CNN-generated images are surprisingly easy to spot, even for models trained on a single CNN architecture. Researchers at UC Berkeley, Adobe Research, and the University of Michigan collected a dataset of images generated by 11 different CNN-based image generation models, including ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, and others. They trained a classifier on images generated by ProGAN and found it could generalize well to images generated by other models, including recently released ones like StyleGAN2. This suggests that today's CNN-generated images share common systematic flaws that make them detectable. The study shows that data augmentation, such as simulated post-processing operations, is critical for generalization. Classifiers trained with diverse training images outperform those trained on smaller datasets. The research also found that CNN-generated images are robust to common post-processing operations like JPEG compression, blurring, and resizing. However, some models, like SAN and DeepFake, showed reduced performance when augmented. The study also highlights that CNN-generated images contain low-level artifacts that can be detected by classifiers, even when the images have undergone post-processing. The researchers propose a new dataset, ForenSynths, consisting of images generated by 11 different models, and a new evaluation metric for detecting CNN-generated images. They compare their results with previous works and find that their models generalize better to other architectures, except for CycleGAN and StarGAN. The study also shows that CNN-generated images have distinct frequency patterns, which can be used for detection. However, some models like DeepFake do not contain obvious artifacts due to extensive post-processing. The findings suggest that while CNN-generated images can be detected, the challenge remains as generation methods evolve. The study emphasizes the importance of robustness to post-processing and the need for further research to improve detection capabilities in real-world scenarios. The results indicate that training on today's generators can generalize well to future ones, given similar underlying building blocks. The study also highlights the importance of considering both low-level and high-level features for detection.CNN-generated images are surprisingly easy to spot, even for models trained on a single CNN architecture. Researchers at UC Berkeley, Adobe Research, and the University of Michigan collected a dataset of images generated by 11 different CNN-based image generation models, including ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, and others. They trained a classifier on images generated by ProGAN and found it could generalize well to images generated by other models, including recently released ones like StyleGAN2. This suggests that today's CNN-generated images share common systematic flaws that make them detectable. The study shows that data augmentation, such as simulated post-processing operations, is critical for generalization. Classifiers trained with diverse training images outperform those trained on smaller datasets. The research also found that CNN-generated images are robust to common post-processing operations like JPEG compression, blurring, and resizing. However, some models, like SAN and DeepFake, showed reduced performance when augmented. The study also highlights that CNN-generated images contain low-level artifacts that can be detected by classifiers, even when the images have undergone post-processing. The researchers propose a new dataset, ForenSynths, consisting of images generated by 11 different models, and a new evaluation metric for detecting CNN-generated images. They compare their results with previous works and find that their models generalize better to other architectures, except for CycleGAN and StarGAN. The study also shows that CNN-generated images have distinct frequency patterns, which can be used for detection. However, some models like DeepFake do not contain obvious artifacts due to extensive post-processing. The findings suggest that while CNN-generated images can be detected, the challenge remains as generation methods evolve. The study emphasizes the importance of robustness to post-processing and the need for further research to improve detection capabilities in real-world scenarios. The results indicate that training on today's generators can generalize well to future ones, given similar underlying building blocks. The study also highlights the importance of considering both low-level and high-level features for detection.