CNN-generated images are surprisingly easy to spot... for now

View on arXiv ← Back to list

Authors: Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, Alexei A. Efros

Published: 2019-12-23 18:58:58+00:00

AI Summary

This paper investigates the creation of a universal detector for differentiating real images from those generated by Convolutional Neural Networks (CNNs). The authors demonstrate that a standard image classifier, trained on images from a single CNN generator (ProGAN), surprisingly generalizes well to unseen architectures and datasets, suggesting common systematic flaws in current CNN-generated images.

Abstract

In this work we ask whether it is possible to create a universal detector for telling apart real images from these generated by a CNN, regardless of architecture or dataset used. To test this, we collect a dataset consisting of fake images generated by 11 different CNN-based image generator models, chosen to span the space of commonly used architectures today (ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, cascaded refinement networks, implicit maximum likelihood estimation, second-order attention super-resolution, seeing-in-the-dark). We demonstrate that, with careful pre- and post-processing and data augmentation, a standard image classifier trained on only one specific CNN generator (ProGAN) is able to generalize surprisingly well to unseen architectures, datasets, and training methods (including the just released StyleGAN2). Our findings suggest the intriguing possibility that today's CNN-generated images share some common systematic flaws, preventing them from achieving realistic image synthesis. Code and pre-trained networks are available at https://peterwang512.github.io/CNNDetection/ .

Key findings

The classifier trained on ProGAN images achieved surprisingly high accuracy in detecting fake images from other unseen CNN generators, indicating shared artifacts across different models. Data augmentation significantly improved the classifier's generalization performance. The model also generalized to the recently released StyleGAN2, suggesting the potential for detecting future CNN-generated images.

Approach

The researchers trained a ResNet-50 image classifier on a large dataset of real and fake images generated by a ProGAN model. They employed data augmentation techniques simulating common image post-processing operations. The trained classifier was then evaluated on a diverse set of CNN-generated images from 11 different models to assess its generalization ability.

Datasets

ForenSynths dataset (containing images from 11 different CNN-based image generator models, including ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, and others) and LSUN, ImageNet, CelebA, COCO, GTA, and FaceForensics++ datasets.

Model(s)

ResNet-50 (pre-trained on ImageNet)

Author countries

USA

← Previous