race and gender, as explicit constraints rather than soft utility goals. Given a real face image, we first invert it into StyleGAN2’s single-W latent space. We then represent protected attributes with learned linear latent boundaries and restrict identity-changing edits to directions that remain parallel to those boundaries. This allows the method to induce controlled identity drift while preserving selected demographic attributes to first order. We evaluate this approach using demographic preservation rates, black-box and open-source face verification attacks, recognizer-only TPR/TNR baselines, OSR–TNR rank agreement, and subgroup fairness metrics. The results show that demographic preservation is feasible but not automatic: unconstrained StyleGAN baselines can achieve strong privacy while inheriting demographic traits from auxiliary identities, whereas constrained editing improves demographic stability and makes subgroup-level privacy analysis more interpretable.
Together, these contributions provide both evaluation tools and algorithmic methods for improving face obfuscation systems. The dissertation shows that privacy should not be assessed by average obfuscation success alone; instead, robust and fair obfuscation requires joint evaluation of privacy, utility, demographic preservation, recognizer baselines, and subgroup-level disparities.
Keywords
Face obfuscation, Visual privacy protection, Face de-identification, DeObfuscation fairness, Demographic bias, StyleGAN, Latent space editing, Face recognition, Attribute preservation, Obfuscation robustness
Disciplines
Artificial Intelligence and Robotics | Computer Sciences | Information Security | Other Computer Sciences
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Moosavi Khorzooghi, Seyyed Mohammad Sadegh, "Evaluating and Advancing the Robustness and Fairness of Face Obfuscation Systems" (2026). Computer Science and Engineering Dissertations. 12.
https://mavmatrix.uta.edu/cse_dissertations2/12
Included in
Artificial Intelligence and Robotics Commons, Information Security Commons, Other Computer Sciences Commons
Comments
This dissertation was completed at The University of Texas at Arlington under the supervision of Dr. Shirin Nilizadeh.