**DiffuseTrace: A Transparent and Flexible Watermarking Scheme for Latent Diffusion Model** **Authors:** Liangqi Lei **Abstract:** Latent Diffusion Models (LDMs) have revolutionized various applications, but their potential for illegal utilization raises ethical concerns. Adding watermarks to LDM outputs is crucial for copyright tracking and mitigating risks associated with AI-generated content. However, existing watermarking techniques are susceptible to evasion and require retraining when the watermark message needs to be changed. To address these issues, we propose DiffuseTrace, a novel watermarking technique that embeds invisible watermarks semantically into all generated images for future detection. DiffuseTrace establishes a unified representation of initial latent variables and watermark information through an encoder-decoder model. The watermark is embedded into the initial latent variables and integrated into the sampling process. Extraction is achieved by reversing the diffusion process and using the decoder. DiffuseTrace does not rely on fine-tuning of the diffusion model components and can be seamlessly integrated into various LDMs. Experiments validate the effectiveness and flexibility of DiffuseTrace, demonstrating its superior performance in combating attacks based on variational autoencoders and diffusion models. **Keywords:** Latent Diffusion Model, Model Watermarking, Copyright Protection, Security **Introduction:** LDMs have significantly advanced the synthesis of photorealistic content, impacting text-to-image, image editing, in-painting, super-resolution, content creation, and video synthesis. However, the potential for malicious use, such as generating insulting or offensive images, poses significant risks. Watermarking is a vital technique to track copyright and mitigate these risks. Existing methods often suffer from trade-offs between watermark robustness and image quality, and they are susceptible to post-processing attacks. DiffuseTrace embeds watermarks into the latent variables of the model, ensuring semantic consistency and image quality without post-processing. It can be seamlessly integrated into various LDMs and is robust against state-of-the-art attacks. **Background:** Diffusion models transition data from a true distribution to stochastic noise and reverse this process through iterative denoising. The forward process involves progressively diffusing the data distribution towards the noise distribution. The backward process, or diffusion inversion, reconstructs the noise representation from an image. DiffuseTrace leverages this property to embed watermarks. **Problem Formulation:** DiffuseTrace aims to embed robust multi-bit watermarks into the initial latent variables of LDMs. It ensures semantic consistency and image quality while being flexible and robust against various attacks. The scheme includes a watermark encoder and decoder, with the encoder constructing a unified representation of watermark information and latent variables. The decoder extracts the watermark from the latent variables. **Proposed Watermarking Scheme:** DiffuseTrace pre-trains an encoder-decoder structure to embed and extract watermarks. The encoder encodes the initial latent variables, and the decoder extracts the watermark. The scheme is robust against image**DiffuseTrace: A Transparent and Flexible Watermarking Scheme for Latent Diffusion Model** **Authors:** Liangqi Lei **Abstract:** Latent Diffusion Models (LDMs) have revolutionized various applications, but their potential for illegal utilization raises ethical concerns. Adding watermarks to LDM outputs is crucial for copyright tracking and mitigating risks associated with AI-generated content. However, existing watermarking techniques are susceptible to evasion and require retraining when the watermark message needs to be changed. To address these issues, we propose DiffuseTrace, a novel watermarking technique that embeds invisible watermarks semantically into all generated images for future detection. DiffuseTrace establishes a unified representation of initial latent variables and watermark information through an encoder-decoder model. The watermark is embedded into the initial latent variables and integrated into the sampling process. Extraction is achieved by reversing the diffusion process and using the decoder. DiffuseTrace does not rely on fine-tuning of the diffusion model components and can be seamlessly integrated into various LDMs. Experiments validate the effectiveness and flexibility of DiffuseTrace, demonstrating its superior performance in combating attacks based on variational autoencoders and diffusion models. **Keywords:** Latent Diffusion Model, Model Watermarking, Copyright Protection, Security **Introduction:** LDMs have significantly advanced the synthesis of photorealistic content, impacting text-to-image, image editing, in-painting, super-resolution, content creation, and video synthesis. However, the potential for malicious use, such as generating insulting or offensive images, poses significant risks. Watermarking is a vital technique to track copyright and mitigate these risks. Existing methods often suffer from trade-offs between watermark robustness and image quality, and they are susceptible to post-processing attacks. DiffuseTrace embeds watermarks into the latent variables of the model, ensuring semantic consistency and image quality without post-processing. It can be seamlessly integrated into various LDMs and is robust against state-of-the-art attacks. **Background:** Diffusion models transition data from a true distribution to stochastic noise and reverse this process through iterative denoising. The forward process involves progressively diffusing the data distribution towards the noise distribution. The backward process, or diffusion inversion, reconstructs the noise representation from an image. DiffuseTrace leverages this property to embed watermarks. **Problem Formulation:** DiffuseTrace aims to embed robust multi-bit watermarks into the initial latent variables of LDMs. It ensures semantic consistency and image quality while being flexible and robust against various attacks. The scheme includes a watermark encoder and decoder, with the encoder constructing a unified representation of watermark information and latent variables. The decoder extracts the watermark from the latent variables. **Proposed Watermarking Scheme:** DiffuseTrace pre-trains an encoder-decoder structure to embed and extract watermarks. The encoder encodes the initial latent variables, and the decoder extracts the watermark. The scheme is robust against image