A server-side privacy protection data generation method and system

Abstract

The application relates to a server-side privacy protection data generation method and system, which comprises the following steps: a user uploads a task demand to a server, wherein the task demand comprises a task requirement and a data type; the server generates a prompt statement according to the task demand of the user; the server obtains a data set according to the prompt statement and generates a prediction result of a sample in the data set; the data set and the prediction result of the sample in the data set are sent to the user; the user obtains the prediction result of the sample in the data set by using a local small model, obtains the prediction accuracy of the local small model according to the prediction result sent by the server, obtains a loss value by using the prediction accuracy and the loss of the local small model; each sample is scored according to the loss value; if the score of the sample is greater than a threshold value, the sample is marked as a sample needing to be corrected; the server trains a generator by using the sample needing to be corrected, and the generator is regenerated by using the corrected generator.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, specifically to a server-side method and system for generating privacy-preserving data. Background Technology

[0002] Enhancing the performance of small user-side models in federated learning is fundamental to ensuring the efficiency of federated learning and designing stable models, playing a crucial role in improving model adaptability in dynamic data environments. Using a large model to guide the learning of small models and enhancing the generalization ability of small models through privacy-preserving data is a potential approach. However, the time-varying nature of user data distribution makes it difficult for small models to continuously adapt to new data, impacting model performance. How to improve the performance of small models in dynamic data environments while protecting user data privacy is a critical issue that urgently needs to be addressed in federated learning. Summary of the Invention

[0003] To address the above problems, in a first aspect of the present invention, a server-side privacy-preserving data generation method is provided, the method comprising the following steps:

[0004] Users upload task requirements to the server, which include task specifications and data types; the server generates prompt statements based on the user's task requirements.

[0005] The server retrieves the dataset based on the prompt and generates prediction results for the samples in the dataset; it then sends the dataset and the prediction results for the samples in the dataset to the user.

[0006] Users use a local small model to obtain prediction results for samples in the dataset, and obtain the prediction accuracy of the local small model based on the prediction results sent by the server. They then use the prediction accuracy and the loss of the local small model to obtain the loss value, and score each sample based on the loss value.

[0007] If a sample's score is greater than a threshold, the sample is marked as needing correction. The server then uses the samples that need correction to train the generator and regenerates samples using the corrected generator.

[0008] Preferably, the step of obtaining the loss value using prediction accuracy and the loss of the local small model specifically involves:

[0009] The loss function is defined as follows:

[0010] in, It is the prediction accuracy of the small model for the sample (x,y). Where C is the total number of categories, y c It is the one-hot encoding of the real label. It is the one-hot encoding of the prediction results of the small model, LKD (p S (x),p L (x) represents the knowledge distillation loss. p L (x) is the probability distribution of the local small model's predictions for the input x, p S (x) The probability distribution of the prediction of the large model in the server for the input x.

[0011] Preferably, the server trains the generator using samples that need correction, specifically:

[0012] The server will need to correct the dataset D. a ={x a ,y a Input the pre-set generator G au The loss function is obtained as follows:

[0013]

[0014] in, To reconstruct the loss, To generate data in a local small model Confidence loss on For the data that needs to be corrected (x) a ,y a After generator G aux The generated new data, θ aux For generator parameters, Let z be the perturbation data, λ be the random noise, and λ be the weighting coefficient.

[0015] Preferably, the server generates a prompt statement based on the user's task requirements, specifically as follows:

[0016] The server assembles the user's task description into a prompt statement: descrup = {prompt} + {task} + {type}, where prompt is the prompt added by the server, task is the task requirement, and type is the data type.

[0017] Preferably, the server obtains the dataset according to the prompt statement, specifically as follows:

[0018] The server-side large model provides a recommended dataset name based on the prompt and selects the corresponding dataset from the pre-integrated data pool on the server side; if the dataset does not exist on the server side, it downloads the dataset from the Internet and selects whether to add it to the data pool as needed.

[0019] In a second aspect of the present invention, a server-side privacy-preserving data generation system is provided, the system comprising the following modules:

[0020] The task upload module allows users to upload their task requirements to the server. These requirements include task specifications and data types. The server then generates prompts based on the user's task requirements.

[0021] The dataset acquisition module retrieves the dataset based on the prompts and generates prediction results for the samples in the dataset; it then sends the dataset and the prediction results for the samples in the dataset to the user.

[0022] The data evaluation module allows users to obtain prediction results for samples in the dataset using a local small model, and to obtain the prediction accuracy of the local small model based on the prediction results sent by the server. The loss value is obtained by using the prediction accuracy and the loss of the local small model. Each sample is scored based on the loss value.

[0023] In the generator training module, if a sample's score is greater than a threshold, the sample is marked as a sample that needs correction. The server then uses the sample that needs correction to train the generator and uses the corrected generator to regenerate the sample.

[0024] Preferably, the step of obtaining the loss value using prediction accuracy and the loss of the local small model specifically involves:

[0025] The loss function is defined as follows:

[0026] in, It is the prediction accuracy of the small model for the sample (x,y). Where C is the total number of categories, y c It is the one-hot encoding of the real label. It is the one-hot encoding of the prediction results of the small model, L KD (p S (x),p L (x) represents the knowledge distillation loss. p L (x) is the probability distribution of the local small model's predictions for the input x, p S (x) The probability distribution of the prediction of the large model in the server for the input x.

[0027] Preferably, the server trains the generator using samples that need correction, specifically:

[0028] The server will need to correct the dataset D. a ={x a ,y a Input the pre-set generator G aux The loss function is obtained as follows:

[0029]

[0030] in, To reconstruct the loss, To generate data in a local small model Confidence loss on For the data that needs to be corrected (x) a ,y a After generator G aux The generated new data, θ aux For generator parameters, Let z be the perturbation data, λ be the random noise, and λ be the weighting coefficient.

[0031] Preferably, the server generates a prompt statement based on the user's task requirements, specifically as follows:

[0032] The server assembles the user's task description into a prompt statement: descrup = {prompt} + {task} + {type}, where prompt is the prompt added by the server, task is the task requirement, and type is the data type.

[0033] Preferably, the server obtains the dataset according to the prompt statement, specifically as follows:

[0034] The server-side large model provides a recommended dataset name based on the prompt and selects the corresponding dataset from the pre-integrated data pool on the server side; if the dataset does not exist on the server side, it downloads the dataset from the Internet and selects whether to add it to the data pool as needed.

[0035] This invention optimizes the generated data by setting up a data generator on the server side, combining the filtering and enhancement of public datasets with a user feedback mechanism, thereby effectively improving the performance of the user-side small model. Furthermore, during the data generation process, the prediction results of the user-side small model and the guidance information from the large model are combined. This collaborative mechanism not only ensures the high quality and high relevance of the generated data but also significantly enhances the privacy data processing capabilities of the large model without compromising user data privacy, providing an efficient and reliable solution for data augmentation in federated learning. Attached Figure Description

[0036] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0037] In the embodiments of this invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0038] It is understood that the term "embodiment" used throughout the specification means that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this application. Therefore, various embodiments throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It is understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0039] In this invention, unless otherwise specified, the same or similar parts between the various embodiments can be referred to each other. In the various embodiments of this invention, and in the various implementation methods / methods / implementations within each embodiment, unless otherwise specified or logically conflicting, the terminology and / or descriptions between different embodiments and between the various implementation methods / methods / implementations within each embodiment are consistent and can be mutually referenced. The technical features in different embodiments and the various implementation methods / methods / implementations within each embodiment can be combined according to their inherent logical relationships to form new embodiments, implementation methods, methods, or implementation approaches. The embodiments described below do not constitute a limitation on the scope of protection of this application.

[0040] Example 1

[0041] like Figure 1 The server-side privacy-preserving data generation method shown includes the following steps:

[0042] S1, the user uploads the task requirements to the server, the task requirements include task specifications and data types; the server generates prompt statements based on the user's task requirements.

[0043] The user sends the specific requirements of the task and the data type to be processed to the server. After receiving the user's task requirements, the server generates a prompt statement based on this information, which tells the server the required data and its data type.

[0044] S2, the server retrieves the dataset according to the prompt and generates prediction results for the samples in the dataset; the dataset and the prediction results for the samples in the dataset are then sent to the user.

[0045] The server uses prompts to retrieve a dataset from its stored data pool or generates its own dataset based on the prompts. This dataset contains multiple samples. The server uses a pre-trained large model to process each sample in the retrieved dataset and obtains a corresponding prediction result. For image classification tasks, the prediction result might be which category the image belongs to. The server sends the original dataset and the prediction results for the samples in that dataset to the user.

[0046] S3: The user uses the local small model to obtain the prediction results of the samples in the dataset, and obtains the prediction accuracy of the local small model based on the prediction results sent by the server. The user then uses the prediction accuracy and the loss of the local small model to obtain the loss value, and scores each sample based on the loss value.

[0047] The user uses a local small model to predict each sample in the dataset sent by the server, obtaining the prediction result from the local small model. The user then compares the prediction result of the local small model with the prediction result sent by the server or the true label of the sample. The accuracy of the local small model's prediction is evaluated by comparing the consistency between the two.

[0048] The user further combines the loss generated by the local small model during the prediction process with the previously calculated prediction accuracy to calculate a comprehensive loss value. This loss value takes into account both the model's prediction error and uncertainty. Based on the calculated loss value, the user scores each sample in the dataset.

[0049] In one embodiment, obtaining the loss value using prediction accuracy and the loss of the local small model specifically involves:

[0050] The loss function is defined as follows:

[0051] in, It is the prediction accuracy of the small model for the sample (x,y). Where C is the total number of categories, y c It is either the one-hot encoding of the true label or the prediction result of samples in the dataset. It is the one-hot encoding of the prediction results of the small model, L KD (p S (x),p L (x) represents the knowledge distillation loss. p L (x) is the probability distribution of the local small model's predictions for the input x, p S (x) The probability distribution of the prediction of the large model in the server for the input x.

[0052] S4. If the score of a sample is greater than the threshold, the sample is marked as a sample that needs to be corrected. The server then uses the sample that needs to be corrected to train the generator and uses the corrected generator to regenerate the sample.

[0053] After users score each sample, if a sample's score exceeds a certain threshold, then that sample needs correction. Upon receiving this feedback, the server uses these samples to train a generator model. After training, the generator model's capabilities are improved and optimized. The server then uses this corrected generator model to regenerate a new batch of samples or to regenerate samples based on the original samples of the samples to be corrected.

[0054] In one embodiment, the server trains the generator using samples that need correction, specifically:

[0055] The server will need to correct the dataset D. a ={x a ,y a Input the pre-set generator G aux The loss function is obtained as follows:

[0056]

[0057] in, To reconstruct the loss, To generate data in a local small model Confidence loss on For the data that needs to be corrected (x) a ,y a After generator G aux The generated new data, θ aux For generator parameters, Let z be the perturbation data, λ be the random noise, and λ be the weighting coefficient.

[0058] Example 2

[0059] Step 1: The user uploads the task requirements.

[0060] Users upload their task requirements to the server, including task specifications (e.g., image classification, text generation, image retrieval) and data type (e.g., image, text). The server assembles the user's task description into a specific description descriptor:

[0061] descrp={prompt}+{task}+{type}

[0062] For example, for the sentiment classification task, dcp = "Recommend a dataset for me. The task requirement is: binary classification of sentiment (positive vs. negative), and the data type is: text". For the image classification task, dcp = "Recommend a dataset for me. The task requirement is: ten categories of daily necessities, and the data type is: image".

[0063] Step 2: The server selects and outputs a public dataset.

[0064] 1) Public Dataset Selection. The large model P recommends dataset names based on the task description dcp and selects the appropriate dataset from the pre-integrated data pool on the server side. If the dataset is not available on the server side, it is downloaded from the internet, and the model chooses whether to add the dataset as needed. Let the public dataset be denoted as... in This represents a set of task descriptions.

[0065] 2) Output representation of public datasets. For samples from public datasets... The large model outputs its corresponding probability distribution p L (x). For classification tasks, p L (x) represents the category distribution; for the prediction of the next word, p L (x) represents the predicted probability distribution of the words in the bag-of-words. Let x be all inputs and their corresponding prediction distributions p. L The data corresponding to (x) is constructed into a distillation dataset {x, p}. L (x)}. Since the distillation dataset is model-independent, it can be applied to different user-side models. To compress the dataset size, for natural language models, only the top k largest prediction probabilities can be selected as p. L (x). The server will store the dataset {x, p}. L (x)} is sent to the corresponding user according to the requirement description.

[0066] Step 3: The user generates a data evaluation and uploads it to the server.

[0067] 1) Prediction by the small model. The user inputs x into the local small model S. i Obtain the output probability distribution With predictive labels For classification tasks, S represents i The predicted class probability distribution; for autoregressive tasks, This represents the conditional probability distribution for word prediction.

[0068] 2) Definition of loss function. Loss function in This refers to the prediction accuracy of the small model for samples (x, y), typically achieved using cross-entropy loss.

[0069] Where C is the total number of categories, y c It is the one-hot encoding of the real label. It is the class probability predicted by the small model.

[0070] L KD (p S (x),p L (x) is the knowledge distillation loss, used to measure the consistency of the predictive power between small and large models, and is usually expressed as KL divergence.

[0071] Where β is the weight of the knowledge distillation loss, used to balance prediction accuracy and consistency.

[0072] 3) User-side rating. Users rate their scores based on the loss function. Each generated data sample is scored. The lower the score, the higher the quality of the generated data and the better it matches the user's needs.

[0073] Step 4: The server filters data based on user ratings.

[0074] Data sorting and filtering. The server sorts the dataset based on user ratings and sets a quality assessment threshold θ. Data with scores below the threshold θ is sent directly to the corresponding user; this data is considered high-quality and meets the user's needs. Data with scores above the threshold θ is marked by the server as needing correction.

[0075] Step 5: Server-enhanced data generation.

[0076] 1) Assisted generator training. The server inputs the data that needs correction into the pre-set generator G. aux Used to further optimize generated data based on user feedback.

[0077] 2) Data correction. The server uses the corrected generator G. aux Regenerate data {x ′ ,y}, and send the corrected data to the user for verification.

[0078] Step 6: The server sends the generated data to the corresponding user.

[0079] The user side does not process all generated data equally. Instead, each generated data sample is scored based on a loss function. The lower the score, the higher the quality of the generated data and the better it matches the user's needs. By defining a loss function, including prediction accuracy and knowledge distillation loss, the consistency of the predictive capabilities of small and large models is measured, thereby evaluating the generated data.

[0080] The server sorts the generated data based on user ratings and sets a quality assessment threshold θ. High-quality data with ratings below threshold θ is sent directly to the user, while data with ratings above threshold θ is marked as needing correction. Classifying data quality based on user feedback improves the efficiency and accuracy of data filtering, ensuring that only high-quality data that meets user needs is sent directly to the user.

[0081] The server inputs the data requiring correction into a pre-configured auxiliary generator, further optimizing the generated data based on user feedback. The auxiliary generator uses a loss function derived from generative adversarial networks (GANs). By training the auxiliary generator, the server can generate data that better meets user needs. After data correction, the server sends the corrected data to the user for verification, ensuring data quality and its alignment with user requirements. The principle is to leverage user feedback to drive generator optimization, thereby dynamically improving the quality of the generated data and ensuring a high degree of consistency with user needs.

[0082] Example 3

[0083] Suppose an e-commerce platform needs to classify user reviews by sentiment. In this case, the server chooses to use a privacy data augmentation method based on a small model guided by a large model for processing.

[0084] Step 1: The user uploads the task requirements.

[0085] The e-commerce platform uploads task requirements to the server. The task category is sentiment classification (positive / negative binary classification), and the data type is text (customer reviews). The server assembles the user's task description into a specific description descriptor p, where descriptor p = ...

[0086] {prompt}+{task}+{type} specifically means "Recommend a dataset for me. The task requirement is: binary classification of positive and negative sentiment. The data type is: text".

[0087] Step 2: The server selects and outputs a public dataset.

[0088] 1) Public Dataset Selection. The server recommends the IMDb movie review dataset (containing extensive text sentiment annotations) based on the task description and selects the appropriate dataset from a pre-integrated data pool on the server side. If the server does not have the dataset, it downloads it from the internet, and the server chooses whether to add the dataset as needed. Let the public dataset be denoted as... in This represents a set of task descriptions.

[0089] 2) Output representation of public datasets. For samples from the public dataset IMDb... Where y is the sample label, and the large model outputs its corresponding probability distribution p. L (x). For classification tasks, p L (x) represents the category distribution; for the prediction of the next word, p L (x) represents the predicted probability distribution of the words in the bag-of-words. Let x be all inputs and their corresponding prediction distributions p. L The data corresponding to (x) is constructed into a distillation dataset {x, p}. L (x)}. Since the distillation dataset is model-independent, it can be applied to different user-side models. To compress the dataset size, for natural language models, only the top 5 largest prediction probabilities can be selected as p. L (x). The server will store the dataset {x, p}. L (x)}(including original comment x + probability distribution p) L (x) is sent to the e-commerce platform.

[0090] Step 3: The user generates a data evaluation and uploads it to the server.

[0091] 1) Prediction using a small model. The e-commerce platform inputs x into the local small model S. i For example, the lightweight BERT model performs sentiment classification to obtain the output probability distribution. With predictive labels For classification tasks, S represents i The predicted class probability distribution; for autoregressive tasks, This represents the conditional probability distribution for word prediction.

[0092] 2) Definition of loss function. Loss function in This refers to the prediction accuracy of the small model for samples (x, y). Preferably, cross-entropy loss is used.

[0093] Where C is the total number of categories, y c It is the one-hot encoding of the real label. It is the class probability predicted by the small model.

[0094] L KD (p S (x),p L (x) represents the knowledge distillation loss, used to measure the consistency of the predictive capabilities of small and large models. Preferably, KL divergence is used.

[0095] Here, β is the weight of the knowledge distillation loss, used to balance prediction accuracy and consistency.

[0096] 3) User-side rating. E-commerce platforms use a loss function... Each generated data sample is scored. The lower the score, the higher the quality of the generated data and the better it matches the user's needs.

[0097] Step 4: The server filters data based on user ratings.

[0098] Data sorting and filtering. The server sorts the dataset based on the e-commerce platform's ratings and sets a quality assessment threshold θ. Data with ratings below the threshold θ is sent directly to the e-commerce platform; this data is considered high-quality and meets user needs. Data with ratings above the threshold θ is marked by the server as needing correction.

[0099] Step 5: Server-enhanced data generation.

[0100] 1) Assisted generator training. The server inputs the data that needs correction into the pre-set generator G. aux Used to further optimize generated data based on user feedback.

[0101] 2) Data correction. The server uses the corrected generator G. aux Regenerate data {x ′ ,y}, and send the corrected data to the e-commerce platform for verification.

[0102] Step 6: The server will send the generated data to the corresponding user.

[0103] The above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0104] The steps of the methods or algorithms described in the embodiments of this application can be directly embedded in hardware, software units executed by a processor, or a combination of both. The software units can be stored in RAM, flash memory, ROM, EPROM, EEPROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium in the art. Exemplarily, the storage medium can be connected to the processor so that the processor can read information from and write information to the storage medium. Optionally, the storage medium can also be integrated into the processor. The processor and storage medium can be housed in an ASIC.

[0105] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0106] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of the claims and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A server-side privacy-preserving data generation method, characterized by, The method comprises the following steps: The user uploads a task requirement to the server, the task requirement comprising a task requirement and a data type; the server generates a prompt statement according to the task requirement of the user; The server acquires a data set according to the prompt statement and generates a prediction result of a sample in the data set; the data set and the prediction result of the sample in the data set are sent to the user; The user obtains the prediction result of the sample in the data set by using a local small model, obtains a prediction accuracy of the local small model according to the prediction result sent by the server, obtains a loss value by using the prediction accuracy and a loss of the local small model, and scores each sample according to the loss value; The generated data is sorted according to the score on the user side, high-quality data with a score lower than a threshold value is directly sent to the user, if the score of a sample is greater than the threshold value, the sample is marked as a sample needing to be corrected, the server trains a generator by using the sample needing to be corrected, and the sample is regenerated by using the corrected generator, after the data is corrected, the server sends the corrected data to the user for verification; The server trains the generator by using the sample needing to be corrected, specifically as follows: The server will need to correct the data set Input pre-set generator The loss function is obtained: ； where, is the reconstruction loss, is the confidence loss of the generated data on the local small model , is the data that needs to be corrected by the generator , is the generator parameter, is the perturbed data, z is random noise, is the weight coefficient.

2. The method of claim 1, wherein, The loss value is obtained by using the prediction accuracy and the loss of the local small model, specifically as follows: The loss function is defined as: ; wherein, is a small model the prediction loss for a sample , wherein is the total number of classes, is the one-hot encoding of the true label, is the one-hot encoding of the small model prediction, is the knowledge distillation loss, , is the prediction probability distribution of the local small model for input x, is the prediction probability distribution of the server large model for input x.

3. The method of claim 1, wherein, The server generates a prompt statement according to the task requirement of the user, specifically as follows: The server assembles the user's task description into a prompt statement: where prompt is the prompt added by the server, task is the task requirement, and type is the data type.

4. The method of claim 1, wherein, The server acquires a data set according to the prompt statement, specifically as follows: The server-side large model gives a recommended data set name according to the prompt statement, and selects a corresponding data set from a pre-integrated data pool on the server side; if the server side does not exist, the data set is downloaded from the Internet and whether to add to the data pool is selected according to the requirement.

5. A server-side privacy-preserving data generation system, characterized by, The system comprises the following modules: A requirement uploading module, the user uploads a task requirement to the server, the task requirement comprising a task requirement and a data type; the server generates a prompt statement according to the task requirement of the user; A data set acquiring module, the server acquires a data set according to the prompt statement and generates a prediction result of a sample in the data set; the data set and the prediction result of the sample in the data set are sent to the user; A data evaluation module, the user obtains the prediction result of the sample in the data set by using a local small model, obtains a prediction accuracy of the local small model according to the prediction result sent by the server, obtains a loss value by using the prediction accuracy and a loss of the local small model, and scores each sample according to the loss value; A generator training module, the generated data is sorted according to the score on the user side, high-quality data with a score lower than a threshold value is directly sent to the user, if the score of a sample is greater than the threshold value, the sample is marked as a sample needing to be corrected, the server trains a generator by using the sample needing to be corrected, and the sample is regenerated by using the corrected generator, after the data is corrected, the server sends the corrected data to the user for verification; The server trains the generator by using the sample needing to be corrected, specifically as follows: The server will need to correct the data set Input pre-set generator , get loss function: ； in, To reconstruct the loss, To generate data in a local small model Confidence loss on For data that needs correction After generator The generated new data, For generator parameters, For perturbation data, z is random noise. These are the weighting coefficients.

6. The system of claim 5, wherein, The loss value is obtained by using the prediction accuracy and the loss of the local small model, specifically as follows: The loss function is defined as: ; wherein is the prediction accuracy of the small model on the sample , wherein is the total number of classes, is the one-hot encoding of the true label, is the one-hot encoding of the small model prediction result, is the knowledge distillation loss, , is the prediction probability distribution of the local small model for the input x, is the prediction probability distribution of the large model in the server for the input x.

7. The system of claim 5, wherein, The server generates a prompt statement according to the task requirement of the user, specifically as follows: The server assembles the user's task description into a prompt statement: where prompt is the prompt added by the server, task is the task requirement, and type is the data type.

8. The system of claim 5, wherein, The server acquires a data set according to the prompt statement, specifically as follows: The server-side large model gives a recommended data set name according to the prompt statement, and selects the corresponding data set from the server-side pre-integrated data pool; if the server side does not exist, the data set is downloaded from the Internet and whether to add to the data pool is selected according to the demand.

Images