Data co-processing method and device and computer equipment

Abstract

The invention relates to a data co-processing method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring original data, encrypting the original data to obtain target encrypted data, and registering to a central server to receive a to-be-trained initial model corresponding to a task issued by the central server; training the initial model based on the target encrypted data to obtain a local gradient; performing homomorphic encryption processing on the sensitive information in the local gradient to obtain an encryption gradient, and uploading the encryption gradient to a central server, so that the central server performs global training based on the encryption gradient uploaded by each participating terminal to obtain a trained target global model; and receiving the target global model issued by the central server, and processing based on the target global model and the local client data of the participating terminals to obtain a task classification result.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data collaborative processing method, apparatus, and computer equipment. Background Technology

[0002] With the digital transformation of the insurance industry, various member units within the industry have accumulated massive amounts of customer data related to products. When multiple units collaborate, customer data is not shared. Generally, customer data is anonymized before it is used. Key information is not fully retained, resulting in lower accuracy of models trained based on anonymized data. Summary of the Invention

[0003] Therefore, it is necessary to provide a data collaborative processing method, apparatus, and computer equipment that can improve data utilization and training accuracy in response to the above-mentioned technical problems.

[0004] Firstly, this application provides a data collaborative processing method applied to participating terminals; the method includes:

[0005] The process involves acquiring raw data, encrypting the raw data to obtain target encrypted data, and registering with the central server to receive the initial model to be trained corresponding to the task issued by the central server.

[0006] The initial model is trained based on the target encrypted data to obtain the local gradient; the sensitive information in the local gradient is homomorphically encrypted to obtain the encrypted gradient, and the encrypted gradient is uploaded to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model.

[0007] The system receives the target global model sent by the central server and processes it based on the target global model and the local customer data of the participating terminals to obtain the task classification result.

[0008] In some embodiments, encrypting the original data to obtain target encrypted data includes:

[0009] Based on the original data, determine the label corresponding to the original data;

[0010] Based on the tags, the original data is format-normalized to generate globally unified data. The privacy fields in the globally unified data are then homomorphically encrypted, and the non-sensitive fields are differentially encrypted to obtain the target encrypted data.

[0011] In some embodiments, uploading the encrypted gradients to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain a trained target global model includes:

[0012] The encrypted gradient is uploaded to the central server so that the central server can generate an updated global model if it determines that the training completion conditions are not met.

[0013] The system receives an updated global model from the central server, updates the initial model based on the updated global model to obtain an updated initial model, and re-executes the step of training the initial model based on the target encrypted data to obtain local gradients based on the updated initial model, until the central server obtains a target global model that meets the training completion conditions.

[0014] In some embodiments, the processing of local customer data of the participating terminal based on the target global model to obtain task classification results includes:

[0015] The model parameters of the updated initial model are updated based on the target global model to obtain the target model. The local customer data of the participating terminal is processed by the target model to obtain the task classification result. The task classification result is used to determine the objects to be recommended.

[0016] In some embodiments, the method further includes:

[0017] Obtain incremental data and define the incremental encrypted data corresponding to the incremental data;

[0018] The incremental encryption data is processed by the target model to obtain an incremental encryption gradient, and the incremental encryption gradient is sent to the central server so that the central server updates the target global model based on each incremental encryption gradient to obtain an updated global model.

[0019] The system receives the updated global model returned by the central server and updates the target model based on the updated global model to obtain an updated target model. The updated target model is used to determine the task classification result corresponding to the local customer data containing the incremental data.

[0020] Secondly, this application provides a data collaborative processing method applied to a central server; the method includes:

[0021] The system receives registration requests from each participating terminal and sends an initial model to be trained corresponding to the task to each participating terminal, so that each participating terminal can train the initial model based on the target encrypted data to obtain the encryption gradient; the target encrypted data is obtained by the participating terminal after encrypting the collected raw data.

[0022] Global training is performed based on the encrypted gradients uploaded by each participating terminal to obtain a trained target global model; the target global model is then distributed to each participating terminal so that each participating terminal can process its local customer data to obtain task classification results.

[0023] In some embodiments, the step of performing global training based on the encrypted gradients uploaded by each of the participating terminals to obtain a trained target global model includes:

[0024] The decryption module configured on the central server decrypts each encryption gradient to obtain the decryption gradient;

[0025] If the training completion condition is not met, a weighted average is performed based on the data volume of each participating terminal to obtain an updated global model. The updated global model is then sent back to each participating terminal so that each participating terminal receives the updated initial model. The step of re-training the initial model based on the target encrypted data to obtain the encryption gradient is then performed. This continues until the central server determines that the training completion condition is met based on each decryption gradient, and a trained target global model is obtained.

[0026] In some embodiments, the method further includes:

[0027] The task loss data is calculated using the cross-entropy loss function, and the privacy protection loss data is calculated using the decryption gradient, i.e., the encryption gradient; and a weighted result is calculated based on the task loss data and the privacy protection loss data, the weighted result being the joint loss data;

[0028] If the weighted result is determined to be less than a preset threshold, then the training completion condition is determined to be met; or, if the weighted result is determined to be greater than or equal to the preset threshold, then the training completion condition is determined not to be met.

[0029] Thirdly, this application also provides a data collaborative processing apparatus for use in participating terminals, comprising:

[0030] The first acquisition module is used to acquire raw data, encrypt the raw data to obtain target encrypted data, and register with the central server to receive the initial model to be trained corresponding to the task issued by the central server.

[0031] The first training module is used to train the initial model based on the target encrypted data to obtain the local gradient; to perform homomorphic encryption on the sensitive information in the local gradient to obtain the encrypted gradient, and to upload the encrypted gradient to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model.

[0032] The first receiving module is used to receive the target global model sent by the central server, and process it based on the target global model and the local customer data of the participating terminal to obtain the task classification result.

[0033] Fourthly, this application also provides a data collaborative processing apparatus for use in a central server, comprising:

[0034] The second receiving module is used to receive registration requests from each participating terminal and send the initial model to be trained corresponding to the task to each participating terminal, so that each participating terminal can train the initial model based on the target encrypted data to obtain the encryption gradient; the target encrypted data is obtained by the participating terminal after encrypting the collected raw data.

[0035] The second training module is used to perform global training based on the encrypted gradients uploaded by each participating terminal to obtain a trained target global model; and to distribute the target global model to each participating terminal so that each participating terminal can process its local customer data to obtain task classification results.

[0036] Fifthly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0037] The process involves acquiring raw data, encrypting the raw data to obtain target encrypted data, and registering with the central server to receive the initial model to be trained corresponding to the task issued by the central server.

[0038] The initial model is trained based on the target encrypted data to obtain the local gradient; the sensitive information in the local gradient is homomorphically encrypted to obtain the encrypted gradient, and the encrypted gradient is uploaded to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model.

[0039] The system receives the target global model sent by the central server and processes it based on the target global model and the local customer data of the participating terminals to obtain the task classification result.

[0040] Sixthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0041] The process involves acquiring raw data, encrypting the raw data to obtain target encrypted data, and registering with the central server to receive the initial model to be trained corresponding to the task issued by the central server.

[0042] The initial model is trained based on the target encrypted data to obtain the local gradient; the sensitive information in the local gradient is homomorphically encrypted to obtain the encrypted gradient, and the encrypted gradient is uploaded to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model.

[0043] The system receives the target global model sent by the central server and processes it based on the target global model and the local customer data of the participating terminals to obtain the task classification result.

[0044] In a seventh aspect, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0045] The process involves acquiring raw data, encrypting the raw data to obtain target encrypted data, and registering with the central server to receive the initial model to be trained corresponding to the task issued by the central server.

[0046] The initial model is trained based on the target encrypted data to obtain the local gradient; the sensitive information in the local gradient is homomorphically encrypted to obtain the encrypted gradient, and the encrypted gradient is uploaded to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model.

[0047] The system receives the target global model sent by the central server and processes it based on the target global model and the local customer data of the participating terminals to obtain the task classification result.

[0048] The aforementioned data collaborative processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product include the following steps: acquiring raw data, encrypting the raw data to obtain target encrypted data, and registering with a central server to receive an initial model to be trained corresponding to a task issued by the central server; training the initial model based on the target encrypted data to obtain local gradients; homomorphically encrypting sensitive information in the local gradients to obtain encrypted gradients, and uploading the encrypted gradients to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain a trained target global model; receiving the target global model issued by the central server and processing it based on the target global model and the local client data of the participating terminals to obtain task classification results. By employing this method, privacy data in the raw data can be encrypted at the field level, ensuring that the raw data locally on the participating terminals is not shared with other devices, thereby achieving accurate training of the global model and improving model performance and task execution accuracy. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is an application environment diagram of the data collaborative processing method in one embodiment;

[0051] Figure 2 This is a flowchart illustrating a data collaborative processing method in one embodiment;

[0052] Figure 3 This is a flowchart illustrating the training steps in one embodiment;

[0053] Figure 4a This is a flowchart illustrating the data collaborative processing method in another embodiment;

[0054] Figure 4b This is a schematic diagram illustrating the interaction process between the life insurance node and the federal coordinator in one embodiment.

[0055] Figure 5 This is a structural block diagram of a data collaborative processing device in one embodiment;

[0056] Figure 6 This is a structural block diagram of a data collaborative processing device in one embodiment;

[0057] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0060] The data collaborative processing method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, multiple participating terminals 102 communicate with the central server 104 via a network. Each participating terminal 102 can obtain its own local raw data and process it to obtain the target encrypted data. Each participating terminal 102 can then register with the central server 104 and receive the initial model corresponding to the task returned by the central server 104. Based on this, the central server 104 can collaborate with the multiple participating terminals 102 to train the model, so that the central server 104 obtains a trained global target model, and each participating terminal 102 obtains a trained target model, resulting in a task classification result.

[0061] The participating terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, and projection equipment. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The central server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0062] In one exemplary embodiment, such as Figure 2As shown, a data collaborative processing method is provided, which can be applied to... Figure 1 Taking the participating terminals in the example, the explanation includes:

[0063] Step 202: Obtain the raw data and encrypt it to obtain the target encrypted data. Also, register with the central server to receive the initial model to be trained corresponding to the task issued by the central server.

[0064] The participating terminals can be multiple devices that need to collaboratively process data during the task, i.e., devices associated with the task. For example, in the case of identifying target users, the participating terminals could be various types of insurance companies, banks, etc., within the insurance industry; different insurance types can include property insurance, life insurance, etc. The raw data can be basic data collected locally by the participating devices. For example, if the participating terminal is a life insurance company, the corresponding raw data could be age, policy type, and annual premium; if the participating terminal is a property insurance company, the corresponding raw data could be vehicle value, number of car insurance uses, age, gender, and premium; if the participating terminal is a bank-level wealth management institution, the corresponding raw data could be investment preferences, etc. The central server can be a server that collaboratively processes and coordinates the encrypted gradients uploaded by each participating terminal; the initial model to be trained for the task can be a neural network model, machine learning model, etc., that has not yet started training and is used to process the task, such as an XGBoost classifier, etc.

[0065] Specifically, the data collaborative processing system may include a central server and multiple participating terminals. For each participating terminal, it can collect local raw data—that is, the raw data generated by the terminal during business processing. The participating terminal can then determine the tags corresponding to each piece of raw data and encrypt it based on these tags to obtain the target encrypted data. The participating terminals can then send registration requests to the central server. The central server can register the terminals based on the received registration requests and distribute the initial model to be trained for the task to the participating terminals. This completes the initialization of the data collaborative system's architecture.

[0066] Step 204: Train the initial model based on the target encrypted data to obtain the local gradient. Perform homomorphic encryption on the sensitive information in the local gradient to obtain the encrypted gradient, and upload the encrypted gradient to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model.

[0067] The local gradient is related to the specific type of the initial model to be trained. When the initial model is an XGBoost classifier, the corresponding local gradient can be the XGBoost tree structure update amount. Sensitive information can be information about sensitive fields, such as the gradient associated with the customer ID. Homomorphic encryption can be implemented through a preset homomorphic encryption algorithm, such as the Paillier homomorphic encryption algorithm. A federated coordinator can be configured in the central server.

[0068] Specifically, for each participating terminal, it can train the initial model to be trained based on the target encrypted data generated in the above steps, obtain the local gradient of the initial model, and perform Paillier homomorphic encryption on the sensitive information in the local gradient to generate an encrypted gradient. In this way, the participating terminal can upload the encrypted gradient to the central server. The central server can receive encrypted gradients sent by multiple participating terminals, and can decrypt each encrypted gradient. For example, the central server can decrypt the encrypted gradient using the private key in the federated coordinator to obtain decrypted data. Based on the decrypted data from each participating terminal, the central server trains the global model to obtain an updated global model. In this way, the central server can calculate the joint loss function and, if it determines that the joint loss function does not meet the training completion condition, send the updated global model back to each participating terminal. Each participating terminal can update its local initial model based on the received global update model to obtain the updated initial model. Based on the updated initial model, the step of training the initial model based on the target encrypted data in the above embodiment to obtain the local gradient is re-executed, that is, the joint training iteration of the initial model and the global model is performed until the central server determines that the training completion condition is met based on the joint loss function in the current training round. Then, the central server can determine the updated global model in the current training round as the trained target global model.

[0069] Step 206: Receive the target global model sent by the central server, and process it based on the target global model and the local customer data of the participating terminals to obtain the task classification result.

[0070] Among them, the local customer data of the participating terminal can be customer-related data that the participating terminal can directly collect. For example, it can be the financial preference data of customers that can be collected by participating nodes such as banks, or financial products with interactive behavior, such as click operation, browsing operation, and purchase operation.

[0071] Specifically, after obtaining the trained global target model, the central server can resend the trained global target model to each participating terminal. This allows each participating terminal to update the parameters of its newly updated initial model based on the global target model, generating a trained initial model, which is the target model corresponding to each participating terminal. Based on this, participating terminals can collect local customer data and process it using the target model. The output of the target model is then determined as the task classification result. For example, if the task is to determine whether a customer is a labeled customer, the corresponding task classification result is either that the customer in the local customer data is a labeled customer, or that the customer in the local customer data is not a labeled customer. Optionally, the task classification result can be a predicted probability / confidence that the customer in the local customer data is a labeled customer. The participating terminal can compare this predicted probability with a prediction threshold. If the predicted probability exceeds the prediction threshold, the participating terminal can determine that the customer in the local customer data is a labeled customer; if the predicted probability is less than or equal to the prediction threshold, the participating terminal can determine that the customer in the local customer data is not a labeled customer.

[0072] In the aforementioned data collaborative processing method, raw data is acquired and encrypted to obtain target encrypted data. The data is then registered with a central server to receive the initial model to be trained for the task assigned by the central server. The initial model is trained based on the target encrypted data to obtain local gradients. Sensitive information in the local gradients is homomorphically encrypted to obtain encrypted gradients, which are then uploaded to the central server. The central server then performs global training based on the encrypted gradients uploaded by each participating terminal to obtain a trained target global model. The target global model is received from the central server and processed based on the target global model and the local client data of the participating terminals to obtain the task classification result. By employing this method, privacy data in the raw data can be encrypted at the field level. This ensures that the raw data on the participating terminals is not shared with other devices, enabling accurate training of the global model and improving model performance and task execution accuracy.

[0073] In some embodiments, the specific processing steps of "encrypting the original data to obtain the target encrypted data" include:

[0074] Based on the original data, the corresponding labels are determined. The original data is then normalized based on the labels to generate globally unified data. Homomorphic encryption is applied to the privacy fields within the globally unified data, and differential privacy processing is applied to the non-sensitive fields to obtain the target encrypted data.

[0075] Specifically, participating terminals collect raw customer-related data locally. For this raw data, the participating terminals can determine tags for each data point, i.e., identifiers for each data point. These identifiers include general tags across participating terminals and unique tags specific to each participating terminal. For example, general tags may include age, gender, name, etc., while unique tags are features associated with the terminal functions or application functions of the participating terminal that collected the data. For instance, the unique tag for a life insurance participating terminal might be policy type, while the unique tag for a banking participating terminal might be financial preference data, etc. Based on this, after determining the tags corresponding to the raw data, the participating terminals can perform format conversion processing on the data corresponding to the general tags collected in the participating terminal, based on the general form of the general tags, to obtain converted raw data. The participating terminals can also perform normalization encoding or interval mapping processing on the raw data to obtain processed raw data. Based on this, the participating terminals can generate globally unified data from the converted and processed raw data.

[0076] In this way, participating terminals can determine whether each tag belongs to a privacy field / sensitive field, and perform Paillier homomorphic encryption on the original data under tags that belong to privacy fields, and perform differential privacy processing on the original data under tags that do not belong to privacy fields / sensitive fields. Thus, participating terminals can obtain the target encrypted data. The differential privacy processing refers to the process of processing the original data using the Laplace differential privacy perturbation algorithm, where the differential privacy parameter ε=1. Tags belonging to privacy fields can include ID numbers, card numbers, etc., while tags not belonging to privacy fields can include age, premium, investment product name, etc.

[0077] In this embodiment, by performing homomorphic encryption and differential privacy perturbation, the risk of privacy data leakage can be further reduced, and the storage and usage security of the collected data can be further improved.

[0078] In one embodiment, such as Figure 3 As shown, the specific processing steps of the step "uploading the encrypted gradients to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model" include:

[0079] Step 302: Upload the encrypted gradient to the central server so that the central server can generate an updated global model if it determines that the training completion conditions are not met.

[0080] Step 304: Receive the updated global model sent by the central server, update the initial model based on the updated global model to obtain the updated initial model, and re-execute the step of training the initial model based on the target encrypted data to obtain the local gradient, until the central server obtains the target global model that meets the training completion conditions.

[0081] The central server is equipped with a federated coordinator and a privacy protection module. The privacy protection module includes a homomorphic encryption engine controller and a differential privacy controller. Optionally, the differential privacy parameter ε in the differential privacy controller is 1.

[0082] Specifically, in the data collaboration system, each participating terminal related to the task can execute the steps in the above embodiments to obtain the encryption gradient corresponding to each participating terminal. In this way, the central server can receive encryption gradients sent by multiple participating terminals. The central server can decrypt each encryption gradient; for example, the central server can decrypt the encryption gradient using the private key in the federated coordinator to obtain decrypted data. Based on the decrypted data from each participating terminal, the global model in the central server is trained to obtain an updated global model. The central server can also calculate the joint loss function, and if it determines that the joint loss function does not meet the training completion conditions, it sends the updated global model back to each participating terminal. Each participating terminal can update its local initial model based on the received global update model to obtain an updated initial model. Based on this updated initial model, the step of training the initial model based on the target encrypted data in the above embodiments to obtain the local gradient is re-executed. That is, the participating terminal can perform the next round of iteration processing based on the updated initial model, i.e., the participating terminal can train the updated initial model based on the target encrypted data to obtain the local gradient. Sensitive information in the local gradient is homomorphically encrypted to obtain an encrypted gradient, which is then uploaded to the central server again. This allows the central server to receive encrypted gradients from multiple participating terminals in the current round, decrypt each encrypted gradient, and obtain decrypted data. Based on the decrypted data from each participating terminal, the updated global model in the central server is trained to obtain the updated global model for the current training round. The central server can then calculate the joint loss function. If the joint loss function satisfies the training completion condition, the central server can identify the updated global model in the current training round as the trained target global model, thus obtaining the target global model that meets the training completion condition.

[0083] In this embodiment, a privacy-enhanced data collaboration system is used to enable joint training and iteration of local models of multiple participating terminals and global models in the central server, thereby improving the training efficiency and accuracy of the models.

[0084] In one embodiment, the specific processing steps for the step "processing the local customer data of participating terminals based on the target global model to obtain task classification results" are as follows:

[0085] The model parameters of the updated initial model are updated based on the target global model to obtain the target model. The local customer data of the participating terminals is then processed using the target model to obtain the task classification results. The task classification results are used to determine the objects to be recommended.

[0086] The objects to be recommended can be exclusive permissions corresponding to tagged customers or exclusive recommendation information / data.

[0087] Specifically, after obtaining the trained global target model, the central server can resend the trained global target model to each participating terminal. Each participating terminal updates the model parameters of the updated initial model in the current training round based on the global target model, generating a trained initial model, which is the target model corresponding to each participating terminal. Based on this, the participating terminal can collect local customer data and process the customer data through the target model to determine the output of the target model as the task classification result. For example, if the task is to determine whether a customer is a labeled customer, the corresponding task classification result is that the customer corresponding to the local customer data is a labeled customer, or the customer corresponding to the local customer data is not a labeled customer. Optionally, the task classification result can be the predicted probability / confidence that the customer corresponding to the local customer data is a labeled customer. The participating terminal can compare the predicted probability with a prediction threshold. If the predicted probability exceeds the prediction threshold, the participating terminal can determine that the customer corresponding to the local customer data is a labeled customer; if the predicted probability is less than or equal to the prediction threshold, the participating terminal can determine that the customer corresponding to the local customer data is not a labeled customer.

[0088] In this way, the participating terminal can, upon determining that the target customer corresponding to the local customer data is a tagged customer, push the pre-defined recommendation data corresponding to the tagged customer to that target customer. Alternatively, the participating terminal can, upon determining that the target customer is a tagged customer, grant that target customer exclusive permissions based on pre-defined exclusive permissions or exclusive recommendation information for the tagged customer, and push the exclusive recommendation information to that target customer. For example, exclusive permissions could be lifetime upgrade benefits in a life insurance participating terminal; or high-level service benefits in a bank participating terminal, etc., and the corresponding recommendation data could be high-level wealth management products, etc.

[0089] In this embodiment, by jointly training the global model and the local model of the participating terminal, and by directly converting the model output results into task classification results, the model performance can be significantly improved, and the ability of the model to connect with business can also be enhanced.

[0090] In one embodiment, the data collaborative processing method further includes:

[0091] Acquire incremental data and determine the corresponding incremental encrypted data. Process the incremental encrypted data using the target model to obtain the incremental encryption gradient, and send the incremental encryption gradient to the central server. The central server then updates the target global model based on each incremental encryption gradient to obtain the updated global model.

[0092] The system receives the updated global model returned by the central server and updates the target model based on the updated global model to obtain the updated target model. The updated target model is used to determine the task classification result corresponding to the local customer data containing incremental data.

[0093] Incremental data can be either the original data that has changed or an indicator of the original data that has changed. For example, the indicator can be a data distribution indicator, a real-time data change indicator, KL divergence, a population stability index, etc.

[0094] Specifically, for any participating terminal, the participating terminal can monitor the local raw data. If it is determined that the current raw data has changed, the participating terminal can obtain the changed data, that is, collect incremental data, identify the label of the incremental data, determine the label of the incremental data, and perform format normalization processing on the incremental data based on the label to generate globally unified data. Then, homomorphic encryption processing is performed on the privacy field data in the globally unified data, and differential privacy processing is performed on the non-sensitive field data to obtain incremental encrypted data.

[0095] Participating terminals can train the target model to obtain incremental local gradients, and perform homomorphic encryption on sensitive information in these incremental local gradients to obtain incremental encrypted gradients. These encrypted gradients are then uploaded to the central server. The central server uses the encrypted gradients uploaded by the participating terminal to incrementally train the target global model, obtaining an updated global model. The updated global model is then distributed to the participating terminal that uploaded the encrypted gradients, enabling the participating terminal to incrementally update its target model based on the received updated global model, obtaining an updated target model. Based on this updated target model, the participating terminal can determine and process the local customer data corresponding to the incremental data, obtaining the task classification result for the customer corresponding to the incremental data.

[0096] In this embodiment, when the original data changes, the changed data can be feature aligned and encrypted, thereby enabling the updating of the local model and the global model based on incremental data, as well as the adaptation to the latest business data. This reduces the model training cost while improving the model's adaptability to changing scenarios.

[0097] In one exemplary embodiment, such as Figure 4a As shown, a data collaborative processing method is provided, which can be applied to... Figure 1 Taking the central server in the middle as an example, the explanation includes:

[0098] Step 402: Receive registration requests from each participating terminal and distribute the initial model to be trained corresponding to the task to each participating terminal, so that each participating terminal can train the initial model based on the target encrypted data to obtain the encryption gradient. The target encrypted data is obtained by the participating terminals after encrypting the collected raw data.

[0099] Step 404: Perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model. Then, distribute the target global model to each participating terminal so that each terminal can process its local customer data and obtain the task classification result.

[0100] Specifically, for each participating terminal, it can train the initial model to be trained based on the target encrypted data generated in the above steps, obtain the local gradient of the initial model, and perform Paillier homomorphic encryption on the sensitive information in the local gradient to generate an encrypted gradient. In this way, the participating terminal can upload the encrypted gradient to the central server. The central server can receive encrypted gradients sent by multiple participating terminals, and can decrypt each encrypted gradient. For example, the central server can decrypt the encrypted gradient using the private key in the federated coordinator to obtain decrypted data. Based on the decrypted data from each participating terminal, the central server trains the global model to obtain an updated global model. In this way, the central server can calculate the joint loss function and, if it determines that the joint loss function does not meet the training completion condition, send the updated global model back to each participating terminal. Each participating terminal can update its local initial model based on the received global update model to obtain the updated initial model. Based on the updated initial model, the step of training the initial model based on the target encrypted data in the above embodiment to obtain the local gradient is re-executed, that is, the joint training iteration of the initial model and the global model is performed until the central server determines that the training completion condition is met based on the joint loss function in the current training round. Then, the central server can determine the updated global model in the current training round as the trained target global model.

[0101] Specifically, after obtaining the trained global target model, the central server can resend this model to each participating terminal. Each terminal then updates its initial model parameters based on this new global target model, generating its own trained initial model – thus obtaining the target model for each terminal. Based on this, each participating terminal can collect its local customer data, process it using the target model, and determine the output of the target model as the task classification result.

[0102] In this embodiment, privacy data in the original data can be encrypted at the field level. This ensures that the original data on the participating terminal is not shared with other devices, thereby enabling accurate training of the global model and improving model performance and task execution accuracy.

[0103] In one embodiment, the specific processing steps of "performing global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model" include:

[0104] The decryption module configured on the central server decrypts each encryption gradient to obtain the decrypted gradient. If the training completion condition is not met, a weighted average is performed based on the amount of data from each participating terminal to obtain an updated global model. This updated global model is then sent back to each participating terminal so that they receive the updated initial model and re-execute the step of training the initial model based on the target encrypted data to obtain the encryption gradient. This process continues until the central server determines that the training completion condition is met based on the decrypted gradients, resulting in a trained target global model.

[0105] The central server is configured with a decryption module that can be a homomorphic encryption engine controller and a differential privacy controller. The central server is also configured with a federated coordinator.

[0106] Specifically, the central server can receive encrypted gradients sent by each participating terminal and decrypt each encrypted gradient through the homomorphic encryption engine controller and the differential privacy controller to obtain decrypted data; alternatively, the central server can also decrypt the encrypted gradient using the private key in the federated coordinator to obtain decrypted data. Based on the decrypted data from each participating terminal, the global model in the central server is trained to obtain an updated global model. The central server can also calculate the joint loss function, and if it determines that the joint loss function does not meet the training completion conditions, it sends the updated global model back to each participating terminal. Each participating terminal can update its local initial model based on the received global update model to obtain an updated initial model. Based on the updated initial model, the step of training the initial model based on the target encrypted data in the above embodiment to obtain the local gradient is re-executed. That is, the participating terminal can perform the next round of iteration processing based on the updated initial model, i.e., the participating terminal can train the updated initial model based on the target encrypted data to obtain the local gradient. Sensitive information in the local gradient is homomorphically encrypted to obtain an encrypted gradient, which is then uploaded to the central server again. This allows the central server to receive encrypted gradients from multiple participating terminals in the current round, decrypt each encrypted gradient, and obtain decrypted data. Based on the decrypted data from each participating terminal, the updated global model in the central server is trained to obtain the updated global model for the current training round. The central server can then calculate the joint loss function. If the joint loss function satisfies the training completion condition, the central server can identify the updated global model in the current training round as the trained target global model, thus obtaining the target global model that meets the training completion condition.

[0107] In one example, the step "performing a weighted average based on the data volume of each participating terminal to obtain an updated global model" can be as follows: After the central server obtains the encryption gradients corresponding to the encryption gradients sent by each participating terminal, it can determine the data volume of each participating terminal based on these encryption gradients, and then perform a weighted average on the data volume of each participating terminal. This weighted average is used as the model parameters in the global model, and the updated global model is obtained based on the weighted average of the data volume of each participating terminal. The weight of the data volume of each participating terminal can be determined based on the ratio between the number of customers of that participating terminal and the total number of customers.

[0108] In this embodiment, a privacy-enhanced data collaboration system is used to enable joint training and iteration of local models of multiple participating terminals and global models in the central server, thereby improving the training efficiency and accuracy of the models.

[0109] In one embodiment, the data collaborative processing method further includes:

[0110] The task loss data is calculated using the cross-entropy loss function, and the privacy-preserving loss data is calculated using the decrypted gradient (i.e., encrypted gradient). A weighted result is then calculated based on the task loss data and the privacy-preserving loss data.

[0111] If the weighted result is less than a preset threshold, then the training completion condition is met. Alternatively, if the weighted result is greater than or equal to the preset threshold, then the training completion condition is not met.

[0112] The cross-entropy loss function is used to calculate the cross-entropy loss of the task, such as the cross-entropy loss data of the task clustering results, to measure the accuracy of task clustering. The privacy-preserving loss data can be the penalty data for abnormal gradients calculated by simulating gradient leakage attacks through a generative adversarial network (GAN), and this penalty data is identified as the privacy-preserving loss data. The preset threshold can be the minimum threshold of the loss value corresponding to a pre-configured loss function. This embodiment does not limit the specific value of the preset threshold, which can be determined by those skilled in the art based on the actual application scenario. For example, the specific value can be 0.05.

[0113] Specifically, the central server can determine the predicted clustering result based on the sample clustering result corresponding to the original data and the output result of the global model. It then calculates the task loss data by using the cross-entropy loss function to obtain the sample clustering result and the predicted clustering result. Correspondingly, the central server can use a pre-trained generator in the GAN network to generate gradients based on random noise, and a discriminator to judge the authenticity of the generated gradient, outputting the probability value that the generated gradient is real noise. The GAN is pre-trained using sample gradients and their real labels to obtain a trained GAN. The generator in the trained GAN is used to generate a gradient by decrypting the gradient (e.g., random noise). The discriminator judges the authenticity of the generated gradient and outputs the real probability value corresponding to the output gradient. Based on the generated gradient and the decryption gradient, the adversarial loss data of the GAN is calculated, and the privacy-preserving loss data is calculated based on this adversarial loss data and a preset penalty coefficient. That is, the product between the adversarial loss data and the preset penalty coefficient is determined as the privacy-preserving loss data.

[0114] The central server can perform a weighted calculation on the task loss data and the privacy protection loss data. The weight of the privacy protection loss data can be a balancing coefficient, which can be determined based on the importance of privacy protection or the actual application scenario. If the weighted result is greater than or equal to a preset threshold, the training completion condition is determined not to be met. If the weighted result is less than the preset threshold, the training completion condition is determined to be met.

[0115] In this embodiment, the loss data of model training is accurately calculated by using a joint loss function of multiple tasks, which can further improve model accuracy and thus improve model training efficiency.

[0116] For example, the data collaborative processing method provided in this embodiment is a cross-institutional insurance data privacy collaborative analysis method and federated learning system. It can jointly construct local and global models tailored to the characteristics of the insurance industry. By homomorphically encrypting sensitive fields and differentially protecting non-sensitive fields, it achieves field-level encryption, preventing the leakage of privacy data at participating nodes and further enhancing data security. It also addresses privacy issues; achieves vertical data alignment through heterogeneous feature standardization (such as binning mapping and normalization); improves model accuracy through multi-task loss functions; and enables dynamic training of incremental data, reducing training costs and adapting to dynamically changing business data. This results in effects such as original data remaining within its domain, high-precision model training, and business results output by the model that are directly understandable to participating nodes. The system in this embodiment uses federated learning as a layered architecture framework. Through data preprocessing, privacy architecture construction, multi-task modeling, business implementation, and dynamic updates, it protects privacy data during cross-institutional collaborative processing in the insurance industry. The system comprises a central server and multiple participating nodes, including: a data preprocessing module (deployed on each participating node) for feature sorting, heterogeneous feature alignment, and field-level encryption of local data from participating nodes / terminals; a federated coordinator (deployed on the central server) responsible for task scheduling, privacy policy configuration, and parameter aggregation; a local training module (deployed on each participating node) for executing local model training and uploading encrypted gradients; a model output module (deployed on the central server) for converting the global model into business results and supporting API calls; and a security audit module (deployed on the central server) for recording operation logs and verifying privacy compliance.

[0117] The following describes a specific application scenario in a concrete embodiment. The classification task can be a collaborative effort between life insurance institutions, property insurance institutions, and banks to differentiate tagged customers. These tagged customers can be pre-defined with distinguishing criteria, such as high-level customers. These criteria could include holding more than a first threshold of products, annual premiums exceeding a second threshold, and bank assets exceeding a third threshold, etc. The specific implementation steps of the above data collaborative processing method are described in detail below with reference to this concrete embodiment:

[0118] Step 1: Data preprocessing. Each participating node provides local raw data. For example, local raw data for life insurance may include age, policy type, and annual premium; local raw data for property insurance may include vehicle value and number of car insurance claims; and for banks, it may include bank assets and investment preferences, etc. Specifically, each participating node determines the characteristics of its local raw data and jointly defines common labels (such as age) with other participating nodes, as well as institution-specific labels for each participating node (such as policy type in life insurance). Features with format differences (such as life insurance "age" being a numerical value) are also addressed. The system generates a global feature space by binning data (e.g., dividing numerical ages into "18-30", "31-50", and "51+") or normalizing data (e.g., scaling deposit balances to the 0-1 range). Sensitive fields (e.g., "ID number" and "bank card number") are stored using Paillier homomorphic encryption. Non-sensitive fields (e.g., "age" and "premium") are perturbed using Laplace differential privacy perturbation (ε=1, to control privacy leakage risk). Based on this, target encrypted data is generated, and a registration request is sent to the federated coordinator for registration.

[0119] Step 2: Establish a privacy-enhanced federated architecture. The central server deploys a privacy protection module, which includes a homomorphic encryption engine, a differential privacy controller, and is configured with an encryption algorithm (Paillier) and differential privacy parameters (ε=1). Specifically, participating nodes (life insurance, property insurance, and banking) on ​​the central server register with the federated coordinator and confirm the task type. Based on this, the initialization of the system structure can be achieved.

[0120] Step 3, Multi-task Joint Modeling and Training: An improved FedAvg algorithm is used to introduce a multi-objective loss function (main task loss + privacy protection loss), and parameter aggregation with "data not leaving the domain" is achieved through encrypted gradients. Specifically, during task initialization, the federated coordinator distributes the initial model to each participating node; local training (each participating node) trains the model using local encrypted data, calculates local gradients (e.g., the tree structure update amount in XGBoost), and re-encrypts sensitive information in the gradients (e.g., gradients associated with customer IDs) using Paillier homomorphic encryption before uploading to the federated coordinator. The federated coordinator decrypts the encrypted gradients using its private key, performs a weighted average based on data volume (weight = number of institutional customers / total number of customers), and generates a global model update; the updated global model is then sent back to each participating node so that each node can update its local model. Steps 2-3 are repeated until the model loss (main task loss + privacy protection loss) is below a threshold (e.g., 0.05).

[0121] Among them, the model loss L total It can be calculated using the following formula:

[0122] ;

[0123] L task The primary task loss is the cross-entropy loss from customer segmentation, which measures the accuracy of segmentation; L privacy The privacy-preserving loss is calculated by simulating gradient leakage attacks using a generative adversarial network (GAN) and calculating a penalty term for abnormal gradients. λ is a balance coefficient that can be adjusted according to the scenario, and can optionally be 0.1. Under the condition that the training is completed, a convergent global clustering model stored in the federated coordinator is obtained.

[0124] The model output module transforms the global model into a business-understandable result. Each participating node inputs local customer data, and the model outputs a judgment result on whether the customer is a tagged customer (e.g., yes / no). Based on the tags, each institution generates business actions: life insurance companies push target benefits, property insurance companies recommend target products, banks offer higher-level benefits, and so on. Each participating node can obtain each tagged customer and implement corresponding unique services for each tagged customer.

[0125] In addition, this system can achieve dynamic incremental training when conditions for incremental training are triggered. When the data distribution of an institution changes (such as the addition of new pension insurance customers, a change in customer balance exceeding 20%, or the detection of data distribution deviation by the security audit module (KL divergence > 0.1, or a shift in the group stability index PSI, etc.); participating nodes only extract the incremental data of the changed portion, perform feature alignment and encryption; the local training module updates the local model with the incremental data and uploads the incremental encrypted gradient; the federated coordinator aggregates the incremental gradient and updates the global model (without retraining the full dataset); the updated global model is obtained and adapted to the latest business data.

[0126] In one example, this could be an architecture diagram of a data collaborative processing system provided in this embodiment, including a security audit module, a central server (configured with a federated coordinator), and multiple participating nodes (participating terminals). Each participating node could include life insurance nodes, property insurance nodes, and bank nodes, etc. The federated coordinator includes a task scheduling module, a local training module, a model aggregation module, and a business output module. The specific execution process of this system could include:

[0127] Each participating node obtains standardized encrypted data. The central server distributes the initial model B and task schedule to each participating node. The participating nodes perform local training through the local training module to obtain encrypted gradients, and then upload the encrypted gradients to the federated coordinator. The federated coordinator processes the encrypted gradients of each participating node through the model aggregation module, generates an updated global model, and distributes the global model. Based on this, the participating nodes can obtain the task processing results through the business output module.

[0128] Life insurance features can include numerical age, which is binned and mapped to a global feature space, resulting in a unified feature representation. For example, numerical age can be mapped to multiple age ranges: 18-30, 31-50, and 50+. Similarly, policy type can also be binned and mapped to a global feature space, resulting in a unified feature representation. Examples include health insurance, property insurance, and life insurance. For bank features in the bank participation node, these can include numerical deposits and investment preferences. Numerical deposits are normalized to generate a 0-1 range. Investment preferences can also be binned and mapped to a global feature space, resulting in a unified feature representation. Examples include conservative, aggressive, and balanced preferences.

[0129] like Figure 4b As shown, taking a life insurance node as an example, the interaction process between the life insurance node and the federated coordinator is described in detail:

[0130] Life insurance nodes upload encrypted data to the federated coordinator, which then distributes an initial model to the life insurance nodes. The life insurance nodes train their local models based on the initial model, obtaining encrypted gradients. Life insurance nodes can also upload encrypted gradients to the federated coordinator. The federated coordinator receives encrypted gradients from each participating node, decrypts and aggregates them to obtain a global model. The federated coordinator then distributes the global model to the life insurance nodes. Optionally, during incremental training, life insurance nodes can upload incremental encrypted gradients to the federated coordinator. The federated coordinator can update the global model based on these encrypted gradients, obtaining an updated global model, which is then distributed to the participating nodes.

[0131] Optionally, the fixed weight λ in the multi-task loss weight can be replaced with dynamic weights (e.g., adjusted according to the training stage, initially focusing on the main task and later on privacy protection); Federated Learning (FL) refers to distributed machine learning technology that allows multiple participants to jointly train a global model without exchanging original data; Vertical Federated Learning represents a type of federated learning where there is a lot of overlap in the data features of the participants and little overlap in the samples; Homomorphic Encryption (HE): an encryption method that allows addition, multiplication, and other operations on encrypted data, and the correct result can be obtained without decryption (e.g., Paillier encryption); Differential Privacy (DP): a technique that protects the privacy of individual data by adding controllable noise to the data (ε is the privacy budget, the smaller the value, the stronger the privacy protection); Global Feature Space: a unified feature representation generated by all participating institutions after feature alignment and standardization (e.g., "age" is uniformly binned); Incremental Training: a training method that updates the model based on an existing global model using only newly added or changed local data, avoiding full retraining.

[0132] The data collaborative processing method provided in this embodiment uses homomorphic encryption for sensitive fields and differential privacy for non-sensitive fields to achieve enhanced privacy protection through field-level encryption, ensuring that the original data does not leave the domain. Experiments show that the risk of original data leakage is ≤0.01%. The heterogeneous feature standardization method aligns the data features of different institutions into a global space, improving collaborative modeling efficiency by 50%. The model output is directly converted into business results such as a "high-value customer list," and the AUC (area under the curve) of the joint modeling reaches 0.92, which is significantly improved compared to single-institution modeling (life insurance 0.85, property insurance 0.81, and banking 0.79). Incremental training only requires uploading the gradient of the changed data, reducing the cost of repeated training by 70%.

[0133] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0134] Based on the same inventive concept, this application also provides a data collaborative processing apparatus for implementing the data collaborative processing method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more data collaborative processing apparatus embodiments provided below can be found in the limitations of the data collaborative processing method described above, and will not be repeated here.

[0135] In one exemplary embodiment, such as Figure 5 As shown, a data collaborative processing device 500 is provided, applied to participating terminals, including:

[0136] The first acquisition module 502 is used to acquire raw data, encrypt the raw data to obtain target encrypted data, and register with the central server to receive the initial model to be trained corresponding to the task issued by the central server.

[0137] The first training module 504 is used to train the initial model based on the target encrypted data to obtain the local gradient; to perform homomorphic encryption on the sensitive information in the local gradient to obtain the encrypted gradient, and to upload the encrypted gradient to the central server so that the central server can perform global training based on the encrypted gradient uploaded by each participating terminal to obtain the trained target global model.

[0138] The first receiving module 506 is used to receive the target global model sent by the central server, and process it based on the target global model and the local customer data of the participating terminals to obtain the task classification result.

[0139] In some embodiments, the first acquisition module is specifically used for:

[0140] Based on the original data, determine the label corresponding to the original data;

[0141] Based on the tags, the original data is format-normalized to generate globally unified data. The privacy fields in the globally unified data are then homomorphically encrypted, and the non-sensitive fields are differentially encrypted to obtain the target encrypted data.

[0142] In some embodiments, the first training module is specifically used for:

[0143] The encrypted gradient is uploaded to the central server so that the central server can generate an updated global model if it determines that the training completion conditions are not met.

[0144] The system receives an updated global model from the central server, updates the initial model based on the updated global model to obtain an updated initial model, and re-executes the step of training the initial model based on the target encrypted data to obtain local gradients based on the updated initial model, until the central server obtains a target global model that meets the training completion conditions.

[0145] In some embodiments, the first receiving module is specifically used for:

[0146] The model parameters of the updated initial model are updated based on the target global model to obtain the target model. The local customer data of the participating terminal is processed by the target model to obtain the task classification result. The task classification result is used to determine the objects to be recommended.

[0147] In some embodiments, the apparatus further includes:

[0148] The second acquisition module is used to acquire incremental data and determine the incremental encrypted data corresponding to the incremental data;

[0149] The first processing module is used to process the incremental encrypted data through the target model to obtain the incremental encryption gradient, and send the incremental encryption gradient to the central server so that the central server updates the target global model based on each of the incremental encryption gradients to obtain the updated global model.

[0150] The third receiving module is used to receive the updated global model returned by the central server, and update the target model based on the updated global model to obtain an updated target model. The updated target model is used to determine the task classification result corresponding to the local customer data containing the incremental data.

[0151] In one exemplary embodiment, such as Figure 6 As shown, a data collaborative processing device 600 is provided, applied to a central server, comprising:

[0152] The second receiving module 602 is used to receive registration requests from each participating terminal and send the initial model to be trained corresponding to the task to each participating terminal, so that each participating terminal can train the initial model based on the target encrypted data to obtain the encryption gradient; the target encrypted data is obtained by the participating terminal after encrypting the collected raw data.

[0153] The second training module 604 is used to perform global training based on the encrypted gradients uploaded by each participating terminal to obtain a trained target global model; and to distribute the target global model to each participating terminal so that each participating terminal can process its local customer data to obtain task classification results.

[0154] In some embodiments, the second training module is specifically used for:

[0155] The decryption module configured on the central server decrypts each encryption gradient to obtain the decryption gradient;

[0156] If the training completion condition is not met, a weighted average is performed based on the data volume of each participating terminal to obtain an updated global model. The updated global model is then sent back to each participating terminal so that each participating terminal receives the updated initial model. The step of re-training the initial model based on the target encrypted data to obtain the encryption gradient is then performed. This continues until the central server determines that the training completion condition is met based on each decryption gradient, and a trained target global model is obtained.

[0157] In some embodiments, the apparatus further includes:

[0158] The first calculation module is used to calculate task loss data through the cross-entropy loss function, and to calculate privacy protection loss data through the decryption gradient, i.e. the encryption gradient.

[0159] The first determining module is configured to determine that the training completion condition is met if the sum of the task loss data and the privacy protection loss data is less than a preset threshold; or,

[0160] The second determining module is used to determine that the training completion condition is not met if the sum of the task loss data and the privacy protection loss data is greater than or equal to a preset threshold.

[0161] Each module in the aforementioned data collaborative processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0162] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores training data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a data collaborative processing method.

[0163] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0164] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0165] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0166] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0167] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0168] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0169] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0170] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A data collaborative processing method, characterized in that, Applied to participating terminals; the method includes: The process involves acquiring raw data, encrypting the raw data to obtain target encrypted data, and registering with the central server to receive the initial model to be trained corresponding to the task issued by the central server. The initial model is trained based on the target encrypted data to obtain the local gradient; the sensitive information in the local gradient is homomorphically encrypted to obtain the encrypted gradient, and the encrypted gradient is uploaded to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model. The system receives the target global model sent by the central server and processes it based on the target global model and the local customer data of the participating terminals to obtain the task classification result.

2. The method according to claim 1, characterized in that, The step of encrypting the original data to obtain the target encrypted data includes: Based on the original data, determine the label corresponding to the original data; Based on the tags, the original data is format-normalized to generate globally unified data. The privacy fields in the globally unified data are then homomorphically encrypted, and the non-sensitive fields are differentially encrypted to obtain the target encrypted data.

3. The method according to claim 1, characterized in that, The step of uploading the encrypted gradient to the central server, so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal, to obtain a trained target global model, includes: The encrypted gradient is uploaded to the central server so that the central server can generate an updated global model if it determines that the training completion conditions are not met. The system receives an updated global model from the central server, updates the initial model based on the updated global model to obtain an updated initial model, and re-executes the step of training the initial model based on the target encrypted data to obtain local gradients based on the updated initial model, until the central server obtains a target global model that meets the training completion conditions.

4. The method according to claim 3, characterized in that, The process of processing the local customer data of the participating terminals based on the target global model to obtain task classification results includes: The model parameters of the updated initial model are updated based on the target global model to obtain the target model. The local customer data of the participating terminal is processed by the target model to obtain the task classification result. The task classification result is used to determine the objects to be recommended.

5. The method according to claim 4, characterized in that, The method further includes: Obtain incremental data and define the incremental encrypted data corresponding to the incremental data; The incremental encryption data is processed by the target model to obtain an incremental encryption gradient, and the incremental encryption gradient is sent to the central server so that the central server updates the target global model based on each incremental encryption gradient to obtain an updated global model. The system receives the updated global model returned by the central server and updates the target model based on the updated global model to obtain an updated target model. The updated target model is used to determine the task classification result corresponding to the local customer data containing the incremental data.

6. A data collaborative processing method, characterized in that, Applied to a central server; the method includes: The system receives registration requests from each participating terminal and sends an initial model to be trained corresponding to the task to each participating terminal, so that each participating terminal can train the initial model based on the target encrypted data to obtain the encryption gradient; the target encrypted data is obtained by the participating terminal after encrypting the collected raw data. Global training is performed based on the encrypted gradients uploaded by each participating terminal to obtain a trained target global model; the target global model is then distributed to each participating terminal so that each participating terminal can process its local customer data to obtain task classification results.

7. The method according to claim 6, characterized in that, The step of performing global training based on the encrypted gradients uploaded by each of the participating terminals to obtain the trained target global model includes: The decryption module configured on the central server decrypts each encryption gradient to obtain the decryption gradient; If the training completion condition is not met, a weighted average is performed based on the data volume of each participating terminal to obtain an updated global model. The updated global model is then sent back to each participating terminal so that each participating terminal receives the updated initial model. The step of re-training the initial model based on the target encrypted data to obtain the encryption gradient is then performed. This continues until the central server determines that the training completion condition is met based on each decryption gradient, and a trained target global model is obtained.

8. The method according to claim 7, characterized in that, The method further includes: The task loss data is calculated using the cross-entropy loss function, and the privacy protection loss data is calculated using the decryption gradient, i.e., the encryption gradient; and a weighted result is calculated based on the task loss data and the privacy protection loss data, the weighted result being the joint loss data; If the weighted result is determined to be less than a preset threshold, then the training completion condition is determined to be met; or, if the weighted result is determined to be greater than or equal to the preset threshold, then the training completion condition is determined not to be met.

9. A data collaborative processing device, characterized in that, The device is applied to a participating terminal and includes: The first acquisition module is used to acquire raw data, encrypt the raw data to obtain target encrypted data, and register with the central server to receive the initial model to be trained corresponding to the task issued by the central server. The first training module is used to train the initial model based on the target encrypted data to obtain the local gradient; to perform homomorphic encryption on the sensitive information in the local gradient to obtain the encrypted gradient, and to upload the encrypted gradient to the central server so that the central server can perform global training based on the encrypted gradients uploaded by each participating terminal to obtain the trained target global model. The first receiving module is used to receive the target global model sent by the central server, and process it based on the target global model and the local customer data of the participating terminal to obtain the task classification result.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images