Identity Preservability and Detectability of IDInvert GAN Model

Abstract

GAN inversion is a type of generative adversarial networks (GAN) models that can regenerate realistic images from real face photos and further perform image manipulation. While GAN inversion models can be useful for many purposes, it can be abused to generate harmfully fake contents also. This paper evaluates the performance of the recent In-domain GAN inversion model (IDInvert) regarding identity preservability and detectability of its generated face images. The experiments are conducted to answer the two main questions; how well IDInvert can imitate real face photos and how well existing image classification techniques can distinguish its generated images from the real ones. The results show that generated images do not preserve personal identity and thus significantly loss similarity to their reference photos. In addition, common machine learning classifiers can modestly distinguish these generated images from real photos with 0.87 accuracy. This indicates that the recent IDInvert model’s ability to imitate real faces is not yet perfect and hazardous, and its generated images are still simply detected.