A model training method, device and equipment based on federated transfer learning

Abstract

The present application discloses a model training method, apparatus, and device based on federated transfer learning, comprising: receiving a training data set sent by a main computing node, obtained by performing knowledge distillation processing on an original data set, wherein the training data set includes training data corresponding to each original data in the original data set, and each training data respectively includes: any one or more labels among category labels, result labels, and cause labels; performing category training, result training, and cause training on an initial model based on the original data set and the training data set, respectively, to obtain an initial category loss, an initial result loss, and an initial cause loss; determining an initial total loss based on any one or more losses among the initial category loss, the initial result loss, and the initial cause loss, and adjusting parameters of the initial model based on the initial total loss to obtain a current model, until a predetermined training stop condition is met to obtain a target model. The present application can improve the accuracy of model training.

Description

Technical Field

[0001] The present invention relates to the field of computers, and in particular to a model training method, apparatus and device based on federated transfer learning. Background Art

[0002] Transfer learning is a learning process that leverages similarities between data, tasks, or models to apply models learned in previous domains to new domains. Federated transfer learning is a new transfer learning paradigm that uses a federated learning framework. It aims to address the problem of knowledge being dispersed across different data owners through collaborative modeling across multiple institutions, while protecting the original model from being leaked.

[0003] However, existing model training methods based on federated transfer learning suffer from the problems of inaccurate and unreliable model training. Summary of the Invention

[0004] In view of this, the present invention provides a model training method, device and equipment based on federated transfer learning, the main purpose of which is to solve the current problems of inaccurate and unreliable model training.

[0005] To solve the above problems, this application provides a model training method based on federated transfer learning, which is applied to slave computing nodes, including:

[0006] Receive a training data set sent by the master computing node, which is obtained by performing knowledge distillation processing on the original data set of the slave computing node, wherein the training data set includes training data corresponding to each original data in the original data set, and each training data includes: any one or more labels of a category label, a result label, and a cause label;

[0007] Based on the original data set and the training data set, the initial model is trained on categories, results, and causes, respectively, to obtain the initial category loss, initial result loss, and initial cause loss;

[0008] Based on any one or more of the initial category loss, the initial result loss, and the initial cause loss, the initial total loss is determined, and the parameters of the initial model are adjusted based on the initial total loss to obtain the current model, until the predetermined training stop condition is met, the target model is obtained.

[0009] Optionally, before receiving the training data set sent by the master computing node, the method further includes:

[0010] Receiving a first public key sent by the master computing node, encrypting the local original data set based on the first public key to generate a first encrypted data set;

[0011] The first encrypted data set is sent to the main computing node, so that the main computing node decrypts the first encrypted data set based on a first private key corresponding to the first public key to obtain the original data set.

[0012] Optionally, before receiving the training data set sent by the master computing node, the method further includes:

[0013] Pre-generating a second public key and a second private key corresponding to the second public key;

[0014] Sending the second public key to the main computing node, so that the main computing node encrypts the training data set based on the second public key to obtain a second encrypted data set;

[0015] The receiving, sent by the master computing node, of a training data set obtained by performing knowledge distillation processing on an original data set of a slave computing node specifically includes:

[0016] receiving a second encrypted data set sent by the primary computing node;

[0017] The second encrypted data set is decrypted based on the second private key to obtain a training data set.

[0018] Optionally, the initial model is subjected to category training, result training, and cause training based on the original dataset and the training dataset, respectively, to obtain an initial category loss, an initial result loss, and an initial cause loss, specifically including:

[0019] Based on the training data set, a category training set containing category labels, a result training set containing result labels, and a cause training set containing cause labels are constructed respectively;

[0020] Based on the original data set and the category training set, the initial model is trained to obtain the initial category loss;

[0021] Based on the original data set and the result training set, the initial model is trained to obtain the initial result loss;

[0022] Based on the original data set and the reason training set, the initial model is trained to obtain the initial reason loss.

[0023] Optionally, the determining of the initial total loss based on any one or more of the initial category loss, the initial result loss, and the initial cause loss specifically includes:

[0024] Determine a first coefficient corresponding to the category label, a second coefficient corresponding to the result label, and a third coefficient corresponding to the cause label;

[0025] An initial total loss is determined based on the initial category loss, the initial result loss, the initial cause loss, the first coefficient, the second coefficient, and the third coefficient.

[0026] Optionally, after obtaining the current model, the method further includes:

[0027] Based on the current loss corresponding to the current model, determine whether the predetermined training stop condition is met;

[0028] Alternatively, based on the number of training rounds corresponding to the current model, it is determined whether a predetermined training stop condition is met.

[0029] To solve the above problems, this application provides a model training method based on federated transfer learning, which is applied to the main computing node, including:

[0030] Receive the original data set sent from the computing node;

[0031] Performing knowledge distillation on the original data set based on a predetermined prompt word template and a predetermined large model to obtain a training data set;

[0032] The training data set is sent to the corresponding slave computing node, so that the slave computing node can perform model training based on the training data set and the original data set to obtain the target model.

[0033] Optionally, before receiving the original data set sent from the computing node, the method further includes:

[0034] Pre-generating a first public key and a first private key corresponding to the first public key;

[0035] Sending the first public key to the slave computing node, so that the slave computing node encrypts the original data set based on the first public key to obtain a first encrypted data set;

[0036] The receiving of the original data set sent from the computing node specifically includes:

[0037] receiving a first encrypted data set sent from a computing node;

[0038] The first encrypted data set is decrypted based on the first private key to obtain the original data set.

[0039] Optionally, performing knowledge distillation on the original data set based on a predetermined prompt word template and a predetermined large model to obtain a training data set specifically includes:

[0040] Based on a predetermined prompt word template and each original data in the original data set, a predetermined large model is used to generate training data corresponding to each original data to obtain a training data set;

[0041] Each training data includes: any one or more of category labels, result labels, and cause labels.

[0042] Optionally, before performing knowledge distillation processing on the original dataset based on the predetermined prompt word template and the predetermined large model, the method further includes:

[0043] Pre-generating a prompt word template based on a predetermined prompt mode;

[0044] The predetermined prompting methods include any one or more of the following: a zero-sample prompting method, a small number of sample prompting methods, and a thinking chain prompting method.

[0045] Optionally, before sending the training data set to the corresponding slave computing node, the method further includes:

[0046] receiving a second public key sent from the computing node;

[0047] Encrypting the training data set based on the second public key to obtain a second encrypted data set;

[0048] The sending of the training data set to the corresponding slave computing node specifically includes:

[0049] The second encrypted data set is sent to the slave computing node, so that the slave computing node decrypts the second encrypted data set based on the second private key to obtain the training data set.

[0050] To solve the above problems, this application provides a model training device based on federated transfer learning, including:

[0051] A first receiving module is configured to receive a training data set sent by the master computing node, which is obtained by performing knowledge distillation processing on the original data set of the slave computing node, wherein the training data set includes training data corresponding to each original data in the original data set, and each training data includes: any one or more labels of a category label, a result label, and a cause label;

[0052] The training module is used to perform category training, result training, and cause training on the initial model based on the original data set and the training data set, respectively, to obtain the initial category loss, initial result loss, and initial cause loss;

[0053] The acquisition module is used to determine the initial total loss based on any one or more of the initial category loss, the initial result loss, and the initial cause loss, and adjust the parameters of the initial model based on the initial total loss to obtain the current model, until the predetermined training stop condition is met, and the target model is obtained.

[0054] To solve the above problems, this application provides a model training device based on federated transfer learning, including:

[0055] A second receiving module is used to receive the original data set sent from the computing node;

[0056] A processing module is used to perform knowledge distillation on the original data set based on a predetermined prompt word template and a predetermined large model to obtain a training data set;

[0057] The second sending module is used to send the training data set to the corresponding slave computing node, so that the slave computing node can perform model training based on the training data set and the original data set to obtain a target model.

[0058] To solve the above problems, the present application provides an electronic device, which includes at least a memory and a processor, wherein a computer program is stored on the memory, and when the processor executes the computer program on the memory, it implements the steps of any of the above-mentioned model training methods based on federated transfer learning.

[0059] The present application discloses a model training method, apparatus and device based on federated transfer learning. By receiving a training data set sent by a main computing node and obtained by knowledge distillation processing of the original data set, the slave computing node can subsequently perform multi-task learning on the local initial model based on the training data set and the local original data set, that is, perform cause training / attribution training, result training and category training respectively, and then determine the final loss according to the loss of each training, and optimize and adjust the model parameters based on the final loss, so that the target model obtained by the final training is more accurate and reliable.

[0060] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0062] Figure 1 This is a flowchart of a model training method based on federated transfer learning according to an embodiment of the present application;

[0063] Figure 2 This is a flowchart of a model training method based on federated transfer learning according to another embodiment of the present application;

[0064] Figure 3 This is a structural block diagram of a model training device based on federated transfer learning in another embodiment of the present application;

[0065] Figure 4 This is a structural block diagram of a model training device based on federated transfer learning in another embodiment of the present application;

[0066] Figure 5 This is a structural block diagram of an electronic device according to another embodiment of the present application. DETAILED DESCRIPTION

[0067] Various aspects and features of the present application are described herein with reference to the accompanying drawings.

[0068] It should be understood that various modifications may be made to the embodiments of the present application. Therefore, the above description should not be considered as limiting, but merely as an example of an embodiment. Other modifications within the scope and spirit of the present application will occur to those skilled in the art.

[0069] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0070] These and other characteristics of the present application will become apparent from the following description of a preferred form of embodiment given as a non-limiting example with reference to the accompanying drawings.

[0071] It should also be understood that although the present application has been described with reference to certain specific examples, those skilled in the art will readily be able to implement many other equivalent forms of the present application.

[0072] The above and other aspects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings.

[0073] Specific embodiments of the present application will be described hereinafter with reference to the accompanying drawings; however, it should be understood that the embodiments described are merely examples of the present application and may be implemented in a variety of ways. Familiar and / or repetitive functions and structures are not described in detail to avoid obscuring the present application with unnecessary or redundant details. Therefore, the specific structural and functional details described herein are not intended to be limiting, but rather serve merely as a basis and representative basis for the claims to teach those skilled in the art to variously utilize the present application with substantially any suitable detailed structure.

[0074] This specification may use the phrases "in one embodiment," "in another embodiment," "in yet another embodiment," or "in other embodiments," which may all refer to one or more of the same or different embodiments according to the present application.

[0075] The embodiment of the present application provides a model training method based on federated transfer learning, which can be applied to the slave computing node / participant of federated learning. Specifically, the slave computing node can be a financial institution. Figure 1 As shown, the method in this embodiment includes the following steps:

[0076] Step S101: receiving a training data set sent by a master computing node, which is obtained by performing knowledge distillation processing on the original data set of the slave computing node. The training data set includes training data corresponding to each original data in the original data set, and each training data includes: any one or more labels of a category label, a result label, and a cause label;

[0077] During the specific implementation of this step, the slave computing nodes can pre-deploy the components of the federated transfer learning collaborative network in the local data center and transmit data with the master computing node. Specifically, each slave computing node can first send the original data set to the master computing node, so that the master computing node can perform knowledge distillation processing on the original data set of each slave computing node respectively, obtain a training data set corresponding to each original data set, and then the master computing node sends the training data set to the corresponding slave computing node. Among them, when the master computing node performs knowledge distillation processing on the original data set, each original data in the original data set and a predetermined prompt word template can be input into a predetermined large model, and the large model can be used to output training data containing three label dimensions of cause label, result label and category label according to the prompt word template, thereby obtaining a training data set corresponding to the original data set.

[0078] Step S102: Based on the original data set and the training data set, the initial model is respectively subjected to category training, result training, and cause training to obtain an initial category loss, an initial result loss, and an initial cause loss;

[0079] In this step, the training data corresponding to each piece of original data in the original data set can be determined based on the training data set. Since the training data contains the category label, result label and cause label of the original data, each piece of original data and the category label corresponding to the original data can be input into the initial model local to the computing node, and the initial model is used to output the prediction result, and the category loss is calculated based on the prediction result, that is, the initial category loss is obtained; similarly, each piece of original data and the result label corresponding to the original data can be input into the initial model local to the computing node, and the initial model is used to output the prediction result, and the result loss is calculated based on the prediction result, that is, the initial result loss is obtained; similarly, each piece of original data and the cause label corresponding to the original data can be input into the initial model local to the computing node, and the initial model is used to output the prediction result, and the cause loss is calculated based on the prediction result, that is, the initial cause loss is obtained.

[0080] Step S103, based on any one or more of the initial category loss, the initial result loss, and the initial cause loss, the initial total loss is determined, and the parameters of the initial model are adjusted based on the initial total loss to obtain the current model, until the predetermined training stop condition is met, and the target model is obtained.

[0081] In this step, an initial total loss can be determined based on one or more of the initial class loss, initial result loss, and initial cause loss, as well as the coefficients / hyperparameters corresponding to each loss type. Finally, the model parameters of the initial model are adjusted based on the initial total loss to obtain the current model. In the specific implementation of this step, whether a predetermined training stopping condition is met can be determined based on the current total loss of the current model, or can also be determined based on the number of training rounds.

[0082] The model training method based on federated transfer learning in this embodiment receives the training data set sent by the main computing node, which is obtained by knowledge distillation processing of the original data set. Subsequently, the slave computing node can perform multi-task learning on the local initial model based on the training data set and the local original data set, that is, cause training / attribution training, result training and category training are performed respectively, and then the final loss is determined according to the loss of each training, and the model parameters are optimized and adjusted based on the final loss, so that the target model obtained by the final training is more accurate and reliable.

[0083] Based on the above embodiment, another embodiment of the present application provides a model training method based on federated transfer learning, which is applied to slave computing nodes. The overall process of model training is as follows:

[0084] Step S201: Installation and initialization of the federated transfer learning framework.

[0085] In this step, each financial institution can be used as a slave computing node / participant, and the components of the federated transfer learning collaborative network can be pre-deployed in the local data center of each slave computing node. Similarly, financial infrastructure can be used as a master computing node / leader, and the components of the federated transfer learning collaborative network can be pre-deployed in the local data center of the master computing node. The framework of the federated learning collaborative network can be implemented using other software products with similar functions, such as FATE, SecretFlow, PaddleFL, TensorFlowFederated, and Pysyft. This federated transfer learning network has a star topology, and each slave computing node only interacts bidirectionally with the master computing node.

[0086] Step S202, key interaction;

[0087] In this step, the main computing node can pre-generate an asymmetric key pair, namely a first key pair, which includes a first public key pk center And the first private key sk center , then the first public key pk center Sent to the slave computing node, so that the slave computing node can receive the first public key pk sent by the master computing node center .

[0088] Similarly, each slave computing node will also pre-generate an asymmetric key pair, namely the second key pair, which includes the second public key pk i node And the second private key sk i node Where i represents the i-th slave computing node, and then the slave computing node i can use the second public key pk i node Sent to the main computing node to complete the key interaction.

[0089] Step S203: encrypting the local original data set based on the first public key to generate a first encrypted data set;

[0090] In this step, the slave computing node can use the first public key of the master computing node to encrypt the original data set to prevent local data leakage and protect data privacy.

[0091] Step S204: Send the first encrypted data set to the main computing node, so that the main computing node decrypts the first encrypted data set based on the first private key corresponding to the first public key to obtain the original data set;

[0092] In this step, after receiving the first encrypted data, the main computing node can use the first private key sk center The first encrypted data is decrypted to obtain the original data set of the slave computing node. Specifically, since the master computing node receives the first encrypted data sent by each slave computing node, the slave computing node can configure a slave computing node identifier for each first encrypted data / original data set to facilitate the master computing node to distinguish between the first encrypted data / original data sets.

[0093] Step S205, receiving the second encrypted data set sent by the main computing node;

[0094] In this step, the second encrypted data set is obtained by the master computing node encrypting the training data set based on the second public key, and the training data set is obtained by the master computing node performing knowledge distillation processing on the original data set of the slave computing node.

[0095] That is, after the master computing node obtains the original data set by decrypting it based on the first private key, it can perform knowledge distillation on the original data set to obtain the training data set. Then, the master computing node can use the second public key Pk i node , encrypting the training data set to obtain a second encrypted data set, and sending the second encrypted data set to the slave computing node, so that the slave computing node can receive the second encrypted data set.

[0096] Step S206: decrypt the second encrypted data set based on the second private key to obtain a training data set;

[0097] In this step, the slave computing node can use the second private key sk i node The second encrypted data set is decrypted to obtain a training data set corresponding to the local original data set.

[0098] Step S207, performing category training, result training, and cause training on the initial model based on the original data set and the training data set, respectively, to obtain an initial category loss, an initial result loss, and an initial cause loss;

[0099] In the specific implementation of this step, the training process of cause training, result training and category training is as follows:

[0100] 1. Category training: Construct a category training set containing category labels based on the training dataset; perform category training on the initial model based on the original dataset and the category training set to obtain the initial category loss.

[0101] That is, the computing node uses the original conversation data / raw data and [label] labels / category labels as input, and the category data as output to train the locally deployed small model / initial model, such as:

[0102] prompt_CoT_Few_shot = Q: Please analyze the category of this task [label]: "In the second quarter of 2024, the scale of risky bonds was 15.276 billion yuan, a year-on-year decrease of approximately 76%, and credit risk continued to converge. Is this good news for the bond market?"

[0103] A: "Sentiment analysis."

[0104] Therefore, the initial category loss can be calculated based on the category data and category labels output by the small model / initial model. That is, the initial category loss of node i is calculated as follows:

[0105]

[0106] in, is the category loss in category training, N i is the number of data in the category training set of the i-th slave computing node, CL is the label prediction loss, which is used to measure the category predicted by the model With the real category The cross entropy loss between .

[0107] 2. Result training: Construct a result training set of result labels based on the training dataset; perform result training on the initial model based on the original dataset and the result training set to obtain the initial result loss;

[0108] That is, the slave computing node can use the original conversation data / raw data and [result] labels / result labels as input, and the result data as output to perform result training on the locally deployed small model / initial model, such as:

[0109] prompt_CoT_Few_shot = Q: "In the second quarter of 2024, the scale of risky bonds was 15.276 billion yuan, a year-on-year decrease of approximately 76%. Credit risk continues to converge. Is this good news for the bond market?"

[0110] A: "Based on these data, the bond market performance in the second quarter of 2024 can be considered positive public opinion."

[0111] Therefore, the initial result loss can be calculated based on the result data and result labels output by the small model / initial model. That is, the initial result loss of node i is calculated as follows:

[0112]

[0113] in, is the result loss in the result training, N i is the number of data in the training set of the result from the i-th computing node, and CL is the label prediction loss, which is used to measure the results of model prediction. With real results The cross entropy loss between .

[0114] 3. Cause training / attribution training: Construct a cause training set containing cause labels based on the training dataset; perform cause training on the initial model based on the original dataset and the cause training set to obtain the initial cause loss.

[0115] That is, the computing node uses the original conversation data / raw data and the [reason] label / reason label as input, and the attribution data / reason data as output to train the small model / initial model deployed locally on the computing node, such as:

[0116] prompt_CoT_Few_shot = Q: "In the second quarter of 2024, the scale of risky bonds was 15.276 billion yuan, a year-on-year decrease of approximately 76%, and credit risk continued to converge. Is this good news for the bond market?" The category [label] of this task is sentiment analysis, and the result [result] is "Based on these data, the performance of the bond market in the second quarter of 2024 can be considered positive public opinion." Please analyze the reasons for this result.

[0117] A: "This data shows the following positive signals: 1. A significant decline in the scale of defaults: Compared with the same period last year, the scale of defaults has decreased significantly, indicating that credit risk in the bond market has been effectively controlled, with the frequency and scale of default events decreasing. 2. Convergence of credit risk: The convergence of credit risk means that the uncertainty faced by market participants has decreased, which helps to enhance investor confidence and attract more funds to flow into the bond market. 3. Enhanced market stability: A reduction in defaults is generally seen as enhanced market stability, which is of great significance for maintaining the healthy development of the bond market. Therefore, despite the existence of defaults and extensions, the significant decline in the scale of defaults and the convergence of credit risk indicate that the market is moving in a healthier direction."

[0118] Therefore, the initial cause loss can be calculated based on the cause data and cause labels output by the small model / initial model. as follows:

[0119]

[0120] in, is the cause loss in attribution training, N i is the number of data in the training set from the i-th slave computation node, and CL is the label prediction loss, which is used to measure the attribution of model predictions. and true attribution The cross entropy loss between .

[0121] Step S208, determining the initial total loss based on any one or more of the initial cause loss, the initial result loss, and the initial category loss;

[0122] During the specific implementation of this step, the first coefficient corresponding to the category label, the second coefficient corresponding to the result label, and the third coefficient corresponding to the cause label can be determined; based on the initial category loss, initial result loss, initial cause loss, the first coefficient, the second coefficient, and the third coefficient, the initial total loss is determined.

[0123] Specifically, taking the example of determining the total loss based on the three types of losses: category loss, result loss, and cause loss, the initial total loss calculation formula is as follows:

[0124]

[0125] in, is the total loss of the i-th slave computing node in the chain distillation process, 、 、 All of them are configurable and adjustable hyperparameters / coefficients, which can be flexibly adjusted by the slave computing nodes to meet For example, select , , .

[0126] During the chain distillation process, the initial model / small model not only predicts the task output results, but also generates corresponding attributions and limits the task category according to the business related to the computing node, thereby improving the reasoning and generation explanation capabilities of the small model.

[0127] Step S209: Adjust the parameters of the initial model based on the initial total loss to obtain the current model, and obtain the target model when the predetermined training stop condition is met.

[0128] In this step, after obtaining the current model, it can be determined whether the predetermined training stop condition is met based on the current total loss corresponding to the current model; or, it can be determined whether the predetermined training stop condition is met based on the number of training rounds corresponding to the current model.

[0129] That is, after obtaining the current model, the current model can be used to perform category training, result training, and cause training using the original data set and the training data set, respectively, to obtain the current category loss, the current result loss, and the current cause loss. Then, based on any one or more of the current category loss, the current result loss, and the current cause loss, the current total loss is determined, and the current total loss is compared with the predetermined loss threshold. When it is determined that the current total loss is less than or equal to the predetermined loss threshold, or when the number of model training rounds is greater than the predetermined round threshold, it is determined that the predetermined training stop condition is met, and the current model can be used as the target model. Conversely, when it is determined that the current total loss is greater than the predetermined loss threshold and the number of training rounds is less than the predetermined round threshold, the current model continues to be adjusted based on the current total loss until the current total loss of the adjusted current model is less than or equal to the predetermined loss threshold and the number of training rounds is greater than the predetermined round threshold, and the model training is stopped.

[0130] In this embodiment, after each slave compute node completes local mini-model / initial model training, it can be deployed and used. These mini-models have reasoning and explanation generation capabilities close to those of the large model in their respective business domains, without requiring interaction with the master compute node. Specifically, the master compute node can use the error dataset returned by the slave compute node as input to fine-tune the instructions of the large model.

[0131] The method in this embodiment receives a training data set sent by the main computing node, which is obtained by knowledge distillation processing of the original data set. Subsequently, the computing node can perform multi-task learning on the local initial model based on the training data set and the local original data set, that is, cause training / attribution training, result training and category training are performed respectively, and then the final total loss is determined according to the loss of each training, so that the determination of the total loss is more reasonable and accurate. Subsequently, the model parameters can be accurately optimized and adjusted based on the total loss, so that the target model obtained by the final training is more accurate and reliable.

[0132] Another embodiment of the present application provides a model training method based on federated transfer learning, which can be applied to the main computing node / leading party of federated learning. Specifically, the main computing node can be a financial infrastructure. Figure 2 As shown, the method in this embodiment includes the following steps:

[0133] Step S301, receiving an original data set sent from a computing node;

[0134] In this step, the master computing node can pre-deploy the components of the federated transfer learning collaborative network in the local data center and transmit data with the slave computing nodes, so that the master computing node can receive the original data sets sent by each slave computing node.

[0135] Step S302: performing knowledge distillation on the original dataset based on a predetermined prompt word template and a predetermined large model to obtain a training dataset;

[0136] In this step, the master computing node can pre-deploy a large model locally for performing knowledge distillation. Based on the large model, the master computing node can perform knowledge distillation on the original datasets of each slave computing node, obtaining training datasets corresponding to each original dataset. Specifically, when performing knowledge distillation on the original dataset, the master computing node can input each piece of raw data in the original dataset and a predetermined prompt word template into the predetermined large model. The large model then outputs training data containing three label dimensions: cause label, result label, and category label, based on the prompt word template, thereby obtaining training datasets corresponding to the original datasets.

[0137] Specifically, category labels [label] may include: sentiment analysis, text generation, machine translation, text classification, entity relationship extraction, text summarization, semantic role labeling, and dialogue systems, etc. The specific data format can be:

[0138] Sentiment Analysis (SA): This task involves determining the sentiment of input text. For example, analyzing public opinion about bond market news.

[0139] Text Generation (TG): Tasks that generate text based on a given prompt are classified as "TG." For example, a monthly report on interbank bond market statistics and analysis might be generated based on statistical data.

[0140] Machine Translation (MT): MT tasks involve translating from one language into another. For example, the translation of news about Euroclear's private blockchain-based central bank digital currency and French government bond settlements.

[0141] Text Classification (TC): TC is a task that categorizes text into predefined categories. For example, it can identify whether the text content of several scanned documents belongs to a category such as a subscription agreement, trust deed, agency agreement, legal opinion, due diligence materials, or roadshow materials.

[0142] Entity Relation Extraction (ERE): The task of extracting entities and their relationships from text is categorized as "ERE." For example, extracting key elements from a prospectus for a publicly issued corporate bond.

[0143] Text Summarization (TS): The task category for generating text summaries is "TS." For example, generating a summary of "xxx Report."

[0144] Semantic Role Labeling (SRL): The task category of labeling the arguments of verbs in sentences and their semantic roles is "SRL".

[0145] Dialogue System (DS): The task category for building a dialogue system is "DS." For example, an intelligent customer service dialogue system.

[0146] Specifically, the reason tag / attribution tag [reason] may include, for example: 1. Significant decline in default scale: Compared to the same period last year, the scale of defaults has decreased significantly, indicating that credit risk in the bond market has been effectively controlled, with the frequency and scale of default events decreasing. 2. Convergence of credit risk: Convergence of credit risk means reduced uncertainty facing market participants, which helps to strengthen investor confidence and attract more funds to the bond market. 3. Enhanced market stability: A reduction in defaults is generally seen as increased market stability, which is of great significance for maintaining the healthy development of the bond market. Therefore, despite the presence of defaults and extensions, the significant decline in default scale and the convergence of credit risk indicate that the market is moving in a healthier direction.

[0147] Specifically, the result tag [result] may include, for example: the bond market performance in the second quarter of xx can be regarded as positive public opinion.

[0148] In this step, by designing the output of the three dimensions of "cause, effect, and category" and the unique conversation data of the relevant business fields of each slave computing node, it is possible to extract knowledge of certain specific business fields based on the large model deployed by the master computing node.

[0149] Step S303: Send the training data set to the corresponding slave computing node, so that the slave computing node performs model training based on the training data set and the original data set to obtain a target model.

[0150] In this step, each slave computing node can pre-configure a corresponding node identifier for each original data set, so that the master computing node can send the training data set obtained by knowledge distillation processing to the corresponding slave computing node according to the node identifier, so that the subsequent slave computing node can perform model training based on the local original data set and training data set.

[0151] The model training method based on federated transfer learning in this embodiment obtains a training data set by performing knowledge distillation on the original data set, and sends the training data set to the corresponding slave computing node. Subsequently, the slave computing node can perform multi-task learning on the local initial model based on the training data set and the local original data set, that is, perform cause training / attribution training, result training and category training respectively, and then determine the final loss according to the loss of each training, and optimize and adjust the model parameters based on the final loss, so that the target model obtained by the final training is more accurate and reliable.

[0152] Another embodiment of the present application provides a model training method based on federated transfer learning, which is applied to the master computing node. The overall process of model training is as follows:

[0153] Step S401: Installation and initialization of the federated transfer learning framework.

[0154] In this step, the financial infrastructure can be used as the master computing node / leader, and the components of the federated transfer learning collaborative network can be pre-deployed in the master computing node's local data center. Similarly, each financial institution can be used as a slave computing node / participant, and the components of the federated transfer learning collaborative network can be pre-deployed in the local data center of each slave computing node. The framework of the federated learning collaborative network can be implemented using other software products with similar functions, such as FATE, SecretFlow, PaddleFL, TensorFlowFederated, and Pysyft. The federated transfer learning network has a star topology, and each slave computing node only interacts bidirectionally with the master computing node.

[0155] Step S402, key interaction;

[0156] In this step, the main computing node can pre-generate an asymmetric key pair, namely a first key pair, which includes a first public key pk center And the first private key sk center , then the first public key pk center Sent to the slave computing node for the slave computing node to use based on the first public key pk center The original data set is encrypted to obtain a first encrypted data set.

[0157] Similarly, each slave computing node will also pre-generate an asymmetric key pair, namely the second key pair, which includes the second public key pk i node And the second private key sk i node Where i represents the i-th slave computing node, and then the slave computing node i can use the second public key pk i node Sent to the master computing node. That is, the master computing node receives the second public key sent by the slave computing node, thereby completing the key exchange.

[0158] Step S403, receiving a first encrypted data set sent from a computing node;

[0159] In this step, the first encrypted data is obtained from the computing node based on the first public key pk center The original data set is encrypted. That is, after completing the key interaction, the slave node will encrypt the data based on the first public key pk center The original data set is encrypted to obtain first encrypted data, and then the first encrypted data is sent to the main computing node.

[0160] Step S404: decrypt the first encrypted data set based on the first private key to obtain the original data set;

[0161] In this step, after receiving the first encrypted data, the main computing node will decrypt the first encrypted data based on the first private key held by itself, thereby obtaining the original data set.

[0162] Step S405: Based on the predetermined prompt word template and each raw data in the original data set, a predetermined large model is used to generate training data corresponding to each raw data to obtain a training data set; wherein each training data includes: any one or more of a category label, a result label, and a cause label;

[0163] In this step, during the specific implementation process, a prompt word template can be pre-generated based on a predetermined prompt method; wherein the predetermined prompt method includes any one or more of the following: a zero-sample prompt method, a small number of sample prompt methods, and a thought chain prompt method.

[0164] The zero-sample prompt method may be, for example:

[0165] prompt_CoT_Zero_shot = [input text]+'Please reason step by step and draw a conclusion. Reasoning needs to include the following three steps: first determine the task category of this input, then generate the output result of this input, and finally explain the reason for this output result.'

[0166] response_CoT_Zero_shot = openai.Completion.create(

[0167] model="text-davinci-004",

[0168] prompt=prompt_CoT_Zero_shot,

[0169] max_tokens=2500, )

[0171] Examples of small sample prompts include:

[0172] prompt_CoT_Few_shot = Q: "In the second quarter of 2024, the scale of risky bonds was 15.276 billion yuan, a year-on-year decrease of approximately 76%. Credit risk continues to converge. Is this good news for the bond market?"

[0173] A: "[label] Sentiment analysis. [result] Judging from these data, the bond market performance in the second quarter of 2024 can be considered positive. [reason] This data shows the following positive signals: 1. Significant decline in defaults: Compared with the same period last year, the scale of defaults has decreased significantly, indicating that credit risk in the bond market has been effectively controlled, with the frequency and scale of defaults decreasing. 2. Convergence of credit risk: Convergence of credit risk means reduced uncertainty facing market participants, which helps to strengthen investor confidence and attract more funds to the bond market. 3. Enhanced market stability: A reduction in defaults is generally seen as increased market stability, which is of great significance for maintaining the healthy development of the bond market. Therefore, despite the existence of defaults and extensions, the significant decline in defaults and the convergence of credit risk indicate that the market is moving in a healthier direction."

[0174] Q:'+ [input text]+'

[0175] A:'

[0176] response_CoT_Few_shot = openai.Completion.create(

[0177] model="text-davinci-004",

[0178] prompt=prompt_CoT_Few_shot,

[0179] max_tokens=2500, )

[0181] Specifically, when the main computing node performs knowledge distillation processing on the original data set, it can input each original data in the original data set and a predetermined prompt word template into a predetermined large model, and use the large model to output training data containing three label dimensions: category label, result label, and cause label according to the prompt word template, thereby obtaining a training data set corresponding to the original data set.

[0182] Step S406: encrypt the training data set based on the second public key to obtain a second encrypted data set;

[0183] In this step, after the master computing node obtains the training data set, it can use the second public key pk i node The training data set is encrypted to obtain a second encrypted data set. By encrypting the training data set, leakage of the training data set can be prevented and data privacy can be protected.

[0184] Step S407: Send the second encrypted data set to the slave computing node, so that the slave computing node decrypts the second encrypted data set based on the second private key to obtain a training data set;

[0185] In this step, after obtaining the second encrypted data, the master computing node can send the second encrypted data to the corresponding slave computing node, so that the slave computing node can perform category training, result training and cause training on the initial model based on the original data set and the training data set, respectively, to obtain the initial category loss, the initial result loss and the initial cause loss. At the same time, based on any one or more of the initial cause loss, the initial result loss and the initial category loss, the initial total loss is determined, and based on the initial total loss, the parameters of the local initial model of the slave computing node are optimized and adjusted to obtain the current model, until the predetermined training stop condition is met and the target model is obtained.

[0186] During the specific implementation of this embodiment, after obtaining the training dataset, a manual or AI audit process can be used to control incorrect outputs caused by problems such as hallucinations in the large model. Specifically, the training dataset can be audited manually or by AI to obtain audit information containing erroneous data and the causes of the errors in the training dataset. Furthermore, the main computing node can also receive the audit information for the training dataset and then adjust the large model based on the training dataset and the corresponding audit information to improve the accuracy of the knowledge distillation processing performed by the large model.

[0187] Specifically, the master computing node can also receive error data and error causes for the training data set fed back by each slave computing node. The master computing node can then use methods such as fine-tuning the large model, improving the encoder and decoder, enhancing the input data with external information, and adding post-processing links to improve the hallucination problem of its large model based on the fed-back error data.

[0188] The method in this embodiment obtains a training data set by performing knowledge distillation on the original data set, and sends the training data set to the corresponding slave computing node. Subsequently, the slave computing node can perform multi-task learning on the local initial model based on the training data set and the local original data set, that is, perform cause training / attribution training, result training and category training respectively, and then determine the final loss according to the loss of each training, and optimize and adjust the model parameters based on the final loss, so that the target model obtained by the final training is more accurate and reliable.

[0189] Another embodiment of the present application provides a model training device based on federated transfer learning, such as Figure 3 As shown, including:

[0190] The first receiving module 11 is configured to receive a training data set sent by the master computing node, which is obtained by performing knowledge distillation processing on the original data set of the slave computing node. The training data set includes training data corresponding to each original data in the original data set, and each training data includes: any one or more labels of a category label, a result label, and a cause label;

[0191] A training module 12 is configured to perform category training, result training, and cause training on the initial model based on the original data set and the training data set, respectively, to obtain an initial category loss, an initial result loss, and an initial cause loss;

[0192] The acquisition module 13 is used to determine the initial total loss based on any one or more of the initial category loss, the initial result loss, and the initial cause loss, and adjust the parameters of the initial model based on the initial total loss to obtain the current model, until the predetermined training stop condition is met, and the target model is obtained.

[0193] In the specific implementation process of this embodiment, the first receiving module is further used to: receive the first public key sent by the main computing node;

[0194] The model training device based on federated transfer learning also includes a first encryption module and a first sending module;

[0195] The first encryption module is used to: encrypt the local original data set based on the first public key to generate a first encrypted data set;

[0196] The first sending module is used to send the first encrypted data set to the main computing node, so that the main computing node decrypts the first encrypted data set based on the first private key corresponding to the first public key to obtain the original data set.

[0197] In a specific implementation of this embodiment, the model training device based on federated transfer learning further includes a generation module, which is used to pre-generate a second public key and a second private key corresponding to the second public key; the first sending module is further used to send the second public key to the master computing node, so that the master computing node encrypts the training data set based on the second public key to obtain a second encrypted data set;

[0198] The first receiving module is specifically configured to: receive a second encrypted data set sent by the main computing node; and decrypt the second encrypted data set based on a second private key to obtain a training data set.

[0199] During the specific implementation of this embodiment, the training module is specifically used to: construct a category training set containing category labels, a result training set containing result labels, and a cause training set containing cause labels based on the training data set; perform category training on the initial model based on the original data set and the category training set to obtain an initial category loss; perform result training on the initial model based on the original data set and the result training set to obtain an initial result loss; perform cause training on the initial model based on the original data set and the cause training set to obtain an initial cause loss.

[0200] During the specific implementation of this embodiment, the acquisition module is specifically used to: determine the first coefficient corresponding to the category label, the second coefficient corresponding to the result label, and the third coefficient corresponding to the cause label; based on the initial category loss, the initial result loss, the initial cause loss, the first coefficient, the second coefficient, and the third coefficient, determine the initial total loss.

[0201] During the specific implementation of this embodiment, the model training device based on federated transfer learning also includes a determination module, which is used to: determine whether a predetermined training stop condition is met based on the current loss corresponding to the current model; or determine whether a predetermined training stop condition is met based on the number of training rounds corresponding to the current model.

[0202] The device in this embodiment receives the training data set sent by the main computing node, which is obtained by knowledge distillation processing of the original data set. Subsequently, the slave computing node can perform multi-task learning on the local initial model based on the training data set and the local original data set, that is, perform cause training / attribution training, result training and category training respectively, and then determine the final loss according to the loss of each training, and optimize and adjust the model parameters based on the final loss, so that the target model obtained by the final training is more accurate and reliable.

[0203] Another embodiment of the present application provides a model training device based on federated transfer learning, such as Figure 4 As shown, including:

[0204] The second receiving module 21 is used to receive the original data set sent from the computing node;

[0205] A processing module 22 is configured to perform knowledge distillation on the original data set based on a predetermined prompt word template and a predetermined large model to obtain a training data set;

[0206] The second sending module 23 is used to send the training data set to the corresponding slave computing node, so that the slave computing node can perform model training based on the training data set and the original data set to obtain a target model.

[0207] In the specific implementation process of this embodiment, the model training device based on federated transfer learning further includes a key generation module, and the key generation module is used to: pre-generate a first public key and a first private key corresponding to the first public key;

[0208] The second sending module is further configured to send the first public key to the slave computing node, so that the slave computing node encrypts the original data set based on the first public key to obtain a first encrypted data set;

[0209] The second receiving module is specifically configured to: receive a first encrypted data set sent from a computing node; and decrypt the first encrypted data set based on a first private key to obtain an original data set.

[0210] During the specific implementation of this embodiment, the processing module is specifically used to: based on a predetermined prompt word template and each original data in the original data set, use a predetermined large model to generate training data corresponding to each original data to obtain a training data set; wherein each training data includes: any one or more labels among: category labels, result labels and cause labels.

[0211] During the specific implementation of this embodiment, the model training device based on federated transfer learning also includes a template generation module, which is used to: pre-generate a prompt word template based on a predetermined prompt method; wherein the predetermined prompt method includes any one or more of the following: zero-sample prompt method, small-sample prompt method, and thought chain prompt method.

[0212] In a specific implementation of this embodiment, the second receiving module is further configured to: receive a second public key sent from a computing node;

[0213] The model training device based on federated transfer learning further includes a second encryption module, which is used to: encrypt the training data set based on the second public key to obtain a second encrypted data set;

[0214] The second sending module is specifically configured to send the second encrypted data set to the slave computing node, so that the slave computing node decrypts the second encrypted data set based on the second private key to obtain the training data set.

[0215] The device in this embodiment obtains a training data set by performing knowledge distillation on the original data set, and sends the training data set to the corresponding slave computing node. The slave computing node can then perform multi-task learning on the local initial model based on the training data set and the local original data set, that is, perform cause training / attribution training, result training and category training respectively, and then determine the final loss according to the loss of each training, and optimize and adjust the model parameters based on the final loss, so that the target model obtained by the final training is more accurate and reliable.

[0216] Another embodiment of the present application provides an electronic device, such as Figure 5 As shown, it at least includes a memory 1 and a processor 2. The memory 1 stores a computer program. When the processor 2 executes the computer program on the memory 1, it implements the following method steps:

[0217] Step 1: Receive a training data set sent by the master computing node, which is obtained by performing knowledge distillation processing on the original data set of the slave computing node. The training data set includes training data corresponding to each original data in the original data set. Each training data includes: any one or more labels of a category label, a result label, and a cause label;

[0218] Step 2: Based on the original data set and the training data set, the initial model is trained on categories, results, and causes to obtain the initial category loss, initial result loss, and initial cause loss.

[0219] Step 3: Based on any one or more of the initial category loss, initial result loss, and initial cause loss, determine the initial total loss, and adjust the parameters of the initial model based on the initial total loss to obtain the current model, until the predetermined training stop condition is met, and the target model is obtained.

[0220] Alternatively, implement the following method steps:

[0221] Step 1: Receive the original data set sent from the computing node;

[0222] Step 2: Perform knowledge distillation on the original dataset based on the predetermined prompt word template and the predetermined large model to obtain a training dataset;

[0223] Step 3: Send the training data set to the corresponding slave computing node, so that the slave computing node can perform model training based on the training data set and the original data set to obtain the target model.

[0224] The specific implementation process of the above method steps can be found in any of the above-mentioned embodiments of the model training method based on federated transfer learning, and this embodiment will not be repeated here.

[0225] The electronic device in this embodiment receives the training data set sent by the main computing node, which is obtained by knowledge distillation processing of the original data set. Subsequently, the slave computing node can perform multi-task learning on the local initial model based on the training data set and the local original data set, that is, perform cause training / attribution training, result training and category training respectively, and then determine the final loss according to the loss of each training, and optimize and adjust the model parameters based on the final loss, so that the target model obtained by the final training is more accurate and reliable.

[0226] The above embodiments are merely exemplary embodiments of the present application and are not intended to limit the scope of the present application. The scope of protection of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and scope of protection of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present application.

Claims

1. A model training method based on federated transfer learning, applied to a slave computing node, characterized in that: include: receiving a training data set sent by the master computing node, which is obtained by performing knowledge distillation processing on the original data set of the slave computing node, wherein the training data set includes training data corresponding to each original data in the original data set, each training data including: a category label, a result label, and a cause label; the original data is conversation data; Based on the original dataset and the training dataset, the initial model is trained on chain distillation categories, chain distillation results, and chain distillation causes, respectively, to obtain the initial category loss, initial result loss, and initial cause loss; Based on the initial category loss, initial result loss, and initial cause loss, the chain distillation method is used to determine the initial total loss, and the parameters of the initial model are adjusted based on the initial total loss to obtain the current model. When the predetermined training stop condition is met, the target model is obtained. The training data set uses a predetermined large model to output training data of three label dimensions including a cause label, a result label, and a category label according to a prompt word template, so as to obtain a training data set corresponding to the original data set; Before receiving the training data set sent by the main computing node, the method further includes: Receiving a first public key sent by the master computing node, encrypting the local original data set based on the first public key to generate a first encrypted data set; Sending the first encrypted data set to the main computing node, so that the main computing node decrypts the first encrypted data set based on a first private key corresponding to the first public key to obtain the original data set; Before receiving the training data set sent by the main computing node, the method further includes: Pre-generating a second public key and a second private key corresponding to the second public key; Sending the second public key to the main computing node, so that the main computing node encrypts the training data set based on the second public key to obtain a second encrypted data set; The receiving, sent by the master computing node, of a training data set obtained by performing knowledge distillation processing on an original data set of a slave computing node specifically includes: receiving a second encrypted data set sent by the primary computing node; The second encrypted data set is decrypted based on the second private key to obtain a training data set.

2. The method according to claim 1, wherein The initial model is respectively subjected to chain distillation category training, chain distillation result training, and chain distillation cause training based on the original data set and the training data set to obtain the initial category loss, the initial result loss, and the initial cause loss, specifically including: Based on the training data set, a category training set containing category labels, a result training set containing result labels, and a cause training set containing cause labels are constructed respectively; Based on the original data set and the category training set, the initial model is trained to obtain the initial category loss; Based on the original data set and the result training set, the initial model is trained to obtain the initial result loss; Based on the original data set and the reason training set, the initial model is trained to obtain the initial reason loss.

3. The method according to claim 1, wherein The chain distillation method is used to determine the initial total loss based on the initial category loss, initial result loss, and initial cause loss, specifically including: Determine a first coefficient corresponding to the category label, a second coefficient corresponding to the result label, and a third coefficient corresponding to the cause label; An initial total loss is determined based on the initial category loss, the initial result loss, the initial cause loss, the first coefficient, the second coefficient, and the third coefficient.

4. The method according to any one of claims 1 to 3, wherein After obtaining the current model, the method further includes: Based on the current loss corresponding to the current model, determine whether the predetermined training stop condition is met; Alternatively, based on the number of training rounds corresponding to the current model, it is determined whether a predetermined training stop condition is met.

5. The model training method based on federated transfer learning according to claim 1, applied to a master computing node, is characterized in that: include: Receive the original data set sent from the computing node; Performing knowledge distillation on the original data set based on a predetermined prompt word template and a predetermined large model to obtain a training data set; The training data set is sent to the corresponding slave computing node, so that the slave computing node can perform model training based on the training data set and the original data set to obtain the target model.

6. A model training device based on federated transfer learning, characterized in that: include: A first receiving module is configured to receive a training dataset sent by a master computing node, obtained by performing knowledge distillation processing on an original dataset of a slave computing node. The training dataset includes training data corresponding to each original data in the original dataset, each training data including a category label, a result label, and a cause label. The original data is conversation data. The training dataset is configured to output training data including three label dimensions, namely, a cause label, a result label, and a category label, based on a prompt word template using a predetermined large model, thereby obtaining a training dataset corresponding to the original dataset. The first receiving module is further configured to: receive a first public key sent by the master computing node; The model training device based on federated transfer learning also includes a first encryption module and a first sending module; The first encryption module is used to: encrypt the local original data set based on the first public key to generate a first encrypted data set; The first sending module is used to: send the first encrypted data set to the main computing node, so that the main computing node decrypts the first encrypted data set based on the first private key corresponding to the first public key to obtain the original data set; The model training device based on federated transfer learning further includes a generation module, wherein the generation module is used to pre-generate a second public key and a second private key corresponding to the second public key; the first sending module is further used to send the second public key to the main computing node, so that the main computing node encrypts the training data set based on the second public key to obtain a second encrypted data set; The first receiving module is specifically configured to: receive a second encrypted data set sent by the main computing node; decrypt the second encrypted data set based on the second private key to obtain a training data set; The training module is used to perform chain distillation category training, chain distillation result training, and chain distillation cause training on the initial model based on the original dataset and the training dataset, respectively, to obtain the initial category loss, initial result loss, and initial cause loss; The acquisition module is used to determine the initial total loss using the chain distillation method based on the initial category loss, the initial result loss, and the initial cause loss, and adjust the parameters of the initial model based on the initial total loss to obtain the current model until the predetermined training stop condition is met to obtain the target model.

7. A model training device based on federated transfer learning according to claim 6, characterized in that: include: A second receiving module is used to receive the original data set sent from the computing node; A processing module is used to perform knowledge distillation on the original data set based on a predetermined prompt word template and a predetermined large model to obtain a training data set; The second sending module is used to send the training data set to the corresponding slave computing node, so that the slave computing node can perform model training based on the training data set and the original data set to obtain a target model.

8. An electronic device, characterized in that: The system comprises at least a memory and a processor, wherein a computer program is stored on the memory, and when the processor executes the computer program on the memory, the processor implements the steps of the model training method based on federated transfer learning as described in any one of claims 1 to 5.

Images