Multi-node dense state regression model training method
By employing a multi-node dense-state regression model training method and utilizing homomorphic encryption and blockchain technology, the security and efficiency issues of data sharing and model training in the insurance risk control field are solved, enabling efficient and secure model training and data sharing among multiple insurance companies.
Patent Information
- Application Number
- CN202510913917.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-03
- Publication Date
- 2025-10-28
AI Technical Summary
In the field of insurance risk control, existing technologies struggle to effectively utilize high-order machine learning models due to poor interpretability. Traditional univariate linear regression models also suffer from complex data security management and high computational costs when collaborating among multiple insurance companies, and lack effective data sharing and modeling training mechanisms.
A multi-node dense-state regression model training method is adopted, which utilizes homomorphic encryption and blockchain technology to achieve secure data sharing and model training through smart contracts. This ensures secure data transmission and computation among all parties. The least squares method is combined for model training, and key management and decryption auditing mechanisms are provided.
It enables remote collaborative modeling and training among multiple insurance companies, reducing costs, improving model training efficiency and data transmission security, and ensuring model interpretability and data privacy protection.
Smart Images
Figure BDA0005480949900000034 
Figure BDA0005480949900000035 
Figure BDA0005480949900000036
Abstract
Description
Technical Field
[0001] This invention relates to risk control models, and more particularly to a method for training risk control models in the insurance industry. Background Technology
[0002] Simple linear regression is a frequently used regression model. Due to its ease of use, relatively accurate results, and excellent interpretability, it is often used in the field of insurance risk control. Because the insurance risk control field highly values model interpretability, while more advanced machine learning techniques such as neural networks can achieve better predictive results, their poor interpretability prevents their application in insurance risk control.
[0003] Specifically, in the insurance sector, non-traditional products such as new energy vehicle insurance and mobile phone insurance have emerged in recent years. Assessing the risk level of these new products has become a problem hindering the industry's development. Building risk control models for new insurance products requires a large amount of historical business data, including policyholder information, operational status (e.g., mileage and accident history of a particular vehicle model and common road conditions in new energy vehicle insurance), and claims data. However, this data requires time to accumulate, making it impossible to build a high-quality risk control model in the short term.
[0004] Because the data used in the modeling is direct business data from each company, possessing high value and being considered confidential by each company, the current industry solution to this problem involves multiple insurance companies sending their respective modeling engineers. These engineers, carrying copies of the data, gather at a pre-agreed location to collaboratively complete the modeling of multiple univariate linear regression models. After the modeling is finished, all data is publicly deleted. This collaborative modeling model is highly complex; furthermore, several additional security management issues need to be addressed during the modeling process. A typical problem is which unit has the capability to decrypt the data in an online collaborative environment. A second problem is on which device the specific calculations take place. A third problem is who provides the modeling code. Consequently, the modeling training cycle is relatively long and the cost is high. Summary of the Invention
[0005] This invention aims to address the aforementioned shortcomings of existing technologies by providing a method for training multi-node dense-state regression models. This invention allows various organizations to confidently aggregate data without requiring personnel to gather, while simultaneously training better-performing models to enhance business operations.
[0006] This invention is implemented as follows: a multi-node dense-state regression model training method, comprising:
[0007] S1. Each participating party first specifies a keyid, and each party uses this key to encrypt its own data;
[0008] S2. All parties invoke the secret-state modeling smart contract, which aggregates the secret-state data from all parties to complete the training of the regression model;
[0009] S3. The participants call the decryption algorithm to decrypt the intercept and slope;
[0010] S4. Locally validate model quality;
[0011] S5. Audit, check if any participants have maliciously used the key in the system;
[0012] The storage method of the specified key ID is designed in accordance with the "GM / T0028-2014 Security Technical Requirements for Cryptographic Modules". For homomorphic encryption keys, they need to be used as session keys for access control and management. When generating the key, a homomorphic encryption key generation algorithm is used.
[0013] The multi-node dense-state regression model training method uses a smart contract to implement a homomorphic encryption algorithm. The encrypted data input by the user will be encrypted by the smart contract, and the encryption process is open and transparent. The blockchain system will generate an encryption record that represents the user's account, the key number, and uses this number to confirm which key was used in the encryption process and the time when the encryption algorithm was called. This encryption record is used to audit the user's use of the key afterward.
[0014] The multi-node dense-state regression model training method described above uses smart contracts for encryption, which can cause data to leave the data provider's node, leading to data leakage. Therefore, a secondary encryption technique is used to protect the data. The process of the secondary encryption technique is as follows:
[0015] Level 1 encryption: The user generates a random number r of the same length as the message m locally, and calculates padding = m + r;
[0016] Second-level encryption: The user calls the homomorphic encryption smart contract to encrypt the padding, obtaining e = (m + r); the smart contract returns e to the user; the user calls the homomorphic subtraction operation to obtain c = HeSub(e, r), which is the ciphertext that directly encrypts the message m.
[0017] The multi-node dense-state regression model training method, wherein the dense-state smart contract modeling process is written by the user using a smart contract language, and executed by the blockchain through code based on a homomorphic encryption-based regression model training scheme.
[0018] The multi-node dense-state regression model training method, wherein the audit is to audit the logs, allows all participants in the system to check the execution status of the smart contract and the execution status of the decryption algorithm at any time, preventing malicious participants in the system from decrypting the original data.
[0019] The multi-node dense-state regression model training method, wherein the training of the regression model includes a dense-state regression model training method based on the least squares method, with inputs: (x1, y1), (x2, y2)..., (x... i y i ), ...(x n y n ); where x and y are both ciphertexts obtained through homomorphic encryption;
[0020] Output: Model α0 and α1 are ciphertext.
[0021] The multi-node dense-state regression model training method, wherein obtaining the ciphertexts α0 and α1 includes:
[0022] Calculation of dense state mean:
[0023] The operator HeAvg(x1,...,x) is calculated using the average value of homomorphic encryption. n The means of the independent and dependent variables were obtained separately. and
[0024]
[0025] Calculation of dense-state slope and intercept:
[0026] Slope: The slope α1 is calculated using the addition (HeAdd), subtraction (HeSub), multiplication (HeMul), and division (HeDiv) operations from homomorphic encryption.
[0027] (1) Repeat the process n times, i = 1, ..., n.
[0028]
[0029] (2) Let nSum = p1, and execute the loop for n-1 rounds, i = 2, ..., n.
[0030] nSum = HeAdd(nSum, p) i )
[0031] (3) Repeat the process n times, i = 1, ..., n
[0032]
[0033] (4) Let dSum = dl The loop executes n-1 rounds, where i = 2, ..., n.
[0034] dSum = HeAdd(dSum, d i )
[0035] (5) α1 = HeDiv(nSum, dSum)
[0036] Intercept: The intercept α0 can be calculated using the subtraction HeSub and multiplication HeMul methods in homomorphic encryption.
[0037]
[0038] This invention allows each participating organization to develop a training contract using homomorphic encryption's basic computational functions after confirming modeling requirements, and then deploy the contract to the blockchain. In addition to the training contract, the blockchain also includes key generation, encryption, and decryption contracts, providing encryption and decryption capabilities for the training process. By combining homomorphic encryption with blockchain technology, this invention enables the training of multi-node encrypted regression models. Participants can remotely model and train without needing to gather in person, resulting in low costs. Furthermore, it prevents malicious decryption of original data, providing evidence for subsequent accountability and significantly improving the reliability and security of data transmission and computation. Detailed Implementation
[0039] This invention relates to a method for testing the independence of dense-state data and a method for training multi-node dense-state regression models.
[0040] S1. Each participating party first designates a key, and each party uses this key to encrypt its own data;
[0041] S2. All parties invoke the secret-state modeling smart contract, which aggregates the secret-state data from all parties to complete the training of the regression model;
[0042] S3. The participants call the decryption algorithm to decrypt the intercept and slope;
[0043] S4. Locally validate model quality;
[0044] S5. Audit, check if any participants have maliciously used the key in the system;
[0045] The storage method of the specified key is designed in accordance with the "GM / T0028-2014 Security Technical Requirements for Cryptographic Modules". For homomorphic encryption keys, they need to be used as session keys for access control and management. When generating the key, a homomorphic encryption key generation algorithm is used.
[0046] Training scheme for regression models based on homomorphic encryption
[0047] Regarding the training method of univariate linear regression models, this invention designs a dense-state regression model training method based on the least squares method, and the training process is as follows.
[0048] Input: (x1, y1), (x2, y2)..., (x i y i ), ...(x n y n ). Here, x and y are both ciphertexts obtained through homomorphic encryption.
[0049] Output: Model α0 and α1 are ciphertext.
[0050] process:
[0051] Calculation of dense state mean:
[0052] The operator HeAvg(x1, ..., x) is calculated using the average value of homomorphic encryption. n The means of the independent and dependent variables were obtained separately. and
[0053]
[0054] Calculation of dense state slope and intercept
[0055] Slope: The slope α1 is calculated using the addition (HeAdd), subtraction (HeSub), multiplication (HeMul), and division (HeDiv) operations from homomorphic encryption.
[0056] (1) Repeat the process n times, i = 1, ..., n.
[0057]
[0058] (2) Let nSum = p1, and execute the loop for n-1 rounds, i = 2, ..., n.
[0059] nSum = HeAdd(nSum, p) i )
[0060] (3) Repeat the process n times, i = 1, ..., n
[0061]
[0062] (4) Let dSum = d1, and repeat the process for n-1 rounds, i = 2, ..., n.
[0063] dSum = HeAdd(dSum, d i )
[0064] (5) α1 = HeDiv(nSum, dSum)
[0065] Intercept: The intercept α0 can be calculated using the subtraction HeSub and multiplication HeMul methods in homomorphic encryption.
[0066]
[0067] Multi-node regression model training system combining homomorphic encryption and blockchain
[0068] The system possesses five core functions: key management, homomorphic encryption, homomorphic decryption, smart contract modeling, and log auditing. The execution flow of each function is as follows:
[0069] Key Management: The requirements for the key management module are designed in accordance with the "GM / T0028-2014 Security Technical Requirements for Cryptographic Modules". For homomorphic encryption keys, they need to be used as session keys for access control and management. When generating keys, the homomorphic encryption key generation algorithm HeKeyGen() is used.
[0070] Homomorphic encryption: This method uses smart contracts to implement homomorphic encryption algorithms. User-inputted data is encrypted by the smart contract, and the encryption process is transparent. The blockchain system generates an encrypted record (userid, keyid, time), where userid represents the user's account, keyid represents the key number (which identifies the key used in the encryption process), and time represents the time the encryption algorithm was invoked. This record allows for post-event auditing of the user's key usage. When using smart contracts for encryption, data may leave the data provider's node, potentially leading to data leakage. To address this issue, secondary encryption technology is needed to protect the data. The secondary encryption process is as follows:
[0071] Level 1 encryption: The user generates a random number r of the same length as the message m locally, and calculates padding = m + r.
[0072] Second-level encryption: The user invokes a homomorphic encryption smart contract to encrypt the padding, obtaining e = (m + r). The smart contract returns e to the user. The user then performs a homomorphic subtraction operation to obtain c = HeSub(e, r), which is the ciphertext directly encrypted with respect to message m.
[0073] By using two-level encryption, smart contracts can securely encrypt data and ensure homomorphism.
[0074] Homomorphic decryption: After the model has been trained, a homomorphic encryption algorithm is called to decrypt α1 and α2, obtaining the coefficients and intercepts a and b of the plaintext state. During the decryption process, the execution of the smart contract is open and transparent, and all nodes in the blockchain can check whether the decryption algorithm has been legally invoked. Simultaneously, the system generates a record (userid, keyid, ct, m), where userid represents the user's account, keyid represents the key number, ct represents the ciphertext being decrypted, and m represents the decryption result. Through this record, all parties in the system can check whether the decryption algorithm has only been used to decrypt the ciphertext containing the intercepts and slopes, and not to decrypt the original data.
[0075] Smart contract modeling is a process in which users write code in a smart contract language (e.g., Solidity) to implement a regression model training scheme based on homomorphic encryption, which is then executed by the blockchain.
[0076] Log auditing: All participants in the system can check the execution status of smart contracts and decryption algorithms at any time to prevent malicious participants from decrypting the original data. If a malicious participant decrypts the original data, the system will provide a clear record of the amount, content, and time of the decrypted data, allowing other participants to protect their legitimate rights through legal means.
Claims
1. A training method for a multi-node dense-state regression model, characterized in that, S1. Each participating party first specifies a keyid, and each party uses this key to encrypt its own data; S2. All parties invoke the secret-state modeling smart contract, which aggregates the secret-state data from all parties to complete the training of the regression model; S3. The participants call the decryption algorithm to decrypt the intercept and slope; S4. Locally validate model quality; S5. Audit, check if any participants have maliciously used the key in the system; The specified key ID storage method is designed in accordance with the "GM / T0028-2014 Security Technical Requirements for Cryptographic Modules". For homomorphic encryption keys, they need to be used as session keys for access control and management. When generating the key, a homomorphic encryption key generation algorithm is used.
2. The multi-node dense-state regression model training method according to claim 1, characterized in that, The smart contract is invoked to implement the homomorphic encryption algorithm. The encrypted data input by the user will be encrypted by the smart contract, and the encryption process is open and transparent. The blockchain system will generate an encryption record, which represents the user's account, the key number, and the key used in the encryption process. The encryption record is used to audit the user's use of the key afterward.
3. The multi-node dense-state regression model training method according to claim 1, characterized in that, When using smart contracts for encryption, data leaves the data provider's node, potentially leading to data leakage. Double encryption is used to protect this data; the process of double encryption is as follows: Level 1 encryption: The user generates a random number r of the same length as the message m locally, and calculates padding = m + r; Second-level encryption: The user calls the homomorphic encryption smart contract to encrypt the padding, obtaining e = (m + r); the smart contract returns e to the user; the user calls the homomorphic subtraction operation to obtain c = HeSub(e, r), which is the ciphertext that directly encrypts the message m.
4. The multi-node dense-state regression model training method according to claim 1, characterized in that, The aforementioned encrypted smart contract modeling process is written by the user using a smart contract language, and executed by the blockchain through code based on a regression model training scheme using homomorphic encryption.
5. The multi-node dense-state regression model training method according to claim 1, characterized in that, The audit refers to the auditing of logs. All participants in the system can check the execution status of smart contracts and decryption algorithms at any time to prevent malicious participants in the system from decrypting the original data.
6. The multi-node dense-state regression model training method according to claim 1, characterized in that, The training of the regression model includes a dense-state regression model training method based on the least squares method, with inputs: (x1, y1), (x2, y2)..., (x... i y i ), ...(x n y n ); where x and y are both ciphertexts obtained through homomorphic encryption; Output: Model α0 and α1 are ciphertext.
7. The multi-node dense-state regression model training method according to claim 6, characterized in that... The acquisition of the ciphertexts α0 and α1 includes: Calculation of dense state mean: The operator HeAvg(x1, ..., x) is calculated using the average value of homomorphic encryption. n The means of the independent and dependent variables were obtained separately. and Calculation of dense-state slope and intercept: Slope: The slope α1 is calculated using the addition (HeAdd), subtraction (HeSub), multiplication (HeMul), and division (HeDiv) operations from homomorphic encryption. (1) Repeat the process n times, i = 1, ..., n. (2) Let nSum = p1, and execute the loop for n-1 rounds, i = 2, ..., n. nSum=HeAdd(nSum,p i ) (3) Repeat the process n times, i = 1, ..., n (4) Let dSum = d1, and repeat the process for n-1 rounds, i = 2, ..., n. dSum=HeAdd(dSum,d i ) (5) α1 = HeDiv(nSum, dSum) Intercept: The intercept α0 can be calculated using the subtraction HeSub and multiplication HeMul methods in homomorphic encryption.