Cross-domain insurance data collaborative computing data processing method and system based on federated learning

Abstract

The application provides a cross-domain insurance data collaborative computing data processing method and system based on federated learning, and relates to the technical field of insurance data processing. The method comprises the following steps: performing privacy intersection processing on locally stored insurance-related data by each participant, completing the encrypted alignment of cross-domain common samples without leaking the original plaintext, and obtaining a ciphertext alignment dataset; extracting local features based on the ciphertext alignment dataset, mapping the local features to a high-dimensional feature circle manifold, and constructing a feature distribution circle with the mean of the feature distribution as the center and the variance of the feature distribution as the radius; determining the domain offset existing in the local features, mapping the domain offset to a feature drift straight line; solving the tangent equation of the feature drift straight line and the feature distribution circle to obtain the coordinates of the tangent point and the tangent direction vector. The application can realize high-robustness and high-generalization cross-domain insurance collaborative modeling.

Description

Technical Field

[0001] This invention relates to the field of insurance data processing technology, and in particular to a cross-domain insurance data collaborative computing data processing method and system based on federated learning. Background Technology

[0002] With the increasing digital transformation and data security compliance requirements in the insurance industry, cross-institutional and cross-regional collaborative modeling of insurance data has become a necessity for scenarios such as auto insurance pricing, risk assessment, and claims prediction. Federated learning, with its characteristics of data remaining local and controllable privacy and security, has become the mainstream technical solution for cross-domain collaborative computing of insurance data. It can achieve encrypted alignment of shared samples through privacy intersection, and then complete the global model construction through a process of local training, encrypted gradient uploading, and coordination node aggregation.

[0003] However, most existing federated learning frameworks suffer from the following technical shortcomings: cross-domain insurance data exhibits a certain degree of domain shift and feature drift, making it impossible to achieve accurate and dynamic distribution alignment in the dense feature space. Influenced by regional underwriting rules, customer group structure, vehicle type distribution, and differences in driving environment, the local feature distributions of different insurance institutions naturally shift, leading to a series of problems: due to the inconsistency between local features and the global baseline distribution, local model training overfits to intra-domain features, making it difficult for the global model to converge and significantly reducing generalization ability; moreover, gradient aggregation is subject to gradient direction conflicts caused by inter-domain feature shifts, and traditional weighted aggregation cannot adaptively correct for the shift contribution, reducing the accuracy and stability of risk prediction and pricing models.

[0004] The aforementioned shortcomings are particularly prominent in typical scenarios such as joint pricing of auto insurance and cross-regional claims risk control. Even if sample privacy alignment is achieved, it is difficult to build a robust cross-domain collaborative model due to the misalignment of feature spaces, which restricts the effectiveness of federated learning in the insurance field. Summary of the Invention

[0005] This invention provides a data processing method and system for cross-domain insurance data collaborative computation based on federated learning, which can achieve highly robust and highly generalizable cross-domain insurance collaborative modeling.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] Firstly, a cross-domain insurance data collaborative computation data processing method based on federated learning, the method comprising:

[0008] By performing privacy-preserving intersection processing on the locally stored insurance-related data by each participating party, the encrypted alignment of cross-domain shared samples is completed without disclosing the original plaintext, resulting in a encrypted aligned dataset.

[0009] Based on the dense-state aligned dataset, local features are extracted locally and mapped onto a high-dimensional feature circle manifold. A feature distribution circle is constructed with the mean of the feature distribution as the center and the variance of the feature distribution as the radius. The domain offsets existing in the circle are determined according to the local features and mapped to feature drift lines.

[0010] Solve for the equations of the tangents between the characteristic drift line and the characteristic distribution circle to obtain the coordinates of the tangent point and the tangent direction vector;

[0011] The feature correction offset is calculated based on the coordinates of the tangent point and the tangent direction vector, which is used to dynamically calibrate local features and generate refined feature embeddings after distribution alignment.

[0012] Local model training is initiated using refined feature embeddings with distribution alignment, loss is calculated after forward propagation, dense gradient parameters are generated through backpropagation, and the dense gradient parameters are encrypted and uploaded to the coordinating node.

[0013] The system receives encrypted gradient parameters uploaded by each participant, and after decryption or recombination, it performs adaptive weighting and direction correction on the contribution of each gradient according to the preset inter-domain difference metric to generate the corrected aggregate gradient.

[0014] The coordinating node updates the global model parameters based on the corrected aggregated gradient, and distributes the updated global model parameters back to each participant in encrypted form so that each participant can update its local model until a cross-domain collaborative computing model resistant to domain offset is obtained.

[0015] Secondly, a cross-domain insurance data collaborative computing and processing system based on federated learning includes:

[0016] The encrypted sample alignment module is used by each participant to perform privacy intersection processing on the insurance-related data stored locally. It completes the encrypted alignment of cross-domain shared samples without revealing the original plaintext, and obtains the encrypted aligned dataset.

[0017] The local feature modeling and domain offset detection module is used to extract local features based on the dense state aligned dataset, map the local features to a high-dimensional feature circle manifold, construct a feature distribution circle with the feature distribution mean as the center and the feature distribution variance as the radius; determine the domain offset existing in the circle based on the local features, and map the domain offset to a feature drift line;

[0018] The high-dimensional feature geometry correction module is used to solve the equation of the tangent line between the feature drift line and the feature distribution circle to obtain the coordinates of the tangent point and the tangent direction vector; based on the coordinates of the tangent point and the tangent direction vector, the feature correction offset is calculated to dynamically calibrate local features and generate refined feature embeddings after distribution alignment.

[0019] The model training and dense gradient upload module is used to start local model training with refined feature embedding after distribution alignment, calculate the loss after forward propagation, generate dense gradient parameters through back propagation, and upload the encrypted dense gradient parameters to the coordinating node.

[0020] The gradient adaptive aggregation correction module is used to receive the encrypted gradient parameters uploaded by each participant, and after decryption or recombination, adaptively weight and correct the direction of each gradient contribution according to the preset inter-domain difference metric to generate the corrected aggregate gradient.

[0021] The global model collaborative update module is used to coordinate nodes to update global model parameters based on the corrected aggregated gradient. By distributing the updated global model parameters in encrypted form back to each participant, each participant can update its local model until a cross-domain collaborative computing model resistant to domain offset is obtained.

[0022] The above-described solution of the present invention has at least the following beneficial effects:

[0023] Because it employs privacy-preserving intersection and homomorphic encryption to align dense-state samples, it overcomes the security issue of plaintext leakage of cross-domain insurance data, thus achieving accurate alignment of shared samples under compliance. Because it uses a high-dimensional feature circle manifold to construct a feature distribution circle and maps domain offset to feature drift lines, it overcomes the difficulty in quantifying and representing domain offset, thus achieving intuitive modeling of local feature distribution. Because it uses solving the tangent equation and calculating the feature correction offset to dynamically calibrate local features, it overcomes the problems of feature drift and distribution misalignment in dense-state space, thus generating stable and unified refined feature embeddings. Because it uses inter-domain difference metrics to adaptively weight and correct the gradient direction, it overcomes the problems of gradient direction conflict and low accuracy of traditional aggregation, thus improving the convergence speed and generalization ability of the global model. Because it uses global model dense-state updates and local iterative training, it overcomes the problem of poor robustness in cross-domain collaborative modeling, thus obtaining a domain offset-resistant and high-precision cross-domain insurance collaborative computation model. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a cross-domain insurance data collaborative computing data processing method based on federated learning, provided by an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of a cross-domain insurance data collaborative computing data processing system based on federated learning, provided by an embodiment of the present invention. Detailed Implementation

[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0027] like Figure 1 As shown, an embodiment of the present invention proposes a cross-domain insurance data collaborative computing data processing method based on federated learning, the method comprising the following steps:

[0028] Step 1: Each participant performs privacy intersection processing on the locally stored insurance-related data to complete the encrypted alignment of cross-domain shared samples without disclosing the original plaintext, thus obtaining the encrypted aligned dataset.

[0029] Step 2: Extract local features based on the dense-state aligned dataset, map the local features onto a high-dimensional feature circle manifold, and construct a feature distribution circle with the feature distribution mean as the center and the feature distribution variance as the radius; determine the domain offset existing in the circle based on the local features, and map the domain offset to a feature drift line;

[0030] Step 3: Solve for the equations of the tangents between the feature drift line and the feature distribution circle to obtain the coordinates of the tangent point and the tangent direction vector;

[0031] Step 4: Calculate the feature correction offset based on the tangent point coordinates and tangent direction vector. This offset is used to dynamically calibrate local features and generate refined feature embeddings after distribution alignment.

[0032] Step 5: Start local model training with refined feature embedding after distribution alignment, calculate loss after forward propagation, generate dense gradient parameters through backpropagation, and upload the encrypted dense gradient parameters to the coordinating node.

[0033] Step 6: Receive the encrypted gradient parameters uploaded by each participant. After decryption or recombination, adaptively weight and correct the direction of each gradient contribution according to the preset inter-domain difference metric to generate the corrected aggregate gradient.

[0034] Step 7: The coordinating node updates the global model parameters based on the corrected aggregated gradient. The updated global model parameters are then distributed back to each participant in encrypted form so that each participant can update its local model until a cross-domain collaborative computing model resistant to domain offset is obtained.

[0035] In this embodiment of the invention, precise encrypted alignment of cross-domain shared samples can be achieved without disclosing the original plaintext of insurance data, ensuring data privacy and compliance; it can effectively quantify the domain offset and feature drift of cross-domain insurance data, and achieve feature distribution alignment in the dense space through dynamic feature correction, avoiding overfitting of the local model; it can alleviate directional conflicts in the gradient aggregation process, improve the convergence speed and cross-domain generalization ability of the global model, and finally construct a cross-domain insurance collaborative computing model that is resistant to domain offset and has high precision, adapting to practical application scenarios such as auto insurance pricing and claims risk control, and promoting the efficient implementation of federated learning in the insurance field.

[0036] In a preferred embodiment of the present invention, step 1 above may include:

[0037] Step 1.1: Each participating party performs standardized preprocessing on the sample identifiers in the locally stored insurance-related data to generate a unified set of inter-participation sample identifiers. Specifically, each participating party is a cross-domain data holder, such as an insurance institution or insurance data service provider in a different region. Each participating party independently holds local insurance-related data and is only responsible for processing its own local data, without disclosing any original data to other participating parties. Each participating party first extracts the sample identifiers from the locally stored insurance-related data. The sample identifiers are information used to uniquely identify insurance samples, and may include policy numbers, insured de-identification identifiers, vehicle identification codes, etc. Then, these sample identifiers are standardized preprocessed. Abnormal identifiers with abnormal formats, missing characters, and duplicate redundancies are first removed, and then processed according to a preset unified format specification. The preset unified format specification specifically includes unified character encoding as UTF-8, unified identifier length as a fixed number of digits, unified character case as lowercase, and unified removal of spaces and special symbols to ensure that the sample identifiers of all participating parties are completely consistent in format, ultimately generating a unified set of inter-participation sample identifiers without abnormalities or redundancy.

[0038] Step 1.2: Encrypt each sample identifier in the set of sample identifiers to be intersected using a preset homomorphic encryption method to generate a corresponding encrypted sample identifier. Combine all encrypted sample identifiers to form a local encrypted identifier dataset. Specifically, this includes: using a preset homomorphic encryption method, which specifically selects a partial homomorphic encryption method that supports encrypted comparison and does not disclose the original data. The implementation process and principle are as follows: Each participant generates its own public key and private key locally. The public key is only used for this encryption operation, and the private key is properly kept by the participant locally and not disclosed to external parties. During encryption, each sample identifier in the set of sample identifiers to be intersected is first converted into an encryptable string format. Then, the preset homomorphic encryption method is called, and the locally generated public key is used to encrypt each sample identifier item by item. During the encryption process, only the string of the sample identifier is irreversibly converted into encrypted form without changing the correspondence between the identifiers. The entire process is completed on the participant's local terminal, and no original sample identifiers or intermediate encryption process data are transmitted to other participants or coordinating nodes.

[0039] After encryption, a secret sample identifier is generated that corresponds one-to-one with each original sample identifier. The secret sample identifier is only a string of characters without definite meaning, and the original sample identifier information cannot be obtained by reverse deduction. After encryption, each participant organizes and combines all secret sample identifiers one by one according to the original order of the sample identifiers in the set of sample identifiers to be intersected, forming a local secret identifier dataset that contains only secret identifier information, does not involve any original plaintext data, and can be used for subsequent secret comparison.

[0040] Step 1.3 involves each participating party uploading its local encrypted identifier dataset to the coordination node. The coordination node then performs an encrypted comparison operation on all received encrypted identifier datasets to filter out the encrypted sample identifiers common to all participating parties, forming an encrypted intersection identifier set. Specifically, each participating party uploads its locally generated encrypted identifier dataset to the coordination node via an encrypted secure channel. The coordination node is the core scheduling and aggregation node of the entire cross-domain insurance data collaborative computing data processing system. It is independent of each participating party and has strong security protection capabilities, with built-in data encryption transmission module, encrypted comparison module, secure storage module, and access control module.

[0041] The functions and relationships of each module are as follows: The access control module is the basic protection module, responsible for verifying the legitimate identities of each participant, allowing only authorized participants to access the node and interact with data, preventing unauthorized node intrusion, and providing a security prerequisite for all subsequent operations; the data encryption transmission module works in conjunction with the access control module, connecting to an encrypted security channel when participants upload data, performing double encryption protection on the transmitted encrypted identifier dataset, and simultaneously verifying the integrity and legitimacy of data transmission in real time, ensuring that the data is not stolen or tampered with during transmission, and transmitting the verified complete encrypted identifier dataset to the encrypted comparison module; As a core functional module, the state comparison module receives the dataset of encrypted state identifiers of all participants transmitted by the data encryption transmission module. Without performing any decryption operations, it performs item-by-item comparison, matching, and deduplication operations to filter out the encrypted state sample identifiers common to all participants. After completing the comparison, the results are transmitted to the secure storage module. The secure storage module is responsible for temporarily encrypting and storing the set of encrypted state intersection identifiers output by the encrypted state comparison module. It also stores the legitimate access information of each participant (without containing any original plaintext data) and works in conjunction with the access control module to allow only authorized modules within the coordination node to access the stored data, preventing data leakage.

[0042] Each module works in concert, with the coordination node specifically responsible for receiving encrypted data uploaded by each participant, performing encrypted comparison operations, and distributing relevant encrypted results. It does not participate in the decryption of any original data, nor does it store any participant's original plaintext data. During the upload process, the encrypted security channel employs end-to-end encryption, providing double encryption protection for the data transmission link and real-time verification of data transmission integrity to prevent theft, tampering, or leakage of encrypted identifier data during transmission. After receiving the encrypted identifier datasets uploaded by all participants, the coordination node first performs integrity verification through the data encryption transmission module. After confirming no missing or tampered entries, the encrypted comparison module compares, matches, and removes duplicates from each encrypted identifier dataset, verifying the existence of each encrypted sample identifier in all participants' datasets. It then filters out encrypted sample identifiers that are present in all participants' datasets and, according to the set unified sorting rules (consistent with the sorting rules for sample identifiers in step 1.1), organizes these shared encrypted identifiers to form a complete, ordered set of encrypted intersection identifiers containing only shared encrypted identifiers. After organization, the secure storage module temporarily encrypts and stores the collected encrypted identifiers.

[0043] Step 1.4: The coordinating node distributes the set of encrypted intersection identifiers to each participant. Each participant, based on the received set of encrypted intersection identifiers, decrypts it locally and maps it back to the original sample identifier. This is used to extract the complete data record corresponding to the original sample identifier from the local insurance-related data, ultimately obtaining an encrypted alignment dataset containing only shared samples. Specifically, the coordinating node synchronously distributes the generated set of encrypted intersection identifiers to each participant through an encrypted secure channel. End-to-end encryption protection is also used during the distribution process, and the integrity and consistency of the distributed data are verified in real time to ensure that each participant can receive the set of encrypted intersection identifiers completely and securely, and to prevent data leakage or tampering during the distribution process. After receiving the set of encrypted intersection identifiers, each participant verifies the integrity and legality of the received data to confirm that the data is complete. If the encryption key is lost, tamper-proof, and originates from a legitimate coordinating node, a dedicated decryption key is securely stored locally. This dedicated decryption key is a key generated locally by each participant during the homomorphic encryption operation in step 1.2. It corresponds one-to-one with the public key used during encryption and is held independently by the corresponding participant. It is securely stored locally using encrypted storage and is not disclosed or transmitted to any external node (including the coordinating node and other participants). It is specifically used to decrypt the encrypted state identifier data generated by the participant, ensuring the uniqueness and security of the decryption operation. The encrypted state intersection identifier set is decrypted on the local terminal. The entire decryption process is completed locally by the participant and no decrypted information is transmitted to the coordinating node, other participants, or any external node. The decryption key and intermediate decryption process data are also not disclosed.

[0044] After decryption, each encrypted identifier is mapped back to its corresponding original sample identifier. The restoration process strictly follows the encryption mapping relationship in step 1.2 to ensure that the restored original sample identifiers are accurate. Based on the restored original sample identifiers, complete data records corresponding to each original sample identifier are accurately extracted from the locally stored insurance-related data. During the extraction process, only valid data fields related to cross-domain collaborative computing are selected, and redundant and invalid data fields unrelated to modeling are removed to ensure that the extracted data is accurate and efficient. Finally, a encrypted aligned dataset is obtained that contains only samples shared by all participants, does not leak any original plaintext throughout the process, has a unified data format, and can be used for feature extraction.

[0045] In a preferred embodiment of the present invention, step 2 above may include:

[0046] Step 2.1: Based on the obtained dense-state aligned dataset, each participant extracts features from the original data fields of each sample record in the dataset, generating a multi-dimensional local feature vector composed of multiple feature dimensions. These multiple feature dimensions include at least one or more of the following: driving behavior feature dimension, vehicle operating status feature dimension, historical claims feature dimension, and road violation feature dimension. All local feature vectors are then aggregated to form a local feature set. Specifically, each participant, based on the obtained dense-state aligned dataset, first preprocesses each sample record in the dataset, removing invalid fields with missing data or abnormal fluctuations, and retaining those relevant to cross-domain collaboration. Calculate the relevant valid raw data fields, and then perform targeted feature extraction on the valid raw data fields of each sample record. The multiple feature dimensions can be specifically divided into driving behavior feature dimension, vehicle operating status feature dimension, historical claims feature dimension, and road violation feature dimension. The driving behavior feature dimension can extract features such as the number of times of rapid acceleration, the number of times of emergency braking, and the average driving speed. The vehicle operating status feature dimension can extract features such as vehicle usage time, frequency of malfunctions, and maintenance records. The historical claims feature dimension can extract features such as the number of claims, the amount of claims, and the reason for claims. The road violation feature dimension can extract features such as the number of violations, the type of violation, and the location of the violation.

[0047] After extraction, all feature dimension values ​​for each sample are standardized to ensure that the feature values ​​of different dimensions are on the same order of magnitude, avoiding the impact of dimensional value differences on subsequent modeling results. Then, according to the set fixed feature dimension order (such as driving behavior feature dimension, vehicle operating status feature dimension, historical claims feature dimension, and road violation feature dimension), the feature dimension values ​​of each sample are arranged one by one and integrated into an ordered numerical sequence, with each value corresponding to a feature dimension. This generates a multidimensional local feature vector composed of multiple feature dimensions, ensuring that each sample corresponds to a unique multidimensional local feature vector, and that the value at each position in the vector corresponds one-to-one with the feature dimension without confusion. Finally, the local feature vectors corresponding to all local samples are summarized one by one to form a local local feature set that can completely reflect the local common sample features.

[0048] Step 2.2 involves performing statistical analysis on the local feature set to calculate the average value of all local feature vectors in the set across each feature dimension, and combining these average values ​​to generate a mean vector. Simultaneously, the covariance matrix of the local feature set is calculated, and the variance values ​​corresponding to each feature dimension are extracted from the diagonal of the covariance matrix as the distribution variance of each dimension. Specifically, this includes: performing a comprehensive and detailed statistical analysis on the locally generated local feature set; first, traversing all local feature vectors in the local feature set to confirm that the total number of feature dimensions for each vector is exactly the same and the order of each dimension is also completely consistent, ensuring a unified benchmark for subsequent statistical calculations; then, according to the set fixed feature dimension order, performing independent numerical statistics and calculations for each feature dimension; first, for a single feature dimension, traversing the corresponding values ​​of all valid local feature vectors under that dimension, identifying and removing extreme outliers that significantly deviate from the normal distribution range of the dimensions' values ​​to avoid outliers interfering with the accuracy of the statistical results; then, accumulating the remaining valid values ​​one by one, recording intermediate values ​​during the accumulation process to ensure the accuracy of the accumulation result, and obtaining the total value of that feature dimension.

[0049] The process begins by counting the number of valid feature vectors participating in the accumulation under the stated dimension. Vectors corresponding to samples with no valid values ​​after outlier removal are removed to ensure the validity of the denominator. Then, the sum of the values ​​in that dimension is divided by the number of valid feature vectors participating in the statistics to calculate the average value under that feature dimension. During the calculation, a preset number of decimal places are retained to ensure the accuracy of the average value. Following the same process and standards, the average values ​​of all feature dimensions are calculated sequentially. After the calculation is completed, the average values ​​of all feature dimensions are combined one by one in a fixed order according to the set parameters to form a mean vector that can accurately represent the center of the overall distribution of local features. This mean vector can intuitively reflect the average level of local common samples in each feature dimension, providing data support for determining the center of the feature distribution circle.

[0050] Simultaneously, a complete covariance analysis was conducted on the local feature set. First, the order of all feature dimensions was confirmed. Then, each feature dimension was paired (including combinations within the same dimension). For each pair of feature dimensions, the corresponding values ​​of all valid local feature vectors under each dimension were extracted. The deviation of each sample's value in each dimension from the average value of the corresponding dimension was calculated. The two deviations were then multiplied and summed. The sum was divided by the number of valid samples participating in the statistics to obtain the covariance between the two dimensions. All the covariance results of the pairwise combinations were then arranged into a matrix according to the order of the feature dimensions, constructing a matrix that fully reflects the different features. The covariance matrix represents the degree of correlation and linear correlation between dimensions. The number of rows and columns in this covariance matrix is ​​consistent with the number of feature dimensions of the local feature vectors. The values ​​on the main diagonal of the matrix represent the covariance of each feature dimension with itself. The values ​​are extracted sequentially from the main diagonal of the covariance matrix. These values ​​correspond to the variance of each feature dimension and can directly characterize the dispersion of local feature values ​​under the corresponding dimension. The larger the variance, the more dispersed the feature distribution of that dimension is, and the smaller the variance, the more concentrated the feature is. These variance values ​​on the main diagonal are uniformly organized and retain a preset precision as the distribution variance corresponding to each feature dimension, which is used to determine the radius of the feature distribution circle along each axis.

[0051] Step 2.3: Based on the calculated mean vector and the distribution variance of each dimension, each participating party constructs a feature distribution circle that reflects the local feature distribution pattern on a preset high-dimensional feature circular manifold, using the point represented by the mean vector as the center and the distribution variance of each dimension as the radius along each axis. Specifically, this includes: first, determining that the high-dimensional feature circular manifold and feature distribution circle in this invention are symbolic names, but their essence is adaptation. n High-dimensional feature manifolds and insurance features n A high-dimensional feature distribution hypersphere is constructed step by step and precisely on a pre-defined high-dimensional feature circle manifold based on the calculated mean vector and the distribution variance of each dimension. The pre-defined high-dimensional feature circle manifold is a high-dimensional spatial geometric structure adapted to multiple feature dimensions of insurance (driving behavior, vehicle operating status, etc.). Its number of dimensions is completely consistent with the number of feature dimensions of local feature vectors. Each dimension corresponds to an insurance feature dimension, and each point in the space corresponds to a complete set of multi-dimensional local feature vectors. It can accurately carry the distribution modeling of multi-dimensional features without relying on additional tools. It can be constructed simply through coordinate mapping and geometric association.

[0052] The specific implementation process is as follows: The obtained mean vector is mapped according to a preset high-dimensional feature space mapping rule. This mapping rule is implemented as follows: the average value of each feature dimension is used as the coordinate value of the corresponding high-dimensional coordinate axis. The average value of each feature dimension in the mean vector corresponds one-to-one with a coordinate axis of the high-dimensional feature circular manifold. Let the number of feature dimensions of the local feature vector be... n ( n ≥1 corresponds to driving behavior, vehicle operating status, etc. in insurance data. n (each feature dimension), the calculated mean vector is ,in For the first i The average value of each feature dimension. Indicates the first i Feature values ​​in each feature dimension; the high-dimensional space coordinate mapping formula is specifically as follows: The mean values ​​of all feature dimensions are sequentially used as coordinate parameters for the corresponding coordinate axes. Substituting these values ​​into the mapping formula, a unique mapping point O on the high-dimensional feature circular manifold is calculated. This mapping point O is then determined as the center of the feature distribution circle. This center precisely corresponds to the overall distribution center of the local features, intuitively reflecting the average level of the local common samples across all feature dimensions, thus laying a core benchmark for constructing the feature distribution circle. For each feature dimension, the distribution variance corresponding to that dimension is processed separately. First, the distribution variance is standardized and calibrated to eliminate the influence of differences in variance values ​​across different feature dimensions. Let the first... i The variance of the distribution of each feature dimension is After calibration, the radius of that dimension is obtained. (The calibration process involves taking the square root of the variance to obtain the standard deviation, then normalizing it within the [0,1] interval to ensure the radius matches the coordinate scale.) The calibrated radius... As the radius of the feature dimension along the corresponding axis of the high-dimensional feature circular manifold, the size of the radius is positively correlated with the degree of dispersion of the feature dimension. The stronger the dispersion, the larger the corresponding radius, and the weaker the dispersion, the smaller the corresponding radius. This accurately represents the discrete range of local features under each feature dimension, ensuring that the distribution characteristics of each dimension can be fully reflected by the size of the radius.

[0053] Then, using the determined center O as a reference, the radii corresponding to each feature dimension are... Extending along the corresponding axes of the high-dimensional feature circular manifold, determine the endpoint positions of the radius in each dimension one by one. i The coordinates of the two endpoints of each dimension are respectively and Then, the endpoints are connected through high-dimensional spatial coordinate association. The specific implementation process is as follows: taking the coordinates of the center O as the reference, the high-dimensional coordinates of the two radius endpoints in each dimension are obtained, based on the high-dimensional hypersphere equation. The coordinates of all endpoints are substituted into the equation for verification. Following the principle of symmetry in the same dimension and linkage in different dimensions, the coordinates of all endpoints are correlated and calculated to ensure that the endpoints of adjacent dimensions can form a closed surface with smooth connection. Then, the endpoints of all dimensions are connected in a fixed order to form a closed circular structure that conforms to the hypersphere equation. Finally, the closed circular structure is verified to confirm that the circular structure can completely cover the mapping positions of most effective local feature vectors in the local feature set, and that the correlation between the center and each radius and the connection logic of each endpoint are accurate. In this way, a feature distribution circle that can accurately reflect the distribution pattern of local features, fit the characteristics of insurance data, and intuitively show the distribution differences of each feature dimension can be constructed. This feature distribution circle can clearly present the distribution range and concentration trend of local features, providing a solid foundation for domain offset detection, feature drift line construction, and feature correction.

[0054] In a preferred embodiment of the present invention, step 2 above may include:

[0055] Step 2.4: Obtain the global baseline feature distribution parameters pre-distributed by the coordinating node. The global baseline feature distribution parameters include the global baseline mean vector and the variance of each dimension of the global baseline. Specifically, this involves actively obtaining the global baseline feature distribution parameters pre-distributed by the coordinating node through an encrypted secure channel. The global baseline feature distribution parameters are pre-calculated by the coordinating node before the cross-domain collaborative computing starts, based on the encrypted alignment datasets uploaded by all participants in the past, industry-standard insurance feature benchmark data, and combined with the actual needs of cross-domain insurance collaborative modeling (such as auto insurance risk assessment and pricing). After comprehensive statistical analysis, outlier removal, and standardization calibration, they are specifically used as a unified benchmark standard for aligning the local feature distribution of each participant with the global feature distribution, ensuring that the domain offset detection of all participants has a unified reference basis.

[0056] Specifically, the global benchmark feature distribution parameters include the global benchmark mean vector and the global benchmark variance of each dimension. The global benchmark mean vector corresponds to the overall average level of the shared samples of all participants in each feature dimension (driving behavior, vehicle operating status, historical claims, and traffic violations). The global benchmark variance of each dimension corresponds to the overall dispersion of the shared samples of all participants in each feature dimension. The number and order of the feature dimensions of the global benchmark feature distribution parameters are completely consistent with the number and order of the dimensions of the local feature vectors of each participant to avoid misalignment in subsequent comparisons. After receiving the parameters, each participant performs integrity verification to confirm that there are no missing dimensions or numerical errors in the parameters. Then, it performs legality verification by verifying the digital signature of the coordination node to confirm that the parameters come from a legitimate coordination node and have not been tampered with. After the verification is passed, the global benchmark feature distribution parameters are encrypted and stored locally, separately from the local mean vector and distribution variance, specifically for the calculation and comparison of the domain offset later.

[0057] Step 2.5: Compare the calculated mean vector with the obtained global benchmark mean vector to calculate the difference between the two. Use this difference as an offset vector, where the offset vector includes the offset direction and offset distance in each feature dimension. Specifically, this includes: calling the locally encrypted and stored, calculated local mean vector, and then retrieving the obtained and verified global benchmark mean vector. Perform dimension consistency verification on the two mean vectors to confirm that the number of feature dimensions and the dimension arrangement order are completely consistent, avoiding dimension misalignment that could lead to distorted comparison results. After the verification is passed, compare the two mean vectors one by one for each feature dimension in the same feature dimension order. For each feature dimension, first extract the average value of the current dimension in the local mean vector and the benchmark average value of the same index dimension in the global benchmark mean vector, and then use the local dimension average value as the basis for comparison. The difference is calculated by subtracting the global baseline dimension average. During the process, the values ​​are preserved in terms of accuracy and abnormal fluctuations are filtered out to ensure that the difference results for each dimension are stable, reliable, and accurately reflect the degree of deviation. After the calculation is completed, the difference values ​​corresponding to all feature dimensions are integrated one by one in strict accordance with the preset feature dimension order to form an ordered difference sequence, which is the offset vector. This offset vector can accurately reflect the deviation between the local feature distribution and the global baseline feature distribution. The offset vector includes the offset direction and offset distance of each feature dimension. A positive difference value indicates that the local average value of that feature dimension is higher than the global baseline average value, and the offset direction is positive. A negative difference value indicates that the local average value of that feature dimension is lower than the global baseline average value, and the offset direction is negative. The absolute value of the difference is the offset distance of that feature dimension. The larger the absolute value, the more severe the domain offset of that dimension.

[0058] Step 2.6: Based on the obtained offset vector, starting from the center of the constructed feature distribution circle, construct a straight line on the high-dimensional feature circular manifold along the offset direction indicated by the offset vector. This straight line is determined as the feature drift line. Specifically, this includes: first, determining the center of the constructed feature distribution circle, which is the unique mapping point of the local mean vector on the high-dimensional feature circular manifold, accurately representing the overall distribution center of the local features; confirming that the high-dimensional coordinate parameters of the center are accurate and completely consistent with the coordinates obtained in Step 2.3; then, extracting the obtained offset vector, confirming the offset direction, offset distance, and offset component ratio of each feature dimension, integrating the offset information of all dimensions, confirming the overall offset direction indicated by the offset vector, ensuring that the overall direction is consistent with the offset directions of each dimension, and that there is no conflict in the dimensional directions; and recording the multi-dimensional direction parameters of the offset vector for accurate determination of the subsequent straight line extension direction.

[0059] Starting from the center of the feature distribution circle, and combining the spatial coordinate rules of the high-dimensional feature circular manifold, the axial distribution characteristics of each feature dimension, and the multidimensional directional information of the offset vector, a straight line is constructed step by step on the high-dimensional feature circular manifold strictly according to the overall offset direction indicated by the offset vector. During construction, the high-dimensional coordinates of the circle center are first fixed as the starting coordinates of the straight line, and coordinate markings are made. Then, calculations are performed based on the offset components of the offset vector in each feature dimension. For each feature dimension, the offset component of that dimension is used as the vertical increment of the straight line on the corresponding high-dimensional coordinate axis, and a set unit length is used as the horizontal increment. The slope of the straight line on that coordinate axis is calculated by the ratio of the vertical increment to the horizontal increment. Finally, the slopes of all dimensions are integrated, combined with the direction of the high-dimensional space. The synthesis rule (which refers to synthesizing multiple single-dimensional offset directions into a unified spatial direction based on the offset components corresponding to each feature dimension and according to the relative proportion and positive / negative direction of each offset component in the high-dimensional space. This rule can integrate the offset contributions of all feature dimensions to form a unique and stable overall offset direction, ensuring that the extension trend of the line in the high-dimensional space is consistent with the offset of all dimensions) determines the extension angle of the line on each high-dimensional coordinate axis one by one. When calculating the angle, the positive direction of the coordinate axis is used as the reference. If the offset component is positive, the angle is the positive angle; if the offset component is negative, the angle is the negative angle. At the same time, the extension trend of the line is determined according to the absolute value of the slope of each dimension. The larger the absolute value of the slope of the dimension, the higher the weight of the influence on the overall extension direction of the line.

[0060] After calculation, the extension angle of each dimension is compared with the offset direction of the corresponding dimension of the offset vector through dimensional calibration. If the angle deviation exceeds the preset threshold, the slope of that dimension is adjusted until the angle is perfectly matched, ensuring that the spatial direction of the line is completely consistent with the multidimensional synthesis direction of the offset vector, avoiding deviation of the overall direction due to a single dimension deviation. At the same time, the line extends synchronously in both positive and negative directions with the center of the circle as the midpoint. The extension length is reasonably set in combination with the offset distance of each feature dimension. One side points to the actual direction in which the local feature drifts relative to the global benchmark, covering the actual range of the current drift, and the other side points to the compensation direction opposite to the drift, reserving the reverse adjustment path required for feature correction, ensuring that the line can completely cover the offset trend and correction requirements of the local feature distribution.

[0061] After construction, the straight line is fully verified. First, it is verified whether the straight line passes precisely through the center of the feature distribution circle, and the starting coordinates of the straight line are completely consistent with the center coordinates. Then, it is verified whether the direction of the straight line and the multidimensional composite direction of the offset vector are completely matched without any deviation. Finally, it is verified whether the extension direction of the straight line conforms to the geometric structure of the high-dimensional feature circle manifold and is compatible with the axial distribution of each feature dimension. After verification, the straight line that passes through the center of the circle, extends along the direction of the offset vector, and covers the two-way path of drift and compensation is finally determined as the feature drift line, providing an accurate and precise directional reference for solving the tangent equation, calculating the feature correction offset, and realizing the feature dynamic calibration.

[0062] In a preferred embodiment of the present invention, step 3 above may include:

[0063] Step 3.1: Based on the constructed feature distribution circle, establish the standard equation of this feature distribution circle on the high-dimensional feature circle manifold using its center coordinates and the distribution variance of each dimension. Specifically, this includes: performing equation establishment operations based on the already constructed feature distribution circle. To ensure the continuity and consistency of the high-dimensional spatial geometric description, the standard equation of the feature distribution circle established in this step (which also applies to the entire invention) is essentially a high-dimensional hypersphere equation. It is completely consistent with the high-dimensional hypersphere equation used in Step 2.3 to verify the closed structure of the feature distribution circle in terms of geometric meaning, parameter definition, and expression form. The only difference is the application scenario, which has been changed from construction verification to solution operation. The specific process is as follows: using the previously determined center coordinates of the feature distribution circle... and the radius of each feature dimension after standardization and calibration. In a high-dimensional circular manifold, any point in space at the th... i Dynamic coordinates on each feature dimension Coordinates of the center of the circle in that dimension Perform a difference operation, square the result, and then divide by the radius of the corresponding dimension. Square the result of the feature distribution circle, then sum the results of all feature dimensions sequentially and make the sum equal to 1. This establishes the standard equation of the feature distribution circle in high-dimensional space. ,in n This represents the total number of feature dimensions corresponding to the local feature vectors. For the first in higher-dimensional space i With dynamic coordinate variables in multiple dimensions, this equation can completely and uniquely describe the geometric position, morphological structure, and distribution range of the feature distribution circle on the high-dimensional feature circle manifold.

[0064] Step 3.2: Based on the determined feature drift line, establish the parametric equation of this feature drift line using its starting point and direction vector. Specifically, this includes: establishing the corresponding parametric equation based on the determined feature drift line, with the geometric information determined in Step 2.6 serving as the sole basis for the establishment process. Specifically, this involves: determining the center coordinates of the feature distribution circle... As the fixed starting point of the characteristic drift line, the offset vector used is followed. As the direction vector of the line, where Indicates the first i Offset components in each feature dimension, introducing independent real-valued parameters. s Used to identify the positions of different points on a straight line, and to represent any point on the feature drift line in high-dimensional space. i Dynamic coordinates in each dimension Represented as the coordinates of the center of the circle in that dimension, along with the components and parameters of the direction vector. s The sum of products, here Compared with step 3.1 The references are completely identical, both referring to the first [unit] in a higher-dimensional space. i The dynamic coordinate variables of each feature dimension are limited here to the corresponding dimension coordinates of any point on the feature drift line; expressions for each dimension are generated one by one according to this rule, and finally the parametric equations of the feature drift line in the high-dimensional feature circular manifold are established. ( i= 1 , 2 ,…,n )parameter s Each value uniquely corresponds to a spatial point on the straight line, which can completely describe the extended trajectory and spatial attitude of the feature drift line.

[0065] Step 3.3: Simultaneously solve the standard equation of the established feature distribution circle and the parametric equation of the established feature drift line to obtain the parameter values ​​that satisfy the condition of tangency between the line and the circle. Substitute these parameter values ​​into the line parametric equation to calculate the precise coordinates of the tangency point in the feature space. Specifically, this includes: first, substituting the obtained feature drift line parametric equation dimension by dimension into the established standard equation of the feature distribution circle, and then calculating the dynamic coordinate variables in the equation. Replace all with After the replacement is complete, first check each dimension. After simplification, we get Then, substitute the simplified result back into the equation and expand it to obtain a result containing only independent parameters. s The quadratic equation in one variable is: After further organization, it can be represented as Based on the core geometric condition that a line is tangent to a hypersphere in high-dimensional space, when the two have one and only one common point, the simultaneous equations have a unique solution, and the discriminant of the corresponding quadratic equation must be zero. Based on this constraint, the discriminant of the rearranged quadratic equation is calculated and solved to obtain the unique parameter value that satisfies the tangency condition. After the solution is completed, the parameter values ​​will be... Substitute each element back into the parametric equation of the feature drift line, and calculate the coordinate values ​​corresponding to each feature dimension, i.e., the coordinates of the tangent point in each dimension. Then, the coordinates of the tangent points of all feature dimensions are combined one by one according to the preset dimensional order, and finally the precise coordinates of the tangent points of the line and the feature distribution circle on the high-dimensional feature circle manifold are obtained. This coordinate provides a precise and unique spatial location basis for the subsequent calculation of the tangent direction vector.

[0066] Step 3.4: Based on the calculated precise coordinates of the tangent point and the center coordinates of the characteristic distribution circle, determine the direction passing through the tangent point and perpendicular to the line connecting the center of the circle to the tangent point, and generate the tangent direction vector. Specifically, this includes: based on the calculated precise coordinates of the tangent point... and the coordinates of the center of the defined characteristic distribution circle First, calculate the difference dimension by dimension to obtain the line vector connecting the center of the circle to the point of tangency. Then, in the high-dimensional feature circular manifold, we solve for the spatial vector that satisfies the perpendicular condition with the connecting vector, i.e., the vector whose inner product is zero. We then standardize and normalize the vectors that satisfy the condition to eliminate the influence of the magnitude and retain the direction information, finally generating the tangent direction vector that is tangent to the feature distribution circle at the tangent point. ,in τ This is the overall identifier for the tangent direction vector. The tangent direction vector is represented at the th... i Components in each feature dimension i= 1 , 2 ,…,n ( n (Still the total number of feature dimensions, consistent with the previous text), this tangent direction vector provides key directional constraints for subsequent feature correction and offset calculation.

[0067] In a preferred embodiment of the present invention, step 4 above may include:

[0068] Step 4.1: Based on the calculated precise coordinates of the tangent point and the extracted local feature vector, calculate the spatial distance from the local feature vector to the tangent point in the feature space, and use this spatial distance as the first correction component. Specifically, this includes retrieving the calculated precise coordinates of the tangent point on the high-dimensional feature circular manifold from the locally encrypted storage of the calculation results. Simultaneously, the original local feature vectors to be aligned are extracted from the local feature database and denoted as... ,in Specifically, this indicates that the original local feature vector is in the th... i The specific components in each feature dimension; under the spatial metric system of a high-dimensional feature circular manifold, according to the calculation rules of Euclidean distance, the corresponding dimension coordinates of the original local feature vector and the tangent point are subtracted one by one to obtain the first... i Coordinate difference in each dimension Then, the coordinate differences in each dimension are squared to eliminate the effect of the negative sign, thus merging all... n The squares of each feature dimension are summed, and then the square root of the sum is taken. The result is the global Euclidean distance from the original local feature vector to the tangent point in the feature space. This distance is formally defined as the first correction component, denoted as . This component can accurately quantify the spatial deviation of the original local feature vector relative to the target tangent point, providing a basic measurement basis for subsequent correction.

[0069] Step 4.2: Based on the generated tangent direction vector and the determined offset vector, calculate the cosine of the angle between the tangent direction vector and the offset vector. Multiply the cosine of the angle by the magnitude of the offset vector to obtain the second correction component. Specifically, this includes: extracting the standardized tangent direction vector generated in step 3.4 from the local calculation result library. and the offset vector determined in step 2.5 that characterizes the deviation between the local and global references. This ensures that the number and order of the feature dimensions of the two vectors are completely matched; in the high-dimensional feature space, following the rules for calculating the inner product of vectors, the dot product of the two vectors is calculated first. Then calculate the magnitude of the tangent direction vector. With the magnitude of the offset vector Since the tangent direction vector has been normalized, its magnitude is... Therefore, dividing the dot product result by the magnitude of the offset vector yields the cosine of the angle between the two vectors. This cosine value accurately reflects the degree of coincidence between the tangent direction and the global domain drift direction; finally, the calculated cosine value of the included angle is compared with the magnitude of the offset vector. The result of the multiplication operation is denoted as the second correction component. This component specifically characterizes the magnitude of the component in the original domain offset that can be effectively corrected along the tangential direction, thus achieving a precise measurement of the effectiveness of the correction direction.

[0070] Step 4.3 involves weighted fusing of the first correction component and the second correction component to generate a feature correction offset. Specifically, this includes: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] With the obtained second correction component A weighted fusion operation is performed to generate a comprehensive feature correction offset, with a preset weight coefficient for the first correction component. α =0.6, weighting coefficient of the second correction component β =0.4, which satisfies α+β The normalization constraint is set to 1, and this ratio is adapted to the weight requirements of spatial distance deviation and directional offset deviation of the insurance features; two independent non-negative weight coefficients are preset. α and β ,in α This is the weighting coefficient for the first correction component, used to adjust the contribution of spatial distance deviation to the total correction. β The weighting coefficients for the second correction component are used to adjust the contribution of the orientation alignment deviation to the total correction, and the two weighting coefficients satisfy the normalization constraint. α+β =1, ensuring numerical stability of the fusion process; the first correction component With corresponding weight coefficients α Multiplying them together yields the first weighted correction term. Then the second correction component With corresponding weight coefficients β Multiplying them together yields the second weighted correction term. The two weighted correction terms are summed to generate a feature correction offset that comprehensively reflects both spatial position deviation and directional drift deviation. This offset achieves an organic combination of two correction dimensions, ensuring the comprehensiveness, accuracy, and scene adaptability of subsequent feature correction operations.

[0071] Step 4.4 involves using a feature correction offset to perform a vector translation operation on the extracted local feature vectors, so that the translated feature vectors approximate the tangent point position on the feature distribution circle, resulting in a refined feature embedding with distribution alignment. Specifically, this includes using the generated comprehensive feature correction offset Δ to perform a vector translation operation on the extracted original local feature vectors. Perform precise vector translation operations; based on the coordinate transformation rules of the high-dimensional feature circular manifold, determine the direction of the translation operation as the direction from the original local feature vector to the tangent point, which is consistent with the target of feature correction; for each feature dimension component of the original local feature vector... All of them perform the same translation operation, changing the original components of this dimension. The new coordinate components after translation in each dimension are obtained by summing the offset Δ with the comprehensive feature correction. This ensures that all feature dimensions undergo synchronous and balanced correction; and that all... n The new coordinate components after each feature dimension translation are combined one by one according to the preset feature dimension order to form a new feature vector, which is defined as the refined feature embedding after distribution alignment, denoted as . This refined feature embedding can accurately approximate the tangent point position on the high-dimensional feature circular manifold, effectively eliminating the domain offset phenomenon of local feature distribution relative to the global benchmark, realizing the unified alignment of local feature distribution across participants, and providing high-quality feature data support for collaborative modeling tasks such as cross-domain insurance risk assessment and accurate pricing.

[0072] In a preferred embodiment of the present invention, step 5 above may include:

[0073] Step 5.1: Based on the obtained distribution-aligned refined feature embeddings, use them as input layer data for the local model. The local model is a preset initial model for risk prediction. Simultaneously, initialize the local model parameters and initiate the forward propagation process. Specifically, this includes: retrieving the obtained distribution-aligned refined feature embeddings from local encrypted storage, ensuring that the refined feature embeddings for each sample have eliminated domain offsets and meet the requirements for cross-participant feature distribution alignment; then using these refined feature embeddings one by one as input layer data for the local model, ensuring the input data is complete and error-free. The local model is a preset model for insurance risk prediction. The initial model is derived from the Multilayer Perceptron (MLP) architecture and improved to meet the core requirements of cross-participant feature alignment and privacy protection in this invention. The improved model retains the advantages of the MLP architecture, such as simple structure, adaptability to multi-dimensional features, and high training efficiency. At the same time, it optimizes the network hidden layer structure and activation function, solving the problems of overfitting and unreasonable feature weight allocation in traditional MLP in insurance multi-feature scenarios. It can more accurately capture insurance risk association features in refined feature embedding and is suitable for the application scenarios of cross-participant collaborative risk prediction and data privacy protection in this invention, such as cross-institutional risk assessment in auto insurance and health insurance.

[0074] The specific construction and training process of this initial model is as follows: the number of hidden layer nodes in the model follows... n The dimensional input layer features correspond to the first hidden layer 2. n One node, second hidden layer n The principle for setting each node, combined with the dimension of refined feature embedding in this invention, namely as described above... nThe system incorporates several feature dimensions, covering core insurance features such as driving behavior, vehicle operating status, claims history, and health indicators. This data forms the initial model's network structure, with the number of input layer nodes and the dimensions of refined feature embeddings. n To maintain consistency, a refined feature embedding after distribution alignment is received; 2 to 3 hidden layers are set, with the number of hidden layer nodes reasonably set according to the feature dimension, and the ReLU activation function is used to perform non-linear transformation on the input refined features to extract risk-related high-order features. At the same time, a Dropout layer is added to prevent the model from overfitting, adapting to the complex data scenarios with multiple participants and multiple features in this invention; the output layer has 1 node and uses the Sigmoid activation function to output the risk occurrence probability of the sample (adapting to the binary classification or probability prediction requirements of insurance risk prediction).

[0075] After the model is built, it is pre-trained using local historical insurance sample data that has not participated in cross-domain alignment. During the pre-training process, the actual risk labels of the samples (such as whether a claim has occurred and the claim amount level) are used as supervision. The loss function consistent with the subsequent training is used to gradually adjust the initial parameters of the model so that the model has basic risk prediction capabilities. After the pre-training is completed, training is stopped and the model at this time is used as the initial model. Each participant initializes and fine-tunes all network parameters of the initial model, sets reasonable initial value ranges for the parameters, and ensures that the parameter values ​​meet the needs of the insurance risk prediction scenario and are not abnormal. After initialization, the forward propagation process of the local model is started to prepare for risk prediction calculation.

[0076] Step 5.2: Through the forward propagation process, the refined feature embedding after distribution alignment is calculated layer by layer using the current local model to obtain the predicted risk value for each sample. Specifically, after each participant confirms that the local model initialization is complete and forward propagation has started, the refined feature embedding after distribution alignment input in Step 5.1 is used as the starting point for calculation, and the calculation is passed layer by layer according to the network hierarchy of the improved multilayer perceptron. First, the components of each dimension of the refined feature embedding are passed to the first hidden layer. The input feature data is linearly weighted and summed with the preset weight matrix of this layer, and a bias term is added. That is, a linear combination is completed through the weighted processing formula. The weighted processing formula is as follows: A = W × X + b ,in, A It is a linear combination of the corresponding network layers; W This is the weight matrix for the corresponding level. X Input features for this level, b The corresponding level bias term is then transformed nonlinearly using the ReLU activation function to extract basic risk association features. At the same time, the Dropout mechanism is used to prevent overfitting. The output of this layer is used as the input of the next layer, and the above weighting and activation operations are repeated to deeply fuse and extract higher-order risk association features.

[0077] If the model has a third hidden layer, the feature calculation and transfer will continue according to the above rules, gradually improving the abstraction and risk correlation of the features. After the feature extraction and transformation of all hidden layers are completed, the high-order risk feature matrix output by the last hidden layer will be accurately input into the output layer of the model. The output layer will first perform linear weighted summation and bias term superposition on the input high-order feature data, and then input the calculation result into the Sigmoid activation function to stably map the output value to the range of 0 to 1. Finally, a unique numerical result will be generated for each sample. This numerical result is the predicted risk value for that sample. The closer the value is to 1, the higher the probability of occurrence of the insurance risk and the higher the risk level of the sample. The closer the value is to 0, the lower the probability of occurrence of the insurance risk and the lower the risk level of the sample.

[0078] Step 5.3 compares the predicted risk value for each sample with the corresponding true label value. A pre-defined loss function is used to calculate the difference between the predicted risk value and the true label value, generating the loss value for the current batch. Specifically, this involves: retrieving the true label value for each sample. This true label value is pre-labeled and accurately reflects the actual insurance risk of the sample, and its representation dimension is consistent with the predicted risk value, such as risk level or risk probability. Then, the predicted risk value for each sample is compared one-to-one with its corresponding true label value to determine the magnitude of the deviation. Finally, a pre-defined loss function is used, specifically adapted to insurance risk prediction scenarios. The cross-entropy loss function is selected (adapting to the 0-1 probability range of the predicted risk value, accurately measuring the difference between the predicted and true values). Its specific mathematical expression is:

[0079] ;

[0080] In the formula This represents the loss value for the current batch. M This represents the total number of samples in the current training batch. For the first j The true label value of each sample For the first j The predicted risk value of each sample is calculated; the difference between the predicted risk value and the true label value of each sample is calculated one by one using the loss function, and then the average value of the difference results of all samples in the current training batch is taken to obtain the loss value of the current batch. The loss value can effectively reflect the prediction accuracy of the current local model.

[0081] Step 5.4: Based on the loss value of the current batch, perform backpropagation calculation to determine the gradients corresponding to each network parameter in the local model layer by layer, and generate the original gradient parameter set. Specifically, after obtaining the calculated loss value of the current batch, immediately start the backpropagation calculation process of the local model. The backpropagation process is in the exact opposite direction to the forward propagation process, strictly starting from the output layer of the model and tracing back layer by layer to the input layer to ensure that the gradient calculation of each layer can accurately correlate with the change of the loss value. During the calculation process, each participant will determine the gradient corresponding to each network parameter in the local model (including the weight matrix and bias term of each hidden layer and the weight matrix and bias term of the output layer) according to the specific magnitude and trend of the loss value of the current batch and the correlation between the loss value and each network parameter.

[0082] The core function of the gradient is to accurately characterize the influence of each network parameter on the loss value. The larger the absolute value of the gradient, the more significant the impact of a small change in that parameter on the model's loss value, and the more obvious the impact on the model's prediction accuracy. A positive gradient indicates that the loss value will increase when the parameter increases, while a negative gradient indicates that the loss value will decrease when the parameter increases. After the gradients corresponding to all network parameters are calculated, according to the set parameter order, that is, the order in which the parameters are called in the forward propagation, starting from the parameters corresponding to the input layer, and then proceeding to the parameters of each hidden layer and the output layer, all gradients are sorted and classified one by one to form a complete set of original gradient parameters. This set completely contains the gradient information of all network parameters of the local model without omissions or misalignments, and is an important core basis for updating local model parameters and cross-participant model collaborative optimization.

[0083] Step 5.5 involves encrypting the original gradient parameter set using a preset homomorphic encryption method to obtain the encrypted gradient parameters. Specifically, to strictly ensure the privacy and security of the gradient parameters, prevent the leakage of original gradient information, protect the local data privacy of each participant, and align with the core requirements of cross-participant collaboration and privacy protection in this invention, each participant will use a preset homomorphic encryption method: a homomorphic encryption method based on the Paillier algorithm is selected. This method is suitable for the cross-participant gradient aggregation requirements of this invention, and can complete collaborative operations such as gradient summation and weighting without decryption. The specific encryption process is as follows: each gradient value in the original gradient parameter set is preprocessed, converting the gradient value into an integer format that conforms to the requirements of the Paillier algorithm to avoid numerical deviation; then, the set encryption key is called to perform Paillier encryption on each preprocessed gradient value, converting the original gradient value into ciphertext through public key encryption, strictly preserving the numerical correlation characteristics of the gradient values ​​during the encryption process; all encrypted individual gradient ciphertexts are integrated according to the order of the original gradient parameter set to form complete encrypted gradient parameters.

[0084] The entire encryption process strictly follows the encryption rules of the Paillier algorithm, ensuring that the encrypted encrypted gradient parameters do not reveal any numerical information, trend of change, or parameter correlation of the original gradient. At the same time, it takes into account the needs of subsequent collaborative computation, ensuring that when the subsequent coordinating nodes aggregate the gradients of all participants, they can directly perform the relevant calculations without decryption. This ensures both privacy and security, and does not affect the efficiency of cross-participant model collaborative training. After all encryption operations are completed, complete and secure encrypted gradient parameters are obtained, ensuring that the gradient data is in a secure encrypted state throughout the subsequent transmission and computation process.

[0085] Step 5.6 involves uploading the encrypted gradient parameters to the coordination node via a secure channel. Specifically, each participant uploads the encrypted encrypted gradient parameters to the coordination node through a pre-established encrypted secure channel. This encrypted secure channel has undergone security testing and verification beforehand, possessing anti-tampering, anti-theft, and anti-monitoring functions, effectively resisting various security risks during data transmission. During the upload process, the channel's security is verified in real time; a security verification is performed after each data segment is transmitted to promptly identify any anomalies during transmission, preventing the encrypted gradient parameters from being illegally tampered with, stolen, or lost during transmission, ensuring that the encrypted gradient parameters are transmitted securely, completely, and without deviation to the coordination node. After receiving and confirming that the encrypted gradient parameters are correct, the coordination node will use them for subsequent cross-participant gradient aggregation and model parameter updates, providing secure, reliable, and accurate gradient data support for cross-participant model collaborative training, ensuring the security and effectiveness of the entire collaborative training process.

[0086] In a preferred embodiment of the present invention, step 6 above may include:

[0087] Step 6.1: The coordinating node receives encrypted gradient parameters uploaded by each participant via a secure channel. These encrypted gradient parameters are the uploaded, encrypted gradient data. Specifically, the coordinating node is in a standby receiving state, constantly monitoring the encrypted gradient parameter upload requests from each participant to ensure timely response to each participant's upload operation. When each participant uploads the encrypted encrypted gradient parameters via the established secure encrypted channel, the coordinating node first performs a security verification on the upload request. Only after confirming that the request originates from a legitimate participant and that there are no signs of illegal tampering or forgery does the coordinating node formally receive the encrypted gradient parameters uploaded by each participant. These encrypted gradient parameters are the gradient data obtained by each participant after encrypting the original gradient parameters. The entire process is encrypted, effectively preventing data leakage during reception and transmission, perfectly meeting the core requirements of cross-participant collaboration and privacy protection in this invention.

[0088] Step 6.2: Decrypt the received encrypted gradient parameters from each participant using the decryption process corresponding to the encryption method to restore the original gradient parameters of each participant. Specifically, after confirming that all encrypted gradient parameters uploaded by all participants have been received, the coordinating node sequentially performs source legality verification, data integrity verification, format standardization verification, and length consistency verification on each encrypted gradient parameter. The verification process compares the participant identifier, data length, and field structure one by one with the pre-agreed transmission standard to ensure that the encrypted gradient parameters have not been lost, corrupted, tampered with, or forged during transmission. Only after all verifications pass can the decryption process begin. Subsequently, the Paillier homomorphic decryption algorithm, which is strictly corresponding to the encryption method of the participants, is used. This algorithm is a probabilistic homomorphic encryption algorithm based on the large integer factorization problem and is specifically adapted to this invention. The core requirement for privacy protection and collaborative computing among participants in this invention is the support for additive homomorphism and scalar multiplication homomorphism. It enables aggregation operations such as gradient summation and weighting without decryption. The decryption process only requires the coordinating node to hold a dedicated decryption private key, and the decryption result is unique and secure, effectively preventing the leakage of ciphertext information while accurately restoring plaintext data. This meets the decryption requirements of the encrypted gradient parameters in this invention. Then, using the securely managed dedicated decryption private key, each encrypted gradient parameter is decrypted element-by-element according to the parameter order. During the decryption process, the accurate mapping between ciphertext and plaintext is ensured, ultimately restoring the unencrypted original gradient parameters of each participant. This includes key information such as the gradient values ​​and gradient proportions of all network parameters, providing a real, accurate, and directly computable data foundation for subsequent gradient weighting, angle correction, and aggregation calculations.

[0089] Step 6.3: Obtain a preset inter-domain difference metric. This preset inter-domain difference metric is used to quantify the degree of gradient direction deviation caused by differences in data distribution among different participants. Specifically, after restoring the original gradient parameters of all participants, the preset inter-domain difference metric is retrieved. This preset inter-domain difference metric is a quantitative indicator specifically constructed by combining the feature alignment process of this invention, the data distribution characteristics of cross-participants, and the insurance risk prediction scenario. It is mainly used to measure the degree of gradient direction deviation and gradient value difference among different participants caused by factors such as uneven distribution of local samples, residual domain offset, differences in feature distribution, and local data noise. The preset inter-domain difference metric incorporates the local data scale, feature alignment error, historical gradient fluctuation, and model training stability of each participant when setting it. It can objectively and stably reflect the consistency level of the gradient of each participant with the global gradient trend, and can provide a reliable, unified, and physically meaningful measurement basis for weighting coefficient calculation, gradient direction correction, and aggregated gradient generation.

[0090] Step 6.4: Based on the preset inter-domain difference metric, calculate the weighting coefficients of the original gradient parameters of each participant. The weighting coefficients are inversely proportional to the inter-domain difference metric; that is, the greater the inter-domain difference, the lower the gradient contribution weight of the participant. Specifically, this includes: using the retrieved preset inter-domain difference metric as the core calculation basis, calculating the corresponding adaptive weighting coefficients for each participant one by one, ensuring that the weighting coefficients accurately match the degree of inter-domain difference for each participant; let the... The inter-domain difference measure for each participant is: (wherein the inter-domain difference measure) By calculating the first u The weighting coefficients are obtained by calculating the Wasserstein distance between the local feature distributions of each participant and the global baseline feature distribution. The weighting coefficients and the inter-domain difference measure strictly satisfy an inverse relationship, and their calculation relationship can be expressed as follows: The core purpose of this design is to achieve precise control over the gradient contributions of different participants. The greater the inter-domain difference metric, the higher the deviation of the gradient of that participant from the global gradient trend, and the smaller its corresponding weighting coefficient. Consequently, the gradient of that participant contributes less to the subsequent aggregation process. This approach can effectively suppress the interference of participants with large domain offsets on the global gradient, prevent abnormal gradients from affecting the accuracy of aggregated gradients, and ensure the stability of subsequent global model training.

[0091] Step 6.5: For the original gradient parameters of each participant obtained through decryption, extract their gradient direction vectors respectively. Perform angle correction on each gradient direction vector according to a preset inter-domain difference metric to minimize the angle between the corrected gradient direction and the sum of the original gradient vectors of all participants. Specifically, after completing the decryption and verification of the original gradient parameters of each participant, the coordination node accurately extracts the corresponding gradient direction vector from the original gradient parameters of each participant. Let the first... The original gradient direction vectors of each participant are This vector can fully characterize the direction and trend of gradient change of the participant; using a preset inter-domain difference metric as a reference standard, a targeted angle correction operation is performed on the gradient direction vector of each participant. The core objective of the correction is to make the corrected gradient direction vector... The sum of the original gradient vectors of all participants The angle between them is minimized; the entire correction process takes vector direction alignment as the core principle, focusing on eliminating gradient direction offset caused by factors such as domain differences and uneven data distribution, so that the gradient directions of all participants are as close as possible to the unified global optimization direction, laying a consistent foundation for subsequent gradient aggregation.

[0092] Step 6.6 involves multiplying the calculated weighting coefficients by the angle-corrected gradient parameters to obtain the weighted corrected gradient for each participant. Specifically, after completing the weighting coefficient calculation and gradient direction correction, the coordination node multiplies the weighting coefficients corresponding to each participant. The gradient obtained after correction of the corresponding participant angle. Element-wise precise multiplication is performed; during the calculation, it is ensured that the weighting coefficient of each participant is accurately applied to its corresponding corrected gradient, without parameter misalignment or calculation deviation, ultimately yielding the... u The gradient after weighted correction by each participant This operation achieves a precise match between gradient magnitude and the degree of difference between domains. It retains the effective gradient information of participants with small domain differences and high gradient reliability, while further weakening the gradient influence of participants with large domain offsets, effectively improving the stability and reliability of global gradient aggregation.

[0093] Step 6.7: Sum and average the weighted corrected gradients of all participants to generate the corrected aggregate gradient. Specifically, after obtaining the weighted corrected gradients of all participants, the coordinating node sums and aggregates the weighted corrected gradients of all participants. During the summation process, the gradient data of each participant is checked one by one to ensure no omissions or calculation errors, and to completely summarize the effective gradient information of all participants. The total gradient value obtained by summing is then divided by the total number of participants. The averaging operation is completed, and finally the corrected aggregated gradient is generated. The corrected aggregation gradient is calculated as follows:

[0094] The corrected aggregated gradient fully integrates the effective gradient information of all participants. At the same time, through the dual operations of weighting and angle correction, it effectively weakens the adverse effects of domain offset. The gradient direction is unified and the values ​​are accurate and reliable. It can provide accurate and robust gradient guidance for the parameter update of the global model, ensuring the effect of cross-participant collaborative training, which is in line with the core requirements of the insurance risk prediction of this invention.

[0095] In a preferred embodiment of the present invention, step 7 above may include:

[0096] Step 7.1: Obtain the corrected aggregated gradient through the coordination node, combine the corrected aggregated gradient with the current global model parameters, and update the global model parameters according to the set optimizer update rules to generate updated global model parameters. Specifically, this includes: immediately retrieving and obtaining the corrected aggregated gradient after the aggregated gradient is generated, performing a secondary verification on the aggregated gradient, checking the gradient's numerical range, parameter dimensions, and arrangement order one by one, confirming that the aggregated gradient is complete, unbiased, and without outliers, and has undergone weighting coefficient adjustment and gradient direction angle correction, accurately reflecting the gradient collaborative optimization direction of all participants, and meeting the requirements for global model parameter updates; then retrieving the currently used global model parameters, which include the weight matrices and bias terms of all hidden layers in the global model, as well as the weight matrices and bias terms of the output layer, and are completely consistent with the local model parameter structure of each participant.

[0097] The corrected aggregated gradient is then mapped one-to-one with the current global model parameters, and the global model parameter update operation is executed strictly according to the preset optimizer update rule. This optimizer update rule is adapted to the model training characteristics of the insurance risk prediction scenario, and momentum gradient descent is selected. The SGD (Simultaneous Gradient Descent) optimization rule's core is to combine the current aggregated gradient with historical update momentum during parameter updates. This ensures that the parameter update direction aligns with the gradient guidance while suppressing parameter oscillations and accelerating convergence. The specific execution logic is as follows: First, a momentum term is calculated based on historical parameter update records. This momentum term reflects the trend of previous parameter updates. Then, the current corrected aggregated gradient is fused with the momentum term to obtain the final parameter update amount. A fixed learning rate is set to control the update magnitude. The learning rate value is preset based on the convergence characteristics of the insurance risk prediction model to avoid excessively large update magnitudes leading to model instability or excessively small update magnitudes leading to slow convergence. Differentiated learning rate decay strategies are used for different types of parameters (weight matrix, bias term), with the learning rate decaying slightly faster for the weight matrix than for the bias term, adapting to the learning patterns of insurance risk characteristics. During the update process, the values ​​of each parameter before and after the update are recorded in real time for easy traceability and anomaly detection, ensuring that each global model parameter is accurately adjusted, ultimately generating the updated global model parameters.

[0098] Step 7.2: The updated global model parameters are encrypted using a preset homomorphic encryption method to generate encrypted global model parameters. Specifically, after generating the updated global model parameters, to fully protect the privacy and security of the parameters during cross-participant distribution, prevent leakage of global model parameters, and avoid unauthorized acquisition of core model information from each participant, a preset homomorphic encryption method identical to that in Steps 5.5 and 6.2 is used: namely, a homomorphic encryption method based on the Paillier algorithm, to fully encrypt the updated global model parameters. During the encryption process, the updated global model parameters are first preprocessed, converting the parameter values ​​into symbols. The system uses an integer format that meets the encryption requirements of the Paillier algorithm, eliminating invalid values ​​and correcting format deviations to ensure that the parameter format is fully compatible with the encryption algorithm. It calls a preset encryption public key, which is consistent with the public key used by each participant when encrypting the original gradient parameters. This public key is securely stored and subject to access control. The system encrypts the global model parameters element by element according to the parameter type and order. During the encryption process, it ensures that the ciphertext and plaintext of each parameter can be accurately mapped. The encryption process is error-free and without omissions, and finally generates complete encrypted global model parameters. This ensures that the global model parameters remain encrypted during subsequent distribution, effectively preventing unauthorized access and fully protecting the privacy of the model parameters.

[0099] Step 7.3 involves distributing the encrypted global model parameters to each participant via a secure channel. This includes: after completing the encryption of the global model parameters and generating the encrypted global model parameters, performing a dual verification on the encrypted global model parameters. Firstly, verifying the integrity of the encrypted parameters to confirm that all global model parameters have been encrypted without missing, incorrect, or duplicated parameters. Secondly, verifying the encryption effectiveness by randomly selecting a portion of the encrypted parameters for trial decryption testing to confirm that the encryption effect meets the preset requirements and can effectively resist unauthorized decryption. After successful verification, the coordinating node communicates with each participant via a pre-established encrypted secure channel. The channel used for transmitting state gradient parameters is consistent and has anti-tampering, anti-theft, and anti-monitoring functions. It has also undergone prior security testing and verification. The dense state global model parameters are distributed to each participant one by one. During the distribution process, the coordinating node monitors the security status of the channel in real time. After each transmission of a segment of dense state parameters, a security check is performed to verify the integrity and consistency of the data and promptly investigate any abnormalities in the transmission process. This prevents the dense state global model parameters from being illegally tampered with, stolen, lost, or misplaced during transmission, ensuring that the dense state parameters can be transmitted safely, completely, and without deviation to each participant, thus guaranteeing the security and reliability of model parameter distribution.

[0100] Step 7.4: Each participant receives the distributed encrypted global model parameters and decrypts them using the decryption process corresponding to the encryption method to restore the updated plaintext global model parameters. Specifically, each participant remains in parameter receiving standby mode, monitors the parameter distribution signal of the coordination node in real time, and ensures timely response to the coordination node's distribution operation. Upon receiving the encrypted global model parameters distributed by the coordination node via the encrypted secure channel, the participant first performs dual verification on the received encrypted parameters. On the one hand, the participant verifies the legitimacy of the parameter source by comparing it with the coordination node's unique identifier to confirm that the parameters originate from a legitimate coordination node, thus preventing unauthorized parameters from accessing the system. On the other hand, the participant verifies the integrity of the parameters by checking the length, format, and field structure of each encrypted parameter to confirm that there are no missing, misaligned, or tampered parameters.

[0101] After successful verification, each participant employs a decryption process that strictly corresponds to the encryption method used by the coordinating node. This decryption method is based on the Paillier algorithm, which calls a dedicated decryption private key that is pre-stored securely locally. This private key corresponds one-to-one with the public key used by the coordinating node during encryption and is managed by designated personnel. The call to this private key requires authentication and permission verification. The encrypted global model parameters are decrypted element by element in the parameter order. The decryption process strictly follows the decryption rules of the Paillier algorithm, accurately reconstructing the plaintext parameters corresponding to the ciphertext. There are no numerical deviations or order errors throughout the process, ensuring that the updated global model parameter plaintext can be completely and accurately restored.

[0102] Step 7.5: Based on the obtained updated global model parameter plaintext, replace and update the locally stored local model parameters to complete the synchronization of the local model in this round. Specifically, this includes: after successfully restoring the updated global model parameter plaintext, immediately retrieving the currently used local model parameters. These parameters include the weight matrices and bias terms of all hidden layers and the weight matrix and bias terms of the output layer in the local model, and are completely consistent with the structure and dimensions of the global model parameter plaintext. Each participant compares the local model parameters with the global model parameter plaintext element by element and dimension by dimension to confirm that the parameter types, dimension specifications, and arrangement order of the two are completely matched, with no mismatches. To avoid local model update failure due to parameter mismatch, after verification, the corresponding local model parameters are replaced one by one with the updated global model parameters in plaintext according to the parameter correspondence. During the replacement process, the values ​​of each parameter before and after replacement are recorded in real time to ensure that the parameter replacement is without misalignment, omission, or duplication. After the replacement is completed, the integrity of the local model parameters is checked again to confirm that the local model parameters have been fully synchronized with the updated global model parameters, so that the local model and the global model are highly consistent. This ensures that the local models of all participants can be trained based on the unified global parameters in the next round of training iteration, and is fully prepared for the next round of training iteration.

[0103] Step 7.6: Determine whether the current iteration round has reached the set maximum number of iteration rounds or whether the global model meets the set convergence conditions. If it has not reached the maximum number of iteration rounds and has not converged, repeat the current iteration training operation based on the updated global model parameters. If it has reached the maximum number of iteration rounds or has converged, stop the iteration and determine the current global model as a cross-domain collaborative computing model that resists domain offset. Specifically, this includes: after completing the local model parameter synchronization, immediately sending a parameter synchronization completion signal to the coordination node. After receiving synchronization completion feedback from all participants, the coordination node formally starts the iteration judgment process to ensure the accuracy and comprehensiveness of the iteration judgment. The iteration judgment mainly includes... It contains two core dimensions. On the one hand, the coordination node retrieves the current iteration cycle record and compares it with the set maximum number of iteration cycles to determine whether the current iteration cycle has reached the set maximum number of iteration cycles. On the other hand, it retrieves the training data of the global model to determine whether the global model meets the set convergence conditions. These convergence conditions are pre-set based on the actual needs of insurance risk prediction and are mainly based on indicators such as the prediction accuracy of the global model, the fluctuation of the loss value, and the magnitude of gradient change. When the prediction accuracy of the global model reaches the set threshold, the loss value tends to stabilize, and the fluctuation range is within the preset range, the global model is determined to have reached the convergence state.

[0104] If the current iteration round has not reached the maximum number of iteration rounds and the global model has not met the convergence condition, each participant will repeat the iteration training operation based on the updated global model parameters to continue cross-participant collaborative training and continuously optimize model performance. If the current iteration round has reached the maximum number of iteration rounds or the global model has met the convergence condition, all iteration training operations will be stopped immediately, and the updated global model will be determined as a cross-domain collaborative computing model resistant to domain offset. This model can effectively adapt to the differences in data distribution among participants, resist the adverse effects of domain offset, accurately complete the insurance risk prediction task, and effectively meet the core requirements of cross-participant collaboration, privacy protection, and resistance to domain offset of this invention.

[0105] like Figure 2 As shown, embodiments of the present invention also provide a cross-domain insurance data collaborative computing data processing system based on federated learning, comprising:

[0106] The encrypted sample alignment module is used by each participant to perform privacy intersection processing on the insurance-related data stored locally. It completes the encrypted alignment of cross-domain shared samples without revealing the original plaintext, and obtains the encrypted aligned dataset.

[0107] The local feature modeling and domain offset detection module is used to extract local features based on the dense state aligned dataset, map the local features to a high-dimensional feature circle manifold, construct a feature distribution circle with the feature distribution mean as the center and the feature distribution variance as the radius; determine the domain offset existing in the circle based on the local features, and map the domain offset to a feature drift line;

[0108] The high-dimensional feature geometry correction module is used to solve the equation of the tangent line between the feature drift line and the feature distribution circle to obtain the coordinates of the tangent point and the tangent direction vector; based on the coordinates of the tangent point and the tangent direction vector, the feature correction offset is calculated to dynamically calibrate local features and generate refined feature embeddings after distribution alignment.

[0109] The model training and dense gradient upload module is used to start local model training with refined feature embedding after distribution alignment, calculate the loss after forward propagation, generate dense gradient parameters through back propagation, and upload the encrypted dense gradient parameters to the coordinating node.

[0110] The gradient adaptive aggregation correction module is used to receive the encrypted gradient parameters uploaded by each participant, and after decryption or recombination, adaptively weight and correct the direction of each gradient contribution according to the preset inter-domain difference metric to generate the corrected aggregate gradient.

[0111] The global model collaborative update module is used to coordinate nodes to update global model parameters based on the corrected aggregated gradient. By distributing the updated global model parameters in encrypted form back to each participant, each participant can update its local model until a cross-domain collaborative computing model resistant to domain offset is obtained.

[0112] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0113] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A cross-domain insurance data collaborative computing data processing method based on federated learning, characterized in that, The method includes: By performing privacy-preserving intersection processing on the locally stored insurance-related data by each participating party, the encrypted alignment of cross-domain shared samples is completed without disclosing the original plaintext, resulting in a encrypted aligned dataset. Based on the dense-state aligned dataset, local features are extracted locally and mapped onto a high-dimensional feature circle manifold. A feature distribution circle is constructed with the mean of the feature distribution as the center and the variance of the feature distribution as the radius. The domain offsets existing in the circle are determined according to the local features and mapped to feature drift lines. Solve for the equations of the tangents between the characteristic drift line and the characteristic distribution circle to obtain the coordinates of the tangent point and the tangent direction vector; The feature correction offset is calculated based on the coordinates of the tangent point and the tangent direction vector, which is used to dynamically calibrate local features and generate refined feature embeddings after distribution alignment. Local model training is initiated using refined feature embeddings with distribution alignment, loss is calculated after forward propagation, dense gradient parameters are generated through backpropagation, and the dense gradient parameters are encrypted and uploaded to the coordinating node. The system receives encrypted gradient parameters uploaded by each participant, and after decryption or recombination, it performs adaptive weighting and direction correction on the contribution of each gradient according to the preset inter-domain difference metric to generate the corrected aggregate gradient. The coordinating node updates the global model parameters based on the corrected aggregated gradient, and distributes the updated global model parameters back to each participant in encrypted form so that each participant can update its local model until a cross-domain collaborative computing model resistant to domain offset is obtained.

2. The cross-domain insurance data collaborative computing data processing method based on federated learning according to claim 1, characterized in that, By performing privacy-preserving intersection processing on the locally stored insurance-related data by each participating party, the encrypted alignment of cross-domain shared samples is completed without revealing the original plaintext, resulting in a encrypted aligned dataset, including: Each participant performs standardized preprocessing on the sample identifiers in the insurance-related data stored locally to generate a unified set of sample identifiers to be intersected. Each sample identifier in the set of sample identifiers to be intersected is encrypted using a preset homomorphic encryption method to generate a corresponding encrypted sample identifier, and all encrypted sample identifiers are combined to form a local encrypted identifier dataset. Each participant uploads its local secret state identifier dataset to the coordination node, which then performs a secret state comparison operation on all received secret state identifier datasets to filter out the secret state sample identifiers common to all participants, thus forming a secret state intersection identifier set. The coordinating node distributes the set of secret intersection identifiers to each participant. Each participant decrypts the received set of secret intersection identifiers locally and maps it back to the original sample identifier. This is used to extract the complete data record corresponding to the original sample identifier from the local insurance-related data, and finally obtains a secret alignment dataset containing only the common samples.

3. The cross-domain insurance data collaborative computing data processing method based on federated learning according to claim 2, characterized in that, Based on the dense-state aligned dataset, local features are extracted locally and mapped onto a high-dimensional feature circle manifold. A feature distribution circle is constructed with the mean of the feature distribution as its center and the variance of the feature distribution as its radius, including: Based on the obtained close-state aligned dataset, each participant extracts features from the original data fields of each sample record in the dataset to generate a multi-dimensional local feature vector composed of multiple feature dimensions. The multiple feature dimensions include at least one or more of the following: driving behavior feature dimension, vehicle operating status feature dimension, historical claims feature dimension, and road violation feature dimension. All local feature vectors are then aggregated to form a local feature set. Statistical analysis is performed on the local feature set to calculate the average value of all local feature vectors in the set on each feature dimension, and the average values ​​are combined to generate a mean vector; at the same time, the covariance matrix of the local feature set is calculated, and the variance value corresponding to each feature dimension is extracted from the diagonal of the covariance matrix as the distribution variance of each dimension. Based on the calculated mean vector and the distribution variance of each dimension, each participant constructs a feature distribution circle that reflects the local feature distribution pattern on a preset high-dimensional feature circle manifold, with the point represented by the mean vector as the center and the distribution variance of each dimension as the radius on each axis.

4. The cross-domain insurance data collaborative computing data processing method based on federated learning according to claim 3, characterized in that, Based on the local features, determine the domain offsets present therein, and map the domain offsets to feature drift lines, including: Obtain the global benchmark feature distribution parameters pre-distributed by the coordination node. The global benchmark feature distribution parameters include the global benchmark mean vector and the variance of each dimension of the global benchmark. The calculated mean vector is compared with the obtained global benchmark mean vector to calculate the difference between the mean vector and the global benchmark mean vector. The difference is used as the offset vector, wherein the offset vector includes the offset direction and offset distance in each feature dimension. Based on the obtained offset vector, starting from the center of the constructed feature distribution circle, a straight line is constructed on the high-dimensional feature circle manifold along the offset direction indicated by the offset vector, and this straight line is determined as the feature drift line.

5. The cross-domain insurance data collaborative computing data processing method based on federated learning according to claim 4, characterized in that, Solving for the equations of the tangents between the characteristic drift line and the characteristic distribution circle yields the coordinates of the tangent point and the tangent direction vector, including: Based on the constructed characteristic distribution circle, the standard equation of this characteristic distribution circle is established on the high-dimensional characteristic circle manifold through its center coordinates and the distribution variance of each dimension. Based on the defined characteristic drift line, the parametric equation of this characteristic drift line is established using its starting point and direction vector; By simultaneously establishing the standard equation of the characteristic distribution circle and the parametric equation of the characteristic drift line, the parameter values ​​that satisfy the condition of tangency between the line and the circle are solved. The parameter values ​​are then substituted into the parametric equation of the line to calculate the precise coordinates of the tangency point in the characteristic space. Based on the calculated precise coordinates of the tangent point and the center coordinates of the characteristic distribution circle, the direction passing through the tangent point and perpendicular to the line connecting the center of the circle to the tangent point is determined, and the tangent direction vector is generated.

6. The cross-domain insurance data collaborative computing data processing method based on federated learning according to claim 5, characterized in that, The feature correction offset is calculated based on the tangent point coordinates and the tangent direction vector, which is used for dynamic calibration of local features to generate a refined feature embedding after distribution alignment, including: Based on the calculated precise coordinates of the tangent point and the extracted local feature vector, the spatial distance from the local feature vector to the tangent point in the feature space is calculated, and this spatial distance is used as the first correction component. Based on the generated tangent direction vector and the determined offset vector, the cosine value of the angle between the tangent direction vector and the offset vector is calculated, and the cosine value of the angle is multiplied by the magnitude of the offset vector to obtain the second correction component. The first correction component and the second correction component are weighted and fused to generate the feature correction offset; The extracted local feature vectors are translated using a feature correction offset to make the translated feature vectors approximate the tangent point on the feature distribution circle, thus obtaining a refined feature embedding after distribution alignment.

7. The cross-domain insurance data collaborative computing data processing method based on federated learning according to claim 6, characterized in that, Local model training is initiated using refined feature embeddings with distribution alignment. Loss is calculated after forward propagation, and dense gradient parameters are generated via backpropagation. These dense gradient parameters are then encrypted and uploaded to the coordinating node, including: Based on the refined feature embedding obtained after distribution alignment, it is used as the input layer data of the local model. The local model is a preset initial model for risk prediction. At the same time, the local model parameters are initialized and the forward propagation process is started. Through the forward propagation process, the refined feature embedding after distribution alignment is calculated layer by layer using the current local model to obtain the predicted risk value for each sample. The predicted risk value for each sample is compared with the corresponding true label value. The difference between the predicted risk value and the true label value is calculated using a set loss function to generate the loss value for the current batch. Based on the loss value of the current batch, backpropagation is performed to determine the gradient corresponding to each network parameter in the local model layer by layer, and the original gradient parameter set is generated. The original gradient parameter set is encrypted using a preset homomorphic encryption method to obtain dense gradient parameters; The dense gradient parameters are uploaded to the coordination node via a secure channel.

8. The cross-domain insurance data collaborative computing data processing method based on federated learning according to claim 7, characterized in that, The system receives encrypted gradient parameters uploaded by each participant. After decryption or recombination, it adaptively weights and corrects the direction of each gradient contribution based on a preset inter-domain difference metric, generating a corrected aggregate gradient, including: The coordinating node receives encrypted gradient parameters uploaded by each participant via a secure channel. These encrypted gradient parameters are the uploaded gradient data. The received encrypted gradient parameters of each participant are decrypted using a decryption process corresponding to the encryption method to restore the original gradient parameters of each participant. Obtain a preset inter-domain difference metric, which is used to quantify the degree of gradient direction deviation between different participants due to differences in data distribution; Based on the preset inter-domain difference metric, the weighting coefficients of the original gradient parameters of each participant are calculated. The weighting coefficients are inversely proportional to the inter-domain difference metric, meaning that the greater the inter-domain difference, the lower the gradient contribution weight of the participant. For the original gradient parameters of each participant obtained from decryption, their gradient direction vectors are extracted respectively. The angle of each gradient direction vector is corrected according to the preset inter-domain difference metric so as to minimize the angle between the corrected gradient direction and the sum of the original gradient vectors of all participants. Multiply the calculated weighting coefficients by the angle-corrected gradient parameters to obtain the weighted corrected gradient for each participant; The weighted gradients of all participants are summed and averaged to generate the corrected aggregate gradient.

9. The cross-domain insurance data collaborative computing data processing method based on federated learning according to claim 8, characterized in that, The coordinating node updates the global model parameters based on the corrected aggregated gradient, and distributes the updated global model parameters back to each participant in encrypted form, enabling each participant to update its local model until a cross-domain collaborative computation model resistant to domain offset is obtained, including: The corrected aggregated gradient is obtained through the coordination node, and then combined with the current global model parameters. The global model parameters are updated according to the set optimizer update rules to generate the updated global model parameters. The updated global model parameters are encrypted using a preset homomorphic encryption method to generate encrypted global model parameters. The dense global model parameters are distributed to all participants via a secure channel; Each participant receives the distributed encrypted global model parameters, decrypts them using the decryption process corresponding to the encryption method, and restores the updated plaintext global model parameters. Based on the obtained updated global model parameter plaintext, the locally stored local model parameters are replaced and updated to complete the synchronization of the local model in this round. Determine whether the current iteration round has reached the set maximum number of iteration rounds or whether the global model meets the set convergence conditions. If it has not reached the maximum number of iteration rounds or has not converged, repeat the current iteration training operation based on the updated global model parameters. If it has reached the maximum number of iteration rounds or has converged, stop the iteration and determine the current global model as a cross-domain collaborative computing model that resists domain offset.

10. A cross-domain insurance data collaborative computing and processing system based on federated learning, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, include: The encrypted sample alignment module is used by each participant to perform privacy intersection processing on the insurance-related data stored locally. It completes the encrypted alignment of cross-domain shared samples without revealing the original plaintext, and obtains the encrypted aligned dataset. The local feature modeling and domain offset detection module is used to extract local features based on the dense state aligned dataset, map the local features to a high-dimensional feature circle manifold, construct a feature distribution circle with the feature distribution mean as the center and the feature distribution variance as the radius; determine the domain offset existing in the circle based on the local features, and map the domain offset to a feature drift line; The high-dimensional feature geometry correction module is used to solve the equation of the tangent line between the feature drift line and the feature distribution circle to obtain the coordinates of the tangent point and the tangent direction vector; based on the coordinates of the tangent point and the tangent direction vector, the feature correction offset is calculated to dynamically calibrate local features and generate refined feature embeddings after distribution alignment. The model training and dense gradient upload module is used to start local model training with refined feature embedding after distribution alignment, calculate the loss after forward propagation, generate dense gradient parameters through back propagation, and upload the encrypted dense gradient parameters to the coordinating node. The gradient adaptive aggregation correction module is used to receive the encrypted gradient parameters uploaded by each participant, and after decryption or recombination, adaptively weight and correct the direction of each gradient contribution according to the preset inter-domain difference metric to generate the corrected aggregate gradient. The global model collaborative update module is used to coordinate nodes to update global model parameters based on the corrected aggregated gradient. By distributing the updated global model parameters in encrypted form back to each participant, each participant can update its local model until a cross-domain collaborative computing model resistant to domain offset is obtained.

Images