Face recognition method and system based on fully homomorphic encryption

Abstract

The invention discloses a face recognition method and system based on fully homomorphic encryption, and belongs to the field of face recognition, and the method comprises the steps: obtaining face data, encrypting the face data through a fully homomorphic encryption technology, and obtaining encrypted face data; performing parameter optimization on the PolyProtect model on the basis of a Bayesian optimization method to obtain a PolyProtect model with optimal parameters; performing multivariate polynomial transformation on the encrypted face data based on the PolyProtect model of the optimal parameter to obtain face embedding after transformation; and recognizing a human face based on the transformed face embedding. The key parameters of PolyProtect are optimized by considering the characteristics of the FHE, so that the balance between the privacy protection degree and the model precision is realized, and the privacy protection performance and the recognition efficiency of face recognition are improved.

Description

Technical Field

[0001] The present invention belongs to the field of face recognition, and in particular relates to a face recognition method and system based on fully homomorphic encryption. Background Art

[0002] In modern society, face recognition technology has been widely used in many fields such as personalized services, law enforcement, border security and smartphone access. Traditional face recognition systems mainly rely on deep neural networks to extract salient features from face images. These features are represented as embeddings in the latent space and are usually stored as templates in the face recognition system. However, these embedded templates are prone to data leakage risks and can even be used to reconstruct the original face image in some cases, threatening personal privacy.

[0003] To address this challenge, existing technologies have adopted a variety of template protection schemes. Among them, a technology called PolyProtect converts face embeddings into more secure templates by using multivariate polynomials and user-specific parameters. A key parameter in PolyProtect technology is m, which determines the complexity and security of the polynomial transformation. However, while PolyProtect introduces security, it also brings about a trade-off between accuracy and security. Specifically, increasing the overlap parameter (overlap) can improve the accuracy of face recognition, but at the same time it will reduce the security of the template.

[0004] In addition, fully homomorphic encryption (FHE) is a powerful encryption technology that allows arbitrary computations to be performed on encrypted data while ensuring that the computation results can only be decrypted by the authorized party with the private key. In the field of face recognition, FHE has been used to encrypt face embeddings and perform face matching in the encrypted domain to protect the privacy of face templates. However, existing FHE-based face recognition methods usually do not fully consider how to optimize the determination of the parameter m to maximize the privacy protection performance while maintaining recognition accuracy.

[0005] In summary, the existing technologies have made some progress in privacy protection of face recognition, but there are still problems such as incomplete security coverage, insufficient parameter optimization, and trade-off between accuracy and security. Therefore, the present invention proposes a face recognition method and system combining fully homomorphic encryption and PolyProtect, and at the same time achieves a balance between model accuracy and security through parameter optimization, thereby improving the privacy protection performance and recognition efficiency of face recognition. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a face recognition method and system based on fully homomorphic encryption, comprising:

[0007] Acquire facial data, and encrypt the facial data using a fully homomorphic encryption technology to obtain encrypted facial data;

[0008] The parameters of the PolyProtect model are optimized based on the Bayesian optimization method to obtain the PolyProtect model with the best parameters;

[0009] Performing a multivariate polynomial transformation on the encrypted face data based on the PolyProtect model with the optimal parameters to obtain a transformed face embedding;

[0010] The face is recognized based on the transformed facial embedding.

[0011] Preferably, the process of obtaining encrypted facial data includes:

[0012] Perform ciphertext calculation on the face data through the fully homomorphic encryption CKKS scheme to obtain a ciphertext vector;

[0013] The encrypted face data is obtained by summing the elements of the ciphertext vector using a preset algorithm and approximating it using a polynomial of an inverse square root function in the encryption domain.

[0014] Preferably, the process of obtaining the PolyProtect model with optimal parameters comprises:

[0015] Defining the search range of each parameter in the PolyProtect model and defining the evaluation index for each parameter combination;

[0016] The performance of unevaluated parameters is predicted by using Gaussian process as a surrogate model to obtain parameter combinations;

[0017] Evaluate the parameter combination using the evaluation index to obtain the parameter combination with the highest score;

[0018] The parameter combination with the highest score is fed back to the PolyProtect model to obtain the PolyProtect model with the optimal parameters.

[0019] Preferably, each parameter in the PolyProtect model includes: the number of terms in the polynomial, the number of overlapping elements and the range of coefficients.

[0020] Preferably, the expression of the evaluation index is:

[0021] score=privacy_weight×(1-privacy_leak)+efficiency_w eight×(1-normalized_time);

[0022] Among them, privacy_leak represents the classification accuracy of sensitive attributes; normalized_time is the normalized calculation time; privacy_weight and efficiency_weight are the weights of privacy protection and calculation efficiency.

[0023] Preferably, the Gaussian process is used to predict the score distribution of the parameter, and the expression is:

[0024] f(x)~GP(m(x),k(x,x'));

[0025] Among them, m(x) is the mean function, k(x,x') is the covariance function, are the evaluated parameters and the corresponding scores.

[0026] Preferably, the PolyProtect model based on the optimal parameters performs a multivariate polynomial transformation on the encrypted facial data to obtain the transformed facial embedding, which includes: converting the encrypted facial data into a facial embedding vector, mapping the facial embedding vector to a single element in another real feature vector through a coefficient polynomial equation and an exponential polynomial equation, calculating the value at a specific point through the PolyProtect model with the optimal parameters, and generating the transformed facial embedding.

[0027] On the other hand, the present invention also provides a face recognition system based on fully homomorphic encryption, comprising:

[0028] A data acquisition module is used to acquire face data, encrypt the face data by using a fully homomorphic encryption technology, and obtain encrypted face data;

[0029] The optimization module is used to optimize the parameters of the PolyProtect model based on the Bayesian optimization method to obtain the PolyProtect model with the best parameters;

[0030] A transformation module, configured to perform a multivariate polynomial transformation on the encrypted face data based on the PolyProtect model with the optimal parameters to obtain a transformed face embedding;

[0031] A recognition module is used to recognize a human face based on the transformed facial embedding.

[0032] On the other hand, the present invention also provides an electronic device, comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that the face recognition method based on fully homomorphic encryption is implemented when the processor executes the computer program.

[0033] On the other hand, the present invention also provides a computer-readable storage medium, which stores a computer program, and is characterized in that when the computer program is executed by a processor, the face recognition method based on fully homomorphic encryption is implemented.

[0034] Compared with the prior art, the present invention has the following advantages and technical effects:

[0035] In view of the privacy leakage problem existing in the current face recognition system, especially the risk of sensitive information being leaked due to attacks when storing and transmitting facial feature templates, the present invention designs a face recognition method and system combining fully homomorphic encryption and PolyProtect scheme. The method aims to encrypt the facial feature templates through fully homomorphic encryption technology to ensure that the facial recognition operation is performed in an encrypted state, while preventing unauthorized users from obtaining any sensitive information about facial features. In addition, the present invention also constructs an efficient and secure face recognition system through Bayesian optimization. The system can achieve high-precision face recognition function under the premise of protecting personal privacy, and find an optimal balance between security and recognition accuracy, so as to meet the needs of various application scenarios such as personalized services, law enforcement, border security, and smartphone access. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 The figure is a schematic diagram of a method flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0039] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0040] First, the technical terms involved in the following embodiments are explained below:

[0041] (1) Fully Homomorphic Encryption

[0042] Fully Homomorphic Encryption (FHE) is an advanced encryption technology that allows arbitrary calculations to be performed on encrypted data and obtains the same encrypted result as the plaintext calculation result without decrypting the data first. This technology allows the separation of data processing rights and data ownership, thereby achieving flexible data processing while ensuring data security. The core feature of FHE lies in its homomorphism, that is, when the encrypted data is subjected to operations such as addition and multiplication, the result of the operation is consistent with the result of the plaintext after the same operation is performed and then encrypted. This feature makes FHE have broad application prospects in cloud computing, big data analysis and other fields, because it allows users to use the computing power of cloud services for data processing and analysis without decrypting the data. In addition, FHE also has strict theoretical security guarantees and can resist a variety of cryptographic attacks. Therefore, it plays an important role in protecting data privacy and ensuring data security. With the continuous development of technology, FHE is expected to be applied and promoted in more fields.

[0043] (2) PolyProtect Solution

[0044] PolyProtect is a protection scheme for embedded features in face recognition systems. The scheme converts face embedding features into more secure templates through polynomial transformations, which are more difficult to reverse engineer or link to the original face embeddings while retaining enough information for recognition. In the PolyProtect scheme, the face embeddings are first preprocessed to ensure that they are suitable for polynomial transformations. Subsequently, using a carefully designed algorithm, the face embeddings are mapped to a polynomial space, where each embedding corresponds to a polynomial defined by user-specific parameters. These parameters play a crucial role in the scheme. They not only affect the form and complexity of the polynomial, but also directly affect the security and practicality of the final template. Even if an attacker obtains multiple templates protected by PolyProtect, they still find it difficult to restore the original facial information due to the complexity of the polynomial and the diversity of user-specific parameters. This irreversibility and unlinkability are the key to the PolyProtect scheme's ability to significantly improve the privacy protection capabilities of face recognition systems.

[0045] (3) Bayesian Optimization

[0046] Bayesian optimization is an advanced black-box function optimization method. It uses Bayesian statistical principles to efficiently find the global optimal solution under limited computing resources by constructing a probability model of the objective function (usually a Gaussian process model). The core of this method is that it not only relies on the values ​​of the objective function at known points, but also makes full use of the correlation information between these values, as well as the prior knowledge that the objective function may have, such as smoothness and continuity. In each iteration, the Bayesian optimization algorithm selects a point that is most likely to improve the quality of the current optimal solution for evaluation based on the currently constructed probability model. This point is usually called the "collection point" or "next evaluation point". After each evaluation, the proxy model is updated to make its prediction more accurate. In order to select the collection point, the Bayesian optimization algorithm adopts a strategy called "collection function" or "utility function". This function weighs the balance between exploration (looking for better solutions in unknown areas) and utilization (fine search in known high-performance areas), that is, while intensively searching in the most promising areas, it also explores areas that have not yet been evaluated. By continuously iteratively updating the probability model and selecting the collection points, the Bayesian optimization algorithm can gradually approach the global optimal solution. Bayesian optimization uses the previous evaluation results to guide the next search, and can usually find the optimal parameter combination with fewer evaluation times.

[0047] Embodiment 1

[0048] like Figure 1 As shown, this embodiment provides a face recognition method and system based on fully homomorphic encryption, including:

[0049] Fully homomorphic encryption part

[0050] Fully homomorphic encryption (FHE) is an advanced encryption technology. Its core is to allow direct computational operations on encrypted data without prior decryption. This feature allows homomorphic encryption to find a clever balance between data privacy protection and computing requirements. Assume that the public key is pk, the private key is sk, the encryption mechanism is D, and the decryption mechanism is E. If any plaintext x and y are encrypted as x'=E(x,pk) and y'=E(y,pk) respectively, if another operator · can be found such that xοy=D(x'·y',sk) holds, then the cryptographic system is considered to be homomorphic with respect to the selected operator (addition or multiplication), denoted as ο. In other words, FHE operates on the encrypted data, and the result obtained when decrypted using the private key is the same. Given c i =E(x i ,pk), i=1,2,…,K, FHE can calculate c=g(c1,c2,…,c K ) so that D(c,sk)=f(x1,x2,…,x K) is established. This embodiment uses the CKKS scheme in FHE to encrypt face embedding, supports floating-point and complex operations, and introduces unique encoding methods and noise management technologies, which have been incorporated into the implementation of multiple homomorphic encryption libraries, including IBM's HElib, Microsoft's SEAL, and the most comprehensive general lattice cryptographic library PALISADE. The addition and multiplication operations in the PolyProtect scheme can directly use the CKKS scheme for ciphertext calculations. However, the original CKKS scheme cannot be directly applied to the element summation and cosine similarity calculation of the simd ciphertext vector. In order to solve this problem, this embodiment uses Algorithm 1 to sum the elements of the ciphertext vector.

[0051]

[0052]

[0053] Since the calculation of cosine similarity requires an inverse square root function, and FHE has limitations in realizing nonlinear functions, this embodiment uses a polynomial of the inverse square root function to approximate in the encrypted domain. Polynomial approximation is a commonly used numerical calculation method. By selecting a suitable polynomial function, the target function can be approximated with high accuracy within a certain range. In order to obtain a better approximation effect, this embodiment limits the input of the inverse square root function to the range of (0,1). This is because the change of the inverse square root function is relatively gentle within this interval, and the effect of polynomial approximation is better. If the input exceeds this range, the approximation accuracy may drop significantly. The performance of this approximation is measured by the relative error of 2000 random points in the range of (0,1). For each random point, this embodiment calculates the relative error between the polynomial approximation value and the true inverse square root value. By statistically analyzing the relative errors of 2000 random points, this embodiment can comprehensively evaluate the performance of the approximate polynomial. Generally speaking, the smaller the relative error, the higher the accuracy of the approximate polynomial, and when calculating the inverse square root function in the encrypted domain, a result closer to the true value can be obtained.

[0054] On the basis of weighing the computational depth and accuracy, this embodiment selects two polynomials of degree 6 and degree 8 as candidates. Lower-order polynomials (such as degree 6) have shallower computational depths, which means that the computational efficiency in FHE is higher, but the approximate accuracy may be relatively low. Although higher-order polynomials (such as degree 8) have deeper computational depths and relatively lower computational efficiency, they can provide higher approximate accuracy. According to the needs of specific application scenarios, the appropriate polynomial order can be flexibly selected. If the computational efficiency is required to be high, a lower-order polynomial can be selected; if the computational accuracy is required to be high, a higher-order polynomial can be selected. Through this trade-off, this embodiment can maximize the approximate accuracy of the inverse square root function while ensuring a certain computational efficiency.

[0055] PolyProtect Solution Parts:

[0056] The PolyProtect scheme uses a multivariate polynomial with user-specific parameters to transform the face embedding feature values. Given a face embedding vector V = [v1,v2,…,v n ], mapping V to another real eigenvector P = [p1, p2, ..., p k ] (where k < n), P is the PolyProtect template of V. m (m < < n) consecutive elements from V are obtained by the coefficient polynomial equation C = [c1, c2, ..., c m ] and the exponential polynomial equation E=[e1,e2,…,e m ] is mapped to a single element in P, where C and E are user-defined, non-zero, and different for each user of the face recognition system. The first m elements of V are mapped to p1 by formula (1):

[0057]

[0058] Another important user-defined parameter is overlap, which refers to the number of common elements from V used in consecutive values ​​of P. When overlap is 0, V is unique in each set. The minimum and maximum overlap values ​​are 0 and m-1, respectively. The p2 mappings for overlaps of 0 and m-1 are given by equations (2) and (3), respectively:

[0059]

[0060] Once the polynomial P is constructed, PolyProtect can generate templates by computing the value of P at specific points (or sets of points). These points are usually selected based on user-specific parameters (such as overlap) to ensure the diversity and security of the template. For each selected point x i , calculate p i =p(x i ) and use these values ​​as part of the template. In actual implementations, multiple such polynomials may be used to generate a more complex template.

[0061] During the recognition or matching phase, the system calculates a new facial embedding using the same polynomial and parameters and compares the result with the stored template. Due to the complexity of the polynomial transformation, it is difficult for an attacker to reverse engineer the original facial embedding even if they obtain multiple templates. In this way, PolyProtect aims to improve the security of the template and prevent attackers from using leaked templates to reconstruct the original face image.

[0062] However, in the case of a full disclosure threat model, when the attacker has complete knowledge of the PolyProtect method, including its algorithm, user-specific parameters, etc., if the face embedding is not encrypted, the hacker can infer the soft biometric information from the PolyProtect template. Therefore, this embodiment encrypts the face embedding through FHE and then applies the PolyProtect scheme for template protection, which can further ensure that even under the full disclosure threat model, the attacker cannot easily obtain useful facial information.

[0063] Based on the above two descriptions, the system architecture of this embodiment is as follows: Figure 1 As shown, the goal is to ensure that face embeddings do not leak any soft biometric information (including gender, age, and race) to unauthorized users, so that deep learning-based face recognition can proceed safely. The main steps are: first, receive the user's encrypted face embeddings, which are generated by a pre-trained face recognition model and encrypted locally on the user side using the user's proprietary private key through the fully homomorphic encryption (FHE) algorithm; then, a template protection algorithm PolyProtect algorithm is given to enhance the security of the encrypted template and form an additional layer of protection. The resulting encrypted PolyProtect template can be securely stored on the user terminal and used for tasks such as face recognition or extraction of soft biometrics as needed. The results of these tasks are transmitted to the user in encrypted form and can only be decrypted using the user's private key.

[0064] In this process, the parameter settings of PolyProtect will affect the amount of calculation. The main aspects are as follows:

[0065] (1) The influence of m value:

[0066] The m value determines the number of terms in the polynomial. The larger the m value, the higher the complexity of the polynomial operation and the corresponding increase in the amount of calculation. Therefore, it is very important to choose a suitable m value and strike a balance between security and efficiency.

[0067] (2) Impact of overlap:

[0068] Overlap determines the number of overlapping elements between polynomial terms. The larger the overlap, the more original information is retained after the polynomial transformation. While security decreases, the computational complexity will also increase. Therefore, overlap is also a parameter that balances security and efficiency.

[0069] (3) Influence of coefficient range C:

[0070] The range of coefficient values ​​[-C, C] affects the proportion of large number operations in polynomial operations. The larger C is, the larger the numerical range of the polynomial result is. Large number operations require higher calculation accuracy, which leads to more calculation overhead.

[0071] Based on the above analysis, this embodiment designs a detailed scheme based on Bayesian optimization to find the best parameter combination of PolyProtect. The specific steps are as follows:

[0072] Define the search space:

[0073] First, this embodiment needs to define the search range of each parameter. According to the previous analysis, this embodiment can set the following search space:

[0074] m (number of terms in the polynomial): integer, range [2,8];

[0075] Overlap (number of overlapping elements): integer, range [0,4]

[0076] C (coefficient range): integer, range [5,50]

[0077] Define the evaluation metrics:

[0078] Define an evaluation index to measure the performance of each parameter combination. Considering that this embodiment hopes to minimize the computational overhead while ensuring security, a weighted scoring function is designed:

[0079] score=privacy_weight×(1-privacy_leak)+efficiency_weight×(1-normalized_time)(4)

[0080] Among them, privacy_leak represents the classification accuracy of sensitive attributes (such as gender, age, and race), which reflects the degree of privacy leakage; normalized_time is the normalized calculation time, which reflects the calculation efficiency and can be obtained by dividing the actual time by the maximum time in the search space; privacy_weight and efficiency_weight are the weights of privacy protection and calculation efficiency, which can be adjusted according to actual needs. For example, if privacy protection is more important, privacy_weight can be set higher.

[0081] To optimize:

[0082] In the optimization process, this embodiment uses Gaussian Process as a proxy model to predict the performance of unevaluated parameters. Specifically, given a set of evaluated parameters and corresponding scores Gaussian processes can predict the rating distribution for any parameter x:

[0083] f(x)~GP(m(x),k(x,x')) (5)

[0084] Wherein, m(x) is the mean function, and k(x, x') is the covariance function. Generally, this embodiment uses a zero mean function m(x)=0 and a square exponential covariance function:

[0085]

[0086] Among them, σ and l are hyperparameters that control the amplitude and smoothness of the Gaussian process.

[0087] In each round of optimization, this embodiment selects the most promising parameter combination x according to the current Gaussian process model. * This selection process can be achieved by maximizing an acquisition function:

[0088] x * = argmax x a(x) (7)

[0089] Commonly used acquisition functions include expected improvement (EI) and upper confidence bound (UCB). Taking EI as an example, it measures the expected improvement of the evaluation at parameter x:

[0090] EI(x)=E[max(f(x)-f*,0)] (8)

[0091] Among them, f * is the highest score among the currently evaluated parameters. Evaluate the selected parameter combination x * After that, this embodiment adds the result to the existing data points and updates the Gaussian process model. This process is repeated until a preset number of evaluations or a threshold value of performance improvement is reached.

[0092] Finally, this embodiment selects the one with the highest score among all evaluated parameter combinations as the best choice:

[0093]

[0094] The above is a parameter selection method based on Bayesian optimization. This method uses Gaussian process to model the parameter space and captures the correlation between different parameters through mean and covariance. At the same time, the acquisition function guides the optimization process to balance between exploration and utilization and find a good parameter combination within a limited number of evaluations.

[0095] In practical applications, this embodiment can use existing Bayesian optimization libraries, such as GPyOpt and BoTorch, to simplify the implementation of the algorithm. These libraries provide flexible interfaces that allow users to customize search spaces, scoring functions, and acquisition functions, and provide a variety of visualization tools to facilitate the analysis and interpretation of results. Through this Bayesian optimization-based method, this embodiment can systematically and efficiently explore the parameter space of PolyProtect and find the parameter combination that strikes the best balance between privacy protection and computational efficiency. This not only improves the actual performance of PolyProtect, but also provides a general solution for parameter tuning of other similar privacy protection algorithms.

[0096] On the other hand, this embodiment also provides a face recognition system based on fully homomorphic encryption, including:

[0097] A data acquisition module is used to acquire face data, encrypt the face data by using a fully homomorphic encryption technology, and obtain encrypted face data;

[0098] The optimization module is used to optimize the parameters of the PolyProtect model based on the Bayesian optimization method to obtain the PolyProtect model with the best parameters;

[0099] A transformation module, configured to perform a multivariate polynomial transformation on the encrypted face data based on the PolyProtect model with the optimal parameters to obtain a transformed face embedding;

[0100] A recognition module is used to recognize a human face based on the transformed facial embedding.

[0101] On the other hand, this embodiment also provides an electronic device, comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that the processor implements the face recognition method based on fully homomorphic encryption when executing the computer program.

[0102] On the other hand, this embodiment also provides a computer-readable storage medium, which stores a computer program, and is characterized in that when the computer program is executed by a processor, the face recognition method based on fully homomorphic encryption is implemented.

[0103] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A face recognition method based on fully homomorphic encryption, characterized in that: include: Acquire facial data, and encrypt the facial data using a fully homomorphic encryption technology to obtain encrypted facial data; The parameters of the PolyProtect model are optimized based on the Bayesian optimization method to obtain the PolyProtect model with the best parameters; Performing a multivariate polynomial transformation on the encrypted face data based on the PolyProtect model with the optimal parameters to obtain a transformed face embedding; The face is recognized based on the transformed facial embedding.

2. The method according to claim 1, characterized in that The process of obtaining encrypted face data includes: Perform ciphertext calculation on the face data through the fully homomorphic encryption CKKS scheme to obtain a ciphertext vector; The encrypted face data is obtained by summing the elements of the ciphertext vector using a preset algorithm and approximating it using a polynomial of an inverse square root function in the encryption domain.

3. The method according to claim 1, characterized in that The process of obtaining the PolyProtect model with optimal parameters includes: Defining the search range of each parameter in the PolyProtect model and defining the evaluation index for each parameter combination; The performance of unevaluated parameters is predicted by using Gaussian process as a surrogate model to obtain parameter combinations; Evaluate the parameter combination using the evaluation index to obtain the parameter combination with the highest score; The parameter combination with the highest score is fed back to the PolyProtect model to obtain the PolyProtect model with the optimal parameters.

4. The method according to claim 3, characterized in that: Each parameter in the PolyProtect model includes: the number of terms in the polynomial, the number of overlapping elements and the range of coefficients.

5. The method according to claim 3, characterized in that: The expression of the evaluation index is: score=privacy_weight×(1-privacy_leak)+efficiency_w eight×(1-normalized_time); Among them, privacy_leak represents the classification accuracy of sensitive attributes; normalized_time is the normalized calculation time; privacy_weight and efficiency_weight are the weights of privacy protection and calculation efficiency.

6. The method according to claim 3, characterized in that The Gaussian process is used to predict the score distribution of the parameter, and the expression is: f(x)~GP(m(x),k(x,x')); Among them, m(x) is the mean function, k(x,x') is the covariance function, are the evaluated parameters and the corresponding scores.

7. The method according to claim 1, characterized in that The PolyProtect model based on the optimal parameters performs a multivariate polynomial transformation on the encrypted facial data to obtain the transformed facial embedding, which includes: converting the encrypted facial data into a facial embedding vector, mapping the facial embedding vector to a single element in another real feature vector through a coefficient polynomial equation and an exponential polynomial equation, calculating the value at a specific point through the PolyProtect model with the optimal parameters, and generating the transformed facial embedding.

8. A face recognition system based on fully homomorphic encryption, characterized in that: include: A data acquisition module is used to acquire face data, encrypt the face data by using a fully homomorphic encryption technology, and obtain encrypted face data; The optimization module is used to optimize the parameters of the PolyProtect model based on the Bayesian optimization method to obtain the PolyProtect model with the best parameters; A transformation module, configured to perform a multivariate polynomial transformation on the encrypted face data based on the PolyProtect model with the optimal parameters to obtain a transformed face embedding; A recognition module is used to recognize a human face based on the transformed facial embedding.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method described in any one of claims 1 to 7 is implemented.

Images