A Passive User Authentication Method and System for Smart Glasses Based on Breathing Voiceprints from a Nose Pad Contact Microphone

By employing a nose pad contact microphone sensing scheme and a multi-branch feature extraction network, the problems of insufficient passivity and weak attack resistance in smart glasses user authentication are solved. This achieves passive user authentication with high signal-to-noise ratio, lightweight design, and environmental robustness, making it suitable for continuous identity authentication and privacy protection in smart glasses.

CN123116243APending Publication Date: 2026-11-17FUZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610952782.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing smart glasses user authentication technologies suffer from insufficient passivity, limited signal quality, and weak resistance to attacks, especially failing to effectively utilize the unique sensor position of the nose pad.

Method used

A dual-sided contact microphone sensing scheme with a nose pad is adopted to capture air-conducted turbulent sound and bone-conducted breathing vibration through direct mechanical coupling between the nose pad and the nasal bone. A multi-branch feature extraction and embedding network architecture is designed to extract breathing voiceprint features using high-frequency turbulence, low-frequency bone conduction, handcrafted physical features, and bilateral asymmetry. Attack resistance is improved by placing a rejection-type training strategy.

Benefits of technology

It achieves high signal-to-noise ratio passive user authentication, and features high signal-to-noise ratio, lightweight design, and robustness to environmental and breathing patterns. It can effectively resist spoofing attacks and is suitable for lightweight deployment in smart glasses.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116243A_ABST
    Figure CN123116243A_ABST
Patent Text Reader

Abstract

This invention relates to a passive user authentication method and system for smart glasses based on breath voiceprints from nose pad contact microphones, belonging to the fields of biometric recognition and wearable computing technology. MEMS microphones are embedded in the left and right nose pads of the smart glasses, and through direct mechanical coupling between the nose pads and the nasal bone, air-conducted turbulent sound and bone-conducted vibration signals are simultaneously acquired. The system performs breathing event segmentation and normalization preprocessing on the dual-channel audio streams, extracting four complementary features: high-frequency Mel spectrum, low-frequency Mel spectrum, handcrafted physical features, and asymmetric Mel spectrum. A multi-branch embedding network is used to fuse and encode a 256-dimensional identity embedding vector. During training, an angle-interval loss function containing placement rejection classes is introduced, enabling the system to distinguish between normal wearing and non-wearing states without additional sensors. This invention achieves passive, contactless authentication without user cooperation, possessing strong anti-attack capabilities and low computational overhead, and is suitable for privacy protection in smart glasses and AR / VR headsets.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biometric recognition and wearable computing, and particularly relates to an intelligent glasses passive user authentication method and system based on a nose pad contact type microphone breathprint. BACKGROUND

[0002] Smart glasses are a new generation of mobile computing platforms that are rapidly gaining popularity. Commercial products such as Meta Ray-Ban, Apple VisionPro, and Google Glass Enterprise are used by users throughout the day, providing notification push, voice assistant, navigation, augmented reality overlay, and camera applications, among other services. While providing these services, smart glasses also continuously collect a large amount of privacy-sensitive data, including personal messages, payment credentials, enterprise resources, and continuous audio and video streams. If the device is lost or stolen, the user's entire digital life will be at risk of exposure. Therefore, smart glasses urgently need a passive, unobtrusive, and continuous identity authentication mechanism. However, existing technologies have significant shortcomings in achieving the above three goals, mainly in the following aspects.

[0003] PIN codes or pattern unlocking require touch surfaces or virtual keyboards, which most smart glasses lack and cannot adapt to. Face recognition and iris recognition require inward-facing cameras that face the wearer's face, which contradicts the design of smart glasses' outward-facing optical systems and adds extra hardware burden to the limited volume of the glasses frame. Voice authentication requires the user to speak actively, which is essentially a non-passive method and has a significant decline in accuracy in noisy public environments, which are the most common scenarios for smart glasses. Gait, gesture, and other behavioral biometrics require the user to perform specific actions, which cannot work at all when the user is sitting or stationary. The above methods cannot simultaneously meet the passive, unobtrusive, and continuous authentication requirements.

[0004] In recent years, acoustic-based authentication schemes have gained attention due to their requirement of only speakers and microphones, hardware already available or easily integrated into smart glasses. For example, the SkullConduct scheme transmits audio signals through the eyeglass frame to the skull, using the modulation of the skull's frequency response to identify the user, achieving an accuracy of 97.5% across 10 users. The SonicID scheme transmits inaudible ultrasound waves to scan the user's facial shape, achieving an accuracy of 96.6%. Another scheme embeds an ultrasonic transducer in the nose pads of the glasses, emitting ultrasonic probe signals and analyzing the modulation of reflected signals by the geometry of the nasal cavity tissue, achieving an equal error rate of 0.09 across 11 users. However, these active acoustic schemes all require continuous signal transmission, which not only consumes energy but can also produce perceptible vibrations or buzzing even in the ultrasonic band, affecting the user's wearing experience. Furthermore, continuous signal transmission increases device power consumption, posing a challenge to the limited battery capacity of smart glasses.

[0005] Breathing, as a passive biometric, offers significant advantages: each wearer involuntarily breathes continuously while wearing the glasses, generating a continuous flow of acoustic signals. Studies have shown that nasal airflow patterns can distinguish individuals, achieving an accuracy rate of 96.8% in 97 participants, and the breath fingerprint remained stable over 3.5 months. However, existing breath authentication schemes primarily rely on non-contact microphones. For example, the BreathPrint scheme uses a smartphone microphone to capture three types of nasal breathing sounds, achieving an accuracy rate of over 94% in 10 users; the BreathID scheme utilizes respiratory biometrics for speaker identification, achieving a false recognition rate of 0.04% in 50 users. However, these non-contact schemes can only capture airborne breathing sounds, resulting in limited signal-to-noise ratios and severe performance degradation under ambient noise. More importantly, breathing signals collected in open environments are easily intercepted by nearby malicious microphones, creating conditions for replay attacks and posing serious security risks.

[0006] To improve the signal-to-noise ratio (SNR) of respiratory signals, existing solutions utilize inward-facing microphones in commercial headphones to capture bone conduction respiratory acoustic signals, achieving an accuracy of 95.17% per respiratory cycle in 35 participants. However, the in-ear placement is far from the nasal airway, resulting in an initial SNR of only 1.82 dB. This necessitates a complex multi-stage noise reduction pipeline, including bandpass filtering followed by wavelet noise reduction based on genetic algorithm-optimized parameters. Furthermore, in-ear solutions require users to wear additional headphones, increasing the hardware burden and making them unsuitable for scenarios where only smart glasses are worn. More importantly, the single-microphone configuration in the in-ear system cannot capture information about the asymmetry between the two nasal cavities, lacking natural liveness detection capabilities against spoofing attacks.

[0007] Besides acoustic solutions, there are also breathing authentication schemes based on WiFi channel state information and those based on smartphone frequency-modulated continuous wave sonar. The former requires fixed WiFi infrastructure and cannot be used in mobile scenarios; the latter requires the user to maintain a specific posture and take deep breaths in front of the phone, which does not meet the requirements of being passive and continuous. While multimodal sensor fusion solutions can improve performance, they increase system complexity and hardware requirements.

[0008] In summary, the core deficiency of existing technologies lies in the fact that no solution utilizes the unique sensing location naturally provided by smart glasses—the nose pad. The nose pad is the only part of smart glasses that maintains continuous, mechanically stable contact with the wearer's skin, and it is located directly above the nasal airway, the main outlet for respiratory airflow. This location offers three sensing advantages: First, being directly above the nasal airway outlet, it can capture the complete turbulent airflow spectrum modulated by the wearer's specific nasal geometry; second, its direct mechanical contact with the nasal bone and lateral nasal cartilage provides a low-impedance bone conduction vibration transmission path, enabling the acquisition of bone conduction respiratory vibrations below 500 Hz that are rich in identity information; third, the bilateral configuration of the left and right nose pads can capture anatomical asymmetry caused by septal deviation and differences in turbinate morphology. This asymmetry is stable and individual-specific during normal wear but disappears when the glasses are removed from the nose, serving as a natural liveness detection cue. Existing technologies have failed to discover and utilize these sensing advantages, resulting in significant deficiencies in signal quality, attack resistance, and system convenience. Summary of the Invention

[0009] The purpose of this invention is to address the problems of insufficient passivity, limited signal quality, and weak attack resistance in existing smart glasses user authentication technologies. This invention provides a passive user authentication method and system for smart glasses based on a nose pad contact microphone and breath voiceprint. The system involves the following key research work:

[0010] (1) A double-sided contact microphone sensing scheme for the nose pad is proposed. The direct mechanical coupling between the nose pad and the nasal bone is used to capture both air-conducted turbulent sound and bone-conducted respiratory vibration. The average signal-to-noise ratio reaches 18.3 dB, and high-quality respiratory signals can be obtained without complex noise reduction.

[0011] (2) Design a multi-branch feature extraction and embedding network architecture to extract respiratory voiceprint features from four complementary dimensions: high-frequency turbulence, low-frequency bone conduction, manual physical features and bilateral asymmetry, and fuse them into a 256-dimensional identity embedding vector through a four-branch parallel network.

[0012] (3) A placement-aware training strategy is proposed, which introduces a placement rejection class into the angle interval loss function, so that the network can distinguish between normal wearing and non-wearing states without changing the inference process, thereby resisting spoofing attacks.

[0013] (4) Construct a complete registration-verification-template update pipeline to achieve continuous authentication judgment every 2.5 seconds, and adapt to gradual drift such as wearing posture through a triple-gated template update mechanism.

[0014] To achieve the above objectives, the technical solution of the present invention is: a passive user authentication method and system for smart glasses based on the breathing voiceprint of a nose pad contact microphone, comprising:

[0015] Hardware deployment steps: Embed contact microphones in the nose pads on both sides of the smart glasses, with the sensing surface of the contact microphones facing the wearer's skin to form a direct mechanical coupling with the nasal bone and lateral nasal cartilage.

[0016] Training steps: Collect nasal breathing recordings of registered users and non-contact breathing recordings of speakers from a calibration set that do not overlap; preprocess the recordings to obtain breathing segments; extract four types of complementary features from the breathing segments; input the four types of complementary features into a multi-branch embedding network for processing; use the angle interval loss function, including the placement rejection class, to perform end-to-end training on the multi-branch embedding network to obtain a trained embedding network model.

[0017] Registration steps: Collect the target user's breathing conversation, extract 256-dimensional embedding vectors using a trained embedding network model, and calculate the centroid of all embedding vectors as the target user's identity template for storage;

[0018] Verification steps: Continuously segment the real-time dual-channel audio stream during the user's wearing of smart glasses, extract the embedding vector of each audio segment using a trained embedding network model, calculate the similarity with the stored identity template, and aggregate the similarity scores of multiple segments within the decision window to make an identity determination.

[0019] Update steps: Update the identity template using a confidence gating mechanism.

[0020] Furthermore, the preprocessing includes respiratory event segmentation and signal normalization;

[0021] The respiratory event segmentation specifically includes: averaging the signals from the left and right microphone channels and then performing bandpass filtering; calculating the short-time energy envelope of the filtered signal and performing smoothing; detecting the energy peak position based on the session average energy threshold, and extracting a respiratory segment of fixed duration centered on the peak position;

[0022] The signal normalization specifically involves performing root mean square normalization processing independently on each respiratory segment.

[0023] Furthermore, the four types of complementary features include:

[0024] High-frequency Mel-ray spectral characteristics: Logarithmic Mel-ray spectra of the left and right channel signals are extracted in the frequency band from 1kHz to 4kHz and stacked along the channel axis;

[0025] Low-frequency Mel-frequency spectral characteristics: Logarithmic Mel-frequency spectra of the left and right channel signals in the 0 to 1 kHz frequency band are extracted and stacked along the channel axis;

[0026] Artificial physical characteristics: composed of Mel frequency cepstral coefficients, cross-power spectral density asymmetry characteristics, and subband energy ratio characteristics;

[0027] Asymmetric Mel-frequency spectral characteristics: Dual-channel characteristics are constructed by calculating the difference between the Mel-frequency spectra of the left and right channels and the Mel-frequency spectra of the time-domain signals of the left and right channels.

[0028] Furthermore, the extraction process of the handcrafted physical features specifically includes:

[0029] Extract the Mel frequency cepstral coefficients of the left and right channels in the frequency band from 1kHz to 4kHz;

[0030] Calculate the cross-power spectral density amplitude of the left and right channels in the frequency band from 150Hz to 1000Hz, and then normalize the right channel's self-power spectral density and resample it to a preset number of frequency points.

[0031] The signal energy of the left and right channels is calculated according to the ratio of sub-band energy to the total segment energy for each of the six preset frequency sub-bands.

[0032] Furthermore, the multi-branch embedding network includes four branches that process in parallel and a fusion layer;

[0033] The first branch processes high-frequency Mel spectrum features, sequentially passing through multi-scale convolutional layers, channel attention modules, two-dimensional convolutional modules, one-dimensional temporal convolutional modules, and attention statistical pooling layers to output a 64-dimensional vector;

[0034] The second branch processes low-frequency Mel-frequency spectral features, using a lightweight convolutional neural network structure that includes a two-dimensional convolutional module, a channel attention module, and a one-dimensional temporal convolutional module to output a 64-dimensional vector.

[0035] The third branch processes handcrafted physical features and outputs a 64-dimensional vector through a two-layer fully connected network.

[0036] The fourth branch processes asymmetric Mel-frequency spectral features, using the same topology as the second branch, but its activation function is LeakyReLU;

[0037] The fusion layer concatenates the 64-dimensional vectors output from the four branches into a 256-dimensional vector, which is then mapped by a fully connected layer and normalized by L2 to output the final identity embedding vector.

[0038] Furthermore, the attention statistical pooling layer is configured to calculate the weighted mean and weighted standard deviation of the variable-length temporal features using the learned attention weights, and then concatenate the two and project them into a fixed-length vector.

[0039] Furthermore, in the training step, the rejection class is composed of non-contact breathing recordings from a calibration set of speakers that do not overlap, which is used to enable the network to learn the acoustic feature distribution in the non-wearing state, so that it can distinguish between normal wearing and non-wearing states without modifying the process during the inference stage.

[0040] Furthermore, in the verification step, the audio stream is segmented using a sliding window method, with a window length of 1.0 second and a frame shift of 0.5 seconds. Within a decision window containing 10 consecutive segments, the maximum value of the cosine similarity is selected as the window-level score and compared with a preset threshold to make a decision to accept or reject.

[0041] Furthermore, the confidence gating mechanism includes three gating conditions:

[0042] High confidence check: Select the embedding vector with the highest similarity to the template within the decision window and determine whether it exceeds the preset update threshold;

[0043] Drift anchoring: Determines whether the similarity between the current identity template and the initial registration template is higher than the drift anchoring threshold;

[0044] Rate limiting: Limits the number of template updates within a single wearable session to a preset limit;

[0045] When all three gating conditions are met, the identity template is updated using an exponential moving average method.

[0046] This invention also provides a passive user authentication system for smart glasses based on the breathing voiceprint of a nose pad contact microphone, comprising:

[0047] The sensing module is equipped with contact microphones installed on the left and right nose pads of the smart glasses to collect dual-channel respiratory acoustic signals.

[0048] The processing module is configured to execute the smart glasses passive user authentication method based on the breathing voiceprint of the nose pad contact microphone as described above.

[0049] The storage module is used to store the parameters of the trained multi-branch embedding network model and the user's identity template;

[0050] The decision module is used to output the identity authentication result based on the similarity score calculated in real time.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] This invention, as a passive user authentication method and system for smart glasses based on the breathing voiceprint of a nose pad contact microphone, has the following significant advantages:

[0053] First, it is completely passive and imperceptible. This system uses the wearer's involuntary, continuous breathing as the authentication signal source, without requiring the user to perform any active actions (such as speaking, looking, entering a password, etc.). It completes an identity verification approximately every 2.5 seconds without the user's awareness, truly achieving a continuous and uninterrupted authentication experience.

[0054] Secondly, it offers high signal quality and simple preprocessing. The nose-pad contact microphone, through direct mechanical coupling with the nasal bone, achieves an average signal-to-noise ratio of 18.3 dB, more than 10 dB higher than non-contact solutions. This high signal-to-noise ratio eliminates the need for complex multi-stage noise reduction pipelines; simple root-mean-square normalization is sufficient to preserve the complete spectral characteristics, significantly reducing signal processing complexity and computational overhead.

[0055] Third, multi-dimensional feature complementarity. This system extracts respiratory voiceprint features from four complementary dimensions: high-frequency turbulence, low-frequency bone conduction, handcrafted physical descriptors, and bilateral asymmetry. It fully utilizes the unique advantage of simultaneously acquiring both air conduction and bone conduction acoustic paths at the nose pad position, resulting in stronger identity differentiation capabilities compared to single-feature schemes.

[0056] Fourth, inherent attack resistance. By employing a placement-based training strategy, a rejection class is introduced during the training phase, enabling the network to distinguish between normal wearing and non-wearing states without increasing inference overhead. Experiments show that the system effectively resists random attacks (FAR 1.21%), replay attacks (FAR 0.70%), and spoofing attacks (FAR 1.15%), with FARs for all three attacks below 3%.

[0057] Fifth, it is lightweight and easy to deploy. The model has only 758K (2.9 MB) parameters, and the latency of a single-fragment pure CPU inference is about 204ms, which is far below the real-time processing requirement of 500ms. It is suitable for deployment on low-power ARM processors in commercial smart glasses without the need for dedicated acceleration hardware.

[0058] Sixth, robust to environmental and breathing modes. The system maintains stable performance in four acoustic environments ranging from 35 to 55 dB and in three breathing modes: normal, deep, and rapid. It does not require users to use it in specific environments or adopt specific breathing methods, and has good practical adaptability.

[0059] The main application scenarios of this invention include: continuous identity authentication and privacy protection for smart glasses, user identity management for augmented reality / mixed reality head-mounted devices, access control and permission management for enterprise-level smart glasses, and other scenarios that require passive biometric authentication on nose-contact wearable devices. Attached Figure Description

[0060] Figure 1 This is a diagram of the overall system framework of the present invention.

[0061] Figure 2 This is a diagram illustrating the deployment of a hardware prototype.

[0062] Figure 3 A comparison of the signal characteristics of contact and non-contact microphones.

[0063] Figure 4 This is a schematic diagram of respiratory event detection under three respiratory modes.

[0064] Figure 5 This is a diagram of a multi-branch embedded network structure.

[0065] Figure 6 User-by-user authentication accuracy for 20 registered users.

[0066] Figure 7 FAR, FRR, and EER under different conditions.

[0067] Figure 8 Setup for replay attack experiments.

[0068] Figure 9 For the spoofing attack experiment setup: the non-contact microphone is located 10 cm away from the nose.

[0069] Figure 10 Participant-by-participant FAR for three attack types. Detailed Implementation

[0070] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0071] This invention provides a passive user authentication method and system for smart glasses based on the breathing voiceprint of a nose pad contact microphone, comprising:

[0072] Hardware deployment steps: Embed contact microphones in the nose pads on both sides of the smart glasses, with the sensing surface of the contact microphones facing the wearer's skin to form a direct mechanical coupling with the nasal bone and lateral nasal cartilage.

[0073] Training steps: Collect nasal breathing recordings of registered users and non-contact breathing recordings of speakers from a calibration set that do not overlap; preprocess the recordings to obtain breathing segments; extract four types of complementary features from the breathing segments; input the four types of complementary features into a multi-branch embedding network for processing; use the angle interval loss function, including the placement rejection class, to perform end-to-end training on the multi-branch embedding network to obtain a trained embedding network model.

[0074] Registration steps: Collect the target user's breathing conversation, extract 256-dimensional embedding vectors using a trained embedding network model, and calculate the centroid of all embedding vectors as the target user's identity template for storage;

[0075] Verification steps: Continuously segment the real-time dual-channel audio stream during the user's wearing of smart glasses, extract the embedding vector of each audio segment using a trained embedding network model, calculate the similarity with the stored identity template, and aggregate the similarity scores of multiple segments within the decision window to make an identity determination.

[0076] Update steps: Update the identity template using a confidence gating mechanism.

[0077] This invention also provides a passive user authentication system for smart glasses based on the breathing voiceprint of a nose pad contact microphone, comprising:

[0078] The sensing module is equipped with contact microphones installed on the left and right nose pads of the smart glasses to collect dual-channel respiratory acoustic signals.

[0079] The processing module is configured to execute the smart glasses passive user authentication method based on the breathing voiceprint of the nose pad contact microphone as described above.

[0080] The storage module is used to store the parameters of the trained multi-branch embedding network model and the user's identity template;

[0081] The decision module is used to output the identity authentication result based on the similarity score calculated in real time.

[0082] The following is a detailed implementation process of the present invention.

[0083] This invention constructs a passive user authentication system for smart glasses based on the breathing voiceprint of a nose pad contact microphone. The overall architecture is divided into three collaborative phases: training phase, registration phase, and verification phase. Figure 1 As shown, the three stages sequentially complete model learning, user identity modeling, and real-time identity determination, forming a complete end-to-end authentication chain.

[0084] At the hardware level, this system embeds a microelectromechanical system (MEMS) contact microphone in each of the left and right nose pads of the smart glasses. This microphone makes direct mechanical contact with the wearer's nasal bone and lateral nasal cartilage, simultaneously capturing air-conducted nasal turbulence sound and bone-conducted respiratory vibration signals. The dual-microphone configuration further captures the anatomical asymmetry between the left and right nasal cavities. The average signal-to-noise ratio of the nose pad contact microphones reaches 18.3 dB, approximately 10 dB higher than that of non-contact microphones. Therefore, complex spectral noise reduction processing is unnecessary; simple root-mean-square normalization is sufficient to preserve the signal's spectral characteristics.

[0085] During the training phase, nasal breathing recordings from registered users and non-contact breathing recordings from a non-overlapping calibration set are jointly input into the preprocessing pipeline. The preprocessing pipeline sequentially performs breathing event segmentation, signal normalization, and multi-branch feature extraction, generating four sets of complementary features. Subsequently, the multi-branch embedding network processes the four sets of features separately and fuses them into a 256-dimensional embedding vector. The network is trained end-to-end using the angular interval loss function (ArcFace) and incorporates a placement rejection class, enabling the network to acquire intrinsic placement awareness without altering the inference process, thereby resisting spoofing attacks.

[0086] During the registration phase, users wearing glasses provide multiple breathing sessions. Each session undergoes complete preprocessing and embedding extraction, resulting in a set of 256-dimensional embedding vectors. The centroids of all embedding vectors are stored in the system as the user's identity template.

[0087] During the verification phase, the system continuously segments the real-time audio stream using a sliding window approach. After feature extraction and embedding calculation, each window segment is compared with a stored identity template to calculate a cosine similarity score. The system aggregates the scores of multiple segments within a decision window, takes the maximum score as the window-level score, and compares it with a preset threshold to make an acceptance or rejection decision. The decision window generates an authentication result approximately every 2.5 seconds. Furthermore, the system employs a confidence-gated template update mechanism, using a triple gating system of high-confidence checks, drift anchoring, and rate limiting to prevent imposters from embedding themselves into the template while tracking gradual drifts such as wearing posture.

[0088] 1. Hardware Design

[0089] The core hardware of this invention is a contact microphone embedded in the nose pad of smart glasses, constructing a dual-sided respiratory acoustic sensing platform. For example... Figure 2 As shown, this system embeds a miniature MEMS contact microphone in each of the nose pads on the left and right sides of a standard eyeglass frame. The microphone sensing surface is placed facing the wearer's skin, so that it forms a direct mechanical coupling with the nasal bone and lateral nasal cartilage.

[0090] The nose pad is the only part of the smart glasses that maintains continuous, mechanically stable contact with the wearer's skin, and it is located directly above the nasal airway. This position allows the contact microphone to simultaneously acquire breathing signals through two complementary paths: the first is the air conduction path, which captures the turbulent airflow sound radiating from the nostrils, with frequencies mainly concentrated in the 1 to 4 kHz range, and its spectral characteristics are modulated by the wearer's specific nasal geometry (including nasal septum morphology, turbinate size, and nasal passage cross-sectional area); the second is the bone conduction path, which captures breathing vibrations below 500 Hz through the low-impedance mechanical transmission channel of the nasal bone, and these low-frequency components encode vocal tract resonance and craniofacial bone structure features.

[0091] The left and right nose pad microphones form a dual-sided configuration. Due to the natural anatomical asymmetry of the human nasal cavity (originating from septal deviation and differences in the shape of the left and right turbinates, and regulated by the nasal cycle), the signals collected by the left and right microphones exhibit systematic differences in spectral amplitude and phase. This asymmetry is stable and individual-specific when the wearer is wearing the glasses normally, but when the glasses are removed from the nose (e.g., held in hand or placed on a table), the difference in signals between the two channels approaches zero. This physical characteristic provides the system with a natural placement sensing capability, allowing it to distinguish between normal wearing and non-wearing states without the need for additional sensors.

[0092] Experimental measurements show that the average signal-to-noise ratio (SNR) of the nose-touch contact microphone reaches 18.3 dB, while that of the non-contact microphone located 10 cm from the nose is only 7.6 dB. The contact microphone offers a SNR advantage of over 10 dB. Figure 3 As shown in (a), breathing events are clearly visible in the time-frequency plot of the contact microphone, with spectral energy extending to above 4 kHz; while the non-contact microphone can only capture faint low-frequency traces. Figure 3 (b) compares the power spectral density of the two microphones, showing that the contact microphone has a power advantage of about two orders of magnitude in the bone conduction band below 500 Hz. Figure 3 (c) compares the cross-power spectral density asymmetry ratios of the left and right channels. The contact microphone exhibits a structured frequency-dependent asymmetric mode, while the asymmetry ratio of the non-contact microphone is close to zero. The aforementioned high signal-to-noise ratio allows this system to retain complete spectral characteristics without the need for complex spectral noise reduction pipelines, requiring only simple root-mean-square normalization.

[0093] 2. Signal preprocessing

[0094] Signal preprocessing, as the front-end processing step of the system, is responsible for converting the raw audio captured by the nose microphone into a standardized feature representation suitable for subsequent embedded network processing. This step includes two steps: respiratory event segmentation and signal normalization.

[0095] 2.1 Respiratory event segmentation

[0096] The identifying information from breathing sounds is concentrated at the peak of airflow. The amplitude of the breathing sound is related to the airflow rate. The power-law relationship means that the acoustic energy and biometric information carried during high airflow moments are far greater than those during resting intervals. This invention employs a peak detection method based on energy envelopes to locate respiratory events, with the specific steps as follows.

[0097] First, the signals from the left and right nose microphone channels are averaged to form a mixed channel signal. A 6th-order Butterworth bandpass filter with a cutoff frequency of 1 kHz to 8 kHz is applied to the mixed channel to suppress out-of-band noise. Then, a short-time Fourier transform (STFT) is calculated on the filtered signal, with parameters set as follows: number of FFT points. The frame shift is 256 sampling points, and the sampling rate is 48 kHz. The frame-level energy envelope is obtained by summing the squares of the spectral amplitudes of each frame. (Using width...) A Gaussian kernel of approximately 0.15 seconds is used to smooth the energy envelope to suppress minor fluctuations. Finally, peak detection is performed using the session average energy as the lowest height threshold to locate the energy peak positions for all breathing events.

[0098] like Figure 4 As shown, this detection method can accurately locate respiratory events in three modes: normal breathing, deep breathing, and rapid breathing. Normal breathing generates approximately 8 events within 14 seconds, with a peak interval of approximately 1.75 seconds; deep breathing generates approximately 6 events within 16 seconds, with a peak interval of approximately 3 seconds, and higher energy per breath; rapid breathing generates approximately 12 events within 16 seconds, with a peak interval of approximately 1.33 seconds, and lower energy per breath, but all exceeding the detection threshold. The same set of detection parameters can be used universally without adjustment for different breathing types.

[0099] During the training and registration phases, a 1.0-second window is extracted as a breathing segment centered on each detected peak to maximize the identity information content in each segment. During the continuous verification phase, a fixed sliding window method (window length 1.0 second, frame shift 0.5 seconds, overlap rate 50%) is used to continuously segment the audio stream to ensure that each part of the signal is evaluated.

[0100] 2.2 Signal Normalization

[0101] The nose-pad contact microphone, with its high signal-to-noise ratio of 18.3 dB, eliminates the need for complex spectral noise reduction processing. This system performs root mean square (RMS) normalization independently on each 1.0-second segment, eliminating gain fluctuations caused by coupling stress variations while preserving the spectral shape of the signal.

[0102] 3. Multi-branch feature extraction

[0103] Each 1.0 second breath segment after preprocessing is fed into a multi-branch feature extraction module, generating four sets of complementary features. The four sets of features are designed along two axes: frequency dimension (high frequency and low frequency) and analysis perspective (user authentication and placement status verification).

[0104] 3.1 Branch A – High-frequency Mel-frequency spectral characteristics

[0105] Branch A targets the 1 to 4 kHz frequency band, which is dominated by air-conducted nasal turbulence sound. Log-Melograms are calculated independently for the left and right nose bridge microphone channels, with parameters set as follows: FFT points. The frame shift is 10 ms, the window length is 25 ms, and the Mel filter bank contains 40 filters, covering a frequency range of 1 to 4 kHz. The spectrograms of the two channels are stacked along the channel axis to form a shape... The tensor, in which This is the number of time frames in a 1.0 second clip.

[0106] 3.2 Branch B – Low-frequency Mel-frequency spectral characteristics

[0107] Branch B targets the 0–1 kHz frequency band, which is dominated by bone conduction vibrations transmitted through the nasal bones. These low-frequency components encode vocal tract resonances and craniofacial bone structure features, complementing the turbulent cues in branch A. The feature extraction process is identical to branch A, except that the frequency range of the Mel filter bank is changed to 0–1 kHz. The output shape is also the same. .

[0108] 3.3 Branch C – Manual Physical Characteristics

[0109] Branch C aggregates the three types of hand-designed physical descriptors into a 62-dimensional feature vector.

[0110] The first category is Mel-frequency cepstral coefficients (MFCC), which have 18 dimensions. The 0th to 8th order cepstral coefficients are extracted from the left and right channels in the 1 to 4 kHz frequency band, resulting in 9 dimensions per channel, for a total of... Dimension. MFCC compactly encodes the spectral envelope information.

[0111] The second category is the asymmetric cross-power spectral density feature, which has 32 dimensions. The cross-power spectral density amplitudes of the left and right channels in the 150–1000 Hz frequency band are calculated and normalized using the right channel's auto-power spectral density.

[0112]

[0113] Among them, molecules The cross-power spectral density amplitude of the left and right nose bridge microphone signals at frequency f is used to characterize the correlation coupling strength between the left and right channels; denominator This represents the self-power spectral density of the right channel at the same frequency, serving as a normalization reference to offset the effects of overall breath sound intensity, wearing tightness, and acquisition gain variations on the asymmetry ratio.

[0114] The resulting asymmetry ratio curve was resampled to 32 evenly spaced frequency points to capture the frequency-dependent bilateral asymmetry introduced by the nasal septum geometry and the differential coupling between the left and right nasal bridges.

[0115] The third category is the sub-band energy ratio characteristic, which has 12 dimensions. The signal energy is calculated separately for six frequency sub-bands: 0 to 150 Hz, 150 to 500 Hz, 500 Hz to 1 kHz, 1 to 2 kHz, 2 to 4 kHz, and 4 to 8 kHz. Each channel has 6 dimensions, totaling... The energy of each subband is represented as the ratio of that subband's energy to the total energy of the segment, providing a normalized spectral distribution characteristic.

[0116] Three types of features are concatenated to form a 62-dimensional vector: .

[0117] 3.4 Branch D – Asymmetric Mel-spectral characteristics

[0118] Branch D is specifically used for state verification, encoding the bilateral asymmetric modes of the left and right channels in the 1 to 4 kHz frequency band. This branch constructs two channels:

[0119]

[0120]

[0121] Where Ch0 represents the frequency-domain asymmetric Mel channel, and Ch1 represents the time-domain asymmetric Mel channel obtained after difference; mel(·) represents the same logarithmic Mel spectrum extraction operation as branch A; mic L and mic R These represent the time-domain signals of the left and right nose pad contact microphones, respectively. The first channel calculates the frequency domain asymmetry (by extracting the Mel spectra separately and then subtracting them), and the second channel calculates the time-domain asymmetry (by subtracting in the time domain and then extracting the Mel spectra), i.e., mel(mic). L ) - mel(mic R The expression mel(mic) indicates that the Mel spectra of the left and right channels are extracted separately and then subtracted to describe the difference between the two sides in the frequency domain; L - mic RThe expression indicates that the difference between the left and right channels is first calculated in the time domain, and then the Mel spectrum is extracted to describe the temporal asymmetry caused by the difference in mechanical coupling. When the glasses are worn normally on the nose, the wearer's unique nasal geometry produces a stable, individual-specific asymmetric pattern between the left and right nose pads; when the glasses are removed from the nose, the differential coupling disappears, and both channels approach zero. This physical distinction is the basis for placing sensory rejection. The output shape is... .

[0122] 4. Multi-branch embedded network

[0123] The four sets of features are processed in parallel by a multi-branch embedding network. Each branch produces a 64-dimensional embedding vector. The outputs of the four branches are concatenated into a 256-dimensional vector, which is then mapped by a fusion layer to obtain the final 256-dimensional identity embedding. The overall network structure is as follows: Figure 5 As shown.

[0124] 4.1 Branch A Network

[0125] Branch A is the deepest branch and is designed to capture high-frequency Mel input. The multi-scale time-frequency mode in the model consists of the following five types of building blocks.

[0126] Multi-scale convolutional layer: three parallel two-dimensional convolutional kernels (sizes are respectively) , and The three convolutional kernels act simultaneously on the input, each producing 32 output channels. These three sets of outputs are concatenated along the channel axis to form a 96-channel feature map, which is then batch normalized and activated by ReLU. The use of three different widths of convolutional kernels allows the network to simultaneously capture fine, medium, and coarse-grained temporal structures.

[0127] Channel Attention Module (SE Module): First, it compresses the spatial dimension into channel descriptors through global average pooling, and then passes them through two bottleneck networks (compression ratio). That is, first Dimensions dropped to The dimension is activated by ReLU and then expanded back. The input feature map is then activated by a Sigmoid algorithm to generate channel-wise gated weights, which are then multiplied element-wise by the input feature map to recalibrate the channel responses.

[0128] Two-dimensional convolutional modules: Each module contains one A two-dimensional convolutional layer (padded with 1s) is followed by batch normalization and ReLU activation.

[0129] One-dimensional temporal convolution module: Each module contains a one-dimensional convolutional layer with a kernel size of 3 (padded with 1), followed by batch normalization and ReLU activation, operating along the time axis to model temporal dependencies.

[0130] Attention Statistical Pooling (ASP): Aggregates variable-length temporal sequences into fixed-length vectors using a learned attention mechanism. A small one-dimensional convolutional scoring network computes frame-by-frame attention weights using Tanh and Softmax. Then calculate the weighted mean. and weighted standard deviation :

[0131]

[0132]

[0133]

[0134] In ASP formulas, Let represent the hidden feature vector of the t-th time frame. This represents the attention score given by the scoring network to the output of this time frame. exp(·) is an exponential function, and the summation term in the denominator normalizes the scores for all time frames t'. This represents the weight of frame t in statistical pooling; Indicates by The weighted time series mean Indicates by The weighted time series standard deviation, This represents the element-wise squared deviation of each frame relative to the weighted mean. The square root is used to restore the weighted variance to the standard deviation.

[0135] Will After splicing, a 64-dimensional branch embedding is obtained by linear projection.

[0136] The complete process of branch A is as follows: the input first passes through a multi-scale convolutional layer, then through the SE module; then through a 128-channel 2D convolutional module and max pooling, then through the SE module and adaptive average pooling; then through two consecutive 256-channel 1D temporal convolutional modules to model long-range temporal dependencies; finally, the ASP layer aggregates the time dimension into a 64-dimensional vector.

[0137] 4.2 Branch B Network

[0138] Branch B processes low-frequency Mel inputs. It employs a lightweight CNN architecture. It contains two stages of 2D convolutional modules (32 and 64 channels respectively), each followed by an SE module and max pooling. After adaptive average pooling, a 128-channel 1D temporal convolutional module captures the temporal context, and finally, an ASP layer generates 64-dimensional branch embeddings. All building blocks are consistent with the definitions in branch A.

[0139] 4.3 Branch C Network

[0140] The 62-dimensional handcrafted feature vectors are processed by a two-layer fully connected network. The first layer maps the 62 dimensions to 128 dimensions, followed by batch normalization, ReLU activation, and Dropout. The second layer maps 128 dimensions to 64 dimensions, followed by batch normalization and ReLU activation, outputting a 64-dimensional branch embedding.

[0141] 4.4 Branching D-network

[0142] Branch D shares the same network topology as branch B, but with one modification: all ReLU activations are replaced with LeakyReLU (negative slope). Because the asymmetric features (branch D input) are signed (negative values ​​indicate which channel is dominant), standard ReLU discards the sign information, while LeakyReLU retains negative activations, thus preserving the directional asymmetric information.

[0143] 4.5 Feature Fusion and Embedding

[0144] 64-dimensional output of four branches The vectors are concatenated into a 256-dimensional vector, passed through a fully connected layer (from 256 to 256 dimensions), batch normalized, ReLU activated, and then L2 normalized to obtain the final embedding vector.

[0145]

[0146] in This indicates a splicing operation. This indicates that the fully connected layer is followed by batch normalization and ReLU activation.

[0147] 4.6 Training based on angle margin loss

[0148] The network is trained end-to-end using the additive angular margin (ArcFace) loss function. For the embedding vectors... and their corresponding real categories The loss function is defined as:

[0149]

[0150] Where N represents the number of samples in a training batch, i represents the i-th training sample in the batch, and the pre-calculation factor 1 / N is used to average the classification loss of all samples in the batch, so that the loss value does not change with the batch size. Embedded vector With the The angle between class weights Scaling factor The angular interval is used. ArcFace applies geodesic interval constraints on the hypersphere, causing embeddings of the same identity to cluster tightly and embeddings of different identities to be fully separated.

[0151] 4.7 Placement Perception Training

[0152] To enable the network to inherently possess placement rejection capability, a placement rejection class is added to the ArcFace classification head. The training samples for this class come from non-contact recordings in a calibration set where speakers do not overlap with any registered users. This design leverages a key physical property: the acoustic difference between nose pad wearing and not wearing is a physical phenomenon, dependent on the mechanical coupling path rather than the wearer's identity. Therefore, the placement rejection class trained on the non-overlapping speaker set can generalize to unseen users. Since the placement rejection class only participates in training as an additional identity class in ArcFace, the inference process requires no modification: breathing segments in the non-wearing state naturally map to the rejection region of the embedding space, exhibiting low cosine similarity to any registered user template.

[0153] 5. Registration and Verification Mechanism

[0154] 5.1 User Registration

[0155] During the registration phase, users wear smart glasses to provide multiple breathing sessions. Each session undergoes a complete preprocessing pipeline (breathing event segmentation, normalization, multi-branch feature extraction) and embedding network computation, resulting in a set of 256-dimensional embedding vectors. The system calculates the centroids of all embedded vectors as the user's identity template:

[0156]

[0157] 5.2 Continuous Validation

[0158] During the verification phase, the system continuously segments the real-time audio stream using a sliding window with a window length of 1.0 second and a frame shift of 0.5 seconds. Each segment, after feature extraction and embedding network computation, has its cosine similarity calculated with the stored identity template.

[0159]

[0160] in, This represents the cosine similarity between the i-th verification fragment and the user template. This represents the 256-dimensional embedding vector obtained by the multi-branch embedding network for this segment. This represents the identity template that user u received during the registration phase. This represents the dot product of the two. represents the L2 norm of each; the denominator is used to normalize the dot product so that the score reflects only the similarity of the embedding direction and is not affected by the vector magnitude.

[0161] To suppress transient errors, the system aggregates within the decision window. The scores for each consecutive segment are calculated using a window-level score:

[0162]

[0163] when The system accepts the current wearer's request; otherwise, it rejects it. The frame shift is 0.5 seconds. With this configuration, the system generates an authentication decision approximately every 2.5 seconds. Threshold Determined at the same error rate operation point.

[0164] 5.3 Confidence Gating Template Update

[0165] Wearing posture, skin moisture, and nasal congestion can cause gradual shifts in the embedding distribution over time. To track these shifts and prevent imposters from being embedded in the template, this system employs a confidence-gated template update mechanism, using a triple-gating system for progressive filtering:

[0166] First layer of gating – high confidence check: Select the embedding with the highest cosine similarity to the template within the current decision window. Only when the similarity exceeds a strict update threshold (above the verification threshold) Updates are only attempted when the system is highly confident in them, ensuring that only embeddings that the system is highly confident in are considered for inclusion.

[0167] Second level of gating – drift anchoring: requires the current template With the initial registration template The cosine similarity between them is still higher than the drift anchoring threshold. This prevents the template from deviating too far from the registration point due to the accumulation of continuous marginal updates.

[0168] The third layer of gating—rate limiting: the number of updates within each wearable session cannot exceed [a certain limit]. Secondly, to prevent uncontrolled adaptation in adversarial scenarios.

[0169] Once all three gating conditions are met, the template is updated using an exponential moving average:

[0170]

[0171] It is an adjustable parameter.

[0172] Larger (closer to 1): Updates rely more on the old template, the impact of new embeddings is minimal, and template changes are very slow → Suitable for stable environments, effectively suppressing noise and short-term fluctuations.

[0173] The smaller (closer to 0): the more it relies on the new embedding during updates, and the old template is quickly replaced → suitable for scenarios with rapid environmental changes, it can quickly adapt to the current state, but it is also more susceptible to being skewed by outliers.

[0174] The updated template has been re-normalized to unit length.

[0175] 6. Experiment and Evaluation

[0176] 6.1 Experimental Prototype and Data Acquisition

[0177] Since commercial smart glasses do not provide raw multi-channel audio interfaces to third-party applications, this invention constructs a dedicated research prototype for data acquisition. The prototype hardware includes... Figure 2 As shown, it consists of a USB dual-channel audio driver board and two electret condenser microphone capsules (4.0 mm in diameter, 1.5 mm in height, sensitivity...). The system consists of two microphones (dB). Two microphones are mounted on the left and right nose pads of the commercial eyeglasses frame, respectively, and are coupled to the wearer's skin via soft foam padding. An audio driver board is mounted on the top of the frame and connects to a laptop via USB to record audio at a 48 kHz sampling rate and 16-bit precision.

[0178] This invention recruited 20 volunteers (12 men and 8 women, aged 18 to 45 years) as registered users of the main dataset. Each participant performed three breathing activities (normal breathing, deep breathing, and rapid breathing). Data acquisition spanned three recording days, approximately one week apart. On each recording day, one session was recorded in four acoustic environments: a quiet laboratory (approximately 35 dB), an office (approximately 45 dB), a campus public area (approximately 50 dB), and a dining hall (approximately 55 dB). Each session contained approximately 30 breathing cycles. Each participant generated a total of [data missing]. One session, approximately 1080 breathing cycles. In addition, seven non-overlapping individuals were recruited to record non-contact breathing data, specifically for placement rejection training.

[0179] Training, registration, and threshold calibration all use data from day 1, while data from days 2 and 3 are fully retained for testing, ensuring that the evaluation measures generalization ability across days and environments.

[0180] 6.2 Evaluation Indicators

[0181] This invention uses the following four standard biometric verification indicators:

[0182] Authentication Accuracy (AAR): The percentage of legitimate breath samples that are correctly accepted by the system.

[0183] False Acceptance Rate (FAR): The percentage of imposters or attack samples that are mistakenly accepted as legitimate by the system.

[0184] False Rejection Rate (FRR): The proportion of legitimate breath samples that are falsely rejected by the system.

[0185] Equal Error Rate (EER): The error rate when FAR equals FRR, summarizing the verification accuracy with a single value.

[0186] 6.3 Overall Certification Performance

[0187] With all 20 registered users, 8 test sessions, and three breathing types, the system achieved an overall EER of 3.11%, AAR of 95.85%, FAR of 0.66%, and FRR of 4.15%. Figure 6 As shown, the AAR of all 20 users exceeded 92%, with a mean of 95.85% (standard deviation 1.53%), indicating that the system has consistent validation performance among users with different nasal anatomical structures and breathing habits.

[0188] like Figure 7 As shown in (a), the free anomaly (FAR) remained low at approximately 0.66% across all four noise environments, while the free rate of return (FRR) increased from 1.07% in the quiet laboratory to 5.38% in the restaurant. The energy efficiency ratio (EER) in the quiet laboratory was as low as 0.75%, while the EER in the three noisier environments remained stable between 3.81% and 3.85%. Figure 7 As shown in (b), the performance of the three breathing types is similar, with EER ranging from 2.73% to 3.63% and FAR ranging from 0.5% to 0.7%, indicating that the system does not require users to adopt a specific breathing pattern during verification.

[0189] 6.4 Security Analysis

[0190] This invention evaluates three types of progressively escalating attacks.

[0191] Random attack: After obtaining the victim's glasses, the attacker wears them normally and breathes naturally, relying on the attacker's respiratory tract geometry to randomly generate similar acoustic features. This invention also recruits 20 volunteers with non-overlapping speakers, each recording 5 sessions. Each session is compared with all 20 registered templates (a total of 400 attacker-victim combinations). Across all 49,420 decision windows, the random attack FAR was 1.21%. Figure 10 As shown in (a), the FAR of all 20 attackers is below 3%.

[0192] Replay attack: The attacker records the victim's breathing signal through the nose pads and replays it on a silicone head model via a bone conduction transducer. The silicone head is wearing the victim's prototype glasses. Figure 8 As shown, a silicone head was used instead of a living person to eliminate interference from the participant's own breathing. Ten registered users were selected as victims, and each person underwent five replay sessions (a total of 1625 decision windows). The replay attack FAR was 0.70%. Figure 10 As shown in (b), the FAR of all 10 victims did not exceed 1.5%. The system’s resistance stemmed from a triple physical mismatch: the acoustic impedance difference between silicone and living bone, the inability of a single transducer to reproduce bilateral nasal asymmetry, and the mechanical vibrations not generating actual nasal airflow.

[0193] Impersonation attack: An attacker places glasses close to the nose of an unconscious victim (e.g., asleep) to capture real biometric signals, but the glasses are not properly worn on the nose. Figure 9 As shown, this invention uses a non-contact microphone recording, synchronously recorded 10 cm from the nose, as a controlled proxy for spoofing attacks. This recording preserves the victim's genuine respiratory acoustic characteristics but lacks bone conduction coupling and bilateral asymmetry. The spoofing attack FAR was 1.15% across 5059 decision windows. Figure 10 As shown in (c), the FAR of all 20 users is less than 3%.

[0194] 6.5 Computational Efficiency

[0195] Pure CPU inference tests were performed on a typical laptop (Intel Core i7-10750H, 16 GB RAM). The total processing latency for each 1.0-second segment was approximately 204 ms, with preprocessing (four-branch feature extraction) accounting for 186.3 ms, model forward propagation for 16.3 ms, and cosine scoring less than 0.1 ms. This latency is significantly lower than the 500 ms segment frame shift, ensuring real-time streaming processing. The model has 758K parameters (2.9 MB in float32 format), suitable for deployment on low-power ARM processors in commercial smart glasses.

[0196] This invention discloses a passive user authentication method and system for smart glasses based on breath voiceprints from nose pad contact microphones. It collects breath signals by embedding dual contact microphones in the nose pads and combines multi-branch feature extraction with a deep embedding network to achieve identity verification. This method completes end-to-end passive authentication from breath signal acquisition to identity determination without relying on active signal transmission or requiring user interaction. Its main steps are as follows:

[0197] Step 1. Hardware Deployment: Embed a MEMS contact microphone in each of the left and right nose pads of the smart glasses frame, with the microphone sensing surface facing the wearer's skin, so that it forms a direct mechanical coupling with the nasal bone and lateral nasal cartilage. The dual microphones simultaneously record dual-channel audio at a sampling rate of 48 kHz.

[0198] Step 2. Model Training: Collect nasal breathing recordings from multiple users and non-contact breathing recordings from a calibration set where speakers do not overlap. Perform breathing event segmentation (based on energy envelope peak detection, truncating a 1.0-second window) and root mean square normalization on the recordings. Extract four sets of complementary features from each segment: high-frequency Mel spectrum (1–4 kHz), low-frequency Mel spectrum (0–1 kHz), 62-dimensional handcrafted physical features (MFCC, cross-power spectral density asymmetry, subband energy ratio), and asymmetric Mel spectrum. Input these four sets of features into a multi-branch embedding network. Each branch generates a 64-dimensional embedding, which is then concatenated into a 256-dimensional vector. After a fusion layer and L2 normalization, the final embedding is obtained. The network is trained using an additive angular spacing (ArcFace) loss function, which includes a placement rejection class to give the network placement awareness.

[0199] Step 3. User Registration: The user wears smart glasses and breathes normally; the system records multiple breathing sessions. Each session is preprocessed and extracted by an embedding network to generate a set of 256-dimensional embedding vectors. The centroids of all embedding vectors are calculated as the user's identity template and stored.

[0200] Step 4. Continuous Verification: When the user wears the smart glasses daily, the system continuously segments the real-time audio stream using a sliding window with a 1.0-second window and a 0.5-second frame shift. After feature extraction and embedding calculation, each segment is compared with the stored identity template to calculate cosine similarity. The system takes the maximum cosine score within the decision window (10 consecutive segments) as the window-level score, compares it with a preset threshold, and then makes an acceptance or rejection decision. The system outputs the authentication result approximately every 2.5 seconds.

[0201] Step 5. Template Update: The system automatically updates the identity template using a confidence gating mechanism to adapt to gradual shifts such as wearing posture. Updates require passing three gating mechanisms: high-confidence check (cosine similarity exceeds the update threshold), drift anchoring (the similarity between the current template and the initial template is higher than the drift threshold), and rate limiting (the number of updates per session does not exceed a maximum). After all three gating mechanisms are satisfied, the template is updated using an exponential moving average and re-normalized to unit length.

[0202] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A passive user authentication method and system for smart glasses based on the breathing voiceprint of a nose pad contact microphone, characterized in that, include: Hardware deployment steps: Embed contact microphones in the nose pads on both sides of the smart glasses, with the sensing surface of the contact microphones facing the wearer's skin to form a direct mechanical coupling with the nasal bone and lateral nasal cartilage. Training steps: Collect nasal breathing recordings of registered users and non-contact breathing recordings of speakers from a calibration set that do not overlap; preprocess the recordings to obtain breathing segments; extract four types of complementary features from the breathing segments; input the four types of complementary features into a multi-branch embedding network for processing; use the angle interval loss function, including the placement rejection class, to perform end-to-end training on the multi-branch embedding network to obtain a trained embedding network model. Registration steps: Collect the target user's breathing conversation, extract 256-dimensional embedding vectors using a trained embedding network model, and calculate the centroid of all embedding vectors as the target user's identity template for storage; Verification steps: Continuously segment the real-time dual-channel audio stream during the user's wearing of smart glasses, extract the embedding vector of each audio segment using a trained embedding network model, calculate the similarity with the stored identity template, and aggregate the similarity scores of multiple segments within the decision window to make an identity determination. Update steps: Update the identity template using a confidence gating mechanism.

2. The passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone according to claim 1, characterized in that, The preprocessing includes respiratory event segmentation and signal normalization; The respiratory event segmentation specifically includes: averaging the signals from the left and right microphone channels and then performing bandpass filtering; calculating the short-time energy envelope of the filtered signal and performing smoothing; detecting the energy peak position based on the session average energy threshold, and extracting a respiratory segment of fixed duration centered on the peak position; The signal normalization specifically involves performing root mean square normalization processing independently on each respiratory segment.

3. The passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone according to claim 1, characterized in that, The four types of complementary features include: High-frequency Mel-ray spectral characteristics: Logarithmic Mel-ray spectra of the left and right channel signals are extracted in the frequency band from 1kHz to 4kHz and stacked along the channel axis; Low-frequency Mel-frequency spectral characteristics: Logarithmic Mel-frequency spectra of the left and right channel signals in the 0 to 1 kHz frequency band are extracted and stacked along the channel axis; Artificial physical characteristics: composed of Mel frequency cepstral coefficients, cross-power spectral density asymmetry characteristics, and subband energy ratio characteristics; Asymmetric Mel-frequency spectral characteristics: Dual-channel characteristics are constructed by calculating the difference between the Mel-frequency spectra of the left and right channels and the Mel-frequency spectra of the time-domain signals of the left and right channels.

4. The passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone according to claim 3, characterized in that, The extraction process of the manual physical features specifically includes: Extract the Mel frequency cepstral coefficients of the left and right channels in the frequency band from 1kHz to 4kHz; Calculate the cross-power spectral density amplitude of the left and right channels in the frequency band from 150Hz to 1000Hz, and then normalize the right channel's self-power spectral density and resample it to a preset number of frequency points. The signal energy of the left and right channels is calculated according to the ratio of sub-band energy to the total segment energy for each of the six preset frequency sub-bands.

5. The passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone according to claim 3, characterized in that, The multi-branch embedding network includes four parallel processing branches and a fusion layer; The first branch processes high-frequency Mel spectrum features, sequentially passing through multi-scale convolutional layers, channel attention modules, two-dimensional convolutional modules, one-dimensional temporal convolutional modules, and attention statistical pooling layers to output a 64-dimensional vector; The second branch processes low-frequency Mel-frequency spectral features, using a lightweight convolutional neural network structure that includes a two-dimensional convolutional module, a channel attention module, and a one-dimensional temporal convolutional module to output a 64-dimensional vector. The third branch processes handcrafted physical features and outputs a 64-dimensional vector through a two-layer fully connected network. The fourth branch processes asymmetric Mel-frequency spectral features, using the same topology as the second branch, but its activation function is LeakyReLU; The fusion layer concatenates the 64-dimensional vectors output from the four branches into a 256-dimensional vector, which is then mapped by a fully connected layer and normalized by L2 to output the final identity embedding vector.

6. The passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone according to claim 5, characterized in that, The attention statistical pooling layer is configured to calculate the weighted mean and weighted standard deviation of variable-length temporal features using the learned attention weights, and then concatenate the two and project them into a fixed-length vector.

7. The passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone according to claim 1, characterized in that, In the training step, the rejection class consists of non-contact breathing recordings from a calibration set of speakers that do not overlap, which is used to enable the network to learn the acoustic feature distribution in the non-wearing state, so that it can distinguish between normal wearing and non-wearing states without modifying the process during the inference stage.

8. The passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone according to claim 1, characterized in that, In the verification step, the audio stream is segmented using a sliding window method with a window length of 1.0 second and a frame shift of 0.5 seconds. Within a decision window containing 10 consecutive segments, the maximum value of the cosine similarity is selected as the window-level score and compared with a preset threshold to make a decision to accept or reject.

9. The passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone according to claim 1, characterized in that, The confidence gating mechanism includes three gating conditions: High confidence check: Select the embedding vector with the highest similarity to the template within the decision window and determine whether it exceeds the preset update threshold; Drift anchoring: Determines whether the similarity between the current identity template and the initial registration template is higher than the drift anchoring threshold; Rate limiting: Limits the number of template updates within a single wearable session to a preset limit; When all three gating conditions are met, the identity template is updated using an exponential moving average method.

10. A passive user authentication system for smart glasses based on the breathing voiceprint of a nose pad contact microphone, characterized in that, include: The sensing module is equipped with contact microphones installed on the left and right nose pads of the smart glasses to collect dual-channel respiratory acoustic signals. The processing module is configured to perform the passive user authentication method for smart glasses based on the breathing voiceprint of a nose pad contact microphone as described in any one of claims 1 to 9. The storage module is used to store the parameters of the trained multi-branch embedding network model and the user's identity template; The decision module is used to output the identity authentication result based on the similarity score calculated in real time.