Generative Model Watermarking Based on Human Visual System

Intellectual property protection of deep neural networks is receiving attention from more and researchers, the latest research applies model watermarking to generative models for image processing. However, existing methods designed do not take into account effects different channels sample images on watermarking. As a result, performance still limited. To tackle this problem, in paper, we first analyze embedding watermark information channels. Then, based characteristics human visual system (HVS), introduce two HVS-based methods, which are realized RGB color space YUV respectively. In space, embedded R B fact that HVS sensitive G channel. DCT domain U V brightness changes. Experimental results demonstrate effectiveness proposed work, improves fidelity be protected has good universality compared with previous methods.