Privacy protection SURF image feature extraction method based on secret sharing in cloud environment

Abstract

The invention discloses a privacy protection SURF image feature extraction method based on secret sharing in a cloud environment. According to the method, a data owner, a cloud server and three participants are included. A data owner randomly divides a private image into two addition random shares at a local end and respectively sends the two addition random shares to two non-collusive servers. And after receiving the respective image share, cooperatively operating a series of secure interactive operation protocols based on additive secret sharing to cooperatively complete an SURF feature extraction algorithm so as to obtain the respective SURF feature descriptor random share sum, and returning the SURF feature descriptor random share sum to the data owner. And the data owner can recover the real SURF feature descriptor through simple addition. On the premise of effectively protecting the image content privacy of the data owner, the completeness, efficiency and accuracy of the cloud SURF step are improved, and the calculation and communication burden of the data owner is effectively reduced.

Description

Technical Field

[0001] This invention relates to the technical fields of multimedia information security and image processing, and specifically to a privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment. Background Technology

[0002] With the advent of the big data era, massive amounts of multimedia data, especially image data such as medical images, personal mobile phone photos, and satellite image databases, are being generated at an exponential rate. For resource-constrained data owners, the cost of storing and processing large amounts of image data has become extremely high. Fortunately, cloud computing allows data owners to outsource expensive data storage and complex image data processing to cloud servers, utilizing customizable computing services to handle specific tasks. For images containing sensitive information (such as faces, credit card security codes, age, salary, etc.), data owners should not outsource images to the unprotected cloud to prevent privacy breaches. To address privacy concerns, data owners can encrypt or perturb images locally before outsourcing them to the cloud. However, the use of image encryption renders traditional image processing solutions ineffective on encrypted images. Encrypted image processing has attracted widespread attention and is contributing to the development of cloud-based privacy-preserving computing.

[0003] In recent years, many technologies such as homomorphic encryption, secure multi-party computation, and order-preserving encryption have been applied to privacy-preserving computing. In the field of image processing, image feature extraction is a crucial step, removing redundant information and extracting quantitative descriptors that describe image features. For data owners with large amounts of image data but limited resources, there are insufficient computing resources to support local image feature extraction. The high computational cost of feature extraction forces image owners to outsource it to cloud computing.

[0004] It is evident that image feature extraction in the cloud environment has wide applications, but it also raises privacy and security concerns when using the storage and computing resources of cloud servers: the image data itself may contain the private information of the image owner; and image owners face the risk of privacy leakage when using clouds located in different trust domains.

[0005] In conclusion, it is essential to propose an image feature extraction method that balances privacy and security, feature extraction accuracy, and operational efficiency in a cloud computing environment. Speeded-Up Robust Feature (SURF) is a widely used feature extraction algorithm, an enhanced version of the well-known feature extractor SIFT.

[0006] Chinese invention application No. 201410031154.7 discloses "A method for extracting SURF features from encrypted images". The method includes: firstly, constructing a Paillier encryption system and a DGK encryption system, and generating corresponding public and private keys; the user end uses the Paillier encryption system to encrypt the image with the generated public key, and then sends the encrypted image to the server end; the server end extracts SURF feature points from the encrypted image; the server end corrects the extracted SURF feature points; and the server end extracts SURF feature descriptors. Summary of the Invention

[0007] To address the technical problems existing in the background technology, this invention provides a privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment. The technical solution adopted by this invention is as follows: The first aspect of this invention provides a privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment, the method comprising: The original privacy image is divided into a first additive random share and a second additive random share; The first additive random share and the second additive random share are respectively sent to the preset first cloud server and the second cloud server for integral image calculation to obtain the corresponding first integral image random share and second integral image random share. Based on the first integral image random share and the second integral image random share, scale space is constructed through the first cloud server and the second cloud server to obtain the first scale space additive random share and the second scale space additive random share. By taking two adjacent points in their respective scale space addition random shares as input, the first cloud server and the second cloud server cooperate to run a secure comparison protocol to obtain the comparison results of adjacent points in the scale space, and take the extreme points that are greater than the preset threshold as candidate key points. The first cloud server and the second cloud server cooperate to run the security matrix inverse protocol and the security matrix multiplication protocol to perform Taylor interpolation fitting on the candidate key points, and delete the candidate key points with excessive offset to obtain the final key points; The first cloud server and the second cloud server cooperate to run a secure division protocol and a secure comparison protocol to securely calculate the gradient angle of sampling points in the neighborhood of the key point. Then, the key point orientation is calculated by cooperating to run a secure multiplication protocol and a secure comparison protocol. By cooperating with the first cloud server and the second cloud server to run the Secure Absolute Value Protocol, a multi-dimensional floating-point feature vector is calculated as the SURF feature descriptor. After the calculation is completed, the first feature descriptor vector addition random share and the second feature descriptor vector addition random share are obtained respectively. The first cloud server and the second cloud server return the random share of the first feature descriptor vector addition and the random share of the second feature descriptor vector addition to the data owner, and the data owner recovers the true SURF feature descriptor through addition.

[0008] As a preferred approach, the method for dividing the original privacy image into a first additive random share and a second additive random share includes: Data owner The original privacy image is stored locally. Randomly divided into first additive random shares Second addition random share ,satisfy The first additive random share To maintain the original privacy image The second additive random share of the same size random number matrix. The image is disturbed by adding random numbers.

[0009] As a preferred embodiment, the method of sending the first additive random share and the second additive random share to a preset first cloud server and a preset second cloud server for integral image calculation to obtain the corresponding first integral image share and second integral image share includes: First Cloud Server calculate integral image Second cloud server calculate integral image Because it satisfies ,Right now and yes The addition of random shares, therefore satisfying the condition... ,Right now and yes The random fraction of the integral image, where that is The integral image.

[0010] As a preferred embodiment, the method for constructing a scale space based on the first integral image random share and the second integral image random share through a first cloud server and a second cloud server, and obtaining the first scale space additive random share and the second scale space additive random share, includes: First Cloud Server Second cloud server Random share of their respective integral images and Use different scales An approximate box filter with a 2D Gaussian convolution kernel is used to compute the Hessian matrix for each pixel. ; Obtain the additive random shares of the Hessian matrix C2 obtains the additive random shares of the Hessian matrix. ; and Each Matrix Addition Random Shares and As input, the cooperative operation of the security matrix determinant protocol ,get Additive random shares of the matrix determinant get , get ,satisfy ,in yes The determinant, i.e., the scale space; the scale space is a three-dimensional floating-point space. These are the x and y coordinates of the image, respectively. For the scale coordinates of the self-box filter; and After collaboratively calculating all pixels After performing the matrix determinant, we obtain the first-scale space additive random share. Second-scale spatial additive random share .

[0011] As a preferred embodiment, the method of using two adjacent points from their respective scale-space additive random shares as input, and cooperating to run a secure comparison protocol to obtain the comparison results of adjacent points in the scale space, and selecting extreme points greater than a preset threshold as candidate key points, includes: First Cloud Server Second cloud server Using two adjacent points in their respective scale space shares as input, a cooperative operational safety comparison protocol is established. This yields the comparison results of adjacent points in the scale space; a result of 1 indicates that the former is larger, a result of 0 indicates that the two are equal, and a result of -1 indicates that the latter is larger. and The collaboration compares each point in the scale space with its 26 neighboring points to detect machine extrema in the scale space, and selects extrema points that are greater than a predefined threshold as candidate key points.

[0012] As a preferred approach, the method for obtaining the final key points by having the first cloud server and the second cloud server collaboratively run a security matrix inverse protocol and a security matrix multiplication protocol to perform Taylor interpolation fitting on candidate key points and remove candidate key points with excessive offsets includes: Interpolation fitting is performed on the positions of candidate keypoints to obtain their precise locations: Let the coordinates of the candidate key points be... The value of this point in scale space is ,set up Then the Taylor expansion of that point is:

[0013] Differentiate equation (eq_1) and set the derivative to 0 to find the exact location of the extreme point. Set precise location With discrete position The offset is Solving for:

[0014] Due to the first cloud server Second cloud server Holding additive random shares in the scale space, thus obtaining the matrix sum matrix The addition of random shares; and First, according to formula (eq_2), the cooperative running security matrix inversion protocol is executed. To obtain inverse matrix The addition of random shares, then the cooperative operation of the secure matrix multiplication protocol. The offset of the extreme point position is calculated. Additive random shares, hold , hold ; and Through cooperative operation of the security comparison protocol offset Keypoints exceeding the preset value are considered unstable and are deleted.

[0015] As a preferred embodiment, the method for securely calculating the gradient angle of sampling points in the neighborhood of a key point by having the first cloud server and the second cloud server cooperate to run a secure division protocol and a secure comparison protocol, and then calculating the direction of the key point by cooperating to run a secure multiplication protocol and a secure comparison protocol, includes: In the image Around each key point on Within a circular neighborhood of the radius To ensure uniform sampling at intervals, Calculate the sampling points based on the scale coordinates of the key point. direction and Directional gradient and First Cloud Server Using the first integral image share ,use Wavelet filter obtained and Second cloud server With the help of the second integral image share ,use Wavelet filter obtained and ;in and yes Additive random shares, ; and yes Additive random shares, ; and Using the gradient shares in both directions as input, a cooperative safe division protocol is run. The additive fractions of the gradient tangent values ​​are obtained respectively. and Then, they will cooperate to run a secure comparison protocol. The gradient tangent is compared with the tangent of a preset angle to determine the angle range corresponding to the gradient direction of the sampling point. exist Set the size to within the range The sliding step size is A fan-shaped sliding window, where the statistical gradient direction falls within the sampling points of the window. direction and Cumulative gradient in direction and , and Obtain the additive random share of the cumulative gradient. get and , get and The relationship satisfies , ; Each sliding window direction and The cumulative gradient in the direction forms a direction vector. The longest direction vector in all sliding windows will be... The angle between axes serves as the direction of the key point; and Cooperative Operation of Safe Multiplication Protocol To calculate the square of the direction vector length ; and Obtain the sum of the squared lengths of the direction vector by random addition, and then... Compare the lengths of different direction vectors and obtain the longest direction vector; at last and Cooperative Operation Security Division Agreement The additive fractions of the gradient tangent values ​​are obtained respectively. and Then, cooperate to run a secure comparison protocol. Addition of the tangent value of the gradient direction and The gradient direction of the key point is compared with the tangent value of the preset angle to determine the angle interval corresponding to the gradient direction of the key point; the interval in which the key point direction is located is obtained, and then the midpoint of the interval is used as the key point direction.

[0016] As a preferred embodiment, the method of calculating a multi-dimensional floating-point feature vector as a SURF feature descriptor by having the first cloud server and the second cloud server cooperate to run a Secure Absolute Value Protocol, and obtaining the first feature descriptor vector plus random share and the second feature descriptor vector plus random share after the calculation, includes: For each keypoint assigned an orientation, a 64-dimensional floating-point feature vector is computed as the keypoint's feature descriptor; uniform sampling is performed within a square neighborhood around each keypoint in the image, rotated to the same angle as the keypoint's orientation. Each pixel divides the square neighborhood into equal parts. Each sub-region contains [number] sub-regions. 1 sampling point; for each sub-region, calculate the gradient of the sampling point parallel to the direction of the key point. and gradient perpendicular to the key point direction First Cloud Server pass ,use Wavelet filter obtained and Second cloud server pass ,use Wavelet filter obtained and ,in and yes Additive random shares, ; and yes Additive random shares, ; and Cooperative Operational Security Absolute Value Protocol ,get and absolute value and The added share, C1 gets and C2 obtained and ,in and yes Additive random shares, ; and yes Additive random shares, Then, statistical analysis is performed on all sampling points within the sub-region. , , and The summation, C1 obtains C2 obtained For each of the 16 sub-regions, each server includes 64 additive random shares of cumulative sums. These 64 values ​​are concatenated into a 64-dimensional vector according to their sub-region positions. Holding Vector , Holding Vector ;Will and These are respectively used as random shares of the first feature descriptor vector addition and random shares of the second feature descriptor vector addition.

[0017] A second aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment.

[0018] A third aspect of the present invention provides a computer device, including a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of the aforementioned privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment.

[0019] Compared with the prior art, the beneficial effects of this invention are: This invention employs additive secret sharing to protect image privacy. By implementing the complete steps of SURF feature extraction on two cloud servers through a secure multi-party computation protocol based on additive secret sharing, it aims to protect the privacy of data owners and reduce their computational burden, while ensuring that the extracted SURF features are as close as possible to the original SURF features. Attached Figure Description

[0020] Figure 1 This is a flowchart of the privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment provided in this embodiment; Figure 2 This is a flowchart illustrating the operational framework of the privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment provided in this embodiment. Figure 3 This is a diagram showing the keypoint extraction results of the SURF image feature extraction method provided in this embodiment on a public dataset; Figure 4 This is a matching result diagram of the image feature point matching experiment using the feature descriptor extracted by the SURF image feature extraction method provided in this embodiment. Detailed Implementation The accompanying drawings are for illustrative purposes only and should not be construed as limiting the invention. It should be understood that the described embodiments are merely some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of the embodiments of this application.

[0021] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the embodiments of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0022] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0023] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. The invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Example 1 Please refer to Figure 1 as well as Figure 2 This embodiment provides a privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment. The method includes: S1: Divide the original privacy image into a first additive random share and a second additive random share; In one specific embodiment, the method for dividing the original privacy image into a first additive random share and a second additive random share includes: Data owner The original privacy image is stored locally. Randomly divided into first additive random shares Second addition random share ,satisfy The first additive random share To maintain the original privacy image The second additive random share of the same size random number matrix. The image is disturbed by adding random numbers.

[0026] It should be noted that in step S1, the data is sent to... of It is a random number matrix with the same size as the original image, therefore Unable to obtain any information containing the original image's private content. Sent to of It is an image that has been scrambled by adding random numbers. The privacy content of the original image cannot be directly obtained. If To attempt to recover the original image, you will need to perform... Using brute-force calculations to recover the image is practically impossible.

[0027] S2: Send the first additive random share and the second additive random share to the preset first cloud server and second cloud server respectively for integral image calculation to obtain the corresponding first integral image random share and second integral image random share. In a specific embodiment, the method for sending the first additive random share and the second additive random share to a preset first cloud server and a preset second cloud server respectively for integral image calculation to obtain the corresponding first integral image share and second integral image share includes: First Cloud Server calculate integral image Second cloud server calculate integral image Because it satisfies ,Right now and yes The addition of random shares, therefore satisfying the condition... ,Right now and yes The random fraction of the integral image, where that is The integral image.

[0028] It should be noted that in step S2, and Based on the received image, an integral image of the image is constructed by adding random shares, which is the addition random shares of the integral image of the original image. Therefore, compared with S1, the server does not obtain additional information in step S2.

[0029] S3: Based on the first integral image random share and the second integral image random share, scale space is constructed through the first cloud server and the second cloud server to obtain the first scale space additive random share and the second scale space additive random share; In a specific embodiment, the method for constructing a scale space based on the first integral image random share and the second integral image random share through a first cloud server and a second cloud server, and obtaining the first scale space additive random share and the second scale space additive random share, includes: First Cloud Server Second cloud server Random share of their respective integral images and Use different scales An approximate box filter with a 2D Gaussian convolution kernel is used to compute the Hessian matrix for each pixel. ; Obtain the additive random shares of the Hessian matrix C2 obtains the additive random shares of the Hessian matrix. ; and Each Matrix Addition Random Shares and As input, the cooperative operation of the security matrix determinant protocol ,get Additive random shares of the matrix determinant get , get ,satisfy ,in yes The determinant, i.e., the scale space; the scale space is a three-dimensional floating-point space. These are the x and y coordinates of the image, respectively. For the scale coordinates of the self-box filter; and After collaboratively calculating all pixels After performing the matrix determinant, we obtain the first-scale space additive random share. Second-scale spatial additive random share .

[0030] It should be noted that in step S3, and pass The additive random fractions of the scale space are calculated. It is by A composite computing protocol, and The security of the data has been proven, therefore the data exchanged between servers is secure and will not leak information about the original image's private content.

[0031] S4: The first cloud server and the second cloud server take two adjacent points in their respective scale space addition random shares as input, cooperate to run a security comparison protocol, obtain the comparison results of adjacent points in the scale space, and take the extreme points that are greater than the preset threshold as candidate key points. In a specific embodiment, the method of using two adjacent points from their respective scale-space additive random shares as input, and cooperating to run a secure comparison protocol to obtain the comparison results of adjacent points in the scale space, and selecting extreme points greater than a preset threshold as candidate key points, includes: First Cloud Server Second cloud server Using two adjacent points in their respective scale space shares as input, a cooperative operational safety comparison protocol is established. This yields the comparison results of adjacent points in the scale space; a result of 1 indicates that the former is larger, a result of 0 indicates that the two are equal, and a result of -1 indicates that the latter is larger. and The collaboration compares each point in the scale space with its 26 neighboring points to detect machine extrema in the scale space, and selects extrema points that are greater than a predefined threshold as candidate key points.

[0032] It should be noted that in step S4, and pass The results of the comparison of the magnitude relationship between different points in the scale space were calculated. The security of the data has been proven, therefore the data exchanged between servers is secure and will not leak information about the original image's private content.

[0033] S5: The first cloud server and the second cloud server cooperate to run the security matrix inverse protocol and the security matrix multiplication protocol to perform Taylor interpolation fitting on the candidate key points, and delete the candidate key points with excessive offset to obtain the final key points; In one specific embodiment, the method for obtaining the final key points by having the first cloud server and the second cloud server collaboratively run a security matrix inverse protocol and a security matrix multiplication protocol to perform Taylor interpolation fitting on candidate key points and delete candidate key points with excessive offsets includes: Interpolation fitting is performed on the positions of candidate keypoints to obtain their precise locations: Let the coordinates of the candidate key points be... The value of this point in scale space is ,set up Then the Taylor expansion of that point is:

[0034] Differentiate equation (eq_1) and set the derivative to 0 to find the exact location of the extreme point. Set precise location With discrete position The offset is Solving for:

[0035] Due to the first cloud server Second cloud server Holding additive random shares in the scale space, thus obtaining the matrix sum matrix The addition of random shares; and First, according to formula (eq_2), the cooperative running security matrix inversion protocol is executed. To obtain inverse matrix The addition of random shares, then the cooperative operation of the secure matrix multiplication protocol. The offset of the extreme point position is calculated. Additive random shares, hold , hold ; and Through cooperative operation of the security comparison protocol offset Keypoints exceeding the preset value are considered unstable and are deleted.

[0036] It should be noted that in step S5, the server... , and The interaction involves calculating the offset of extreme point locations and comparing whether the offset exceeds 1. Since the security of these secure calculation protocols has been proven, the data exchanged between servers is secure and will not reveal the content of the original image. After step S5 is completed, the server obtains additional information about the discrete locations of image keypoints in scale space. However, the server cannot obtain the specific locations; these keypoint locations alone can only reveal some basic shapes in the image, not the specific content. The original content of the image is effectively protected.

[0037] S6: The first cloud server and the second cloud server cooperate to run a secure division protocol and a secure comparison protocol to securely calculate the gradient angle of the sampling points in the neighborhood of the key point, and then calculate the direction of the key point by cooperating to run a secure multiplication protocol and a secure comparison protocol. In one specific embodiment, the method for securely calculating the gradient angle of sampling points in the neighborhood of a key point by having the first cloud server and the second cloud server cooperate to run a secure division protocol and a secure comparison protocol, and then calculating the direction of the key point by cooperating to run a secure multiplication protocol and a secure comparison protocol, includes: In the image Around each key point on Within a circular neighborhood of the radius To ensure uniform sampling at intervals, Calculate the sampling points based on the scale coordinates of the key point. direction and Directional gradient and First Cloud Server Using the first integral image share ,use Wavelet filter obtained and Second cloud server With the help of the second integral image share ,use Wavelet filter obtained and ;in and yes Additive random shares, ; and yes Additive random shares, ; and Using the gradient shares in both directions as input, a cooperative safe division protocol is run. The additive fractions of the gradient tangent values ​​are obtained respectively. and Then, they will cooperate to run a secure comparison protocol. The gradient direction tangent value and By comparing the determined values, we can determine which 10-degree interval the gradient direction of the sampling point falls within. exist Set the size to within the range The sliding step size is A fan-shaped sliding window, where the statistical gradient direction falls within the sampling points of the window. direction and Cumulative gradient in direction and , and Obtain the additive random share of the cumulative gradient. get and , get and The relationship satisfies , ; Each sliding window direction and The cumulative gradient in the direction forms a direction vector. The longest direction vector in all sliding windows will be... The angle between axes serves as the direction of the key point; and Cooperative Operation of Safe Multiplication Protocol To calculate the square of the direction vector length ; and Obtain the sum of the squared lengths of the direction vector by random addition, and then... Compare the lengths of different direction vectors and obtain the longest direction vector; at last and Cooperative Operation Security Division Agreement The additive fractions of the gradient tangent values ​​are obtained respectively. and Then, cooperate to run a secure comparison protocol. Addition of the tangent value of the gradient direction and respectively with By comparing the determined values, the angle interval corresponding to the gradient direction of the key point is determined; the interval in which the key point direction is located is obtained, and then the midpoint of the interval is used as the key point direction.

[0038] It should be noted that in step S6, and pass , and Perform interaction and record the direction of key points. Because... , and The security of the algorithm has been proven, therefore the data exchanged between servers is secure and does not leak the content of the original image. In this step, the server obtains additional information about the direction of keypoints, but this direction is the median of the interval containing the true direction, an approximation of the keypoint directions extracted by the original SURF algorithm. The server cannot obtain additional image privacy content based on this.

[0039] Please refer to Figure 3 , Figure 3This is a graph showing the keypoint extraction results of the SURF image feature extraction method on a public dataset. The first row contains four typical scene images from the Caltech256 public dataset. Figure 3 (a) and Figure 3 (b) consists of two images of different sizes. Texture image, Figure 3 (c) is of size Highly structured images, Figure 3 (d) is of size Images with poor structure. Figure 3 (h) shows the feature points extracted by the original SURF algorithm. Figure 3 (l) Demonstrates the feature points extracted by the privacy-preserving SURF method of the present invention. It can be seen that the privacy-preserving scheme of the present invention maintains the same number of feature points extracted as the original SURF algorithm.

[0040] S7: The first cloud server and the second cloud server cooperate to run the Secure Absolute Value Protocol to calculate a multi-dimensional floating-point feature vector as a SURF feature descriptor. After the calculation is completed, the first feature descriptor vector addition random share and the second feature descriptor vector addition random share are obtained respectively. In one specific embodiment, the method of calculating a multi-dimensional floating-point feature vector as a SURF feature descriptor by cooperating between the first cloud server and the second cloud server to run a Secure Absolute Value Protocol includes: For each keypoint assigned an orientation, a 64-dimensional floating-point feature vector is computed as the keypoint's feature descriptor; uniform sampling is performed within a square neighborhood around each keypoint in the image, rotated to the same angle as the keypoint's orientation. Each pixel divides the square neighborhood into equal parts. Each sub-region contains [number] sub-regions. 1 sampling point; for each sub-region, calculate the gradient of the sampling point parallel to the direction of the key point. and gradient perpendicular to the key point direction First Cloud Server pass ,use Wavelet filter obtained and Second cloud server pass ,use Wavelet filter obtained and ,in and yes Additive random shares, ; and yes Additive random shares, ; and Cooperative Operational Security Absolute Value Protocol ,get and absolute value and The added share, C1 gets and C2 obtained and ,in and yes Additive random shares, ; and yes Additive random shares, Then, statistical analysis is performed on all sampling points within the sub-region. , , and The summation, C1 obtains C2 obtained For each of the 16 sub-regions, each server includes 64 additive random shares of cumulative sums. These 64 values ​​are concatenated into a 64-dimensional vector according to their sub-region positions. Holding Vector , Holding Vector ;Will and These are respectively used as random shares of the first feature descriptor vector addition and random shares of the second feature descriptor vector addition.

[0041] It should be noted that in step S7, and pass Interact with the system to calculate feature descriptors for key points. It is by and A composite computing protocol, and and The security of the data has been proven, therefore the data exchanged between servers is secure and will not leak information about the original image's private content.

[0042] S8: The first cloud server and the second cloud server return the random share of the first feature descriptor vector addition and the random share of the second feature descriptor vector addition to the data owner, and the data owner recovers the real SURF feature descriptor through addition; In one specific embodiment and Will and Return to the data owner . Recovering the true SURF feature descriptor through addition .

[0043] Please refer to Figure 4 , Figure 4 This is the matching result image of the image feature point matching experiment using feature descriptors extracted by the SURF image feature extraction method. The experiment connects the feature points corresponding to the 10 pairs of feature descriptors with the closest Euclidean distance. Figure 4 (a) is the matching result image using the original SURF algorithm. Figure 4 (b) is a matching result diagram using the privacy-preserving SURF method of the present invention.

[0044] It should be noted that this invention employs a secure computing protocol based on additive secret sharing, which runs on two or more servers. The protocol supports direct operations on floating-point numbers without conversion to binary representation, resulting in high efficiency. The secure computing protocol consists of three parties: the image owner... cloud servers and O divides the privacy data and the on-demand generated random number into two additive random shares and sends them separately. , 𝑖= 1,2. Then the two servers collaborate to calculate and generate a random share of the result by addition, and return the random share to . The result can be recovered through simple addition. The following is a brief introduction to the secure computing protocol used in this invention. In the protocol, all values ​​of 𝑖 are either 1 or 2, identifying the two servers. and its additive random shares. For example Private data The additive random share is and ,but Please note that the protocol is described below using only one or a pair of inputs, but in reality, all protocols support simultaneous operations on multiple or more pairs of packaged input data.

[0045] Safe Multiplication Agreement: .for Private data and random numbers ,in C𝑖 holds their additive share In collaborative execution After that, the server get Addition shares .

[0046] Safe division protocol: .for Private data Non-zero random numbers , Hold their additive shares In collaborative execution After that, the server get Addition shares . Security Comparison Protocol: .for Private data 𝑥, 𝑦 and positive random number 𝑟, Hold their additive shares In collaborative execution After that, the server get symbols Specifically, if the server receives 1, then... If the server receives -1, then there is If the server receives 0, then there is .

[0047] Security rational exponentiation protocol: .for Specified real number 𝑎 and private data , hold and Addition shares In collaborative execution After that, the server get Addition shares .

[0048] Secure matrix multiplication protocol: .for Private matrix , and random number matrix ,in , Hold their additive shares In collaborative execution After that, the server get Addition shares .

[0049] Secure matrix inversion protocol: .for Private matrix , hold Addition shares And generate a random number matrix. In collaborative execution After that, the server get Addition shares .

[0050] Security Matrix Determinant Protocol: .for Private matrix , Hold its additive random shares In collaborative execution After that, the server get Additive random share .

[0051] Security Absolute Value Protocol: .for Private data , hold Addition shares In collaborative execution After that, the server get The absolute value of |k| and its additive share .

[0052] In one specific embodiment, the actual effects of the present invention are shown in the table below: Table 1. Matching accuracy of source and target images under different lighting conditions.

[0053] Table 2. Matching accuracy of source and target images with different levels of blur.

[0054] Table 3. Matching accuracy of source and target images from different perspectives.

[0055] Specifically, in this embodiment of the invention, the effectiveness of feature extraction is comprehensively evaluated using the keypoint matching accuracy of image matching experiments. Threshold T The Euclidean distance is defined as the ratio of the Euclidean distance between the descriptor vector in the target image and the descriptor vector in the source image that is closest to it, to the Euclidean distance between the two descriptors. When the actual ratio is less than a threshold T, the keypoint corresponding to the descriptor with the closest Euclidean distance is considered to match the keypoint corresponding to the descriptor in the source image. For example, for the descriptor vector in the source image... In the target image Euclidean nearest descriptor The Euclidean distance is The second closest descriptor The Euclidean distance is ,if Then it is believed and The corresponding keypoints are a pair of matching keypoints. Precision is defined as the ratio of correctly matched keypoints among all matching keypoints.

[0056] Using source and target images with different lighting, different blur levels, and different viewpoints, the accuracy under different thresholds was compared, as detailed in Tables 1, 2, and 3. In the encrypted domain, the matching accuracy of the proposed method is extremely close to that of the original SURF in the plaintext domain under most conditions, satisfying privacy requirements while ensuring descriptor accuracy.

[0057] Example 2 This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment as described in Embodiment 1.

[0058] Example 3 This embodiment provides a computer device, including a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, it implements the steps of the privacy-preserving SURF image feature extraction method based on secret sharing in a cloud environment as described in Embodiment 1.

[0059] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for privacy-preserving SURF image feature extraction based on secret sharing in a cloud environment, characterized in that, The method comprises: dividing the original privacy image into a first additive random share and a second additive random share; sending the first additive random share and the second additive random share to a preset first cloud server and a second cloud server respectively for integral image calculation to obtain corresponding first integral image random share and second integral image random share; based on the first integral image random share and the second integral image random share, constructing a scale space through the first cloud server and the second cloud server, and obtaining a first scale space additive random share and a second scale space additive random share; through the first cloud server and the second cloud server, taking two adjacent points in the respective scale space additive random share as input, cooperating to run a secure comparison protocol, obtaining a comparison result of adjacent points in the scale space, and taking an extreme value point greater than a preset threshold as a candidate key point; through the first cloud server and the second cloud server, cooperating to run a secure matrix inverse protocol and a secure matrix multiplication protocol to perform Taylor interpolation fitting on the candidate key point, and deleting the candidate key point with an offset exceeding a limit to obtain a final key point; through the first cloud server and the second cloud server, cooperating to run a secure division protocol and a secure comparison protocol to securely calculate the gradient angle of a sampling point in the neighborhood of the key point, and then cooperating to run a secure multiplication protocol and a secure comparison protocol to calculate the direction of the key point; through the first cloud server and the second cloud server, cooperating to run a secure absolute value protocol to calculate a multi-dimensional floating-point feature vector as a SURF feature descriptor, and after the calculation, obtaining a first feature descriptor vector additive random share and a second feature descriptor vector additive random share respectively; through the first cloud server and the second cloud server, returning the first feature descriptor vector additive random share and the second feature descriptor vector additive random share to the data owner, and the data owner recovering the real SURF feature descriptor through addition.

2. The secret sharing based privacy preserving SURF image feature extraction method in cloud environment according to claim 1, characterized in that, The method for dividing the original privacy image into a first additive random share and a second additive random share comprises: Data owner In the local end, the original privacy image Is randomly divided into a first additive random share And a second additive random share , satisfying The first additive random share Is a random number matrix with the same size as the original privacy image The second additive random share Is a disturbed image after adding random numbers. 3.The cloud-based secret sharing based privacy-preserving SURF image feature extraction method of claim 1, wherein, The method for sending the first additive random share and the second additive random share to a preset first cloud server and a second cloud server respectively for integral image calculation to obtain corresponding first integral image random share and second integral image random share comprises: first cloud server computing integral image , second cloud server computing integral image ; since , i.e. and are additive shares of , it follows that , i.e. and are integral image shares of , where is the integral image of .

4. The secret sharing based privacy preserving SURF image feature extraction method in cloud environment according to claim 1, characterized in that, The method for constructing a scale space based on the first integral image random share and the second integral image random share through the first cloud server and the second cloud server, and obtaining a first scale space additive random share and a second scale space additive random share comprises: First cloud server and second cloud server in the respective integral image random shares and approximate box filter using different scales of two-dimensional Gaussian convolution kernel, calculate the Hessian matrix of each pixel ; obtain the additive random share of the Hessian matrix , C2obtain the additive random share of the Hessian matrix ; and obtain matrix addition random share and as input, cooperatively run the secure matrix determinant protocol , obtain matrix determinant addition random share obtain , obtain , satisfy wherein is determinant, i.e. scale space; the scale space is a three-dimensional floating-point number space, respectively the horizontal coordinate, the vertical coordinate of the image, is the scale coordinate of the self-box filter; and after cooperatively calculating the matrix determinant of all pixels , obtain the first scale space addition random share and the second scale space addition random share .

5. The secret sharing based privacy preserving SURF image feature extraction method in cloud environment according to claim 1, characterized in that, The method for taking two adjacent points in the respective scale space additive random share as input through the first cloud server and the second cloud server, cooperating to run a secure comparison protocol, obtaining a comparison result of adjacent points in the scale space, and taking an extreme value point greater than a preset threshold as a candidate key point comprises: The first cloud server and the second cloud server Cooperatively run a secure comparison protocol with two adjacent points in each scale space share as input , get the comparison result of adjacent points in scale space; the result is 1, indicating that the former is large, the result is 0, indicating that the two are equal, and the result is -1, indicating that the latter is large; and Cooperatively compare each point in the scale space with its 26 adjacent points to detect local extreme points in the scale space, and select the extreme points greater than a predefined threshold as candidate key points.

6. The secret sharing based privacy preserving SURF image feature extraction method in cloud environment according to claim 1, characterized in that, The method for obtaining the final key point by the first cloud server and the second cloud server cooperating to run a secure matrix inversion protocol and a secure matrix multiplication protocol to perform Taylor interpolation fitting on the candidate key points and deleting the candidate key points with offset exceeding the limit includes: Interpolation fitting is performed on the positions of the candidate key points to obtain the accurate positions of the key points: Let the coordinates of the candidate keypoint be and let the value of the point in the scale space be Let Then the Taylor expansion of the point is Take the derivative of (eq_1) and set the derivative equal to zero to solve for the exact position of the extremum , let the exact position be offset from the discrete position by , and solve for Since the first cloud server and the second cloud server hold additive secret shares of the scale space, they obtain additive secret shares of the matrix and the matrix ; and First, according to formula (eq_2), the cooperative operation of the secure matrix inversion protocol obtains the inverse matrix of The additive secret share of the matrix, then the cooperative operation of the secure matrix multiplication protocol , the additive secret share of the extreme point position offset is calculated, holds , holds ; and Running a safety comparison protocol by cooperation , offset Key points greater than a preset value are considered unstable key points and are deleted.

7. The secret sharing based privacy preserving SURF image feature extraction method in cloud environment according to claim 1, characterized in that, The method for calculating the gradient angle of the sampling points in the neighborhood of the key point by the first cloud server and the second cloud server cooperating to run a secure division protocol and a secure comparison protocol, and then calculating the direction of the key point by cooperating to run a secure multiplication protocol and a secure comparison protocol includes: around each key point on the image uniformly sampled with interval within a circular neighborhood of radius , the scale coordinate of the key point, the direction and direction gradient and , the first cloud server obtains and using a wavelet filter by means of the first integral image share and , the second cloud server obtains and using a wavelet filter by means of the second integral image share and ; wherein and are additive random shares of ; and are additive random shares of ;​ and The gradient shares of two directions are taken as inputs, and a secure division protocol is cooperatively operated , respectively obtaining addition shares of tangent values of gradient directions and Then, a secure comparison protocol is cooperatively operated , the tangent values of gradient directions are compared with tangent values of preset angles, so that the angle interval corresponding to the gradient direction of the sampling point is determined; Within the range of a fan-shaped sliding window is set, the cumulative gradient of the sample points falling within the window is counted direction and direction and , and obtain the additive random share of the cumulative gradient, obtain and , obtain and , the relationship satisfies , ; Each sliding window direction and The cumulative gradient in the direction forms a direction vector. The longest direction vector in all sliding windows will be... The angle between the axes serves as the direction of the key point; and Cooperative Operation of Safe Multiplication Protocol To calculate the square of the direction vector length ; and Obtain the sum of the squared lengths of the direction vector by random addition, and then... Compare the lengths of different direction vectors and obtain the longest direction vector; Finally and Cooperatively running a secure division protocol , respectively, obtain the addition shares of the tangent value of the gradient direction and Then cooperatively running a secure comparison protocol Compare the addition shares of the tangent value of the gradient direction and with the tangent value of the preset angle respectively, so as to determine the angle interval corresponding to the gradient direction of the key point; obtain the interval where the key point direction is located, and then adopt the interval value as the key point direction.

8. The secret sharing based privacy preserving SURF image feature extraction method in cloud environment according to claim 1, characterized in that, The method for calculating a multi-dimensional floating-point number feature vector as a SURF feature descriptor by the first cloud server and the second cloud server cooperating to run a secure absolute value protocol, and obtaining the first feature descriptor vector addition random share and the second feature descriptor vector addition random share after the calculation includes: For each keypoint assigned an orientation, a 64-dimensional floating-point feature vector is computed as the keypoint's feature descriptor; uniform sampling is performed within a square neighborhood around each keypoint in the image, rotated to the same angle as the keypoint's orientation. Each pixel divides the square neighborhood into equal parts. Each sub-region contains [number] sub-regions. 1 sampling point; for each sub-region, calculate the gradient of the sampling point parallel to the direction of the key point. and gradient perpendicular to the key point direction First Cloud Server pass ,use Wavelet filter obtained and Second cloud server pass ,use Wavelet filter obtained and ,in and yes Additive random shares, ; and yes Additive random shares, ; and Cooperatively running a secure absolute value protocol , obtain and absolute values and additive shares of, C1 obtains and , C2 obtains and , wherein and are additive random shares of ; and are additive random shares of ; then accumulate all the sampling points in the sub-region , , and , C1 obtains , C2 obtains ; for 16 sub-regions, each server includes 64 additive random shares of the accumulated sum, and the 64 values are connected in order of sub-region position into a 64-dimensional vector holds the vector , holds the vector ; and and are respectively a first feature descriptor vector additive random share and a second feature descriptor vector additive random share.

9. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the privacy protection SURF image feature extraction method based on secret sharing in the cloud environment as claimed in any one of claims 1 to 8.

10. A computer device, comprising: The computer program is executed by the processor to realize the steps of the privacy protection SURF image feature extraction method based on secret sharing in the cloud environment as claimed in any one of claims 1 to 8.

Images