Conventional respiration identity authentication method based on sonar perception
By using sonar signals to sense breathing on a smartphone and combining wavelet transform and deep learning techniques to construct a hybrid neural network model, the device dependence and environmental sensitivity problems of existing breathing authentication methods are solved, and efficient, environment-independent identity authentication is achieved.
Patent Information
- Application Number
- CN202511274637.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-08
- Publication Date
- 2026-01-27
AI Technical Summary
Existing breath-based continuous identity authentication methods rely on dedicated equipment, and the variability of the environment affects performance. The authentication frequency is insufficient to meet the needs of high-security scenarios, and they cannot achieve efficient and environment-independent identity authentication.
Commercial smartphones are used to transmit sonar signals to sense user breathing. The respiratory cycle is segmented by combining empirical wavelet transform and deep learning techniques. A BiLSTM-CNN hybrid neural network model is constructed to achieve dual-modal feature extraction and identity authentication, and an environment-independent identity preservation mechanism is designed.
It achieves efficient identity authentication without the need for dedicated equipment, adapts to changing environments, improves authentication frequency and accuracy, and meets the identity authentication needs of high-security scenarios.
Smart Images
Figure CN121412966A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of biometric authentication methods, specifically relating to a conventional respiratory identity authentication method based on sonar perception. Background Technology
[0002] To ensure system security, traditional one-time authentication methods, such as passwords or fingerprints, typically authenticate a user's identity upon initial access to the system. However, once a user has successfully authenticated, these methods become ineffective, failing to re-authenticate the user's identity and thus failing to effectively detect any unauthorized data access or data tampering after login. More importantly, if an attacker successfully obtains login credentials, they can use these credentials to bypass security measures and launch attacks on the authentication system. Therefore, one-time authentication methods are clearly insufficient for high-security scenarios, especially when dealing with continuous and dynamic security threats.
[0003] To address this deficiency, continuous authentication methods have emerged and gained widespread attention for their ability to continuously monitor and authenticate user identity after the initial login. Among them, continuous biometric authentication methods are gradually becoming the mainstream identity authentication method for protecting mobile terminal devices due to their security, convenience, and user-friendliness. Several breath-based continuous authentication methods have been developed. While these methods can achieve breath-based continuous authentication, they still have some limitations. First, these methods rely on dedicated equipment and are not suitable for everyday life. Second, the authentication environment is highly variable. Factors such as building structure, furniture layout, and the placement of sensors can significantly affect the performance of breath-based continuous authentication methods. Third, these methods typically require analyzing a complete respiratory cycle. Considering a respiratory rate of 12-16 breaths per minute, the authentication frequency of these methods is unlikely to exceed 0.3 Hz. For high-security scenarios, this authentication frequency may be insufficient, as high-security scenarios require minimizing the time interval between adjacent authentications. Summary of the Invention
[0004] To address the problems existing in existing methods, this invention proposes a more practical contactless continuous identity authentication method based on conventional breathing. This method is applicable to smartphones, is independent of the environment, and can achieve more efficient authentication.
[0005] This invention is specifically implemented through the following technical solutions: A conventional respiratory identity authentication method based on sonar perception, comprising an identity authentication phase and an identity retention phase; The specific steps of the identity authentication phase are as follows: S1. The acoustic module of the smartphone emits sonar signals to sense the chest movement of the user during normal breathing and records the breathing signals simultaneously. S2. Design an energy-based signal synchronization algorithm to synchronize the received signal and the transmitted signal obtained in step S1. S3. Calculate the short-time average amplitude sequence of the received signal after synchronization, and develop a respiratory cycle segmentation method suitable for low signal-to-noise ratio environments by combining empirical wavelet transform and deep learning technology to segment the respiratory signal and obtain the "inhalation-exhalation" segment. S4. Use a bandpass filter and a high-pass filter to perform bandpass and high-pass filtering on the "inhalation-exhalation" segment obtained in step S3 to obtain the breathing sound segment and the chest movement segment. S5. Extract acoustic and spectral features from each respiratory sound signal and its corresponding chest movement cycle obtained in step S4, design an identity authentication model to support the input of two modal features, and complete the initial login verification. The specific steps of the identity preservation phase are as follows: S6. The synchronized transmitted signal and received signal obtained in step S2 are processed by frame division. In order to improve the peak value and sidelobe ratio of the signal, the chest cavity motion signal is windowed. The windowed chest cavity motion signal is mixed with the transmitted signal and then low-pass filtered to extract the intermediate frequency IF signal containing the identity information. S7. Considering that smartphones may introduce hardware noise, the moving median method is used to identify outliers in the IF signal obtained in step S6, and then the outliers are replaced with the previous non-outliers. In order to further eliminate the large pulse interference that users may generate due to coughing, sneezing or talking, the pulse ratio IR of each frame of the IF signal is calculated. If the IR exceeds 10, the frame is discarded; otherwise, the frame is retained. S8. Calculate the power spectral density (PSD) of each IF signal frame according to the Welch method, and then calculate the relative correlation coefficient (RCC) between adjacent PSD frames to measure the continuity of identity. When the RCC exceeds 0.85, it indicates that the current breathing comes from the same user, thus maintaining the initial identity information. Otherwise, the identity maintenance phase terminates, and identity authentication is re-executed based on the two modalities of chest movement and breathing sound.
[0006] Preferably, the sonar signal emitted by the smartphone in step S1 is in the form of a sinusoidal frequency-modulated continuous wave with a frequency range of 20kHz-24kHz.
[0007] Preferably, the respiratory signal in step S1 includes the sonar echo signal of chest cavity movement and the respiratory sound signal.
[0008] Preferably, the signal synchronization algorithm in step S2 includes: performing a short-time Fourier transform on the chest cavity motion-related signal, extracting the center frequency power sequence, and determining the starting position of the chirp signal with the first trough.
[0009] Preferably, the deep learning technology described in step S3 is implemented using a neural network model with LSTM as the basic architecture.
[0010] Preferably, the acoustic features extracted in step S5 are Mel frequency cepstral coefficients (MFCC) and gamma-ton frequency cepstral coefficients (GFCC), and the spectral features are Buck spectral features, including spectral flatness, spectral flow, spectral slope, and harmonic ratio; the identity authentication model adopts a hybrid neural network architecture that combines convolutional neural networks and long short-term memory networks (CNN-LSTM).
[0011] Preferably, in step S6, the windowing process involves using the Hanning window function for signal preprocessing when windowing the acquired thoracic motion signal.
[0012] This invention discloses a conventional respiratory identity authentication method based on sonar sensing. Compared with existing technologies, this invention has significant advantages: it only requires a commercial smartphone to emit sonar signals to sense the user's chest movement, eliminating the need for dedicated hardware and significantly reducing deployment costs. Combining Empirical Wavelet Transform (EWT) and deep learning techniques, a robust respiratory cycle segmentation method suitable for low signal-to-noise ratio environments is proposed. Based on the acoustic and spectral characteristics of respiratory signals, a BiLSTM-CNN hybrid neural network model is constructed to achieve high-precision bimodal initial identity authentication. To further improve authentication efficiency, an environment-independent identity retention mechanism is designed. This mechanism ensures continuous verification of the user's identity by real-time tracking and analysis of mid-frequency signals affected by chest cavity movement. Attached Figure Description
[0013] Figure 1 This is a flowchart of a conventional respiratory identity authentication method based on sonar perception according to the present invention; Figure 2 This is the sequence-to-sequence regression model of the present invention; Figure 3 This invention provides a dual-modal identity authentication model. Detailed Implementation
[0014] The present invention will be further described below with reference to the accompanying drawings to provide a better understanding of the technical solution.
[0015] This invention provides a conventional respiratory authentication method based on sonar sensing. A smartphone speaker emits a linear frequency modulated sine wave (FMCW) signal with a frequency range of 20-24 kHz, simultaneously recording respiratory sounds and sonar echoes caused by chest cavity movement. By extracting the dual-modal features of respiratory sounds and chest cavity movement, the user's initial login identity is identified. After successful initial authentication, the system enters an identity retention phase, continuously monitoring and analyzing changes in the mid-frequency signal affected by chest cavity movement to continuously verify the user's authorization status.
[0016] Specifically, such as Figure 1 As shown, the provided sonar-based conventional respiratory authentication method includes the following steps.
[0017] Identity verification stage: S1. The acoustic module of the smartphone emits sonar signals to sense the chest movement of the user during normal breathing and simultaneously records the breathing signals (including the sonar echo signal of the chest movement and the breathing sound signal).
[0018] In this step, to improve the ability to detect minute respiratory movements, the present invention optimizes the linear FMCW sonar signal by introducing a sine FMCW signal. Its frequency at time t is given by the following formula:
[0019] Among them, f c B and T represent the lower limit of the frequency, the sweep bandwidth, and the sweep period, respectively. Its phase expression is:
[0020] Considering the sampling rate of smartphones (most commercial smartphones have a sampling rate of 48kHz) and to avoid being heard by the human ear, the frequency range of this invention is set to [20kHz, 24kHz]. The sweep period of the chirp signal is set to 50ms.
[0021] S2. An energy-based signal synchronization algorithm was designed to synchronize the received signal and the transmitted signal obtained in step S1.
[0022] In this step, the received signal is first subjected to a 20kHz high-pass filter, retaining only the signal related to chest cavity movement. Then, a short-time Fourier transform (STFT) is performed on the first second of the received signal to obtain its time-frequency energy distribution. Based on this characteristic, the center power sequence of the received signal is extracted, and the first trough is determined as the starting position of the chirp signal.
[0023] S3. Calculate the short-time average amplitude sequence of the synchronized thoracic motion signal, and develop a respiratory cycle segmentation method suitable for low signal-to-noise ratio environments by combining empirical wavelet transform and deep learning technology to segment the respiratory signal and obtain the "inhalation-exhalation" segment.
[0024] This method combines empirical wavelet transform (EWT) and deep learning techniques. First, the short-time average amplitude (STAA) of the thoracic motion signal is calculated to obtain the energy variation pattern of respiration. Then, the EWT is used to decompose the STAA sequence into multi-resolution analysis (MRA) components of different frequencies. Subsequently, a sequence-to-sequence regression model is trained to generate a pseudo-STAA sequence to compensate for amplitude differences in the original STAA sequence, enabling it to more accurately match the temporal characteristics of the optimal MRA components.
[0025] The structure of the model is as follows: Figure 2 As shown, the feature extraction module consists of three LSTM layers and three FC layers, each containing 128 hidden units. The last LSTM layer is followed by a LeakyReLU activation function to enhance the model's non-linear processing capabilities.
[0026] After obtaining the pseudo-STAA sequence output by the model, the correlation between this sequence and each MRA component is calculated to determine the optimal MRA component. The formula for calculating the correlation is as follows:
[0027] In this context, A and B represent the pseudo-STAA sequence and the MRA component, respectively. and This represents the mean and standard deviation of A, while and Let B represent the mean and standard deviation. N represents the sequence length. Finally, the MRA component with the highest correlation to the pseudo-STAA sequence is selected as the optimal MRA component, and the respiratory cycle is precisely segmented based on its valley information.
[0028] S4. Apply a bandpass filter of 1500-5000Hz and a high-pass filter of 20kHz to the "inhalation-exhalation" segment obtained in step S3 to obtain the breathing sound segment and the chest movement segment.
[0029] S5. Extract acoustic features (MFCC and GFCC) and Barker spectral features (spectral flatness, spectral flow, spectral slope, and harmonic ratio) from each respiratory sound signal obtained in step S4 and its corresponding thoracic motion cycle. These features are then input into a system such as... Figure 3 The identity authentication model shown completes the initial login verification. The feature extraction module of this model mainly consists of two BiLSTM layers and two CNN layers.
[0030] Identity maintenance phase: S6. The synchronized transmitted and received signals obtained in step S2 are divided into 2-second frames. To improve the peak-to-side-lobe ratio of the signal, the Hanning window shown below is applied to the thoracic motion signal.
[0031]
[0032] Where N is the window length and n is the sample index. Then, the windowed chest motion signal is mixed with the transmitted signal:
[0033] in, This represents the transmitted signal at time t. This represents the windowed thoracic motion signal at time t.
[0034] The mixed signal is processed by a low-pass filter to extract the IF signal containing the identity information.
[0035] S7. Considering the potential introduction of hardware noise by smartphones, the moving median method is used to identify outliers in the IF signal obtained in step S6, and then the outliers are replaced with the previous non-outlier values. To further eliminate large pulse interference that users may generate due to coughing, sneezing, or talking, the pulse ratio (IR) of each frame of the IF signal is calculated:
[0036] If the IR exceeds 10, discard the frame; otherwise, retain the frame.
[0037] S8. Calculate the PSD of each IF signal frame using the Welch method. This method includes segmenting the IF signal, applying a window function, calculating the Discrete Fourier Transform, and finally averaging the power spectrum of each segment. Then, calculate the relative correlation coefficient (RCC) between adjacent PSD frames to measure the continuity of the respiratory signal.
[0038] Where x and y represent two adjacent PSD frames, and y and x are the mean values, respectively, and n is the frame length. ,in and Let x and y represent the rank values of x and y at the i-th sample, respectively. When the RCC exceeds 0.85, it indicates that the current breathing comes from the same user, thus maintaining the initial identity information. Otherwise, the identity preservation phase terminates, and identity authentication is re-performed based on the two modalities of chest movement and breathing sound.
[0039] In practical applications, this invention can manifest as a system, method, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions on it for causing a processor to implement aspects of the invention.
[0040] The foregoing does not limit the scope of the present invention. It should be noted that, for those skilled in the art, various changes, improvements, additions or substitutions can still be made without departing from the essential content of the present invention. All such changes should be considered as part of the present invention and included in its protection scope.
Claims
1. A conventional respiratory authentication method based on sonar perception, characterized in that, This method includes an identity authentication phase and an identity preservation phase; The specific steps of the identity authentication phase are as follows: S1. The acoustic module of the smartphone emits sonar signals to sense the chest movement of the user during normal breathing and records the breathing signals simultaneously. S2. Design an energy-based signal synchronization algorithm to synchronize the received signal and the transmitted signal obtained in step S1. S3. Calculate the short-time average amplitude sequence of the received signal after synchronization, and develop a respiratory cycle segmentation method suitable for low signal-to-noise ratio environments by combining empirical wavelet transform and deep learning technology to segment the respiratory signal and obtain the "inhalation-exhalation" segment. S4. Use a bandpass filter and a high-pass filter to perform bandpass and high-pass filtering on the "inhalation-exhalation" segment obtained in step S3 to obtain the breathing sound segment and the chest movement segment. S5. Extract acoustic and spectral features from each respiratory sound signal and its corresponding chest movement cycle obtained in step S4, design an identity authentication model to support the input of two modal features, and complete the initial login verification. The specific steps of the identity preservation phase are as follows: S6. The synchronized transmitted signal and received signal obtained in step S2 are processed by frame division. In order to improve the peak value and sidelobe ratio of the signal, the chest cavity motion signal is windowed. The windowed chest cavity motion signal is mixed with the transmitted signal and then low-pass filtered to extract the intermediate frequency IF signal containing the identity information. S7. Considering that smartphones may introduce hardware noise, the moving median method is used to identify outliers in the IF signal obtained in step S6, and then the outliers are replaced with the previous non-outliers. In order to further eliminate the large pulse interference that users may generate due to coughing, sneezing or talking, the pulse ratio IR of each frame of the IF signal is calculated. If the IR exceeds 10, the frame is discarded; otherwise, the frame is retained. S8. Calculate the power spectral density (PSD) of each IF signal frame according to the Welch method, and then calculate the relative correlation coefficient (RCC) between adjacent PSD frames to measure the continuity of identity. When the RCC exceeds 0.85, it indicates that the current breathing comes from the same user, thus maintaining the initial identity information. Otherwise, the identity maintenance phase terminates, and identity authentication is re-executed based on the two modalities of chest movement and breathing sound.
2. The conventional respiratory authentication method based on sonar perception as described in claim 1, characterized in that, In step S1, the sonar signal emitted by the smartphone is in the form of a sinusoidal frequency modulated continuous wave with a frequency range of 20kHz-24kHz.
3. The conventional respiratory authentication method based on sonar perception as described in claim 1, characterized in that, The respiratory signals in step S1 include sonar echo signals of chest cavity movement and respiratory sound signals.
4. The conventional respiratory authentication method based on sonar perception as described in claim 1, characterized in that, The signal synchronization algorithm described in step S2 includes: performing a short-time Fourier transform on the chest cavity motion-related signal, extracting the center frequency power sequence, and determining the starting position of the chirp signal using the first trough.
5. The conventional respiratory authentication method based on sonar perception as described in claim 1, characterized in that, The deep learning technique described in step S3 is a neural network using LSTM as its core.
6. The conventional respiratory authentication method based on sonar perception as described in claim 1, characterized in that, The acoustic features extracted in step S5 are Mel frequency cepstral coefficients (MFCC) and gamma-ton frequency cepstral coefficients (GFCC), and the spectral features are Buck spectral features, including spectral flatness, spectral flow, spectral slope, and harmonic ratio. The identity authentication model adopts a hybrid neural network architecture that combines convolutional neural networks and long short-term memory networks (CNN-LSTM).
7. The conventional respiratory authentication method based on sonar perception as described in claim 1, characterized in that, The windowing process in step S6 involves applying a Hanning window to the thoracic motion signal.