Identity authentication system of wearable device based on wireless earphone external microphone

Abstract

The invention discloses an identity authentication system of wearable equipment based on a wireless earphone external microphone. The identity authentication system is characterized by comprising the following steps: step 1, customizing an earphone to obtain original audio data; step 2, signal noise reduction and enhancement; 3, constructing a user model; 4, a lightweight self-adaptive multi-window sliding detection algorithm is carried out; 5, interference events are screened out; step 6, feature extraction and data set expansion; and 7, model training and incremental learning. Compared with the prior art, the identity authentication system has the advantages that the head movement signal of a wearer is collected through the necessary external ear microphone on the wireless earphone, stable biological characteristics are extracted, finally training is carried out through the neural network, a legal user is recognized, the identity authentication system can accurately recognize in different environments, and the user experience is improved. And various cheating attacks can be resisted.

Description

Technical Field

[0001] The present invention relates to the field of privacy protection and information security technology, and in particular to an identity authentication system for a wearable device based on a wireless headset and an external microphone. Background Art

[0002] Using the sensing function of wearable devices for identity authentication has been a research focus for many years. At present, the identity authentication systems based on hardware design on wearable devices can be roughly divided into two categories: one is behavior-based authentication, which relies on the uniqueness of human behavioral characteristics, such as human walking gait, arm swing, typing pattern, etc. This solution is mostly implemented by using the IMU equipped on the wearable device, including accelerometers, gyroscopes, etc.; the other is physiological feature-based authentication, which captures the unique individual physiological characteristics of humans, including ear canal shape, heartbeat, breathing sound, teeth biting sound, pupil reaction, etc. This type of solution is mostly based on the in-ear microphone equipped on high-end wireless headphones or by applying external stimulation.

[0003] However, the above authentication schemes inevitably have certain limitations. On the one hand, although the solution based on the in-ear microphone of the headset can capture the physiological characteristics of the user through bone conduction, the occlusion effect will amplify the low frequency of the signal, thereby reducing the interference of environmental noise, but these enhancements also bring certain physiological interference signals, such as signals from heartbeat and breathing. In addition, it must be admitted that not all commercial Bluetooth headsets are equipped with in-ear microphones. The necessity of additional hardware configuration undoubtedly brings additional burdens to headset manufacturers. On the other hand, the solution based on the inertial measurement unit (IMU) on the headset (including accelerometers and gyroscopes) is designed to capture the user's behavior for authentication. Although this method effectively circumvents the influence of environmental noise, the high sensitivity of the IMU may cause the authentication result to be easily interfered by other behaviors (such as gait).

[0004] Head movement is one of the most common behaviors in human daily life. At the same time, head movement is the simultaneous action of bones and muscles, mainly determined by the innate physiological musculoskeletal structure and acquired daily habits. This also means that head movement has a certain uniqueness, and can therefore be regarded as a unique physiological signal. At the same time, considering that the use of wireless headphones is already relatively common. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a device security and privacy protection for wearable devices in response to the deficiencies raised in the above-mentioned background technology. The external ear microphone on the wireless headset is used to capture the sound effects caused by air friction generated during the movement of the user's head, and combined with deep learning to mine the inherent biometric features, a wearable device identity authentication system based on the external microphone of the wireless headset is implemented to realize user identity authentication.

[0006] In order to solve the above technical problems, the technical solution provided by the present invention is: an identity authentication system of a wearable device based on a wireless headset external microphone, characterized in that it comprises the following steps:

[0007] Step 1: Customize headphones to obtain original audio data

[0008] Step 2: Signal noise reduction and enhancement

[0009] Step 3: User model construction

[0010] Step 4: Lightweight adaptive multi-window sliding detection algorithm

[0011] Step 5: Screening out interference events

[0012] Step 6: Feature extraction and dataset expansion

[0013] Step 7: Model training and incremental learning

[0014] Furthermore, the step one is specifically to obtain raw head movement data without any processing by customizing the earphones.

[0015] Furthermore, the step 2 specifically uses a lightweight high-pass filter and a multi-band spectrum subtraction method to reduce noise and enhance the signal, thereby filtering out a large amount of interference signals.

[0016] Furthermore, the step three is specifically to construct a user model based on the sound signal generated by the head movement.

[0017] Furthermore, the step four is specifically to design a lightweight adaptive multi-window sliding detection algorithm to segment the head motion signal.

[0018] Furthermore, the step five specifically uses the Fast-DTW algorithm to match the spectrum flatness feature of the segmentation event with the preset head movement event template.

[0019] Furthermore, the step six specifically designs a novel biometric feature to represent the differences between different individuals, and uses a variational autoencoder to expand the limited registration data. The biometric features include head movement features, head movement conduction features, and body structure difference features.

[0020] Furthermore, the step seven is to train the Resnet50 neural network to implement the binary classification task, and create a comparative learning model, realize incremental learning through the experience replay strategy, continuously update the model parameters, and maintain the stability of the biological characteristics.

[0021] After adopting the above method, the present invention has the following advantages: the head movement signal of the wearer is collected by the external ear microphone required on the wireless headset, stable biometric features are extracted, and finally the neural network is trained to identify the legitimate user. This identity authentication system can accurately identify in different environments and can resist various deception attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the basic structure of the identity authentication system;

[0023] Figure 2 Design of a custom-made headset prototype and schematic diagram of the experimenter wearing the headset;

[0024] Figure 3 False acceptance rate (FAR) of identity authentication experiments conducted on 30 volunteers wearing custom headphones;

[0025] Figure 4 False rejection rate (FRR) of identity authentication experiments conducted on 30 volunteers wearing custom headphones;

[0026] Figure 5 When 10 volunteers wearing headphones were authenticated, the system's false acceptance rate (FAR) in the face of imitation attacks and replay attacks;

[0027] Figure 6 Statistics of the system's false acceptance rate (FAR) and false rejection rate (FRR) after the user wore the customized headphones for one week. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0029] A wearable device identity authentication system based on a wireless headset external microphone comprises the following steps:

[0030] Step 1: Customize the earphones to obtain original audio data, specifically, customize the earphones to obtain original head movement data without any processing.

[0031] Currently, all common commercial Bluetooth headsets on the market will process the audio data collected by the microphone through certain noise reduction algorithms. The algorithms used by these headset manufacturers are different and not open source. Therefore, in order to avoid losing key information in the head movement signal during the processing of these noise reduction algorithms, it is necessary to obtain the unprocessed head movement signal data for analysis and processing. At the same time, in order to ensure the most basic audio playback and sound recording functions of the headset, the customized headset is equipped with a speaker and a microphone, and the microphone is installed as far away from the earplug as possible to simulate the effect of the outer ear microphone.

[0032] Step 2: Signal noise reduction and enhancement, specifically using a lightweight high-pass filter and multi-band spectrum subtraction to reduce noise and enhance the signal to filter out a large amount of interference signals.

[0033] By analyzing the signal spectrum, it is found that most of the energy of the head motion signal is concentrated in the high-frequency part above 10kHz, while most of the noise signals in the real environment (including but not limited to air-conditioning noise, music noise, gait, talking, etc.) are distributed in the low-frequency part of 2kHz. Therefore, most of the interference noise in the collected signal is filtered out by the high-pass filter.

[0034] At the same time, due to the high noise in the real environment, the head motion signal obtained after noise reduction is often very weak, so multi-band spectral subtraction is used to enhance the weak head motion signal. Multi-band spectral subtraction is a technology used for noise suppression. By dividing the signal into multiple frequency sub-bands (or bands) and applying the spectral subtraction formula respectively, the noise can be subtracted more accurately. Suppose the spectrum range corresponding to each band k is f∈[f k-1 , f k ], then apply the spectral subtraction formula on each band k:

[0035]

[0036] in is the estimate of the clean signal in the kth band, Y k (f) is the noisy signal of the kth band, N k (f) is the noise estimate of the kth band, α k is the over-subtraction factor for the kth band. The final output signal is obtained by splicing the processing results of each band.

[0037] Step 3: User model construction, specifically constructing a user model based on the acoustic signal generated by head movement.

[0038] In the system, the head and neck are conceptualized as a unified entity, with the connection between the neck and sternum as the pivot point, thus constructing a single pivot model. The entire head movement process can be regarded as the collective rotation of the head and neck around this pivot. During the head movement, the entire rotation process can also be interpreted as the angular velocity change process in the field of motion mechanics. From the beginning to the end of the movement, the angular velocity gradually increases from zero to the peak value, and then decelerates back to zero. The speed is proportional to the air friction. As the speed fluctuates, the air friction will naturally fluctuate. This change in air friction will inevitably cause the microphone diaphragm to continue to vibrate, thereby generating an acoustic signal.

[0039] The head movement process is intrinsically linked to changes in angular velocity, and this process is significantly affected by the complex bone structure of the neck. Medical and physiological studies have shown that the human neck is composed of seven cervical vertebrae of different shapes and sizes, located between the base of the skull and the thoracic vertebrae. During head movement, the neck muscles are responsible for supporting and moving the head. The proportion of muscle fiber types (slow-twitch fibers and fast-twitch fibers) varies from person to person, and differences in muscle strength and endurance will affect athletic performance. When the central nervous system sends signals to control the movement of the head and neck, the neck muscles contract and relax to generate tension, controlling the sliding and movement of the cervical vertebrae in different directions. The shape and stiffness of the cervical spine directly affect the stability and speed of head movement. For example, a longer cervical spine may require more force to achieve the same rotation speed, while a more flexible cervical spine structure can support a wider range of rotation, although maintaining the speed of rotation may require more force.

[0040] In addition to differences in physiological structure, there are also significant differences in personalized characteristics between different users. For example, long-term posture habits (such as sitting for a long time at the office or frequently looking down at the phone) will have a long-term impact on the neck muscles and joints, thereby changing the movement pattern. In addition, individual health conditions (such as musculoskeletal diseases, neurological diseases, and other health problems) will also have a certain impact on the movement ability of the head and neck. Therefore, even if the same head movement posture is performed, the rotation angle and movement direction of different individuals will still be different, resulting in different sound field changes inside the headphones, which in turn affects the spectral characteristics of the sound signal. Based on these innate differences in physiological structures and personalized acquired movement habits, personalized models for different users are comprehensively considered and portrayed.

[0041] Step 4: A lightweight adaptive multi-window sliding detection algorithm is specifically designed to segment the head motion signal.

[0042] Given that the duration of head motion events is not completely fixed, it is necessary to determine the start point (SP) and end point (EP) of the event. Recognizing the suddenness and instantaneousness of head motion, an adaptive segmentation algorithm based on double sliding windows is proposed. This method dynamically adjusts the energy threshold according to the changing conditions to ensure accurate detection and segmentation of events while ensuring lightweight detection.

[0043] Signal framing: In order to separate valid events from the input audio data stream, three sliding windows W1 are used in the data stream. end 、W2 end , W start First, W1 end and W2 end The signal is framed with a step of 0.5 seconds, and then the average energy of the signal within the window needs to be calculated. After analyzing all the collected head motion signals, it is observed that the duration of most events is about 2 seconds. In order to accurately detect EP, W1 end and W2 end The entire event needs to be covered, so W1 end and W2 end The size of is set to 3 seconds, which is slightly longer than the duration of all head movements. Once the EP of a head movement event is detected, W start It will start from EP and start at W1 end Search SP from right to left and put W start The size is set to 0.5 seconds.

[0044] EP detection, W1 end and W2 end represents two adjacent windows of the same size. Specifically, W1 end The starting point of sliding is W2 end Advance by one window length. Two windows slide simultaneously on the input audio stream. When the user moves his head, the valid event enters W1 first end Because it is located at W2 end Before. This leads to W1 end Energy level E1 end Keep increasing until W1 end Completely surround the event. At this point, E1 end Reaching the maximum value, W1 end and W2 end The energy ratio between them also reaches a peak. Subsequently, the event gradually exits W1 end and enter W2 end , causing the ratio to slowly decrease. The moment corresponding to the peak of the ratio thus marks the EP of the motion event.

[0045] SP detection, using W startTo detect SP, a simple peak detection algorithm is used. However, the threshold is not a preset fixed value, but is adaptively adjusted according to the changing context information. Specifically, considering that the duration of the effective event is very short and the background noise changes little during this short time, the context information at SP and EP should be similar. Therefore, the energy around EP for about 0.5 seconds is calculated as E thresh Used to detect SP. Considering W1 end Complete coverage of the entire incident, W start The sliding range can be in W1 end When E start >E thresh When the SP of the event can be detected.

[0046] Signal Segmentation,After detecting EP and SP, the segmentation results of the original signal are obtained.,But in order to generate fixed-length network input data, all,segmentation results are standardized to a fixed length of 2 seconds.,For segments shorter than 2 seconds, time domain interpolation technology is used to,extend their length. On the contrary, for segments longer than 2 seconds,,redundant parts are eliminated starting from SP.

[0047] Step 5: Interference event screening, specifically using the Fast-DTW algorithm to match the spectral flatness characteristics of the segmented event with the preset head motion event template.

[0048] Although a large amount of low-frequency noise has been successfully filtered out, there are still many high-frequency audio events in real life. This results in some non-head motion events being mixed in the audio events segmented by the detection algorithm. In order to prevent these interference events from affecting the authentication effect of the model (for example, causing frequent authentication failures), by comparing the spectral flatness features of head motion events with common audio events in real scenes (such as gait, speech, music, etc.), it is found that the spectral flatness of head events is highly consistent, so the Fast-DTW algorithm is used for matching for the first time. Fast-DTW is a fast dynamic time warping algorithm, which is an approximate algorithm of DTW with linear time complexity and is suitable for real-time systems. The spectral flatness features of the segmented event are matched with the preset head motion event template. If the distance returned by the match is lower than the threshold set by experience, the event is confirmed to be a head motion event. This method helps to effectively eliminate interference and improve the authentication accuracy of the model.

[0049] Step 6: Feature extraction and data set expansion. Specifically, a novel biometric feature is designed to represent the differences between different individuals, and a variational autoencoder is used to expand the limited registration data. The biometric features include head movement features, head movement conduction features, and body structure difference features.

[0050] Feature extraction,Considering that the user's head movement behavior is often complex and changeable, it is necessary to extract effective and stable physiological features from the segmented signal fragments to effectively identify different users.,Since statistical features cannot popularize a wide range of human biological features, and using deep learning technology to directly process head signals will produce biased biological features, a novel,biometric representation is designed.

[0051] a) Head motion features: Mel frequency cepstral coefficients (MFCC) describe the dynamics between different frequency bands of the sound produced by head motion. MFCC is extracted from the collected head motion signals by combining linear and nonlinear features, which has been proven to be effective in various acoustic signal pattern recognition tasks. Specifically, 16 MFCC features (a total of 16×2 features) are extracted from the acoustic signal of each channel.

[0052] b) Head motion conduction features: The acoustic signals generated by head movement contain unique information about the conduction characteristics of the neck bones and muscle tissue. In order to characterize the sound propagation channel, a commonly used acoustic feature is used for sound resonance analysis, namely the linear prediction coefficient (LPC). LPC abstracts the complex propagation path from head movement to the ear as a channel and characterizes the rendering of the acoustic signal in the channel. Specifically, an external ear microphone is used to capture the acoustic signal generated by the user's head movement, and 16th-order LPC features are extracted from the data of each ear (a total of 16×2 features).

[0053] c) Body structure difference features: The cross-correlation coefficient (CC) of left and right ear signals varies between different individuals, mainly due to differences in head anatomical structure, head movement pattern, auditory physiology, external environment, device wearing position and movement coordination. These factors jointly affect the similarity of the left and right ear sound signals collected by the outer ear microphone during head movement, resulting in differences in the cross-correlation coefficient between individuals. By analyzing these differences, we can better understand and distinguish the head movement patterns of different individuals. Specifically, 16-order CC features (a total of 16×2 features) are extracted from the data of each ear.

[0054] In order to avoid building multiple classifiers for heterogeneous features, they are combined into a complete biometric representation. Specifically, the user's head motion signal is first segmented into non-overlapping windows of 2 seconds. This window is selected because it is statistically found that a single head motion posture of a user usually does not exceed 2s. Then, the Hamming window is applied to segment the data into multiple short frames (e.g., 32 overlapping 100 millisecond frames), and three sets of features are extracted from each frame. These features are grouped into three matrices, corresponding to the MFCC and LPC of the left ear, the MFCC and LPC of the right ear, and the CC of both ears. Then, the values ​​of these features are scaled to integers between 0 and 255, and they are combined into RGB images as subsequent network inputs, where the three matrices correspond to the red, green, and blue channels, respectively.

[0055] Dataset augmentation,In order to minimize the user's identity authentication burden, only a small amount of registration data is required from the user during the registration process. It is well known that neural networks usually require a large amount of training data to mitigate the risk of overfitting and enhance adaptability to new data. Therefore, in order to improve the performance, generalization and robustness of neural networks, data augmentation is necessary even if the data set is limited. Variational Autoencoder (VAE) is a widely used generative model that learns data distribution from latent space and generates new samples. Specifically, VAE maps Mel-spectrograms to points in a low-dimensional latent space, compressing the data into mean and variance vectors. Random sampling from these vectors produces latent vectors, which are then mapped back to the original data space by the decoder to generate new samples. VAE introduces a reconstruction loss to minimize the difference between the input data and the data generated by the decoder, and introduces a KL divergence regularization term to ensure that the latent representation follows a Gaussian distribution, making it easier to sample and generate.

[0056] LOSS VAE =LOSS recon +LOSS KL

[0057] The loss function of VAE is composed of the reconstruction function LOSS recon and KL divergence LOSS KL This loss function is minimized during training in order to constrain the distribution of the latent space while maintaining the quality of the reconstruction.

[0058]

[0059] LOSS recon is the decoder output The MSE loss between the encoder input X reflects the difference between the VAE network generation result and the input data.

[0060]

[0061] LOSS KL The latent distribution q(z|x) used to regularize the encoder output is close to the standard normal distribution p(z) = N(0,I), making the latent space suitable for generation. Where d is the dimension of the latent space, μ refers to the mean vector, and σ refers to the standard deviation vector.

[0062] Step 7: Model training and incremental learning, specifically training the Resnet50 neural network to implement binary classification tasks, creating a comparative learning model, implementing incremental learning through the experience replay strategy, continuously updating model parameters, and maintaining the stability of biological features.

[0063] For model training, the Resnet50 network is used to implement the binary classification task (legitimate users and illegal users). The last fully connected layer of Resnet50 is replaced with a new layer tailored for binary classification. Subsequently, a Softmax function is added to normalize the output to obtain the prediction probability. Finally, a decision is made based on the set threshold. Samples exceeding the threshold can be judged as belonging to the positive class. At the same time, in order to improve the accuracy and processing speed of the model, the following optimizations are made to the network.

[0064] L2 regularization and Dropout: In order to reduce overfitting and improve the generalization ability of the model, L2 regularization and Dropout layers are introduced in Resnet50. L2 regularization controls the complexity of the model by adding a penalty term of the sum of squares of all weight parameters to the loss function. The Dropout layer reduces the interdependence between neurons by randomly "discarding" some neurons (i.e. setting their output to zero) during training, thereby preventing overfitting and improving the generalization ability of the model.

[0065] Attention mechanism: By adding the SE module (Squeeze-and-Excitation) to ResNet50, the feature representation capability is significantly enhanced, effectively improving the classification accuracy. The SE module adaptively adjusts the channel weights of the feature map through global average pooling, fully connected layers, and reweighting steps, so that the model pays more attention to important features and suppresses irrelevant features.

[0066] Loss function: In order to solve the problem of imbalance between positive and negative samples in the training set, the FocalLoss loss function is introduced. By assigning smaller weights to easy-to-classify samples and larger weights to difficult-to-classify samples, the impact of easy-to-classify samples on the loss is reduced, making the model pay more attention to difficult-to-classify samples. This mechanism can improve the performance of the model on unbalanced data sets, especially when processing small class samples, and significantly improve the classification accuracy.

[0067] Model pruning and quantization: In order to reduce the model size and improve the model's inference speed, the original model is pre-trained and then pruned based on the size of the weight value, removing neurons with smaller weights, and then fine-tuning the pruned model to restore the original performance. At the same time, in order to reduce the model's computing time and energy consumption, quantization-aware training is adopted, that is, the model weights and activations are quantized during the training phase and mapped to a lower-precision representation while maintaining the model performance.

[0068] Incremental learning, when a batch of samples is sent to the user classifier, the predicted probability of the user is calculated. Then, the predicted probabilities of positive and negative samples are sorted, and the predicted probabilities of these samples are compared with the threshold σ. Samples exceeding σ are stored in the experience replay memory buffer, where the initial training samples are also saved. A priority queue is used to build this memory buffer. The priority score of the sample is calculated by the following formula. By assigning higher priorities to new data and samples with higher predicted probabilities, the model is promoted to quickly adapt to emerging patterns. This strategy can also alleviate the model's prediction uncertainty for samples that are difficult to distinguish, which ensures a more robust and adaptable learning process. As new data is introduced, samples with lower priorities will be filtered out to prevent wasting computing resources due to excessive buffer size.

[0069]

[0070] Among them, α and β are weight parameters used to balance the influence of prediction probability and τ. τ is the timestamp of the sample. The newer the data, the higher τ is.

[0071] Then a uniform sampling strategy is implemented in the experience replay memory, which involves uniform sampling from different parts of the buffer. The new data is mixed with the initial samples in a ratio of 8:2 for training. After each round of training, the model parameters are updated, which allows the model to acquire new knowledge while retaining the memory of previous knowledge. In contrastive learning, the contrastive loss function LOSS is used. cl To train the model, similar samples are brought closer together and different samples are moved away from each other.

[0072]

[0073] where d = || a n -b n ||2 represents the Euclidean distance between the features of the two samples, y is the label of whether the two samples match, y=1 means that the two samples are similar or matched, y=0 means they do not match, and margin is the set threshold.

[0074] Combined with Figure 1-6 , the specific experiments are as follows:

[0075] according to Figure 1 As shown, the authentication system is divided into two stages:

[0076] During the registration stage, the user wears customized headphones to record a small amount of head movements. After the system denoises and enhances the recorded audio data, it uses the designed multi-sliding window detection algorithm to segment out valid head movement events. It then uses the VAE network to use a small amount of registration data to generate a rich training data set to train the neural network model.

[0077] During the authentication stage, the system also needs to reduce noise and enhance the collected audio data, then extract valid events, make predictions through the trained model, and determine whether the user is a legitimate user based on the set threshold. At the same time, in order to ensure that the detection effect of the model remains stable during long-term use, during the authentication stage, the system will enter samples with higher prediction probabilities into the training sample library and perform incremental training on the model.

[0078] Experimental setup

[0079] In order to verify the effectiveness of the authentication system for different scenarios and attacks, the present invention recruited 30 volunteers aged between 18 and 25, asked them to wear customized headphones, and verbally described the head movement tasks they needed to perform before the experiment. These tasks include nodding, raising the head, and turning the head left and right. Each participant completed 6 trials (including 3 in the laboratory environment, 1 in the external environment, 1 with headphone audio noise, and 1 with motion noise), and each volunteer was required to perform each type of action 30 times.

[0080] Three indicators are set to indicate the reliability of the effect:

[0081] False Acceptance Rate (FAR): describes the probability that the system successfully identifies a non-legitimate user;

[0082] False Rejection Rate (FRR): It indicates the probability that the recognition system rejects a legitimate user;

[0083] The experiment content includes the following three aspects:

[0084] (1) Evaluate the overall performance of the model;

[0085] (2) The effects of different attack methods (imitation attack, replay attack);

[0086] (3) Stability of the model during long-term use.

[0087] Experimental Results

[0088] Figure 2The custom-made headset device is shown, and how users wear the custom-made headset to conduct experiments. Each participant (i.e., P1) is designated as a legitimate user in turn, while the remaining participants (i.e., P2-P30) are regarded as potential attackers. This experimental setup simulates a random attack scenario where the attacker tries to deceive the system with his own data. Figure 3 and Figure 4 The evaluation results are presented, which show that the average FAR and FRR of 30 participants are 3.25% and 5.31% respectively, which shows that the authentication system effectively classifies users based on the captured biometrics and achieves a high level of authentication performance.

[0089] Figure 5 The following shows how the system performs when it is attacked by impersonation and replay attacks:

[0090] Imitation attack: 10 participants were randomly assigned as legitimate users. Then, the remaining participants were asked to observe the head movement patterns of these legitimate users and then try to imitate four different head postures. The results showed that the FAR of all participants was less than 2.7%, and the average FAR was 1.9%.

[0091] Replay attack: A microphone was strategically placed near 10 participants to capture the sounds they made while performing various head movements. These recorded sounds were then replayed through a speaker to launch the attack, and the results showed that the FAR for all participants was less than 1.6%, with an average FAR of 1.19%.

[0092] Figure 6 The authors describe how the system’s authentication performance remains stable over long-term use. They conducted a 14-day study with five randomly selected participants, during which their head movements were continuously recorded. Specifically, the FAR remained consistently below 3.6% and averaged just 2.3%, while the FRR remained below 4.2%, with an average of 2.9%.

[0093] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not limited to the present invention. Any technician familiar with or not in the field can make many possible changes and modifications to the technical solution of the present invention by using the technical content disclosed above without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention should fall within the scope of protection of the technical solution of the present invention.

[0094] The present invention and its implementation methods are described above, and such description is not restrictive, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by it, and does not deviate from the purpose of the invention, and does not creatively design a structure and implementation method similar to the technical solution, they should all fall within the protection scope of the present invention.

Claims

1. An identity authentication system for wearable devices based on wireless earphones and external microphones, characterized in that: It includes the following steps: Step 1: Customize headphones to obtain original audio data Step 2: Signal noise reduction and enhancement Step 3: User model construction Step 4: Lightweight adaptive multi-window sliding detection algorithm Step 5: Screening out interference events Step 6: Feature extraction and dataset expansion Step 7: Model training and incremental learning.

2. The identity authentication system of a wearable device based on a wireless headset and an external microphone according to claim 1, characterized in that: The step 1 specifically involves customizing headphones to obtain raw head movement data that has not been processed in any way.

3. The identity authentication system of a wearable device based on a wireless headset and an external microphone according to claim 1, characterized in that: The step 2 specifically uses a lightweight high-pass filter and multi-band spectrum subtraction to reduce noise and enhance the signal, thereby filtering out a large amount of interference signals.

4. The identity authentication system for wearable devices based on wireless earphones and external microphones according to claim 1, characterized in that: The step three is specifically to build a user model based on the acoustic signal generated by the head movement.

5. The identity authentication system of a wearable device based on a wireless headset and an external microphone according to claim 1, characterized in that: The step four is specifically to design a lightweight adaptive multi-window sliding detection algorithm to segment the head motion signal.

6. The identity authentication system of a wearable device based on a wireless headset and an external microphone according to claim 1, characterized in that: The step five specifically uses the Fast-DTW algorithm to match the spectrum flatness feature of the segmented event with the preset head movement event template.

7. The identity authentication system of a wearable device based on a wireless headset and an external microphone according to claim 1, characterized in that: The step six specifically designs a novel biometric feature to represent the differences between different individuals, and uses a variational autoencoder to expand the limited registration data. The biometric features include head movement features, head movement conduction features, and body structure difference features.

8. The identity authentication system of a wearable device based on a wireless headset and an external microphone according to claim 1, characterized in that: The step seven is to train the Resnet50 neural network to implement the binary classification task, create a comparative learning model, implement incremental learning through the experience replay strategy, continuously update the model parameters, and maintain the stability of the biological characteristics.

Images