Novel smart phone identity verification method based on finger musculoskeletal features

Abstract

The invention relates to a novel smart phone identity verification method based on finger musculoskeletal features, and the method comprises the steps: S1, controlling a linear motor to generate a short-time frequency sweeping vibration signal after receiving a click on a screen; s2, capturing an IMU signal through a built-in accelerometer, and normalizing the IMU signal; S3, respectively extracting a time domain feature, a frequency domain feature and a time sequence feature from the normalized IMU signal; s4, integrating the time domain features, the frequency domain features and the time sequence features into feature vectors in a dimension raising and weighted fusion mode; s5, the feature vectors are input into a trained verification model, a verification result is obtained, and a loss function of the verification model comprises domain alignment loss. Compared with the prior art, the method has the advantages that feature change caused by environment and time migration can be overcome, and high-robustness and high-generalization authentication is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of user identity security authentication, and in particular to a novel smart phone identity verification method based on finger muscle-skeletal characteristics. BACKGROUND

[0002] In modern society, smart phones have become an indispensable tool in people's daily life. With the popularity of smart phones, protecting user privacy and security has become an important issue. Existing smart phone identity verification methods mainly include passwords, pattern locks, fingerprint recognition and facial recognition, etc. These traditional identity authentication methods, although providing security to some extent, are at risk of being attacked by illegal users.

[0003] Firstly, passwords and pattern locks are vulnerable to forgetting or being cracked, while fingerprint recognition and facial recognition technologies, although improving security, can be affected by fake fingerprints and fake headsets in the case of false identification. In noisy environments, authentication methods that rely on acoustic or visual information may be disturbed, reducing the accuracy of verification. In recent years, biometric-based identity verification technology has been widely researched. Studies have shown that the muscle-skeletal characteristics of each person's fingers differ, and this characteristic can be used as a unique biometric feature to identify users. However, existing finger biometric authentication technologies usually rely on complex sensors or expensive hardware devices, which limits their practical application. Therefore, how to maintain high accuracy while providing a simple and convenient identity verification method is an important challenge in the current technical field. The present application proposes a novel smart phone identity verification method that uses the built-in linear vibration motor and accelerometer of the smart phone to capture the user's finger's response to the micro-vibration signal, extract time-frequency domain features and LSTM hidden layer features, and use Transformer to build a model for user authentication. This method can still achieve high-accuracy user authentication in relatively noisy environments, with significant technical advantages and application prospects.

[0004] For example, Chinese patent CN115481380A discloses a continuous identity authentication method based on sliding vibration signals, which belongs to the technical field of mobile computing applications. When the touch screen detects a human sliding action, the vibration motor generates a specific vibration signal, and the accelerometer receives a unique vibration response affected by the finger sliding. Because each person's finger has unique characteristics such as shape, size, bone density, and muscle distribution, it will generate a unique individual difference in vibration response. This method designs a new vibration signal generation mechanism, including the combination of two different types of vibration signals and the addition of random frequency components. The invention extracts different user biometric features from the received two signals, and designs a neural network to eliminate the influence of behavioral characteristics in the user's biometric features. Finally, continuous user identity authentication is achieved by using behavior-independent biometric features. It uses vibration signals for continuous authentication, which to some extent solves the limitations of authentication, but the existing technology including the above method generally faces a difficult problem: domain drift. Even the same legitimate user, in different times, different environments (such as quiet indoor and noisy outdoor) or different holding postures, the biometric feature signals generated will change significantly, resulting in inconsistent feature distribution. This domain drift can seriously reduce the accuracy and generalization of the authentication system, making the model frequently fail in real-world scenarios. SUMMARY

[0005] The purpose of the present application is to provide a new type of smartphone identity verification method based on finger muscle and skeletal characteristics, which uses the built-in linear vibration motor and accelerometer of the smartphone, captures the user's finger response to the micro-vibration signal, extracts time-frequency domain features and LSTM hidden layer features, and uses Transformer to build a model for user authentication. By introducing the domain alignment mechanism, this technical problem is solved.

[0006] The purpose of the present application can be achieved by the following technical solutions: A new type of smartphone identity verification method based on finger muscle and skeletal characteristics, comprising: Step S1: After receiving a click on the screen, control the linear motor to generate a short-time swept-frequency vibration signal; Step S2: Capture the IMU signal through the built-in accelerometer, and normalize the IMU signal: Step S3: Extract time domain features, frequency domain features, and time series features from the normalized IMU signal respectively; Step S4: Integrate the time domain features, frequency domain features, and time series features into a feature vector through dimensionality increase and weighted fusion; Step S5: inputting the feature vector into a trained verification model to obtain a verification result, wherein a domain alignment loss is contained in a loss function of the verification model.

[0007] The sweep vibration signal has a duration of 250 milliseconds and a frequency of 70 Hz to 250 Hz.

[0008] The step S1 comprises: Step S1-1: obtaining a posture classification of a user holding the smart phone; Step S1-2: determining a first detection area based on the posture classification, wherein the first detection area is a part of the screen; Step S1-3: after a click exists only in the first detection area, controlling the linear motor to generate a short-time sweep vibration signal.

[0009] In the normalization process of the step S2, positive values and negative values are normalized separately, and the value of 0 is not changed.

[0010] The time domain features include integral absolute value, average absolute value, variance, root mean square, standard deviation, signal amplitude area, skewness, kurtosis, energy and entropy.

[0011] The frequency domain features include: Peak values and corresponding peak heights extracted from a frequency spectrum generated based on fast Fourier transform; Peak values and corresponding peak heights extracted from discrete cosine transform data; Peak values and corresponding peak heights extracted from wavelet transform coefficients; Peak values and corresponding peak heights extracted from power spectral density data.

[0012] The extraction process of the time sequence feature comprises: Training a long short-term memory network using third training samples, wherein the third training samples include normalized IMU data and labeled results; Extracting hidden layer features as time sequence features based on the trained long short-term memory network.

[0013] The step S4 comprises: Step S4-1: concatenating the time domain features and the frequency domain features after being standardized respectively to obtain time-frequency domain statistical features; Step S4-2: using a multi-layer perception machine to increase the dimension of the standardized time-frequency domain statistical features and the frequency domain features to the same dimension as the time sequence features; Step S4-3: weighting and fusing the increased time-frequency domain statistical features and the time sequence features to obtain a feature vector: Wherein: Xfinal For feature vectors, X LSTM As a time series feature, X stats_mapped For the time-frequency domain statistical features after dimensionality upgrade, α These are learnable fusion weight parameters.

[0014] The loss function of the validation model is: Loss = ContrastiveLos s+ Alignloss + CrossEntropyLoss in: Loss To verify the model's loss function, ContrastiveLos s represents the total comparison loss. Alignloss For domain alignment loss, CrossEntropyLoss For cross-entropy loss, N The number of sample pairs, y i For the first i Labels for each sample pair D 2 (·) represents the square of the distance between sample pairs in the embedding space, and D(·) represents the distance between sample pairs in the embedding space. x i For the first i The input of one sample in a pair of samples, x i 'For the first i The input of one sample in a sample pair, m The preset boundary threshold, d As the dimension of the feature, X s Characteristics of the source domain X t Features of the target domain C For the number of categories, For the first i The model prediction results for each sample pair Cov (·) is the covariance matrix.

[0015] A novel smartphone authentication device based on finger muscle and skeletal features includes a memory, a processor, and a program stored in the memory, characterized in that the processor executes the program to implement the method described above.

[0016] Compared with the prior art, the present invention has the following advantages: by using the linear vibration motor and accelerometer built into the smartphone, the time-frequency domain features and LSTM hidden layer features are extracted by capturing the user's finger response to the tiny vibration signal, and the Transformer is used to build a model for user authentication. Furthermore, by introducing a domain alignment mechanism, this technical pain point is solved. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the technical principle of the present invention; Figure 2 This is a time-domain diagram of the swept-frequency vibration signal of the present invention; Figure 3 This is the frequency domain diagram of the swept vibration signal of the present invention; Figure 4 This is a schematic diagram of positive and negative sample pairs in this invention; Figure 5 This is a schematic diagram of the main steps of the method of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0019] A novel smartphone authentication method based on finger muscle and skeletal features, such as Figure 1 and Figure 5 As shown, it includes: Step S1: Upon receiving a tap on the screen, control the linear motor to generate a short-duration sweep frequency vibration signal, including: Step S1-1: Obtain the user's posture classification when holding the smartphone; Step S1-2: Determine the first detection region based on pose classification, wherein the first detection region is a part of the screen; Step S1-3: Once a click occurs only in the first detection area, control the linear motor to generate a short-term sweep frequency vibration signal.

[0020] In this example, a swept-frequency vibration signal is generated using a linear vibration motor configured on a smartphone. The smartphone model is OPPO Find X6. The vibration signal is modulated using the RichTap library, with a frequency range of 70Hz-250Hz and a duration of 250ms. Figure 2 and Figure 3As shown. The frequency range and sweep pattern of the signal have been optimized to fully stimulate the user's finger response, thereby capturing sufficient information for user feature extraction. This application uses a "click" interaction to trigger authentication. Compared to continuous actions such as "swiping," a "click" is a standardized, transient impulse response test.

[0021] Furthermore, for pose classification, this embodiment includes: With the left hand supporting the bottom, that is, the three fingers of the left hand are located at the lower part of the back of the phone, and the little finger of the left hand is located at the bottom of the phone's frame. The height of the top of the phone is higher than the height of the bottom. At this time, the first detection area is a rectangle, with its right boundary being the right boundary of the phone's screen, the distance between the left and right boundaries being the pre-configured first length, approximately the length of a thumb, the lower boundary being the lower boundary of the phone's screen, and the upper boundary being the horizontal center line or one-third line of the screen.

[0022] The right-hand bottom support is mirror-symmetrical to the left-hand bottom support, and the axis of symmetry is the vertical center line of the phone screen.

[0023] With the left hand supporting the phone, the four fingers of the left hand are positioned slightly below the back of the phone, and the left thumb is positioned on the left side of the phone's frame. In this position, the first detection area is similar to and concentrically positioned with the phone screen.

[0024] The right hand provides vertical support, mirroring the left hand's horizontal support. The axis of symmetry is the vertical midline of the phone screen.

[0025] With the left hand flat on the ground, that is, the four fingers of the left hand are located on the lower part of the back of the phone, and the angle between the center line of the phone screen and the horizontal plane is less than 5°, the first detection area is a rectangle, with its left boundary being the left boundary of the phone screen, its right boundary being the right boundary of the phone screen, its lower boundary being the lower boundary of the phone screen, and its upper boundary being the horizontal center line of the screen. The right hand supports the object horizontally, which is a mirror image of the left hand supporting the object horizontally. The axis of symmetry of the object is the vertical center line of the phone screen.

[0026] The specific attitude classification can be obtained using existing technologies, such as through IMU signals or by setting a conductive sheet on the mobile phone for classification detection. For details, please refer to the design of CN 104679358 A. To avoid obscuring the purpose of this invention, it will not be described in detail here.

[0027] Through the above posture classification, some interference can be filtered out by designing a first detection area, thereby reducing power consumption. In addition, if the power consumption of the program itself is reduced after optimization, posture classification is not necessary.

[0028] Step S2: Capture the IMU signal using the built-in accelerometer and normalize the IMU signal: In this embodiment, the accelerometer built into the smartphone is used to capture the user's finger response to a sweeping vibration signal. Specifically, when the user taps the smartphone screen at a specific location, the accelerometer records the acceleration changes of the user's finger in three axes (X, Y, and Z). This data is recorded in the form of three-axis accelerometer data, containing the dynamic response of the user's finger to the vibration signal. This three-axis acceleration data reflects the unique response pattern generated by the user's finger during operation. This pattern can accurately capture the interaction characteristics between the user and the smartphone, providing an important basis for the subsequent user authentication process. This method significantly reduces the interference of user behavior characteristics (such as swiping speed, duration, and pressure changes) from the data acquisition source, allowing the acquired response signal to more purely focus on reflecting the user's inherent finger musculoskeletal characteristics. The user holds the smartphone with their left hand and taps any location on the phone screen with their right index finger; this authentication action is more in line with the user's daily smartphone usage habits. The accelerometer receives feedback and stores the vibration signal as a .txt file containing three-axis (X, Y, and Z axis) acceleration data.

[0029] In addition, during the normalization process, positive and negative values ​​are normalized separately, without changing the value of 0.

[0030] Step S3: Extract time-domain features, frequency-domain features, and time-series features from the normalized IMU signal; Time-domain characteristics include integral absolute value, mean absolute value, variance, root mean square, standard deviation, signal amplitude area, skewness, kurtosis, energy, and entropy.

[0031] Frequency domain characteristics include: Peak values ​​and corresponding peak heights extracted from the spectrum generated by the Fast Fourier Transform; Peak values ​​and corresponding peak heights extracted from discrete cosine transform data; Peak values ​​and corresponding peak heights extracted from wavelet transform coefficients; Peak values ​​and their corresponding heights are extracted from power spectral density data.

[0032] The process of extracting temporal features includes: The Long Short-Term Memory (LSTM) network is trained using a third training sample, which includes normalized IMU data and labeled results. Based on the trained Long Short-Term Memory network, hidden layer features are extracted as temporal features.

[0033] Step S4: Integrate the time-domain features, frequency-domain features, and time-series features into a feature vector through dimensionality increase and weighted fusion, including: Step S4-1: Standardize the time-domain features and frequency-domain features respectively, and then concatenate them to obtain the time-frequency domain statistical features; Step S4-2: Use a multilayer perceptron to upgrade the standardized time-frequency domain statistical features and frequency domain features to the same dimension as the time-series features. In this embodiment, it is 128 dimensions, and the time-frequency domain statistical features before the upgrade are 108 dimensions. Step S4-3: Weighted fusion of the upgraded time-frequency domain statistical features and time-series features to obtain the feature vector: in: X final For feature vectors, X LSTM As a time series feature, X stats_mapped For the time-frequency domain statistical features after dimensionality upgrade, α The fusion weights are learnable parameters, with an initial value of 0.5.

[0034] The backpropagation algorithm automatically adjusts the weight values, enabling the model to automatically balance the contributions of details (from statistical features) and stability (from LSTM features) to achieve optimal overall certification performance under the challenges of noisy environments and variable domains. The backpropagation algorithm automatically adjusts the weight values ​​so that the fused features maximize model performance. During training, α is continuously updated, eventually converging to the optimal value α = 0.8209, meaning that LSTM features contribute approximately 82.09% and statistical features contribute approximately 17.91%. The weight update rule using gradient descent is as follows: in, For learning rate, This is the gradient of the loss function with respect to the fusion weight parameters.

[0035] This step is the core of this application. Its purpose is to fuse the extracted fine-grained statistical features with the extracted robust depth features to achieve complementary advantages. Fine-grained statistical features provide rich signal details (such as energy and specific frequency peaks) and have high discriminative potential, but are susceptible to noise and motion artifacts. Robust depth features, through temporal learning using an LSTM model, have filtered out most random interference, capturing stable and abstract biological signatures with strong robustness.

[0036] Step S5: Input the feature vector into the trained validation model to obtain the validation result. The loss function of the validation model includes domain alignment loss.

[0037] The improvement in this step mainly lies in the Transformer model and the domain alignment loss. Traditional CNN models struggle to effectively process the generated feature vectors. The fused feature vectors have a complex internal structure, containing both macroscopic statistical information and microscopic temporal information. This application employs the Transformer model, leveraging its powerful self-attention mechanism to effectively capture the long-distance dependencies within these highly complex features.

[0038] More importantly, the necessity of domain alignment loss lies in addressing the "domain drift" problem mentioned in the background art (i.e., feature changes caused by different environments, times, and grip postures). This application introduces domain alignment loss (Align_Loss) during training. This loss function forces the model to learn domain-invariant features, minimizing the feature distribution differences between data from different domains (different collection batches). Through "domain alignment" optimization, this invention significantly improves the generalization and robustness of the authentication system in real-world, ever-changing scenarios, a key problem that existing technologies have failed to solve.

[0039] In this embodiment, the loss function for validating the model is: Loss = ContrastiveLos s+ Alignloss + CrossEntropyLoss in: Loss To verify the model's loss function, ContrastiveLos s represents the total comparison loss. Alignloss For domain alignment loss, CrossEntropyLoss For cross-entropy loss, N The number of sample pairs, y i For the first i The labels of each sample pair (1 for similarity, 0 for dissimilarity). D 2 (·) represents the square of the distance between sample pairs in the embedding space, and D(·) represents the distance between sample pairs in the embedding space. x i For the first i The input of one sample in a pair of samples, x i 'For the first i The input of one sample in a sample pair, m The preset boundary threshold, d As the dimension of the feature, X sCharacteristics of the source domain X t Features of the target domain C For the number of categories, For the first i The model prediction results for each sample pair Cov (·) is the covariance matrix.

[0040] For total comparison loss, such as Figure 4 As shown, the data is divided into positive and negative sample pairs. Data pairs within the same class are positive sample pairs, and data pairs from different classes are negative sample pairs. The contrast loss is calculated to bring positive sample pairs closer together and negative sample pairs further apart. The Transformer model is trained by combining the three losses into a single overall loss, which is used to learn the similarities and differences between classes, domains, and samples.

[0041] The experimental verification process for this application is as follows: 1) After initial registration or identity verification, users will undergo a secondary verification step to ensure the accuracy of their identity. In this step, the user needs to tap the smartphone screen again. During this process, the smartphone's linear vibration motor emits a specific swept-frequency vibration signal with a predetermined frequency variation pattern. This vibration signal is designed to stimulate the user's finger response to the vibration, thereby acquiring the user's biometric data. The smartphone's built-in accelerometer will begin operating simultaneously with the user tapping the screen. It will accurately capture the dynamic response of the user's finger to the emitted vibration signal. The accelerometer records signals including the acceleration changes of the user's finger in three-dimensional space, measured along the X, Y, and Z axes. These measurements provide detailed data on how the user's finger responds to the vibration signal. These datasets are combined into a comprehensive set of response signal data, reflecting the characteristics of the user's finger when interacting with the vibration signal. This signal data includes not only the acceleration changes of the finger in different directions but also minute variations that may occur during the transmission of the vibration signal. By analyzing this acceleration data, the system can capture the unique biometric characteristics of the user's finger during operation, providing crucial evidence for user identity verification.

[0042] 2) When processing the collected response data, the first step is to normalize it to ensure it falls within a uniform range. The main purpose of normalization is to eliminate data bias caused by different acquisition conditions (such as device sensitivity, environmental noise, and user finger contact status), making the data more consistent and comparable. This process involves linearly transforming each data point to ensure all data are within the same range, for example, scaling the data to the interval [0, 1] for easier subsequent analysis. After normalization, further mathematical processing is performed to extract key information. First, time-domain features are extracted from the normalized data. These features include absolute value, variance, and entropy. These features help describe the basic statistical properties and fluctuations of the signal. For example, absolute value reflects the amplitude variation of the signal, variance measures the range of signal fluctuations, and entropy describes the complexity and randomness of the signal. Next, the normalized data is further processed through frequency domain analysis. Using the Fast Fourier Transform (FFT), the time-domain signal is converted into a frequency-domain signal, generating spectral data. Spectral data reveals the energy distribution of the signal at different frequencies, showcasing its frequency characteristics and periodic components. These frequency domain features are crucial for understanding the signal's periodicity and frequency variations, reflecting the response patterns of a user's finger to specific vibration signals. By combining time-domain and frequency-domain features, the system can comprehensively describe the user's finger's response to vibration signals. These features collectively provide detailed dynamic characteristics of the user's finger during the response process, enabling the system to accurately capture the user's unique biometrics, thus providing reliable data support for subsequent authentication and classification.

[0043] 3) The current user's data is processed through a trained LSTM model to obtain the hidden layer features. In this step, a Long Short-Term Memory (LSTM) network model is used to extract the contextual dependencies of the time series from the normalized accelerometer data. First, the collected accelerometer data is standardized to ensure that its numerical range is on the same scale before being input into the LSTM network for processing. The LSTM model can effectively capture long-term dependencies in time series data, solving the gradient vanishing problem of traditional RNNs when processing long sequences. Through the internal gating mechanism of the LSTM network, the model can learn and store important information step by step according to the time steps of the input data, while forgetting irrelevant information. The accelerometer data is passed through the forward propagation of the LSTM model to obtain a series of hidden layer states. These hidden layer states contain the dependencies and feature expressions between each time step in the input data, thus generating hidden layer features containing contextual dependency information. In the specific implementation, each timestamp data of the accelerometer is input into each time step of the LSTM model. After passing through several LSTM layers, the last hidden state vector output is the contextual feature representing the entire time series. This feature vector not only contains the raw information from the accelerometer data but also incorporates dynamic changes and patterns in the time series, serving as the foundation for subsequent feature fusion and user identification. Through this process, the LSTM model effectively extracts the temporal information and context-dependent features from the accelerometer data. These features will serve as important information sources in subsequent feature fusion and classification processes, helping to distinguish the behavioral characteristics of different users.

[0044] 4) The extracted features are integrated into a feature vector using feature dimensionality upscaling and linear weighting to obtain the final feature representation. Specifically, firstly, the context-dependent features extracted from the accelerometer data by the LSTM network are processed and used as a representation of temporal features. Secondly, the temporal and frequency domain features are upscaled through a fully connected layer to ensure that their dimensions match the LSTM hidden layer features, forming a unified feature space. Finally, a linear weighting method is used to fuse the upscaled temporal and frequency domain features with the LSTM hidden layer features into a 128-dimensional feature vector, which serves as the user's identity information representation data. The goal of this process is to fuse multiple types of features into a unified representation that can be used for user authentication.

[0045] 5) The user's features are compared with the trained model. Based on the classification results and threshold determination, it is judged whether the user's data belongs to a legitimate category and whether the user's identity belongs to a legitimate category. If the data matches, authentication is successful, and the user is identified as a legitimate user within the system. If the data does not match, authentication fails, and the user is identified as an illegitimate user. This matching verification ensures that only authorized users can enter the system, effectively preventing unauthorized users from accessing it. Conversely, if the data does not match, the system will fail authentication. The system will identify the user as an illegitimate user and automatically deny their access rights. This mechanism ensures the security of the system and prevents unauthorized users from gaining access through impersonation or other means. The entire process, through efficient data comparison and classification algorithms, guarantees the accuracy of user authentication and the overall security of the system.

[0046] In this example, 100 volunteers participated in the experiment. Each volunteer held a smartphone in their left hand and tapped randomly on the screen with their right index finger. Each volunteer tapped the screen 100 times per batch, collecting data from 5 batches, for a total of 500 taps. Each tap triggered a vibration signal from the smartphone's motor, while an accelerometer recorded the user's finger's response to the vibration. Users did not need to maintain a specific posture while tapping; they only needed to ensure their left hand was stable and their right index finger completed the tapping action. The tapped screen locations were also randomly selected to ensure experimental diversity. The results showed a classification accuracy of 99.4% for the 100 users. This high accuracy indicates that our method can effectively distinguish users by extracting the musculoskeletal features of the user's fingers from the vibration signal. In practical applications, after registering in the system, users successfully authenticated and entered the system every time they performed the authentication process again. Meanwhile, unregistered users (i.e., attackers) were accurately identified as unauthorized users through similar tapping actions and were unable to access the system. This demonstrates that our system can effectively utilize subtle features in accelerometer data to achieve high-accuracy user authentication.

[0047] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A novel smartphone identity verification method based on finger musculoskeletal features, characterized in that, The application relates to a method for detecting a touch on a screen. Step S1: after receiving a touch on the screen, a short sweep vibration signal is generated by controlling a linear motor; Step S2: an IMU signal is captured by a built-in accelerometer, and the IMU signal is normalized; Step S3: time domain features, frequency domain features and time sequence features are extracted from the normalized IMU signal; Step S4: the time domain features, the frequency domain features and the time sequence features are integrated into a feature vector through dimensionality lifting and weighted fusion; Step S5: the feature vector is input into a trained verification model to obtain a verification result, wherein a loss function of the verification model contains a domain alignment loss.

2. The novel smart phone identity verification method based on finger musculoskeletal features according to claim 1, characterized in that, The sweep vibration signal has a duration of 250 milliseconds and a frequency of 70 Hz to 250 Hz.

3. The novel smart phone identity verification method based on finger musculoskeletal features according to claim 1, characterized in that, The step S1 comprises: Step S1-1: obtaining a posture classification of a user holding a smart phone; Step S1-2: determining a first detection area based on the posture classification, wherein the first detection area is a part of the screen; Step S1-3: after a touch exists only in the first detection area, a short sweep vibration signal is generated by controlling the linear motor.

4. The novel smart phone identity verification method based on finger musculoskeletal features according to claim 1, characterized in that, In the normalization process of the step S2, positive values and negative values are normalized separately, and the value of 0 is not changed.

5. The novel smart phone identity verification method based on finger musculoskeletal features according to claim 1, characterized in that, The time domain features include integral absolute value, average absolute value, variance, root mean square, standard deviation, signal amplitude area, skewness, kurtosis, energy and entropy.

6. The novel smart phone identity verification method based on finger musculoskeletal features according to claim 1, characterized in that, The frequency domain features comprise: Peak values and corresponding peak heights extracted from a frequency spectrum generated based on fast Fourier transform; Peak values and corresponding peak heights extracted from discrete cosine transform data; Peak values and corresponding peak heights extracted from wavelet transform coefficients; Peak values and corresponding peak heights extracted from power spectral density data.

7. The novel smart phone identity verification method based on finger musculoskeletal features according to claim 1, characterized in that, The extraction process of the time sequence features comprises: Training a long short-term memory network by using third training samples, wherein the third training samples comprise normalized IMU data and labeled results; Based on the trained long short-term memory network, hidden layer features are extracted as time sequence features.

8. The novel smart phone identity verification method based on finger musculoskeletal features according to claim 1, characterized in that, The step S4 comprises: Step S4-1: after the time domain features and the frequency domain features are standardized respectively, time-frequency domain statistical features are obtained by splicing; Step S4-2: the standardized time-frequency domain statistical features and frequency domain features are lifted to the same dimension as the time sequence features by using a multilayer perception machine; Step S4-3: the lifted time-frequency domain statistical features and the time sequence features are weighted and fused to obtain a feature vector: wherein: X final is a feature vector, X LSTM is a time sequence feature, X stats_mapped is a statistical feature in the time-frequency domain after dimensionality increase, α is a learnable fusion weight parameter.

9. The novel smart phone identity verification method based on finger musculoskeletal features according to claim 1, characterized in that, The loss function of the verification model is: Loss = ContrastiveLos s+ Alignloss + CrossEntropyLoss in: Loss To verify the model's loss function, ContrastiveLos s represents the total comparison loss. Alignloss For domain alignment loss, CrossEntropyLoss For cross-entropy loss, N The number of sample pairs, y i For the first i Labels for each sample pair D 2 (·) represents the square of the distance between sample pairs in the embedding space, and D(·) represents the distance between sample pairs in the embedding space. x i For the first i The input of one sample in a pair of samples, x i 'For the first i The input of one sample in a sample pair, m The preset boundary threshold, d As the dimension of the feature, X s Characteristics of the source domain X t Features of the target domain C For the number of categories, For the first i The model prediction results for each sample pair Cov (·) is the covariance matrix. 10.A novel smart phone identity verification device based on finger musculoskeletal features, comprising a memory, a processor, and a program stored in the memory, characterized in that, The processor implements the method according to any one of claims 1-9 when executing the program. The application relates to a method for detecting a touch on a screen.

Images