FaceBERT: Face De-Identification Using VQGAN and BERT

Abstract

This paper presents FaceBERT, a face de-identification technique based on image generation to solve the infringement of portrait rights. Using the VQGAN model, it learns a quantized codebook that expresses an image in block units, encodes the image using the codebook, and then trains the BERT model. As a result of an experiment using FFHQ and CelebA-HQ, it was confirmed that the face images generated using FaceBERT were natural de-identified face images and different from the originals.