A Dual-Protection Framework for Copyright Protection and Image Editing Using Multi-Label Conformal Prediction

TMLR Paper6722 Authors

Abstract: Recent advances in diffusion models have significantly enhanced image editing capabilities, raising serious concerns about copyright protection. Traditional watermarks often fail to withstand diffusion-based edits, making image protection challenging. To address this, we propose a method that embeds an imperceptible perturbation in images, serving as a watermark while simultaneously disrupting the output of latent diffusion models. Our approach employs a Score Estimator trained on select latent embeddings to embed the watermark by minimizing the score function. We then apply conformal inference to compute p-values for watermark detection. To distort the output of latent diffusion models, we shift watermarked image embeddings away from the distribution mean, distorting unauthorized generations. Experiments demonstrate our framework's superior performance in watermark detection, imperceptibility, and robustness against attacks, offering a comprehensive approach to protect images against latent diffusion models.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Charles_Xu1

Submission Number: 6722