A large model-based risk monitoring management system and method
By using multimodal feature fusion and dimensionality reduction modeling, the problem of easy forgery of individual biometric features is solved, achieving efficient and reliable identity verification, which is applicable to scenarios such as financial payment and government services.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU RED NET TECH CO LTD
- Filing Date
- 2026-01-28
- Publication Date
- 2026-06-02
Smart Images
Figure CN122134099A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of risk monitoring technology, and more specifically, to a risk monitoring management system and method based on a large model. Background Technology
[0002] In scenarios with extremely high requirements for the security and reliability of identity authentication, such as financial payments, government services, and smart access control, the theft, forgery, or impersonation of user identity information may lead to serious problems such as property loss, leakage of permissions, and security risks. Therefore, risk monitoring is necessary to prevent unauthorized users from breaking through the identity authentication defenses by using false biometrics.
[0003] Existing methods typically employ dedicated sensors to periodically collect raw biometric data belonging to the same user, such as fingerprints, faces, irises, and voiceprints. Then, feature extraction algorithms are used to extract biometric vectors for each biometric feature (such as the minutiae vector of fingerprints, the depth vector of faces, the texture vector of irises, and the Mel-frequency cepstral coefficient vector of voiceprints). The similarity between different biometric vectors at the same time and the baseline biometric vectors retained during the registration phase of the corresponding legitimate user is calculated. The identity is then verified one by one, thus initially achieving identity verification and risk screening.
[0004] However, in the process of verifying different biometric vectors one by one, although the method of verifying multiple biometric vectors one by one can theoretically increase the difficulty of identity theft, criminals may still try to break through the verification defense by forging a single modality of biometrics (such as copying fingerprint templates, using high-definition facial photos, etc.).
[0005] However, due to the large dimensional differences and strong modal independence of individual biometric vectors, and the relatively isolated verification standards for different modal features, there is a lack of comprehensive verification of the correlation between multimodal features. At the same time, individual biometric vectors are susceptible to interference from the collection environment and the upgrading of forgery technology, and are at risk of being illegally tampered with or counterfeited. Therefore, criminals can bypass the overall identity verification mechanism by forging biometric vectors of some modalities to pass through individual verification links, thereby causing serious risks such as the theft of user identity information, infringement of property security, and illegal abuse of government service authority. This cannot meet the high requirements for identity authentication security and reliability in scenarios such as financial payment and government services.
[0006] In view of this, we propose a risk monitoring and management system and method based on a large model. Summary of the Invention
[0007] The purpose of this invention is to solve the above-mentioned problems.
[0008] To achieve the above objectives, the present invention provides a risk monitoring and management system based on a large model, comprising:
[0009] The multimodal feature fusion module is used to convert multiple biometric vectors of a legitimate user into a unique initial fusion vector;
[0010] The dimensionality reduction modeling and feature library module is used to learn the global distribution pattern and feature correlation of the initial fusion vectors of all legitimate users, build a feature dimensionality reduction model, use the feature dimensionality reduction model to transform the initial fusion vectors of legitimate users into benchmark low-dimensional feature vectors, establish a benchmark feature library for legitimate users, and generate the corresponding verification low-dimensional feature vectors for the users to be verified.
[0011] The repeated acquisition feature optimization module constructs an initial fusion vector corresponding to the user to be verified by using an element-level fusion method when repeatedly acquiring the same type of biometric features; and calculates the difference in dimension vectors between the same feature dimensions, and uses the difference in dimension vectors to adjust the feature dimensionality reduction model to generate the result of verifying the low-dimensional feature vector.
[0012] As a further improvement to this technical solution, the multimodal feature fusion module receives multiple biometric vectors from all legitimate users and concatenates all biometric vectors of each legitimate user in a fixed order to obtain the corresponding initial fusion vector.
[0013] As a further improvement to this technical solution, the dimensionality reduction modeling and feature library module uses the initial fusion vector in the multimodal feature fusion module to build a feature dimensionality reduction model. Specifically, for the same feature dimension in all initial fusion vectors, the feature mean of the same feature dimension is calculated; the feature value of each feature dimension in each initial fusion vector is retrieved, the feature mean of the corresponding feature dimension is matched, the difference between the feature value and the feature mean is calculated to obtain the corresponding central feature value, and the initial fusion vector is transformed into the corresponding centralized fusion vector.
[0014] The centralized fusion vectors of all legitimate users together constitute a multi-user fusion vector set;
[0015] Calculate the covariance matrix of the multi-user fusion vector set;
[0016] Decompose the direction vector and information weights for each pair of related dimensions in the covariance matrix;
[0017] Set the target dimension reduction dimension, select the direction vector corresponding to the top information weight, and concatenate the selected direction vectors to form the projection matrix.
[0018] The central eigenvalue of each feature dimension of the single-user fusion vector is multiplied by the transpose of the projection matrix to obtain the dimensionality-reduced baseline low-dimensional feature vector.
[0019] As a further improvement to this technical solution, the dimensionality reduction modeling and feature library module generates a verification low-dimensional feature vector corresponding to the user to be verified: it retrieves multiple biometric feature vectors corresponding to the user to be verified, and the multimodal feature fusion module constructs multiple biometric feature vectors as corresponding initial fusion vectors; it outputs the initial fusion vector to the feature dimensionality reduction model after it has been built, and the feature dimensionality reduction model uses the feature mean to transform the initial fusion vector into a centered fusion vector, and finally uses the projection matrix transpose to obtain the verification low-dimensional feature vector corresponding to the user to be verified.
[0020] As a further improvement to this technical solution, the covariance matrix in the dimensionality reduction modeling and feature library module is specifically as follows:
[0021] Retrieve any two feature dimensions with different physical meanings and data distribution characteristics from the multi-user fusion vector set to form a pair of related dimensions; use the central eigenvalues corresponding to the feature dimensions to calculate the covariance between each pair of related dimensions, which is used to quantify the degree of linear correlation between the two feature dimensions. If the absolute value of the covariance is larger, it indicates that the overlap between the two feature dimensions is higher and the redundancy is stronger. If the covariance is close, it indicates that the two feature dimensions are relatively independent.
[0022] As a further improvement to this technical solution, the target dimensionality reduction dimension is set in the dimensionality reduction modeling and feature library module:
[0023] An initial dimension reduction dimension equal to the number of direction vectors is preset, and a dimension feature vector is constructed. The initial dimension reduction dimension is gradually reduced to obtain multiple temporary dimension reduction dimensions until the temporary dimension reduction dimension equals 1. Each time the dimension reduction is reduced, the direction vector with the smallest information weight is removed, and the dimension feature vector is reconstructed based on the remaining direction vectors.
[0024] After the reduction, the information weights corresponding to each directional vector in the temporary dimensionality reduction dimension are calculated again. Specifically, the scalars corresponding to each directional vector are added together and used as the denominator. The scalars of each directional vector are then used as the numerators, and the numerators are divided by the denominators to obtain the information weights of the corresponding directional vectors.
[0025] Cross-calculate the vector similarity between the two feature vectors, and select the temporary dimension reduction dimension that minimizes the vector similarity and makes them all different as the target dimension reduction dimension.
[0026] As a further improvement to this technical solution, the repeated acquisition feature optimization module includes a fusion vector construction unit and a dimensionality reduction weight adjustment unit;
[0027] The fusion vector construction unit uses an element-level fusion method to construct the initial fusion vector corresponding to the user to be verified. Specifically, it receives multiple biometric vectors corresponding to biometric A, compares the feature vectors with the same feature dimension among the multiple biometric vectors, selects the smallest feature vector of each feature dimension among the multiple biometric vectors to form a unique biometric vector of biometric A, and constructs the corresponding initial fusion vector through the biometric vector.
[0028] The dimension reduction weight adjustment unit calculates the difference in dimension vectors between the same feature dimensions; defines the same feature dimensions as the core dimensions, increases the information weight of the corresponding direction vector in the projection matrix based on the difference in dimension vectors, and then the feature dimension reduction model uses the increased information weights and the biometric vectors in the fusion vector construction unit to generate the verification low-dimensional feature vector corresponding to the user to be verified.
[0029] As a further improvement to this technical solution, the element-level fusion method in the fusion vector construction unit is specifically as follows:
[0030] Receive the mean values of features corresponding to different feature dimensions in the feature dimensionality reduction model; sequentially calculate the differences between multiple mean vectors of all feature dimensions and their corresponding mean values in each biological feature vector;
[0031] By comparing the differences between multiple mean vectors of the same feature dimension, the feature vector corresponding to the largest mean vector difference is selected as the feature vector of the corresponding feature dimension in the biological feature vector. Then, the feature vectors of all feature dimensions are integrated to form the corresponding biological feature vector.
[0032] As a further improvement to this technical solution, the dimension vector difference in the dimension reduction weight adjustment unit is the difference between the largest and smallest eigenvectors under the same feature dimension; specifically, if the feature dimension... and feature dimensions In the direction vector When inside, the direction vector The directional difference is: + ;
[0033] It calculates the proportion corresponding to the difference in each direction, and defines the proportion of the difference in direction as the amplification factor of the corresponding direction vector. The percentage is: ,in, This includes the number of core dimension direction vectors;
[0034] Using an increase factor Increase direction vector Information weight To obtain information weights The feature dimensionality reduction model then employs information weights. The fusion vector construction unit generates a low-dimensional verification feature vector corresponding to the user to be verified from the biometric feature vector; and information weights are used. After successful identity verification, the user's baseline feature library is updated based on the biometric feature vector corresponding to the multimodal feature fusion module.
[0035] A risk monitoring and management method based on a large model includes the following steps:
[0036] S1. Convert multiple biometric vectors of legitimate users into a unique initial fusion vector;
[0037] S2. Build a feature dimensionality reduction model, use the feature dimensionality reduction model to transform the initial fusion vector of the legitimate user into the benchmark low-dimensional feature vector, establish a benchmark feature library of legitimate users, and generate the corresponding verification low-dimensional feature vector of the user to be verified.
[0038] S3. When repeatedly collecting the same type of biometric features, use the element-level fusion method to construct the initial fusion vector corresponding to the user to be verified; and calculate the difference in dimension vectors between the same feature dimensions, and use the difference in dimension vectors to adjust the feature dimensionality reduction model to generate the result of verifying the low-dimensional feature vector.
[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0040] In this risk monitoring and management system and method based on a large model, a multimodal feature fusion module transforms multiple biometric vectors of legitimate users into a unique initial fusion vector. This eliminates the problems of significant differences in dimensionality and strong heterogeneity in numerical distribution and physical meaning of biometric vectors from different modalities, and avoids the situation of inconsistent training data formats and information overlap and redundancy caused by the dispersion and independence of multimodal features. Subsequently, the dimensionality reduction modeling and feature library module uses the initial fusion vector from the multimodal feature fusion module to build a feature dimensionality reduction model. Through the feature dimensionality reduction model, a benchmark feature library for legitimate users is established, and a verification low-dimensional feature vector corresponding to the user to be verified is generated. This can effectively remove feature redundancy and reduce data dimensionality while retaining core identity features, avoiding the risk of overfitting of the feature dimensionality reduction model caused by the curse of dimensionality. At the same time, it achieves feature dimension alignment and standardization, providing an efficient and unified feature carrier for subsequent similarity calculation. This can significantly improve the computational efficiency and robustness of identity verification and meet the real-time verification needs of scenarios such as financial payment and smart access control.
[0041] To address the issue of repeated collection of the same biometric vectors due to factors such as device response delays and accidental touches during biometric data acquisition, the fusion vector construction unit employs an element-level fusion method to construct the biometric vector corresponding to the user to be verified. This method can accurately select the key information with the highest identity differentiation among each feature dimension, offsetting random collection errors and avoiding distortion of the fusion vector caused by multiple sets of repetitive data. The dimensionality reduction weight adjustment unit further adjusts the information weights required for verifying low-dimensional feature vectors in the feature dimensionality reduction model. This solves the pain point of traditional dimensionality reduction models that prioritize retaining globally high-variance features and easily suppress local effective identity biases. It dynamically strengthens the influence of the directional vectors corresponding to the core dimensions and weakens the weights of irrelevant feature dimensions.
[0042] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description
[0043] Figure 1 This is a schematic diagram of the overall module of the present invention;
[0044] Figure 2 This is a schematic diagram of the repeatable acquisition feature optimization module of the present invention.
[0045] The meanings of the labels in the diagram are as follows:
[0046] 100. Multimodal feature fusion module; 200. Dimensionality reduction modeling and feature library module; 300. Repeated feature acquisition optimization module; 310. Fusion vector construction unit; 320. Dimensionality reduction weight adjustment unit. Detailed Implementation
[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0048] refer to Figure 1 , Figure 2 As shown, a risk monitoring and management system based on a large model includes:
[0049] The multimodal feature fusion module 100 is used to convert multiple biometric vectors of a legitimate user into a unique initial fusion vector;
[0050] The dimensionality reduction modeling and feature library module 200 is used to learn the global distribution pattern and feature correlation of the initial fusion vectors of all legitimate users, build a feature dimensionality reduction model, use the feature dimensionality reduction model to transform the initial fusion vectors of legitimate users into benchmark low-dimensional feature vectors, establish a benchmark feature library for legitimate users, and generate the corresponding verification low-dimensional feature vectors for the users to be verified.
[0051] When repeatedly collecting the same type of biometric features, the repeated feature acquisition optimization module 300 uses an element-level fusion method to construct an initial fusion vector corresponding to the user to be verified; and calculates the difference in dimension vectors between the same feature dimensions, and uses the difference in dimension vectors to adjust the feature dimensionality reduction model to generate the result of verifying the low-dimensional feature vector.
[0052] In the implementation of the above embodiments, the multimodal feature fusion module 100 converts multiple biometric vectors of legitimate users into a unique initial fusion vector, realizing the initial integration and standardization of multimodal features and providing a unified data foundation for subsequent model training and feature processing. The dimensionality reduction modeling and feature library module 200 learns the global distribution law and feature correlation based on the initial fusion vector, builds a feature dimensionality reduction model, and converts and generates a benchmark low-dimensional feature vector and a verification low-dimensional feature vector. This removes feature redundancy, reduces computational costs, and retains core identity recognition, establishing a reliable benchmark feature library. When repeatedly collecting the same type of biometric features, the repeated collection feature optimization module 300 constructs an accurate initial fusion vector through element-level fusion, and then adjusts the output result of the dimensionality reduction model based on the difference in dimensional vectors of the same feature dimension. This is used to solve the problems of multimodal feature heterogeneity, high-dimensional redundancy, and effective deviation reduction caused by repeated collection, significantly improving the accuracy, robustness, and efficiency of identity verification.
[0053] In scenarios with extremely high requirements for identity authentication security and reliability, such as financial payments, government services, and smart access control, the theft, forgery, or impersonation of user identity information can lead to serious problems such as property loss, access control leakage, and national security risks. While multi-source biometric recognition can integrate complementary information from modalities such as fingerprints, faces, and irises to improve security, different modal biometric vectors exhibit significant dimensional differences and strong heterogeneity in numerical distribution and physical meaning. Furthermore, individual biometric features are easily forged. Therefore, to avoid the problem of isolated verification standards and lack of cross-modal correlation verification during the verification of multiple biometric vectors, which allows criminals to successfully verify by forging some modal features, the multi-modal feature fusion module 100 is used to convert multiple biometric vectors of legitimate users into a unique initial fusion vector. This eliminates information redundancy between multiple biometric vectors, strengthens feature complementarity, and integrates scattered identity feature information using a unique initial fusion vector. This retains key data reflecting the core recognizability of the user's identity from multiple biometric vectors while avoiding the limitations of a single biometric vector due to interference from the collection environment and differences in device accuracy. The specific working principle is as follows:
[0054] The system receives multiple biometric vectors from all legitimate users and concatenates them in a fixed order (e.g., fingerprint vector, face depth vector, iris texture vector, acoustic vector) to obtain the corresponding initial fusion vector. The expression for the initial fusion vector of the k-th legitimate user is as follows:
[0055] ;
[0056] in, This represents the biometric vector of the kth legitimate user and the mth type (where m is the category number of the biometric feature, such as m=1 for fingerprints, m=2 for faces, etc., for a total of M types of biometric features). The transpose of the biometric feature vector of the k-th legitimate user and the m-th class (converting column vectors to row vectors for easier concatenation; D is the total dimension of the initial fused vector, satisfying...) ; It belongs to the D-dimensional real space.
[0057] When the multimodal feature fusion module 100 splices all the biometric vectors of each legitimate user in a fixed order, it is used to integrate the multiple scattered biometric vectors into a unique initial fusion vector. In this way, while preserving multiple biometric vectors to the greatest extent, it lays the foundation for subsequent feature dimensionality reduction and the construction of a feature dimensionality reduction model, and avoids multiple biometric vectors being scattered and independent.
[0058] Furthermore, by using the initial fusion vector, when building the feature dimensionality reduction model in the dimensionality reduction model and feature library module 200, the feature dimensionality reduction model can be used to learn the global distribution pattern and feature correlation relationship between multiple biometric vectors in the initial fusion vector. This eliminates the limitations of a single biometric vector being easily affected by the collection environment and having weak anti-spoofing ability. Thus, while retaining the core identity recognition, feature redundancy is accurately removed and data dimensionality is reduced. This can avoid the risk of model overfitting caused by the curse of dimensionality and achieve security.
[0059] Furthermore, when building a feature dimensionality reduction model in the dimensionality reduction model and feature library module 200, the feature dimensionality reduction model is used to learn the global distribution pattern and feature correlation between multiple biometric vectors in the initial fusion vector. Since the initial fusion vector has integrated the complementary information of multiple biometric vectors such as fingerprint, face, and iris, it breaks the isolation of a single biometric vector. This makes the feature dimensionality reduction model no longer rely on a single biometric vector, but judges based on the inherent correlation of the initial fusion vector. Even if an illegal user forges a single modality feature, it cannot meet the inherent correlation verification requirements between multimodal features. This eliminates the limitations of a single biometric vector being easily affected by the collection environment and having weak anti-forgery ability. While retaining the core identity recognition, it accurately removes feature redundancy and reduces data dimensionality, which can avoid the risk of model overfitting caused by the curse of dimensionality and achieve security.
[0060] Regarding the specific splicing order, since the feature dimensionality reduction model in the dimensionality reduction modeling and feature library module 200 can learn the linear correlation and global distribution law between different feature dimensions in the initial fusion vector, the splicing order can be flexibly set based on priorities such as the stability of biometric features, anti-forgery ability, and core discriminability. As long as the splicing order of all legitimate users and users to be verified is consistent, the feature dimensionality reduction model can stably identify the correlation and redundant information between cross-modal features.
[0061] The multimodal feature fusion module 100 contains multiple biometric vectors corresponding to all legitimate users. Each biometric vector is obtained by collecting data from a standardized acquisition device during the registration phase of a legitimate user, parsing it using a multimodal feature extraction algorithm, and optimizing it through a feature preprocessing workflow. Furthermore, each biometric vector corresponds to a different feature dimension, for example:
[0062] The fingerprint feature vector extracted using a deep convolutional network has 128 feature dimensions; the face depth feature vector extracted using FaceNet has 512 feature dimensions; the iris texture feature vector extracted using high-resolution iris coding has 1024 feature dimensions; and the acoustic feature vector extracted using MFCC coefficients has 39 feature dimensions.
[0063] Furthermore, considering that although the multimodal feature fusion module 100 can convert multiple biometric vectors of legitimate users into a unique initial fusion vector, achieving preliminary integration of multimodal features, it only splices the feature vectors of each modality in a fixed order during the conversion process, without effectively handling the information overlap, redundancy, and linear correlation between different modal features. This leads to high-dimensional redundancy problems, resulting in low efficiency when directly verifying identity based on the high-dimensional initial fusion vector. Therefore, the dimensionality reduction modeling and feature library module 200 is used to learn the global distribution pattern and feature correlation of all legitimate users' initial fusion vectors, build a feature dimensionality reduction model with core feature extraction and redundancy removal capabilities, and establish a legitimate user benchmark feature library based on the converted benchmark low-dimensional feature vector; and generate the corresponding verification low-dimensional feature vector for the user to be verified.
[0064] The dimensionality reduction modeling and feature library module 200 uses the initial fusion vectors in the multimodal feature fusion module 100 to build a feature dimensionality reduction model. The specific working principle is as follows: For the same feature dimension in all initial fusion vectors, calculate the feature mean of the same feature dimension; retrieve the feature value of each feature dimension in each initial fusion vector, match the feature mean of the corresponding feature dimension, calculate the difference between the feature value and the feature mean to obtain the corresponding central feature value, and transform the initial fusion vector into the corresponding centralized fusion vector;
[0065] And based on the centralized fusion vectors of all legitimate users, a multi-user fusion vector set is constructed (the multi-user fusion vector set is a D×N matrix (rows correspond to feature dimensions, columns correspond to legitimate users)), the specific expression is:
[0066] ;
[0067] Where N is the total number of legitimate users; the central eigenvalue of the j-th row and k-th column in the multi-user fusion vector set. ; Let be the feature mean of the j-th feature dimension.
[0068] Furthermore, considering the differences in numerical distribution patterns and physical meanings of biometric features originating from different modalities such as fingerprints, faces, irises, and acoustics, leading to information overlap or redundancy, the covariance matrix of the multi-user fusion vector set is calculated based on this set. Specifically:
[0069] Retrieve any two feature dimensions with different physical meanings and data distribution characteristics from the multi-user fusion vector set to form a pair of related dimensions; use the central eigenvalues corresponding to the feature dimensions to calculate the covariance between each pair of related dimensions, which is used to quantify the degree of linear correlation between the two feature dimensions. If the absolute value of the covariance is larger, it indicates that the overlap between the two feature dimensions is higher and the redundancy is stronger. If the covariance is close to 0, it indicates that the two feature dimensions are relatively independent.
[0070] The element at any position in the covariance matrix corresponds to the covariance of the associated dimension pair. Specifically, the element in the i-th row and j-th column of the covariance matrix... The covariance between the i-th feature dimension and the j-th feature dimension in the corresponding multi-user fusion vector set is:
[0071] ;
[0072] in, Let the central feature value of the k-th legitimate user and the i-th feature dimension be denoted as . Let be the central feature value of the kth legitimate user and the jth feature dimension;
[0073] Decompose the direction vector of each correlation dimension pair in the covariance matrix (specifically, the concentrated distribution direction of the correlation dimension pair, clarifying the information aggregation and propagation path of the correlation dimension pair), and the information weight (specifically, the quantitative representation of the linear correlation between the correlation dimension pairs, intuitively reflecting the effective information content carried by the corresponding correlation dimension pair):
[0074] ;
[0075] in:
[0076] The eigenvector matrix, For the corresponding scalar The direction vector;
[0077] It is an eigenvalue diagonal matrix. Let be a scalar of the i-th direction vector, and Specifically, the m-th direction vector The corresponding information weights are: ;
[0078] The above covariance matrix decomposition:
[0079] Based on the eigenvalue decomposition property of real symmetric matrices in linear algebra, the covariance matrix is eigenvalued, and the decomposition process strictly follows... Where C is the covariance matrix, V is the eigenvector matrix, and each column... The feature vector is the direction vector of the correlation dimension. This is an eigenvalue diagonal matrix, with diagonal elements... Let be a scalar of the corresponding eigenvector and satisfy Each direction vector obtained by decomposition is the concentrated distribution direction of the corresponding associated dimension pair. The absolute value of each component reflects the contribution of the corresponding original feature dimension to that direction. The component sign reflects the linear correlation direction of the associated dimension pair (positive sign indicates change in the same direction, negative sign indicates change in opposite direction). The eigenvalue quantifies the total amount of effective information carried by the corresponding direction vector. Then, the information weight corresponding to each direction vector is calculated. The information weight is a quantitative representation of the linear relationship between the associated dimensions. The larger the weight value, the higher the proportion of effective identity information carried by the corresponding associated dimension. Conversely, the lower the weight value, the higher the proportion of redundant information.
[0080] After decomposing the direction vectors and information weights of each correlation dimension pair in the covariance matrix: set the target dimensionality reduction dimension, select the direction vectors corresponding to the top-ranking information weights, and concatenate the selected direction vectors to form the projection matrix. Where K is the target dimensionality reduction dimension.
[0081] The central eigenvalue of each feature dimension of the single-user fusion vector is multiplied by the transpose of the projection matrix to obtain the dimensionality-reduced baseline low-dimensional feature vector. Specifically, the baseline low-dimensional feature vector of the kth legitimate user is: ;
[0082] The target dimensionality reduction dimension in the feature dimensionality reduction model is used to achieve dimensionality compression while preserving the core distinguishing information of biometric features, avoiding redundant data from occupying storage and computing resources. Therefore, when the target dimensionality reduction dimension is equal to the number of direction vectors (i.e., all feature directions are fully preserved), due to the differences in the original distribution and scalar weights of different biometric feature vectors, the multiple low-dimensional feature vectors obtained by transposing the projection matrix will still maintain their uniqueness and will not be completely identical. At the same time, in order to balance the accuracy of identity verification and the system's computational efficiency, and to ensure that the feature vectors after dimensionality reduction do not lose key distinguishing information while minimizing the dimensionality scale to improve processing speed, the working principle of setting the target dimensionality reduction dimension in the dimensionality reduction modeling and feature library module 200 is as follows:
[0083] An initial dimension reduction dimension equal to the number of direction vectors is preset, and a dimension feature vector is constructed. The initial dimension reduction dimension is gradually reduced to obtain multiple temporary dimension reduction dimensions until the temporary dimension reduction dimension equals 1. Each time the dimension reduction is reduced, the direction vector with the smallest information weight is removed, and the dimension feature vector is reconstructed based on the remaining direction vectors.
[0084] After the reduction, the information weights corresponding to each directional vector in the temporary dimensionality reduction dimension are calculated again. Specifically, the scalars corresponding to each directional vector are added together and used as the denominator. The scalars of each directional vector are then used as the numerators, and the numerators are divided by the denominators to obtain the information weights of the corresponding directional vectors.
[0085] Cross-calculate the vector similarity between the two feature vectors. Select the temporary dimension reduction dimension (which can be the initial dimension reduction dimension or any temporary dimension reduction dimension generated during the reduction process) that minimizes the difference in vector similarity between the two vectors as the target dimension reduction dimension. The specific formula for calculating the vector similarity is as follows:
[0086] Let the two feature vectors be:
[0087] Dimensional feature vector 1 / 2 dimensional feature vector ,in If we reduce the dimensionality of the target vector, then the vector similarity is:
[0088] ;
[0089] in, For dimensional feature vectors 1 / 2 dimensional feature vector The inner product; For dimensional feature vectors The L2 norm; For dimensional feature vectors The L2 norm.
[0090] The dimensionality reduction modeling and feature library module 200 generates verification low-dimensional feature vectors corresponding to the users to be verified: multiple biometric feature vectors corresponding to the users to be verified are obtained, and the multimodal feature fusion module 100 constructs multiple biometric feature vectors as corresponding initial fusion vectors; the initial fusion vectors are output to the feature dimensionality reduction model after construction, and the feature dimensionality reduction model uses feature mean transformation to convert the initial fusion vectors into centered fusion vectors. Finally, the projection matrix transpose is used to obtain the verification low-dimensional feature vectors corresponding to the users to be verified.
[0091] The dimensionality reduction modeling and feature library module 200 constructs low-dimensional feature vectors (benchmark low-dimensional feature vectors and validation low-dimensional feature vectors) to remove redundant dimensions in the feature vectors and eliminate linear or non-linear correlations between dimensions while retaining the core distinguishability information of multiple biometric features.
[0092] Furthermore, the feature dimensionality reduction model decomposes the two different feature dimensions in each correlation dimension pair in the covariance matrix. It is not only for the same feature type, but also covers feature dimensions of different feature types such as fingerprint, face, iris, and acoustic. By quantifying the linear correlation between different types of features, it accurately identifies information overlap and redundancy between cross-type features. This allows the feature dimensionality reduction process to break through the limitations of a single modality, achieve deep fusion and effective screening of multi-source cross-modal features, retain core feature dimensions with greater global recognizability, and eliminate interference caused by irrelevant correlations between different modalities. Ultimately, this allows the low-dimensional feature vector to more comprehensively and stably carry the core information of user identity.
[0093] Furthermore, because the two different feature dimensions in the associated dimension pair encompass cross-type dimensions from different modalities such as fingerprints, faces, irises, and acoustics, as well as feature dimensions within the same feature type, and because these feature dimensions originate from different acquisition methods and carry significantly different physical meanings and data distribution patterns, some feature dimensions exhibit high-intensity information overlap and redundancy (large absolute covariance values), while other feature dimensions are relatively independent and carry core identification information (covariance close to 0). The inherent differences in the effective information content among these feature dimensions are precisely quantified into scalars corresponding to each directional vector during the decomposition of the covariance matrix, resulting in scalars corresponding to each directional vector exhibiting… Significant gradient differences exist, and the information weights based on scalar computation decrease accordingly. Furthermore, most of the core user identity identification information is concentrated in a few high-information-weight directional vectors, while low-information-weight directional vectors often correspond to redundant information between cross-type features or minor features that are easily disturbed. Therefore, in setting the target dimensionality reduction dimension, the feature dimensions corresponding to these low-information-weight directional vectors can be selectively removed. Core identity information can be fully preserved without retaining all original feature dimensions, thereby further reducing the number of selected feature dimensions. This maximizes the compression of data scale and reduces storage and computing costs while ensuring the accuracy and robustness of subsequent identity verification and risk monitoring.
[0094] Furthermore, when the multimodal feature fusion module 100 constructs the initial fusion vector corresponding to the user to be verified, due to the random influence of multiple factors such as the user's physiological state, environmental interference, and device accuracy during biometric collection, the same type of biometric features may be repeatedly collected for the same user to be verified (for example, after the user to be verified successfully verifies through basic identity credentials, during biometric collection, the face image may be repeatedly captured due to reasons such as the response delay of the collection device, accidental touch of the operation interface, or the user not leaving the collection area in time, resulting in the generation of multiple face depth feature vectors). This makes it impossible to clearly know the corresponding biometric feature vector required for dimensionality reduction. The purpose of dimensionality reduction in the dimensionality reduction modeling and feature library module 200 is to compress feature vector redundancy and improve computational efficiency while retaining the core recognizability of different feature dimensions in multiple biometric feature vectors. However, when the feature vector values under the same feature dimension are different, if the multimodal feature fusion module 100 constructs multiple different initial fusion vectors corresponding to the user to be verified, the feature dimensionality reduction model will then perform dimensionality reduction on the multiple different initial fusion vectors multiple times to obtain multiple verification low-dimensional feature vectors.
[0095] This necessitates subsequent identity verification based on similarity matching using multiple low-dimensional feature vectors. However, during the verification process, some vectors may match successfully while others fail, leading to a situation where the system cannot provide a clear identity verification result based on a unified matching standard. This not only causes the verification process to stall and degrades the user experience but also risks misjudgment due to conflicting matching results from different vectors. It may misjudge legitimate feature fluctuations of normal users as abnormal and reject the verification, or it may mistakenly approve the verification because some successful results ignore the feature deviations of abnormal users.
[0096] Furthermore, during dimensionality reduction, feature dimensionality reduction models typically follow the core logic of preserving the maximum overall variance of the vectors and weakening local feature fluctuations. Essentially, they map the initial fusion vectors to a low-dimensional space through linear projection, prioritizing the preservation of the common global feature directions of the initial fusion vectors. This naturally suppresses local feature differences caused by repeated collection of data from individual users. Therefore, if the effective deviations contained in multiple initial fusion vectors (i.e., feature differences that reflect the core distinguishability of user identity, such as subtle fluctuations in the core ridge of fingerprints or individual unique deviations in the ratio of eye distance in the face) belong to local low-variance features (i.e., features with smaller information weights when building the feature dimensionality reduction model with the feature library module 200) rather than the high-variance directions of the global data, then the effective deviations will be unreasonably reduced by the global optimization objective of the feature dimensionality reduction model. This leads to a weakening of the identity distinguishability between multiple low-dimensional feature vectors, making it impossible to clearly define which are normal feature fluctuations caused by repeated collections and which are feature deviations of abnormal users. Ultimately, this causes the matching degree judgment standard during user identity verification to lose accuracy, resulting in a significant increase in the probability of normal users being misjudged as abnormal or abnormal users being misjudged as normal.
[0097] Therefore, to avoid the above situation, when the biometric feature A corresponding to the user to be verified is repeatedly collected, the repeated collection feature optimization module 300 includes a fusion vector construction unit 310 for constructing the initial fusion vector corresponding to the user to be verified, and a dimensionality reduction weight adjustment unit 320 for adjusting the feature dimensionality reduction model to generate the verification low-dimensional feature vector strategy. The specific working principle is as follows:
[0098] The fusion vector construction unit 310 uses an element-level fusion method to construct the initial fusion vector corresponding to the user to be verified. The specific working principle is as follows:
[0099] Receive multiple biometric vectors corresponding to biometric feature A, compare the feature vectors with the same feature dimension among the multiple biometric vectors, select the smallest feature vector for each feature dimension among the multiple biometric vectors to form a unique biometric vector for biometric feature A, and construct the corresponding initial fusion vector through the biometric vector, specifically:
[0100] Receive the mean values of features corresponding to different feature dimensions in the feature dimensionality reduction model; sequentially calculate the differences between multiple mean vectors of all feature dimensions and their corresponding mean values in each biological feature vector;
[0101] By comparing the differences between multiple mean vectors of the same feature dimension, the feature vector corresponding to the largest mean vector difference is selected as the feature vector of the corresponding feature dimension in the biometric feature vector. Then, the feature vectors of all feature dimensions are integrated to form the corresponding biometric feature vector. Then, the multimodal feature fusion module 100 is used to construct the initial fusion vector corresponding to the user to be verified, thereby maximizing the retention of feature vectors that deviate from the global mean in each feature dimension, forming a biometric feature vector that is both complete and unique. This avoids the situation where multiple identical biometric feature vectors are disordered and cause the initial fusion vector to be messy, thereby improving the effectiveness of subsequent feature dimensionality reduction and the accuracy of identity verification.
[0102] Dimensionality reduction weight adjustment unit 320 calculates the difference in dimensionality vectors between features of the same dimension (i.e., the difference between the largest and smallest eigenvectors under the same feature dimension), specifically the feature dimension. Dimensional vector difference for:
[0103] ;
[0104] Where t is the feature dimension. Number of times the data was collected;
[0105] Define the same feature dimension as the core dimension, increase the information weight of the corresponding direction vector in the projection matrix based on the difference of dimension vectors, and then the feature dimensionality reduction model uses the increased information weight and the biometric vector in the fusion vector construction unit 310 to generate the verification low-dimensional feature vector corresponding to the user to be verified. The specific working principle is: retrieve the feature dimension corresponding to each direction vector in the projection matrix and match the direction vector including the core dimension.
[0106] The difference between the core dimensions of the added direction vectors is used to obtain the corresponding direction difference. Specifically, if the feature dimension... and feature dimensions In the direction vector When inside, the direction vector The directional difference is: + ;
[0107] Calculate the proportion corresponding to the difference in each direction, and define the proportion of the direction difference as the amplification factor of the corresponding direction vector. The percentage is: ,in, This includes the number of core dimension direction vectors;
[0108] Using an increase factor Increase direction vector Information weight To obtain information weights Then, the feature dimensionality reduction model uses information weights. The biometric feature vector in the fusion vector construction unit 310 generates the verification low-dimensional feature vector corresponding to the user to be verified; and information weights are used. After successful identity verification, the user's baseline feature library is updated based on the biometric feature vector corresponding to the multimodal feature fusion module 100.
[0109] Increase information weight Information weight This is used to strengthen the core identity features in the feature dimensionality reduction model. The fusion vector construction unit 310 has fused the feature vectors collected multiple times corresponding to the same type of biometric features (such as fingerprint minutia vectors collected multiple times and face depth feature vectors from multiple frames) to obtain a unique initial fusion vector. During the fusion process, the feature vectors of the same feature dimension in multiple biometric vectors are fused. Specifically, the fusion is carried out by the minimum feature vector of each feature dimension in multiple biometric vectors. The minimum feature vector of each same feature dimension is the difference between the maximum mean vector of the feature dimension and the mean of the corresponding feature. The maximum mean vector difference can fully filter out the core identity dimension in the multiple collected feature vectors and eliminate redundant and easily fluctuating non-critical dimensions, laying an accurate foundation for subsequent feature processing.
[0110] Meanwhile, during the generation and verification of low-dimensional feature vectors, the linear correlation between feature dimensions is quantified based on the covariance matrix. Core dimensions with high information weights are retained first, while overlapping and redundant information is removed. Furthermore, the inherent correlations of multimodal features (including linear and nonlinear correlations) are integrated to ensure that the low-dimensional vectors are both concise and fully convey identity differentiation information. The maximum mean vector difference is constrained by the physical characteristics of biometric features (such as the lifelong stability of fingerprint minutiae and iris patterns) and the outlier removal mechanism after repeated data collection, preventing the occurrence of boundless extreme values and increasing the coefficient. It only specifically strengthens the weights of the core identity dimension, preventing the feature dimensionality reduction model from becoming overly sensitive to irrelevant fluctuations; therefore, it is aimed at increasing the coefficients. There is no need to set corresponding threshold limits, so that it can be clearly determined whether the user to be verified is a legitimate registered user during the subsequent verification process, depending on the information weight of the core identity dimension. As the coefficient increases After enhancement, the similarity between the low-dimensional feature vector of a legitimate user and its own feature vector in the benchmark feature library will remain in the high range due to the high matching of core features. Even if an illegitimate user forges a certain modality feature, it cannot meet the inherent correlation verification requirements between the core dimensions of multiple modalities, and the similarity will be significantly lower than the legitimate threshold. Accurate differentiation can be achieved without limiting the threshold.
[0111] In summary:
[0112] In this example, the multimodal feature fusion module 100 first constructs initial fusion vectors corresponding to multiple biometric vectors, fully integrating multiple biometric vectors into a unique initial fusion vector. This provides a rich and comprehensive data source for the subsequent dimensionality reduction modeling and feature library module 200 to build a feature dimensionality reduction model. Subsequently, the dimensionality reduction modeling and feature library module 200 uses the initial fusion vectors of multiple verified users to build a feature dimensionality reduction model. This model is used to learn the global distribution pattern and core feature mapping relationship of the initial fusion vectors, ensuring that the feature dimensionality reduction model has basic high-dimensional data dimensionality reduction capabilities and feature extraction accuracy. In scenarios where the same type of biometric feature is repeatedly collected, the fusion vector construction unit 310 calculates the mean deviation between each feature dimension in each collected vector and the corresponding mean vector, based on the mean vector calculated by the dimensionality reduction modeling and feature library module 200 when building the feature dimensionality reduction model. The feature value corresponding to the largest mean deviation is selected and integrated to form a unique initial fusion vector corresponding to the repeatedly collected same type of biometric feature. This avoids the influence of factors such as the collection environment (e.g., light, humidity), equipment accuracy fluctuations, and subtle changes in the user's physiological state during repeated collection, which can lead to random errors and redundancy in different biometric vectors during repeated collection. Information is used to filter and retain the most distinctive key information in the same feature dimension, avoiding the vector clutter problem caused by repeated biometric vectors. At the same time, the multimodal feature fusion module 100 integrates them again to form a unique initial fusion vector, simplifying the subsequent data processing flow and reducing redundant calculations. The dimensionality reduction weight adjustment unit 320 retrieves the dimensionality vector difference of the same feature dimension in the same feature dimension of repeated collection of the same type of biometric features (i.e., the difference between the maximum and minimum feature values under the same feature dimension, quantifying the feature fluctuation amplitude of the dimension). Based on the dimensionality vector difference, the weight amplification coefficient of the corresponding direction vector is calculated, and the information weight in the feature dimensionality reduction model is adjusted again. This avoids the problem that the feature dimensionality reduction model in the above-mentioned dimensionality reduction model and feature library module 200 follows the traditional construction logic of prioritizing the retention of the global maximum variance of the data, which over-focuses on the common feature direction shared by all users and ignores the local low variance features unique to each user. At the same time, it is affected by the intra-class changes during repeated collection of biometric features (such as the pressure and direction difference of fingerprint collection, and the light and expression fluctuations of face collection), which leads to the unreasonable suppression of the effective bias that is crucial to distinguishing user identity, ultimately resulting in insufficient distinguishability of the verification low-dimensional feature vector of the user.
[0113] Specifically, by specifically adjusting the internal weight allocation of the feature dimensionality reduction model, the influence of the directional vector carrying the core differences in user identity is dynamically amplified, while the weights of irrelevant feature dimensions that do not have identity differentiation value are actively weakened. This ultimately achieves precise optimization of difference enhancement and noise suppression, ensuring that the unique feature differences of user identity are accurately preserved during dimensionality reduction, clearly defining normal collection fluctuations and abnormal feature deviations. It also avoids the inherent limitations of traditional dimensionality reduction models that prioritize retaining globally high-variance features and suppressing local effective biases. Through deep linkage between dimensionality difference and weight adjustment, it solves the problem of feature vector clutter caused by repeated collection and effectively avoids the weakening of identity differentiation due to intra-class changes. Furthermore, after successful verification, the baseline feature library is updated based on the optimized biometric vectors, further improving the model's adaptability to natural changes in user identity features. Ultimately, this significantly improves the accuracy, robustness, and scenario adaptability of identity verification, perfectly meeting the needs of scenarios with extremely high security requirements for identity authentication, such as financial payments and government services.
[0114] The precise optimization of difference enhancement and noise suppression enables the unique characteristics of user identity to be accurately preserved during the dimensionality reduction process, and clearly defines the normal collection fluctuations and abnormal feature deviations.
[0115] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A risk monitoring and management system based on a large model, characterized in that, include: A multimodal feature fusion module (100) is used to convert multiple biometric vectors of a legitimate user into a unique initial fusion vector; The dimensionality reduction modeling and feature library module (200) is used to learn the global distribution pattern and feature correlation of the initial fusion vectors of all legitimate users, build a feature dimensionality reduction model, use the feature dimensionality reduction model to transform the initial fusion vectors of legitimate users into benchmark low-dimensional feature vectors, establish a benchmark feature library for legitimate users, and generate the corresponding verification low-dimensional feature vectors for the users to be verified. The repeated acquisition feature optimization module (300) constructs the initial fusion vector corresponding to the user to be verified by using the element-level fusion method when repeatedly acquiring the same type of biometric features; and calculates the difference in dimension vectors between the same feature dimensions, and adjusts the feature dimensionality reduction model through the difference in dimension vectors to generate the result of verifying the low-dimensional feature vector.
2. The risk monitoring and management system based on a large model according to claim 1, characterized in that: The multimodal feature fusion module (100) receives multiple biometric vectors from all legitimate users and concatenates all biometric vectors of each legitimate user in a fixed order to obtain the corresponding initial fusion vector.
3. The risk monitoring and management system based on a large model according to claim 1, characterized in that: The dimensionality reduction modeling and feature library module (200) uses the initial fusion vector in the multimodal feature fusion module (100) to build a feature dimensionality reduction model. Specifically, for the same feature dimension in all initial fusion vectors, the feature mean of the same feature dimension is calculated; the feature value of each feature dimension in each initial fusion vector is retrieved, the feature mean of the corresponding feature dimension is matched, the difference between the feature value and the feature mean is calculated, the corresponding central feature value is obtained, and the initial fusion vector is converted into the corresponding centralized fusion vector. The centralized fusion vectors of all legitimate users together constitute a multi-user fusion vector set; Calculate the covariance matrix of the multi-user fusion vector set; Decompose the direction vector and information weights for each pair of related dimensions in the covariance matrix; Set the target dimension reduction dimension, select the direction vectors corresponding to the top multiple information weights, and concatenate the selected direction vectors to form the projection matrix. The central eigenvalue of each feature dimension of the single-user fusion vector is multiplied by the transpose of the projection matrix to obtain the dimensionality-reduced baseline low-dimensional feature vector.
4. The risk monitoring and management system based on a large model according to claim 3, characterized in that: The dimensionality reduction modeling and feature library module (200) generates a verification low-dimensional feature vector corresponding to the user to be verified: it retrieves multiple biometric feature vectors corresponding to the user to be verified, and the multimodal feature fusion module (100) constructs multiple biometric feature vectors as corresponding initial fusion vectors; it outputs the initial fusion vector to the feature dimensionality reduction model after it has been built, and the feature dimensionality reduction model uses the feature mean to transform the initial fusion vector into a centered fusion vector, and finally uses the projection matrix transpose to obtain the verification low-dimensional feature vector corresponding to the user to be verified.
5. The risk monitoring and management system based on a large model according to claim 4, characterized in that: The covariance matrix in the dimensionality reduction modeling and feature library module (200) is specifically as follows: Retrieve any two feature dimensions with different physical meanings and data distribution characteristics from the multi-user fusion vector set to form a pair of related dimensions; calculate the covariance between each pair of related dimensions using the central eigenvalues corresponding to the feature dimensions, which is used to quantify the degree of linear correlation between the two feature dimensions. If the absolute value of the covariance is larger, it indicates that the overlap between the two feature dimensions is higher and the redundancy is stronger. If the covariance is close to 0, it indicates that the two feature dimensions are relatively independent.
6. The risk monitoring and management system based on a large model according to claim 5, characterized in that: The target dimensionality reduction dimension is set in the dimensionality reduction modeling and feature library module (200): An initial dimension reduction dimension equal to the number of direction vectors is preset, and a dimension feature vector is constructed. The initial dimension reduction dimension is gradually reduced to obtain multiple temporary dimension reduction dimensions until the temporary dimension reduction dimension equals 1. Each time the dimension reduction is reduced, the direction vector with the smallest information weight is removed, and the dimension feature vector is reconstructed based on the remaining direction vectors. After the reduction, the information weights corresponding to each directional vector in the temporary dimensionality reduction dimension are calculated again. Specifically, the scalars corresponding to each directional vector are added together and used as the denominator. The scalars of each directional vector are then used as the numerators, and the numerators are divided by the denominators to obtain the information weights of the corresponding directional vectors. Cross-calculate the vector similarity between the two feature vectors, and select the temporary dimension reduction dimension that minimizes the vector similarity and makes them all different as the target dimension reduction dimension.
7. A risk monitoring and management system based on a large model according to claim 6, characterized in that: The repeated acquisition feature optimization module (300) includes a fusion vector construction unit (310) and a dimensionality reduction weight adjustment unit (320). The fusion vector construction unit (310) uses the element-level fusion method to construct the initial fusion vector corresponding to the user to be verified. Specifically, it receives multiple biometric vectors corresponding to biometric A, compares the feature vectors with the same feature dimension among the multiple biometric vectors, selects the smallest feature vector of each feature dimension among the multiple biometric vectors, forms a unique biometric vector of biometric A, and constructs the corresponding initial fusion vector through the biometric vector. The dimension reduction weight adjustment unit (320) calculates the difference in dimension vectors between the same feature dimensions; defines the same feature dimensions as the core dimensions, increases the information weight of the corresponding direction vector in the projection matrix based on the difference in dimension vectors, and then the feature dimension reduction model uses the increased information weights and the biometric vector in the fusion vector construction unit (310) to generate the verification low-dimensional feature vector corresponding to the user to be verified.
8. A risk monitoring and management system based on a large model according to claim 7, characterized in that: The element-level fusion method in the fusion vector construction unit (310) is specifically as follows: Receive the mean values of features corresponding to different feature dimensions in the feature dimensionality reduction model; sequentially calculate the differences between multiple mean vectors of all feature dimensions and their corresponding mean values in each biological feature vector; By comparing the differences between multiple mean vectors of the same feature dimension, the feature vector corresponding to the largest mean vector difference is selected as the feature vector of the corresponding feature dimension in the biological feature vector. Then, the feature vectors of all feature dimensions are integrated to form the corresponding biological feature vector.
9. A risk monitoring and management system based on a large model according to claim 8, characterized in that: In the dimension reduction weight adjustment unit (320), the dimension vector difference is the difference between the largest and smallest eigenvectors under the same feature dimension; specifically, if the feature dimension and feature dimensions In the direction vector When inside, the direction vector The directional difference is: + ; Calculate the proportion corresponding to the difference in each direction, and define the proportion of the direction difference as the amplification factor of the corresponding direction vector. The percentage is: ,in, This includes the number of core dimension direction vectors; Using an increase factor Increase direction vector Information weight To obtain information weights The feature dimensionality reduction model then employs information weights. , and the biometric vector in the fusion vector construction unit (310) generates the verification low-dimensional feature vector corresponding to the user to be verified; Using information weights After successful identity verification, the user baseline feature library is updated based on the biometric feature vector corresponding to the multimodal feature fusion module (100).
10. A risk monitoring and management method based on a large model, applied to a risk monitoring and management system based on a large model according to any one of claims 1-9, characterized in that, Includes the following steps: S1. Convert multiple biometric vectors of legitimate users into a unique initial fusion vector; S2. Build a feature dimensionality reduction model, use the feature dimensionality reduction model to transform the initial fusion vector of the legitimate user into the benchmark low-dimensional feature vector, establish a benchmark feature library of legitimate users, and generate the corresponding verification low-dimensional feature vector of the user to be verified. S3. When repeatedly collecting the same type of biometric features, use the element-level fusion method to construct the initial fusion vector corresponding to the user to be verified; and calculate the difference in dimension vectors between the same feature dimensions, and use the difference in dimension vectors to adjust the feature dimensionality reduction model to generate the result of verifying the low-dimensional feature vector.