Retrievable encrypted image reconstruction method based on cyclic consistency generative adversarial network

Abstract

The invention discloses a searchable encrypted image reconstruction method based on a cyclic consistency generative adversarial network. The method comprises the following steps: extracting entity image features from an entity data set; on the basis of entity image features, feature importance is measured through GN layer weight vectors, and two different subsets are obtained; based on two different subsets, the two subsets are encrypted through encryption modes with different security strengths, and a target domain is obtained; based on the target domain, realizing encryption and decryption operations of the features through a cyclic consistency generative adversarial network, and obtaining corresponding encryption features; and calculating the Euclidean distance between the features, sorting and outputting a retrieval result, then obtaining corresponding decryption features according to the retrieval result, and obtaining a reconstructed image through a decoder composed of deconvolution layers. Compared with other searchable encrypted image reconstruction methods, the method provided by the invention has the advantages that the performance is obviously improved, the ciphertext has relatively high security and good retrieval performance, and the whole method has better interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and natural language generation, and in particular to a retrievable encrypted image reconstruction method based on a cyclic consistency generative adversarial network. Background Art

[0002] With the development of cloud computing and intelligent technologies, forms such as images and videos have become important carrier media for information dissemination, and retrievable encrypted image reconstruction has become an important research topic. Retrievable encrypted image reconstruction aims to encrypt the original image, retrieve similar images in the database by inputting the encrypted image, and reconstruct the retrieved image into the original image. The encrypted image will hide the information of the original image. How to securely encrypt the image and retain some features of the original image unchanged for retrieval has become one of the main challenges in retrievable encrypted image reconstruction. Retrievable encrypted image reconstruction is a special encryption method that allows users to perform search operations on encrypted data, promotes the innovation of encryption algorithms, provides a new solution for the secure storage and privacy protection of image data, and has important theoretical research value; moreover, retrievable encrypted image reconstruction has broad application prospects. For example, the retrieval of medical privacy images, the rapid retrieval of targets by intelligent monitoring systems, the retrieval of encrypted data in e-commerce, digital rights management, etc. can all apply the retrievable encrypted image reconstruction technology.

[0003] The current mainstream retrievable encrypted image reconstruction methods only support retrieval based on keywords or simple features, and it is difficult to extract effective image features to achieve good retrieval accuracy. The reason is that these image features are extracted manually through traditional methods, and the extracted image features cannot represent the essential features of the image, resulting in low retrieval accuracy. In addition, some retrievable encryption methods focus on good retrieval accuracy of images while ignoring the security of ciphertexts, resulting in attackers being able to infer some information of the original image from these ciphertext images, thus destroying the privacy of the image. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and propose a retrievable encrypted image reconstruction method based on a cyclic consistency generative adversarial network, which can achieve reliable retrievable encrypted image reconstruction.

[0005] To achieve the above purpose, the technical solution provided by the present invention is as follows: A retrievable encrypted image reconstruction method based on a cyclic consistency generative adversarial network, including the following steps:

[0006] 1) Extract entity image features from the entity dataset;

[0007] 2) Based on the entity image features, measure the importance of each channel of the entity image features through the weight vector in the Group Normalization (GN) layer, and select the features conducive to image retrieval from the entity image features according to the importance to form an important feature subset, and the remaining features form an unimportant feature subset, obtaining two different subsets;

[0008] 3) Based on the two different subsets, encrypt these two subsets respectively through encryption methods with different security intensities to obtain two corresponding encrypted subsets, and then splice these two encrypted subsets into a whole to obtain the target domain;

[0009] 4) Based on the target domain, realize the encryption and decryption of features through the Cycle-Consistent Generative Adversarial Network (Cycle-GAN) to obtain the corresponding encrypted features. In order to ensure that the mean change of the important feature subset after encryption for retrieval is as small as possible, the important feature subset is divided into blocks, and the mean square error loss that measures the mean difference between the block features before and after encryption of the important feature subset is additionally added to the loss function for training Cycle-GAN, thereby constraining the change of the mean;

[0010] 5) Calculate the Euclidean distance between the encrypted features, sort according to the distance, select the top r features as the final retrieval results, then use Cycle-GAN to decrypt according to the retrieval results to obtain the corresponding decrypted features, and then use the decrypted features to reconstruct the original image.

[0011] Furthermore, in step 1), a convolutional neural network is used to extract entity image features. The specific process is as follows: Given an entity dataset, each image in the entity dataset is scaled to a size of 256×256, and then the entity image features are extracted through a convolutional neural network to obtain a total of N E samples, and v Ei is the i-th entity image feature.

[0012] Furthermore, in step 2), measure the importance of each channel of the entity image features through the weight vector in the Group Normalization (GN) layer, and select the features conducive to image retrieval from the entity image features according to the importance size to form an important feature subset a total of N S features, and v Sj is the j-th feature, and the remaining features form an unimportant feature subset a total of N T features, and v Tk is the k-th feature. Specifically, V S and V T are obtained in the following way:

[0013] For the image feature v Ei ∈R B×C×H×W, where R represents real numbers, B represents batches, and C represents the number of channels of v Ei , and H and W are the height and width of v Ei respectively; First, normalize the image feature v Ei , and the calculation method is as follows:

[0014]

[0015] In the formula, v iout is the normalized feature, μ and σ are the mean and standard deviation of v Ei respectively, ε is a small positive number added for division stability, and γ and β are trainable parameters in the group normalization GN layer; The parameter γ ∈ R C can measure the pixel variance of each channel of v Ei , reflect the richness of spatial information, and have different degrees of influence on retrieval, that is, reflect the different importance of each channel; Normalize γ to obtain the normalized weight vector W γ ∈ R C , and the calculation method is as follows:

[0016]

[0017] In the formula, W γ is a weight vector, w a represents the a-th element of W γ , γ a represents the a-th element of γ, and γ a represents the b-th element of γ. Select a suitable threshold TH and apply it to W γ , to obtain the corresponding mask M, and the calculation is as follows:

[0018]

[0019] In the formula, m n represents the n-th element of M, and w n represents the n-th element of W γ . After obtaining the mask M, it can be applied to the feature V E to obtain the important feature subset V S and the non-important feature subset V T .

[0020] Further, in step 3), first encrypt the important feature subset V S through the encrypted TPE by retaining the thumbnail. Specifically: First, divide each feature v S in V Sj into multiple feature blocks of the same size, and then randomly swap the positions of the features within the block to implement the feature permutation operation. In addition, generate three pseudo-random sequences U P, U Q and U Z , U P and U Q are used to represent the positions of the features to be changed, and U Z represents the change value of the feature. Using these three pseudo-random sequences, the encryption operation of replacing the feature values of each feature v S in the important feature subset V Sj is performed, and the mean value within each feature block is ensured to remain unchanged. The calculation method is as follows:

[0021]

[0022] In the formula, c represents the c-th feature block of the feature v Sj , idx represents the index position; traversing the three pseudo-random sequences in turn, the TPE encryption is performed on each feature block of the feature v Sj , and then the TPE encryption is performed on the entire important feature subset V S to obtain the corresponding encrypted subset; secondly, the chaotic sequence generated by the 3D hyperchaotic system is used to encrypt the non-important feature subset V T to enhance its security. The chaotic sequence has randomness, and its calculation method is as follows:

[0023]

[0024] In the formula, a1, b1, c1 are hyperparameters, u, v, z are the three variables of the chaotic system, represents the result of iteration. Set the initial values of u, v, z, and then iterate 500 + I t times to obtain three sequences, and then truncate the first 500 numbers of the sequences to obtain three chaotic sequences Z t with a length of I M , Z P , Z Q ; Use the sequence Z M to achieve the permutation of the feature positions of each feature v T in the non-important feature subset V Tk , and use Z P and Z Q to complete the operation of replacing the feature values of each feature v Tk . Each element of the sequence Z P represents the position of the feature to be changed, and each element of the sequence Z Q represents the change value of the feature to be changed; traversing the above three chaotic sequences in turn, the corresponding permutation and value replacement operations are performed on the feature v Tk , and the non-important feature subset V TEach feature performs a similar operation to obtain a corresponding encrypted subset; the above two different encrypted subsets are spliced together in the original order to obtain the target domain V ET 。

[0025] Further, in step 4), by inputting the target domain V ET and the entity image feature V E into the Cycle-GAN (Cycle Generative Adversarial Network) together, the encryption and decryption operations of the features are realized. Specifically: First, the first generator G J of the Cycle-GAN encrypts the entity image feature V E to obtain the corresponding encrypted feature, realizing the encryption operation of the feature. Then, the second generator G L decrypts the encrypted feature, realizing the decryption operation of the feature; in order to ensure that the mean change of the important feature subset after encryption for retrieval is as small as possible, the important feature subset is divided into blocks, and the mean square error loss L bmse measuring the mean difference between the block features before and after encryption of the important feature subset is additionally added to the loss function for training the Cycle-GAN, bmse so as to constrain the change of the mean. The definition of L

[0026]

[0027] is as follows: b In the formula, N represents the number of feature blocks, Avg(.) represents the function of calculating the mean, ET represents the kr-th block of V represents the kr-th block of V E 。

[0028] Further, in step 5), all the encrypted features are input into a neural network composed of multiple fully connected layers to output a low-dimensional vector. Calculate the distance between the low-dimensional vector corresponding to a certain encrypted feature and the low-dimensional vectors corresponding to other encrypted features, and sort them in ascending order of distance. Select the first r corresponding encrypted features as the final retrieval result. According to the obtained retrieval result, use the decryption network of the Cycle-GAN to decrypt to obtain the corresponding decrypted feature, and then the decrypted feature is input into a decoder composed of transposed convolutional layers to reconstruct the corresponding image.

[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0030] 1. The present invention proposes a basic framework for a retrievable encrypted image reconstruction method. First, a general entity dataset is introduced, and entity image features are learned based on the entity dataset. The Group Normalization (GN) layer is used to obtain two subsets by selecting features with high information richness from the entity image features using the weight vector that can measure the importance of feature channels. Then, the two subsets are encrypted through two different encryption methods to obtain the target domain. Finally, the encryption and decryption operations of the features are realized based on the Cycle-Consistent Generative Adversarial Network (Cycle-GAN).

[0031] 2. The present invention proposes to use the weight vector of the GN layer to measure the importance of features. By introducing an entity common dataset and training the GN layer, a weight vector representing each channel of the feature is obtained. This weight vector reflects the spatial information richness contained in different feature channels and has different importance for retrieval. In this way, an important feature subset and an unimportant feature subset are obtained from the image features, which solves the problem of difficult separation of feature mixing in the existing methods.

[0032] 3. The present invention proposes a two-level security encryption method based on the Cycle-Consistent Generative Adversarial Network (Cycle-GAN), aiming to ensure the security and good retrieval accuracy of the encrypted ciphertext features after feature encryption. For the important feature subset, thumbnail retention encryption (TPE) is used to obtain the corresponding encrypted subset, ensuring that the mean within the block remains unchanged to enable retrieval of the ciphertext. For the unimportant feature subset, a three-dimensional hyperchaotic system is used for encryption, and the pseudo-random sequence generated by the chaotic system is used to confuse and permute the features to obtain the corresponding encrypted subset to ensure the security of the ciphertext. These two encrypted subsets are concatenated into a whole as the target domain of Cycle-GAN, and the network is trained to realize the conversion from the source feature domain to the target domain, thus realizing a retrievable encryption method with two-level security.

[0033] 4. Compared with other retrievable image encryption methods, the method of the present invention has significantly improved performance. The encrypted ciphertext can have high security and good retrieval performance, and the overall method has better interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is the logic flow chart of the method of the present invention.

[0035] Figure 2 is the logic flow chart of feature encryption proposed by the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] The present invention will be further described in detail below in conjunction with the embodiments and the drawings, but the embodiments of the present invention are not limited thereto.

[0037] Such as Figure 1 and Figure 2As shown in the figure, this embodiment discloses a retrievable encrypted image reconstruction method based on a cycle-consistent generative adversarial network, including the following steps:

[0038] 1) Introduce a general entity dataset and extract features from entity images. Use a convolutional neural network to extract entity image features. The specific process is as follows: Given an entity dataset, scale each image in the entity dataset to a size of 256×256, and then extract entity image features through a convolutional neural network to obtain a total of N E samples, where v Ei is the i-th entity image feature.

[0039] 2) Based on the entity image features extracted in the previous step, measure the importance of each channel of the entity image features through the weight vector in the group normalization GN layer, and select the features conducive to image retrieval from the entity image features according to the importance size to form an important feature subset a total of N S features, where v Sj is the j-th feature, and the remaining features form an unimportant feature subset a total of N T features, where v Tk is the k-th feature. Specifically, obtain V S and V T as follows:

[0040] For the image feature v Ei ∈R B×C×H×W , where R represents real numbers, B is the batch, C is the number of channels of v Ei , and H and W are the height and width of v Ei respectively. First, standardize the image feature v Ei , and the calculation method is as follows:

[0041]

[0042] In the formula, v iout is the standardized feature, μ and σ are the mean and standard deviation of v Ei respectively, ε is a small positive number added for division stability, and γ and β are trainable parameters in the group normalization GN layer. The parameter γ∈R C can measure the pixel variance of each channel of v Ei , which can reflect the richness of spatial information and has different degrees of influence on retrieval, that is, it reflects the different importance of each channel. Normalize γ to obtain the normalized weight vector W γ ∈R C , and the calculation method is as follows:

[0043]

[0044] W γ is a weight vector, w a represents the a-th element of W γ γ, where γ a represents the a-th element of γ, γ a represents the b-th element of γ. Select a suitable threshold TH and apply it to W γ , to obtain the corresponding mask M, which is calculated as follows:

[0045]

[0046] In the formula, m n represents the n-th element of M, and w n represents the n-th element of W γ . After obtaining the mask M, we can apply it to the feature V E to obtain the important feature subset V S and the unimportant feature subset V T .

[0047] 3) Based on the two feature subsets obtained in the previous step, encrypt the important feature subset V S by retaining the encrypted TPE through thumbnails. Specifically, first divide each feature v S in V Sj into multiple feature blocks of the same size, and then randomly swap the positions of the features within the blocks to achieve the feature permutation operation. In addition, generate three pseudo-random sequences U P , U Q and U Z , U P and U Q are used to represent the positions of the features to be changed, and U Z represents the change value of the feature. Use these three pseudo-random sequences to implement the encryption operation of replacing the feature values of each feature v S in the important feature subset V Sj and ensure that the mean value within each feature block remains unchanged. The calculation method is as follows:

[0048]

[0049] c represents the c-th feature block of the feature v Sj , and idx represents the index position. Traverse these three pseudo-random sequences in turn, perform TPE encryption on each feature block of the feature v Sj , and then perform TPE encryption on the entire important feature subset V S to obtain the corresponding encrypted subset. Secondly, use the chaotic sequence generated by the 3D hyperchaotic system to encrypt the unimportant feature subset V TEncrypt it to enhance its security. The chaotic sequence has randomness, and its calculation method is as follows:

[0050]

[0051] Here, a1, b1, c1 are hyperparameters, and u, v, z are the three variables of this chaotic system. represents the result of iteration. Set the initial values of u, v, z, and then iterate 500 + I t times to obtain three sequences. Then truncate the first 500 numbers of the sequences to obtain three chaotic sequences Z t with a length of I M , Z P , Z Q . Use the sequence Z M to implement the permutation of the feature positions of each feature v T in the unimportant feature subset V Tk . Use Z P and Z Q to complete the eigenvalue replacement operation for each feature v Tk . Each element of the sequence Z P represents the position of the feature to be changed, and each element of the sequence Z Q represents the change value of the feature to be changed. Traverse these three chaotic sequences in turn to perform the corresponding permutation and value replacement operations on the feature v Tk . Perform similar operations on each feature in the unimportant feature subset V T to obtain the corresponding encrypted subset. Concatenate the above two different encrypted subsets together in the original order to obtain the target domain V ET .

[0052] 4) Based on the target domain obtained in the previous step, by inputting the target domain V ET and the entity image feature V E into the Cycle-Consistent Generative Adversarial Network Cycle-GAN together, implement the encryption and decryption operations of the features. Specifically, first encrypt the entity image feature V J through the first generator G E of the Cycle-Consistent Generative Adversarial Network Cycle-GAN to obtain the corresponding encrypted feature, thus implementing the encryption operation of the feature. Decrypt the encrypted feature through the second generator G L to implement the decryption operation of the feature. To ensure that the mean change of the important feature subset after encryption for retrieval is as small as possible, divide the important feature subset into blocks, and additionally add the mean squared error loss L bmse that measures the mean difference between the block features before and after encryption of the important feature subset to the loss function for training Cycle-GAN, thereby constraining the change of the mean. Lbmse is defined as follows:

[0053]

[0054] where N b represents the number of feature blocks, Avg(.) represents the function of calculating the mean, represents the kr-th block of V ET and represents the kr-th block of V E too.

[0055] 5) By calculating the Euclidean distance between the encrypted features, sorting them according to the distance, selecting the top r features as the final retrieval results, and using the decryption network of Cycle-GAN to decrypt the corresponding decrypted features according to the retrieval results, and then reconstructing the original image using the decrypted features. Specifically, all the encrypted features are input into a neural network composed of multiple fully connected layers to output a low-dimensional vector, calculate the distance between the low-dimensional vector corresponding to a certain encrypted feature and the low-dimensional vectors corresponding to other encrypted features, sort them from small to large according to the distance, and select the top r corresponding encrypted features as the final retrieval results. According to the obtained retrieval results, use the decryption network of Cycle-GAN to decrypt the corresponding decrypted features, and then input the decrypted features into a decoder composed of transposed convolution layers to reconstruct the corresponding image.

[0056] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and shall be included in the protection scope of the present invention.

Claims

1. A retrievable encryption image reconstruction method based on a cycle-consistent generative adversarial network, characterized in that It includes the following steps: 1) Extract entity image features from the entity dataset; 2) Based on the entity image features, measure the importance of each channel of the entity image features through the weight vector in the group normalization GN layer, and select the features conducive to image retrieval from the entity image features according to the importance to form an important feature subset, and the remaining features form an unimportant feature subset, obtaining two different subsets; 3) Based on the two different subsets, encrypt the two subsets respectively through encryption methods with different security intensities to obtain two corresponding encrypted subsets, and then splice the two encrypted subsets into a whole to obtain the target domain; 4) Based on the target domain, realize the encryption and decryption of features through the cycle-consistent generative adversarial network Cycle-GAN to obtain the corresponding encrypted features. In order to ensure that the mean change of the important feature subset for retrieval after encryption is as small as possible, the important feature subset is divided into blocks, and the mean square error loss that measures the mean difference between the block features before and after encryption of the important feature subset is additionally added to the loss function for training Cycle-GAN, so as to constrain the change of the mean; 5) Calculate the Euclidean distance between the encrypted features, sort according to the distance, select the top r features as the final retrieval result, then use Cycle-GAN to decrypt according to the retrieval result to obtain the corresponding decrypted features, and then use the decrypted features to reconstruct the original image.

2. The retrievable encryption image reconstruction method based on a cycle-consistent generative adversarial network according to claim 1, wherein In step 1), a convolutional neural network is used to extract entity image features. The specific process is as follows: Given an entity dataset, each image in the entity dataset is scaled to a size of 256×256, and then the entity image features are extracted through the convolutional neural network to obtain a total of N E samples, where v Ei is the i-th entity image feature.

3. The retrievable encrypted image reconstruction method based on a cycle-consistent generative adversarial network according to claim 2, wherein In step 2), the importance of each channel of the entity image features is measured by the weight vector in the group normalization GN layer, and features conducive to image retrieval are selected from the entity image features according to the importance size to form an important feature subset. There are a total of N S features, where v Sj is the j-th feature, and the remaining features form a non-important feature subset. There are a total of N T features, where v Tk is the k-th feature. Specifically, V S and V T are obtained in the following way: For the image feature v Ei ∈R B×C×H×W , where R represents real numbers, B is the batch, C is the number of channels of v Ei , H and W are the height and width of v Ei respectively; first, standardize the image feature v Ei , and the calculation method is as follows: where v iout is the standardized feature, μ and σ are the mean and standard deviation of v Ei respectively, ε is a small positive number added for division stability, and γ and β are trainable parameters in the group normalization GN layer; the parameter γ ∈ R C can measure the pixel variance of each channel of v Ei , reflecting the richness of spatial information and having different degrees of influence on retrieval, that is, reflecting the different importance of each channel; normalizing γ to obtain the normalized weight vector W γ ∈ R C , and the calculation method is as follows: Where, W γ is a weight vector, w a represents the a-th element of W γ and γ a represents the a-th element of γ, and γ a represents the b-th element of γ. Select a suitable threshold TH and apply it to W γ to obtain the corresponding mask M, which is calculated as follows: where m n represents the n-th element of M, and w n represents the n-th element of W γ . After obtaining the mask M, it can be applied to the feature V E to obtain the important feature subset V S and the non-important feature subset V T .

4. The retrievable encryption image reconstruction method based on a cycle-consistent generative adversarial network according to claim 3, wherein In step 3), we first use the thumbnail-preserving encrypted TPE to encrypt the important feature subset V S To encrypt, first: S Each feature v in Sj It is divided into multiple feature blocks of the same size, and then the features in the block are randomly exchanged to achieve feature permutation operation. In addition, three pseudo-random sequences U are generated by a pseudo-random number generator. P , U Q and U Z , U P and U Q Used to indicate the position of the feature to be changed, U Z Represents the change value of the feature, and uses these three pseudo-random sequences to implement the important feature subset V S Each feature v Sj The encryption operation of feature value replacement and ensuring that the mean value in each feature block remains unchanged is calculated as follows: Wherein, c represents the c-th feature block of the feature v Sj and idx represents the index position; after traversing the three pseudo-random sequences in turn, perform TPE encryption on each feature block of the feature v Sj and then perform TPE encryption on the entire important feature subset V S to obtain the corresponding encrypted subset; secondly, encrypt the unimportant feature subset V T using the chaotic sequence generated by the 3D hyperchaotic system to enhance its security. The chaotic sequence has randomness, and its calculation method is as follows: Wherein, a1, b1, and c1 are hyperparameters, and u, v, and z are three variables of the chaotic system. Denote the result of iteration. Set the initial values of u, v, and z, and then iterate 500 + I t times to obtain three sequences. Then truncate the first 500 numbers of the sequences to obtain three chaotic sequences Z t with length I. M 、Z P 、Z Q ; Use the sequence Z M to implement the permutation of the feature positions of each feature v T in the unimportant feature subset V Tk , and use Z P and Z Q to complete the eigenvalue replacement operation for each feature v Tk . Each element of the sequence Z P represents the position of the feature to be changed, and each element of the sequence Z Q represents the change value of the feature to be changed. Traverse the above three chaotic sequences in turn, perform the corresponding permutation and value replacement operations on the feature v Tk , and perform similar operations on each feature in the unimportant feature subset V T to obtain the corresponding encrypted subset; Concatenate the above two different encrypted subsets together in the original order to obtain the target domain V ET .

5. The retrievable encryption image reconstruction method based on the cyclic consistency generative adversarial network according to claim 4, wherein In step 4), by taking the target domain V ET and the entity image feature V E and jointly inputting them into the Cycle-Consistent Generative Adversarial Network (Cycle-GAN), the encryption and decryption operations of the features are realized. Specifically: First, through the first generator G J of the Cycle-Consistent Generative Adversarial Network (Cycle-GAN), the entity image feature V E is encrypted to obtain the corresponding encrypted feature, realizing the encryption operation of the feature. Then, through the second generator G L the encrypted feature is decrypted, realizing the decryption operation of the feature; In order to ensure that the mean change of the important feature subset after encryption for retrieval is as small as possible, the important feature subset is divided into blocks, and the mean square error loss L bmse measuring the mean difference between the block features before and after encryption of the important feature subset is additionally added to the loss function for training Cycle-GAN, bmse so as to constrain the change of the mean. The definition of L bmse is as follows: where N b represents the number of feature blocks, and Avg(.) represents the function of calculating the mean value, represents the kr-th block of V ET , and represents the kr-th block of V E .

6. The retrievable encrypted image reconstruction method based on a cyclic consistency generative adversarial network according to claim 5, wherein In step 5), input all the encrypted features into a neural network composed of multiple fully connected layers, output a low-dimensional vector, calculate the distance between the low-dimensional vector corresponding to a certain encrypted feature and the low-dimensional vectors corresponding to other encrypted features, and sort them in ascending order of distance, select the top r corresponding encrypted features as the final retrieval result, according to the obtained retrieval result, use the decryption network of Cycle-GAN to decrypt to obtain the corresponding decrypted features, and then input the decrypted features into a decoder composed of transposed convolutional layers to reconstruct the corresponding image.

Images