Financial institution business transaction authentication system based on audio authentication

Abstract

The invention discloses a financial institution business transaction authentication system based on audio authentication, which is characterized in that an audio acquisition and preprocessing module acquires user voice sample audio, and a voiceprint feature extraction and recognition module extracts multi-dimensional voiceprint features from the preprocessed audio by using a deep convolutional neural network; calculating a similarity score of the real-time voiceprint and the registered voiceprint template through a feature vector comparison algorithm; and the dynamic authentication decision module dynamically combines the authentication factors according to the voiceprint matching score and the risk score, adjusts the authentication intensity according to the real-time risk level, and triggers different levels of security authentication processes. According to the invention, through multi-modal risk assessment combining acoustic environment analysis and behavior feature analysis, fraudulent behaviors can be comprehensively identified; the dynamic authentication decision can flexibly adjust the authentication process and factor combination according to the real-time risk to balance the safety and experience; and the authentication knowledge base continuously optimizes the risk assessment model to ensure the recognition capability of the system on the novel fraud mode.

Description

Technical Field

[0001] This invention relates to the field of secure transaction authentication systems for financial institutions, and in particular to a business transaction authentication system for financial institutions based on audio authentication. Background Technology

[0002] In the realm of financial institution transactions, various authentication technologies are widely used to ensure account security. Currently, mainstream authentication methods include password authentication, SMS dynamic code authentication, and biometric authentication. Among these, biometric authentication is gaining popularity due to its uniqueness and convenience. Voiceprint authentication, as a type of biometric authentication, has begun to be used in some financial institution business scenarios due to its advantages of requiring no additional hardware and being easy to operate. It primarily involves collecting the user's voice signal, extracting voiceprint features, and comparing them with a preset template to complete transaction identity verification.

[0003] However, existing voiceprint authentication technologies still have significant shortcomings. On the one hand, most systems only collect a single audio signal, which is easily affected by environmental background noise and equipment interference, leading to inaccurate voiceprint feature extraction and reduced authentication accuracy. On the other hand, they lack the ability to comprehensively identify fraudulent activities, making it difficult to effectively detect attacks such as recording playback and simulated voice. Furthermore, the authentication process is fixed and cannot dynamically adjust the authentication strength according to transaction risks. This results in either over-authentication affecting user experience or under-authentication leaving security vulnerabilities. In addition, the risk assessment models of existing systems are outdated and cannot adapt to changes in new fraud patterns. Summary of the Invention

[0004] To address the shortcomings of the existing technology, this invention provides a financial institution business transaction authentication system based on audio authentication.

[0005] The objective of this invention is achieved as follows:

[0006] The financial institution business transaction authentication system based on audio authentication includes an audio acquisition and preprocessing module, a voiceprint feature extraction and recognition module, a multimodal risk assessment module, a dynamic authentication decision module, and an authentication knowledge base management module.

[0007] The audio acquisition and preprocessing module includes a multi-channel audio acquisition unit and an audio quality enhancement unit. The multi-channel audio acquisition unit is used to simultaneously acquire user voice samples, ambient background sounds, and device interaction audio. The audio quality enhancement unit uses an adaptive noise reduction algorithm to perform noise reduction and feature enhancement processing on the original audio signal.

[0008] The voiceprint feature extraction and recognition module includes a deep voiceprint feature extraction unit and a real-time voiceprint matching unit. The deep voiceprint feature extraction unit uses a deep convolutional neural network to extract multi-dimensional voiceprint features from the preprocessed audio. The real-time voiceprint matching unit calculates the similarity score between the real-time voiceprint and the registered voiceprint template through a feature vector comparison algorithm.

[0009] The multimodal risk assessment module integrates an acoustic environment analysis unit and a behavioral characteristic analysis unit;

[0010] The dynamic authentication decision module includes a multi-factor authentication engine and a risk adaptive strategy unit. The multi-factor authentication engine dynamically combines authentication factors based on voiceprint matching scores and risk scores. The risk adaptive strategy unit adjusts the authentication strength according to the real-time risk level and triggers different levels of security authentication processes.

[0011] The authentication knowledge base management module includes a voiceprint feature library and a fraud pattern library. The voiceprint feature library stores user voiceprint feature vectors and update history, while the fraud pattern library accumulates known fraudulent audio features and attack patterns.

[0012] As a preferred embodiment of the financial institution business transaction authentication system based on audio authentication described in this invention, the multi-channel audio acquisition unit includes a directional microphone, an omnidirectional microphone, and an anti-interference microphone. The directional microphone is used to acquire the user's directional voice signal, the omnidirectional microphone is used to acquire ambient background sound, and the anti-interference microphone is used to filter out device interaction audio after electromagnetic interference. The sampling frequencies of the audio acquisition devices are kept synchronized.

[0013] As a preferred embodiment of the audio-based financial institution transaction authentication system of the present invention, the real-time voiceprint matching unit employs a deep feature matching algorithm based on metric learning, and the formula for calculating its similarity metric is as follows:

[0014] S=σ(ω1·cos(θ)+ω2·exp(-||v1-v2|| 2 / δ)+ω3·tanh(v1 T Mv2))

[0015] Where v1 and v2 represent two voiceprint feature vectors to be compared, θ is the angle between the two vectors, and M is a learnable metric matrix. δ is an adjustable weight coefficient, σ is a scale parameter, and σ represents the sigmoid activation function.

[0016] As a preferred embodiment of the audio-based financial institution business transaction authentication system of the present invention, the adaptive noise reduction algorithm includes a noise estimation sub-algorithm, a noise suppression sub-algorithm, and a signal recovery sub-algorithm.

[0017] The noise estimation sub-algorithm determines the noise type and intensity by statistically analyzing the time-domain energy distribution of the original audio signal;

[0018] The noise suppression sub-algorithm suppresses noise signals based on the minimum mean square error criterion;

[0019] The signal recovery sub-algorithm uses spectrum compensation technology to repair the speech detail features lost during the noise reduction process.

[0020] As a preferred embodiment of the financial institution business transaction authentication system based on audio authentication described in this invention, the deep convolutional neural network adopts a network structure combining multiple convolutional layers and multiple fully connected layers, with each convolutional layer connected to a batch normalization layer and a ReLU activation function layer.

[0021] The fully connected layer progressively compresses the feature dimension and outputs a multi-dimensional voiceprint feature vector with a fixed dimension.

[0022] The feature vector comparison algorithm uses the cosine similarity algorithm. When the calculated similarity score reaches a preset threshold, it is determined that the real-time voiceprint and the registered voiceprint template are initially matched.

[0023] As a preferred embodiment of the financial institution business transaction authentication system based on audio authentication described in this invention, the acoustic environment analysis unit is used to detect background noise characteristics and recording device fingerprints to identify recording playback fraudulent behavior, and the behavioral feature analysis unit analyzes the speech rate, tone dynamics and interactive behavior patterns in the user's speech to generate a comprehensive fraud risk score.

[0024] The acoustic environment analysis unit generates a recording device fingerprint by extracting device background noise characteristics, sampling rate deviation characteristics and signal distortion characteristics from the audio signal. When the matching degree between the recording device fingerprint and the device fingerprint when the user registers is lower than a preset threshold, a recording playback fraud warning is triggered.

[0025] The behavioral feature analysis unit identifies abnormal behavior as situations where the amplitude of speech rate fluctuations, the frequency of abrupt changes in tone, or the delay in interactive response exceeds the normal range. Furthermore, the comprehensive fraud risk score adopts a preset scoring system, and when the score reaches the high-risk threshold, it is determined to be a high-risk transaction.

[0026] As a preferred embodiment of the audio-based financial institution transaction authentication system of the present invention, the multi-factor authentication engine includes voiceprint factors, dynamic password factors, and device binding factors. Different combinations of authentication factors are activated to complete authentication based on different combinations of voiceprint matching scores and comprehensive fraud risk scores.

[0027] When the voiceprint matching score is high and the overall fraud risk score is low, only the voiceprint factor is used to complete the authentication.

[0028] When the voiceprint matching score and the comprehensive fraud risk score are at a medium level, the voiceprint factor and dynamic password factor are used together to complete the authentication.

[0029] When the voiceprint matching score is low or the overall fraud risk score is high, the voiceprint factor, dynamic password factor and device binding factor are combined to complete the authentication.

[0030] The risk adaptive strategy unit has three levels of authentication strength: low, medium, and high, which correspond to different numbers of authentication steps. The high-level authentication process requires an additional manual review.

[0031] As a preferred embodiment of the financial institution business transaction authentication system based on audio authentication described in this invention, the voiceprint feature library adopts a distributed database storage architecture, which partitions and stores user voiceprint feature vectors according to user account IDs, and automatically triggers voiceprint feature updates according to a preset period. During the update, multiple batches of historical feature vectors are retained for traceability and comparison.

[0032] The machine learning algorithm is the gradient boosting tree algorithm. After accumulating a predetermined number of new fraudulent audio feature data, the fraud pattern library triggers a model training update using the gradient boosting tree algorithm.

[0033] In a second aspect, some embodiments of the present invention provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any of the implementations of the first aspect above.

[0034] Thirdly, some embodiments of the present invention provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0035] This invention effectively addresses the shortcomings of existing technologies, significantly improving the security and practicality of transaction authentication for financial institutions. Through multi-channel audio acquisition and adaptive noise reduction processing, it accurately acquires pure voice signals, ensuring the accuracy of voiceprint feature extraction. Combined with multimodal risk assessment based on acoustic environment analysis and behavioral feature analysis, it comprehensively identifies fraudulent behavior. Dynamic authentication decision-making flexibly adjusts the authentication process and factor combinations based on real-time risk, balancing security and user experience. The authentication knowledge base continuously optimizes the risk assessment model, ensuring the system's ability to identify new fraud patterns and providing more reliable authentication guarantees for financial institutions' transactions. Attached Figure Description

[0036] Figure 1 This is a flowchart of the financial institution business transaction authentication system based on audio authentication in Example 1.

[0037] Figure 2 This is a schematic diagram of the authentication and identification of the financial institution business transaction authentication system based on audio authentication in Example 1. Detailed Implementation

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0040] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0041] Example 1

[0042] Reference Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a financial institution business transaction authentication system based on audio authentication, comprising:

[0043] The system includes an audio acquisition and preprocessing module 100, a voiceprint feature extraction and recognition module 200, a multimodal risk assessment module 300, a dynamic authentication decision module 400, and an authentication knowledge base management module 500.

[0044] The audio acquisition and preprocessing module 100 includes a multi-channel audio acquisition unit 101 and an audio quality enhancement unit 102. The multi-channel audio acquisition unit 101 is used to simultaneously acquire user voice samples, ambient background sounds and device interaction audio. The audio quality enhancement unit 102 uses an adaptive noise reduction algorithm to perform noise reduction and feature enhancement processing on the original audio signal.

[0045] The multi-channel audio acquisition unit 101 includes a directional microphone, an omnidirectional microphone, and an anti-interference microphone. The directional microphone is used to acquire the user's directional voice signal, the omnidirectional microphone is used to acquire ambient background sound, and the anti-interference microphone is used to filter out device interaction audio after electromagnetic interference. The sampling frequencies of the audio acquisition devices are kept synchronized.

[0046] The voiceprint feature extraction and recognition module 200 includes a deep voiceprint feature extraction unit 201 and a real-time voiceprint matching unit 202. The deep voiceprint feature extraction unit 201 uses a deep convolutional neural network to extract multi-dimensional voiceprint features from the preprocessed audio. The real-time voiceprint matching unit 202 calculates the similarity score between the real-time voiceprint and the registered voiceprint template through a feature vector comparison algorithm.

[0047] The real-time voiceprint matching unit 202 employs a deep feature matching algorithm based on metric learning, and its similarity metric is calculated using the following formula:

[0048] S=σ(ω1·cos(θ)+ω2·exp(-||v1-v2|| 2 / δ)+ω3·tanh(v1 T Mv2))

[0049] Where v1 and v2 represent two voiceprint feature vectors to be compared, θ is the angle between the two vectors, and M is a learnable metric matrix. δ is an adjustable weight coefficient, σ is a scale parameter, and σ represents the sigmoid activation function.

[0050] The multimodal risk assessment module 300 integrates an acoustic environment analysis unit 301 and a behavioral feature analysis unit 302. The acoustic environment analysis unit 301 is used to detect background noise features and recording device fingerprints to identify recording playback fraud. The behavioral feature analysis unit 302 analyzes the speech rate, tone dynamics, and interactive behavior patterns in the user's speech to generate a comprehensive fraud risk score.

[0051] The acoustic environment analysis unit 301 generates a recording device fingerprint by extracting the device's background noise characteristics, sampling rate deviation characteristics, and signal distortion characteristics from the audio signal. When the matching degree between the recording device fingerprint and the device fingerprint when the user registers is lower than a preset threshold, a recording playback fraud warning is triggered.

[0052] The behavioral feature analysis unit 302 determines abnormal behavior when the fluctuation range of speech rate, the frequency of abrupt changes in tone, or the delay in interactive response exceeds the normal range. The comprehensive fraud risk score adopts a preset scoring system, and when the score reaches the high-risk threshold, it is determined to be a high-risk transaction.

[0053] Adaptive noise reduction algorithms include noise estimation sub-algorithms, noise suppression sub-algorithms, and signal recovery sub-algorithms;

[0054] The noise estimation sub-algorithm determines the noise type and intensity by statistically analyzing the time-domain energy distribution of the original audio signal;

[0055] The noise suppression sub-algorithm suppresses noise signals based on the minimum mean square error criterion;

[0056] The signal recovery sub-algorithm uses spectrum compensation technology to repair the speech detail features lost during the noise reduction process.

[0057] The dynamic authentication decision module 400 includes a multi-factor authentication engine 401 and a risk adaptive strategy unit 402. The multi-factor authentication engine 401 dynamically combines authentication factors based on voiceprint matching scores and risk scores. The risk adaptive strategy unit 402 adjusts the authentication strength according to the real-time risk level and triggers different levels of security authentication processes.

[0058] The authentication factors of the multi-factor authentication engine 401 include voiceprint factors, dynamic password factors, and device binding factors. Different combinations of authentication factors are used to complete authentication based on different combinations of voiceprint matching scores and comprehensive fraud risk scores.

[0059] When the voiceprint matching score is high and the overall fraud risk score is low, only the voiceprint factor is used to complete the authentication.

[0060] When the voiceprint matching score and the comprehensive fraud risk score are at a medium level, the voiceprint factor and dynamic password factor are used together to complete the authentication.

[0061] When the voiceprint matching score is low or the overall fraud risk score is high, the voiceprint factor, dynamic password factor and device binding factor are combined to complete the authentication.

[0062] The risk adaptive strategy unit 402 has three levels of authentication strength: low, medium, and high, which correspond to different numbers of authentication steps. The high-level authentication process requires an additional manual review.

[0063] The authentication knowledge base management module 500 includes a voiceprint feature library 501 and a fraud pattern library 502. The voiceprint feature library 501 stores user voiceprint feature vectors and update history, while the fraud pattern library 502 accumulates known fraudulent audio features and attack patterns.

[0064] The voiceprint feature library 501 adopts a distributed database storage architecture, which stores user voiceprint feature vectors in partitions according to user account IDs, and automatically triggers voiceprint feature updates at preset cycles. During the update, multiple batches of historical feature vectors are retained for traceability and comparison.

[0065] The machine learning algorithm is the gradient boosting tree algorithm. After accumulating a predetermined number of new fraudulent audio feature data, the fraud pattern library 502 triggers a model training update of the gradient boosting tree algorithm.

[0066] Deep convolutional neural networks employ a network structure that combines multiple convolutional layers and multiple fully connected layers, with each convolutional layer followed by a batch normalization layer and a ReLU activation function layer.

[0067] The fully connected layer progressively compresses the feature dimension and outputs a multi-dimensional voiceprint feature vector with a fixed dimension.

[0068] The feature vector comparison algorithm uses the cosine similarity algorithm. When the calculated similarity score reaches a preset threshold, it is determined that the real-time voiceprint and the registered voiceprint template are initially matched.

[0069] Example 2

[0070] The second embodiment of the present invention differs from the first embodiment in that it further includes an experimental preparation and implementation process:

[0071] The experiment selected three high-frequency financial institution business scenarios: "offline counter transfers at financial institutions," "payment via mobile financial institution business APP," and "ATM withdrawal verification." Each scenario included 60 valid authentication tests, for a total of 180 tests. The subjects were divided into two groups: the experimental group consisted of the "Multi-Module Collaborative Audio Authentication System" (i.e., this invention), and the control group consisted of the existing technology "Single Voiceprint Template Comparison System."

[0072] In terms of hardware preparation, the experimental group was equipped with a multi-channel audio acquisition kit (including a directional microphone, an omnidirectional microphone, and an anti-interference microphone), a preprocessing server (equipped with an adaptive noise reduction algorithm), a deep neural network computing terminal (deploying a 5-layer convolutional + 3-layer fully connected voiceprint extraction model), and a risk assessment terminal (integrating acoustic environment and behavioral feature analysis modules); the control group was only equipped with a single directional microphone and a basic voiceprint comparison terminal (without noise reduction and risk assessment functions). In terms of software preparation, the experimental group imported the registered voiceprint templates of 100 testers (each person contained 8 different speech samples, covering normal, slow, and fast speech speeds), and constructed a fraud sample library (including 40 sets of recorded playback samples, 30 sets of simulated speech samples, and 30 sets of device tampering samples); the control group only imported the basic voiceprint templates of the same testers (each person had 3 standard speech speed samples).

[0073] Audio Acquisition and Preprocessing Stage: In each business scenario, testers initiate transaction authentication according to standard procedures. The experimental group synchronously acquires data using multi-channel microphones: directional microphones acquire user voice, omnidirectional microphones acquire ambient background sound, and anti-interference microphones acquire device interaction audio. After acquisition, the audio quality enhancement unit activates the adaptive noise reduction algorithm, first calculating the temporal energy distribution of the original audio using a noise estimation sub-algorithm. The noise estimation sub-algorithm formula is:

[0074]

[0075] Where P is the energy value, N is the number of sampling points, and x i For the amplitude of the i-th sampling point, the noise type is determined; then, the suppression coefficient is calculated based on the minimum mean square error criterion using a noise suppression sub-algorithm; finally, the speech details are restored using a signal recovery sub-algorithm, and the preprocessed audio is output. The control group simply inputs the raw audio collected by the directional microphone directly into the speakerprint comparison module without noise reduction processing.

[0076] Voiceprint feature extraction and matching stage: The experimental group inputs the preprocessed audio into the deep voiceprint feature extraction unit, which processes it through a deep convolutional neural network: convolutional layers (with kernel sizes of 3×3 and 5×5 alternately set) extract local acoustic features, and batch normalization layers standardize the data.

[0077] Multimodal risk assessment and dynamic authentication decision-making stage: The multimodal risk assessment module of the experimental group is activated: the acoustic environment analysis unit extracts the device fingerprint from the audio and compares it with the registered device fingerprint; the behavioral feature analysis unit counts the user's speech rate and the number of abrupt changes in tone, and combines it with the voiceprint matching score to generate a comprehensive fraud risk score through a weighted formula.

[0078] R = 0.4S + 0.3E + 0.3B

[0079] Where S represents the voiceprint matching score, E represents the environmental risk score, and B represents the behavioral risk score, the dynamic authentication decision module adjusts the process based on the R value: when R < 0.3, only voiceprint authentication is required; when 0.3 ≤ R ≤ 0.6, dynamic password authentication is added; when R > 0.6, dynamic password authentication plus manual review is added. The control group has no risk assessment step; authentication is directly passed only when the voiceprint matching score is ≥ 0.7, otherwise it is rejected.

[0080] Table 1 Comparison of Authentication Performance in Offline Counter Transfer Scenarios

[0081]

[0082]

[0083] Table 2 Comparison of Authentication Performance in Payment Scenarios of Mobile Financial Institution Business Apps

[0084]

[0085]

[0086] Table 3 Comparison of Authentication Performance in ATM Withdrawal Verification Scenarios

[0087]

[0088] Table 4 Comparison of Adaptive Noise Reduction Effects in Different Scenarios

[0089]

[0090]

[0091] Table 5 Comparison of Dynamic Authentication Decision Efficiency

[0092]

[0093]

[0094] Table 6. Comparison of Accuracy of Risk Assessment Models After Knowledge Base Optimization

[0095]

[0096]

[0097] As shown in Tables 1-3, the scenario-based authentication data reveals that the experimental group significantly outperformed the control group in all business scenarios in terms of core performance indicators. The advantages became more pronounced with increasing environmental complexity, directly addressing the shortcomings of existing single-voiceprint authentication technologies, such as weak anti-interference capabilities and poor authentication stability. In the relatively low-interference counter scenario, the experimental group achieved an authentication success rate of 96.7%, 18.4 percentage points higher than the control group's 78.3%, with a false recognition rate of 1.7%, only 27.4% of the control group's 6.2%. In the more interference-prone ATM scenario, the experimental group maintained an authentication success rate of 92.5%, while the control group's plummeted to 65.0%, with the false recognition rate gap widening further to 5.5 percentage points. The core reason for this difference lies in the experimental group's multi-channel acquisition and adaptive noise reduction technology. Table 4 shows that the experimental group achieved a signal-to-noise ratio improvement of 13.8 dB in the counter scenario, 23 times that of the control group (0.6 dB); a speech clarity score of 95.2 points, 12.6 points higher than the control group's 82.6 points; and a feature extraction completeness of 98.6%, ensuring the accuracy of subsequent voiceprint matching.

[0098] From a technical perspective, the control group relied solely on a single directional microphone to collect raw audio without any noise reduction processing. This resulted in environmental noise being directly superimposed on the speech signal, severely interfering with voiceprint features and achieving a matching similarity of only 0.61, leading to numerous false rejections of legitimate users. In contrast, the experimental group used multiple microphones to collect speech, ambient sound, and device sound separately. Then, through an adaptive noise reduction algorithm with noise estimation and suppression sub-algorithms, speech and noise were effectively separated, maintaining an average voiceprint matching similarity above 0.75 and ensuring authentication stability. This "multi-source acquisition + intelligent noise reduction" technical solution overcomes the limitations of existing technologies that rely on "single acquisition + no noise reduction," demonstrating the invention's ingenuity in anti-interference design.

[0099] The experimental group's performance in fraud detection addresses the shortcomings of existing technologies, such as their inability to identify multiple types of fraud and rigid authentication processes. Tables 1-3 show that the experimental group maintained a recognition rate of over 91.7% for recorded playback fraud and over 88.3% for analog voice fraud, while the corresponding indicators for the control group were all below 52.5%. Taking the ATM scenario as an example, the experimental group's recorded playback fraud recognition rate was 91.7%, 45 percentage points higher than the control group's 46.7%; the analog voice fraud recognition rate was 88.3%, 45.8 percentage points higher than the control group's 42.5%. This advantage stems from the experimental group's multimodal risk assessment module: the acoustic environment analysis unit can extract the unique sampling rate deviation of the recording equipment, the behavioral feature analysis unit can capture abnormal fluctuations in the speech rate of analog speech, and then combine the voiceprint matching score with a weighted formula:

[0100] R = 0.4S + 0.3E + 0.3B

[0101] The system calculates a comprehensive risk score to accurately identify fraudulent activities; while the control group judges fraud based solely on voiceprint matching scores, and only judges a sample as passing if the matching score is ≥0.7, resulting in a large number of fraudulent activities being missed.

[0102] Table 5 shows the dynamic authentication decision data, further demonstrating the novelty of this invention. The experimental group dynamically adjusted the authentication process based on the risk score R: in low-risk scenarios (R<0.3), only one voiceprint authentication step was required, with a user operation time of 1.6 seconds and a transaction completion rate of 99.2%, balancing efficiency and user experience; in medium-risk scenarios (0.3≤R≤0.6), two authentication steps were enabled, with a risk misjudgment rate of only 1.2%, 6.6 percentage points lower than the control group's 7.8%; in high-risk scenarios (R>0.6), three authentication steps were enabled, achieving an abnormal transaction interception rate of 96.0% (compared to only 58.0% in the control group). The control group used a fixed one-step authentication process, with no additional verification regardless of risk level, resulting in no user experience advantage in low-risk scenarios (satisfaction score of 90.2, lower than the experimental group's 94.5), and significant security vulnerabilities in high-risk scenarios (transaction completion rate of 83.3%, lower than the experimental group's 97.8%). This "risk-adaptive" decision-making logic breaks through the limitations of existing technologies' "rigid processes," achieving a balance between security and user experience.

[0103] The data in Table 6 demonstrates that the experimental group's authentication knowledge base management module addresses the shortcomings of existing technologies, such as "lagging risk assessment model updates and inability to adapt to new types of fraud." The initial model in the experimental group achieved a known fraud recognition rate of 88.5% and an unknown fraud recognition rate of 78.3%. After optimization using machine learning algorithms, the known fraud recognition rate increased to 96.8%, the unknown fraud recognition rate increased to 90.5%, and the model prediction error rate decreased from 5.8% to 1.8%. When faced with new fraud samples, the experimental group achieved an adaptation rate of 92.3%, quickly learning new fraud characteristics. In contrast, the control group, lacking model optimization capabilities, only saw a slight increase in the known fraud recognition rate from 55.2% to 56.3%, and a new fraud adaptation rate of only 45.8%, failing to cope with the iteration of fraud methods.

[0104] From a technical perspective, the experimental group's fraud pattern library continuously accumulates fraud features through a gradient boosting tree algorithm, while its voiceprint feature library retains users' historical feature vectors for retrospective comparison, enabling the model to dynamically adapt to changes in user voiceprints and updates in fraud methods. The control group, lacking a knowledge base iteration mechanism, had fixed model parameters, leading to a continuous decline in authentication accuracy over time due to voiceprint feature drift and the emergence of new fraud methods. This "continuously optimized" knowledge base management scheme ensures the system maintains high authentication accuracy over the long term, demonstrating the invention's innovative and forward-looking technological approach.

[0105] In summary, the comparison of the data in the six tables shows that the present invention, from audio acquisition, feature extraction, risk assessment to decision optimization, addresses the shortcomings of existing single voiceprint authentication technologies in terms of anti-interference, anti-fraud, and model iteration. All performance indicators are significantly better than existing technologies, demonstrating outstanding creativity and novelty.

[0106] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. The terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, unless otherwise explicitly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A financial institution business transaction authentication system based on audio authentication, characterized in that, It includes an audio acquisition and preprocessing module (100), a voiceprint feature extraction and recognition module (200), a multimodal risk assessment module (300), a dynamic authentication decision module (400), and an authentication knowledge base management module (500). The audio acquisition and preprocessing module (100) includes a multi-channel audio acquisition unit (101) and an audio quality enhancement unit (102). The multi-channel audio acquisition unit (101) simultaneously acquires user voice samples, ambient background sounds, and device interaction audio. The audio quality enhancement unit (102) uses an adaptive noise reduction algorithm to perform noise reduction and feature enhancement processing on the original audio signal. The voiceprint feature extraction and recognition module (200) includes a deep voiceprint feature extraction unit (201) and a real-time voiceprint matching unit (202). The deep voiceprint feature extraction unit (201) uses a deep convolutional neural network to extract multi-dimensional voiceprint features from the preprocessed audio. The real-time voiceprint matching unit (202) calculates the similarity score between the real-time voiceprint and the registered voiceprint template through a feature vector comparison algorithm. The multimodal risk assessment module (300) integrates an acoustic environment analysis unit (301) and a behavioral characteristic analysis unit (302). The dynamic authentication decision module (400) includes a multi-factor authentication engine (401) and a risk adaptive strategy unit (402). The multi-factor authentication engine (401) dynamically combines authentication factors based on voiceprint matching scores and risk scores. The risk adaptive strategy unit (402) adjusts the authentication strength based on the real-time risk level and triggers different levels of security authentication processes. The authentication knowledge base management module (500) includes a voiceprint feature library (501) and a fraud pattern library (502). The voiceprint feature library (501) stores user voiceprint feature vectors and update history, and the fraud pattern library (502) accumulates known fraudulent audio features and attack patterns.

2. The financial institution business transaction authentication system based on audio authentication according to claim 1, characterized in that, The multi-channel audio acquisition unit (101) includes a directional microphone, an omnidirectional microphone, and an anti-interference microphone. The directional microphone acquires the user's directional voice signal, the omnidirectional microphone acquires the ambient background sound, and the anti-interference microphone filters out electromagnetic interference from the device's interactive audio. The sampling frequencies of the audio acquisition devices are kept synchronized.

3. The financial institution business transaction authentication system based on audio authentication according to claim 1, characterized in that, The real-time voiceprint matching unit (202) adopts a deep feature matching algorithm based on metric learning, and the formula for calculating its similarity measure is as follows: ， in and Let represent the two voiceprint feature vectors to be compared, θ be the angle between the two vectors, and M be the learnable metric matrix. δ is an adjustable weight coefficient, σ is a scale parameter, and σ represents the sigmoid activation function.

4. The financial institution business transaction authentication system based on audio authentication according to claim 1, characterized in that, The adaptive noise reduction algorithm includes a noise estimation sub-algorithm, a noise suppression sub-algorithm, and a signal recovery sub-algorithm. The noise estimation sub-algorithm determines the noise type and intensity by statistically analyzing the time-domain energy distribution of the original audio signal; The noise suppression sub-algorithm suppresses noise signals based on the minimum mean square error criterion; The signal recovery sub-algorithm uses spectrum compensation technology to repair the speech detail features lost during the noise reduction process.

5. The financial institution business transaction authentication system based on audio authentication according to claim 1, characterized in that, The deep convolutional neural network adopts a network structure that combines multiple convolutional layers and multiple fully connected layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function layer. The fully connected layer progressively compresses the feature dimension and outputs a multi-dimensional voiceprint feature vector with a fixed dimension. The feature vector comparison algorithm uses the cosine similarity algorithm. When the calculated similarity score reaches a preset threshold, it is determined that the real-time voiceprint and the registered voiceprint template are initially matched.

6. The financial institution business transaction authentication system based on audio authentication according to claim 1, characterized in that, The acoustic environment analysis unit (301) detects background noise characteristics and recording device fingerprints to identify recording playback fraud. The behavior feature analysis unit (302) analyzes the speech rate, tone dynamics and interactive behavior patterns in the user's speech to generate a comprehensive fraud risk score. The acoustic environment analysis unit (301) generates a recording device fingerprint by extracting the device background noise characteristics, sampling rate deviation characteristics and signal distortion characteristics in the audio signal. When the matching degree between the recording device fingerprint and the device fingerprint when the user registers is lower than a preset threshold, a recording playback fraud warning is triggered. The behavioral feature analysis unit (302) determines abnormal behavior when the fluctuation range of speech rate, the frequency of abrupt changes in tone, or the delay in interactive response exceeds the normal range. The comprehensive fraud risk score adopts a preset scoring system, and when the score reaches the high-risk threshold, it is determined to be a high-risk transaction.

7. The financial institution business transaction authentication system based on audio authentication according to claim 1, characterized in that, The multi-factor authentication engine (401) includes voiceprint factors, dynamic password factors, and device binding factors. Different combinations of authentication factors are used to complete authentication based on different combinations of voiceprint matching scores and comprehensive fraud risk scores. When the voiceprint matching score is high and the overall fraud risk score is low, only the voiceprint factor is used to complete the authentication. When the voiceprint matching score and the comprehensive fraud risk score are at a medium level, the voiceprint factor and dynamic password factor are used together to complete the authentication. When the voiceprint matching score is low or the overall fraud risk score is high, the voiceprint factor, dynamic password factor and device binding factor are combined to complete the authentication. The risk adaptive strategy unit (402) has three levels of authentication strength: low, medium and high, which correspond to different numbers of authentication steps. The high-level authentication process requires an additional manual review.

8. The financial institution business transaction authentication system based on audio authentication according to claim 1, characterized in that, The voiceprint feature library (501) adopts a distributed database storage architecture, which stores user voiceprint feature vectors in partitions according to user account IDs, and automatically triggers voiceprint feature updates according to a preset period. During the update, multiple batches of historical feature vectors are retained for traceability and comparison. The machine learning algorithm is the gradient boosting tree algorithm. After accumulating a predetermined number of new fraudulent audio feature data, the fraud pattern library (502) triggers a model training update of the gradient boosting tree algorithm.

9. An electronic device, characterized in that... include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the system as described in any one of claims 1-8.

10. A computer-readable storage medium having executable instructions stored thereon, characterized in that... When executed by the processor, this instruction causes the processor to implement the system as described in any one of claims 1-8.

Images