CNN-Generated Images Are Surprisingly Easy to Spot… for Now

By Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, Alexei A. Efros. 2019

1,553 citations recorded by the source corpus.

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.

Subfields: Computer Vision and Pattern Recognition, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Safety, Robustness & Interpretability

Citations in the corpus

Denoising Diffusion Probabilistic Models (2020, background): “Fortunately, CNN-generated images currently have subtle flaws that allow detection [62], but improvements in generative models may make this more difficult.”
Analyzing and Improving the Image Quality of StyleGAN (2019, neutral): “At present, classifier-based methods can quite reliably detect generated images, regardless of their exact origin [29, 45, 40, 51, 41].”
On Buggy Resizing Libraries and Surprising Subtleties in FID Calculation (2021, background): “Despite these advances, generative methods continue to be detectable [10, 74], and discriminative networks continue to be sensitive to small perturbations, such as shifts [3, 22] and JPEG compression [29].”
Generative Negative Text Replay for Continual Vision-Language Pretraining (2022, methodology): “In addition, the feature-based distillation methods [15] are usually designed for uni-modal CNNs and rarely take into account the characteristics of transformer-based multi-modal pre-training, and the model inversion [51] aims to generate images as positive examples for certain classes using a froze”
Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs (2020, background): “First, the transferability of the feature extractor of the discriminator could be applied for the universal detector of generated images [41].”
Artificial Fingerprinting for Generative Models: Rooting Deepfake Attribution in Training Data (2020, methodology): “(3) Our solution outperforms [50] with clear margins.”
Focal Frequency Loss for Image Reconstruction and Synthesis (2020, background): “Recent studies [59, 72, 22] in media forensics have shown some notable periodic patterns in the frequency spectra of manipulated images, which may be consistent with artifacts in the spatial domain.”
Spatial-Phase Shallow Learning: Rethinking Face Forgery Detection in Frequency Domain (2021, background): “Under this setting, these approaches achieve excellent performance on some public datasets [48, 20, 37, 26, 45].”
DeepFake Detection by Analyzing Convolutional Traces (2020, background): “[42] that is actually able to reach very high results by simply using a discriminator trained on one family of GANs and using it to infer if images are real or generated from other types of GANs.”
Lips Don't Lie: A Generalisable and Robust Approach to Face Forgery Detection (2020, methodology): “(2) CNN-aug [59]: we employ a ResNet-50 with JPEG compression and Gaussian blurring augmentations, both with probability of 0.1.”
Media Forensics and DeepFakes: An Overview (2020, methodology): “Instead, in [197] a careful pre- and postprocessing and data augmentation are applied to improve transferability.”
Learning Self-Consistency for Deepfake Detection (2020, background): “To detect the whole image synthesis, recent research [58, 36, 13, 52] observes that GANgenerated images contain specific cues that can be easily detected, and the trained models have exhibited good generalization ability across different synthesis methods.”
What makes fake images detectable? Understanding properties that generalize (2020, methodology): “[33] generalizes across a wide variety of datasets simply by adding various levels of augmentation.”
A Continual Deepfake Detection Benchmark: Dataset, Methods, and Essentials (2022, methodology): “Similarly, [76] uses ResNet-50 [26] pretrained with ImageNet, and further train it in a binary classification setting for deepfake detection.”
GenImage: A Million-Scale Benchmark for Detecting AI-Generated Image (2023, methodology): “4 GenImage Benchmark”
Language-Driven Image Style Transfer (2021, methodology): “To mitigate this issue, we can apply techniques from image forensics [88, 89, 90] to detect the authenticity of an image.”