Open Zero-Watermarking Approach to Prevent the Unauthorized Use of Images in Deep Learning

The recent emergence of generative AI has led to a surge of new technologies that have had massive repercussions on society. In particular, AI image generators have created some major issues, such as deepfake images and image theft. Specifically with image theft, there lies the issue of getting permission to use images in a deep learning model. Therefore, we propose a threefold contribution. First, we offer a tool to provide image owners with the capabilities to mark their images with a specific watermark that states that they do not wish their images to be used for training imaging models. To this end, we propose two different simple watermark algorithms. The second contribution is a deep learning model to identify the existence of a watermark in an image. The accuracy result was 97.3% for the classification of images marked with the rolling average watermark algorithm. However, the ring circle algorithm was overfitted. Our third contribution is an open solution that can be replicated or extended by practitioners.