Data generation method and device, electronic equipment, storage medium and product

Abstract

The invention discloses a data generation method and device, electronic equipment, a storage medium and a product. The method comprises the steps that a to-be-processed data set is acquired; a target data set is obtained according to the to-be-processed data set, the target data set comprises a first target data set and a second target data set, the first target data set comprises data obtained after data dimension reduction processing and / or data sampling processing is carried out on data in a privacy data set, and the second target data set comprises data obtained after data dimension reduction processing and / or data sampling processing is carried out on the data in the privacy data set; the data included in the second target data set is different from that included in the first target data set; and processing the target data set by using a randomness constraint enhancement model to generate a target synthetic data set.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a data generation method, device, electronic device, storage medium and product. Background Art

[0002] With the development of artificial intelligence technology, large models have become an important tool in many fields, capable of implementing many complex tasks, such as natural language processing, image recognition, and recommendation systems. However, the training of large models usually requires massive amounts of data, which may contain users' private information. Privacy protection has become one of the important issues in the research and application of large models. Privacy protection aims to ensure that users' private information will not be used or leaked by bad parties when using large models for training and reasoning.

[0003] In the related art, a privacy-preserving data generation method based on Generative Adversarial Networks (GAN) is used for training and data generation. The method includes the following steps: constructing a pre-autoencoder; constructing a GAN model; and applying the autoencoder to data generation of GAN. The method uses a pre-trained autoencoder to map a given data set to a low-dimensional continuous space, and performs adversarial learning through the generator in the low-dimensional space and the discriminator in the original data space to obtain a generative model with simulated real data. This method can effectively solve the problem of lack of training data sets due to restrictions on obtaining training data under privacy protection policies.

[0004] However, the data synthesized by the GAN model cannot fully conform to the statistical characteristics of real data and cannot protect privacy information. Summary of the invention

[0005] The present application provides a data generation method, device, electronic device, storage medium and product to solve the problem in the related art that the data synthesized by the GAN model cannot fully conform to the statistical characteristics of the real data and cannot protect privacy information.

[0006] The technical solution of this application is implemented as follows:

[0007] A data generation method, comprising:

[0008] Get the data set to be processed;

[0009] According to the data set to be processed, a target data set is obtained, wherein the target data set includes: a first target data set and a second target data set, wherein the first target data set includes data obtained after performing data dimension reduction processing and / or data sampling processing on the data in the privacy data set, and the second target data set includes data different from the first target data set;

[0010] The target data set is processed using a random constraint enhancement model to generate a target synthetic data set.

[0011] A data generating device, comprising:

[0012] The second acquisition module is used to acquire the data set to be processed;

[0013] A second processing module is used to obtain a target data set according to the data set to be processed, wherein the target data set includes: a first target data set and a second target data set, wherein the first target data set includes data obtained by performing data dimension reduction processing and / or data sampling processing on the data in the privacy data set, and the second target data set includes data different from the first target data set;

[0014] The second processing module is used to process the target data set using a random constraint enhancement model to generate a target synthetic data set.

[0015] An electronic device, comprising:

[0016] Processor and communication interface; wherein,

[0017] The processor is configured to:

[0018] Acquire a data set to be processed through the communication interface; obtain a target data set based on the data set to be processed, and the target data set includes: a first target data set and a second target data set, the first target data set includes data obtained after data dimension reduction processing and / or data sampling processing is performed on the data in the privacy data set, and the second target data set is different from the data included in the first target data set; use a random constraint enhancement model to process the target data set to generate a target synthetic data set.

[0019] A storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the data generation method as described above.

[0020] A computer product comprises a computer program, wherein when the computer program is executed by a processor, the steps of the above-mentioned data generation method are implemented.

[0021] The present application provides a data generation method, which obtains a data set to be processed; obtains a target data set according to the data set to be processed, and the target data set includes: a first target data set and a second target data set, the first target data set includes data obtained after data dimension reduction and / or data sampling processing of the data in the privacy data set, and the second target data set is different from the data included in the first target data set; the target data set is processed using a random constraint enhancement model to generate a target synthetic data set. The present application proposes a random constraint enhancement model, which can generate samples with more diversity and creativity, and generate a more realistic target synthetic data set that conforms to the statistical characteristics of real data; at the same time, before processing the data set to be processed, the data set to be processed is converted into a target data set and then input into the random constraint enhancement model for processing, thereby achieving privacy protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flowchart of a data generation method provided in an embodiment of the present application;

[0023] Figure 2 A flowchart of a model training method provided in an embodiment of the present application;

[0024] Figure 3 A flowchart of data processing provided in an embodiment of the present application;

[0025] Figure 4 An architectural diagram of a data processing system provided in an embodiment of the present application;

[0026] Figure 5 A schematic diagram of the structure of a model training device provided in an embodiment of the present application;

[0027] Figure 6 A schematic diagram of the structure of a data generation device provided in an embodiment of the present application;

[0028] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0030] It should be understood that the "embodiments of the present application" or "the aforementioned embodiments" mentioned throughout the specification mean that specific features, structures or characteristics related to the embodiments are included in at least one embodiment of the present application. Therefore, "in the embodiments of the present application" or "in the aforementioned embodiments" appearing throughout the specification may not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. In the various embodiments of the present application, the size of the sequence number of the above-mentioned processes does not mean 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 the present application. The above-mentioned sequence numbers of the embodiments of the present application are for description only and do not represent the advantages and disadvantages of the embodiments.

[0031] It should be pointed out that the terms "first\second\third" involved in the embodiments of the present application are merely to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0032] Reference Figure 1 As shown, the data generation method provided by the present application comprises the following steps:

[0033] Step 101: Obtain a data set to be processed.

[0034] In practical applications, the data set to be processed is used as real data in actual scenarios. Before the application processes the data set to be processed, the random constraint enhancement model can be trained to improve the diversity and creativity of the samples generated by the random constraint enhancement model when processing the data set to be processed.

[0035] Step 102: Obtain a target data set based on the data set to be processed, where the target data set includes: a first target data set and a second target data set, where the first target data set includes data obtained by performing data dimensionality reduction processing and / or data sampling processing on the data in the privacy data set, and the second target data set includes different data from the first target data set.

[0036] Step 103: Process the target data set using the random constraint enhancement model to generate a target synthetic data set.

[0037] In practical applications, the generator and discriminator network structures are defined for the Randomness Constraint Enhanced GAN (RCE-GAN), and their parameters are initialized. By introducing randomness constraints, RCE-GAN can make the generated target synthetic datasets more diverse and creative.

[0038] The present application provides a data generation method, which obtains a data set to be processed; obtains a target data set according to the data set to be processed, and the target data set includes: a first target data set and a second target data set, the first target data set includes data obtained after data dimension reduction and / or data sampling processing of the data in the privacy data set, and the second target data set is different from the data included in the first target data set; the target data set is processed using a random constraint enhancement model to generate a target synthetic data set. The present application proposes a random constraint enhancement model, which can generate samples with more diversity and creativity, and generate a more realistic target synthetic data set that conforms to the statistical characteristics of real data; at the same time, before processing the data set to be processed, the data set to be processed is converted into a target data set and then input into the random constraint enhancement model for processing, thereby achieving privacy protection.

[0039] In practical applications, the trained generator and discriminator models can be used to process the dataset to be processed to generate new samples. Since the real dataset is used to train the generator and discriminator models of RCE-GAN during the training process, the generator can generate synthetic data similar to the real data, and the discriminator is responsible for distinguishing between the real data and the synthetic data. Through adversarial training, the generator gradually learns to generate high-quality target synthetic datasets.

[0040] The present application provides a data generation method, which obtains a data set to be processed; obtains a target data set according to the data set to be processed, and the target data set includes: a first target data set and a second target data set, the first target data set includes data obtained after data dimension reduction and / or data sampling processing of the data in the privacy data set, and the second target data set is different from the data included in the first target data set; the target data set is processed using a trained random constraint enhancement model to generate a target synthetic data set. In the early stage of this application, a large model sample data set (i.e., a real data set) in the same field as the data set to be processed is used to train the generator and discriminator models of RCE-GAN. Through adversarial training, the generator gradually learns to generate high-quality target synthetic data sets.

[0041] In some embodiments, step 103 processes the target data set using the random constraint enhancement model to generate a target synthetic data set, including:

[0042] The processing data set is processed using the random constraint enhancement model, temperature regulator and target loss function to generate a target synthetic data set. The temperature regulator is used to adjust the discreteness of the generated target synthetic data set, and the target loss function is used to measure the distance between the target synthetic data set and the target data set.

[0043] In practical applications, in the process of generating the target synthetic data set, a temperature regulator (TC) and a target loss function can be introduced to process the data set to be processed, so as to optimize the quality and diversity of the generated target synthetic data set to a certain extent.

[0044] Among them, the temperature regulator is used to adjust the discreteness of the generated target synthetic data set, and can also be regarded as used to adjust the temperature of the target synthetic data set.

[0045] In practical applications, the temperature parameter is defined. Here, Softmax is used as the temperature function, i.e., the temperature regulator.

[0046]

[0047] The temperature parameter T is usually a positive number that measures the diversity of the generated data. A larger temperature value will result in more dispersed generated data, while a smaller temperature value will result in more concentrated generated data.

[0048] Here, the output of the generator is normalized. For each element i of the generator output, first calculate its exponential value exp(z i / T), and then divided by the sum of all exponential values ​​∑ j exp(z j / T) to obtain the standardized probability distribution q i , where j is the index used to iterate over all the elements output by the generator.

[0049] In practical applications, adjust the standardized probability distribution. Divide each element of the standardized probability distribution by the value of the temperature parameter T. A larger temperature value will amplify the probability of each element, thereby increasing diversity; a smaller temperature value will reduce the probability of each element, making the generated data closer to the real data distribution.

[0050] Among them, the target loss function is used to measure the distance between the target synthetic dataset and the target dataset.

[0051] In practical applications, for each generated sample, sampling is performed according to the adjusted probability distribution. Random sampling can be performed according to the probability of the generated sample in the adjusted probability distribution, so that the synthetic data can be more diverse or closer to the original data distribution to obtain the final generated data.

[0052] In actual application, the following loss function is used as the target loss function

[0053]

[0054] Where m represents the dimension of the data or the number of samples; i represents an index variable, which is used to traverse each element or each sample. y represents the feature vector of the original image (for example, the vector obtained by converting the grayscale value of each pixel in the image). Represents the predicted value of the image feature vector generated by a neural network.

[0055] The objective loss function can be used to measure the distance between the generated data and the original data, such as mean square error or cross entropy. During the optimization process, the temperature parameter and the weight of the objective function can be adjusted to balance diversity and authenticity.

[0056] In practical applications, the trained RCE-GAN generator is used in combination with the temperature regulator and the objective function to generate synthetic data, which has similar distribution characteristics to the real data set.

[0057] In some embodiments, step 103 processes the target data set using the random constraint enhancement model. Before generating the target synthetic data set, the random constraint enhancement model is trained by the following steps:

[0058] Reference Figure 2 As shown, the model training method provided in this application includes the following steps:

[0059] Step 201: During each iterative training process, a large model sample data set and a first random noise are obtained.

[0060] In practical application, the generator and discriminator network structures are defined for the randomness constraint enhanced GAN (RCE-GAN), and their parameters are initialized. This application introduces an additional randomness constraint, namely the first random noise, into the input of the generator. In practical application, this can be achieved by adding a random noise vector to the input of the generator (i.e., a large model sample data set, which can be represented in the form of a vector), and connecting the vector to the original input vector (i.e., the vector corresponding to the large model sample data set).

[0061] Step 202: Input the large model sample data set and the first random noise into the generator of the randomness constraint enhancement model to generate a synthetic sample data set.

[0062] In practical applications, by introducing random constraints, RCE-GAN can make synthetic sample datasets (also known as generated samples) more diverse and creative.

[0063] Step 203: Acquire a synthetic sample data set connected to a second random noise and a large model sample data set connected to a second random noise, where the second random noise is noise randomly sampled from a predefined random noise distribution.

[0064] In practical applications, in each training iteration, some noise (here represented by noise vectors) is randomly sampled from a predefined random noise distribution and connected to the large model sample dataset or synthetic sample dataset. The discriminator needs to distinguish between the synthetic sample dataset connected with random noise and the large model sample dataset. This can force the generator to have a certain degree of randomness when generating samples, thereby increasing the diversity of samples.

[0065] Step 204: input the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise into the discriminator of the randomness constraint enhancement model to obtain a distinction result.

[0066] In practical application, the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise are input into the discriminator of the random constraint enhancement model, and the discriminator performs distinction to obtain a distinction result.

[0067] Step 205: Update the parameters of the generator and the discriminator according to the differentiation result until the iteration stop condition is met and the training is stopped.

[0068] In actual application, the iteration stop condition includes but is not limited to reaching the preset number of iterations. The parameters of the generator and the discriminator are updated according to the differentiation results until the iteration stop condition is met, the iteration is stopped, the model training is completed, and the trained random constraint enhancement model is obtained.

[0069] The present application obtains a large model sample data set and a first random noise during each iterative training process; inputs the large model sample data set and the first random noise into a generator of a randomness constraint enhancement model to generate a synthetic sample data set; obtains a synthetic sample data set connected to a second random noise and a large model sample data set connected to the second random noise, wherein the second random noise is noise randomly sampled from a predefined random noise distribution; inputs the synthetic sample data set connected to the second random noise and the large model sample data set connected to the second random noise into a discriminator of the randomness constraint enhancement model to obtain a distinction result; and updates the parameters of the generator and the discriminator according to the distinction result until the iteration stop condition is met, thereby obtaining a trained randomness constraint enhancement model. This application proposes a randomness constraint enhancement model. By introducing a randomness constraint, namely, a first random noise, the randomness constraint enhancement model can generate samples that are more diverse and creative. At the same time, in each training iteration, some noise, namely, a second random noise, is randomly sampled from a predefined random noise distribution and connected to a synthetic sample data set and a large model sample data set. The discriminator distinguishes between the synthetic sample data set connected to the second random noise and the original sample data set. This can force the generator to have a certain degree of randomness when generating samples, further increasing the diversity of the synthetic sample data set.

[0070] In some embodiments, the large model sample data set includes: a first sample data set and a second sample data set, the first sample data set includes data obtained after data dimension reduction and / or data sampling processing is performed on the data in the privacy data set, and the second sample data set includes different data from the first sample data set.

[0071] In practical applications, the large model sample data set is divided into two parts: the first sample data set (processed privacy data set) and the second sample data set (real data set). The real data set can be used directly to train the model. This application further performs data dimension reduction and data sampling on the privacy data set for training the model. Among them, data dimension reduction refers to reducing the number and identifiability of sensitive features through methods such as feature selection or principal component analysis. Data sampling refers to reducing the amount and identifiability of data by using random sampling or specific sampling methods.

[0072] In actual application, the data included in the large model sample data set comes from the following fields: research institutions, healthcare institutions, financial institutions, data service providers, etc. The data collection has been authorized by the users.

[0073] In some embodiments, the above-mentioned model training method also includes: during each iterative training process, adding a random regularization term to the loss function of the random constraint enhancement model to constrain the complexity of the random constraint enhancement model.

[0074] In practical applications, in addition to the adversarial loss function, a random regularization term is introduced into the loss function during training. This random regularization term can be achieved by minimizing the difference between the generated samples and the random noise, and the difference can be measured using the L1 or L2 norm.

[0075] In some embodiments, step 204 inputs the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise into the discriminator of the randomness constraint enhancement model, and before obtaining the distinction result, it also includes:

[0076] The first random noise is passed to the discriminator to enable the search discriminator to evaluate the randomness of the synthetic sample data set.

[0077] In practical applications, when training the generator, the first random noise is passed to the discriminator, so that the discriminator can identify the randomness in the generated samples and further improve the diversity of the generated samples by evaluating the randomness of the generated samples.

[0078] In a realistic scenario, refer to Figure 3 The flowchart of data processing shown in the figure is as follows: the first stage: data preparation; in this stage, the large model sample data set is divided into two parts: the first sample data set (the processed privacy data set) and the second sample data set (the real data set), where the first sample data set is the data set obtained after data dimension reduction and data sampling processing on the privacy data set. The second stage: RCE-GAN training; in this stage, by introducing random constraints, RCE-GAN can make the generated samples more diverse and creative, and through the random regularization term, the diversity of the generated samples can be controlled to a certain extent. The third stage: data synthesis using the temperature regulator, objective function, and trained RCE-GAN model; in this stage, the quality and diversity of the generated synthetic data can be optimized to a certain extent by combining the trained RCE-GAN generator and temperature regulator TC, as well as optimizing the objective function and iterative training.

[0079] In a feasible scenario, combining RCE-GAN with a large model solution can provide a more powerful and efficient generative model. Figure 4The data processing system architecture diagram shown in FIG. 1 includes an RCE-GAN training subsystem, a large model subsystem, and a secure multi-party computing (SMC) subsystem. The RCE-GAN training subsystem includes an RCE-GAN training module for performing RCE-GAN model training; and a data synthesis module for generating a target synthetic data set using a trained RCE-GAN model. The large model subsystem includes a distributed training module, a high-performance computing module, a pre-training and transfer learning module, a model compression and acceleration module, and a model fusion and integration module.

[0080] Distributed training module: Generative adversarial networks usually require a lot of computing resources and data to train complex generative models. Combined with the large model solution, the training process of RCE-GAN can be distributed to multiple computing nodes or distributed systems for efficient training. Each node can be responsible for training part of the data or model parameters, and then update the generative model through a communication protocol (such as Http).

[0081] High-performance computing module: During the training process of adversarial generative networks, a large amount of computing resources are usually required to process complex models and large-scale data. In combination with large model solutions, high-performance computing platforms or cloud services can be used to provide powerful computing capabilities to accelerate the training process of adversarial generative networks, while improving the performance of the generative model and the quality of the generated samples.

[0082] Model compression and acceleration module: For large generative models, model compression and acceleration are key to improving performance and efficiency. By combining large model solutions, model compression techniques such as pruning and quantization can be applied to reduce the size and computational complexity of the generative model. This can not only speed up the training and inference process, but also reduce the resource requirements of the generative model.

[0083] Pre-training and transfer learning module: The large model solution can use massive data for pre-training and transfer learning, and use part of the weights or features of the trained large model for training the adversarial generative network. This can provide better initialization parameters and stronger initial generation effects, thereby accelerating the training process of the generative model.

[0084] Model fusion and integration module: By combining different large models, the model fusion and integration strategy can be implemented to improve the performance and robustness of the generated model. Multiple generated models can be trained at the same time, and their generated results can be combined through model fusion techniques such as voting and weight fusion, thus providing more diverse and realistic generated samples.

[0085] By combining the generative adversarial network with the large model solution, the training and generation capabilities of the generative model can be enhanced, the quality and diversity of the generated samples can be improved, and it can be flexibly combined with specific application scenarios.

[0086] This application applies secure multi-party computing technology to large model data privacy protection. In the SMC subsystem, there are multiple participants, such as data owners, data analysts, and result verifiers. The participants collaborate through encryption technology to complete the computing task, and multiple participants can safely collaborate to train models and generate synthetic data without directly exposing individual sensitive data or model details. To complete the computing task, the SMC subsystem includes a data partitioning and local training module, an asynchronous communication and model parameter update module, a secure computing protocol and model evaluation module, a model trade-off and protection layer design module, and an evaluation and verification privacy protection module.

[0087] The above data partitioning and local training modules: First, the participants divide the original data into multiple parts, each of which is maintained by one participant. This ensures that sensitive data is not concentrated in one location, reducing the risk of privacy leakage. Then, each participant uses RCE-GAN to train the generative network locally to generate realistic synthetic data. Local training ensures the absolute control of the participants over the data while avoiding direct sharing of data.

[0088] The above asynchronous communication and model parameter update module: Participants interact with other participants using asynchronous communication protocols, but do not share original data. By exchanging the training and generation results of the generative network, participants can collect different data creation and distribution characteristics from other parties, thereby improving the performance of the generative network. Participants can use secure multi-party computing technology to encrypt or perturb model parameters, and then jointly calculate the aggregated or updated model parameters to protect model privacy.

[0089] The above secure computing protocol and model evaluation module: Participants can use secure multi-party computing technology to calculate model parameters. The secure computing protocol can ensure that calculations are performed without exposing individual data or model parameters, and guarantee the correctness and privacy protection of the results. By entrusting the calculation of model parameters to the secure computing protocol, participants can jointly obtain updated model parameters while ensuring the privacy of the model.

[0090] The above model trade-offs and protection layer design module: To protect the privacy of the model, participants can extract model sensitive information into a specific protection layer and perform additional encryption or perturbation processing on it. This can prevent malicious attackers from inferring the original data through model parameters. By designing appropriate protection layers and privacy measures, the privacy of the model can be protected while maintaining model performance.

[0091] The privacy protection module for the above evaluation and verification: In the process of collaborative training of models and generation of synthetic data, secure multi-party computing technology can be used to protect the privacy of evaluation and verification during the training process. Through secure computing protocols, participants can jointly verify the quality of generated data and model performance without exposing individual data or model details. This can meet compliance requirements while ensuring that specific data or model details cannot be inferred.

[0092] In traditional secure multi-party computing, participants need to collaborate through multiple communications and calculations, which leads to low computing efficiency. In order to solve the problem of low efficiency, this application takes the following measures to improve computing efficiency:

[0093] Compression technology: In secure multi-party computing, data transmission is an important performance bottleneck. Compression technology can effectively reduce the amount of data transmitted, thereby improving efficiency. This can include the use of compression algorithms such as LZ77, DEFLATE, etc. to reduce the size of data. In addition, data compression technology can also be used to reduce communication overhead. Compression technology can improve the efficiency of secure multi-party computing by reducing the amount of data transmission and improving communication bandwidth utilization.

[0094] Parallel computing: Decomposing a secure multi-party computing task into multiple subtasks and executing them in parallel on multiple computing resources can improve computing efficiency. Through parallel computing, more computing tasks can be processed in the same amount of time, thus speeding up the entire computing process. By allocating computing resources and making full use of parallel computing capabilities, the time and cost of computing can be significantly reduced.

[0095] Optimizing protocol design: Protocol design is a key factor affecting the efficiency of secure multi-party computing. Optimizing protocol design can reduce the complexity of communication and calculation, thereby improving efficiency. This application selects a more efficient algorithm, the Shamir secret sharing protocol, which divides the input value into multiple parts and distributes them to the participants. The participants perform calculations together without exposing the original data, which can provide faster and more efficient secure multi-party computing.

[0096] The present application may also provide a data privacy protection system, which includes three main components: a data generator, an SMC computing engine, and a result verification module. The data generator is responsible for generating synthetic data, the SMC computing engine protects data privacy through collaborative computing, and the result verification module is used to verify the correctness of the computing results.

[0097] From the above, it can be seen that the large model data privacy protection method proposed in this application uses generative adversarial networks and secure multi-party computing. The improved model RCE-GAN of the generative adversarial network is trained and synthesized with similar high-quality synthetic data, and SMC is introduced to protect the privacy of the training generation process, thereby protecting the privacy data. This application uses the improved generative adversarial network RCE-GAN to confront and optimize the privacy data to generate realistic forged data. Combined with the large model solution, RCE-GAN is applied to the privacy protection of large model data, and finally used in combination with the secure multi-party computing framework to enhance the privacy protection capability of data generation.

[0098] This application has at least the following beneficial effects: for the low-quality synthetic data of the adversarial generative network, which leads to the instability and inefficiency of training; this application proposes a random constraint enhanced GAN, and introduces a temperature controller and an objective function, which effectively improves the diversity and creativity of the generated samples, and improves the stability of data generation. Combining RCE-GAN with a large model solution can enhance the training and generation capabilities of the generative model and provide a more powerful and efficient generative model. SMC technology is applied to the data privacy protection of the RCE-GAN large model. By combining SMC and RCE-GAN, privacy protection can be achieved in the process of generating data to ensure that the private data of the participants is not leaked. Improve data sharing and cooperation: Since data privacy is effectively protected, participants can share data and cooperate in calculations more confidently. This helps to promote data cooperation and sharing and accelerate the progress of innovation and research. The private input of the participants is protected and will not be known to other parties, thereby ensuring compliance and data security.

[0099] The present application also provides a model training device. Figure 5 As shown, the model training device 500 includes:

[0100] A first acquisition module 501 is used to acquire a large model sample data set and a first random noise during each iterative training process;

[0101] A first processing module 502 is used to input the large model sample data set and the first random noise into the generator of the randomness constraint enhancement model to generate a synthetic sample data set;

[0102] A first acquisition module 501 is used to acquire a synthetic sample data set connected to a second random noise and a large model sample data set connected to the second random noise, where the second random noise is noise randomly sampled from a predefined random noise distribution;

[0103] The first processing module 502 is used to input the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise into the discriminator of the randomness constraint enhancement model to obtain a distinction result;

[0104] The first processing module 502 is used to update the parameters of the generator and the discriminator according to the differentiation result until the iteration stop condition is met to obtain the trained random constraint enhancement model.

[0105] In some embodiments, the large model sample data set includes: a first sample data set and a second sample data set, the first sample data set includes data obtained after data dimension reduction and / or data sampling processing is performed on the data in the privacy data set, and the second sample data set includes different data from the first sample data set.

[0106] In some embodiments, the first processing module 502 is used to add a random regularization term to the loss function of the stochastic constraint enhancement model during each iterative training process to constrain the complexity of the stochastic constraint enhancement model.

[0107] In some embodiments, the first processing module 502 is used to pass the first random noise to the discriminator, so that the search discriminator evaluates the randomness of the synthetic sample data set.

[0108] The present application also provides a data generating device. Figure 6 As shown, the data generating device 600 includes:

[0109] The second acquisition module 601 is used to acquire a data set to be processed;

[0110] A second processing module 602 is used to obtain a target data set according to the data set to be processed, where the target data set includes: a first target data set and a second target data set, where the first target data set includes data obtained by performing data dimension reduction processing and / or data sampling processing on the data in the privacy data set, and the second target data set includes data different from the first target data set;

[0111] The second processing module 602 is used to process the target data set using the random constraint enhancement model to generate a target synthetic data set.

[0112] In some embodiments, the second processing module 602 is used to process the processing data set using the random constraint enhancement model, the temperature regulator and the target loss function to generate a target synthetic data set. The temperature regulator is used to adjust the discreteness of the generated target synthetic data set, and the target loss function is used to measure the distance between the target synthetic data set and the target data set.

[0113] In some embodiments, the second processing module 602 is used to obtain a large model sample data set and a first random noise during each iteration of training;

[0114] Inputting the large model sample data set and the first random noise into the generator of the randomness constraint enhanced model to generate a synthetic sample data set;

[0115] Acquire a synthetic sample data set connected to a second random noise and a large model sample data set connected to the second random noise, wherein the second random noise is noise randomly sampled from a predefined random noise distribution;

[0116] Inputting the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise into the discriminator of the randomness constraint enhancement model to obtain a distinction result;

[0117] The parameters of the generator and the discriminator are updated according to the differentiation result until the iteration stopping condition is met and the training is stopped.

[0118] In some embodiments, the large model sample data set includes: a first sample data set and a second sample data set, wherein the first sample data set includes data obtained after data dimension reduction and / or data sampling processing is performed on the data in the privacy data set, and the second sample data set includes different data from the first sample data set.

[0119] In some embodiments, the second processing module 602 is used to add a random regularization term to the loss function of the stochastic constraint enhancement model during each iterative training process to constrain the complexity of the stochastic constraint enhancement model.

[0120] In some embodiments, the second processing module 602 is used to pass the first random noise to the discriminator, so that the search discriminator evaluates the randomness of the synthetic sample data set.

[0121] An embodiment of the present application provides an electronic device, referring to Figure 7 As shown, the electronic device 700 includes: a processor 701, a memory 702 and a communication bus 703, wherein: the communication bus 703 is used to realize the communication connection between the processor 701 and the memory 702;

[0122] The processor 701 is used to execute the model training program stored in the memory 702 to implement the following steps:

[0123] During each iterative training process, a large model sample data set and a first random noise are obtained;

[0124] Inputting the large model sample data set and the first random noise into the generator of the randomness constraint enhancement model to generate a synthetic sample data set;

[0125] Acquire a synthetic sample data set connected to a second random noise and a large model sample data set connected to the second random noise, wherein the second random noise is a noise randomly sampled from a predefined random noise distribution;

[0126] Inputting the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise into the discriminator of the randomness constraint enhancement model to obtain a distinction result;

[0127] The parameters of the generator and the discriminator are updated according to the differentiation results until the iteration stop condition is met, and the trained random constraint enhancement model is obtained.

[0128] In some embodiments of the present application, the large model sample data set includes: a first sample data set and a second sample data set, the first sample data set includes data obtained after data dimensionality reduction and / or data sampling processing is performed on the data in the privacy data set, and the second sample data set includes different data from the first sample data set.

[0129] In some embodiments of the present application, the processor 701 is used to execute the model training program stored in the memory 702 to implement the following steps: during each iterative training process, a random regularization term is added to the loss function of the random constraint enhancement model to constrain the complexity of the random constraint enhancement model.

[0130] In some embodiments of the present application, the processor 701 is used to execute the model training program stored in the memory 702 to implement the following steps: passing the first random noise to the discriminator so that the search discriminator evaluates the randomness of the synthetic sample data set.

[0131] An embodiment of the present application provides an electronic device, referring to Figure 7 As shown, the electronic device 700 includes: a processor 701, a memory 702 and a communication bus 703, wherein: the communication bus 703 is used to realize the communication connection between the processor 701 and the memory 702;

[0132] The processor 701 is used to execute the data generation program stored in the memory 702 to implement the following steps:

[0133] Get the data set to be processed;

[0134] According to the data set to be processed, a target data set is obtained, where the target data set includes: a first target data set and a second target data set, where the first target data set includes data obtained by performing data dimension reduction processing and / or data sampling processing on the data in the privacy data set, and the second target data set includes data different from the first target data set;

[0135] The target dataset is processed using the random constraint enhancement model to generate a target synthetic dataset.

[0136] In some embodiments of the present application, the processor 701 is used to execute the data generation program stored in the memory 702 to implement the following steps: using the random constraint enhancement model, the temperature regulator and the target loss function to process the processing data set to generate a target synthetic data set, the temperature regulator is used to adjust the discreteness of the generated target synthetic data set, and the target loss function is used to measure the distance between the target synthetic data set and the target data set.

[0137] In some embodiments of the present application, the processor 701 is used to execute the data generation program stored in the memory 702 to implement the following steps: in each iterative training process, obtain a large model sample data set and a first random noise;

[0138] Inputting the large model sample data set and the first random noise into the generator of the randomness constraint enhanced model to generate a synthetic sample data set;

[0139] Acquire a synthetic sample data set connected to a second random noise and a large model sample data set connected to the second random noise, wherein the second random noise is noise randomly sampled from a predefined random noise distribution;

[0140] Inputting the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise into the discriminator of the randomness constraint enhancement model to obtain a distinction result;

[0141] The parameters of the generator and the discriminator are updated according to the differentiation result until the iteration stopping condition is met and the training is stopped.

[0142] In some embodiments of the present application, the large model sample data set includes: a first sample data set and a second sample data set, the first sample data set includes data obtained after data dimensionality reduction and / or data sampling processing is performed on the data in the privacy data set, and the second sample data set includes different data from the first sample data set.

[0143] In some embodiments of the present application, the processor 701 is used to execute the data generation program stored in the memory 702 to implement the following steps: during each iterative training process, a random regularization term is added to the loss function of the stochastic constraint enhancement model to constrain the complexity of the stochastic constraint enhancement model.

[0144] In some embodiments of the present application, the processor 701 is used to execute the data generation program stored in the memory 702 to implement the following steps:

[0145] The first random noise is passed to the discriminator so that the search discriminator evaluates the randomness of the synthetic sample data set.

[0146] The processor can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., among which the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0147] It should be noted that, for the description of the same steps and the same contents in this embodiment as those in other embodiments, reference can be made to the description in other embodiments and will not be repeated here.

[0148] An embodiment of the present application provides a computer storage medium storing one or more programs, which can be executed by one or more processors to implement the following. Figure 1 Steps shown.

[0149] It should be noted that, for the description of the same steps and the same contents in this embodiment as those in other embodiments, reference can be made to the description in other embodiments and will not be repeated here.

[0150] It should be noted that the above-mentioned computer storage medium / memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM) and other memories; it can also be various terminals including one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0151] The embodiment of the present application provides a computer product, including a computer program, which can be executed by a processor 701 of an electronic device 700 to complete the following steps: Figure 1 Steps shown.

[0152] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0153] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0154] In addition, all functional units in the embodiments of the present application can be integrated into one processing module, or each unit can be a separate unit, or two or more units can be integrated into one unit; the above integrated unit can be implemented in the form of hardware or in the form of hardware plus software functional units. A person of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium, which, when executed, executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memory (ROM), random access memory (RAM), disks or optical disks, and other media that can store program codes.

[0155] The methods disclosed in several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0156] The features disclosed in several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0157] The features disclosed in several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0158] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A data generation method, characterized in that: include: Get the data set to be processed; According to the data set to be processed, a target data set is obtained, wherein the target data set includes: a first target data set and a second target data set, wherein the first target data set includes data obtained by performing data dimension reduction processing and / or data sampling processing on the data in the privacy data set, and the second target data set includes data different from the first target data set; The target data set is processed using a random constraint enhancement model to generate a target synthetic data set.

2. The method according to claim 1, characterized in that The step of processing the target data set by using the random constraint enhancement model to generate a target synthetic data set includes: The target data set is processed using a random constraint enhancement model, a temperature regulator and a target loss function to generate the target synthetic data set, wherein the temperature regulator is used to adjust the discreteness of the generated target synthetic data set, and the target loss function is used to measure the distance between the target synthetic data set and the target data set.

3. The method according to claim 1, characterized in that Before the target data set is processed by the random constraint enhancement model to generate the target synthetic data set, the random constraint enhancement model is trained by the following steps: During each iterative training process, a large model sample data set and a first random noise are obtained; Inputting the large model sample data set and the first random noise into the generator of the randomness constraint enhanced model to generate a synthetic sample data set; Acquire a synthetic sample data set connected to a second random noise and a large model sample data set connected to the second random noise, wherein the second random noise is noise randomly sampled from a predefined random noise distribution; Inputting the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise into the discriminator of the randomness constraint enhancement model to obtain a distinction result; The parameters of the generator and the discriminator are updated according to the differentiation result until the iteration stopping condition is met and the training is stopped.

4. The method according to claim 3, characterized in that The large model sample data set includes: a first sample data set and a second sample data set, the first sample data set includes data obtained after data dimension reduction and / or data sampling processing is performed on the data in the privacy data set, and the second sample data set includes different data from the first sample data set.

5. The method according to claim 3, characterized in that: Also includes: During each iterative training process, a random regularization term is added to the loss function of the stochastic constraint enhancement model to constrain the complexity of the stochastic constraint enhancement model.

6. The method according to claim 3, characterized in that Before inputting the synthetic sample data set connected with the second random noise and the large model sample data set connected with the second random noise into the discriminator of the randomness constraint enhancement model to obtain the distinction result, the method further includes: The first random noise is transmitted to the discriminator so that the search discriminator evaluates the randomness of the synthetic sample data set.

7. A data generating device, characterized in that: include: The second acquisition module is used to acquire the data set to be processed; A second processing module is used to obtain a target data set according to the data set to be processed, wherein the target data set includes: a first target data set and a second target data set, wherein the first target data set includes data obtained by performing data dimension reduction processing and / or data sampling processing on the data in the privacy data set, and the second target data set includes data different from the first target data set; The second processing module is used to process the target data set using a random constraint enhancement model to generate a target synthetic data set.

8. An electronic device, characterized in that: include: Processor and communication interface; wherein, The processor is configured to: Acquire a data set to be processed through the communication interface; obtain a target data set based on the data set to be processed, and the target data set includes: a first target data set and a second target data set, the first target data set includes data obtained after data dimension reduction processing and / or data sampling processing is performed on the data in the privacy data set, and the second target data set is different from the data included in the first target data set; use a random constraint enhancement model to process the target data set to generate a target synthetic data set.

9. A storage medium, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the data generation method according to any one of claims 1 to 6.

10. A computer product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the data generation method according to any one of claims 1 to 6 are implemented.

Images