**CS294-26 Final Research Project: An Empirical Study of GAN Watermarking** By Neerja Thakkar Overview ======== AI-synthesized content, known as deep fakes, are becoming increasingly accessible and indistinguishable from real content, and pose a significant threat to society. Existing detection methods face ever-improving deep fake technology at an increasingly large scale. We explore an alternate approach to the problem of deep fakes: GAN watermarking. Initial work has shown that a GAN trained on watermarked data will synthesize watermarked content. We empirically study the impact of different training schemes and image perturbations on watermark robustness and propose future avenues of study. !(does_not_exist/1.jpeg width=300)!(does_not_exist/7.jpeg width=300)!(does_not_exist/15.jpeg width=300) These people do not exist. They were generated from StyleGAN2, and appear perceptually indistinguishable from images of real people. In order to obtain these images, no code was needed – they were simply taken fromthispersondoesnotexist.com. Method =============================================================================== We implement the method of Yu et al., namely, using a neural steganography algorithm to create a watermarked dataset, and then training a GAN on the watermarked data. We investigate different training schemes and perturb generated images to study watermark robustness. ![An overview of the method of Yu et al.](Yu_method.png width=700) We find that when training DCGAN, the watermark does transfer to generated images with bitwise accuracy of around 83%, much higher than the 50% accuracy expected for random guessing. For full details of experiments and results, see the paper. Paper and Presentation ========================= The project writeup is [here](CS294_26_Final_Project_Write_Up.pdf) and the presentation video can be found [here](https://www.youtube.com/watch?v=D2-BUrOxf04&ab_channel=NeerjaThakkar).