Image Encryption Algorithms Based On Adversarial Neural Network And Chaotic System

With the continuous development of digital technology,digital image information has been widely used in all aspects of daily life,such as communications,education,medicine due to its strong comprehensiveness and intuitiveness.At the same time,digital images may be illegally stolen or tampered by attackers during the process of storage and transmission,which will greatly harm the interests of legal information owners.Therefore,it is very important to develop a safe and reliable digital image encryption scheme.Recent years,various signal processing methods have been used in image encryption,such as Fourier transform,discrete cosine transform,Mellin transform,chaotic map and their fractional versions.However,there are few reports of deep learning-based image encryption in literature.This dissertation presents two image encryption algorithms based on adversarial neural networks,and conducts relevant simulation experiments and analysis.The final results verify that these two algorithms have a high security.The main research contents are as follows.(1)To overcome the vulnerability to known-plaintext attack or chosen-plaintext attack of a linear image encryption system,a new approach for image encryption is proposed based on adversarial neural cryptography combined with SHA-256 controlled chaotic systems.In this scheme,the optimal network model is first obtained by training an adversarial neural network(ANN),and then the trained network model is used to achieve a noise-like intermediate image.Subsequently,the XOR operation based on a logistic map is performed on the intermediate image to obtain the final ciphertext.The intrinsic non-linearity of the neural network guarantees the ability of the proposed system to resist common attacks.The plaintext dependent SHA-256 controlled logistic map greatly improves the diffusion performance so that the encryption system can resist differential attacks.Numerical simulation results prove the reliability,effectiveness,and security of the proposed scheme.(2)An image encryption algorithm based on Kronecker inner product matrix over a finite field and adversarial neural network is designed.In this scheme,the secure hash function SHA-256 is used to highly correlate the plaintext image with the key,and the hash value generated by SHA-256 is normalized to control the Logistic-Sine chaotic mapping,and then the plaintext image is mapped to the finite field,and the pixels are scrambled and diffused based on the Kronecker inner product matrix over a finite field.Finally,the adversarial neural network is performed for further scrambling.Similarly,as the neural networks have intrinsic non-linearity characteristics,this means that the encryption system of the algorithm has highly nonlinear characteristics.In addition,the look-up table method is used to construct the addition and multiplication operations based over G F(2 ~8)finite fields,which can effectively improve the operation speed.At the same time,relying on the secure hash function SHA-256 of the plaintext image to control the Logistic-Sine mapping greatly improves the diffusion and security of the encryption system.Various tests are performed on the image encryption algorithms involved,and the simulation results prove that the algorithms have high security and sufficient key sensitivity,and have good resistance to many common attacks.