Abstract: To address the issue of image authenticity verification in deepfake detection and model copyright protection, this paper proposed DeWM (Decoder-driven WaterMarking for Diffusion Model), a high-quality and robust watermarking method for diffusion models. First, a decoder-driven watermark embedding network was proposed, enabling direct sharing of encoder and decoder features to produce watermarks with high robustness and imperceptibility. Second, a fine-tuning strategy was designed to fine-tune the pre-trained diffusion model’s decoder, embedding a specific watermark into all generated images without altering the model architecture or diffusion process. Experimental results showed that compared with the Stable Signature method on the COCO dataset, even when the watermark bit-length is increased to 64 bits, the Peak Signal-to-Noise Ratio(PSNR) and Structure Similarity Index Measure(SSIM) of the watermarked images generated improved by 14.87% and 9.41%, respectively. Moreover, the average bit extraction accuracy under common image transformation attacks—such as cropping, JPEG compression, brightness adjustment—and image reconstruction is enhanced by more than 3%, which demonstrates significantly improved robustness.

Key words: deepfake detection, active detection, image watermark, diffusion model, Artificial Intelligence Generative Content &#40

摘要： 为解决模型版权保护和深度伪造检测中图像真实性验证的问题，提出了高质量和高鲁棒的扩散模型水印方法DeWM(Decoder-driven WaterMarking for Diffusion Model)。首先，提出了一种由解码器驱动的水印嵌入网络，实现了编码器和解码器特征的直接共享，从而生成有高鲁棒性和不可见性的水印；其次，设计了一种微调策略，对预训练扩散模型的解码器进行微调，使生成的所有图像隐含特定水印，在不改变模型架构和扩散过程的前提下，实现了简单且有效的水印嵌入。实验结果表明，在COCO数据集上与潜在扩散模型水印方法Stable Signature相比，即使在水印位数提高至64位的情况下，所提方法生成的水印图像的峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)与结构相似性指数(Structure Similarity Index Measure,SSIM)分别增加了14.87%和9.41%；所提方法针对裁剪、JEPG压缩、亮度调整等常见图像变换攻击以及图像重建攻击的水印提取的位精度平均提升了3%以上，鲁棒性显著提高。

关键词: 深伪检测, 主动检测, 图像水印, 扩散模型, 人工智能生成内容