Financial data multi-party security sharing method based on privacy protection

Abstract

The invention relates to the technical field of data sharing, in particular to a financial data multi-party security sharing method based on privacy protection. The method comprises the following steps: obtaining training parameters of internal financial data of each financial institution, wherein the training parameters are calculated by different security calculation protocols; according to the numerical value change and discreteness analysis of the current training parameter and the historical parameter, determining a numerical value anomaly index; determining an initial aggregation parameter; determining a direction change normal index according to the current and historical vector change of the training parameter and the initial aggregation parameter; determining a credible index of a current training parameter of each financial institution in combination with a numerical value anomaly index and the direction change normal index; and carrying out weighted analysis on the current federated model aggregation in combination with the credible index until a global optimum is achieved, and outputting shared data. According to the method, the weight of the financial institution can be adaptively adjusted, the accuracy and reliability of the whole federation model are improved, and the finally output shared data result is more real and accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of data sharing, and in particular to a method for secure multi-party sharing of financial data based on privacy protection. Background Art

[0002] In the digital age, data has become a critical resource driving social progress and development. The circulation and sharing of data often carries the risk of privacy leaks. Financial institutions hold vast amounts of sensitive data, and the sharing of this data can lead to the disclosure of privacy or commercial secrets. Multi-party secure computing (MPC) enables the secure sharing and computation of data among multiple parties without compromising data privacy. This effectively breaks down data barriers and enables the sharing of financial data while protecting data privacy, leading to its widespread adoption.

[0003] Related technologies use federated models to share financial data. This approach relies on training parameters uploaded locally by financial institutions. If the obtained parameters are inaccurate, such as due to tampering with parameters by a financial institution or loss or modification of the submitted training parameters during transmission, the parameters obtained during the federated model aggregation process will deviate significantly from the true optimal parameters. This will affect the federated model's optimization process and, consequently, the accuracy of the shared results. Summary of the Invention

[0004] To address the technical problem that errors in training parameters of financial institutions cause the parameters obtained during the federated model aggregation process to deviate significantly from the true optimal parameters, thereby affecting the accuracy of the shared results, the present invention provides a method for secure multi-party sharing of financial data based on privacy protection. The technical solutions adopted are as follows:

[0005] The present invention proposes a method for secure multi-party sharing of financial data based on privacy protection, the method comprising:

[0006] The central server obtains training parameters obtained by calculating the internal financial data of each financial institution using different security calculation protocols, wherein the training parameters are parameter vectors;

[0007] Determine the numerical abnormality index of the current training parameter based on the obtained numerical changes and discreteness analysis of the current training parameter and historical parameters of the same financial institution;

[0008] Based on the mean aggregation method, the federated model aggregates the training parameters of each financial institution to obtain the initial aggregation parameters; based on the vector changes between the current and historical training parameters and the initial aggregation parameters, the normal change indicators of the direction of the current aggregation analysis are determined;

[0009] Combined with the numerical anomaly index and the normal direction change index, the credibility index of the current training parameter of each financial institution is determined; combined with the credibility index, a weighted analysis of the current federated model aggregation is performed to obtain the current target aggregation parameter, and the target aggregation parameter is returned to the local client of each financial institution for federated model iteration until the global optimum is reached and the shared data is output.

[0010] Furthermore, the determination of the numerical abnormality index of the current training parameter based on the obtained numerical variation and discreteness analysis of the current training parameter and the historical parameters of the same financial institution includes:

[0011] The training parameters obtained from any secure computation protocol calculation by the same financial institution are used as the current parameters, and the training parameters obtained from the previous calculation are used as the historical parameters. The reverse difference analysis is performed on two adjacent historical parameters to obtain a historical reverse difference sequence containing all historical parameters, wherein the reverse difference analysis is performed by randomly selecting the difference between the previous and next historical parameters of two adjacent historical parameters.

[0012] According to the changing characteristics of the linear fitting of the values ​​of the elements in the historical inverse difference sequence, the downward trend characteristic index of the current parameter is determined;

[0013] According to the downward trend characteristic indicator and the discrete changes in the values ​​of the current parameter and the historical parameter, the numerical abnormality indicator of the current training parameter is determined.

[0014] Furthermore, the changing characteristics of the linear fitting based on the values ​​of the elements in the historical inverse difference sequence to determine the downward trend characteristic index of the current parameter includes:

[0015] Perform straight line fitting on the elements in the historical inverse difference sequence based on the least squares method to obtain a fitting straight line;

[0016] The slope of the fitted line is taken as the fitting slope, and the sum of the distances between the points corresponding to all elements and the fitted line is taken as the correction determination coefficient;

[0017] The fitting slope is normalized to obtain a slope influence index, and the ratio of the correction determination coefficient to the slope influence index is calculated and normalized to obtain a downward trend characteristic index.

[0018] Furthermore, based on the downward trend characteristic indicator and the discrete changes in the values ​​of the current parameter and the historical parameter, the numerical abnormality indicator of the current training parameter is determined, including:

[0019] Perform inverse difference analysis on the set of current parameters and all historical parameters to obtain the current inverse difference sequence containing the current parameters and all historical parameters;

[0020] The difference between the coefficient of variation of all data in the historical difference sequence and the coefficient of variation of all data in the current difference sequence is used as the numerator, and the downward trend characteristic index is used as the denominator to obtain the trend analysis parameter;

[0021] The trend analysis parameters are normalized to obtain numerical anomaly indicators.

[0022] Furthermore, the normal indicator of the direction change of the current aggregation analysis is determined based on the vector change of the training parameters and the initial aggregation parameters between the current and historical times, including:

[0023] According to the vector difference between the training parameters and the aggregation parameters obtained in the same training, the target vector of the corresponding training parameters is obtained;

[0024] Calculate the cosine of the angle between the target vector of the current training parameter and the target vector of each historical parameter;

[0025] The mean of all angle cosine values ​​is normalized and used as a normal indicator of directional change in the current sub-aggregation analysis.

[0026] Furthermore, combining the numerical abnormality indicator and the direction change normality indicator, a credibility indicator of the current training parameter of each financial institution is determined, including:

[0027] The difference between the normal direction change index and the abnormal value index is calculated, and the maximum and minimum values ​​are normalized as a credible index.

[0028] Furthermore, a weighted analysis of the current federated model aggregation is performed in combination with the trust indicators to obtain the current target aggregation parameters, including:

[0029] In the weighted aggregation process of the current federated model, the trustworthy indicators are directly used as numerical weights to obtain the current target aggregation parameters.

[0030] Furthermore, the method further comprises:

[0031] The local client of the financial institution obtains the target aggregation parameters output by the central server;

[0032] The training parameters of the current local client obtained through the federated model processing are iterated toward the target aggregation parameters and used as the training parameters for the next iterative training.

[0033] Furthermore, it also includes: the local client of the financial institution obtains shared data, and optimizes the federated model of the local client of the financial institution based on the shared data to obtain a local optimized model.

[0034] Furthermore, it also includes: the local client of the financial institution collects relevant internal financial data, unifies the format of the relevant financial data and performs local homomorphic encryption to obtain the first training parameters.

[0035] The present invention has the following beneficial effects:

[0036] Due to the errors in training parameters by financial institutions, the existing technology causes the parameters obtained during the federated model aggregation process to seriously deviate from the actual optimal parameters, affecting the accuracy of the shared results.

[0037] The present application scheme adaptively sets the trust indicators of each financial institution in each training, and uses the trust indicators as weights for weighted analysis. The specific weighted analysis process is: the training parameters calculated by different security computing protocols of the same financial institution are analyzed for numerical changes and discreteness, so as to accurately determine the numerical characteristic analysis of the parameters of different iterative training of the security computing protocols of the same financial institution; then, combined with the directional changes of the training parameters, the normal directional change indicators of each aggregation analysis are determined, that is, the directional change characteristic analysis is realized; finally, the two analysis methods of parameter numerical characteristic analysis and directional change characteristic analysis are combined to determine the trust indicators of each training parameter of each financial institution, and then the trust indicators of the same training parameters are combined for weight analysis. Compared with the direct mean weight analysis in the existing federal model, this scheme realizes weight adjustment by combining data fluctuations, so that the data of financial institutions with abnormal performance are given lower weights, thereby improving the accuracy and reliability of the overall federal model, making the final output shared data results more real and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 A flowchart of a method for secure multi-party sharing of financial data based on privacy protection provided by one embodiment of the present invention;

[0040] Figure 2 A schematic diagram of a model architecture provided by an embodiment of the present invention;

[0041] Figure 3 A flowchart of a method for secure multi-party sharing of financial data based on privacy protection provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0042] To further illustrate the technical means and effectiveness of the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effectiveness of a privacy-preserving multi-party financial data secure sharing method proposed by the present invention. In the following description, different references to "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0043] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0044] The following describes in detail a specific solution of a method for secure multi-party sharing of financial data based on privacy protection provided by the present invention with reference to the accompanying drawings.

[0045] It should be noted that the implementation scenario of this solution is the financial scenario, and each financial institution is analyzed. For data security reasons, each financial institution needs to independently deploy the corresponding federated model of the local client of the financial institution to avoid data leakage. Therefore, this solution is applied to the central server and the local client of the financial institution. Figure 2 , Figure 2 This is a schematic diagram of the model architecture provided by an embodiment of the present invention. The central server integrates the federated models of the local clients of all financial institutions and performs weight analysis, weighting each financial institution's local client to obtain the target aggregation parameters, and returns them to the corresponding financial institution's local client. The financial institution's local client responds to the target aggregation parameters, performs iterative optimization based on the federated model calculation, and outputs training parameters to the central server until the overall federated model converges and shared data is obtained.

[0046] See also Figure 1 , which shows a flow chart of a method for secure multi-party sharing of financial data based on privacy protection provided by one embodiment of the present invention, the method comprising:

[0047] S101: Obtain training parameters obtained by calculating the internal financial data of each financial institution using different security calculation protocols, wherein the training parameters are parameter vectors.

[0048] In the digital age, data has become a key resource for promoting social progress and development. The circulation and sharing of data are often accompanied by the risk of privacy leakage. Financial institutions are institutions with strong privacy requirements. In addition, due to the industry characteristics of financial institutions themselves, data sharing and circulation are also required between different financial institutions. Financial institutions have a large amount of sensitive data. The sharing of this data may lead to the leakage of privacy or business secrets. Therefore, when analyzing financial institutions, secure encryption and secure sharing are usually required.

[0049] This solution mainly uses multi-party secure computing to achieve data privacy protection. Multi-party secure computing can achieve secure sharing and calculation of multi-party data without leaking data privacy, thereby effectively breaking down barriers between data and realizing the sharing of financial data while protecting data privacy.

[0050] Based on the above description, in some embodiments of the present application, the implementation scenario can be specifically, for example, a credit assessment scenario, where a user's credit rating needs to be assessed when applying for financial services. Because data between financial institutions is not interoperable, a user may have an abnormal credit risk at one financial institution and then apply for a financial product at another financial institution. Because this financial institution does not have abnormal risk data for this user from other financial institutions, it will assess the user as having good credit, resulting in an inaccurate assessment of the user's true credit rating. Therefore, it is necessary to utilize multi-party secure computing to achieve the sharing of credit data while protecting privacy.

[0051] In this application, data security is shared through a federated model. The federated model requires model training. During the model training process, the central server first acquires data (i.e., training parameters). The training parameters in the embodiments of this application can be, for example, financial data collected internally by various financial institutions, such as basic customer information, transaction information, credit ratings, and other information. The training parameters are obtained through localized encryption and are multi-dimensional vector parameters.

[0052] From the above, we can see that since each financial institution reads financial data locally, obtains the corresponding user credit risk level through financial data, and uploads the training parameters, the protection and sharing of the private financial data of each financial institution can be achieved.

[0053] Secure aggregation is a crucial process for achieving privacy-preserving financial data sharing. The accuracy with which the federated model utilizes the training parameters provided by various financial institutions affects the federated model's ultimate data sharing results, which is the secure aggregation process described above. To avoid biasing individual financial institutions, averaging the parameters is often used to achieve secure aggregation.

[0054] The above requirements require that each financial institution provide accurate training parameters. If the obtained training parameters are inaccurate, for example, due to the tampering of the parameters by a certain financial institution through malicious competitive behavior, the submitted parameters are inaccurate. The malicious modification of the parameters by one party causes the parameters obtained in the aggregation process to deviate seriously from the true optimal parameters. This seriously affects the optimization process of the central server, thereby affecting the accuracy of the shared results. However, it is impossible to accurately judge whether the parameters have actually been tampered with. It may also be a normal result of the federal model error on the financial institution. Therefore, in this case, the embodiment of the present application needs to adjust the weight of the model training for the parameters provided by each financial institution. Please refer to the subsequent embodiments for details.

[0055] S102: Determine a numerical abnormality index of the current training parameter based on the obtained numerical change and discreteness analysis of the current training parameter and historical parameters of the same financial institution.

[0056] First, the numerical values ​​are analyzed and the training parameters of each financial institution are aligned to ensure that the training parameters representing the same attributes at the same position can be operated.

[0057] The central server collects training parameters from each financial institution and analyzes them based on their numerical values. This process is performed incrementally during each iteration, resulting in a parameter sequence with distinct directionality. Therefore, under normal circumstances, changes in training parameters for each financial institution exhibit a specific trend. Therefore, if a financial institution's training parameter submission exhibits significant, irregular changes, it indicates that the training parameters are likely abnormal. In this embodiment of the present invention, the central server can determine numerical anomaly indicators based on numerical change and discreteness analysis.

[0058] Record the parameters submitted by each financial institution through the secure computing protocol each time, represents the training parameters submitted by the i-th financial institution for the jth time. Each submitted parameter is represented by a parameter vector. The j-th submission also reflects the j-th training of the federated model. During the initial security aggregation process, the parameter weights of each financial institution are initialized to the same value, i.e., the federated model parameters are obtained by mean aggregation. Only after this, are the weights of the training parameters of each financial institution adjusted.

[0059] Furthermore, in some embodiments of the present invention, the numerical anomaly index of the current training parameter is determined based on the obtained numerical changes and discreteness analysis of the current training parameter and historical parameters of the same financial institution, including: taking the training parameter calculated by any security calculation protocol of the same financial institution as the current parameter, and the training parameter calculated previously as the historical parameter; calculating two adjacent historical parameters for inverse difference analysis to obtain a historical inverse difference sequence containing all historical parameters, wherein the inverse difference analysis is to randomly select the difference between the previous historical parameter and the next historical parameter in two adjacent historical parameters; determining the downward trend characteristic index of the current parameter based on the change characteristics of the numerical values ​​of the elements in the historical inverse difference sequence; determining the numerical anomaly index of the current training parameter based on the downward trend characteristic index and the numerical discrete changes of the current parameter and the historical parameter.

[0060] If the current training parameters are the kth training parameters, then the current training parameters are At this time, the historical parameters of the i-th financial institution are:

[0061]

[0062] The corresponding inverse difference operation is Thus, we get the historical inverse difference series

[0063] First, a two-dimensional coordinate system can be set up, with the elements in the historical inverse difference sequence as the vertical coordinate and time as the horizontal coordinate, to determine the coordinate point of each element in the coordinate system. Then, a straight line fitting is performed based on the least squares method, wherein the linear fitting of the least squares method is a fitting method well known to relevant technical personnel in this field, and no further limitation or elaboration is made on this.

[0064] Under normal circumstances, the overall elements show a downward trend, indicating that the training parameters are moving normally towards the optimal direction after sharing, and gradually approaching the optimal parameters with the iteration of the algorithm; on the contrary, if there are obvious parameter fluctuations, or the difference in parameters becomes larger with each generation, the possibility of parameter anomaly is greater.

[0065] In some embodiments of the present invention, the slope of the fitting straight line is used as the fitting slope, and the sum of the distances between the points corresponding to all elements and the fitting straight line is used as the correction determination coefficient; the fitting slope is normalized to obtain a slope influence index, and the ratio of the correction determination coefficient to the slope influence index is calculated and normalized as a downward trend characteristic indicator.

[0066] It can be understood that the smaller the slope, the more obvious the downward trend of the training parameters, that is, the more consistent with the normal optimal fit change, and the correction determination coefficient represents the fitting difference between the element and the fitting straight line. The smaller the value, the more consistent with the normal downward trend. Therefore, in the embodiment of the present invention, the ratio of the correction determination coefficient to the slope influence index is directly calculated and normalized as a downward trend characteristic indicator.

[0067] It should be noted that in order to ensure that the calculation results are meaningful, when performing fractional operations in the embodiments of the present invention, when encountering a situation where the denominator is 0, it is necessary to add a parameter adjustment factor greater than 0 to the denominator to prevent the denominator from being 0. The value of the parameter adjustment factor is set by the implementer according to actual conditions, and this application does not impose any special restrictions.

[0068] The downward trend characteristic index represents the downward trend characteristics. The larger the value of the downward trend characteristic index is, the more the historical parameters conform to the normal training characteristics.

[0069] Furthermore, in some embodiments of the present invention, the numerical anomaly index of the current training parameter is determined based on the downward trend characteristic index and the numerical discrete changes of the current parameter and the historical parameter, including: performing inverse differential analysis on the current parameter and the set consisting of all historical parameters to obtain a current inverse differential sequence containing the current parameter and all historical parameters; taking the difference between the coefficient of variation of all data in the historical differential sequence and the coefficient of variation of all data in the current differential sequence as the numerator and the downward trend characteristic index as the denominator to obtain the trend analysis parameter; and normalizing the trend analysis parameter to obtain the numerical anomaly index.

[0070] The current inverse difference series includes By analyzing the changes in data discreteness in the current inverse difference sequence containing the current parameters and the historical difference sequence excluding the current parameters, the impact of the current parameters on the overall discreteness can be determined.

[0071] In embodiments of the present invention, discreteness analysis can be performed using the coefficient of variation. The difference between the coefficient of variation of all data in the historical differential sequence and the coefficient of variation of all data in the current differential sequence is used as the impact of the current parameter on overall discreteness. The larger the value, the greater the impact of the current parameter, that is, the more abnormal the current parameter. Therefore, combined with the above-mentioned downward trend characteristic indicator, the larger the value, the more consistent the historical parameter with normal training characteristics. The difference in the coefficient of variation can be used as the numerator and the downward trend characteristic indicator as the denominator to obtain the trend analysis parameter. The trend analysis parameter is then normalized to obtain the numerical anomaly indicator. The numerical anomaly indicator can accurately characterize the abnormal effects of the training parameters submitted by the financial institution during different changes in the training parameters themselves.

[0072] S103: Perform federated model aggregation on each training parameter of the same financial institution based on mean aggregation to obtain initial aggregation parameters; determine the normal indicator of the direction change of the current aggregation analysis based on the vector change between the current and historical training parameters and the initial aggregation parameters.

[0073] Among them, in order to conduct weight analysis on different financial institutions in the current training, a mean weight analysis can be performed first, and the weight can be adjusted according to the mean weight analysis result to realize the adaptive adjustment function of the weight.

[0074] That is, first perform the federated model aggregation of mean aggregation to obtain the initial aggregation parameters, and then conduct a specific analysis of the changes in the current aggregation direction.

[0075] Furthermore, in some embodiments of the present invention, the normal indicator of the direction change of the current aggregation analysis is determined based on the vector change of the training parameters and the initial aggregation parameters between the current time and the history, including: obtaining the target vector of the corresponding training parameter based on the vector difference between the training parameters and the aggregation parameters obtained in the same training; calculating the cosine value of the angle between the target vector of the current training parameter and the target vector of each historical parameter; and normalizing the mean of all the cosine values ​​of the angle as the normal indicator of the direction change of the current aggregation analysis.

[0076] In assessing customer credit risk, if Financial Institution A primarily targets young entrepreneurs, its customer data has high income but low stability, resulting in a higher overall risk. Other financial institutions target mature enterprises, whose customers have stable income but less expansion needs, resulting in a lower overall risk. This results in significant parameter fluctuations during federated learning for Financial Institution A to accommodate the instability of customer credit risk. Meanwhile, parameter updates for Financial Institution B are relatively stable. Parameter changes during normal training iterations for Financial Institution A are perceived as modifications, when in reality, they are normal parameter fluctuations caused by its own data distribution.

[0077] The initial aggregate parameters obtained by mean aggregation of the federated model represent the overall optimization direction. Under normal circumstances, although there are data differences in the parameters of each financial institution in their respective local training, when the parameters are encrypted and uploaded to the federated model, the optimization direction after the aggregated parameters provided by each financial institution are shared is obtained. The parameter optimization of all financial institutions should be oriented towards this optimization direction.

[0078] Therefore, if the parameter optimization of a financial institution differs significantly from that of other financial institutions, it indicates that the parameters of that financial institution have abnormal behavior and are likely to be tampered with. This allows us to correct the possibility of abnormality.

[0079] Among them, the larger the cosine value of the angle between the two target vectors, the smaller the angle between the two target vectors and the more consistent the directions. In other words, the mean of all the cosine values ​​of the angles is normalized and used as a normal indicator of the direction change of the current aggregation analysis.

[0080] S104: Determine the trustworthy index of the current training parameters of each financial institution by combining the numerical anomaly index and the normal directional change index; perform weighted analysis of the current federated model aggregation by combining the trustworthy index to obtain the current target aggregation parameter, and return the target aggregation parameter to the local client of each financial institution for federated model iteration until the global optimum is reached and the shared data is output.

[0081] Furthermore, in some embodiments of the present invention, the numerical abnormality index and the normal direction change index are combined to determine the credibility index of the current training parameters of each financial institution, including: calculating the difference between the normal direction change index and the numerical abnormality index, and normalizing the maximum and minimum values ​​as the credibility index.

[0082] Among them, the trust index represents the degree of trust in the current training parameters of the corresponding financial institution. Combined with the direction change characteristics and numerical anomaly characteristics, the trust index is analyzed to obtain the trust index, making the trust index more reliable.

[0083] Furthermore, in some embodiments of the present invention, a weighted analysis of the current federated model aggregation is performed in combination with trusted indicators to obtain the current target aggregation parameters, including: in the current federated model weighted aggregation process, the trusted indicators are directly used as numerical weights to obtain the current target aggregation parameters.

[0084] The larger the value of the trustworthy indicator, the more normal the corresponding financial institution parameters are, and its weight can be increased. The smaller the value of the trustworthy indicator, the greater the anomaly may be in the direction fluctuation and value change. Therefore, the trustworthy indicator is directly used as the numerical weight, and the current target aggregation parameter is obtained by weighting. It should be noted that in this weighting process, the sum of all trustworthy indicators is 1.

[0085] The embodiment of the present invention adaptively sets the trustworthy index of each financial institution in each training, and uses the trustworthy index as a weight for weighted analysis. The specific weighted analysis process is as follows: the training parameters calculated by different security computing protocols of the same financial institution are analyzed for numerical changes and discreteness, so as to accurately determine the numerical characteristic analysis of the parameters of different iterative trainings of the security computing protocols of the same financial institution; then, the normal directional change index of each aggregation analysis is determined in combination with the directional change of the training parameters, that is, the directional change characteristic analysis is realized; finally, the two analysis methods of parameter numerical characteristic analysis and directional change characteristic analysis are combined to determine the trustworthy index of each training parameter of each financial institution, and then the trustworthy index of the same training parameter is combined for weight analysis. Compared with the direct mean weight analysis in the existing federated model, this solution realizes weight adjustment by combining data fluctuations, so that the data of financial institutions with abnormal performance are given a lower weight, thereby improving the accuracy and reliability of the overall federated model, and making the final output shared data results more real and accurate.

[0086] Furthermore, in some embodiments of the present invention, a local client of a financial institution can collect relevant internal financial data, standardize the format of the relevant financial data, and locally homomorphically encrypt it to obtain the initial training parameters. In these embodiments of the present invention, the collection of financial data is authorized by the relevant users, and the collection process does not violate relevant laws and regulations or violate public order and good morals.

[0087] See also Figure 3 , Figure 3 A flowchart of a method for secure multi-party sharing of financial data based on privacy protection provided in a second embodiment of the present invention includes:

[0088] S301: The local client of the financial institution obtains the target aggregation parameters output by the central server.

[0089] The overall model training in the embodiment of the present invention is an iterative process, that is, the local client of the financial institution iteratively optimizes the federated model in response to the target aggregation parameters, and outputs the training parameters to the central server until the overall federated model converges and shared data is obtained. Therefore, it is necessary to first obtain the target aggregation parameters output by the central server. The central server and the local client of the financial institution can be connected via wired or wireless connection to transmit data information, and the target aggregation parameters can be transmitted through the connection between the central server and the local client of the financial institution, that is, the local client of the financial institution obtains the target aggregation parameters output by the central server via wired or wireless means.

[0090] S302: Iterate the training parameters of the current local client obtained through the federation model processing toward the target aggregate parameters, and use them as the next training parameters for iterative training.

[0091] Each financial institution obtains training parameters based on the local data set through training on its local data. After sharing the training parameters through a multi-party security agreement, the parameters returned by the federated model differ from the training parameters of the local training. The subsequent parameter adjustments of the financial institution should be made in the direction of the global optimality after sharing.

[0092] After obtaining the target aggregation parameters, the local client uses its own decryption key to decrypt the encrypted parameters. After the federated model obtains and processes the data, the financial institution performs local homomorphic encryption on the data to obtain the training parameters.

[0093] It should be noted that all financial institutions jointly determine the federated model architecture. In this embodiment, a federated neural network is used, and the error function is mean squared error. Each financial institution performs local model training on locally encrypted data (e.g., performing gradient calculations on encrypted customer credit data) to obtain training parameters, which are also stored in encrypted form.

[0094] The updated parameters are obtained and used as the training parameters for the next round of training. The above process is repeated to optimize the overall federated model until the recall rate of the federated model is greater than 99%. The shared data is obtained and the federated model of the local client is optimized based on the shared data to obtain the local optimized model.

[0095] After training is completed, each financial institution can input the user data to be analyzed into the federated model and output the credit rating results after data sharing.

[0096] In the embodiment of the present invention, by locally acquiring target aggregation parameters and iterating the model, the overall data calculation can be performed on the local client of the financial institution, avoiding the error effects caused by multiple data transmissions, improving the data processing stability of the local client, and at the same time enhancing the local data security of the financial institution, effectively breaking down the data barriers between financial institutions, and realizing the sharing of financial data under the premise of protecting data privacy.

[0097] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0098] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A method for multi-party secure sharing of financial data based on privacy protection, characterized in that: The method comprises: The central server obtains training parameters obtained by calculating the internal financial data of each financial institution using different security calculation protocols, wherein the training parameters are parameter vectors; Determine the numerical abnormality index of the current training parameter based on the obtained numerical changes and discreteness analysis of the current training parameter and historical parameters of the same financial institution; Based on the mean aggregation method, the federated model aggregates the training parameters of each financial institution to obtain the initial aggregation parameters; based on the vector changes between the current and historical training parameters and the initial aggregation parameters, the normal change indicators of the direction of the current aggregation analysis are determined; Combined with the numerical anomaly index and the normal direction change index, the credibility index of the current training parameter of each financial institution is determined; combined with the credibility index, a weighted analysis of the current federated model aggregation is performed to obtain the current target aggregation parameter, and the target aggregation parameter is returned to the local client of each financial institution for federated model iteration until the global optimum is reached and the shared data is output.

2. The method for secure multi-party sharing of financial data based on privacy protection according to claim 1, characterized in that: The method of determining the numerical abnormality index of the current training parameter based on the obtained numerical changes and discreteness analysis of the current training parameter and the historical parameters of the same financial institution includes: The training parameters obtained from any secure computation protocol calculation by the same financial institution are used as the current parameters, and the training parameters obtained from the previous calculation are used as the historical parameters. The reverse difference analysis is performed on two adjacent historical parameters to obtain a historical reverse difference sequence containing all historical parameters, wherein the reverse difference analysis is performed by randomly selecting the difference between the previous and next historical parameters of two adjacent historical parameters. According to the changing characteristics of the linear fitting of the values ​​of the elements in the historical inverse difference sequence, the downward trend characteristic index of the current parameter is determined; According to the downward trend characteristic indicator and the discrete changes in the values ​​of the current parameter and the historical parameter, the numerical abnormality indicator of the current training parameter is determined.

3. The method for secure multi-party sharing of financial data based on privacy protection according to claim 2, characterized in that: The method of determining the downward trend characteristic index of the current parameter based on the change characteristics of the linear fitting of the values ​​of the elements in the historical inverse difference sequence includes: Perform straight line fitting on the elements in the historical inverse difference sequence based on the least squares method to obtain a fitting straight line; The slope of the fitted line is taken as the fitting slope, and the sum of the distances between the points corresponding to all elements and the fitted line is taken as the correction determination coefficient; The fitting slope is normalized to obtain a slope influence index, and the ratio of the correction determination coefficient to the slope influence index is calculated and normalized to obtain a downward trend characteristic index.

4. The method for secure multi-party sharing of financial data based on privacy protection according to claim 2, characterized in that: Determining the numerical abnormality indicator of the current training parameter based on the downward trend characteristic indicator and the discrete numerical changes of the current parameter and the historical parameter includes: Perform inverse difference analysis on the set of current parameters and all historical parameters to obtain the current inverse difference sequence containing the current parameters and all historical parameters; The difference between the coefficient of variation of all data in the historical difference sequence and the coefficient of variation of all data in the current difference sequence is used as the numerator, and the downward trend characteristic index is used as the denominator to obtain the trend analysis parameter; The trend analysis parameters are normalized to obtain numerical anomaly indicators.

5. The method for secure multi-party sharing of financial data based on privacy protection according to claim 1, characterized in that: Determining the normal indicator of the direction change of the current aggregation analysis based on the vector change of the training parameters and the initial aggregation parameters between the current and historical times includes: According to the vector difference between the training parameters and the aggregation parameters obtained in the same training, the target vector of the corresponding training parameters is obtained; Calculate the cosine of the angle between the target vector of the current training parameter and the target vector of each historical parameter; The mean of all angle cosine values ​​is normalized and used as a normal indicator of directional change in the current sub-aggregation analysis.

6. The method for secure multi-party sharing of financial data based on privacy protection according to claim 1, characterized in that: Combining the numerical abnormality indicator and the direction change normality indicator, a credibility indicator of the current training parameter of each financial institution is determined, including: The difference between the normal direction change index and the abnormal value index is calculated, and the maximum and minimum values ​​are normalized as a credible index.

7. The method for secure multi-party sharing of financial data based on privacy protection according to claim 1, characterized in that: Combined with the trust indicators, a weighted analysis of the current federated model aggregation is performed to obtain the current target aggregation parameters, including: In the weighted aggregation process of the current federated model, the trustworthy indicators are directly used as numerical weights to obtain the current target aggregation parameters.

8. The method for multi-party secure sharing of financial data based on privacy protection according to claim 1, characterized in that: The method further comprises: The local client of the financial institution obtains the target aggregation parameters output by the central server; The training parameters of the current local client obtained through the federated model processing are iterated toward the target aggregation parameters and used as the training parameters for the next iterative training.

9. The method for secure multi-party sharing of financial data based on privacy protection according to claim 8, characterized in that: Also includes: The local client of the financial institution obtains the shared data, and optimizes the federated model of the local client of the financial institution based on the shared data to obtain a local optimized model.

10. A method for secure multi-party sharing of financial data based on privacy protection as claimed in claim 9, characterized in that: Also includes: The local client of the financial institution collects relevant internal financial data, unifies the format of the relevant financial data, and performs local homomorphic encryption to obtain the first training parameters.

Images