Privacy protection image retrieval method and device based on multiple keys, electronic equipment and storage medium

Abstract

This invention provides a privacy-preserving image retrieval method, apparatus, electronic device, and storage medium based on multi-key technology. The invention utilizes a trusted center to assign independent keys to image owners and query users, and distributes key conversion parameters to a cloud server. It extracts image feature vectors using a ResNet50 model, constructs a hierarchical sparse graph index, and configures local statistics for graph nodes. Image features, image data, and node summaries are encrypted, and a verifiable structure supporting dynamic updates is constructed using chameleon hashing. During retrieval, the query user generates a query trapdoor, and the cloud server performs a graph index search within a unified comparison domain after key conversion, first obtaining candidate results. This invention avoids the security risks associated with shared keys through a multi-key mechanism, improves retrieval efficiency and accuracy through graph indexing and two-stage hybrid distance retrieval, and ensures the correctness and consistency of retrieval results and index status through a dynamic verifiable structure.

Description

Technical Field

[0001] This invention relates to the fields of cloud computing and data security, and image processing technology, and in particular to a privacy-preserving image retrieval method, apparatus, electronic device, and storage medium based on multiple keys. Background Technology

[0002] With the rapid development of cloud computing, the Internet of Things, and multimedia technologies, a large amount of image data is outsourced to cloud servers for storage and management. Content-based image retrieval technology can quickly find target images similar to the query image from large-scale image data, and has wide application value in scenarios such as medical diagnosis, security monitoring, social networks, and smart terminals. However, image data often contains sensitive privacy information such as facial attributes, identity information, or scene content. If uploaded directly to cloud servers in plaintext, it can easily lead to the leakage of user privacy. Although traditional encryption methods can protect the security of image content, they make it difficult for the cloud to directly perform similar image retrieval. Therefore, how to achieve usable encrypted image retrieval while protecting image privacy has become an important research direction in related fields.

[0003] Existing privacy-preserving image retrieval schemes still have many shortcomings in practical applications. First, many existing schemes are mainly designed for single-user or shared-key scenarios. When multiple query users use the same query key, if one user's key is leaked, it will affect the retrieval security of other legitimate users. If separate retrieval indexes are built for different users, it will significantly increase the computational and storage overhead for image owners. Therefore, unified retrieval under multi-user non-shared-key conditions remains an urgent problem to be solved. Second, existing schemes often struggle to balance retrieval efficiency and security. One type of scheme, while having high retrieval efficiency, suffers from problems such as similarity leakage and insufficient resistance to attacks. Another type of scheme, while having strong security, is difficult to apply to large-scale image data scenarios due to complex encrypted calculations and inefficient index structures. Third, cloud servers are often considered honest but may return incomplete or tampered retrieval results. Existing result verification methods mostly rely on structures such as Merkle hash trees or accumulators. However, when image data is dynamically added, deleted, or updated, there are often problems such as high update costs, long verification paths, or inefficiency in supporting dynamic scenarios, making it difficult to meet the requirements of practical applications for the correctness, consistency, and dynamic verifiability of retrieval results.

[0004] Therefore, in real-world scenarios involving multiple users, massive image outsourcing, and dynamic updates, there is an urgent need for an image retrieval method that can support multiple keys, efficient retrieval, privacy protection, and dynamic verifiability, in order to improve the security, practicality, and deployability of encrypted image retrieval systems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a privacy-preserving image retrieval method, apparatus, electronic device, and storage medium based on multiple keys.

[0006] In a first aspect, the present invention provides a privacy-preserving image retrieval method based on multiple keys, comprising:

[0007] The Trust Center assigns index encryption matrices and query key matrices to image owners and query users respectively, and distributes key conversion parameters to the cloud server;

[0008] Image owners use the ResNet50 model to extract image features from image datasets, obtaining a set of image feature vectors, and then construct a hierarchical sparse graph index based on the set of image feature vectors.

[0009] Each image in the image dataset is encrypted using the AES algorithm to obtain an encrypted image set;

[0010] Configure local diagonal covariance statistics for graph nodes in the hierarchical sparse graph index, and encrypt the encrypted images of each graph node in the hierarchical sparse graph index in sequence to obtain encrypted node representations; generate summary messages based on encrypted node representations and local diagonal covariance statistics; and process the summary messages using chameleon hashing to construct a verifiable structure that supports dynamic updates.

[0011] The querying user extracts image features from the image data to be retrieved, obtains the feature vector of the image to be retrieved, and encrypts the feature vector of the image to be retrieved using the querying user's query key matrix, thus obtaining a searchable trapdoor;

[0012] The cloud server maps the searchable query trapdoors to a unified comparison domain according to the key conversion parameters, performs candidate search in the hierarchical sparse graph index, and reorders them using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence.

[0013] The query allows users to verify a set of similar encrypted images and their corresponding sets of verification evidence. Once the verification is successful, the set of similar encrypted images is decrypted to obtain the similar images.

[0014] Preferably, the step of constructing a hierarchical sparse graph index based on the image feature vector set includes:

[0015] Each image feature vector is mapped to a corresponding graph node, and Euclidean distance is used to measure the similarity between image feature vectors;

[0016] Based on the similarity between image feature vectors, the graph nodes are organized into different levels, and representative graph nodes are selected as entry nodes to form a hierarchical graph structure.

[0017] In the same layer of the hierarchical graph structure, local nearest neighbor edges are established for each graph node, and cross-layer jump edges are established for each graph node between adjacent layers to form a hierarchical sparse graph connection structure.

[0018] Between adjacent layers, there are graph nodes. Establish cross-level jump edges pointing to nearby nodes in the next level; specifically: for the... Graph nodes in the layer Select the closest one within the same layer. Each graph node forms a local nearest neighbor set. ,Right now:

[0019] ;

[0020] In the formula, Indicates the first The number of local nearest neighbors of a node in a layer graph. This indicates selecting the first one with the smallest distance. Each graph node; its jump target is represented as:

[0021] ;

[0022] In the formula, Represents graph nodes The target node for the jump in the adjacent next level;

[0023] The hierarchical sparse graph index is obtained based on the local nearest neighbor edge set at the same level and the cross-level jump edge set. ,Right now:

[0024] ;

[0025] In the formula, Represents the set of local nearest neighbor edges at the same level. This represents the set of cross-level jump edges.

[0026] Preferably, configuring local diagonal covariance statistics for graph nodes in the hierarchical sparse graph index includes:

[0027] For the first in the hierarchical sparse graph index Let there be a number of graph nodes, and let the set of image feature vectors associated with their neighborhood be denoted as . ;

[0028] Based on the local statistical distribution of image feature vectors in each dimension within the neighborhood, a local diagonal covariance statistic is configured for the corresponding graph node. ,Right now:

[0029] ;

[0030] ;

[0031] In the formula, Indicates the first Local diagonal covariance statistics of each graph node; The dimension of the image feature vector; Indicates the first The image feature vector within the neighborhood of each graph node is at the th... Local variance in the dimension; Indicates the first The image feature vector within the neighborhood of each graph node is at the th... Local mean on the dimension.

[0032] Encrypted images of each graph node in a hierarchical sparse graph index The encryption process is performed sequentially to obtain the encrypted node representation; including:

[0033] Use the index encryption matrix Encrypted images Encryption is performed to obtain the encrypted node representation. ,Right now:

[0034] ;

[0035] In the formula, For the first The encrypted node representation corresponding to each graph node; Indicates the first The scaling factor of the feature vector of each graph node; Indicates the first The noise vector of each graph node.

[0036] Preferably, after completing the node feature encryption, according to the first... The image identifier, encrypted node digest, local diagonal covariance digest, adjacency relation digest, version identifier, and associated ciphertext image digest corresponding to each graph node are used to generate the [number]th [node]. Summary message of each graph node ,Right now:

[0037]

[0038] In the formula, For summary functions; Indicates the first The image identifier corresponding to each graph node; Representation of encrypted nodes Abstract; Represents the local diagonal covariance statistic Abstract; Indicates the first The version identifier corresponding to each graph node; Represents a summary of adjacency relationships; Indicates the first Adjacency information of each graph node; Represents a digest of the associated encrypted image; For the first The ciphertext image associated with each graph node.

[0039] Preferably, chameleon hashing is used to process the digest message to construct a verifiable structure that supports dynamic updates, including:

[0040] Using Chameleon Hash to the first Summary message of each graph node Processing is performed to obtain the graph node verification values. ;

[0041] Verification values ​​for each graph node Associating layers to construct verifiable nodes that support dynamic updates, including:

[0042] For the first By aggregating the verification values ​​of graph nodes in the layer, we can obtain the first... Layer verification value ;

[0043] Then, perform inter-layer correlation on the validation values ​​of each layer to obtain the global validation value corresponding to the entire hierarchical sparse graph index. .

[0044] Preferably, the querying user extracts image features from the image data to be retrieved, obtaining a feature vector of the image to be retrieved. This feature vector is then encrypted using the querying user's query key matrix to obtain a searchable trapdoor; including:

[0045] The ResNet50 model is used to extract image features from the image data to be retrieved, resulting in the feature vector of the image to be retrieved.

[0046] Feature vector of the image to be retrieved Perform normalization processing;

[0047] For the Each query user uses the query key matrix corresponding to that user. The normalized feature vector of the image to be retrieved is encrypted to generate a query trapdoor;

[0048] Query trap Send to the cloud server.

[0049] Preferably, the cloud server uses the key conversion parameters to determine the searchable query trapdoors. Mapped to a unified comparison domain, candidate search is performed in the hierarchical sparse graph index, and reordering is performed using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence; including:

[0050] The cloud server received the first A query trap sent by a query user Then, according to the key conversion parameters Query trap Mapping to the unified comparison domain yields a transformation trapdoor. ;

[0051] After completing the unified comparison domain mapping, the cloud server starts from the entry node of the hierarchical sparse graph index. Starting from this point, perform a candidate search layer by layer along local nearest neighbor edges within the same layer and jump edges across layers, and select and transform trapdoors at each layer. The nearest graph node serves as the entry point for the next search layer, continuing until the lowest layer yields the candidate graph node set. ;

[0052] For candidate graph node set Candidate images in the dataset, combined with candidate image feature vectors The local diagonal covariance statistic of its corresponding graph node Candidate images are calculated using adaptive hybrid distance. Image to be retrieved Second-stage reordering scores ;

[0053] The second-stage reordering score corresponding to each candidate image. Sort by size from smallest to largest and select the top... The results are used as the final search results to obtain a set of similar encrypted images. ;

[0054] The cloud server generates a set of similar encrypted images based on a verifiable structure that supports dynamic updates. Corresponding set of verification evidence .

[0055] Secondly, the present invention provides a privacy-preserving image retrieval system based on multiple keys, comprising:

[0056] The key distribution module uses a trusted center to distribute index encryption matrices and query key matrices to image owners and query users respectively, and distributes key conversion parameters to the cloud server.

[0057] The hierarchical sparse graph index construction module uses the ResNet50 model to extract image features from the image dataset, obtains a set of image feature vectors, and constructs a hierarchical sparse graph index based on the set of image feature vectors.

[0058] The image encryption module is used to encrypt each image in the image dataset using the AES algorithm, resulting in an encrypted image set.

[0059] A verifiable structure building module is used to configure local diagonal covariance statistics for graph nodes in the hierarchical sparse graph index, and to encrypt the encrypted images of each graph node in the hierarchical sparse graph index in sequence to obtain encrypted node representations; to generate summary messages based on the encrypted node representations and local diagonal covariance statistics; and to process the summary messages using chameleon hashing to build a verifiable structure that supports dynamic updates.

[0060] The trapdoor generation module is used to extract image features from the image data to be retrieved, obtain the feature vector of the image to be retrieved, and encrypt the feature vector of the image to be retrieved using the query key matrix of the querying user to obtain a searchable trapdoor.

[0061] The image retrieval module is used to map searchable query trapdoors to a unified comparison domain according to the key conversion parameters, perform candidate search in the hierarchical sparse graph index, and reorder them using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence.

[0062] The verification and decryption module is used to verify the set of similar encrypted images and the corresponding set of verification evidence. After verification, the set of similar encrypted images is decrypted to obtain the similar images.

[0063] Thirdly, the present invention provides an electronic device including at least one processor;

[0064] and a memory communicatively connected to the at least one processor;

[0065] The memory stores a computer program that can be executed by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the multi-key-based privacy-preserving image retrieval method.

[0066] The beneficial effects of this invention are as follows:

[0067] 1. This invention employs a multi-key allocation and key conversion method to assign independent query keys to different query users. By using key conversion parameters, the query trapdoors generated by different users are mapped to a unified comparison domain to complete the retrieval. This avoids the problem of shared keys affecting the retrieval security of other legitimate users due to the leakage of a single user's key. At the same time, it avoids the additional computation and storage overhead caused by image owners repeatedly building retrieval indexes for different users.

[0068] 2. This invention utilizes a pre-trained convolutional neural network model, hierarchical sparse graph indexing, and adaptive hybrid distance reordering to perform efficient and accurate encrypted retrieval of image data, enabling in-depth extraction of high-level semantic features of images and making image content representation more accurate.

[0069] 3. By constructing a hierarchical graph connection structure, the present invention can gradually narrow the search range along the graph structure from the entry node during retrieval, significantly reducing the overhead of full comparison under massive image data; furthermore, by combining the local diagonal covariance statistic to reorder the candidate results, it can better distinguish fine-grained similar images.

[0070] 4. This invention adopts a dynamic verifiable structure based on chameleon hashing to uniformly associate graph node summaries, adjacency relationships, and version status, and provides corresponding verification evidence while returning search results. This enables query users to determine whether the results returned by the cloud server have been tampered with, whether they are consistent with the current graph index structure, and whether they correspond to the latest updated version. When image data is added, deleted, or modified, only the relevant verification information of the affected nodes needs to be updated, without rebuilding the entire verification structure, thereby reducing dynamic maintenance costs. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the method of Embodiment 1 of the present invention;

[0072] Figure 2 This is a schematic diagram of the process framework of the method in Embodiment 1 of the present invention;

[0073] Figure 3 This is a schematic diagram illustrating the generation process of the hierarchical sparse graph index in Embodiment 1 of the present invention;

[0074] Figure 4 This is a structural framework diagram of the system in Embodiment 2 of the present invention. Detailed Implementation

[0075] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0076] like Figure 1 and 2 As shown, this embodiment provides a privacy-preserving image retrieval method based on multiple keys, including:

[0077] S1) The Trusted Center allocates the index encryption matrix and query key matrix to the image owner and the query user, respectively, and distributes the key conversion parameters to the cloud server;

[0078] In this embodiment, the trusted center assigns an index encryption matrix to the image owner. , for the first Each query user is assigned an independent query key matrix corresponding to their identity. ; and based on the query key matrix Generate the first Key conversion parameters for each query user , For index encryption matrix The inverse matrix, and the key transformation parameters. Send to the cloud server.

[0079] S2) The image owner uses the ResNet50 model to extract image features from the image dataset, obtaining a set of image feature vectors. And based on the image feature vector set Constructing a hierarchical sparse graph index; the aforementioned image feature vector set Represented as: In the formula, Indicates the first Image feature vectors of the images; This indicates the number of images in the image dataset.

[0080] In this embodiment, the ResNet50 model effectively alleviates the degradation problem in deep network training through residual learning mechanism, and can extract high-level semantic features with strong discriminative ability, thereby improving the accuracy of similar image retrieval.

[0081] In this embodiment, each image feature vector Mapped to the corresponding graph nodes This yields the graph node set. And the Euclidean distance is used to measure the similarity between image feature vectors, that is:

[0082] ;

[0083] In the formula, Image feature vector and The distance between them; the smaller the distance, the more similar the two are. The dimension of the image feature vector; Indicates the first The image feature vector at the th ... Components in the dimension;

[0084] Based on the similarity between image feature vectors, the graph nodes are organized into different levels, and representative graph nodes are selected as entry nodes. This forms a hierarchical graph structure; Figure 3 The diagram illustrates the generation process of the hierarchical sparse graph index in this embodiment. The image feature vectors first complete node mapping and then perform hierarchical organization to provide a structural foundation for subsequent connection construction.

[0085] Within the same layer of the hierarchical graph structure, local nearest neighbor edges are established for each graph node, and cross-layer jump edges are established for each graph node between adjacent layers, thus forming a hierarchical sparse graph connection structure, as detailed below:

[0086] For the Graph nodes in the layer Select the closest one within the same layer. Each graph node forms a local nearest neighbor set. ,Right now:

[0087] ;

[0088] In the formula, Indicates the first The number of local nearest neighbors of a node in a layer graph. This indicates selecting the first one with the smallest distance. Each graph node;

[0089] Between adjacent layers, there are graph nodes. Establish a cross-level jump edge pointing to a nearby node in the next level. The jump target can be represented as:

[0090] ;

[0091] In the formula, Represents graph nodes The target node for the jump in the adjacent next layer.

[0092] The hierarchical sparse graph index is obtained based on the local nearest neighbor edge set within the same layer and the jump edge set across layers. ,Right now:

[0093] ;

[0094] In the formula, Represents the set of local nearest neighbor edges at the same level. This represents the set of cross-level jump edges.

[0095] By maintaining local connections between similar nodes through local nearest neighbor edges in the same layer, and then using cross-layer jump edges to achieve rapid transitions between different layers, the search range can be narrowed layer by layer from the entry node during retrieval, without having to perform a linear traversal of all feature vectors.

[0096] S3) Encrypt each image in the image dataset using the AES algorithm to obtain an encrypted image set;

[0097] In this embodiment, for the first Zhang Image Based on the image encryption key Encryption is performed to obtain an encrypted image, namely:

[0098] ;

[0099] In the formula, For the first Zhang encrypted image; This indicates encryption operations based on the AES pseudo-random generator of spatiotemporal chaos.

[0100] S4) Configure local diagonal covariance statistics for graph nodes in the hierarchical sparse graph index, and encrypt each graph node in the hierarchical sparse graph index to obtain an encrypted node representation; generate a summary message based on the encrypted node representation and the local diagonal covariance statistics; and process the summary message using chameleon hashing to construct a verifiable structure that supports dynamic updates.

[0101] In this embodiment, for the first in the hierarchical sparse graph index Let there be a number of graph nodes, and let the set of image feature vectors associated with their neighborhood be denoted as . ,Right now:

[0102] ;

[0103] in, Indicates the first The number of image feature vectors in the neighborhood of each graph node.

[0104] Based on the local statistical distribution of image feature vectors in each dimension within the neighborhood, a local diagonal covariance statistic is configured for the corresponding graph node. ,Right now:

[0105] ;

[0106] In the formula, Indicates the first Local diagonal covariance statistics of each graph node; The dimension of the image feature vector; Indicates the first The image feature vector within the neighborhood of each graph node is at the th... Local variance in the dimension;

[0107] ;

[0108] in, Indicates the first The image feature vector within the neighborhood of each graph node is at the th... Local mean on the dimension;

[0109] By introducing local diagonal covariance statistics at the graph node level, the subsequent re-ranking process no longer relies solely on global distance information, but can simultaneously reflect the dispersion of features in each dimension within the neighborhood of the current graph node, thereby enabling fine-grained differences between candidate results to be further distinguished.

[0110] In this embodiment, an index encryption matrix is ​​used. Encrypted images Encryption is performed to obtain the encrypted node representation. ,Right now:

[0111] ;

[0112] In the formula, For the first The encrypted node representation corresponding to each graph node; Indicates the first The scaling factor of the feature vector of each graph node; Indicates the first The noise vector of each graph node, and .

[0113] After completing the node feature encryption, according to the first... The image identifier, encrypted node digest, local diagonal covariance digest, adjacency relation digest, version identifier, and associated ciphertext image digest corresponding to each graph node are used to generate the [number]th [node]. Summary message of each graph node ,Right now:

[0114] ;

[0115] In the formula, For summary functions; Indicates the first The image identifier corresponding to each graph node; Representation of encrypted nodes Abstract; Represents the local diagonal covariance statistic Abstract; Indicates the first The version identifier corresponding to each graph node; Represents a summary of adjacency relationships; Indicates the first Adjacency information of each graph node; Represents a digest of the associated encrypted image; For the first The ciphertext image associated with each graph node.

[0116] By unifying node content, local statistics, graph connectivity, version status, and associated ciphertext images into the digest message. In the subsequent verification, it is possible to simultaneously check whether the content of the search results is correct, whether the graph index structure is consistent, and whether the node version is valid.

[0117] In this embodiment, the chameleon hash is used to... Summary message of each graph node Processing is performed to obtain the graph node verification values. ,Right now:

[0118] ;

[0119] in, Indicates the first Chameleon hash values ​​of each graph node; This represents the chameleon hash operation. Indicates the first Summary message of each graph node, This represents the random hash factor corresponding to the node in the graph.

[0120] For dynamically updated scenarios, when the first... The summary message for each graph node is provided by Change to At that time, random factors can be regenerated. , making This allows for controlled updates without altering existing validation value relationships. Consequently, when image data, graph connectivity, or node versions change, only the summary message and random factor corresponding to the affected nodes need adjustment, rather than requiring a complete reconstruction of the entire validation structure.

[0121] The verification values ​​of each graph node are associated layer by layer to construct a verifiable structure that supports dynamic updates, namely:

[0122] Let the first The set of graph node verification values ​​in the layer is ,Right now:

[0123] ;

[0124] in, Indicates the first The number of graph nodes in the layer;

[0125] For the first By aggregating the verification values ​​of graph nodes in the layer, we can obtain the first... Layer verification value ,Right now:

[0126] ;

[0127] In the formula, .

[0128] Then, perform inter-layer correlation on the validation values ​​of each layer to obtain the global validation value corresponding to the entire hierarchical sparse graph index. ,Right now:

[0129] ;

[0130] in, This indicates the highest level number in the hierarchical sparse graph index. This represents the global verification value for the entire verifiable structure.

[0131] When the cloud server returns a similar encrypted image, it also returns verification evidence for the corresponding graph node and a comparison with the global verification value. With the relevant verification path, the querying user can verify whether the graph node corresponding to the search result has been tampered with, and whether the search result is consistent with the current graph index status, based on the verification evidence.

[0132] S5) The query user extracts image features from the image data to be retrieved, obtains the feature vector of the image to be retrieved, and encrypts the feature vector of the image to be retrieved using the query user's query key matrix to obtain a searchable trapdoor; including the following steps:

[0133] S51) Using the ResNet50 model to extract image features from the image data to be retrieved, the feature vector of the image to be retrieved is obtained. ;in, The dimension of the image feature vector. Indicates the feature vector of the image to be retrieved at the th... Components in the dimension;

[0134] S52), Feature vector of the image to be retrieved Normalization is performed to reduce the impact of differences in feature scales between different images on the retrieval results, i.e.:

[0135] ;

[0136] In the formula, Represents the feature vector of the image to be retrieved The second norm; This is the normalized feature vector of the image to be retrieved;

[0137] After normalization, the feature vector of the image to be retrieved retains the original semantic information while the numerical range of each dimension is more stable, which facilitates candidate search and similarity reordering in the subsequent unified comparison domain.

[0138] S53), for the first Each query user uses the query key matrix corresponding to that user. The normalized feature vector of the image to be retrieved is encrypted to generate a query trapdoor. ,Right now:

[0139] ;

[0140] In the formula, Indicates the first A query trap generated by a user. Indicates the first Query key matrix for each query user The inverse matrix, Indicates the first The expansion factor corresponding to each query user Indicates the first The noise vector corresponding to each query user. ; Indicates the transpose operation;

[0141] In this embodiment, the expansion factor Used to adjust the magnitude of query features, noise vector This is used to perturb the query features, preventing the cloud server from directly recovering the original query features when performing a unified comparison domain retrieval later; simultaneously, different query users use their own independent query key matrices. A query trapdoor is generated, thus enabling the independent generation of query requests from multiple users without sharing the query key.

[0142] S54), query trapdoor Send to the cloud server.

[0143] S6) The cloud server will use the key conversion parameters to determine the searchable query trapdoors. Mapped to a unified comparison domain, candidate search is performed in the hierarchical sparse graph index, and reordering is performed using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence; specifically as follows:

[0144] S61), Cloud server receives the first A query trap sent by a query user Then, according to the key conversion parameters Query trap Mapping to the unified comparison domain yields a transformation trapdoor. ,Right now:

[0145] ;

[0146] In the formula, For the first Transition traps for individual query users in the unified comparison domain; Indicates the first The key conversion parameters corresponding to each query user; Indicates the first A query trap generated by a user.

[0147] S62) After completing the unified comparison domain mapping, the cloud server starts from the entry node of the hierarchical sparse graph index. Starting from this point, perform a candidate search layer by layer along local nearest neighbor edges within the same layer and jump edges across layers, and select and transform trapdoors at each layer. The nearest graph node serves as the entry point for the next search layer, continuing until the lowest layer yields the candidate graph node set. ,Right now:

[0148] ;

[0149] in, Indicates the first A set of candidate graph nodes corresponding to each query user. Indicates the number of candidate graph nodes;

[0150] S63) For the candidate graph node set Candidate images in the dataset, combined with candidate image feature vectors The local diagonal covariance statistic of its corresponding graph node The second-stage re-ranking score is calculated using adaptive hybrid distance. ;

[0151] Let the candidate image be The corresponding image feature vector is The local diagonal covariance statistic corresponding to the node in the graph is denoted as . Candidate images are calculated using adaptive hybrid distance. Image to be retrieved Second-stage reordering scores ,Right now:

[0152] ;

[0153] In the formula, Representing candidate images The second-stage re-ranking score between the image and the image to be retrieved. Indicates the first The adaptive weight coefficients corresponding to each graph node satisfy the following conditions: ; The Euclidean distance term, representing the feature vector of the candidate image and the feature vector of the image to be retrieved, is used to describe the global similarity between the candidate image and the image to be retrieved. Indicates the local diagonal covariance statistic The Mahalanobis distance term is used to describe the degree of difference between the candidate image and the image to be retrieved under the local statistical structure of the current graph node;

[0154] S64) The second-stage reordering scores corresponding to each candidate image. Sort by size from smallest to largest and select the top... The results are used as the final search results to obtain a set of similar encrypted images. ,Right now:

[0155] ;

[0156] Indicates returning to the first A set of similar encrypted images for each query user; Indicates the number of similar images returned; Indicates the first The returned encrypted image;

[0157] S65) The cloud server generates a set of similar encrypted images based on a verifiable structure that supports dynamic updates. Corresponding set of verification evidence ,Right now:

[0158] ;

[0159] In the formula, This represents the set of verification evidence corresponding to the returned result. Represents encrypted images Corresponding verification evidence.

[0160] S7) The query user verifies the set of similar encrypted images and the corresponding set of verification evidence. After verification, the similar encrypted images are decrypted to obtain the similar images, as follows:

[0161] S71) Query the set of similar encrypted images returned by the cloud server to the user. and the corresponding set of verification evidence Based on the aforementioned set of verification evidence, the correctness of the returned results, the consistency of the graph index structure, and the validity of the version are verified; specifically:

[0162] The query user is based on the returned verification path and graph node summary message. Graph node verification value and version identifier Each graph node corresponding to the similar encrypted image is verified to determine whether the similar encrypted image returned by the cloud server is consistent with the candidate search and reordering results, and to confirm whether the graph node and its adjacency relationship have been tampered with in the current index state; if the verification is successful, the returned result is accepted; if the verification fails, the returned result is rejected.

[0163] S72) After successful verification, query the user's collection of similar encrypted images. Decryption yields a set of similar images. .

[0164] To further verify the effectiveness of this embodiment, the accuracy of image retrieval will be explained. The Corel-10k image dataset is used as the image data set of the image owners, and then 1000 images are randomly selected from it as the set of images to be retrieved by the query user. The Corel-10k image dataset contains 10,000 images, with 100 image categories, and each category contains 100 similar images. As an indicator of accuracy, among them, This indicates the number of similar encrypted images in the ciphertext state returned by the cloud server. Indicates the number of similar images. The closer the value is to 1, the higher the accuracy of the search.

[0165] Table 1. Comparison of image retrieval accuracy based on multi-key mechanism under different numbers of returned images

[0166]

[0167] Table 1 compares the accuracy of image retrieval under multi-key conditions as the number of returned images increases from 10 to 50. Table 1 shows that the image retrieval accuracy is not significantly affected after introducing independent key generation, key conversion, and a unified comparison domain retrieval mechanism in this embodiment, indicating that the multi-key security mechanism and the similar image retrieval function are effectively decoupled. This means that while ensuring secure retrieval using independent keys by different query users, high retrieval accuracy can still be maintained, thus ensuring good practicality and deployability of this solution in real-world multi-user scenarios.

[0168] Table 2. Comparison of image retrieval accuracy under different image feature extraction methods

[0169]

[0170] Table 2 compares the accuracy of image retrieval using two image feature extraction methods as the number of returned images increases from 10 to 50. Here, ConvNeXt represents feature extraction from image data using the ConvNeXt convolutional neural network model, and CLD represents extracting the color layout descriptor of the image as the image feature vector. Table 2 shows that the overall retrieval accuracy obtained using the ConvNeXt image feature extraction method is higher than that obtained using the CLD image feature extraction method. This indicates that the convolutional neural network feature extraction method used in this invention can obtain higher-level semantic features with stronger discriminative power. Compared with traditional manual features such as color layout descriptors, it can more accurately represent the true content of the image, thereby improving the accuracy of similar image retrieval and better meeting the application needs of this invention in large-scale image scenarios.

[0171] Example 2

[0172] like Figure 4 As shown, this embodiment provides a privacy-preserving image retrieval system based on multiple keys, including:

[0173] The key distribution module 100 uses a trusted center to distribute the index encryption matrix and the query key matrix to the image owner and the query user respectively, and distributes the key conversion parameters to the cloud server.

[0174] The hierarchical sparse graph index construction module 200 uses the ResNet50 model to extract image features from the image dataset, obtains a set of image feature vectors, and constructs a hierarchical sparse graph index based on the set of image feature vectors.

[0175] The image encryption module 300 is used to encrypt each image in the image data set using the AES algorithm to obtain an encrypted image set;

[0176] The verifiable structure construction module 400 is used to configure local diagonal covariance statistics for graph nodes in the hierarchical sparse graph index, and to encrypt the encrypted images of each graph node in the hierarchical sparse graph index in sequence to obtain encrypted node representations; to generate summary messages based on the encrypted node representations and local diagonal covariance statistics; and to process the summary messages using chameleon hashing to construct a verifiable structure that supports dynamic updates.

[0177] The trapdoor generation module 500 is used to extract image features from the image data to be retrieved, obtain the feature vector of the image to be retrieved, and encrypt the feature vector of the image to be retrieved using the query key matrix of the querying user to obtain a searchable trapdoor.

[0178] The image retrieval module 600 is used to map searchable query trapdoors to a unified comparison domain according to the key conversion parameters, perform candidate search in the hierarchical sparse graph index, and reorder them using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence.

[0179] The verification and decryption module 700 is used to verify the set of similar encrypted images and the corresponding set of verification evidence. After verification, the set of similar encrypted images is decrypted to obtain the similar images.

[0180] In this embodiment, the key distribution module 100 utilizes a trusted center to distribute index encryption matrices and query key matrices to the image owner and query user respectively, and distributes key conversion parameters to the cloud server, including:

[0181] The trusted center assigns an index encryption matrix to the image owner. , for the first Each query user is assigned an independent query key matrix corresponding to their identity. ; and based on the query key matrix Generate the first Key conversion parameters for each query user and convert the key parameters Send to the cloud server, that is:

[0182] ;

[0183] in, For index encryption matrix The inverse matrix.

[0184] In this embodiment, the hierarchical sparse graph index construction module 200 constructs a hierarchical sparse graph index based on the image feature vector set, including:

[0185] Each image feature vector Mapped to the corresponding graph nodes This yields the graph node set. And the Euclidean distance is used to measure the similarity between image feature vectors, that is:

[0186] ;

[0187] In the formula, Image feature vector and The distance between them; the smaller the distance, the more similar the two are. The dimension of the image feature vector; Indicates the first The image feature vector at the th ... Components in the dimension;

[0188] Based on the similarity between image feature vectors, the graph nodes are organized into different levels, and representative graph nodes are selected as entry nodes. This forms a hierarchical graph structure.

[0189] Within the same layer of the hierarchical graph structure, local nearest neighbor edges are established for each graph node, and cross-layer jump edges are established for each graph node between adjacent layers, thus forming a hierarchical sparse graph connection structure, as detailed below:

[0190] For the Graph nodes in the layer Select the closest one within the same layer. Each graph node forms a local nearest neighbor set. ,Right now:

[0191] ;

[0192] In the formula, Indicates the first The number of local nearest neighbors of a node in a layer graph. This indicates selecting the first one with the smallest distance. Each graph node;

[0193] Between adjacent layers, there are graph nodes. Establish a cross-level jump edge pointing to a nearby node in the next level. The jump target can be represented as:

[0194] ;

[0195] In the formula, Represents graph nodes The target node for the jump in the adjacent next layer.

[0196] The hierarchical sparse graph index is obtained based on the local nearest neighbor edge set within the same layer and the jump edge set across layers. ,Right now:

[0197] ;

[0198] In the formula, Represents the set of local nearest neighbor edges at the same level. This represents the set of cross-level jump edges.

[0199] In this embodiment, the image encryption module 300 uses the AES algorithm to encrypt each image in the image data set, resulting in an encrypted image set, including:

[0200] For the Zhang Image Based on the image encryption key Encryption is performed to obtain an encrypted image, namely:

[0201] ;

[0202] In the formula, For the first Zhang encrypted image; This indicates an AES-based encryption operation.

[0203] In this embodiment, the verifiable structure building module 400 configures local diagonal covariance statistics for graph nodes in the hierarchical sparse graph index, including:

[0204] For the first in the hierarchical sparse graph index Let there be a number of graph nodes, and let the set of image feature vectors associated with their neighborhood be denoted as . ,Right now:

[0205] ;

[0206] in, Indicates the first The number of image feature vectors in the neighborhood of each graph node.

[0207] Based on the local statistical distribution of image feature vectors in each dimension within the neighborhood, a local diagonal covariance statistic is configured for the corresponding graph node. ,Right now:

[0208] ;

[0209] In the formula, Indicates the first Local diagonal covariance statistics of each graph node; The dimension of the image feature vector; Indicates the first The image feature vector within the neighborhood of each graph node is at the th... Local variance in the dimension;

[0210] ;

[0211] in, Indicates the first The image feature vector within the neighborhood of each graph node is at the th... Local mean on the dimension.

[0212] In this embodiment, the verifiable structure construction module 400 encrypts the images of each graph node in the hierarchical sparse graph index. The encryption process is performed sequentially to obtain the encrypted node representation, including:

[0213] Use the index encryption matrix Encrypted images Encryption is performed to obtain the encrypted node representation. ,Right now:

[0214] ;

[0215] In the formula, For the first The encrypted node representation corresponding to each graph node; Indicates the first The scaling factor of the feature vector of each graph node; Indicates the first The noise vector of each graph node, and .

[0216] In this embodiment, after completing the node feature encryption, the verifiable structure construction module 400, according to the first... The image identifier, encrypted node digest, local diagonal covariance digest, adjacency relation digest, version identifier, and associated ciphertext image digest corresponding to each graph node are used to generate the [number]th [node]. Summary message of each graph node ,Right now:

[0217]

[0218] In the formula, For summary functions; Indicates the first The image identifier corresponding to each graph node; Representation of encrypted nodes Abstract; Represents the local diagonal covariance statistic Abstract; Indicates the first The version identifier corresponding to each graph node; Represents a summary of adjacency relationships; Indicates the first Adjacency information of each graph node; Represents a digest of the associated encrypted image; For the first The ciphertext image associated with each graph node.

[0219] In this embodiment, the verifiable structure construction module 400 uses chameleon hashing to process the digest message to construct a verifiable structure that supports dynamic updates; including:

[0220] Using Chameleon Hash to the first Summary message of each graph node Processing is performed to obtain the graph node verification values. ,Right now:

[0221] ;

[0222] in, Indicates the first Chameleon hash value of each graph node; This represents the chameleon hash operation. Indicates the first Summary message of each graph node, This represents the random hash factor corresponding to the node in the graph.

[0223] The verification values ​​of each graph node are associated layer by layer to construct a verifiable structure that supports dynamic updates, namely:

[0224] Let the first The set of graph node verification values ​​in the layer is ,Right now:

[0225] ;

[0226] in, Indicates the first The number of graph nodes in the layer;

[0227] For the By aggregating the verification values ​​of graph nodes in the layer, we can obtain the first... Layer verification value ,Right now:

[0228] ;

[0229] In the formula, 。;

[0230] Then, perform inter-layer correlation on the validation values ​​of each layer to obtain the global validation value corresponding to the entire hierarchical sparse graph index. ,Right now:

[0231] ;

[0232] in, This indicates the highest level number in the hierarchical sparse graph index. This represents the global verification value for the entire verifiable structure.

[0233] When the cloud server returns a similar encrypted image, it also returns verification evidence for the corresponding graph node and a comparison with the global verification value. With the relevant verification path, the querying user can verify whether the graph node corresponding to the search result has been tampered with, and whether the search result is consistent with the current graph index status, based on the verification evidence.

[0234] In this embodiment, the trapdoor generation module 500 extracts image features from the image data to be retrieved, obtains a feature vector of the image to be retrieved, and encrypts the feature vector of the image to be retrieved using the query key matrix of the querying user, thereby obtaining a searchable trapdoor; including:

[0235] The ResNet50 model is used to extract image features from the image data to be retrieved, resulting in the feature vector of the image to be retrieved. ;in, The dimension of the image feature vector. Indicates the feature vector of the image to be retrieved at the th... Components on the dimension;

[0236] Feature vector of the image to be retrieved Normalization is performed to reduce the impact of differences in feature scales between different images on the retrieval results;

[0237] After normalization, the feature vector of the image to be retrieved retains the original semantic information while the numerical range of each dimension is more stable, which facilitates candidate search and similarity reordering in the subsequent unified comparison domain.

[0238] For the Each query user uses the query key matrix corresponding to that user. The normalized feature vector of the image to be retrieved is encrypted to generate a query trapdoor. ,Right now:

[0239] ;

[0240] In the formula, Indicates the first A query trap generated by a user. Indicates the first Query key matrix for each query user The inverse matrix, Indicates the first The expansion factor corresponding to each query user Indicates the first The noise vector corresponding to each query user. ; Indicates the transpose operation;

[0241] Query trap Send to the cloud server.

[0242] In this embodiment, the image retrieval module 600 maps searchable query trapdoors to a unified comparison domain according to the key conversion parameters, performs candidate search in the hierarchical sparse graph index, and re-ranks them using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence; including:

[0243] The cloud server received the first A query trap sent by a user. Then, according to the key conversion parameters Query trap Mapping to the unified comparison domain yields a transformation trapdoor. ,Right now:

[0244] ;

[0245] In the formula, For the first Transition traps for individual query users in the unified comparison domain; Indicates the first The key conversion parameters corresponding to each query user; Indicates the first A query trap generated by a user.

[0246] After completing the unified comparison domain mapping, the cloud server starts from the entry node of the hierarchical sparse graph index. Starting from this point, perform a candidate search layer by layer along local nearest neighbor edges within the same layer and jump edges across layers, and select and transform trapdoors at each layer. The nearest graph node serves as the entry point for the next search layer, continuing until the lowest layer yields the candidate graph node set. ,Right now:

[0247] ;

[0248] in, Indicates the first A set of candidate graph nodes corresponding to each query user. Indicates the number of candidate graph nodes;

[0249] Since the query trapdoors generated by different query users can be mapped to a unified comparison domain through their respective key conversion parameters, the cloud server does not need to know the original query key of the query user to complete the candidate search on the same set of hierarchical sparse graph indexes. At the same time, the hierarchical graph connection structure enables the search process to shrink from the entry node to the local similar region layer by layer, thereby avoiding the linear traversal of the full feature vector.

[0250] For candidate graph node set Candidate images in the dataset, combined with candidate image feature vectors The local diagonal covariance statistic of its corresponding graph node The second-stage re-ranking score is calculated using adaptive hybrid distance. ;

[0251] Let the candidate image be The corresponding image feature vector is The local diagonal covariance statistic corresponding to the node in the graph is denoted as . Adaptive hybrid distance calculation is used to calculate candidate images. Image to be retrieved Second-stage reordering scores ,Right now:

[0252] ;

[0253] In the formula, Representing candidate images The second-stage re-ranking score between the image and the image to be retrieved. Indicates the first The adaptive weight coefficients corresponding to each graph node satisfy the following conditions: ; The Euclidean distance term, representing the feature vector of the candidate image and the feature vector of the image to be retrieved, is used to describe the global similarity between the candidate image and the image to be retrieved. Represents the local diagonal covariance statistic The Mahalanobis distance term is used to describe the degree of difference between the candidate image and the image to be retrieved under the local statistical structure of the current graph node;

[0254] The second-stage reordering score corresponding to each candidate image. Sort by size from smallest to largest and select the top... The results are used as the final search results to obtain a set of similar encrypted images. ,Right now:

[0255] ;

[0256] Indicates returning to the first A set of similar encrypted images for each query user; Indicates the number of similar images returned; Indicates the first The returned encrypted image;

[0257] The cloud server generates a set of similar encrypted images based on a verifiable structure that supports dynamic updates. Corresponding set of verification evidence ,Right now:

[0258] ;

[0259] In the formula, This represents the set of verification evidence corresponding to the returned result. Represents encrypted images Corresponding verification evidence.

[0260] Example 3

[0261] This embodiment provides an electronic device, including at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to cause the electronic device to perform the privacy-preserving image retrieval method based on multiple keys as described in Embodiment 1.

[0262] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A privacy-preserving image retrieval method based on multiple keys, characterized in that, include: The Trust Center assigns index encryption matrices and query key matrices to image owners and query users respectively, and distributes key conversion parameters to the cloud server; Image owners use the ResNet50 model to extract image features from image datasets, obtaining a set of image feature vectors, and then construct a hierarchical sparse graph index based on the set of image feature vectors. Each image in the image dataset is encrypted using the AES algorithm to obtain an encrypted image set; Configure local diagonal covariance statistics for graph nodes in the hierarchical sparse graph index, and encrypt the encrypted images of each graph node in the hierarchical sparse graph index in sequence to obtain encrypted node representations; generate summary messages based on encrypted node representations and local diagonal covariance statistics; and process the summary messages using chameleon hashing to construct a verifiable structure that supports dynamic updates. The querying user extracts image features from the image data to be retrieved, obtains the feature vector of the image to be retrieved, and encrypts the feature vector of the image to be retrieved using the querying user's query key matrix, thus obtaining a searchable trapdoor; The cloud server maps the searchable query trapdoors to a unified comparison domain according to the key conversion parameters, performs candidate search in the hierarchical sparse graph index, and reorders them using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence. The query allows users to verify a set of similar encrypted images and their corresponding sets of verification evidence. Once the verification is successful, the set of similar encrypted images is decrypted to obtain the similar images.

2. The privacy-preserving image retrieval method based on multiple keys according to claim 1, characterized in that: The construction of a hierarchical sparse graph index based on the image feature vector set includes: Each image feature vector is mapped to a corresponding graph node, and Euclidean distance is used to measure the similarity between image feature vectors; Based on the similarity between image feature vectors, the graph nodes are organized into different levels, and representative graph nodes are selected as entry nodes to form a hierarchical graph structure. In the same layer of the hierarchical graph structure, local nearest neighbor edges are established for each graph node, and cross-layer jump edges are established for each graph node between adjacent layers to form a hierarchical sparse graph connection structure. Between adjacent layers, there are graph nodes. Establish cross-level jump edges pointing to nearby nodes in the next level; specifically: for the... Graph nodes in the layer Select the closest one within the same layer. Each graph node forms a local nearest neighbor set. ,Right now: ； In the formula, Indicates the first The number of local nearest neighbors of a node in a layer graph. This indicates selecting the first one with the smallest distance. Each graph node; its jump target is represented as: ； In the formula, Represents graph nodes The target node for the jump in the adjacent next level; The hierarchical sparse graph index is obtained based on the local nearest neighbor edge set at the same level and the cross-level jump edge set. ,Right now: ； In the formula, Represents the set of local nearest neighbor edges on the same level. This represents the set of cross-level jump edges.

3. The privacy-preserving image retrieval method based on multiple keys according to claim 1, characterized in that: Each image in the image dataset is encrypted using the AES algorithm, resulting in an encrypted image set, including: For the Zhang Image Based on the image encryption key Encryption is performed to obtain an encrypted image, namely: ； In the formula, For the first Zhang encrypted image; This indicates encryption operations based on the AES pseudo-random generator of spatiotemporal chaos.

4. The privacy-preserving image retrieval method based on multiple keys according to claim 1, characterized in that: Encrypted images of each graph node in a hierarchical sparse graph index The encryption process is performed sequentially to obtain the encrypted node representation; including: Use the index encryption matrix Encrypted images Encryption is performed to obtain the encrypted node representation. ,Right now: ； In the formula, For the first The encrypted node representation corresponding to each graph node; Indicates the first The scaling factor of the feature vector of each graph node; Indicates the first The noise vector of each graph node.

5. The privacy-preserving image retrieval method based on multiple keys according to claim 4, characterized in that: After completing the node feature encryption, according to the first... The image identifier, encrypted node digest, local diagonal covariance digest, adjacency relation digest, version identifier, and associated ciphertext image digest corresponding to each graph node are used to generate the [number]th [node]. Summary message of each graph node ,Right now: In the formula, For summary functions; Indicates the first The image identifier corresponding to each graph node; Representation of encrypted nodes Abstract; Represents the local diagonal covariance statistic Abstract; Indicates the first The version identifier corresponding to each graph node; Represents a summary of adjacency relationships; Indicates the first Adjacency information of each graph node; Represents a digest of the associated encrypted image; For the first The ciphertext image associated with each graph node.

6. The privacy-preserving image retrieval method based on multiple keys according to claim 5, characterized in that: Chameleon hashing is used to process digest messages to build a verifiable structure that supports dynamic updates, including: Using Chameleon Hash to the first Summary message of each graph node Processing is performed to obtain the graph node verification values. ; Verification values ​​for each graph node Associating layers to construct verifiable nodes that support dynamic updates, including: For the first By aggregating the verification values ​​of graph nodes in the layer, we can obtain the first... Layer verification value ; Then, perform inter-layer correlation on the validation values ​​of each layer to obtain the global validation value corresponding to the entire hierarchical sparse graph index. .

7. The privacy-preserving image retrieval method based on multiple keys according to claim 1, characterized in that: The querying user extracts image features from the image data to be retrieved, obtaining a feature vector of the image to be retrieved. This feature vector is then encrypted using the querying user's query key matrix, resulting in a searchable trapdoor; including: The ResNet50 model is used to extract image features from the image data to be retrieved, resulting in the feature vector of the image to be retrieved. Feature vector of the image to be retrieved Perform normalization processing; For the Each query user uses the query key matrix corresponding to that user. The normalized feature vector of the image to be retrieved is encrypted to generate a query trapdoor; Query trap Send to the cloud server.

8. The privacy-preserving image retrieval method based on multiple keys according to claim 7, characterized in that: The cloud server will use the key conversion parameters to identify searchable query trapdoors. Mapped to a unified comparison domain, candidate search is performed in the hierarchical sparse graph index, and reordering is performed using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence; including: The cloud server received the first A query trap sent by a query user Then, according to the key conversion parameters Query trap Mapping to the unified comparison domain yields a transformation trapdoor. ; After completing the unified comparison domain mapping, the cloud server starts from the entry node of the hierarchical sparse graph index. Starting from this point, perform a candidate search layer by layer along local nearest neighbor edges within the same layer and jump edges across layers, and select and transform trapdoors at each layer. The nearest graph node serves as the entry point for the next search layer, continuing until the lowest layer yields the candidate graph node set. ; For candidate graph node set Candidate images in the dataset, combined with candidate image feature vectors The local diagonal covariance statistic of its corresponding graph node Candidate images are calculated using adaptive hybrid distance. Image to be retrieved Second-stage reordering scores ; The second-stage reordering score corresponding to each candidate image. Sort by size from smallest to largest and select the top... The results are used as the final search results to obtain a set of similar encrypted images. ; The cloud server generates a set of similar encrypted images based on a verifiable structure that supports dynamic updates. Corresponding set of verification evidence .

9. A privacy-preserving image retrieval system based on multiple keys, characterized in that, include: The key distribution module uses a trusted center to distribute index encryption matrices and query key matrices to image owners and query users respectively, and distributes key conversion parameters to the cloud server. The hierarchical sparse graph index construction module uses the ResNet50 model to extract image features from the image dataset, obtains a set of image feature vectors, and constructs a hierarchical sparse graph index based on the set of image feature vectors. The image encryption module is used to encrypt each image in the image dataset using the AES algorithm, resulting in an encrypted image set. A verifiable structure building module is used to configure local diagonal covariance statistics for graph nodes in the hierarchical sparse graph index, and to encrypt the encrypted images of each graph node in the hierarchical sparse graph index in sequence to obtain encrypted node representations; to generate summary messages based on the encrypted node representations and local diagonal covariance statistics; and to process the summary messages using chameleon hashing to build a verifiable structure that supports dynamic updates. The trapdoor generation module is used to extract image features from the image data to be retrieved, obtain the feature vector of the image to be retrieved, and encrypt the feature vector of the image to be retrieved using the query key matrix of the querying user to obtain a searchable trapdoor. The image retrieval module is used to map searchable query trapdoors to a unified comparison domain according to the key conversion parameters, perform candidate search in the hierarchical sparse graph index, and reorder them using an adaptive hybrid distance to obtain a set of similar encrypted images and a corresponding set of verification evidence. The verification and decryption module is used to verify the set of similar encrypted images and the corresponding set of verification evidence. After verification, the set of similar encrypted images is decrypted to obtain the similar images.

10. An electronic device comprising at least one processor; and a memory communicatively connected to the at least one processor; in, The memory stores a computer program executable by the at least one processor, characterized in that the computer program is executed by the at least one processor to cause the electronic device to perform the privacy-preserving image retrieval method based on multiple keys as described in any one of claims 1-8.

Images