An invisible and robust watermarking scheme using convolutional neural networks

摘要

Although zero-watermarking schemes based on image features demonstrate perfect imperceptibility, they are exposed to the weak robust problem. This is because the zero-watermark information construction and the watermark information detection depend on XOR logic operation. In this study, a watermarking scheme that constructs a zero-watermark image by superimposing the color style of the host image on the content of the watermark logo using convolutional neural networks is proposed. A stylized image is generated by iterating between the style of the host image and the content of a watermark logo added with a timestamp. The stylized image is then encrypted via Arnold transform and is registered in the intellectual property rights as the zero-watermark image. The image copyright is verified using the convolutional neural network whose loss function is defined as the difference between the original watermark logo and its output image. Its training dataset consists of many image pairs, each of which is composed of the host image under an attack and the decrypted zero-watermark image. Experimental results show that the proposed zero-watermarking scheme is highly robust to both common image processing and geometric attacks, and its performance far surpasses those of the existing zero-watermarking methods and non-zero-watermarking methods.