A cross-institution tenant information processing method and system based on federated learning

Abstract

This invention discloses a cross-institutional tenant information processing method and system based on federated learning, relating to the housing rental industry. The method includes: S1, a coordinator initiates a joint modeling task, where multiple participants perform privacy set intersection based on locally stored tenant identifiers to obtain a shared tenant sample set; S2, the coordinator distributes initial global model parameters to each participant, and each participant trains its local model using its locally held feature data based on the shared tenant sample set, and calculates its local gradient; S3, each participant encrypts its local gradient and uploads it to the coordinator, which securely aggregates the received encrypted gradients in encrypted form, updates the global model parameters, and distributes them to each participant, iterating until the model converges; this improves the accuracy of the system output results.

Description

Technical Field

[0001] This invention relates to the housing rental industry, specifically to a cross-institutional tenant information processing method and system based on federated learning. Background Technology

[0002] In the current housing rental industry, for tenant credit assessment and risk prediction (such as the risk of arrears, identification of group rental behavior, etc.), the existing mainstream technical solutions mainly rely on the centralized data warehouse model or the traditional offline data exchange model. Specifically, the existing technical solutions usually take the following steps: (1) Data collection and aggregation: Large rental platforms, third-party credit reporting agencies or banks take the lead and sign offline data sharing agreements, requiring participating parties such as apartment operators, property companies, and financial institutions to export their original business data (including tenant identity information, historical payment records, default records, water, electricity and gas usage data, etc.). (2) Data cleaning and integration: The leading party imports the collected multi-source heterogeneous data into its self-built centralized big data platform or data warehouse. In this process, technical personnel need to perform a lot of manual or scripted operations to unify the data format, field definition and time granularity of different sources. (3) Sample alignment and modeling: In a centralized environment, tenants' plaintext ID numbers, mobile phone numbers and other unique identifiers are used to perform multi-table association (Join operation) to build a training dataset containing all features. Subsequently, data scientists train traditional machine learning models (such as logistic regression, XGBoost, random forest, etc.) or deep learning models on the complete dataset to output credit scores or risk probabilities. (4) Model deployment and application: The trained model is deployed on the server of the host company. When a new tenant applies for a room, the system calls the model, inputs the tenant's existing data on the host company, and outputs the risk assessment results.

[0003] The above-mentioned solutions have the following drawbacks: First, data silos lead to single model features and limited accuracy. Due to the strict prohibition of "raw data leaving the country / domain" by laws and regulations, as well as the commercial concerns of various parties regarding the leakage of core data assets, most small and medium-sized apartment operators and third-party institutions refuse to upload raw data to centralized platforms. This results in existing models often only being able to be trained based on the single-dimensional data of the leading party (such as payment records alone), lacking the complementarity of cross-domain features. Second, plaintext identity alignment poses a serious privacy leakage risk. When performing multi-source data fusion, existing technologies must rely on plaintext user unique identifiers (such as ID card numbers and mobile phone numbers) for sample matching. In this process, participating parties must expose their entire customer list to the leading party or other parties; thirdly, there is a lack of fair profit distribution and incentive mechanisms, making it difficult to form a large-scale alliance. In the traditional offline data exchange model, data contributors (such as small apartment operators) lose control of the data after handing it over, and it is difficult to quantify the specific contribution of their data to the final model's effectiveness; fourthly, model updates are lagging behind and cannot adapt to dynamic risk changes. Traditional models usually adopt periodic (such as monthly or quarterly) batch data extraction and model retraining methods. Due to the long data flow chain and high coordination costs, the model is unable to absorb the latest default behaviors or abnormal living patterns in real time, resulting in lagging risk control measures. Summary of the Invention

[0004] This invention provides a cross-institutional tenant information processing method and system based on federated learning to achieve a balance between privacy protection and data value, and improve the efficiency and applicability of joint modeling.

[0005] This invention provides the following technical solution: a cross-institutional tenant information processing method based on federated learning, comprising:

[0006] S1. The coordinator initiates a joint modeling task, and multiple participants perform privacy set intersection based on locally stored tenant identifiers to obtain a shared tenant sample set.

[0007] S2. The initial global model parameters are distributed to each participating party in the coordination direction. Each participating party, based on the shared tenant sample set, uses the feature data it holds locally to train the local model and calculate the local gradient.

[0008] S3. Each participant encrypts its local gradient and uploads it to the coordinator. The coordinator securely aggregates the received encrypted gradients in encrypted form, updates the global model parameters, and distributes them to each participant. The training continues iteratively until the model converges.

[0009] As a further improvement of the present invention, in S1, the heterogeneous identifiers are preprocessed before the intersection of the privacy sets, including:

[0010] Each participating party calls its local identifier mapping rule base to convert the original local tenant identifiers into a unified intermediate identifier format;

[0011] Encryption is performed based on the data type and security level of the participants;

[0012] The coordinating or participating parties perform intersection calculations based on the processed set of encrypted identifiers, and only output the indices of shared samples.

[0013] As a further improvement of the present invention, during the encryption process, for a massive set of identifiers, a pre-screening technique based on a counting Bloom filter is first used to reduce the amount of data involved in the core encryption calculation; then, an elliptic curve cryptography algorithm is used to encrypt the identifiers to generate an encrypted identifier set.

[0014] As a further improvement of the present invention, the encryption steps of the local gradient are as follows:

[0015] Based on historical training data or the current gradient distribution, dynamically determine the gradient magnitude threshold, and set gradients below the threshold to zero to achieve sparsity.

[0016] Low-bit quantization technology is used to map full-precision floating-point gradients to fixed-point representations with a finite number of bits, thereby reducing the amount of data transmitted.

[0017] Homomorphic encryption is used to encrypt the compressed gradient.

[0018] As a further improvement of the present invention, during or after federated training, the coordinator calculates the multi-dimensional contribution of each participant to the improvement of the global model performance based on the gradient information uploaded by each participant; the multi-dimensional contribution and related training process information are recorded on the blockchain for notarization; based on the contribution notarized on the blockchain, dynamic weight adjustment or industry digital rights allocation for each participant is automatically executed through smart contracts.

[0019] As a further improvement to this invention, the multi-dimensional contribution is calculated using the Shapley value marginal contribution algorithm, with the following steps:

[0020] The coordinator first determines the model's base performance when no participants contribute; then, it collects all local gradients uploaded by all participants.

[0021] The gradient of a certain participant is added to / removed from the current participant set in turn, the model is evaluated in turn, the performance changes are compared, and the marginal contribution value of that participant to the improvement of model performance under different combinations is obtained.

[0022] Based on the marginal contributions of all subsets, and following the weighting rules of Shapley values, the comprehensive Shapley contribution value for each party is obtained through approximate calculation.

[0023] A federated learning-based cross-institutional tenant information processing system, applicable to the aforementioned federated learning-based cross-institutional tenant information processing method, includes:

[0024] Multiple participating clients are deployed in the intranet environment of each data holder. Each participating party only accesses the original data in this local environment, and the original data does not leave the local area.

[0025] The coordinator server is responsible for initializing the global model, safely aggregating gradient parameters, distributing global model updates, and executing the coordination logic in the improved privacy set intersection protocol.

[0026] As a further improvement of the present invention, each client includes:

[0027] The local data preprocessing and storage module is used to process and maintain local heterogeneous data;

[0028] The local model trainer is used to perform local forward and backward propagation calculations;

[0029] The gradient compression module is used to compress local gradients;

[0030] The homomorphic encryption / decryption module is used to encrypt gradients and decrypt the distributed model parameters.

[0031] As a further improvement of the present invention, the coordinating server includes:

[0032] A sample alignment engine for coordinating and executing an improved privacy-preserving set intersection protocol;

[0033] The secure aggregation module is used to securely aggregate cryptographic gradients from each participant in a ciphertext state and manage the updates and distribution of the global model.

[0034] The contribution calculation module is used to calculate the Shapley value contribution of each participant based on gradient information.

[0035] As a further improvement of the present invention, it also includes a blockchain evidence storage network for recording key hash values, contribution scores of participants, and incentive allocation records during the federated learning process, ensuring that the process is tamper-proof and traceable.

[0036] The present invention has the following beneficial effects:

[0037] 1. This invention, through improved privacy-preserving intersection technology and homomorphic encryption mechanism, ensures that sensitive data such as the tenant's original identity information and transaction records are always kept locally throughout the entire modeling lifecycle and never leave the data holder's intranet in plaintext form. Even the coordinator or intermediary can only access the encrypted gradient or intersection results and cannot reconstruct the original data.

[0038] 2. Sparsity processing is achieved through dynamic gradient magnitude thresholding, combined with low-bit quantization for further data compression. This dual approach reduces the total amount of gradient data, significantly lowering network bandwidth consumption and transmission time. Using dynamic thresholding to adaptively select high-contribution gradients, rather than fixed thresholds or random discarding, maximizes the retention of effective gradient information crucial for model updates. Combined with low-bit quantization to control accuracy loss, data lightweighting is achieved while ensuring that model training convergence speed and final accuracy are largely unaffected, balancing compression efficiency and model performance. Because it only transmits key gradient information, it adapts to the uneven network infrastructure of participants in the housing rental industry, enabling more small and medium-sized institutions to join the alliance with low barriers to entry, significantly reducing communication bandwidth consumption and improving the efficiency of large-scale concurrent training.

[0039] 3. Due to the use of distributed training and encrypted aggregation, even if a participating node is compromised or the coordinator is maliciously invaded, the attacker will not be able to obtain the original data or complete model parameters of other parties, effectively defending against data theft and model reverse engineering attacks; the physical architecture of distributed data storage combined with the logical protection of homomorphic encryption builds a defense-in-depth system, avoiding the single point of failure risk of a centralized data warehouse collapsing once it is leaked.

[0040] 4. By quantifying contributions based on Shapley values ​​and storing them on the blockchain, the system can objectively evaluate the value of each party's data. On the one hand, participants with high-quality data and significant contributions can obtain higher weight or commercial rewards, thereby incentivizing all parties to continuously provide high-quality data, which is beneficial for building a global model. On the other hand, this invention transforms abstract data contributions into verifiable and tamper-proof on-chain credentials, establishing a foundation of trust, making long-term cross-institutional cooperation possible, and helping to create an industry-level "credit-based rental" standard. Attached Figure Description

[0041] Figure 1 This is a screenshot of the homepage of the coordinating server.

[0042] Figure 2 This is another screenshot of the homepage of the coordinating server.

[0043] Figure 3 This is a diagram of the interface for federated training management in the coordinating server.

[0044] Figure 4 This is a graphical interface diagram for incentivizing contributions within the coordinating server.

[0045] Figure 5 This is another interface diagram for the contribution incentive in the coordinating server.

[0046] Figure 6 A graphical interface diagram for finding the intersection of privacy sets in the coordinating server.

[0047] Figure 7 Another graphical interface diagram for finding the intersection of privacy sets in the coordinating server.

[0048] Figure 8 Another graphical interface diagram for finding the intersection of privacy sets in the coordinating server.

[0049] Figure 9 This is a training interface diagram for a vertical federated learning model in the coordinating server.

[0050] Figure 10 Another training interface diagram for the vertical federated learning model in the coordinator server.

[0051] Figure 11 This is a diagram of the interface for the Federated Learning Contribution Incentive Center in the coordinating server.

[0052] Figure 12 This is another screenshot of the interface for the Federated Learning Contribution Incentive Center in the coordinator server.

[0053] Figure 13 This is another screenshot of the interface for the Federated Learning Contribution Incentive Center in the coordinator server.

[0054] Figure 14 This is another screenshot of the interface for the Federated Learning Contribution Incentive Center in the coordinator server.

[0055] Figure 15 A flowchart for finding the intersection of privacy sets.

[0056] Figure 16 This is a block diagram of the module architecture for the participating client.

[0057] Figure 17 This is a flowchart illustrating the iterative training process of a longitudinal federated learning model. Detailed Implementation

[0058] The technical solutions of the embodiments of this specification will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of this specification and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of this specification.

[0059] To facilitate understanding of this invention, the following terms are now explained:

[0060] Shapley: The Shapley value is a classic method in game theory used to fairly distribute the benefits of cooperation. Its core is to calculate the average marginal contribution of each participant in all possible alliances.

[0061] Example 1

[0062] Please see Figures 1-17 As shown, a cross-institutional tenant information processing method based on federated learning includes:

[0063] S1. The coordinator initiates a joint modeling task, and multiple participants perform privacy set intersection based on locally stored tenant identifiers to obtain a shared tenant sample set.

[0064] In S1, heterogeneous identifiers are preprocessed before intersection of privacy sets, including:

[0065] Each participating party calls its local identifier mapping rule base to convert the original local tenant identifiers into a unified intermediate identifier format;

[0066] Based on the data types and security levels of the participants, an asymmetric encryption process is adopted: For high-sensitivity identifier sets from public security and financial institutions, a high-security privacy intersection protocol based on blind signatures or zero-knowledge proofs is used; for massive identifier sets from property management and apartment operators, a pre-screening technique based on counting Bloom filters is first used to reduce the amount of data involved in the core encryption calculation; finally, elliptic curve cryptography is used to encrypt the identifiers, generating an encrypted identifier set. This approach satisfies the data privacy protection requirements of different security levels while avoiding resource waste caused by uniform encryption, achieving a balance between security and execution efficiency. The Bloom filter pre-screening significantly eliminates non-candidate data before the core encryption intersection, significantly reducing the amount of data involved in elliptic curve cryptography and intersection calculations, lowering computational overhead and communication bandwidth consumption, making it particularly suitable for efficient intersection in scenarios with massive numbers of tenant identifiers. Using elliptic curve cryptography to encrypt identifiers, while ensuring high-strength encryption security, has advantages over traditional encryption algorithms such as shorter key length, faster computation speed, and lower resource consumption, making it more suitable for lightweight secure computation in multi-party federated scenarios.

[0067] The coordinating or participating parties perform intersection calculations based on the processed set of encrypted identifiers, and only output the indices of shared samples.

[0068] Privacy Set Intersection (PSI) Calculation: Suppose the participants hold sets X and Y respectively. Using an ECDH-type privacy set intersection protocol, the original sample elements are first hashed, and then the shared tenant sample set S is obtained through double-ended encrypted verification. The core calculation formula is as follows: .

[0069] In the above, the system only outputs the tenant sample index that is shared by all parties, and neither party can know the information of the non-shared samples, nor can it deduce the other party's complete user list.

[0070] Taking a joint modeling task for tenant default risk initiated by a city's housing rental safety joint management platform (the coordinator) as an example, the participants include: the city's public security bureau's population data center (public security agency), the retail finance department of a local joint-stock commercial bank (financial institution), the city's property service association (property management agency), and a leading chain long-term rental apartment operating company (apartment operator). The four parties each store local tenant identity data and must align the shared tenant samples without leaking the original data or exposing the full user list.

[0071] First, each participating party calls its local pre-built identifier mapping rule library to convert the heterogeneous original tenant identifiers in terms of format and type into unified intermediate identifiers:

[0072] Original identifiers from the public security authorities: ID card number, household registration code, and actual population registration number;

[0073] Original identifiers for financial institutions: ID card number, bank-registered mobile phone number, and credit user code;

[0074] Original identifiers from the property management company: owner / tenant name, mobile phone number, and apartment number;

[0075] Apartment operator's original identifiers: contracted mobile phone number, anonymized ID number, check-in order ID;

[0076] After mapping and transformation, a hash-type intermediate identity identifier without plaintext privacy information is uniformly generated, thus completing the standardization of multimodal identifier format.

[0077] Secondly, the system executes an asymmetric encryption process based on the data sensitivity of the participants:

[0078] Highly sensitive identifiers for public security and financial institutions: A high-security privacy intersection protocol combining blind signatures and zero-knowledge proofs is directly used for preprocessing to verify the legality of the identifiers without disclosing plaintext;

[0079] The massive low-sensitivity identifiers of property and apartment operators (approximately 160,000 for property and 90,000 for apartments) are first pre-screened using a count Bloom filter to remove invalid and duplicate identifiers, reducing the amount of data involved in the encryption calculation by approximately 35%.

[0080] All participants ultimately use the Elliptic Curve Cryptography (ECC) algorithm to encrypt the intermediate identifier, generating a set of encrypted identifiers that only support ciphertext matching.

[0081] Finally, the coordinator gathers the encrypted identifiers from all four parties and performs the intersection calculation of the privacy sets in ciphertext:

[0082] The entire calculation process does not decrypt any data or transmit plaintext information.

[0083] The system ultimately outputs only a list of numerical indexes of shared tenant samples. The indexes are serial numbers with no actual semantic meaning and are not associated with any privacy information such as identity, address, or finance.

[0084] Each participating party can only obtain the records of the corresponding shared tenants in its own database through the local index mapping relationship. They cannot know the non-shared sample information of other parties, nor can they deduce the complete tenant list of the other party.

[0085] The four parties completed the alignment of shared tenant samples, meeting the basic data requirements for joint modeling. At the same time, it ensured that: highly sensitive government and financial data were not leaked, the full list of tenants for properties and apartments was not exposed, and only shared information was available, strictly complying with data privacy and security requirements.

[0086] S2. The initial global model parameters are distributed to each participating party in the coordination direction. Each participating party, based on the shared tenant sample set, uses the feature data it holds locally to train the local model and calculate the local gradient.

[0087] The local gradient is the first-order partial derivative of the model loss function with respect to the global model parameters. This gradient is calculated locally by each participant and is based only on the shared tenant sample set and local feature data, without involving non-shared samples or original privacy data. The formula for calculating the local gradient is: ,in,

[0088] w: Initial global model parameters issued to each participating party in the coordination direction;

[0089] S: The set of shared tenant samples obtained by intersection calculation of privacy sets;

[0090] D i : Feature data held locally by the i-th participant;

[0091] L(w;D i S): Under the global model w, based on local features D iThe model loss function constructed with the shared samples S;

[0092] g i : The local gradient calculated by the i-th participant.

[0093] S3. Each participant encrypts its local gradient and uploads it to the coordinator. The coordinator securely aggregates the received encrypted gradients in encrypted form, updates the global model parameters, and distributes them to each participant. The training continues iteratively until the model converges.

[0094] The encryption steps for the local gradient are as follows:

[0095] Based on historical training data or the current gradient distribution, the gradient magnitude threshold is dynamically determined to retain effective gradient information that plays a key role in model updates. Gradients below the threshold are then set to zero to achieve sparsity. This achieves data lightweighting while ensuring the model training convergence speed and final accuracy.

[0096] Low-bit quantization technology is used to map full-precision floating-point gradients into fixed-point representations with a finite number of bits, thereby reducing the amount of data transmitted and the storage pressure during intermediate transmission.

[0097] Homomorphic encryption is used to encrypt the compressed gradient, thereby alleviating the computing burden on terminals or edge nodes.

[0098] Gradient sparsification, quantization, and homomorphic encryption computation: The local gradient g obtained above... i Dynamic amplitude threshold sparsification, low-bit quantization compression, and homomorphic encryption are performed sequentially to obtain the encrypted ciphertext gradient, calculated using the following formula:

[0099] ,

[0100] τ(g i ) is based on local gradient g i The dynamically generated amplitude threshold is distributed adaptively, Δ is the low-bit quantization step size, round(·) is the floor function, HE(·) is the homomorphic encryption function, and Enc(g i (final) represents the final encrypted gradient uploaded to the coordinator, used for subsequent encrypted aggregation and global model updates.

[0101] Taking the following scenario as an example, the specific implementation process of S2 and S3 is as follows:

[0102] Coordinating party: Housing rental platform operation center;

[0103] Participant A: A rental agency in a certain city (holding the tenant's performance characteristics);

[0104] Participant B: Credit reporting agency (holding the tenant's credit profile);

[0105] Shared tenant sample: 1000 shared tenant IDs were obtained through privacy requests;

[0106] Global Model: A logistic regression model used for predicting tenant default risk;

[0107] Model dimensions: The weight vector has 8 dimensions.

[0108] First, the coordinator generates an initial global model, including weight vectors, bias terms, and learning rate, and sends the parameters to participants A and B through a secure channel, whereby both parties load the model locally.

[0109] Secondly, each participating party conducts local training based on the shared tenant sample:

[0110] Taking the first iteration as an example:

[0111] Participant A (Leasing Agency)

[0112] Local characteristics: 4 dimensions including rent payment ratio, number of historical overdue payments, length of lease, and income stability.

[0113] Forward inference and loss calculation were performed on a sample of 1000 tenants to obtain the local gradient.

[0114] Participant B (Credit Reporting Agency)

[0115] Local characteristics: four dimensions including credit score, debt ratio, number of inquiries, and overdue records.

[0116] Forward inference and loss calculation were performed on a sample of 1000 tenants to obtain the local gradient.

[0117] Finally, the participants first clear the small gradients to zero, quantize and compress them, and then homomorphically encrypt and upload them; the coordinator adds the ciphertext gradients directly without decryption, and then updates the global model after decryption, repeating the process until the model is stable.

[0118] S4. The trained model is deployed on the rental platform's server. When a new tenant applies for check-in, the system calls the model and outputs risk assessment data about the tenant.

[0119] During or after federated training, the coordinator calculates the multi-dimensional contribution of each participant to the performance improvement of the global risk control model based on the gradient information uploaded by each participant; the multi-dimensional contribution and related training process information are recorded on the blockchain for notarization; based on the contribution notarized on the blockchain, dynamic weight adjustment or industry digital rights allocation for each participant is automatically executed through smart contracts.

[0120] The multi-dimensional contribution is calculated using the Shapley value marginal contribution algorithm, and the calculation steps are as follows:

[0121] The coordinator first determines the model's base performance when no participants contribute; then, it collects all local gradients uploaded by all participants.

[0122] The gradient of a certain participant is added to / removed from the current participant set in turn, the model is evaluated in turn, the performance changes are compared, and the marginal contribution value of that participant to the improvement of model performance under different combinations is obtained.

[0123] Based on the marginal contributions of all subsets, and following the weighting rules of Shapley values, the comprehensive Shapley contribution value for each party is obtained through approximate calculation.

[0124] For participant i, its Shapley value ϕ i The formula for calculating (v) is:

[0125] ;

[0126] Where N = {1, 2, ..., n}: the set of all participating parties.

[0127] n: Total number of participants

[0128] S: Any subset that does not include participant i.

[0129] v(S): The value / contribution (model performance, benefits, etc.) generated by the subset S alone.

[0130] By quantifying contributions based on Shapley values ​​and storing them on the blockchain, the system can objectively evaluate the value of each party's data. On the one hand, participants with high-quality data and significant contributions can obtain higher weight or commercial rewards, thereby incentivizing all parties to continuously provide high-quality data, which is beneficial for building a global model. On the other hand, this invention transforms abstract data contributions into verifiable and tamper-proof on-chain credentials, establishing a foundation of trust, making long-term cross-institutional cooperation possible, and helping to create an industry-level "credit-based rental" standard.

[0131] Example 2

[0132] A federated learning-based cross-institutional tenant information processing system, applicable to the aforementioned federated learning-based cross-institutional tenant information processing method, includes:

[0133] Multiple participating client applications are deployed within the intranet environment of each data holder. Each participating party accesses the original data only within this local environment; the original data does not leave the local machine. Each client includes:

[0134] The local data preprocessing and storage module is used to process and maintain local heterogeneous risk control data;

[0135] The local model trainer is used to perform local forward and backward propagation calculations;

[0136] The gradient compression module is used to compress local gradients;

[0137] The homomorphic encryption / decryption module is used to encrypt gradients and decrypt the distributed model parameters.

[0138] The coordinator server is responsible for global model initialization, secure aggregation of gradient parameters, global model update distribution, and executing the coordination logic in the improved privacy set intersection protocol. The coordinator server includes:

[0139] A sample alignment engine for coordinating and executing an improved privacy-preserving set intersection protocol;

[0140] The secure aggregation module is used to securely aggregate encrypted gradients from all participants in encrypted form and manage the updates and distribution of the global risk control model.

[0141] The contribution calculation module is used to calculate the Shapley value contribution of each participant based on gradient information.

[0142] A blockchain-based evidence storage network is used to record key hash values, participant contribution scores, and incentive distribution records during the federated learning process, ensuring that the process is tamper-proof and traceable.

[0143] Because of the use of distributed training and encrypted aggregation, even if a participating node is compromised or the coordinator is maliciously invaded, attackers cannot obtain the original data or complete model parameters of other parties, effectively preventing data theft and model reverse engineering attacks. The physical architecture of distributed data storage combined with the logical protection of homomorphic encryption builds a defense-in-depth system, avoiding the single point of failure risk of a centralized data warehouse collapsing once it is leaked.

[0144] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.

Claims

1. A cross-institutional tenant information processing method based on federated learning, characterized in that, include: S1. The coordinator initiates a joint modeling task, and multiple participants perform privacy set intersection based on locally stored tenant identifiers to obtain a shared tenant sample set; S2. The initial global model parameters are distributed to each participating party in the coordination direction. Each participating party, based on the shared tenant sample set, uses the feature data it holds locally to train the local model and calculate the local gradient. S3. Each participant encrypts its local gradient and uploads it to the coordinator. The coordinator securely aggregates the received encrypted gradients in encrypted form, updates the global model parameters, and distributes them to each participant. The training continues iteratively until the model converges.

2. The cross-institutional tenant information processing method based on federated learning according to claim 1, characterized in that, In S1, heterogeneous identifiers are preprocessed before intersection of privacy sets, including: Each participating party calls its local identifier mapping rule base to convert the original local tenant identifiers into a unified intermediate identifier format; Encryption is performed based on the data type and security level of the participants; The coordinating or participating parties perform intersection calculations based on the processed set of encrypted identifiers, and only output the indices of common samples.

3. The cross-institutional tenant information processing method based on federated learning according to claim 2, characterized in that, During encryption processing, for a massive set of identifiers, a pre-screening technique based on a counting Bloom filter is first used to reduce the amount of data involved in the core encryption calculation; then, an elliptic curve cryptography algorithm is used to encrypt the identifiers to generate an encrypted identifier set.

4. The cross-institutional tenant information processing method based on federated learning according to claim 1, characterized in that, The encryption steps for local gradients are as follows: Based on historical training data or the current gradient distribution, dynamically determine the gradient magnitude threshold, and set gradients below the threshold to zero to achieve sparsity. Low-bit quantization technology is used to map full-precision floating-point gradients to fixed-point representations with a finite number of bits, thereby reducing the amount of data transmitted. Homomorphic encryption is used to encrypt the compressed gradient.

5. The cross-institutional tenant information processing method based on federated learning according to claim 1, characterized in that, During or after federated training, the coordinator calculates the multi-dimensional contribution of each participant to the improvement of the global model performance based on the gradient information uploaded by each participant; the multi-dimensional contribution and related training process information are recorded on the blockchain for notarization; based on the contribution notarized on the blockchain, dynamic weight adjustment or industry digital rights allocation for each participant is automatically executed through smart contracts.

6. The cross-institutional tenant information processing method based on federated learning according to claim 5, characterized in that, The multi-dimensional contribution is calculated using the Shapley value marginal contribution algorithm, and the calculation steps are as follows: The coordinator first determines the model's base performance when no participants contribute; then, it collects the local gradients uploaded by all participants. The gradient of a certain participant is added to / removed from the current participant set in turn, the model is evaluated in turn, the performance changes are compared, and the marginal contribution value of that participant to the improvement of model performance under different combinations is obtained. Based on the marginal contributions of all subsets, and following the weighting rules of Shapley values, the comprehensive Shapley contribution value for each party is obtained through approximate calculation.

7. A cross-institutional tenant information processing system based on federated learning, applicable to the cross-institutional tenant information processing method based on federated learning as described in any one of claims 1-6, characterized in that, include: Multiple participating clients are deployed in the intranet environment of each data holder. Each participating party only accesses the original data in this local environment, and the original data does not leave the local area. The coordinator server is responsible for initializing the global model, safely aggregating gradient parameters, distributing global model updates, and executing the coordination logic in the improved privacy set intersection protocol.

8. The cross-institutional tenant information processing system based on federated learning according to claim 7, characterized in that, Each client includes: The local data preprocessing and storage module is used to process and maintain local heterogeneous data; The local model trainer is used to perform local forward and backward propagation calculations; The gradient compression module is used to compress local gradients; The homomorphic encryption / decryption module is used to encrypt gradients and decrypt the distributed model parameters.

9. The cross-institutional tenant information processing system based on federated learning according to claim 7, characterized in that, The coordinator servers include: A sample alignment engine for coordinating and executing an improved privacy-preserving set intersection protocol; The secure aggregation module is used to securely aggregate cryptographic gradients from each participant in a ciphertext state and manage the updates and distribution of the global model. The contribution calculation module is used to calculate the Shapley value contribution of each participant based on gradient information.

10. The cross-institutional tenant information processing system based on federated learning according to claim 7, characterized in that, It also includes a blockchain-based evidence storage network to record key hash values, participant contribution scores, and incentive allocation records during the federated learning process, ensuring that the process is tamper-proof and traceable.

Images