An efficient encrypted image retrieval method based on neural network
By extracting image features through convolutional neural networks and principal component analysis, and constructing an index tree using ISODATA clustering, the problem of insufficient retrieval accuracy and efficiency in existing encrypted image retrieval schemes is solved, achieving efficient and accurate encrypted image retrieval.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAN AVIATION COMPUTING TECH RES INST OF AVIATION IND CORP OF CHINA
- Filing Date
- 2022-12-28
- Publication Date
- 2026-05-19
AI Technical Summary
Existing encrypted image retrieval schemes, while protecting image privacy, lack sufficient retrieval accuracy and efficiency. Traditional manual feature methods are cumbersome and incomplete, while common indexing technologies have high computational complexity or large storage requirements, making it difficult to achieve efficient and accurate encrypted image retrieval.
Convolutional neural networks are used to extract image feature vectors, combined with principal component analysis for dimensionality reduction, and an index tree is constructed using ISODATA clustering. Encryption is performed using system keys and symmetric keys. Airborne query users generate secure trapdoors for querying, and ground servers calculate secure Euclidean distances for retrieval.
It improves the accuracy and efficiency of encrypted image retrieval, enabling efficient and accurate image retrieval while protecting privacy.
Smart Images
Figure CN116244453B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network security technology, specifically to a highly efficient encrypted image retrieval method based on neural networks. Background Technology
[0002] With the widespread adoption of digital imaging equipment and the rapid development of internet technology, image transmission has become a crucial method of information dissemination. The security of image transmission between ground support and airborne users is particularly noteworthy. In recent years, to prevent the decryption of sensitive information in air-to-ground transmitted images, images can be encrypted before being sent to a server, thus preventing attackers from obtaining the plaintext information. However, encryption hides all image features, rendering traditional image retrieval mechanisms ineffective. Therefore, how to efficiently and accurately retrieve images of interest from a set of encrypted images while protecting privacy during air-to-ground image transmission is a critical issue in air-to-ground image transmission.
[0003] Currently, most encrypted image retrieval schemes use artificial features such as SIFT and Fisher's features. However, the process of manually constructing features is cumbersome and can only extract features from one aspect of the image, which is not comprehensive or accurate. Compared with artificial features, features learned by neural networks are more comprehensive and perform better in image retrieval tasks. Encrypted query techniques in encrypted image retrieval schemes mainly include searchable encryption, homomorphic encryption, and secure k-nearest neighbor query. Searchable encryption has good applications in keyword-based encrypted text retrieval, but it is difficult to apply to encrypted image retrieval. Homomorphic encryption can guarantee retrieval accuracy, but it introduces a lot of computational and communication overhead. Secure k-nearest neighbor query can achieve better performance in encrypted image retrieval. Indexing techniques in encrypted image retrieval mainly include inverted indexes, hash indexes, and tree indexes. Inverted indexes are suitable for bag-of-visual-words, but the large amount of clustering calculation will affect search efficiency. Hash indexes mainly use locality-sensitive hashing, which has extremely high computational complexity and requires pre-storing the original features of the database images, which will occupy a lot of storage space. The retrieval efficiency of common tree indexes KD-trees and R-trees decreases significantly with the increase of data dimensionality. Therefore, there is still room for improvement in the retrieval accuracy and efficiency of existing encrypted image retrieval schemes. Summary of the Invention
[0004] In view of this, the embodiments of this application provide a highly efficient encrypted image retrieval method based on neural networks, which has the characteristics of high retrieval accuracy, high retrieval efficiency and strong security.
[0005] This application provides the following technical solution: a highly efficient encrypted image retrieval method based on neural networks, comprising:
[0006] Step 1: A trusted third party generates a system key and a symmetric key, forming a key pair that is distributed to the ground image owner and the airborne query user;
[0007] Step 2: The ground image owner uses a convolutional neural network to extract the feature vector of each image, and then uses principal component analysis to reduce the dimensionality of the extracted high-dimensional feature vectors. The resulting feature vectors of a specified dimension are used as the global feature representation of the image.
[0008] Step 3: The ground image owner divides the image feature vectors into multiple clusters and builds an index tree from top to bottom;
[0009] Step 4: The ground image owner encrypts the original image dataset using a symmetric key and sends the encrypted image set to the ground server;
[0010] Step 5: The ground image owner encrypts the index tree using the system key, associates the encrypted index with the leaf nodes to generate a secure index tree, and sends the secure index tree and the encrypted image set to the ground server.
[0011] Step 6: The airborne query user obtains the feature vector of the image to be queried, encrypts the feature vector using the system key to generate a security trapdoor, and sends the security trapdoor to the ground server for querying.
[0012] Step 7: After receiving a valid request from the airborne query user, the ground server retrieves the security index tree, calculates the secure Euclidean distance between the security trapdoor and the security index, obtains the top k security indexes with the highest similarity, and returns the k encrypted images corresponding to these k security indexes as the search results to the airborne query user.
[0013] Step 8: After receiving the encrypted result, the onboard query user decrypts the encrypted result using the symmetric key to obtain the final similar image.
[0014] According to one embodiment, step 1, in which the trusted third party generates the system key and the symmetric key, specifically includes:
[0015] A trusted third party randomly selects a 2×2m bit invertible matrix M to generate a key SK = {M, M...} -1}, where M -1 It is the inverse matrix of M, SK = {M, M -1} is the system key used for encrypting the vector; a trusted third party also randomly selects a symmetric key K for encrypting the image, forming the key pair {SK,K}.
[0016] According to one embodiment, step 2, the process of obtaining a feature vector of a specified dimension, specifically includes:
[0017] The ground image owner uses a convolutional neural network to extract features from each image in the image database. The one-dimensional vector output by the linear layer of the network is used as the image feature vector. Then, principal component analysis is used to reduce the dimensionality of the high-dimensional vector. An orthogonal algorithm is used on the principal components of the data to remove mutually influential elements and find new, uncorrelated elements to represent the original data, thus obtaining a feature vector of a specified dimension.
[0018] According to one embodiment, in step 3, the ground image owner uses the dynamic clustering algorithm ISODATA to divide the image feature vector into multiple clusters.
[0019] According to one embodiment, step 3, the process of constructing the index tree, specifically includes:
[0020] The ground image owner divides the image feature vector into multiple clusters, uses the center point of the cluster as the child node of the root node of the index tree, and uses the data points belonging to the cluster as the leaf nodes under the child node, thus constructing the index tree from top to bottom.
[0021] According to one embodiment, in step 6, the onboard query user uses a convolutional neural network and principal component analysis to obtain the feature vector of the image to be queried.
[0022] According to one embodiment, in step 6, the security trapdoor does not reveal information about the queried image, but can be used to search for similar images on the security index tree.
[0023] According to one embodiment, in step 7, the closer the secure Euclidean distance between the secure trapdoor and the secure index, the higher the similarity.
[0024] This invention provides a highly efficient encrypted image retrieval method based on neural networks. It employs convolutional neural networks and principal component analysis to extract features from image sets, thereby improving retrieval accuracy, and constructs an index tree based on ISODATA clustering to enhance retrieval efficiency. Compared to other retrieval methods, this highly efficient encrypted image retrieval method based on neural networks offers both high retrieval accuracy and high retrieval efficiency. Attached Figure Description
[0025] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 This is a flowchart of an efficient encrypted image retrieval method based on neural networks according to an embodiment of the present invention;
[0027] Figure 2This is an architecture diagram of a high-efficiency encrypted image retrieval system based on neural networks, according to an embodiment of the present invention. Detailed Implementation
[0028] The embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0029] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments, providing a clear and complete description of the technical solutions of the present invention. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0030] like Figures 1-2 As shown, this embodiment of the invention provides an efficient encrypted image retrieval method based on neural networks, including:
[0031] Step 1: A trusted third party generates a system key and a symmetric key, forming a key pair that is distributed to the ground image owner and the airborne query user;
[0032] Specifically, a trusted third party randomly selects a 2×2m-bit invertible matrix M to generate the key SK = {M, M...} -1}, where M -1 It is the inverse matrix of M, SK = {M, M -1 {SK,K} is the system key used to encrypt the vector. The trusted third party also randomly selects a symmetric key K for encrypting the image. The trusted third party then sends the key pair {SK,K} to the ground data owner and the airborne query user.
[0033] Step 2: The ground image owner uses a convolutional neural network to extract the feature vector of each image, and then uses principal component analysis to reduce the dimensionality of the extracted high-dimensional feature vectors. The resulting feature vectors of a specified dimension are used as the global feature representation of the image.
[0034] Specifically, the ground image owner uses a convolutional neural network to extract features from each image in the image database. The one-dimensional vector output by the linear layer of this network is used as the image feature vector. Then, Principal Component Analysis (PCA) is used to reduce the dimensionality of the high-dimensional vector. An orthogonal algorithm is applied to the principal components of the data to remove mutually influential elements and find new, uncorrelated elements to represent the original data. These new elements are linear combinations of the original data elements, minimizing the data dimensionality while maximizing the variance after mapping, thus retaining sufficient original data information to obtain a feature vector p of a specified dimension. i={p i1,i2 ,…, im}
[0035] Step 3: The ground image owner divides the image feature vectors into multiple clusters and builds an index tree from top to bottom;
[0036] Specifically, the ground image owner uses the ISODATA clustering algorithm to divide the image dataset features into K clusters, and the clustering results represent ( 1,2 ,…, K The index tree is constructed from top to bottom by taking the center point generated by the cluster as the child node of the root node and the data points belonging to the cluster as the leaf nodes under the child node.
[0037] Step 4: The ground image owner encrypts the original image dataset using a symmetric key and sends the encrypted image set to the ground server;
[0038] Specifically, the ground image owner uses a symmetric encryption key K to encrypt the image dataset M = {1,2,...,} n}, thus obtaining the encrypted image set C = {c1,2,..., n The outsourcing is done to ground servers.
[0039] Step 5: The ground image owner encrypts the index tree using the system key, associates the encrypted index with the leaf nodes to generate a secure index tree, and sends the secure index tree and the encrypted image set to the ground server.
[0040] Specifically, in order to protect the index of each node, the ground image owner encrypts it using the system key SK, and associates the encrypted index with the leaf nodes to generate a secure index tree I. The encrypted image set C and the secure index tree I are then outsourced to the ground server.
[0041] Step 6: The airborne query user obtains the feature vector of the image to be queried, encrypts the feature vector using the system key to generate a security trapdoor, and sends the security trapdoor to the ground server for querying.
[0042] Specifically, the onboard query user also uses convolutional neural networks and principal component analysis to extract the feature vector q of the image of interest. { q1,2,…, m The system key SK is used to encrypt the query feature vector to generate a secure trapdoor T, which is then sent to the server. The trapdoor does not reveal information about the query image, but it can be used to search for similar images on the secure index tree I.
[0043] Step 7: After receiving a valid request from the airborne query user, the ground server retrieves the security index tree, calculates the secure Euclidean distance between the security trapdoor T and the security index I, obtains the top k security indices with the highest similarity, and returns the k encrypted images corresponding to these k security indices as the search results to the airborne query user.
[0044] Step 8: After receiving the encrypted result, the onboard query user decrypts the encrypted result using the symmetric key to obtain the final similar image.
[0045] This invention provides an efficient encrypted image retrieval method based on neural networks. It uses convolutional neural networks and principal component analysis to extract features from the image set to improve retrieval accuracy, and constructs an index tree based on ISODATA clustering to improve retrieval efficiency.
[0046] refer to Figure 1 In one specific embodiment, the execution flow of each step of the efficient encrypted image retrieval method based on neural networks in this embodiment is as follows:
[0047] 1. Key set generation
[0048] The specific steps are described in detail below:
[0049] Step 101: The trusted third party first generates a random (+2)×(+2) invertible matrix M. Using M... -1 Let SK denote the inverse matrix of M. SK = {M, M...} -1} is the system key, used to encrypt image feature vectors.
[0050] Step 102: A trusted third party randomly selects a symmetric encryption key K to encrypt the image.
[0051] Step 103: Send the keys {SK,K} to the ground data owner and the airborne retrieval user respectively.
[0052] 2 Image Feature Extraction
[0053] The specific steps are described in detail below:
[0054] Step 201: The image owner uses a VGG19 convolutional neural network to extract image features. This network consists of 16 convolutional layers and 3 fully connected layers, using consecutive 3x3 convolutional kernels instead of large kernels, and setting all pooling sizes to 2x2. This increases network depth while reducing the number of parameters. The one-dimensional vector output from the linear layers of this network is used as the image feature vector to extract features from each image in the image dataset.
[0055] Step 202: Dimensionality reduction is performed using Principal Component Analysis (PCA). An orthogonal algorithm is applied to the principal components of the data to remove mutually influencing elements and find new, uncorrelated elements to represent the original data. These new elements are linear combinations of the original data elements, minimizing the data dimensionality while maximizing the variance after mapping, thus retaining sufficient original data information to obtain m for each image. i eigenvectors p of a specified dimension i ={p i1 , i2 ,…, im}
[0056] 3. Index Tree Generation
[0057] Step 301: The ground image owner uses the ISODATA clustering algorithm to classify the image features, scans all sample points, and if the number of points within the radius of a sample point is greater than or equal to the minimum value, the sample point is set as the core point to form a temporary cluster. Then, connected temporary clusters are merged to form a cluster until all temporary clusters are processed, forming K clusters.
[0058] Step 302: After dividing the image feature vector into K clusters, construct a hierarchical index tree from top to bottom. Take the center point generated by the cluster as the child node of the root node of the index tree, and take the data point belonging to the cluster as the leaf node under the child node. Each leaf node corresponds to an image. During the query phase, perform the retrieval operation on the index tree to find the image related to the query image, thereby narrowing the retrieval scope and avoiding traversing the entire image feature library during the search phase.
[0059] 4. Image set encryption
[0060] The specific steps are described in detail below:
[0061] Step 401: The ground image owner uses the symmetric encryption key K from the security key set to encrypt the image dataset M, generating an encrypted image set C.
[0062] Step 402: The ground image owner outsources the encrypted image set C to the server.
[0063] 5. Index Tree Encryption
[0064] Step 501: The ground image owner will assign the feature vector p i Expand into Where, γ, α 1,2 ,…, m-1 ∈Z p It is a random number chosen by the image owner for each vector.
[0065] Step 502: The ground image owner encrypts each feature vector using a key, the encryption formula being as follows: in, It is for each An integer error vector randomly selected from a discrete Gaussian.
[0066] Step 503: The ground image owner associates the encrypted index with the leaf nodes to generate a secure index tree. The image owner then combines the encrypted image set with the secure index tree. They were all outsourced to cloud servers.
[0067] 6. Safety Trapdoor Generation
[0068] The specific steps are described in detail below:
[0069] Step 601: The onboard query user uses a convolutional neural network and principal component analysis to extract the feature vector q = {q} of the image of interest. 1,2 ,…, m}
[0070] Step 602: The onboard query user expands the feature vector q of the query image using the following formula: Where γ is a positive integer, It is a randomly selected number.
[0071] Step 603: The onboard query user uses a key to encrypt the feature vector of the query image. The encryption formula is as follows: in, and They are and The transpose of , For each The generated random integer error vector.
[0072] Step 604: The onboard query user sends a security trapdoor to the ground server. The trapdoor cannot reveal information about the queried image, but it can be used to search for the image on the security index tree.
[0073] 7 Server Search
[0074] Step 701: After receiving the secure index and secure trapdoor, the ground server first calculates the secure Euclidean distance between the secure trapdoor and the child nodes of the secure index tree, and then calculates and compares the secure Euclidean distance between the secure trapdoor and the leaf nodes in the cluster. Given two encrypted image features... and an encrypted query vector and The Euclidean distance can be calculated using the following formula, the correctness of which can be proven by the properties of matrix multiplication. According to Comp... ia and Compib The ground server can easily obtain the Euclidean distance comparison result from the formula. Because γ is a positive number, if the output is positive, it means the image... With query image Closer distance indicates higher similarity; otherwise, it indicates that the image is similar. With query image The similarity is higher.
[0075] The formula for calculating safe Euclidean distance is as follows:
[0076]
[0077] Step 702: The ground server calculates and sorts the secure Euclidean distance between the trapdoor and the secure index tree, and returns the encrypted query results to the airborne query user. This results are the encrypted image results corresponding to the top k most similar indexes, i.e., the top k most similar ciphertext images. The encrypted query results are then returned to the image query user.
[0078] 8 Image Decryption
[0079] Step 801: The onboard query user receives the encrypted query result R sent by the cloud server. q ′ .
[0080] Step 802: Decrypt it using a symmetric key to obtain the similar image result R. q .
[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A highly efficient encrypted image retrieval method based on neural networks, characterized in that, include: Step 1: A trusted third party generates a system key and a symmetric key, forming a key pair that is distributed to the ground image owner and the airborne query user; Step 2: The ground image owner uses a convolutional neural network to extract the feature vector of each image, and then uses principal component analysis to reduce the dimensionality of the extracted high-dimensional feature vectors. The resulting feature vectors of a specified dimension are used as the global feature representation of the image. Step 3: The ground image owner divides the image feature vectors into multiple clusters and builds an index tree from top to bottom; Step 4: The ground image owner encrypts the original image dataset using a symmetric key and sends the encrypted image set to the ground server; Step 5: The ground image owner encrypts the index tree using the system key, associates the encrypted index with the leaf nodes to generate a secure index tree, and sends the secure index tree and the encrypted image set to the ground server. Step 6: The airborne query user obtains the feature vector of the image to be queried, encrypts the feature vector using the system key to generate a security trapdoor, and sends the security trapdoor to the ground server for querying. Step 7: After receiving a valid request from the airborne query user, the ground server retrieves the security index tree, calculates the secure Euclidean distance between the security trapdoor and the security index, obtains the top k security indexes with the highest similarity, and returns the k encrypted images corresponding to these k security indexes as the search results to the airborne query user. Step 8: After receiving the encrypted result, the onboard query user decrypts the encrypted result using the symmetric key to obtain the final similar image; In step 1, the process of a trusted third party generating the system key and symmetric key specifically includes: Trusted third party randomly selected Invertible matrix of bits Generate key ,in yes The inverse matrix, The system key is used for encrypting the vector; a trusted third party also randomly selects a symmetric key. Used to encrypt images, forming the key pair ; Step 2, the process of obtaining the feature vector of the specified dimension specifically includes: The ground image owner uses a convolutional neural network to extract features from each image in the image database. The one-dimensional vector output by the linear layer of the network is used as the image feature vector. Then, principal component analysis is used to reduce the dimensionality of the high-dimensional vector. An orthogonal algorithm is used on the principal components of the data to remove mutually influential elements and find new, uncorrelated elements to represent the original data, thus obtaining a feature vector of a specified dimension.
2. The efficient encrypted image retrieval method based on neural networks according to claim 1, characterized in that, In step 3, the ground image owner uses the dynamic clustering algorithm ISODATA to divide the image feature vectors into multiple clusters.
3. The efficient encrypted image retrieval method based on neural networks according to claim 1, characterized in that, Step 3, the process of building the index tree, specifically includes: The ground image owner divides the image feature vector into multiple clusters, uses the center point of the cluster as the child node of the root node of the index tree, and uses the data points belonging to the cluster as the leaf nodes under the child node, thus constructing the index tree from top to bottom.
4. The efficient encrypted image retrieval method based on neural networks according to claim 1, characterized in that, In step 6, the onboard query user uses a convolutional neural network and principal component analysis to obtain the feature vector of the image to be queried.
5. The efficient encrypted image retrieval method based on neural networks according to claim 1, characterized in that, In step 6, the security trapdoor does not reveal information about the queried image, but can be used to search for similar images on the security index tree.
6. The efficient encrypted image retrieval method based on neural networks according to claim 1, characterized in that, In step 7, the closer the secure Euclidean distance between the secure trapdoor and the secure index, the higher the similarity.