Image Watermarking Covert Attack Technology Based On Deep Learning

With the rapid development of Internet technology,image information acquisition becomes more convenient,and massive images can be freely stored and transmitted.However,how to protect the security of images becomes a key issue.In this regard,digital watermarking technology provides effective support for image security issues such as copyright protection and data monitoring.However,in the process of transmitting watermarked images,they may be affected by watermark attack techniques,which in turn affect the correct extraction and integrity of watermark information by watermark algorithms.The emergence of watermark attack techniques makes it possible to verify,test and optimize robust watermark algorithms.An excellent robust watermark algorithm needs to prove its ability to resist malicious attacks through watermark attack techniques,thereby promoting the optimization and improvement of the anti-attack ability of watermark algorithms.However,the current watermark attack methods mainly focus on traditional attack methods,and most watermark algorithms have the ability to resist these attacks,which leads to the inability of the existing evaluation system to comprehensively and effectively evaluate the performance of watermark methods.In addition,although the current watermark attack methods can interfere with the extraction of watermark information,they cause a large degree of damage to the watermarked images,reduce the quality of the images,and thus affect the practical value of the images.Therefore,it is imperative to research and develop highly covert new watermark attack methods.This paper focuses on the research on the concealment of watermark attack techniques,aiming to design a method that can effectively remove the watermark information in the image while minimizing the damage to the watermarked image.The main work of this paper is as follows:(1)Proposing a covert digital watermark attack method based on image denoising residual networks.This method utilizes an image denoising residual network to design a covert watermark attack method that effectively attacks the watermark information contained in grayscale images.The residual network takes a watermarked image as input and outputs a low-frequency residual feature image that contains most of the watermark information.By removing this feature image from the original image,the effect of attacking the watermark information can be achieved.In the attacked image,the majority of the watermark information is successfully removed while preserving the details and other features of the original image to the maximum extent.Experimental results demonstrate that this attack method can effectively remove most of the watermark information and outperforms other traditional watermark attack methods.(2)Proposing a black-box covert attack method based on watermark inversion.This method addresses two common watermark attack scenarios: unknown watermark embedding algorithms and known watermark embedding algorithms,by proposing black-box and white-box watermark attack schemes.Under the black-box attack mode,a watermark inversion attack method based on non-subsampled contourlet transform is employed to attack unknown watermark algorithms.Under the white-box attack mode,a watermark inversion attack method based on residual channel attention network is used to attack known watermark algorithms.This method not only extends the study of singlechannel grayscale images to the domain of three-channel color images but also introduces a perceptual loss function to ensure that the watermarked image after the attack has good visual quality.Experimental results demonstrate that this method achieves excellent results in both attack modes,proving to be a novel and effective covert watermark attack method for color images.