Abstract
Traditional image watermarking technology focuses on the robustness and imperceptibility of the copyright information embedded in the cover image. However, in addition to copyright theft, the cover images stored and transmitted in the open network environment is facing the threat of being identified and retrieved by deep neural network (DNN) with malicious purpose, which is a new privacy threat. Therefore, it is essential to protect the copyright and the privacy of cover image simultaneously. In this paper, a novel cover-concealed image watermarking (CCIW) is proposed, which combines conditional generative adversarial networks with channel attention mechanisms to generate adversarial examples of the cover image containing invisible copyright information. This method can effectively prevent privacy leakage and copyright infringement simultaneously, since the cover image cannot be collected and processed by DNNs without permission, and the embedded copyright information is hardly to be removed. The experimental results show that the proposed method achieved a success rate of adversarial attack over 98% on the Caltech256 dataset, and the generated adversarial examples have good image quality. The accuracy of copyright information extraction is close to 100%, and it also exhibits good robustness in different noise environments.