A Secure and Efficient Watermarking Scheme for Latent Diffusion Models in Model Distribution Scenarios

Researchers propose a secure and efficient watermarking scheme for latent diffusion models in model distribution scenarios. The scheme uses watermark randomness and watermark-model association as constraints to prevent watermark leakage and escape, decouples watermark injection from the security mechanism, fine-tunes only the VAE to reduce training overhead, and uses a verification strategy based on the watermark distribution to enhance robustness. Experiments show that the method consistently outperforms six baselines in effectiveness and robustness against ten image processing attacks and adversarial attacks, and enhances security in distribution scenarios.