ADDW: Adaptive dynamic dual-authentication watermarking in black-box models

Model watermarking can be used to protect the intellectual property of deep learning models. However, existing model watermarking schemes often suffer from limited robustness and concealment. Moreover, copyright infringement issues are easily overlooked. In this paper, we propose an adaptive dynamic dual-authentication watermarking method for black-box models, which enhances model copyright protection and prevents fraudulent ownership claims in black-box scenarios. Specifically, we exploit both the edge contour structures of images and core semantic information to adaptively generate dynamic watermarks that are highly correlated with image contents. Subsequently, we develop an identity watermarking scheme based on edge contour coloring to embed the copyright owner’s exclusive color identity information. With the guidance of attention features, the scheme collaboratively determines optimal embedding positions for the identity information. Extensive experiments on four datasets and three network models demonstrate that, compared with four baseline methods, the proposed watermarking scheme achieves superior imperceptibility and robustness while maintaining a stable watermark extraction success rate.