Deep Learning Based Image Watermarking Attacking

In the era of the mobile internet,a large number of digital images are produced and widely spread.In the process of sharing and transmission of digital images,it is often necessary to determine the ownership or source of these images.In order to pro-tect the copyright or trace the source of these digital images effectively,many robust digital watermarking algorithms were applied.At the same time,in order to measure the robustness of watermarking algorithms,watermark attacking algorithms have also attracted the attention of scholars.They can promote each other and complement each other.After years of development,robust digital watermarking algorithms have been able to effectively resist conventional attacks.However,since the end of last century,watermark attacking algorithms have been stagnant.A good watermark attacking algo-rithm can often inspire researchers to design a more reliable and robust watermarking algorithm to protect the digital image,so the attacking algorithms are also of great sig-nificance in the watermarking algorithm research system.The main purpose of watermark attacking algorithms is to destroy the extraction of watermark.In the process of attacking,it is usually assumed that the attacker can use all the information except the secret key to destroy the extraction process of watermark information.The most effective way of watermark attacking is called removal attack,which aims to completely remove the watermark information in the host image.Tradi-tional removal attacks destroy the watermark information,but also degrade the image quality.After years of development,the watermark algorithms are robust to traditional removal attacks.In recent years,neural network algorithms have shown a strong ability in the field of image restoration.The deep convolution neural network can learn the prior distribution of the image and recover the image effectively.Therefore,the robust-ness of traditional watermarking algorithms is facing new challenges in the era of the neural network.In this thesis,we designed and trained neural network models as the removal attack to test the robustness of watermarking algorithms.The specific research scheme includes the following two points:(1)Attack scheme of robust watermarking algorithm based on the convolu-tional neural networkIn order to remove the watermark and maintain the image quality,this thesis uses convolutional neural network model to remove the robust watermark.According to the knowledge of the watermark algorithm,the noise image with different similarity to the watermarked image distribution is generated.In the white box attack of the known wa-termarking algorithm,noisy images with a high degree of similarity to the distribution of watermarked images are generated by using all algorithm knowledge except the se-cret key.In the black box attack of the unknown watermarking algorithm,according to the commonness of robust watermarking algorithms,the noisy image is generated by the template with high value of medium and low frequency components.Then noise images and corresponding original images are used as image pairs to train the model.By using the simplified data manifold and the effective network structure,we can re-duce the upper bound of loss risk of the model.The experimental results show that the convolutional neural network model trained in this scheme can effectively remove the watermark information from the image without degrading the image quality.(2)Attack scheme of robust watermarking algorithm based on the generative adversarial networkIn this thesis,we propose a robust watermarking algorithm attack scheme based on generative adversarial network.In the white box attack,the watermark is removed by truncating the watermark coefficient to a fixed value.In the black box attack,the watermark information is destroyed by high-intensity noise.In the process of removing the watermark,it will inevitably lose the image details.In order to make up for this part of the detail information,we use the structure of the generative adversarial network to complete the learning of the lost information.The generator is used to generate image details,and the discriminator is used to supervise that the generated details conform to the original image distribution.The experimental results show that the scheme can effectively destroy the watermark information under the condition of ensuring the visual quality of the image.