Face-Image Source Generator Identification

Abstract

Recent advances in deep networks and specifically, Generative Adversarial Networks, have introduced new ways of manipulating and synthesizing “fake” images. Concerns have been raised as to the sinister use of these images, and accordingly challenges have been raised to detect “fake” from “real” images. In this study we address a slightly different problem in image forensics. Rather than discriminating real from fake, we attempt to perform “Source Generator Identification”, i.e. determine the source generator of the synthesized image. In this study we focus on face images. We exploit the specific characteristics associated with each fake face image generator and introduce a face generator representation space (the profile space) which allows a study of the distribution of face generators, their distinctions as well as allows estimating probability of images arising from the same generator.