Hadamard transform and improved MinHash-based face template protection method and system
Through Hadamar transformation and improving MinHash algorithm to generate face templates, the contradiction between identification accuracy and security, generation efficiency and security in the prior art is solved, and the effects of high recognition accuracy, high security and high generation efficiency are achieved.
Patent Information
- Application Number
- CN202510643096.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-19
- Publication Date
- 2025-08-12
AI Technical Summary
The existing face template protection methods have contradictions between recognition accuracy and security, and there is contradiction between generation efficiency and security, making it difficult to achieve high recognition accuracy, high security and high generation efficiency at the same time.
Face templates are generated by Hadamar transform and improved MinHash algorithm. Face features are protected through local Hadamar transform and improved MinHash algorithm, low-dimensional feature vectors are generated and copied and expanded into high-dimensional feature vectors. The improved MinHash algorithm is used to generate binary hash codes, and match them with Jaccard similarity calculation.
Two-layer protection for face templates is achieved, recognition accuracy and generation efficiency are improved, the security of face templates is ensured, and contradictions in the existing technology are solved.
Smart Images
Figure CN120472515A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of biometric recognition technology, and in particular to a face template protection method and system based on Hadamard transform and improved MinHash. Background Art
[0002] Existing face template protection methods typically use cryptographic algorithms or irreversible transformation methods to protect the original facial features. Cryptographic algorithms offer high security, but their complex computational process can significantly impact the efficiency of face template generation. Furthermore, due to the precision of cryptographic algorithms and the fuzzy and variable nature of biometric features, face template protection methods based on cryptographic algorithms suffer from low recognition accuracy, making them difficult to implement. Compared to traditional cryptographic algorithms, irreversible transformation methods offer higher generation efficiency. Furthermore, user-specific transformation parameters enable such methods to meet the requirements of revocability and unlinkability. If a face template is leaked or attacked, new transformation parameters can be used to generate a new face template for the user, preventing the misuse of the user's identity. However, irreversible transformation methods inevitably lose some original facial feature information when transforming facial features from the original feature domain to the secure, protected domain, impacting the recognition accuracy of the face template.
[0003] Therefore, the existing face template protection methods have the contradiction between recognition accuracy and security, and the contradiction between generation efficiency and security. How to design a face template protection method with high security, high recognition accuracy and high generation efficiency is a problem that needs to be solved urgently. Summary of the Invention
[0004] To this end, the present invention provides a face template protection method and system based on Hadamard transform and improved MinHash, which solve the contradiction between recognition accuracy and security, and the contradiction between generation efficiency and security in existing face template protection methods.
[0005] According to the design scheme provided by the present invention, on the one hand, a face template protection method based on Hadamard transform and improved MinHash is provided, comprising:
[0006] Collecting a facial image of a registered user, generating a facial template of the registered user using a facial template protection method, and storing the facial template of the registered user in a protected facial template database, wherein the facial template protection method generates the facial template by extracting facial image features, performing a local Hadamard transform, and using an improved MinHash algorithm;
[0007] Collect the facial image of the user to be authenticated, generate the target facial template of the user to be authenticated through the facial template protection method, and perform similarity matching between the target facial template and the registered user facial template in the protected facial template database to determine whether the facial image of the user to be authenticated is a registered user facial image.
[0008] As the face template protection method based on Hadamard transform and improved MinHash of the present invention, further, generating a face template through the face template protection method includes:
[0009] Extract facial image features and form facial feature vectors;
[0010] Performing a local Hadamard transform on the facial feature vector using a local Hadamard matrix to transform the facial feature vector into a protected low-dimensional feature vector, and copying and expanding the low-dimensional feature vector into a high-dimensional feature vector;
[0011] The high-dimensional feature vector is input into the improved MinHash algorithm, and the improved MinHash algorithm is used to generate a binary hash code of the high-dimensional feature vector, so as to obtain the corresponding face template based on the binary hash code.
[0012] As the face template protection method based on Hadamard transform and improved MinHash of the present invention, further, extracting face image features includes:
[0013] Align and crop the collected facial images to obtain facial images of fixed size;
[0014] A fixed-size face image is input into a deep neural network model, and the deep neural network model is used to extract facial features of the face image and generate a facial feature vector. The deep neural network model is obtained by training the face recognition network ArcFace using a face image dataset based on the additive angular margin loss function. The additive angular margin loss function guides network training optimization by maximizing the inter-class distance and minimizing the intra-class distance in the feature space.
[0015] As the face template protection method based on Hadamard transform and improved MinHash of the present invention, further, the local Hadamard transform includes:
[0016] Generate a Hadamard matrix of dimension N by a recursive algorithm, where N is a power of 2;
[0017] Randomly extract r rows from the Hadamard matrix to obtain the local Hadamard matrix;
[0018] Perform matrix multiplication on the local Hadamard matrix and the facial eigenvector to obtain the eigenvector after local Hadamard transformation.
[0019] As a face template protection method based on Hadamard transform and improved MinHash of the present invention, further, the face template is obtained by using the improved MinHash algorithm, including:
[0020] The eigenvector obtained by local Hadamard transform is replicated n times to obtain a high-dimensional eigenvector;
[0021] Performing a permutation operation on the high-dimensional feature vector by randomly permuting a seed to obtain a permuted feature vector, wherein the permuted feature vector is composed of the high-dimensional feature vector pointed to by the element at the position corresponding to the random permutation seed, and the random permutation seed is composed of the element index value of the high-dimensional feature vector;
[0022] Flip the permuted eigenvector to obtain a flipped eigenvector;
[0023] A binary hash code representing the face template is generated by comparing the sizes of the elements at the same position of the permuted feature vector and the flipped feature vector.
[0024] As the face template protection method based on Hadamard transform and improved MinHash of the present invention, further, comparing the sizes of the elements at the same position of the permuted feature vector and the flipped feature vector, including:
[0025] The elements at the same position of the permuted feature vector and the flipped feature vector are grouped and represented, and the index value of the element position in each group is marked, so as to generate a binary hash code by extracting the index value of the minimum element in each group.
[0026] As the face template protection method based on Hadamard transform and improved MinHash of the present invention, further, similarity matching is performed on the target face template with the registered user face template in the protected face template database, including:
[0027] The similarity between the target face template and the registered user face template in the protected face template database is calculated by calculating the total number of hash code values at the same position between the face templates based on the Jaccard similarity;
[0028] Determine whether the target face template matches the registered user face template in the protected face template database based on the similarity.
[0029] On the other hand, the present invention also provides a face template protection system based on Hadamard transform and improved MinHash, comprising: a registration module and a recognition module, wherein:
[0030] A registration module is configured to collect a facial image of a registered user, generate a facial template of the registered user using a facial template protection method, and store the facial template of the registered user in a protected facial template database. The facial template protection method generates the facial template by extracting facial image features, performing a local Hadamard transform, and employing an improved MinHash algorithm.
[0031] The recognition module is used to collect the facial image of the user to be authenticated, generate the target facial template of the user to be authenticated through the facial template protection method, and perform similarity matching between the target facial template and the registered user facial template in the protected facial template database to determine whether the facial image of the user to be authenticated is a facial image of a registered user.
[0032] Beneficial effects of the present invention:
[0033] The present invention implements a two-layer protection for facial templates used in face recognition based on the Hadamard transform and an improved MinHash algorithm, offering high security. Both protection methods effectively improve the recognition accuracy and generation efficiency of facial templates. First, the local Hadamard transform not only has low computational complexity and is easy to implement, but also exhibits high similarity preservation. Therefore, the local Hadamard transform can ensure both the recognition accuracy and generation efficiency of facial templates. Furthermore, the local Hadamard matrix used in the local Hadamard transform is a rank-deficient matrix, so its matrix multiplication with the original facial features is a one-to-many mapping. Even if the local Hadamard matrix is leaked, the original facial features cannot be deduced from the numerous solutions. Therefore, the local Hadamard transform provides a first layer of protection for the original facial features. Furthermore, the improved MinHash algorithm comprehensively extracts the size relationships between elements in the facial features, resulting in a richer set of extracted features, effectively ensuring the recognition accuracy of facial templates. Furthermore, the improved MinHash algorithm only requires a single hash function calculation to generate the final binary facial template, significantly improving the efficiency of facial template generation. The improved MinHash algorithm nonlinearly maps facial features into binary hash codes, further hiding the original facial features and providing a second layer of protection. Therefore, this face template protection method based on the Hadamard transform and improved MinHash improves the recognition accuracy and generation efficiency of face templates while ensuring the security of face templates. This resolves the contradictions between recognition accuracy and security, as well as between generation efficiency and security, that exist in existing face template protection methods. It has promising application prospects in the field of face recognition authentication and authorization. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 This is a flowchart of the face template protection method based on Hadamard transform and improved MinHash in the embodiment;
[0035] Figure 2 This is a schematic diagram of the overall framework of the face template protection method based on Hadamard transform and improved MinHash in the embodiment;
[0036] Figure 3 Schematic diagram of a face template generated based on the improved MinHash algorithm when n=2 and r=6 in the embodiment. DETAILED DESCRIPTION
[0037] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below with reference to the accompanying drawings and technical solutions.
[0038] In view of the contradiction between recognition accuracy and security, and between generation efficiency and security in existing face template protection methods, the embodiments of the present invention, such as Figure 1 As shown in the figure, a face template protection method based on Hadamard transform and improved MinHash is provided, which specifically includes the following contents:
[0039] S101. Collect a facial image of a registered user, generate a facial template of the registered user using a facial template protection method, and store the facial template of the registered user in a protected facial template database, wherein the facial template protection method generates the facial template by facial image feature extraction, local Hadamard transform, and improved MinHash algorithm;
[0040] S102. Collect a facial image of the user to be authenticated, generate a target facial template of the user to be authenticated through a facial template protection method, and perform similarity matching between the target facial template and the registered user facial template in the protected facial template database to determine whether the facial image of the user to be authenticated is a registered user facial image.
[0041] like Figure 2 In the illustrated method architecture, facial images are first collected for user registration and authentication, and facial feature vectors are extracted from these images using a pre-trained deep neural network model. Next, a local Hadamard transform is used to irreversibly transform the high-dimensional facial feature vector into a protected low-dimensional feature vector, which is then replicated and expanded into a high-dimensional feature vector. Finally, a binary hash code (i.e., the final facial template) is generated from the replicated and expanded high-dimensional feature vector using an improved MinHash algorithm. A successful match between the registration and query face images is determined by calculating the Jaccard similarity between the registration and query templates.
[0042] The face template protection method is used to generate a face template, which can be designed to include:
[0043] Extract facial image features and form facial feature vectors;
[0044] Performing a local Hadamard transform on the facial feature vector using a local Hadamard matrix to transform the facial feature vector into a protected low-dimensional feature vector, and copying and expanding the low-dimensional feature vector into a high-dimensional feature vector;
[0045] The high-dimensional feature vector is input into the improved MinHash algorithm, and the improved MinHash algorithm is used to generate a binary hash code of the high-dimensional feature vector, so as to obtain the corresponding face template based on the binary hash code.
[0046] Specifically, extracting facial image features may include:
[0047] Align and crop the collected facial images to obtain facial images of fixed size;
[0048] A fixed-size face image is input into a deep neural network model, and the deep neural network model is used to extract facial features of the face image and generate a facial feature vector. The deep neural network model is obtained by training the face recognition network ArcFace using a face image dataset based on the additive angular margin loss function. The additive angular margin loss function guides network training optimization by maximizing the inter-class distance and minimizing the intra-class distance in the feature space.
[0049] The user's face image is collected through a camera or other device, and the deep neural network model ArcFace is used to extract facial feature vectors from the face image. First, all face images are aligned and cropped to 112×112 size by MTCNN; then, the ArcFace model pre-trained on the MS-Celeb-1M dataset is used to extract 512-dimensional facial feature vectors. ArcFace is a deep learning-based face recognition model that uses a novel additive angular margin loss function to improve the model's discriminative ability by increasing the angular margin. The core idea of ArcFace is to maximize the inter-class distance and minimize the intra-class distance in feature space, thereby learning more discriminative facial features.
[0050] Specifically, the local Hadamard transform can be designed to include:
[0051] Generate a Hadamard matrix of dimension N by a recursive algorithm, where N is a power of 2;
[0052] Randomly extract r rows from the Hadamard matrix to obtain the local Hadamard matrix;
[0053] Perform matrix multiplication on the local Hadamard matrix and the facial eigenvector to obtain the eigenvector after local Hadamard transformation.
[0054] A Hadamard matrix of dimension N (N is a power of 2) can be generated recursively: in represents the Kronecker product, and
[0055] Local Hadamard Matrix It is obtained by randomly extracting r rows from the Hadamard matrix H. First, a row vector A=[a1,a2,...,a r ], where r<N. Each element a in the row vector A i ∈{1,2,...,N-1,N} are all positive integers, and each element is different from each other, which represents the row index of the Hadamard matrix H. Therefore, a local Hadamard matrix with a dimension of r×N can be generated by a random row vector A
[0056] After generating the local Hadamard matrix, the local Hadamard matrix Original facial features Perform matrix multiplication operations,
[0057]
[0058] That is, the face feature dimension extracted by Arcface model is 512. Through matrix multiplication operation, the generated Hadamard matrix needs to be 512×512, and the dimension of the local Hadamard matrix should be r×512. Therefore, the transformed features generated by the local Hadamard transform of the original face features are The dimension is r.
[0059] Specifically, the improved MinHash algorithm is used to obtain the face template, which can be designed to include:
[0060] The eigenvector obtained by local Hadamard transform is replicated n times to obtain a high-dimensional eigenvector;
[0061] Performing a permutation operation on the high-dimensional feature vector by randomly permuting a seed to obtain a permuted feature vector, wherein the permuted feature vector is composed of the high-dimensional feature vector pointed to by the element at the position corresponding to the random permutation seed, and the random permutation seed is composed of the element index value in the high-dimensional feature vector;
[0062] Flip the permuted eigenvector to obtain a flipped eigenvector;
[0063] A binary hash code representing the face template is generated by comparing the sizes of the elements at the same position of the permuted feature vector and the flipped feature vector.
[0064] For low-dimensional features Will Copying n times can get high-dimensional features with dimension nr Contains more abundant original facial feature information, which is crucial to ensure the recognition accuracy of this method. Randomly permuting the dimension of the seed π and The same is generated by randomly arranging non-repeating integers from 1 to nr, representing The index value of the element in . The features after random permutation It is pointed to by the element at the corresponding position of π The elements in .
[0065] Will Flip it to get Right now The original MinHash algorithm can only generate a 1-bit hash code through a random permutation operation. In this embodiment, the original MinHash algorithm is improved. and The size relationship of elements at the same position generates a binary hash code. The specific description of the algorithm can be summarized as follows: First, and Elements at the same position are grouped together Each group The position index values of the two elements in are recorded as 0 and 1 respectively. Then, take each group The index value of the smallest element in is used as the hash code in represents the minimum element index in each group. This demonstrates that the improved MinHash algorithm can generate nr-bit hash codes through a single random permutation operation, which is crucial for improving the efficiency of face template generation. Furthermore, the improved MinHash algorithm comprehensively exploits the size relationships between feature elements after transformation, resulting in richer extracted features and effectively improving face template recognition accuracy. Figure 3 An example of generating a face template based on the improved MinHash algorithm when n=2 and r=6 is shown.
[0066] Specifically, similarity matching is performed between the target face template and the registered user face template in the protected face template database, which can be designed to include:
[0067] The similarity between the target face template and the registered user face template in the protected face template database is calculated by calculating the total number of hash code values at the same position between the face templates based on the Jaccard similarity;
[0068] Determine whether the target face template matches the registered user face template in the protected face template database based on the similarity.
[0069] By calculating the similarity between the registration template and the query template generated by the user during the registration and authentication stages, it can be determined whether the user can pass the authentication. e and query template T q The specific calculation formula can be expressed as follows:
[0070]
[0071] Where nr represents T e and T q The length of XNOR(T e ,T q ) represents the exclusive OR operation, when T e and T q If the hash code values of the corresponding positions are the same, 1 is output, and if they are different, 0 is output. The calculation is T e and T q The total number of hash code values that are the same at the same location. S(T e ,T q )∈[0,1], the closer its value is to 1, the more T e and T q The higher the similarity, the higher the value, and vice versa.
[0072] Furthermore, based on the above method, an embodiment of the present invention also provides a face template protection system based on Hadamard transform and improved MinHash, comprising: a registration module and a recognition module, wherein:
[0073] A registration module, configured to collect a facial image of a registered user, generate a facial template of the registered user using a facial template protection method, and store the facial template of the registered user in a protected facial template database. The facial template protection method generates the facial template by extracting facial image features, performing a local Hadamard transform, and employing an improved MinHash algorithm.
[0074] The recognition module is used to collect the facial image of the user to be authenticated, generate the target facial template of the user to be authenticated through the facial template protection method, and perform similarity matching between the target facial template and the registered user facial template in the protected facial template database to determine whether the facial image of the user to be authenticated is a facial image of a registered user.
[0075] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.
[0077] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.
[0078] Those skilled in the art will appreciate that all or part of the steps in the above method can be performed by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a magnetic disk, or an optical disk. Alternatively, all or part of the steps in the above embodiment can be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or software functional modules. The present invention is not limited to any specific combination of hardware and software.
[0079] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.
Claims
1. A face template protection method based on Hadamard transform and improved MinHash, characterized in that: Include: Collecting a facial image of a registered user, generating a facial template of the registered user using a facial template protection method, and storing the facial template of the registered user in a protected facial template database, wherein the facial template protection method generates the facial template by extracting facial image features, performing a local Hadamard transform, and using an improved MinHash algorithm; Collect the facial image of the user to be authenticated, generate the target facial template of the user to be authenticated through the facial template protection method, and perform similarity matching between the target facial template and the registered user facial template in the protected facial template database to determine whether the facial image of the user to be authenticated is a registered user facial image.
2. The face template protection method based on Hadamard transform and improved MinHash according to claim 1, characterized in that: Generate a face template through the face template protection method, including: Extract facial image features and form facial feature vectors; Performing a local Hadamard transform on the facial feature vector using a local Hadamard matrix to transform the facial feature vector into a protected low-dimensional feature vector, and copying and expanding the low-dimensional feature vector into a high-dimensional feature vector; The high-dimensional feature vector is input into the improved MinHash algorithm, and the improved MinHash algorithm is used to generate a binary hash code of the high-dimensional feature vector, so as to obtain the corresponding face template based on the binary hash code.
3. The face template protection method based on Hadamard transform and improved MinHash according to claim 1 or 2, characterized in that: Extract facial image features, including: Align and crop the collected facial images to obtain facial images of fixed size; A fixed-size face image is input into a deep neural network model, and the deep neural network model is used to extract facial features of the face image and generate a facial feature vector. The deep neural network model is obtained by training the face recognition network ArcFace using a face image dataset based on the additive angular margin loss function. The additive angular margin loss function guides network training optimization by maximizing the inter-class distance and minimizing the intra-class distance in the feature space.
4. The face template protection method based on Hadamard transform and improved MinHash according to claim 1 or 2, characterized in that: Local Hadamard transform, including: Generate a Hadamard matrix of dimension N by a recursive algorithm, where N is a power of 2; Randomly extract r rows from the Hadamard matrix to obtain the local Hadamard matrix; Perform matrix multiplication on the local Hadamard matrix and the facial eigenvector to obtain the eigenvector after local Hadamard transformation.
5. The face template protection method based on Hadamard transform and improved MinHash according to claim 1 or 2, characterized in that: The improved MinHash algorithm is used to generate face templates, including: The eigenvector obtained by local Hadamard transform is replicated n times to obtain a high-dimensional eigenvector; Performing a permutation operation on the high-dimensional feature vector by randomly permuting a seed to obtain a permuted feature vector, wherein the permuted feature vector is composed of the high-dimensional feature vector pointed to by the element at the position corresponding to the random permutation seed, and the random permutation seed is composed of the element index value in the high-dimensional feature vector; Flip the permuted eigenvector to obtain a flipped eigenvector; A binary hash code representing the face template is generated by comparing the sizes of the elements at the same position of the permuted feature vector and the flipped feature vector.
6. The face template protection method based on Hadamard transform and improved MinHash according to claim 5, characterized in that: Compare the size of the elements at the same position of the permuted eigenvector and the flipped eigenvector, including: The elements at the same position of the permuted feature vector and the flipped feature vector are grouped and represented, and the index value of the element position in each group is marked, so as to generate a binary hash code by extracting the index value of the minimum element in each group.
7. The face template protection method based on Hadamard transform and improved MinHash according to claim 1, characterized in that: Perform similarity matching between the target face template and the registered user face template in the protected face template database, including: The similarity between the target face template and the registered user face template in the protected face template database is calculated by calculating the total number of hash code values at the same position between the face templates based on the Jaccard similarity; Determine whether the target face template matches the registered user face template in the protected face template database based on the similarity.
8. A face template protection system based on Hadamard transform and improved MinHash, characterized in that: Contains: registration module and identification module, among which, A registration module, configured to collect a facial image of a registered user, generate a facial template of the registered user using a facial template protection method, and store the facial template of the registered user in a protected facial template database. The facial template protection method generates the facial template by extracting facial image features, performing a local Hadamard transform, and employing an improved MinHash algorithm. The recognition module is used to collect the facial image of the user to be authenticated, generate the target facial template of the user to be authenticated through the facial template protection method, and perform similarity matching between the target facial template and the registered user facial template in the protected facial template database to determine whether the facial image of the user to be authenticated is a facial image of a registered user.
9. An electronic device, characterized in that: include: at least one processor, and a memory coupled to the at least one processor; The memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that The computer program stored in the computer-readable storage medium can implement the method according to any one of claims 1 to 7 when the computer program is executed.