Deceptively realistic deep synthetic images are widely disseminated across the internet, raising significant concerns about misinformation and content authenticity. From the perspective of responsible model publishers and credible regulators, this study introduces a novel robust generative image watermarking method that significantly advances detection and tracing of deep synthetic images. Unlike existing approaches that focus primarily on watermark embedding, our method uniquely combines adversarial training with specially designed cropping and resizing manipulations to substantially enhance watermark robustness. Furthermore, we introduce a residual signal prediction mechanism through the encoder, coupled with a discriminator-based quality enhancement approach, which significantly improves the quality of generated images while maintaining watermark integrity. With minimal impact on the original performance of generative models, this approach embeds robust generative image watermarks that demonstrate superior resilience against common disturbances such as compression, cropping, and resizing encountered on social media platforms, achieving over 75% bit-wise accuracy even under extreme distortions, thereby facilitating effective tracing and detection of synthetic images at the model level.
Sun et al. (Mon,) studied this question.