KR20200051903A - Fake fingerprint detection method and system - Google Patents

Abstract

The present invention relates to a method and a system for fingerprint recognition capable of fake fingerprint detection and, more specifically, to a method and a system for fingerprint recognition capable of fake fingerprint detection which allows a system to learn the difference between a biometric fingerprint and a fake fingerprint by using deep learning in fake fingerprint detection, divide and read a fake fingerprint when reading the fingerprint by learned results, add up divided reading results to determine a fake status, and use a gram matrix to increase detection accuracy. The system comprises: a sensor module to acquire bioinformation of a user; a feature extraction unit to perform quality measurement and data preprocessing of the bioinformation, extract features, and determine a fake status; a database unit to store a bioinformation template for authentication; and a matching unit to generate a matching score of the template and the acquired bioinformation and perform authentication. The feature extraction unit divides the acquired bioinformation into a plurality of patches, then extracts the features, and determines the fake status.

Description

Fingerprint recognition method and system capable of detecting forged fingerprints {FAKE FINGERPRINT DETECTION METHOD AND SYSTEM}

The present invention relates to a fingerprint recognition method and system capable of detecting a forged fingerprint, and more specifically, a system learns a difference between a biometric fingerprint and a forged fingerprint using deep learning in detecting a forged fingerprint, and reads the forged fingerprint through the learned result. It relates to a fingerprint recognition method and system capable of detecting a forgery fingerprint that improves the accuracy of detection by using a Gram matrix to determine whether or not to falsify by dividing the read and summing the results of the divided read.

Recently, due to the increase in the biometric authentication system on mobile devices and the case of manipulation of access records using forged fingerprints, discussions on the detection of forged fingerprints have been revitalized. Existing counterfeit fingerprint detection methods can be divided into software-based methods and hardware-based methods. The hardware-based method has a disadvantage in that an additional sensor is required to check whether the body is biometric, and the software-based method does not require an additional sensor, but has a disadvantage in that the performance of determining whether the counterfeit is counterfeit is poor compared to the hardware method.

Recently, the convolutional neural network (CNN), which is widely used in the field of image processing, is showing the best performance in the field of image classification and detection, and is widely used in many industries. If the high classification performance of the convolutional neural network method can be applied to the detection of counterfeit fingerprints, there will be a room to offset the relatively low recognition performance of the software-based counterfeit fingerprint detection.

However, in general, the learning of the convolutional neural network requires a large amount of fingerprint images and the number of parameters to be learned requires an optimization process to be applied to an embedded system having a small memory capacity and low computational performance. The process of creating a fake fingerprint for learning takes a lot of time, and a large number of parameters require more memory and processing time. In addition, since the existing convolutional neural network technology is designed for classification and detection of general images with significantly different data distribution characteristics from forgery fingerprints, it is necessary to design a network suitable for forgery fingerprint detection.

Accordingly, an object of the present invention relates to an optimal design method of a convolutional neural network for detecting software-based forgery fingerprints, and a fingerprint recognition method capable of detecting forgery fingerprints to provide a high classification performance of a convolutional neural network to an embedded system. And systems.

According to the present invention, it is possible to perform forged fingerprint detection with higher accuracy than the existing forged fingerprint detection method.

In addition, it is possible to provide a forgery fingerprint detection technology capable of maintaining accuracy while reducing the number of variables to be calculated than the existing forgery fingerprint detection method.

Meanwhile, the effects of the present invention are not limited to the above-mentioned effects, and various effects may be included within a range obvious to those skilled in the art from the following description.

1 is an example of a forged fingerprint detection system according to an embodiment of the present invention.
2 shows a method of applying CNNs to forgery fingerprint detection according to an embodiment of the present invention.
3 is an example of image transformation using a gram matrix.
4 illustrates a process of generating a gram matrix according to an embodiment of the present invention.
5 illustrates a process of generating a K × K gram matrix using the Gram-K module.
6 shows a CNNs structure using a Gram matrix according to an embodiment of the present invention.
7 illustrates the structure of a Gram matrix-based CNNs model according to an embodiment of the present invention.
8 shows hyper-parameters used in the end-to-end gram model according to an embodiment of the present invention.
9 is a comparison of the performance of the method applied to the Gram module and the existing method.
10 is a comparison of the detection performance of the Gram-128 model.
11 shows the DET curve of the Gram-128 model.
12 is a graph showing the detection error rate [0%, 20%] section of the DET curve of Gram-128.
13 illustrates a patch-based counterfeit fingerprint detection method according to an embodiment of the present invention.
14 illustrates the overall structure of a patch-based fully connected network to which an optimum threshold value is applied according to an embodiment of the present invention.
15 shows the number of patches obtained for each data.
16 is a diagram illustrating a proposed patch-based CNNs model structure using a Fire module according to an embodiment of the present invention.
17 shows a patch-based CNNs structure according to an embodiment of the present invention.
18 is a comparison of an existing method and a method according to an embodiment of the present invention.
FIG. 19 shows changes in the values of Ferrfake and Ferrlive according to threshold values of LivDet2015 CrossMatch data.
20 illustrates hyperparameters used for learning a patch-based network according to an embodiment of the present invention.
21 is a 32 × 32 patch-based counterfeit fingerprint detection performance comparison.
22 is a graph comparing 32 × 32 patch-based counterfeit fingerprint detection performance.
23 is a 48 × 48 patch-based counterfeit fingerprint detection performance comparison.
24 is a graph comparing 48 × 48 patch-based counterfeit fingerprint detection performance.
25 is a 64 × 64 patch-based counterfeit fingerprint detection performance comparison.
26 is a graph comparing 64 × 64 patch-based counterfeit fingerprint detection performance.
27 compares performance according to patch sizes.
28 is a graph comparing performance according to patch sizes.
Fig. 29 compares the results of applying the voting method and the optimal threshold.
30 compares the performance of the optimal threshold method and the voting method.
31 compares performance with a model to which data enhancement is applied.
32 is a comparison of performance when data augmentation is applied.
33 shows a DET curve of a method according to an embodiment of the present invention.
Fig. 34 compares the processing speed of the modified SqueezeNet (unit: ms).
35 compares the processing speed of SqueezeNet and a modified SqueezeNet according to an embodiment of the present invention.
36 is a comparison of the lattice method of the CNNs structure and the speed of a complete convolution method according to an embodiment of the present invention.
37 is a result of partially observing the characteristics of a forgery fingerprint using full convolution according to an embodiment of the present invention.
38 is an example of a case in which all of the same counterfeit fingerprints are classified as biological fingerprints.
39 illustrates the results of generalization performance evaluation of methods according to an embodiment of the present invention.
40 is a flowchart of a method for detecting a forged fingerprint according to an embodiment of the present invention.

Hereinafter, the 'alignment grid volume rendering method and apparatus' according to the present invention will be described in detail with reference to the accompanying drawings. The described embodiments are provided so that those skilled in the art can easily understand the technical idea of the present invention, and the present invention is not limited thereby. In addition, matters expressed in the accompanying drawings may be different from those actually implemented in schematic drawings to easily describe embodiments of the present invention.

On the other hand, each component represented below is only an example for implementing the present invention. Accordingly, other components may be used in other implementations of the present invention without departing from the spirit and scope of the present invention.

In addition, each component may be implemented solely in the configuration of hardware or software, but may also be implemented in a combination of various hardware and software components that perform the same function. Also, two or more components may be implemented together by one hardware or software.

In addition, the expression 'includes' certain components, as an expression of 'open', simply refers to the existence of the components, and should not be understood as excluding additional components.

1 is an example of a forged fingerprint detection system according to an embodiment of the present invention.

Referring to FIG. 1, a forgery fingerprint detection system according to an embodiment of the present invention may be included in a fingerprint recognition system.

The forgery fingerprint detection system according to an embodiment of the present invention may include a sensor module 110, a feature extraction unit 120, a matching unit 130, and a database unit 140.

The sensor module 110 may acquire biometric information of each individual. The sensor module 110 may recognize and obtain a user's fingerprint as biometric information. The biometric information may include fingerprints, voices, irises, DNA, face types, and the like.

The feature extraction unit 120 may measure the quality of the bio-information and perform data preprocessing, feature extraction, and determination of forgery. In the feature extraction unit 120, the size of the patch may have one of 32X32, 48X48, and 64X64 pixels.

Preferably, the feature extraction unit 120 may have a size of the patch 48X48 pixels.

The feature extraction unit 120 may determine whether to counterfeit the plurality of patches individually.

The feature extracting unit 120 may determine whether the entire biometric information is counterfeit or biometric information by synthesizing the result of determining whether the plurality of patches are forged.

The feature extracting unit 120 may determine whether to falsify based on a threshold value determined by learning.

The feature extraction unit 120 may extract a gram matrix from the obtained biometric information and use it to determine whether or not it is forged.

The feature extraction unit 120 may have a size of the gram matrix of 128x128.

The matching unit 130 may generate a template stored in the database unit 140 and a matching score of biometric features obtained by the feature extraction unit 120. The matching unit 130 may determine whether to authenticate according to the matching score of the biometric information. The matching unit 130 may perform the operation of authenticating the user according to the matching score of the biometric information.

The database unit 140 may store biometric data as a template through a biometric registration process. The database unit 140 may store the results learned by the feature extraction unit 120. The database unit 140 may store data for authentication and data for falsification.

2 shows a method of applying CNNs to forgery fingerprint detection according to an embodiment of the present invention.

In counterfeit fingerprint studies, the models that trained the classifier after extracting texture information using a local binary pattern (LBP) or a Gabor filter often showed good performance. The matching unit 130 according to an embodiment of the present invention may operate regardless of the input size of the fingerprint and may use an end-to-end CNN model using a Gram matrix having a parameter size of about 1.2 MB. The matching unit 130 according to an embodiment of the present invention has a total number of parameters of the CNN model of 308,554, and when one parameter is 4 bytes, the size may be about 1.2 MB. The feature extraction unit 120 according to an embodiment of the present invention detects a feature suitable for forgery fingerprint detection, and the matching unit 130 can introduce a Gram module to configure a network independent of the size of the input fingerprint. have. The CNN model used by the matching unit 130 may receive a fingerprint image as an input without a pre-processing process, and finally distinguish between a forgery and a living body. In the fingerprint recognition system capable of detecting a forged fingerprint according to an embodiment of the present invention, since the average detection error of the forged fingerprint is 2.61%, an end-to-end CNN model having high performance with a small number of parameters can be constructed.

In the case where CNNs are used to detect counterfeit fingerprints, most systems operate in the form of (a) and (b) of FIG. 2. In the case of (a) of FIG. 2, the fingerprint is divided into patches and processed. Figure 2 (a) is a method of determining whether a final forgery is obtained by dividing a fingerprint into small patches of 16 × 16 or 32 × 32 and applying it to CNNs and post-processing the result. In the case of (b) of FIG. 2, the method applied in the general classification is applied to the fingerprint as it is. CNNs are applied after the fingerprint is cut to fit the network's input size or applied to enlarge / reduce. To this end, a method such as image segmentation may be applied. In the case of (a) of FIG. 2, pre-processing of cutting fingerprints into patches before applying CNNs and a post-processing process for synthesizing patch results are required. In the case of (b), there is no room for information loss due to image reduction. It exists and likewise requires a pretreatment process such as segmentation. In the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention, the CNNs model using the Gram matrix operates in the form of (c) of FIG. 2 which is end-to-end and is the most of the CNNs-based methods presented so far. Use a small number of parameters.

3 is an example of image transformation using a gram matrix.

Referring to FIG. 3, a fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention uses a Gram matrix, and in order to utilize this, it is necessary to understand style transfer. The style delivery method has attracted many people's attention by extracting the artist's painting style using CNNs and applying it to a general picture to replace the artist's own style as shown in FIG. 3. The essence of this method is to change only the style while maintaining the content information of the input photo. Introduced to extract styles from artists' paintings is the Gram matrix, which calculates the correlation coefficient of each channel of the CNNs tensor to form a matrix. The difference between the Gram matrix obtained from the artist's picture and the Gram matrix of the image to be transformed is made as a loss function, and instead of updating the parameters, the pixels of the input image are updated. At the same time, if the iterative update of the input image described above is performed while maintaining the content information of the original input, the effect shown in FIG. 3 can be obtained.

The forgery fingerprint detection system according to an embodiment of the present invention focuses on a style transfer idea, extracts a Gram matrix from a fingerprint, and then applies it to the forgery fingerprint classification. That is, it is assumed that the style defined in the style delivery method is the texture of the image, and the Gram matrix is applied to extract it.

The Gram module used for style delivery is defined as a texture network, and an example using it is defined as a network using the Gram matrix described above as a texture network and connected with GAN (Generative Adversarial Networks) to create higher-level artistic photos. There is something I used to pay. Another example suggested a CNNs model that rotated a tensor inside a network to obtain the rotation-invariant features through CNNs, and then combined it with their own designed pooling network. This is to apply object recognition robust to rotation rather than to acquire texture information. The forgery fingerprint detection system according to an embodiment of the present invention is used as a classification model rather than a generation model using a Gram matrix to make an artistic photograph. In addition, the forgery fingerprint detection system according to an embodiment of the present invention designed a network in a form of reducing the number of parameters to twice or more than SqueezeNet, and introduced a Gram module that makes the size of the tensor the same in the network, to input the image size. It operates regardless of design conditions and satisfies all design constraints.

4 illustrates a process of generating a gram matrix according to an embodiment of the present invention.

Referring to FIG. 4, if a three-dimensional tensor having a height, width, and channel size of H × W × C exists in the Gram matrix, it can be made into a two-dimensional (HW) × matrix, and then a correlation matrix of each row component can be created. have. In order to generalize the process of creating the Gram matrix, assuming that the 2D matrix transformation result of (HW) × which is an intermediate step in FIG. 4 is,, this may mean the second layer of the neural network inserted for generalization. Assuming that is the first row vector of the matrix, each element of the 2-dimensional Gram matrix of the layer _{, j,} can be calculated through the dot product as shown in Equation 1.

[Equation 1]

The Gram matrix is a symmetric matrix, which is used to extract the style of artistic paintings. The style feature can be interpreted as an overall texture independent of position through the Gram operation, and this feature can also be used to detect counterfeit fingerprints.

5 shows a process of generating a K × K gram matrix using the Gram-K module, and FIG. 6 illustrates a CNNs structure using a Gram matrix according to an embodiment of the present invention.

5 and 6, a Gram module is defined to include a Gram matrix in CNNs. The Gram module may include a 1 × 1 convolution layer, a tanh nonlinear activation function, and a Gram layer. The Gram module controls the number of 1 × 1 convolution filters to create a Gram matrix of the desired size. According to an embodiment of the present invention, a Gram module in which the number of 1 × 1 filters is K may be defined as Gram-K. 5 shows an example of obtaining a K × K Gram matrix as a result of Gram-K. Using Gram-K, you can configure CNNs that accept any size input. The fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention applies a Fire module together to design CNNs having a small number of parameters and may include a structure as shown in FIG. 6. The first Fire module, fire2-128, has 128 Expand layer filters. Since the SR (Squeeze Ratio) is set to 0.125, the number of 1 × 1 filters in the Squeeze layer may include 16, 8 times less than the indicated 128. In addition, PCT, which is the ratio of the number of 1x1 filters and 3x3 filters in the Expand layer, was set to 0.5 to have the same number. This is the basic SR and PCT values applied in the original SqueezeNet, and the same SR and PCT can be applied to all Fire modules in FIG. 6. In the first layer of FIG. 6, conv1-96 / 2, 96 7 × 7 filters having a moving size of 2 were used. The last convolution layer, conv7-2, is two 1x1 filters. The maxpool / 2 present in FIG. 6 refers to maxpooling having a filter movement 2 of 3 × 3 size, and avgpooling means global average pooling. Gram-128 is a Gram module and the size of the output tensor can be 128 × 128 × 1.

7 illustrates the structure of a Gram matrix-based CNNs model according to an embodiment of the present invention.

Referring to FIG. 7, three Gram matrices obtained from each Gram-128 module are stacked in the channel direction to form a 128 × 128 × 3 form, which in turn becomes the input of the Fire module. The fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention can process input of any size without a process such as segmentation by applying a Gram module and operate in an end-to-end manner. The total number of parameters in the network is 308,554, which is more than twice the number of parameters in SqueezeNet, 738,594. When the size of the fingerprint image input in FIG. 7 is K × K, the structure of the CNNs model using the Gram matrix is expressed in detail.

7, it can be seen that input images of various sizes (rectangular input is also possible) are transformed to a size of 128 × 128 × 1 when passing through the Gram module. The concatenation part of FIG. 7 is a part connecting three results of the Gram module, and it can be seen that the size is 128 × 128 × 3.

8 shows hyper-parameters used in the end-to-end Gram model according to an embodiment of the present invention.

Referring to FIG. 8, a fingerprint recognition system capable of detecting forged fingerprints according to an embodiment of the present invention may utilize NVIDA GTX 1080. The parameter applied to learning was 0.0005, the batch size was 8, and the number of generations (epoch) was 80. Adamax was applied as the optimization method, and 10% of the learning LivDet data was used for verification. For verification data, fingerprints from other people were used as much as possible, and the data distribution was matched identically. In addition, when learning, random mirror symmetry is applied in batch units, and the loss value of verification data is checked for each generation, and if the loss value does not decrease during 4 generations, the learning rate is doubled. 8 shows the hyperparameters used in the experiment.

FIG. 9 compares the performance of the method to which the Gram module is applied and the existing method, and FIG. 10 compares the detection performance of the Gram-128 model.

9 and 10, in order to evaluate forgery fingerprint detection performance of the model using the proposed Gram matrix, the (Average Classification Error) ACE result of the forgery fingerprint detection is shown in FIG. 9. While the average performance of the existing technology is 3.52%, the average ACE of the fingerprint recognition system capable of detecting forgery fingerprints according to an embodiment of the present invention was 2.61%. Data augmentation is the result of randomly applying additional up and down symmetry to the mirror symmetry as before. Looking at the average value of the ACE, it shows slightly worse performance than the one not applied. The processing time required to determine the forgery fingerprint appears in proportion to the size of the input image and shows an average of 21 ms, which proves that it is processed faster than SqueezeNet of the segmentation-based method. Fig. 10 shows the results of Fig. 9 graphically. The CNNs model using the Gram module of the forgery fingerprint detection system according to an embodiment of the present invention showed better detection results with a smaller number of parameters than SqueezeNet. It can be seen that the average time required to process a single fingerprint image was about twice as fast as the SqueezeNet method with segmentation applied, and if the time required for segmentation was measured together, the speed improvement result would appear at a larger rate. .

FIG. 11 shows the DET curve of the Gram-128 model, and FIG. 12 shows the [0%, 20%] section of the detection error of the DET curve of the Gram-128.

Referring to FIG. 11, if a DET (Detection Error Tradeoff) curve of the Gram-128 model of FIG. 9 is drawn, the same result as in FIG. 11 can be obtained. In the case of the Italdata sensor, it can be confirmed that Ferrlive hardly decreases. In fact, the ACE value of the 2011 Italdata sensor of FIG. 9 is a value calculated by Ferrlive being 10% and Ferrfake being 0%. 12, the error range is enlarged and expressed as [0%, 20%] so that the trend of the change can be better identified.

13 illustrates a patch-based counterfeit fingerprint detection method according to an embodiment of the present invention.

Referring to FIG. 13, a fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention may include a CNNs model for determining whether a fingerprint is forged by using a patch of a fingerprint image. When a patch is used instead of the entire fingerprint image, a network according to the patch size, not the size of the fingerprint image, can be constructed, thereby establishing a network independent of the size of the input image. There are generally several advantages to using a patch:

1) A network independent of the size of a fingerprint can be built.

 Regardless of the size of the fingerprint, it is possible to build a network according to the size of the patch, so it can process fingerprint images of various sizes.

2) The size of the training data can be increased.

 Since the fingerprint is divided into several patches for learning, the number of data used for learning can be increased.

3) Local characteristics of counterfeit fingerprints can be identified.

 Since the forgery fingerprint is determined based on the local patch without using the entire fingerprint, errors that may occur when using the entire fingerprint can be reduced.

4) By reducing the number of pooling applied to CNNs, information loss can be prevented.

 Since the size of the patch coming into the network input is small, it is possible to perform less pooling, thereby reducing information loss.

5) It is easy to understand the characteristics of counterfeit fingerprints through visualization.

 It provides an analysis tool that can visually check which part of the fingerprint is judged to be a forgery or a biometric fingerprint through the patch.

6) It can be applied in places such as mobile fingerprint sensors.

 Since the mobile fingerprint sensor has a very small area compared to the conventional optical fingerprint input sensor, it acquires an image of a small fingerprint area.

While most of the forgery fingerprint classification constitutes a classifier of two types of biometric fingerprints and forgery fingerprints, the fingerprint recognition system capable of detecting forgery fingerprints according to an embodiment of the present invention includes three types of biometrics, forgeries, and backgrounds. Classify as. Through this, it is possible to integrate the most frequently used segmentation process into CNNs as a pre-processing process for fingerprints, thereby allowing forgery detection at a faster rate. When determining the forgery fingerprint, the background class is ignored, and the results classified as the living body and the forgery are synthesized to determine whether the forgery is counterfeit. This procedure is shown in FIG. 13. The size of the patch used in the fingerprint recognition system capable of detecting forgery fingerprints according to an embodiment of the present invention may be 32 × 32, 48 × 48, or 64 × 64. Through this, it is possible to confirm a change in the performance of detecting a forgery fingerprint according to the patch size. The patch-based method used in the feature extraction unit 120 of the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention has the following differentiation from the CNNs-based forgery fingerprint detection methods using the existing patch.

 Fingerprints are not simply two categories of forgery and living body, but are divided into living body, forgery, and background to integrate preprocessing processes such as segmentation into CNNs.

 Unlike existing methods using only one patch size, the size of the patch used for learning was experimented with 32 × 32, 48 × 48, and 64 × 64 to determine the appropriate patch size for the network.

 After cutting the fingerprints in units of patches, it was shown that it is possible to perform better in a completely convolutional way than the conventional method of applying CNNs to each patch.

 When determining the final forgery, it was shown that the method of determining the optimal threshold value from the learning data and using it to determine the final forgery showed better performance than the existing method of determining the final forgery through voting.

 A design with a limited number of parameters can include a CNNs model that maintains high performance with a small number of parameters.

 It has been shown that the local characteristics of counterfeit fingerprints can be visualized using the results processed in a completely convolutional manner.

As a result of the experiment, the proposed method showed an average detection error of 1.35% in LivDet counterfeit fingerprint data with 536,143 parameters (approximately 2.0 MB), indicating that a high-performance full-convolution method with few parameters is applicable to counterfeit fingerprint detection. In addition, the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention uses a patch that is not affected by the size of the input fingerprint and processes a forgery fingerprint of any size because there is no complete connection layer inside the network. Can be. In addition, since the processed results are classified into counterfeit, biological, and background, there is an advantage that no pre-processing such as segmentation is necessary.

14 illustrates the overall structure of a patch-based fully connected network to which an optimum threshold value is applied according to an embodiment of the present invention.

14, the entire structure of the feature extraction unit 120 of the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention may be as shown in FIG. 14. The feature extraction unit 120 first applies segmentation to learn CNNs, and then extracts bio, counterfeit, and background patches based on the region. The extracted patches are used for learning CNNs, and the completed learning CNNs can be used to find an optimal threshold with a complete convolution structure. After removing the results classified as the background from the results of the full convolution network, the patch results can be synthesized to find the optimal threshold value for distinguishing the forgery and the living body. In the test process, after applying the trained network, the result is compared with the optimal threshold value to determine whether or not it is forged. Each process is explained in detail later.

In the feature extraction unit 120, a patch obtained from LivDet data should have a background class, but background data does not exist in LivDet. Therefore, it is necessary to create a background class directly on the acquired patch. Each patch was checked by eye, and the background class was not assigned, and the segmentation method was used to automatically classify the background. The method of acquiring the learning patch is divided into a patch in a grid form that does not allow overlapping after segmenting the fingerprint image. If the segmented areas in each part of the divided patch appear below a certain percentage, the patch is classified as a background class. For example, the 32 × 32 patch has a total number of 3232 = 1024 pixels, of which 800/1024 = 0.78 when there are 800 pixels segmented by fingerprint, and if this value is less than 0.4, it is classified as a background class. . When the value for determining the ratio of the segmented region was R, the value of R was set to 0.4 for the 32 × 32 patch. The R value was set differently according to the size of the patch. For 48 × 48, R × 1/2 and 64 × 64 were used for R × 1/3. This is because the larger the patch, the more fingerprints are included in the background class, so the adaptively changes the R value.

15 shows the number of patches obtained for each data.

Referring to FIG. 15, in general, when a fingerprint is divided into patches, the number of classes in the background is very small. Therefore, the background patch was set to a similar number to other patches through data augmentation. The number of patches obtained according to each size is shown in FIG. 15. The number in the front of the Background column in FIG. 15 is the number of background batches acquired and the number in the rear part is the total number of patches obtained through data enhancement. It can be seen that as the number of patches increases, the number of backgrounds acquired is very insufficient. The data augmentation method used to increase the number of background classes is random rotation, size, and translation.

FIG. 16 shows a proposed patch-based CNNs model structure using a Fire module according to an embodiment of the present invention, and FIG. 17 shows a patch-based CNNs structure according to an embodiment of the present invention.

16 and 17, in the feature extraction unit 120, the CNNs model can apply a fire module used in SqueezeNet to be suitable for processing a patch, as shown in FIG. Since SqueezeNet is designed to process images with input size of 224 × 224, it is rapidly reduced to 55 × 55 by applying convolution and max spooling with a movement interval of 2 immediately upon the start of the network. When applying the patch method, the size is small, so there is no need to quickly reduce the size. Therefore, the proposed structure can take the modified structure of SqueezeNet as shown in FIG. 17 when the patch size is 32 × 32.

While the first convolution of the existing SqueezeNet was a 7 × 7 convolution of a moving 2 value, the structure in the feature extraction unit 120 performs a convolution of 3 × 3 moving 1. This is because the size of the input is small, so there is no need to apply a large filter such as 7 × 7, and the number of parameters can be reduced.

The feature extraction unit 120 also does not perform pooling to reduce the loss of information to the fire2 module, and the existing SqueezeNet performs a total of 5 poolings, whereas the proposed SqueezeNet transformation performs only a total of 3 poolings. . Leaky ReLU is applied as the activation function used in the network, and batch normalization can be performed before applying the activation function. The total number of parameters of the SqueezeNet modified in the feature extraction unit 120 is 536,143, which is less than the SqueezeNet of 738,594.

18 is a comparison of an existing method and a method according to an embodiment of the present invention.

18 (a) and (b), a method of using an existing patch extracts patches after making an input image into a grid. This method of processing will be referred to as a grid method for convenience. The lattice method determines whether each of the patches is forged, and then determines whether or not the final forgery is done by a voting method, and this method is as shown in FIG. Unlike this, in the feature extraction unit 120, the method of learning CNNs using the patches is the same, but in the test, it is processed in a completely convolutional form without extracting the patches, as shown in FIG. 18 (b). The result processed by the method of FIG. 18 (b), which is a processing method in the feature extraction unit 120, is a forgery, a living body, a 3D tensor of a background, and fingerprints using the probability of living body and forgery, except for the result classified as the background. The overall bio-probability can be determined.

The probability of the living body determined by the feature extraction unit 120 is compared with the optimal threshold value obtained from the learning data to determine whether the final fingerprint is living body. When the fingerprint is processed by the full convolution method rather than the lattice method, a larger number of forgery evaluations can be performed. If the performance of CNNs for classifying patches is good, the more patches are processed, the more network ensemble effects are generated, and the better results can be obtained. In addition, since the background is included in the class, the entire fingerprint can be processed without segmentation.

The feature extraction unit 120 simply counts the number of forged and bio-categorized patches and classifies them into many, instead of ignoring patches classified in the background and normalizing the probability of forgery and bio-failure of each patch, and then synthesizing these probabilities. How to do. The method for determining the optimum threshold value of the feature extraction unit 120 may be performed as follows. Among the three-dimensional tensors obtained by using the full convolution method, the part with the highest probability of background is removed. The rest of the results, except for the background, are normalized again using only forgery and bio-probability. This is equal to the probability of bioavailability at the local location of the fingerprint, excluding the background. The average bio-probability at each location thus obtained is the bio-score of the final fingerprint. This process was described in the following algorithm.

The method for determining the optimum threshold value in the feature extraction unit 120 is determined as the minimum average classification error (ACE) value in the training data. After performing the above algorithm on the entire training data, bio-probabilities are calculated using a complete convolution method. The process of acquiring the optimal door touch value when the score calculated through the algorithm is set to is calculated through the following equations.

[Equation 2]

[Equation 3]

[Equation 4]

Equation 2 means a ratio in the case where the Livescore calculated through the algorithm is s, but is classified as a bio-fingerprint beyond the threshold value t despite being a fake fingerprint. Equation 3, on the contrary, refers to a ratio classified as forgery even though it is a biological fingerprint. Equation (4) means that we find that we can determine the average of Ferrfake and Ferrlive to a minimum. The determined t * is applied when testing the performance of counterfeit fingerprints. Ferrfake and Ferrlive refer to the error rate of forgery fingerprints and the error rate of biological fingerprints, respectively.

19 shows changes in the values of Ferrfake and Ferrlive according to threshold values of LivDet2015 CrossMatch data.

19 is an example of a change in values of Ferrfake and Ferrlive according to threshold values of LivDet2015 CrossMatch data.

20 illustrates hyperparameters used for learning a patch-based network according to an embodiment of the present invention.

Referring to FIG. 20, in the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention, the GPU used NVIDA GTX 1080, and the parameters applied to learning were set to a learning rate of 0.0005, a batch size of 64, and a generation number of 60. Can be. Adamax was applied as the optimization method, and 10% of the learning LivDet data was used for verification. By verifying the loss value of verification data for each generation, if the loss value does not decrease during 4 generations, the learning rate is doubled. The hyperparameters used in the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention are the same as the previous experiment except for the batch size and the number of generations, and the learning data is randomly mirrored in batch units. The hyperparameters used in the experiment are summarized in FIG. 20.

Two methods were performed: when the patch was divided into a grid method that does not allow overlapping, and when the patch was applied as a full convolution method. Both the full convolution method and the lattice method use the same parameters. Both methods were tested for 32 × 32, 48 × 48, and 64 × 64 patch sizes to verify performance by changing patch sizes. For the performance evaluation of the proposed patch-based network, the results of applying the patch-based method using SqueezeNet are presented together.

21 is a 32 × 32 patch-based counterfeit fingerprint detection performance comparison, and FIG. 22 is a graph comparing the 32 × 32 patch-based counterfeit fingerprint detection performance.

In the case of the 32 × 32 patch, an ACE value of forgery fingerprint detection as shown in FIG. 21 was obtained, and this was graphically represented in FIG. 22. The method proposed in FIG. 21 is expressed as Modified SqueezeNet because the fire module of SqueezeNet is used. In the legend of FIG. 22, Squeeze-Grid and Squeeze-FCC mean the results of the grid method and the full convolution method when SqueezeNet is applied, respectively. In addition, M-Squeeze-Grid and M-Squeeze-FCC refer to the results when the CNNs structure used in the feature extraction unit 120 of the present invention is tested by a grid method and a full convolution method. 21, it can be seen that the network of the present invention exhibits better performance than when the general SqueezeNet is applied. In the case of SqueezeNet, the grid method performed better than the full convolution method.

It is considered that in the case of SqueezeNet, the classification result of each patch is not good because the learning is progressed after the size has been drastically reduced, and more errors have occurred because the test is performed with a completely convolutional structure different from the learning. In the case of the Modified SqueezeNet used by the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention of FIG. 21, it showed better performance when performed using a full convolution layer. It can be estimated that the result of more patches can be obtained in a good performance state.

23 is a 48 × 48 patch-based counterfeit fingerprint detection performance comparison, Figure 24 is a 48 × 48 patch-based counterfeit fingerprint detection performance comparison.

25 is a 64 × 64 patch-based counterfeit fingerprint detection performance comparison, and FIG. 26 is a graph comparing the 64 × 64 patch-based counterfeit fingerprint detection performance.

In the case of the 48 × 48 patch, the trend is similar to that of 32 × 32, but it can be confirmed through FIGS. 23 and 24 that the performance has been improved on average compared to 32 × 32. The 64 × 64 patch test results are shown in FIGS. 25 and 26.

FIG. 27 compares performance according to patch size, and FIG. 28 is a graph comparing performance according to patch size.

Since the number of data used in the experiment is large, the performance according to the patch size on average is shown in FIG. 27, and the graph is shown in FIG. 28. Overall, it can be seen that it is better to transform SqueezeNet to the proposed structure than to use SqueezeNet as it is. In addition, among the models used by the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention, the best performance is obtained when the CNNs structure is applied in a completely convolutional form and the patch size is 48 × 48. I can confirm it

FIG. 29 compares the results of applying the voting method and the optimal threshold value, and FIG. 30 compares the performance of the optimal threshold method and the voting method.

29 and 30, the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention determines an optimal threshold value of patches from learning data. In this embodiment, the optimal threshold value is determined and the result when applying it and the number of results of the classified patches are simply counted, and the voting method and performance comparison are performed to classify the results in more ways, and the results are compared. It is summarized in 29.

Looking at the results, it can be seen that, on average, in most cases, full convolution is used, and it is helpful to improve performance by determining and applying an optimal threshold value for each data rather than a simple voting method. In FIG. 30, the average ACE of the optimal threshold method and the voting method is represented graphically. When the optimal threshold value is applied, it can be seen that the simple voting method tends to decrease in performance as the patch size increases, and when the optimal threshold value is applied, it can be seen that it shows the best performance in the model trained at 48 × 48 size.

FIG. 31 compares performance with a model to which data augmentation is applied, and FIG. 32 compares performance when data augmentation is applied.

Referring to FIGS. 31 and 32, the performance when data augmentation is applied to a fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention is illustrated in FIG. 31. The applied data augmentation method is a method of applying vertical symmetry in addition to arbitrary mirror symmetry. The FCN in FIG. 31 is a proposed model without data augmentation, and Aug-FCN is a model with data augmentation. After learning the 48 × 48 patch that showed the best performance and applying the full convolution method, it can be seen that data augmentation through vertical symmetry on average does not greatly help performance improvement. FIG. 32 comprehensively expresses the results of FIG. 31 through a cumulative graph. It can be seen that images acquired in a controllable environment such as fingerprints have difficulty in exerting a large effect on data augmentation.

33 shows a DET curve of a method according to an embodiment of the present invention.

Referring to FIG. 33, among the embodiments of the fingerprint recognition system capable of detecting forgery fingerprints according to an embodiment of the present invention, the DET (Detection Error) of the model that learned the 48 × 48 patch of FIG. 31 through data augmentation Trandeoff) curve, you can get the result shown in Figure 33.

Fig. 34 compares the processing speed of the modified SqueezeNet (unit: ms).

Referring to FIG. 34, the grid method is performed by cutting the patches without allowing the overlapping of fingerprints and grouping the patches in batches. Full convolution is calculated in the form of a general convolution, and the final result appears in a two-dimensional form. The feature extraction unit 120 of the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention shows a somewhat slower processing speed than the circular SqueezeNet, which is shown in FIG. 34.

The complete convolution method using the optimal threshold value in the feature extraction unit 120 is represented by FCN in FIG. 34 and can process one fingerprint at an average rate of about 124 ms. When experimenting with LivDet data, the fingerprint image is processed at a rate of about 37ms when it is the smallest, and at the rate of about 397ms when it is the largest. SqueezeNet quickly reduces the size of the input tensor at the beginning of the layer, so it can be processed faster than the proposed model structure.

FIG. 35 compares the processing speed of SqueezeNet and the modified SqueezeNet according to an embodiment of the present invention, and FIG. 36 compares the speed of the lattice method and the complete convolution method of the CNNs structure according to an embodiment of the present invention. will be.

35 and 36, when full convolution is applied, since the convolution operation is performed regardless of the size of the patch, the processing speed is the same even if the size of the patch used for learning changes. In the case of processing in a lattice method, the larger the size of the patch, the smaller the number of patches obtained from the fingerprint, so that a faster processing speed can be obtained. The result of comparing the average processing speed of the original SqueezeNet and the proposed network for each patch size is shown in FIG. 35. The M-Squeeze-Grid of FIG. 35 applies the proposed method as a grid structure, and the M-Squeeze-Conv. Applies the proposed method in a complete convolution form. Since the average processing speed is faster than the lattice method except 64 × 64, it can be seen that it is more advantageous to process in a full convolution method in terms of classification performance and time. Unlike the existing forgery fingerprint algorithm, the proposed method does not require a process such as segmentation by including the background as a class, so the processing speed of the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention will be actually processed faster. . In the fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention, the processing speed when the lattice method and the complete convolution method are applied is represented as a cumulative graph for each data, and is represented in FIG. 36.

37 is a result of partially observing the characteristics of a forgery fingerprint using full convolution according to an embodiment of the present invention.

Referring to FIG. 37, when a fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention tests a learned parameter by a complete convolution method, it is possible to confirm a detection result at a specific location of each fingerprint. 37 shows the results of application of the full convolution method. The two fingerprints in FIG. 37 are examples of visualization. Green indicates the probability of forgery, red indicates the probability of bio-fingerprint, and blue indicates the probability of the background, and these are respectively recorded as probability in the RGB channel. Since the fingerprint is not simply classified into a forgery and a living body, but classified into a forgery, a living body, and a background, a result as shown in FIG. 37 can be obtained. Visualization is an enlargement of the result of the reduced size through the complete convolution method to the original image size through linear interpolation. Therefore, the fingerprint of the Sagem sensor of LivDet2011 having a size of 384 × 352, which is the upper fingerprint of FIG. 37, appears smeared. The lower fingerprint in FIG. 37 is a fingerprint obtained from the Italdata sensor of LivDet2011, and the original size is 480 × 640. The black line in FIG. 37 is shown for comparison with the grid type processing method, and shows a size of 48 × 48.

In general, when a person judges whether to falsify by looking at a fingerprint, there is a tendency to judge that it is a forgery when black and white lumps appear to be crushed and unnaturalness on the outside of the fingerprint. FIG. 37 shows that this analysis, which can be experienced by a person, is somewhat possible using visualization through full convolution. In the case of the fingerprint on the upper part of FIG. 37, it can be seen that the fingerprint reacted as a living body with a red color in the place where the fingerprint appeared relatively clearly, and it was judged as a forgery in the blurry part and the outline of the upper part of the fingerprint. Also, if you look closely at the background part, it can be confirmed that the background part is not blue because there are many remaining fingerprints remaining in the sensor, so there are many gray parts. It can be seen that the fingerprint at the bottom of FIG. 37 is also found to be mainly forged in the black smear and outer portions of the fingerprint. Also, because it has a relatively clean background, it can be confirmed that the background part is completely blue.

38 is an example of the case in which all of the same counterfeit fingerprints are classified as biometric fingerprints.

Referring to FIG. 38, when looking at the error result of the data, it can be seen that the result is not constant. Although it is possible to evaluate the performance of the network to a certain extent by averaging the detection errors of each data, it is not easy to find the consistency of the results when applied in practice. This is because the fingerprint data is not made with a unique finger, but the same person creates multiple forgeries with the finger, indicating a large detection error even if the parameter is slightly changed. FIG. 38 is an example of a case in which a false fingerprint is mistaken as a bio fingerprint in Italdata of LivDet 2011. In fact, all the fingerprints present in FIG. 38 are fingerprints of the same finger. As can be seen in FIG. 38, if errors with similar probability are properly classified due to minute changes in parameters and thresholds during the learning process, fingerprints such as FIG. 38 may all be accurately classified, and the detection rate may be discontinuous. It is caused by raising.

In the experiments to date, only counterfeit fingerprints made of the same kind of material were used for both learning and testing of counterfeit fingerprints. However, in general, in a fingerprint recognition system capable of detecting a forgery fingerprint, an attack of a forgery fingerprint made of a material not used for learning may occur.

LivDet2015 includes counterfeit fingerprints consisting of substances not used for learning in the test data. Therefore, we can test how high the generalization performance of the proposed CNNs models is.

39 shows a result of generalization performance evaluation of methods according to an embodiment of the present invention.

Referring to FIG. 39, data not used in the learning of LivDet2015 was tested in order to examine the generalization possibility of models used by a fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention. The results are shown in FIG. 38. For the objectification of performance, the results of the organizations participating in the LivDet2015 competition are also displayed. The first row of FIG. 39 means the sensor of LivDet2015, and the shaded column is the name of the group or individual participating in the competition.

In order to measure generalization performance, not only the patch-based method, but also the segmentation-based region-based method and the Gram matrix-based method were tested together. The proposed segmentation-based SqueezeNet is shown in Are-SQZ in FIG. 39, the Gram matrix-based model is Gram, and the patch-based model is PatchmSQZ. The fingerprint recognition system capable of detecting a forgery fingerprint according to an embodiment of the present invention uses a model having the best performance in the previous experiment, and it is expressed in bold in FIG. 39. Looking at the results, we can see that all of the proposed methods can classify counterfeit fingerprints that are not used for learning with high accuracy.

In the fingerprint recognition system capable of detecting forgery fingerprints according to an embodiment of the present invention, a system for applying high classification performance of CNNs to software-based forgery fingerprint detection is proposed. In general, the study of applying CNNs to forgery fingerprint detection simply applies the CNNs model used for general object classification, but in the case of fingerprint images, there is a need to construct a network suitable for this because the characteristics of the acquired image and the process and images are different. . In addition, in order to be applied to the authentication system of a mobile terminal, the number of parameters of CNNs must be small, and in order to be used universally, a network capable of operating regardless of the size of the input image must be constructed. In order to build a network that satisfies this, in the fingerprint recognition system capable of detecting forged fingerprints according to an embodiment of the present invention, a CNNs model was constructed with two approaches: a region-based forgery fingerprint detection and a patch-based forgery fingerprint detection method.

In the case of the first domain-based forgery detection, there are two methods: segmentation and Gram matrix. The method of using segmentation extracted six fixed-sized 224 × 224 areas from the segmented fingerprint area without image reduction, and then applied SqueezeNet and Separable SqueezeNet to calculate the average of the classified results as the forgery fingerprint detection result. As a result, it showed an average error of 2.96% and 3.82%, respectively, and it was able to prove that it showed good performance corresponding to the result of the existing State-Of-The-Art with a small number of parameters. In the method using the Gram module, we present an end-to-end CNNs model capable of processing input images of any size. The Gram module is composed of 1x1 convolution, tanh activation function, and Gram layer, and can convert the input into a Gram matrix of the desired size. As a result of combining the Gram module and the fire module, it was proved that an optimized CNNs model of End-to-End with an average error of 2.61% was designable.

In the second patch-based forgery fingerprint detection, the fingerprint image was divided into 32 × 32, 48 × 48, and 64 × 64 sized patches, and modified SqueezeNet suitable for the patch was applied, and the results of the forgery fingerprint detection according to each patch size were presented. We propose a model with less information loss while reducing the number of parameters by reducing the pooling stage of the SqueezeNet structure for processing patches from 5 to 3 and reducing the size of the convolution filter. The generated model operates in a completely convolutional form, shows a two-dimensional estimation result, and sets an optimal threshold from training data to improve performance. As a result of experiments, we found the optimal threshold value by learning the 48 × 48 patch using data augmentation, and then classified the forgery fingerprint in the form of full convolution. Could.

It is possible to process fingerprints of various sizes without changing the size of the input fingerprint image, which is a design constraint proposed by the fingerprint recognition system capable of detecting forgery fingerprints according to an embodiment of the present invention, and satisfies the design condition that the number of parameters is less than SqueezeNet. . In the future, we will visualize the characteristics of counterfeit fingerprints for each material through analysis of experimental results and test the operation on mobile devices. In addition, since it was confirmed that data augmentation does not lead to a large performance improvement, it is possible to improve the performance of detecting forgery fingerprints by applying the Generative Adversarial Networks model.

40 is a flowchart of a fingerprint recognition method capable of detecting forged fingerprints according to an embodiment of the present invention.

Referring to FIG. 40, a fingerprint recognition method capable of detecting a forgery fingerprint according to an embodiment of the present invention may include obtaining a user's biometric information (S4010).

In step S4010, the sensor module 110 may acquire biometric information of each individual. The sensor module 110 may recognize and obtain biometric information of the user's fingerprint. The biometric information may include fingerprints, voices, irises, DNA, face types, and the like.

The fingerprint recognition method capable of detecting a forgery fingerprint according to an embodiment of the present invention may include the step of performing quality measurement and data pre-processing of the biometric information, extracting features, and determining whether to falsify (S4020).

In step S4020, the feature extraction unit 120 may measure the quality of the bio-information and perform data pre-processing, feature extraction and forgery determination. In the feature extraction unit 120, the size of the patch may have one of 32X32, 48X48, and 64X64 pixels. Preferably, the feature extraction unit 120 may have a size of the patch 48X48 pixels.

In step S4020, the feature extraction unit 120 may determine whether to counterfeit the plurality of patches individually. The feature extracting unit 120 may determine whether the entire biometric information is counterfeit or biometric information by synthesizing the result of determining whether the plurality of patches are forged. The feature extracting unit 120 may determine whether to falsify based on a threshold value determined by learning. The feature extraction unit 120 may extract a gram matrix from the obtained biometric information and use it to determine whether or not it is forged. The feature extraction unit 120 may have a size of the gram matrix of 128x128.

The fingerprint recognition method capable of detecting a forgery fingerprint according to an embodiment of the present invention may include generating a matching score of a pre-stored template and the obtained biometric information and performing authentication (S4030).

In step S4030, the matching unit 130 may generate a template stored in the database unit 140 and a matching score of the biometric features acquired by the feature extraction unit 120. The matching unit 130 may determine whether to authenticate according to the matching score of the biometric information. The matching unit 130 may perform the operation of authenticating the user according to the matching score of the biometric information.

The fingerprint recognition method capable of detecting a forgery fingerprint according to an embodiment of the present invention may include storing a template using biometric information for authentication (S4040).

In step S4040, the database unit 140 may store biometric data as a template through a biometric registration process. The database unit 140 may store the results learned by the feature extraction unit 120. The database unit 140 may store data for authentication and data for falsification.

So far, the present invention has been focused on the preferred embodiments. Those skilled in the art to which the present invention pertains will appreciate that the present invention may be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered in terms of explanation, not limitation. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the equivalent range should be interpreted as being included in the present invention.

Claims (13)

A sensor module for acquiring user biometric information;
A feature extraction unit that performs quality measurement and data pre-processing of the biometric information, extracts features, and determines whether to falsify;
A database unit storing a biometric information template for authentication; And
It includes; a matching unit for generating a matching score of the template and the obtained biometric information and performing authentication;
The feature extracting unit is a fingerprint recognition system capable of detecting a forgery fingerprint, characterized in that after dividing the obtained biometric information into a plurality of patches, it is determined whether to extract the feature and forgery. According to claim 1,
The feature extraction unit,
The size of the patch has a size of any one of 32X32, 48X48, 64X64 pixels Fingerprint identification system capable of detecting a forgery fingerprint.
According to claim 2,
The feature extraction unit,
A fingerprint recognition system capable of detecting a forgery fingerprint, wherein the patch has a size of 48X48 pixels. According to claim 1,
The feature extraction unit,
A fingerprint recognition system capable of detecting a forgery fingerprint that individually determines whether to falsify a plurality of the patches. The method of claim 4,
The feature extraction unit,
A fingerprint recognition system capable of detecting a forgery fingerprint by determining whether the whole of the biometric information is counterfeit or biometric information by synthesizing the result of determining whether or not the plurality of patches are forged. The method of claim 5,
The feature extraction unit,
A fingerprint recognition system capable of detecting a forgery fingerprint based on a threshold determined by learning to determine whether or not it has been forged. According to claim 1,
The feature extraction unit,
A fingerprint recognition system capable of detecting a forgery fingerprint, characterized in that a gram matrix is extracted from the obtained biometric information and used to determine whether or not it has been forged. The method of claim 7,
The feature extraction unit,
A fingerprint recognition system capable of detecting a forgery fingerprint, wherein the size of the gram matrix is 128x128. Step (a) of obtaining the user's biometric information;
Step (b) of performing the quality measurement and data pre-processing of the bio-information, extracting features, and determining whether to falsify;
(C) generating a matching score of the pre-stored template and the obtained biometric information and performing authentication; And
Including (d) step of storing the template using biometric information for authentication;
In step (b), the obtained biometric information is divided into a plurality of patches, and then feature extraction and forgery determination are performed.
The method of claim 9,
Step (b) is,
A fingerprint recognition method capable of detecting a forgery fingerprint that individually determines whether to falsify a plurality of the patches. The method of claim 10,
Step (b) is,
A fingerprint recognition method capable of detecting a forgery fingerprint by deciding whether to falsify information or biometric information of the entire biometric information by synthesizing the results of determining whether or not the plurality of the patches are forged.
The method of claim 9,
Step (b) is,
A fingerprint recognition method capable of detecting a forgery fingerprint, characterized by extracting a gram matrix from the obtained biometric information and using it to determine whether or not it is forged.
The method of claim 12,
Step (b) is,
A fingerprint recognition method capable of detecting forged fingerprints, wherein the size of the gram matrix is 128x128.

Images