Privacy preservation through makeup transfer for facial feature obfuscation

Privacy preservation through makeup transfer for facial feature obfuscation

연구 분야: Analysis

학회: The Journal of Supercomputing

Automated facial recognition can infer sensitive attributes from facial images without consent, posing substantial privacy risks. Existing adversarial perturbation methods often degrade visual fidelity and compromise identity utility. We propose makeup-transfer obfuscation GAN (MTO-GAN), which uses makeup-inspired perturbations to obfuscate soft biometric traits while preserving realism and recognition performance. Methodologically, (i) an entropy-increase perspective motivates the use of adversarial noise; (ii) a density-ratio-to-probabilistic-classification reformulation with a Siamese objective estimates and mitigates domain shift while alleviating conflicts with cycle consistency; and (iii) a lightweight domain regularization module based on RRDB denoises and harmonizes features to stabilize the cycle. To address the challenges of large-scale facial image privacy computation and extreme computational load in deep learning, we employ GPU-accelerated parallel inference to meet throughput and latency requirements. Experiments across four public face datasets and five face recognizers show that MTO-GAN drives age and race predictability toward random-guessing levels while largely preserving identity verification, and it improves perceptual quality over prior perturbation approaches. Overall, MTO-GAN achieves a favorable balance among privacy protection, visual fidelity, and identity utility.