A Robust Image Watermarking Approach for Application in AI Generated Images

In this work, we present a novel approach to enhance the robustness of image watermarking through the integration of Convolutional Neural Network (CNN) and Generative Adversarial Network (GAN). The utilization of CNN facilitates the encoding and decoding of watermarks within images, while GAN and subsequent image post-processing techniques ensure robustness against unlawful removal of watermarks. Given the increasing significance of watermarking for data integrity and privacy, development of a robust watermarking algorithm is of critical importance. To this end, we propose to incorporate a pre-trained CNN model to decode watermarks associated with the output of the generator of the GAN model. We argue that the transferability of watermarks from training images to the generated images ensures accurate retrieval of watermark from generated images. We further enhance the efficiency and robustness of our proposed watermark embedding process through fine-tuning and data augmentation.Experimental results on the CelebA dataset validate our approach, achieving 93% accuracy in watermark detection, a Fréchet Inception Distance (FID) score of 14.09, and approximately $80 \%$ accuracy under heavy perturbations (noise standard deviation $=10$). These results demonstrate the adaptability of our method across diverse GAN architectures, tasks, and image resolutions, as well as its resilience to post-processing techniques such as JPEG compression, noise addition, and blurring.