CWAN: Covert Watermarking Attack Network

Abstract

Digital watermarking technology is widely used in today’s copyright protection, data monitoring, and data tracking. Digital watermarking attack techniques are designed to corrupt the watermark information contained in the watermarked image (WMI) so that the watermark information cannot be extracted effectively or correctly. While traditional digital watermarking attack technology is more mature, it is capable of attacking the watermark information embedded in the WMI. However, it is also more damaging to its own visual quality, which is detrimental to the protection of the original carrier and defeats the purpose of the covert attack on WMI. To advance watermarking attack technology, we propose a new covert watermarking attack network (CWAN) based on a convolutional neural network (CNN) for removing low-frequency watermark information from WMI and minimizing the damage caused by WMI through the use of deep learning. We import the preprocessed WMI into the CWAN, obtain the residual feature images (RFI), and subtract the RFI from the WMI to attack image watermarks. At this point, the WMI’s watermark information is effectively removed, allowing for an attack on the watermark information while retaining the highest degree of image detail and other features. The experimental results indicate that the attack method is capable of effectively removing the watermark information while retaining the original image’s texture and details and that its ability to attack the watermark information is superior to that of most traditional watermarking attack methods. Compared with the neural network watermarking attack methods, it has better performance, and the attack performance metrics are improved by tens to hundreds of percent in varying degrees, indicating that it is a new covert watermarking attack method.