Watermarking Image Processing Models for Intellectual Property Protection

Deep learning has been successful in various computer vision tasks. However, training deep models is computationally expensive and requires significant high-quality data, making a good pre-trained model highly valuable. However, there is a growing concern about the security of pre-trained deep models. When these models are shared or deployed widely, they become exposed to the risk of illegal theft. This raises concerns about how to protect the intellectual property (IP) of model owners. One solution is to use deep model watermarking. Most recent research has focused on protecting classification networks. However, image processing models are still under-researched, and existing methods lack generality. This paper presents a novel model watermarking method to protect image processing models that also supports black-box verification. We watermark the target model by training it on watermarked training data. The watermarked model learns to embed the watermark into its output images, which can be extracted to realize copyright protection. Our method has been extensively tested, and the results demonstrate its fidelity, uniqueness, and robustness.