A full-matter database privacy protection method and device

Abstract

The application discloses a full-mesh database privacy protection method and device, and relates to the technical field of data encryption; the method extracts the first depth feature dataset of the depth feature vector of the query segment data through a preset model; the query segment hash code and the query segment feature are obtained through the first depth feature dataset; the query segment hash code and the query segment feature are uploaded to a cloud server to determine target encrypted data, and the original data is obtained by decrypting the target encrypted data through a first key. The generated query segment hash code is used to search the encrypted dataset corresponding to the query segment hash code in the preset database, the query efficiency and response speed are improved, and all data is prevented from being operated, and the target encrypted data is determined through the query segment feature, so that the required data is searched without obtaining the data content, the resources consumed by data searching are reduced under the condition of ensuring data security.

Description

Technical Field

[0001] This invention belongs to the field of data encryption technology, specifically relating to a method and apparatus for protecting privacy in a fully encrypted database. Background Technology

[0002] Fully Homomorphic Encryption (FHE) is an encryption technique that allows computations to be performed on encrypted data without first decrypting it. This means that various operations and analyses can be performed on the data while it remains encrypted, thus protecting the data's privacy and security.

[0003] With the rapid development of online multimedia information and intelligent applications, many enterprises and individuals choose to store local data on cloud servers. However, cloud servers, as third parties, are not entirely trustworthy. This causes users to lose physical control over their data when outsourcing it to the cloud, raising concerns about data security and privacy, especially for sensitive voice data. To address this issue, data is typically encrypted and then stored in the cloud in ciphertext. However, when users need to retrieve encrypted data stored in the cloud, they must first download all the data and then decrypt it before they can retrieve it. This operation not only consumes significant amounts of network bandwidth, computing resources, and local storage space but can also lead to server congestion. Summary of the Invention

[0004] The purpose of this invention is to solve the problem that when users need to retrieve encrypted data stored in the cloud, they must first download all the data and then decrypt it before they can retrieve it. This operation not only consumes a lot of network bandwidth, computing resources and local storage space, but may also cause server congestion. Therefore, this invention proposes a fully encrypted database privacy protection method and device.

[0005] In a first aspect of this invention, a fully encrypted database privacy protection method is first proposed, the method being applied to a client:

[0006] The first deep feature dataset is obtained by extracting deep feature vectors from the query fragment data using a preset model;

[0007] The first deep feature dataset obtains the query fragment hash code and query fragment features corresponding to the query fragment data through a hash generation method and a query feature generation method;

[0008] The query segment hash code and the query segment feature are uploaded to the cloud server; the cloud server then searches a preset database for an encrypted dataset corresponding to the query segment hash code based on the query segment hash code. For each encrypted dataset, a target encrypted dataset is determined based on the query segment feature and the complete original feature corresponding to the encrypted audio data. The target encrypted dataset is then sent to the client; the target encrypted dataset is obtained by encrypting the original data with a first key; the query segment data is a segment from the original data.

[0009] The target encrypted data is obtained, and the original data is obtained by decrypting the target encrypted data using the first key.

[0010] Optionally, the original data is audio data, and the deep feature data of the query segment data extracted through a preset model includes:

[0011] The query segment data is opened from the preset target library to obtain the speech waveform data of the query segment data;

[0012] The speech waveform data is processed by removing all parts that do not contain actual speech to obtain multiple speech waveform data segments; the parts that do not contain actual speech are silence segments and background noise segments.

[0013] Calculate the MFCC coefficients of the speech frames in each speech waveform data segment to obtain multiple sets of feature data;

[0014] All group feature data are fed into the CNN model to obtain deep feature data.

[0015] Optionally, the hash generation method specifically includes:

[0016] The deep feature vectors in the deep feature dataset are sorted by size to obtain the feature sorting set;

[0017] Determine the depth feature vector value corresponding to the median position in the feature sorting set, and denot it as the target judgment value;

[0018] Using the target judgment value as the judgment condition, all deep feature vectors in the deep feature dataset are hashed to obtain the target hash code; the deep feature vector values ​​in the deep feature dataset that are greater than the target judgment value are recorded as 1, and the deep feature vector values ​​that are less than or equal to the target judgment value are recorded as 0.

[0019] Optionally, the query feature generation method specifically includes: processing the first deep feature dataset using a second key to obtain query fragment features; the second key includes a first random number and a preset matrix;

[0020] For the first deep feature dataset, all first deep feature data vectors are transformed to a first target dimension using the first random number to obtain a first target feature vector; the first target dimension is the dimension of the first deep feature data vectors in the first deep feature dataset plus 3;

[0021] The first target fragment feature is obtained by encrypting the first target feature vector using a preset matrix; the number of rows and columns of the preset matrix is ​​the same as the dimension of the first deep feature data vector in the first deep feature dataset.

[0022] Optionally, the method applied to the client further includes:

[0023] The second deep feature dataset is obtained by extracting deep feature vectors from the original data using a preset model.

[0024] The second deep feature dataset obtains the complete hash code and complete original features corresponding to the original data through a hash generation method and a complete feature generation method;

[0025] The original data is encrypted using the first key to obtain encrypted data;

[0026] The encrypted data is then bound to the complete hash code and the complete original feature and uploaded to the cloud server.

[0027] Optionally, the complete feature generation method specifically includes: processing the second deep feature dataset using a third key to obtain the original complete features; the third key includes a second random number, a third random number, and a preset inverse matrix;

[0028] For the second deep feature dataset, all second deep feature data vectors are transformed to a second target dimension using the second random number and the third random number to obtain a second target feature vector; the second target dimension is the dimension of the second deep feature data vector in the second deep feature dataset plus 3;

[0029] The original complete features are obtained by encrypting the second target feature vector using a preset inverse matrix; the number of rows and columns of the preset inverse matrix is ​​the same as the dimension of the second deep feature data vector in the inverse deep feature dataset.

[0030] In a second aspect of this invention, a fully encrypted database privacy protection method is proposed, which is applied to a cloud server:

[0031] Obtain the hash code of the query fragment, and search a preset database for an encrypted dataset corresponding to the hash code of the query fragment; the encrypted dataset includes encrypted data, a complete hash code, and complete original features;

[0032] The query fragment features are obtained, and the target distance between the query fragment features and the complete original features corresponding to each encrypted data in the encrypted dataset is calculated using Euclidean distance. The encrypted data corresponding to the minimum target distance is recorded as the target encrypted data, and the target encrypted data is sent to the client.

[0033] In a third aspect of this invention, a fully encrypted database privacy protection device is proposed, characterized in that the device is applied to a client and includes:

[0034] The first deep feature generation module is used to extract the deep feature vectors of the query fragment data through a preset model to obtain the first deep feature dataset.

[0035] The first deep feature processing module is used to obtain the query fragment hash code and query fragment features corresponding to the query fragment data from the first deep feature dataset through a hash generation method and a query feature generation method.

[0036] The query module is used to upload the query segment hash code and the query segment features to the cloud server; so that the cloud server can search for the encrypted dataset corresponding to the query segment hash code in a preset database; for each encrypted data, the target encrypted data is determined based on the query segment features and the complete original features corresponding to the encrypted audio data; and the target encrypted data is sent to the client; the target encrypted data is obtained by encrypting the original data with a first key; and the query segment data is a segment of the original data.

[0037] The decryption module is used to obtain the target encrypted data and decrypt the target encrypted data using the first key to obtain the original data.

[0038] In a fourth aspect of this invention, a fully encrypted database privacy protection device is proposed, characterized in that the device is applied to a cloud server:

[0039] An encrypted dataset lookup module is used to obtain the hash code of the query fragment and search for an encrypted dataset corresponding to the hash code of the query fragment in a preset database; the encrypted dataset includes encrypted data, a complete hash code, and complete original features;

[0040] The target encrypted data determination module is used to obtain the query fragment features, calculate the target distance between the query fragment features and the complete original features corresponding to each encrypted data in the encrypted dataset using Euclidean distance, determine the encrypted data corresponding to the minimum target distance as the target encrypted data, and send the target encrypted data to the client.

[0041] The beneficial effects of this invention are:

[0042] This invention proposes a fully encrypted database privacy protection method. It extracts deep feature vectors from query fragment data using a pre-defined model to obtain a first deep feature dataset. This first deep feature dataset is then used to generate query fragment hash codes and query fragment features corresponding to the query fragment data using a hash generation method and a query feature generation method. The query fragment hash codes and features are uploaded to a cloud server. The cloud server then uses the query fragment hash codes to search for encrypted datasets corresponding to those hash codes in the pre-defined database. For each encrypted data set, the target encrypted data is determined based on the query fragment features and the complete original features corresponding to the encrypted audio data. This target encrypted data is then sent to the client. The target encrypted data is then obtained and decrypted using a first key to retrieve the original data. By using the generated query fragment hash codes to search for encrypted datasets corresponding to those hash codes in the pre-defined database, query efficiency and response speed are improved. This avoids operating on all data before determining the target encrypted data using query fragment features, ensuring that the required data can be retrieved without accessing the data content itself. This avoids the problem of users having to download and decrypt all data before retrieving encrypted data stored in the cloud, thus reducing the resources consumed in data retrieval while ensuring data security. Attached Figure Description

[0043] The invention will now be further described with reference to the accompanying drawings.

[0044] Figure 1 A flowchart of a fully encrypted database privacy protection method is provided as an embodiment of the present invention;

[0045] Figure 2 A flowchart of another fully encrypted database privacy protection method is provided for embodiments of the present invention;

[0046] Figure 3 A schematic diagram of a fully encrypted database privacy protection device is provided as an embodiment of the present invention;

[0047] Figure 4 This invention provides a schematic diagram of another fully encrypted database privacy protection device. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and B can represent: A alone, A and B simultaneously, and B alone. Furthermore, descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" can explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0049] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] This invention provides a method for protecting the privacy of a fully encrypted database. See also... Figure 1 , Figure 1 This is a flowchart illustrating a fully encrypted database privacy protection method provided in an embodiment of the present invention. The method is applied to a client and specifically includes the following steps:

[0051] S101, the first deep feature dataset is obtained by extracting the deep feature vector of the query fragment data through the preset model.

[0052] S102, the first deep feature dataset obtains the query fragment hash code and query fragment feature corresponding to the query fragment data through the hash generation method and the query feature generation method.

[0053] S103, upload the query segment hash code and query segment features to the cloud server; so that the cloud server can search the preset database for the encrypted dataset corresponding to the query segment hash code based on the query segment hash code, and for each encrypted data, determine the target encrypted data based on the query segment features and the complete original features corresponding to the encrypted audio data, and send the target encrypted data to the client.

[0054] S104: Obtain the target encrypted data, and decrypt the target encrypted data using the first key to obtain the original data.

[0055] The target encrypted data is obtained by encrypting the original data with the first key; the query fragment data is a fragment of the original data.

[0056] The fully encrypted database privacy protection method provided by this invention improves query efficiency and response speed by using the generated query fragment hash code to find the encrypted dataset corresponding to the query fragment hash code in a preset database. It avoids operating on all data and then determines the target encrypted data through query fragment features, ensuring that the required data can be retrieved without obtaining the data content. This avoids the problem that when users need to retrieve encrypted data stored in the cloud, they must first download all data and then decrypt it before they can perform the retrieval. It reduces the resources consumed by data retrieval while ensuring data security.

[0057] In one implementation, a pre-defined model is used to extract deep feature vectors from the query fragment data, thereby obtaining a first deep feature dataset. This helps to extract high-level feature representations of the data.

[0058] In one implementation, the target encrypted data is obtained by searching the fragment hash code, which can narrow down the scope of verification that needs to be performed by querying fragment features. This improves speed while ensuring that the data is not leaked.

[0059] In one embodiment, the original data is audio data, and the extraction of deep feature data of the query segment data through a preset model includes:

[0060] Open the query segment data from the preset target library to obtain the speech waveform data of the query segment data;

[0061] Data removal is performed on all parts of the speech waveform data that do not contain actual speech to obtain multiple speech waveform data segments; the parts that do not contain actual speech are the silence segment and the background noise segment.

[0062] Calculate the MFCC coefficients of the speech frames in each speech waveform data segment to obtain multiple sets of feature data;

[0063] All group feature data are fed into the CNN model to obtain deep feature data.

[0064] In one implementation, the preset target library is a library that can load speech files and obtain the original speech waveforms. The preset target library can be any one of Librosa, PyAudio, Soundfile, TensorFlow IO, Essentia, Kaldi, and Soundfile.

[0065] In one implementation, multiple speech waveform data segments are obtained by removing all parts of the speech waveform data that do not contain actual speech. All segments that do not contain actual speech are selected and deleted, which reduces the dimensionality of subsequent hash generation and also ensures that the hash value generated by querying the segment data corresponds to the hash value generated by the original data.

[0066] In one implementation, the CNN model consists of 8 layers, with the first 5 being convolutional layers. Each convolutional layer contains 8 convolutional filters and uses ReLU as the activation function. Batch normalization is performed in convolutional layers 2, 4, and 5 to improve fitting speed. A dropout regularization method with a value of 0.25 is also introduced to prevent overfitting. The formula for extracting deep feature data is: Where V represents depth feature data, m i h represents the weights of the i-th recurrent convolutional layer and pooling layer. i σ represents the MFCC coefficients input in the i-th iteration, σ is the bias parameter to prevent overfitting, and n is the number of iterations.

[0067] In one embodiment, the hash generation method specifically includes:

[0068] The deep feature vectors in the deep feature dataset are sorted by size to obtain the feature sorting set;

[0069] Determine the deep feature vector value corresponding to the median position in the feature sorting set, and denote it as the target judgment value;

[0070] The target hash code is obtained by performing a hash transformation on all deep feature vectors in the deep feature dataset using the target judgment value as the judgment condition. Deep feature vector values ​​in the deep feature dataset that are greater than the target judgment value are recorded as 1, and deep feature vector values ​​that are less than or equal to the target judgment value are recorded as 0.

[0071] In one implementation, the deep feature vectors in the deep feature dataset are transformed into irreversible target hash codes, thereby better protecting the privacy and security of the data, reducing the complexity and dimensionality of the data, and simplifying subsequent data retrieval.

[0072] In one embodiment, the query feature generation method specifically includes: processing a first deep feature dataset using a second key to obtain query fragment features; the second key includes a first random number and a preset matrix;

[0073] For the first deep feature dataset, all first deep feature data vectors are transformed to the first target dimension using a first random number to obtain the first target feature vector; the first target dimension is the dimension of the first deep feature data vector in the first deep feature dataset plus 3;

[0074] The first target fragment features are obtained by encrypting the first target feature vector using a preset matrix; the number of rows and columns of the preset matrix are the same as the dimension of the first deep feature data vector in the first deep feature dataset.

[0075] In one implementation, the following embodiments are only used to illustrate the application of the query feature generation method in this invention, including a second key K2{S1, M}, where S1 is a first random number, M is a preset matrix, and a first deep feature dataset T{t1, t2, ..., ti}; transforming all first deep feature data vectors to a first target dimension to obtain a first target feature vector (reference formula). The first target fragment features are then obtained by encrypting the first target feature vector using a preset matrix.

[0076] In one embodiment, the method applied to the client further includes:

[0077] The second deep feature dataset is obtained by extracting deep feature vectors from the original data using a preset model.

[0078] The second deep feature dataset obtains the complete hash code and complete original features corresponding to the original data through hash generation method and complete feature generation method;

[0079] The original data is encrypted using the first key to obtain encrypted data;

[0080] The encrypted data is then bound to the complete hash code and the complete original characteristics before being uploaded to the cloud server.

[0081] One implementation method involves encrypting the original data using a first key to obtain encrypted data. Specifically, the first key includes a preset first sequence, a preset second sequence, and a preset third sequence. Encrypting the original data using the first key specifically includes: dividing the original data into non-overlapping segments to obtain M speech frames, each containing N sampling points; performing a discrete cosine transform on each speech frame to obtain a coefficient domain matrix; the coefficient domain matrix has a size of M*N; performing row scrambling on the coefficient domain matrix using the preset first sequence, followed by column scrambling on the scrambled coefficient domain matrix using the preset second sequence to obtain a target coefficient domain matrix; concatenating all target coefficient domain matrices in frame order, and then performing an inverse discrete cosine transform to obtain target data; the target data is a one-dimensional time-domain signal; and performing a bitwise XOR operation between the target data and the preset third sequence to obtain encrypted data.

[0082] In one embodiment, the complete feature generation method specifically includes: processing the second deep feature dataset using a third key to obtain the original complete features; the third key includes a second random number, a third random number, and a preset inverse matrix;

[0083] For the second deep feature dataset, all second deep feature data vectors are transformed to the second target dimension using the second and third random numbers to obtain the second target feature vector; the second target dimension is the dimension of the second deep feature data vector in the second deep feature dataset plus 3;

[0084] The original complete features are obtained by encrypting the second target feature vector using a preset inverse matrix; the number of rows and columns of the preset inverse matrix is ​​the same as the dimension of the second deep feature data vector in the inverse deep feature dataset.

[0085] In one implementation, the embodiments described below are only used to illustrate the application of the query feature generation method in this invention, including the third key K2{S2, S3, M}. -1 S2 is the second random number, S3 is the third random number, and M is the third random number. -1 Given a predefined inverse matrix and a second deep feature dataset R{r1, r2, ..., rn}, transform all second deep feature data vectors to the second target dimension to obtain the second target feature vector (reference formula). The first target fragment features are then obtained by encrypting the first target feature vector using a preset matrix.

[0086] Based on the same inventive concept, this invention also provides another method for protecting the privacy of a fully encrypted database. See also Figure 2 , Figure 2 A flowchart illustrating another fully encrypted database privacy protection method provided in this embodiment of the invention, the method being applied to a cloud server:

[0087] S201, Obtain the query fragment hash code and search for the encrypted dataset corresponding to the query fragment hash code in the preset database.

[0088] S202, obtain the query fragment features, calculate the target distance between the query fragment features and the complete original features corresponding to each encrypted data in the encrypted dataset using Euclidean distance, determine the encrypted data corresponding to the minimum target distance and record it as the target encrypted data, and send the target encrypted data to the client.

[0089] The encrypted dataset includes encrypted data, complete hash codes, and complete original features;

[0090] In one implementation, Euclidean distance is used to calculate the target distance between the query fragment features and the complete original features corresponding to each encrypted data in the encrypted dataset. This allows the search for the original data most similar to the query fragment, avoiding the need for customers to download all data and decrypt it before retrieving encrypted data stored in the cloud, thus improving efficiency.

[0091] Based on the same inventive concept, this invention also provides a fully encrypted database privacy protection device. See also Figure 3 , Figure 3 This is a schematic diagram of a fully encrypted database privacy protection device provided in an embodiment of the present invention. The device is applied to a client:

[0092] The first deep feature generation module is used to extract the deep feature vectors of the query fragment data through a preset model to obtain the first deep feature dataset.

[0093] The first deep feature processing module is used to obtain the query fragment hash code and query fragment features corresponding to the query fragment data from the first deep feature dataset through the hash generation method and the query feature generation method.

[0094] The query module is used to upload the query segment hash code and query segment features to the cloud server; so that the cloud server can search the preset database for the encrypted dataset corresponding to the query segment hash code based on the query segment hash code; for each encrypted data, the target encrypted data is determined based on the query segment features and the complete original features corresponding to the encrypted audio data, and the target encrypted data is sent to the client; the target encrypted data is obtained by encrypting the original data with the first key; the query segment data is a segment in the original data;

[0095] The decryption module is used to obtain the target encrypted data and decrypt the target encrypted data using the first key to obtain the original data.

[0096] Based on the same inventive concept, embodiments of the present invention also provide another fully encrypted database privacy protection device. See also Figure 4 , Figure 4 This is a schematic diagram of another fully encrypted database privacy protection device provided in an embodiment of the present invention. The device is applied to a cloud server.

[0097] In one embodiment, the device for use on a client also includes:

[0098] The encrypted dataset search module is used to obtain the hash code of the query fragment and search for the encrypted dataset corresponding to the hash code of the query fragment in the preset database; the encrypted dataset includes encrypted data, complete hash code and complete original features;

[0099] The target encrypted data determination module is used to obtain query fragment features, calculate the target distance between the query fragment features and the complete original features corresponding to each encrypted data in the encrypted dataset using Euclidean distance, determine the encrypted data corresponding to the minimum target distance as the target encrypted data, and send the target encrypted data to the client.

[0100] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for protecting privacy in a fully encrypted database, characterized in that, The method is applied to the client: The first deep feature dataset is obtained by extracting deep feature vectors from the query fragment data using a preset model; The first deep feature dataset obtains the query fragment hash code and query fragment features corresponding to the query fragment data through a hash generation method and a query feature generation method; Upload the query fragment hash code and the query fragment features to the cloud server; The cloud server searches a preset database for an encrypted dataset corresponding to the query segment hash code based on the query segment hash code. For each encrypted dataset, a target encrypted dataset is determined based on the query segment features and the complete original features corresponding to the encrypted audio data. The target encrypted dataset is then sent to the client. The target encrypted dataset is obtained by encrypting the original data using a first key. The query segment data is a segment from the original data; Obtain the target encrypted data, and decrypt the target encrypted data using the first key to obtain the original data; The query feature generation method specifically includes: processing the first deep feature dataset using a second key to obtain query fragment features; the second key includes a first random number and a preset matrix; For the first deep feature dataset, all first deep feature data vectors are transformed to a first target dimension using the first random number to obtain a first target feature vector; the first target dimension is the dimension of the first deep feature data vectors in the first deep feature dataset plus 3; The first target fragment feature is obtained by encrypting the first target feature vector using a preset matrix; the number of rows and columns of the preset matrix is ​​the same as the dimension of the first deep feature data vector in the first deep feature dataset. The method also applies to the client side: The second deep feature dataset is obtained by extracting deep feature vectors from the original data using a preset model; The second deep feature dataset obtains the complete hash code and complete original features corresponding to the original data through a hash generation method and a complete feature generation method; The original data is encrypted using the first key to obtain encrypted data; The encrypted data is bound to the complete hash code and the complete original feature and then uploaded to the cloud server. The method is applied to cloud servers: Obtain the hash code of the query fragment, and search a preset database for an encrypted dataset corresponding to the hash code of the query fragment; the encrypted dataset includes encrypted data, a complete hash code, and complete original features; The query fragment features are obtained, and the target distance between the query fragment features and the complete original features corresponding to each encrypted data in the encrypted dataset is calculated using Euclidean distance. The encrypted data corresponding to the minimum target distance is determined and recorded as the target encrypted data. The target encrypted data is then sent to the client.

2. The method for protecting privacy in a fully encrypted database according to claim 1, characterized in that, The original data is audio data, and the first deep feature dataset for extracting the query segment data using a preset model includes: The query segment data is opened from the preset target library to obtain the speech waveform data of the query segment data; The speech waveform data is processed by removing all parts that do not contain actual speech to obtain multiple speech waveform data segments; the parts that do not contain actual speech are silence segments and background noise segments. Calculate the MFCC coefficients of the speech frames in each speech waveform data segment to obtain multiple sets of feature data; The first deep feature dataset is obtained by feeding all the group feature data into the CNN model.

3. The method for protecting privacy in a fully encrypted database according to claim 1, characterized in that, The hash generation method specifically includes: The feature sorting set is obtained by sorting the depth feature vectors in the first depth feature dataset by size; Determine the depth feature vector value corresponding to the median position in the feature sorting set, and denot it as the target judgment value; Using the target judgment value as the judgment condition, all depth feature vectors in the first depth feature dataset are hashed to obtain the target hash code; depth feature vector values ​​in the first depth feature dataset that are greater than the target judgment value are recorded as 1, and depth feature vector values ​​that are less than or equal to the target judgment value are recorded as 0.

4. The method for protecting privacy in a fully encrypted database according to claim 1, characterized in that, The complete feature generation method specifically includes: processing the second deep feature dataset using a third key to obtain complete original features; the third key includes a second random number, a third random number, and a preset inverse matrix; For the second deep feature dataset, all second deep feature data vectors are transformed to a second target dimension using the second random number and the third random number to obtain a second target feature vector; the second target dimension is the dimension of the second deep feature data vector in the second deep feature dataset plus 3; The second target feature vector is encrypted using a preset inverse matrix to obtain the complete original features; the number of rows and columns of the preset inverse matrix is ​​the same as the dimension of the second deep feature data vector in the second deep feature dataset.

5. A fully encrypted database privacy protection device, characterized in that, The device is used in the client application including: The first deep feature generation module is used to extract the deep feature vectors of the query fragment data through a preset model to obtain the first deep feature dataset. The first deep feature processing module is used to obtain the query fragment hash code and query fragment features corresponding to the query fragment data from the first deep feature dataset through a hash generation method and a query feature generation method. The query module is used to upload the query fragment hash code and the query fragment features to the cloud server; The cloud server searches a preset database for an encrypted dataset corresponding to the query segment hash code based on the query segment hash code. For each encrypted dataset, a target encrypted dataset is determined based on the query segment features and the complete original features corresponding to the encrypted audio data. The target encrypted dataset is then sent to the client. The target encrypted dataset is obtained by encrypting the original data using a first key. The query segment data is a segment from the original data; The decryption module is used to obtain the target encrypted data and decrypt the target encrypted data using the first key to obtain the original data. The query feature generation method specifically includes: processing the first deep feature dataset using a second key to obtain query fragment features; the second key includes a first random number and a preset matrix; For the first deep feature dataset, all first deep feature data vectors are transformed to a first target dimension using the first random number to obtain a first target feature vector; the first target dimension is the dimension of the first deep feature data vectors in the first deep feature dataset plus 3; The first target fragment feature is obtained by encrypting the first target feature vector using a preset matrix; the number of rows and columns of the preset matrix is ​​the same as the dimension of the first deep feature data vector in the first deep feature dataset. The device also includes, for use with the client: The second deep feature dataset is obtained by extracting deep feature vectors from the original data using a preset model; The second deep feature dataset obtains the complete hash code and complete original features corresponding to the original data through a hash generation method and a complete feature generation method; The original data is encrypted using the first key to obtain encrypted data; The encrypted data is then bound to the complete hash code and the complete original feature before being uploaded to the cloud server.

Images