Fake Image Attribution

The ArtiFact dataset is a large-scale image dataset that aims to include a diverse collection of real and synthetic images from multiple categories, including Human/Human Faces, Animal/Animal Faces, Places, Vehicles, Art, and many other real-life objects. The dataset comprises 8 sources that were carefully chosen to ensure diversity and includes images synthesized from 25 distinct methods, including 13 GANs, 7 Diffusion, and 5 other miscellaneous generators. The dataset contains 2,496,738 images, comprising 964,989 real images and 1,531,749 fake images. To ensure diversity across different sources, the real images of the dataset are randomly sampled from source datasets containing numerous categories, whereas synthetic images are generated within the same categories as the real images. Captions and image masks from the COCO dataset are utilized to generate images for text2image and inpainting generators, while normally distributed noise with different random seeds is used for noise2image generators. The dataset is further processed to reflect real-world scenarios by applying random cropping, downscaling, and JPEG compression, in accordance with the IEEE VIP Cup 2022 standards. The ArtiFact dataset is intended to serve as a benchmark for evaluating the performance of synthetic image detectors under real-world conditions. It includes a broad spectrum of diversity in terms of generators used and syntheticity, providing a challenging dataset for image detection tasks. * Total number of images: 2,496,738 * Number of real images: 964,989 * Number of fake images: 1,531,749 * Number of generators used for fake images: 25 (including 13 GANs, 7 Diffusion, and 5 miscellaneous generators) * Number of sources used for real images: 8 * Categories included in the dataset: Human/Human Faces, Animal/Animal Faces, Places, Vehicles, Art, and other real-life objects * Image Resolution: 200 x 200

This dataset was created to test whether it's possible to build a general-purpose detector that can tell real images apart from fake ones generated by convolutional neural networks (CNNs), no matter which model or dataset was used to create the fake images. To do this, the authors collected fake images generated by 11 different CNN-based image generation models. These models represent a wide range of current image synthesis techniques and include: ProGAN StyleGAN BigGAN CycleGAN StarGAN GauGAN DeepFakes Cascaded Refinement Networks (CRN) Implicit Maximum Likelihood Estimation (IMLE) Second-order Attention Super-Resolution (SOAT-SR) Seeing-in-the-Dark (SID) The dataset includes fake images from each of these models and a set of real images, allowing for binary classification (real vs. fake). The study found that a standard image classifier (like a convolutional neural network) trained on fake images from just one generator (ProGAN) was able to detect fake images from other, completely different generators with surprising accuracy. This suggests that many CNN-generated images, even from different architectures, share common flaws that can be learned and detected. The dataset is useful for research in detecting synthetic media, improving image forensics, and understanding the weaknesses in current generative models. Code and pre-trained models were made available by the authors (https://github.com/chuangchuangtan/GANGen-Detection).