WO2022160691A1

WO2022160691A1 - Reliable user authentication method and system based on mandibular biological features

Info

Publication number: WO2022160691A1
Application number: PCT/CN2021/114402
Authority: WO
Inventors: 刘建伟; 宋文帆; 沈乐明; 韩劲松; 任奎
Original assignee: 浙江大学
Priority date: 2021-02-01
Filing date: 2021-08-24
Publication date: 2022-08-04
Also published as: CN112949403A; CN112949403B

Abstract

Disclosed is a reliable user authentication method based on mandibular biological features. The method comprises: acquiring, by means of an inertial measurement unit, vibration signals of six axes that are generated when a user makes a sound with his/her throat; preprocessing the vibration signals of each axis into a gradient array; and inputting the gradient array into a mandibular biological feature extractor so as to obtain mandibular biological features, and registering and authenticating the user according to the mandibular biological features. In the present invention, vibration signals are generated by a user making a sound with his/her throat, and a mandible is driven to vibrate; and in the present invention, vibration signals comprising biological features are collected by means of earphones. In the present invention, an inertial measurement unit is used to capture vibration features of a mandible; positive vibration features and negative vibration features are separated by using gradient symbols; and mandibular biological features are extracted from vibration signals by using a dual-branch deep neural network. Biological feature vectors are converted into cancelable biological feature vectors by using a Gaussian matrix, so as to avoid replay attacks.

Description

A reliable user authentication method and system based on mandibular biometrics

technical field

The invention belongs to the field of user authentication, and in particular relates to a reliable user authentication method that uses an inertial measurement unit in an earphone to collect vibration signals containing mandibular biological features and extracts the biological features with a deep neural network.

Background technique

User authentication technology is widely deployed in security and privacy related applications in people's daily life. For example, the door lock of the user's house, the unlocking of the smartphone, and the identity verification of the high-speed rail plane and other transportation stations. Existing user authentication technologies can be divided into biometric-based user authentication technologies and non-biometrics-based identity authentication technologies according to whether or not biometrics are used. Both of the existing authentication technologies have their own drawbacks.

Non-biometric based authentication techniques often use knowledge or equipment as an authentication credential. For example, the password or pattern unlocking of a mobile phone is a knowledge credential. The user's ID card and ID card are device credentials, and these two authentication methods are simple and easy to use. However, the knowledge-based authentication method requires the user to remember sufficiently complex knowledge for sufficient security, which brings inconvenience to the user. Device-based authentication has a big security risk if the device is lost, because anyone who has the device can be authenticated.

In order to solve the inconvenient and insecure defects of non-biometric-based authentication technology, biometric-based authentication technology has been proposed and widely studied. Existing biometric-based user authentication technologies include in vivo biometric-based and in vitro biometric-based authentication technologies. For example, fingerprints, facial features are in vitro features because they can be captured from the body surface. Brain wave and heartbeat signatures are in vivo biosignatures because they are extracted from organ dynamics in the body. Although the authentication technology based on in vitro features is very convenient in collecting features, fingerprints and facial features are easily stolen and copied by attackers, and its security is not as secure as the authentication technology based on in vivo biometrics. However, the authentication technology based on in vivo biometrics is very inconvenient when collecting features, and the features themselves are not stable enough. For example, brain wave acquisition devices are cumbersome and not suitable for long-term wear. Heartbeat characteristics are easily affected by factors such as mood movement. Therefore, there is an urgent need for a reliable user authentication technology that can capture stable biometrics with a simple acquisition method to realize authentication.

With the rapid development of hardware devices, inertial measurement units can already be deployed in existing headset devices. This makes it possible to acquire the vibration signal of the mandible with the inertial measurement unit. In addition, the headset will become an important computing platform for the next generation, and there are already headsets that deploy deep neural networks for real-time language translation. Therefore, it is feasible to extract mandibular biometrics (in vivo features) from vibration signals using deep neural networks in headphones. The invention utilizes the inertial measurement unit in the earphone to collect the vibration signal of the mandible and uses the deep neural network to extract the biometrics of the mandible in the body, and proposes a reliable authentication method based on the biometrics of the mandible.

SUMMARY OF THE INVENTION

In order to solve the problems in the prior art, the present invention proposes a user authentication method that uses a simple inertial measurement unit in a headset to collect in-vivo vibration signals, and uses a deep neural network to extract stable and reliable in-vivo biometrics.

In order to achieve the above purpose, the technical scheme adopted in the present invention is:

A method for reliable user authentication based on mandibular biometrics, comprising the following steps:

The inertial measurement unit is used to obtain the six-axis vibration signals generated by the user when the throat utters; the inertial measurement unit includes an accelerometer and a gyroscope. By closing the mouth and making an 'emm' sound, the user can cause the mandible to vibrate, which travels along the mandible to the ear, where it is picked up by the centralized inertial measurement unit.

The vibration signal of each axis is removed outliers, filtered, normalized by dispersion, and the gradient value is calculated and divided into positive and negative vibration signals, and finally spliced to form a gradient array.

The gradient array is input to the mandibular biometric extractor, the mandibular biometrics are obtained, and the user is registered and authenticated according to the mandibular biometrics.

Further, the vibration signals of the six axes are N sampling points intercepted from the vibration starting point, and N is not less than 60.

Further, the vibration starting point is determined by the following method:

Divide the original signal into a window of ten values, and calculate the standard deviation of all signal values within the window. If the standard deviation of a certain window is greater than 250, and the standard deviation of the following two windows is greater than 100, the first signal value of the window with the standard deviation of 250 is determined as the starting point of vibration.

Further, the removal of abnormal values is specifically: using the MAD detection algorithm to find out abnormal values, and for each abnormal value, at the same time, the average value of the previous two normal values and the back two normal values of the abnormal value is used to replace, with achieve the purpose of noise reduction. Further, high-pass filtering is performed by using a Butterworth filter with a cutoff frequency of 20 Hz, because the frequency of occurrence of people is generally higher than 150 Hz, and the frequency caused by body motion is generally lower than 10 Hz. This filtering approach removes low-frequency components that are not related to mandibular biometrics.

Further, it also includes performing interpolation processing on the positive and negative vibration signals to keep the dimensions of the gradient array consistent.

Further, the mandibular biometric feature extractor is a trained neural network, a classifier, and the like. Preferably, the mandible biological feature extractor is a double-branched deep neural network, and the positive gradient feature and the negative gradient feature are respectively input into the neural network to obtain a biological feature vector with dimension (1, 512). Training is done by the system manufacturer.

Further, after obtaining the biological features of the mandible, it also includes:

Multiply the mandible biometrics by a Gaussian matrix to convert the mandibular undo biometrics. If the Gaussian matrix is denoted as G, and the biometric vector output by the feature extractor is denoted as M ₁ , the converted biometric vector of the mandible can be calculated by the following formula:

M=M ₁ ×G.

Once the revocable biometric template in the headset is stolen, the user can change a Gaussian matrix G and generate a new revocable biometric template. The attack is rejected due to the low similarity between the old template and the new template.

Further, the process of registering the user according to the biometric features of the mandible is as follows:

The registration is completed by saving the obtained biometric features of the registered user's mandible.

The process of authenticating the user according to the mandibular biometrics is as follows: calculating the similarity between the mandible biometrics of the authenticated user and the mandibular biometrics saved during registration, if the similarity is greater than a threshold, accept the authentication, otherwise reject. The cosine algorithm is used to calculate the similarity, that is, the cosine similarity between the new revocable mandible biometric vector and the revocable biometric vector template is calculated. If the obtained similarity is greater than the acceptance threshold set in advance, the authentication will be accepted. Otherwise, the authentication is rejected.

Among them, the user only needs to speak for 0.2 seconds to provide 60 vibration sampling points. Because the sampling rate of the inertial measurement unit can be 350HZ.

Among them, the formula for using dispersion normalization to process the data (amplitude) of each axis is as follows:

Wherein v' _i , v _i , v _min , v _max are the normalized vibration data value, the vibration data value before normalization, the maximum value in the axis, and the minimum value, respectively.

Among them, the formula for obtaining the gradient for each axis is as follows:

where g _i , v′ _i , |t _i+1 -t _i | are the ith gradient value, the ith normalized vibration data value, and the difference between the ith data and the i+1th data, respectively Time difference.

The present invention also provides an authentication system for the above-mentioned reliable user authentication method based on mandibular biometric features, including:

The inertial measurement unit is used to obtain the six-axis vibration signal generated by the user's vocalization in the throat.

The signal processing module is used for removing outliers, filtering, normalizing the dispersion and calculating the gradient value of the vibration signal of each axis, dividing it into positive and negative vibration signals, and finally splicing to form a gradient array.

Mandible biometric extractor for extracting mandibular biometrics based on gradient arrays.

The registration and authentication module is used to store the mandibular biometrics generated during user authentication, and perform authentication based on the mandibular biometrics during user authentication.

The above authentication system can be a stand-alone system or built into an existing device including an inertial measurement unit, such as a headset, and the headset manufacturer embeds the trained mandibular biometric extraction network and the registration and authentication module into the headset system when manufacturing the headset. available in. It can thus be used for device authentication.

Compared with the existing user authentication technology, the present invention generates a vibration signal through the user's throat to vibrate and drives the mandible to vibrate. The present invention captures the vibration features of the mandible by using the inertial measurement unit, separates the positive and negative vibration features by the sign of the gradient, and uses the dual-branch deep neural network to extract the mandible biological features in the vibration signal. Transform biometric vectors into revocable biometric vectors using a Gaussian matrix to prevent replay attacks.

Description of drawings

Fig. 1 is the authentication flow chart of the present invention;

Fig. 2 is a schematic diagram of a mandibular vibration model;

Fig. 3 is a structure diagram of a dual-branch neural network;

Figure 4 is a schematic diagram of authentication performance;

Detailed ways

Aiming at the situation that the existing user authentication technology cannot satisfy both ease of use and security, the present invention proposes a method to capture safe, reliable and high discrimination using an inertial measurement unit on a wearable device, such as an earphone. biometrics for authentication. Figure 2 shows the vibration model of the mandible. Assuming that the force generating the positive vibration signal is F _P , the mass of the mandible is m, and the damping and elastic coefficients of the body components on both sides of the mandible are (c ₁ , c ₂ , respectively). ) and (k ₁ , k ₂ ), then according to Newton's second law, we can get:

F _P (t)=mx″(t)+c ₁ x′(t)+(k ₁ +k ₂ )x(t),

where x(t) is the positive vibration displacement of the mandible. After the Fourier transform of this equation, the following formula is obtained:

where w, X _P (w), i, and F _P (0) represent the frequency components of the vibration signal, the frequency spectrum of the vibration signal, the imaginary unit, and the instantaneous force that produces positive vibration, respectively. After the vibration is transmitted to the earphone, the spectrum of the vibration signal can be expressed as:

where a and d are the attenuation coefficient and propagation distance of the vibration signal, respectively. Similar to the propagation law of the positive vibration signal, the spectrum of the negative vibration signal when it reaches the earphone can be expressed as:

So one vibration period (displacement of the mandible from the center position to the positive edge and from the center position to the negative edge) can be expressed as:

It can be found that m, c ₁ , c ₂ , k ₁ , and k ₂ in this formula are all biological features of the mandible, which are distinguishable in different people. Therefore, capturing vibration signals containing mandibular biometrics from the headset's inertial measurement unit is a viable authentication method. The method of the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

A reliable user authentication method based on mandibular biometrics, its brief process is shown in Figure 1, which is completed in the following five steps:

Step 1) The user makes a sound for 0.2 seconds with the throat. The vibration of the throat vibrates the user's jawbone. The vibration signal is captured by an inertial measurement unit in the headset as it travels along the mandible to the headset. The vibration signal contains the biometric features of the user's mandible.

Step 2) Preprocess the original vibration signal collected by the inertial measurement unit to remove noise and environment-independent components in the signal data, and finally obtain a signal array.

Divide the original signal into a window every ten values, and calculate the standard deviation of all signal values within the window. If the standard deviation of a window is greater than 250, and the standard deviation of the next two windows is greater than 100, the vibration is considered to start from the first data of the window with the standard deviation of 250. And select consecutive 60 data values for the vibration signal of each axis.

Leverage the MAD algorithm to find outliers in each axis caused by hardware imperfections or human motion. Each outlier is then replaced with the average of the two normal values before it and the two normal values after it.

The data of each axis were processed by high-pass filtering with Butterworth filter, and the cut-off frequency was 20 Hz to eliminate the low-frequency components irrelevant to the biological features of the mandible and make the data more pure.

Since the initial vibration value of each axis is different, it is necessary to normalize the dispersion for each axis. Otherwise, the effect of the axis with the smaller starting value will be masked by the effect of the axis with the larger starting value in the subsequent processing. The normalized six-axis data are spliced to form a signal array of dimension (6, 60).

Step 3) Calculate the positive and negative gradient features

The signal array obtained by step 1) and step 2) is difficult to distinguish between positive vibration and negative vibration, so the gradient of each axis can be obtained and the positive and negative vibration can be separated according to the sign of the gradient. Gradients greater than or equal to 0 are positive vibrations, and other gradients are negative vibrations.

The obtained positive and negative gradients are linearly interpolated separately, so that the gradients in each direction of each axis contain 30 values. Concatenate the gradients of all axes to form a gradient array of dimension (2, 6, 30).

Step 4) Input the gradient matrix obtained in step 3) into the mandible biometric extractor to obtain a mandible biometric vector. This vector is multiplied by a Gaussian matrix and stored in the headset as a revocable authentication template.

In this embodiment, the mandibular biometric feature extractor adopts a trained neural network, and its structure is shown in FIG. 3 , including two convolution branches inputting positive gradient and negative gradient respectively. Each convolutional branch consists of three convolutional layers. Each convolutional layer is followed by a batch normalization function and a ReLU activation function. The outputs of the convolutional layers are concatenated and fed into two fully connected layers. The output of the first fully connected layer is the mandible biometric vector extracted from the gradient feature array.

Step 5) The user makes the earphone capture the vibration signal of the mandible by 0.2 seconds of throat vibration. The vibration signal is processed through steps 2) and 3) and then input to the feature extractor to obtain a new mandibular biological feature vector. After converting the biometric vector into a revocable biometric vector, the similarity calculation is performed with the template in the headset. If the similarity is greater than 0.57, the authentication is accepted; otherwise, the authentication is rejected.

The present invention proposes a safe and reliable user authentication method that can be implemented on earphones, aiming at the situation that the security and ease of use cannot be guaranteed at the same time in the existing user authentication technology. The invention does not require physical interaction between the user and the authentication device, only needs to emit throat vibration, the inertial measurement unit in the earphone can capture the vibration signal containing the biological features of the mandible, and uses multi-step preprocessing and deep network to extract high-discrimination Mandible biometrics, using Gaussian matrix and similarity calculation to achieve anti-replay security authentication method. The false rejection rate and false acceptance rate in the experimental data of 30 individuals are shown in Figure 4, and the equal error rate is less than 2%.

Obviously, the above-mentioned embodiments are only examples for clear description, and are not intended to limit the implementation manner. For those of ordinary skill in the art, changes or modifications in other different forms can also be made on the basis of the above description. All implementations need not and cannot be exhaustive here. However, the obvious changes or changes derived from this are still within the protection scope of the present invention.

Claims

A reliable user authentication method based on mandibular biometric features, comprising the following steps:

Use the inertial measurement unit to obtain the six-axis vibration signals generated by the user when the user vocalizes in the throat;

The vibration signal of each axis is removed outliers, filtered, normalized by dispersion, and the gradient value is calculated to divide it into positive and negative vibration signals, and finally spliced to form a gradient array;

The gradient array is input to the mandibular biometric extractor, the mandibular biometrics are obtained, and the user is registered and authenticated according to the mandibular biometrics.
The reliable user authentication method based on mandibular biometrics according to claim 1, wherein the vibration signals of the six axes are N sampling points taken after the vibration starting point, and N is not less than 60.
The reliable user authentication method based on mandibular biometric features according to claim 2, wherein the vibration starting point is determined by the following method:

Divide the original signal into a window every ten values, and calculate the standard deviation of all signal values in the window; if the standard deviation of a window is greater than 250, and the standard deviation of the next two windows is greater than 100, the standard deviation is determined as The first signal value of the window of 250 is the vibration start point.
The method for reliable user authentication based on mandibular biometric features according to claim 1, wherein the removing outliers is specifically: using a MAD detection algorithm to find outliers, and for each outlier, using the abnormality at the same time The value is replaced by the mean of the two normal values in front of the value and the two normal values in the back.
The reliable user authentication method based on mandibular biometric features according to claim 1, wherein the cut-off frequency of the filtering process is 20 Hz.
The method for reliable user authentication based on mandibular biometric features according to claim 1, further comprising performing interpolation processing on the positive and negative vibration signals to keep the dimensions of the gradient array consistent.
The reliable user authentication method based on mandibular biometrics according to claim 1, wherein the mandibular biometrics extractor is a trained neural network and a classifier.
The reliable user authentication method based on mandibular biometrics according to claim 1, characterized in that, after obtaining mandibular biometrics, the method further comprises:

Multiply the mandible biometrics by a Gaussian matrix to convert the mandibular undo biometrics.
The reliable user authentication method based on mandibular biometrics according to claim 1, wherein the process of registering a user according to mandibular biometrics is:

The registration is completed by saving the obtained biometric features of the registered user's mandible;

The process of authenticating the user according to the mandibular biometrics is as follows: calculating the similarity between the mandible biometrics of the authenticated user and the mandibular biometrics saved during registration, if the similarity is greater than a threshold, accept the authentication, otherwise reject.
The authentication system for a reliable user authentication method based on mandibular biometrics according to any one of claims 1-9, characterized in that, comprising:

Inertial measurement unit, used to obtain the six-axis vibration signal generated by the user when vocalizing in the throat;

The signal processing module is used for removing outliers, filtering, normalizing the dispersion and calculating the gradient value of the vibration signal of each axis, dividing it into positive and negative vibration signals, and finally splicing to form a gradient array;

A mandibular biometric extractor for extracting mandibular biometrics based on gradient arrays;

The registration and authentication module is used to store the mandibular biometrics generated during user authentication, and perform authentication based on the mandibular biometrics during user authentication.