Image Encryption And Its Analysis Based On Cycle-Consistent Generative Adversarial Network

With the rapid development of mobile Internet and social software,images are widely used in network information dissemination because of their intuitive,visualizable and vivid characteristics.Therefore,image information leakage has become a common problem.One of the effective methods to ensure the security of image transmission is image encryption technology.Image encryption can transform the image into completely different meaningless images by randomly transforming the position and value of image pixels.The traditional image encryption methods are mostly manual design methods,which involve much complex expert knowledge in the design process.However,the existing deep learning image encryption methods are faced with insufficient performance in terms of security,generalization ability,encryption and decryption efficiency and so on.Taking digital image as the research object,this paper studies and analyzes the image encryption method based on cyclic consistent generative adversarial network to solve various common problems in deep learning image encryption.The main contributions are summarized as follows:1.An image asymmetric encryption method based on cycle-consistent generative adversarial network is proposed.Taking image encryption as a migration of image style,an autoencoder image encryption model is constructed.Firstly,an encoder-decoder framework is proposed to simulate the process of image encryption and decryption.Based on the generator network of c cycle-consistent generative adversarial network,dense residual network is used to simplify the structure of neural network as the main part of encryption network and decryption network.Secondly,the high uncertainty of neural network parameter initialization and training makes the model parameters different after each training.The parameters of the encryption network and the target random image can be regarded as the public key,and the parameters of the decryption network can be regarded as the private key.Finally,an adjustable weight loss function model is proposed,which combines the encryption loss function,decryption loss function and total variational loss function to train the model.Experiments show that this method can complete image encryption and decryption in multiple scenes,and has high decryption quality and randomness of encrypted image.The combination training of encryption loss,decryption loss and total variation loss is the key technology to realize the ideal image encryption and recovery results.Adding decryption total variation loss will help to improve the quality of decrypted image.2.Analyze the avalanche effect of cycle-consistent generative adversarial network.Although the ciphertext image generated by the image encryption method based on cycle-consistent generative adversarial network is visually hidden,the ciphertext change intensity under the select plaintext attack is low,and the ability to resist select plaintext attack is weak.From the perspective of neural network structure and operation process,this paper analyzes the causes of weak avalanche effect of cycle-consistent generative adversarial network generator.Firstly,the operation and input-output relationship contained in the cycle-consistent generative adversarial network generator are analyzed,and then the influence of layer-level output and each part of the network on the avalanche effect is observed through statistical experiments.The defect characteristics and reasons of the low avalanche effect of the cycle-consistent generative adversarial network generator as an encrypted network are analyzed.It is pointed out that with the deepening of convolutional neural network,the avalanche effect increases slowly,the generation ability decreases,the resource consumption is large and the encryption efficiency is low.Therefore,it is difficult to improve the avalanche effect of the cycle-consistent generative adversarial network generator under the condition of ensuring the encryption and decryption quality only by relying on the structure deepening of the neural network.3.An image encryption method based on cycle-consistent generative adversarial network and combined with traditional diffusion algorithm is proposed to make up for the weak anti select plaintext attack ability of the image encryption method based on cycle-consistent generative adversarial network,and enhance the avalanche effect and ciphertext image security.Firstly,the image is scrambled and diffused by neural network,and then the pixels are further diffused by bit-XOR diffusion algorithm.Experimental results show that this method will greatly enhance the avalanche effect,improve the plaintext sensitivity,and resist differential attacks more effectively.In addition,this method can improve the information entropy of ciphertext image,reduce the correlation and randomness of adjacent pixels,and has the ability to resist a variety of attacks.Compared with the recently proposed image encryption method based on deep learning,this method has advantages in encryption and decryption efficiency and quality,key space,and resistance to selective plaintext attack.In addition,data loss or noise attack will leave traces in the decrypted image obtained by this method,which increases the traceability of data loss and noise attack.