A user authentication method and system based on dynamic deformation of ear canal

Abstract

The application provides a user authentication method and system based on ear canal dynamic deformation, relates to the technical field of biometric identification and mobile terminals, and comprises the following steps: in the case that a user wears and uses an in-ear earphone, collecting an original pressure sequence generated during the interaction between the user and the earphone in real time; performing voice activity detection, empirical mode decomposition and normalization processing on the original pressure sequence to obtain a standardized sequence; inputting the standardized sequence into a pre-trained user authentication model to obtain an authentication result output by the pre-trained user authentication model; the pre-trained user authentication model is composed of an Inception feature extraction module, a cross-attention mechanism and a global pooling layer; when the authentication result is 1, it is confirmed that the user authentication is successful; when the authentication result is 0, it is confirmed that the user authentication fails; through the above technical scheme, the convenience and stability of earphone-type devices in the individual authentication process can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biometric identification and mobile terminals, and in particular to a user authentication method and system based on dynamic deformation of an ear canal. BACKGROUND

[0002] With the wide application of smart wearable devices, especially the continuous enhancement of terminal functions of wireless earphones, smart earbuds and the like, such devices are increasingly playing a key role in sensitive scenarios such as voice interaction, contactless payment, smart home control and the like. In order to ensure safe use, how to implement a user authentication mechanism in earbud-type devices that takes into account convenience, security and low power consumption has become an important research direction in the current technical field. At present, user authentication schemes in earphone-type devices can be mainly divided into the following three categories.

[0003] The first category is an authentication method based on voiceprint. This type of method mainly collects user voice signals through a microphone and extracts individual voice features for identification. Since it does not require additional sensors and is easy to integrate, it has become the most widely used authentication method. However, voiceprint features are easily disturbed by background noise, device differences and other factors, and there is a risk of being attacked by voice imitation, voice synthesis (such as TTS, voice conversion, GAN generation) and the like, making it difficult to ensure identification accuracy and security in an open environment.

[0004] The second category is an authentication method based on interactive behavior. This type of method generates unique signals by having the user actively complete specific actions (such as tooth clenching, cheek sliding, facial muscle movement) for individual identification. This method can achieve high recognition rates in controlled environments, but it generally relies on user cooperation, is complex to interact with, is prone to operation failure, and has poor user experience, limiting its practicality in daily use scenarios.

[0005] The third category is an authentication method based on physiological structure or physiological signals. This type of scheme uses user's inherent biological structure (such as ear canal shape) or passive physiological signals (such as heartbeat, respiration) as features, which has strong privacy and individual uniqueness. Common schemes include obtaining ear canal geometric features for identification based on ultrasonic reflection, extracting photoplethysmogram based on ear canal blood flow variation signals, and collecting bone conduction signals or heartbeat sound and the like weak physiological sound waves for authentication. Although this type of method has certain recognition accuracy, it generally has many technical limitations. First, most of them rely on active acoustic excitation signals (such as ultrasound), which conflict with normal audio playback functions, resulting in increased power consumption. Second, the hardware design is complex and requires high earbud structure. Finally, the response speed is slow and the stability is poor, making it difficult to adapt to the rapid authentication needs in natural interaction scenarios.

[0006] Some studies have attempted to use the changes in mandibular movement caused during speech to indirectly capture the dynamic changes of the ear canal to achieve authentication. However, most of such solutions rely on IMU (inertial measurement unit), accelerometer or bone conduction sensor for data acquisition, and the signal is easily disturbed by head movement and sensitive to wearing posture, with low stability. SUMMARY

[0007] The embodiments of the present application provide a user authentication method and system based on dynamic deformation of an ear canal, which solves the technical problems of poor robustness, high user cooperation degree, high hardware integration, and difficulty in resisting imitation attacks in the individual authentication process in the prior art, and improves the convenience and stability of individual authentication.

[0008] To achieve the above-mentioned purpose, the technical solutions of the embodiments of the present application are as follows.

[0009] In a first aspect, the embodiments of the present application provide a user authentication method based on dynamic deformation of an ear canal, which comprises:

[0010] In the case that a user wears and uses an in-ear earphone, real-time collection of an original pressure sequence generated during the interaction between the user and the in-ear earphone;

[0011] Voice activity detection, empirical mode decomposition and normalization processing of the original pressure sequence to obtain a standardized sequence;

[0012] Input of the standardized sequence into a pre-trained user authentication model to obtain an authentication result output by the pre-trained user authentication model; the authentication result is 1 or 0; the pre-trained user authentication model is composed of an Inception feature extraction module, a cross-attention mechanism and a global pooling layer, includes a 768-dimensional feature vector, and outputs the authentication result through a full connection layer and a Sigmoid function; the pre-trained user authentication model is deployed to a terminal connected with the in-ear earphone after being processed by lightening;

[0013] When the authentication result is 1, it is confirmed that the user authentication is successful; when the authentication result is 0, it is confirmed that the user authentication fails.

[0014] In a possible implementation, the real-time collection of the ear pressure signal generated during the interaction between the user and the in-ear earphone comprises:

[0015] Real-time collection of the pressure signals of the left and right ear canals during the interaction between the user and the in-ear earphone by the built-in pressure sensor of the in-ear earphone, and generation of the original pressure sequence with a time stamp.

[0016] In a possible implementation, the voice activity detection, empirical mode decomposition and normalization processing on the original pressure sequence to obtain the standardized sequence comprises:

[0017] The voice activity detection is performed on the original pressure sequence to intercept the pressure sequence corresponding to the existing interaction;

[0018] The empirical mode decomposition is performed on the pressure sequence to eliminate the baseline drift introduced by the height or temperature change in the pressure sequence;

[0019] The Min-Max normalization processing is performed on the pressure sequence after the empirical mode decomposition to eliminate the amplitude range difference in different interaction contents, to obtain the standardized sequence.

[0020] In a possible implementation, the voice activity detection, empirical mode decomposition and normalization processing on the original pressure sequence to obtain the standardized sequence comprises:

[0021] The voice segment in the original pressure sequence is extracted by the WebRTC algorithm, and the pressure sequence is synchronously intercepted.

[0022] In a possible implementation, the empirical mode decomposition is performed on the pressure sequence to eliminate the baseline drift introduced by the height or temperature change in the pressure sequence, comprising:

[0023] The original pressure sequence is decomposed into a plurality of intrinsic mode functions (IMF) and a low-frequency residual component, and the IMF is retained to reconstruct a drift-free signal in [2-5] to eliminate low-frequency interference.

[0024] In a possible implementation, the Min-Max normalization processing is performed on the pressure sequence after the empirical mode decomposition to eliminate the amplitude range difference in different interaction contents, to obtain the standardized sequence, comprising:

[0025] The Min-Max normalization processing is performed by the following formula to uniformly scale the pressure sequence to the interval [-1, 1] to obtain the standardized sequence:

[0026] ; wherein, Pnorm(t) represents the pressure value of the normalized sequence at time t, P(t) represents the value of the original pressure sequence at time t, Pmin represents the minimum value of the original pressure sequence in the entire time sequence, used to determine the lower limit of the pressure sequence; Pmax represents the maximum value of the original pressure sequence in the entire time sequence, used to determine the upper limit of the pressure sequence. ​​​​

[0027] In a possible implementation, the Inception feature extraction module comprises at least a one-dimensional convolution sub-module and a two-dimensional convolution sub-module; the one-dimensional convolution sub-module and the two-dimensional convolution sub-module, and the global pooling layer are configured to extract time domain features and frequency domain features of the standardized sequence;

[0028] The cross-attention mechanism is configured to fuse the time domain features and the frequency domain features through a bidirectional attention mechanism to calculate attention weights, and update the fused time domain features and the fused frequency domain features through residual connection to output a 768-dimensional feature vector;

[0029] The fully connected layer and the Sigmoid function are configured to output an authentication result.

[0030] In a possible implementation, the user authentication model is pre-trained in the following manner:

[0031] Sample data is taken as input of a preset network model, and the preset network model is trained;

[0032] In a case where a difference between a sample authentication result output by the preset network model and an actual authentication result satisfies a preset evaluation condition, it is determined that the preset network model is trained, and the preset network model is taken as the pre-trained user authentication model;

[0033] The sample data is a stress signal obtained by performing time axis alignment, rhythm transformation, and stress signal synchronization on an audio signal; a difference between the sample authentication result and the actual authentication result is evaluated through a cross-entropy loss function and a supervised contrast loss function; a data enhancement strategy is used in the training process to further improve the generalization of the preset network model; the data enhancement strategy is one or more of time stretching, random time offset, and Gaussian noise injection.

[0034] In a possible implementation, the pre-trained user authentication model is processed in the following manner to be lightweight:

[0035] The pre-trained user authentication model is compressed to be below 5 MB through an ONNX tool chain, and FP16 mixed precision quantization is adopted to reduce the calculation load;

[0036] Based on the hardware architecture optimization operator of the terminal, hardware acceleration is performed by using a CoreML / NNAPI framework to realize model lightweight.

[0037] In a second aspect, the embodiments of the present application provide a user authentication system based on dynamic deformation of ear canal, which is used to execute the user authentication method based on dynamic deformation of ear canal in the first aspect, and comprises an in-ear earphone and a terminal.

[0038] The pressure sensor is configured to collect a raw pressure sequence generated during interaction between the user and the in-ear earphone in real time when the user wears and uses the in-ear earphone.

[0039] The terminal is configured to perform voice activity detection, empirical mode decomposition and normalization processing on the raw pressure sequence to obtain a standardized sequence, input the standardized sequence into a pre-trained user authentication model to obtain an authentication result output by the pre-trained user authentication model, wherein the authentication result is 1 or 0, the pre-trained user authentication model comprises an Inception feature extraction module, a cross-attention mechanism and a global pooling layer, includes a 768-dimensional feature vector, and outputs the authentication result through a full connection layer and a Sigmoid function, and the pre-trained user authentication model is deployed to the terminal connected with the in-ear earphone after being processed in a light manner, the user authentication is confirmed to be successful when the authentication result is 1, and the user authentication is confirmed to be failed when the authentication result is 0.

[0040] The one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0041] The present application uses the speech perception pressure sequence generated by the dynamic deformation of the ear canal as the biological feature, has strong intra-individual consistency and difference, and eliminates baseline drift through empirical mode decomposition, and is robust in complex moving scenes such as walking and climbing stairs; the time-frequency domain features are fused by combining convolution and cross-attention mechanism, the speaking rhythm change problem is effectively solved, and the model generalization ability is improved; a hybrid loss function is designed and hard negative sample mining is introduced, the feature space distinguishability is enhanced, and the performance is stable to speech imitation attack; the model is deployed to the terminal after being processed in a light manner, the real-time authentication is completed by natural speaking, and the convenience and security are considered. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application or the prior art. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the related art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0043] Figure 1A flowchart of a user authentication method based on dynamic deformation of an ear canal is provided for an embodiment of the present application.

[0044] Figure 2 A block diagram of a user authentication system based on dynamic deformation of an ear canal is provided for an embodiment of the present application.

[0045] Figure 3 A block diagram of another user authentication system based on dynamic deformation of an ear canal is provided for an embodiment of the present application.

[0046] Figure 4 A schematic diagram of a pressure sequence SPS is provided for an embodiment of the present application.

[0047] Figure 5 A schematic diagram of an IMF decomposition sequence is provided for an embodiment of the present application.

[0048] Figure 6 A schematic diagram of an original signal and a reconstructed signal is provided for an embodiment of the present application.

[0049] Figure 7 A schematic diagram of an end-to-end authentication model architecture is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] In the description of the embodiments of the present application, it should be noted that the terms “center”, “upper”, “lower”, “left”, “right”, “vertical”, “horizontal”, “inner”, “outer” and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the embodiments of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. The terms “first”, “second”, “third” are only for the purpose of description, and cannot be understood as indicating or implying relative importance. In addition, the terms “mounting”, “connecting”, “connecting” should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be direct connection, or indirect connection through an intermediate medium, or the communication between two elements inside. For those skilled in the art, the specific meaning of the above terms in the embodiments of the present application can be understood according to the specific circumstances.

[0052] Firstly, the application scenario of the present application is introduced. The present application can be applied to the non-aware verification scene of in-ear (or earplug) earphones, and is particularly suitable for application scenarios such as mobile payment and smart home control, which have high requirements for security and portability.

[0053] From the prior art, it can be seen that existing researches mainly focus on the static geometric differences or sound wave conduction characteristics of ear canals, and the potential of real-time dynamic deformation of ear canals caused by the movement of articulatory organs such as jaws and tongues during speech and the corresponding inner ear pressure change signals has not been fully tapped. Such signals have natural individual differences and stable repeatability, and are a kind of dynamic biometric features that have not been fully utilized.

[0054] Therefore, the present application provides a user authentication method and system based on ear canal dynamic deformation. In the case that a user wears and uses an in-ear earphone, the original pressure sequence generated during the interaction between the user and the in-ear earphone is collected in real time. The original pressure sequence is subjected to speech activity detection, empirical mode decomposition and normalization processing to obtain a standardized sequence. The standardized sequence is input into a pre-trained user authentication model to obtain an authentication result output by the pre-trained user authentication model. The authentication result is 1 or 0. The pre-trained user authentication model is composed of an Inception feature extraction module, a cross-attention mechanism and a global pooling layer, includes a 768-dimensional feature vector, and outputs the authentication result through a fully connected layer and a Sigmoid function. The pre-trained user authentication model is deployed to a terminal connected to the in-ear earphone after being subjected to light-weight processing. When the authentication result is 1, it is confirmed that the user authentication is successful. When the authentication result is 0, it is confirmed that the user authentication fails. Through the above technical solution, the speech perception pressure sequence generated by the dynamic deformation of the ear canal is used as a biometric feature, which has strong intra-individual consistency and difference. The baseline drift is eliminated through empirical mode decomposition, and the robustness is significantly improved in complex mobile scenarios such as walking and climbing stairs. The time-frequency domain features are fused by combining convolution and cross-attention mechanism, effectively solving the problem of speech rhythm change and improving the model generalization ability. A hybrid loss function is designed and hard negative samples are introduced to enhance the feature space discriminability and show stability to speech imitation attacks. The model is deployed to the terminal after being subjected to light-weight processing, realizing real-time authentication that can be completed by natural speaking, and balancing convenience and security.

[0055] Figure 1 A flowchart of a user authentication method based on ear canal dynamic deformation provided by an embodiment of the present application is shown in FIG. 1. Figure 1 As shown in FIG. 1, the method can include the following steps.

[0056] S101, in the case that a user wears and uses an in-ear earphone, the original pressure sequence generated during the interaction between the user and the in-ear earphone is collected in real time.

[0057] For example, the interaction period can refer to a period during which the user speaks to the in-ear earphone.

[0058] S102, voice activity detection, empirical mode decomposition and normalization processing are performed on the original pressure sequence to obtain a standardized sequence.

[0059] S103, the standardized sequence is input into a pre-trained user authentication model to obtain an authentication result output by the pre-trained user authentication model.

[0060] The authentication result is 1 or 0; the pre-trained user authentication model is composed of an Inception feature extraction module, a cross-attention mechanism and a global pooling layer, includes a 768-dimensional feature vector, and outputs the authentication result through a fully connected layer and a Sigmoid function; the pre-trained user authentication model is deployed to a terminal connected to the in-ear earphone after being lightened.

[0061] For example, the terminal can be a mobile terminal, such as a mobile phone, a computer, etc., which is not limited here. The pre-trained user authentication model can be an end-to-end authentication model (BaroIDNet), which can improve the feature discrimination ability through a hybrid loss function and hard negative sample mining. The authentication result can be authentication success or authentication failure, and the authentication result can also be authentication probability, i.e. legal or illegal, etc., which is not limited here.

[0062] S104, when the authentication result is 1, it is confirmed that the user authentication is successful; when the authentication result is 0, it is confirmed that the user authentication fails.

[0063] Through the above technical solution, the speech perception pressure sequence generated by the dynamic deformation of the ear canal is used as a biological feature, which has strong intra-individual consistency and difference, and the baseline drift is eliminated through empirical mode decomposition, which is robust in complex moving scenes such as walking and climbing stairs; the convolution and cross-attention mechanism are combined to fuse time-frequency domain features, effectively solving the problem of speaking rhythm change and improving the model generalization ability; a hybrid loss function is designed and hard negative sample mining is introduced to enhance the feature space distinguishability and show stability to voice imitation attacks; the model is deployed to the terminal after being lightened, realizing real-time authentication that can be completed by natural speaking, and balancing convenience and security.

[0064] In one possible implementation, the ear pressure signal generated during the interaction between the user and the in-ear earphone is collected in real time, including: through the built-in pressure sensor of the in-ear earphone, the pressure signals of the left and right ear canals during the interaction between the user and the in-ear earphone are collected in real time, and the original pressure sequence with timestamp is generated.

[0065] In an example, the pressure sensor can be a dual Bosch BMP390 MEMS air pressure sensor (sampling rate 108 Hz, accuracy ±0.06 hPa), which synchronously collects left and right ear canal pressure signals. and generates a time-stamped original pressure sequence with a time-stamping accuracy of 1 ms, ensuring accurate collection and time alignment of the pressure signal. The pressure sensor is tightly fitted to the ear canal wall through a flexible silicone catheter with an inner diameter of 1.5 mm, ensuring efficient conduction of the pressure signal.

[0066] In a possible implementation, the original pressure sequence is subjected to voice activity detection, empirical mode decomposition, and normalization processing to obtain a standardized sequence, including: voice activity detection on the original pressure sequence to intercept the pressure sequence corresponding to the presence of interaction; empirical mode decomposition of the pressure sequence to eliminate baseline drift introduced by height or temperature changes in the pressure sequence; Min-Max normalization processing of the pressure sequence after empirical mode decomposition to eliminate amplitude range differences in different interaction contents to obtain the standardized sequence.

[0067] In a possible implementation, the original pressure sequence is subjected to voice activity detection to intercept the pressure sequence corresponding to the presence of interaction, including: extracting the voice segment in the original pressure sequence through the WebRTC algorithm, and synchronously intercepting the pressure sequence.

[0068] In a possible implementation, the pressure sequence is subjected to empirical mode decomposition to eliminate baseline drift introduced by height or temperature changes in the pressure sequence, including: decomposing the original pressure sequence into multiple intrinsic mode functions (IMF) and low-frequency residual components, and retaining the IMF to reconstruct a drift-free signal in the range of [2-5] to eliminate low-frequency interference.

[0069] In a possible implementation, the pressure sequence after empirical mode decomposition is subjected to Min-Max normalization processing to eliminate amplitude range differences in different interaction contents to obtain the standardized sequence, including: Min-Max normalization processing through the following formula to uniformly scale the pressure sequence to the interval [-1, 1] to obtain the standardized sequence: ; wherein, represents the pressure value after normalization at time , represents the value of the original pressure sequence at time , represents the minimum value of the original pressure sequence in the entire time sequence, used to determine the lower limit of the pressure sequence; represents the maximum value of the original pressure sequence The maximum value throughout the entire time series is used to determine the upper limit of the stress series.

[0070] For example, the WebRTC speech activity detection algorithm can be used to extract audio segments and simultaneously capture stress signals. ( ); the original signal is decomposed using Empirical Mode Decomposition (EMD). It is decomposed into 7 intrinsic mode functions (IMFs) and low-frequency residual components. ,Right now Preserve the IMF in [2-5] for drift-free signal reconstruction. To eliminate low-frequency interference. Among them, Represents the first corresponding to local oscillations at different frequencies. Each intrinsic modulus function, and This represents the final residual component, which captures global low-frequency trends (such as baseline drift); by removing This allows for the extraction of target signals without baseline drift. Finally, the Min-Max normalization formula is used: The pressure signal is normalized, thereby scaling the pressure signal uniformly to the [-1, 1] interval.

[0071] In one possible implementation, the Inception feature extraction module includes at least a one-dimensional convolutional submodule and a two-dimensional convolutional submodule; the one-dimensional convolutional submodule, the two-dimensional convolutional submodule, and the global pooling layer are used to extract the temporal and frequency domain features of the normalized sequence; the cross-attention mechanism is used to fuse the temporal and frequency domain features through a bidirectional attention mechanism and calculate attention weights; and the fused temporal and frequency domain features are updated through residual connections to output a 768-dimensional feature vector; the fully connected layer and the Sigmoid function are used to output the authentication result.

[0072] For example, this can be achieved through parallel feature construction – temporal processing: normalizing the signals from the left and right ears. After concatenation, the input is a 3-layer Conv1D network (kernel size [10, 5, 3], stride 2, activation function GELU) to extract pronunciation rhythm features. ; Frequency domain processing can be achieved through parallel feature construction: STFT (window length 128, step size 32) is performed on the left and right ear signals to generate a spectrogram. Frequency domain formant features were extracted using a 2-layer Conv2D network. .

[0073] The multi-scale feature extraction can be performed by a one-dimensional convolution submodule and a two-dimensional convolution submodule: three Inception feature extraction modules (convolution kernel dynamically adjusted to [k, k / 2, k / 4]) are used to model the time-frequency domain features in multiple scales. Each module captures global trends and local details through parallel branches (different convolution kernel sizes + pooling), and outputs features , balancing the computational complexity and feature expression ability. The cross-domain dependency of the time-frequency domain features is modeled through a bidirectional attention mechanism, taking the time domain features as the query (Query, ), the frequency domain features as the key (Key, ) and the value (Value, ), and then calculating the attention weight through the following formula: . Wherein, is a learnable projection matrix that maps the features to a 64-dimensional unified space, and the fused time-frequency domain features are updated through a residual connection.

[0074] In one possible implementation, the user authentication model is pre-trained by the following method: taking the sample data as the input of the preset network model, training the preset network model; if the difference between the sample authentication result output by the preset network model and the actual authentication result satisfies the preset evaluation condition, it is determined that the preset network model is trained, and the preset network model is taken as the pre-trained user authentication model; wherein the sample data is a pressure signal obtained by aligning the time axis, transforming the rhythm and synchronizing the pressure signal of the audio signal; the difference between the sample authentication result and the actual authentication result is evaluated by using cross-entropy loss function and supervised contrast loss function; in the training process, a data enhancement strategy is used to further improve the generalization of the preset network model; the data enhancement strategy is one or more of time stretching, random time offset and Gaussian noise injection.

[0075] In an example, the sample data includes original sample data and pseudo sample data. The preset evaluation condition can be set by the user, for example, it can be that the difference is less than 1.5%.

[0076] Since there is a time sequence deviation between the pressure signal and the audio signal (such as the mandibular adjustment action before pronunciation, which will cause the pressure signal to be ahead of or lag behind the audio), time axis alignment needs to be achieved through resampling, and the specific method is as follows: the time point of the pressure signal is mapped and aligned with the time point of the audio signal , and the formula is as follows: , wherein, and are the lengths of the pressure signal and the audio signal respectively. If For non-integer values, calculate the corresponding pressure value using linear interpolation: Finally, a resampled pressure signal of the same length as the audio signal is obtained. That is, the original sample data.

[0077] A phase vocoder is used to time-stretch / compress the audio signal (rhythm shift) to simulate different speaking rhythms (normal, slow, fast) while maintaining the pitch. Specifically, the audio signal... The spectrum is obtained by performing a short-time Fourier transform (STFT). Through parameters (Step size 0.1, 7 possible values, excluding) Scaling the timeline (from the original rhythm) generates a modified spectrum. Six audio samples with different rhythms were obtained. Take values ​​of 0.7, 0.8, 0.9, 1.1, 1.2, and 1.3.

[0078] The audio signal after rhythm change Aligned pressure signal Perform synchronous transformation; specifically, for each ,Will Scale at the same rhythm to generate corresponding pressure signals. Ultimately, synchronous sample pairs are formed. .

[0079] By employing time stretching technology (parameters), each original sample data is used to generate 6 times the amount of pseudo-sample data, expanding the data scale to 7 times (including the original sample data), thereby improving the model's generalization ability and enhancing its robustness to input data under different speaking rhythms.

[0080] For example, the cross-entropy loss function and the supervised contrastive loss function can be combined to form a hybrid loss function, which evaluates the difference between the sample authentication result and the actual authentication result. That is, a hybrid loss function. ( ),in Let cross-entropy be the loss function. The supervised contrastive loss function.

[0081] Cross-entropy loss function The formula used to directly optimize classification decisions is: ,in, Indicates batch size. It is the predicted output. It is the Sigmoid activation function, and This indicates the actual label.

[0082] Supervised contrastive loss For forced same-class feature compactness and different-class feature separation, combined with hard negative sample mining (memory block size , temperature ), the formula is as follows: , wherein, represents the L2 normalized feature vector of the anchor sample, represents the feature vector of the positive sample, represents the sample index set in the current batch, contains the same positive sample as the anchor point , is a temperature parameter for controlling the concentration of feature distribution, and and and represent the hard negative samples in the current batch and the historical negative sample memory library, respectively.

[0083] In the supervised learning loss function, memory library negative sample mining is used. Specifically, the memory library stores stores historical negative sample features, and each iteration selects negative samples from and the current batch, calculates the similarity of sample features and positive and negative sample features, enhances the discrimination ability of BaroIDNet for difficult-to-distinguish samples, and promotes the model to learn more discriminative feature representations until the loss converges, while saving the model parameters to a pt format file.

[0084] Training parameters: Adam optimizer (learning rate 0.0001, batch size 128), train for 50 rounds until the error rate of the validation set is less than 1.5% convergence; In the training process, data augmentation strategies (including time stretching, random time offset, Gaussian noise injection, etc.) are used to further improve the generalization of the model.

[0085] In one possible implementation, the pre-trained user authentication model is processed in a lightweight manner by: compressing the pre-trained user authentication model to below 5MB using the ONNX tool chain, and using FP16 mixed precision quantization to reduce the computational load; optimizing operators based on the hardware architecture of the terminal, and using the CoreML / NNAPI framework for hardware acceleration to achieve model lightweight.

[0086] In an example, the model parameters of the pre-trained user authentication model can be prevented from being tampered with by the terminal native security mechanism (such as CoreML model signature), and the device security module is combined to realize trusted interaction of authentication results, meeting the privacy protection requirements. In addition, the lightweight model can be deployed to the terminal, and the ear pressure signals collected by the earphone in real time are used for authentication.

[0087] Figure 2A block diagram of a user authentication system based on dynamic deformation of ear canal is provided for an embodiment of the present application. The user authentication system based on dynamic deformation of ear canal 200 is used to execute the above-mentioned user authentication method based on dynamic deformation of ear canal. The system 200 includes an in-ear earphone 210 and a terminal 220. The in-ear earphone 210 is internally provided with a pressure sensor 211 and is connected to the terminal 220. The pressure sensor 211 is used to collect the original pressure sequence generated during the interaction between the user and the in-ear earphone in real time when the user wears and uses the in-ear earphone. The terminal 220 is used to perform voice activity detection, empirical mode decomposition and normalization processing on the original pressure sequence to obtain a standardized sequence. The standardized sequence is input into a pre-trained user authentication model to obtain an authentication result output by the pre-trained user authentication model. The authentication result is 1 or 0. The pre-trained user authentication model is composed of an Inception feature extraction module, a cross-attention mechanism and a global pooling layer, includes a 768-dimensional feature vector, and outputs the authentication result through a fully connected layer and a Sigmoid function. The pre-trained user authentication model is deployed to the terminal connected to the in-ear earphone after being subjected to light-weight processing. When the authentication result is 1, it is confirmed that the user authentication is successful. When the authentication result is 0, it is confirmed that the user authentication fails.

[0088] In an example, the in-ear earphone can be matched / connected to the terminal through Bluetooth, which is not limited here.

[0089] Based on the above, the present application proposes a user authentication method and system based on dynamic deformation of ear canal. A micro pressure sensor embedded in the cavity of an in-ear earphone captures the pressure changes in the user's closed ear canal during speaking (i.e., speech-induced pressure sequence SPSs). In the present application, a corresponding data preprocessing method (including voice activity detection, baseline drift suppression and normalization) is provided to preprocess the detected SPSs. Then, through the proposed pseudo sample generation, the SPSs under different speaking rhythms are simulated based on the rhythm deformation conversion according to the synchronous sampling of the speech signal, thereby enhancing the robustness and generalization ability of the model. The present application also designs an end-to-end identity authentication model named BaroIDNet, which directly outputs the legitimacy of the user authentication result (0: illegal / 1: legal) according to the input (SPS and corresponding spectrogram). In addition, the present application also optimizes and deploys the trained BaroIDNet for mobile terminals, thereby realizing efficient and secure identity verification.

[0090] Therefore, the present application includes the following key technical points:

[0091] Key point 1: This application is based on an embedded MEMS microbarometer (Bosch BMP390) to build an ear canal pressure sensing system. It uses the pressure sequence generated by the dynamic deformation of the ear canal caused by natural speech to perform identity authentication. No additional sensors or user interaction is required, and it balances privacy, low power consumption, and user experience. It is suitable for real-time, non-invasive individual identification application scenarios.

[0092] Key point 2: This application proposes a baseline drift elimination method based on empirical mode decomposition (EMD). It adaptively separates the low-frequency drift component in the ear canal pressure sequence, effectively eliminating sensor baseline drift caused by altitude changes, elevator motion, or environmental temperature. This method does not require pre-defined basis functions and has high adaptability, ensuring the stability of the pressure signal and the accuracy of the authentication system in mobile scenarios.

[0093] Key point 3: This application proposes a pseudo-sample generation method based on rhythm transformation. It uses a phase vocoder to stretch and compress the speech signal, generating speech samples of various rhythm styles, and simultaneously constructing corresponding ear canal pressure sequence pseudo-samples to enhance the robustness of the model to speech rhythm changes. This method significantly expands the rhythm diversity of the original data, improving the generalization ability and stability of the identity authentication system.

[0094] Key point 4: This application proposes a deep neural network architecture (BaroIDNet) that integrates multi-scale convolution and cross-domain attention mechanisms. It extracts features in the time domain and frequency domain of the ear canal pressure sequence in parallel and dynamically fuses them using cross-attention mechanisms, effectively capturing global trends and local dynamic features during speech production.

[0095] Key point 5: This application proposes a hybrid loss function design that combines cross-entropy and supervised contrastive learning, combining hard negative sample mining and memory mechanisms to strengthen the clustering of same-class samples and the separation of different-class samples in the feature space.

[0096] To clearly illustrate the technical implementation path of this application, the following describes the specific implementation process of the user authentication method and system based on ear canal dynamic deformation, combined with the pre-training, pseudo-sample enhancement, and fine-tuning optimization ideas of the model, and aided by the accompanying drawings.

[0097] Figure 3 Another block diagram of a user authentication system based on ear canal dynamic deformation provided by the embodiment of the application is shown. It includes all necessary modules and processing procedures. The following details each step involved in the technical solution of the present application:

[0098] Step 1: Pre-training data collection.

[0099] (1) Construction and configuration of data acquisition device: a pair of self-developed in-ear earphones are constructed, each of which is embedded with a micro MEMS pressure sensor (Bosch BMP390) for detecting air pressure changes in the ear canal; Arduino Nano 33 BLESense is used as a controller, which is equipped with a BLE communication module to realize the synchronous acquisition and Bluetooth transmission of voice and SPSs; the microphone collects voice signals with a sampling rate of 44.1 kHz; the pressure sensor collects SPS signals with a sampling rate of 108 Hz; a set of matching Python application (running on PC or smart phone) is developed to realize real-time receiving and storing of signals through serial or Bluetooth connection to the acquisition device.

[0100] (2) Construction of voice command library and standardization of instructions: a standard corpus of 30 common voice commands is constructed, which is divided into three length intervals: short commands (2-3 words): such as "Hey Siri", "Call Nancy", etc.; medium commands (4-5 words): such as "Turnoff Wi-Fi", "Find my way home", etc.; long commands (6-7 words): such as "Can you translate this into Chinese". Each participant selects 10 sentences from the corpus, maintaining natural pronunciation habits.

[0101] Examples, Figure 4 (a) The example shows the original audio signal captured by the microphone and the corresponding voice perception pressure sequence SPS recorded by the pressure sensor when a certain user says "Navigate to my office". Figure 4 (b) The example shows the SPS collected by a certain user A and user B when they say the command "Video my mom", showing obvious individual differences.

[0102] (3) As many volunteers as possible are recruited (with a moderate male-to-female ratio, aged 21-50), and each participant completes a repeated reading task within a cycle (such as 30 days) to collect rich corpus and corresponding SPS samples. All acquisition processes are standardized to avoid abnormal behavior affecting data quality.

[0103] (4) Daily use scenario training sample collection: in each daily collection, each command is read 10 times, including the following rhythm and pronunciation intensity control dimensions: speed: slow (2 times), normal (2 times), fast (2 times); volume: normal volume (4 times), low volume (1 time), high volume (1 time); a total of 300 data samples are collected for each user for each command, including voice signals and SPS pairs. In the later stage of collection (the last 15 days), 3 kinds of mobile scenarios are arranged for sampling, 5 days for each, specifically walking on flat ground, going up / down stairs, and elevator lifting.

[0104] (5) Data synchronization and annotation specification: All samples contain raw audio signals, left and right ear SPSs, and command labels, user IDs, collection dates, and environmental labels; video recordings are used synchronously during the collection process for reference in attack simulation; each data naming structure is uniform, facilitating automatic batch processing and subsequent model training calls.

[0105] Step 2: Data preprocessing.

[0106] (1) Voice activity detection: This step aims to extract the user's effective pronunciation segment from continuous audio signals and use it for time segmentation of SPS signals. The WebRTC VAD (Voice Activity Detection) tool is used for endpoint detection of audio streams; correspondingly, the SPS segments aligned with the pronunciation period are cut out to form a set of segment pairs: . Wherein, is the audio sequence of the segment, and Figure 3 is its corresponding SPS sequence. The flowchart arrow "speech signal → application → time synchronization" in

[0107] (2) Time synchronization of speech and pressure signals: Since the ear canal pressure change has the characteristic of preceding the speech waveform, the original SPS signal has a time advance and lag relative to the audio, so boundary expansion and linear interpolation alignment are required. Given the audio start and end times to , the audio sequence is: , then the SPS synchronization expansion is:

[0108] . Wherein, the parameter controls the context expansion window (empirical setting, default 0.1-0.3 seconds). The synchronization effect is shown in Figure 4 (a), where the SPS sequence changes before and after speech, and needs to be aligned for model input.

[0109] (3) Baseline drift suppression: In mobile scenarios (such as taking the elevator or climbing the stairs), due to the sensitivity of the air pressure sensor to environmental height and temperature, there may be significant low-frequency drift in the SPS sequence, affecting model stability.

[0110] This application introduces the Empirical Mode Decomposition (EMD) algorithm to eliminate baseline drift, and the specific process is as follows.

[0111] (3.1) EMD decomposition. Find all local maxima and minima in the SPS signal ; respectively, interpolate to form upper and lower envelopes and ; compute local mean: ; extract oscillation component: ; repeat the above process until the intrinsic mode function (IMF) condition is met. Finally, the signal is decomposed into: where, represents the th intrinsic mode function corresponding to the local oscillation of different frequencies, while denotes the final residual component, which captures the global low-frequency trend (e.g., baseline drift); by removing , the target signal without baseline drift can be extracted. .

[0112] (3.2) Reconstructing the de-drifted signal: Select the IMFs that represent the effective signal (e.g., IMF[2] to IMF[5]) and reconstruct the SPS without drift: ; this process is illustrated in Figure 5 with an example of IMF decomposition sequence, and Figure 6 compares the original signal with the reconstructed signal after drift removal.

[0113] (3.3) Normalization: To eliminate the differences in stress amplitude under different environments, the present application performs a minimum-maximum normalization operation on each SPS sequence: , which standardizes the signal to the interval [-1, 1], ensuring that the model receives input of uniform amplitude scale, facilitating network convergence.

[0114] Step 3: Pseudo-sample generation.

[0115] (1) Speech rhythm transformation: The present application uses a phase vocoder to adjust the time-domain rhythm of the speech signal, allowing it to be stretched or compressed in time without changing the pitch of the speech. The specific steps are as follows: given the original speech signal , it is represented by short-time Fourier transform (STFT) as: , the rhythm transformation is achieved by modifying the phase relationship and reconstructing the audio at the transformation ratio , where represents stretching (slowing down), represents compression (speeding up), , excluding 1.0 (original rhythm).

[0116] (2) SPS synchronization extension: Since the original SPS sequence is time-aligned with the original audio , it can be mapped to the new time axis by linear interpolation to generate a pseudo-speech-perceived stress sequence with consistent rhythm​ The interpolation formula is as follows: The length of the speech-perceived stress sequence is adjusted to the length of the variable-speed audio using linear interpolation, keeping the SPS shape feature relatively consistent and only stretching or compressing the time axis.

[0117] (3) Construct an augmented sample set: each group of original samples Multiple pseudo-samples can be generated: Finally, each original sample can be expanded to 6 pseudo-samples through this method, with diverse rhythm distribution, enhancing the model's learning ability for rhythm changes.

[0118] Step 4: Build an authentication model.

[0119] (1) Build an end-to-end identity authentication model (BaroIDNet), with the model architecture as shown in Figure 7 , which can be specifically divided into the following components.

[0120] (1.1) Model input: BaroIDNet network receives two input modalities: speech-perceived stress signal (SPS): input is a pre-processed single-channel time series; shape is , T is the number of time points; the input signal has been removed from the baseline drift and normalized at the front end; spectrogram (Spectrogram): obtained from the SPS sequence by short-time Fourier transform (STFT); the spectrogram is further logarithmized to form a Log-Magnitude Spectrogram; the shape is usually , where F is the number of frequency channels, , and T is the number of frames.

[0121] (1.2) Dual-branch feature extraction: BaroIDNet designs two parallel branches to extract features from the two inputs, processing the time domain and frequency domain respectively: time domain branch (SPS feature extraction): uses a 3-layer one-dimensional convolutional network (1D CNN), each layer containing a convolution layer (Conv1D), different convolution kernel sizes (e.g. 3, 5, 7) to extract multi-scale local changes; batch normalization (BatchNorm); activation function (ReLU); maximum pooling (MaxPool) to downsample, outputting a compressed time-domain feature tensor ; frequency domain branch (Spectrogram feature extraction): uses a 3-layer two-dimensional convolutional network (2D CNN), with convolution kernels of , , each layer also configured with BatchNorm, ReLU, and MaxPool, and the flattened feature map is recorded as .

[0122] (1.3) Cross-domain cross-attention fusion module: to enhance the collaborative expression ability between time-frequency features, BaroIDNet introduces a cross-modal attention mechanism to complete feature alignment and fusion in the following way: the outputs of the two branches are projected into the same dimensional space; the cross-attention weight matrix is calculated respectively, as follows: and V, wherein, , , Q, K, V are the query, key, and value matrices after linear transformation, is the attention dimension scaling factor, and the enhanced time-frequency fusion vector is obtained; BaroIDNet uses bidirectional Cross-Attention (Time→Freq, Freq→Time) for feature interaction.

[0123] (1.4) Discrimination and output layer: the fused features are input to a two-layer fully connected network for dimension compression; the output layer is a Sigmoid function, and the output is the legality probability ; the final judgment rule is: the user is judged to be legal.

[0124] Model configuration and explanation: the network parameters are about 1.2M, which can be adapted to mobile deployment; all convolutional layers use ReLU activation function and BatchNorm to accelerate convergence; Dropout is added between fully connected layers to prevent overfitting.

[0125] Step 5: model pre-training.

[0126] (1) Training task definition and sample organization: this application adopts a "one person one model (one-vs-all)" training strategy, that is, a binary classification model is trained for each registered user, and the positive sample pair is the voice-SPS pair of the user, and the negative sample is the voice-SPS pair produced by other users reading the same command. Each training sample input is the preprocessed SPS, the corresponding spectrogram, and the legality label .

[0127] (2) Training data preparation and pseudo-sample expansion: the training data comes from the general training samples generated in step 3.

[0128] (3) Input features and model architecture binding: from step 4.

[0129] (4) Loss function: in order to consider the prediction accuracy and feature separability, this application designs the following loss function: cross entropy loss (Cross Entropy Loss): wherein, N represents the training batch size,​ represents the predicted output, represents the Sigmoid activation function, and represents the true value label.

[0130] Supervised Contrastive Loss: where, represents the feature vector of the anchor sample after L2 normalization, represents the feature vector of the positive sample, represents the sample slice set in the current batch, contains the anchor positive samples with the same identity, is a temperature parameter that controls the concentration of the feature distribution, and and represent hard negative samples and memory banks, respectively.

[0131] Finally, the total loss function used to train BaroIDNet is: where, .

[0132] (5) Training configuration and optimization strategy: Select Adam optimizer; initial learning rate 0.001; upper limit of training rounds 100; use early stopping mechanism to terminate when the validation accuracy no longer improves; in the contrast learning strategy, construct positive and negative sample pairs in each batch, introduce a hard negative sample mining mechanism, preferentially select heterogeneous samples similar to the anchor sample features, and improve the boundary discrimination ability; when the mixed loss converges, save the Pytorch model parameters.

[0133] Step 6: BaroIDNet authentication model fine-tuning.

[0134] (1) User registration data collection: When the user first uses the system of the present application, prompt the user to read a number of randomly specified commands (e.g. 3 times). In this process, the system synchronously collects the user's voice signal and the corresponding speech perception stress sequence (SPS). Each command is repeated multiple times (e.g. 3-5 times for each command) to enhance the stability of the fine-tuning samples. The collected data undergoes the same preprocessing process as the pre-training phase.

[0135] (2) Model initialization and fine-tuning settings: Load the BaroIDNet model parameters pre-trained on the general dataset, freeze the main structure of the model (convolutional layers and attention modules), and only update the weights of the end fully connected layers, or partially unfreeze the feature layer parameters according to the device computing power.

[0136] (3) Fine-tuning optimization strategy: the training target is to identify the "current user" as a positive class and other users as negative classes, using a shared negative sample pool; the mixed loss function used in pre-training is retained; the learning rate is reduced to 0.0001 to prevent overfitting; the number of training rounds is reduced to 20 rounds; the entire process is completed offline on the server microservice for one-time fine-tuning, and the fine-tuning result will be used for mobile optimization and deployment in step 7.

[0137] Step 7: BaroIDNet deployment on mobile terminal.

[0138] (1) Model export and format conversion. The fine-tuned BaroIDNet model is exported to ONNX format by PyTorch; ONNX Graph Optimizer is used for graph fusion and operator rewriting to compress the model size and improve inference efficiency, generating the final deployment model file for terminal calling.

[0139] (2) Integration of terminal inference framework. The present application supports the following mobile terminal running frameworks: Android: ONNX RuntimeMobile, NNAPI; iOS: CoreML (converted through ONNX). The model can be deployed on CPU, GPU or NPU (Neural Network Processing Unit), and the running platform is automatically selected according to the device configuration.

[0140] (3) Local identity authentication inference process. The user wears the earphone and speaks naturally, and the system real-time acquires the ear canal pressure sequence. Data preprocessing is performed on the local terminal, and forward inference is performed using the local ONNX model to output whether the user is legitimate.

[0141] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other, and each embodiment focuses on the difference from other embodiments.

[0142] The above examples are only used to illustrate the technical solutions of the present application, and are not limited to the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the present application.

Claims

1. A user authentication method based on dynamic deformation of an ear canal, characterized by, The method comprises: In the case that a user wears and uses an in-ear earphone, real-time collection of an original pressure sequence generated during interaction of the user with the in-ear earphone; Voice activity detection, empirical mode decomposition and normalization processing of the original pressure sequence to obtain a standardized sequence; Input of the standardized sequence into a pre-trained user authentication model to obtain an authentication result output by the pre-trained user authentication model; the authentication result is 1 or 0; the pre-trained user authentication model is composed of an Inception feature extraction module, a cross-attention mechanism and a global pooling layer, includes a 768-dimensional feature vector, and outputs the authentication result through a full connection layer and a Sigmoid function; the pre-trained user authentication model is deployed to a terminal connected with the in-ear earphone after light-weight processing; When the authentication result is 1, it is confirmed that the user authentication is successful; when the authentication result is 0, it is confirmed that the user authentication fails.

2. The method of claim 1, wherein, The real-time collection of the ear pressure signal generated during the interaction of the user with the in-ear earphone comprises: Real-time collection of the pressure signal of the left and right ear canals during the interaction of the user with the in-ear earphone through the built-in pressure sensor of the in-ear earphone, and generation of the original pressure sequence with a time stamp.

3. The method of claim 2, wherein, The voice activity detection, empirical mode decomposition and normalization processing of the original pressure sequence to obtain a standardized sequence comprise: Voice activity detection of the original pressure sequence to intercept the pressure sequence corresponding to the existence of interaction; Empirical mode decomposition of the pressure sequence to eliminate baseline drift introduced in the pressure sequence due to height or temperature changes; Min-Max normalization processing of the pressure sequence after empirical mode decomposition to eliminate amplitude range differences in different interaction contents to obtain the standardized sequence.

4. The method of claim 3, wherein, The voice activity detection of the original pressure sequence to intercept the pressure sequence corresponding to the existence of interaction comprises: Extraction of the voice segment in the original pressure sequence through a WebRTC algorithm, and synchronous interception of the pressure sequence.

5. The method of claim 4, wherein, The empirical mode decomposition of the pressure sequence to eliminate baseline drift introduced in the pressure sequence due to height or temperature changes comprises: Decomposition of the original pressure sequence into a plurality of intrinsic mode functions (IMF) and a low-frequency residual component, and preservation of the IMF in [2-5] to reconstruct a drift-free signal to eliminate low-frequency interference.

6. The method of claim 5, wherein, The Min-Max normalization processing of the pressure sequence after empirical mode decomposition to eliminate amplitude range differences in different interaction contents to obtain the standardized sequence comprises: Min-Max normalization processing through the following formula to uniformly scale the pressure sequence to the [-1, 1] interval to obtain the standardized sequence: ; wherein, denotes the pressure value, after normalization, at time , represents the value of the original pressure sequence at time , denotes the minimum value of the original pressure sequence over the entire time sequence, used to determine the lower limit of the pressure sequence; represents the maximum value of the original pressure sequence over the entire time sequence, used to determine the upper limit of the pressure sequence.

7. The method of claim 6, wherein, The Inception feature extraction module at least includes a one-dimensional convolution submodule and a two-dimensional convolution submodule; the one-dimensional convolution submodule and the two-dimensional convolution submodule, and the global pooling layer are used to extract time domain features and frequency domain features of the standardized sequence; The cross-attention mechanism is used to fuse the time domain features and the frequency domain features through a bidirectional attention mechanism to calculate attention weights; And, the fused time domain feature and frequency domain feature are updated by a residual connection to output a 768-dimensional feature vector; The full connection layer and the Sigmoid function are used to output an authentication result.

8. The method of claim 7, wherein, The user authentication model is pre-trained in the following manner: Sample data is taken as input of a preset network model, and the preset network model is trained; In a case where a difference between a sample authentication result and an actual authentication result output by the preset network model satisfies a preset evaluation condition, it is determined that the preset network model is trained, and the preset network model is taken as the pre-trained user authentication model; The sample data is a pressure signal obtained after time axis alignment, rhythm transformation and pressure signal synchronization of an audio signal; a difference between the sample authentication result and the actual authentication result is evaluated by using a cross-entropy loss function and a supervised contrast loss function; a data enhancement strategy is used in the training process to further improve the generalization of the preset network model; the data enhancement strategy is one or more of time stretching, random time offset and Gaussian noise injection.

9. The method of claim 8, wherein, The pre-trained user authentication model is processed in the following manner: The pre-trained user authentication model is compressed to below 5MB by using an ONNX tool chain, and FP16 mixed precision quantization is adopted to reduce the calculation load; Based on the hardware architecture optimization operator of the terminal, hardware acceleration is performed by using a CoreML / NNAPI framework to realize model lightweight.

10. A user authentication system based on dynamic deformation of an ear canal, characterized by, The system is used to perform the user authentication method based on ear canal dynamic deformation according to any one of claims 1 to 9, and the system comprises an in-ear earphone and a terminal, the in-ear earphone is internally provided with a pressure sensor and is connected to the terminal; The pressure sensor is used to collect a raw pressure sequence generated during interaction between a user and the in-ear earphone in real time in a case where the user wears and uses the in-ear earphone; The terminal is used to perform voice activity detection, empirical mode decomposition and normalization processing on the raw pressure sequence to obtain a standardized sequence; the standardized sequence is input into a pre-trained user authentication model to obtain an authentication result output by the pre-trained user authentication model; the authentication result is 1 or 0; the pre-trained user authentication model is composed of an Inception feature extraction module, a cross-attention mechanism and a global pooling layer, includes a 768-dimensional feature vector, and outputs the authentication result through a full connection layer and a Sigmoid function; the pre-trained user authentication model is deployed to the terminal connected to the in-ear earphone after being processed in a lightweight manner; in a case where the authentication result is 1, it is confirmed that the user authentication is successful; in a case where the authentication result is 0, it is confirmed that the user authentication fails.

Images