Public service model training system based on split federated learning

Abstract

The embodiment of the invention provides a public service model training system based on split federated learning, which comprises the following steps: a plurality of service terminals generate question data features of local service question data by using an embedded layer, obtain question coding features fed back by the service terminals, input the question coding features to a head layer for processing, obtain predicted answer data, and send the predicted answer data to the head layer; updating a head layer based on the target answer data and the predicted answer data, sending head layer parameters to an aggregation server, and sending a head layer gradient to the server; the server updates the coding layer according to the head layer gradient, and feeds back the coding layer gradient to the plurality of service terminals; the plurality of service terminals construct a privacy gradient based on the coding layer gradient, update the embedded layer by using the privacy gradient, and send embedded layer parameters to the aggregation server; and the aggregation server optimizes the public service model based on the head layer parameter and the embedded layer parameter sent by each service terminal to obtain a target public service model.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of artificial intelligence technology, and in particular to a public service model training system based on split federated learning. Background Art

[0002] With the advancement of computing and internet technologies, the amount of data related to public services has expanded dramatically across various sectors, increasing the intensity and complexity of the work of relevant departments. Efficiently developing large public service models while protecting data privacy is crucial. Traditional federated learning requires clients to share complete model parameters, exposing the model architecture. Gradient interactions also pose the risk of leaking raw data. Public service departments have limited terminal computing power, making it difficult to train large models with full parameters. Furthermore, the frequent global parameter transmission required by traditional federated learning leads to a significant increase in communication overhead. Therefore, an effective solution is urgently needed to address these issues. Summary of the Invention

[0003] In view of this, embodiments of this specification provide a public service model training system based on split federated learning. One or more embodiments of this specification also involve a data processing method, a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product to address technical deficiencies in the prior art.

[0004] According to a first aspect of an embodiment of this specification, a public service model training system based on split federated learning is provided, comprising a plurality of service terminals, a server, and an aggregation server, wherein the plurality of service terminals deploy a header layer and an embedding layer of a public service model, and the server deploys an encoding layer of the public service model, including: The multiple service terminals generate question data features of local service question data using the embedding layer, obtain question coding features fed back by the server after processing the question data features using the coding layer, input the question coding features into the header layer for processing to obtain predicted answer data, update the header layer based on the target answer data of the local service question data and the predicted answer data, send header layer parameters of the updated header layer to the aggregation server, and send the header layer gradient of the updated header layer to the server; The server updates the coding layer according to the header layer gradient sent by each service terminal, and feeds back the updated coding layer gradient to the multiple service terminals; The multiple service terminals construct a privacy gradient based on the coding layer gradient, update the embedding layer using the privacy gradient, and send the embedding layer parameters of the updated embedding layer to the aggregation server; The aggregation server optimizes the public service model based on the header layer parameters and embedding layer parameters sent by each service terminal until a target public service model that meets the training stop condition is obtained.

[0005] Optionally, the server initializes the public service model and splits the initialized public service model to obtain the header layer, the embedding layer and the coding layer, and sends the first model parameters corresponding to the header layer and the embedding layer to each service terminal for deploying the header layer and the embedding layer at each service terminal.

[0006] Optionally, the multiple service terminals process the local service problem data using the embedding layer to obtain initial problem data features, add noise features to the initial problem data features to obtain problem data features, and send the problem data features to the service end; The server processes the problem data features using the coding layer to obtain problem coding features, and feeds back the problem coding features to the multiple service terminals.

[0007] Optionally, the multiple service terminals determine the target answer data corresponding to the local service question data, calculate the loss value based on the target answer data and the predicted answer data, update the head layer according to the loss value, and obtain the head layer parameters and head layer gradient.

[0008] Optionally, the multiple service terminals receive the coding layer gradient fed back by the server, clip the coding layer gradient to obtain an intermediate coding layer gradient, and add standard noise to the intermediate coding layer gradient to obtain a privacy gradient.

[0009] Optionally, the aggregation server receives header layer parameters and embedding layer parameters sent by each service terminal, performs weighted average calculation on the header layer parameters and embedding layer parameters sent by each service terminal, obtains global header layer parameters and global embedding layer parameters, and optimizes the public service model based on the global header layer parameters and the global embedding layer parameters until a target public service model that meets the training stop conditions is obtained.

[0010] Optionally, the aggregation server sends the global header layer parameters and the global embedding layer parameters to the multiple service terminals; The multiple service terminals update the embedding layer based on the global header layer parameters, and update the embedding layer based on the global embedding layer parameters, and perform the step of generating problem data features of local service problem data using the embedding layer according to the updated header layer and the updated embedding layer, until a target public service model that meets the training stop conditions is obtained.

[0011] According to a second aspect of an embodiment of this specification, a data processing method is provided, which is applied to a target service terminal among multiple service terminals, including: Receive service problem data submitted by the user for the target service corresponding to the target service terminal; Inputting the service problem data into a target public service model for processing, wherein the target public service model is determined by the system; Determine service answer data according to the processing result, and display the service answer data to the user.

[0012] According to a third aspect of an embodiment of this specification, there is provided a data processing device, applied to a target service terminal among a plurality of service terminals, comprising: A receiving module configured to receive service problem data submitted by a user for a target service corresponding to a target service terminal; a processing module configured to input the service problem data into a target public service model for processing, wherein the target public service model is determined by the above system; The display module is configured to determine service answer data according to the processing result and display the service answer data to the user.

[0013] According to a fourth aspect of the embodiments of this specification, a computing device is provided, including: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the above-mentioned data processing method are implemented.

[0014] According to a fifth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the instructions are executed by a processor, the steps of the above-mentioned data processing method are implemented.

[0015] According to a sixth aspect of the embodiments of this specification, a computer program product is provided, comprising a computer program or instructions, which implement the steps of the above-mentioned data processing method when executed by a processor.

[0016] The public service model training system based on split federated learning provided in this embodiment includes multiple service terminals, a server and an aggregation server. The multiple service terminals deploy the header layer and embedding layer of the public service model, and the server deploys the encoding layer of the public service model, thereby reducing the model training burden of the public service terminal by splitting the public service model. On this basis, multiple service terminals can use the embedding layer to generate question data features of local service question data, obtain question coding features fed back by the server after processing the question data features using the coding layer, input the question coding features into the header layer for processing, obtain predicted answer data, update the header layer based on the target answer data and predicted answer data of the local service question data, send the header layer parameters of the updated header layer to the aggregation server, and send the header layer gradient of the updated header layer to the server; thereafter, the server can update the coding layer according to the header layer gradient sent by each service terminal, and feed back the coding layer gradient of the updated coding layer to multiple service terminals; further, multiple service terminals construct a privacy gradient based on the coding layer gradient, and use the privacy gradient to update the embedding layer, and send the embedding layer parameters of the updated embedding layer to the aggregation server; finally, the aggregation server can optimize the public service model based on the header layer parameters and embedding layer parameters sent by each service terminal until the target public service model that meets the training stop conditions is obtained. Under the premise of ensuring the security of data and model information, the model training burden of the service terminal is reduced, and the trained model has stronger predictive capabilities, so that after being deployed in public service scenarios, it can quickly respond to user questions and provide answers, thereby improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a schematic diagram of the structure of a public service model training system based on split federated learning provided by an embodiment of this specification; Figure 2 This is a schematic diagram of model training in a public service model training system based on split federated learning provided by one embodiment of this specification; Figure 3 is a flow chart of a data processing method provided by one embodiment of this specification; Figure 4 This is a schematic diagram of the structure of a data processing device provided by one embodiment of this specification; Figure 5 This is a structural block diagram of a computing device provided by one embodiment of this specification. DETAILED DESCRIPTION

[0018] The following description sets forth many specific details to facilitate a thorough understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0019] The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms "a," "the," and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0020] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0021] In addition, 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, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0022] This specification provides a public service model training system based on split federated learning. One or more embodiments of this specification also involve a data processing method, a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, each of which is described in detail in the following embodiments.

[0023] With the accelerated digital transformation of public services, large-scale public service models (e.g., for public welfare demand classification and cross-departmental data collaboration) have become key tools for enhancing the intelligent delivery capabilities of public services. However, public service data is highly sensitive (e.g., citizen privacy, financial information, and special documents) and distributed (stored across provinces, municipalities, and departments), posing significant challenges to traditional centralized training. To address this issue, federated learning (FL) for large government models is a key approach. Federated learning (FL) is an emerging distributed machine learning paradigm designed to address issues such as data privacy, data silos, and communication efficiency. However, traditional FL carries privacy risks. For example, clients must download the complete set of large model parameters, leading to model information leakage (e.g., exposing the architecture of customized models). Furthermore, traditional FL faces computing power and efficiency bottlenecks. For example, public service departments have limited computing power on their terminals, making it difficult to locally train large models with hundreds of billions of parameters. Therefore, an effective solution is urgently needed to address these issues.

[0024] The public service model training system based on split federated learning provided in this embodiment includes multiple service terminals, a server and an aggregation server. The multiple service terminals deploy the header layer and embedding layer of the public service model, and the server deploys the encoding layer of the public service model, thereby reducing the model training burden of the public service terminal by splitting the public service model. On this basis, multiple service terminals can use the embedding layer to generate question data features of local service question data, obtain question coding features fed back by the server after processing the question data features using the coding layer, input the question coding features into the header layer for processing, obtain predicted answer data, update the header layer based on the target answer data and predicted answer data of the local service question data, send the header layer parameters of the updated header layer to the aggregation server, and send the header layer gradient of the updated header layer to the server; thereafter, the server can update the coding layer according to the header layer gradient sent by each service terminal, and feed back the coding layer gradient of the updated coding layer to multiple service terminals; further, multiple service terminals construct a privacy gradient based on the coding layer gradient, and use the privacy gradient to update the embedding layer, and send the embedding layer parameters of the updated embedding layer to the aggregation server; finally, the aggregation server can optimize the public service model based on the header layer parameters and embedding layer parameters sent by each service terminal until the target public service model that meets the training stop conditions is obtained. Under the premise of ensuring the security of data and model information, the model training burden of the service terminal is reduced, and the trained model has stronger predictive capabilities, so that after being deployed in public service scenarios, it can quickly respond to user questions and provide answers, thereby improving the user experience.

[0025] See also Figure 1 , Figure 1The following is a schematic diagram of a public service model training system based on split federated learning according to an embodiment of the present specification. The public service model training system 100 based on split federated learning includes multiple service terminals 110, a server 120, and an aggregation server 130. The multiple service terminals 110 deploy the header layer and embedding layer of the public service model, and the server 120 deploys the encoding layer of the public service model, including: The multiple service terminals 110 generate question data features of local service question data using the embedding layer, obtain question encoding features fed back by the server after processing the question data features using the encoding layer, input the question encoding features into the header layer for processing to obtain predicted answer data, update the header layer based on the target answer data of the local service question data and the predicted answer data, and send header layer parameters of the updated header layer to the aggregation server, as well as the header layer gradient of the updated header layer to the server; The server 120 updates the coding layer according to the header layer gradient sent by each service terminal, and feeds back the updated coding layer gradient to the multiple service terminals; The plurality of service terminals 110 construct a privacy gradient based on the coding layer gradient, update the embedding layer using the privacy gradient, and send the embedding layer parameters of the updated embedding layer to the aggregation server; The aggregation server 130 optimizes the public service model based on the header layer parameters and embedding layer parameters sent by each service terminal until a target public service model that meets the training stop condition is obtained.

[0026] Specifically, multiple service terminals refer to terminal devices deployed in public service departments. Different public service departments provide different public services, and the data they hold is private data compared to other parts. A server refers to a server that provides computing resources. An aggregation server refers to a server that aggregates model parameters and updates the model parameters of each service terminal. Both the aggregation server and the server are third-party servers trusted by multiple service terminals.

[0027] Accordingly, the public service model is a model applied to public service scenarios, which can answer public service questions input by users. It can be associated with livelihood appeal scenarios, policy consultation scenarios, rules and regulations scenarios, etc., and supports users to understand public service information through question-and-answer interactions with the public service model, such as responses to materials required to sue others, responses to materials required for household registration transfer, responses to traffic accident handling procedures, etc. This embodiment does not impose any restrictions here. Accordingly, the public service model can be divided into an embedding layer, an encoding layer, and a header layer. The embedding layer is used to embed data, the encoding layer is used to encode data, and the header layer is used to predict the answer results. In order to save computing resources of the service terminal while ensuring data security, the embedding layer and the header layer can be deployed to the service terminal, and the encoding layer can be deployed to the server, thereby realizing multi-party joint training of the public service model. At the same time, it can not only ensure the model's prediction ability, but also avoid data leakage and leakage of model parameters, thereby improving security.

[0028] Accordingly, local service question data specifically refers to the service question data held by each service terminal, which is related to its own public service content. For example, if Department A is mainly responsible for tax matters, then the service question data it holds is associated with the tax dimension. If Department B is mainly responsible for provident fund matters, then the service question data it holds is associated with the provident fund dimension. Accordingly, the question data features are the vector representations corresponding to the local service question data. The question encoding features specifically refer to the encoding features obtained after the server uses the encoding layer to encode the question data features; the predicted answer data specifically refers to the predicted answer obtained after the head layer predicts the local question data. The target answer data is the true answer corresponding to the local service question data, that is, the sample label. The privacy gradient specifically refers to the gradient information of the hidden local information obtained after the service terminal processes the encoding layer gradient. The target public service model specifically refers to the model obtained after training, which can be deployed to each service terminal to implement service question answering.

[0029] Based on this, the public service model training system based on split federated learning provided in this embodiment includes multiple service terminals, a server and an aggregation server. Multiple service terminals deploy the header layer and embedding layer of the public service model, and the server deploys the encoding layer of the public service model, thereby reducing the model training burden of the public service terminal by splitting the public service model. On this basis, multiple service terminals can use the embedding layer to generate question data features of local service question data, obtain question coding features fed back by the server after processing the question data features using the coding layer, input the question coding features into the header layer for processing, obtain predicted answer data, update the header layer based on the target answer data and predicted answer data of the local service question data, send the header layer parameters of the updated header layer to the aggregation server, and send the header layer gradient of the updated header layer to the server; thereafter, the server can update the coding layer according to the header layer gradient sent by each service terminal, and feed back the coding layer gradient of the updated coding layer to multiple service terminals; further, multiple service terminals construct a privacy gradient based on the coding layer gradient, and use the privacy gradient to update the embedding layer, and send the embedding layer parameters of the updated embedding layer to the aggregation server; finally, the aggregation server can optimize the public service model based on the header layer parameters and embedding layer parameters sent by each service terminal until the target public service model that meets the training stop conditions is obtained. Under the premise of ensuring the security of data and model information, the model training burden of the service terminal is reduced, and the trained model has stronger predictive capabilities, so that after being deployed in public service scenarios, it can quickly respond to user questions and provide answers, thereby improving the user experience.

[0030] See also Figure 2As shown in the schematic diagram, the public service model training system based on split federated learning provided in this embodiment can be implemented using Transformer. It includes a head layer, an encoder layer, and an embedding layer. The embedding layer and the head layer can be deployed on the terminals corresponding to each department, and the encoding layer can be deployed on the main server. During model training, the parameters of the public service model can be initialized first, and the model can be split according to the above deployment relationship. The service terminals of the public service department are responsible for calculating the embedding layer and the head layer, and the server is responsible for calculating the encoding layer. During the training phase, the service terminals of the public service department can calculate the forward propagation of the embedding layer, add Laplace noise, and send the result to the server. The server can calculate the forward propagation of the encoding layer and return the result to the terminal corresponding to the corresponding public service department. The service terminal corresponding to the public service department can calculate the loss based on the server output and local labels, and perform backpropagation only on the head layer to update the head layer parameters. The updated head layer parameters can then be sent to the aggregation server, and the head layer gradient can be returned to the server. The server receives the header layer gradient and performs backpropagation only on the coding layer, updating the server coding layer parameters, calculating the coding layer gradient, and returning the coding layer gradient to the service terminal of the public service department. The service terminal of the public service department can clip the gradient returned by the server and inject calibration noise to obtain the privacy gradient. After updating the local model, the model parameters are sent to the aggregation server. At this point, the aggregation server can aggregate the parameters submitted by each service terminal and update the global public service model. The updated model parameters are then distributed to each service terminal and the above operation is repeated until the model converges. The public service model that meets the usage requirements can be obtained and deployed in actual application scenarios.

[0031] For example, Department A is responsible for housing matters, and Department B is responsible for provident fund matters. When departments A and B jointly train a public service model that can provide users with answers to housing questions (such as provident fund loan questions, housing procedures, and housing tax payments), they can split the model into an embedding layer, an encoding layer, and a header layer. The embedding and header layers can then be deployed to the service terminals of Departments A and B, respectively, while the encoding layer is deployed on the server. During the model training phase, Department A can use the embedding layer to generate question data features for local service question data (e.g., what documents are required to buy a house in City C). Meanwhile, Department B can use the embedding layer to generate question data features for local service question data (e.g., what is the interest rate for provident fund loans). The question data features can then be submitted to the server, where they are encoded using the encoding layer. The resulting encoded features are fed back to Department A's service terminal as encoded feature a, and to Department B's service terminal as encoded feature b.

[0032] Furthermore, department A's service terminal can input the encoded feature a into the header layer for processing to obtain the predicted answer a. It can then calculate the loss value of the predicted answer a and the label corresponding to the local problem data (e.g., proof of tax payment required to buy a house in City C). The header layer deployed by department A's service terminal can then be updated. The updated header layer parameter a can then be sent to the aggregation service segment, and the header layer gradient a can be sent to the server. Similarly, after processing, department B's service terminal will send the updated header layer parameter b to the aggregation service segment, and the header layer gradient b to the server.

[0033] Furthermore, after receiving the header layer gradients from departments A and B, the server can update the coding layer and then feed the coding layer gradients back to the service terminals of departments A and B. Department A's service terminal can calculate the privacy gradient based on the coding layer gradients and use it to update the embedding layer. The resulting embedding layer parameter a is then sent to the aggregation server. Similarly, department B's service terminal sends the embedding layer parameter b to the aggregation server.

[0034] On this basis, the aggregation server performs a weighted average calculation on the header-layer parameters and embedding-layer parameters submitted by Departments A and B, respectively, to obtain global header-layer parameters and global embedding-layer parameters. These global header-layer parameters and global embedding-layer parameters are then fed back to the service terminals of Departments A and B. The above process is repeated until the model converges. This results in a public service model that supports answering users' housing questions (such as provident fund loans, housing procedures, and housing tax payments). After deploying this model to the service terminals of Departments A and B, users can use it to inquire about housing, provident fund loans, and housing tax payments. For example, a user asks the question "I want to buy a house in City S. Can the provident fund I paid in City D be used in City S?" at the service terminal of Departments A or B? The question is fed into the trained model for processing, and the answer is fed back to the user: "I can use the provident fund I paid in City D to buy a house in City S. I need to have paid taxes for * years to buy a house in City S..." This is then displayed to the user.

[0035] In summary, under the premise of ensuring the security of data and model information, the model training burden of the service terminal is reduced, and the trained model has stronger predictive capabilities, so that after being deployed in public service scenarios, it can quickly respond to user questions and provide answers, thereby improving the user experience.

[0036] In one or more implementations of this embodiment, the server initializes the public service model and splits the initialized public service model to obtain the header layer, the embedding layer and the coding layer, and sends the first model parameters corresponding to the header layer and the embedding layer to each service terminal for deploying the header layer and the embedding layer at each service terminal.

[0037] Specifically, the first model parameter refers to the model parameters of the embedding layer and the header layer in the public service model, which are used to be sent to each service terminal to realize the construction of the embedding layer and the header layer in each service terminal for joint training of the public service model.

[0038] Based on this, in order to reduce the computing power pressure of the corresponding service terminals of the public service departments, the server can initialize the public service model and split the initialized public service model to obtain the header layer, embedding layer and coding layer. At this time, the first model parameters corresponding to the header layer and embedding layer can be sent to each service terminal for deploying the header layer and embedding layer in each service terminal, while the coding layer can be deployed locally on the server.

[0039] In practical applications, to reduce the computing pressure on the service terminals of public service departments, the service terminals of public service departments can be responsible only for calculating and updating the embedding layer and header layer of the public service model, and the encoding layer with a heavier computational burden can be deployed on the server side. The embedding layer can be initialized using random initialization or pre-trained word vectors. For the initialization of the encoding layer, if the public service model is newly constructed, standard Transformer initialization can be used. If it is based on some pre-trained parameters, the corresponding weights can be loaded and the uncovered parts can be expanded. The header layer can be initialized randomly according to the task.

[0040] Furthermore, model splitting and training initialization can be completed before model training begins, after which a set number of distributed training rounds can begin. In specific implementations, the number of service terminals from all public service departments participating in training can be set to C, and the proportion of service terminals from these public service departments selected for global training in each round can be set to F. Therefore, in each training round, the server can randomly select service terminals from F×C public service departments for training, allowing the selected service terminals from these public service departments to participate in subsequent model training.

[0041] In summary, by splitting the public service model and deploying it on different terminals, the model splitting process can effectively reduce the computing pressure of the service terminal, thereby improving the training efficiency of the model and reducing the computing cost of the service terminal.

[0042] In one or more implementations of this embodiment, the multiple service terminals process the local service problem data using the embedding layer to obtain initial problem data features, add noise features to the initial problem data features to obtain problem data features, and send the problem data features to the server; The server processes the problem data features using the coding layer to obtain problem coding features, and feeds back the problem coding features to the multiple service terminals.

[0043] Specifically, the initial problem data feature specifically refers to the vector expression obtained after processing by the embedding layer, and the noise feature specifically refers to the Laplace noise added to the problem data feature.

[0044] Based on this, any service terminal can use the embedding layer to process the local service problem data to obtain the initial problem data features, and then add noise features to the initial problem data features to obtain the problem data features, and then send the problem data features to the server; after the server receives the problem data features uploaded by each service terminal, it can use the encoding layer to process each problem data feature separately, and after obtaining the problem coding features corresponding to each problem data feature, it can feed back the problem coding features to the corresponding service terminals for subsequent model training processing.

[0045] In specific implementation, the feed-forward calculation of local problem service data on the service terminal of the public service department can stop when it reaches the previous layer L of Layernorm, based on the following formula (1): (1) The output of layer L can be injected with Laplace noise, where Represents the private version , represents the privacy budget. The sensitivity of function A is defined as the maximum output change caused by different inputs under the given input and output distance measurement, as shown in the following formula (2), where the distance measurement between input and output adopts p and q norms respectively: (2) The output of the added noise is fed forward through the rest of the embedding layer of the service terminal model in the public service department, resulting in the output of the embedding layer, namely the problem data features. These features can then be sent to the server. After receiving the problem data features submitted by each service terminal in parallel, the server can input each feature into the encoding layer for forward computation. The resulting encoded features are then fed back to each service terminal for subsequent model training.

[0046] Continuing with the above example, when department A uses the embedding layer to process local service problem data (such as what materials are needed to buy a house in city C), it can combine the above formulas (1) and (2) to complete the construction of the problem data features. Similarly, department B also constructs the problem data features according to the same process, and then submits the problem data features to the server to encode the features using the encoding layer. After obtaining the encoded features, the encoded feature a is fed back to the service terminal of department A, and the encoded feature b is fed back to the service terminal of department B for use in subsequent model training.

[0047] In summary, adding noise to the forward calculation of the embedding layer can achieve differential privacy and thus achieve the purpose of privacy protection, while also improving the robustness of the trained model.

[0048] In one or more implementations of this embodiment, the multiple service terminals determine the target answer data corresponding to the local service question data, calculate the loss value based on the target answer data and the predicted answer data, update the head layer according to the loss value, and obtain the head layer parameters and head layer gradient.

[0049] Specifically, the loss value is the loss value obtained by combining the target answer data and the predicted answer data with a set loss function. It is used to update the head layer parameters. The choice of loss function can be set according to actual needs, such as cross entropy loss function, contrast loss function, etc.

[0050] Based on this, after each service terminal uses the header layer to predict the predicted answer data corresponding to the local question data, it can determine the target answer data corresponding to the local service question data. At this time, the loss value can be calculated based on the target answer data and the predicted answer data, and the header layer can be updated according to the loss value to obtain the header layer parameters and the header layer gradient, so as to send the header layer parameters to the aggregation server and send the header layer gradient to the server for subsequent model training.

[0051] In one or more implementations of this embodiment, the multiple service terminals receive the coding layer gradient fed back by the server, clip the coding layer gradient to obtain an intermediate coding layer gradient, add standard noise to the intermediate coding layer gradient, and obtain a privacy gradient.

[0052] Specifically, the intermediate coding layer gradient refers to the gradient obtained after clipping the coding layer gradient, and the standard noise refers to the interference noise added to the intermediate coding layer gradient to generate the privacy gradient to avoid model information leakage.

[0053] Based on this, after the server calculates the coding layer gradient, it will send the coding layer gradient to each service terminal. The service terminal can receive the coding layer gradient fed back by the server and clip the coding layer gradient to obtain the intermediate coding layer gradient. On this basis, standard noise can be added to the intermediate coding layer gradient to obtain the privacy gradient for use in subsequent model parameter updates.

[0054] During specific implementation, after the service terminal of the public service department receives the problem coding features fed back by the server, it can process the problem coding features through the header layer to obtain the prediction results. At this time, the loss calculation can be performed in combination with the labels corresponding to the local problem data and the prediction results, and backpropagation can be performed to update the header layer. After that, the updated header layer parameters can be uploaded to the aggregation service segment, and the header layer gradient is returned to the server.

[0055] Furthermore, after receiving the header layer gradient returned by the corresponding service terminal of each public service department, the server can perform backpropagation to calculate the coding layer gradient and return it to the service terminal of the corresponding public service department.

[0056] Furthermore, after receiving the coding layer gradient, the service terminal of the public service department can calculate the back propagation. Specifically, assuming that 𝜎 represents the noise intensity and C' represents the gradient norm boundary value. At time 𝑡, the service terminal 𝑘 of the public service department receives the fault parameter from the server. , and thus back propagates for each local sample x i Calculating gradients , as shown in formula (3): (3) Then, based on the following formulas (4) and (5), the 2-norm of each gradient can be clipped one by one, and calibration noise can be injected to obtain the privacy gradient Wherein, formulas (4) and (5) are as follows: (4) (5) Finally, the local model parameters of the service terminal k of the public service department can be updated as , and Send it to the aggregation server for parameter aggregation.

[0057] Continuing with the above example, the service terminal of department A can input the encoding feature a into the head layer for processing to obtain the predicted answer a. It can then calculate the loss value of the predicted answer a and the label corresponding to the local question data (such as tax proof materials are required to buy a house in City C), and update the head layer deployed by the service terminal of department A. After that, the updated head layer parameter a can be sent to the aggregation service segment, and the head layer gradient a can be sent to the server. Similarly, after processing, the service terminal of department B will send the updated head layer parameter b to the aggregation service segment and the head layer gradient b to the server. After receiving the head layer gradients of departments A and B, the server can update the encoding layer and then feed back the encoding layer gradient of the encoding layer to the service terminals of departments A and B. The service terminal of department A can calculate the privacy gradient by combining the encoding layer gradient using the above formulas (3) to (5). Then, the privacy gradient can be used to update the locally deployed model to obtain the embedding layer parameter a after the model update, and send it to the aggregation server. Similarly, the service terminal of department B sends the embedded layer parameter b to the aggregation server.

[0058] In summary, by updating only the embedding layer and header layer on the service terminal and deploying the encoding layer on the server side, the computing pressure on the service terminal can be effectively reduced.

[0059] In one or more implementations of this embodiment, the aggregation server receives the header layer parameters and embedding layer parameters sent by each service terminal, performs weighted average calculation on the header layer parameters and embedding layer parameters sent by each service terminal, obtains global header layer parameters and global embedding layer parameters, and optimizes the public service model based on the global header layer parameters and the global embedding layer parameters until a target public service model that meets the training stop conditions is obtained.

[0060] Specifically, the global header layer parameters refer to the header layer parameters obtained by weighted averaging the header layer parameters submitted by each service terminal, and the global embedding layer parameters refer to the embedding layer parameters obtained by weighted averaging the embedding layer parameters submitted by each service terminal.

[0061] Based on this, after receiving the header layer parameters and embedding layer parameters of each service terminal, the aggregation server can perform weighted average calculation on the header layer parameters and embedding layer parameters sent by each service terminal respectively, so as to obtain the global header layer parameters and global embedding layer parameters from the weighted average calculation results. Thereafter, the public service model can be optimized based on the global header layer parameters and the global embedding layer parameters until the target public service model that meets the training stop conditions is obtained.

[0062] In one or more implementations of this embodiment, the aggregation server sends the global header layer parameters and the global embedding layer parameters to the multiple service terminals; The multiple service terminals update the embedding layer based on the global header layer parameters, and update the embedding layer based on the global embedding layer parameters, and perform the step of generating problem data features of local service problem data using the embedding layer according to the updated header layer and the updated embedding layer, until a target public service model that meets the training stop conditions is obtained.

[0063] Based on this, after the aggregation server calculates the global header layer parameters and the global embedding layer parameters, it can send the parameters to each service terminal. After receiving the global header layer parameters and the global embedding layer parameters, the service terminal can update the embedding layer based on the global header layer parameters and the global embedding layer parameters. Thereafter, the step of generating problem data features of local service problem data using the embedding layer can be performed based on the updated header layer and the updated embedding layer, until the target public service model that meets the training stop condition is obtained.

[0064] In specific implementation, after each round of local training of the service terminal of the public service department is completed, the service terminal of the public service department can update the embedded layer parameters and header layer parameters Upload to the aggregation server, which can calculate the number of service terminals n of each public service department. k The formula for calculating global parameters and weighted average calculation can be implemented by formulas (6) and (7): (6) (7) When a new round of training begins, the aggregation server can use the weighted average global embedding layer parameters and header layer parameters The data is distributed to the terminals of each public service department, overwriting their local model parameters and conducting a new round of model training until the model converges and can be deployed for use.

[0065] Continuing with the above example, the aggregation server performs a weighted average calculation on the header-layer parameters and embedding-layer parameters submitted by Departments A and B, respectively, to obtain global header-layer parameters and global embedding-layer parameters. These global header-layer parameters and global embedding-layer parameters are then fed back to the service terminals of Departments A and B. The above process is repeated until the model converges. This results in a public service model that supports answering users' housing questions (such as questions about provident fund loans, housing procedures, and housing tax payments). Deployed on the service terminals of Departments A and B, this model allows users to inquire about housing, provident fund loans, and housing tax payments. For example, a user asks the question "I want to buy a house in City S. Can the provident fund I paid in City D be used in City S?" at the service terminal of Departments A or B? The question is fed into the trained model for processing, and the answer is fed back to the user: "I can use the provident fund I paid in City D to buy a house in City S. To buy a house in City S, you must have paid taxes for * years..." This is then displayed to the user.

[0066] The public service model training system based on split federated learning provided in this embodiment includes multiple service terminals, a server and an aggregation server. The multiple service terminals deploy the header layer and embedding layer of the public service model, and the server deploys the encoding layer of the public service model, thereby reducing the model training burden of the public service terminal by splitting the public service model. On this basis, multiple service terminals can use the embedding layer to generate question data features of local service question data, obtain question coding features fed back by the server after processing the question data features using the coding layer, input the question coding features into the header layer for processing, obtain predicted answer data, update the header layer based on the target answer data and predicted answer data of the local service question data, send the header layer parameters of the updated header layer to the aggregation server, and send the header layer gradient of the updated header layer to the server; thereafter, the server can update the coding layer according to the header layer gradient sent by each service terminal, and feed back the coding layer gradient of the updated coding layer to multiple service terminals; further, multiple service terminals construct a privacy gradient based on the coding layer gradient, and use the privacy gradient to update the embedding layer, and send the embedding layer parameters of the updated embedding layer to the aggregation server; finally, the aggregation server can optimize the public service model based on the header layer parameters and embedding layer parameters sent by each service terminal until the target public service model that meets the training stop conditions is obtained. Under the premise of ensuring the security of data and model information, the model training burden of the service terminal is reduced, and the trained model has stronger predictive capabilities, so that after being deployed in public service scenarios, it can quickly respond to user questions and provide answers, thereby improving the user experience.

[0067] See also Figure 3 , Figure 3A flow chart of a data processing method provided according to an embodiment of the present specification is shown, which is applied to a target service terminal among multiple service terminals and specifically includes the following steps.

[0068] Step S302: receiving service problem data submitted by a user for a target service corresponding to a target service terminal.

[0069] Step S304: input the service problem data into a target public service model for processing, wherein the target public service model is determined by the above system.

[0070] Step S306: Determine service answer data according to the processing result, and display the service answer data to the user.

[0071] The data processing method provided in this embodiment can be applied to any public service scenario. For example, when a user interacts with a target public service model through question-and-answer interaction and asks about the documents required for issuing a certain certificate, the model can output a list of required documents for the user's reference. For another example, when a user asks about the materials required to apply for a residence permit in City A, the model can output information about the required materials. In specific implementation, the public service scenario can be set according to actual needs, and this embodiment does not impose any restrictions on this.

[0072] Based on this, after receiving the service problem data submitted by the user for the target business corresponding to the target service terminal, the service problem data can be input into the target public service model for processing. At this time, the service answer data can be determined according to the processing results, and the service answer data can be displayed to the user for user use.

[0073] During specific implementation, the deployment of the target public service model can also be completed according to the deployment relationship in the model training phase, thereby saving the computing resources of the service terminal and realizing the joint completion of the model application by multiple parties. It is not only convenient for users to use, but also can reduce the resource consumption of the service terminal, thereby saving the operating costs of the service terminal.

[0074] Corresponding to the above method embodiment, this specification also provides a data processing device embodiment, Figure 4 FIG1 shows a schematic diagram of the structure of a data processing device provided by an embodiment of this specification. Figure 4 As shown, the device is applied to a target service terminal among multiple service terminals, including: The receiving module 402 is configured to receive service problem data submitted by a user for a target service corresponding to a target service terminal; a processing module 404 configured to input the service problem data into a target public service model for processing, wherein the target public service model is determined by the above system; The display module 406 is configured to determine service answer data according to the processing result and display the service answer data to the user.

[0075] The above is a schematic diagram of a data processing device according to this embodiment. It should be noted that the technical solution of the data processing device and the technical solution of the above-mentioned data processing method are based on the same concept. For details not described in detail in the technical solution of the data processing device, please refer to the description of the technical solution of the above-mentioned data processing method.

[0076] Figure 5 The block diagram of a computing device 500 according to one embodiment of the present disclosure is shown. Components of the computing device 500 include, but are not limited to, a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and a database 550 is used to store data.

[0077] The computing device 500 also includes an access device 540 that enables the computing device 500 to communicate via one or more networks 560. Examples of such networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 540 may include one or more of any type of network interface (e.g., a network interface card (NIC)) whether wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, or a near field communication (NFC) interface.

[0078] In one embodiment of the present specification, the above components of the computing device 500 and Figure 5 Other components not shown in the figure may also be connected to each other, for example, via a bus. Figure 5The computing device structure block diagram shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art may add or replace other components as needed.

[0079] Computing device 500 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, personal digital assistant, laptop computer, notebook computer, netbook computer, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smartwatch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or personal computer (PC). Computing device 500 can also be a mobile or stationary server.

[0080] The processor 520 is configured to execute the following computer-executable instructions, which implement the steps of the above-mentioned data processing method when executed by the processor.

[0081] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above-mentioned data processing method are of the same concept. For details not described in detail in the technical scheme of the computing device, please refer to the description of the technical scheme of the above-mentioned data processing method.

[0082] An embodiment of the present specification further provides a computer-readable storage medium storing computer-executable instructions, which implement the steps of the above-mentioned data processing method when executed by a processor.

[0083] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the above-mentioned data processing method are based on the same concept. For details not described in detail in the technical scheme of the storage medium, please refer to the description of the technical scheme of the above-mentioned data processing method.

[0084] An embodiment of the present specification further provides a computer program product, including a computer program or instructions, which implement the steps of the above-mentioned data processing method when executed by a processor.

[0085] The above is a schematic solution of a computer program product of this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the above-mentioned data processing method are based on the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the above-mentioned data processing method.

[0086] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0087] The computer instructions include computer program code, which may be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content of the computer-readable medium may be appropriately increased or decreased based on the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0088] It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0089] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0090] The preferred embodiments disclosed above are intended only to help illustrate this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of the embodiments described herein. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification.

Claims

1. A public service model training system based on split federated learning, characterized by: The system comprises a plurality of service terminals, a server and an aggregation server, wherein the plurality of service terminals deploy a header layer and an embedding layer of a public service model, and the server deploys a coding layer of the public service model, including: The multiple service terminals generate question data features of local service question data using the embedding layer, obtain question coding features fed back by the server after processing the question data features using the coding layer, input the question coding features into the header layer for processing to obtain predicted answer data, update the header layer based on the target answer data of the local service question data and the predicted answer data, send header layer parameters of the updated header layer to the aggregation server, and send the header layer gradient of the updated header layer to the server; The server updates the coding layer according to the header layer gradient sent by each service terminal, and feeds back the updated coding layer gradient to the multiple service terminals; The multiple service terminals construct a privacy gradient based on the coding layer gradient, update the embedding layer using the privacy gradient, and send the embedding layer parameters of the updated embedding layer to the aggregation server; The aggregation server optimizes the public service model based on the header layer parameters and embedding layer parameters sent by each service terminal until a target public service model that meets the training stop condition is obtained.

2. The public service model training system based on split federated learning according to claim 1 is characterized in that: The server initializes the public service model and splits the initialized public service model to obtain the header layer, the embedding layer and the coding layer, and sends the first model parameters corresponding to the header layer and the embedding layer to each service terminal for deploying the header layer and the embedding layer at each service terminal.

3. The public service model training system based on split federated learning according to claim 1 is characterized in that: The multiple service terminals process the local service problem data using the embedding layer to obtain initial problem data features, add noise features to the initial problem data features to obtain problem data features, and send the problem data features to the service end; The server processes the problem data features using the coding layer to obtain problem coding features, and feeds back the problem coding features to the multiple service terminals.

4. The public service model training system based on split federated learning according to claim 1 is characterized in that: The multiple service terminals determine the target answer data corresponding to the local service question data, calculate the loss value based on the target answer data and the predicted answer data, update the head layer according to the loss value, and obtain the head layer parameters and head layer gradient.

5. The public service model training system based on split federated learning according to claim 1 is characterized in that: The multiple service terminals receive the coding layer gradient fed back by the server, clip the coding layer gradient to obtain an intermediate coding layer gradient, and add standard noise to the intermediate coding layer gradient to obtain a privacy gradient.

6. The public service model training system based on split federated learning according to claim 1 is characterized in that: The aggregation server receives the header layer parameters and embedding layer parameters sent by each service terminal, performs weighted average calculation on the header layer parameters and embedding layer parameters sent by each service terminal, obtains global header layer parameters and global embedding layer parameters, and optimizes the public service model based on the global header layer parameters and the global embedding layer parameters until a target public service model that meets the training stop conditions is obtained.

7. The public service model training system based on split federated learning according to claim 6 is characterized in that: The aggregation server sends the global header layer parameters and the global embedding layer parameters to the multiple service terminals; The multiple service terminals update the embedding layer based on the global header layer parameters, and update the embedding layer based on the global embedding layer parameters, and perform the step of generating problem data features of local service problem data using the embedding layer according to the updated header layer and the updated embedding layer, until a target public service model that meets the training stop conditions is obtained.

8. A data processing method, characterized in that: The target service terminal applied to multiple service terminals includes: Receive service problem data submitted by the user for the target service corresponding to the target service terminal; inputting the service problem data into a target public service model for processing, wherein the target public service model is determined by the system of any one of claims 1 to 7; Determine service answer data according to the processing result, and display the service answer data to the user.

9. A data processing device, characterized in that: The target service terminal applied to multiple service terminals includes: A receiving module configured to receive service problem data submitted by a user for a target service corresponding to a target service terminal; a processing module configured to input the service problem data into a target public service model for processing, wherein the target public service model is determined by the system of any one of claims 1 to 7; The display module is configured to determine service answer data according to the processing result and display the service answer data to the user.

10. A computing device, characterized in that include: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method according to claim 8 are implemented.

11. A computer-readable storage medium, characterized in that It stores computer-executable instructions, which implement the steps of the method according to claim 8 when executed by a processor.

12. A computer program product, characterized in that The method comprises a computer program or an instruction, which implements the steps of the method according to claim 8 when the computer program or the instruction is executed by a processor.

Images