Do GANs Leave Artificial Fingerprints?

Abstract

In the last few years, generative adversarial networks (GAN) have shown tremendous potential for a number of applications in computer vision and related fields. With the current pace of progress, it is a sure bet they will soon be able to generate high-quality images and videos, virtually indistinguishable from real ones. Unfortunately, realistic GAN-generated images pose serious threats to security, to begin with a possible flood of fake multimedia, and multimedia forensic countermeasures are in urgent need. In this work, we show that each GAN leaves its specific fingerprint in the images it generates, just like real-world cameras mark acquired images with traces of their photo-response non-uniformity pattern. Source identification experiments with several popular GANs show such fingerprints to represent a precious asset for forensic analyses.

Attributing Fake Images to GANs: Learning and Analyzing GAN Fingerprints

“…Comparison to baselines. In terms of attribution accuracy, our method consistently outperforms three baseline methods (including a very recent one [45]) on two datasets under a variety of experimental conditions. In terms of feature representation, our fingerprints show superior distinguishability across image sources compared to inception features [52].…”

“…during each image acquisition procedure [24]. Recently, Marra et al [45] visualize GAN fingerprints based on PRNU, and show their application to GAN source identification. We replace their hand-crafted fingerprint formulation with a learning-based one, decoupling model fingerprint from image fingerprint, and show superior performances in a variety of experimental conditions.…”

“…Alternatively to our attribution network in Section 3.1 where fingerprints are implicitly represented in the feature domain, we describe a model similar in spirit to Marra et al [45] to explicitly represent them in the image domain. But in contrast to their hand-crafted PRNU-based representation, we modify our attribution network architecture and indicates pixel-wise multiplication of two normalized images.…”

Detecting GAN generated Fake Images using Co-occurrence Matrices

“…Several methods have been proposed in the area of image forensics over the past years [25,26,27,28,29]. Recent approaches have focused on applying deep learning based methods to detect tampered images [30,31,32,33,34,9,35] The detection of GAN images is a new area in image forensics and there are very few papers in this area [36,37,38,39,40,41,42,43,44,45]. Related fields also include detection of computer generated (CG) images [46,47,48,49].…”

Establishing an evaluation metric to quantify climate change image realism *

“…However, these modifications do not evaluate the visual fidelity between different modes, only within them in the case of Fréchet Joint Distance [18], which limits their application in multimodal settings such as ours. Methods for detecting artifacts [20] and artificial fingerprints [21] in Table 1: Pearson's r Correlation Coefficient. Results of Pearson's r and bootstrap 95% confidence intervals between human HYPE-Style scores and all automated methods across different layers of an ImageNet-pretrained Inception-V3 model, including the three pooling layers (pool 1, pool 2, pool 3) and the layer preceding the auxiliary classifier (pre-aux).…”