Financial equipment security authentication method, system, equipment and medium

Abstract

The invention relates to a financial equipment security authentication method, system and device and a medium, and the method comprises the steps: carrying out the feature extraction and matching of user data collected in multiple modes, and generating a matching score and a quality score of each mode; uncertainty quantification and evidence conversion are carried out based on scores and scores, and a basic probability distribution function is constructed; carrying out dynamic weighting on probability distribution in combination with a sensor state and historical performance, and generating a preprocessing evidence set; through conflict adaptive evidence fusion processing, calculating conflict quality between evidences and redistributing conflict values according to a weight ratio, and generating a global fusion evidence body; and finally, the authentication confidence is calculated according to the fusion result, and a binary authentication instruction is generated through comparison with a dynamic risk threshold. The method can effectively solve the problem of evidence conflict in multi-modal authentication, dynamically adapts to system state change, and remarkably improves the accuracy and reliability of financial equipment security authentication.

Description

Technical Field

[0001] This invention relates to the field of equipment security technology, specifically to a method, system, device, and medium for security authentication of financial equipment. Background Technology

[0002] With the rapid development of electronic financial services, the security authentication mechanisms of financial devices such as ATMs and smart POS terminals are increasingly becoming a crucial link in ensuring the safety of user assets. Currently, authentication methods employing multimodal biometric fusion are gaining widespread attention due to their higher reliability. This method integrates multiple biometric information such as fingerprints, voiceprints, and facial features to form a composite judgment criterion, thereby overcoming the limitations of single-modal authentication, which is susceptible to environmental interference or malicious forgery. This fusion technology aims to leverage the complementarity of various modalities to provide a more robust identity verification barrier for financial transactions in complex application scenarios.

[0003] However, existing multimodal fusion authentication methods still face significant challenges in practical deployment. When authentication results from different modalities are inconsistent, traditional methods struggle to effectively quantify and handle such evidence conflicts, often employing static weights or simple voting mechanisms, which can lead to misjudgments at key decision points. Especially in complex situations involving sensor performance fluctuations, uneven data quality, or interference with some modalities, existing solutions lack the ability to dynamically assess evidence uncertainty and adjust decision-making strategies based on real-time system status. This not only reduces the accuracy of authentication results but may also affect the real-time response performance of financial devices due to computational redundancy in handling conflicting evidence. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide a financial device security authentication method, system, device and medium that can accurately handle multimodal evidence conflicts, dynamically adapt to changes in system state and ensure the reliability of authentication decisions.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for security authentication of financial devices, comprising the following steps:

[0007] S1: Extract and match features from user data collected from at least two authentication modalities, calculate the similarity and quality of the extracted features with the pre-stored registration templates, and generate a local matching score and feature quality score for each modality.

[0008] S2: Perform uncertainty quantification and evidence transformation on the local matching score and feature quality score. By weighted fusion of score and quality value, the uncertainty measure of each modality is obtained. Based on this measure, the matching result is converted into a standard evidence structure, and the basic probability assignment function of each modality is generated.

[0009] S3: Dynamically weight the basic probability allocation function based on sensor status and historical performance, and combine the real-time collected sensor health indicators and the long-term statistical authentication success rate to assign weight coefficients to each piece of evidence, generating a preprocessed evidence set containing weighted evidence bodies and dynamic weight sets.

[0010] S4: Perform conflict-adaptive evidence fusion on the weighted evidence body in the preprocessed evidence set and the dynamic weight set. Generate a global fused evidence body by calculating the conflict quality between evidence and redistributing the conflict values ​​to each proposition according to the weight ratio.

[0011] S5: Perform authentication decision processing on the globally fused evidence body, calculate the confidence level of the fused evidence for the legitimate user proposition, compare it with the preset dynamic risk threshold, and generate binary authentication instructions. The binary authentication instructions are used to control the financial device to execute or terminate transactions.

[0012] Secondly, the present invention provides a security authentication system for financial devices, which is configured with the following modules:

[0013] The multimodal feature matching module is used to extract and match features from user data collected from at least two authentication modalities, calculate the similarity and quality of the extracted features with the pre-stored registration template, and generate a local matching score and feature quality score for each modality.

[0014] The uncertainty evidence conversion module is used to quantify the uncertainty and convert the local matching score and feature quality score into evidence. It calculates the uncertainty measure of each modality by weighted fusion of score and quality value, and converts the matching result into a standard evidence structure based on the measure, generating the basic probability assignment function for each modality.

[0015] The evidence dynamic weighting module is used to dynamically weight the basic probability allocation function based on sensor status and historical performance. It combines real-time collected sensor health indicators and long-term statistical authentication success rate to assign weight coefficients to each piece of evidence, generating a preprocessed evidence set containing weighted evidence bodies and dynamic weight sets.

[0016] The conflict adaptive fusion module is used to perform conflict adaptive evidence fusion on the weighted evidence body in the preprocessed evidence set and the dynamic weight set. It generates a global fused evidence body by calculating the conflict quality between the evidence and redistributing the conflict values ​​to each proposition according to the weight ratio.

[0017] The fusion evidence authentication decision module is used to process the global fusion evidence body for authentication decisions, calculate the confidence level of the fusion evidence for legitimate user propositions, compare it with the preset dynamic risk threshold, and generate binary authentication instructions. The binary authentication instructions are used to control financial devices to execute or terminate transactions.

[0018] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned financial device security authentication methods.

[0019] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-mentioned financial device security authentication methods.

[0020] In summary, the financial device security authentication method provided in this application effectively solves the problems of evidence conflict and dynamic adaptability in multimodal financial device authentication by constructing a complete evidence fusion decision chain. This method achieves precise quantification of the reliability of various features through collaborative analysis and quality assessment of multimodal biometrics, establishing a high-quality data foundation for subsequent evidence fusion. Furthermore, by transforming the matching results into a standard evidence structure and introducing uncertainty metrics, the system can accurately characterize the credibility of each modality's authentication results, thus overcoming the shortcomings of traditional methods that ignore evidence uncertainty. Based on this, by dynamically weighting the evidence through the fusion of real-time sensor status and historical performance data, the system can adaptively adjust the weights of each modality according to the system's operating status, effectively addressing complex scenarios such as sensor performance fluctuations or temporary failures. Finally, by employing an improved evidence theory fusion mechanism, which calculates conflict quality and intelligently redistributes it according to weights, the method can thoroughly solve the decision-making problem when multimodal evidence contradicts each other, avoiding the risk of misjudgment that may result from simple weighted fusion. Ultimately, by linking the fusion results with dynamic risk thresholds to make decisions, a flexible authentication strategy that matches the transaction risk level can be formed, ensuring the security of high-risk transactions while also taking into account the user experience in low-risk scenarios.

[0021] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a security authentication method for financial devices provided in this application embodiment;

[0023] Figure 2 This is a schematic diagram of the process for generating a globally fused evidence body provided in an embodiment of this application;

[0024] Figure 3 This is a schematic diagram of the structure of a financial device security authentication system provided in another embodiment of this application. Detailed Implementation

[0025] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0027] In one embodiment, such as Figure 1 As shown, a security authentication method for financial devices is provided. This embodiment illustrates the method applied to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0028] S1: Extract and match features from user data collected from at least two authentication modalities, calculate the similarity and quality of the extracted features with the pre-stored registration templates, and generate a local matching score and feature quality score for each modality.

[0029] Specifically, the system selects at least two complementary biometric authentication modalities, preferably a combination of fingerprint, face, and voiceprint, to meet the portability and security requirements of financial devices. It collects raw user data through dedicated sensors integrated into the financial devices. Fingerprint data is collected using a capacitive fingerprint sensor with anti-interference filtering capabilities, filtering out signal distortion caused by changes in skin humidity and minor dirt. Face data is collected using an infrared binocular camera to obtain facial images containing depth information, mitigating the risk of two-dimensional image forgery. Voiceprint data is collected through a dual-microphone array to capture user voice commands, simultaneously recording environmental noise parameters for subsequent quality assessment. In the feature extraction stage, fingerprint features are extracted using the Minutiae point extraction algorithm to extract key features such as endpoints and bifurcation points in the fingerprint image. Through coordinate normalization and rotation invariance processing, a fixed-dimensional fingerprint feature vector is generated. Face features are extracted based on a lightweight CNN network, focusing on the depth features of key areas such as the eyes, nose, and mouth, and outputting a fixed-dimensional face feature vector. The network has been specifically trained for financial scenarios and is able to adapt to changes in facial expressions and slight occlusion. Voiceprint features are extracted using MFCC to extract the spectral features of speech, combined with GMM modeling, to generate a fixed-dimensional voiceprint feature vector, eliminating feature shifts caused by changes in speech rate and tone.

[0030] During the similarity calculation and registration template matching process, pre-stored registration templates are stored in the encrypted security chip of the financial device, using an encrypted algorithm for storage. Template retrieval requires a security verification process built into the chip. The system selects the corresponding similarity calculation method based on the matching characteristics of each modality feature: Mahalanobis distance is used for fingerprint features, cosine similarity for face features, and Euclidean distance for voiceprint features. All similarity results are normalized to a fixed interval with a starting value of 0 and an ending value of 1. The trend of numerical change is positively correlated with the matching degree. In the feature quality scoring stage, the system constructs a multi-dimensional feature quality evaluation system. The scores for each modality are normalized to a fixed interval with a starting value of 0 and an ending value of 1. Fingerprint quality assessment is based on a weighted calculation of three dimensions: image clarity, feature point integrity, and acquisition area coverage. Each dimension is assigned a weight according to its impact on the matching result. Face quality assessment is based on a weighted calculation of three dimensions: pose deviation angle, illumination uniformity, and occlusion rate. The weight of each dimension is determined based on its impact on authentication accuracy in actual application scenarios. Voiceprint quality assessment is based on a weighted calculation of three dimensions: signal-to-noise ratio, speech duration, and distortion. The weight of each dimension is set according to the reliability requirements of voiceprint feature extraction. Through the above process, the system outputs a local matching score and feature quality score for each modality.

[0031] S2: Perform uncertainty quantification and evidence transformation on the local matching score and feature quality score. By weighted fusion of score and quality value, the uncertainty measure of each modality is obtained. Based on this measure, the matching result is converted into a standard evidence structure, and the basic probability assignment function of each modality is generated.

[0032] Specifically, the system constructs a standard evidence structure based on DS evidence theory, transforming matching results and quality information into a fusionable evidence body. In the uncertainty quantification stage, the system uses information entropy theory to quantify the uncertainty of each modality's data, defining the uncertainty measure as a weighted entropy value combination. The weighting coefficients are determined through multiple scenario tests to balance the impact of matching stability and quality reliability on the uncertainty measure. The entropy calculation function follows the basic definition of information entropy, calculating the corresponding entropy values ​​using local matching scores and feature quality scores respectively. These entropy values ​​are then weighted and summed using the weighting coefficients to obtain the uncertainty measure for each modality. The range of this measure is constrained by the entropy calculation rules, and its numerical trend is positively correlated with data uncertainty.

[0033] In the standard evidence structure construction process, the system defines an identification framework containing three propositions: the current user is a legitimate registered user, the current user is an illegitimate user, and the user's identity cannot be determined. This framework conforms to the normative requirements of DS evidence theory, providing a unified proposition space for subsequent evidence fusion. In the basic probability allocation function generation stage, the system converts local matching scores and feature quality scores into basic probability allocation values ​​for each proposition based on uncertainty metrics. In the conversion rules, the basic probability allocation value for the legitimate user proposition is obtained by multiplying the complement of the uncertainty metric by the weighted sum of the matching score and quality score; the basic probability allocation value for the illegitimate user proposition is obtained by multiplying the complement of the uncertainty metric by the weighted sum of the complement of the matching score and quality score; and the basic probability allocation value for the uncertain proposition is directly taken from the uncertainty metric. The matching score weight is used to prioritize the core role of the matching result in evidence conversion. The conversion process strictly follows the normative constraints of DS evidence theory, ensuring that the sum of the basic probability allocation values ​​for each proposition equals 1. Through the above conversion rules, the system generates the basic probability allocation function corresponding to each modality, which is the standard evidence body for that modality.

[0034] S3: The basic probability allocation function is dynamically weighted based on sensor status and historical performance. Combined with real-time collected sensor health indicators and long-term statistical authentication success rate, weight coefficients are assigned to each piece of evidence to generate a preprocessed evidence set containing weighted evidence bodies and dynamic weight sets.

[0035] Specifically, the system assigns dynamic weights to each piece of evidence by analyzing real-time device status and long-term performance statistics, improving the scenario adaptability of fusion decision-making. During sensor health indicator collection, the system collects health parameters for each modality of sensor in real time. For fingerprint sensors, these parameters include contact pressure, signal-to-noise ratio, hardware temperature, and response time. For face cameras, they include frame rate, image resolution, lens cleanliness, and data transmission error rate. For voiceprint microphones, they include audio sampling rate, distortion, and environmental noise suppression. The system transmits the collected parameters to the device's main control unit in real time via a secure bus, and the transmitted data uses a verification algorithm to ensure integrity. During the historical performance statistical analysis phase, the system builds an authentication log database locally to store authentication records for each modality. Statistical dimensions include core performance indicators and stability indicators. Core performance indicators include the probability of a legitimate user successfully authenticating, the probability of an illegitimate user falsely authenticating, and the probability of a legitimate user falsely being rejected. Stability indicators include the number of consecutive successful authentications, the average authentication response time, and the frequency of abnormal interruptions. The database is stored in encrypted form, with read access only granted to the authentication algorithm module, prohibiting external access.

[0036] Preferably, the system can use a Logistic regression model to calculate the weight coefficients of each modality's evidence body. The model input consists of normalized values ​​of sensor health indicators and normalized values ​​of historical performance indicators. The normalized values ​​of sensor health indicators are processed using the Min-Max normalization method, and the normalized values ​​of historical performance indicators are processed after assigning corresponding weights according to the influence degree of each core performance indicator. The model output is the weight coefficient, which has a starting value of 0 and an ending value of 1, and the sum of the weight coefficients of all modalities equals 1. The sensor state and historical performance jointly determine the magnitude of the weight coefficient; the better the state and the better the performance of the modality, the larger the corresponding weight coefficient. During the construction of the preprocessed evidence set, the system associates the basic probability allocation function of each modality with the corresponding dynamic weight coefficient to form a preprocessed evidence set, which includes weighted evidence bodies and a dynamic weight set.

[0037] S4: Perform conflict-adaptive evidence fusion on the weighted evidence body in the preprocessed evidence set and the dynamic weight set. By calculating the conflict quality between evidence and redistributing the conflict values ​​to each proposition according to the weight ratio, a globally fused evidence body is generated.

[0038] Specifically, for conflicting evidence in the preprocessed evidence set, the system can employ an adaptive fusion strategy to achieve effective synthesis, avoiding the paradox of traditional fusion algorithms. In the evidence conflict quality calculation stage, the system uses an improved conflict metric to calculate the conflict quality between any two pieces of evidence. This metric balances conflict intensity and evidence consistency. It calculates the conflict quality of each pair of evidence pieces by summing the products of the basic probability assignments of corresponding propositions, combined with evidence similarity and consistency compensation coefficients. Evidence similarity is calculated as the ratio of the sum of the products of the basic probability assignments of each proposition to the square root of the sum of the squares of the basic probability assignments of each evidence piece itself. The system averages the conflict quality of all pairwise evidence pieces to obtain the global conflict quality. The global conflict quality range starts at 0 and ends at 1, and is divided into three intervals: low conflict, medium conflict, and high conflict. During the execution of the adaptive conflict fusion algorithm, the system employs differentiated fusion strategies for different conflict intervals.

[0039] Within the low-conflict range, the system directly employs the weighted DS synthesis rule. It exponentially powers the basic probability allocation values ​​of the corresponding propositions using the weight coefficients of each piece of evidence, then multiplies the results of the exponentiation of all evidence pieces. Finally, it normalizes the sum of the product results to obtain the basic probability allocation values ​​of each proposition in the globally fused evidence body. Within the medium-conflict range, the system introduces a conflict allocation factor. First, it calculates the weight percentage of each piece of evidence, obtained by the ratio of the weight coefficient of a single piece of evidence to the sum of the weight coefficients of all pieces of evidence. Then, it calculates the total conflict value, obtained by the difference between 1 and the sum of the products of the basic probability allocation values ​​of the corresponding propositions of all evidence pieces. The total conflict value is then allocated to each piece of evidence according to its weight percentage, correcting the basic probability allocation values ​​of the legitimate user propositions for each piece of evidence. The corrected evidence pieces are then fused using the DS synthesis rule to obtain the globally fused evidence body.

[0040] Within the high-conflict interval, the system adds an evidence credibility screening step to the medium-conflict interval processing flow. The credibility of each piece of evidence is calculated using a likelihood function, which is the sum of the basic probability allocation values ​​for legitimate user propositions and uncertain propositions. Credibility is calculated by multiplying the basic probability allocation value of each proposition by its corresponding likelihood function value. Evidence pieces whose credibility does not meet the set requirements are downweighted. The downweighted weight coefficient is obtained by multiplying the original weight coefficient by the credibility value. The downweighted weight coefficients of all evidence pieces are then normalized. The conflict allocation and synthesis process for the medium-conflict interval is repeated using the corrected weight coefficients to generate a globally fused evidence body. This evidence body contains the fused basic probability allocation values ​​for each proposition.

[0041] S5: Perform authentication decision processing on the globally fused evidence body, calculate the confidence level of the fused evidence for the legitimate user proposition, compare it with the preset dynamic risk threshold, and generate binary authentication instructions. The binary authentication instructions are used to control the financial device to execute or terminate transactions.

[0042] Specifically, the system makes security authentication decisions based on a globally fused evidence body and outputs device control commands. In the confidence level calculation phase, the system calculates the confidence level of the legitimate user proposition based on the globally fused evidence body, using a joint evaluation of the trust function and the likelihood function. The trust function is the basic probability allocation value of the legitimate user proposition in the globally fused evidence body, directly reflecting the system's absolute confidence that the current user is a legitimate registered user. The likelihood function is the sum of the basic probability allocation values ​​of the legitimate user proposition and the basic probability allocation values ​​of the uncertain proposition, reflecting the system's maximum possible confidence that the current user is a legitimate registered user. The trust function serves as the core indicator for decision-making, while the likelihood function is used to assist in risk assessment. The difference between the likelihood function and the trust function reflects the degree of uncertainty in the decision; a larger difference indicates a higher degree of uncertainty regarding the authentication result.

[0043] Preferably, the system dynamically adjusts the decision threshold based on the risk level of the financial transaction scenario. The threshold is calculated in real time by the device's main control unit based on transaction parameters, including transaction amount, user security level, and device environment status. Changes in transaction amount affect threshold adjustment, with the amount scale and threshold adjustment trend being positively correlated. User security level is determined based on the user's historical transaction behavior, with the cumulative number of historically normal transaction records affecting security level classification, and security level and threshold adjustment trend being negatively correlated. Device environment status is determined through the device's built-in positioning and network security detection functions, with the security level of the area where the device is located and the network attack risk level affecting threshold adjustment, and risk level and threshold adjustment trend being positively correlated. The threshold adjustment range is within a fixed interval, ensuring a balance between risk control and ease of use.

[0044] Furthermore, the system compares the confidence level of legitimate user propositions with a dynamically adjusted threshold and generates a corresponding binary authentication command based on the comparison result. When the confidence level is greater than or equal to the threshold, the system generates a transaction permission command. This command is encrypted and transmitted to the transaction control unit of the financial device via a secure communication link. Upon receiving the command, the transaction control unit initiates the transaction execution process, which includes secondary verification of user identity, authorization of fund transfer, and encrypted storage of transaction records. The transaction records contain a fused evidence digest. When the confidence level is less than the threshold, the system generates a transaction termination command. Upon receiving the command, the transaction control unit immediately disconnects the transaction link and clears sensitive data from the current session. Sensitive data includes temporarily stored biometric vectors and intermediate matching results. Simultaneously, the system provides authentication failure information to the user through the device's display and audio output components. The feedback information does not include a specific reason for the failure. The system encrypts the failure record and writes it to the security log. The failure record includes information such as conflict quality, confidence level, and transaction parameters for subsequent system optimization and anomaly tracing.

[0045] In summary, the financial device security authentication method provided in this application effectively solves the problems of evidence conflict and dynamic adaptability in multimodal financial device authentication by constructing a complete evidence fusion decision chain. This method achieves precise quantification of the reliability of various features through collaborative analysis and quality assessment of multimodal biometrics, establishing a high-quality data foundation for subsequent evidence fusion. Furthermore, by transforming the matching results into a standard evidence structure and introducing uncertainty metrics, the system can accurately characterize the credibility of each modality's authentication results, thus overcoming the shortcomings of traditional methods that ignore evidence uncertainty. Based on this, by dynamically weighting the evidence through the fusion of real-time sensor status and historical performance data, the system can adaptively adjust the weights of each modality according to the system's operating status, effectively addressing complex scenarios such as sensor performance fluctuations or temporary failures. Finally, by employing an improved evidence theory fusion mechanism, which calculates conflict quality and intelligently redistributes it according to weights, the method can thoroughly solve the decision-making problem when multimodal evidence contradicts each other, avoiding the risk of misjudgment that may result from simple weighted fusion. Ultimately, by linking the fusion results with dynamic risk thresholds to make decisions, a flexible authentication strategy that matches the transaction risk level can be formed, ensuring the security of high-risk transactions while also taking into account the user experience in low-risk scenarios.

[0046] In one embodiment, the local matching score of the financial device security authentication method provided by the present invention includes a fingerprint local matching score and a voiceprint local matching score, and the feature quality score includes a fingerprint feature quality score and a voiceprint feature quality score. S1 includes:

[0047] S11: Perform grayscale conversion and noise filtering on the image data acquired from the fingerprint recognition modality, construct the fingerprint feature point topology through orientation field estimation and ridge extraction algorithms, and generate fingerprint feature vectors.

[0048] Specifically, the system receives raw image data acquired from the fingerprint recognition modality. By integrating the brightness information of each channel, it converts the multi-channel image into a single-channel grayscale image. This process follows the basic principles of image grayscale conversion, achieved through linear combination of channel information, ensuring that the grayscale values ​​accurately reflect the texture details of the original image. Subsequently, the system performs noise filtering, employing a neighborhood pixel-based smoothing algorithm to correct abnormal pixels in the grayscale image. Abnormal pixels are determined by their grayscale differences from surrounding pixels. During the filtering process, the boundary features of fingerprint ridges and valleys are preserved, avoiding texture distortion. In the orientation field estimation stage, the system divides the grayscale image into several non-overlapping image blocks. It statistically analyzes the grayscale gradient direction of pixels within each block, determining the overall direction of the ridges in that region through the distribution of gradient vector directions, thus constructing a global fingerprint orientation field model. This orientation field model provides directional guidance for subsequent ridge extraction.

[0049] During ridge extraction, the system uses an adaptive threshold segmentation algorithm to convert grayscale images into binary images based on the ridge orientation determined by the orientation field model, thus separating ridges from valleys. Morphological processing is then used to remove isolated noise and minor breaks in the binary image, repairing discontinuous ridge sections. In the feature point topology construction stage, the system identifies the endpoints and bifurcation points of ridges in the binary image, records the coordinates, orientation information, and connectivity of each feature point, forming a fingerprint feature point topology. Based on this topology, the system extracts fixed-dimensional feature parameters, arranges them in a preset order to generate fingerprint feature vectors, and the dimensional design of the feature vectors meets the efficiency and accuracy requirements of subsequent matching calculations.

[0050] S12: Perform detailed feature matching on the fingerprint feature vector and the pre-stored fingerprint registration template, calculate the feature point overlap degree under the minimum rotation and translation parameters, and generate the fingerprint local matching score.

[0051] Specifically, the system retrieves a pre-stored fingerprint registration template from an encrypted storage medium. The fingerprint registration template and the currently extracted fingerprint feature vector use the same feature extraction standard to ensure consistency in the matching process. After feature matching processing is initiated, the system performs geometric calibration on the current fingerprint feature vector and the fingerprint registration template. By finding corresponding reference points in the two sets of feature points, the system calculates rotation and translation parameters. These parameters are solved based on the coordinate transformation relationship between the reference points. The system uses an iterative optimization algorithm to determine the optimal parameters that minimize the rotation and translation deviation, aligning the current feature vector and the registration template in the same coordinate system and eliminating the influence of differences in finger placement angle and position during the acquisition process. After calibration, the system performs detailed feature matching processing, comparing the type, coordinates, orientation, and topological relationships of adjacent feature points in the current feature vector and the registration template one by one, and counting the number of feature points that meet the matching conditions.

[0052] In the feature point overlap calculation stage, the system generates a feature point overlap index based on the ratio of the number of successfully matched feature points to the total number of feature points in the registered template, combined with the geometric consistency parameter of feature point matching. The geometric consistency parameter is calculated by comprehensively considering the distance and angle deviations between matched feature point pairs. The system maps the feature point overlap index to a fingerprint local matching score. The mapping rule is set based on the statistical regularity of feature matching to ensure that the matching score can objectively reflect the similarity between the current fingerprint and the registered template. The value range of the matching score follows a normalization standard to maintain a consistent scale with the matching scores of other modalities, facilitating the fusion processing of multimodal data.

[0053] S13: Perform sharpness and contrast analysis on the original image acquired by the fingerprint recognition modality, evaluate the effective area ratio and ridge continuity of the image, and generate a fingerprint feature quality score.

[0054] Specifically, the system acquires the original image from the fingerprint recognition modality. In the sharpness analysis process, the system calculates the gray-level gradient amplitude of the original image. By statistically analyzing the average value and distribution of the overall gray-level gradient, the system assesses the image's sharpness. The gray-level gradient amplitude is calculated from the gray-level difference between adjacent pixels, and the concentration of the gradient distribution directly reflects the sharpness of image details. In the contrast analysis stage, the system statistically analyzes the gray-level value distribution range of the original image, calculates the difference between the maximum and minimum gray-level values, and analyzes the distribution pattern of the gray-level histogram. The contrast of the image is assessed by the dispersion of the gray-level values. The contrast index reflects the gray-level difference between ridges and valleys; a greater difference is more conducive to feature extraction.

[0055] During the effective region proportion assessment, the system identifies and segments the effective region containing fingerprint texture based on the contour features of the fingerprint image, excluding background areas without texture at the image edges and invalid areas that were not fully captured. The effective region proportion is calculated as the ratio of the number of pixels in the effective region to the total number of pixels in the image. In the ridge continuity assessment stage, the system tracks the ridges in the binarized image, recording the number and length of ridge breakpoints. Ridge continuity is comprehensively evaluated using two indicators: breakpoint density and average break length. Breakpoint density is the ratio of the number of breakpoints to the total ridge length. The system weights and sums the four assessment indicators—clarity, contrast, effective region proportion, and ridge continuity—according to preset weights to generate a fingerprint feature quality score. The weight allocation is determined based on the influence of each indicator on the accuracy of subsequent matching results. The score is normalized to a unified range to maintain consistency with quality scores from other modalities.

[0056] S14: Perform frame segmentation and pre-emphasis processing on the audio data collected from the voiceprint recognition modality, extract the vocal tract and vocal cord excitation features through linear predictive coding cepstral coefficients, and generate a voiceprint feature vector.

[0057] Specifically, the system receives raw audio data acquired from voiceprint recognition modality, divides the continuous audio signal into several short time frames according to a preset time interval, and sets overlapping regions between frames. The overlap ratio is determined according to the short-term stationarity requirements of the audio signal to ensure smooth feature transitions between frames and avoid information loss. After framing, the system performs pre-emphasis processing, filtering each frame of audio signal through a high-pass filter to compensate for high-frequency attenuation during transmission and acquisition. The cutoff frequency of the filter is set according to the frequency characteristics of the human voice signal, enhancing the high-frequency vocal cord excitation characteristics while preserving the main frequency components of the human voice.

[0058] In the feature extraction stage, the system uses a linear predictive coding algorithm to process the pre-emphasized audio frames. By establishing a linear prediction model, the system predicts the value of the current sampling point using a linear combination of the previous sampling points, and solves for the prediction coefficients of the model. These coefficients reflect the resonance characteristics of the vocal tract. The system converts the linear prediction coefficients into linear predictive coding cepstral coefficients, which are obtained by performing a discrete cosine transform on the prediction coefficients. This effectively compresses the data dimensionality while preserving the key features of the vocal tract. Furthermore, the system extracts the fundamental frequency information of the audio signal. The fundamental frequency information is determined through spectral analysis, reflecting the vibration frequency characteristics of the vocal cords. The linear predictive coding cepstral coefficients and the fundamental frequency features are combined according to preset rules to construct a voiceprint feature vector containing vocal tract and vocal cord excitation features. The dimensionality design of the feature vector balances recognition accuracy and computational efficiency, meeting the needs of subsequent matching processing.

[0059] S15: Perform dynamic time warping matching on the voiceprint feature vector and the pre-stored voiceprint registration template, calculate the minimum cumulative distance between the test speech and the registered speech in the feature space, and generate a local voiceprint matching score.

[0060] Specifically, the system retrieves a pre-stored voiceprint registration template, which is a voiceprint feature vector collected during the user registration phase and generated through the same processing procedure. After the matching process starts, the system can use a dynamic time warping algorithm to solve the problem of inconsistent duration between the current test speech and the registered speech. This algorithm constructs a feature space distance matrix, where the matrix elements are the distances between the current voiceprint feature vector and the corresponding frame feature vector of the registration template. The distance calculation uses common metrics such as Euclidean distance or Manhattan distance. The system searches for an optimal path in the distance matrix. The path is constrained by endpoint constraints, monotonicity constraints, and slope constraints to ensure that the path reflects the optimal matching relationship between the two sequence features. The sum of the distances of all elements on the path is the minimum cumulative distance, which reflects the similarity between the current voiceprint feature vector and the registration template. The smaller the distance, the higher the similarity. The system converts the minimum cumulative distance into a local voiceprint matching score. The conversion rule is determined by the mapping relationship between distance and matching degree. During the mapping process, the feature difference distribution between different speakers is considered to ensure that the matching score can objectively reflect the reliability of identity matching. The score results are normalized to a unified range and maintain scale consistency with the local matching score of fingerprint modality, which facilitates the subsequent fusion of multimodal data.

[0061] S16: Perform signal-to-noise ratio and harmonicity analysis on the raw audio collected by the voiceprint recognition modality, evaluate the background noise intensity and fundamental frequency trajectory stability, and generate a voiceprint feature quality score.

[0062] Specifically, the system acquires raw audio data from the voiceprint recognition modality. It separates the speech signal from the background noise in the audio data and uses a speech activity detection algorithm to identify speech segments and non-speech segments within the audio frame. The non-speech segments are considered background noise samples. The system calculates the signal energy of the speech segment and the noise energy of the background noise segment. Both signal and noise energy are obtained by summing the squares of the sampled values ​​of the corresponding audio segments. The signal-to-noise ratio (SNR) is the ratio of signal energy to noise energy, reflecting the degree of interference of background noise on the speech signal.

[0063] In the harmonic analysis phase, the system performs spectral analysis on the speech segment, extracts the fundamental frequency and harmonic frequency components, calculates the multiple relationship between the harmonic frequencies and the fundamental frequency, and assesses the integrity and stability of the harmonic components. The harmonicity index reflects the regularity of vocal cord vibration. Background noise intensity assessment is achieved by calculating the average noise energy, directly reflecting the strength of environmental noise. In the fundamental frequency trajectory stability assessment, the system tracks the fundamental frequency value of each frame of the speech segment, calculates the standard deviation and coefficient of variation of the fundamental frequency value. The standard deviation represents the degree to which the fundamental frequency value deviates from the average value, and the coefficient of variation is the ratio of the standard deviation to the average value. Both indicators jointly reflect the stability of the fundamental frequency trajectory. The system weights and sums the four evaluation indicators—signal-to-noise ratio, harmonicity, background noise intensity, and fundamental frequency trajectory stability—according to preset weights to generate a voiceprint feature quality score. The weight allocation is determined based on the degree of influence of each indicator on the accuracy of voiceprint matching. The score result is normalized to a unified interval, consistent with the feature quality score of fingerprint modality.

[0064] In one embodiment, step S2 of the financial device security authentication method provided by the present invention specifically includes the following steps:

[0065] S21: Perform linear weighted fusion processing on the local matching score and feature quality score, and calculate the confidence of the matching score by discounting it in combination with the quality score to generate an uncertainty metric for each modality.

[0066] Specifically, the system acquires the local matching score and feature quality score for each modality, performs linear weighted fusion, and sets the weighting rule based on the degree of influence of the two scores on the final credibility assessment. This process is implemented using the following formula:

[0067]

[0068] in, Let i be the uncertainty metric for the i-th mode. The local matching score for the i-th modality. The feature quality score for the i-th mode is given. and These are the weighting coefficients for the local matching score and the feature quality score, respectively, and satisfy the following conditions: The system first multiplies the local matching score by its corresponding weighting coefficient, then multiplies the feature quality score by its corresponding weighting coefficient, and finally sums the two products to obtain a weighted fusion value. Based on this weighted fusion value, the system performs a credibility discount calculation, generating an uncertainty measure by the difference between 1 and the weighted fusion value. The magnitude of the uncertainty measure is inversely related to the credibility of the matching score; the higher the weighted fusion value, the smaller the uncertainty measure. This result provides a quantitative basis for the subsequent conversion of the matching results into a standard evidence structure, ensuring that the evidence conversion process reflects the reliability of the data itself.

[0069] S22: Perform basic probability allocation processing on the uncertainty metric and local matching score, and allocate the matching score to the propositions of legitimate users and illegitimate users according to the uncertainty ratio to generate the initial probability allocation.

[0070] Specifically, the system calls the uncertainty metric value and local matching score of each modality, and proportionally allocates the effective probability quality (after deducting uncertainty) to legitimate user propositions and illegitimate user propositions. Preferably, the allocation formula is as follows:

[0071]

[0072]

[0073] in, Let be the initial probability assignment value of the i-th modal evidence for the legitimate user proposition A. Let A represent the initial probability assignment of the i-th modality evidence to the proposition B concerning illegitimate users, where A represents the proposition that the current user is a legitimate registered user, and B represents the proposition that the current user is an illegitimate user. This represents the effective probability quality ratio after deducting uncertainty. The system first calculates the effective probability quality ratio using the difference between 1 and the uncertainty metric. Then, it multiplies this ratio by the local matching score to obtain the initial probability allocation value for legitimate user propositions. Simultaneously, it multiplies the effective probability quality ratio by the complement of the local matching score to obtain the initial probability allocation value for illegitimate user propositions. The initial probability allocation process strictly follows the correlation between the matching score and the uncertainty metric, ensuring that the probability allocation for legitimate and illegitimate user propositions objectively reflects the matching results and reliability of the modal data.

[0074] S23: Perform uncertainty normalization on the initial probability assignment, allocate the remaining probability mass to the globally uncertain proposition, complete the evidence body construction, and generate the basic probability assignment function.

[0075] Specifically, the system sums the initial probability assignments of legitimate user propositions and illegitimate user propositions to obtain a total initial probability assignment. Based on this total, the system performs uncertainty normalization to ensure that the sum of the probability assignments of all propositions meets the normative requirements of the DS evidence theory. This process is implemented using the following formula:

[0076]

[0077] in, For the i-th modal evidence regarding the globally uncertain proposition The probability allocation value, This represents propositions where user identity cannot be determined. The system first calculates the sum of the initial probability assignments for legitimate and illegitimate user propositions. Then, it obtains the probability assignment value for globally uncertain propositions by subtracting 1 from this sum. This value is equal to the uncertainty metric, forming a logical closed loop. Through normalization, the system ensures... The system satisfies the normative constraints of the basic probability assignment function. At this point, the system has completed the construction of the evidence body, generating a basic probability assignment function that includes probability assignment values ​​for legitimate user propositions, illegitimate user propositions, and globally uncertain propositions.

[0078] In one embodiment, step S3 of the financial device security authentication method provided by the present invention specifically includes the following steps:

[0079] S31: Real-time status monitoring of the fingerprint sensor and microphone, collection of signal strength and self-test error code data, and generation of sensor health status score.

[0080] Specifically, the system continuously monitors the fingerprint sensor and microphone in real time, covering the entire device operation cycle to ensure the continuity and integrity of data acquisition. The system collects signal strength data from the fingerprint sensor at preset time intervals. This signal strength data is acquired through the sensor's built-in signal detection circuit, reflecting the signal transmission efficiency between the sensor and the target object. Simultaneously, it collects self-test error code data from both types of sensors. These error codes are generated by the sensor's internal diagnostic module, recording status information such as hardware malfunctions and data transmission failures. The system quantizes the collected signal strength data and self-test error code data, mapping the signal strength data to a uniform-scale intensity quantization value and converting the self-test error code data into error quantization values. The error quantization value is positively correlated with the frequency and severity of errors. Preferably, this process is implemented using the following formula:

[0081]

[0082] in, Score the health status of the i-th type of sensor. The signal strength quantization value. For error quantization value, For signal strength weighting coefficients, The error quantization value is weighted by a coefficient, and the sum of the two is 1. The system first multiplies the signal strength quantization value by the corresponding weight coefficient, then multiplies the error quantization value by the corresponding weight coefficient, and sums the two products to obtain the sensor health status score, which directly reflects the current operational reliability of the sensor.

[0083] S32: Perform a moving average calculation on the sensor health status score and historical certification records, and generate a dynamic trust discount factor by combining the number of recent successful certifications and sensor stability indicators.

[0084] Specifically, the system retrieves sensor health status scores and historical authentication records. These records contain authentication results and sensor operating parameters for each time period. The system can use a moving average algorithm to process the sensor health status scores. The length of the moving average time window is set according to the temporal distribution characteristics of the authentication business to ensure that it reflects recent trends in sensor status changes. During the moving average calculation, the system sums the health status scores within the window and then divides them by the number of scores within the window to obtain the moving average value of the health status.

[0085] Simultaneously, the system extracts the number of recent successful authentications and sensor stability indicators from historical authentication records. The sensor stability indicator is calculated based on the dispersion of health status scores within a window, with the dispersion obtained by accumulating the deviations of each score from the moving average. The system then integrates the health status moving average, the quantified value of the number of recent successful authentications, and the sensor stability indicator according to a preset ratio to generate a dynamic trust discount factor. The formula for calculating the dynamic trust discount factor is as follows:

[0086]

[0087] in, Let i be the dynamic trust discount factor corresponding to the i-th type of sensor. The moving average of health status This is a quantified value based on the number of recent successful authentications. For sensor stability indicators, , The weighting coefficients are 1, and the sum of the three is 1. The system calculates the weighted sum of each indicator according to the above formula to obtain the dynamic trust discount factor. The value of this factor ranges from 0 to 1, and its magnitude is positively correlated with the sensor's recent operational reliability and certification performance.

[0088] S33: Apply a dynamic confidence discount factor to the basic probability assignment function to perform evidence discounting, scale the probability assignment values ​​of each proposition according to the confidence ratio, and generate a weighted evidence body.

[0089] Specifically, the system acquires the basic probability allocation function and the dynamic confidence discount factor, and adjusts the probability allocation values ​​of each proposition according to the confidence level to reduce the impact of low-confidence sensor data on the fusion results. This process is achieved through the following formula:

[0090]

[0091]

[0092]

[0093] in, , , These are, respectively, the i-th modal evidence for legitimate users (proposition A), illegitimate users (proposition B), and globally uncertain propositions after discounting. The probability allocation value, As a dynamic trust discount factor, , , These are the original basic probability assignment values.

[0094] Preferably, the system first multiplies the dynamic trust discount factor by the original probability allocation values ​​of legitimate user propositions and illegitimate user propositions respectively to obtain the discounted probability allocation value of the corresponding proposition; then, by subtracting the dynamic trust discount factor from 1 and adding the product of the dynamic trust discount factor and the original global uncertain proposition probability allocation value, the system obtains the discounted global uncertain proposition probability allocation value. After the discounting process, the system ensures... This satisfies the normative requirements of the basic probability allocation function and generates a weighted evidence body that incorporates sensor trust information, thereby improving the reliability of the evidence.

[0095] S34: Normalize and weight the dynamic trust discount factor and uncertainty measure, calculate the relative importance of each piece of evidence in the fusion process, generate a dynamic weight set, and integrate the weighted evidence to generate a preprocessed evidence set.

[0096] Specifically, the system calls upon the dynamic confidence discount factor and uncertainty metric of each modality, performs normalized weighting, and determines the relative importance of each piece of evidence in the fusion process. This process is implemented through the following formula:

[0097]

[0098] in, The dynamic weights of the evidence for the i-th modality are... Let i be the dynamic trust discount factor for the i-th mode. Let i be the uncertainty metric for the i-th mode. The total number of certified modalities is represented by the sum of the products of the dynamic confidence discount factor and the corresponding uncertainty measure complement for all modalities.

[0099] Furthermore, the system first calculates the product of the dynamic confidence discount factor and the uncertainty measure complement for each modality, then sums the product results for all modalities to obtain a normalized denominator. The product result for a single modality is divided by the normalized denominator to obtain the dynamic weight for that modality, and the sum of the dynamic weights for all modalities is 1. The dynamic weights reflect the relative importance of each piece of evidence; modalities with higher dynamic confidence discount factors and lower uncertainty measures have larger dynamic weights. The system associates the weighted evidence bodies of each modality with their corresponding dynamic weights to form a dynamic weight set. Then, it integrates all weighted evidence bodies and the dynamic weight set to generate a preprocessed evidence set. This evidence set provides structured and standardized input data for subsequent conflict-adaptive evidence fusion, ensuring that the fusion process fully considers the reliability and importance of each piece of evidence.

[0100] In one embodiment, step S4 of the financial device security authentication method provided by the present invention specifically includes the following steps:

[0101] S41: Perform pairwise orthogonal sum calculation on the weighted evidence body, calculate the intersection and union probabilities of all evidence propositions using Dempster's combination rule, and generate preliminary fusion results and conflict quality.

[0102] Specifically, the system retrieves weighted evidence from the preprocessed evidence set, performs pairwise orthogonal sum calculations in sequence, and synthesizes the evidence according to Dempster's combination rule. The synthesis formula is as follows:

[0103]

[0104] in, Two weighted evidence bodies and The initial probability assignment value for proposition A after synthesis, where X and Y are the proposition sets of the two pieces of evidence, respectively. The intersection operation X∩Y=A means that the corresponding probability assignment values ​​are multiplied and summed only if the intersection of the propositions of the two pieces of evidence is A. The conflict quality K is calculated using the following formula:

[0105]

[0106] in, The propositions of two pieces of evidence have no overlap. The corresponding probability assignment values ​​are multiplied and summed to obtain a quantified value of the degree of conflict. The system first selects any two weighted pieces of evidence and performs orthogonal sum calculation to obtain an intermediate composite piece of evidence. Then, the above calculation is repeated with the next weighted piece of evidence until all weighted pieces of evidence are synthesized pairwise, generating a preliminary fusion result covering all propositions. At the same time, the conflict quality K obtained by the final calculation is recorded. This K value reflects the overall degree of conflict among all weighted pieces of evidence.

[0107] S42: Perform proportional allocation processing on conflict quality and dynamic weight set, allocate conflict values ​​according to the total support of each proposition in the weighted evidence body, and generate conflict allocation function.

[0108] Specifically, the system acquires the conflict quality K and the dynamic weight set in the preprocessed evidence set, and performs conflict ratio allocation processing. The conflict allocation function calculates the conflict allocation value for each proposition according to preset rules. The allocation function for legitimate user propositions is implemented through the formula:

[0109]

[0110] in, This represents the conflict allocation function, which assigns the conflict quality to the value of a legitimate user proposition A. To represent the dynamic weights of the i-th mode, For the first The weighted evidence body of each modality is assigned a value to the legitimate user proposition. Let be the weighted evidence body of the i-th modality, and its assignment value to the illegal user proposition.

[0111] Further, the system first calculates the product of the weight of each modality and the probability assignment value of that modality to A, and sums the products to obtain the weighted total support of A; then it calculates the product of the weight of each modality and the sum of the probability assignment values ​​of that modality to A and B, and sums the products to obtain the joint weighted total support of A and B; the ratio of the two is the proportional coefficient of A in conflict allocation, which is multiplied by the conflict quality K to obtain the conflict allocation value of A. Similarly, the conflict allocation function of the illegal user proposition B is calculated in the same way. Through this process, a complete conflict allocation function is generated, realizing the reasonable allocation of conflict quality according to weight and proposition support. Preferably, the expression of the conflict allocation function of the illegal user proposition B is:

[0112]

[0113] in, Assign a value to proposition B for the legitimate user to determine the conflict quality.

[0114] S43: Perform final synthesis processing on the preliminary fusion results and conflict assignment function, merge the conflict assignment values ​​into the probability assignment of the corresponding proposition, and generate a global fusion evidence body.

[0115] Specifically, the system retrieves the preliminary fusion results and conflict assignment function for final synthesis processing, ensuring that the probability assignment values ​​of each proposition meet the normative requirements of DS evidence theory. The synthesis process is performed separately for each proposition, and the final probability assignment value of legitimate user proposition A is calculated using the formula:

[0116]

[0117] in, This represents the probability assignment value of A to the globally fused evidence body. The probability assignment value for A is the initial fusion result. Assigning values ​​to A for the conflict assignment function. The final probability assignment value for the illegal user proposition B is as follows: Calculate the final probability assignment value of the globally uncertain proposition Ω according to Calculations show that since all conflicting qualities have been allocated to A and B, Ω does not participate in conflict allocation. After the system completes the synthesis of all propositions, verification is performed. To verify its validity, the system ensures that the normative constraints of the basic probability assignment function are met. Upon successful verification, the system generates a globally fused evidence body containing the final probability assignment values ​​for the three propositions A, B, and Ω. This evidence body integrates the valid information from all weighted evidence bodies and effectively resolves evidence conflicts, providing precise evidentiary support for subsequent authentication decisions.

[0118] In one embodiment, step S5 of the financial device security authentication method provided by the present invention specifically includes the following steps:

[0119] S51: Calculate the trust function and likelihood function for the globally fused evidence body, extract the minimum support probability and maximum probability for the legitimate user proposition, and generate the trust interval.

[0120] Specifically, the system calculates the trust function and likelihood function based on the basic probability allocation values ​​in the globally fused evidence body, adhering primarily to the confidence measurement rules of the DS evidence theory. The trust function is calculated using the following formula:

[0121]

[0122] in, The trust function value for proposition A of legitimate users is obtained by globally fusing evidence. Let be the basic probability assignment value for proposition A of the legitimate user in the globally fused evidence body. This value directly reflects the minimum support probability of the system for the legitimate user's identity. The formula for calculating the likelihood function is as follows:

[0123]

[0124] in, The likelihood function value of globally fused evidence for proposition A of legitimate users. For globally uncertain propositions in a globally fused evidence body The basic probability allocation value reflects the system's maximum possible probability of supporting a legitimate user's identity. The system first extracts the basic probability allocation values ​​of the legitimate user proposition and the globally uncertain proposition from the global fused evidence body, substitutes them into the two formulas mentioned above to complete the calculation, and then constructs a trust interval by using the trust function value as the lower limit of the interval and the likelihood function value as the upper limit of the interval. The span of this interval directly corresponds to the degree of uncertainty of the authentication result.

[0125] S52: Perform decision factor synthesis processing on the trust interval, and calculate the weighted sum of trust and uncertainty by combining the system security policy preference coefficient to generate the global authentication confidence.

[0126] Specifically, the system calls the upper and lower limits of the trust interval and combines them with preset system security policy preference coefficients to synthesize decision factors, thereby integrating trust and uncertainty information. The formula for calculating the global authentication confidence level is as follows:

[0127]

[0128] in, For global authentication confidence, The system security policy preference coefficient, determined by the security policy configuration of the financial device, is used to balance the explicit support reflected in the trust function with the risk compensation corresponding to uncertainty. The system first obtains the preset preference coefficient in the security policy, calculates the difference between the likelihood function and the trust function to obtain a measure of uncertainty in the authentication result, and then subtracts this uncertainty measure from 1 to obtain the uncertainty compensation coefficient. Subsequently, the trust function value is multiplied by the preference coefficient, and the uncertainty compensation coefficient is multiplied by (1 - preference coefficient). The two products are then summed to obtain the global authentication confidence score. This confidence score comprehensively considers explicit trust and potential uncertainty, forming a single-dimensional quantitative authentication indicator.

[0129] S53: Perform multi-dimensional risk assessment processing on transaction environment data and historical risk control records, analyze the risk level of transaction amount, geographical location and behavior pattern, and generate dynamic risk thresholds.

[0130] Specifically, the system collects current transaction environment data and historical risk control records to comprehensively analyze key factors affecting transaction security. Transaction environment data includes the current transaction amount, the geographical location of the transaction, and the user's current operational behavior patterns. Historical risk control records include records of risk events from past transactions and records of historical device authentication anomalies. The system performs risk level analysis on transaction amounts, classifying risk levels based on the potential loss corresponding to the amount; it assesses the risk of geographical location information, comparing the current location with the user's historically frequently used transaction areas to analyze the risks arising from location differences; and it analyzes behavioral pattern data, comparing the interaction frequency and command sequence of current operations with the degree of consistency with the user's historical behavioral patterns to identify abnormal behavioral characteristics.

[0131] Furthermore, the system integrates the risk assessment results from the three dimensions according to preset weights. The weight allocation is determined based on the priority of each dimension's impact on transaction security, and a comprehensive risk assessment value is obtained through weighted summation. Based on this comprehensive risk assessment value, the system generates a dynamic risk threshold. The threshold value is positively correlated with the change in the comprehensive risk assessment value, ensuring that the higher the risk level, the stricter the threshold, providing a judgment standard adapted to the current transaction risk for certification decisions.

[0132] S54: Perform a numerical comparison between the global authentication confidence level and the dynamic risk threshold, and map the Boolean judgment result of whether the confidence level exceeds the threshold to the device execution instruction to generate a binary authentication instruction.

[0133] Specifically, the system obtains the quantified values ​​of the global authentication confidence level and the dynamic risk threshold, and performs numerical comparison processing. First, it compares the global authentication confidence level with the dynamic risk threshold to determine whether the confidence level exceeds the threshold. When the global authentication confidence level is greater than or equal to the dynamic risk threshold, the Boolean judgment result is true; when the global authentication confidence level is less than the dynamic risk threshold, the Boolean judgment result is false. The system establishes a mapping relationship between the Boolean judgment result and the device execution instruction: a true result corresponds to a transaction permission instruction, and a false result corresponds to a transaction termination instruction.

[0134] Based on this mapping relationship, the system converts the Boolean judgment result into binary code, generating a binary authentication command. One code corresponds to transaction permission, and the other to transaction termination. This binary authentication command is transmitted to the transaction control unit via the secure communication link of the financial device. The encoding format of the command conforms to the communication protocol requirements of the device control interface, ensuring that the transaction control unit can accurately parse and execute the corresponding operation, thereby achieving secure control over the financial device's transaction process.

[0135] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0136] Based on the same inventive concept, this application also provides a financial device security authentication system for implementing the aforementioned financial device security authentication method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the financial device security authentication system provided below can be found in the limitations of the financial device security authentication method described above, and will not be repeated here.

[0137] Preferably, such as Figure 3 As shown, the present invention provides a financial device security authentication system 600, which is configured with the following modules:

[0138] The multimodal feature matching module 610 is used to extract and match features from user data collected from at least two authentication modalities, calculate the similarity and quality of the extracted features with the pre-stored registration template, and generate a local matching score and feature quality score for each modality.

[0139] The uncertainty evidence conversion module 620 is used to quantify uncertainty and convert evidence into local matching scores and feature quality scores. It calculates the uncertainty measure of each modality by weighted fusion of score values ​​and quality values, and converts the matching results into a standard evidence structure based on the measure, generating a basic probability allocation function for each modality.

[0140] The evidence dynamic weighting module 630 is used to dynamically weight the basic probability allocation function based on sensor status and historical performance. It combines real-time collected sensor health indicators and long-term statistical authentication success rate to assign weight coefficients to each piece of evidence, generating a preprocessed evidence set containing weighted evidence bodies and dynamic weight sets.

[0141] The conflict adaptive fusion module 640 is used to perform conflict adaptive evidence fusion on the weighted evidence body in the preprocessed evidence set and the dynamic weight set. It generates a global fused evidence body by calculating the conflict quality between the evidence and redistributing the conflict values ​​to each proposition according to the weight ratio.

[0142] The fusion evidence authentication decision module 650 is used to perform authentication decision processing on the global fusion evidence body, calculate the confidence level of the fusion evidence for the legitimate user proposition, compare it with the preset dynamic risk threshold, and generate binary authentication instructions. The binary authentication instructions are used to control the financial device to execute or terminate transactions.

[0143] Preferably, the multimodal feature matching module 610 provided in this application is configured with the following units:

[0144] The fingerprint feature extraction unit is used to perform grayscale and noise filtering on the image data acquired from the fingerprint recognition modality, and to construct the fingerprint feature point topology through orientation field estimation and ridge extraction algorithms to generate fingerprint feature vectors.

[0145] The fingerprint matching score unit is used to perform detailed feature matching processing on the fingerprint feature vector and the pre-stored fingerprint registration template, calculate the feature point coincidence degree under the minimum rotation and translation parameters, and generate the fingerprint local matching score.

[0146] The fingerprint quality scoring unit is used to perform sharpness and contrast analysis on the original images acquired by the fingerprint recognition modality, evaluate the effective area ratio and ridge continuity of the image, and generate a fingerprint feature quality score.

[0147] The voiceprint feature extraction unit is used to perform frame segmentation and pre-emphasis processing on the audio data collected from the voiceprint recognition modality, and to extract the vocal tract and vocal cord excitation features through linear predictive coding cepstral coefficients to generate voiceprint feature vectors.

[0148] The voiceprint matching score unit is used to perform dynamic time warping matching processing on the voiceprint feature vector and the pre-stored voiceprint registration template, calculate the minimum cumulative distance between the test speech and the registered speech in the feature space, and generate the local voiceprint matching score.

[0149] The voiceprint quality scoring unit is used to perform signal-to-noise ratio and harmonic analysis on the raw audio collected by the voiceprint recognition modality, evaluate the background noise intensity and fundamental frequency trajectory stability, and generate a voiceprint feature quality score.

[0150] Preferably, the uncertainty evidence conversion module 620 provided in this application is configured with the following units:

[0151] The score-quality weighted fusion unit is used to perform linear weighted fusion processing on the local matching score and the feature quality score, and to calculate the confidence of the matching score by discounting it in combination with the quality score, thereby generating an uncertainty metric for each modality.

[0152] The basic probability allocation unit is used to perform basic probability allocation processing on the uncertainty metric and the local matching score, and to allocate the matching score to the legitimate user and illegitimate user propositions according to the uncertainty ratio, thereby generating the initial probability allocation.

[0153] The probability allocation normalization unit is used to normalize the uncertainty of the initial probability allocation, allocate the remaining probability mass to the globally uncertain proposition, complete the evidence body construction, and generate the basic probability allocation function.

[0154] Preferably, the evidence dynamic weighting module 630 provided in this application is configured with the following units:

[0155] The real-time sensor status monitoring unit is used to monitor the fingerprint sensor and microphone in real time, collect signal strength and self-test error code data, and generate a sensor health status score.

[0156] The dynamic trust calculation unit is used to perform a moving average calculation on the sensor health status score and historical authentication records, and to generate a dynamic trust discount factor by combining the number of recent successful authentications and the sensor stability index.

[0157] The evidence confidence discount unit is used to apply a dynamic confidence discount factor to the basic probability assignment function to perform evidence discounting, scaling the probability assignment values ​​of each proposition according to the confidence ratio, and generating a weighted evidence body.

[0158] The dynamic weight generation and evidence integration unit is used to normalize and weight the dynamic trust discount factor and uncertainty measure, calculate the relative importance of each piece of evidence in the fusion process, generate a dynamic weight set, and integrate the weighted evidence to generate a preprocessed evidence set.

[0159] Preferably, the conflict adaptive fusion module 640 provided in this application is configured with the following units:

[0160] The evidence orthogonal sum calculation and conflict quality generation unit is used to perform pairwise orthogonal sum calculation on the weighted evidence body, calculate the intersection and union probabilities of all evidence propositions through Dempster's combination rule, and generate preliminary fusion results and conflict quality.

[0161] The conflict quality proportional allocation unit is used to proportionally allocate conflict quality and dynamic weight set. It allocates conflict values ​​according to the total support of each proposition in the weighted evidence body and generates a conflict allocation function.

[0162] The final synthesis unit of the fusion result is used to perform final synthesis processing on the preliminary fusion result and the conflict assignment function, and merge the conflict assignment value into the probability assignment of the corresponding proposition to generate the global fusion evidence body.

[0163] Preferably, the fusion evidence authentication decision module 650 provided in this application is configured with the following units:

[0164] The trust likelihood function calculation unit is used to perform trust function and likelihood function calculations on the global fused evidence body, extract the minimum support probability and maximum probability of the legitimate user proposition, and generate the trust interval.

[0165] The decision factor synthesis and confidence generation unit is used to synthesize decision factors for the trust interval, and calculate the weighted sum of trust and uncertainty by combining the system security policy preference coefficient to generate the global authentication confidence.

[0166] The multi-dimensional risk assessment and dynamic threshold generation unit is used to perform multi-dimensional risk assessment on transaction environment data and historical risk control records, analyze the risk level of transaction amount, geographical location and behavior pattern, and generate dynamic risk thresholds.

[0167] The confidence threshold comparison and authentication instruction generation unit is used to perform numerical comparison processing between the global authentication confidence level and the dynamic risk threshold, and to map the Boolean judgment result of whether the confidence level exceeds the threshold into a device execution instruction to generate a binary authentication instruction.

[0168] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described financial device security authentication method.

[0169] In one embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described financial device security authentication method.

[0170] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0171] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0172] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for security authentication of financial devices, characterized in that, Includes the following steps: S1: Extract and match features from user data collected from at least two authentication modalities, calculate the similarity and quality of the extracted features with the pre-stored registration templates, and generate a local matching score and feature quality score for each modality. S2: Perform uncertainty quantification and evidence transformation on the local matching score and the feature quality score, obtain the uncertainty measure of each modality by weighted fusion of score value and quality value, and convert the matching result into a standard evidence structure based on the measure to generate the basic probability allocation function of each modality; S3: The basic probability allocation function is dynamically weighted based on sensor status and historical performance. Combined with real-time collected sensor health indicators and long-term statistical authentication success rate, each piece of evidence is assigned a weight coefficient to generate a preprocessed evidence set containing a weighted evidence body and a dynamic weight set. S4: Perform conflict-adaptive evidence fusion on the weighted evidence body and dynamic weight set in the preprocessed evidence set. Calculate the conflict quality between evidence and redistribute the conflict values ​​to each proposition according to the weight ratio to generate a globally fused evidence body. S5: Perform authentication decision processing on the globally fused evidence body, calculate the confidence level of the fused evidence for the legitimate user proposition, compare it with the preset dynamic risk threshold, and generate a binary authentication instruction. The binary authentication instruction is used to control the financial device to execute or terminate the transaction.

2. The method according to claim 1, characterized in that, The local matching score includes a fingerprint local matching score and a voiceprint local matching score, and the feature quality score includes a fingerprint feature quality score and a voiceprint feature quality score. S1 includes: S11: Perform grayscale conversion and noise filtering on the image data acquired from the fingerprint recognition modality, construct the fingerprint feature point topology through orientation field estimation and ridge extraction algorithms, and generate fingerprint feature vectors; S12: Perform detailed feature matching processing on the fingerprint feature vector and the pre-stored fingerprint registration template, calculate the feature point overlap degree under the minimum rotation and translation parameters, and generate a local fingerprint matching score; S13: Perform sharpness and contrast analysis on the original image acquired by the fingerprint recognition modality, evaluate the effective area ratio and ridge continuity of the image, and generate a fingerprint feature quality score. S14: Perform frame segmentation and pre-emphasis processing on the audio data collected from the voiceprint recognition modality, extract the vocal tract and vocal cord excitation features through linear predictive coding cepstral coefficients, and generate voiceprint feature vectors; S15: Perform dynamic time warping matching processing on the voiceprint feature vector and the pre-stored voiceprint registration template, calculate the minimum cumulative distance between the test speech and the registered speech in the feature space, and generate a local voiceprint matching score. S16: Perform signal-to-noise ratio and harmonicity analysis on the original audio collected by the voiceprint recognition modality, evaluate the background noise intensity and fundamental frequency trajectory stability, and generate a voiceprint feature quality score.

3. The method according to claim 1, characterized in that, S2 includes: S21: Perform linear weighted fusion processing on the local matching score and the feature quality score, and calculate the confidence of the matching score by discounting it in combination with the quality score to generate an uncertainty metric for each modality; S22: Perform basic probability allocation processing on the uncertainty metric and the local matching score, and allocate the matching score to the propositions of legitimate users and illegitimate users according to the uncertainty ratio to generate an initial probability allocation; S23: Perform uncertainty normalization processing on the initial probability allocation, allocate the remaining probability mass to the globally uncertain proposition, complete the evidence body construction, and generate the basic probability allocation function.

4. The method according to claim 1, characterized in that, S3 includes: S31: Real-time status monitoring of the fingerprint sensor and microphone, collection of signal strength and self-test error code data, and generation of sensor health status score; S32: Perform a moving average calculation on the sensor health status score and historical authentication records, and generate a dynamic trust discount factor by combining the recent number of successful authentications and the sensor stability index. S33: Apply the dynamic confidence discount factor to the basic probability allocation function to perform evidence discounting, scale the probability allocation value of each proposition according to the confidence ratio, and generate a weighted evidence body; S34: Normalize and weight the dynamic trust discount factor and the uncertainty metric, calculate the relative importance of each piece of evidence in the fusion process, generate a dynamic weight set, and integrate the weighted evidence body to generate a preprocessed evidence set.

5. The method according to claim 1, characterized in that, S4 includes: S41: Perform pairwise orthogonal sum calculation on the weighted evidence body, calculate the intersection and union probabilities of all evidence propositions using the Dempster combination rule, and generate preliminary fusion results and conflict quality; S42: Perform proportional allocation processing on the conflict quality and the dynamic weight set, and allocate conflict values ​​according to the total support of each proposition in the weighted evidence body, thereby generating a conflict allocation function; S43: Perform final synthesis processing on the preliminary fusion result and the conflict allocation function, merge the conflict allocation value into the probability allocation of the corresponding proposition, and generate a global fusion evidence body.

6. The method according to claim 1, characterized in that, The expression for the conflict allocation function is: in, This represents the conflict assignment function, that is, the value assigned by the conflict assignment function to legitimate user propositions. To represent the dynamic weights of the i-th mode, For the first The weighted evidence body of each modality is assigned a value to the legitimate user proposition. Let be the weighted evidence body of the i-th modality, and its assignment value to the illegal user proposition.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: Calculate the trust function and likelihood function for the global fusion evidence body, extract the minimum support probability and maximum probability for the legitimate user proposition, and generate a trust interval; S52: Perform decision factor synthesis processing on the trust interval, and calculate the weighted sum of trust and uncertainty by combining the system security policy preference coefficient to generate the global authentication confidence. S53: Perform multi-dimensional risk assessment processing on transaction environment data and historical risk control records, analyze the risk level of transaction amount, geographical location and behavior pattern, and generate the dynamic risk threshold; S54: Perform a numerical comparison between the global authentication confidence level and the dynamic risk threshold, and map the Boolean judgment result of whether the confidence level exceeds the threshold into a device execution instruction to generate a binary authentication instruction.

8. A security authentication system for financial devices, characterized in that, The system includes: The multimodal feature matching module is used to extract and match features from user data collected from at least two authentication modalities, calculate the similarity and quality of the extracted features with the pre-stored registration template, and generate a local matching score and feature quality score for each modality. The uncertainty evidence conversion module is used to perform uncertainty quantification and evidence conversion on the local matching score and the feature quality score. It calculates the uncertainty measure of each modality by weighted fusion of the score and quality value, and converts the matching result into a standard evidence structure based on the measure, generating a basic probability allocation function for each modality. The evidence dynamic weighting module is used to dynamically weight the basic probability allocation function based on sensor status and historical performance, and combine the real-time collected sensor health indicators and the long-term statistical authentication success rate to assign weight coefficients to each piece of evidence, generating a preprocessed evidence set containing weighted evidence bodies and dynamic weight sets. The conflict adaptive fusion module is used to perform conflict adaptive evidence fusion on the weighted evidence body and dynamic weight set in the preprocessed evidence set. It generates a global fused evidence body by calculating the conflict quality between evidence and redistributing the conflict values ​​to each proposition according to the weight ratio. The fusion evidence authentication decision module is used to perform authentication decision processing on the globally fused evidence body, calculate the confidence level of the fused evidence for legitimate user propositions, compare it with a preset dynamic risk threshold, and generate a binary authentication instruction. The binary authentication instruction is used to control the financial device to execute or terminate a transaction.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images