Abstract:

Deep neural networks have been successfully used in the field of digital watermarking in recent years.An encoder that embeds the watermark,a noise layer that simulates the attacks,and a decoder that retrieves the watermark make up a typical deep learning watermarking model.The commonly used end-to-end training approach requires that the attacks involved in training should be conductible,and this requirement limits the robustness capability of watermarking models in the face of real differentiable attacks.In addition,in a complete image,which usually contains both smooth and rough textures,current watermarking models seldom utilize the texture information directly to realize the embedding process of the watermark.To address the above problem,this paper proposes an iterative watermarking model guided by the image texture.An image texture attention module is introduced so that the model can guide the embedding process of the watermark,and according to the roughness of the image texture it can improve the imperceptibility of the watermark.To make the learning of unguided attacks,a two-stage iterative training approach is adopted.In order to optimize the imperceptibility of the model,the first stage involves conducting joint end-to-end encoder-decoder training without attacks,along with the inclusion of the image texture attention module.In the second stage,the decoder involved in differentiable attacks is trained independently,and the decoder with strong robustness is built by learning from any real attack distribution.Through cooperative training of encoders and decoders,which realizes a superior balance of watermarking imperceptibility and robustness,the two-stage training is eventually carried out repeatedly to the model's global optimum.According to experimental findings,the proposed approach outperforms popular deep learning watermarking techniques in terms of both imperceptibility and robustness.

Key words: deep learning, image watermarking, texture, robustness, imperceptibility