Flexible Semantic Watermarking for Robust Diffusion Model Detection and Tracing

The excellent capability and wide application of Diffusion Models have lead to great concerns about malicious usage.Watermarking is regarded as a promising solution to identify the model authenticity and trace the origin of the image, thus alleviating the potential misuse risks.However, existing works fall short in properties such as fidelity and watermark capacity, thus decreasing the robustness and flexibility of watermarking methods.In response to the limitations of existing methods, this paper proposes a multi-bit latent space watermarking scheme for watermark detection and identification.We design the watermark scheme based on the principle of the diffusion sampling process, which enables watermarking generated content without modifying the original workflow or adjusting the pretrained weights of the diffusion model.Experiments have demonstrated that the proposed method achieves a good balance in terms of watermark effectiveness, fidelity and robustness.Additionally, our method can be adapted for different sampling settings of diffusion model, indicating great flexibility and versatility in practical applications.