基于卷积神经网络的扩散模型潜在水印嵌入算法OA

The existing digital watermarking algorithms have problems with weak robustness,insufficient invisibility,or poor generation quality when applied to diffusion models.To address these issues,this paper proposes a latent watermark embedding algorithm for diffusion models based on convolutional neural networks.Firstly,a watermark encoder is design to convert one-dimensional watermark information into two-dimensional features,which can adapt to the diffusion process while ensure stable embedding of the watermark.Subsequently,a pre-trained watermark encoder is obtained by conducting independent training on the frozen image encoder-decoder.Finally,the complete embedding of watermark information is achieved by introducing the dynamic process of forward noise addition and multi-step reverse denoising recovery,as well as constraining the predicted noise.Experimental results demonstrate that the proposed algorithm maintains high extraction accuracy under various common noise attacks,while noise attacks exert negligible impact on the visual quality of generated images.