Did You Use My GAN to Generate Fake? Post-hoc Attribution of GAN Generated Images via Latent Recovery
Syou Hirofumi, Kazuto Fukuchi, Youhei Akimoto, Jun Sakuma · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
This study proposes a method that enables attribution of GAN-generated images to the GAN model that generated the images. Existing attribution methods (e.g., model watermark) require preprossessing on the model before model publication to attain high attribution performance. This study proposes a post-hoc attribution method that does not require preprocessing before model publication. Our attribution method is designed based on the fact that latent recovery can attain better image recovery if images to be attributed are generated by the source model. Our experimental evaluation shows that our post-hoc attribution method attains almost the same attribution performance as existing methods that require preprocessing if more than five images are available for attribution.