Research On Digital Watermarking Technology Based On Neural Network

Digital watermarking technology is a technology that embeds specific digital signals into digital products to protect the integrity of copyright.It is an effective method to protect information security.Neural network technology is a technology that imitates the structure and function of biological neural network to generate mathematical model.It has excellent self-study function and optimization ability.It is feasible to add neural network to digital watermarking technology for copyright protection.This paper selects and improves the generative countermeasure network in the classical neural network,and uses the watermark data set to generate the countermeasure samples needed for the follow-up research.The improvements include: combining the high-level class loss and low-level feature loss,and using the antagonistic label to promote the antagonistic classification.The improved countermeasure sample has the advantage of low degree of change to the original image.The subsequent research mainly proposes to use the countermeasure samples to attack in two aspects: watermark detection and watermark removal.As for the former,this paper selects three mainstream neural network target detection algorithms:Fast-RCNN,Yolo V3 and SSD300,uses the watermark data set to train and generate the detection model,and tests it through the countermeasure samples.The map of the three target detection models is reduced to less than 50%.The countermeasure samples have good attack effect on the target detection model and can be migrated at the same time.For the watermark removal attack,this paper selects the classic neural network denoising algorithm noise2 noise algorithm,and improves it by introducing residual module and adding hole convolution to generate an algorithm model suitable for watermark removal.The watermark data set and countermeasure samples are used to test the watermark removal.The average watermark removal rate of the countermeasure sample is about 20% lower than that of the watermark data set,which shows that the countermeasure sample has a certain attack effect on the watermark removal model.