A Deep Learning Based Attack For The Chaosbased Image Encryption

In recent years,deep learning has made relatively great breakthroughs in image processing and speech recognition,but its application in the field of cryptography and information security is still relatively small.This thesis first applied deep learning to cracking chaotic encrypted images.Compared with the traditional attack method based on cryptography,the method in this thesis does not need to manually analyze the vulnerability of the algorithm to infer the encryption key,but directly attacks the encryption algorithm itself in a key-independent manner,and the non-linear mapping ability of the neural network can extract the hidden features of the ciphertext image,thus ensuring the validity of the model.The trained attack model can automatically reconstruct the plain image from the cipher image with high fidelity.The main contents include:(1)We propose an attack model based on pix2pix,and the chaotic image encryption algorithm: Arnold's cat map,Baker's map,Baker-sub,The chaos-based image encryption algorithm,Two-dimensional logistic chaotic map,and non-chaotic images encryption algorithms: XOR-OTP-RC4-PIX,XOR-OTP-RC4-MSB,XOR-OTP-CSTD-PIX,XOR-OTP-CSTD-MSB,XOR-Followers conducted attack experiments.In the experiment,it is basically possible to completely crack the relatively simple encryption algorithms,such as XOR-OTP-RC4-MSB and XOR-Followers,the cracking rate of these two algorithms in the case of static key and dynamic key encryption has reached about 95%;For simple encryption algorithms,such as Arnold's cat map and The chaos-based image encryption algorithm can crack most encrypted content.The cracking rate of both algorithms under static key encryption has reached more than 94%,and in the case of "One-Time-Pad",the cracking rate of Arnold's cat map is 67.56%,and the cracking rate of The chaos-based image encryption algorithm is 44.14%;For medium difficulty encryption algorithms,such as Baker's map,only part of the content can be cracked.the cracking rate of under the static keys is 95.76%,and the cracking rate in the case of "One-Time-Pad" is only 20.04%;The proposed method is unable to crack the encryption algorithm with high difficulty and strong security,such as Two-dimensional logistic chaotic map.The experimental results prove that the proposed method has certain effectiveness,and preliminarily verifies that deep learning may directly crack the image encryption algorithm.(2)An attack model Decrypt Net based on a fully convolutional neural network is proposed.The experimental results of this model eliminate the "dilation" phenomenon in the pix2pix attack model and improve the cracking performance.Among them,the cracking rate of the two encryption algorithms Arnold's cat map and The chaos-based image encryption algorithm has been improved considerably.In the case of " One-Time-Pad ",the cracking rate of Arnold's cat map has increased by 8.8 percentage points,reaching 76.42%,the crack rate of The chaos-based image encryption algorithm increased by 11.7 percentage points to 55.85%,and the cracking rates of Baker's map,Baker-sub,XOR-OTP-RC4-PIX,XOR-OTP-RC4-MSB,XOR-OTP-CSTD-PIX,XOR-OTP-CSTD-MSB,XOR-Followers were also improved.The experimental results prove that the Decrypt Net attack model is more effective in cracking encrypted images than the pix2pix attack model.