Research Of Image Information Hiding Based On GAN

Image information hiding,as the main technical means of secret communication,transmits secret information without being detected by third parties through information hiding algorithms,and extracts secret information using inverse methods,so as to achieve information security transmission.Although the existing image information hiding based on traditional methods and deep learning has better effect and can ensure the quality of images,the traditional methods are limited with the artificially designed features,while the deep learning based methods only start from the network construction and ignore the fact that the deeper the network is,the more likely it is to lead to inaccurate feature extraction or learn some redundant features,so there are shortcomings in image quality,the amount of embedded information and the resistance to information Therefore,there are shortcomings in image quality,embedded information and resistance to information hiding analysis.To address the above problems,this paper firstly proposes a generative adversarial network model based on void space pyramid pooling,which is mainly void space pyramid model and codec network.Since the generative adversarial network model has certain limitations in global feature extraction,and the interdependence between network structures tends to cause a certain redundancy in feature extraction,a codec network is designed,which is used for feature extraction through discriminators and codec networks,and the cavity space pyramid pooling model is added to combine multiple scales of feature information to reconstruct,which well compensates for the semantic information the lack of detailed features,and uses generative adversarial approach to train the model,thus improving the security of image information hiding.Based on the proposed method,an attention mechanism is embedded and a combined attention coding and decoding GAN image information hiding network architecture is proposed.The main idea is to make the generator network extract information-rich features from the input image and retain the complete structural features of the input image object in order to improve the resolution of the generated image,and use the attention model for feature extraction in the generator network to improve the robustness of the network feature extraction.In addition,by constructing a feature fusion model between codec networks,the features of different layers are organically fused,and the attention mechanism is used to learn their features so that the model focuses more on important features and suppresses irrelevant and redundant features.Through experiments and simulations of adversarial training under various noises,and comparing them with the original method,we finally verify the effectiveness of each model in the proposed method and the rationality of the proposed method,which has certain advantages over the existing methods,reduces the loss of encoder and overall model and improves the accuracy of decoder.