Multi-category peasant household identity joint verification method and system based on federated learning and national secret algorithm

Abstract

The invention discloses a multi-category peasant household identity joint verification method and system based on federated learning and a national secret algorithm, and relates to the technical field of agricultural informatization and financial science and technology crossing, and the method comprises the steps: constructing a feature set through multi-source data cleaning and association labeling; a longitudinal federated learning architecture is adopted to cooperatively train a farmer classification model, a local model of a big data development center of the Ministry of Agricultural Rural processes land circulation scale features, a local model of a communication operator processes production behavior features, gradients are encrypted through an SM4 national cryptographic algorithm, and global parameters are aggregated and updated; constructing a differentiated feature system based on evidence weight WOE and score card conversion, and performing dynamic weight evaluation on scale peasant households, production type peasant households and operation type peasant households; and binding the identity and the peasant household type by using joint hash encryption, and outputting hierarchical risk rating. The problems of high misjudgment rate, potential safety hazards of plaintext transmission and insufficient model classification adaptability caused by data islands in the prior art are solved, and accurate identity verification under cross-mechanism data is realized.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of agricultural informatization and financial technology, and in particular to a method and system for joint verification of the identities of multiple types of farmers based on federated learning and national cryptographic algorithms. Background Technology

[0002] With the deepening of agricultural modernization, agricultural financial services have placed higher demands on the accuracy, security, and efficiency of farmer identity verification and credit assessment. However, the current farmer identity verification and credit assessment system still faces many technical bottlenecks, which seriously restrict the inclusiveness and risk control capabilities of agricultural financial services.

[0003] Currently, agricultural data is scattered across different government departments and commercial institutions, lacking cross-institutional data collaboration and sharing mechanisms. Each institution can only train identity recognition models based on incomplete local data, resulting in incomplete model sample coverage and an inability to accurately capture the multi-dimensional characteristics of farmers. Furthermore, plaintext transmission is used during identity verification without data encryption. The farmer credit assessment model adopts a uniform fixed-weight algorithm framework and fails to customize risk indicators based on the differentiated characteristics of large-scale farmers, ordinary production-oriented farmers, and ordinary management-oriented farmers. In addition, the dynamic behavior trajectories of farmers are not collected and integrated in real time.

[0004] Therefore, a multi-category farmer identity joint verification method and system based on federated learning and national cryptographic algorithms is proposed to solve the above problems. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for joint verification of the identities of multiple types of farmers based on federated learning and national cryptographic algorithms, so as to solve the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms, comprising the following steps: Multi-source data collection and preprocessing: Collect relevant data of farmers from the database of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs, the family farm registration system, the village collective organization system and the interface of telecommunications operators, clean the data, handle outliers, and label farmers as large-scale farmers, ordinary production-type farmers, ordinary business-type farmers or non-farm types according to preset rules. Feature engineering and construction of a differentiated feature system: Feature selection and construction are performed on the preprocessed data. Through correlation analysis, discrimination test, recursive feature elimination and stepwise regression, the final feature set is selected, and a differentiated feature system is constructed based on the selection results. Federated learning model training and scorecard generation: Construct a vertical federated learning framework, including a central coordinator and local models of at least two participating institutions; each participating institution uses local data encrypted with the national cryptographic SM4 algorithm and its corresponding farmer type labels to train the model under a differentiated feature subset; iteratively optimize the global model through a secure aggregation mechanism, and generate scorecards based on the output of the logistic regression model; Identity verification and result output: Receive the ID card number and mobile phone number of the user to be verified, encrypt them using the national cryptographic algorithm, and verify their consistency with the data of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs; after successful verification, call the federated score card model to calculate the user's risk score; based on the score result and the farmer type, output the risk score and identity verification result.

[0007] Preferably, the preset rules include: Those who meet the criteria of having a family farm owner tag of 1 or having an inflow of land area greater than 100 mu are marked as large-scale farmers; Those who meet the criteria of having 1 village collective member identifier and having met rural production standards for ≥180 days are labeled as ordinary production-type farmers. Those who meet the rural production standards for ≥240 days are classified as ordinary operating farmers; Those who do not meet any of the above conditions will be marked as non-agricultural households.

[0008] Preferably, the feature filtering includes: Correlation analysis: Pearson correlation coefficient and Spearman correlation coefficient were used to calculate the correlation between features and farmer types, and weakly correlated features with an absolute value of correlation coefficient <0.3 were removed; Discrimination test: The differences in the mean characteristics of different types of farmers were tested by analysis of variance, and characteristics with a p-value < 0.05 were retained; Recursive feature elimination: Combining the model feature importance score, the features with the lowest contribution are gradually eliminated; Stepwise regression screening: The final feature set is screened by using forward / backward stepwise regression, based on the Wald chi-square test and the logic that the direction of the feature coefficients conforms to business common sense. Collinearity control: Calculate the variance inflation factor (VIF) of the features, remove highly collinear variables with VIF ≥ 10, and construct a differentiated feature system based on the screening results.

[0009] Preferably, the process for generating the scorecard includes: Sample definition: Positive samples Y=1 are those identified as genuine farmers by authoritative institutions, and negative samples Y=0 are those of non-farmers or those with falsified identities. The basic model includes an intercept term. Variable introduction: Variables are introduced in descending order of characteristic IV values. Wald test is used to ensure the statistical significance of variables. Variables with P > 0.05 or negative coefficients are removed. Scorecard Conversion: The linear combination of β×WOE output from logistic regression is converted into a percentage scorecard. The conversion parameters are: baseline score = 60 points, baseline odds = 2 (i.e., the initial ratio of positive to negative samples is 2:1), PDO = 10, and odds doubles for every 10 points increase. The conversion formula is: .

[0010] Preferably, the model training process includes: Data encryption: Local characteristic data is encrypted using the SM4 algorithm, and the key is stored and managed through a hardware security module and is forcibly rotated every quarter; Model initialization: The central coordinator distributes initial model parameters to each participating organization; Local training: Each institution trains sub-models based on encrypted data and type labels, generating encrypted gradients; among them, the contribution of features to farmer types is calculated by using SHAP values, and the weights of differentiated features are dynamically adjusted. Secure aggregation: Each institution transmits encrypted gradients to the central coordinator via the TLS 1.3 protocol. The coordinator uses a federated averaging algorithm to aggregate gradients and update the global model. Iterative convergence: Repeat the local training to safe aggregation process until the model meets the deployment criteria.

[0011] Preferably, the identity verification process includes: The SM4 algorithm is used to encrypt identity-sensitive information. Joint hash encryption is implemented for ID card number, mobile phone number and farmer type identifier. The hash values ​​of the three are bound by the SHA-256 algorithm to enhance the correlation between identity and behavioral characteristics. Request the data interface of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs to verify the consistency of ID card number, mobile phone number, and farmer type; After the consistency verification is passed, the federal scorecard model is invoked to calculate the risk score of farmers based on the differentiated feature system.

[0012] A joint verification system for farmer identity models based on federated learning and national cryptographic algorithms includes the following modules: Multi-source data acquisition module: Connects to the data interface of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs to collect data from farmers; Data preprocessing module: performs record deduplication, box plot outlier handling, farmer type labeling, and multi-source data association, outputting a feature dataset with classification labels; Feature engineering module: Enables correlation analysis, IV value screening, RFE feature elimination, stepwise regression screening, and VIF collinearity control to build a differentiated feature system; Federated Model Architecture Module: Deploys a vertical federated learning framework, including a central coordinator and local models of participating institutions; Encryption training module: integrates SM4 encryption algorithm, HSM key management and TLS1.3 transmission protocol to realize data encryption, key rotation and secure gradient transmission; The scorecard generation module generates percentage scorecards and sets binning rules based on the output of the logistic regression model. Identity verification module: Performs joint hash encryption of ID card number, mobile phone number, and type identifier, data consistency verification and risk score calculation by the Big Data Development Center of the Ministry of Agriculture and Rural Affairs; Results output module: Outputs identity verification results, farmer type, and risk label based on the scoring level; Model monitoring module: Periodically calculates PSI and VIF values ​​to ensure model stability and feature effectiveness.

[0013] Preferably, the multi-source data acquisition module uses an API interface and incremental synchronization to acquire data, synchronizing the update records of the database of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs once per hour and synchronizing the data of the telecommunications operator once per day.

[0014] Preferably, the key rotation in the encryption training module adopts a dual-key transition mechanism: the old and new keys are used simultaneously for 30 days before the rotation to ensure data compatibility; after the rotation is completed, the old key is destroyed immediately.

[0015] Preferably, the result output module supports multiple output formats: outputting identity verification results, risk scores, and farmer types in JSON format to financial institutions; and outputting verification result notifications in SMS format to farmers.

[0016] The present invention has the following beneficial effects: 1. In this invention, a vertical federated learning framework combined with full-link encryption using national cryptographic algorithms is adopted. This enables institutions such as the Big Data Development Center of the Ministry of Agriculture and Rural Affairs and telecommunications operators to collaboratively train a unified model without sharing sensitive original data, simply by exchanging intermediate model parameters through encryption. This not only breaks through the traditional data silos and integrates multi-dimensional information such as land, operation, and behavior to improve model accuracy, but also fundamentally eliminates the risk of data leakage and meets increasingly stringent data security regulations.

[0017] 2. In this invention, multi-category dynamic labeling is performed through a preset rule engine, and a differentiated feature system matching it is constructed. In feature engineering, the most distinctive feature subsets are selected for different types of farmers, so that the model can not only verify the authenticity of the identity, but also accurately identify the specific type of the farmer, providing financial institutions with decision-making basis that is far richer than traditional methods.

[0018] 3. In this invention, the output of the federated learning model is transformed into a standard scorecard format. Intuitive parameters such as benchmark score and PDO are used to map probability predictions into percentage scores and clear risk labels. At the same time, the system integrates a continuous monitoring mechanism for model stability and feature effectiveness, making the model's decision logic transparent and auditable. The output scores can be directly used to formulate business strategies, achieving a safe, reliable, and efficient seamless integration of artificial intelligence with core business risk control processes. Attached Figure Description

[0019] Figure 1 This is a flowchart of the multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms of the present invention; Figure 2 This is a vertical federated learning architecture diagram of the multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms of the present invention. Figure 3 This is a flowchart illustrating the identity verification process of the multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms of this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-3 This invention provides a technical solution: a multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms, comprising the following steps: Multi-source data collection and preprocessing: Collect relevant data of farmers from the database of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs, the family farm registration system, the village collective organization system and the interface of telecommunications operators, clean the data, handle outliers, and label farmers as large-scale farmers, ordinary production-type farmers, ordinary business-type farmers or non-farm types according to preset rules. Data sources include authoritative databases from the Big Data Development Center of the Ministry of Agriculture and Rural Affairs and behavioral data from telecommunications operators, specifically including: The National Rural Land Contractual Management Rights Confirmation and Registration Database of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs includes farmers' names, ID numbers, contract numbers, land plot coordinates, contracted area, contracted years, land transfer records (inflow / outflow area, transfer time), family member information, etc., and is used to verify the authenticity of farmers' land contractual rights and land transfer status. The Family Farm Registration System of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs includes information such as the business type of family farm (e.g., planting, breeding, and integrated planting and breeding), registered capital, legal representative's ID number, and registration time. It is used to identify large-scale farmers (operating area ≥ 100 mu) and commercial farmers (whose main purpose is commercial operation). The Village Collective Organization Database of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs contains information on village collective members (such as name, ID number, and time of joining) and shareholding status (such as whether they hold village collective shares), which is used to help determine the stability of the village collective and the status of farmers as production-oriented farmers (most village collective members are production-oriented farmers). Telecommunications operator data includes the number of days farmers met rural production standards (≥20 days of activity in rural areas per month is considered meeting production standards) and the number of months farmers met rural living standards (≥15 days of activity in rural areas per month is considered meeting living standards) in the past year, which are used to distinguish the behavioral characteristics of production-oriented farmers (more days meeting production standards) and management-oriented farmers (more months meeting living standards). To ensure data quality, this step involves the following preprocessing of the collected raw data: Deduplication of records: Using the ID number as the unique key, merge duplicate records such as multiple land contracting contracts and family farm registration information of the same farmer. For example, the same farmer may own multiple contracted lands, and their contracted area and land transfer area need to be merged into a total value. Outlier handling: Outlier identification is based on box plot method, and the range of outliers is defined as follows: ( It is the first quartile. It is the third quartile. (The interquartile range is used). For example, if a farmer's contracted area is 1,000 mu (far higher than the average contracted area of ​​50 mu for farmers in the same area), it is judged as an outlier and replaced with the median of the contracted area of ​​farmers in the region. Farmer type labeling: Farmers are classified according to the authoritative definition in the "Announcement on Continuing the Implementation of the Policy of Exempting Financial Institutions from Value-Added Tax on Loan Interest Income to Farmers" and in combination with data on business scale, registration type, and production / operation behavior. Data association: Using the ID card number as the unique identifier, land contract data, family farm data, village collective data and telecommunications operator data are linked to construct a multi-dimensional farmer identity feature dataset with classification labels. For example, the land contract area, registered capital, village collective member identification, and number of days of production compliance of farmers with ID card number = 11010119900101XXXX are linked into a complete record. Feature engineering and construction of a differentiated feature system: Feature selection and construction are performed on the preprocessed data. Through correlation analysis, discrimination test, recursive feature elimination and stepwise regression, the final feature set is selected, and a differentiated feature system is constructed based on the selection results. Federated learning model training and scorecard generation: Construct a vertical federated learning framework, including a central coordinator and local models of at least two participating institutions; each participating institution uses local data encrypted with the national cryptographic SM4 algorithm and its corresponding farmer type labels to train the model under a differentiated feature subset; iteratively optimize the global model through a secure aggregation mechanism, and generate scorecards based on the output of the logistic regression model; To address the data silo problem, this invention employs a vertical federated learning architecture. This architecture includes a central coordinator and multiple participating institutions' local models (such as the Big Data Development Center model of the Ministry of Agriculture and Rural Affairs and telecommunications operator models), with specific responsibilities as follows: Central coordinator: Responsible for initial model parameter distribution, encrypted gradient aggregation, and global model update. It does not access any raw data and achieves model coordination only through encrypted gradient information. The local model developed by the Big Data Development Center of the Ministry of Agriculture and Rural Affairs: Based on local data from the national rural land contracting and management rights confirmation and registration database, the family farm registration system, and the village collective organization database, feature sub-models were trained for large-scale farmers and production-oriented farmers; Local model for telecommunications operators: Based on the behavioral data of telecommunications operators (such as the number of days that meet production standards, the number of months that meet living standards, and the identification of agricultural benefit packages), feature sub-models are trained for production-oriented farmers and management-oriented farmers; Financial Institution Local Model (Optional): If financial institutions participate in the collaboration, they can train a risk assessment sub-model based on their historical credit data to further improve the model's risk discrimination capabilities. The core advantage of vertical federated learning lies in the fact that it does not share raw data, but only model parameters and encrypted gradients, which enables cross-institutional data collaboration while ensuring data privacy. Identity verification and result output: Receive the ID card number and mobile phone number of the user to be verified, encrypt them using the national cryptographic algorithm, and verify their consistency with the data of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs; after successful verification, call the federated score card model to calculate the user's risk score; based on the score result and farmer type, output the risk score and identity verification result. This step uses an encrypted identity verification process to call the federal scoring card model to output the identity verification result and risk level, achieving real-time and accurate identification. To ensure the security of identity information transmission, this invention employs the following encryption method: SM4 Encrypted Transmission: Sensitive information (such as ID card number, contract number, and number of days of production compliance) is encrypted using the SM4 algorithm in CTR mode. The key is managed by HSM to ensure that the information cannot be stolen or tampered with during transmission. Joint hash encryption: Implement joint hash encryption using ID card number, mobile phone number, and farmer type identifier, and generate the hash value using the SHA-256 algorithm. The formula is: ; in, For string concatenation, joint hash encryption strongly binds identity information to farmer type, preventing cross-type identity misuse (such as ordinary production-type farmers misusing the identity of large-scale farmers). Result Output Definition: This invention, based on a federated model, extracts corresponding features to calculate risk scores and determine farmer types, classifying identity verification results into four categories as follows: Large-scale farmers: If the risk score of the large-scale farmer model meets the standard, the output will be a large-scale farmer whose identity verification has been passed; Ordinary production-type farmers: If the risk score of the ordinary production-type farmer model meets the standard and the identity verification is passed, the farmer is defined as an ordinary production-type farmer. Ordinary operating farmers: If the risk score of the ordinary operating farmer model meets the standard and the identity verification is passed, the farmer is defined as an ordinary operating farmer. Non-farmers: The risk scores of the large-scale farmer model, the ordinary production-type farmer model, and the ordinary management-type farmer model all fail to meet the standards, resulting in an identity verification failure, and they are defined as non-farmers.

[0022] The preset rules include: Those who meet the criteria of having a family farm owner tag of 1 or having an inflow of land area greater than 100 mu are marked as large-scale farmers; Those who meet the criteria of having 1 village collective member identifier and having met rural production standards for ≥180 days are labeled as ordinary production-type farmers. Those who meet the rural production standards for ≥240 days are classified as ordinary operating farmers; Those who do not meet any of the above conditions will be marked as non-agricultural households.

[0023] Features are the core of the model. This invention adopts the principles of accurate identification, comprehensive coverage, and objective correlation, developing features from multiple dimensions and filtering them based on four principles: relevance, discriminative power, availability, and stability. The specific steps are as follows: Feature filtering includes: Correlation analysis: Pearson correlation coefficient (applicable to continuous features) and Spearman correlation coefficient (applicable to discrete features) were used to calculate the correlation between features and farmer type identifiers, and weakly correlated features with an absolute value of correlation coefficient below 0.3 were removed. For example, the correlation coefficient between farmer name and farmer type was 0.05, which is a weakly correlated feature and was removed; the correlation coefficient between the number of days meeting rural production standards in the past year and production-type farmers was 0.72, which is a strongly correlated feature and was retained. Discrimination test: The differences in the mean characteristics of different types of farmers were tested using analysis of variance (ANOVA). Characteristics with a p-value < 0.05 (i.e., the differences in characteristics between different types of farmers are statistically significant) were retained. Assuming that the mean area of ​​land inflow is 80 mu for large-scale farmers and 10 mu for production-type farmers, the ANOVA test p-value < 0.01 indicates that this characteristic has significant discrimination between large-scale farmers and production-type farmers, and is therefore retained. Recursive Feature Elimination (RFE): Combining the feature importance score of the logistic regression model, the features with the lowest contribution are gradually eliminated; assuming that the initial feature set contains 10 features, the agricultural benefit package identifier (importance score 0.02) with the lowest contribution is gradually eliminated through the RFE algorithm, and finally 9 features are retained; Stepwise regression screening: A combination of forward stepwise regression (starting from an empty model and gradually introducing significant variables) and backward stepwise regression (starting from the full model and gradually removing insignificant variables) is used to screen the final feature set based on the Wald chi-square test (P<0.05 admission criteria), ensuring that the direction of the feature coefficients aligns with common business sense. For example, if the coefficient for the number of days meeting rural production standards in the past year is positive (the more days meeting production standards, the more likely the farmer is to be a production-oriented farmer), it aligns with business logic and is retained. If a feature coefficient is negative (e.g., the coefficient for registered capital is -0.1, meaning the higher the registered capital, the less likely the farmer is to be a production-oriented farmer), but business logic requires that higher registered capital is more likely to be a large-scale farmer, then the farmer type orientation of this feature needs to be verified. If it belongs to the characteristics of large-scale farmers, it is retained; otherwise, it is removed. Collinearity control: The variance inflation factor (VIF) of the features is calculated, and highly collinear variables with VIF ≥ 10 are eliminated. Based on the screening results, a differentiated feature system is constructed. It is assumed that the VIF value of the contracted area and the land transfer area is 12.5, indicating that there is a high degree of collinearity between the two. Since the land transfer area can better reflect the operating scale of large-scale farmers, the contracted area is eliminated and the land transfer area is retained.

[0024] The process of generating scorecards includes: Sample definition: Positive samples Y=1 are those identified as genuine farmers by authoritative institutions, and negative samples Y=0 are those of non-farmers or those with falsified identities. The basic model includes an intercept term. Variable introduction: Variables are introduced in descending order of characteristic IV values. Wald test is used to ensure the statistical significance of variables. Variables with P > 0.05 or negative coefficients are removed. Scorecard Conversion: The linear combination of β×WOE output from logistic regression is converted into a percentage scorecard. The conversion parameters are: baseline score = 60 points, baseline odds = 2 (i.e., the initial ratio of positive to negative samples is 2:1), PDO = 10, and odds doubles for every 10 points increase. The conversion formula is: ; This invention uses a logistic regression model combined with stepwise regression to train an interpretable scorecard model. The specific steps are as follows: Model algorithm selection: Logistic regression is a commonly used interpretable model in the financial field. It is suitable for binary classification (farmer / non-farmer) and multi-class classification (large-scale farmer / production-type farmer / operation-type farmer) problems. Its output logistic probability can be directly converted into a score, which has good business transparency. Training data partitioning: The associated dataset is divided into training set and test set according to the proportion. The training set is used for model parameter estimation, and the test set is used for model performance verification. Feature WOE transformation: Converting continuous features (such as inflow land area) and discrete features (such as village collective member identification) into evidence weights. WOE reflects the ability of feature binning to distinguish between different types of farmers. The calculation formula is as follows: ; in, For feature binning, positive samples represent the target farmer type (e.g., large-scale farmers), and negative samples represent non-target farmer types. For example, if the inflow land area is divided into three bins: [0,50), [50,100), and [100,+∞), the proportion of positive samples for large-scale farmers in the [100,+∞) bin is 0.8, and the proportion of negative samples is 0.1. The expression is ln(0.8 / 0.1) = 2.08. Value Calculation: Information Value ( This is used to measure the ability of a feature to distinguish between different types of farmers. The calculation formula is as follows: ; The higher the IV value, the stronger the distinguishing ability of the feature. This invention retains features with an IV value ≥ 0.1 (features with an IV value < 0.1 have weak distinguishing ability and are removed). For example, the IV value of the inflow land area is 0.45, which is a high distinguishing feature and is retained; the IV value of the number of months of rural living standards met is 0.12, which is a medium distinguishing feature and is retained. Stepwise regression training: Features are introduced in descending order of IV value. The statistical significance of the features is ensured by the Wald chi-square test (P<0.05). Features with P>0.05 or whose coefficient direction does not conform to common business sense are removed. For example, after introducing the inflow land area, the Wald test P value is <0.01 and the coefficient is positive (the larger the inflow land area, the more likely it is to be a large-scale farmer), which conforms to business logic and is retained. After introducing the agricultural benefit package label, the Wald test P value is 0.12, which does not meet the significance requirement and is removed. Scoring card transformation: a linear combination of logistic regression outputs To convert to a percentage-based rating card, the conversion formula is: ; in: Baseline score: set at 60 points (corresponding to medium risk); (Points to Double Odds): Set to 10 points (i.e., for every 10 points added, ...) Doubled); benchmark Set to 2 (i.e., the baseline score corresponds to) (2:1) The probability ratio output by logistic regression. ,in The probability that the target farmer type is present. For example, a certain size of farm If the probability is 4 (meaning the probability of being a large-scale farmer is 80%), then the score is calculated as follows: ; Binning definition: The binning boundary of the scorecard is set with reference to business knowledge (such as the Sichuan agricultural cycle: two crops a year, and the number of days meeting the production standard is ≥80 days as normal) and WOE monotonicity (the WOE value of the binning increases or decreases monotonically as the characteristic value increases) to ensure the business interpretability of binning.

[0025] The model training process includes: Data encryption: Local data is encrypted using the SM4 national cryptographic algorithm (compliant with GB / T 32907-2016 standard). The encryption mode is CTR (counter mode), the key length is 256 bits, and the key is managed through the hardware security module (HSM) and rotated quarterly to ensure data security. Model initialization: The central coordinator distributes initial model parameters (such as the intercept term of logistic regression) to each participating institution. With coefficient ); Local training: Each participating institution trains a sub-model based on encrypted local data and farmer type labels, calculates encrypted gradients (using homomorphic encryption technology to encrypt the gradients), and dynamically adjusts feature weights through SHAP values ​​(SHapley Additives Planations)—SHAP values ​​reflect the contribution of a feature to the model output. If the SHAP value of a feature decreases, its weight is reduced, and vice versa. Secure aggregation: Each participating institution transmits encrypted gradients to the central coordinator via the TLS 1.3 protocol. The coordinator aggregates the gradients using secure multi-party computation (SMPC) technology and updates the global model parameters. Iterative convergence: Repeat the local training to safe aggregation process until the model converges and meets the deployment criteria, which is the final federated scorecard model.

[0026] The identity verification process includes: The SM4 algorithm is used to encrypt identity-sensitive information. Joint hash encryption is implemented for ID card number, mobile phone number and farmer type identifier. The hash values ​​of the three are bound by the SHA-256 algorithm to enhance the correlation between identity and behavioral characteristics. Request the data interface of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs to verify the consistency of ID card number, mobile phone number, and farmer type; After the consistency verification is passed, the federated scorecard model is invoked to calculate the risk score of farmers based on the differentiated feature system. The specific steps of the identity verification process are as follows: Input information collection: The user enters the ID card number and mobile phone number to be verified; Encryption request generation: The system generates a combined hash value of the ID card number, mobile phone number, and farmer type identifier (initially empty), and encrypts it using the SM4 algorithm; Consistency verification: The encrypted request is sent to the data interface of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs to verify the consistency between the ID card number and the mobile phone number (i.e., whether the mobile phone number belongs to the farmer with the ID card number). Farmer type identification: After the consistency verification is passed, the federated scorecard model is called to identify the farmer type (scale / production type / operation type). Risk score calculation: Based on the identified farmer type, extract the corresponding features (such as the area of ​​land inflow and the number of days of production compliance for large-scale farmers) and calculate a percentage score; Output results: The authentication result is output based on the score.

[0027] A joint verification system for farmer identity models based on federated learning and national cryptographic algorithms includes the following modules: Multi-source data acquisition module: Connects to the data interface of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs to collect data from farmers; Data preprocessing module: performs record deduplication, box plot outlier handling, farmer type labeling, and multi-source data association, outputting a feature dataset with classification labels; Feature engineering module: Enables correlation analysis, IV value screening, RFE feature elimination, stepwise regression screening, and VIF collinearity control to build a differentiated feature system; Federated Model Architecture Module: Deploys a vertical federated learning framework, including a central coordinator and local models of participating institutions; Encryption training module: integrates SM4 encryption algorithm, HSM key management and TLS1.3 transmission protocol to realize data encryption, key rotation and secure gradient transmission; The scorecard generation module generates percentage scorecards and sets binning rules based on the output of the logistic regression model. Identity verification module: Performs joint hash encryption of ID card number, mobile phone number, and type identifier, data consistency verification and risk score calculation by the Big Data Development Center of the Ministry of Agriculture and Rural Affairs; Results output module: Outputs identity verification results, farmer type, and risk label based on the scoring level; Model monitoring module: Periodically calculates PSI and VIF values ​​to ensure model stability and feature effectiveness.

[0028] The multi-source data acquisition module uses an API interface and incremental synchronization to acquire data, synchronizing the update records of the Ministry of Agriculture and Rural Affairs Big Data Development Center database once an hour and synchronizing the data of telecommunications operators once a day.

[0029] The key rotation in the encryption training module adopts a dual-key transition mechanism: the old and new keys are used simultaneously for 30 days before the rotation to ensure data compatibility; after the rotation is completed, the old key is destroyed immediately.

[0030] The results output module supports multiple output formats: outputting identity verification results, risk scores, and farmer types in JSON format to financial institutions; and outputting verification result notifications in SMS format to farmers.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for joint verification of the identities of multiple types of farmers based on federated learning and national cryptographic algorithms, characterized in that, Includes the following steps: Multi-source data collection and preprocessing: Collect relevant data of farmers from the database of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs, the family farm registration system, the village collective organization system and the interface of telecommunications operators, clean the data, handle outliers, and label farmers as large-scale farmers, ordinary production-type farmers, ordinary business-type farmers or non-farm types according to preset rules. Feature engineering and construction of a differentiated feature system: Feature selection and construction are performed on the preprocessed data. Through correlation analysis, discrimination test, recursive feature elimination and stepwise regression, the final feature set is selected, and a differentiated feature system is constructed based on the selection results. Federated learning model training and scorecard generation: Construct a vertical federated learning framework, including a central coordinator and local models of at least two participating institutions; each participating institution uses local data encrypted with the national cryptographic SM4 algorithm and its corresponding farmer type labels to train the model under a differentiated feature subset; iteratively optimize the global model through a secure aggregation mechanism, and generate scorecards based on the output of the logistic regression model; Identity verification and result output: Receive the ID card number and mobile phone number of the user to be verified, encrypt them using the national cryptographic algorithm, and verify their consistency with the data of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs; after successful verification, call the federated score card model to calculate the user's risk score; based on the score result and the farmer type, output the risk score and identity verification result.

2. The multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms according to claim 1, characterized in that, The preset rules include: Those who meet the criteria of having a family farm owner tag of 1 or having an inflow of land area greater than 100 mu are marked as large-scale farmers; Those who meet the criteria of having 1 village collective member identifier and having met rural production standards for ≥180 days are labeled as ordinary production-type farmers. Those who meet the rural production standards for ≥240 days are classified as ordinary operating farmers; Those who do not meet any of the above conditions will be marked as non-agricultural households.

3. The multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms according to claim 1, characterized in that, The feature filtering includes: Correlation analysis: Pearson correlation coefficient and Spearman correlation coefficient were used to calculate the correlation between features and farmer types, and weakly correlated features with an absolute value of correlation coefficient <0.3 were removed; Discrimination test: The differences in the mean characteristics of different types of farmers were tested by analysis of variance, and characteristics with a p-value < 0.05 were retained; Recursive feature elimination: Combining the model feature importance score, the features with the lowest contribution are gradually eliminated; Stepwise regression screening: The final feature set is screened by using forward / backward stepwise regression, based on the Wald chi-square test and the logic that the direction of the feature coefficients conforms to business common sense. Collinearity control: Calculate the variance inflation factor (VIF) of the features, remove highly collinear variables with VIF ≥ 10, and construct a differentiated feature system based on the screening results.

4. The multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms according to claim 1, characterized in that, The process for generating the scorecard includes: Sample definition: Positive samples Y=1 are those identified as genuine farmers by authoritative institutions, and negative samples Y=0 are those of non-farmers or those with falsified identities. The basic model includes an intercept term. Variable introduction: Variables are introduced in descending order of characteristic IV values. Wald test is used to ensure the statistical significance of variables. Variables with P > 0.05 or negative coefficients are removed. Scorecard Conversion: The linear combination of β×WOE output from logistic regression is converted into a percentage scorecard. The conversion parameters are: baseline score = 60 points, baseline odds = 2 (i.e., the initial ratio of positive to negative samples is 2:1), PDO = 10, and odds doubles for every 10 points increase. The conversion formula is: .

5. The multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms according to claim 1, characterized in that, The model training process includes: Data encryption: Local characteristic data is encrypted using the SM4 algorithm, and the key is stored and managed through a hardware security module and is forcibly rotated every quarter; Model initialization: The central coordinator distributes initial model parameters to each participating organization; Local training: Each institution trains sub-models based on encrypted data and type labels, generating encrypted gradients; among them, the contribution of features to farmer types is calculated by using SHAP values, and the weights of differentiated features are dynamically adjusted. Secure aggregation: Each institution transmits encrypted gradients to the central coordinator via the TLS 1.3 protocol. The coordinator uses a federated averaging algorithm to aggregate gradients and update the global model. Iterative convergence: Repeat the local training to safe aggregation process until the model meets the deployment criteria.

6. The multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms according to claim 1, characterized in that, The identity verification process includes: The SM4 algorithm is used to encrypt identity-sensitive information. Joint hash encryption is implemented for ID card number, mobile phone number and farmer type identifier. The hash values ​​of the three are bound by the SHA-256 algorithm to enhance the correlation between identity and behavioral characteristics. Request the data interface of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs to verify the consistency of ID card number, mobile phone number, and farmer type; After the consistency verification is passed, the federal scorecard model is invoked to calculate the risk score of farmers based on the differentiated feature system.

7. A joint verification system for farmer identity models based on federated learning and national cryptographic algorithms, comprising the multi-category farmer identity joint verification method based on federated learning and national cryptographic algorithms as described in any one of claims 1-6, characterized in that, Includes the following modules: Multi-source data acquisition module: Connects to the data interface of the Big Data Development Center of the Ministry of Agriculture and Rural Affairs to collect data from farmers; Data preprocessing module: performs record deduplication, box plot outlier handling, farmer type labeling, and multi-source data association, outputting a feature dataset with classification labels; Feature engineering module: Enables correlation analysis, IV value screening, RFE feature elimination, stepwise regression screening, and VIF collinearity control to build a differentiated feature system; Federated Model Architecture Module: Deploys a vertical federated learning framework, including a central coordinator and local models of participating institutions; Encryption training module: integrates SM4 encryption algorithm, HSM key management and TLS1.3 transmission protocol to realize data encryption, key rotation and secure gradient transmission; The scorecard generation module generates percentage scorecards and sets binning rules based on the output of the logistic regression model. Identity verification module: Performs joint hash encryption of ID card number, mobile phone number, and type identifier, data consistency verification and risk score calculation by the Big Data Development Center of the Ministry of Agriculture and Rural Affairs; Results output module: Outputs identity verification results, farmer type, and risk label based on the scoring level; Model monitoring module: Periodically calculates PSI and VIF values ​​to ensure model stability and feature effectiveness.

8. The multi-category farmer identity joint verification system based on federated learning and national cryptographic algorithms according to claim 7, characterized in that: The multi-source data acquisition module uses an API interface and incremental synchronization to acquire data, synchronizing the update records of the Ministry of Agriculture and Rural Affairs Big Data Development Center database once an hour and synchronizing the data of telecommunications operators once a day.

9. The multi-category farmer identity joint verification system based on federated learning and national cryptographic algorithms according to claim 8, characterized in that, The key rotation in the encryption training module adopts a dual-key transition mechanism: the old and new keys are used simultaneously for 30 days before the rotation to ensure data compatibility; after the rotation is completed, the old key is destroyed immediately.

10. The multi-category farmer identity joint verification system based on federated learning and national cryptographic algorithms according to claim 8, characterized in that, The output module supports multiple output formats: outputting identity verification results, risk scores, and farmer types in JSON format to financial institutions; and outputting verification result notifications in SMS format to farmers.

Images