Robust and imperceptible image watermarks in stable-diffusion image editing models

Abstract

With the rapid development of generative models, generative image editing has significantly enriched people’s lives but has also introduced ethical challenges, such as the fake news and misinformation. This dissertation proposes a robust watermarking framework designed for developers of Stable Diffusion based image editing models. The research aims to develop a watermarking method for not only embedding invisible and robust watermarks in every edited image, allowing developers for source detection and tracing, but also improving the quality of the generated images as much as possible, which means ensure the invisibility of the watermark to enhance the user experience. The method employs a pretrained robust encoder for watermarking training dataset and a decoder for bit string extraction after watermarked images generated by Stable Diffusion Model. The latent decoder of the editing model is fine-tuned, incorporating a discriminator and adversarial training to enhance watermark imperceptibility and image quality. The watermark robustness under various of attacks and visual qualities of watermarked edited images are evaluated, showing that our method can reach nearly 100% of extracted bit accuracy, maintaining superior image quality as well. Through experiments, it is demonstrated that our method outperforms previous watermark-in-generation methods on image quality and watermark invisibility, while preserving a certain level of bit extraction accuracy.