Research And Implementation Of Image Information Hiding Algorith,m Based On Generative Adversarial Network

With the development of modern technology,multimedia information grows and spreads rapidly.At the same time,the security and privacy of multimedia information have gradually attracted people’s attention.Thus,how to effectively protect it has become an important issue to be solved urgently.Information hiding is a technology that hides secret information(text,images,etc.)in public media information,making it difficult for people to perceive its existence with intuitive vision and hearing.Without affecting the normal use of digital carriers,it can realize secret communication,copyright protection,bill anti-counterfeiting,product identification and tampering traceability.Scientific research has shown that image information hiding algorithms based on deep learning are superior to traditional algorithms in many aspects,and are scalable and easy to deploy.Among them,generative adversarial network,which uses game learning to complete information embedding and extraction task,is widely used in the field of information hiding.However,the existing information hiding network is too slow in training.Its model is complex and unstable,and the embedded information of digital carrier is low and the result is poor.Therefore,how to reduce the complexity of the model and improve the information hiding.capacity and stability of the image while ensuring the availability of the carrier based on the generative adversarial network has become a key research problem.In response to the above problems,this thesis proposes two information hiding algorithms based on generative adversarial networks,uses deep learning method to implement the algorithms,and builds an optimized visual demonstration system.The main work in the thesis is as follows:(1)A high-capacity information hiding algorithm based on an atrous spatial pyramid is proposed.The algorithm uses the idea of adversarial learning to achieve end-to-end highcapacity information hiding.An enhanced Dense Spatial Pyramid Pooling module is used in the Encoder-Decoder network to improve the image information capacity and the accuracy of information extraction.Meanwhile,a lightweight and efficient residual discriminant network is proposed to optimize the encoder network.In addition,multiple update strategies are used to train the model to improve the stability of the network.Experiments show that the algorithm can actively learn the structure and visual redundancy of the image,adaptively embed more information in the redundant area,achieve the visual and perceptual consistency of the original image and the steganographic image,and effectively resist steganography detection of analytical tools.(2)A color image information hiding algorithm based on self-attention module is proposed.The algorithm selects the U-shaped network as the image coding module to perform multi-scale learning of image features.At the same time,by means of skip connection,it completes the fusion of high and low-level semantic features and realizes the adaptive embedding of color images.The self-attention mechanism is used to expand the feature receptive field,improve the representation ability of the encoding network,the quality of the generated image and the recovery quality of the hidden image.In order to reduce the redundancy of the generated network,a lightweight structure is used to build the network,and a new multi-loss function is proposed to integrate pixel and feature-level network learning to speed up network convergence.Experiments show that the algorithm can improve the quality of steganographic images and restored images when the model is embedded in color images while keeping the network complexity low.(3)Optimize the proposed information hiding algorithm and build a visual demonstration system.Design the application system,embed the above two proposed information hiding algorithms,and visualize the algorithm.In order to improve the training speed of the network and meet the application requirements of users,the network structure in the first proposed algorithm is optimized to reduce the amount of network parameters and calculations,reduce the complexity of the model,and reduce the training time.According to the needs of the user,the selection of image data and embedded information is carried out.Through this system,the embedding and extraction of information are completed,and the protection of image data or the communication of secret information is realized.