Face biometric template protection method based on reusable security sketch and electronic device

Abstract

The present application relates to a kind of face biometric template protection method and electronic equipment based on reusable security sketch, method includes the following steps: using the face neural network based on additive angle interval loss function, the output vector of face neural network is obtained and mapped to preset space, as face biometric template;For the preset space, the positive detection rate and the false acceptance rate of alternative security sketch under different binary error correction code are calculated, the selection of security sketch is completed, and the public string corresponding to the face biometric template is obtained by security sketch generation algorithm;The picture to be identified is input into the face neural network, and the output vector is mapped to the preset space, to obtain a new face biometric template, based on the public string, the recovery of face biometric template is realized by security sketch recovery algorithm.Compared with prior art, the present application can realize the face biometric template protection based on security sketch with reusability, and is simple and easy to use.

Description

Technical Field

[0001] This invention relates to the field of biometric feature extraction and recognition technology, and in particular to a method and electronic device for protecting facial biometric templates based on reusable security profiles. Background Technology

[0002] Biometric identification technology, as an identity authentication technology, has been widely used in the identity recognition systems of various organizations and institutions. With the development of artificial intelligence technology, the accuracy and recall rates of biometric identification technologies have significantly improved, greatly enhancing the security and usability of identity verification. However, the arbitrary collection, misuse, and leakage of identity information still occur frequently, greatly threatening the privacy of human biometric information and reducing the security of biometric-based identity recognition systems. How to protect the privacy and security of human biometric information has become an increasingly important and urgent problem to be solved.

[0003] Biometric Template Protection (BTP) technology aims to address the privacy concerns of biometric information. Its goal is to securely store a person's biometric template information while maintaining the highest possible recognition rate.

[0004] Biological template protection technologies targeting neural networks can be divided into two main categories: neural network-based protection algorithms and non-neural network-based protection algorithms. Compared to neural network-based protection algorithms, non-neural network-based protection algorithms offer higher interpretability and security, and have received more attention and application.

[0005] Non-neural network-based protection algorithms can be categorized into three types based on their techniques and underlying assumptions: homomorphic encryption-based, hash-based, and random transformation-based. Among these, homomorphic encryption-based biometric template protection algorithms, due to their reliance on a trusted third party and strong security assumptions, have limited practicality. Random transformation-based protection algorithms, due to their complex design and niche nature, are difficult to analyze effectively, and compared to mature cryptographic algorithms and components, often fail to achieve the security levels anticipated by their designers, raising concerns about their overall security.

[0006] Hash-based protection algorithms leverage the irreversibility of cryptographic hash functions to protect biometric templates. A cryptographic hash algorithm produces a definite hash value from the input, which is computationally irreversible and collision-resistant. Since each extraction of the biometric template is slightly different, an error correction mechanism is needed to obtain a stable output. Protection algorithms using this approach include fuzzy commitments and fuzzy extractors. Both fuzzy commitments and fuzzy extractors are based on a key component: a secure sketch.

[0007] Some face neural networks output in a high-dimensional spherical space, making security frameworks based on binary error-correcting codes and lattice erasure codes unsuitable. This necessitates mapping the output space to the corresponding space, leading to a decrease in recognition performance. While security frameworks based on error-correcting codes in spherical space offer higher recognition performance and security, they are not reusable.

[0008] In summary, there is currently a lack of a method for protecting facial biometric templates to solve or partially solve the aforementioned problems. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a face biometric template protection method and electronic device based on a reusable security framework. It is oriented towards a face neural network with an additive angular interval loss function and realizes face biometric template protection based on a reusable security framework.

[0010] The objective of this invention can be achieved through the following technical solutions:

[0011] One aspect of the present invention provides a method for protecting facial biometric templates based on reusable security profiles, comprising the following steps:

[0012] Using a face neural network based on additive angular interval loss function, the output vector of the face neural network is obtained and mapped to a preset space as a face biometric template;

[0013] For the preset space, the positive detection rate and false acceptance rate of the candidate security profiles under different binary error correction codes are calculated to complete the selection of security profiles. The public string corresponding to the face biometric template is obtained by the security profile generation algorithm.

[0014] The image to be identified is input into the face neural network and the output vector is mapped to the preset space to obtain a new face biometric template. Based on the public string, the face biometric template is restored through a secure summary recovery algorithm.

[0015] As a preferred technical solution, the mapping process of mapping the output vector to a preset space includes:

[0016] The modulo operation is performed on the output vector, and it is mapped to a preset space through the arctangent function.

[0017] As a preferred technical solution, the output vector is mapped to a preset space using the following formula:

[0018]

[0019] in, Let f(·) be the output vector of the face neural network, and let f(·) denote the mapping function, tan -1This represents the arctangent function.

[0020] As a preferred technical solution, the algorithm for generating the security summary is as follows:

[0021]

[0022] Where SS() represents the generation operator, This represents the vector that is first mapped to the preset space, and mod represents the modulo operation. The public string generated from the security summary, It is a binary error-correcting code. n is the dimension of the output vector of the face neural network.

[0023] As a preferred technical solution, the security summary recovery algorithm is as follows:

[0024]

[0025] Where Rec() represents the recovery operator, This represents the vector that is mapped to the preset space a second time. The public string generated by the security summary is represented by DECODE(), which represents the error correction algorithm operator, and mod represents the modulo operation. n is the dimension of the output vector of the face neural network.

[0026] As a preferred technical solution, the error correction algorithm is implemented using the following formula:

[0027]

[0028] Where DECODE() represents the error correction algorithm operator. This represents the vector that is mapped to the preset space a second time. The public string generated from the security summary, It is a binary error-correcting code. n is the dimension of the output vector.

[0029] As a preferred technical solution, during the training process of the face neural network, the vectors of the preset space are embedded into the original high-dimensional spherical space of the face neural network using the following formula:

[0030]

[0031] Where g() is the embedding operator, Vectors in a predefined space.

[0032] As a preferred technical solution, the binary error-correcting code is a linear code.

[0033] In another aspect, an electronic device is provided, comprising: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the aforementioned face biometric template protection method based on a reusable security framework.

[0034] In another aspect, the present invention provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the aforementioned face biometric template protection method based on a reusable security profile.

[0035] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0036] (1) Achieving reusable face biometric template protection based on a security framework: The security framework used in this invention ensures that the linear difference between templates extracted multiple times from the same face does not expose more information about the original face biometric template, thus the security framework is reusable. In practice, analyzing the linear difference between templates extracted multiple times from the same face requires an extremely large number of faces and involves a very large amount of computation, making it difficult to perform. It can be assumed that the linear difference in the preset space will not expose more information about the original face biometric template. Therefore, this invention has updatable / reusable features.

[0037] (2) Simple and easy to use: During the training process of the face neural network, the vectors of the preset space can be embedded into the original high-dimensional spherical space of the face neural network without modifying the specific neural network structure. Only a simple strengthening of the output space of the neural network is needed to continue using the original neural network structure. The changes to the original recognition algorithm are minor, and the recognition positive detection rate and false acceptance rate can be obtained with the same level as the original algorithm. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of a face biometric template protection method based on a reusable security framework, as illustrated in the embodiment. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0040] Example 1

[0041] To address the problems existing in the aforementioned prior art, this embodiment provides a face biometric template protection method based on reusable security profiles. Without modifying the loss function and specific network structure of the face neural network (e.g., ArcFace), a reusable security profile based on binary error correction codes is constructed on the output template space of the face neural network. This maintains recognition efficiency comparable to the original algorithm while ensuring security and reusability.

[0042] The security framework of this method is based on error-correcting codes. It transforms the biometric template onto random codewords, and this transformation is made public. When a user needs to restore their registered biometric template, they must re-extract their biometric template, transform it using the public transformation, correct it using the error-correcting codes to obtain the corresponding codewords, and finally perform an inverse transformation to restore the biometric template. The error-correcting property of the security framework can transform the user's biometric template to a deterministic, error-free output, which can then be used as input for a hash algorithm.

[0043] See Figure 1 This method mainly includes the following steps:

[0044] S1, Generation of the facial biometric protection template. Specifically includes:

[0045] S101, Modify the output template space of the face neural network to complete the training of the face neural network.

[0046] Based on ArcFace or other neural network frameworks that use the additive angular margin loss function, the network output vector is The vector is converted by the modulo operation. Mapped to a higher-dimensional spherical space As a template vector.

[0047] In this method, the mapping space of ArcFace's network output vector is modified by using the arctangent function to transform the output space R. n Mapped to new space As a new template space (i.e., the preset template space), a human face biometric template is obtained. The specific mapping function f(·) is shown below:

[0048]

[0049] Due to space It can be done through functions It is embedded into a high-dimensional spherical space. Therefore, during model training, the network structure and loss function used by ArcFace can be reused to achieve similar convergence results and recognition accuracy.

[0050] During training, the template space is embedded into a high-dimensional spherical space, which can be obtained by performing modulo operations on the output vector of the neural network every two dimensions.

[0051] S102, for the new template space Design and select a security outline.

[0052] In the new template space Based on different fault tolerance distances t as thresholds, appropriate binary error correction codes are selected. A security profile is obtained, and the positive detection rate and false acceptance rate are calculated on the test set based on whether the output of the Rec algorithm is equal to the original input. Based on the calculated positive detection rate and false acceptance rate, and considering the trade-off between security and usability, an appropriate security profile is selected.

[0053] Pick For the new template space The security profile generation algorithm SS for this method is shown below:

[0054]

[0055] in, For [n1,k,t] binary error-correcting codes SS() represents the generator operator. This represents the vector that is first mapped to the preset space, and mod represents the modulo operation. The public string generated from the security summary, n is the dimension of the output vector of the face neural network.

[0056] S2, the process of restoring the human face biometric template.

[0057] The image to be detected is acquired and input into the modified and trained face neural network. The output vector is then mapped to a new template space to obtain a new face biometric template. The original biological template is obtained by using the Rec algorithm for recovering the security summary for error correction:

[0058]

[0059] Where Rec() represents the recovery operator, This represents the vector mapped to the preset space a second time. DECODE() represents the error correction algorithm operator, and mod represents the modulo operation. n is the dimension of the output vector of the face neural network. For [n1,k,t] binary error-correcting codes Where n1 is the input length, t is the error correction distance, and k is the actual length that can be encoded. The binary error-correcting code can be implemented using BCH codes or other types of binary error-correcting codes. The DECODE algorithm is based on binary error-correcting codes. The general error correction algorithm is defined as follows:

[0060]

[0061] The specific implementation of the error correction algorithm can be achieved through enumeration or by using... and The difference is converted into a binary code according to the sign of the corresponding cosine value, and then obtained by calling the general error correction algorithm of the binary error correction code.

[0062] The error-correcting capability of binary error-correcting codes ensures that when vectors are mapped to a preset space twice, the error correction capability is correct. and When they are sufficiently close, the output vector of the recovery algorithm Rec is similar to the original input vector. Equal. This ensures the correctness and usability of the security framework.

[0063] This security outline can provide up to k bits of security. Combined with fuzzy commitments and cryptographic hash functions, this scheme can provide usable security in a computationally incomputable sense.

[0064] During recognition, the similarity between two faces can be obtained by calculating the cosine similarity of the output template in the embedded high-dimensional spherical space.

[0065] This method meets the objectives of protecting biological templates that are irreversible and updatable / reusable, specifically:

[0066] Irreversibility. Irreversibility requires that, given a protected template processed by the algorithm, it is computationally impossible to recover the original template. This means an attacker cannot recover the user's original biometric information within a finite timeframe, thus ensuring the privacy of the user's biometric information.

[0067] Renewability / Reusability. While maintaining irreversibility, protected templates registered through the algorithm should be revocable. Users should be able to re-register and generate protected templates multiple times, and attackers, after obtaining protected templates multiple times, should be unable to recover the original template in a computationally meaningful way. Renewability allows users' biometric templates to be registered and used multiple times, improving the reusability of biometric algorithms.

[0068] The effectiveness of this method is verified below with a specific embodiment.

[0069] Model training:

[0070] Based on the ArcFace face recognition neural network model framework, the network architecture was set to IResNet50 with an output dimension of 512, a batch size of 128, a learning rate of 0.1, an additive angular margin loss function, and a hyperparameter α = 0.5. The model was trained on the MS1M-RetinaFace face dataset on two RTX 4090 GPUs for a total of 10 epochs. Using cosine similarity as the measurement standard, the trained model achieved an accuracy of 99.84% on the LFW test set.

[0071] Choose the appropriate security outline:

[0072] The input to the security profile is set as the output template space of the neural network model.

[0073] Using the LFW face dataset as the test set, the binary error-correcting code for the security profile was selected as the (255, 47, 42) BCH code, resulting in a positive detection rate of 79.8% and a false acceptance rate of 0%. Approximately 47 bits of security were obtained.

[0074] Using the FEI face dataset as the test set, the binary error-correcting code for the security profile was selected as a combination of 128 two-dimensional repeating codes. With a false acceptance rate of 0%, a positive detection rate of 15% was obtained. By allowing the security profile to undergo multiple error correction attempts, the positive detection rate can be improved to 60%; estimation shows that this scheme can achieve approximately 54 bits of security.

[0075] This method has the following advantages:

[0076] (1) Simple to use and high performance. No need to modify the specific neural network structure, only a simple strengthening of the output space of the neural network is required to continue using the original neural network structure. The changes to the original recognition algorithm are minor, and the recognition positive detection rate and false acceptance rate can be obtained with the same as the original algorithm.

[0077] (2) Updatable / Reusable. The security profile used in this method satisfies the linear property when the binary error-correcting code is linear. Assuming that the linear difference between templates extracted multiple times from the same face does not reveal more information about the original template, the security profile is reusable. In practice, analyzing the linear difference between templates extracted multiple times from the same face requires an extremely large number of faces and involves a very large computational load, making it difficult to perform. Therefore, this assumption can be considered valid. Thus, this invention is updatable / reusable.

[0078] Example 2

[0079] This embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the face biometric template protection method based on reusable security profile as described in Embodiment 1.

[0080] Example 3

[0081] This embodiment provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the face biometric template protection method based on a reusable security profile as described in Embodiment 1.

[0082] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for protecting facial biometric templates based on reusable security profiles, characterized in that, Includes the following steps: Using a face neural network based on additive angular interval loss function, the output vector of the face neural network is obtained and mapped to a preset space as a face biometric template; For the preset space, the positive detection rate and false acceptance rate of the candidate security profiles under different binary error correction codes are calculated to complete the selection of security profiles, and the public string corresponding to the face biometric template is obtained by the security profile generation algorithm. The image to be identified is input into the face neural network and the output vector is mapped to the preset space to obtain a new face biometric template. Based on the public string, the face biometric template is restored through a secure summary recovery algorithm. The output vector is mapped to a preset space using the following formula: in, The output vector of the face neural network. Represents a mapping function. Represents the arctangent function; The algorithm for generating the security summary is as follows: in, Indicates the generation operator, This represents the vector that is first mapped to the preset space. This represents the modulo operation. The public string generated from the security summary, It is a binary error-correcting code. , This refers to the dimension of the output vector of the face neural network; During the training process of the face neural network, the vectors of the preset space are embedded into the original high-dimensional spherical space of the face neural network using the following formula: in, For embedding operators, Vectors within a predefined space; The binary error-correcting code is a linear code.

2. The method for protecting facial biometric templates based on reusable security profiles according to claim 1, characterized in that, The algorithm for recovering the security summary is as follows: in, Indicates the recovery operator, This represents the vector that is mapped to the preset space a second time. The public string generated from the security summary, This represents the error correction algorithm operator. This represents the modulo operation. , This represents the dimension of the output vector of the face neural network.

3. The method for protecting facial biometric templates based on reusable security outlines according to claim 2, characterized in that, The error correction algorithm is implemented using the following formula: in, This represents the error correction algorithm operator. This represents the vector that is mapped to the preset space a second time. The public string generated from the security summary, It is a binary error-correcting code. , This represents the dimension of the output vector.

4. An electronic device, characterized in that, include: One or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the face biometric template protection method based on any one of claims 1-3.

5. A computer-readable storage medium, characterized in that, It includes one or more programs that are executed by one or more processors of an electronic device, the one or more programs including instructions for performing the face biometric template protection method based on any one of claims 1-3.

Images