Biometric fingerprint recognition method and product based on dual-task learning

Abstract

The present invention discloses a biometric fingerprint recognition method based on dual-task learning, which utilizes a dual-task neural network to extract fingerprint forced attack features from normal fingerprint images and forced fingerprint images, and simultaneously uses the same network to extract user individual features, utilizes a neural network to build and train a fingerprint forced press recognition dual-task model, utilizes the trained model and a conventional fingerprint feature recognition method to perform parallel detection on the acquired fingerprint, and only when the conditions of user information being in the database and forced press detection being normal are met at the same time, the device is allowed to complete fingerprint unlocking, thereby supplementing the vulnerability of previous fingerprint recognition devices facing biometric fingerprint forced press recognition attacks, and simultaneously adding a basic individual recognition function that shares weights with forced press recognition, thereby providing strong protection for biometric fingerprint recognition while completing basic individual recognition, thus saving resource consumption.

Description

Technical Field

[0001] The present invention belongs to the technical field of biometric fingerprint recognition security, and specifically relates to a biometric fingerprint recognition method and product based on dual-task learning. Background Art

[0002] The rapid development of biometric fingerprint recognition technology has achieved unprecedented success in user privacy and security, and its widespread commercial use has significantly improved the way people protect their privacy and property. Fingerprint technology is used to protect the privacy of mobile phones and in door locks in every household. However, fingerprint recognition technology still has vulnerabilities. Currently, large-scale commercial fingerprint recognition devices only extract the ridge features of a user's fingerprint to distinguish legitimate users, but fail to account for situations where an unauthorized intruder grabs the legitimate user's hand and forces them to unlock the device. With current fingerprint recognition technology, the fingerprint module can be unlocked regardless of whether the fingerprint is forcibly pressed by an attacker or pressed voluntarily by the user.

[0003] Patent document CN202210170301.3 discloses how to implement forced press detection, but it is limited to manual feature extraction and has certain accuracy limitations. Furthermore, patent document CN202210170301.3 does not consider user identification. Therefore, how to detect forced presses while ensuring fingerprint identity recognition has become a hot topic among those skilled in the art. Summary of the Invention

[0004] The present invention is aimed at the existing problems and discloses a biometric fingerprint recognition method based on dual-task learning. This method uses a dual-task neural network to extract fingerprint forced attack features from two types of fingerprint images, and uses the same network to extract user individual features. Finally, the trained model and conventional fingerprint feature recognition method are used to perform parallel detection on the acquired fingerprints. At the same time, the conditions of user information being in the database and the forced press detection being normal are met before the device is allowed to complete fingerprint unlocking. This makes up for the vulnerability of previous fingerprint recognition devices facing biometric fingerprint forced press recognition attacks, and at the same time adds a basic individual recognition function that shares weights with forced press recognition, providing strong protection for biometric fingerprint recognition while completing basic individual recognition and saving resource consumption. This patent document uses the dual labels of personal identity labels and forced action labels in the fingerprint database, so that the established neural network model can train two tasks at the same time. The shared weights of the dual-task model not only add the user authentication function to the algorithm, but also save additional computing resources for model establishment.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: a biometric fingerprint recognition method based on dual-task learning, comprising the following steps:

[0006] S1, normal data acquisition: record legal and non-forced fingerprints on the fingerprint unlocking device, and mark each normal fingerprint with a double label of personal identity and non-forced;

[0007] S2, forced data acquisition: when a biometric fingerprint unlocking operation is recorded on a fingerprint unlocking device, the attacker forces the unlocking fingerprint to be used, and each forced fingerprint is labeled with both personal identity and forced;

[0008] S3, feature enhancement: filtering the forced fingerprint image obtained in step S2 and enhancing the forced fingerprint features in the fingerprint; the fingerprint feature enhancement is achieved by Fourier transform, wavelet transform and OTSU segmentation algorithm;

[0009] S4, dual-task model establishment: A dual-task model for fingerprint forced press recognition is constructed using a neural network. The model includes an input layer, 48 hidden layers, and an output layer. The hidden layers are composed of pooling layers, convolutional layers, and fully connected layers. The output layer includes two types: a user identifier for outputting the user identity corresponding to the fingerprint image, and a forced action identifier for determining whether the fingerprint image has been subjected to a forced press attack. The fingerprint image processed in step S3 is used as the input of the model. The features extracted by the network are subjected to dual-task classification, and the two labels of the image are used to simultaneously perform forced attack recognition and user identity recognition.

[0010] S5, model parameter optimization: using the established fingerprint library with dual labels to train the fingerprint forced press recognition model in step S4, so as to continuously optimize and update the model parameters;

[0011] S6, Identification Processing: Fingerprint identity recognition and forced fingerprint attack recognition are processed in parallel. While identifying the fingerprint identity, it is detected whether the fingerprint is forced. If the fingerprint is in a forced state, the fingerprint recognition will be rejected regardless of whether the fingerprint is in the database, and a warning will be given through the network report; if the fingerprint is in a normal state, fingerprint identity recognition will be carried out normally, completing the dual tasks of defending against fingerprint forced press attacks and identity recognition.

[0012] As an improvement of the present invention, the fingerprint enhancement in step S3 is specifically as follows:

[0013] Input an M×N fingerprint image, represented by L-level grayscale as [1,2,…,L-1], N i Represents the number of pixels with grayscale i, and the total number of pixels in the image is:

[0014]

[0015] The probability of gray level i appearing is:

[0016]

[0017] Select a threshold k (0 < k < k - 1), which divides the input image into two classes, namely C1 and C2. C1 consists of pixels with gray values in the range [0, k], and C2 consists of pixels with gray values in the range [k + 1, L - 1]; then the probability that a pixel is classified in C1 is:

[0018]

[0019] Similarly, the probability that a pixel is classified in C2 is:

[0020]

[0021] The average gray value of C1 is:

[0022] [[ID=1�]]

[0023] Similarly, the average gray value of C2 is:

[0024]

[0025] The global variance is: <000\062>

[0027] The greater the average difference between m1(k) and m2(k), the greater it is, and the image is segmented and feature enhanced based on this.

[0028] As another improvement of the present invention, the filtering in step S3 at least includes average filtering, Gaussian filtering, and median filtering.

[0029] As another improvement of the present invention, in step S5, the optimization of the model parameters depends on the minimization of the dual-task model joint loss function, that is:

[0030]

[0031] Where, respectively represent the loss function of personal identity recognition and the loss function of forced recognition, W is the model weight, and the noise parameters are σ1 and σ2.

[0032] To achieve the above purpose, the technical solution adopted by the present invention is also: a product of a biometric fingerprint recognition method based on dual-task learning, including a computer program, and when the computer program is executed by a processor, it realizes the steps of the biometric fingerprint recognition method based on dual-task learning.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] (1) This method uses a dual-task model to extract fine-grained features of each user from the database based on the user's daily unlocking habits rather than fixed thresholds or established criteria, preprocesses the images, and enhances the fine-grained features.

[0035] (2) This method utilizes the dual labels of personal identity labels and forced action labels in the fingerprint database, so that the established neural network model can train two tasks at the same time. The shared weights of the dual-task model not only add the user authentication function to the algorithm, but also save the computing resources of additional model building, realize multi-user recognition and forced press detection, and enhance the security of biometric fingerprint recognition.

[0036] (3) The present invention provides a biometric fingerprint recognition method based on dual-task learning. Compared with the existing technology, it performs forced action detection and user identification at the same time. It is safer and more efficient than single fingerprint recognition and single forced action detection. When establishing a database, this method marks each picture with a dual label of personal identity and non-forced, and uses a joint loss function to continuously update and optimize the model parameters during training. At the same time, the fingerprint image is preprocessed, and a segmentation algorithm is used to make the forced features more obvious. The test results show that the present invention can effectively detect whether the user is legitimate and whether the user is forced when performing biometric fingerprint recognition, preventing illegal users from unlocking the fingerprint while also preventing illegal attackers from using coercive means to hold down the hand of a legitimate user to unlock the fingerprint. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flowchart of the steps of the biometric fingerprint recognition method based on dual-task learning of the present invention;

[0038] Figure 2 This is a workflow diagram of the biometric fingerprint recognition method product based on dual-task learning of the present invention;

[0039] Figure 3 It is a performance test diagram of the method of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0041] Example 1

[0042] Biometric fingerprint recognition methods based on dual-task learning, such as Figure 1 As shown, the following steps are included:

[0043] Step (1) Legitimate users A and B input a set of legally obtained non-forced pressed fingerprints on a fingerprint unlocking device, and attach dual labels of personal identity and non-forced to these fingerprints. When legitimate users A and B actively conduct biometric fingerprint recognition, it is required that the fingerprint recognition can be completed normally.

[0044] Step (2) When legitimate users A and B perform biometric fingerprint unlocking operations, attacker C forces them to perform fingerprint unlocking, inputs the fingerprint images of legitimate users who have suffered forced pressing attacks, and attaches dual labels of personal identity and forced to these fingerprints. When C grabs the hands of legitimate users A and B and forcibly presses the fingers used by A and B for fingerprint recognition on the fingerprint recognition device, legitimate users A and B will resist to a certain extent.

[0045] Step (3) Preprocess the fingerprint images and enhance the fingerprint features of forced fingerprints;

[0046] The methods used for image filtering in the above steps include average filtering, Gaussian filtering, and median filtering. The algorithms used for fingerprint feature enhancement include Fourier transform, wavelet transform, and OTSU segmentation. The method for enhancing to distinguish forced pressing is as follows:

[0047] First, input a fingerprint image of M×N, and represent it with L-level grayscale as [1, 2, …, L - 1]. N i Represents the number of pixels with grayscale i. The total number of pixels in the image can be expressed as:

[0048]

[0049] The probability of the occurrence of grayscale i is:

[0050]

[0051] Assume that a threshold k (0 < k < k - 1) is selected, which divides the input image into two categories, C1 and C2. Among them, C1 consists of pixels with grayscale values in the range of [0, k], and C2 consists of pixels with grayscale values in the range of [k + 1, L - 1]. The probability that a pixel is classified in C1 is:

[0052]

[0053] Similarly, the probability that a pixel is classified in C2 is:

[0054]

[0055] The grayscale mean value of C1 is:

[0056]

[0057] Similarly, the grayscale mean value of C2 is:

[0058]

[0059] Assume that the global variance is:

[0060]

[0061] The larger the average difference between m1(k) and m2(k), The larger the image is, the more image segmentation and feature enhancement can be performed.

[0062] Step (4) Use neural network to build fingerprint forced press recognition dual-task model:

[0063] The first layer of the forced fingerprint press recognition dual-task neural network model is the fingerprint image input layer; the hidden layer is composed of a pooling layer, a convolutional layer, and a fully connected layer; the output layer has two different output layers, one for outputting the user corresponding to the fingerprint image, and the other for determining whether the fingerprint image has been subjected to a forced press attack.

[0064] Step (5) uses the established fingerprint library with dual labels to train the fingerprint forced press recognition model to obtain model parameters. The optimization of the model parameters depends on minimizing the joint loss function of the dual-task model. In this scenario, the joint loss function is:

[0065]

[0066] Denote the loss function for task 1 of person identification and task 2 of forced identification, respectively. W represents the model weight, and the noise parameters are σ1 and σ2. The weights and parameters of the training model are balanced by minimizing the joint loss function.

[0067] In step (6), conventional fingerprint recognition classification and forced fingerprint attack recognition are processed in parallel. While identifying whether the fingerprint is in the database, it is also detected whether the fingerprint is forced. If the fingerprint is in a forced state, the fingerprint recognition will be rejected regardless of whether the fingerprint is in the database, and a warning will be reported through the network. If the fingerprint is in a normal state, the fingerprint recognition will be carried out normally, and the defense against the forced fingerprint attack will be finally completed.

[0068] Example 2

[0069] A product based on a dual-task learning biometric fingerprint recognition method includes a computer program that, when executed by a processor, implements the following Figure 2The steps of the dual-task learning-based biometric fingerprint recognition method are shown. First, the user presses the fingerprint module to obtain a fingerprint image. The neural network extracts fine-grained features and then identifies the fingerprint image, determining the user to whom the image belongs and whether the user is being forced. If the fingerprint image is a non-forced fingerprint and belongs to a user in the database, unlocking is allowed. If the fingerprint image is a forced fingerprint or the fingerprint image is a non-forced fingerprint but does not belong to a user in the database, unlocking is denied.

[0070] Test Case

[0071] The test case database is from 71 users who have given permission. The age range of these users is 20 to 75 years old. It contains a total of 5112 fingerprint images with a size of 160*160. Among them, there are 2556 forced fingerprint images and non-forced fingerprint images. There are 4544 fingerprint images used for training and 568 fingerprint images used for testing. The final forced attack detection results are shown in the attached figure. Figure 3 After 300 rounds of training, the joint loss function approaches 0 and the model converges. Finally, the accuracy of the forced action identifier is 99.6% and the accuracy of the user identifier is 82.5%.

[0072] Therefore, the method in this case utilizes the dual labels of personal identity labels and forced action labels in the fingerprint database, so that the established neural network model can train two tasks at the same time. The shared weights of the dual-task model not only add the user authentication function to the algorithm, but also save the computing resources for establishing additional models, realize multi-user recognition and forced press detection, and enhance the security of biometric fingerprint recognition.

[0073] It should be noted that the above content merely illustrates the technical idea of the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A biometric fingerprint recognition method based on dual-task learning, characterized in that: It includes the following steps: S1, Normal data acquisition: Enter legal and non-forced pressed fingerprints on the fingerprint unlocking device, and assign a dual label of personal identity and non-forced to each normal fingerprint; S2, Forced data acquisition: When entering a biometric fingerprint unlocking operation on the fingerprint unlocking device, enter the unlocking fingerprints forced by the attacker, and assign a dual label of personal identity and forced to each forced fingerprint; S3, Feature enhancement: Filter the forced fingerprint images obtained in step S2, and enhance the fingerprint features affected by force in the fingerprint; The fingerprint feature enhancement is achieved through Fourier transform, wavelet transform and OTSU segmentation algorithm; S4, Dual-task model establishment: Use a neural network to build a dual-task model for fingerprint forced pressing recognition. The model includes an input layer, forty-eight hidden layers and an output layer. The hidden layer is composed of a pooling layer, a convolutional layer and a fully connected layer; There are two types of output layers. One is a user identifier for outputting the user identity corresponding to the fingerprint image, and the other is a forced action identifier for judging whether the fingerprint image is subject to a forced pressing attack; Use the fingerprint image processed in step S3 as the input of the model, perform dual-task classification on the features extracted by the network, and use the two labels of the image to perform forced attack recognition and user identity recognition simultaneously; S5, Model parameter optimization: Use the established fingerprint database with dual labels to train the fingerprint forced pressing recognition model in step S4. Among them, the user identifier authenticates the identity through the fingerprint image features of the user, and the forced action identifier continuously optimizes and updates the model parameters by identifying the action types to improve the overall prediction accuracy of the model; S6, Recognition processing: Fingerprint identity recognition and forced fingerprint attack recognition are processed in parallel. While recognizing the fingerprint identity, detect whether the fingerprint is forced. If the fingerprint is in a forced state, reject the fingerprint recognition regardless of whether the fingerprint is in the database and give a warning through network reporting; If the fingerprint is in a normal state, perform fingerprint identity recognition normally to complete the dual tasks of resisting forced pressing attacks on fingerprints and identity recognition.

2. The dual-task learning-based biometric fingerprint recognition method according to claim 1, wherein: The fingerprint enhancement in step S3 is specifically: Input an M×N fingerprint image, represented by L-level grayscale as [1,2,…,L-1], N i Represents the number of pixels with grayscale i, and the total number of pixels in the image is: The probability of the occurrence of gray level i is: Select a threshold k (0 < k < k - 1), which divides the input image into two categories, namely C1 and C2. C1 consists of pixels with gray values in the range [0, k], and C2 consists of pixels with gray values in the range [k + 1, L - 1]; Then the probability that a pixel is classified in C1 is: Similarly, the probability that a pixel is classified in C2 is: The gray mean value of C1 is: Similarly, the gray mean value of C2 is: The global variance is: The larger the average difference between m1(k) and m2(k), The larger the image is, the more image segmentation and feature enhancement can be performed.

3. The biometric fingerprint recognition method based on dual-task learning according to claim 1, characterized in that: The filtering in step S3 includes at least average filtering, Gaussian filtering and median filtering.

4. The biometric fingerprint recognition method based on dual-task learning according to claim 2 or 3, characterized in that: In step S5, the optimization of the model parameters depends on the minimization of the joint loss function of the dual-task model, that is: in, They represent the loss function of personal identification and forced identification respectively, W is the model weight, and the noise parameters are σ1 and σ2.

5. The dual-task learning-based biometric fingerprint recognition method according to claim 4, wherein: In the model parameter optimization of step S5, the model learning rate is 0.01 and the number of training rounds is 300 rounds.

6. Products, including computer programs, based on a dual-task learning biometric fingerprint recognition method, characterized by: When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-5 above.

Images