A blockchain storage method and system based on audio authenticity identification technology

Abstract

The present application relates to the technical field of audio discrimination, and particularly relates to a blockchain storage method and system based on audio authenticity discrimination technology; an audio feature extraction model is used to extract audio features of training audio data; an audio authenticity discrimination model is constructed, and the audio authenticity discrimination model is trained by using the audio features of the training audio data; whether target audio data is real human voice is determined by using the trained audio authenticity discrimination model, and when the target audio data is real human voice, the target audio data is sent to an audio management server for storage, so that the safety of the stored audio data is ensured; storage code is sent to a blockchain server for storage, so that the storage code can be prevented from being tampered with; before the target audio data is accessed on a user terminal, authenticity verification of whether the target audio data is tampered with is required, so that the authenticity and safety of the target audio data are further ensured.

Description

Technical Field

[0001] This invention relates to the field of audio authentication technology, specifically to a blockchain-based evidence storage method and system for audio authenticity authentication technology. Background Technology

[0002] As society becomes increasingly information-driven, information storage solutions often face various cybersecurity risks and threats. Audio, as a crucial carrier of information, is particularly vulnerable to content security threats. For example, audio biometrics is widely used in identity verification scenarios, such as social media platforms like WeChat. However, with the development of AI (Artificial Intelligence) technology, an increasing amount of AI-generated audio information is emerging. This generated audio information can be highly convincing, severely reducing the reliability of audio as a biometric means. Therefore, there is an urgent need for a technological solution capable of identifying whether audio is AI-generated. Summary of the Invention

[0003] The main objective of this invention is to provide a blockchain-based evidence storage method and system based on audio authenticity verification technology, aiming to solve the current urgent need for a technical solution that can identify whether audio is generated by AI.

[0004] The technical solution proposed in this invention is as follows:

[0005] A blockchain-based evidence storage method based on audio authenticity verification technology is applied to a blockchain-based evidence storage system based on audio authenticity verification technology. The system includes an audio verification server, an audio management server, a blockchain server, and a user terminal. The audio verification server, blockchain server, and user terminal are all communicatively connected to the audio management server. The method includes:

[0006] The audio identification server constructs an audio feature extraction model and acquires training audio data. The audio feature extraction model then extracts the audio features from the training audio data.

[0007] The audio authentication server constructs an audio authenticity authentication model and trains the model by using the audio features of the training audio data.

[0008] The audio authentication server acquires the target audio data that needs to be authenticated, and uses the trained audio authentication model to determine whether the target audio data is real human voice audio.

[0009] If so, the audio identification server generates a unique evidence storage code corresponding to the target audio data, sends the evidence storage code to the blockchain server for storage, and sends the target audio data to the audio management server for storage;

[0010] The user terminal sends a request to the audio management server to access the target audio information;

[0011] The audio management server determines whether the user terminal has permission to access the target audio data based on the request information.

[0012] If the user has permission to access the target audio data, the audio management server will verify the authenticity of the target audio data and send the verified target audio data to the user terminal.

[0013] Preferably, the audio identification server constructs an audio feature extraction model and acquires training audio data, extracting audio features from the training audio data using the audio feature extraction model, including:

[0014] The audio identification server preprocesses the training audio data using an audio feature extraction model, including:

[0015] The audio feature extraction model applies a linear gain to the training audio data, thereby enhancing the amplitude of the high-frequency signals in the training audio data. The formula for calculating the applied linear gain is as follows:

[0016] ,

[0017] In the formula, The original speech signal at the nth sampling point of the training audio data; The speech signal at the nth sample point of the training audio data after applying a linear gain; This is the pre-emphasis coefficient;

[0018] The audio feature extraction model performs frame segmentation and windowing operations on the training audio data, including:

[0019] The audio feature extraction model determines the starting position of the first audio frame of the training audio data based on the preset frame length and preset frame shift, and uses it as the starting frame position;

[0020] The audio feature extraction model determines the boundary of each audio frame based on the starting frame position, preset frame length, and preset frame shift of the training audio data, and adds a window function at the boundary of each audio frame.

[0021] For subsequent audio frames of the training audio data, the audio feature extraction model adds a window function starting from the end position of the previous audio frame and extracts the next frame.

[0022] The audio feature extraction model performs audio frame extraction calculations on the training audio data based on a preset frame length and a preset frame shift;

[0023] The audio identification server normalizes each audio frame of the training audio data using an audio feature extraction model.

[0024] Preferably, the audio identification server preprocesses the training audio data using an audio feature extraction model, and then further includes:

[0025] The audio identification server extracts cepstral coefficients from the training audio data using an audio feature extraction model, including:

[0026] The audio feature extraction model performs a Fast Fourier Transform (FFT) on the preprocessed training audio data to convert the time-domain signal of the training audio data into a frequency-domain signal. The calculation formula for the FFT is as follows:

[0027] ,

[0028] In the formula, The frequency domain signal of the sampling points for training audio data; The frequency of the training audio data; The sampling points are the training audio data; N is the total number of sampling points in the training audio data.

[0029] The audio feature extraction model squares the frequency domain signal of the training audio data to obtain the power spectrum of the audio frame of the training audio data;

[0030] The audio feature extraction model takes the logarithm of the power spectrum to obtain the logarithmic power spectrum;

[0031] The audio feature extraction model performs discrete cosine transform on the logarithmic power spectrum to obtain the cepstral coefficients of the training audio data.

[0032] Preferably, the audio identification server extracts cepstral coefficients from the training audio data using an audio feature extraction model, and then further includes:

[0033] The audio identification server extracts linear frequencies from the training audio data using an audio feature extraction model, including:

[0034] The audio feature extraction model determines the frequency components present in the training audio data based on the energy value at the peak frequency of the logarithmic power spectrum. For each frequency component, the linear frequency of the training audio data is calculated using the following formula:

[0035] ,

[0036] In the formula, The linear frequencies of the training audio data; k is the index of the frequency component.

[0037] Preferably, the audio authenticity detection model includes a first Gaussian model and a second Gaussian model; the training audio data includes first audio data representing real human voices and second audio data representing fake human voices; the audio authentication server constructs the audio authenticity detection model and trains it using the audio features of the training audio data, including:

[0038] The audio identification server initializes the first Gaussian model and the second Gaussian model;

[0039] The audio identification server inputs the audio features of the first audio data into the first Gaussian model for training, and inputs the audio features of the second audio data into the second Gaussian model for training, so as to obtain the audio authenticity identification model that has been trained.

[0040] Preferably, the audio authentication server acquires the target audio data that needs to be authenticated, and determines whether the target audio data is genuine human voice audio by using a trained audio authentication model, including:

[0041] The audio authentication server inputs the target audio data into the trained audio authenticity authentication model to obtain the score output by the first Gaussian model. and the score output by the second Gaussian model. ;

[0042] The audio identification server calculates the score difference. ;

[0043] When the score difference If the value is greater than 0, the audio identification server determines that the target audio data is real human voice audio;

[0044] When the score difference If the value is less than 0, the audio identification server determines that the target audio data is fake human voice audio.

[0045] Preferably, the audio management server includes speakers; the audio identification server calculates the score difference. And then it includes:

[0046] When the score difference If the value is 0, the audio identification server marks the target audio data as audio to be identified and sends it to the audio management server;

[0047] The audio management server plays the audio to be identified through a speaker and obtains the human verification information corresponding to the input audio to be identified. The human verification information corresponding to the audio to be identified is sent to the audio identification server. The human verification information is used to indicate whether the audio to be identified is a real human voice audio.

[0048] The audio verification server determines whether the target audio data is genuine human voice audio based on manually verified information.

[0049] Preferably, the audio authentication server generates a unique evidence storage code corresponding to the target audio data, sends the evidence storage code to the blockchain server for storage, and sends the target audio data to the audio management server for storage, including:

[0050] The audio authentication server calculates the hash value of the target audio data based on a hash algorithm and uses the calculated hash value as the unique evidence storage code corresponding to the target audio data.

[0051] The audio authentication server sends the unique evidence code corresponding to the target audio data to the blockchain server for storage.

[0052] The audio authentication server obtains the authorized identity information corresponding to the target audio data, wherein the user terminal corresponding to the authorized identity information has the permission to access the corresponding target audio data;

[0053] The request information includes a request code and the current identity information of the user currently operating the user terminal. The audio management server determines whether the user terminal has permission to access the target audio data based on the request information, including:

[0054] When the blockchain server contains a notarized code that matches the request code in the request information, and the current identity information in the request information is included in the authorized identity information corresponding to the target audio data corresponding to the request code, the audio management server determines that the user terminal has the right to access the target audio data.

[0055] If the user has permission to access the target audio data, the audio management server verifies the authenticity of the target audio data and sends the verified target audio data to the user terminal, including:

[0056] If the server has permission to access the target audio data, it will mark the target audio data corresponding to the request code in the request information as data to be accessed.

[0057] The audio management server calculates the hash value of the data to be accessed based on a hash algorithm and determines whether the hash value of the data to be accessed is consistent with the evidence storage code corresponding to the data to be accessed.

[0058] If they match, the audio management server determines that the data to be accessed has passed the authenticity verification and sends the data to the user terminal.

[0059] If there is a discrepancy, the audio management server determines that the data to be accessed has not passed the authenticity verification and prohibits any user terminal from accessing the data to be accessed.

[0060] Preferably, the audio authentication server generates a unique evidence storage code corresponding to the target audio data, sends the evidence storage code to the blockchain server for storage, and sends the target audio data to the audio management server for storage, further comprising:

[0061] The audio identification server obtains the total number of audio frames contained in the target audio data and randomly generates a preset number of distinct selected values ​​corresponding to the target audio data. ,in, This indicates the i-th selected value, where the selected value is less than the total number of audio frames contained in the target audio data;

[0062] The audio identification server arranges the audio frames in the target audio data according to their chronological order, and then sequentially identifies the first audio frame in the target audio data. Each audio frame is muted, where 1≤i≤I, and I is a preset number;

[0063] The audio authentication server calculates the hash value of the target audio data after the noise reduction process based on a hash algorithm, and uses the calculated hash value as the unique evidence storage code corresponding to the target audio data.

[0064] The audio authentication server establishes a correspondence between the evidence storage code corresponding to the same target audio data and a preset number of selected values, and then sends them together to the blockchain server for storage.

[0065] If the user has permission to access the target audio data, the audio management server verifies the authenticity of the target audio data and sends the verified target audio data to the user terminal. The method also includes:

[0066] If the server has permission to access the target audio data, it will mark the target audio data corresponding to the request code in the request information as data to be accessed.

[0067] The audio management server obtains and requests the selected value corresponding to the certificate number that matches the encoding from the blockchain server. After arranging the audio frames in the target audio data in chronological order, the data to be accessed is then sequentially judged. Is the sound intensity of each audio frame 0?

[0068] If so, the audio management server determines that the data to be accessed has passed the authenticity verification and sends the data to the user terminal;

[0069] If not, the audio management server determines that the data to be accessed has failed the authenticity verification and prohibits any user terminal from accessing the data.

[0070] This invention also proposes a blockchain-based evidence storage system based on audio authenticity verification technology, and applies a blockchain-based evidence storage method based on audio authenticity verification technology; the system includes an audio verification server, an audio management server, a blockchain server, and a user terminal; the audio verification server, the blockchain server, and the user terminal are all communicatively connected to the audio management server.

[0071] The above technical solution can achieve the following beneficial effects:

[0072] This invention proposes a blockchain-based evidence storage method based on audio authenticity verification technology, providing a technical solution for identifying whether audio is AI-generated. First, an audio feature extraction model is constructed to extract audio features from training audio data. Then, an audio authenticity verification model is built and trained using the audio features from the training audio data. Afterward, the trained audio authenticity verification model determines whether the target audio data is genuine human voice audio. Only when the target audio data is confirmed to be genuine human voice audio is it sent to the audio management server for storage, thus ensuring the security of the stored audio data. The evidence storage code uniquely corresponds to the target audio data, and the evidence storage code is sent to the blockchain server for storage, preventing tampering and ensuring its uniqueness and security. Subsequently, before a user terminal accesses the target audio data, it needs to undergo authenticity verification to ensure it has not been tampered with. Only target audio data that passes authenticity verification is sent to the user terminal. In other words, this solution performs two verifications on the target audio data: genuine human voice verification before storage, and another authenticity verification when the user terminal accesses the target audio data, further ensuring the authenticity and security of the target audio data. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0074] Figure 1 This is a flowchart of the steps in the first embodiment of a blockchain evidence storage method based on audio authenticity verification technology proposed in this invention.

[0075] Figure 2 This is a schematic diagram of a blockchain-based evidence storage system based on audio authenticity verification technology proposed in this invention. Detailed Implementation

[0076] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0077] This invention proposes a blockchain-based evidence storage method and system based on audio authenticity verification technology.

[0078] As attached Figure 1-2 As shown, in the first embodiment of the blockchain evidence storage method based on audio authenticity verification technology proposed in this invention, the method is applied to a blockchain evidence storage system based on audio authenticity verification technology; the system includes an audio verification server, an audio management server, a blockchain server, and a user terminal; the audio verification server, the blockchain server, and the user terminal are all communicatively connected to the audio management server; this embodiment includes the following steps:

[0079] Step S110: The audio identification server constructs an audio feature extraction model (specifically, an audio feature extraction model based on a linear frequency scale), acquires training audio data, and extracts audio features from the training audio data through the audio feature extraction model.

[0080] Specifically, this step proposes an audio feature extraction model based on a linear frequency scale. This model segments the audio file into multiple frames at the frame bit level for frequency and cepstral calculation, providing higher frequency resolution, better preservation of the time-domain signal, and more accurate capture of the frequency components of the sound signal. This helps improve the accuracy of audio feature extraction and demonstrates better robustness when dealing with noise and distortion.

[0081] Step S120: The audio authentication server constructs an audio authenticity authentication model (specifically, an audio authenticity authentication model based on a Gaussian mixture model), and trains the audio authenticity authentication model by using the audio features of the training audio data.

[0082] Specifically, this step proposes an audio authenticity detection model based on Gaussian Mixture Models (GMMs). GMMs can flexibly represent complex probability distributions, and by adjusting the parameters of the Gaussian distribution, they can adapt to different types of sound feature distributions. This model trains two Gaussian models using processed fake audio features and real audio features. These two models are designed to identify fake and real audio respectively, thereby achieving audio authentication. This model demonstrates high accuracy in audio authenticity detection.

[0083] Step S130: The audio identification server obtains the target audio data that needs to be identified as genuine or fake, and determines whether the target audio data is real human voice audio by using the trained audio authenticity identification model.

[0084] Specifically, the security of the target audio data is determined by judging whether it is real human voice audio. In addition to human voice audio, the audio data may also be AI-generated simulated human voice audio. In order to ensure the reliability of audio as a biometric means, it is necessary to judge the authenticity of the target audio data, that is, to judge whether the target audio data is generated by AI.

[0085] If so, proceed to step S140: The audio identification server generates a unique evidence storage code corresponding to the target audio data, sends the evidence storage code to the blockchain server for storage, and sends the target audio data to the audio management server for storage.

[0086] Specifically, the target audio data is only sent to the audio management server for storage after it is confirmed that the target audio data is indeed real human voice audio, thereby ensuring the security of the stored audio data; the evidence storage code is uniquely associated with the target audio data, and the evidence storage code is sent to the blockchain server for storage, which can prevent the evidence storage code from being tampered with, thereby ensuring the uniqueness and security of the evidence storage code.

[0087] Step S150: The user terminal sends a request to the audio management server to access the target audio information.

[0088] Step S160: The audio management server determines whether the user terminal has permission to access the target audio data based on the request information.

[0089] Step S170: If the user has permission to access the target audio data, the audio management server verifies the authenticity of the target audio data and sends the target audio data that has passed the authenticity verification to the user terminal.

[0090] Specifically, before a user terminal accesses the target audio data, the target audio data needs to be verified for authenticity (whether it has been tampered with), and only the target audio data that passes the authenticity verification is sent to the user terminal; that is, this solution performs two verifications on the target audio data: before storage, it performs real human voice identification, and only when the target audio data is a real human voice is it stored; the user terminal performs another authenticity verification when accessing the target audio data; the two verifications can further ensure the authenticity and security of the target audio data.

[0091] This invention proposes a blockchain-based evidence storage method based on audio authenticity verification technology, providing a technical solution for identifying whether audio is AI-generated. First, an audio feature extraction model is constructed to extract audio features from training audio data. Then, an audio authenticity verification model is built and trained using the audio features from the training audio data. Afterward, the trained audio authenticity verification model determines whether the target audio data is genuine human voice audio. Only when the target audio data is confirmed to be genuine human voice audio is it sent to the audio management server for storage, thus ensuring the security of the stored audio data. The evidence storage code uniquely corresponds to the target audio data, and the evidence storage code is sent to the blockchain server for storage, preventing tampering and ensuring its uniqueness and security. Subsequently, before a user terminal accesses the target audio data, it needs to undergo authenticity verification to ensure it has not been tampered with. Only target audio data that passes authenticity verification is sent to the user terminal. In other words, this solution performs two verifications on the target audio data: genuine human voice verification before storage, and another authenticity verification when the user terminal accesses the target audio data, further ensuring the authenticity and security of the target audio data.

[0092] Furthermore, this invention presents an audio feature extraction model based on a linear frequency scale. This model utilizes the LFCC (Linear Frequency Filter Bank) audio feature extraction algorithm, replacing the Mel filter bank in MFCC (Mel Frequency Cepstral Coefficients) with a linearly distributed filter bank to achieve audio feature extraction. It also presents an audio authenticity verification model based on a Gaussian mixture model, capable of capturing subtle differences between AI-generated fake human voice audio and genuine human voice audio, thus distinguishing between AI-generated audio and genuine human audio. Simultaneously, this invention proposes a blockchain-based audio authenticity verification system framework. The verified audio data is stored as evidence via the blockchain, leveraging the traceability and tamper-proof characteristics of on-chain data to ensure the security of the authentic audio and facilitate subsequent retrieval.

[0093] In the second embodiment of the blockchain evidence storage method based on audio authenticity verification technology proposed in this invention, based on the first embodiment, step S110 includes the following steps:

[0094] Step S210: The audio discrimination server preprocesses the training audio data using an audio feature extraction model, including the following steps:

[0095] Step S211: The audio feature extraction model applies a linear gain to the training audio data, enhancing the amplitude of the high-frequency signals to compensate for the attenuation of high-frequency signals in the vocal tract. The aim is to improve the clarity and intelligibility of the speech signal (especially in the high-frequency range). The formula for calculating the linear gain is as follows:

[0096] ,

[0097] In the formula, The original speech signal of the nth sampling point of the training audio data is given, where n is the sampling point of the training audio data. The speech signal at the nth sample point of the training audio data after applying a linear gain; This is the pre-emphasis coefficient, which is usually taken as 0.97.

[0098] Step S220: The audio feature extraction model performs frame segmentation and windowing operations on the training audio data, including the following steps:

[0099] Step S221: The audio feature extraction model determines the starting position of the first audio frame of the training audio data based on the preset frame length and preset frame shift, and uses it as the starting frame position; usually, the starting position of the first frame is randomly selected, or optimized according to the characteristics of the speech signal.

[0100] Specifically, framing refers to dividing a continuous speech signal into a series of short and discrete signal segments for subsequent feature extraction and analysis. Framing involves two parameters: a preset frame length and a preset frame shift. The frame length refers to the number of sampling points in each frame, typically set to 20-40 milliseconds, depending on the speech signal's sampling rate and processing requirements. Shorter frame lengths capture better speech signal details but increase computational complexity. Frame shift refers to the overlap between consecutive frames, typically set to 50%-70%, meaning each new frame overlaps with the previous one by 50%-70%. Smaller frame shifts better capture the continuity of the speech signal. Windowing involves adding a window function (in this embodiment, a Hamming window) at the boundary of each audio frame to reduce energy and spectral leakage between frames. Windowing improves the signal's spectral characteristics, making the signal's spectrum more uniform and richer.

[0101] Step S222: The audio feature extraction model determines the boundary of each audio frame based on the starting frame position, preset frame length, and preset frame shift of the training audio data, and adds a window function at the boundary of each audio frame.

[0102] Step S223: For subsequent audio frames of the training audio data, the audio feature extraction model adds a window function starting from the end position of the previous audio frame and extracts the next frame. Each new audio frame maintains a 50% to 70% overlap with the previous audio frame to ensure continuity.

[0103] Step S224: The audio feature extraction model performs audio frame extraction calculations on the training audio data according to the preset frame length and preset frame shift. Since the length of each audio frame is fixed, each extracted frame is a short-time, discrete signal segment, which is convenient for subsequent feature extraction and analysis.

[0104] Step S230: The audio identification server normalizes each audio frame of the training audio data through the audio feature extraction model to reduce the amplitude difference between different frames, thereby improving the stability and robustness of feature extraction and analysis. Commonly used normalization methods include linear normalization and logarithmic normalization.

[0105] In the third embodiment of the blockchain evidence storage method based on audio authenticity authentication technology proposed in this invention, based on the second embodiment, after step S210, the following steps are further included:

[0106] Step S310: The audio identification server extracts cepstral coefficients from the training audio data using an audio feature extraction model, including the following steps:

[0107] Step S311: The audio feature extraction model performs a Fast Fourier Transform (FFT) on the preprocessed training audio data to convert the time-domain signal of the training audio data into a frequency-domain signal. The calculation formula for the FFT is as follows:

[0108] ,

[0109] In the formula, The frequency domain signal of the sampling points for training audio data; The frequency of the training audio data; represents the sampling points of the training audio data; N represents the total number of sampling points of the training audio data, i.e., the length of the training audio data.

[0110] Step S312: The audio feature extraction model squares the frequency domain signal of the training audio data to obtain the power spectrum of the audio frame of the training audio data.

[0111] Specifically, the power spectrum can reflect the energy distribution of a frame.

[0112] Step S313: The audio feature extraction model takes the logarithm of the power spectrum to obtain the logarithmic power spectrum.

[0113] Specifically, the logarithmic power spectrum is obtained, thereby reducing the dynamic range and enhancing the stability of signals of different amplitudes.

[0114] Step S314: The audio feature extraction model performs discrete cosine transform on the logarithmic power spectrum to obtain the cepstral coefficients of the training audio data.

[0115] Specifically, the Discrete Cosine Transform (DFT) is a transform related to the Fourier Transform. It is similar to the Discrete Fourier Transform (DFT) but uses only real numbers; the cepstral coefficients reflect the frequency distribution of the signal.

[0116] In the fourth embodiment of the blockchain evidence storage method based on audio authenticity authentication technology proposed in this invention, based on the third embodiment, after step S310, the following steps are further included:

[0117] Step S410: The audio discrimination server performs linear frequency extraction on the training audio data using an audio feature extraction model, including the following steps:

[0118] Step S411: The audio feature extraction model determines the frequency components present in the training audio data based on the energy value at the peak frequency (or a specific frequency) on the logarithmic power spectrum. For each frequency component, the linear frequency of the training audio data is calculated using the following formula:

[0119] ,

[0120] In the formula, The linear frequencies of the training audio data; k is the index of the frequency component.

[0121] Specifically, the cepstral coefficients obtained by solving using linear frequencies are the extraction results of the linear frequency filter bank. They can reflect the frequency distribution and linear probability information of the signal and are an important feature of audio signals.

[0122] In the fifth embodiment of the blockchain evidence storage method based on audio authenticity verification technology proposed in this invention, based on the fourth embodiment, the audio authenticity verification model includes a first Gaussian model and a second Gaussian model; the training audio data includes first audio data of real human voices and second audio data of fake human voices, and the number of first audio data and second audio data is multiple; step S120 includes the following steps:

[0123] Step S510: The audio identification server initializes the first Gaussian model and the second Gaussian model.

[0124] Specifically, the purpose of initializing the first and second Gaussian models is to provide a reasonable starting point so that the expectation maximization algorithm can iterate from this starting point and eventually converge to the global optimum. Before formal training, 10% of the first audio data is selected to initialize the first Gaussian model, and 10% of the second audio data is selected to initialize the second Gaussian model.

[0125] Step S520: The audio identification server inputs the audio features of the first audio data into the first Gaussian model for training, and inputs the audio features of the second audio data into the second Gaussian model for training, so as to obtain the audio authenticity identification model that has been trained.

[0126] Specifically, during training, the Expectation-Maximization (EM) algorithm is used iteratively, with a maximum of 100 iterations for both the first and second Gaussian models. The covariance matrices of both models are diagonal covariance matrices. Considering the difference in the number of features between real and fake audio (real human voices are more stable than AI-generated sounds, resulting in fewer features), the Gaussian distribution number is set to 512 for the first model and 768 for the second. After training, the first Gaussian model can recognize real human voices, while the second Gaussian model can recognize fake audio.

[0127] In the sixth embodiment of the blockchain evidence storage method based on audio authenticity authentication technology proposed in this invention, based on the fifth embodiment, step S130 includes the following steps:

[0128] Step S610: The audio identification server inputs the target audio data into the trained audio authenticity identification model to obtain the score value output by the first Gaussian model. and the score output by the second Gaussian model. .

[0129] Step S620: The audio identification server calculates the score difference. .

[0130] Step S630: When the score difference If the value is greater than 0, the audio identification server determines that the target audio data is real human voice audio.

[0131] Step S640: When the score difference If the value is less than 0, the audio identification server determines that the target audio data is fake human voice audio.

[0132] Specifically, because the first Gaussian model has the ability to recognize real human voice audio, the first Gaussian model outputs a score value for the target audio data. The higher the score, the greater the probability that the target audio data is a real human voice; similarly, the higher the score of the second Gaussian model for the target audio data... The higher the score, the greater the probability that the target audio data is a fake human voice (e.g., an AI-generated voice); therefore, by comparing the score differences... This allows for the judgment of the authenticity of the audio; if the score difference is significant... A score greater than 0 indicates that the first Gaussian model has a higher similarity to the target audio data, and the target audio data is presumed to be real human voice audio; if the score difference is... A value less than 0 indicates that the second Gaussian model has a higher similarity to the target audio data, and the target audio data is presumed to be fake or forged audio.

[0133] In the seventh embodiment of the blockchain evidence storage method based on audio authenticity authentication technology proposed in this invention, based on the sixth embodiment, the audio management server includes a speaker; step S620, followed by the following steps:

[0134] Step S710: When the score difference If the value is 0, the audio identification server marks the target audio data as audio to be identified and sends it to the audio management server.

[0135] Step S720: The audio management server plays the audio to be identified through the speaker and obtains the manual verification information corresponding to the input audio to be identified. The manual verification information corresponding to the audio to be identified is sent to the audio identification server. The manual verification information is used to indicate whether the audio to be identified is a real human voice audio.

[0136] Step S730: The audio identification server determines whether the target audio data is real human voice audio based on the manual verification information.

[0137] Specifically, when the score difference A value of 0 indicates that the audio authenticity identification model has not been able to clearly identify whether the target audio data is a real human voice. It is necessary to listen to the target audio data manually and then make a judgment to determine whether the target audio data is a real human voice audio.

[0138] In the eighth embodiment of the blockchain evidence storage method based on audio authenticity authentication technology proposed in this invention, based on the sixth embodiment, step S140 includes the following steps:

[0139] Step S810: The audio authentication server calculates the hash value of the target audio data based on a hash algorithm, and uses the calculated hash value as the unique evidence storage code corresponding to the target audio data.

[0140] Specifically, a hash algorithm can convert a piece of data into a flag value (i.e., a hash value), and this flag value is closely related to every byte of the source data. Hash algorithms also have the characteristic that it is difficult to find a reverse pattern.

[0141] Step S820: The audio authentication server sends the unique evidence code corresponding to the target audio data to the blockchain server for storage.

[0142] Specifically, after the evidence storage code is sent to the blockchain for storage, it can be prevented from being tampered with, ensuring the authenticity and security of the evidence storage code.

[0143] Step S830: The audio identification server obtains the authorized identity information corresponding to the target audio data. The user terminal corresponding to the authorized identity information has the right to access the corresponding target audio data, and the authorized identity information is manually entered by the administrator after being determined.

[0144] Specifically, after sending the target audio data to the audio management server for storage, the audio identification server will also obtain the authorized identity information corresponding to the target audio data. Through the authorized identity information, it can know which user terminals can access the stored target audio data, and prepare for verification when user terminals access the target audio data in the future.

[0145] In addition, the request information includes a request code and the current identity information of the user currently operating the user terminal. Step S160 includes the following steps:

[0146] Step S840: When there is a certificate of authenticity code in the blockchain server that is consistent with the request code in the request information, and the current identity information in the request information is included in the authorized identity information corresponding to the target audio data corresponding to the request code, the audio management server determines that the user terminal has the right to access the target audio data.

[0147] Step S170 includes the following steps:

[0148] Step S850: If the server has permission to access the target audio data, the audio management server marks the target audio data corresponding to the request code in the request information as data to be accessed.

[0149] Step S860: The audio management server calculates the hash value of the data to be accessed based on a hash algorithm, and determines whether the hash value of the data to be accessed is consistent with the evidence storage code corresponding to the data to be accessed.

[0150] Specifically, under normal circumstances, the hash value of the data to be accessed should be consistent with the storage code corresponding to the data to be accessed. If they are inconsistent, it proves that the data to be accessed has been tampered with or damaged during the storage of the audio management server and is no longer the original audio data, that is, it has not passed the authenticity verification.

[0151] Step S870: If they match, the audio management server determines that the data to be accessed has passed the authenticity verification and sends the data to be accessed to the user terminal.

[0152] Step S880: If there is a discrepancy, the audio management server determines that the data to be accessed has failed the authenticity verification and prohibits any user terminal from accessing the data to be accessed.

[0153] Step S170 further includes the following steps:

[0154] Step S891: If the server has permission to access the target audio data, the audio management server marks the target audio data corresponding to the request code in the request information as data to be accessed.

[0155] Step S892: The audio management server inputs the data to be accessed into the trained audio authenticity identification model to obtain the score value output by the first Gaussian model, which is then marked as the first verification value. The score output by the second Gaussian model is obtained and marked as the second validation value. .

[0156] Step S893: The audio management server obtains the score value output by the first Gaussian model, which is obtained by pre-verification of the data to be accessed through the audio authenticity identification model before storing it on the audio management server. and the score output by the second Gaussian model. .

[0157] Step S894: The audio management server determines whether the following conditions are met: First verification value The score output by the first Gaussian model Consistent, and the second verification value The score output by the second Gaussian model Consistent.

[0158] If so, proceed to step S895: The audio management server determines that the data to be accessed has passed the authenticity verification and sends the data to be accessed to the user terminal.

[0159] If not, proceed to step S896: The audio management server determines that the data to be accessed has failed the authenticity verification and prohibits any user terminal from accessing the data to be accessed.

[0160] In this embodiment, the data to be accessed is verified again using an audio authenticity verification model based on a Gaussian model to obtain the first verification value. Second verification value If the data to be accessed has not been tampered with during storage, then the following should be satisfied: First verification value The score output by the first Gaussian model Consistent, and the second verification value The score output by the second Gaussian model Consistency; therefore, based on this, the authenticity of the data to be accessed can be verified to see if it has been tampered with (i.e., secondary verification), further improving data security.

[0161] In the ninth embodiment of the blockchain evidence storage method based on audio authenticity authentication technology proposed in this invention, based on the first embodiment, step S140 further includes the following steps:

[0162] Step S910: The audio identification server obtains the total number of audio frames contained in the target audio data, and randomly generates a preset number (e.g., 3) of distinct selected values ​​corresponding to the target audio data. ,in, This indicates the i-th selected value, which is less than the total number of audio frames contained in the target audio data.

[0163] Specifically, for example in this embodiment, the total number of audio frames of the target audio data is 600, and the three randomly generated selected values ​​are 23, 146, and 278.

[0164] Step S920: After arranging the audio frames in the target audio data according to their chronological order, the audio identification server sequentially processes the audio frames in the target audio data. Each audio frame is muted, where 1 ≤ i ≤ I, and I is a preset number.

[0165] Specifically, in this embodiment, the 23rd, 146th, and 278th audio frames in the target audio data are muted sequentially.

[0166] Step S930: The audio authentication server calculates the hash value of the target audio data after the noise reduction process based on the hash algorithm, and uses the calculated hash value as the unique evidence storage code corresponding to the target audio data.

[0167] Step S940: The audio authentication server establishes a correspondence between the evidence storage code corresponding to the same target audio data and a preset number of selected values, and then sends them together to the blockchain server for storage.

[0168] Specifically, the evidence storage code and these three selected values ​​are stored in the blockchain server at the same time.

[0169] Step S170 further includes the following steps:

[0170] Step S950: If the server has permission to access the target audio data, the audio management server marks the target audio data corresponding to the request code in the request information as data to be accessed.

[0171] Step S960: The audio management server obtains the selected value corresponding to the certificate number that matches the requested encoding from the blockchain server. After arranging the audio frames in the target audio data in chronological order, the data to be accessed is then sequentially judged. Is the sound intensity of each audio frame 0?

[0172] Specifically, under normal circumstances (assuming the data to be accessed is stored on the audio management server without being tampered with or damaged), after arranging the audio frames in the target audio data in chronological order, the sound intensity of the 23rd, 146th, and 278th audio frames in the data to be accessed should all be 0. If this condition is not met, it means that the data to be accessed has been tampered with or damaged during storage on the audio management server, and is no longer the original audio data, i.e., it has not passed the authenticity verification.

[0173] If so, proceed to step S970: The audio management server determines that the data to be accessed has passed the authenticity verification and sends the data to be accessed to the user terminal.

[0174] If not, proceed to step S980: The audio management server determines that the data to be accessed has failed the authenticity verification and prohibits any user terminal from accessing the data to be accessed.

[0175] Specifically, this embodiment provides another technical solution for verifying the authenticity of target audio data.

[0176] This invention also proposes a blockchain-based evidence storage system based on audio authenticity verification technology, and applies a blockchain-based evidence storage method based on audio authenticity verification technology; the system includes an audio verification server, an audio management server, a blockchain server, and a user terminal; the audio verification server, the blockchain server, and the user terminal are all communicatively connected to the audio management server.

[0177] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0178] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A blockchain-based evidence storage method based on audio authenticity verification technology, characterized in that, A blockchain-based evidence storage system for audio authenticity verification technology is applied; the system includes an audio verification server, an audio management server, a blockchain server, and a user terminal; The audio authentication server, blockchain server, and user terminal are all communicatively connected to the audio management server; the method includes: The audio identification server constructs an audio feature extraction model and acquires training audio data. The audio feature extraction model then extracts the audio features from the training audio data. The audio authentication server constructs an audio authenticity authentication model and trains the model by using the audio features of the training audio data. The audio authentication server acquires the target audio data that needs to be authenticated, and uses the trained audio authentication model to determine whether the target audio data is real human voice audio. If so, the audio identification server generates a unique evidence storage code corresponding to the target audio data, sends the evidence storage code to the blockchain server for storage, and sends the target audio data to the audio management server for storage; The user terminal sends a request to the audio management server to access the target audio information; The audio management server determines whether the user terminal has permission to access the target audio data based on the request information. If the user has permission to access the target audio data, the audio management server will verify the authenticity of the target audio data and send the verified target audio data to the user terminal. The request information includes a request code and the current identity information of the user currently operating the user terminal. The audio management server determines whether the user terminal has permission to access the target audio data based on the request information, including: When the blockchain server contains a notarized code that matches the request code in the request information, and the current identity information in the request information is included in the authorized identity information corresponding to the target audio data corresponding to the request code, the audio management server determines that the user terminal has the right to access the target audio data. If the user has permission to access the target audio data, the audio management server verifies the authenticity of the target audio data and sends the verified target audio data to the user terminal, including: If the server has permission to access the target audio data, it will mark the target audio data corresponding to the request code in the request information as data to be accessed. The audio management server calculates the hash value of the data to be accessed based on a hash algorithm and determines whether the hash value of the data to be accessed is consistent with the evidence storage code corresponding to the data to be accessed. If they match, the audio management server determines that the data to be accessed has passed the authenticity verification and sends the data to the user terminal. If there is a discrepancy, the audio management server determines that the data to be accessed has not passed the authenticity verification and prohibits any user terminal from accessing the data to be accessed.

2. The blockchain-based evidence storage method based on audio authenticity verification technology according to claim 1, characterized in that, The audio identification server constructs an audio feature extraction model and acquires training audio data. The audio feature extraction model then extracts audio features from the training audio data, including: The audio identification server preprocesses the training audio data using an audio feature extraction model, including: The audio feature extraction model applies a linear gain to the training audio data, thereby enhancing the amplitude of the high-frequency signals in the training audio data. The formula for calculating the applied linear gain is as follows: ， In the formula, The original speech signal at the nth sampling point of the training audio data; The speech signal at the nth sample point of the training audio data after applying a linear gain; This is the pre-emphasis coefficient; The audio feature extraction model performs frame segmentation and windowing operations on the training audio data, including: The audio feature extraction model determines the starting position of the first audio frame of the training audio data based on the preset frame length and preset frame shift, and uses it as the starting frame position; The audio feature extraction model determines the boundary of each audio frame based on the starting frame position, preset frame length, and preset frame shift of the training audio data, and adds a window function at the boundary of each audio frame. For subsequent audio frames of the training audio data, the audio feature extraction model adds a window function starting from the end position of the previous audio frame and extracts the next frame. The audio feature extraction model performs audio frame extraction calculations on the training audio data based on a preset frame length and a preset frame shift; The audio identification server normalizes each audio frame of the training audio data using an audio feature extraction model.

3. The blockchain-based evidence storage method based on audio authenticity verification technology according to claim 2, characterized in that, The audio identification server preprocesses the training audio data using an audio feature extraction model, and then further includes: The audio identification server extracts cepstral coefficients from the training audio data using an audio feature extraction model, including: The audio feature extraction model performs a Fast Fourier Transform (FFT) on the preprocessed training audio data to convert the time-domain signal of the training audio data into a frequency-domain signal. The calculation formula for the FFT is as follows: ， In the formula, The frequency domain signal of the sampling points for training audio data; The frequency of the training audio data; The sampling points are the training audio data; N is the total number of sampling points in the training audio data. The audio feature extraction model squares the frequency domain signal of the training audio data to obtain the power spectrum of the audio frame of the training audio data; The audio feature extraction model takes the logarithm of the power spectrum to obtain the logarithmic power spectrum; The audio feature extraction model performs discrete cosine transform on the logarithmic power spectrum to obtain the cepstral coefficients of the training audio data.

4. A blockchain-based evidence storage method based on audio authenticity verification technology according to claim 3, characterized in that, The audio identification server extracts cepstral coefficients from the training audio data using an audio feature extraction model, and then further includes: The audio identification server extracts linear frequencies from the training audio data using an audio feature extraction model, including: The audio feature extraction model determines the frequency components present in the training audio data based on the energy value at the peak frequency of the logarithmic power spectrum. For each frequency component, the linear frequency of the training audio data is calculated using the following formula: ， In the formula, The linear frequencies of the training audio data; k is the index of the frequency component.

5. A blockchain-based evidence storage method based on audio authenticity verification technology according to claim 4, characterized in that, The audio authenticity detection model includes a first Gaussian model and a second Gaussian model; the training audio data includes first audio data representing real human voices and second audio data representing fake human voices; the audio authentication server constructs the audio authenticity detection model and trains it using the audio features of the training audio data, including: The audio identification server initializes the first Gaussian model and the second Gaussian model; The audio identification server inputs the audio features of the first audio data into the first Gaussian model for training, and inputs the audio features of the second audio data into the second Gaussian model for training, so as to obtain the audio authenticity identification model that has been trained.

6. A blockchain-based evidence storage method based on audio authenticity verification technology according to claim 5, characterized in that, The audio authentication server acquires the target audio data that needs to be authenticated, and determines whether the target audio data is genuine human voice audio using a trained audio authentication model, including: The audio authentication server inputs the target audio data into the trained audio authenticity authentication model to obtain the score output by the first Gaussian model. and the score output by the second Gaussian model. ; The audio identification server calculates the score difference. ; When the score difference If the value is greater than 0, the audio identification server determines that the target audio data is real human voice audio; When the score difference If the value is less than 0, the audio identification server determines that the target audio data is fake human voice audio.

7. A blockchain-based evidence storage method based on audio authenticity verification technology according to claim 6, characterized in that, The audio management server includes speakers; the audio identification server calculates the score difference. And then it includes: When the score difference If the value is 0, the audio identification server marks the target audio data as audio to be identified and sends it to the audio management server; The audio management server plays the audio to be identified through a speaker and obtains the human verification information corresponding to the input audio to be identified. The human verification information corresponding to the audio to be identified is sent to the audio identification server. The human verification information is used to indicate whether the audio to be identified is a real human voice audio. The audio verification server determines whether the target audio data is genuine human voice audio based on manually verified information.

8. A blockchain-based evidence storage method based on audio authenticity verification technology according to claim 6, characterized in that, The audio authentication server generates a unique evidence storage code corresponding to the target audio data, and sends the evidence storage code to the blockchain server for storage. The target audio data is then sent to the audio management server for storage, including: The audio authentication server calculates the hash value of the target audio data based on a hash algorithm and uses the calculated hash value as the unique evidence storage code corresponding to the target audio data. The audio authentication server sends the unique evidence code corresponding to the target audio data to the blockchain server for storage. The audio authentication server obtains the authorized identity information corresponding to the target audio data, whereby the user terminal corresponding to the authorized identity information has the permission to access the corresponding target audio data.

9. A blockchain-based evidence storage method based on audio authenticity verification technology according to claim 1, characterized in that, The audio authentication server generates a unique evidence storage code corresponding to the target audio data, sends the evidence storage code to the blockchain server for storage, and sends the target audio data to the audio management server for storage. The system also includes: The audio identification server obtains the total number of audio frames contained in the target audio data and randomly generates a preset number of distinct selected values ​​corresponding to the target audio data. ,in, This indicates the i-th selected value, where the selected value is less than the total number of audio frames contained in the target audio data; The audio identification server arranges the audio frames in the target audio data according to their chronological order, and then sequentially identifies the first audio frame in the target audio data. Each audio frame is muted, where 1≤i≤I, and I is a preset number; The audio authentication server calculates the hash value of the target audio data after the noise reduction process based on a hash algorithm, and uses the calculated hash value as the unique evidence storage code corresponding to the target audio data. The audio authentication server establishes a correspondence between the evidence storage code corresponding to the same target audio data and a preset number of selected values, and then sends them together to the blockchain server for storage. If the user has permission to access the target audio data, the audio management server verifies the authenticity of the target audio data and sends the verified target audio data to the user terminal. The method also includes: If the server has permission to access the target audio data, it will mark the target audio data corresponding to the request code in the request information as data to be accessed. The audio management server obtains and requests the selected value corresponding to the certificate number that matches the encoding from the blockchain server. After arranging the audio frames in the target audio data in chronological order, the data to be accessed is then sequentially judged. Is the sound intensity of each audio frame 0? If so, the audio management server determines that the data to be accessed has passed the authenticity verification and sends the data to the user terminal; If not, the audio management server determines that the data to be accessed has failed the authenticity verification and prohibits any user terminal from accessing the data.

10. A blockchain-based evidence storage system based on audio authenticity verification technology, characterized in that, The blockchain-based evidence storage method based on audio authenticity verification technology as described in any one of claims 1-9 is described below; the system includes an audio verification server, an audio management server, a blockchain server, and a user terminal; the audio verification server, the blockchain server, and the user terminal are all communicatively connected to the audio management server.

Images