A Deep Breath Authentication Method Based on Sonar Sensing

Abstract

A deep breath-based identity authentication method using sonar perception, the method comprising the following steps: the user holds the smart device or places the smart device on the desktop to take a deep breath, and the acoustic module built into the smart device emits sonar signals in a specific modulation manner and frequency, and receives the echo signals and breathing sound signals of the combined chest and abdomen movement of the user. By extracting the breathing sound signals, the combined chest and abdomen movement signals, and the correlation sequences between the breathing sound signals and the combined chest and abdomen movement signals, a multi-stream identity authentication model is trained to achieve user identity authentication. To prevent advanced replay attacks, based on the characteristics that the combined chest and abdomen movement signals are affected by the environment and the device during secondary playback, and combined with the vibration data collected by the acceleration sensor built into the smart device, an liveness detection mechanism is designed. This method is simple to operate and low in cost, and can be used as an effective supplement to the mainstream authentication method to improve the security and reliability of identity authentication.

Description

Technical Field

[0001] The present invention belongs to the technical field of identity authentication, and particularly relates to a deep - breath identity authentication method based on sonar perception. Background Art

[0002] With the rapid development of ubiquitous computing and mobile computing, intelligent devices have become the most important connection points between humans, the network, and the physical space. At the same time, their portability and accessibility pose challenges to user authentication, which is regarded as an important access mechanism to ensure the reliability and security of this connection point. On the other hand, the rich sensing and computing resources in intelligent devices have also prompted many researchers to implement local authentication by identifying users' biometric features. In fact, some biometric authentication methods have been widely studied and applied to intelligent devices, such as fingerprints, irises, faces, and voices. However, the rich resources in intelligent devices are a double - edged sword for voice - and face - based authentication, because personal voice and facial data are easily copied and leaked, which increases the possibility of replay attacks and spoofing attacks. In addition, the latest large - scale artificial intelligence models bring new security risks by using face features and voice recordings to synthesize false information for telecom fraud.

[0003] Currently, some researchers have attempted to use breathing, which is difficult to detect and imitate, for identity authentication. Some methods rely only on breathing sounds for user identity authentication. Although they achieve good authentication performance, they are not applicable to noisy environments and are vulnerable to advanced replay attacks and spoofing attacks. In addition, most of the methods using breathing sounds for user identity authentication are wearable or contact - type methods, making users less comfortable. There are also some methods using dedicated devices (such as millimeter - wave radars and Wi - Fi) for non - contact user identity authentication, which increases additional overhead. Besides, neither of these two methods considers the influence of personal physiological states on breathing patterns. Changes in physiological states may directly lead to user authentication failures. Summary of the Invention

[0004] Aiming at the problems existing in the existing methods, the present invention provides a deep - breath identity authentication method based on sonar perception. This method verifies the user's identity in a non - contact manner, overcomes the defect that existing breathing authentication methods are difficult to resist advanced replay and spoofing attacks, and effectively reduces the influence of personal physiological states and environmental noise on user identity authentication. It is specifically realized through the following technical solutions:

[0005] A deep - breath identity authentication method based on sonar perception includes the following steps:

[0006] S1. The user holds the smart device according to their preference or places it on the tabletop. The built-in speaker of the smart device emits sonar signals in a pre-set modulation method and frequency. At the same time, the embedded microphone in the smart device receives the sonar echo signals and breathing sound signals of the user's chest and abdomen movements.

[0007] S3. Denoise the received signals obtained in step S1 through a high-pass filter to obtain the chest and abdomen combined movement signals.

[0008] S6. Calculate the energy sequence of the chest and abdomen combined movement signals obtained in step S2, and train a model to determine the start and end time endpoints of deep breathing.

[0009] S9. Based on the deep breathing endpoints obtained in step S3, determine the activity range of deep breathing in the received signals obtained in step S1.

[0010] S12. Respectively perform high-pass and band-pass filtering on the received signals processed by endpoints in step S4 to obtain the chest and abdomen combined movement signals and the corresponding breathing sound signals.

[0011] S15. Train a living body monitoring classifier based on the chest and abdomen combined movement signals and the vibration data of the smart device caused by hand holding. This living body monitoring classifier is used to identify whether the source of deep breathing is a real person or a replay attacker. If the living body monitoring classifier detects that the deep breathing does not come from a real person, the following steps will no longer be carried out.

[0012] S18. Extract features from the chest and abdomen combined movement signals and breathing sound signals, calculate the correlation between the chest and abdomen combined movement signals and the corresponding breathing sound signals, and obtain the correlation sequence.

[0013] S21. Design a multi-stream identity authentication model to support the input of the three-factor features of the chest and abdomen combined movement signals, breathing sound signals, and correlation sequence obtained in step S7, and perform identity authentication on the user.

[0014] Further, the sonar signal emitted by the smart device in step S1 is in the form of a frequency-modulated continuous wave, and the frequency range is 20 kHz - 24 kHz, where each ultrasonic chirp signal is 50 ms.

[0015] Further, the cut-off frequency of the high-pass filter in step S2 is 20 kHz.

[0016] Further, the method for determining the start and end time points of deep breathing in step S3 is specifically as follows: Calculate the short-time energy and short-time average amplitude sequence of the signal, and then input it into a sequence-to-sequence model based on a gated recurrent unit to judge the activity range of deep breathing.

[0017] Further, in step S5, the cut-off frequency of the high-pass filter is 20 kHz, and the cut-off frequencies of the band-pass filter are set to 500 - 3000 Hz.

[0018] Further, the method for detecting replay attacks in step S6 is as follows: First, synchronously obtain the combined chest and abdomen motion signals and capture the vibration data of the smart device caused by the slight movement of the hand using the built-in acceleration sensor of the smart device; then extract the standard deviation, shape factor, and pulse factor features of the combined chest and abdomen motion signals, as well as the maximum amplitude, signal amplitude area, and average absolute value of the smart device vibration data; finally, input these six-dimensional features into a support vector classification model for identifying replay attackers.

[0019] Further, in step S7, the method for calculating the correlation is to measure the correlation between the combined chest and abdomen motion signal and the corresponding breathing sound signal through mutual information.

[0020] The method of the present invention has significant advantages compared with existing breathing authentication methods: By introducing sonar signals to sense the combined chest and abdomen motion of the user, it not only effectively reduces the interference of environmental noise on the authentication result, but also reduces the risk of simulation attacks. In addition, the present invention uses the combined chest and abdomen motion signal and the vibration data of the smart device caused by hand holding to train a live body monitoring classifier to identify and defend against advanced replay attacks, further improving the robustness of the system. The present invention also fully considers the influence of personal physiological states on user authentication, thereby ensuring the reliability of the authentication process under different physiological conditions. The sonar signals used in the present invention are in the inaudible range of the human ear and cause no interference to the user, having good user-friendliness. More importantly, this method can be directly deployed on commercial smart devices without additional hardware support, thus greatly reducing the implementation cost and hardware burden and having broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flowchart of the present invention;

[0022] Figure 2 is a network diagram of feature extraction and identity authentication of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The following further describes the present invention with reference to the accompanying drawings of the specification to better understand the technical solution.

[0024] A deep - breath identity authentication method based on sonar perception of the present invention uses an acoustic module built in intelligent devices such as smart phones and tablets to transmit FMCW sonar signals, and simultaneously captures the echo signals and breathing sound signals of the combined chest and abdomen movement caused by the user's deep breath to verify the user's identity. To enhance the security of the system, the present invention uses the combined chest and abdomen movement signals and the vibration data of the intelligent device caused by the user's hand - holding to train a living body monitoring classifier to identify and defend against advanced replay attacks.

[0025] Specifically, as Figure 1 shown, the provided deep - breath identity authentication method based on sonar perception includes the following steps:

[0026] S1. The user holds the intelligent device according to their preference or places it on the desktop. The built - in speaker of the intelligent device emits sonar signals in a pre - set modulation method and frequency, and at the same time, its embedded microphone receives the sonar echo signals and breathing sound signals of the user's chest and abdomen movement.

[0027] In this step, the sonar signal emitted by the intelligent device is in the form of a frequency - modulated continuous wave, and the frequency range is 20kHz - 24kHz, where each chirp signal is 50ms. The time - domain representation of the chirp signal is as follows:

[0028]

[0029] Where , N ∈ Z, A represents the amplitude of the transmitted signal, fc=(F l + F h ) / 2 represents the center frequency, B = F h - F l represents the bandwidth, T s represents the scanning period.

[0030] To reduce the power leakage caused by discontinuous frequencies, the present invention applies a cosine window to each chirp signal:

[0031] ,

[0032] Where M is the chirp length, and α is the cosine fraction, which is used to control the length of the edge transition region of the window function.

[0033] S2. The received signal obtained in step S1 is denoised by a high - pass filter to obtain the combined chest and abdomen movement signal. The cut - off frequency of the high - pass filter is set to 20kHz.

[0034] S3. The deep breathing activity is usually accompanied by an increase in signal energy and a change in signal amplitude. Therefore, in order to detect the start and end points of deep breathing, the short-time energy and short-time average amplitude sequences of the combined thoracoabdominal motion signal obtained in step S2 are calculated. Next, a sequence-to-sequence classification model based on gated recurrent units is trained. Then, these two energy sequences are input into the classification model to obtain an output sequence of 0 and 1. 0 indicates no deep breathing, and 1 indicates deep breathing.

[0035] S4. Determine the activity range of deep breathing in the received signal obtained in step S1 based on the output sequence obtained in step S3.

[0036] S5. The received signal processed at the end points in step S4 is respectively subjected to high-pass and band-pass filtering to obtain the combined thoracoabdominal motion signal and the corresponding respiratory sound signal. The cut-off frequency of the high-pass filter is set to 20 kHz. Since there is interference in the direct connection path from the speaker to the microphone in the signal, the cut-off frequency of the band-pass filter is set to 500 - 3000 Hz.

[0037] Since the filter mainly processes the frequency domain of the signal. Therefore, filtering may introduce some extreme peaks or valleys in the time domain. In order to process these extreme values, the median absolute deviation algorithm is used to detect outliers. Then, the outliers are replaced with the last non-outlier value.

[0038] The threat of advanced replay attacks exists in breath-based authentication, that is, an attacker may invade the user's device, steal the original recording containing breath characteristics, and replay these recordings on other devices to deceive the authentication system. Theoretically, when the stolen deep breathing audio re-enters the authentication system: First, an analog signal is generated through the audio digital-to-analog converter (DAC, in the case of a smartphone) of the attacker's device; then, this analog signal is amplified by an audio amplifier and converted into a sound wave signal through a speaker, propagated through the air, and finally sampled by the smart device equipped with this authentication system. However, during this transmission process, the replayed deep breathing audio is subject to two types of interference: one is device-related interference (such as non-linear distortion of the audio amplifier), and the other is environment-related interference (such as multipath interference), resulting in the time-domain signal becoming flat. In addition, when users use smartphones, they usually hold the device or place it on the desktop. In the held state, deep breathing will naturally cause slight movements of the hand, which are difficult to detect by the naked eye and difficult to precisely replicate during the attack. However, these slight movements will be reflected by the phone's vibration and accurately captured by the accelerometer built into the smartphone.

[0039] Therefore, in order to enhance the system's ability to resist advanced replay attacks, a liveness monitoring classifier is trained based on the combined chest and abdomen movement signals and the vibration data of the smart device caused by hand holding (collected by the accelerometer built into the smart device).

[0040] In this step, first, the combined chest and abdomen movement signals are synchronously obtained, and the vibration data of the smart device caused by the micro-movement of the hand is captured using the accelerometer built into the smart device; then, the standard deviation, shape factor, and pulse factor features of the combined chest and abdomen movement signals, as well as the maximum amplitude, signal amplitude area, and mean absolute value of the smart device vibration data, are extracted; a support vector classification model is trained based on these six-dimensional features to identify replay attackers. If it is detected that the deep breath does not come from a real person, the following steps will not be carried out.

[0041] S7. Extract 18-dimensional features from the combined chest and abdomen movement signals (breathing duration, inhalation time, exhalation time, inhalation and exhalation interval, inhalation and exhalation time ratio, breathing average value, inhalation average value, exhalation average value, inhalation and exhalation mean ratio, breathing root mean square, inhalation root mean square, exhalation root mean square, inhalation and exhalation root mean square ratio, inhalation kurtosis, exhalation kurtosis, breathing standard deviation, breathing shape factor, and breathing pulse factor). Extract the Bark spectrum of the breathing sound signal, and calculate the spectral centroid, spectral entropy, spectral roll-off point, and spectral spread of the Bark spectrum. Calculate the mutual information value of the root mean square envelopes of the combined chest and abdomen movement signals and the corresponding breathing sound signals at each time frame. The frame length is set to 1200, 2400, and 4800, and the frame shift is set to 1200.

[0042] S8. Design a multi-stream identity authentication model to support the input of the three-factor features of the combined chest and abdomen movement signals, breathing sound signals, and correlation sequences obtained in S7, and authenticate the user's identity. As Figure 2 shown, the network consists of two parts: feature extraction and identity authentication. In the feature extraction part, the features of the combined chest and abdomen movement signals pass through four fully connected layer networks; the feature sequences and mutual information sequences of the breathing sound signals pass through a combined network of a convolutional neural network (CNN) and a bidirectional LSTM. After the features of the three input streams are concatenated, they are further compressed through self-attention, CNN, and fully connected layers to extract the features related to identity authentication. Finally, classification is performed through the Softmax layer to output the user's identity ID.

[0043] In practical applications, the present invention can be embodied as a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to implement various aspects of the present invention.

Claims

1. A deep breathing identity authentication method based on sonar perception, characterized in that: The method comprises the following steps: S1. The user holds the smart device according to his / her preferences. The built-in speaker of the smart device transmits sonar signals with a preset modulation mode and frequency. At the same time, the built-in microphone of the smart device receives the sonar echo signals and breathing sound signals of the user's chest and abdominal movements. S2, denoising the received signal obtained in step S1 by using a high-pass filter to obtain a chest-abdomen combined motion signal; S3, calculating the energy sequence of the chest-abdomen joint motion signal obtained in step S2, and training the model to determine the start and end time endpoints of deep breathing; S4, determining the activity range of deep breathing in the received signal obtained in step S1 based on the deep breathing endpoint obtained in step S3; S5, filtering the received signal after endpoint processing in step S4 by high-pass filter and band-pass filter respectively to obtain a chest-abdomen joint motion signal and a corresponding breathing sound signal; S6. Training a liveness monitoring classifier based on the chest-abdomen joint motion signal and the vibration data of the smart device caused by hand gripping. The liveness monitoring classifier is used to identify whether the source of the deep breathing is a real person or a replay attacker. If the liveness monitoring classifier detects that the deep breathing is not from a real person, the next step will not be performed. The specific method for detecting replay attacks is as follows: first, synchronously obtain the chest-abdomen joint motion signal and use the built-in accelerometer of the smart device to capture the smart device vibration data caused by hand micro-movement; then extract the standard deviation, shape factor and pulse factor characteristics of the chest-abdomen joint motion signal and the maximum amplitude, signal amplitude area and average absolute value of the smart device vibration data; finally, input these six-dimensional features into the support vector classification model to identify the replay attacker; S7, extracting features from the chest-abdomen combined motion signal and the respiratory sound signal, calculating the correlation between the chest-abdomen combined motion signal and the corresponding respiratory sound signal, and obtaining a correlation sequence; S8. Design a multi-stream identity authentication model to support the input of the three-factor features of the chest-abdomen joint motion signal, breathing sound signal and correlation sequence obtained in step S7 to authenticate the user.

2. A deep breathing identity authentication method based on sonar perception as claimed in claim 1, characterized in that: The sonar signal emitted by the smart device in step S1 is in the form of a frequency modulated continuous wave with a frequency range of 20kHz-24kHz, wherein each ultrasonic chirp signal is 50ms.

3. A deep breathing identity authentication method based on sonar perception as claimed in claim 1, characterized in that: The method for determining the start and end time points of deep breathing in step S3 is specifically: calculating the short-time energy and short-time average amplitude sequence of the signal, and then inputting them into a sequence-to-sequence model based on a gated recurrent unit to determine the activity range of deep breathing.

4. A deep breathing identity authentication method based on sonar perception as claimed in claim 1, characterized in that: In step S5, the cut-off frequency of the high-pass filter is 20 kHz, and the cut-off frequency of the band-pass filter is set to 500-3000 Hz.

5. A deep breathing identity authentication method based on sonar perception as claimed in claim 1, characterized in that: The method for calculating the correlation in step S7 is to measure the chest-abdomen joint motion signal and the corresponding breathing sound signal through mutual information.

Images