Financial data processing method and system based on machine learning and storage medium

Abstract

The invention discloses a financial data processing method and system based on machine learning, and a storage medium. The method comprises the following steps: encrypting customer credit data of a financial institution through Paillier homomorphic encryption and extracting a risk feature vector; performing secret sharing segmentation on the local gradient parameters and then safely aggregating the local gradient parameters into global gradient parameters through a BGW protocol; on the basis of the privacy level, global gradient parameter differential privacy noise is injected and subjected to federal training to obtain a cross-institution credit evaluation result; calculating risk characteristics and evaluation results through homomorphic encryption to obtain an encrypted credit score, and performing secure multi-party calculation and decryption to generate a customer credit report; and block chain supervision compliance verification is carried out to generate an audit record, and a privacy budget optimization scheme is dynamically adjusted according to the business emergency degree. The technical problems that in an existing financial data processing method, information is incomplete due to data islands, data leakage risks are caused by insufficient privacy protection mechanisms, and an effective supervision compliance verification mechanism is lacked are solved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a financial data processing method, system, and storage medium based on machine learning. Background Art

[0002] Existing financial data processing methods primarily rely on traditional centralized data analysis and statistical learning techniques. Each financial institution typically conducts independent credit risk assessments and customer profiling based on its own data. These methods employ conventional machine learning algorithms such as support vector machines, decision trees, and neural networks to model and analyze customer transaction records, credit histories, and financial status. These methods use historical data to train predictive models to assess a customer's credit risk level. At the same time, some financial data collaboration solutions based on federated learning have emerged, attempting to achieve cross-institutional data value mining while protecting data privacy.

[0003] However, existing technologies have significant limitations and shortcomings. First, traditional centralized approaches face serious data silos. Financial institutions are unable to share customer data due to competitive relationships and regulatory constraints, resulting in a limited data perspective for a single institution and difficulty in comprehensively and accurately assessing a customer's true credit status. Second, while existing federated learning solutions claim to protect privacy, actual implementation still presents security risks such as gradient leakage and model inversion attacks, failing to meet the strict data protection requirements of the financial industry. Furthermore, existing methods lack effective regulatory compliance verification mechanisms, making it difficult to ensure operational compliance and auditability during distributed collaboration.

[0004] Based on the above analysis, deeper technical challenges are gradually emerging in existing technologies: how to achieve true cross-institutional data collaborative processing while strictly protecting each institution's sensitive data; how to establish a secure computing framework that can resist various privacy attacks while maintaining model performance; how to achieve real-time regulatory compliance verification and dynamic privacy budget management in the distributed financial data processing process; and how to build a comprehensive financial data processing solution that integrates data encryption, secure computing, privacy protection, and compliance auditing. These progressive technical challenges require the development of a new approach to financial data collaborative processing based on the integration of multiple cryptographic technologies. Summary of the Invention

[0005] This application provides a financial data processing method, system and storage medium based on machine learning, which are used to solve the technical problems in existing financial data processing methods, such as incomplete information caused by data silos, data leakage risks caused by insufficient privacy protection mechanisms, and the lack of effective regulatory compliance verification mechanisms.

[0006] In a first aspect, the present application provides a financial data processing method based on machine learning, which includes: encrypting the customer credit data of each financial institution through the Paillier homomorphic encryption algorithm to obtain an encrypted credit data set, and extracting features from the encrypted credit data set according to the type of financial business to obtain an encrypted risk feature vector; splitting the local gradient parameters corresponding to the encrypted risk feature vector through a secret sharing method to obtain gradient shares, and securely aggregating the gradient shares according to the BGW protocol to obtain global gradient parameters; injecting noise into the global gradient parameters through a differential privacy mechanism according to the privacy level of the encrypted credit data set to obtain a privacy-preserving gradient, and performing federal training on the privacy-preserving gradient to obtain a cross-institutional credit assessment result; performing credit risk prediction on the encrypted risk feature vector and the cross-institutional credit assessment result through homomorphic encryption calculation to obtain an encrypted credit score, and decrypting the encrypted credit score according to secure multi-party computing to obtain a customer credit report; verifying the regulatory compliance of the cross-institutional credit assessment result and the customer credit report through a blockchain system to obtain an audit record, and dynamically adjusting the budget of the privacy-preserving gradient according to the urgency of the business to obtain a budget optimization plan.

[0007] In a second aspect, the present application provides a financial data processing system based on machine learning, wherein the financial data processing system based on machine learning comprises: The encryption module is used to encrypt the customer credit data of each financial institution using the Paillier homomorphic encryption algorithm to obtain an encrypted credit data set, and then extract features from the encrypted credit data set based on the financial business type to obtain an encrypted risk feature vector; The segmentation module is used to segment the local gradient parameters corresponding to the encrypted risk feature vector through a secret sharing method to obtain gradient shares, and then perform secure aggregation calculation on the gradient shares according to the BGW protocol to obtain the global gradient parameters; The injection module is used to inject noise into the global gradient parameters through the differential privacy mechanism according to the privacy level of the encrypted credit dataset to obtain the privacy-preserving gradient, and then perform federated training on the privacy-preserving gradient to obtain the cross-institutional credit assessment results; The prediction module is used to perform credit risk prediction using homomorphic encryption to calculate the encrypted risk feature vector and the cross-institutional credit assessment results, thereby obtaining an encrypted credit score. The encrypted credit score is then decrypted using secure multi-party computation to obtain a customer credit report. The verification module is used to verify the regulatory compliance of cross-institutional credit assessment results and customer credit reports through the blockchain system, obtain audit records, and dynamically adjust the budget of the privacy protection gradient according to the urgency of the business to obtain a budget optimization plan.

[0008] In a third aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned financial data processing method based on machine learning.

[0009] The technical solution provided in this application uses the Paillier homomorphic encryption algorithm to encrypt customer credit data from various financial institutions and extract encrypted risk feature vectors. This fundamentally solves the data silo problem in traditional financial data processing, enabling institutions to participate in collaborative computing without exposing their original data, effectively overcoming the data sharing difficulties caused by privacy concerns in existing technologies. The local gradient parameters corresponding to the encrypted risk feature vectors are segmented using a secret sharing method and securely aggregated using the BGW protocol, ensuring the secure transmission and aggregation of gradient information during distributed machine learning training and avoiding the risks of gradient leakage and model inversion attacks that may occur in traditional federated learning. A privacy-level differential privacy mechanism is used to inject noise into the global gradient parameters and perform federated training, achieving an optimal balance between privacy protection and model performance, ensuring that cross-institutional credit assessment results are both highly accurate and meet strict privacy protection requirements. The encrypted risk feature vectors and cross-institutional credit assessment results are used to predict credit risk through homomorphic encryption and then decrypted through secure multi-party computation to generate customer credit reports. This ensures end-to-end privacy protection for the entire credit assessment process while providing a more comprehensive and accurate risk assessment than single-institution assessments. Ultimately, regulatory compliance verification is performed through the blockchain system, and the privacy budget is dynamically adjusted according to business urgency, which not only ensures the compliance and auditability of the financial data processing process, but also realizes the intelligent management of privacy protection strategies.

[0010] Positive and beneficial effects: The algorithmic features adopted by this invention make a decisive contribution to the effectiveness of the solution. The additive homomorphic property of the Paillier homomorphic encryption algorithm enables financial institutions to directly perform aggregate calculations of risk indicators in a ciphertext state, avoiding the security risks of traditional methods that require decrypting data for calculation. This algorithmic feature directly supports the secure collaborative processing of cross-institutional financial data. The secure multi-party computation feature of the BGW protocol ensures Byzantine fault tolerance during the gradient aggregation process, ensuring the correctness of the calculation results even in the presence of malicious nodes. This feature is of great significance to the high security requirements of the financial sector. The mathematical rigor of the differential privacy mechanism provides quantifiable privacy protection guarantees, and the adjustability of its epsilon parameter allows financial institutions to flexibly adjust the protection strength according to the privacy requirements of different business scenarios, meeting the financial regulatory requirements for precise control of the degree of privacy protection. The immutability and distributed consensus mechanism of blockchain technology provide a complete audit trail for the financial data processing process, ensuring that regulators can effectively supervise the cross-institutional collaboration process. The organic combination of these algorithmic features creates a secure and efficient financial data collaborative processing ecosystem. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a schematic diagram of an embodiment of a financial data processing method based on machine learning in an embodiment of the present application; Figure 2 This is a schematic diagram of an embodiment of a financial data processing system based on machine learning in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The embodiments of the present application provide a method, system, and storage medium for processing financial data based on machine learning. The terms "first," "second," "third," "fourth," and so on (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the data used in this way can be interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products, or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of the financial data processing method based on machine learning includes: Step S101: Encrypt the customer credit data of each financial institution using the Paillier homomorphic encryption algorithm to obtain an encrypted credit dataset, and extract features from the encrypted credit dataset based on the financial business type to obtain an encrypted risk feature vector; Step S102: The local gradient parameters corresponding to the encrypted risk feature vector are divided using a secret sharing method to obtain gradient shares, and the gradient shares are securely aggregated and calculated according to the BGW protocol to obtain the global gradient parameters; Step S103: inject noise into the global gradient parameter using a differential privacy mechanism based on the privacy level of the encrypted credit dataset to obtain a privacy-preserving gradient, and perform federated training on the privacy-preserving gradient to obtain a cross-institutional credit assessment result; Step S104: Using homomorphic encryption to calculate the encrypted risk feature vector and the cross-institutional credit assessment results, a credit risk prediction is performed to obtain an encrypted credit score, and the encrypted credit score is decrypted using secure multi-party computation to obtain a customer credit report; Step S105: The cross-institutional credit assessment results and customer credit reports are verified for regulatory compliance through the blockchain system to obtain audit records, and the budget of the privacy protection gradient is dynamically adjusted according to the business urgency to obtain a budget optimization plan.

[0014] It is understandable that the execution subject of this application can be a financial data processing system based on machine learning, or a terminal or server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0015] Specifically, the Paillier homomorphic encryption algorithm is a homomorphic encryption method that supports addition operations. Financial institutions first identify the data type of customer credit data and classify the transaction amount, account balance, credit score, and risk level. A public-private key pair is then generated by selecting a large prime number. The public key is used to encrypt customer data, while the private key is retained by the local institution. During the encryption process, the original data is converted into ciphertext, and the ciphertext credit data is grouped according to personal credit business and corporate credit business. Homomorphic operation rules allow addition and multiplication calculations to be performed in the ciphertext state, generating composite ciphertext features that contain risk indicators. The Z-score transformation algorithm standardizes the composite ciphertext features by calculating the mean and variance, ultimately obtaining an encrypted risk feature vector that maintains the statistical characteristics of the original data but is fully encrypted.

[0016] Local gradient calculation generates weight gradients and bias gradients based on an encrypted risk feature vector. The secret sharing method divides the gradient parameters into multiple shares using a Lagrange interpolation polynomial, with each share distributed to different financial institution nodes. The BGW protocol is a secure multi-party computation protocol that performs share verification and consistency checks on weight gradient shares and bias gradient shares from multiple nodes to ensure data integrity. The threshold secret sharing mechanism requires a preset number of nodes to reconstruct the original gradient. The polynomial coefficients are calculated using the Lagrange interpolation formula to reconstruct the weight gradient and bias gradient values. The weighted average algorithm assigns weights based on the data contribution of each financial institution, ultimately aggregating and generating global gradient parameters that include global weight gradient parameters and global bias gradient parameters.

[0017] Privacy classification categorizes credit data into high-sensitivity and low-sensitivity levels based on their sensitivity. Differential privacy is a mathematical framework that protects individual privacy by adding calibrated noise to the data. A privacy budget is allocated to weight and bias gradients, with each budget set separately. The Gaussian noise generation algorithm generates noise of a corresponding intensity based on the budget. The noise injection process adds noise to the global gradient parameters, generating a privacy-preserving gradient consisting of a noisy weight gradient and a noisy bias gradient. The federated learning training algorithm iterates the privacy-preserving gradient through multiple rounds, updating the parameters in each round until convergence. Ultimately, a cross-institutional credit assessment result is generated, including the credit rating and default probability.

[0018] The feature fusion algorithm combines the encrypted risk feature vector with the credit rating and default probability to form a fused credit feature vector. A homomorphic encryption neural network performs forward propagation calculations in an encrypted state, generating intermediate calculation results through hidden layer processing and network activation. Homomorphic multiplication performs weight matrix multiplication and bias vector addition on the intermediate results, all performed entirely within the encrypted domain. A homomorphic nonlinear transformation converts the calculated value into an encrypted credit score that includes the customer's credit rating. A secure multi-party computation protocol allows multiple financial institution nodes to collaborate on decryption. Each node provides a decryption share but cannot independently obtain the complete result. The final output is a customer credit report containing the plaintext credit rating, risk rating, and credit limit.

[0019] A hash digest algorithm generates a digital fingerprint for cross-institutional credit assessment results and customer credit reports, creating blockchain-based evidence consisting of the result hash value and the report hash value. Smart contract rules are pre-set, programmatic rules that automatically perform financial regulatory compliance checks and generate verification results that include a compliance status indicator and a violation risk marker. The blockchain consensus mechanism distributes confirmation across multiple financial institution nodes, creating an immutable audit trail through timestamping. Business urgency grading prioritizes scenarios such as risk warnings, routine approvals, and periodic reporting, allocating more privacy budget to high-priority businesses and less to low-priority businesses. A dynamic reallocation process adjusts the budget amounts for the noise weight gradient and noise bias gradient based on actual business needs, generating budget optimization solutions tailored to different business scenarios.

[0020] In a specific embodiment, the process of executing step S101 may specifically include the following steps: Identify and process the data types of customer credit information from various financial institutions to obtain classified credit information data including transaction amounts, account balances, credit scores, and risk levels; Inputting the classified credit information data into the Paillier homomorphic encryption algorithm to perform large prime number selection and public key and private key pair generation processing to obtain encrypted credit information data; performing grouping processing on the encrypted credit investigation data according to the financial business type, namely, personal credit business and corporate credit business, to obtain grouped encrypted data; Performing addition and multiplication calculations on the grouped ciphertext data based on homomorphic operation rules to obtain a composite ciphertext feature containing the risk indicator of the classified credit data; The composite ciphertext feature is normalized after mean and variance calculation using a Z-score transformation algorithm to obtain an encryption risk feature vector based on the grouped ciphertext data.

[0021] Specifically, data type identification processing parses the raw customer credit data input by various financial institutions to identify the specific type and format of the data fields. Transaction amount data includes numerical information such as the customer's transfer records, consumer expenditures, and income flow. Account balance data reflects the customer's financial status at different points in time. Credit score data is a quantitative assessment value calculated based on historical repayment records and credit behavior. Risk level data is a level identifier based on the customer's probability of default and repayment ability. The data type identification process first reads the raw data file and determines the credit data type of each field through field name matching and data format verification. Then, data of the same type is classified and organized to form a classified credit data structure containing four main data types.

[0022] The Paillier homomorphic encryption algorithm's large prime number selection process requires generating two large prime numbers, p and q, where p represents the first large prime number and q represents the second large prime number. The bit lengths of p and q are typically 1024 or 2048 bits. The public-private key pair generation process first calculates n, which equals p multiplied by q. n represents the RSA modulus, the primary component of the public key. Then, λ is calculated, which equals the least common multiple of p minus one multiplied by q minus one. λ represents the Carmichael function value, used for private key calculation. A random integer, g, is selected as the encryption base. g represents a generator, a random integer that satisfies specific mathematical conditions. The public key consists of two parameters, n and g, while the private key consists of two parameters, λ and μ. μ represents the modular inverse element, calculated using λ. The encryption process for classified credit data converts each data element into ciphertext. Transaction amount data is ciphertexted by performing a modulo-n-squared operation on the public keys n and g. Modulo-n-squared represents the square of n. Account balance, credit score, and risk level data undergo the same encryption process. During the encryption process, each original value is converted into a large integer ciphertext. The ciphertext credit data maintains the mathematical relationship of the original data, but the original values ​​cannot be directly read. Financial business type grouping processing divides the ciphertext credit data into different categories based on customer attributes and business characteristics. Personal credit business includes business data for individual customers such as personal consumption loans, mortgage loans, and credit card loans. Corporate credit business covers business data for corporate customers such as corporate working capital loans, equipment purchase loans, and trade financing. The grouping process reads the customer ID and business type fields, classifying the ciphertext credit data belonging to individual customers into the personal credit group and the ciphertext credit data belonging to corporate customers into the corporate credit group, forming a grouped ciphertext data structure differentiated by business type.

[0023] Homomorphic operation rules allow mathematical calculations to be performed directly on ciphertext without decrypting the data. Addition multiplies two ciphertexts to produce a ciphertext result equal to the sum of the corresponding plaintexts. Multiplication multiplies plaintexts by applying exponential operations to ciphertexts. The homomorphic calculation process for grouped ciphertext data first extracts the transaction amount and account balance ciphertexts from each group, then performs homomorphic addition to obtain the customer's total asset ciphertext. The credit score ciphertext and risk level ciphertext are then homomorphically multiplied to obtain the composite risk ciphertext. Composite ciphertext features include various risk indicator ciphertexts generated through homomorphic operations. These indicators reflect the customer's balance sheet, repayment ability, and creditworthiness.

[0024] The Z-score transformation algorithm is a data normalization method that converts data into a standard normal distribution by calculating its mean and standard deviation. The mean calculation process sums all composite ciphertext feature values ​​and divides them by the total number of data points. Due to the characteristics of homomorphic encryption, this calculation is performed directly in the ciphertext state. The variance calculation process squares the difference between each data point and the mean, then averages the difference to obtain the variance. The standard deviation is equal to the square root of the variance. Normalization subtracts the mean from each composite ciphertext feature and divides it by the standard deviation, resulting in standardized data with a mean of zero and a standard deviation of one. The encryption risk feature vector is a collection of standardized ciphertext data after the Z-score transformation. Each element in the vector is a standardized risk indicator calculated based on the grouped ciphertext data.

[0025] In a specific embodiment, the process of executing step S102 may specifically include the following steps: Performing local gradient calculation processing on the encrypted risk feature vector to obtain local gradient parameters including weight gradient and bias gradient; Secretly splitting the weight gradient and bias gradient in the local gradient parameter using Lagrange interpolation polynomials to obtain weight gradient shares and bias gradient shares distributed across multiple financial institution nodes; Based on the BGW protocol, the weight gradient shares and bias gradient shares from multiple financial institution nodes are respectively subjected to share verification and consistency check processing to obtain verified valid weight gradient shares and valid bias gradient shares; According to the threshold secret sharing mechanism, the effective weight gradient share and the effective bias gradient share are respectively subjected to polynomial reconstruction and aggregation operation processing to obtain a weight aggregation gradient result and a bias aggregation gradient result; The weighted aggregation gradient results and the bias aggregation gradient results are weighted according to the data contribution of each financial institution through a weighted average algorithm to obtain a global gradient parameter including a global weight gradient parameter and a global bias gradient parameter.

[0026] Specifically, the local gradient calculation process performs parameter update calculations for the machine learning model based on the standardized risk indicators in the encrypted risk feature vector. The weight gradient represents the partial derivative of the weight parameters connecting different neurons in the model with respect to the loss function, reflecting the degree of influence of the weight parameters on the model prediction error. The bias gradient represents the partial derivative of the bias parameters of each neuron in the model with respect to the loss function, reflecting the adjustment effect of the bias parameters on the model output results. The local gradient calculation process first inputs the encrypted risk feature vector into the local machine learning model of each financial institution, obtains the prediction result through forward propagation calculation, and then calculates the loss value between the prediction result and the true label. Then, the partial derivative of the loss value with respect to each weight parameter and bias parameter is calculated through the backpropagation algorithm to form a local gradient parameter containing weight gradient and bias gradient.

[0027] Lagrange interpolation polynomials are a mathematical tool used to decompose a secret value into multiple shares distributed across different nodes. The secret splitting process first selects a threshold k and the total number of nodes m, where k represents the minimum number of shares required to reconstruct the secret and m represents the total number of nodes participating in the secret sharing, with k less than m. A polynomial of degree k minus one is then constructed, with the constant term being the secret value to be split, and the other coefficients being randomly generated values. The weight gradient secret splitting process uses each weight gradient value as the constant term of the polynomial. The corresponding ordinate value is calculated by selecting m different abscissa points on the polynomial. Each ordinate value is distributed as a weight gradient share to the corresponding financial institution node. The bias gradient secret splitting process uses the same method, constructing each bias gradient value into an independent polynomial and calculating the corresponding bias gradient share. The split weight gradient shares and bias gradient shares are distributed across multiple financial institution nodes, making it impossible for a single node to reconstruct the original gradient value using its own shares.

[0028] The BGW protocol is a secure multi-party computation protocol that uses mathematical verification to ensure the authenticity and validity of the gradient shares submitted by each node. During the share verification process, each financial institution node calculates its own weight gradient share and bias gradient share against a preset verification polynomial, generates verification information, and broadcasts it to other nodes. The consistency check process requires all nodes to cross-validate the received verification information and determine the consistency of the shares by comparing the verification results of different nodes. During the verification process, if the mathematical relationship between a node's share and the shares of other nodes does not conform to the properties of the Lagrange interpolation polynomial, the share is marked as invalid and eliminated. The verified valid weight gradient share and valid bias gradient share are mathematically consistent and meet the prerequisites for subsequent reconstruction calculations.

[0029] The threshold secret sharing mechanism requires at least t valid shares to reconstruct the original secret value, where t is a preset threshold parameter. The polynomial reconstruction process recombines the effective weight gradient shares into the original weight gradient value through the Lagrange interpolation formula. The specific calculation method is to select the coordinate points corresponding to t valid shares, calculate the weight coefficient of each point through the Lagrange basis function, and then multiply the share value by the weight coefficient and sum them to obtain the reconstruction result. The reconstruction process of the bias gradient adopts the same mathematical method, and the effective bias gradient shares are reconstructed by polynomials to obtain the original bias gradient value. The aggregation operation process mathematically merges the reconstructed gradient values ​​from different financial institutions. The weighted aggregate gradient result is obtained by averaging or weighted averaging the reconstructed weight gradient values ​​of all institutions. The biased aggregate gradient result is obtained by performing the same aggregation calculation on the reconstructed bias gradient values ​​of all institutions.

[0030] The weighted average algorithm assigns different weight coefficients based on the data contribution of each financial institution. Data contribution reflects the importance and quality level of the training data provided by each institution in the overall data set, and is usually calculated based on factors such as data volume, data quality score, and historical accuracy. The weight allocation process first calculates the proportion of each financial institution's data contribution to the total contribution, and then applies this proportion as a weight coefficient to the gradient aggregation results of the corresponding institution. The global weight gradient parameter is obtained by multiplying the weight aggregation gradient results of each institution by their corresponding weight coefficients and summing them up. The global bias gradient parameter is obtained by multiplying the bias aggregation gradient results of each institution by their corresponding weight coefficients and summing them up. The global gradient parameter integrates the gradient information of all participating institutions and assigns corresponding influence weights to each institution according to its contribution level.

[0031] In a specific embodiment, the process of performing polynomial reconstruction and aggregation processing on the effective weight gradient share and the effective bias gradient share according to the threshold secret sharing mechanism may specifically include the following steps: Based on the threshold judgment condition, the number of financial institution nodes participating in the aggregate calculation is subjected to threshold verification processing to obtain a valid node set that meets the threshold requirements; The effective weight gradient share in the effective node set is processed by polynomial coefficient calculation using the Lagrange interpolation formula to obtain weight gradient polynomial parameters; Performing a reconstruction operation on the original weight gradient value according to the weight gradient polynomial parameter to obtain a reconstructed weight gradient value; Performing polynomial coefficient calculation processing on the effective bias gradient share in the effective node set through the Lagrange interpolation formula to obtain bias gradient polynomial parameters, and performing reconstruction operation processing on the bias gradient original value according to the bias gradient polynomial parameters to obtain a reconstructed bias gradient value; The reconstructed weight gradient value and the reconstructed bias gradient value are respectively summed up by a vector aggregation algorithm to obtain a weight aggregation gradient result and a bias aggregation gradient result.

[0032] Specifically, the threshold judgment condition is based on the core requirements of the secret sharing mechanism, which stipulates that the number of financial institution nodes participating in the aggregate calculation must reach a preset minimum threshold before gradient reconstruction can be performed. The threshold verification process first counts the total number of financial institution nodes that are currently online and submit valid gradient shares, and then compares it with the preset threshold r. Only when the actual number of nodes is greater than or equal to r, a valid node set that meets the threshold requirements is formed. The valid node set contains all financial institution nodes that have passed the threshold test, and each node holds a valid weight gradient share and a valid bias gradient share verified by the BGW protocol. The mathematical principle of the threshold test is based on the Lagrange interpolation theorem, which requires at least r different coordinate points to uniquely determine an r minus first-degree polynomial. Therefore, the gradient reconstruction calculation cannot be completed when the number of nodes is insufficient.

[0033] The Lagrange interpolation formula is a mathematical method used to reconstruct the coefficients of the original polynomial through known coordinate points. The polynomial coefficient calculation process treats each weight gradient share in the valid node set as a coordinate point on the polynomial. The horizontal coordinate of the coordinate point is the node identifier, and the vertical coordinate is the weight gradient share value. The weight gradient polynomial parameters are calculated using the Lagrange basis function. Each basis function corresponds to the weight gradient share of a node, and the calculation of the basis function involves the product operation of the horizontal coordinates of all other nodes. In the specific calculation process, the Lagrange basis function of the i-th node is equal to the product of the horizontal coordinates of all other nodes minus the horizontal coordinate of the i-th node, divided by the horizontal coordinate of the i-th node minus the product of the horizontal coordinates of all other nodes. The weight gradient polynomial parameters contain all the coefficients of the polynomial, which are obtained by multiplying the weight gradient share of each node with the corresponding basis function and then summing them up.

[0034] The reconstruction operation calculates the original weight gradient value based on the weight gradient polynomial parameters. This calculation process essentially evaluates the polynomial function at the zero position of the horizontal axis. The original weight gradient value is equal to the constant term of the polynomial, and is obtained by multiplying the weight gradient shares of all nodes by the values ​​of their corresponding Lagrange basis functions at zero and then summing them. The reconstructed weight gradient value restores the original weight gradient data obtained by each financial institution in the local gradient calculation. This value reflects the gradient information of the weight parameters with respect to the loss function in the machine learning model. The mathematical correctness of the reconstruction process is based on the uniqueness theorem of Lagrange interpolation, which ensures that the original secret value can be accurately recovered using r or more valid shares.

[0035] The polynomial coefficient calculation process of the bias gradient adopts the same Lagrange interpolation method as the weight gradient, and each bias gradient share in the valid node set is treated as an independent coordinate point. The bias gradient polynomial parameters are obtained through the same basis function calculation method, and each parameter corresponds to a coefficient of the bias gradient polynomial. The reconstruction operation of the original value of the bias gradient is also based on the function value calculation of the polynomial at zero point, and is obtained by multiplying the bias gradient share of each node with the value of the corresponding basis function at zero point and then summing them. The reconstructed bias gradient value restores the gradient information of the bias parameter in the local model of each financial institution, which is used to adjust the output bias of the model. The mathematical process of bias gradient reconstruction is exactly the same as that of weight gradient reconstruction at the algorithm level, but the data objects processed and the final use are different.

[0036] The vector aggregation algorithm is a data merging method used to integrate gradient data from multiple sources into a unified aggregated result. The summation operation numerically adds the reconstructed weight gradient values ​​of all financial institution nodes to produce a weighted aggregate gradient result that includes global weight information. This weighted aggregate gradient result reflects the combined effect of the weight gradient information of all participating institutions and is used in federated learning to update the weight parameters of the global model. The bias gradient summation operation uses the same mathematical operation to numerically add the reconstructed bias gradient values ​​of all nodes to form a biased aggregate gradient result. This biased aggregate gradient result integrates the bias gradient information of all institutions and is used to adjust the bias parameters of the global model.

[0037] In a specific embodiment, the process of executing step S103 may specifically include the following steps: Performing privacy level classification based on the sensitivity classification of the transaction amount, account balance, credit score, and risk level in the encrypted credit investigation data set to obtain a privacy level classification including a high sensitivity level and a low sensitivity level; Based on the privacy level classification, privacy budget allocation processing is performed on the global weight gradient parameter and the global bias gradient parameter in the global gradient parameter to obtain a weight privacy budget and a bias privacy budget; Performing noise injection processing on the global weight gradient parameter and the global bias gradient parameter according to the weight privacy budget and the bias privacy budget using a Gaussian noise generation algorithm to obtain a privacy-preserving gradient including a noise weight gradient and a noise bias gradient; Inputting the privacy-preserving gradient into a federated learning training algorithm for multiple rounds of iterative training to obtain converged federated learning parameters; Based on the federated learning parameters, risk assessment calculations are performed on customer credit data to obtain a cross-institutional credit assessment result including credit rating and default probability.

[0038] Specifically, sensitivity classification is a processing method that categorizes credit data according to the degree of risk of data leakage. The privacy level classification process is based on the sensitivity characteristics of the four data types in the encrypted credit data set. Transaction amount data and account balance data contain direct financial information of customers, which has a serious impact on customer privacy if leaked, and are therefore classified as highly sensitive. Although credit score data is a calculated assessment value, it directly reflects the customer's credit status and is also highly sensitive. Risk level data is a level identifier derived from statistical analysis. Relatively speaking, the privacy risk is low and is classified as low sensitive. Privacy level classification assesses the leakage risk and regulatory requirements of each type of data and divides the four types of credit data into two categories: high sensitivity and low sensitivity. The high sensitivity level includes three types of data: transaction amount, account balance, and credit score, while the low sensitivity level includes one type of data: risk level.

[0039] The privacy budget allocation process is the core step in controlling noise intensity in the differential privacy mechanism. Budgets are allocated to global gradient parameters based on privacy level classification. The global weight gradient parameter corresponds to the weight connection parameter in the model. Because weight parameters directly influence the model's predictive ability and are closely related to highly sensitive credit data, they are allocated a larger privacy budget. The global bias gradient parameter corresponds to the bias adjustment parameter in the model, primarily used to fine-tune the output results. It is highly correlated with less sensitive data and is allocated a relatively smaller privacy budget. Larger weight privacy budgets correspond to lower noise intensity, while smaller bias privacy budgets correspond to higher noise intensity. The mathematical principle of privacy budget allocation is based on the combinatorial nature of differential privacy. The total budget is distributed between weights and biases to ensure that the overall privacy protection level meets the preset requirements.

[0040] The Gaussian noise generation algorithm is a specific implementation of random noise in the differential privacy mechanism. It perturbs the gradient parameters by generating random values ​​that conform to a Gaussian distribution. The noise injection process calculates the corresponding noise standard deviation based on the weight privacy budget. The noise standard deviation is inversely proportional to the privacy budget; a larger privacy budget results in a smaller noise standard deviation. The noise injection process for the global weight gradient parameter directly adds random noise generated from a Gaussian distribution to the weight gradient value, resulting in a noisy weight gradient. The noise injection process for the global bias gradient parameter uses the same mathematical operation, generating Gaussian noise of a corresponding intensity based on the bias privacy budget. This noise is then added to the bias gradient value to produce a noisy bias gradient. The privacy-preserving gradient contains the noised weight and bias gradient information. These gradients maintain the statistical properties of the original data while increasing privacy protection.

[0041] The federated learning training algorithm is a distributed machine learning method that allows multiple participants to collaboratively train models without sharing the original data. Multiple rounds of iterative training process input privacy-preserving gradients into the federated learning framework. Each training round includes steps such as gradient aggregation, parameter updates, and model synchronization. Convergence of iterative training is determined based on the magnitude of the change in the loss function. When the change in the loss function over multiple rounds of training is less than a preset threshold, the training process reaches convergence. Converged federated learning parameters include weights and bias parameters that have been optimized over multiple rounds. These parameters integrate data from all participating institutions while protecting their privacy. The mathematical properties of the federated learning parameters ensure the model's ability to generalize to new data, while limiting the influence of individual data samples on the model through differential privacy mechanisms.

[0042] The risk assessment calculation process uses converged federated learning parameters to perform predictive analysis on customer credit data. This calculation process inputs the customer's multi-dimensional credit characteristics into the trained model for forward propagation. The credit rating is determined by the probability distribution output by the model. The model outputs probability values ​​for multiple credit ratings, and the rating with the highest probability is selected as the customer's final credit rating. The probability of default is calculated using the binary classification output of the model. The model predicts the probability of default within a specific timeframe based on the customer's characteristics. Cross-institutional credit assessment results integrate data from multiple financial institutions, resulting in more accurate and reliable assessments than those from a single institution. Privacy protection mechanisms also ensure that sensitive information from each institution is not leaked.

[0043] In a specific embodiment, the process of executing step S104 may specifically include the following steps: Combining the encrypted risk feature vector with the credit rating and default probability in the cross-institutional credit assessment result through a feature fusion algorithm to obtain a fused credit feature vector; Performing forward propagation calculation processing on the fused credit feature vector based on a homomorphic encryption neural network to obtain an intermediate calculation result including a hidden layer output and a network activation result; Performing weight matrix product and bias vector addition processing on the intermediate calculation result through homomorphic multiplication operation to obtain a credit score calculation value in an encrypted state; Performing a homomorphic nonlinear transformation on the credit score calculation value to obtain an encrypted credit score including the customer credit rating score; The encrypted credit score is collaboratively decrypted among multiple financial institution nodes according to a secure multi-party computing protocol to obtain a customer credit report containing a plaintext credit rating, risk rating, and credit limit.

[0044] Specifically, the feature fusion algorithm is a data combination method that integrates feature data from different sources into a unified feature representation. The combination process performs a data splicing operation on the standardized current asset indicators and comprehensive credit indicators in the encrypted risk feature vector with the credit rating and default probability in the cross-institutional credit assessment results. The credit rating data is converted into numerical features through numerical coding, and the default probability data maintains its original probabilistic numerical form. The fused credit feature vector contains the ciphertext financial indicators from the encrypted risk feature vector and the plaintext assessment indicators from the credit assessment results. This vector integrates the customer's historical financial data and current credit status to form a multi-dimensional credit feature representation. The fusion process arranges the various features in order through vector splicing operations to ensure that the relative position relationship between the features remains stable. The dimension of the fused feature vector is equal to the sum of the dimension of each component feature.

[0045] A homomorphic encryption neural network is a neural network architecture that performs computations in ciphertext, supporting forward propagation of encrypted data without decryption. The forward propagation process feeds the fused credit feature vector into the network's input layer, where a linear transformation is performed through the weighted connections from the input layer to the hidden layer. The hidden layer output is obtained through matrix multiplication in ciphertext, leveraging the multiplicative homomorphic properties of homomorphic encryption to directly operate on the ciphertext. Network activations are generated through homomorphic activation operations that use polynomial approximation to simulate the nonlinear characteristics of traditional activation functions. The intermediate computational results contain the hidden layer data that has undergone linear transformation and activation processing. This data remains in ciphertext but exhibits the mathematical properties of the neural network's forward propagation. The forward propagation mathematical process follows the standard neural network computational process, but all operations are performed in the ciphertext domain, ensuring data privacy.

[0046] Homomorphic multiplication is the core operation in homomorphic encryption that supports ciphertext multiplication, allowing mathematical operations to be performed on encrypted data without exposing the original numerical value. Weight matrix multiplication involves ciphertext multiplication of intermediate calculation results with a pretrained weight matrix containing the connection parameters from the hidden layer to the output layer. Bias vector addition adds bias parameters to the matrix product result through homomorphic addition. The bias vector is used to adjust the numerical offset of the network output. The credit score calculation value in the encrypted state is obtained through a combination of weight product and bias addition. This value reflects the customer's overall credit level but remains in encrypted form. The mathematical correctness of homomorphic operations ensures that the ciphertext calculation results are exactly the same as the plaintext calculation results after decryption. Error control during the calculation process is managed through precision parameter settings and rounding rules.

[0047] Homomorphic nonlinear transformation processing is a computational method that simulates nonlinear activation functions in an encrypted state, achieving complex nonlinear mapping through polynomial approximation techniques. The calculated credit score value is converted into the final credit score output through a homomorphic nonlinear transformation. This transformation process uses preset polynomial coefficients to perform high-order exponentiation and linear combination on the ciphertext data. The encrypted credit score contains the final score result after the nonlinear transformation, which has the numerical and distribution characteristics of the credit score. The accuracy of the polynomial approximation of the nonlinear transformation is controlled through Taylor expansion or Chebyshev approximation methods to ensure the mathematical accuracy of the transformation result. The customer credit rating score is obtained by mapping the continuous credit score value to discrete rating intervals. The mapping relationship is based on the preset scoring threshold and rating criteria.

[0048] The secure multi-party computation protocol is a distributed computing method that allows multiple parties to collaborate on computing tasks without disclosing their private information. Collaborative decryption requires multiple financial institution nodes to participate simultaneously in the decryption process, with each node providing a partial decryption key or decryption share. The decryption process for encrypted credit scores is implemented through a threshold decryption mechanism, requiring a preset number of nodes to successfully decrypt. The plaintext credit rating is determined by the decrypted credit score value and is divided into different credit rating categories based on the score range. The risk rating is based on a comprehensive assessment of the credit rating and other risk factors, reflecting the customer's risk of default in a credit transaction. The credit limit is calculated based on the credit rating, risk rating, and the customer's asset status. The credit limit represents the maximum amount of credit a financial institution is willing to provide to a customer. The customer credit report integrates key information such as the credit rating, risk rating, and credit limit. The report content is based on a comprehensive analysis of data from multiple institutions while protecting the data privacy of each institution.

[0049] In a specific embodiment, the process of executing step S105 may specifically include the following steps: The credit rating and default probability in the cross-institutional credit assessment result and the plaintext credit rating, risk rating and credit limit in the customer credit report are digitally fingerprinted using a hash digest algorithm to obtain blockchain evidence data containing the result hash value and the report hash value; Based on the smart contract rules, the blockchain evidence data is matched with financial regulatory regulations and compliance verification is performed to obtain a compliance verification result including a compliance status identifier and a violation risk marker; The compliance verification results are distributed and confirmed and timestamped across multiple financial institution nodes through a blockchain consensus mechanism, resulting in an audit record that cannot be tampered with. Prioritize risk warning, daily approval, and regular reporting scenarios based on the urgency of financial services, resulting in a business urgency classification that includes high and low priorities. The privacy budgets of the noise weight gradient and the noise bias gradient in the privacy protection gradient are dynamically reallocated based on the business urgency classification to obtain budget optimization solutions for different business scenarios.

[0050] Specifically, a hash digest algorithm is a cryptographic method that converts data of arbitrary length into a fixed-length digital fingerprint, ensuring data integrity verification through irreversible mathematical transformations. The digital fingerprint generation process first serializes the credit rating and default probability data from the cross-institutional credit assessment results according to a predetermined format. This serialization process converts the string representation of the credit rating and the numerical representation of the default probability into a unified byte stream format. The resulting hash value is calculated by applying the SHA-256 hashing algorithm to the serialized credit assessment data. This algorithm performs multiple rounds of bitwise and modular operations on the input data to produce a unique 256-bit hash value. The plaintext credit rating, risk rating, and credit limit data in the customer credit report are also serialized, concatenating the three data types into a continuous byte sequence in field order. The report hash value is calculated by applying the same SHA-256 algorithm to the serialized credit report data, ensuring that any slight change in the data content will result in a significant difference in the hash value. The blockchain-stored evidence data consists of two core elements: the result hash value and the report hash value. These two hash values ​​are stored on the blockchain as digital proof of data integrity.

[0051] Smart contract rules are pre-programmed, self-executing programs on the blockchain that incorporate the specific requirements of financial regulations and compliance check logic. Financial regulation matching processes read hash values ​​from blockchain-stored evidence and compare them against pre-set compliance rules. Compliance verification processes encompass multiple dimensions, including verification of the rationality of credit ratings, numerical range checks for default probability, and regulatory limits on credit limits. Compliance status indicators are generated through the smart contract's logic. A compliance status indicator is identified when all verification items pass, and a non-compliance status is identified when any verification item fails. Violation risk markers are categorized and labeled based on specific violation types, including data anomaly markers, over-limit markers, and rating discrepancies. The compliance verification result integrates the complete information of the compliance status indicator and the violation risk marker. The smart contract automatically records the verification results on the blockchain, creating an immutable proof of compliance.

[0052] The blockchain consensus mechanism is a protocol for multiple nodes in a distributed network to reach consensus on decisions, ensuring that all participating nodes maintain a unified understanding of data status. Distributed confirmation processing requires multiple financial institution nodes to independently verify and vote on compliance verification results. Each node performs confirmation operations based on the execution results of the smart contract and its own verification logic. Timestamping records encapsulate compliance verification results along with the current block generation time into the block data structure. The timestamp uses the UTC standard time format to ensure global consistency. Tamper-proofing is achieved through the blockchain's cryptographic hash chaining and distributed storage mechanism. Each block contains the hash value of the previous block, forming a continuous hash chain. Audit records contain the complete compliance verification process and results data, including key information such as verification time, participating nodes, verification results, and timestamps. Once the audit record is written to the blockchain, it cannot be unilaterally modified or deleted.

[0053] Urgency classification is a grading method based on the timeliness requirements and risk impact of financial services. Priority sorting is based on the urgency characteristics of business scenarios. Risk warning scenarios involve real-time monitoring and immediate response to customer default risks, and have the highest timeliness requirements and are classified as high priority. Daily approval scenarios include routine credit application processing and credit limit adjustments, and have moderate timeliness requirements and are classified as medium priority. Periodic reporting scenarios involve periodic reporting and statistical analysis required by regulators, and have relatively low timeliness requirements and are classified as low priority. Business urgency grading sorts the three scenarios according to their urgency, forming a grading system that includes high priority and low priority. High priority includes risk warning scenarios, low priority includes periodic reporting scenarios, and daily approval scenarios are in the middle.

[0054] Dynamic reallocation is a management method that adjusts the privacy budget allocation ratio based on changing business needs. Privacy budget allocation is calculated based on priority weights based on business urgency. High-priority business scenarios are allocated more privacy budget to reduce noise intensity, while low-priority business scenarios are allocated less privacy budget to accommodate higher noise levels. Budget adjustments for the noise weight gradient are achieved by recalculating the weight coefficients for each business scenario, which are proportional to the business priority. Budget adjustments for the noise bias gradient use the same calculation logic to ensure mathematical consistency between the budget allocations for the weight and bias gradients. The budget optimization plan includes specific budget allocation values ​​for different business scenarios. The plan clearly defines the noise intensity parameters for the weight and bias gradients in each scenario. The total budget allocation remains unchanged, but the internal allocation ratio is dynamically adjusted based on business needs.

[0055] The above describes the financial data processing method based on machine learning in the embodiment of the present application. The following describes the financial data processing system based on machine learning in the embodiment of the present application. Figure 2 In the embodiments of the present application, an embodiment of a financial data processing system based on machine learning includes: The encryption module is used to encrypt the customer credit data of each financial institution using the Paillier homomorphic encryption algorithm to obtain an encrypted credit data set, and then extract features from the encrypted credit data set based on the financial business type to obtain an encrypted risk feature vector; The segmentation module is used to segment the local gradient parameters corresponding to the encrypted risk feature vector through a secret sharing method to obtain gradient shares, and then perform secure aggregation calculation on the gradient shares according to the BGW protocol to obtain the global gradient parameters; The injection module is used to inject noise into the global gradient parameters through the differential privacy mechanism according to the privacy level of the encrypted credit dataset to obtain the privacy-preserving gradient, and then perform federated training on the privacy-preserving gradient to obtain the cross-institutional credit assessment results; The prediction module is used to perform credit risk prediction using homomorphic encryption to calculate the encrypted risk feature vector and the cross-institutional credit assessment results, thereby obtaining an encrypted credit score. The encrypted credit score is then decrypted using secure multi-party computation to obtain a customer credit report. The verification module is used to verify the regulatory compliance of cross-institutional credit assessment results and customer credit reports through the blockchain system, obtain audit records, and dynamically adjust the budget of the privacy protection gradient according to the urgency of the business to obtain a budget optimization plan.

[0056] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the machine learning-based financial data processing method.

[0057] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0058] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a machine learning-based financial data processing device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A financial data processing method based on machine learning, characterized in that: The method comprises: The customer credit data of each financial institution is encrypted using the Paillier homomorphic encryption algorithm to obtain an encrypted credit dataset. Feature extraction is then performed on the encrypted credit dataset based on the type of financial business to obtain an encrypted risk feature vector. The local gradient parameters corresponding to the encrypted risk feature vector are split using a secret sharing method to obtain gradient shares, and the gradient shares are securely aggregated and calculated according to the BGW protocol to obtain the global gradient parameters. Based on the privacy level of the encrypted credit dataset, the global gradient parameters are injected with noise through the differential privacy mechanism to obtain the privacy-preserving gradient. The privacy-preserving gradient is then federated trained to obtain the cross-institutional credit assessment results. Credit risk prediction is performed using homomorphic encryption to combine encrypted risk feature vectors and cross-institutional credit assessment results to obtain an encrypted credit score. The encrypted credit score is then decrypted using secure multi-party computation to obtain a customer credit report. The cross-institutional credit assessment results and customer credit reports are verified for regulatory compliance through the blockchain system to obtain audit records, and the budget of the privacy protection gradient is dynamically adjusted according to the business urgency to obtain a budget optimization plan.

2. The financial data processing method based on machine learning according to claim 1, characterized in that: The customer credit data of each financial institution is encrypted using the Paillier homomorphic encryption algorithm to obtain an encrypted credit dataset, and feature extraction is performed on the encrypted credit dataset based on the financial business type to obtain an encrypted risk feature vector, including: Identify and process the data types of customer credit information from various financial institutions to obtain classified credit information data including transaction amounts, account balances, credit scores, and risk levels; Inputting the classified credit information data into the Paillier homomorphic encryption algorithm to perform large prime number selection and public key and private key pair generation processing to obtain encrypted credit information data; performing grouping processing on the encrypted credit investigation data according to the financial business type, namely, personal credit business and corporate credit business, to obtain grouped encrypted data; Performing addition and multiplication calculations on the grouped ciphertext data based on homomorphic operation rules to obtain a composite ciphertext feature containing the risk indicator of the classified credit data; The composite ciphertext feature is normalized after mean and variance calculation using a Z-score transformation algorithm to obtain an encryption risk feature vector based on the grouped ciphertext data.

3. The financial data processing method based on machine learning according to claim 1, characterized in that: The local gradient parameters corresponding to the encrypted risk feature vector are divided by a secret sharing method to obtain gradient shares, and the gradient shares are securely aggregated and calculated according to the BGW protocol to obtain global gradient parameters, including: Performing local gradient calculation processing on the encrypted risk feature vector to obtain local gradient parameters including weight gradient and bias gradient; Secretly splitting the weight gradient and bias gradient in the local gradient parameter using Lagrange interpolation polynomials to obtain weight gradient shares and bias gradient shares distributed across multiple financial institution nodes; Based on the BGW protocol, the weight gradient shares and bias gradient shares from multiple financial institution nodes are respectively subjected to share verification and consistency check processing to obtain verified valid weight gradient shares and valid bias gradient shares; According to the threshold secret sharing mechanism, the effective weight gradient share and the effective bias gradient share are respectively subjected to polynomial reconstruction and aggregation operation processing to obtain a weight aggregation gradient result and a bias aggregation gradient result; The weighted aggregation gradient results and the bias aggregation gradient results are weighted according to the data contribution of each financial institution through a weighted average algorithm to obtain a global gradient parameter including a global weight gradient parameter and a global bias gradient parameter.

4. The financial data processing method based on machine learning according to claim 3, characterized in that: The effective weight gradient share and the effective bias gradient share are respectively subjected to polynomial reconstruction and aggregation operation processing according to the threshold secret sharing mechanism to obtain a weight aggregation gradient result and a bias aggregation gradient result, including: Based on the threshold judgment condition, the number of financial institution nodes participating in the aggregate calculation is subjected to threshold verification processing to obtain a valid node set that meets the threshold requirements; The effective weight gradient share in the effective node set is processed by polynomial coefficient calculation using the Lagrange interpolation formula to obtain weight gradient polynomial parameters; Performing a reconstruction operation on the original weight gradient value according to the weight gradient polynomial parameter to obtain a reconstructed weight gradient value; Performing polynomial coefficient calculation processing on the effective bias gradient share in the effective node set through the Lagrange interpolation formula to obtain bias gradient polynomial parameters, and performing reconstruction operation processing on the bias gradient original value according to the bias gradient polynomial parameters to obtain a reconstructed bias gradient value; The reconstructed weight gradient value and the reconstructed bias gradient value are respectively summed up by a vector aggregation algorithm to obtain a weight aggregation gradient result and a bias aggregation gradient result.

5. The financial data processing method based on machine learning according to claim 1, characterized in that: The method injects noise into the global gradient parameter through the differential privacy mechanism according to the privacy level of the encrypted credit dataset to obtain the privacy-preserving gradient, and performs federated training on the privacy-preserving gradient to obtain the cross-institutional credit assessment results, including: Performing privacy level classification based on the sensitivity classification of the transaction amount, account balance, credit score, and risk level in the encrypted credit investigation data set to obtain a privacy level classification including a high sensitivity level and a low sensitivity level; Based on the privacy level classification, privacy budget allocation processing is performed on the global weight gradient parameter and the global bias gradient parameter in the global gradient parameter to obtain a weight privacy budget and a bias privacy budget; Performing noise injection processing on the global weight gradient parameter and the global bias gradient parameter according to the weight privacy budget and the bias privacy budget using a Gaussian noise generation algorithm to obtain a privacy-preserving gradient including a noise weight gradient and a noise bias gradient; Inputting the privacy-preserving gradient into a federated learning training algorithm for multiple rounds of iterative training to obtain converged federated learning parameters; Based on the federated learning parameters, risk assessment calculations are performed on customer credit data to obtain a cross-institutional credit assessment result including credit rating and default probability.

6. The financial data processing method based on machine learning according to claim 1, characterized in that: The encrypted risk feature vector and the cross-institutional credit assessment results are used to predict credit risk through homomorphic encryption calculation to obtain an encrypted credit score, and the encrypted credit score is decrypted according to secure multi-party computing to obtain a customer credit report, including: Combining the encrypted risk feature vector with the credit rating and default probability in the cross-institutional credit assessment result through a feature fusion algorithm to obtain a fused credit feature vector; Performing forward propagation calculation processing on the fused credit feature vector based on a homomorphic encryption neural network to obtain an intermediate calculation result including a hidden layer output and a network activation result; Performing weight matrix product and bias vector addition processing on the intermediate calculation result through homomorphic multiplication operation to obtain a credit score calculation value in an encrypted state; Performing a homomorphic nonlinear transformation on the credit score calculation value to obtain an encrypted credit score including the customer credit rating score; The encrypted credit score is collaboratively decrypted among multiple financial institution nodes according to a secure multi-party computing protocol to obtain a customer credit report containing a plaintext credit rating, risk rating, and credit limit.

7. The financial data processing method based on machine learning according to claim 1, characterized in that: The cross-institutional credit assessment results and customer credit reports are verified for regulatory compliance through the blockchain system to obtain audit records. The budget of the privacy protection gradient is dynamically adjusted according to the business urgency to obtain a budget optimization plan, including: The credit rating and default probability in the cross-institutional credit assessment result and the plaintext credit rating, risk rating and credit limit in the customer credit report are digitally fingerprinted using a hash digest algorithm to obtain blockchain evidence data containing the result hash value and the report hash value; Based on the smart contract rules, the blockchain evidence data is matched with financial regulatory regulations and compliance verification is performed to obtain a compliance verification result including a compliance status identifier and a violation risk marker; The compliance verification results are distributed and confirmed and timestamped across multiple financial institution nodes through a blockchain consensus mechanism, resulting in an audit record that cannot be tampered with. Prioritize risk warning, daily approval, and regular reporting scenarios based on the urgency of financial services, resulting in a business urgency classification that includes high and low priorities. The privacy budgets of the noise weight gradient and the noise bias gradient in the privacy protection gradient are dynamically reallocated based on the business urgency classification to obtain budget optimization solutions for different business scenarios.

8. A financial data processing system based on machine learning, characterized in that: For implementing the financial data processing method based on machine learning according to any one of claims 1 to 7, the financial data processing system based on machine learning comprises: The encryption module is used to encrypt the customer credit data of each financial institution using the Paillier homomorphic encryption algorithm to obtain an encrypted credit data set, and then extract features from the encrypted credit data set based on the financial business type to obtain an encrypted risk feature vector; The segmentation module is used to segment the local gradient parameters corresponding to the encrypted risk feature vector through a secret sharing method to obtain gradient shares, and then perform secure aggregation calculation on the gradient shares according to the BGW protocol to obtain the global gradient parameters; The injection module is used to inject noise into the global gradient parameters through the differential privacy mechanism according to the privacy level of the encrypted credit dataset to obtain the privacy-preserving gradient, and then perform federated training on the privacy-preserving gradient to obtain the cross-institutional credit assessment results; The prediction module is used to perform credit risk prediction using homomorphic encryption to calculate the encrypted risk feature vector and the cross-institutional credit assessment results, thereby obtaining an encrypted credit score. The encrypted credit score is then decrypted using secure multi-party computation to obtain a customer credit report. The verification module is used to verify the regulatory compliance of cross-institutional credit assessment results and customer credit reports through the blockchain system, obtain audit records, and dynamically adjust the budget of the privacy protection gradient according to the urgency of the business to obtain a budget optimization plan.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the financial data processing method based on machine learning according to any one of claims 1 to 7.

Images