Adaptive Robust Watermarking Method Based on Deep Neural Networks

Fan Li, Chen Wan, Fangjun Huang

Abstract: Aiming at the problem of digital multimedia piracy and infringement, an adaptive robust watermarking algorithm based on Deep Neural Networks (DNNs) is proposed. In our method, the watermark sequence to be embedded is mapped to a noise pattern first, which has the same dimension as the carrier image. Specifically, the noise pattern is generated adaptively according to the statistical properties of the carrier image, in which the noise intensity corresponding to the texture area of the carrier image is large, and that corresponding to the smooth area is small. Thus, after adding the generated noise pattern to the carrier image, good visual quality can be easily obtained. Furthermore, considering a series of attacks such as adding noise and JPEG compression, the watermark encoder and decoder in our scheme are jointly trained to resist the potential attacks in the physical world. Experimental results demonstrate that better visual quality and higher robustness can be obtained compared with those state-of-the-art algorithms based on DNNs. This means that we have better solved the problem of mutual restriction between visual quality and robustness.