[논문 리뷰] CNN-generated images are surprisingly easy to spot... for now

Sheng-Yu Wang, Oliver Wang외 3명|arXiv (Cornell University)|2019. 12. 23.

Generative Adversarial Networks and Image Synthesis참고 문헌 44인용 수 42

한 줄 요약

A detector trained to identify CNN-generated images from one generator generalizes surprisingly well to unseen CNN generators, given proper data augmentation and diverse training data.

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/ .

연구 동기 및 목표

• Motivate a universal forensics detector for CNN-generated imagery across architectures and datasets.
• Investigate whether artifacts or fingerprints are shared across CNN generators that enable cross-model detection.
• Evaluate how data augmentation and training data diversity affect cross-generator generalization.
• Assess robustness of detectors to post-processing like JPEG compression, blur, and resizing.

제안 방법

• Construct a dataset (ForenSynths) with fake images from 11 CNN-based generators spanning unconditional, conditional, perceptual-loss, and post-processing methods.
• Train a ResNet-50 binary classifier to distinguish real vs. ProGAN-generated images, using real images as negatives.
• Apply extensive image augmentations that simulate post-processing during training to promote generalization.
• Evaluate cross-generator generalization by testing the ProGAN-trained detector on all other generators.
• Test robustness to post-processing (blur, JPEG) at test time to assess real-world applicability.
• Compare to prior methods and analyze frequency-domain artifacts to understand generalization cues.

실험 결과

연구 질문

• RQ1 Can a detector trained on one CNN generator generalize to unseen generators across architectures and datasets?
• RQ2 What roles do training data diversity and augmentation play in cross-model generalization?
• RQ3 How robust are detectors to common post-processing operations encountered in real-world pipelines?
• RQ4 Do CNN-generated images share detectable, transferable artifacts across different synthesis methods?

주요 결과

• A classifier trained on ProGAN images generalizes surprisingly well to many unseen GANs and image-to-image translation models.
• Data augmentation simulating post-processing significantly improves cross-generator generalization and robustness to JPEG, blur, and resizing.
• Increased training diversity (more classes/datasets) improves performance up to a point, with diminishing returns beyond around 16–20 classes.
• StyleGAN2 out-of-the-box remains detectable by the ProGAN-trained detector, indicating transferability to newer generators with similar building blocks.
• On some models (e.g., SAN, DeepFake) augmentation can hurt performance, suggesting model-specific cue differences.