Gender-Adversarial Networks for Face Privacy Preserving

Abstract

Privacy concerns over face recognition systems have attracted extensive attention in various fields. For gender privacy-preserving work, there are two key challenges: 1) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">privacy</i> , i.e., confusing gender classifiers and 2) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">utility</i> , i.e., maintaining its face verification performance. To address both issues, this article develops a novel gender-adversarial network, referred to as Gender-AN, to impart gender privacy to face images. Gender-AN employs an attribute-independent encoder–decoder GAN-based network to perturb the input face image, training with the assistance of the proper facial attributes. The perturbed image is then able to obfuscate gender classifiers while maintaining identity discriminability. To optimize the generator, a multitask-based loss function is utilized, which includes attribute manipulation loss, face matcher loss, adversarial loss, and reconstruction loss functions. This optimization facilitates our model to achieve the generalization, verification preserve, and natural appearance, simultaneously. Extensive experiments confirm the effectiveness of the proposed model in enhancing gender privacy and preserving face verification utility.