An identity authentication method and system based on ultrasonic wave sensing hand features

Abstract

The application discloses an identity authentication method and system based on ultrasonic wave sensing hand features, and the method comprises the following steps: acquiring a hand reflection sensing signal by using an active acoustic sensing technology; performing audio signal preprocessing and target region interception on the hand reflection sensing signal, and extracting a hand biometric feature matrix; inputting the hand biometric feature matrix as a hand authentication feature into a pre-trained authentication neural network, outputting a hand authentication vector by a feature extractor; inputting the hand authentication vector into an identity recognizer, matching the hand authentication vector with a reference vector, and outputting a user identity authentication result; and simultaneously using a continuous learning mechanism to guarantee the long-term performance of the authentication neural network. The application can not only realize a high-security static authentication, but also effectively overcome the problem of accuracy reduction caused by data drift in long-term use.

Description

Technical Field

[0001] This invention belongs to the field of acoustic sensing technology, specifically relating to an identity authentication method and system based on ultrasonic sensing of hand features. Background Technology

[0002] Password authentication is the most common identity verification method in mobile applications, widely used in high-risk areas such as online banking and electronic payments. However, current mainstream auxiliary authentication solutions all have shortcomings: fingerprint recognition is affected by the state of the finger and lacks stability; facial recognition requires the collection of visual physiological features, posing a risk of privacy leakage; and iris recognition devices are expensive and difficult to popularize in consumer-grade mobile devices.

[0003] Active acoustic sensing technology has become a research hotspot due to its advantages of low deployment cost, high versatility, and no privacy leakage risk. It only utilizes the phone's native hardware and does not generate visual physiological features, making it suitable for mobile scenarios. Combined with CIR data (channel impulse response data), the phone can perceive the shape of the hand and the acoustic channel changes caused by micro-movements at the sub-centimeter level, providing support for high-precision recognition.

[0004] However, existing ultrasonic sensing solutions are adapted to continuous, large-amplitude gestures, making it difficult to handle the short duration and small amplitude characteristics of tapping gestures in password input scenarios. When used in conjunction with CIR (Channel Impulse Response) technology for secondary authentication, three major bottlenecks exist: first, poor consistency in biometric features, with diverse user tapping postures and password combinations leading to significant intra-class variations in CIR data; second, a large requirement for registration samples, increasing the burden on users; and third, insufficient long-term stability, with static models unable to adapt to user behavior drift, resulting in decreased authentication accuracy. Therefore, developing a highly secure, low-burden, and robust enhanced authentication method is an urgent problem to be solved in this field. Summary of the Invention

[0005] This invention provides an identity authentication method and system based on ultrasonic sensing of hand features, aiming to solve the problems in the existing technology, such as poor consistency of biometric features due to the diversity of user behavior and password combinations, high user registration costs due to the large number of model training samples, and the inability of static authentication models to adapt to feature drift, which leads to a decrease in accuracy.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The first aspect of this invention provides an identity authentication method based on ultrasonic sensing of hand features, comprising:

[0008] Active acoustic sensing technology is used to identify and acquire hand reflection sensing signals; the hand reflection sensing signals are preprocessed with audio signals and the target region is extracted to extract the hand biometric feature matrix;

[0009] In response to the user registration phase, a probability-adaptive frame-level data augmentation strategy is used to expand the hand biometric feature matrix to obtain hand registration features. All hand registration features are then input into a pre-trained authentication neural network, which includes a feature extractor and an identity recognizer. The feature extractor outputs a reference vector, which is then added with the user's real identity label and stored in the hand feature registration information database.

[0010] In response to the user authentication phase, the hand biometric matrix is ​​input as a hand authentication feature into a pre-trained authentication neural network, and the feature extractor outputs a hand authentication vector; the hand authentication vector is input into the identity recognizer, and the hand authentication vector is matched with a reference vector to output the user identity authentication result;

[0011] In response to successful user authentication, the system dynamically updates the baseline vector in the hand feature registration information database using the hand authentication vector to obtain an online sample set. The authentication neural network is then trained and updated using the online sample set at set intervals.

[0012] Furthermore, active acoustic sensing technology is used to identify and acquire hand reflection sensing signals, specifically including:

[0013] An ultrasonic wave signal, specifically a frequency-modulated continuous wave (FMCW) signal, is transmitted to the user's hand via an audio player on a mobile device. An audio receiver receives the hand reflection signal generated during password input. The formula is as follows:

[0014]

[0015] In the formula, The hand reflection perception signal is the result of superimposed multipath signals. This represents the number of signal propagation paths in a complex space environment. The ultrasonic wave signal sent to the audio player. For audio receiver to receive the first The time it takes for a signal to propagate along a path. For the first The propagation delay of a signal along a path. For the first The attenuation ratio of the signal propagating along the path.

[0016] Furthermore, the frequency range of the FM continuous wave signal is set to 18-22kHz, the duration is set to 0.02 seconds, and the sampling frequency of the audio receiver is set to 48kHz.

[0017] Furthermore, the hand reflection perception signal undergoes audio signal preprocessing and target region extraction to extract the hand biometric feature matrix; specifically including:

[0018] The acquired hand reflection sensing signal is subjected to high-pass filtering to remove low-frequency environmental noise interference, and the filtered hand reflection sensing signal is synchronized in frames using ultrasonic transmission signal to obtain calibration echo signal.

[0019] The channel impulse response is obtained by calculating the cross-correlation between the calibration echo signal frame and the ultrasonic transmission signal frame frame by frame, and a global time series matrix is ​​constructed from the channel impulse response of each frame; the effective feature matrix of the hand is obtained by truncating the global time series matrix based on the set sensing distance range.

[0020] The effective feature matrix of the hand is subjected to adjacent frame difference operation to obtain the difference feature matrix; a soft threshold is set according to the average value of the difference feature matrix; local maxima points in the difference feature matrix are located and filtered according to the soft threshold to obtain the peak position; the difference features of the peak position in the difference feature matrix are replaced with the soft threshold to obtain sparse response features;

[0021] After smoothing the sparse response features, the scale normalization operation is performed using the max-min normalization method to obtain the hand biometric matrix.

[0022] Furthermore, the cross-correlation between the calibration echo signal frame and the ultrasonic transmission signal frame is calculated frame by frame to obtain the channel impulse response. A global time series matrix is ​​constructed from the channel impulse responses of each frame, specifically including:

[0023]

[0024]

[0025] In the formula, For the first Frame channel impulse response, This represents the multipath delay, where k is the index of the sampling point within the calibration echo signal frame and the ultrasonic transmission signal frame, and K is the number of sampling points. For the first The calibration echo signal of the frame, It is an ultrasonic wave signal; For the global time series matrix, This represents the length of the global time series.

[0026] Furthermore, a mean filter with a convolution kernel size of 3×3 is used to smooth the sparse response features.

[0027] Furthermore, a probability-adaptive frame-level data augmentation strategy is used to expand the hand biometric matrix to obtain hand registration features, specifically including:

[0028] The Otsu algorithm (or simply Otsu algorithm) is used to binarize the hand biometric feature matrix to obtain the binary mask matrix for each frame; the enhancement probability of the binary mask matrix for each frame is calculated, expressed by the following formula:

[0029]

[0030] In the formula, Let be the enhancement probability for the j-th frame. Let J be the binary mask matrix of the j-th frame. Let the first dimension of the binary mask matrix of the j-th frame be the size; Let the second dimension of the binary mask matrix of the j-th frame be the size of the second dimension. The index of each element in the binary mask matrix of each frame;

[0031] Affine shearing transformation is applied to the hand biometric feature matrix along the frame axis to obtain hand scaling features, and the random value range of the scaling coefficient is set to [0.8, 1.2]. Using the enhancement probability as the trigger probability, random frame interpolation or discarding operation is performed on the hand scaling features to obtain the first hand extended features.

[0032] Based on enhanced probability The amplitude attenuation factor is calculated using the following formula:

[0033]

[0034] In the formula, The amplitude attenuation factor for the j-th frame; The attenuation coefficient is randomly generated, and its value ranges from [0.6, 1.2].

[0035] The second extended hand feature is obtained by attenuating the hand biometric matrix based on the amplitude attenuation factor; the third extended hand feature is obtained by randomly generating a delay transform factor with a value range of [0.7, 1.5] and performing a delay axis affine shearing transform on the hand biometric matrix based on the delay transform factor.

[0036] The hand registration features are composed of the hand biometric matrix, the first hand extended features, the second hand extended features, and the third hand extended features.

[0037] Furthermore, the authentication neural network includes a feature extractor, a feature projector, a password discriminator, and an identity recognizer. The feature extractor contains three convolutional blocks, each of which sequentially includes a convolutional layer, a batch normalization layer, a spatial dropout layer, a max pooling layer, and a ReLU activation function.

[0038] Furthermore, the hand authentication vector is matched with the baseline vector to output the user authentication result, specifically including:

[0039] Clustering is performed on the baseline vectors in the hand feature registration information database based on the user's real identity tag to determine the first user class center. The formula is as follows:

[0040]

[0041]

[0042] In the formula, For real tags to identify users The One reference vector; For real tags to identify users The set of reference vectors; Calculate the function for cluster centers; For the set of reference vectors Cluster centers;

[0043] Calculate the feature space distance between the hand authentication vector and the center of the first user class; compare the minimum feature space distance with a preset security threshold. If the minimum feature space distance is less than or equal to the preset security threshold, determine that the current user is a legitimate user and output the corresponding identity; otherwise, determine that the user is an illegitimate user and refuse authentication.

[0044] Furthermore, the online sample set is obtained by dynamically updating the baseline vector in the hand feature registration information database using the hand authentication vector, specifically including:

[0045] The second user class center is obtained by superimposing the hand authentication vector with the set of baseline vectors with the same user identity and recalculating the cluster center.

[0046]

[0047] In the formula, For hand authentication vectors, For real tags to identify users The set of reference vectors; Calculate the function for cluster centers; For the second user class center;

[0048] Obtain the maximum intra-class distance and minimum inter-class distance of the baseline vector in the hand feature registration information database; calculate the current distance from the hand authentication vector to the center of the second user class; if the current distance is less than the minimum inter-class distance and the maximum intra-class distance, replace the baseline vector with the largest intra-class distance using the hand authentication vector, and use the updated baseline vector set as the online sample set.

[0049] Furthermore, the training process of the authentication neural network specifically includes:

[0050] An offline hand feature matrix is ​​obtained from an ultrasonic sensing hand feature database; an offline sample set is obtained by expanding the offline hand feature matrix based on a probability-adaptive frame-level data augmentation strategy, wherein the offline sample set is configured with real user identity labels and real password labels;

[0051] The authentication neural network is trained using an offline sample set. The feature extractor outputs a hand recognition vector, which is then input into the identity recognizer to obtain the user's identity recognition result.

[0052] Anchor samples are determined from the offline sample set. The similarity between the anchor samples and offline samples of the same class, as well as the difference between offline samples of different classes, are measured. After temperature coefficient scaling and normalization, the supervised contrast loss is output, expressed by the following formula:

[0053]

[0054] In the formula, To monitor and compare losses; , and For sample index, For anchor point samples, This is the set of offline samples from the previous batch. For offline samples of the same type as anchor samples, For the positive sample set belonging to the same user as the anchor sample, For offline samples other than anchor point samples, This refers to the set of offline samples excluding anchor samples. For temperature parameters;

[0055] Based on the user identity recognition results and the user's real identity label, the identity recognition loss is calculated using the cross-entropy function;

[0056] The hand recognition vector is input into the password recognizer to obtain the password recognition probability. The password recognizer includes a first fully connected layer, a one-dimensional batch normalization layer, a ReLU activation function, a second fully connected layer, and a log-Softmax layer. Based on the password recognition probability and the true password label, the password authentication loss is calculated using the binary cross-entropy function.

[0057] The joint loss is obtained by weighted summation of the supervision comparison loss, identity recognition loss, and password authentication loss.

[0058] Using the joint loss as a supervision signal, the parameter gradient of the certification neural network is calculated through backpropagation, and the parameters are updated along the gradient descent direction using an adaptive momentum optimizer. The process of training the certification neural network using an offline sample set is repeated iteratively until the joint loss converges and outputs the trained certification neural network.

[0059] A second aspect of the present invention provides an identity authentication system based on ultrasonic sensing of hand features, comprising:

[0060] The feature acquisition module uses active acoustic sensing technology to identify and acquire hand reflection sensing signals; it performs audio signal preprocessing and target region extraction on the hand reflection sensing signals to extract the hand biometric feature matrix;

[0061] The user registration module is used to expand the hand biometric feature matrix based on a probability-adaptive frame-level data augmentation strategy to obtain hand registration features. All hand registration features are input into a pre-trained authentication neural network, and the feature extractor outputs a baseline vector. The baseline vector is then added with the real user identity label and stored in the hand feature registration information database.

[0062] The user authentication module is used to input the hand biometric feature matrix as the hand authentication feature into a pre-trained authentication neural network, and the feature extractor outputs the hand authentication vector; the hand authentication vector is input into the identity recognizer, and the hand authentication vector is matched with the reference vector to output the user identity authentication result;

[0063] The update and iteration module is used to dynamically update the baseline vector in the hand feature registration information database using the hand authentication vector to obtain an online sample set. The authentication neural network is trained and updated using the online sample set at set intervals.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] This invention utilizes active acoustic sensing technology to identify and acquire hand reflection sensing signals, effectively avoiding fingerprint residue, dirt, and hygiene and safety hazards associated with contact sensors, significantly improving the convenience and security of user interaction; by performing audio signal preprocessing and target area interception on the hand reflection sensing signals, environmental noise interference is effectively suppressed, the effective time-frequency region of hand reflection features is accurately located, and the signal-to-noise ratio and quality stability of the signal are significantly improved.

[0066] This invention uses a probabilistic adaptive frame-level data augmentation strategy to expand the hand biometric feature matrix to obtain hand registration features. It extracts the mask using the Otsu algorithm and calculates the effective component ratio of each frame as the augmentation probability. It also performs targeted simulations of tapping speed changes on the time axis and gesture changes on the delay axis. This allows for the generation of high-quality variant augmentation data with a very small number of initial registration samples, greatly saving users' registration costs.

[0067] This invention uses a hand biometric matrix as the hand authentication feature input to a pre-trained authentication neural network. A deep feature extractor automatically learns a high-level abstract representation of the hand's acoustic features. The hand authentication vector output by the feature extractor contains rich biometric information. This vector is then accurately matched with a baseline vector by an identity recognizer, achieving end-to-end identity authentication and significantly improving recognition accuracy and authentication efficiency. This process fully utilizes the nonlinear modeling capabilities of deep neural networks to construct a compact and highly discriminative feature distribution in a high-dimensional space, effectively reducing intra-class differences and inter-class confusion. Thus, while ensuring high recognition accuracy, it significantly enhances the system's defense against forgery attacks and its adaptability to complex environments.

[0068] This invention utilizes hand authentication vectors to dynamically update the baseline vectors in the hand feature registration information database to obtain an online sample set. The authentication neural network is trained and updated using the online sample set at set intervals. This not only effectively overcomes the problem of accuracy decline caused by data drift under long-term use, but also quickly adapts to the addition of new users, ensuring the long-term effectiveness of the authentication model. Attached Figure Description

[0069] Figure 1 This is an overall flowchart of the identity authentication method based on ultrasonic sensing of hand features provided in Embodiment 1 of the present invention;

[0070] Figure 2 This is a structural diagram of audio data preprocessing provided in Embodiment 1 of the present invention;

[0071] Figure 3 This is a flowchart of user identity authentication provided in Embodiment 1 of the present invention;

[0072] Figure 4 This is a flowchart of frame-level data enhancement provided in Embodiment 1 of the present invention;

[0073] Figure 5 This is a flowchart of the model continuous evolution steps provided in Embodiment 1 of the present invention;

[0074] Figure 6 This is a flowchart of the offline training process of the authentication neural network provided in Embodiment 1 of the present invention. Detailed Implementation

[0075] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0076] Example 1

[0077] like Figure 1 As shown, this embodiment provides an identity authentication method based on ultrasonic sensing of hand features, including:

[0078] Pre-training the authentication neural network specifically includes:

[0079] The authentication neural network includes a feature extractor, a feature projector, a password discriminator, and an identity recognizer. The feature extractor contains three convolutional blocks, each of which sequentially includes a convolutional layer, a batch normalization layer, a spatial dropout layer, a max pooling layer, and a ReLU activation function.

[0080] An offline hand feature matrix is ​​obtained from an ultrasonic sensing hand feature database; an offline sample set is obtained by expanding the offline hand feature matrix based on a probability-adaptive frame-level data augmentation strategy, wherein the offline sample set is configured with real user identity labels and real password labels;

[0081] like Figure 6 As shown, the authentication neural network is trained using an offline sample set, and the hand recognition vector is output by the feature extractor. The hand recognition vector is then input into the identity recognizer to obtain the user identity recognition result.

[0082] Anchor samples are determined from the offline sample set. The feature projector measures the similarity between the anchor samples and offline samples of the same class, as well as the difference between offline samples of different classes. After temperature coefficient scaling and normalization, the output supervised contrastive loss is expressed as follows:

[0083]

[0084] In the formula, To monitor and compare losses; , and For sample index, For anchor point samples, This is the set of offline samples from the previous batch. For offline samples of the same type as anchor samples, For the positive sample set belonging to the same user as the anchor sample, For offline samples other than anchor point samples, This refers to the set of offline samples excluding anchor samples. For temperature parameters;

[0085] Based on the user identity recognition results and the user's real identity label, the identity recognition loss is calculated using the cross-entropy function;

[0086] The hand recognition vector is input into the password recognizer to obtain the password recognition probability. The password recognizer includes a first fully connected layer, a one-dimensional batch normalization layer, a ReLU activation function, a second fully connected layer, and a log-Softmax layer. Based on the password recognition probability and the true password label, the password authentication loss is calculated using the binary cross-entropy function.

[0087] The joint loss is obtained by weighted summation of the supervision comparison loss, identity recognition loss, and password authentication loss.

[0088] Using the joint loss as a supervision signal, the parameter gradient of the certification neural network is calculated through backpropagation, and the parameters are updated along the gradient descent direction using an adaptive momentum optimizer. The process of training the certification neural network using an offline sample set is repeated iteratively until the joint loss converges and outputs the trained certification neural network.

[0089] This embodiment uses supervised contrastive learning to narrow the feature distance between samples from the same user and widen the feature distance between samples from different users. At the same time, it uses a password discriminator with a gradient inversion layer to perform domain adversarial learning, forcing the feature extractor to remove behavioral differences caused by password input actions. This results in the extraction of highly consistent high-dimensional feature representations that are independent of specific passwords, significantly improving the robustness of the authentication system in complex interaction scenarios.

[0090] The active acoustic sensing technology is used to identify and acquire hand reflection sensing signals, specifically including:

[0091] An ultrasonic wave signal, specifically a frequency-modulated continuous wave (FMCW) signal, is transmitted to the user's hand via an audio player on a mobile device. An audio receiver receives the hand reflection signal generated during password input. The formula is as follows:

[0092]

[0093] In the formula, The hand reflection perception signal is the result of superimposed multipath signals. This represents the number of signal propagation paths in a complex space environment. The ultrasonic wave signal sent to the audio player. For audio receiver to receive the first The time it takes for a signal to propagate along a path. For the first The propagation delay of a signal along a path. For the first The attenuation ratio of the signal propagating along the path.

[0094] In this embodiment, the frequency range of the frequency-modulated continuous wave signal is set to 18-22kHz, the duration is set to 0.02 seconds, and the sampling frequency of the audio receiver is set to 48kHz. The use of a high-frequency FMCW signal of 18-22kHz makes this frequency band almost inaudible to the human ear, thus achieving covert silence verification.

[0095] like Figure 2 As shown, the hand reflex sensing signal undergoes audio signal preprocessing and target region extraction to extract the hand biometric feature matrix; specifically including:

[0096] The acquired hand reflection sensing signal is subjected to high-pass filtering, and the cutoff frequency of the high-pass filter is set to 16kHz, thereby effectively removing low-frequency environmental noise interference.

[0097] The ultrasonic emission signal is used to synchronize the filtered hand reflection sensing signal in frames to obtain the calibration echo signal.

[0098] The channel impulse response is obtained by calculating the cross-correlation between the calibration echo signal frame and the ultrasonic transmission signal frame frame by frame. A global time series matrix is ​​constructed from the channel impulse responses of each frame, specifically including:

[0099]

[0100]

[0101] In the formula, For the first Frame channel impulse response, This represents the multipath delay, where k is the index of the sampling point within the calibration echo signal frame and the ultrasonic transmission signal frame, and K is the number of sampling points. For the first The calibration echo signal of the frame, It is an ultrasonic wave signal; For the global time series matrix, The length of the global time series;

[0102] The effective hand feature matrix is ​​obtained by truncating the global time series matrix based on a set perception distance range. In practice, only the first 80 delay units of the effective hand feature matrix are retained as the target region. This region has an effective coverage of approximately 30cm, which is sufficient to cover the physical space where the hand may appear during password input.

[0103] The difference feature matrix is ​​obtained by performing a difference operation between adjacent frames on the effective feature matrix of the hand, thereby eliminating interference from static backgrounds such as human figures in the environment; the formula for expressing the difference feature matrix is:

[0104]

[0105] In the formula, It is the difference characteristic matrix;

[0106] A soft threshold is set based on the average value of the difference feature matrix. Local maxima in the difference feature matrix are located and filtered based on the soft threshold to obtain peak positions. The difference features of the peak positions in the difference feature matrix are replaced with the soft threshold to obtain sparse response features. In this embodiment, the soft threshold is set to 5 times the average value. If the difference features in the difference feature matrix are greater than the soft threshold, they are determined to be peak positions.

[0107] Mean filtering is used to smooth sparse response features, improving their continuity and stability. In this embodiment, a mean filter with a kernel size of 3×3 is used for smoothing.

[0108] The smoothed sparse response features are normalized using the Min-Max Normalization method to eliminate the negative impact of individual diversity and ensure that the dimensions of the final hand biometric moments are uniform.

[0109] In response to the user registration phase, a probability-adaptive frame-level data augmentation strategy is used to expand the hand biometric matrix to obtain hand registration features, specifically including:

[0110] like Figure 4 As shown, a frame-level data augmentation strategy based on probabilistic adaptive methods expands the hand biometric feature matrix to obtain hand registration features, specifically including:

[0111] The Otsu algorithm (or simply Otsu algorithm) is used to binarize the hand biometric feature matrix to obtain the binary mask matrix for each frame; the enhancement probability of the binary mask matrix for each frame is calculated, expressed by the following formula:

[0112]

[0113] In the formula, Let be the enhancement probability for the j-th frame. Let J be the binary mask matrix of the j-th frame. Let the first dimension of the binary mask matrix of the j-th frame be the size; Let the second dimension of the binary mask matrix of the j-th frame be the size of the second dimension. The index of each element in the binary mask matrix of each frame;

[0114] Affine shearing transformation is applied to the hand biometric feature matrix along the frame axis to obtain hand scaling features, and the random value range of the scaling coefficient is set to [0.8, 1.2]. Using the enhancement probability as the trigger probability, random frame interpolation or discarding operation is performed on the hand scaling features to obtain the first hand extended features.

[0115] Based on enhanced probability The amplitude attenuation factor is calculated using the following formula:

[0116]

[0117] In the formula, The amplitude attenuation factor for the j-th frame; The attenuation coefficient is randomly generated, and its value ranges from [0.6, 1.2].

[0118] The second extended hand feature is obtained by attenuating the hand biometric matrix based on the amplitude attenuation factor; the third extended hand feature is obtained by randomly generating a delay transform factor with a value range of [0.7, 1.5] and performing a delay axis affine shearing transform on the hand biometric matrix based on the delay transform factor.

[0119] The hand registration features are composed of a hand biometric matrix, first hand extended features, second hand extended features, and third hand extended features.

[0120] All hand registration features are input into a pre-trained authentication neural network. The feature extractor outputs a baseline vector, which is then added with the user's real identity label and stored in the hand feature registration information database.

[0121] In this embodiment, the mask is extracted using the Otsu algorithm, and the effective component ratio of each frame is calculated as the enhancement probability. Targeted simulations of tapping speed changes on the time axis and gesture changes on the delay axis are performed. This enables the generation of high-quality variant amplification data with a very small number of initial registration samples, greatly saving users' registration costs.

[0122] In response to the user authentication phase, such as Figure 3 As shown, the hand biometric matrix is ​​input as the hand authentication feature into a pre-trained authentication neural network, and the feature extractor outputs a hand authentication vector; the hand authentication vector is then input into the identity recognizer, which matches the hand authentication vector with a reference vector to output the user identity authentication result, specifically including:

[0123] Clustering is performed on the baseline vectors in the hand feature registration information database based on the user's real identity tag to determine the first user class center. The formula is as follows:

[0124]

[0125]

[0126] In the formula, For real tags to identify users The One reference vector; For real tags to identify users The set of reference vectors; Calculate the function for cluster centers; For the set of reference vectors Cluster centers;

[0127] Calculate the feature space distance between the hand authentication vector and the center of the first user class; compare the minimum feature space distance with a preset security threshold. If the minimum feature space distance is less than or equal to the preset security threshold, determine that the current user is a legitimate user and output the corresponding identity; otherwise, determine that the user is an illegitimate user and refuse authentication.

[0128] In this embodiment, the hand biometric matrix is ​​used as the hand authentication feature input to a pre-trained authentication neural network. A deep feature extractor automatically learns a high-level abstract representation of the hand acoustic features. The hand authentication vector output by the feature extractor contains rich biometric information. After accurate matching between the identity recognizer and the reference vector, end-to-end identity authentication is achieved, which greatly improves the recognition accuracy and authentication efficiency. This process makes full use of the nonlinear modeling capability of deep neural networks to construct a compact and highly discriminative feature distribution in a high-dimensional space, effectively reducing intra-class differences and inter-class confusion. Thus, while ensuring high recognition accuracy, it significantly enhances the system's defense against forgery attacks and its adaptability in complex environments.

[0129] In response to successful user authentication, such as Figure 5 As shown, the online sample set is obtained by dynamically updating the baseline vector in the hand feature registration information database using the hand authentication vector, specifically including:

[0130] The second user class center is obtained by superimposing the hand authentication vector with the set of baseline vectors with the same user identity and recalculating the cluster center.

[0131]

[0132] In the formula, For hand authentication vectors, For real tags to identify users The set of reference vectors; Calculate the function for cluster centers; For the second user class center;

[0133] Obtain the maximum intra-class distance and minimum inter-class distance of the baseline vector in the hand feature registration information database; calculate the current distance from the hand authentication vector to the center of the second user class; if the current distance is less than the minimum inter-class distance and the maximum intra-class distance, replace the baseline vector with the largest intra-class distance using the hand authentication vector, and use the updated baseline vector set as the online sample set;

[0134] At set intervals, the authentication neural network is incrementally retrained using an online sample set. This allows the authentication neural network to continuously adapt to the temporal drift of biometric features and quickly adapt to new users without needing to be trained from scratch.

[0135] Example 2

[0136] This embodiment provides an identity authentication system based on ultrasonic sensing of hand features. The gesture recognition system is used to execute the identity authentication method described in Embodiment 1. The gesture recognition system includes:

[0137] The feature acquisition module uses active acoustic sensing technology to identify and acquire hand reflection sensing signals; it performs audio signal preprocessing and target region extraction on the hand reflection sensing signals to extract the hand biometric feature matrix;

[0138] The user registration module is used to expand the hand biometric feature matrix based on a probability-adaptive frame-level data augmentation strategy to obtain hand registration features. All hand registration features are input into a pre-trained authentication neural network, and the feature extractor outputs a baseline vector. The baseline vector is then added with the real user identity label and stored in the hand feature registration information database.

[0139] The user authentication module is used to input the hand biometric feature matrix as the hand authentication feature into a pre-trained authentication neural network, and the feature extractor outputs the hand authentication vector; the hand authentication vector is input into the identity recognizer, and the hand authentication vector is matched with the reference vector to output the user identity authentication result;

[0140] The update and iteration module is used to dynamically update the baseline vector in the hand feature registration information database using the hand authentication vector to obtain an online sample set. The authentication neural network is trained and updated using the online sample set at set intervals.

[0141] Example 3

[0142] This embodiment provides an electronic terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the identity authentication method described in Embodiment 1.

[0143] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0144] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0145] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0147] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An identity authentication method based on ultrasonic sensing of hand features, characterized in that, include: Active acoustic sensing technology is used to identify and acquire hand reflection sensing signals; the hand reflection sensing signals are preprocessed with audio signals and the target region is extracted to extract the hand biometric feature matrix; In response to the user registration phase, a probability-adaptive frame-level data augmentation strategy is used to expand the hand biometric feature matrix to obtain hand registration features. All hand registration features are then input into a pre-trained authentication neural network. The authentication neural network includes a feature extractor and an identity recognizer. The feature extractor outputs a reference vector, which is then added with the user's real identity label and stored in the hand feature registration information database. In response to the user authentication phase, the hand biometric matrix is ​​input as the hand authentication feature into a pre-trained authentication neural network, and the feature extractor outputs the hand authentication vector; the hand authentication vector is input into the identity recognizer, and the hand authentication vector is matched with the reference vector to output the user identity authentication result. In response to successful user authentication, the system dynamically updates the baseline vector in the hand feature registration information database using the hand authentication vector to obtain an online sample set. The authentication neural network is then trained and updated using the online sample set at set intervals.

2. The identity authentication method based on ultrasonic sensing of hand features according to claim 1, characterized in that, The active acoustic sensing technology is used to identify and acquire hand reflection sensing signals, specifically including: An ultrasonic wave signal, which is an frequency-modulated continuous wave signal, is sent to the user's hand via an audio player on a mobile device. An audio receiver receives the hand reflection signal generated by the user during password input. The formula is as follows: ； In the formula, The hand reflection perception signal is the result of superimposed multipath signals. This represents the number of signal propagation paths in a complex space environment. The ultrasonic wave signal sent to the audio player. For audio receiver to receive the first The time it takes for a signal to propagate along a path. For the first The propagation delay of a signal along a path. For the first The attenuation ratio of the signal propagating along the path.

3. The identity authentication method based on ultrasonic sensing of hand features according to claim 2, characterized in that, The hand reflex sensing signal is preprocessed with audio signals and the target region is extracted to obtain the hand biometric feature matrix; specifically including: The acquired hand reflection sensing signal is subjected to high-pass filtering to remove low-frequency environmental noise interference, and the filtered hand reflection sensing signal is synchronized in frames using ultrasonic transmission signal to obtain calibration echo signal. The channel impulse response is obtained by calculating the cross-correlation between the calibration echo signal frame and the ultrasonic transmission signal frame frame by frame, and a global time series matrix is ​​constructed from the channel impulse response of each frame; the effective feature matrix of the hand is obtained by truncating the global time series matrix based on the set sensing distance range. The effective feature matrix of the hand is subjected to adjacent frame difference operation to obtain the difference feature matrix; a soft threshold is set according to the average value of the difference feature matrix; local maxima points in the difference feature matrix are located and filtered according to the soft threshold to obtain the peak position; the difference features of the peak position in the difference feature matrix are replaced with the soft threshold to obtain sparse response features; After smoothing the sparse response features, the scale normalization operation is performed using the max-min normalization method to obtain the hand biometric matrix.

4. The identity authentication method based on ultrasonic sensing of hand features according to claim 3, characterized in that, The channel impulse response is obtained by calculating the cross-correlation between the calibration echo signal frame and the ultrasonic transmission signal frame frame by frame. A global time series matrix is ​​constructed from the channel impulse responses of each frame, specifically including: ； ； In the formula, For the first Frame channel impulse response, This represents the multipath delay, where k is the index of the sampling point within the calibration echo signal frame and the ultrasonic transmission signal frame, and K is the number of sampling points. For the first The calibration echo signal of the frame, It is an ultrasonic wave signal; For the global time series matrix, This represents the length of the global time series.

5. The identity authentication method based on ultrasonic sensing of hand features according to claim 1, characterized in that, A probabilistic adaptive frame-level data augmentation strategy is used to expand the hand biometric feature matrix to obtain hand registration features, specifically including: The Otsu algorithm is used to binarize the hand biometric feature matrix to obtain the binary mask matrix for each frame; the enhancement probability of the binary mask matrix for each frame is calculated, expressed by the following formula: ； In the formula, Let be the enhancement probability for the j-th frame. Let J be the binary mask matrix of the j-th frame. Let the first dimension of the binary mask matrix of the j-th frame be the size; Let the second dimension of the binary mask matrix of the j-th frame be the size of the second dimension. The index of each element in the binary mask matrix of each frame; Affine shearing transformation is applied to the hand biometric feature matrix along the frame axis to obtain hand scaling features, and the random value range of the scaling coefficient is set to [0.8, 1.2]. Using the enhancement probability as the trigger probability, random frame interpolation or discarding operation is performed on the hand scaling features to obtain the first hand extended features. Based on enhanced probability The amplitude attenuation factor is calculated using the following formula: ； In the formula, The amplitude attenuation factor for the j-th frame; The attenuation coefficient is randomly generated, and its value ranges from [0.6, 1.2]. The second extended hand feature is obtained by attenuating the hand biometric matrix based on the amplitude attenuation factor; the third extended hand feature is obtained by randomly generating a delay transform factor with a value range of [0.7, 1.5] and performing a delay axis affine shearing transform on the hand biometric matrix based on the delay transform factor. The hand registration features are composed of the hand biometric matrix, the first hand extended features, the second hand extended features, and the third hand extended features.

6. The identity authentication method based on ultrasonic sensing of hand features according to claim 1, characterized in that, The feature extractor contains three convolutional blocks, each of which sequentially includes a convolutional layer, a batch normalization layer, a spatially randomized deactivation layer, a max pooling layer, and a ReLU activation function.

7. The identity authentication method based on ultrasonic sensing of hand features according to claim 1, characterized in that, The user authentication result is output by matching the hand authentication vector with the baseline vector, specifically including: Clustering is performed on the baseline vectors in the hand feature registration information database based on the user's real identity tag to determine the first user class center. The formula is as follows: ； ； In the formula, The real label is the user's identity. The One reference vector; The real label is the user's identity. The set of reference vectors; Calculate the function for cluster centers; For the set of reference vectors Cluster centers; Calculate the feature space distance between the hand authentication vector and the center of the first user class; compare the minimum feature space distance with a preset security threshold. If the minimum feature space distance is less than or equal to the preset security threshold, determine that the current user is a legitimate user and output the corresponding identity; otherwise, determine that the user is an illegitimate user and refuse authentication.

8. The identity authentication method based on ultrasonic sensing of hand features according to claim 1, characterized in that, An online sample set is obtained by dynamically updating the baseline vector in the hand feature registration information database using the hand authentication vector, specifically including: The second user class center is obtained by superimposing the hand authentication vector with the set of baseline vectors with the same user identity and recalculating the cluster center. ； In the formula, For hand authentication vectors, The real label is the user's identity. The set of reference vectors; Calculate the function for cluster centers; For the second user class center; Obtain the maximum intra-class distance and minimum inter-class distance of the baseline vector in the hand feature registration information database; calculate the current distance from the hand authentication vector to the center of the second user class; if the current distance is less than the minimum inter-class distance and the maximum intra-class distance, replace the baseline vector with the largest intra-class distance using the hand authentication vector, and use the updated baseline vector set as the online sample set.

9. The identity authentication method based on ultrasonic sensing of hand features according to claim 1, characterized in that, The training process of the authentication neural network specifically includes: An offline hand feature matrix is ​​obtained from an ultrasonic sensing hand feature database; an offline sample set is obtained by expanding the offline hand feature matrix based on a probability-adaptive frame-level data augmentation strategy, wherein the offline sample set is configured with real user identity labels and real password labels; The authentication neural network is trained using an offline sample set. The feature extractor outputs a hand recognition vector, which is then input into the identity recognizer to obtain the user's identity recognition result. Anchor samples are determined from the offline sample set. The similarity between the anchor samples and offline samples of the same class, as well as the difference between offline samples of different classes, are measured. After temperature coefficient scaling and normalization, the supervised contrast loss is output, expressed by the following formula: ； In the formula, To monitor and compare losses; , and For sample index, This is the set of offline samples from the previous batch. For anchor point samples, For offline samples of the same type as anchor samples, For the positive sample set belonging to the same user as the anchor sample, For offline samples other than anchor point samples, This refers to the set of offline samples excluding anchor samples. For temperature parameters; Based on the user identity recognition results and the user's real identity label, the identity recognition loss is calculated using the cross-entropy function; The hand recognition vector is input into the password recognizer to obtain the password recognition probability. The password recognizer includes a first fully connected layer, a one-dimensional batch normalization layer, a ReLU activation function, a second fully connected layer, and a log-Softmax layer. Based on the password recognition probability and the true password label, the password authentication loss is calculated using the binary cross-entropy function. The joint loss is obtained by weighted summation of the supervision comparison loss, identity recognition loss, and password authentication loss. Using the joint loss as a supervision signal, the parameter gradient of the certification neural network is calculated through backpropagation, and the parameters are updated along the gradient descent direction using an adaptive momentum optimizer. The process of training the certification neural network using an offline sample set is repeated iteratively until the joint loss converges and outputs the trained certification neural network.

10. An identity authentication system based on ultrasonic sensing of hand features, characterized in that, include: The feature acquisition module uses active acoustic sensing technology to identify and acquire hand reflection sensing signals; it performs audio signal preprocessing and target region extraction on the hand reflection sensing signals to extract the hand biometric feature matrix; The user registration module is used to expand the hand biometric feature matrix based on a probability-adaptive frame-level data augmentation strategy to obtain hand registration features. All hand registration features are input into a pre-trained authentication neural network, which includes a feature extractor and an identity recognizer. The feature extractor outputs a reference vector, and the reference vector is added with the real user identity label and then stored in the hand feature registration information database. The user authentication module is used to input the hand biometric matrix as the hand authentication feature into a pre-trained authentication neural network, and the feature extractor outputs the hand authentication vector; the hand authentication vector is input into the identity recognizer, and the hand authentication vector is matched with the reference vector to output the user identity authentication result; The update and iteration module is used to dynamically update the baseline vector in the hand feature registration information database using the hand authentication vector to obtain an online sample set. The authentication neural network is trained and updated using the online sample set at set intervals.

Images