A Capsule Network-Based Method for Detecting Finger Vein Impersonation Attacks

Abstract

This paper proposes a method for detecting finger vein spoofing attacks based on capsule networks. Capsule networks are not only suitable for small-sample finger vein datasets, but also, by replacing neurons with vector-represented capsules on top of CNNs, they can better handle spatial information such as relative position and angle, enhancing the network's adaptability to finger offset and rotation scenarios. A Bayesian routing algorithm is proposed, incorporating the differential entropy of the capsules as a consideration in calculating activation values. During final classification, feature capsules with high activation probabilities and high concentration are selected, which helps improve the accuracy of classifying genuine and fake veins. By simulating the uncertainty of capsule parameters, training errors can be reduced, improving recognition accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of biometric recognition, and particularly relates to a finger vein forgery attack detection method based on a capsule network. BACKGROUND

[0002] Finger vein recognition technology is a new biometric technology, and has attracted extensive attention from the academic and industrial circles due to its advantages such as living body recognition, non-contact, and difficulty in being interfered by external factors. In a finger vein recognition system, vein feature information is the only basis for the system to verify the identity of a user, and once the information is leaked, the finger vein recognition system may be threatened by a forgery attack. The so-called forgery attack refers to that an attacker uses a certain way to fake or impersonate a real vein according to the stolen information, so as to cheat the recognition system. Common fake ways include direct display by a smart phone, simulation of a finger by using beeswax and printed vein patterns, etc.

[0003] In view of the above risks, researchers have proposed various forgery attack detection (PAD) methods, which can be divided into two categories of living body signal detection and texture analysis. The method based on living body signal detection mainly distinguishes real and fake veins by detecting the activity or life signal of a finger. The method has high accuracy and reliability, but often needs to rely on additional equipment or consume a large amount of system and computing resources. The method based on texture analysis mainly uses the imaging quality difference between real and fake vein images in the texture and noise level to distinguish them, and has relatively low cost, so it is more practical. This kind of algorithm includes Fourier Spectral Bandwidth Energy (FSBE), Binarized Statistical Image Features (BSIF), Monogenic Scale Space (MSS), etc. However, these methods are not good at processing noise, so the feature extraction is not accurate enough. In recent years, researchers have introduced Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) in deep learning to train and identify real and fake vein images, and have achieved certain results. However, neural networks usually need a large number of training samples to achieve good results, while finger vein forgery attacks generally have only a small sample data set, which cannot meet the requirements. In addition, for the case of finger displacement and rotation, neural networks cannot accurately judge and extract features. SUMMARY

[0004] In view of the above defects, the application introduces a capsule network suitable for small sample data sets, and proposes a finger vein forgery attack detection method based on an improved capsule network. The original capsule network cannot fully extract high-level features due to its shallow network architecture, so it performs poorly on complex sample data sets. The original capsule network is improved in the application, and a Bayesian routing algorithm is designed to improve the accuracy of true and false vein classification by improving the measurement of voting consistency between capsules.

[0005] The method is based on a capsule network architecture suitable for small sample data sets, and the routing algorithm is improved. The original capsule network uses an Expectation Maximization (EM) routing algorithm to map between different levels of capsules. On this basis, the method introduces a variational posterior differential entropy into the activation probability to measure the concentration of data points by referring to the Bayesian posterior probability, and calculates the final voting result by comprehensively considering the activation probability and the concentration of data point classification to improve the routing algorithm and improve the classification accuracy.

[0006] The finger vein forgery attack detection method based on the capsule network comprises the following steps:

[0007] Step 1: data acquisition and preprocessing; collect finger vein images as real data sets;

[0008] The real vein data set is denoised, cropped, labeled and processed, and then an artificial forgery attack vein product is made and placed on the collection device for secondary collection to obtain a fake vein data set. Finally, the collected real and fake data sets are normalized and preprocessed by extracting the region of interest, and are divided into a training set and a test set;

[0009] Step 2: constructing a capsule network model, which includes a convolutional layer, a main capsule layer and a routing layer;

[0010] The convolutional layer is used to obtain the basic features of the image. After the real and fake data sets are input into the convolutional layer, the basic features of the input image are obtained. The main capsule layer encapsulates the obtained basic features into capsules in the form of vectors for output. The routing layer is composed of three convolutional capsule layers, which complete the clustering of features. The feature capsules of the main capsule layer are input into the routing layer, and after passing through the three convolutional capsule layers, the classification results of the real and fake vein images are output;

[0011] Step 3: training the network model;

[0012] The training set is used to train the model, and the test set is used to test the model to obtain the best capsule network model in terms of recognition performance, and the parameters of the model are saved;

[0013] Step 4: true and false vein classification;

[0014] Input the real or fake finger vein image to be identified into the trained network model to obtain an identification result.

[0015] Advantages of the present application:

[0016] (1) The method introduces a capsule network. The capsule network is not only suitable for small sample finger vein data sets, but also uses a vector represented capsule to replace a neuron on the basis of a CNN, so that spatial information such as relative position and angle can be better processed, and the adaptability of the network to finger offset and rotation scenarios is enhanced.

[0017] (2) The method proposes a Bayesian routing algorithm. The method increases the differential entropy of the capsule as a consideration for calculating the activation value, and finally selects a feature capsule with a higher activation probability and a higher concentration degree, which helps to improve the accuracy of real and fake vein classification.

[0018] (3) The EM routing algorithm of the original capsule network uses maximum likelihood estimation to determine the parameters, and when the training data is small, overfitting may occur, resulting in a large training error. The method can reduce the training error and improve the recognition accuracy by simulating the uncertainty of the capsule parameters. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a finger vein forgery attack detection method flowchart in the embodiment of the present application.

[0020] Figure 2 is a capsule network model architecture diagram in the embodiment of the present application.

[0021] Figure 3 is a Bayesian routing algorithm flowchart in the embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical solutions of the present application will be further described in detail below in conjunction with the drawings of the specification.

[0023] The method flowchart in the embodiment is as shown in Figure 1 , and the specific step flowchart is as follows.

[0024] Step 1: Data acquisition and preprocessing.

[0025] Collect the finger vein real image of the volunteer, perform conventional denoising, cropping, and marking processing, and obtain a real data set. For each real image, print two laser films containing the real finger vein pattern and having the same size. After aligning and overlapping, superimpose them on a blank paper to obtain corresponding forgery attack samples, place them on the acquisition device to collect again, and finally obtain a fake data set. Normalize and extract the region of interest of the collected real and fake vein data sets, and divide the data set into a training set and a test set.

[0026] Step 2: Building the network model. The capsule network model built in this embodiment includes three parts: a convolutional layer, a main capsule layer, and a routing layer.

[0027] (1) Convolutional layer:

[0028] The capsule network convolutional layer performs preliminary training on the input image to obtain a feature map. As shown in Figure 2 , a 5x5 convolutional layer is used for training, with 32 convolutional kernels and a step size of 2. A batch normalization (BN) layer and a ReLU activation function are added after the convolutional layer. Batch normalization helps to speed up convergence and can avoid problems such as gradient vanishing and data distribution imbalance. Using ReLU as the activation function reduces the computational load of the backpropagation algorithm and can alleviate overfitting.

[0029] (2) Main capsule layer:

[0030] The main capsule layer encapsulates the basic features obtained by the convolutional layer into capsules in vector form. As shown in Figure 2 , the feature map is passed through two 1x1 convolutional branches to obtain a pose matrix and an activation probability, respectively, and finally form 32 capsules as output, each capsule having a 4x4 pose matrix and an activation probability.

[0031] (3) Routing layer:

[0032] The capsule network routing layer performs feature clustering based on the designed Bayesian routing algorithm. As shown in Figure 2 , the capsule network routing layer is composed of three convolutional capsule layers, each including a capsule convolution module and a Bayesian routing module. The convolution kernel sizes of the three capsule convolution modules are 3x3, 3x3, and 1x1, respectively, with step sizes of 2, 1, and 1, respectively. In the Bayesian routing module, the capsules before and after the routing process are called child capsules and parent capsules, respectively, and contain low-level feature information and high-level feature information, respectively. The function of the Bayesian routing module is to consider the activation probability of the parent capsule and the classification concentration of the child capsule on the parent capsule, to achieve accurate routing of the child capsule to the parent capsule, thereby mapping low-level features to high-level features, and finally classifying the vein image.

[0033] The working process of the capsule network routing layer is to input the feature capsules formed by the main capsule layer as child capsules into Figure 2 the first convolutional capsule layer, perform feature clustering, and output parent capsules containing high-level feature information; then input the output capsules as child capsules into the second convolutional capsule layer, perform feature clustering according to the same routing algorithm, and output the corresponding parent capsules; then input the capsules output by the second convolutional capsule layer into the third convolutional capsule layer, and generate the final class capsules and their probabilities of being true or false vein images according to the routing algorithm.

[0034] In the capsule network routing layer, the routing algorithm is the core. It updates the probability of local feature mapping to the overall feature according to the consistency of the votes of multiple sub-capsules to the parent capsule, and finally obtains the classification probability. The routing algorithm designed in this embodiment adds the differential entropy of the capsule as a consideration for calculating the activation value. The algorithm balances between the activation probability and the concentration of data point classification, thereby completing the routing mapping of the sub-capsule to the parent capsule.

[0035] The design of the routing algorithm is specifically introduced below. In this embodiment, each sub-capsule i has a 4x4 pose matrix M i and an activation probability a i . Each sub-capsule i votes for the parent capsule j by multiplying the pose matrix M i with the trainable weight matrix W ij . The specific steps of the algorithm are as follows, and the flow is as shown in Figure 3 .

[0036] ①Calculate the voting vector and the pose matrix. Initialize the capsule prior parameters, initialize the weight matrix W ij . Multiply the pose matrix M i of the capsule i with the weight matrix W ij to obtain the voting matrix V j|i of the sub-capsule i to the parent capsule j, and the calculation method is as shown in formula (1). The voting matrix is flattened to obtain the voting vector v j|i . The pose matrix of the parent capsule j can be regarded as the weighted average of all sub-capsule voting matrices, and the calculation method is as shown in formula (2), wherein r ij is the probability of the sub-capsule i routing to the parent capsule j. The pose matrix of each layer of capsules is used to calculate the voting matrix of the next layer of capsules.

[0037] V j|i =M i W ij (1)

[0038] M j =∑r ij V j|i (2)

[0039] ②Update the differential entropy of the variational posterior distribution. This embodiment uses the differential entropy H[q * (μ j ,Λ j )] of the Gaussian-Wishart variational posterior distribution q * (μ j ,Λ j ) of the parent capsule j to measure the voting consistency of the sub-capsule to the parent capsule, wherein μ j is the 4x4 pose matrix M jThe flattened 16-dimensional vector, Λ j H[q] are the point estimation parameters of capsule j. * (μ j ,Λ j Direct calculation is difficult, so Λ is generally used. j The logarithmic expected value E[lndet(Λ j The estimation is performed as shown in formula (3).

[0040]

[0041] in, It is the digamma function, and D is v. j|i The dimension of a vector. θ j It is the precision matrix, which represents the mapping of data points in capsule j, and its determinant is det(θ). j This can be used to measure the concentration of data points; the higher the concentration, the higher the voting consistency for capsule j.

[0042] ③ Update the activation value a of capsule j. j The calculation method is shown in equation (4).

[0043] a j =σ(β) α –(β u +E[lnπ j ]+E[lndet(Λ j )])×r j (4)

[0044] Where, β α and β u These are trainable offset parameters, and the mixing coefficients π. j Let E[lnπ] represent the probability that parent capsule j is selected during the voting process. j [] represents the logarithmic expected value of the mixing coefficients of all child capsules with respect to parent capsule j, r j This is the routing weight. σ() is the sigmoid activation function, which maps activation values ​​to numbers between 0 and 1. In equation (4), the calculation of activation values ​​introduces differential entropy, which represents the degree of concentration of data points in the classification, and can obtain more accurate recognition results.

[0045] ④ Select the parent capsule with the highest activation value for activation. That is, output each child capsule to the parent capsule with the highest activation value.

[0046] In each routing process, the activated parent capsule becomes the child capsule of the next convolutional capsule layer. After three routing steps, the final classification result of the input image is obtained.

[0047] Step 3: Train the network model.

[0048] The model is trained with the training set and tested with the test set to obtain the capsule network model with the best recognition performance, and the corresponding parameters are saved.

[0049] Step 4: True and false vein classification.

[0050] The true and false finger vein image to be identified is input into the trained network model to obtain the recognition result.

[0051] To verify the performance of the method of the embodiment, comparative tests were performed on the following data sets.

[0052] (1) IDIAP FVD data set (Tome P, Vanoni M, Marcel S. On the vulnerability of finger vein recognition to spoofing [C]. International Conference of the Biometrics Special Interest Group (BIOSIG), 2014: 1-10.)

[0053] (2) SCUT FVD data set (Qiu X, Tian S, Kang W, et al. Finger vein presentation attack detection using convolutional neural networks [C]. Chinese Conference on Biometric Recognition, 2017: 296-305.)

[0054] (3) Self-made data set, denoted as OUR_SET.

[0055] The performance evaluation index uses the following three:

[0056] APCER (Attack Presentation Classification Error Rate, presentation attack classification error rate): The proportion of imitated biometric features that are incorrectly identified as legitimate users.

[0057] BPCER (BonaFide Presentation Classification Error Rate, bona fide presentation classification error rate): The proportion of legitimate users that are incorrectly identified as imitated biometric features.

[0058] ACER (Average Classification Error Rate): the average of APCER and BPCER, which is calculated as shown in equation (5).

[0059]

[0060] The comparative algorithm is as follows.

[0061] (1) FPNet (Qiu X, Tian S, Kang W, et al. Fingervein presentation attack detection using convolutional neural networks [C]. Chinese Conference on Biometric Recognition, 2017: 296-305.)

[0062] (2) Dynamic routing capsule network (Sabour S, Frosst N, Hinton G E. Dynamic routing between capsules [J]. Advances in Neural Information Processing Systems, 2017, arXiv: 1710.09829.)

[0063] (3) EM routing capsule network (Hinton G E, Sabour S, Frosst N. Matrix capsules with EM routing [C]. International conference on learning representations, 2018: 1-15.)

[0064] The results are shown in Table 1. As can be seen from Table 1, the APCER, BPCER and ACER of the present application on the three data sets are lower than those of other networks, indicating that the present method is superior to the existing method.

[0065] Table 1

[0066]

[0067]

[0068] In summary, the routing algorithm of the capsule network is improved, the accuracy of the final classification result is improved, and the performance of the impersonation detection has obvious advantages compared with other methods. The method is real and effective.

[0069] The above merely describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above embodiments, but any equivalent modifications or changes made by those skilled in the art according to the disclosed content of the present application shall be included in the protection scope recorded in the claims.

Claims

1. A capsule network-based detection method for fake finger vein attacks, characterized in that: The method comprises the following steps: Step 1: data acquisition and pretreatment; collect finger vein images as a real data set; The real vein data set is denoised, cropped, labeled, and then an artificial counterfeit attack vein is made, which is placed on the acquisition device for secondary acquisition to obtain a counterfeit vein data set; finally, the collected real and fake data sets are normalized and preprocessed by extracting the region of interest, and are divided into a training set and a test set; Step 2: a capsule network model is constructed, which comprises a convolutional layer, a main capsule layer and a routing layer; The convolutional layer is used to obtain the basic features of the image; after the real and fake data sets are input into the convolutional layer, the basic features of the input image are obtained; the main capsule layer encapsulates the obtained basic features into capsules in the form of vectors for output; the routing layer is composed of three convolutional capsule layers, which complete the clustering of features; the feature capsules of the main capsule layer are input into the routing layer, and after passing through the three convolutional capsule layers, the classification results of the real and fake vein images are output; In the routing layer of the capsule network, the routing algorithm is the core, which updates the probability of local feature mapping to the overall feature according to the consistency of the voting of multiple sub-capsules to the parent capsule, and finally obtains the classification probability; the differential entropy of the capsule is added in the routing algorithm as the consideration for calculating the activation value, and the algorithm balances between the activation probability and the concentration degree of data point classification, so as to complete the routing mapping of the sub-capsule to the parent capsule; Step 3: training the network model; The model is trained by using the training set, and is tested by using the test set, so as to obtain the capsule network model with the best recognition performance, and the parameters thereof are saved; Step 4: real and fake vein classification; The real and fake finger vein images to be identified are input into the trained network model to obtain the recognition result.

2. The capsule network based method for fake attack detection of finger vein according to claim 1, characterized in that: In step 1, for each real image, two laser films containing real finger vein patterns and having the same size are printed; after being aligned and overlapped, the two laser films are superimposed on a blank paper to obtain corresponding counterfeit attack samples, which are placed on the acquisition device for secondary acquisition, and finally a counterfeit data set is obtained.

3. The capsule network based method for fake attack detection of finger vein according to claim 1, characterized in that: In the capsule network model of step 2, the convolutional layer comprises one 5x5 convolutional layer, one batch normalization layer BN and a linear rectifier function ReLU.

4. The capsule network based method for fake attack detection of finger vein according to claim 1, characterized in that: In the capsule network model of step 2, the main capsule layer is composed of two 1x1 convolutional branch networks; after the feature map output by the previous layer is input into the main capsule layer, a 4x4 pose matrix and an activation probability are obtained, which are combined to form 32 feature capsules as output.

5. The capsule network based method for fake attack detection of finger vein according to claim 1, characterized in that: In the capsule network model of step 2, the routing layer is composed of three convolutional capsule layers, each of which comprises a capsule convolution module and a Bayesian routing module; the convolution kernel sizes of the three capsule convolution modules are 3x3, 3x3 and 1x1 respectively, and the step sizes are 2, 1 and 1 respectively; in the Bayesian routing module, the capsules before and after the routing process are called sub-capsules and parent capsules respectively, which contain low-level feature information and high-level feature information respectively.

6. The capsule network based method for fake attack detection of finger vein according to claim 5, characterized in that: The Bayesian routing module considers the activation probability of the parent capsule and the classification concentration degree of the sub-capsule to the parent capsule, realizes accurate routing of the sub-capsule to the parent capsule, and thus completes the mapping of the low-level feature to the high-level feature.

7. The capsule network based method for fake attack detection of finger vein according to claim 5, characterized in that: The last convolutional capsule layer shares a weight matrix containing spatial dimension information, and the pose matrix and activation probability of the two classes of real and false capsules are calculated by a Bayesian routing algorithm.

8. The capsule network based method for fake attack detection of finger vein according to claim 5, characterized in that: In the capsule network routing layer, each sub-capsule i has a 4x4 pose matrix M i and an activation probability a i Each sub-capsule i votes for a parent capsule j by multiplying the pose matrix M i by a trainable weight matrix W ij , including the following steps: Step a, compute the voting vector and pose matrix; initialize the capsule prior parameters, initialize the weight matrix W ij ; The pose matrix M of capsule i i With weight matrix W ij Multiplying these matrices yields the voting matrix V from child capsule i to parent capsule j. j|i The calculation method is shown in equation (1); flattening the voting matrix yields the voting vector v. j|i The pose matrix of parent capsule j can be regarded as a weighted average of the voting matrices of all child capsules, and its calculation method is shown in equation (2), where r ij It represents the probability that child capsule i is routed to parent capsule j; the pose matrix of each capsule layer is used to calculate the voting matrix of the next capsule layer. V j|i = M i W ij (1) M j =∑r ij V j|i (2) Step b, updating the differential entropy of the variational posterior distribution; using the Gaussian-Wishart variational posterior distribution q * (μ j ,Λ j ) of parent capsule j * ,Λ j ) of parent capsule j j ,Λ j is the 16-dimensional vector flattened from the 4x4 pose matrix M j of capsule j, and Λ j is the point estimate parameter of capsule j; H[q * (μ j ,Λ j )] is estimated by the log-expected value E[lndet(Λ j )] of Λ j , as shown in equation (3) wherein, is the digamma function, D is v j|i dimension of the vector; θ j is the precision matrix, which represents the mapping of data points by capsule j, whose determinant det(θ j ) is used to measure the concentration of data points, the higher the concentration, the higher the consistency of the vote for capsule j; Step c, updating the activation value a of capsule j j , the calculation method is shown in formula (4); a j = σ(β α – (β u + E[lnπ j ]+ E[lndet(Λ j )]) × r j ) (4) where β α and β u are trainable bias parameters, the mixing coefficient π j represents the probability of parent capsule j being chosen in the voting process, E[ln π j ] represents the log expected value of the mixing coefficient for parent capsule j over all child capsules, r j is the routing weight; σ() is the sigmoid activation function, which maps the activation value to a number between 0 and 1. Step d: select the parent capsule with the highest activation value for activation, i.e., output each child capsule to the parent capsule with the highest activation value.

9. The capsule network based method for fake attack detection of finger vein according to claim 8, characterized in that: In each routing process, the activated parent capsule serves as a child capsule for the next convolutional capsule layer; after three routing processes, the final classification result of the input image is obtained.

Images