MAP-GAN: Generative Adversarial Network for Privacy Preserving of Multiple Facial Attributes

The widespread use of face recognition systems has raised concerns about privacy and security. In this paper, we proposed multi-attribute privacy-preserving generative ad-versarial networks, referred to as MAP-GAN, for protecting face privacy at the image level. In the MAP-GAN, a specially designed encoder-decoder with residual structure was used as the generator to reconstruct privacy-preserving images. We introduced the concept of attribute probability score to address the binary attribute flipping problem. Additionally, to ensure the utility of MA-GAN, the
$L_{1}$
distance and the change in information recognition performance were used to constrain the information difference between the privacy-preserving images and the original images. Extensive experiments demonstrate the effectiveness of MAP-GAN in multi-attribute privacy preserving and utility compared to other models.