Intelligent user authentication method, device and software product based on acoustic sensing

The biological and behavioral characteristics of smartphone users are extracted through acoustic sensing technology, solving the problems of traditional identity authentication security risks and user burdens, and achieving efficient and secure dual-device identity authentication, resisting simulation and puppet attacks, enhancing the authentication system robustness.

CN118536102BActive Publication Date: 2024-11-08WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410993112.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-08
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

Traditional smartphone identity authentication methods have security risks and are susceptible to password guessing, stealing, tampering, replaying attacks and puppet attacks. Frequent password input or graphic authentication operations increase the burden on users, making it difficult to take into account efficiency and security.

Method used

Using an intelligent user authentication method based on acoustic sensing, the acoustic signal is sent through the first terminal, the second terminal collects and performs data preprocessing, extracts time-frequency features and CWT transformation features, combines the CNN model to extract features, performs user identity authentication, and uses Users' unique biological and behavioral characteristics are legally authenticated.

Benefits of technology

It realizes efficient and secure dual-device identity authentication, which is difficult to be copied and stolen, has good defense effects against simulated attacks and puppet attacks, improves the robustness and reliability of the authentication system, and is easy to deploy on existing mobile devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118536102B_ABST
    Figure CN118536102B_ABST
Patent Text Reader

Abstract

The present invention discloses an intelligent user authentication method, device and software product based on acoustic sensing, involving entities including a first terminal, a second terminal and a legitimate user; first, the first terminal sends an acoustic signal according to a preset duration and frequency; then the second terminal collects the acoustic signal sent by the first terminal, and performs data preprocessing to eliminate environmental noise and segment the signal; then the acoustic signal features are extracted, including time-frequency features, features based on CWT transformation and features extracted by CNN, and finally spliced ​​as joint features, and finally user identity authentication is performed based on the joint features. The present invention proposes a joint feature representation method based on time-frequency analysis, continuous wavelet transform and deep learning, which can effectively resist simulation, replay and puppet attacks. The present invention is applicable to a variety of authentication scenarios, has high security and high robustness, is easy to deploy and does not require dedicated hardware.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of security authentication, and relates to an intelligent user authentication method, device and software product, and specifically to an intelligent user authentication method, device and software product based on acoustic sensing. Background Art

[0002] With the popularity of smartphones and the increase in the frequency of use, the amount of personal information and sensitive data stored in mobile phones has also increased accordingly. Mobile phone authentication, as a key technology to prevent unauthorized access by illegal users and protect the privacy of personal mobile devices, has been widely studied. At present, smartphone identity authentication mainly relies on traditional authentication technologies such as passwords, graphic authentication locks, fingerprint recognition, and facial recognition.

[0003] However, these traditional authentication methods have a series of problems. First, password and graphic authentication are vulnerable to snooping by others, while fingerprint and facial recognition may be copied or deceived, which poses a security risk. The attack threats faced by traditional authentication can be divided into two categories: one is the common problems faced by conventional authentication systems, including: password guessing, stealing or tampering with templates, and tampering with matching results; the other is the problem faced by fingerprint and facial authentication systems that is different from conventional encryption systems: replay attack, that is, the attacker uses artificial means to bypass the authentication mechanism to achieve the attack. The main method of replay attack is that the attacker uses professional equipment and special materials such as silicone to make fake fingerprints and fake faces for attack. In addition, there is a special attack method called puppet attack, that is, the attacker forcibly uses the victim's fingerprint or face for authentication when the victim is unwilling (the victim is usually in a coma or sleeping state). From the above, it can be seen that the characteristic of puppet attack that is different from general replay attack is that it can effectively attack the defense system based on liveness detection. This makes it more difficult to resist puppet attack than general replay attack. Secondly, frequently entering passwords or graphic authentication operations will increase the burden on users and significantly reduce the user experience. In addition, fingerprint and facial recognition authentication methods use sensitive biometrics of users, which, once leaked, will lead to serious security risks such as identity theft and financial fraud. In recent years, with the increase in cyber attacks and fraud, as well as people's increased awareness of privacy protection, traditional identity authentication methods are facing increasingly severe challenges.

[0004] Therefore, finding a new, safer and more convenient identity authentication method has become an important research topic at present. Some new identity authentication technologies have been applied in today's society. For example, Apple introduced a new authentication function in iOS 14.5 beta, which uses Apple Watch to authenticate iPhone users. When the iPhone detects a nearby Apple Watch via Bluetooth and confirms that it has been paired, if facial recognition cannot be completed because the user is wearing a mask, the system will automatically call the Apple Watch for identity authentication. Similarly, Huawei and Xiaomi have also realized the function of using the bracelet to authenticate users of the same brand of mobile phones at close range through the Bluetooth connection between the bracelet and the mobile phone. Although these new authentication methods are more convenient, they do not consider security, such as: ① The judgment criteria for bracelet authentication of mobile phone users are based on distance (about 3 meters), rather than on the user's unique identity characteristics. As long as the attacker steals the user's bracelet, the authentication can be successfully passed. This authentication method lacks in-depth verification of personal identity, so there is a potential security risk. ② Even if the attacker does not steal the user's bracelet, as long as he approaches the legitimate user wearing the bracelet within the Bluetooth connection range of the bracelet and the mobile phone, the authentication can be successfully passed, which poses a serious security risk.

[0005] In view of the above scenarios and actual application requirements, the existing designs are difficult to balance efficiency and security. Therefore, considering how to achieve efficient and secure dual-device identity authentication technology is an urgent problem to be solved in this field. Summary of the invention

[0006] In view of the above requirements for data security, convenient availability, resistance to replay puppet attacks, etc. and the drawbacks of the above traditional solutions, the present invention proposes an intelligent user authentication method, device and software product based on acoustic sensing.

[0007] The technical solution adopted by the method of the present invention is: an intelligent user authentication method based on acoustic sensing, which is used to perform intelligent authentication of a legitimate user on a second terminal through a first terminal; the intelligent user authentication method comprises the following steps:

[0008] Step 1: The first terminal sends an acoustic signal according to a preset duration and frequency;

[0009] Step 2: The second terminal collects the acoustic signal sent by the first terminal, and performs data preprocessing to eliminate environmental noise and segment the signal to obtain a pulse signal sample, wherein each sample data in the pulse signal sample contains only one valid pulse peak;

[0010] Step 3: Extract acoustic signal features from all pulse signal samples, including time-frequency features, features based on CWT transformation, and features extracted by CNN, and finally concatenate them as joint features;

[0011] Step 4: Perform user identity authentication based on joint features.

[0012] Preferably, in step 1, the built-in universal sensor accelerometer and gyroscope of the first terminal are used to detect the user's hand-raising action; the changes in the accelerometer value are read and recorded in real time, and compared with a preset threshold value. If the threshold value is exceeded, it is determined that the hand-raising action has been triggered; when the triggering action is recognized, the digital signal is converted into an analog signal by calling the DAC interface, thereby driving the speaker to play ultrasonic waves.

[0013] Preferably, the specific implementation of step 2 includes the following sub-steps:

[0014] Step 2.1: Receive the original acoustic signal x ( t ), while transmitting and receiving acoustic signals in a quiet environment, and saving the received acoustic signals in the quiet environment as the same as the original acoustic signals x ( t ) cross-correlated calibration template, and the calibration template and the original acoustic signal x ( t ) to perform cross-correlation calculation, the peak position in the cross-correlation function represents the matching degree of the signal, and the distance between the peaks represents the period T of the signal; t Indicates time;

[0015] Step 2.2: Receive the original acoustic signal x ( t ) The signal is passed through a bandpass filter to remove environmental noise;

[0016] Step 2.3: Original acoustic signal x ( t ) performs Hilbert transform, that is The analytical signal used to generate the original acoustic signal, where Represents the original acoustic signal at time The signal amplitude at ; then extract the upper envelope of the analytical signal, that is Eliminate the burrs of the signal and obtain the peak of each pulse; finally, use moving average filtering to locate the peak position of each valid pulse according to the peak + T and segment each pulse signal sample.

[0017] Preferably, in step 3, the time-frequency feature is a feature statistic calculated in the time domain and the frequency domain from the pulse signal samples obtained after data preprocessing, including:

[0018] Mean ,

[0019] Standard Deviation ,

[0020] Maximum ,

[0021] Minimum ,

[0022] scope ,

[0023] Skewness ,

[0024] Kurtosis ,

[0025] Second quartile ,

[0026] Fourth quartile ,

[0027] Discreteness ,

[0028] energy ,

[0029] Spectral Centroid ,

[0030] Spectrum Diffusion ,

[0031] Spectral skewness ,

[0032] Spectral kurtosis ,

[0033] Spectral power density ,

[0034] Spectral entropy ;

[0035] in, It is i The signal amplitude of the sample data point, n is the number of sample data points in the pulse signal sample, is the standard deviation, is the mean, Is a location or state i The value below is the square value of each sample data point, It is i The frequency corresponding to the sample data points, is the normalized i The amplitude value of the sample data point is calculated as follows: , is the sampling frequency.

[0036] Preferably, in step 3, the features based on CWT transformation first use wavelet basis functions to calculate the CWT amplitude and phase information of the pulse signal samples obtained after data preprocessing; then solve the energy information of the CWT amplitude, and set the signal part with energy below the threshold to zero; finally, use principal component analysis to reduce the dimension and obtain features based on CWT transformation.

[0037] Preferably, in step 3, the features extracted by CNN are first converted into an amplitude spectrum containing time domain and frequency domain information through STFT transformation, and then the amplitude spectrum is converted into a spectrogram; and then the features are extracted using a CNN model.

[0038] Preferably, the CNN model is composed of a 12-layer network connected in sequence; the first layer is a convolutional layer Conv2d plus a LeakyReLu activation function; the second layer is a convolutional layer Conv2d plus a LeakyReLu activation function; the third layer is a pooling layer Pooling, a Dropout layer and a batch normalization BN layer; the fourth layer is a convolutional layer Conv2d plus a LeakyReLu activation function; the fifth layer is a convolutional layer Conv2d plus a LeakyReLu activation function; the sixth layer is a pooling layer P The 7th layer is the convolutional layer Conv2d plus the LeakyReLu activation function; the 8th layer is the convolutional layer Conv2d plus the LeakyReLu activation function; the 9th layer is the pooling layer Pooling, Dropout layer and batch normalization BN layer; the 10th layer is the flattening layer Flatten; the 11th layer is the fully connected layer FC plus the LeakyReLu activation function; the 12th layer is the fully connected layer FC plus the Softmax activation function.

[0039] Preferably, the specific implementation of step 4 includes the following sub-steps:

[0040] Step 4.1: Register;

[0041] The user performs N legal identity authentications. After the registration is completed, these N sets of feature vectors together constitute the legal identity authentication model of the user; where N is a preset value;

[0042] Step 4.2: Authentication;

[0043] Ultrasonic signals are collected through the second terminal, and the user's current bio-behavioral feature vector is generated based on these signals; then the legal identity authentication model generated when the user registers is read from the second terminal, the current user's bio-behavioral feature vector is input into the logistic regression classifier for classification, and the linear combination of the feature vector is mapped to between 0 and 1 through a logical function, indicating the probability that the sample belongs to a legal user. According to this probability, the sample is divided into a positive class or a negative class to determine whether the current user is a legal user, and then decide whether to allow the user to perform the operation.

[0044] The technical solution adopted by the device of the present invention is: an intelligent user authentication device based on acoustic sensing, including a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the intelligent user authentication method based on acoustic sensing when executing the program.

[0045] The technical solution adopted by the software product of the present invention is: a software product, including a computer program, and when the computer program is executed by a processor, the intelligent user authentication method based on acoustic sensing is implemented.

[0046] Compared with the existing dual-device authentication technology, the advantages and positive effects of the present invention are mainly reflected as follows:

[0047] (1) The user characteristics used in the present invention are not only individually unique, but also difficult to copy and steal. The way a user holds a mobile phone contains a wealth of biological and behavioral characteristics, such as hand contour, finger size, muscle and bone density, grip strength and style. These characteristics will cause different degrees of signal attenuation when sound waves propagate through human hands.

[0048] (2) The present invention adopts an innovative feature extraction technology. It not only uses traditional time-frequency features, but also introduces the continuous wavelet transform (CWT) suitable for ultrasonic pulse signals. CWT can provide localized information of the signal in time and frequency, which is particularly important for analyzing the characteristics of ultrasonic pulses. Combined with acoustic signal analysis, the present invention performs deep feature extraction on the spectrogram of ultrasonic pulses, significantly improving the accuracy of the machine learning model and enhancing the robustness and reliability of the authentication system.

[0049] (3) The present invention exhibits good defensive effects when facing new attack methods such as simulation attacks and puppet attacks, and naturally resists theft and replay attacks, effectively rejecting identity authentication requests from illegal users, and can work effectively in various user scenarios.

[0050] (4) The present invention is easy to deploy and does not rely on any dedicated hardware for data measurement. It only requires a built-in speaker, a microphone, and general sensors (such as an accelerometer, a gyroscope, etc.) to achieve efficient data collection and is easy to deploy on existing mobile devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The technical solution of this invention is further described below using embodiments and specific implementation methods. In addition, some drawings are also used in the process of describing the technical solution. For those skilled in the art, other drawings and the intention of the invention can also be obtained based on these drawings without paying creative labor.

[0052] Figure 1 A diagram showing the functions of system modules and the data flow between modules in an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of feature extraction in an embodiment of the present invention;

[0054] Figure 3 The figure is a flowchart of user identity authentication in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the following is combined with the drawings and examples, but it should be noted that these implementation methods are not limitations of the present invention, and any equivalent transformations or substitutions in functions, methods, or structures made by ordinary technicians in the field based on these implementation methods are all within the scope of protection of the present invention.

[0056] Please see Figure 1 This embodiment provides an intelligent user authentication method based on acoustic sensing, which is used to perform intelligent authentication of a legitimate user on a second terminal through a first terminal; the intelligent user authentication method includes the following steps:

[0057] The first terminal is a transmitting terminal of ultrasonic signals. The present invention uses the built-in universal sensor accelerometer and gyroscope of the wristband to detect the user's hand-raising action; the value changes of hardware devices such as the accelerometer are read and recorded in real time, and compared with the preset threshold value. If the threshold value is exceeded, it is determined that the hand-raising action has been triggered. The ultrasonic audio file is pre-stored in the watch in hexadecimal form. When the triggering action is recognized, the digital signal is converted into an analog signal by calling the DAC interface, and then the speaker is driven to play the ultrasonic wave;

[0058] The second terminal is a receiving terminal for ultrasonic signals. When the user raises his hand to trigger the ultrasonic signal generated by the smart watch, the ultrasonic signal is collected by the second terminal after being attenuated by the tissue structure of the user's wrist. The present invention sets the sampling time to 0.5s and the sampling rate fs to 43366.4Hz. A higher sampling rate can capture higher frequency signals and is suitable for signal processing in the ultrasonic frequency band. The minimum and maximum frequencies of the audio signal are set to 18000Hz and 43366.4Hz, respectively. , according to the Nyquist sampling theorem, half of the sampling rate is the highest frequency that can accurately reconstruct the signal. At this sampling rate, the highest frequency is 21683.2Hz;

[0059] A legitimate user is a legally registered user of the intelligent user authentication system. When a user picks up a mobile phone while wearing a watch, it triggers a natural hand-raising action. After the watch recognizes the triggering action, it will send an ultrasonic signal. The signal is attenuated by the human hand and then received by the mobile phone. The received signal carries the user's unique biological and behavioral characteristics. The authentication system uses this feature to accurately authenticate and identify the user.

[0060] The method comprises the following steps:

[0061] Step 1: Acoustic signal design;

[0062] The first terminal will send an acoustic signal according to a specific duration and frequency. The acoustic signal should meet the characteristics of being imperceptible to users, being resistant to environmental noise interference, being deployable in wearable smart devices, and having low system power consumption and response time;

[0063] In this embodiment, a chirp signal is selected in the acoustic signal design stage. The chirp signal is a pulse signal with a duration of 25 milliseconds and a frequency that increases linearly from 16 kHz to 22 kHz. A Hanning window is used to reshape the pulse envelope to improve the signal-to-noise ratio and reduce spectrum leakage. A delay of 25 milliseconds is then left between each pulse to avoid echo overlap from consecutive pulses. In order to increase the authentication speed, the total duration of the entire signal is 0.5 seconds.

[0064] Step 2: Data preprocessing;

[0065] The first terminal will send an acoustic signal to the outside world, which will be collected by the microphone of the second terminal. The collected acoustic signal needs to be pre-processed to eliminate environmental noise and segment the signal.

[0066] In this embodiment, the second terminal collects the chrip signal emitted by the watch and performs data preprocessing; the specific implementation includes the following sub-steps:

[0067] Step 2.1: Eliminate environmental noise. The received original chrip signal is passed through a 18-22 kHz third-order Butterworth bandpass filter to remove environmental noise.

[0068] Step 2.2: Signal segmentation, set x ( t ) is the original received signal, indicating the time t The signal amplitude at Represents the original received signal x ( t ), produces a complex-valued signal, Represents the envelope of the signal; first, transmit and receive the chrip signal in a quiet environment, and save the received chrip signal as the same as the original received signal x ( t ) cross-correlation calibration template, the calibration template and the original acoustic signal are cross-correlated, the peak position in the cross-correlation function represents the matching degree of the signal, and the distance between the peaks represents the period T of the signal; then the original received signal x ( t ) performs Hilbert transform, that is The analytical signal used to generate the original signal, where Represents the original acoustic signal at time The signal amplitude at ; then extract the upper envelope of the analytical signal, that is Eliminate the burrs of the signal and obtain the peak of each pulse; finally, use moving average filtering, peak + T to accurately locate the peak position of each valid pulse and segment each pulse signal sample.

[0069] Step 3: Feature extraction;

[0070] In order to accurately characterize the user's biometric features using the deep features of the data to support subsequent authentication classification, the present invention proposes a feature extraction scheme, which selects time-frequency features, features based on CWT transformation, and features extracted by CNN as joint features to provide strong support for subsequent authentication classification;

[0071] Please see Figure 2 In this embodiment, the input data is respectively subjected to time-frequency feature extraction to obtain a 49-dimensional feature vector, and CWT and CNN feature extraction are respectively subjected to 460-dimensional and 180-dimensional feature vectors. Finally, several vectors are concatenated into a 689-dimensional vector as the result of the feature extraction part;

[0072] The specific implementation includes the following sub-steps:

[0073] Step 3.1: Time-frequency feature extraction, calculate the feature statistics in the time domain and frequency domain from the user's biological behavior feature obtained after data preprocessing. The present invention selects the more discriminative features, including: mean , standard deviation , the maximum value , minimum ,scope , skewness , kurtosis , the second quartile , the fourth quartile , discreteness ,energy , the spectrum centroid , spectrum diffusion , spectral skewness , spectral kurtosis , spectral power density , spectral entropy ;in, It is i sample data points, n is the number of sample data points, is the standard deviation, is the mean, Is a location or state i The value below is the square value of each sample data point, It is i The frequency corresponding to the sample data points, is the normalized i The amplitude value of the sample data point is calculated as follows: , is the sampling frequency.

[0074] Step 3.2: Feature extraction based on CWT transform. Continuous Wavelet Transform (CWT) is used to decompose the wavelet signal. Wavelets are small, highly localized oscillations in time, and the Fourier transform decomposes the signal into infinitely long sines and cosines, thus losing all time position information. The basic function of CWT is to scale and shift the time-localized parent wavelet to construct a time-frequency representation of the signal to provide very good time and frequency positioning. When the signal is considered to be non-stationary, CWT can be used to identify the stationary part of the data stream. The definition of CWT is as follows: , ,

[0075] ,

[0076] ;

[0077] in, s is the scale parameter, is the index variable for summing the time series, is the sampling time interval, is the mother wavelet function, which is used to generate wavelet functions of different scales. is the Fourier transform of the mother wavelet function, is the conjugation operator, is the inverse Fourier transform, It's a signal The Fourier transform result of is expressed in the frequency domain. is the frequency variable, which represents the frequency component after Fourier transform; ω is the angular frequency.

[0078] CWT is the convolution of a data sequence with a scaled and translated version of the parent wavelet, i.e., the ψ function; this convolution can be implemented directly, as in the first equation, or by a fast convolution based on FFT in the second equation; the wavelet function used in the CWT transformation of the present invention is the Morlet wavelet, which is a complex Gaussian windowed sine wave, defined in the time and frequency domains as follows:

[0079] ;

[0080] Where η is the dimensionless time parameter, m is the wave number, i is the imaginary unit, H is the Heaviside function, ω is the angular frequency;

[0081] After using the wavelet basis function to calculate the CWT amplitude and phase information of the data, the energy information is solved. The energy calculation formula is as follows:

[0082] ;

[0083] Where E is the wavelet transform energy, W is the wavelet transform coefficient, a is the scale parameter, and b is the displacement parameter;

[0084] In order to reduce the influence of noise, the signal part with energy below the threshold is set to zero; the threshold is selected based on the mean of the energy multiplied by a small coefficient (here 0.1); after the above transformation, the extracted feature dimension is still relatively large, so the present invention also uses principal component analysis (PCA) for dimensionality reduction;

[0085] Step 3.3: Feature extraction based on convolutional neural network. The convolutional neural network feature extraction method of the present invention is implemented based on two technologies: Short Time Fourier Transform (STFT) and Convolutional Neural Network (CNN). First, the user's biological behavior feature signal is converted into an amplitude spectrum containing time domain and frequency domain information through STFT, and then the amplitude spectrum is converted into a spectrogram for training the CNN model, and then the CNN model is used to extract features. The structure of the CNN model used in the present invention is described as follows: the first layer is a convolutional layer (Conv2d) plus a LeakyReLu activation function, the output format is 62×126×24, and the number of parameters is 1,536; the second layer is also a convolutional layer (Conv2d) plus a LeakyReLu activation function, the output format is 60×124×24; the third layer is a pooling layer (Pooling ling), Dropout and batch normalization (BN), the output format is 30×62×24; the 4th layer is a convolutional layer (Conv2d) plus a LeakyReLu activation function, the output format is 28×60×48; the 5th layer is a convolutional layer (Conv2d) plus a LeakyReLu activation function, the output format is 26×58×48; the 6th layer is a pooling layer (Pooling), Dropout and batch normalization (BN), the output format is 13×29×48; the 7th layer is a convolutional layer (Conv2d) plus a LeakyReLu activation function, the output format is 11×27×1 6; the 8th layer is also a convolutional layer (Conv2d) plus a LeakyReLu activation function, with an output format of 9×25×16; the 9th layer is a pooling layer (Pooling), Dropout and batch normalization (BN), with an output format of 4×12×16; the 10th layer is a flattening layer (Flatten), with an output format of 768; the 11th layer is a fully connected layer (FC) plus a LeakyReLu activation function, with an output format of 180; the 12th layer is a fully connected layer (FC) plus a Softmax activation function, with an output format of 90; for the two-dimensional convolutional (Conv2d) layer and the fully connected (FC) layer The activation function of the present invention uses a leaky random linear rectifier unit (LeakyReLu) to deal with the gradient vanishing problem in the model training stage; for the pooling layer, the present invention uses the maximum pooling method to downsample the input, which controls overfitting and saves computing costs by reducing the number of training parameters; in order to avoid overfitting, the present invention adds a Dropout layer after each pooling layer; in addition, the present invention also includes a batch normalization (BN) layer to normalize the output of the previous layer, thereby accelerating the training of the model and enhancing the stability of the model; the last layer adds a Softmax layer for prediction, and outputs the classification probability distribution of each class;Specifically, the kernel sizes of Conv2d and pooling layers are set to 3×3 and 2×2, respectively, which can well represent nonlinear features; the total model contains 202,974 parameters, of which 202,438 are trainable parameters and 536 are non-trainable parameters; after the CNN model training is completed, combined with the specific test situation, the present invention selects the output of the 11th layer as the feature, that is, each group of user behavior data will extract a 180-dimensional CNN feature vector;

[0086] Step 3.4: Combining the above three features, the present invention first normalizes the feature extraction results of the above three feature extraction methods respectively to prevent the excessive weight of a certain feature value from affecting the classification result, and then splices the normalized time-frequency features, the features extracted after CWT transformation, and the CNN features into a 689-dimensional comprehensive feature vector as the biological behavior feature vector.

[0087] Step 4: Identity authentication;

[0088] The identity authentication module is responsible for building a legitimate user model and performing user identity authentication. The present invention uses a logistic regression classifier to classify the feature vectors of the samples, and at the same time, based on the time principle, performs necessary continuous updates on the user model, so that the authentication system maintains high accuracy and high stability.

[0089] Please see Figure 3 In this embodiment, the identity authentication module uses the user's biological behavior features extracted by the feature extraction module to distinguish between legitimate users and illegal users, and provides two functions: registration and authentication. The specific implementation includes the following sub-steps:

[0090] Step 4.1: Registration. When the user activates the present invention, the present invention requires the user to perform five legal identity verification behaviors. The comprehensive feature vectors generated by these five behaviors will be stored in the mobile phone storage, wherein the feature vector generated by each verification contains 689 dimensions. After the registration is completed, the present invention uses these five sets of comprehensive feature vectors to form a legal identity verification model for the user.

[0091] Step 4.2: Authentication. When a successfully registered user uses the present invention for authentication, the present invention collects ultrasonic signals through the mobile phone microphone and generates the user's current bio-behavioral feature vector based on these signals. Subsequently, the present invention reads the legal identity authentication model generated when the user registers from the mobile phone storage, and inputs the current user's bio-behavioral feature vector into the logistic regression classifier for classification. The logistic regression classifier uses the limited memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS) algorithm to optimize the model parameters, and maps the linear combination of the feature vectors to between 0 and 1 through a logic function (also called a sigmoid function), which represents the probability that the sample belongs to a legal user. Based on this probability, the present invention can classify the sample into a positive class or a negative class, determine whether the current user is a legal user, and then decide whether to allow the user to perform the operation.

[0092] This embodiment further provides an update operation. After the user successfully registers, the present invention will start the update operation of the user's legal identity authentication model by default every month; the update strategy is to select a higher update threshold based on the original authentication threshold to ensure that only high-quality data is used when updating the model, thereby avoiding incorporating data with poor effects into the model, thereby improving the reliability and accuracy of the overall authentication; when the user passes the authentication once and the similarity is greater than the update threshold, the present invention will form a new set of feature vectors that pass the authentication this time and the five feature vectors in the user's existing legal identity authentication model; each time a vector is selected from this set, and its similarity with the other five vectors is compared, the vector with the lowest similarity is discarded, and the remaining five feature vectors are retained as the updated legal identity authentication model, ensuring that the model is always composed of the most representative feature vectors, further improving the accuracy and reliability of the authentication.

[0093] This embodiment also provides an intelligent user authentication device based on acoustic sensing, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the intelligent user authentication method based on acoustic sensing when executing the program.

[0094] This embodiment also provides a software product, including a computer program, which implements the intelligent user authentication method based on acoustic sensing when executed by a processor.

[0095] The first terminal of this embodiment can be a wearable device or a handheld terminal such as a smart bracelet, smart watch, smart ring, etc., and the second terminal can be a smart watch, tablet computer, laptop computer, desktop computer, car computer, vehicle-mounted computer, server, access control terminal, etc., any device that requires verification and login to grant certain permissions.

[0096] The present invention has strong robustness, can resist simulation attacks, replay attacks and puppet attacks, and broadens the device identity authentication method. Traditional dual-device authentication schemes are mostly implemented through Bluetooth, and their authentication standards are mainly based on the distance between devices, lacking authentication of individual deep biometrics. The user characteristics utilized by the present invention are not only individually unique, but also difficult to copy and steal. The operating mechanism of the present invention is transparent and silent, and does not require users to remember credentials, which greatly improves the user experience. The present invention can be deployed on mobile devices, Internet of Things devices, and wearable devices to strengthen existing physiological feature authentication systems or as an independent authentication system, providing a complete, end-to-end authentication solution for enterprises and individual users with security needs.

[0097] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An intelligent user authentication method based on acoustic sensing, used for intelligent authentication of legitimate users on a smart phone through a smart bracelet; characterized in that: The intelligent user authentication method comprises the following steps: Step 1: After the user raises his hand, the smart bracelet sends an acoustic signal according to the preset duration and frequency; Among them, the built-in universal sensor accelerometer and gyroscope of the smart bracelet are used to detect the user's hand-raising action; the changes in the accelerometer value are read and recorded in real time, and compared with the preset threshold. If the threshold is exceeded, it is determined that the hand-raising action has been triggered; when the trigger action is recognized, the digital signal is converted into an analog signal by calling the DAC interface, and then the speaker is driven to play ultrasonic waves; the ultrasonic signal propagates along the arm, and due to the differences in the biological and behavioral characteristics of the human body, the acoustic signal will be differentially attenuated; Step 2: The smart phone collects the acoustic signal sent by the smart bracelet and attenuated after propagating along the arm, and performs data preprocessing to eliminate environmental noise and segment the signal to obtain a pulse signal sample, in which each sample data of the pulse signal sample contains only one valid pulse peak; Step 3: Extract acoustic signal features from all pulse signal samples, including time-frequency features, features based on CWT transformation, and features extracted by CNN, and finally concatenate them as joint features; The features based on CWT transformation first use wavelet basis functions to calculate the CWT amplitude and phase information of the pulse signal samples obtained after data preprocessing; then solve the energy information of the CWT amplitude, and set the signal part with energy below the threshold to zero; finally, use principal component analysis to reduce the dimension and obtain the features based on CWT transformation; Step 4: Perform user identity authentication based on joint features.

2. The intelligent user authentication method based on acoustic sensing according to claim 1, characterized in that: The specific implementation of step 2 includes the following sub-steps: Step 2.1: Receive the original acoustic signal x ( t ), while transmitting and receiving acoustic signals in a quiet environment, and saving the received acoustic signals in the quiet environment as the same as the original acoustic signals x ( t ) cross-correlated calibration template, and the calibration template and the original acoustic signal x ( t ) to perform cross-correlation calculation, the peak position in the cross-correlation function represents the matching degree of the signal, and the distance between the peaks represents the period T of the signal; t Indicates time; Step 2.2: Receive the original acoustic signal x ( t ) The signal is passed through a bandpass filter to remove environmental noise; Step 2.3: Original acoustic signal x ( t ) performs Hilbert transform, that is The analytical signal used to generate the original acoustic signal, where Represents the original acoustic signal at time The signal amplitude at ; then extract the upper envelope of the analytical signal, that is Eliminate the burrs of the signal and obtain the peak of each pulse; Finally, a moving average filter is used to locate the peak position of each valid pulse according to the peak + T, and each pulse signal sample is segmented.

3. The intelligent user authentication method based on acoustic sensing according to claim 1, characterized in that: In step 3, the time-frequency features are characteristic statistics calculated in the time domain and frequency domain from the pulse signal samples obtained after data preprocessing, including: Mean , Standard Deviation , Maximum , Minimum , scope , Skewness , Kurtosis , Second quartile , Fourth quartile , Discreteness , energy , Spectral Centroid , Spectrum Diffusion , Spectral skewness , Spectral kurtosis , Spectral power density , Spectral entropy ; in, It is i The signal amplitude of the sample data points, n is the number of sample data points in the pulse signal sample, is the standard deviation, is the mean, Is a location or state i The value below is the square value of each sample data point, It is i The frequency corresponding to the sample data points, is the normalized i The amplitude value of the sample data point is calculated as follows: , is the sampling frequency.

4. The intelligent user authentication method based on acoustic sensing according to claim 1, characterized in that: In step 3, the features extracted by CNN are first converted into an amplitude spectrum containing time domain and frequency domain information through STFT transformation, and then the amplitude spectrum is converted into a spectrogram; and then the CNN model is used to extract features.

5. The intelligent user authentication method based on acoustic sensing according to claim 4 is characterized in that: The CNN model is composed of a 12-layer network connected in sequence; the first layer is a convolutional layer Conv2d with a LeakyReLu activation function; the second layer is a convolutional layer Conv2d with a LeakyReLu activation function; the third layer is a pooling layer Pooling, a Dropout layer, and a batch normalization BN layer; the fourth layer is a convolutional layer Conv2d with a LeakyReLu activation function; the fifth layer is a convolutional layer Conv2d with a LeakyReLu activation function; the sixth layer is a pooling layer Poo ling, Dropout layer and batch normalization BN layer; the 7th layer is the convolution layer Conv2d plus the LeakyReLu activation function; the 8th layer is the convolution layer Conv2d plus the LeakyReLu activation function; the 9th layer is the pooling layer Pooling, Dropout layer and batch normalization BN layer; the 10th layer is the flattening layer Flatten; the 11th layer is the fully connected layer FC plus the LeakyReLu activation function; the 12th layer is the fully connected layer FC plus the Softmax activation function.

6. The intelligent user authentication method based on acoustic sensing according to any one of claims 1 to 5, characterized in that: The specific implementation of step 4 includes the following sub-steps: Step 4.1: Register; The user performs legal identity authentication N times. Each time the user registers, the user raises his hand. After the smart bracelet detects the hand-raising action, it will emit an ultrasonic signal. The sound wave signal propagates along the arm and attenuates and is received by the mobile phone. Data processing, feature extraction and classifier training are performed. After the registration is completed, these N groups of feature vectors together constitute the user's legal identity authentication model; where N is a preset value; Step 4.2: Authentication; The user raises his hand, and the smartphone collects the attenuated ultrasonic signal, and generates the user's current bio-behavioral feature vector based on these signals. Then, the legal identity authentication model generated when the user registered is read from the smartphone, and the current user's bio-behavioral feature vector is input into the logistic regression classifier for classification. The linear combination of the feature vectors is mapped to between 0 and 1 through a logical function, indicating the probability that the sample belongs to a legitimate user. Based on this probability, the sample is divided into positive or negative classes to determine whether the current user is a legitimate user, and then decide whether to allow the user to perform the operation.

7. An intelligent user authentication device based on acoustic sensing, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the intelligent user authentication method based on acoustic sensing as described in any one of claims 1 to 6 is implemented.

8. A software product comprising a computer program, characterized in that When the computer program is executed by a processor, the intelligent user authentication method based on acoustic sensing as claimed in any one of claims 1 to 6 is implemented.