Real-time attacks on robust watermarking tools in the wild by CNN

摘要

Robust watermarking is a widely used technology to protect image copyright. Robustness, the ability to resist various distortions, is the most important property of robust watermarking algorithm. So to improve the robustness of the watermarking schemes, watermark attacking algorithms also attract much attention. Far from now, the existing watermarking attack methods cannot well balance the removal ability and visual quality. To address this issue, this paper proposed a removal attack by a convolutional neural network (CNN). Considering the speed requirements of real-time attack applications, for short computing time, we use a simple but powerful CNN. According to the amount of knowledge of watermarking, a corresponding dataset of watermark images is constructed. After that, the CNN model is trained to remove watermark with these datasets. The experiments show that the trained model can not only effectively remove the watermark, but also recover the original image without much image quality degradation.