Image multi-hiding encryption method based on deep learning and ghost imaging

Abstract

The invention discloses a multi-image hiding encryption method based on deep learning and ghost imaging. Comprising the following steps: 1, converting a secret image into a spectrogram through two-dimensional fast Fourier transform (FFT); 2, a designed equal-resolution image steganography model ERIH-Net is utilized, the spectrogram and the carrier image serve as input of the ERIH-Net model, and a secret-containing image is generated; 3, sampling the secret-containing image for multiple times through a ghost imaging system to obtain a BD value sequence of the one-dimensional bucket detector; 4, reconstructing a secret-containing image by using the light field information re-modulated by the ghost imaging system and the acquired BD value through a correlation reconstruction algorithm; 5, taking the secret-containing image as the input of an ERIH-Net model image extraction sub-network, and extracting a frequency spectrum image; and 6, decrypting the secret image through two-dimensional inverse Fourier transform (IFFT). According to the invention, the security and information hiding capacity of the optical image encryption system can be improved; by cutting part of high-frequency information in the spectrogram, the transmission quantity of the ciphertext can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of optical image encryption, and in particular to an image multiple hiding encryption method based on deep learning and ghost imaging. Background Art

[0002] Ghost imaging plays an important role in optical image encryption because of its advantages such as non-localized imaging, strong anti-interference and high sensitivity. Information hiding technology is also an important research direction in information security and is now widely used in copyright protection fields such as digital watermarking and optical authentication. However, the current mainstream information hiding technology has the disadvantage of small encryption capacity. For example, digital watermarking technology based on discrete wavelet transform (DWT) can usually only hide 1 / 4 of the size of the carrier image, or even 1 / 8 or less image or text information. Summary of the invention

[0003] In view of the shortcomings in the prior art, the purpose of the present invention is to provide a multiple hidden encryption method for images based on deep learning and ghost imaging. Different from the previous research on applying deep learning to ghost imaging image reconstruction, deep learning is applied to the process of image hidden encryption, which can achieve the hiding and extraction of equal-resolution images. After the hidden image is encrypted by the computational ghost imaging technology and decrypted by the second-order association, the information of the plaintext image can still be extracted from the encrypted image, indicating that the encryption scheme has good robustness, realizes the hidden encryption of equal-resolution images, and combined with the spatial multiplexing encryption method, it greatly improves the information hiding capacity of the hidden encryption system. In order to achieve the above-mentioned purpose and other advantages according to the present invention, a multiple hidden encryption method for images based on deep learning and ghost imaging is provided, including: S1. Convert the plaintext image to be encrypted into a spectrum graph through two-dimensional Fourier transform; S2. Design and train a deep learning image steganography model ERIH-Net for equal-resolution image hiding, which consists of two sub-networks, namely, the image hiding network Hide-Net and the image extraction network Extract-Net. The image hiding network Hide-Net in the pre-trained ERIH-Net is used to hide a plaintext image into a non-secret image, generating a secret image containing the plaintext image feature information; S3, loading 16384 Hadamard matrices through DMD to modulate the light field and generate illumination speckles, and using the modulated illumination speckles to illuminate the secret image; collecting the total light intensity value of the secret image through the bucket detector to obtain a sequence of 16384 light intensity values, i.e., the ciphertext information; S4, using the second-order correlation image reconstruction algorithm to reconstruct the secret image information through the ciphertext sequence and the Hadamard modulation mode (key); S5, using the secret image reconstructed by the second-order correlation as the input of the extraction network Extract-Net to extract the initial secret image information; S6. Decode the spectrum map into plaintext image information of spatial pixel distribution through inverse Fourier transform.

[0004] In the image frequency domain conversion process in step S1, the input image is first converted into a grayscale image; secondly, a two-dimensional fast Fourier transform (FFT) is performed on the grayscale image to obtain a complex matrix; finally, a shift operation is performed on the complex matrix so that the zero frequency component is located at the center of the matrix; Step S2: The medium-resolution image steganography model ERIH-Net performs feature extraction and downsampling through multiple convolutional layers to obtain a high-dimensional feature map; based on the high-dimensional feature map, upsampling and feature reconstruction are performed through a deconvolution layer to obtain an imaging result with the same size as the original image; the resolution of the input plaintext image and the non-secret image is the same, and the size is 128×128×1 in the experiment.

[0005] The optical ghost imaging system in step S3 includes a He-Ne laser with a wavelength of 532.8 nm, a beam expander arranged on one side of the laser, a beam collimating lens, a digital micromirror device DMD, a focusing lens arranged on one side of the DMD, a bucket detector BD without spatial resolution capability, and a computer PC.

[0006] In step S4, the ciphertext is received through a public channel, the key is received through a secure channel, and the image reconstruction algorithm uses a second-order correlation algorithm.

[0007] In step S5, Extract-Net consists of 6 different convolutional layers. The first 5 convolutional layers are connected with Relu activation function, and the last convolutional layer is followed by tanh activation function.

[0008] In step S6, the inverse Fourier transform process is applied to the phase angle information in the secret image spectrum, otherwise the original secret image information cannot be reconstructed. In addition, the transmission capacity can be improved by spatial multiplexing and other methods, and the square area of ​​the secret image spectrum with a distance of 64 pixels from the center is set to 0. This can increase the capacity of the transmitted information while ensuring the image quality. If the boundary of the 0 pixel area is less than 64, the human eye will not be able to distinguish the image information.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) The deep learning-based image steganography model is applied to the camouflage encryption process in computational ghost imaging encryption. Compared with other traditional steganography methods, the image hiding process based on deep learning is simple to operate and has strong generalization ability.

[0010] (2) The equal-resolution image steganography model can realize hiding and extraction between images of the same resolution size. After computational ghost imaging reconstruction, the extraction network of the steganography model can also extract the information of the plaintext image well, indicating that the model has good robustness.

[0011] (3) The combination of encoder-decoder based deep learning image steganography network and computational ghost imaging encryption, as well as the use of two-dimensional fast Fourier transform algorithm, can improve the information hiding capacity, security, and imaging quality of camouflage image encryption. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a system flow chart of the image multiple hiding encryption method based on deep learning and ghost imaging according to the present invention; Figure 2 It is a structural diagram of ERIH-Net of the image multiple hidden encryption method based on deep learning and ghost imaging according to the present invention; Figure 3 A ghost imaging optical path structure diagram of an image multiple hiding encryption method based on deep learning and ghost imaging according to the present invention; Figure 4 A system encryption structure diagram of the image multiple hidden encryption method based on deep learning and ghost imaging according to the present invention; Figure 5 A system decryption structure diagram of the image multiple hidden encryption method based on deep learning and ghost imaging according to the present invention; Implementation

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] Reference Figure 1-5 , an image multiple hiding encryption method based on deep learning and ghost imaging, comprising the following steps: S1, convert the plaintext image into a spectrum through two-dimensional fast Fourier transform (FFT); S2. Use deep learning methods to train several image samples to obtain an equal-resolution image steganography model ERIH-Net based on an encoder-decoder structure. Use the spectrum of the secret image and a non-secret carrier image as the input of the Hide-Net image hiding network of the ERIH-Net model to generate a secret image containing the feature information of the plaintext image. S3, loading a series of Hadamard phase modulation matrices with a Cake-cutting arrangement to modulate the light field through a digital micromirror device DMD to generate a number of illumination speckles; illuminating a dense image through the illumination speckles, calculating the light intensity information within the image space range using a bucket detector device BD that does not have spatial resolution capability, and generating a series of light intensities; S4, reconstructing the encrypted image using the modulated light field information and the collected bucket detector values ​​through a second-order correlation reconstruction algorithm; S5, the reconstructed encrypted image is used as the input of the ERIH-Net model image extraction network Extract-Net, and the spectrum image information of the plaintext image is successfully extracted; S6, is converted into spatially distributed secret image information through two-dimensional inverse Fourier transform (IFFT). Example 1

[0015] Step S1, first convert the secret image into a grayscale image through matlab2018b software and scale it to the size required for the experiment; secondly, perform a two-dimensional fast Fourier transform (FFT) on the grayscale image to obtain a complex matrix; finally, shift the complex matrix so that the zero-frequency component is located at the center of the matrix to obtain the image pixel distribution spectrum.

[0016] Step S2, firstly, a uniform resolution image steganography model ERIH-Net based on encoder-decoder structure is designed on the pytorch deep learning framework using Python language. The structure diagram of ERIH-Net is shown in Figure 2 As shown in the figure, it includes two sub-networks: image hiding network Hide-Net and image extraction network Extract-Net. Hide-Net includes preprocessing convolution layer, downsampling convolution layer, and transposed convolution upsampling layer; Extract-Net consists of 6 convolution layers. The preprocessing block can adjust the size of the secret image to the same resolution as the carrier image, and extract the high-dimensional features of the original image; the convolution upsampling block and the transposed convolution downsampling block can fuse the high-dimensional feature map after the two images are spliced, and finally generate the secret image containing the secret image information through the Tanh activation function. Set the MSE loss function as the loss function of the network, and the carrier image c Encrypted Image c 'The mean square error loss between the secret image s With the extracted secret image s 'The mean square error loss between β The sum of times is used as the total loss function of the equal-resolution image steganography network for training iterations ( β is the weight of the reconstruction error). This network uses the Adam optimizer to accelerate training, and the initial learning rate lrSet to 0.001, the training cycle is 200 rounds, and after the 80th and 150th training, the learning rate will be 1 / 10 of the original. The batch size is set to 8, the image size is 128×128×1, the data set is 7300 Oxford-IIIT_Pets images processed by Matlab2018b software, and the test set is a number of Set12, Fashion Mnist and binary images, all scaled to 128×128 size; the 128×128 pixel size secret image and carrier image are used as the input of Hide-Net, and the pre-trained model parameters are called to hide the secret image, generating a carrier image containing the secret image features.

[0017] The optical path structure diagram of the ghost imaging calculation in step S3 is as follows: Figure 3 As shown in the figure, the beam emitted by the laser is expanded and collimated to illuminate the image of the transmitted object. The beam carrying the amplitude information of the object passes through the object and then irradiates the DMD (Digital Micromirror Device) loaded with a series of Cake-Cutting Hadamard phase modulation matrices. After the phase information of the light field is modulated by the DMD, the reflected light intensity information is collected by the bucket detector (Bucket Detector), which is recorded as D i The DMD loads 16384 random phase modulation matrices in sequence to obtain 16384 bucket detector values, which the sender transmits to the receiver through a public channel as ciphertext, and flattens the 16384 phase modulation matrices into 16384 one-dimensional vectors as keys and transmits them to the receiver through a secure channel.

[0018] Step S4, after the receiver receives the ciphertext and key from the public channel and the secure channel, the corresponding light field distribution information can be calculated according to the Fresnel diffraction theorem and the random phase modulation matrix in the key. The receiver performs a second-order correlation operation on the decrypted light field information and the ciphertext information to obtain the decrypted image information. The compressed sensing reconstruction algorithm has the disadvantages of reducing the feature gradient of the reconstructed image and long imaging time during the decryption process. In order to improve the imaging quality at a low sampling rate, this paper adopts a second-order correlation optimization decryption algorithm to achieve high-quality reconstruction of secret information.

[0019] Step S5, the receiver uses the secret image reconstructed by the second-order correlation algorithm as the input of the extraction network to extract the secret image information reconstructed by the correlation imaging; Step S6, the phase angle information in the secret image spectrum is applied to the inverse Fourier transform process, otherwise the original secret image information cannot be reconstructed. In addition, the transmission capacity can be improved by spatial multiplexing and other methods, and the square area of ​​the secret image spectrum with a distance of 64 pixels from the center is set to 0. This can increase the capacity of the transmitted information while ensuring the image quality. If the boundary of the 0 pixel area is less than 64, the human eye will not be able to distinguish the image information.

[0020] In summary, the current mainstream ghost imaging camouflage image encryption research has the disadvantages of small image hiding capacity and complex image hiding process. In view of this phenomenon, the present invention analyzes various image hiding methods, such as digital watermarking, image steganography and other methods, and finds the advantages of hiding and extracting images based on deep learning image steganography models, which have the advantages of high information hiding capacity, strong generalization ability, and simple operation. Then the second-order correlation image reconstruction method is analyzed, and it is found that compared with other image reconstruction algorithms such as compressed sensing, this method has better imaging quality at a lower sampling rate, excellent performance, good robustness, and strong anti-interference ability, which can improve the image reconstruction effect of computational ghost imaging.

[0021] The number of devices and processing scales described here are used to simplify the description of the present invention, and the application, modification and variation of the present invention will be obvious to those skilled in the art.

[0022] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and implementation modes. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A method for multiple image hiding and encryption based on deep learning and ghost imaging, characterized in that: The following steps are involved: S1. Convert the plaintext image to be encrypted into a spectrum graph through two-dimensional Fourier transform; S2. Design and train an image steganography model, and use the image hiding network Hide-Net in the pre-trained ERIH-Net to hide the spectrum of a secret image into a non-secret image, generating a secret image containing the characteristic information of the spectrum of the plaintext image; S3, loading several Cake-Cutting Hadamard sequence matrices through DMD to modulate the light field and generate illumination speckles, and using the modulated illumination speckles to illuminate the secret image; collecting the total light intensity value of the secret image through the bucket detector, and obtaining a sequence with a corresponding number of light intensity values, i.e., ciphertext information; S4, using the second-order correlation reconstruction algorithm to reconstruct the secret image information through the ciphertext sequence and the Hadamard modulation mode (key); S5, using the decoded secret image as the input of the extraction network Extract-Net to extract the initial secret image spectrum information; S6. Decode the spectrum map into plaintext image information of spatial pixel distribution through inverse Fourier transform.

2. The image multiple hiding encryption method based on deep learning and ghost imaging as claimed in claim 1, characterized in that: In the image frequency domain conversion process in step S1, the input image is first converted into a grayscale image; secondly, a two-dimensional fast Fourier transform (FFT) is performed on the grayscale image to obtain a complex matrix; finally, a shift operation is performed on the complex matrix so that the zero-frequency component is located at the center of the matrix to obtain an image pixel distribution spectrum diagram.

3. The image multiple hiding encryption method based on deep learning and ghost imaging as claimed in claim 1, characterized in that: Step S1 The medium-resolution image steganography model ERIH-Net performs feature extraction and downsampling through multiple convolutional layers to obtain a high-dimensional feature map; based on the high-dimensional feature map, upsampling and feature reconstruction are performed through a deconvolution layer to obtain an imaging result with the same size as the original image. Secondly, the resolution of the input plaintext image and the non-secret image is required to be the same, and the size is 128×128×1 in the experiment.

4. The image multiple hiding encryption method based on deep learning and ghost imaging as claimed in claim 1, characterized in that: In step S3, the optical ghost imaging system includes a He-Ne laser with a wavelength of 532.8 nm, a beam expander Expander arranged on one side of the laser, a beam collimating lens L, a digital micromirror device DMD, a focusing lens L arranged on one side of the DMD, a bucket detector BD without spatial resolution, and a computer PC. Among them, CC Hadamard is a method with better image reconstruction quality among several mainstream Hadamard matrices.

5. The image multiple hiding encryption method based on deep learning and ghost imaging as claimed in claim 1, characterized in that: In step S4, the ciphertext is received through a public channel, the key is received through a secure channel, and the image reconstruction algorithm uses a second-order correlation algorithm. Compared with the second-order correlation algorithm that improves image sparsity through discrete wavelet transform (DWT), it can reconstruct a secret image with good visual quality at a lower sampling rate (70%). As the sampling rate increases, the quality of the reconstructed image improves.

6. The image multiple hiding encryption method based on deep learning and ghost imaging as claimed in claim 1, characterized in that: In step S5, Extract-Net consists of 6 different convolutional layers. The first 5 convolutional layers are connected to the Relu activation function, and the last convolutional layer is followed by the tanh activation function. Compared with the traditional sigmoid function, the tanh activation function converges faster, thereby reducing the number of iterations in the training process. Selecting MSE as the loss function during the training process can improve the invisibility of secret information in encrypted images and the visual effect of the extracted plaintext information.

7. The image multiple hiding encryption method based on deep learning and ghost imaging as claimed in claim 1, characterized in that: The inverse Fourier transform process in step S6 is applied to the phase angle information in the secret image spectrum, otherwise the original secret image information cannot be reconstructed. In addition, the transmission capacity can be increased through methods such as spatial multiplexing. The square area of ​​the secret image spectrum that is 64 pixels away from the center can be set to 0. This can increase the capacity of the transmitted information while ensuring the image quality. If the boundary of the 0-pixel area is less than 64, the human eye will not be able to distinguish the image information.

Images