Deep biological hash network model for vein biological feature recognition and recognition method

Abstract

The invention discloses a deep biological Hash network model for vein biological feature recognition and a recognition method, at least comprising a class center alignment module and a deep biological Hash layer, the center alignment module is composed of a two-stage positioning network, a grid generator, a sampler and a deep super template, predicting all potential changes from finger vein views to finger center views through learning transformation parameters, and realizing feature alignment; the deep biological hash layer generates a protected pseudo hash code by using a system-level token and a hash function, a loss function of the deep biological hash network model in a model learning training process is a mixed loss function, and comprises a classification loss item, a consistency basic positioning loss item and a class center triple loss item; the recognition performance, the logout attribute, the safety and the privacy of the logout biological recognition scheme are jointly ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of biometrics, and mainly relates to a deep biometric hashing network model and an identification method for finger vein biometric identification. Background Art

[0002] Finger vein recognition, as a representative of the second-generation biometric technology, is widely used in the fields of finance, security, medical treatment, etc. due to its high security and stability. With the popularization of biometric applications, the need to protect the biometric data used in the system is increasing. In biometrics, the problem of information leakage seriously affects the privacy and anonymity of users. Although biometric features are more difficult to forge than traditional authentication technologies, once the uniqueness of these features is leaked or stolen, it may lead to the permanent loss of identity, thus restricting the biometric options of individuals.

[0003] Revocable biometric technology allows biometric features to be repeatedly reissued and revoked like traditional keys. Revocable means performing an intentional and repeatable distortion transformation operation on the biometric signal, so that the biometric template can be authenticated in the encrypted domain, thereby realizing biometric protection. According to the ISO / IEC 24745 standard, a biometric template protection scheme must satisfy: (1) Irreversibility: Given a protected template, it should be impossible to reconstruct the original biometric sample; (2) Reversibility: Multiple protected templates can be generated from a given biometric sample; (3) Unlinkability: If two protected templates are generated from the same biometric information and stored in different systems, it should be impossible to determine that they belong to the same object; (4) Performance: Using the BTP scheme should not significantly reduce the recognition performance of the system. In addition, the recognition performance should not be sensitive to the parameters specifying the template protection steps adopted.

[0004] Biohashing is a well-known example of a general cancellable biometric based on salt values, which relies on a heuristic distance-preserving hash function to generate a hash template. The hash template generated from the biometric feature can effectively maintain the similarity structure of the biometric feature and minimize the similarity difference between the original space and the hash space.

[0005] However, the high performance of the biohashing algorithm is based on the assumption that the tokenized random numbers will not be stolen. When these random numbers are stolen, its performance will be worse than the case of only using the original biometric feature. When designing cancellable biometrics, hashing provides irreversibility, the user's random numbers provide revocability and unlinkability, and the distance-preserving property of the hash function maintains the recognition performance.

[0006] The finger vein recognition technology faces several key problems, including the risk of permanent loss of identity due to the leakage of biometric data, especially in the case of biometric hashing technology, where there is a hidden danger of significant performance degradation when the tokenized random numbers are stolen. In addition, projecting the original biometrics into a random subspace may disrupt the distance preservation property and increase security risks. At the same time, the recognition performance depends on the quality of the feature extractor and the feature alignment between samples. However, traditional template storage methods are difficult to implement in actual operation due to cost limitations, and deep learning models are also difficult to effectively learn all covariates when the training samples are limited, thus affecting the recognition effect. Summary of the Invention

[0007] In view of the problems existing in the prior art, the present invention provides a deep biometric hashing network model and recognition method for finger vein biometric recognition, which at least includes a class center alignment module and a deep biometric hashing layer. The center alignment module consists of a two-stage localization network, a grid generator, a sampler, and a deep super template. By learning transformation parameters, it predicts the changes from all potential finger vein views to the finger center view to achieve feature alignment. The deep biometric hashing layer uses a system-level token and a hash function to generate protected pseudo-hash codes. The loss function of the deep biometric hashing network model during the model learning and training process is a mixed loss function, including a classification loss term, a consistency-based localization loss term, and a class center triplet loss term, jointly ensuring the recognition performance, revocable attribute, security, and privacy of the revocable biometric recognition scheme.

[0008] To achieve the above object, the technical solution adopted by the present invention is: A deep biometric hashing network model for finger vein biometric recognition, which at least includes a class center alignment module and a deep biometric hashing layer,

[0009] The center alignment module: consists of a two-stage localization network, a grid generator, a sampler, and a deep super template. By learning transformation parameters, it predicts the changes from all potential finger vein views to the finger center view to achieve feature alignment;

[0010] The deep biometric hashing layer: uses a system-level token and a hash function to generate protected pseudo-hash codes; on the one hand, it converts the user's original features into a protected domain to achieve secure authentication; on the other hand, it assigns the same token to all users to resist the security problems caused by attackers stealing the token;

[0011] The loss function of the deep biometric hashing network model during the model learning and training process is a mixed loss function, including a classification loss term, a consistency-based localization loss term, and a class center triplet loss term.

[0012] As an improvement of the present invention, the two-stage localization network is a convolutional neural network. The first stage consists of a convolutional layer, a pooling layer, and two fully-connected layers, which predict the coordinates of the ROI candidate boxes based on shallow features. The second stage consists of an ROIAlign layer, a convolutional layer, a pooling layer, and two fully-connected layers, which output the transformation parameter vector based on the predicted ROI candidate box coordinates and shallow features output by the first stage.

[0013] The grid generator is used to calculate the corresponding sampling positions of each pixel in the output image in the input image according to the predicted transformation parameters, and generate a coordinate mapping grid for subsequent samplers to perform interpolation calculations, thereby realizing the spatial transformation of the input image.

[0014] The sampler is used to realize the spatial transformation of the input feature map. Its input is the set of sampling points generated by the grid generator and the feature map, and the output is the feature map aligned with the super template.

[0015] The deep super template extracts and fuses the depth features according to multiple finger vein samples during model training to become an intermediate view representing the user's vein features.

[0016] As an improvement of the present invention, in the first stage of the two-stage localization network, the pooling layer uses adaptive average pooling with a size of 2×2 pixels and a stride of 2. The spatial scaling of the ROIAlign layer in the second stage is 0.25, and the output size is uniformly 32×64 pixels. The transformation parameter vector is 6-dimensional, respectively controlling scaling, rotation, and translation.

[0017] As another improvement of the present invention, in the grid generator, the transformation parameter vector u is obtained from the localization network, and u is used to create a sampling grid G for specifying the positions to be sampled in the input mapping to generate the transformed output. Assume T θ is a 2D affine transformation, then the transformation of each pixel point is defined as:

[0018]

[0019] where G i is a point on the regular sampling grid G, is the target coordinate of the regular sampling grid in the output feature map, is the source coordinate in the input feature map defining the sampling point, and A θ is the affine transformation matrix.

[0020] As another improvement of the present invention, in the sampler, a bilinear interpolation sampling kernel is used to obtain the pixel value at a specific position in the output feature map:

[0021]

[0022] Among them, the input feature map F roi has a size of (C, H, W), is the value at the position (n, m) of the cth channel of the input feature map, is the value at the position in the cth channel of the output feature map.

[0023] As another improvement of the present invention, the generation of the deep super template includes template initialization and template update:

[0024] In the template initialization, a threshold for the number of training iterations is set. After the iteration threshold is exceeded, template initialization is performed. All training samples of the user are recognized in the current model, and the ROI feature map of the sample with the highest recognition confidence for each user is used to initialize the deep super template of this user. Finally, a deep super template T = {T c} with the shape of (C, H, W) is obtained, where c ∈ [1, C] and C is the total number of user classes;

[0025] For the template update, it is updated by linearly fusing the ROI feature map after sampling the query sample and the super template, which is expressed as:

[0026]

[0027] Among them, p(x i ) is the final classification confidence of the query sample x i , T c is the deep super template of user c in the current state, T c ′ is the updated template, F roi ′ is the feature map after the sample x i of user c is transformed by the sampler, t is the set confidence threshold, and only when the confidence is greater than t, F roi ′ is allowed to be incorporated into the DST.

[0028] As yet another improvement of the present invention, in the deep bio-hash layer, the original biometric features {x i ∈ R|i = 1, …, k} are mapped using the ReLU function Max(0, x i ) to obtain {x i ′ ∈ R + |i = 1, …, k}; based on the system key seed K, pseudo-random vectors {r i ∈ R|i = 1, …, m} are generated, and Gram–Schmidt processing is performed on {r i ∈ R|i = 1, …, q} to obtain a set of orthogonal pseudo-random vectors {p i ∈ R|i = 1, …, q}; calculate xL The dot product with each orthogonal vector in p i results in y L = <x L |p i >>.

[0029] As a further improvement of the present invention, the hybrid loss function is specifically:

[0030] L = L cls + λ1L loc + λ2L cct

[0031] where L cls is the loss for the classification task, L loc is the loss for the semi-supervised localization task, L cct is the class center triple quantization loss, and λ1 and λ2 are the corresponding weight parameters;

[0032] The class center triple quantization loss L cct is specifically:

[0033]

[0034] where represents the distance from x i to the class center of the y i th class, d i,l represents the distance from the anchor sample x i to the class center of the lth class., N represents the number of samples, k represents the total number of users, and margin represents the margin value.

[0035] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a novel privacy-preserving revocable vein biometric scheme - a deep biometric hashing network (DBHN) model for vein biometric recognition, which is an end-to-end lightweight model based on deep learning technology, providing a new solution for user security and privacy protection and the realization of revocable biometrics. DBHN combines a powerful deep learning model architecture, an effective feature alignment module, and a suitable feature transformation scheme to ensure recognition performance, security, and privacy protection. The class center alignment (CCA) module addresses the within-class variation problem by approximating intermediate views and learning transformation parameters; the deep biometric hashing (DBH) layer generates pseudo-hash codes through system-level TRNs to cope with the security challenge of template random number theft; the final hybrid loss function supervises the training to ensure the preservation of the projected feature distances. Through the above combined innovations, DBHN exhibits good revocable attributes and resilience to security and privacy attacks while providing biometric security. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is the overall structure diagram of the deep biohashing network model of the present invention;

[0037] Figure 2 is the schematic diagram of the deep biohashing layer in the network model of the present invention;

[0038] Figure 3 is the variation diagram of the influence of the hash code length obtained from the simulation experiment of the present invention on the recognition performance;

[0039] Figure 4 is the distribution diagram of the matching scores of the preimage attack obtained from the simulation experiment of the present invention;

[0040] Figure 5 is the global unlinkability analysis diagram obtained from the simulation test of the present invention;

[0041] Figure 6 is the distribution diagram of the matching scores in the stolen scenario obtained from the simulation test of the present invention. Detailed implementation manners

[0042] The present invention will be further clarified below in conjunction with the accompanying drawings and detailed implementation manners. It should be understood that the following detailed implementation manners are only used to illustrate the present invention and not to limit the scope of the present invention.

[0043] Example 1

[0044] A deep biohashing network (DBHN) model for end-to-end cancellable privacy protection for finger vein biometric recognition, as Figure 1 shown. This network model includes a class center alignment module that predicts the changes from all potential finger vein views to the finger center view by learning transformation parameters, achieving fine feature alignment; a deep biohashing layer to address the security issues caused by stolen tokenized random numbers, which utilizes system-level tokens instead of assigning unique tokens to each user; and also provides a hybrid loss function for supervising the learning process of DBHN, including a classification loss term, a consistency-based localization loss term, and a class center triplet loss term.

[0045] The class center alignment (CCA) module is used to address the sensitivity problem of the feature transformation-based template protection method in the DBHN scheme to intra-class variations. First, CCA uses a deep super-template to approximate the intermediate view of each instance biometric feature. Second, by learning the transformation parameter θ, CCA obtains the transformation T θ , to align the query sample with the intermediate view of the corresponding class. Through this alignment, the potential view changes of the intermediate view can be predicted using θ.

[0046] The class center alignment module consists of four components: (1) a two-stage localization network, (2) a grid generator, (3) a sampler, and (4) a depth super-template. First, the CCA includes a depth super-template to approximately simulate the depth intermediate view in the physical sense of the biometrics of each instance. Second, the two-stage localization network sequentially regresses the ROI and learns the transformation parameter μ to obtain the transformation T from the query sample to the intermediate view of the corresponding class. u , so that the change from all potential views to their central view can be predicted through μ. Finally, the grid generator and sampler are used to align the query sample template with its corresponding super-template.

[0047] The two-stage localization network is designed as a simple convolutional neural network. The input of the first stage is the shallow feature F obtained from the shallow feature extractor. S , and the output is the coordinates (x min , y min , x max , y max ) of the ROI candidate box. Specifically, the first stage consists of a convolutional layer, a pooling layer, and two fully connected layers. The pooling layer uses adaptive average pooling with a size of 2×2 pixels and a stride of 2. The predicted ROI candidate box coordinates are regressed by the last fully connected layer.

[0048] The input of the second stage is the ROI candidate box coordinates predicted in the previous stage and the shallow feature F. S , and the output is the transformation parameter vector μ. The second stage consists of an ROIAlign layer, a convolutional layer, a pooling layer, and two fully connected layers. The spatial scaling of the ROIAlign layer is 0.25, and the output size is uniformly 32×64 pixels. The extracted ROI feature F roi is subtracted from the depth super-template of the corresponding class to obtain the change feature between the current query sample and the intermediate view of the corresponding class, and then connected with F roi to obtain the feature F t , which is passed as the input to the pooling layer. The pooling layer is the same as the first stage, and the first fully connected layer uses the rectified linear unit (ReLU) activation function. The transformation parameter vector μ is regressed by the last fully connected layer. For affine transformation, μ is 6-dimensional, controlling scaling, rotation, and translation respectively.

[0049] The grid generator in the class center alignment module obtains the transformation parameter vector u from the localization network. The sampling grid G is created using u, which is a set of points G i used to specify the positions in the input map that should be sampled to generate the transformed output. Assuming T θ is a 2D affine transformation, then the transformation of each pixel point in this case is defined as:

[0050]

[0051] where G i is a point on the regular sampling grid G, and is the target coordinate of the regular sampling grid in the output feature map. is the source coordinate in the input feature map that defines the sampling point. A θ is the affine transformation matrix. The scaling, rotation, translation, and linear distortion of the feature are controlled by these six parameters. The source / target transformation and sampling are equivalent to the standard texture mapping and coordinate system used in graphics.

[0052] The sampler in the class center alignment module takes the set of sampling points T θ (G i ) and the feature map F roi as input to the sampler, and the output is the feature map aligned with the super template. T θ (G i ) Each coordinate defines the spatial position in the input feature map, and the bilinear interpolation sampling kernel is used to obtain the pixel value at a specific position in the output feature map.

[0053]

[0054] Among them, for an input feature map F roi , its size is (C, H, W). is the value at the position (n, m) of the cth channel of the input feature map, is the value at the position in the cth channel of the output feature map. The sampling in Formula 2 is the same for any corresponding channels in the input feature F and F roi to ensure the spatial consistency between channels.

[0055] The depth super template in the class center alignment module is a composite template that acquires and processes multiple finger vein samples of the user during registration (model training), extracts depth features, and fuses them into a template. It represents an intermediate view of the user's vein features and enriches and comprehensively represents the features by continuously fusing samples. Generating the depth super template is divided into two steps: template initialization and template update.

[0056] First, template initialization. Set the threshold of the training iteration times. When it is greater than the iteration threshold, template initialization is performed. During the initialization process, all training samples of the user are recognized in the current model, and the ROI feature map of the sample with the highest recognition confidence for each user is used to initialize the depth super template of this user.

[0057] Finally, the initialized depth super template with the shape of (C, H, W) is obtained as T = {T c}, where \(c\in[1,C]\) and \(C\) is the total number of user categories. After the deep super-template is initialized, \(T\) is used as the approximate intermediate view for each user category to initialize the sample center \(c\) of the class center triple loss. l Initialize it.

[0058] After DST is initialized, it is equivalent to defining an approximate intermediate view of each user's finger vein. When the difference between the query sample and the intermediate view is large, it is very difficult to align the former with the latter. After choosing to align the query sample with the intermediate view through the transformation parameter \(\mu\), it is fused with the latter to update the template. Specifically, it is updated by linearly fusing the ROI feature map sampled from the query sample and the super-template, which is expressed as:

[0059]

[0060] where \(p(x i ) is the final classification confidence of the query sample \(x i \), \(T c \) is the deep super-template of user \(c\) in the current state, \(T c '\) is the updated template, and \(F roi '\) is the feature map after the sample \(x i \) of user \(c\) is transformed by the sampler. \(t\) is the set confidence threshold. Only when the confidence is greater than \(t\), \(F roi '\) is allowed to be incorporated into DST.

[0061] The Deep Biohashing (DBH) layer is used to establish a deep biohashing layer in DBHN to solve the security problem caused by the theft of TRNs in the original biohashing. Instead of assigning unique TRNs to each user, DBH uses a system-level TRN and a heuristic distance-preserving hash function to generate protected pseudo-hash codes, as shown in Figure 2 .

[0062] First, map the original biometric features \(\{x i \in\mathbb{R}|i = 1,\ldots,k\}\) using the ReLU function \(\text{Max}(0,x i )\) to obtain \(\{x i '\in\mathbb{R} + |i = 1,\ldots,k\}\).

[0063] Then, based on the system key \(\text{seed}K\), generate pseudo-random vectors \(\{r i \in\mathbb{R}|i = 1,\ldots,m\}\). Perform Gram - Schmidt processing on \(\{r i \in\mathbb{R}|i = 1,\ldots,q\}\) to obtain a set of orthogonal pseudo-random vectors \(\{p i \in\mathbb{R}|i = 1,\ldots,q\}\).

[0064] Finally, calculate x L and the dot product of each orthogonal vector in p i to obtain y L = <x L |p i >.

[0065] The present invention provides a hybrid loss function for training DBHN, which considers three aspects of classification, localization, and quantization losses, including the classification task loss L cls , the semi-supervised localization task loss L loc and the class center triple quantization loss L cct , which is used to ensure that the feature distances after deep biometric hashing projection remain unchanged. Specifically: L = L cls + λ1L loc + λ2L cct

[0066] where L cct uses the class center as the basis for judging positive and negative samples instead of a single sample. Suppose an anchor point x i ∈R D is given, and there are k classifications in the training set D. Then L cct can be defined as:

[0067]

[0068] where the distance to the y i th class center, i.e., the intra-class distance of x i , d i,l the distance from the anchor sample x i to the lth class center. Therefore, the distance d can be calculated (when l = y i is l ≠ y i is d i,l ).

[0069]

[0070] The class center initialization is not performed using the standard Gaussian distribution, but through standard inference using the deep super-template T under the current model parameters, thereby obtaining the initialization center vector for each category. Only when a few categories do not have an initialized deep super-template, the standard Gaussian distribution is used to initialize the center of that category. So the initialized class centers {c} init are:

[0071]

[0072] where T lThe depth super-template of class l. f(·) represents the continuous approximation encoding before binarization obtained from the standard inference of DBHN after the fine alignment module. Rand(seed l ) represents generating an initialized center with a standard Gaussian distribution for class l. L l represents whether the depth super-template of class l is empty (0 for empty, 1 otherwise). The update of the class center can be calculated by simply averaging the corresponding class features within the mini-batch. Therefore, the center change Δc l of class l can be calculated as:

[0073]

[0074] A deep biohashing network (DBHN) model for end-to-end cancellable privacy-preserving finger vein recognition, including a class center alignment module, a deep biohashing layer, and a hybrid loss function to supervise the learning process of DBHN.

[0075] The learning and training process of DBHN is as follows:

[0076] First, the input requirements include a labeled dataset S, an unlabeled dataset U, the initialization T of the depth super-template init , the total number of training epochs T, and the batch size B.

[0077] The model creation stage defines the student model S(θ) and the teacher model T(θ).

[0078] The training loop starts from t = 1 and ends when reaching the total T. In each epoch, first, the augmented labeled set S w and the unlabeled set U s , U w are generated through data augmentation.

[0079] Then, the teacher model is trained on S w , and its parameters are shared with the student model. Subsequently, the student model is trained on the unlabeled samples U s , using the predictions of the teacher model on U w . The initialization stage sets the total loss L = L cls + L loc , and initializes the depth super-template and class center when reaching T init , while updating the loss function to include additional constraints; when exceeding T initDuring the training phase, the deep super-template is updated according to a predetermined formula. Through these steps, the DBHN model can effectively learn and optimize, improving the performance and security of biometric recognition. During the inference / recognition process, the original finger vein image is directly input into the model to complete identity authentication.

[0080] All simulation experiments provided by the present invention used three publicly available mainstream finger vein datasets, HKPU-FV, SDUMLA-FV, and FV-USM. All training used Adam as the optimization method with a momentum of 0.9. The learning rate and batch size were set to 0.001 and 64 respectively. The DBHN model was trained using an RTX 4090 GPU, an Intel(R) Xeon(R) Gold 5218R CPU @ 2.10 GHz, and 440 GB of RAM for 500 epochs using the Pytorch framework. All datasets were divided into training and test sets at a ratio of 8:2.

[0081] Example 2

[0082] A vein biometric recognition method based on a deep biometric hashing network model, using the deep biometric hashing network model described in Example 1, is divided into a training phase and an identification phase:

[0083] The method in the training phase is as follows: The deep biometric hashing network model adopts a teacher-student semi-supervised framework. First, the labeled dataset (S) and the unlabeled dataset (U) are loaded and the deep super-template is initialized as an empty set. The teacher model and the student model are alternately trained, and the parameters are shared between the student and teacher models. The student model performs self-supervised learning through the feature consistency between strongly augmented samples and the teacher model on weakly augmented samples. When the number of model training epochs exceeds a predetermined threshold, the deep super-template is initialized. In addition, the model is additionally supervised using class center triplet loss. The parameters of the round with the minimum loss are saved to obtain the optimal model.

[0084] The method in the recognition stage is as follows: The recognition stage is based on the parameters saved in the optimal deep biometric hashing network model obtained in the training stage. First, the optimal parameters are loaded into the deep biometric hashing network model, and then the user's original vein image is used as the input of the deep biometric hashing network model. The shallow features of the user's vein image are obtained through the shallow feature extractor. Subsequently, the two-stage localization network is used to predict the ROI region and the affine transformation parameters, and the query features are aligned with the pre-stored deep super template to eliminate the pose difference; the aligned features are input into the deep biometric hashing layer, and a dot product operation is performed with the pseudo-random vector and binarized to generate an irreversible hash code; finally, the hash code is input into the fully connected layer to obtain the recognition confidence score. If the current score is greater than the predetermined threshold, it passes, otherwise the recognition is rejected. The recognition system accepts the user's original vein image and makes a recognition decision. The whole process is an end-to-end cancellable privacy-preserving recognition process.

[0085] Test case

[0086] Test case 1: Relationship between the length m of the deep hash code and the recognition performance

[0087] Set the hash code length m to 16, 32, 64, 128, 256, and 512, and test its recognition performance under normal and stolen conditions. Figure 3 The effect curve of different m on EER under normal conditions is shown, and Table 1 below shows the EER results obtained with different hash code lengths under normal and stolen conditions.

[0088] Table 1

[0089]

[0090] By observing Table 1 and Figure 3 It can be seen that:

[0091] (1) A larger m will significantly reduce the EER, and when m is greater than 200, the EER will not decrease further. Longer hash codes provide more bits for storing richer and more complete feature representations, thus improving the discriminability. The EER of the recognition system based on DBHN is very low on all three datasets.

[0092] (2) In the case of being stolen, for any length of hash code, the EER is almost always 0. This indicates that the distributions of the matching scores of registered users and those of adversaries are clearly distinguishable, even if the adversary has the correct TRN. This can ensure the security of the system after the TRN is stolen, and the adversary cannot access the recognition system by relying on the TRN.

[0093] Test case 2: Performance of the revocable vein recognition system based on the present invention against preimage attacks

[0094] To verify whether the original biometric can be recovered from the transformed biometric vector, assume that in the worst case, the attacker can access all factors (including the transformed template and all knowledge of the DBHN), and attempt to recover the original biometric through a preimage attack.

[0095] Figure 4 Shows the matching attacks on three datasets, the distributions of the matching scores for genuine matches and impostor matches. The results show that the distributions of the pre-attack scores and the impostor scores highly overlap, while the genuine match distribution is clearly distinct. The EER for all three datasets is 0. Further analysis shows that for the solution provided by the present invention, even for the same biometric sample, the real-domain biometric vectors generated in the DBHN model using different tokens are different. This is because the DBH is used to design the DBH layer in the deep learning model and participates in optimizing the weights of the DBHN model, the PA-RAMETER. When the model uses different tokens, the orthonormal matrix used for projection will be different. This will affect the parameter optimization of the feature extraction module in the DBHN to ensure that it automatically selects the best features. As a result, the real-domain feature vectors of the DBHN inputs from different tokens to the DBH layer will be different, thus effectively defending against spoofing attacks, providing an assurance of irreversibility, and exhibiting some unlinkability.

[0096] Test Case 3: Performance of the unlinkability of the revocable vein recognition system based on the present invention

[0097] Unlinkability means that the protected templates generated using different tokens are not linkable, which makes it possible to set different keys to generate unlinked protected templates when a user registers for multiple applications or different databases. An adversary cannot attack by linking the user's protected templates across different applications. This test case evaluates the unlinkability of the cancellable finger vein recognition system based on the DBHN. By setting different keys for the user, the paired and unqualified scores are calculated, and finally the global linkability is calculated to perform cross-matching. Figure 5 Shows the distributions of the mate and unqualified scores and the global linkability scores on each dataset. It can be observed that the DBHN-based recognition system shows minimal linkability only when the mating score is greater than the non-legitimate score (less than 1% area).

[0098] Test Case 4: Performance of the revocability of the revocable vein recognition system based on the present invention

[0099] Revocability means that in the recognition system, the original (stolen) and new (re-released) templates are generated from the same finger vein template, but the two templates are not linked. This test case empirically studies revocability by generating an imposter distribution. Revocability can be empirically verified if (1) the imposter and pseudo-imposter distributions overlap and (2) the genuine and pseudo-imposter distributions are clearly separable. Figure 6 Shows the genuine, imposter, and pseudo-imposter distributions on each dataset. There is a large overlap between the imposter and pseudo-imposter, while the genuine and pseudo-imposter distribution scores are clearly distinguishable. The separability or overlap between the two distributions can be quantitatively estimated by a qualitative index. Table 2 below shows the separability index between the three distributions on each dataset. It can be observed that the separability index between the imposter and pseudo-imposter distributions on each dataset is very low. These demonstrate that the cancellable finger vein recognition system provided by the present invention has good revocability.

[0100] Table 2

[0101]

[0102] In summary, a Deep Biohashing Network (DBHN) model disclosed by the method of the present invention can be used for end-to-end cancellable privacy-preserving finger vein recognition. This network model enhances the resistance to existing security and privacy attacks and meets the properties of cancellable biometrics.

[0103] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements all fall within the protection scope of the claims of the present invention.

Claims

1. A deep biometric hashing network model for vein biometric recognition, characterized in that : It includes at least a class center alignment module and a deep biometric hashing layer. The center alignment module: It consists of a two-stage localization network, a grid generator, a sampler, and a deep super-template. By learning transformation parameters, it predicts the changes from all potential finger vein views to the finger center view to achieve feature alignment. The deep biometric hashing layer: It uses system-level tokens and a hashing function to generate protected pseudo-hash codes. The hashing layer transforms the user's original features into a protected domain to achieve secure authentication. At the same time, the same token is assigned to all users to resist attackers from stealing the token. The loss function of the deep biometric hashing network model during the model learning and training process is a hybrid loss function, including a classification loss term, a consistency-based localization loss term, and a class center triplet loss term.

2. The deep biometric hashing network model for vein biometric recognition according to claim 1, wherein: The two-stage localization network is a convolutional neural network. The first stage consists of a convolutional layer, a pooling layer, and two fully connected layers. According to the shallow features, it predicts the coordinates of the ROI candidate boxes. The second stage consists of a ROIAlign layer, a convolutional layer, a pooling layer, and two fully connected layers. According to the predicted ROI candidate box coordinates and shallow features output by the first stage, it outputs a transformation parameter vector. The grid generator: According to the predicted transformation parameters, it calculates the corresponding sampling positions of each pixel in the output image in the input image, generates a coordinate mapping grid, and realizes the spatial transformation of the input image. The sampler is used to realize the spatial transformation of the input feature map. Its input is the set of sampling points and the feature map generated by the grid generator, and the output is the feature map aligned with the super-template. The deep super-template extracts and fuses the deep features according to multiple finger vein samples during model training to become an intermediate view representing the user's vein features.

3. The deep biometric hashing network model for venous biometric recognition according to claim 2, wherein: In the first stage of the two-stage localization network, the pooling layer uses adaptive average pooling with a size of 2×2 pixels and a stride of 2. The spatial scaling of the ROIAlign layer in the second stage is 0.25, and the output size is uniformly 32×64 pixels. The transformation parameter vector is 6-dimensional, respectively controlling scaling, rotation, and translation.

4. The deep biometric hashing network model for vein biometric recognition according to claim 1, characterized in that: In the grid generator, a transformation parameter vector u is obtained from the localization network, and u is used to create a sampling grid G for specifying the positions to be sampled in the input mapping to generate the transformed output; assume T θ is a 2D affine transformation, and the transformation of each pixel is defined as: where G i is a point on the regular sampling grid G, is the target coordinate of the regular sampling grid in the output feature map, is the source coordinate in the input feature map that defines the sampling point, A θ is the affine transformation matrix.

5. The deep biometric hashing network model for vein biometric recognition according to claim 1, characterized in that: In the sampler, a bilinear interpolation sampling kernel is used to obtain the pixel values at specific positions in the output feature map: Among them, the input feature map F roi has a size of (C, H, W), is the value at the position (n, m) of the cth channel of the input feature map, is the value at the position in the cth channel of the output feature map.

6. The deep biometric hashing network model for vein biometric recognition according to claim 1, characterized in that: The generation of the deep super-template includes template initialization and template update: In the template initialization, a threshold for the number of training iterations is set. After the iteration threshold is exceeded, template initialization is performed. All training samples of the user are recognized in the current model, and the ROI feature map of the sample with the highest recognition confidence for each user is used to initialize the deep super-template of this user. Finally, a deep super-template T of shape (C, H, W) is obtained, where T = {T c}, c ∈ [1, C], and C is the total number of user categories; The template update is updated by linearly fusing the ROI feature map after query sample sampling and the super-template, which is expressed as: Among them, p(x i ) is the final classification confidence of the query sample x i , T c is the deep super-template of user c in the current state, T c ' is the updated template, F roi ' is the feature map after the sample x of user c i is transformed by the sampler, t is the set confidence threshold, and only when the confidence is greater than t, F roi ' is allowed to be incorporated into the DST.

7. The deep biometric hashing network model for vein biometric recognition according to claim 1, characterized in that: In the deep bio - hashing layer, the original biometric features {x i ∈R|i = 1,…,k} are mapped using the ReLU function Max(0, x i ) to obtain {x i ′∈R + |i = 1,…,k}; Based on the system key seed K, generate a pseudo-random vector {r i ∈ R|i = 1, …, m}, and perform Gram–Schmidt processing on {r i ∈ R|i = 1, …, q} to obtain a set of orthogonal pseudo-random vectors {p i ∈ R|i = 1, …, q}; Calculate x L with p i the dot product of each orthogonal vector in, to obtain y L = <x L |p i >.

8. The deep biometric hashing network model for venous biometric recognition according to claim 1, characterized in that: The specific form of the hybrid loss function is: L = L cls + λ1L loc + λ2L cct Among them, L cls is the classification task loss, L loc is the semi-supervised localization task loss, and L cct is the class center triple quantization loss, where λ1 and λ2 are weight parameters; The class center triple quantization loss L cct Specifically: Among them, represents the distance from x i to the center of the y i th class, d i,l represents the distance from the anchor sample x i to the center of the lth class, N represents the number of samples, k represents the total number of users, and margin represents the margin value.

9. A venous biometric recognition method using the deep bio-hash network model as described in claim 1, characterized in that: It is divided into a training stage and an identification stage: The training stage: The deep biometric hashing network model adopts a teacher-student semi-supervised framework, loads the labeled dataset and the unlabeled dataset and initializes the deep super-template as an empty set. The teacher model and the student model are alternately trained and share parameters. The student model performs self-supervised learning through the feature consistency between the strongly augmented samples and the teacher model on the weakly augmented samples. When the number of model training rounds exceeds a predetermined threshold, the deep super-template is initialized, and the parameters of the round with the minimum loss are saved to obtain the optimal model. The recognition stage: Based on the optimal deep biometric hashing network model obtained in the training stage, the original vein image of the user is input. The shallow features of the user's vein image are obtained through the shallow feature extractor. The two-stage localization network is used to predict the ROI region and the affine transformation parameters. After aligning the query features with the pre-stored deep super-template, the features are input into the deep biometric hashing layer, and a dot product operation is performed with the pseudo-random vector and binarized to generate an irreversible hash code. The hash code is input into the fully connected layer to obtain the recognition confidence score. If the current score is greater than the predetermined threshold, the recognition is completed.

Images