A User Authentication Method Based on Wearable Device Motion Sensors

Abstract

This invention relates to a user authentication method based on motion sensors in wearable devices, belonging to the field of mobile computing application technology. This method uses motion sensors in a smart wristband device to collect motion signals generated by the user's wrist when tapping their face, removing other motion segments and retaining only the time period of the tapping action for subsequent authentication. By removing other noise present during the tapping action, a clean tapping signal is obtained. Then, features are extracted from the clean tapping signal to build an authentication model, thereby using the motion sensors in the smart wristband device to collect wrist motion signals when fingers tap the face to achieve user authentication. Compared with existing technologies, this invention achieves low-cost, inconspicuous, convenient, and secure authentication with high accuracy using only common motion sensors found in smart wristband devices.

Description

Technical Field

[0001] This invention relates to a user authentication method, and more particularly to an authentication method based on wrist motion signals generated by finger tapping the face using a motion sensor of a smart wristband device, belonging to the field of mobile computing application technology. Background Technology

[0002] With the rise of the global wearable device market, wearable devices are becoming increasingly common in people's lives. During use, these devices inevitably contain some private and sensitive information, such as contact lists, email addresses, and credit card numbers. Under these circumstances, if these devices are accessed without authorization, a large amount of sensitive information may be at risk of being misused. Therefore, in order to protect personal privacy data, it is crucial to research a secure and effective security authentication scheme for wearable devices.

[0003] Traditional authentication methods such as passwords, PIN codes, and gesture passwords are currently the most common authentication means. However, these methods have low security and are easily stolen, leading to privacy breaches. In contrast, biometric-based schemes have the advantages of being difficult to steal and user-friendly, opening up new avenues for the security of wearable devices and attracting widespread attention. Some existing work authenticates users based on fingerprints, voice, iris scans, electroencephalography (EEG), and electromyography (EMG). However, these methods all require specialized sensors and hardware, resulting in high costs and making them unsuitable for large-scale deployment.

[0004] In recent years, methods for authenticating users using built-in sensors (such as microphones, speakers, photoplethysmography sensors, and electrocardiogram sensors) in commercial wearable devices have attracted widespread research interest due to their advantages such as low cost, good integration, and ease of promotion. However, these methods have high requirements for the testing environment; for example, they can only be used in quiet environments or require the wearer to avoid exercise, making them inconvenient for use in daily life.

[0005] Among numerous built-in sensors, motion sensors possess a strong ability to accurately perceive subtle body movements. Currently, motion sensor-based authentication schemes have achieved high-precision user authentication by capturing gait, handwritten signatures, and tapping rhythms. However, these authentication schemes are limited by stringent scenario requirements and do not consider the possibility of attackers deliberately mimicking user behavior (also known as imitation attacks), thus exhibiting significant security vulnerabilities.

[0006] In summary, existing identity authentication methods all have various shortcomings, and new methods are urgently needed to overcome their limitations. Summary of the Invention

[0007] The purpose of this invention is to overcome the technical shortcomings of existing technologies, such as high cost, inconvenience of use, and susceptibility to imitation attacks, and to creatively propose a user authentication method based on motion sensors in wearable devices. This method utilizes a smart wristband device to collect wrist movement information generated by fingers tapping the face, and establishes a connection between this movement information and the unique wrist movements and vibrations during the tapping process, thereby achieving user authentication.

[0008] The innovation of this invention lies in using motion sensors (such as accelerometers and gyroscopes) in a smart wristband device to collect motion signals (including acceleration and angular velocity signals) generated by the wrist when a user taps their face. Other motion segments (such as picking up objects, grabbing the face, walking, etc.) are removed, retaining only the time period of the tapping action for subsequent authentication. By removing other noise present at the time of the tapping action, a clean tapping signal is obtained. Then, features are extracted from the clean tapping signal to build an authentication model, thereby using the motion sensors in the smart wristband device to collect wrist motion signals when the fingers tap the face to achieve user authentication.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] A user authentication method based on a wearable device's motion sensor includes the following steps:

[0011] Step 1: Use motion sensors to collect motion signals from the user's wrist, including data from a three-axis accelerometer and a three-axis gyroscope, and preprocess the signals to eliminate hardware noise, normalize the data, and segment the wrist motion signals.

[0012] Preprocessing can be achieved using the following methods:

[0013] First, remove the noise components introduced by the hardware from the acquired motion signal and normalize the motion signal.

[0014] Then, the short-time energy of the accelerometer Z-axis signal is calculated, and two different thresholds are used to identify the start and end states of wrist movement.

[0015] Step 2: Remove meaningless segments and motion noise from the wrist motion signal to extract the clean tapping signal.

[0016] Specifically, step 2 includes the following steps:

[0017] Since the wrist motion signal fragments obtained in step 1 may capture face-tapping authentication behavior or other meaningless wrist movements, the real face-tapping action fragments are screened out based on the energy ratio of the 10-20Hz and 0-10Hz components of the accelerometer Z-axis signal and the duration of the wrist motion fragments for subsequent analysis and authentication.

[0018] Then, using a convolutional neural network-based method, motion noise (such as noise caused by walking) that exists simultaneously with the face-tapping action is further eliminated from the obtained face-tapping action fragments to extract a clean face-tapping signal.

[0019] Step 3: Extract three types of features representing the face-tapping action.

[0020] From the clean face-tapping signal obtained in step 2, unique motion features, vibration features, and overall features are extracted, and the three types of feature vectors are merged to construct a feature template.

[0021] Step 4: Verify user identity information.

[0022] Using the feature templates constructed in step 3, single-class and multi-class support vector machine classifiers are trained for single-user devices and multi-user devices, respectively.

[0023] When performing user authentication, the user's face tap signal is collected, and the user's biometric template is obtained using steps 1 to 3. A pre-trained support vector machine classifier is then used for authentication.

[0024] Beneficial effects

[0025] The method of the present invention has the following advantages compared with the prior art:

[0026] 1. This invention achieves low-cost, inconspicuous, convenient, and secure identity authentication using only motion sensors commonly found in smart wristband devices. This invention analyzes the wrist motion signals during face tapping and proves that wrist movements and vibrations related to face tapping can be used to identify the wearer. This invention achieves high-precision authentication results with only one face tapping action during the identity verification process, and due to the complex structure of the wrist, it is difficult for attackers to perform imitation attacks.

[0027] 2. This invention removes other body motion segments by utilizing the energy ratio of different frequency band components of the motion signal and the duration of the signal segment, thus distinguishing the face-tapping action from other behaviors in daily life; and designs a denoising model based on a convolutional neural network to filter out the body motion noise that coexists in the face-tapping signal, thereby improving the robustness and reliability of this invention.

[0028] 3. This invention explores a new feature set based on tapping behavior and wrist structure, combining motion features, vibration features and overall features to comprehensively and reliably analyze wrist movement and wrist vibration during face tapping;

[0029] 4. The identity authentication of this invention is robust and effective, achieving an average accuracy rate of 97.28% in an identity authentication trial involving 24 volunteers. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the authentication method based on a wearable device motion sensor according to the present invention.

[0032] Figure 3 This is a schematic diagram illustrating the segmentation of wrist movements according to an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of the energy ratio of different frequency band signals from two users performing face-tapping actions, as proposed in an embodiment of the present invention; wherein, (a) is the energy ratio of the face-tapping signal of user 1 in different frequency bands; and (b) is the energy ratio of the face-tapping signal of user 2 in different frequency bands.

[0034] Figure 5 The performance of authentication in the embodiments of the present invention;

[0035] Figure 6 This invention demonstrates the performance of identity verification under different number of face taps in its embodiments.

[0036] Figure 7 This invention demonstrates the performance of identity verification in different scenarios according to embodiments of the present invention. Detailed Implementation

[0037] The principles and features of the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0038] Figure 1 A schematic diagram illustrating an embodiment of the present invention is shown. When a user taps their face with their fingers, the resulting motion signal is transmitted through the hand skeleton to the wrist and captured by a motion sensor built into the smart wristband device. Because tapping habits (i.e., trajectory, speed, and force) and wrist structures (i.e., bones, muscles, and joints) are unique to each individual, the motion sensor readings will show individual differences. Therefore, the present invention uses the uniqueness of the motion sensor readings to authenticate the user.

[0039] like Figure 2 As shown, an authentication method based on a wearable device's motion sensor includes the following steps:

[0040] Step 1: Use motion sensors to acquire motion signals from the user's wrist, including data from a three-axis accelerometer and a three-axis gyroscope, and preprocess the signals. The purpose of preprocessing is to eliminate hardware noise, normalize the data, and segment the wrist motion signals.

[0041] The specific methods for preprocessing are as follows:

[0042] Step 1.1: Remove hardware-related noise components from the acquired motion signal and normalize the signal.

[0043] Since the wrist motion signals collected by the motion sensors in smart wristband devices are inevitably affected by the noise of the hardware itself, in this embodiment, a Savitzky-Golay filter is used to smooth the signal, thereby removing hardware noise.

[0044] To overcome the differences in the amplitude range of sensor data across different coordinate axes, this embodiment calculates six motion signals a respectively. x ,a y ,a z ,g x ,g y and g z The mean and standard deviation, a x ,a y ,a z ,g x ,g y and g z These refer to the X, Y, and Z axis data from the accelerometer and the X, Y, and Z axis data from the gyroscope, respectively. The Z-score normalization method is used to normalize each of the six motion signals.

[0045] Step 1.2: Calculate the short-time energy of the accelerometer Z-axis signal and use two different thresholds to identify the start and end of wrist movements.

[0046] Specifically, when the user is stationary, the Z-axis acceleration signal a z The amplitude is relatively low and stable; when a user taps their face, a z The signal will jitter violently. Therefore, the objective of this invention is to use this feature to detect the start and end of wrist movements.

[0047] First, use a sliding window to calculate a at time t. z The short-time energy E(t) of the signal is shown in Equation 1:

[0048]

[0049] Among them, a z (β) is a at time β. z Signal amplitude, h is the Hamming window, and L is the window size.

[0050] Two thresholds are set: H = γ1 × TE and I = γ2 × NE, where γ1 and γ2 are constants. In this embodiment, γ1 and γ2 are both 2. TE is the value of the tapping action a pre-collected. z The variance of the short-time energy of the signal, NE is the variance when there is no action. zThe average short-time energy of the signal. If the E(t) of both the current window and subsequent windows is greater than H, the current window is considered the start of the wrist movement event. When the E(t) of a certain window is less than I, the current window is considered the end of the wrist movement event.

[0051] Figure 3 The results of segmented wrist movements are shown.

[0052] Step 2: Remove meaningless segments and motion noise from the wrist motion signal to extract the clean face-tapping signal.

[0053] Step 2.1: Since the wrist motion signal segments obtained in Step 1.2 may capture face-tapping authentication behavior or other meaningless wrist movements, the real face-tapping action segments are selected based on the energy ratio of the 10-20Hz and 0-10Hz components of the accelerometer Z-axis signal and the duration of the wrist motion segments for subsequent analysis and user authentication.

[0054] Since other body movements (such as picking up objects, scratching the face, and walking) also produce significant signal jitter, it is necessary to filter out this noise segment. This is achieved by analyzing a... z Signal analysis revealed that the durations of face-tapping and other body movements were approximately 0.7 seconds and 1.5 seconds, respectively. When picking up objects, scratching the face, walking, and face-tapping, the average energy ratios of the 10-20Hz and 0-10Hz signal components were 0.18, 0.06, 0.01, and 1.62, respectively, showing significant differences. Therefore, different body movements can be distinguished by combining duration and energy ratio.

[0055] Specifically, calculate a z The duration of each wrist movement segment of the signal and the energy ratio of the 10-20Hz to 0-10Hz components. When the duration is below the threshold TH... time And the energy ratio is higher than the threshold TH energy If the event occurs, the segment is identified as a face-tapping event and is thus preserved, while the remaining segments are considered other body movements and are removed.

[0056] Preferably, in this embodiment of the invention, TH time It should be noted that other settings within the range of [0.8s, 1.2s] are also within the scope of this invention; TH energy It should be noted that other settings belonging to [0.5, 1.5] are also within the scope of this invention.

[0057] Step 2.2: Based on the convolutional neural network, further eliminate motion noise (such as noise caused by walking) that exists simultaneously with the face-tapping action from the face-tapping action fragment obtained in Step 2.1, and extract a clean face-tapping signal.

[0058] Furthermore, in order to achieve user authentication in non-stationary states (e.g., walking), this invention proposes a deep learning-based method to filter body motion noise, as follows:

[0059] First, the amplitude spectra of the face-tapping signal with body motion noise and the clean face-tapping signal are used as the prediction variable signal and the network target signal, respectively, and then fed into the denoising model.

[0060] Then, the denoising model learns the difference between the two signals and outputs the amplitude spectrum of the denoised signal.

[0061] Finally, using the amplitude spectrum of the denoised signal and the phase of the noise signal, a clean time-domain signal of the knocking incident is obtained through inverse short-time Fourier transform. Specifically, the denoising model consists of 10 layers: 4 convolutional layers for feature extraction, 3 batch normalization layers to address numerical instability, and 3 calibrated linear unit layers to overcome the vanishing gradient problem.

[0062] Step 3: Extract three types of features representing the face-tapping action.

[0063] Three types of features are extracted from the clean face-tapping signal obtained in step 2, including motion features, vibration features, and overall features. The three types of feature vectors are merged to construct a feature template.

[0064] Specifically, the motion feature refers to the unique face-tapping behavior corresponding to behavioral habits and wrist structure. In this embodiment, the duration of the face-tapping is selected as a feature, and the average value, average deviation, standard deviation, minimum value, and maximum value of linear acceleration and linear angular velocity are extracted as basic features. In addition, commonly used time-domain features such as peak-to-peak value and root square amplitude of linear acceleration and linear angular velocity can also be used to assist in representing wrist movement. Finally, the feature vector FV1 is obtained as the motion feature.

[0065] The vibration features are categorized into time-domain features, which describe the amplitude level of the signal and its temporal similarity, and frequency-domain features, which describe the magnitude of various frequency components in the signal. In this embodiment, widely used statistical features are extracted to characterize wrist vibration, and a random forest-based feature selection algorithm is used to explore key features to achieve rapid authentication with as few features as possible. Finally, the top 15 features most relevant to individual uniqueness are selected based on the feature ranking results, including maximum value, minimum value, variance, standard deviation, kurtosis, skewness, rectifier mean, root mean square, waveform factor, peak factor, impulse factor, sampling point, maximum amplitude, minimum amplitude, and median amplitude. For each tapping signal segment, a feature vector FV2 is obtained as the vibration feature.

[0066] The overall characteristics are first described by using continuous wavelet transform to sequentially decompose six groups of motion signals:

[0067]

[0068] Among them, W f (s,τ) represents the wavelet coefficient sequence obtained after decomposition, and f sequentially refers to the six sets of motion signals a. x ,a y ,a z ,g x ,g y and g z , It is a wavelet basis function with a scaling factor s and a translation factor τ, where dt represents the integral over time t. In this embodiment, Mexican Hat wavelets and a third-level decomposition are used, thus producing 2 3 =8 frequency bands. It should be noted that other settings utilizing different wavelet bases and decomposition levels are also within the scope of this invention. Then, the energy proportions of the signals in different frequency bands are calculated and constructed as the eigenvector FV3:

[0069] FV3={e 1,1 ,e 1,2 ,…,e i,j}, i∈[1,6], j∈[1,8] (3)

[0070] Among them, e i,j It is the energy ratio of the j-th frequency band on the i-th axis, where i takes values ​​from 1 to 6, representing a x ,a y ,a z ,g x ,g y and g z j ranges from 1 to 8, representing the frequency range from the lowest to the highest frequency range, respectively. Figure 4 The data shows the energy ratio of the face-tapping signals from User 1 and User 2 across 48 frequency bands. It can be seen that there are significant differences in the energy ratio between different users.

[0071] Finally, the entire feature template is represented as FV = {FV1, FV2, FV3}.

[0072] Step 4: Verify user identity information.

[0073] Using the feature template constructed in step 3, single-class and multi-class support vector machine classifiers are trained for single-user devices and multi-user devices, respectively. During user authentication, the user's facial tapping signal is collected, and using steps 1 to 3, the user's biometric template is obtained. The pre-trained support vector machine classifier is then used for authentication.

[0074] In a single-user device scenario, the device is bound to only one user, and there are two authentication results: yes / no legitimate user. A single-class support vector machine model (OC-SVM) is trained.

[0075] In a multi-user device environment, where the device is shared by multiple users, there are multiple authentication results: one is a legitimate user, and the other is an attacker. The multi-class support vector machine (OVO-SVM) model consists of multiple binary classifiers.

[0076] Thus, from step 1 to step 4, user identity authentication has been achieved.

[0077] Example verification

[0078] To verify the performance of this invention, a wristband prototype was developed. The motion sensor in the prototype includes a three-axis accelerometer and a three-axis gyroscope. This prototype is equivalent to a smartwatch or smart bracelet. Furthermore, the sampling rate of the motion sensor in the prototype is set to 150Hz.

[0079] A total of 24 volunteers (12 men and 12 women, aged 21 to 50) were recruited to participate in the experiment. During data collection, the 24 volunteers sat comfortably in chairs, wore the prototype on their dominant hand, and repeatedly tapped their faces 30 times with their fingers according to their own habits. Then, data on each volunteer tapping their face in the same manner was collected one month and two months later, for a total of 2160 samples for analysis and training. The collected samples were manually labeled.

[0080] Accuracy, False Acceptance Rate (FAR), and False Rejection Rate (FRR) are used to evaluate system performance. Accuracy is defined as the ratio of the number of samples that correctly verify the identity using this invention to the total number of samples. False Acceptance Rate is defined as the ratio of the number of samples where this invention incorrectly identifies an attacker as a legitimate user to the total number of attacker samples. False Rejection Rate is defined as the ratio of the number of samples where this invention incorrectly identifies a legitimate user as an attacker to the total number of legitimate user samples.

[0081] First, the overall performance of the invention was tested. Figure 5This paper demonstrates the performance of this invention in identity verification of 24 volunteers under both single-class and multi-class support vector machine (SVM) models. For the single-class SVM model, one user is sequentially selected from the 24 volunteers as the legitimate user, while all other users are considered attackers. The training set consists of 80% of the registration data from the selected legitimate user, and the test set consists of 20% of the registration data from the legitimate users and all data from the remaining 23 volunteers. The accuracy, false acceptance rate (FAP), and false rejection rate of this invention are calculated to be 97.02%, 2.67%, and 3.24%, respectively, based on the average values ​​from the 24 volunteers. For the multi-class SVM model, six users are sequentially selected from the 24 volunteers as legitimate users, while all other users are considered attackers, for a total of four rounds of experiments. The training set includes 80% of the registration data from all legitimate users, and the test set includes 20% of the registration data from all legitimate users and data from all attackers. The average values ​​from the four rounds of experiments show that this invention achieves an accuracy of 97.54%, a false acceptance rate of 1.7%, and a false rejection rate of 3.16%, respectively. Combining these two authentication models, the present invention achieves an average accuracy of 97.28%, an average false acceptance rate of 2.19%, and an average false rejection rate of 3.2%. The results demonstrate that the present invention can accurately authenticate users.

[0082] Then, the performance of the invention was tested with different numbers of taps, demonstrating that the invention can achieve a high authentication accuracy with fewer taps. All volunteers underwent authentication testing with 1 to 4 taps. Figure 6 The results show that the accuracy of this invention increases, while the false acceptance and false rejection rates decrease, as the number of taps on the face increases. This is because more taps result in more features provided by the user. However, the invention achieves a high accuracy of 97% with only one tap. Therefore, this invention utilizes a small amount of data to extract stable personal features from the user.

[0083] Finally, the robustness of the invention in practical applications was tested, proving that the invention can achieve a high authentication accuracy in different scenarios. Figure 7 The accuracy, false acceptance rate, and false rejection rate of this invention were displayed when five volunteers tapped their faces 20 times in three scenarios: sitting, standing, and walking. When the user was walking, the accuracy, false acceptance rate, and false rejection rate were 95.72%, 3.62%, and 4.15%, respectively. The results show that although the accuracy of this invention is lowest when the user is walking, it still maintains good performance. This is because, although the mobility of walking inevitably introduces noise that affects the facial tapping signal, this invention successfully removes it by applying deep learning in the motion noise removal module. This confirms that this invention can reliably authenticate users in different scenarios.

[0084] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is used to explain the present invention. It is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A user identity authentication method based on a wearable device motion sensor, characterized by, The method comprises the following steps: Step 1: collect the motion signal of the user's wrist using a motion sensor, including three-axis accelerometer and three-axis gyroscope data, and pre-process the signal to eliminate hardware noise, normalize the data and segment the wrist motion signal; The pre-processing is realized by the following method: First, remove the noise components in the collected motion signal caused by hardware, and normalize the motion signal; Then, the short-time energy of the accelerometer The short-time energy of the axis signals is calculated and two different thresholds are used to identify the start and end state of a wrist movement. Step 2: remove meaningless segments and motion noise from the wrist motion signal, and extract clean knock face signals; First, according to the accelerometer The energy ratio of 10-20 Hz and 0-10 Hz components of the axis signal, and the duration of the wrist movement segment, are used to screen out the real knock action segment for subsequent analysis and authentication; Then, using a method based on convolutional neural network, further eliminate the motion noise existing simultaneously with the knock face action from the obtained knock face action segment, and extract clean knock face signals; Step 3: extract three types of features representing the knock face action behavior; From the clean knock face signal obtained in step 2, extract motion features, vibration features and overall features, and combine the three types of feature vectors to construct a feature template; The duration of the selected knock is taken as a motion feature, and the average value, average deviation, standard deviation, minimum value and maximum value of linear acceleration and linear angular velocity are taken as basic features, respectively as a motion feature Vibration features, time-domain features describe the amplitude level of the signal and its similarity over time, while frequency-domain features describe the size of various frequency components in the signal; statistical features are extracted to characterize the wrist vibration, and a random forest-based feature selection algorithm is used to explore key features, according to the feature ranking results, the features most relevant to the individual uniqueness are selected, including maximum, minimum, variance, standard deviation, kurtosis, skewness, rectifier average, root mean square, wave factor, peak factor, pulse factor, sampling points, amplitude maximum, amplitude minimum and amplitude median; for each knock face signal segment, a feature vector is obtained as a vibration feature; First, use continuous wavelet transform to decompose 6 groups of motion signals in turn: (2) wherein, denotes the sequence of wavelet coefficients obtained after decomposition, denotes the six groups of motion signals in turn and , is a wavelet basis function with a scale factor and a translation factor , denotes the integral over time t; then, the energy proportions of the signal in different frequency bands are calculated, which are constructed as feature vectors : (3) wherein is the first is the first is the energy proportion of the frequency band, takes the values 1 to 6, respectively, indicating and ; takes the values 1 to 8, respectively, indicating the frequency band with the lowest frequency range to the frequency band with the highest frequency range; Finally, the entire feature template is represented as ; Step 4: verify the user's identity information; Using the feature template constructed in step 3, train single-class and multi-class support vector machine classifiers for single-user devices and multi-user devices, respectively; When performing user identity verification, collect the user's knock face signal, obtain the user's biometric feature template using steps 1 to 3, and use the pre-trained support vector machine classifier for identity verification. 2.The user identity authentication method based on a wearable device motion sensor of claim 1, wherein, In step 1, the average and standard deviation of the six motion signals are calculated respectively and , and respectively refer to the , , three-axis data of the accelerometer and the , , three-axis data of the gyroscope; the six motion signals are normalized respectively using the Z-score normalization method; Calculating time using sliding windows At Short-time energy of a signal As shown in Equation 1: （1） wherein is the time of day signal amplitude, is a Hamming window, is the window size; Two thresholds are set I , and are constants, in this embodiment, and are set to 2; is the variance of the short-time energy of the signal of the knock action previously collected, is the average short-time energy of the signal when no action is performed; if both the current window and the subsequent window have greater than , the current window is considered as the start of a wrist motion event; when the of a certain window is less than , the current window is considered as the end of a wrist motion event. 3.The method of claim 1, wherein, In step 2, the duration of each wrist movement segment and the energy ratio of 10-20 Hz to 0-10 Hz components are calculated When the duration is below a threshold and the energy ratio is above a threshold then the segment is identified as a slap event and is kept, the rest of the segments are considered other body movements and are removed, wherein the duration is in the range [0.8s, 1.2s], the energy ratio is in the range [0.5, 1.5]. 4.The method of claim 1, wherein, In step 2, a deep learning-based method is used to filter body motion noise to realize user authentication in a non-stationary state; First, the amplitude spectrum of the knock face signal with body motion noise and the clean knock face signal is used as the prediction variable signal and the network target signal respectively, and is sent into the denoising model; Then, the denoising model learns the difference between the two signals and outputs the amplitude spectrum of the denoised signal; Finally, the amplitude spectrum of the denoised signal and the phase of the noise signal are used to obtain a clean knock face time domain signal through inverse short-time Fourier transform. 5.The method of claim 1, wherein, In step 4, under the condition of single-user device, the device is bound to only one user, and there are two identity authentication results: yes / no legitimate user, and a single-class support vector machine model is trained; Under the condition of multi-user device, the device is shared by multiple users, and there are multiple identity authentication results: a certain legitimate user / attacker; the multi-class support vector machine model is composed of multiple binary classifiers.

Images