Image Encryption Based On Deep Learning And Its Application In AI System

With the rapid development of artificial intelligence technology,a large amount of personal sensitive image information may be exposed to the public in the process of using AI models,which brings privacy and security issues.Image encryption is a technique that prevents unauthorized users from accessing its content by disguising an image as a noise or nonsensical appearance.Although the traditional image encryption method can protect the content of image data to a certain extent,it must be decrypted by traditional algorithms.This inflexibility makes it difficult to organically integrate with other deep learning intelligent computing models,and traditional image encryption.The combination of the method and other intelligent artificial intelligence systems is likely to cause privacy leakage problems.In view of the above problems,the main work and contributions of this paper are as follows:(1)An agile image encryption method based on deep learning is proposed.The method generates a pseudo-random integer sequence based on chaotic mapping,and then periodically expands the sequence into a binary key plane to participate in the encryption and decryption process,and control the key of the authority.It can be dynamically updated,that is,changing the key does not require retraining the network,and has better flexibility.(2)The multi-scale dilated convolutional network is used to encrypt and decrypt the image,and the target image binary reduction strategy is used when constructing the loss function,that is,the target grayscale image is regarded as the likelihood distribution of the binary image to design Differentiable loss function.Two post-processing methods,obfuscation or diffusion,are proposed to further process the encrypted image,which hides the original information of the image to the greatest extent,and further strengthens the security of the encrypted image.Experimental results show that the proposed encryption method has high security and efficiency,can resist various noise and differential attacks,and is more flexible and convenient than other deep learning-based methods.(3)In order to protect the privacy and security of users in the process of using the AI model,this paper further expands the encryption method,and integrates the proposed encryption method with other intelligent computing deep neural networks(DNN)seamlessly,and proposes a ciphertext domain The intelligent computing DNN model can perform relevant intelligent computing on ciphertext images.The experimental results show that authorized users can predict diseases and extract medical organs in the ciphertext environment without decrypting the encrypted images in advance.