GAN-Generated Face Features Dataset based on Cross-Band Co-occurrences

Published: 12 September 2020| Version 1 | DOI: 10.17632/xt866zxsbc.1
Contributor:
Ehsan Nowroozi

Description

Last-generation GAN models allow generating synthetic images that are visually indistinguishable from natural ones, raising the need to develop tools to distinguish fake and natural images thus contributing to preserving the trustworthiness of digital images. One method recently used for distinguishing GAN-generated from natural images by exploiting inconsistencies among spectral bands, with a specific focus on the generation of synthetic face images. The dataset consists of cross-band co-occurrence matrices, in addition to spatial co-occurrence matrices and without a cross band. If you use a dataset and code please use one of the following corresponding citation (https://arxiv.org/abs/2007.12909).

Files

Steps to reproduce

We consider the detection of StyleGAN2 images. StyleGAN2 has been recently proposed as an improvement of the original StyleGAN1 architecture, and achieves impressive results, being capably to generate synthetic images of extremely high quality. We considered a total of 20000 real (from FFHQ) and 20000 GAN-generated images,. Specifically, CoNet consider the 3 co-occurrence matrices computed on the R, G, and B channel of the images and input such tensor of 3 matrices (refer to CoNet *.npy). In order to exploit the relationships among color bands, we compute cross-band co-occurrences, in addition to the spatial co-occurrences computed on the single-color bands separately. We refer to the network trained in this way as Cross-CoNet (refer to CrossNet *.npy). We also expect cross-band features to be more robust to common post-processing operations, which usually focus more on spatial pixel relationships rather than on cross-band characteristics. Policy Before we are able to offer you access to the database, please agree to the following terms of use. 1) Researcher shall use the Database only for non-commercial research and educational purposes. 2) The dataset is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. 3) If you use a dataset and code please use one of the following corresponding citations. If you also use please use the citation (https://arxiv.org/abs/2007.12909) - bibtex: @misc{barni2020cnn, title={CNN Detection of GAN-Generated Face Images based on Cross-Band Co-occurrences Analysis}, author={Mauro Barni and Kassem Kallas and Ehsan Nowroozi and Benedetta Tondi}, year={2020}, eprint={2007.12909}, archivePrefix={arXiv}, primaryClass={cs.CV} }