Unconstrained biological characteristic verification method, system and equipment fusing AdaMTrans and Neure-MS decoding

Abstract

The invention belongs to the field of information security, and particularly relates to an unconstrained biological characteristic verification method, system and device fusing AdaMTrans and Neure-MS decoding. According to the scheme, a novel AdaMTrans encoder is designed to quantify and binarize biological characteristics, and an adaptive random mask scheme is combined, so that the inconsistency between biological characteristic noise and system error correction capability is reduced. According to the scheme, a 5G-LDPC coding scheme in the communication field is further applied, a Neure-MS decoder based on supervised learning is designed, and compared with a non-learning decoder, the scheme has the advantages that the number of iterations is smaller, and the error correction performance is higher, so that the defect that the performance of an existing scheme is reduced is overcome. The scheme of the invention has obvious advantages for identity recognition based on biological characteristics in an unconstrained scene, improves the reliability of biological characteristic recognition, and effectively alleviates the difference between biological characteristic noise and system error correction capability and the privacy security problem caused by inherent connection between the biological characteristics and the user.

Description

Technical Field

[0001] The present invention belongs to the field of information security, and in particular relates to an unconstrained biometric verification method, system and device integrating AdaMTrans and Neural-MS decoding. Background Art

[0002] With the rapid development of biometric technology and the Internet, biometric technology provides reliable identity authentication by measuring unique features such as fingerprints, facial features, and irises. However, directly using biometric measurements as encryption keys still has the problem of high error rates, such as 10%-30% error rates for iris codes and about 30% error rates for facial features. In addition, strict privacy and security policies also pose new challenges to such solutions.

[0003] Fuzzy commitment is a typical representative scheme for securely storing and retrieving biometric measurements in biometric systems. The scheme has been deployed for a variety of biometric modes, including iris, face, fingerprint, etc. However, fuzzy commitment still has obstacles such as performance degradation and noise gap when encountering unconstrained scenarios. Regarding performance degradation, fuzzy schemes are often applied to binary biometric measurements, such as IrisCode, which requires converting real-valued biometrics into binary vectors. This conversion process inherently involves information loss, resulting in performance degradation. Secondly, regarding the noise gap, the gap between the biometric measurement noise level and the ECC fault tolerance is a key issue when applying fuzzy commitments in biometric recognition.

[0004] Ideally, the accuracy of the extracted biometric information can only be ensured if the intra-class variation of the extracted biometrics is zero and the inter-class variation is maximum. However, the inherent noise and variability in existing biometric measurements make it impossible for ECC to achieve the expected level of error correction for biometric information provided by the same user. For example, in the case of binary facial features, intra-class samples typically show an error rate of about 30%. Artificially increasing the error rate by interleaving binary vectors with constant binary vectors will greatly increase the processing cost and is impractical. Summary of the invention

[0005] In order to solve the problem that the error correction performance of biometric information in existing identity authentication schemes under unconstrained conditions is difficult to achieve the expected, the present invention provides an unconstrained biometric verification method, system, and device that integrates AdaMTrans and Neural-MS decoding.

[0006] The technical solution provided by the present invention is:

[0007] An unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding includes the following processes:

[0008] During the registration phase, obtain the biometric feature image x of the registered user, and extract the feature vector v from x through the feature extractor. Input the feature vector v into the AdaMTrans encoder for feature quantization, feature binarization, and random masking to obtain the corresponding binarized vector b. Randomly generate a key k associated with the current user, and use the 5G LDPC encoder to convert the key k into a codeword c. Perform bit-by-bit XOR on the binarized vector b and the codeword c to obtain the secret message δ. Hash the key k to obtain the hash code hash(k), and store [hash(k),δ] as the verification information of the current registered user in the server.

[0009] In the verification phase, the biometric feature image x′ of the verified user is obtained, and it is processed by the feature extractor and the AdaMTrans encoder in turn to obtain the binary vector b′. The secret message δ of each registered user stored in the server is obtained in turn, and the secret message δ and the binary vector b′ are XORed to obtain the codeword c′. The codeword c′ is input into the pre-trained Neural-MS network based on 5GLDPC, and decoded to obtain a recovery key k′. The recovery key k′ is hashed to obtain the hash code hash(k′).

[0010] If the hash code hash(k′) generated based on the biometric feature image input by the verification user and any registered user and secret message δ in the server can match the hash code hash(k) of the registered user, it indicates that the current user identity authentication is successful. If all verification information in the server fails to match successfully, it indicates that the current user identity authentication has failed.

[0011] As a further improvement of the present invention, the biometric feature image x adopts a face image, a finger vein image, a fingerprint, a palm print, an iris, a palm vein image or other imaged biometric features.

[0012] and / or

[0013] The feature extractor uses a neural network model that is pre-trained for the selected biometric feature type.

[0014] As a further improvement of the present invention, the neural network model includes CNN and its variants.

[0015] and / or

[0016] The key k is generated by a PRNG.

[0017] As a further improvement of the present invention, in the AdaMTrans encoder, the feature quantization operation is implemented by a quantization method based on equal probability intervals; including: calculating the probability density of the feature element; dividing the feature space of each feature element into q equal probability intervals marked by integers according to the obtained probability density; mapping each feature vector element to an integer corresponding to the interval to which it belongs;

[0018] The expression of the quantization process is as follows:

[0019]

[0020] In the above formula, z represents the quantization result of the original feature vector v; QLuT represents the quantization function of the equal probability interval.

[0021] As a further improvement of the present invention, in the AdaMTrans encoder, the feature binarization operation uses the following mapping function to map the quantized feature vector z to a binary feature vector preb:

[0022]

[0023] Among them, BLuT is a pre-set permutation function, represents a replacement seed preset for characterizing the mapping relationship between quantized values ​​and binary values; m represents the bit length of the binary item in BLuT, m=q-1.

[0024] As a further improvement of the present invention, in the AdaMTrans encoder, the method for implementing the random mask operation is:

[0025] (1) Define a function r as follows, which is used to randomly generate a binary mask:

[0026]

[0027] In the above formula, u represents a numerical generating function that generates values ​​between -1 and 1; κ is an adjustment coefficient used to adjust the number of bits of the value 1 contained in the function r.

[0028] (2) Perform an AND operation on the binary feature vector preb and the binary mask r to obtain the binary vector b:

[0029] b=preb AND r

[0030] (3) Calculate the intra-class distance and inter-class distance of each binary vector b after masking respectively;

[0031] (4) Taking the probability that the inter-class distance after masking is greater than the threshold τ as a constraint of not less than 95%, minimizing the intra-class distance D as the optimization goal, iteratively search for the optimal κ;

[0032] (5) A binary mask is generated based on the function r after optimizing the κ value and is used to perform a random masking operation on the binary feature vector preb.

[0033] As a further improvement of the present invention, any two eigenvalues ​​b i and b j The masked distance D satisfies the following calculation formula to reduce the intra-class distance:

[0034]

[0035] As a further improvement of the present invention, the 5G LDPC encoder adopts the base code BG2 defined by the 5G standard.

[0036] and / or

[0037] The Neural-MS network includes VN layer and CN layer. The iterative update formula of VN layer is as follows:

[0038]

[0039] In the above formula, Represents the information from VN to CN in the i-th iteration; Represents the information from CN to VN at the i-1th iteration; represents the log-likelihood of the received information; N(v) represents the neighbor nodes of v, which includes the check nodes adjacent to the variable node v.

[0040] The iterative update formula of the CN layer is as follows:

[0041]

[0042] In the above formula, Represents the information from CN to VN in the i-th iteration; Represents the information from VN to CN in the i-th iteration; N(c) represents the neighbor nodes of c, which includes the variable nodes adjacent to the check node c; α and β are two model parameters to be learned.

[0043] During the training process of the Neural-MS network, the gradient descent algorithm is used to update α and β.

[0044] The present invention also provides an unconstrained biometric verification system, which adopts the unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as described above, to provide users with account registration and identity authentication services based on biometric features. The unconstrained biometric verification system includes: a feature extraction module, a first encoding module, an XOR operation module, a key generation unit, an LDPC encoder, a hash module, a storage module, a Neural-MS decoder, and a pairing unit.

[0045] The feature extraction module uses a trained neural network model and is used to extract a feature vector from the input biometric feature image during the registration phase or the verification phase. The first encoding module includes a quantization unit, a binarization unit, and a mask unit. The first encoding module is used to perform feature quantization, feature binarization, and random masking on the feature vector extracted by the feature extraction module in the registration phase or the verification phase, and then obtain the corresponding binarization vector.

[0046] The XOR operation module is used to perform an XOR operation on two input vectors during the registration phase or the verification phase. The key generation unit is used to generate a unique key for each registered user using PANG during the registration phase. The LDPC encoder is used to encode the input key using the base code BG2 defined in the 5G standard during the registration phase to obtain a codeword. The hash module is used to hash the key during the registration phase or the verification phase to obtain the corresponding hash key. The storage module is used to store the secret letters and hash keys of all registered users; and respond to the server's query request during the verification phase. The Neural-MS decoder is used to recover the corresponding key based on the input codeword during the verification phase.

[0047] The pairing unit obtains the hash code of the secret message and codeword of each registered user in the storage module in turn, and then generates a recovery key based on the biometric features of the verified user and the secret message of each registered user, and hashes the recovery key and matches it with the hash key of the registered user. If the pairing is successful, the identity authentication is successful; if all the registration information is not paired successfully, the identity authentication fails.

[0048] The present invention also provides an unconstrained biometric verification device, which includes a server located in the cloud and a biometric feature collection module located in a plurality of terminal devices; the server is deployed with the unconstrained biometric verification system as described above. The terminal device is connected to the server in communication and is used to send the biometric feature image collected by the biometric feature collection module to the server, and the server then provides the user with account registration and identity authentication services based on biometric features.

[0049] The technical solution provided by the present invention has the following beneficial effects:

[0050] 1. Stronger noise suppression performance

[0051] The AdaMTrans designed by the present invention improves accuracy and reliability by converting continuous feature values ​​into binary form, which is more resilient to changes and noise in biometric data. The conversion process ensures that the noise of the feature is continuously mitigated to a level that can be corrected by LDPC, thereby improving the reliability of feature comparison and matching.

[0052] 2. Stronger error correction performance

[0053] The Neural-MS decoder adopted in the present invention is a neural decoding scheme based on supervised learning, which learns from data and adaptively adjusts its decoding strategy, which is crucial for improving the accuracy of key retrieval and verification by dealing with errors introduced during the acquisition and conversion of noisy biometric data.

[0054] 3. Enhanced reliability and security

[0055] The scheme of the present invention adopts a fifth-generation new radio low-density parity check neural decoding scheme (LDPC) based on supervised learning, which enhances the fault tolerance of the instrument system. Due to the optimized parameters, the decoder can achieve more powerful error correction performance with fewer iterations, thereby improving the reliability and security of the instrument system.

[0056] 4. Higher practical value

[0057] The novel biometric recognition system provided by the solution of the present invention has outstanding performance and strong robustness, and is therefore suitable for key areas such as financial services, Internet of Things services and public services. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of the unconstrained biometric authentication method integrating AdaMTrans and Neural-MS decoding provided in Example 1 of the present invention.

[0059] Figure 2 This is a flowchart of the AdaMTrans encoder designed in Example 1 of the present invention to implement feature encoding.

[0060] Figure 3 This is the distribution image of the frame error rate under different cross-rates and iteration rounds in the verification experiment.

[0061] Figure 4 To verify the frame error rates of various schemes in the experiment after 100 iterations at different crossover rates. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0063] Example 1

[0064] This embodiment provides an unconstrained biometric verification method that integrates AdaMTrans and Neural-MS decoding. The method mainly includes two improvements: first, a new encoder named AdaMTrans is designed, which can quantize and binarize the real-valued feature vector extracted from the biometric feature in a pipeline manner for error correction coding; and an adaptive random masking strategy is introduced to keep the bit error rate of the code consistent with the error correction requirements of the scheme, thereby narrowing the noise gap between the extracted biometric feature and the ECC fault tolerance. Second, a neural network decoding scheme based on the Neural-MS network of 5G LDPC, which is borrowed and improved from the communication field, is integrated. The scheme is a learnable decoder that can improve decoding efficiency while ensuring fault tolerance performance and reduce the scheme's requirements for hardware computing power. This allows the scheme provided in this embodiment to be deployed in local terminals or embedded devices with weak computing power, improving the reliability and security of the identity authentication scheme in such scenarios.

[0065] Specifically, Figure 1 As shown, the unconstrained biometric verification method provided in this embodiment includes the following process:

[0066] 1. Registration Phase

[0067] 1. Obtain the biometric feature image x of the registered user and extract the feature vector v from x through the feature extractor.

[0068] The solution provided in this embodiment is applicable to all identity authentication systems that use imaged biometric features for encryption. Biometric features that can be imaged include face images, finger vein images, fingerprints, palm prints, irises, palm vein images, etc. In fact, almost all biometric features that can be put into practical use at this stage can have their feature information converted into image information. And applied to the solution of this embodiment. On this basis, the feature extractor uses a neural network model that is pre-trained for the selected biometric feature type.

[0069] Taking a face image as a biometric feature image as an example, in this embodiment, feature information in the face image can be extracted through a network model based on deep learning. The network model used includes a convolutional neural network (CNN) and its variants, as well as other public non-CNN face feature extraction schemes.

[0070] In the field of deep learning, a key step in image recognition is to convert the image into a real-valued vector, a process also known as image encoding. Taking CNN as an example, we first need to build a pre-trained convolutional neural network model, such as VGG, ResNet, etc. Then the image is input into the model, and after multiple layers of convolution and pooling operations, a one-dimensional vector is finally obtained. This vector is the encoding of the image, which contains all the information of the image. The feature extraction process of the feature extractor is expressed as follows:

[0071] v=f(x)

[0072] In the above formula, x represents the input original biometric feature image; x∈R H×W×L , H, W and L represent the length, width and number of channels of the original biometric feature image respectively. v represents the extracted feature vector, which is a one-dimensional vector. In this embodiment, v∈R 512 .

[0073] It can be seen that the conversion of image information completed by feature extraction is not only the compression of information, but also the refinement of information. Each convolution layer is learning different features of the image, while the pooling layer is retaining the most important information. In this way, we can convert a high-dimensional image into a low-dimensional real-valued vector, which provides a basis for subsequent feature quantization, binarization and other tasks.

[0074] 2. Input the feature vector v into the AdaMTrans encoder for feature quantization, feature binarization and random masking in sequence to obtain the corresponding binarized vector b.

[0075] The real-valued vector of the biometric feature image extracted by the feature extractor must be converted into a binary string suitable for ECC, and information loss must be avoided during the data conversion process to maintain its discrimination ability and accuracy in the binary domain. To address this issue, this embodiment designs a new AdaMTrans encoder to complete this task. The new encoder has the following three advantages: First, it realizes the conversion of feature information, making it suitable for subsequent information encryption. Second, it retains the resolution and accuracy of biometric feature information in the binary domain. Third, for the subsequent encoding and decoding process between biometric feature information and the key, it can narrow the gap between the bit error rate of binary feature information from the same user and the LDPC error correction capability.

[0076] In detail, in order to more clearly introduce the working principle and performance of the AdaMTrans encoder in this embodiment, the following is combined with Figure 2 , describing the data processing process of the AdaMTrans encoder provided in this embodiment:

[0077] 1. Feature Quantization

[0078] In the AdaMTrans encoder, the feature quantization operation is implemented using a quantization method based on equal probability intervals; including:

[0079] A. Calculate the probability density of feature elements.

[0080] B. Divide the feature space of each feature element into q integer-marked equal probability intervals according to the obtained probability density.

[0081] C. Map each eigenvector element to the integer corresponding to the interval to which it belongs.

[0082] The expression of the quantization process is as follows:

[0083]

[0084] In the above formula, z represents the quantization result of the original feature vector v; QLuT represents the quantization function of the equal probability interval.

[0085] like Figure 2 As shown in the left part of the figure, in a typical scheme, the real-valued vector [0.01, -0.1, 0.08] is quantized into the vector [3, 1, 4] after feature quantization.

[0086] (I) Feature Binarization

[0087] In the AdaMTrans encoder, the feature binarization operation uses the following mapping function to map the quantized feature vector z to a binary feature vector preb:

[0088]

[0089] Among them, BLuT is a pre-set permutation function, represents a replacement seed preset for characterizing the mapping relationship between quantized values ​​and binary values; m represents the bit length of the binary item in BLuT, m=q-1.

[0090] Specifically, Figure 2 As shown in the middle part, in a typical scheme of this embodiment, the permutation seed can be set to: "2→000, 1→001, 4→011, 3→111". Under this condition, the permutation function converts the vector [3, 1, 4] into a binary string "111001011".

[0091] (I) Random Mask

[0092] The essence of the mask operation is to perform a bitwise AND operation with the original signal through a binary mask of the same length as the original signal, so that the information on the bits of the original signal corresponding to the binary mask with a value of 1 is retained, and the information on the corresponding bits with a value of 0 is masked. In this embodiment, in order to reduce the intra-class Hamming distance of the feature information of the extracted biometric feature, thereby narrowing the gap between the bit error rate and the error correction capability of the intra-class binary feature; this embodiment finds the most suitable binary mask r by iterative optimization.

[0093] Specifically, in the AdaMTrans encoder, the method of implementing the random mask operation in this embodiment is:

[0094] (i) Define a function r as follows, which is used to randomly generate a binary mask:

[0095]

[0096] In the above formula, u represents a numerical generating function that generates values ​​between -1 and 1; κ is an adjustment coefficient used to adjust the number of bits of the value 1 contained in the function r.

[0097] (ii) Perform an AND operation on the binary feature vector preb and the binary mask r to obtain the binary vector b:

[0098] b=preb AND r

[0099] (iii) Calculate the intra-class distance and inter-class distance of each binary vector b after masking. i and b j The calculation formula of the masked Hamming distance D is as follows:

[0100]

[0101] By analyzing the above formula, it can be found that the Hamming distance between two eigenvalues ​​after masking mainly depends on the binary mask used, which confirms the correctness of the strategy of optimizing the binary mask r in this embodiment.

[0102] (iv) Taking the probability that the masked inter-class distance is greater than the threshold τ as a constraint of not less than 95%, minimizing the intra-class distance D is the optimization goal, and iteratively searching for the optimal κ.

[0103] The mathematical expression of this optimization problem is as follows:

[0104] searchκ

[0105] minD(b i ,b j )

[0106] stP(||preb iXORpreb j ||1-0.5×||r1>τ)≥0.95

[0107] b=prebANDr

[0108] r=sgn(u-κ)

[0109] u∈{0,1} 512×m

[0110] In this embodiment, various existing optimization algorithms (such as genetic algorithms) can be used to solve the above-constructed optimization problem to find the optimized κ value to ensure that the bit b i and b j The masked distance D is minimized.

[0111] (v) Generate a binary mask based on the function r after optimizing the κ value and use it to perform a random masking operation on the binary feature vector preb.

[0112] exist Figure 2 In the typical scheme shown, the optimized binary mask r is: r=011100111. After the binary string "111001011" is masked using the binary mask r, the binarized vector b finally output by the AdaMTrans encoder is b=011000011.

[0113] It should be emphasized that: in the AdaMTrans encoder of this embodiment, r is determined according to the ten-fold cross-validation protocol on each data set. In practical applications, r can be set to be application-specific, that is, all users in the identity authentication system share the same r and allow it to be disclosed to the public. r can also be user-specific, and each user has a unique r. Using user-specific r will improve the reliability of the authentication system.

[0114] 3. Randomly generate a key k associated with the current user and use the 5G LDPC encoder to convert the key k into a codeword c.

[0115] In this embodiment, the key k is generated by a random number generator PANG. In practical applications, each registered user has a unique key k. The key k serves as auxiliary information in the identity authentication system.

[0116] In the transcoding process of the key k, this embodiment directly draws on the LDPC coding method widely used in the communication field, and uses the base code BG2 defined in the 5G standard for information encoding. The base code has 42 rows and 52 columns in the parity check matrix H, contains 197 non-zero elements, and the BG2 code base rate is 0.2. However, considering that in the decoding stage of the key k, the decoder of the original Neural-MS network is designed for the AWGN channel, in order to enable the newly designed 5G LDPC-based Neural-MS decoder to be applied to the binary symmetric channel of binary operations in the identity authentication scenario of this embodiment, this embodiment proposes the following improvements: setting the lifting factor z: z = 10, setting the coding length to N: N = 52 × 10, and the message length to K: K = 10 × 10, this feature length is the most suitable. At the same time, the binary feature b is divided into m blocks, and the block size is 512. Each block is padded with 8 zeros to adapt to the code length of 520. During the encoding process, m codewords will be generated, thereby achieving a key length of 100 × m. In the implementation, m subkeys are retrieved from m codewords and then concatenated to form the final key.

[0117] 4. Perform bit-by-bit XOR on the binary vector b and the codeword c to obtain the secret message δ; perform hash processing on the key k to obtain the hash code hash(k).

[0118] δ=b XOR c

[0119] In this embodiment, the hash processing process can be implemented using any hash algorithm in the encryption field.

[0120] 5. Use [hash(k),δ] as the verification information of the current registered user and store it in the server.

[0121] In the solution of this embodiment, the stored identity authentication information is [hash(k),δ], where δ is the encrypted information containing the key information and the biometric feature; hash(k) is the hash code of the encrypted key information. Therefore, the security of the verification system is very high and can fully protect the user's privacy.

[0122] 2. In the verification stage

[0123] Corresponding to the registration phase, in the verification phase, we first need to obtain the biometric feature image x′ of the verified user, and then process it through the feature extractor and AdaMTrans encoder in turn to obtain the binary vector b′. The feature extractor and AdaMTrans encoder used in the verification phase are the same set of tools used in the registration phase.

[0124] Next, the secret message δ of each registered user stored in the server is obtained in turn, and the secret message δ and the binary vector b′ are XORed to obtain the codeword c′. The codeword c′ is input into the pre-trained Neural-MS network based on 5G LDPC, and decoded to obtain a recovery key k′.

[0125] In this embodiment, the LDPC code is a linear block code defined by a sparse check matrix, which has the excellent property of being able to approach the Shannon limit, is simple to describe, has great flexibility and low error code characteristics, can realize parallel operation, has low decoding complexity, and has great potential for high-speed decoding. LDPC can be described by a sparse parity matrix H, which contains very few parity check matrices. The 1s in each row starting from H specify which bits are associated with the corresponding parity check matrix. H can be used to create a generator matrix G, which decodes the message α using αG mod 2 based on matrix multiplication in the modulo-2 algorithm. LDPC can also be represented as a sparse Tanner graph, which has two groups of nodes: variable nodes (VNs) and check nodes (CNs). Each row of the parity check matrix represents a CN, and each bit in the codeword represents a VN. In the parity check equation, if a bit is involved, the corresponding VN and CN are connected by an edge.

[0126] On this basis, if Figure 1 As shown, some of the connection structures in the decoder of the Neural-MS network are similar to the structure of a neural network for deep learning. Among them, the input layer of the neural network is a vector of size N, where N is the length of the code. Except for the last layer, the size of the remaining layers in the network is E, where E represents the number of edges in the Tanner graph. The hidden layer ID corresponds to the i-th iteration in the decoding process. Each hidden layer contains two sublayers i and j, representing the updates of the VN layer and the CN layer, respectively. The output layer has N neurons. In hidden layer i, sublayer i outputs the message sent by the associated VN layer to the CN layer along the edge, and sublayer i outputs the message sent by the associated CN layer to the VN layer along the edge. The Neural-MS network based on 5G LDPC includes VN layer and CN layer, and the iterative update formula of the VN layer is as follows:

[0127]

[0128] In the above formula, Represents the information from VN to CN in the i-th iteration; Represents the information from CN to VN at the i-1th iteration; represents the log-likelihood of the received information; N(v) represents the neighbor nodes of v, which includes the check nodes adjacent to the variable node v.

[0129] The iterative update formula of the CN layer is as follows:

[0130]

[0131] In the above formula, Represents the information from CN to VN in the i-th iteration; Represents the information from VN to CN in the i-th iteration; N(c) represents the neighbor nodes of c, which includes the variable nodes adjacent to the check node c; α and β are two model parameters to be learned.

[0132] During the training process of the Neural-MS network, the gradient descent algorithm is used to update α and β, and then the model parameters of the trained Neural-MS network are saved and used to decode the recovered codeword c′ to obtain the recovery key k′.

[0133] Finally, the recovery key k′ is hashed to obtain the hash code hash(k′), and then hash(k′) is compared with the hash(k) in the corresponding registered user's verification information [hash(k),δ], and the following decision is made:

[0134] If the hash code hash(k′) generated based on the biometric feature image input by the verification user and any registered user and secret message δ in the server can match the hash code hash(k) stored in the service during the registration phase of the registered user, it indicates that the current user identity authentication is successful.

[0135] If the verification information of the verified user fails to match that of all registered users in the server, it indicates that the current user identity authentication has failed.

[0136] Example 2

[0137] Based on the solution of Example 1, this embodiment further provides an unconstrained biometric verification system, which adopts the unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as in Example 1 to provide users with account registration and identity authentication services based on biometric features. The unconstrained biometric verification system includes: a feature extraction module, a first encoding module, an XOR operation module, a key generation unit, an LDPC encoder, a hash module, a storage module, a Neural-MS decoder, and a pairing unit.

[0138] The feature extraction module uses a trained neural network model and is used to extract a feature vector from the input biometric feature image during the registration phase or the verification phase. The first encoding module includes a quantization unit, a binarization unit, and a mask unit. The first encoding module is used to perform feature quantization, feature binarization, and random masking on the feature vector extracted by the feature extraction module in the registration phase or the verification phase, and then obtain the corresponding binarization vector.

[0139] The XOR operation module is used to perform an XOR operation on two input vectors during the registration phase or the verification phase. The key generation unit is used to generate a unique key for each registered user using PANG during the registration phase. The LDPC encoder is used to encode the input key using the base code BG2 defined in the 5G standard during the registration phase to obtain a codeword. The hash module is used to hash the key during the registration phase or the verification phase to obtain the corresponding hash key. The storage module is used to store the secret letters and hash keys of all registered users; and respond to the server's query request during the verification phase. The Neural-MS decoder is used to recover the corresponding key based on the input codeword during the verification phase.

[0140] The pairing unit obtains the hash code of the secret message and codeword of each registered user in the storage module in turn, and then generates a recovery key based on the biometric features of the verified user and the secret message of each registered user, and hashes the recovery key and matches it with the hash key of the registered user. If the pairing is successful, the identity authentication is successful; if all the registration information is not paired successfully, the identity authentication fails.

[0141] In practical applications, the unconstrained biometric verification system can be deployed in a computer device, which can be an intelligent terminal capable of executing a program, a tablet computer, a laptop computer, a desktop computer, a rack server, a blade server, a tower server or a cabinet server (including an independent server, or a server cluster composed of multiple servers), etc. The computer device indicated in this embodiment includes at least but is not limited to: a memory and a processor that can be interconnected through a system bus. Among them, the memory (i.e., a readable storage medium) includes a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a disk, an optical disk, etc. In some embodiments, the memory can be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory can also be an external storage device of a computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card, etc. Of course, the memory may also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory is usually used to store the operating system and various application software installed on the computer device. In addition, the memory may also be used to temporarily store various data that have been output or are to be output.

[0142] The processor may be a central processing unit (CPU), a graphics processing unit (GPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor is generally used to control the overall operation of a computer device. In this embodiment, the processor is used to run program codes stored in a memory or process data.

[0143] Example 3

[0144] Based on the solution of Example 2, the present invention also provides an unconstrained biometric verification device, which includes a server located in the cloud and a biometric feature collection module located in a plurality of terminal devices; the server is deployed with the unconstrained biometric verification system as described above. The terminal device is connected to the server in communication and is used to send the biometric feature image collected by the biometric feature collection module to the server, and the server provides the user with account registration and identity authentication services based on biometric features.

[0145] Performance Verification

[0146] In order to evaluate the performance of the unconstrained biometric verification method that integrates AdaMTrans and Neural-MS decoding provided in this embodiment, the technicians developed the following experiment, built a corresponding system, and verified the solution.

[0147] 1. Experimental Content

[0148] 1. Experimental environment and configuration

[0149] The hardware configuration of the computer used in this experiment is: Intel Core i9 10900K CPU with 10 cores and 20 threads, paired with NVIDIA GeForce RTX 3090 GPU with 24GB GDDR6X VRAM.

[0150] The hyperparameters involved in the training process that need to be set include the LSSC interval q, the mask threshold t, and the parameters of different decoders. The parameters are systematically analyzed in the experiment to evaluate their impact on the key performance of the solution.

[0151] 2. Dataset

[0152] Six mainstream datasets were selected in the experiment to implement the performance verification of the scheme, including LFW, CALFW, CPLFW, CFPFP, CFPFF and AgeDB-30. These are all unconstrained face datasets. The experimental results based on unconstrained face datasets are helpful to emphasize the feasibility of the method of this embodiment in practical applications.

[0153] Among them, the LFW database is a well-known unconstrained face dataset, containing 13,233 face images from 5,749 identities. CALFW and CPLFW are cross-age and cross-pose databases, respectively. These two databases are used to analyze the robustness of the scheme and the challenges of similar faces to face recognition. The CFP dataset contains the front and side faces of 500 celebrities. This experiment uses two sub-datasets, one is the CFPFF dataset that only compares front faces, and the other is the CFPFP dataset that compares front and side faces. AgeDB is an artificially counted cross-age database. It contains 16,488 images of 568 celebrities of different ages; AgeDB-30 used in this experiment is one of the sub-datasets, which contains identity information with an age difference of 30 years and is also the most challenging dataset.

[0154] In the experiment, CFPFF and CFPFP contain 3500 real comparisons and 3500 impostor comparisons, LFW, CALFW, CPLFW and AgeDB30 contain 3000 real comparisons and 3000 impostor comparisons. In addition, this experiment is also compared on two widely used color face datasets, color FERET and CMU-PIE. Among them, CMU-PIE also belongs to the semi-restricted category, and the database includes 67 objects and 42 types of illumination changes.

[0155] 3. Feature Extractor

[0156] Four popular pre-trained models are used in this experiment, including Arcface (resnet100 trained on ms1mv3), Face ViT (P8S8 trained on ms1m-retinaface with cosface loss function), Adaface (ir101 trained on webface12m), and Magface (ir100 trained on MS1MV2).

[0157] It should be emphasized that the various neural network model-based feature extractors used in this experiment were tested on a disconnected subject dataset, which naturally belongs to an open set and subject-independent recognition setting.

[0158] 4. Control plan

[0159] This experiment also selected some schemes in the public literature for comparison with this case. The references corresponding to the typical control schemes are as follows:

[0160]

[18] C.Rathgeb,J.Merkle,J.Scholz,B.Tams,and V.Nesterowicz,"Deep facefuzzy vault:Implementation and performance,"Computers&Security,vol.113,p.102539,2022.

[0161]

[37] L.Chen,G.Zhao,J.Zhou,A.T.Ho,and L.-M.Cheng,"Face templateprotection using deep ldpc codes learning,"IET Biometrics,vol.8,no.3,pp.190-197,2019.'

[0162]

[55] A.Kumar Jindal,S.Chalamala,and S.Kumar Jami,"Face templateprotection using deep convolutional neural network,"in Proceedingsof the IEEEconference on computer vision and pattern recognitionworkshops,2018,pp.462-470.

[0163]

[56] G.Mai,K.Cao,X.Lan,and P.C.Yuen,"Secureface:Face templateprotection,"IEEE Transactions on Information Forensics and Security,vol.16,pp.262-277,2020.

[0164]

[57] V.N.Boddeti,"Secure face matching using fully homomorphicencryption,"in 2018IEEE 9th International Conference on Biometrics Theory,Applications and Systems(BTAS).IEEE,2018,pp.1-10.

[0165]

[58]

[0166]

[59] K. Zhang, H. Cui, and Y. Yu, "Facial template protection via lattice-based fuzzy extractors," Cryptology ePrint Archive, 2021.

[0167]

[60] BPGilkalaye, A.Rattani, and R.Derakhshani, "Euclidean-DistanceBased Fuzzy Commitment Scheme for Biometric Template Security," in 2019 7th International Workshop on Biometrics and Forensics (1WBF), May 2019, pp.1-6.

[0168] 5. Evaluation indicators

[0169] This experiment uses the absolute keyword reproduction accuracy of the same subject on different face images to measure the keyword retrieval performance of the system. In addition, this experiment combines the commonly used true match rate (GMR) and false match rate (FMR) to further evaluate the performance of the system. In addition, this experiment also quantifies the difference between the distribution of true subject matching scores (matched) and impostor subject matching scores (non-matched). The combination of these metrics provides a comprehensive evaluation that can effectively measure the performance and security of the authentication scheme.

[0170] 2. Experimental Results and Analysis

[0171] Determining the decoding radius of an LDPC code has a complexity equivalent to an NP-Hard problem of determining the minimum; the linear ECC distance - the performance of iterative decoders used for decoding LDPC codes is usually evaluated numerically through comprehensive Monte Carlo simulations. Monte Carlo simulation results statistically approximate the performance metrics of LDPC codes under specific conditions, such as the bit error rate (BER) at a given signal-to-noise ratio (SNR), and can be averaged by performing multiple simulations under different conditions to make a reliable estimate of the error correction capability of the LDPC code.

[0172] 2.1 Decoder Performance

[0173] In this experiment, the performance of SP, MS and the proposed Neural-MS decoder on a binary symmetric channel (BSC) was evaluated. The BSC was subjected to different crossover rates, which represent the probability of bit flipping. The adopted method involves encoding 10,000 randomly generated data frames (10000K) using an LDPC code encoder. After encoding, different levels of noise determined by the crossover rate are introduced into the codewords. After the introduction of noise, the noisy messages are decoded using SP, MS and Neural-MS decoders. The frame error rate (FER) is used in this evaluation, which is the ratio of received messages containing errors to the total number of messages transmitted. Note that a single bit error within a frame may cause the entire frame to be in error, resulting in the FER being generally higher than the bit error rate for specific code and channel conditions. This observation is particularly beneficial for thoroughly evaluating the error correction capabilities of the decoders considered. In addition, it provides a definite upper bound on the decoding errors, which is a critical factor for successful key retrieval.

[0174] Figure 3 The FER at different crossover rates and iterations is shown. Figure 4 The relationship between FER and different crossover rates after 100 iterations is shown. The results show that the decoding performance of Neural-MS is much higher than that of conventional MS decoding, but comparable to SP. Since SP uses the (inverse) hyperbolic tangent function in the decoding process, it is computationally unfriendly, while Neural-MS strikes a balance between computational overhead and accuracy. In addition, as Figure 4 As shown in the figure, in order to achieve a FER less than 5%, the crossover rate of MS is as high as 15.5%, the crossover rate of Neural-MS is 17.62%, and the crossover rate of SP is as high as 18.52%. Therefore, it can be concluded that the intra-class Hamming distance within 18.52% is suitable for LDPC coding. However, the intra-class distance is usually larger than this FER; this explains the inherent reason why the present invention introduces AdaMTrans.

[0175] 2.2 Performance Comparison

[0176] In this experiment, the performance of related secure face recognition systems on constrained face benchmark datasets, including CMU-PIE and Color FERET, and unconstrained datasets, including LFW and CFP, was compared. In the CMU-PIE dataset, 7140 face images of 68 subjects were used, and on the ColorFERET dataset, the left and right side face samples were excluded and the remaining samples were used for evaluation. For the LFW dataset, the standard 6000 pairs of images were used for evaluation.

[0177] Since many early works did not make their codes public or did not provide unified evaluation results on various datasets, direct comparison is challenging. To address this problem, this experiment used four different datasets and compared the results with a variety of existing methods. The experimental results are shown in the following table:

[0178] Table 1: Performance and security of different schemes on constrained and unconstrained datasets

[0179]

[0180] According to the data in the above table, the fuzzy commitment scheme proposed in the present invention outperforms existing methods in both performance and security on constrained and unconstrained data sets. For example, reference

[60] achieved 36.94% FNMR (false rejection rate) at 0.05% FMR (false match rate), while the method in reference

[58] achieved 99.9% top-1 accuracy. In addition, the fuzzy vault scheme in reference

[18] achieved 1% FNMR at 0.01% FMR, and the fuzzy extractor in reference

[59] achieved 70% FNMR at an FMR of 2.1e-7. Compared with these key generation methods, the method proposed in the present invention can not only achieve better results, but also maintain higher security.

[0181] 2.3 Comparison of key retrieval accuracy

[0182] This experiment compares the key retrieval performance of different schemes under different decoding iterations from 1 to 100. The experimental results are shown in the following table:

[0183] Table 2: Comparison of key retrieval accuracy of schemes using different decoders

[0184]

[0185] 1 Overall average across all face models and datasets.

[0186] 2The best result is shown in bold, and the second-best isunderlined.SP is more time-consuming compared to MS and Neural-MS.

[0187] The following points can be observed from the experimental data: 1. As the number of decoding iterations increases, both GMR (true acceptance rate) and FMR (false acceptance rate) increase. 2. When the number of decoding iterations reaches 20 and above, GMR and FMR begin to stabilize. 3. On the LFW dataset, good performance can be obtained regardless of the feature extractor used, while other datasets such as CALFW, AgeDB30, and CFPFP can also achieve an accuracy of about 90%. It should be noted that all benchmark datasets are unconstrained face datasets, so it can be reasonably inferred that the proposed method is suitable for key management in practice. 4. Choosing an appropriate number of decoding iterations can make the system's accuracy suitable for actual deployment. The operator can choose a smaller number of iterations to avoid a higher false acceptance rate. If FMR is not considered, more iterations help the system achieve a higher GMR.

[0188] Furthermore, by comparing different schemes, it can be found that the performance of conventional MS is the worst because it is a hand-crafted decoding method. Due to the learned weights and offset parameters, the performance of the Neural-MS decoder of the present invention is better than that of the conventional MS decoder and is comparable to that of the SP decoder. Both Neural-MS and SP have achieved good performance on LFW. Specifically, the deep features (GMR@OFMR) of Neural-MS on Magface, Arcface, adface and FaceVit can reach 99.17% (99.5%), 99.33% (99.67%), 99.37% (99.70%), and 98.77% (99.27%). The CFP-FF dataset can also achieve similar performance to LFW.

[0189] 3. Experimental Conclusion

[0190] This paper proposes a new unconstrained biometric verification method that integrates AdaMTrans and Neural-MS decoding. It combines facial feature extraction with Neural-MS-based LDPC decoding, and further enhances the system's ability to generate binary features suitable for LDPC coding through the integration of AdamMTransfor coding. The above experiments show that the performance of the scheme of the present invention on unconstrained face datasets is better than that of existing methods, proving its effectiveness in various applications.

[0191] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding, characterized in that: It includes the following process: In the registration phase, the biometric feature image x of the registered user is obtained, and the feature vector v is extracted from x through the feature extractor; the feature vector v is input into the AdaMTrans encoder for feature quantization, feature binarization and random masking in sequence to obtain the corresponding binarized vector b; a key k associated with the current user is randomly generated, and the key k is converted into a codeword c using the 5G LDPC encoder; the binarized vector b and the codeword c are bit-by-bit XORed to obtain the secret message δ; the key k is hashed to obtain the hash code hash(k), and [hash(k),δ] is used as the verification information of the current registered user and stored in the server; In the verification phase, the biometric feature image x′ of the verified user is obtained, and it is processed by the feature extractor and the AdaMTrans encoder in turn to obtain a binary vector b′; the secret message δ of each registered user stored in the server is obtained in turn, and the secret message δ and the binary vector b′ are XORed to obtain the codeword c′; the codeword c′ is input into the pre-trained Neural-MS network based on 5GLDPC, and decoded to obtain a recovery key k′; the recovery key k′ is hashed to obtain the hash code hash(k′); If the hash code hash(k′) generated based on the biometric feature image input by the verification user and any registered user and secret message δ in the server can match the hash code hash(k) of the registered user, it means that the current user identity authentication is successful. If all verification information in the server fails to match successfully, it means that the current user identity authentication has failed.

2. The unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as claimed in claim 1, characterized in that: The biometric feature image x may be a face image, a finger vein image, a fingerprint, a palm print, an iris, a palm vein image or other imaged biometric features; and / or The feature extractor uses a neural network model that is pre-trained for the selected biometric feature type.

3. The unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as claimed in claim 2, characterized in that: The neural network model includes CNN and its variants; and / or The key k is generated by PANG.

4. The unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as claimed in claim 1, characterized in that: In the AdaMTrans encoder, the feature quantization operation is implemented using a quantization method based on equal probability intervals; including: calculating the probability density of feature elements; dividing the feature space of each feature element into q equal probability intervals marked by integers according to the obtained probability density; mapping each feature vector element to the integer corresponding to the interval to which it belongs; the expression of the quantization process is as follows: In the above formula, z represents the quantization result of the original feature vector v; QLuT represents the quantization function of the equal probability interval.

5. The unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as claimed in claim 4, characterized in that: In the AdaMTrans encoder, the feature binarization operation uses the following mapping function to map the quantized feature vector z to a binary feature vector preb: Among them, BLuT is a pre-set permutation function, represents a replacement seed preset for characterizing the mapping relationship between quantized values ​​and binary values; m represents the bit length of the binary item in BLuT, m=q-1.

6. The unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as claimed in claim 5, characterized in that: In the AdaMTrans encoder, the method to implement the random mask operation is: (1) Define a function r as follows, which is used to randomly generate a binary mask: In the above formula, u represents a numerical generating function that generates values ​​between -1 and 1; κ is an adjustment coefficient used to adjust the number of bits of the value 1 contained in the function r; (2) Perform an AND operation on the binary feature vector preb and the binary mask r to obtain the binary vector b: b=preb AND r (3) Calculate the intra-class distance and inter-class distance of each binary vector b after masking respectively; (4) Taking the probability that the inter-class distance after masking is greater than the threshold τ as a constraint of not less than 95%, minimizing the intra-class distance D as the optimization goal, iteratively search for the optimal κ; (5) A binary mask is generated based on the function r after optimizing the κ value and is used to perform a random masking operation on the binary feature vector preb.

7. The unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as claimed in claim 6, characterized in that: Any two eigenvalues ​​b i and b j The calculation formula of the masked distance D is as follows:

8. The unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as claimed in claim 7, characterized in that: The 5G LDPC encoder adopts the base code BG2 defined in the 5G standard; and / or The Neural-MS network includes a VN layer and a CN layer. The iterative update formula of the VN layer is as follows: In the above formula, Represents the information from VN to CN in the i-th iteration; Indicates the information from CN to VN at the i-1th iteration; l v represents the log-likelihood of the received information; N(v) represents the neighbor nodes of v, which includes the check nodes adjacent to the variable node v. The iterative update formula of the CN layer is as follows: In the above formula, Represents the information from CN to VN in the i-th iteration; represents the information from VN to CN in the i-th iteration; N(c) represents the neighbor nodes of c, which includes the variable nodes adjacent to the check node c; α and β are two model parameters to be learned; During the training process of the Neural-MS network, the gradient descent algorithm is used to update α and β.

9. An unconstrained biometric verification system, characterized in that: The unconstrained biometric verification method integrating AdaMTrans and Neural-MS decoding as claimed in any one of claims 1 to 8 is adopted to provide users with account registration and identity authentication services based on biometric features; the unconstrained biometric verification system comprises: A feature extraction module, which uses a trained neural network model and is used to extract feature vectors from an input biometric feature image during an enrollment phase or a verification phase; A first encoding module, comprising a quantization unit, a binarization unit and a mask unit; the first encoding module is used to sequentially perform feature quantization, feature binarization and random masking on the feature vector extracted by the feature extraction module in a registration phase or a verification phase; thereby obtaining a corresponding binarization vector; An XOR operation module, which is used to perform an XOR operation on two input vectors during a registration phase or a verification phase; A key generation unit, which is used to generate a unique key for each registered user using PRNG during the registration phase; LDPC encoder, which is used to encode the input key using the base code BG2 defined in the 5G standard during the registration phase to obtain a codeword; A hash module, which is used to perform hash processing on the key during the registration phase or the verification phase to obtain the corresponding hash key; The storage module is used to store the secret letters and hash keys of all registered users and respond to the query requests of the server during the verification phase; Neural-MS decoder, which is used to recover the corresponding key from the input codeword during the verification phase; A pairing unit obtains the hash code of the secret message and codeword of each registered user in the storage module in turn, and then generates a recovery key according to the biometric features of the verified user and the secret message of each registered user, and hashes the recovery key and matches it with the hash key of the registered user. If the pairing is successful, the identity authentication is passed; if all the registration information is not paired successfully, the identity authentication fails.

10. An unconstrained biometric verification device, characterized in that: It includes a server located in the cloud and a biometric feature collection module located in each terminal device; the server is deployed with the unconstrained biometric feature verification system as described in claim 9; the terminal device is communicated with the server and is used to send the biometric feature image collected by the biometric feature collection module to the server, and the server then provides users with account registration and identity authentication services based on biometric features.

Images