Preprint 1 Mention
CNN-generated images are surprisingly easy to spot... for now
Sheng-Yu Wang2019
Oliver WangRichard Zhang
Top 10% · 90th Percentile
44 citations · Computer Vision and Pattern Recognition
Open Access

TLDR

Computer-made images created by today's AI tools are still easy for computers to spot, even if they look real to people. This might change as the technology improves.

Summary

1 Study Aim

The main goal of this study is to find out if it is possible to build a single, general-purpose detector that can reliably tell apart real photos from images created by any convolutional neural network (CNN)-based generator, no matter which model or dataset was used. The authors want to know if current CNN-generated images share common flaws that make them detectable, and whether a detector trained on one type of fake image can spot fakes from other, unseen models and tasks. A single detector might be able to spot fake images from many different AI models.

2 Study Design

The researchers collected a large dataset of fake images made by 11 different CNN-based image generator models, covering a wide range of architectures and tasks (such as ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, and others). They trained a standard image classifier (ResNet-50, a type of deep neural network) using only images generated by one model (ProGAN) and real images as examples. The team used various data augmentation techniques (like blurring and JPEG compression) during training to help the classifier learn to ignore simple post-processing tricks. They then tested how well this classifier could detect fake images from the other, unseen generator models, including newly released ones like StyleGAN2. The study also compared their approach to other existing fake image detectors and analyzed how training data diversity and augmentation affected performance. The researchers trained a computer program to spot fake images using one AI model, then checked if it could spot fakes from other models.

3 Findings

The study reveals that a classifier trained on fake images from just one CNN generator (ProGAN) can successfully detect fake images from many other, previously unseen CNN-based models, even when those models use different datasets or architectures. The authors demonstrate that using data augmentation (such as blurring and JPEG compression) during training greatly improves the detector's ability to generalize and remain robust to common image post-processing. Increasing the diversity of training images also helps, but only up to a certain point. The detector even works well on new models like StyleGAN2 and can handle some real-world image changes, though heavy post-processing still poses challenges. However, the authors caution that as image generation technology advances, these detectable flaws may disappear, making future fake images harder to spot. They also note that their method does not work on traditional, non-AI image manipulations (like Photoshop edits), and that a combination of technical, social, and legal strategies will be needed to address visual misinformation. Right now, computers can spot AI-made images because they share hidden flaws, but this may not last as the technology gets better.

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/ .

Referenced In