WO2023065632A1 - 数据脱敏方法、数据脱敏装置、设备及存储介质 - Google Patents

数据脱敏方法、数据脱敏装置、设备及存储介质 Download PDF

Info

Publication number
WO2023065632A1
WO2023065632A1 PCT/CN2022/089872 CN2022089872W WO2023065632A1 WO 2023065632 A1 WO2023065632 A1 WO 2023065632A1 CN 2022089872 W CN2022089872 W CN 2022089872W WO 2023065632 A1 WO2023065632 A1 WO 2023065632A1
Authority
WO
WIPO (PCT)
Prior art keywords
key information
data
vector
generator
trained
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2022/089872
Other languages
English (en)
French (fr)
Inventor
郑旭如
赵盟盟
王磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2023065632A1 publication Critical patent/WO2023065632A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/60—Protecting data
    • G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
    • G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
    • G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/284—Lexical analysis, e.g. tokenisation or collocates
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present application relates to the field of artificial intelligence, and in particular to a data desensitization method, a data desensitization device, computer equipment, and a storage medium.
  • Data desensitization technology is an effective method to solve data security problems and risks.
  • Data desensitization refers to the transformation of key information or personal information according to preset rules or transformations, so that personal identity cannot be identified or key information is hidden.
  • common structured data desensitization methods are desensitization methods based on anonymization technology or scrambling technology.
  • the inventor realized that there is a one-to-one mapping relationship between the desensitized data and the original data in the structured data desensitization method based on anonymization technology or scrambling technology, which makes the desensitized data easy to be reversed, thus The original data is easily restored, which leads to the leakage of private information in the original data, and the data security is poor.
  • the present application provides a data desensitization method, data desensitization device, computer equipment, and storage medium, aiming to solve the problem that existing desensitization methods are easily reversed and private information is easily leaked.
  • the present application provides a data desensitization method, the method comprising:
  • Preprocessing the key information to obtain discrete variables corresponding to the key information includes data discretization or data normalization;
  • conditional loss function Based on a conditional loss function, performing conditional random sampling processing on the discrete variable to obtain a conditional embedding vector and a hidden vector, and splicing the conditional embedding vector and the hidden vector to obtain a concatenated vector;
  • the splicing vector is input to the pre-trained generator for desensitization processing to obtain desensitized data.
  • the present application also provides a data desensitization device, the data desensitization device includes:
  • the key information extraction module is used to obtain user data, and based on the pre-trained key information identification model, perform information identification on the user data to obtain key information;
  • An information processing module configured to preprocess the key information to obtain discrete variables corresponding to the key information, where the preprocessing includes data discretization or data normalization;
  • a vector splicing module configured to perform conditional random sampling processing on the discrete variables based on a conditional loss function to obtain a conditional embedding vector and a hidden vector, and splicing the conditional embedding vector and the hidden vector to obtain a splicing vector;
  • the data desensitization module is used to input the splicing vector into the pre-trained generator for desensitization processing to obtain desensitized data.
  • the present application also provides a computer device, the computer device includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and Realize following steps when executing described computer program:
  • Preprocessing the key information to obtain discrete variables corresponding to the key information includes data discretization or data normalization;
  • conditional loss function Based on a conditional loss function, performing conditional random sampling processing on the discrete variable to obtain a conditional embedding vector and a hidden vector, and splicing the conditional embedding vector and the hidden vector to obtain a concatenated vector;
  • the splicing vector is input to the pre-trained generator for desensitization processing to obtain desensitized data.
  • the present application also provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor implements the following steps:
  • Preprocessing the key information to obtain discrete variables corresponding to the key information includes data discretization or data normalization;
  • conditional loss function Based on a conditional loss function, performing conditional random sampling processing on the discrete variable to obtain a conditional embedding vector and a hidden vector, and splicing the conditional embedding vector and the hidden vector to obtain a concatenated vector;
  • the splicing vector is input to the pre-trained generator for desensitization processing to obtain desensitized data.
  • the data desensitization method, data desensitization device, device, and storage medium disclosed in the embodiments of the present application generate splicing vectors by extracting key information of user data and discrete variables of key information, and use a pre-trained generator to Vectors are desensitized to obtain desensitized data, so that the desensitized data cannot be easily reversed, thereby ensuring that private data is not leaked and improving the security of desensitized data.
  • FIG. 1 is a schematic diagram of a scenario of a data desensitization method provided in an embodiment of the present application
  • Fig. 2 is a schematic flow chart of a data desensitization method provided in the embodiment of the present application
  • Fig. 3 is a schematic block diagram of a data desensitization device provided by an embodiment of the present application.
  • Fig. 4 is a schematic block diagram of a computer device provided by an embodiment of the present application.
  • Data desensitization technology is a data processing technology that can reduce and remove data sensitivity by processing data.
  • the use of data desensitization technology can reduce the risk and harm of data leakage and effectively protect the privacy of user data.
  • users can store, view and share personal medical and health data through their personal digital space, but personal medical data will face the risk of leaking sensitive medical information of users in the process of online medical treatment, online purchase of medicines, outpatient appointments, etc.
  • user data has extremely high authenticity and sensitivity. Once the user's personal sensitive information is leaked, it may pose a potential threat to the user's life.
  • the information in the personal digital space can be used for business-related analysis and processing, while avoiding the leakage of user data.
  • Common structured data desensitization methods are desensitization methods based on anonymization technology or scrambling technology.
  • Common anonymization techniques include k-anonymity, l-diversity, and t-closeness. They generalize the quasi-identifier of a single record to make the data indistinguishable in the entire data set, thereby achieving desensitization. Effect. Scrambling-based technology is to add noise to the record, such as adding additive or multiplicative noise to continuous values, so as to achieve the effect of desensitization.
  • this application provides a data desensitization method, which can be applied in the server, specifically in multiple fields such as finance and medical treatment.
  • a pre-trained generator is obtained. Extract the sensitive information of the user data, and use the pre-trained generator to desensitize the sensitive information to obtain the desensitized data, so that the desensitized data cannot be easily reversed, thus ensuring that the private data is not leaked , improving the security of desensitized data.
  • the server may be, for example, an individual server or a server cluster.
  • the following embodiments will introduce in detail the data desensitization method applied to the server.
  • the data desensitization method provided in the embodiment of the present application can be applied to the application environment shown in FIG. 1 .
  • the application environment includes a terminal device 110 and a server 120, wherein the terminal device 110 can communicate with the server 120 through a network.
  • the server 120 obtains the user data sent by the terminal device 110, and the server 120 performs key information extraction, information processing, and desensitization processing on the user data to generate desensitized data, and sends the desensitized data to the terminal device 110, so as to realize Data desensitization processing.
  • the server 120 can be an independent server, and can also provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution network (Content Delivery Network, CDN), and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms.
  • the terminal device 110 may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto.
  • the terminal and the server may be connected directly or indirectly through wired or wireless communication, which is not limited in this application.
  • FIG. 2 is a schematic flowchart of a data desensitization method provided by an embodiment of the present application.
  • the data desensitization method can be applied in the server, thereby making the desensitization data unable to be easily reversely deciphered, thereby ensuring that the private data is not leaked, and improving the security of the desensitization data.
  • the data desensitization method includes steps S101 to S104.
  • S101 Acquire user data, and perform information identification on the user data based on a pre-trained key information identification model to obtain key information.
  • the user data is data containing key information, and may specifically include medical data such as medical record data, financial data such as bank account data, and the like.
  • the key information identification model may be a pre-trained BERT-CRF model based on an attention mechanism, which is used to extract key information in user data.
  • the key information is the information that the user needs to desensitize, which is generally the user's private information.
  • the key information can be the height and weight information in the medical record data, or the account balance information and investment information in the bank account data. . It should be noted that all sensitive or private information can be used as key information.
  • AI artificial intelligence
  • digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
  • Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics.
  • Artificial intelligence software technology mainly includes computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning/deep learning.
  • word segmentation processing is performed on the user data to obtain multiple word segmentations; feature extraction is performed on each of the word segmentations to obtain the embedded features of each of the word segmentations;
  • the word meaning prediction is to obtain the meaning corresponding to each of the word segmentations; the plurality of word segmentations are screened according to the meanings corresponding to each of the word segmentations to obtain key information.
  • the embedding features are word embedding features, position embedding features and segmentation embedding features.
  • the word embedding feature is a vector representation of each word segment
  • the position embedding is a vector representation of each word segment position
  • the segmentation embedding feature is used to distinguish two different sentences.
  • the user data can be segmented based on a word segmentation algorithm to obtain multiple word segments.
  • the word segmentation algorithm can be forward maximum matching method, reverse maximum matching method, word segmentation algorithm based on hidden Markov model, condition-based Random field word segmentation algorithm and other algorithms.
  • the word segmentation algorithm based on the hidden Markov model can be used to segment the user data such as the medical record text of the medical record data as "the patient has symptoms such as frequent urination, hunger, anxiety, tremor, etc., and is suspected of diabetes," to obtain multiple corresponding Word segmentation such as frequent urination, hunger, anxiety, tremor, etc.
  • feature extraction can be performed on each of the word segmentations to obtain the embedded features of each of the word segmentations, and based on the word meaning prediction model, the meaning of each of the word segmentations can be predicted according to the embedded features of each of the word segmentations to obtain The word meaning prediction result corresponding to each word segment, and based on the word meaning prediction result corresponding to each word segment, filter the plurality of word segments to obtain key information.
  • text features can be mined to the greatest extent, richer word representations can be extracted, and the shortcomings of traditional word vectors such as Word2vec and Glove that cannot dynamically represent context information and cannot solve polysemy of a word can be eliminated. Therefore, the similarity between each word segment and the preset standard sensitive word can be quickly obtained, and then the corresponding key information can be quickly obtained.
  • the word meaning prediction model is used to predict the similarity between each word segmentation and the preset standard sensitive word segmentation
  • the word meaning prediction model is obtained by training the semantic matching model and the standard sensitive word segmentation database
  • the semantic prediction model may include LSTM Matching model, MV-DSSM model, ESIM model and other models
  • the word meaning prediction result is the similarity between each participle and the standard sensitive participle in the standard sensitive participle database.
  • the word segmentation includes account balance equivalent words of account information and stock trend information equal segmentation words
  • feature extraction can be performed on each of the word segmentations, and the word embedding features, position embedding features and segmentation embedding of each of the word segmentations can be obtained
  • Features based on the LSTM matching model, perform word meaning prediction for each of the word segmentations according to the word embedding features, position embedding features and segmentation embedding features, and obtain the word meaning prediction results corresponding to each word segmentation, and based on the word meaning prediction results corresponding to each word segmentation Filter out the word segmentation corresponding to the stock trend information to obtain key information.
  • the key information is generally continuous data, it is necessary to perform representation conversion between continuous data and discrete data, that is, data preprocessing operations, which is a key step in the input and output of neural networks.
  • the key information is information such as height and weight
  • the key information is continuous data
  • the key information is information such as the number of investment companies
  • the key information is discrete data
  • the discrete variable refers to a variable whose value can be listed in a certain order, and usually takes an integer value, such as the number of employees, the number of factories, the number of machines, and the like.
  • the data normalization processing may include maximum-minimum normalization processing and normalization processing according to a Gaussian mixture model; the data discretization processing may include K-bins discretization processing and regression tree discretization processing.
  • the key information is subjected to maximum and minimum normalization processing to obtain the discrete variable corresponding to the key information; or, the key information is normalized through a Gaussian mixture model to obtain the key information A discrete variable corresponding to the information; or, K-bins discretization processing is performed on the key information to obtain a discrete variable corresponding to the key information; or, a regression tree discretization process is performed on the key information to obtain the key information Corresponding discrete variables.
  • the key information is continuous data, it can be mapped to the range of [0,1] through the maximum and minimum linear transformation, so that the continuous value can be represented by the tanh activation function, and the discrete variable corresponding to the key information can be obtained.
  • the Gaussian mixture model can be used to fit the key information, and the Gaussian component is sampled according to the probability of the Gaussian component of the key information in the mixture model, and the sampled Gaussian component is used to compare the data in the record.
  • the key information is normalized. Then, the key information can be composed of the normalized representation and the one-hot encoding of the Gaussian component, so as to obtain the discrete variables corresponding to the key information.
  • K-bins discretization processing may be performed on the key information to obtain discrete variables corresponding to the key information.
  • the discretization can also be called binning, that is, the key information is divided into various intervals according to certain rules, and each interval is represented by one-hot encoding, so that the key information is represented by a piecewise function containing four intervals Fitting is performed to obtain discrete variables corresponding to the key information.
  • the key information may be discretized using a CART regression tree to obtain discrete variables corresponding to the key information.
  • the CART regression tree can predict continuous data, and its leaf node represents a predicted value.
  • the key information can be converted into discrete values by expressing a series of leaf nodes of the regression tree or regression tree set of key information through one-hot encoding.
  • conditional loss function is a conditional loss function based on the adversarial generation network, and the data items of the loss function based on the adversarial generation network are generated based on conditional probabilities.
  • the original intention is to enable data to be generated according to conditions, so that the same type
  • the distribution of the data to be desensitized and the generated desensitized data is as consistent as possible.
  • the training process can be constrained by predicting the condition variables, so that the values of the condition variables are consistent with the values of the corresponding variables in the generated data, and the effect of data generation can be further optimized.
  • the conditional embedding vector can randomly select a discrete variable that meets the preset condition from a plurality of discrete variables corresponding to the key information with equal probability, and the hidden vector can be selected from the white noise corresponding to the key information
  • the concatenated vector is obtained by concatenating the conditional embedding vector and the latent vector, and is used as an input of the generator.
  • the one-to-one mapping relationship between the desensitized data and the original data is changed, so that the desensitized data is not easy to be reversely cracked, and private information can be obtained.
  • the distributed representation of the discrete variable can be obtained by constructing the probability mass distribution function of each value of the discrete variable, and the distributed representation of the discrete variable is subjected to conditional random sampling processing to obtain the conditional embedding vector and hidden vector.
  • the white noise corresponding to the discrete variable may be converted by a deep neural network to generate a latent vector from the distributed representation of the discrete variable.
  • the conditional embedding vector is converted to obtain a one-hot encoding; the one-hot encoding is concatenated with the hidden vector to obtain a concatenated vector.
  • the one-hot encoding is One-Hot encoding, also known as effective encoding, and its method is to use N-bit state registers to encode N states, each state has its own independent register bit, and at any time, Only one of them is valid. Converting the conditional embedding vector into a one-hot encoding can solve the problem that the discriminator cannot handle attribute data well, and at the same time, it also plays a role in expanding the vector feature to a certain extent.
  • conditional embedding vector can be transformed through a deep neural network to obtain a one-hot encoding, and the splicing vector can be obtained by splicing the one-hot encoding and the latent vector. From this, a concatenated vector that meets the input requirements of the generator can be obtained.
  • the pre-trained generator is generated based on confrontation generation network training, and the desensitized data is data after desensitizing key information in the data to be desensitized.
  • the splicing vector corresponding to the training data is obtained, and the splicing vector is input to the first generator for desensitization processing to obtain desensitized data; based on the desensitized data and training data pair
  • the preset discriminator is trained to obtain a pre-trained discriminator; according to the preset learning rate and the parameters of the pre-trained discriminator, the parameters of the first generator are iteratively updated multiple times to obtain A second generator, and use the second generator as a pre-trained generator. In this way, the parameters of the first generator can be updated iteratively for multiple times through the pre-trained discriminator and the desensitized data, and very real desensitized data can be generated.
  • the training data is a data set to be desensitized for training generator parameters
  • the first generator is a preset untrained generator
  • the second generator is a Generated by multiple iterative updates.
  • the parameters of the first generator and the second generator are different.
  • the prior probability of the discrete variable can be obtained through the distributed representation of the discrete variable, and parameters are sampled from the prior probability as parameters of the first generator.
  • the generator and the discriminator can be trained by the stochastic gradient Hamiltonian Monte Carlo method to obtain a pre-trained generator and a pre-trained discriminator.
  • the preset discriminator is trained based on the desensitized data and training data, and the pre-trained discriminator is obtained by combining the conditional embedding vector with the desensitized data and the Splicing the training data to obtain the first spliced data and the second spliced data, and calculating the similarity between the first spliced data and the second spliced data, and optimizing the loss function according to the similarity between the first spliced data and the second spliced data, And the discriminator is gradient clipped through the loss function to obtain a pre-trained discriminator.
  • the discriminator parameters can be trained by the first generator and the preset discriminator parameters, and the desensitized data can be judged as false as much as possible, thereby adjusting the discriminator parameters, thereby improving the discriminator's ability to desensitize the data Discrimination ability.
  • the posterior probability of the second generator can be calculated through the prior probability of the parameters of the first generator and the pre-trained discriminator, so that the desensitized data should make the discriminator misjudge it as being to be desensitized Sensitive data, so as to adjust the parameters of the generator to generate real desensitized data.
  • the second generator is subjected to noise-increasing processing based on a loss function of statistical information to obtain a pre-trained generator, wherein the first generator, the The parameters of the generator whose parameters have been updated are different from those of the pre-trained generator. In this way, the generation quality and degree of desensitization of the desensitized data can be controlled.
  • the statistical information-based loss function may include a mean-based loss function, a variance-based loss function, and the like.
  • Gaussian noise can be added to the parameters of the second generator, thereby realizing polynomial fitting of a sinusoidal curve, and the Gaussian noise is an error conforming to a Gaussian normal distribution.
  • the specific value of Gaussian noise can be obtained through experiments.
  • an error term may be introduced into the parameters of the second generator, so that the parameters of the second generator may be corrected to obtain a pre-trained generator. Due to the existence of the error term, there is a certain difference between the generated desensitized data and the original data, but the difference is not large, thus avoiding that the desensitized data often differs greatly from the original data and loses the research value, and also ensures Data cannot be easily reversed.
  • the discrete variables of the desensitized data are randomly sampled to obtain the target discrete variables; based on the logistic regression model, the target The discrete variable is predicted to obtain the predicted result of the target discrete variable; and the parameters of the pre-trained generator are adjusted based on the predicted result of the target discrete variable.
  • the parameters of the generator can be adjusted by predicting the discrete variables to achieve a better desensitization effect.
  • the better desensitization effect here means that it can prevent the desensitized data from being reversely cracked, while maintaining the association with the original data.
  • the target discrete variable is randomly sampled from multiple discrete variables of the desensitized data, and at the same time, in order to associate the discrete variables between the desensitized data and the original data, it can generally be considered that the target discrete variable does not change, and the desensitized data and If the difference in the original data is small, the value of the research will not be lost, so it is necessary to ensure the consistency of the target discrete variables.
  • the logistic regression model was used to predict discrete variables.
  • the cross-entropy loss function can be used to judge whether the generated prediction result of the target discrete variable is consistent with the target discrete variable, so as to determine the generation quality of the desensitized data. If the prediction result of the target discrete variable is consistent with the target discrete variable, there is no need to adjust the parameters of the pre-trained generator; if the prediction result of the target discrete variable is inconsistent with the target discrete variable, then determine the prediction result of the target discrete variable The difference with the target discrete variable, and adjust the parameters of the pre-trained generator according to the difference. In this way, the accuracy of the target discrete variable can be determined, and the generated desensitized data can avoid making the original data too different. Since most of the discrete variables of the desensitized data and the original data are the same, removing one of the discrete variables can accurately predict the discrete variable based on the remaining discrete variables.
  • the target discrete variable of the desensitized data is a shoe size of 43
  • the target discrete variable can be predicted through the remaining discrete variables of the desensitized data, such as height and weight, to obtain the shoe size to determine whether the generated shoe size prediction result is consistent with the shoe size of the desensitized data. For example, if the generated shoe size prediction result is a shoe size of 40, the difference is determined to be 1 size. According to the difference The parameters of the pre-trained generator are updated iteratively; if the predicted result of the generated shoe size is size 43, there is no need to adjust the parameters of the pre-trained generator.
  • the server may also send prompt information for prompting the user that the desensitized data has been generated to the terminal device.
  • the manner of prompting information may specifically include an application program (APP) or Email, a short message, a chat tool, such as WeChat, qq, and the like.
  • APP application program
  • Email a short message
  • chat tool such as WeChat, qq, and the like.
  • the server when the desensitized data has been generated, the server will send a prompt message that the desensitized data has been generated to the terminal device to remind the user.
  • FIG. 3 is a schematic block diagram of a data desensitization device provided by an embodiment of the present application.
  • the data desensitization device can be configured in a server to execute the aforementioned data desensitization method.
  • the data desensitization device 200 includes: a key information extraction module 201 , an information processing module 202 , a vector splicing module 203 and a data desensitization module 204 .
  • the key information extraction module 201 is configured to acquire user data, and based on a pre-trained key information identification model, perform information identification on the user data to obtain key information;
  • An information processing module 202 configured to preprocess the key information to obtain discrete variables corresponding to the key information, the preprocessing includes data discretization processing or data normalization processing;
  • a vector splicing module 203 configured to perform conditional random sampling processing on the discrete variables based on a conditional loss function to obtain a conditional embedding vector and a hidden vector, and splicing the conditional embedding vector and the hidden vector to obtain a concatenated vector;
  • a data desensitization module 204 configured to input the splicing vector into a pre-trained generator for desensitization processing to obtain desensitized data
  • the feature extraction module 201 is also used to perform word segmentation processing on the user data to obtain a plurality of word segments; perform feature extraction on each of the word segments to obtain the embedded features of each of the word segments; according to the embedding features of each of the word segments.
  • the feature predicts the meaning of the word to obtain the meaning corresponding to each of the word segmentations; according to the meaning of each of the word segmentations, the plurality of word segmentations are screened to obtain key information.
  • the information processing module 202 is further configured to perform maximum and minimum normalization processing on the key information to obtain discrete variables corresponding to the key information; or, to perform normalization processing on the key information through a Gaussian mixture model to obtain the The discrete variable corresponding to the key information; or, carry out K-bins discretization process on the key information, obtain the discrete variable corresponding to the key information; or, perform regression tree discretization process on the key information, obtain the described key information Discrete variables corresponding to key information.
  • the vector splicing module 203 is further configured to convert the conditional embedding vector to obtain a one-hot encoding; splice the one-hot encoding and the latent vector to obtain a concatenated vector.
  • the generator training module 205 is used to obtain the splicing vector corresponding to the training data, and input the splicing vector to the first generator for desensitization processing to obtain desensitized data; based on the desensitized data and training
  • the data is used to train the preset discriminator to obtain a pre-trained discriminator; according to the preset learning rate and the parameters of the pre-trained discriminator, the parameters of the first generator are iteratively updated multiple times , get the second generator, and use the second generator as a pre-trained generator.
  • the generator training module 205 is further configured to perform noise-increasing processing on the second generator based on a loss function of statistical information to obtain a pre-trained generator, wherein the first generator and the second generator The parameters of the generator and the pre-trained generator are different.
  • the generator training module 205 is also used to randomly sample the discrete variables of the desensitized data to obtain the target discrete variables; based on the logistic regression model, the target discrete variables are processed according to the remaining discrete variables of the desensitized data. Forecasting, obtaining a prediction result of the target discrete variable; adjusting parameters of the pre-trained generator based on the prediction result of the target discrete variable.
  • the methods and devices of the present application can be used in many general-purpose or special-purpose computing system environments or configurations. Examples: personal computers, server computers, handheld or portable devices, tablet-type devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer terminal equipment, network PCs, minicomputers, mainframe computers, including the above A distributed computing environment for any system or device, and more.
  • the above-mentioned method and apparatus can be implemented in the form of a computer program, and the computer program can be run on the computer device as shown in FIG. 4 .
  • FIG. 4 is a schematic diagram of a computer device provided by an embodiment of the present application.
  • the computer device may be a server.
  • the computer device includes a processor, a memory, and a network interface connected through a system bus, where the memory may include a volatile storage medium, a non-volatile storage medium, and an internal memory.
  • Non-volatile storage media can store operating systems and computer programs.
  • the computer program includes program instructions.
  • the processor can be executed to perform any data masking method.
  • the processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.
  • the internal memory provides an environment for running the computer program in the non-volatile storage medium.
  • the processor can execute any data desensitization method.
  • This network interface is used for network communication, such as sending assigned tasks, etc.
  • Those skilled in the art can understand that the structure of the computer equipment is only a block diagram of the partial structure related to the solution of the present application, and does not constitute a limitation to the computer equipment on which the solution of the application is applied.
  • the specific computer equipment may include More or fewer components are shown in the figures, or certain components are combined, or have different component arrangements.
  • the processor may be a central processing unit (Central Processing Unit, CPU), and the processor may also be other general processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the processor is configured to run a computer program stored in the memory to implement the following steps: acquire user data, and perform information identification on the user data based on a pre-trained key information identification model , to obtain the key information; preprocessing the key information to obtain the discrete variable corresponding to the key information, the preprocessing includes data discretization processing or data normalization processing; based on the conditional loss function, the discrete variable Perform conditional random sampling processing to obtain conditional embedding vectors and hidden vectors, and splice the conditional embedding vectors and the hidden vectors to obtain spliced vectors; input the spliced vectors to the pre-trained generator for desensitization , to get desensitized data.
  • the processor is further configured to: perform word segmentation processing on the user data to obtain multiple word segments; perform feature extraction on each of the word segments to obtain embedded features of each of the word segments; The embedding features of each of the word segmentations are used to predict the meaning of the word to obtain the corresponding meaning of each of the word segmentations; according to the meaning of each of the corresponding word segmentations, the plurality of word segmentations are screened to obtain key information.
  • the processor is further configured to: perform maximum and minimum normalization processing on the key information to obtain discrete variables corresponding to the key information; or, normalize the key information through a Gaussian mixture model or, performing K-bins discretization processing on the key information to obtain discrete variables corresponding to the key information; or, performing regression tree discretization on the key information process to obtain the discrete variables corresponding to the key information.
  • the processor is further configured to: convert the conditional embedding vector to obtain a one-hot encoding; concatenate the one-hot encoding and the latent vector to obtain a concatenated vector.
  • the processor is further configured to: acquire a splicing vector corresponding to the training data, and input the splicing vector to the first generator for desensitization processing to obtain desensitized data; based on the desensitized
  • the pre-sensitized data and training data are used to train the preset discriminator to obtain a pre-trained discriminator; according to the preset learning rate and the parameters of the pre-trained discriminator, the first generator is The parameters are iteratively updated multiple times to obtain a second generator, and the second generator is used as a pre-trained generator.
  • the processor is further configured to: perform noise-increasing processing on the second generator based on a loss function of statistical information to obtain a pre-trained generator, wherein the first generator, The parameters of the second generator and the pre-trained generator are different.
  • the processor is further configured to: perform random sampling on the discrete variables of the desensitized data to obtain target discrete variables; Predicting the target discrete variable to obtain a prediction result of the target discrete variable; adjusting parameters of the pre-trained generator based on the prediction result of the target discrete variable.
  • the embodiment of the present application also provides a computer-readable storage medium, and the computer-readable storage medium may be non-volatile or volatile.
  • a computer program is stored on the computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed, any data desensitization method provided in the embodiments of the present application is implemented.
  • the computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as a hard disk or a memory of the computer device.
  • the computer-readable storage medium can also be an external storage device of the computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD ) card, flash memory card (Flash Card), etc.
  • the computer-readable storage medium may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application program required by at least one function, etc.; The data created using the node, etc.
  • Blockchain essentially a decentralized database, is a series of data blocks associated with each other using cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify its Validity of information (anti-counterfeiting) and generation of the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Bioethics (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Computer Hardware Design (AREA)
  • Computer Security & Cryptography (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Storage Device Security (AREA)

Abstract

本申请涉及人工智能领域，尤其涉及数据脱敏方法、数据脱敏装置、设备及存储介质，所述方法包括：获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。由此能够使脱敏数据的无法被轻易逆向破解，从而保证了隐私数据不被泄露，提高了脱敏数据的安全性。

Description

数据脱敏方法、数据脱敏装置、设备及存储介质
本申请要求于2021年10月21日提交中国专利局、申请号为202111229481.X、发明名称为“数据脱敏方法、数据脱敏装置、设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能领域，尤其涉及数据脱敏方法、数据脱敏装置、计算机设备及存储介质。
背景技术
大数据时代下，针对数据进行的攻击频率日益增多，攻击方式也日趋丰富。数据脱敏技术是解决数据安全问题和风险的一种行之有效的方法。数据脱敏是指对关键信息或个人信息按照预设的规则或者变换进行数据变形，从而使得个人身份无法识别或者隐去关键信息。目前常见的结构化数据脱敏方式有基于匿名化技术或者扰乱技术的脱敏方式。
但是发明人意识到基于匿名化技术或扰乱技术的结构化数据脱敏方法中存在脱敏后数据与原数据是一对一的映射关系的问题，导致脱敏后的数据容易被逆向破解，从而被轻易还原出原数据，进而导致原数据中隐私信息的泄露，数据安全性较差。
发明内容
本申请提供了一种数据脱敏方法、数据脱敏装置、计算机设备及存储介质，旨在解决现有的脱敏方式存在容易被逆向破解导致隐私信息容易被泄露的问题。
为实现上述目的，本申请提供一种数据脱敏方法，所述方法包括：
获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；
对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
为实现上述目的，本申请还提供一种数据脱敏装置，所述数据脱敏装置包括：
关键信息提取模块，用于获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；
信息处理模块，用于对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
向量拼接模块，用于基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
数据脱敏模块，用于将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
此外，为实现上述目的，本申请还提供一种计算机设备，所述计算机设备包括存储器和处理器；所述存储器，用于存储计算机程序；所述处理器，用于执行所述的计算机程序并在执行所述的计算机程序时实现如下步骤：
获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别， 得到关键信息；
对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
此外，为实现上述目的，本申请还提供一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时使所述处理器实现如下步骤：
获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；
对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
本申请实施例公开的数据脱敏方法、数据脱敏装置、设备及存储介质，通过提取用户数据的关键信息以及关键信息的离散变量，从而生成拼接向量，并利用预训练好的生成器对拼接向量进行脱敏处理，得到脱敏数据，由此能够使脱敏数据的无法轻易被逆向破解，从而保证了隐私数据不被泄露，提高了脱敏数据的安全性。
附图说明
为了更清楚地说明本申请实施例技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例提供的一种数据脱敏方法的场景示意图；
图2是本申请实施例提供的一种数据脱敏方法的流程示意图；
图3是本申请一实施例提供的一种数据脱敏装置的示意性框图；
图4是本申请一实施例提供的一种计算机设备的示意性框图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
附图中所示的流程图仅是示例说明，不是必须包括所有的内容和操作/步骤，也不是必须按所描述的顺序执行。例如，有的操作/步骤还可以分解、组合或部分合并，因此实际执行的顺序有可能根据实际情况改变。另外，虽然在装置示意图中进行了功能模块的划分，但是在某些情况下，可以以不同于装置示意图中的模块划分。
在本申请说明书和所附权利要求书中使用的术语“和/或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合，并且包括这些组合。
数据脱敏技术是一种可以通过对数据进行处理，从而达到降低和去除数据敏感程度的数据处理技术。采用数据脱敏技术，可以降低数据泄露的风险和危害，有效地保护用户数据的隐私。在互联网或医疗领域，用户可以通过个人数字空间存储、查看和分享个人医疗健康数据，但是个人的医疗数据在网上看病、网上购买药品、门诊预约等流程中会面临用户医疗敏感信息泄露的风险，而在医疗行业用户的数据具有极高的真实性和敏感性特点，一旦用户的个人敏感信息泄露可能会对用户本人造成潜在的生命威胁。借助数据脱敏，个人数字空间中的信息可以被用于业务相关的分析和处理，同时避免用户数据的泄露。
目前常见的结构化数据脱敏方式有基于匿名化技术或者扰乱技术的脱敏方式。常见的匿名化技术有k-匿名、l-多样性和t-closeness等，它们是通过对单条记录的准标识符进行泛化后使得该数据在整个数据集中无法进行区分，从而达到脱敏的效果。基于扰乱技术则是往记录中加入噪声，比如在连续值中加入加性或者乘性噪声，从而达到脱敏的效果。
而基于匿名化技术或扰乱技术的结构化数据脱敏方法中存在脱敏后数据与原数据是一对一的映射关系的问题，导致脱敏后的数据存在被逆向的风险，而且经过脱敏的数据往往与原数据差别较大而失去了研究的价值。
为解决上述问题，本申请提供了一种数据脱敏方法，可以应用在服务器中，具体应用金融、医疗等多个领域，通过不断对生成器参数进行迭代更新，得到预训练好的生成器，提取用户数据的敏感信息，并利用预训练好的生成器对敏感信息进行脱敏处理，得到脱敏数据，由此能够使脱敏数据的无法被轻易逆向破解，从而保证了隐私数据不被泄露，提高了脱敏数据的安全性。
其中，服务器例如可以为单独的服务器或服务器集群。但为了便于理解，以下实施例将以应用于服务器的数据脱敏方法进行详细介绍。
下面结合附图，对本申请的一些实施方式作详细说明。在不冲突的情况下，下述的实施例及实施例中的特征可以相互组合。
如图1所示，本申请实施例提供的数据脱敏方法，可以应用于如图1所示的应用环境中。该应用环境中包含有终端设备110和服务器120，其中，终端设备110可以通过网络与服务器120进行通信。具体地，服务器120获取终端设备110发送的用户数据，服务器120对用户数据进行关键信息提取、信息处理以及脱敏处理后生成脱敏数据，并将该脱敏数据发送给终端设备110，以实现数据脱敏处理。其中，服务器120可以是独立的服务器，也可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、内容分发网络(Content Delivery Network，CDN)、以及大数据和人工智能平台等基础云计算服务的云服务器。终端设备110可以是智能手机、平板电脑、笔记本电脑、台式计算机、智能音箱、智能手表等，但并不局限于此。终端以及服务器可以通过有线或无线通信方式进行直接或间接地连接，本申请在此不做限制。
请参阅图2，图2是本申请实施例提供的一种数据脱敏方法的示意流程图。其中，该数据脱敏方法可以应用在服务器中，由此能够使脱敏数据的无法被轻易逆向破解，从而保证了隐私数据不被泄露，提高了脱敏数据的安全性。
如图2所示，该数据脱敏方法包括步骤S101至步骤S104。
S101、获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息。
其中，所述用户数据为包含关键信息的数据，具体可以包括医疗数据比如病历数据、金融数据比如银行账户数据等。所述关键信息识别模型可以是基于注意力机制的预训练BERT-CRF模型，用于提取用户数据中的关键信息。所述关键信息是用户需要进行脱敏的信息，一般为用户的隐私信息，比如关键信息可以为病历数据中的身高、体重信息等，还可以为银行账户数据中的账户余额信息和投资信息等。需要说明的是，一切敏感信息或隐私信息均可以作为关键信息。
本申请实施例可以基于人工智能技术对相关的数据进行获取和处理。其中，人工智能(Artificial Intelligence，AI)是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法、技术及应用系统。
人工智能基础技术一般包括如传感器、专用人工智能芯片、云计算、分布式存储、大数据处理技术、操作/交互系统、机电一体化等技术。人工智能软件技术主要包括计算机视觉技术、机器人技术、生物识别技术、语音处理技术、自然语言处理技术以及机器学习/深度学习等几大方向。
在一些实施例中，对所述用户数据进行分词处理，得到多个分词；对每个所述分词进行 特征提取，得到每个所述分词的嵌入特征；根据每个所述分词的嵌入特征进行词义预测，得到每个所述分词对应的词义；根据每个所述分词对应的词义对所述多个分词进行筛选，得到关键信息。由此可以能够准确地提取得到关键信息，提高脱敏数据生成的准确性以及安全性。
其中，所述嵌入特征为词嵌入特征，位置嵌入特征和切分嵌入特征。所述词嵌入特征是每个分词的向量表示，所述位置嵌入为每个分词位置的向量表示，所述切分嵌入特征用于区分不同的两句话。
具体地，可以基于分词算法对所述用户数据进行分词处理，得到多个分词，所述分词算法可以为正向最大匹配法、逆向最大匹配法、基于隐马尔可夫模型的分词算法、基于条件随机场的分词算法等算法。
示例性的，可以基于隐马尔可夫模型的分词算法对用户数据比如病历数据的病历文本为“患者存在尿频、易饿、焦虑、震颤等症状，疑似糖尿病，”进行分词，得到对应的多个分词比如尿频、易饿、焦虑、震颤等。
具体地，可以对每个所述分词进行特征提取，得到每个所述分词的嵌入特征，并基于词义预测模型，根据每个所述分词的嵌入特征对每个所述分词进行词义预测，得到每个分词对应的词义预测结果，并基于每个分词对应的词义预测结果对所述多个分词进行筛选，得到关键信息。由此能够最大程度地挖掘文本特征，抽取更丰富的词表示，消除传统词向量比如Word2vec、Glove等无法动态表示上下文信息和无法解决一词多义的缺点。从而可以快速得到每个分词与预设的标准敏感分词的相似度，进而快速得到对应的关键信息。
其中，所述词义预测模型用于预测每个分词与预设的标准敏感分词的相似程度，所述词义预测模型通过语义匹配模型与标准敏感分词数据库进行训练得到，所述语义预测模型可以包括LSTM匹配模型、MV-DSSM模型、ESIM模型等模型，所述词义预测结果为每个分词与标准敏感分词数据库中的标准敏感分词的相似度。
示例性的，比如分词包括账户信息的账户余额等分词以及股票走势信息等分词，可以对每个所述分词进行特征提取，得到每个所述分词的词嵌入特征，位置嵌入特征和切分嵌入特征，基于LSTM匹配模型，根据词嵌入特征，位置嵌入特征和切分嵌入特征对每个所述分词进行词义预测，得到每个分词对应的词义预测结果，并基于每个分词对应的词义预测结果将股票走势信息对应的分词筛选掉，得到关键信息。
S102、对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理。
由于关键信息一般都是连续型数据，因此需要对连续型数据和离散型数据进行表示转换，即数据预处理操作，是神经网络的输入和输出的关键一步。
示例性的，当关键信息为身高、体重等信息时，则该关键信息为连续型数据，当关键信息为投资企业个数等信息时，则该关键信息为离散型数据。
其中，所述离散变量指变量值可以按一定顺序一一列举，通常以整数位取值的变量，比如职工人数、工厂数、机器台数等。具体地，所述数据归一化处理可以包括最大最小归一化处理和根据高斯混合模型进行归一化处理；所述数据离散化处理可以包括K-bins离散化处理和回归树离散化处理。
在一些实施例中，对所述关键信息进行最大最小归一化处理，得到所述关键信息对应的离散变量；或，通过高斯混合模型对所述关键信息进行归一化处理，得到所述关键信息对应的离散变量；或，对所述关键信息进行K-bins离散化处理，得到所述关键信息对应的离散变量；或，对所述关键信息进行回归树离散化处理，得到所述关键信息对应的离散变量。
具体地，若关键信息为连续型数据，可以通过最大最小线性变换被映射到[0,1]的范围内，使得利用tanh激活函数可以表示该连续值，得到所述关键信息对应的离散变量。
具体地，若关键信息为连续型数据，可以通过高斯混合模型对关键信息进行拟合，根据该关键信息在混合模型的高斯分量的概率来采样高斯分量，用采样出来的高斯分量对记录中的关键信息进行归一化表示。那么，关键信息即可由归一化后的表示以及高斯分量的独热编 码共同构成，从而得到所述关键信息对应的离散变量。
具体地，若关键信息为连续型数据，可以对所述关键信息进行K-bins离散化处理，得到所述关键信息对应的离散变量。其中，所述离散化，也可称为分箱，即按一定规则将关键信息分入各个区间中，并用独热编码表示每个区间，从而将关键信息用一个含有四个区间的分段函数进行拟合，得到所述关键信息对应的离散变量。
具体地，若关键信息为连续型数据，可以利用CART回归树对所述关键信息进行离散化处理，得到所述关键信息对应的离散变量。其中，CART回归树可以预测连续型数据，其叶子节点即表示一个预测值。将关键信息的回归树或回归树集合的一系列叶子节点通过独热编码表示，即可将该关键信息转化为离散值。
需要说明的是，若关键信息为离散型数据，则无需进行数据离散化处理或数据归一化处理。
S103、基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量。
其中，所述条件损失函数为基于对抗生成网络的条件损失函数，所述基于对抗生成网络的损失函数的数据项是基于条件概率生成的，其初衷是使得数据能够按条件生成，使得同样类型的待脱敏数据与生成的脱敏数据的分布尽可能一致。然而，因为每次采样出来的条件都可能是不同的变量，因此，基于条件变量下的数据很难得到充分的训练，可以观察到生成的数据中对应的变量的值与其条件变量的值不一致。可通过对条件变量的预测，对训练过程进行约束，使得条件变量的值与生成的数据中对应的变量的值一致，即可进一步优化数据生成的效果。
具体地，所述条件嵌入向量可以随机等概率地从所述关键信息对应的多个离散变量中挑选出一个符合预设条件的离散变量，所述隐向量可以从所述关键信息对应的白噪声采样得到，所述拼接向量是通过所述条件嵌入向量与所述隐向量拼接得到的，用于作为生成器的输入。通过加入隐向量，从而改变了脱敏后数据与原数据一对一的映射关系，从而使脱敏后的数据不容易被逆向破解，得到隐私信息。
具体地，可以通过构建离散变量下各个值的概率质量分布函数，得到该离散变量的分布式表示，并对该离散变量的分布式表示进行条件随机采样处理，得到条件嵌入向量和隐向量。
示例性的，可以通过深度神经网络转换离散变量对应的白噪声，以从该离散变量的分布式表示生成隐向量。
在一些实施例中，对所述条件嵌入向量进行转化处理，得到独热编码；将所述独热编码与所述隐向量进行拼接，得到拼接向量。其中，所述独热编码即One-Hot编码，又称有效编码，其方法是使用N位状态寄存器来对N个状态进行编码，每个状态都由他独立的寄存器位，并且在任意时候，其中只有一位有效。将所述条件嵌入向量转化为独热编码能够解决判别器不好处理属性数据的问题，同时在一定程度上也起到了扩充向量特征的作用。
具体地，可以通过深度神经网络对条件嵌入向量进行转化处理，得到独热编码，并通过独热编码与隐向量进行拼接，得到拼接向量。由此可以得到符合生成器输入需求的拼接向量。
S104、将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
其中，所述预训练好的生成器是基于对抗生成网络训练生成的，所述脱敏数据为对待脱敏数据中关键信息进行脱敏后的数据。
在一些实施例中，获取训练数据对应的拼接向量，并将所述拼接向量输入到第一生成器进行脱敏处理，得到脱敏后的数据；基于所述脱敏后的数据和训练数据对预设的判别器进行训练，得到预训练好的判别器；根据预设的学习率和所述预训练好的判别器的参数，对所述第一生成器的参数进行多次迭代更新，得到第二生成器，并将所述第二生成器作为预训练好的生成器。由此可以通过预训练好的判别器和脱敏后的数据对第一生成器的参数进行多次迭代更新，能够生成非常真实的脱敏数据。之所以要先训练得到预训练好的判别器，再训练生成器，是因为要先拥有一个好的判别器，使得能够教好地区分出待脱敏数据和生成的脱敏数 据之后，才能够更为准确地对生成器的参数进行更新。
其中，所述训练数据为用于训练生成器参数的的待脱敏数据集，所述第一生成器为预设的未经过训练的生成器，所述第二生成器为第一生成器经过多次迭代更新生成的。其中，所述第一生成器和所述第二生成器的参数不同。可以通过离散变量的分布式表示得到该离散变量的先验概率，并在先验概率中采样出参数，以此作为第一生成器的参数。具体地，可以通过随机梯度哈密顿蒙特卡洛方法对生成器和判别器进行训练，得到预训练好的生成器和预训练好的判别器。
具体地，基于所述脱敏后的数据和训练数据对预设的判别器进行训练，得到预训练好的判别器是通过将所述条件嵌入向量分别与所述脱敏后的数据和所述训练数据进行拼接，得到第一拼接数据和第二拼接数据，并根据计算第一拼接数据和第二拼接数据的相似度，并根据第一拼接数据和第二拼接数据的相似度优化损失函数，并对通过该损失函数对判别器进行梯度裁剪，得到预训练好的判别器。
示例性的，可以通过第一生成器和预设的判别器参数对判别器参数进行训练，将脱敏数据尽可能判别为假，从而调整判别器参数，进而提高判别器对于待脱敏数据的判别能力。
示例性的，可以通过第一生成器和预训练好的判别器参数的先验概率计算第二生成器的后验概率，从而使该脱敏数据要尽可能使得判别器误判其为待脱敏数据，从而调整生成器参数，能够生成真实的脱敏数据。
在一些实施例中，得到第二生成器之后，基于统计信息的损失函数，对所述第二生成器进行增噪处理，得到预训练好的生成器，其中，所述第一生成器、所述参数更新后的生成器和所述预训练好的生成器的参数不同。由此可以控制脱敏数据的生成质量和脱敏程度。
其中，所述基于统计信息的损失函数可以包括基于均值的损失函数、基于方差的损失函数等。
具体地，可以对第二生成器的参数加入高斯噪声，由此可以实现多项式拟合正弦曲线，所述高斯噪声是符合高斯正态分布的误差。而高斯噪声的具体数值可以通过实验得到。
示例性的，可以对第二生成器的参数引入误差项，从而对第二生成器的参数进行修正，得到预训练好的生成器。由于误差项的存在，使得生成的脱敏数据与原数据存在一定差别，但差别不大，由此可以避免经过脱敏的数据往往与原数据差别较大而失去了研究的价值，同时也保证数据不会被轻易被逆向。
在一些实施例中，得到脱敏数据之后，对所述脱敏数据的离散变量进行随机采样处理，得到目标离散变量；基于逻辑回归模型，根据所述脱敏数据的其余离散变量对所述目标离散变量进行预测，得到所述目标离散变量的预测结果；基于所述目标离散变量的预测结果调整所述预训练好的生成器的参数。由此可以通过对离散变量进行预测，从而调整生成器的参数，进而达到更好的脱敏效果。这里更好的脱敏效果是指可以防止脱敏后的数据被逆向破解，同时还能保持与原数据的关联。
其中，所述目标离散变量为从脱敏数据的多个离散变量中随机采样得到的，同时为了脱敏数据与原数据的关联离散变量，一般可以认为目标离散变量不发生改变，脱敏数据与原数据差别较小，便不会失去了研究的价值，因此需要保证目标离散变量的一致性。所述逻辑回归模型用于对离散变量进行预测。
具体地，可以利用交叉熵损失函数来判断生成的目标离散变量的预测结果与目标离散变量是否一致，从而确定所述脱敏数据的生成质量。若目标离散变量的预测结果与目标离散变量一致，则无需对所述预训练好的生成器的参数进行调整；若目标离散变量的预测结果与目标离散变量不一致，则确定目标离散变量的预测结果与目标离散变量的差值，根据所述差值调整所述预训练好的生成器的参数。由此可以确定目标离散变量的准确性，避免生成的脱敏数据使原数据差别过大。由于脱敏数据与原数据的大部分离散变量是相同的，因此去除其中一个离散变量，可以根据其余离散变量准确地预测得到该离散变量。
示例性的，若脱敏数据的目标离散变量是鞋码为43码，基于逻辑回归模型，则可以通过 脱敏数据的其余离散变量比如身高、体重等对目标离散变量进行预测，得到对于鞋码的预测结果，判断生成的鞋码的预测结果与脱敏数据的鞋码是否一致，比如若生成的鞋码的预测结果是鞋码为40码，则确定差值为1码，根据所述差值对所述预训练好的生成器的参数进行不断地迭代更新；若生成的鞋码的预测结果是鞋码为43码，则无需对所述预训练好的生成器的参数进行调整。
在一些实施例中，服务器还可以将用于提示用户脱敏数据已经生成的提示信息发送给终端设备。
其中，所述提示信息的方式具体可以包括应用程序(APP)或Email、短信、聊天工具，聊天工具例如微信、qq等。
示例性的，当脱敏数据已经生成时，服务器会发送脱敏数据已经生成的提示信息给终端设备以提醒用户。
请参阅图3，图3是本申请一实施例提供的一种数据脱敏装置的示意性框图，该数据脱敏装置可以配置于服务器中，用于执行前述的数据脱敏方法。
如图3所示，该数据脱敏装置200包括：关键信息提取模块201、信息处理模块202、向量拼接模块203和数据脱敏模块204。
关键信息提取模块201，用于获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；
信息处理模块202，用于对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
向量拼接模块203，用于基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
数据脱敏模块204，用于将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据；
特征提取模块201，还用于对所述用户数据进行分词处理，得到多个分词；对每个所述分词进行特征提取，得到每个所述分词的嵌入特征；根据每个所述分词的嵌入特征进行词义预测，得到每个所述分词对应的词义；根据每个所述分词对应的词义对所述多个分词进行筛选，得到关键信息。
信息处理模块202，还用于对所述关键信息进行最大最小归一化处理，得到所述关键信息对应的离散变量；或，通过高斯混合模型对所述关键信息进行归一化处理，得到所述关键信息对应的离散变量；或，对所述关键信息进行K-bins离散化处理，得到所述关键信息对应的离散变量；或，对所述关键信息进行回归树离散化处理，得到所述关键信息对应的离散变量。
向量拼接模块203，还用于对所述条件嵌入向量进行转化处理，得到独热编码；将所述独热编码与所述隐向量进行拼接，得到拼接向量。
生成器训练模块205，用于获取训练数据对应的拼接向量，并将所述拼接向量输入到第一生成器进行脱敏处理，得到脱敏后的数据；基于所述脱敏后的数据和训练数据对预设的判别器进行训练，得到预训练好的判别器；根据预设的学习率和所述预训练好的判别器的参数，对所述第一生成器的参数进行多次迭代更新，得到第二生成器，并将所述第二生成器作为预训练好的生成器。
生成器训练模块205，还用于基于统计信息的损失函数，对所述第二生成器进行增噪处理，得到预训练好的生成器，其中，所述第一生成器、所述第二生成器和所述预训练好的生成器的参数不同。
生成器训练模块205，还用于对所述脱敏数据的离散变量进行随机采样处理，得到目标离散变量；基于逻辑回归模型，根据所述脱敏数据的其余离散变量对所述目标离散变量进行预测，得到所述目标离散变量的预测结果；基于所述目标离散变量的预测结果调整所述预训练好的生成器的参数。
需要说明的是，所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，上述描述的装置和各模块、单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
本申请的方法、装置可用于众多通用或专用的计算系统环境或配置中。例如：个人计算机、服务器计算机、手持设备或便携式设备、平板型设备、多处理器系统、基于微处理器的系统、机顶盒、可编程的消费终端设备、网络PC、小型计算机、大型计算机、包括以上任何系统或设备的分布式计算环境等等。
示例性的，上述的方法、装置可以实现为一种计算机程序的形式，该计算机程序可以在如图4所示的计算机设备上运行。
请参阅图4，图4是本申请实施例提供的一种计算机设备的示意图。该计算机设备可以是服务器。
如图4所示，该计算机设备包括通过系统总线连接的处理器、存储器和网络接口，其中，存储器可以包括易失性存储介质、非易失性存储介质和内存储器。
非易失性存储介质可存储操作系统和计算机程序。该计算机程序包括程序指令，该程序指令被执行时，可使得处理器执行任意一种数据脱敏方法。
处理器用于提供计算和控制能力，支撑整个计算机设备的运行。
内存储器为非易失性存储介质中的计算机程序的运行提供环境，该计算机程序被处理器执行时，可使得处理器执行任意一种数据脱敏方法。
该网络接口用于进行网络通信，如发送分配的任务等。本领域技术人员可以理解，该计算机设备的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
应当理解的是，处理器可以是中央处理单元(Central Processing Unit，CPU)，该处理器还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。其中，通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
其中，在一些实施方式中，所述处理器用于运行存储在存储器中的计算机程序，以实现如下步骤：获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
在一些实施例中，所述处理器还用于：对所述用户数据进行分词处理，得到多个分词；对每个所述分词进行特征提取，得到每个所述分词的嵌入特征；根据每个所述分词的嵌入特征进行词义预测，得到每个所述分词对应的词义；根据每个所述分词对应的词义对所述多个分词进行筛选，得到关键信息。
在一些实施例中，所述处理器还用于：对所述关键信息进行最大最小归一化处理，得到所述关键信息对应的离散变量；或，通过高斯混合模型对所述关键信息进行归一化处理，得到所述关键信息对应的离散变量；或，对所述关键信息进行K-bins离散化处理，得到所述关键信息对应的离散变量；或，对所述关键信息进行回归树离散化处理，得到所述关键信息对应的离散变量。
在一些实施例中，所述处理器还用于：对所述条件嵌入向量进行转化处理，得到独热编码；将所述独热编码与所述隐向量进行拼接，得到拼接向量。
在一些实施例中，所述处理器还用于：获取训练数据对应的拼接向量，并将所述拼接向 量输入到第一生成器进行脱敏处理，得到脱敏后的数据；基于所述脱敏后的数据和训练数据对预设的判别器进行训练，得到预训练好的判别器；根据预设的学习率和所述预训练好的判别器的参数，对所述第一生成器的参数进行多次迭代更新，得到第二生成器，并将所述第二生成器作为预训练好的生成器。
在一些实施例中，所述处理器还用于：基于统计信息的损失函数，对所述第二生成器进行增噪处理，得到预训练好的生成器，其中，所述第一生成器、所述第二生成器和所述预训练好的生成器的参数不同。
在一些实施例中，所述处理器还用于：对所述脱敏数据的离散变量进行随机采样处理，得到目标离散变量；基于逻辑回归模型，根据所述脱敏数据的其余离散变量对所述目标离散变量进行预测，得到所述目标离散变量的预测结果；基于所述目标离散变量的预测结果调整所述预训练好的生成器的参数。
本申请实施例还提供一种计算机可读存储介质，所述计算机可读存储介质可以是非易失性，也可以是易失性。所述计算机可读存储介质上存储有计算机程序，所述计算机程序中包括程序指令，所述程序指令被执行时实现本申请实施例提供的任一种数据脱敏方法。
其中，所述计算机可读存储介质可以是前述实施例所述的计算机设备的内部存储单元，例如所述计算机设备的硬盘或内存。所述计算机可读存储介质也可以是所述计算机设备的外部存储设备，例如所述计算机设备上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。
进一步地，所述计算机可读存储介质可主要包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需的应用程序等；存储数据区可存储根据区块链节点的使用所创建的数据等。
本申请所指区块链语言模型的存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链(Blockchain)，本质上是一个去中心化的数据库，是一串使用密码学方法相关联产生的数据块，每一个数据块中包含了一批次网络交易的信息，用于验证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到各种等效的修改或替换，这些修改或替换都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以权利要求的保护范围为准。

Claims (20)

  1. 一种数据脱敏方法，其中，所述方法包括：
    获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；
    对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
    基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
    将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
  2. 根据权利要求1所述的方法，其中，所述基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息，包括：
    对所述用户数据进行分词处理，得到多个分词；
    对每个所述分词进行特征提取，得到每个所述分词的嵌入特征；
    根据每个所述分词的嵌入特征进行词义预测，得到每个所述分词对应的词义；
    根据每个所述分词对应的词义对所述多个分词进行筛选，得到关键信息。
  3. 根据权利要求1所述的方法，其中，所述对所述关键信息进行预处理，得到所述关键信息对应的离散变量，包括：
    对所述关键信息进行最大最小归一化处理，得到所述关键信息对应的离散变量；或，
    通过高斯混合模型对所述关键信息进行归一化处理，得到所述关键信息对应的离散变量；或，
    对所述关键信息进行K-bins离散化处理，得到所述关键信息对应的离散变量；或，
    对所述关键信息进行回归树离散化处理，得到所述关键信息对应的离散变量。
  4. 根据权利要求1所述的方法，其中，所述将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量，包括：
    对所述条件嵌入向量进行转化处理，得到独热编码；
    将所述独热编码与所述隐向量进行拼接，得到拼接向量。
  5. 根据权利要求1所述的方法，其中，所述方法还包括：
    获取训练数据对应的拼接向量，并将所述拼接向量输入到第一生成器进行脱敏处理，得到脱敏后的数据；
    基于所述脱敏后的数据和训练数据对预设的判别器进行训练，得到预训练好的判别器；
    根据预设的学习率和所述预训练好的判别器的参数，对所述第一生成器的参数进行多次迭代更新，得到第二生成器，并将所述第二生成器作为预训练好的生成器。
  6. 根据权利要求5所述的方法，其中，所述得到第二生成器之后，所述方法还包括：
    基于统计信息的损失函数，对所述第二生成器进行增噪处理，得到预训练好的生成器，其中，所述第一生成器、所述第二生成器和所述预训练好的生成器的参数不同。
  7. 根据权利要求1所述的方法，其中，所述得到脱敏数据之后，所述方法还包括：
    对所述脱敏数据的离散变量进行随机采样处理，得到目标离散变量；
    基于逻辑回归模型，根据所述脱敏数据的其余离散变量对所述目标离散变量进行预测，得到所述目标离散变量的预测结果；
    基于所述目标离散变量的预测结果调整所述预训练好的生成器的参数。
  8. 一种数据脱敏装置，其中，包括：
    关键信息提取模块，用于获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；
    信息处理模块，用于对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
    向量拼接模块，用于基于条件损失函数，对所述离散变量进行条件随机采样处理，得到 条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
    数据脱敏模块，用于将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
  9. 一种计算机设备，其中，所述计算机设备包括存储器和处理器；
    所述存储器，用于存储计算机程序；
    所述处理器，用于执行所述的计算机程序并在执行所述的计算机程序时实现如下步骤：
    获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；
    对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
    基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
    将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
  10. 根据权利要求9所述的计算机设备，其中，所述处理器在实现所述基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息，用于实现：
    对所述用户数据进行分词处理，得到多个分词；
    对每个所述分词进行特征提取，得到每个所述分词的嵌入特征；
    根据每个所述分词的嵌入特征进行词义预测，得到每个所述分词对应的词义；
    根据每个所述分词对应的词义对所述多个分词进行筛选，得到关键信息。
  11. 根据权利要求9所述的计算机设备，其中，所述处理器在实现所述对所述关键信息进行预处理，得到所述关键信息对应的离散变量，用于实现：
    对所述关键信息进行最大最小归一化处理，得到所述关键信息对应的离散变量；或，
    通过高斯混合模型对所述关键信息进行归一化处理，得到所述关键信息对应的离散变量；或，
    对所述关键信息进行K-bins离散化处理，得到所述关键信息对应的离散变量；或，
    对所述关键信息进行回归树离散化处理，得到所述关键信息对应的离散变量。
  12. 根据权利要求9所述的计算机设备，其中，所述处理器在实现所述将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量，用于实现：
    对所述条件嵌入向量进行转化处理，得到独热编码；
    将所述独热编码与所述隐向量进行拼接，得到拼接向量。
  13. 根据权利要求9所述的计算机设备，其中，所述处理器还用于实现：
    获取训练数据对应的拼接向量，并将所述拼接向量输入到第一生成器进行脱敏处理，得到脱敏后的数据；
    基于所述脱敏后的数据和训练数据对预设的判别器进行训练，得到预训练好的判别器；
    根据预设的学习率和所述预训练好的判别器的参数，对所述第一生成器的参数进行多次迭代更新，得到第二生成器，并将所述第二生成器作为预训练好的生成器。
  14. 根据权利要求9所述的计算机设备，其中，所述处理器在实现所述得到脱敏数据之后，还用于实现：
    对所述脱敏数据的离散变量进行随机采样处理，得到目标离散变量；
    基于逻辑回归模型，根据所述脱敏数据的其余离散变量对所述目标离散变量进行预测，得到所述目标离散变量的预测结果；
    基于所述目标离散变量的预测结果调整所述预训练好的生成器的参数。
  15. 一种计算机可读存储介质，其中，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时使所述处理器实现如下步骤：
    获取用户数据，并基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息；
    对所述关键信息进行预处理，得到所述关键信息对应的离散变量，所述预处理包括数据离散化处理或数据归一化处理；
    基于条件损失函数，对所述离散变量进行条件随机采样处理，得到条件嵌入向量和隐向量，并将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量；
    将所述拼接向量输入到预训练好的生成器进行脱敏处理，得到脱敏数据。
  16. 根据权利要求15所述的计算机可读存储介质，其中，所述处理器在实现所述基于预训练好的关键信息识别模型，对所述用户数据进行信息识别，得到关键信息，用于实现：
    对所述用户数据进行分词处理，得到多个分词；
    对每个所述分词进行特征提取，得到每个所述分词的嵌入特征；
    根据每个所述分词的嵌入特征进行词义预测，得到每个所述分词对应的词义；
    根据每个所述分词对应的词义对所述多个分词进行筛选，得到关键信息。
  17. 根据权利要求15所述的计算机可读存储介质，其中，所述处理器在实现所述对所述关键信息进行预处理，得到所述关键信息对应的离散变量，用于实现：
    对所述关键信息进行最大最小归一化处理，得到所述关键信息对应的离散变量；或，
    通过高斯混合模型对所述关键信息进行归一化处理，得到所述关键信息对应的离散变量；或，
    对所述关键信息进行K-bins离散化处理，得到所述关键信息对应的离散变量；或，
    对所述关键信息进行回归树离散化处理，得到所述关键信息对应的离散变量。
  18. 根据权利要求15所述的计算机可读存储介质，其中，所述处理器在实现所述将所述条件嵌入向量与所述隐向量进行拼接，得到拼接向量，用于实现：
    对所述条件嵌入向量进行转化处理，得到独热编码；
    将所述独热编码与所述隐向量进行拼接，得到拼接向量。
  19. 根据权利要求15所述的计算机可读存储介质，其中，所述处理器还用于实现：
    获取训练数据对应的拼接向量，并将所述拼接向量输入到第一生成器进行脱敏处理，得到脱敏后的数据；
    基于所述脱敏后的数据和训练数据对预设的判别器进行训练，得到预训练好的判别器；
    根据预设的学习率和所述预训练好的判别器的参数，对所述第一生成器的参数进行多次迭代更新，得到第二生成器，并将所述第二生成器作为预训练好的生成器。
  20. 根据权利要求15所述的计算机可读存储介质，其中，所述处理器在实现所述得到脱敏数据之后，还用于实现：
    对所述脱敏数据的离散变量进行随机采样处理，得到目标离散变量；
    基于逻辑回归模型，根据所述脱敏数据的其余离散变量对所述目标离散变量进行预测，得到所述目标离散变量的预测结果；
    基于所述目标离散变量的预测结果调整所述预训练好的生成器的参数。
PCT/CN2022/089872 2021-10-21 2022-04-28 数据脱敏方法、数据脱敏装置、设备及存储介质 Ceased WO2023065632A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111229481.X 2021-10-21
CN202111229481.XA CN113886885B (zh) 2021-10-21 2021-10-21 数据脱敏方法、数据脱敏装置、设备及存储介质

Publications (1)

Publication Number Publication Date
WO2023065632A1 true WO2023065632A1 (zh) 2023-04-27

Family

ID=79004109

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/089872 Ceased WO2023065632A1 (zh) 2021-10-21 2022-04-28 数据脱敏方法、数据脱敏装置、设备及存储介质

Country Status (2)

Country Link
CN (1) CN113886885B (zh)
WO (1) WO2023065632A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116757748A (zh) * 2023-08-14 2023-09-15 广州钛动科技股份有限公司 基于随机梯度攻击的广告点击预测方法
CN117290888A (zh) * 2023-11-23 2023-12-26 江苏风云科技服务有限公司 一种大数据的信息脱敏方法、存储介质及服务器
CN117744127A (zh) * 2024-02-20 2024-03-22 北京佳芯信息科技有限公司 基于数据信息保护的数据加密认证方法及系统
CN117912624A (zh) * 2024-03-15 2024-04-19 江西曼荼罗软件有限公司 一种电子病历共享方法及系统
CN117932676A (zh) * 2024-01-26 2024-04-26 湖北消费金融股份有限公司 基于网络接口访问控制的数据脱敏方法及系统
CN118278051A (zh) * 2024-06-03 2024-07-02 广州青莲网络科技有限公司 一种基于人工智能的数据脱敏方法及系统
CN118748614A (zh) * 2024-07-12 2024-10-08 青岛海高设计制造有限公司 数据处理方法、装置、设备和存储介质
CN118840269A (zh) * 2024-06-24 2024-10-25 大卓智能科技有限公司 图像和视频的脱敏模型的部署方法、装置、设备及介质
CN119808135A (zh) * 2024-11-12 2025-04-11 北京航空航天大学 面向离散时间动态图的隐私保护方法、装置和电子设备

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113886885B (zh) * 2021-10-21 2025-01-17 平安科技（深圳）有限公司 数据脱敏方法、数据脱敏装置、设备及存储介质
CN115514564B (zh) * 2022-09-22 2023-06-16 成都坐联智城科技有限公司 基于数据共享的数据安全处理方法及系统
CN116361858B (zh) * 2023-04-10 2024-01-26 北京无限自在文化传媒股份有限公司 应用ai决策的用户会话资源数据保护方法及软件产品
CN116629984B (zh) * 2023-07-24 2024-02-06 中信证券股份有限公司 基于嵌入模型的产品信息推荐方法、装置、设备和介质
CN119052506B (zh) * 2024-11-01 2025-02-28 南京信息工程大学 基于扩展隐空间和对比学习的生成式隐写方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110188202A (zh) * 2019-06-06 2019-08-30 北京百度网讯科技有限公司 语义关系识别模型的训练方法、装置及终端
CN111143884A (zh) * 2019-12-31 2020-05-12 北京懿医云科技有限公司 数据脱敏方法与装置、电子设备及存储介质
WO2021027533A1 (zh) * 2019-08-13 2021-02-18 平安国际智慧城市科技股份有限公司 文本语义识别方法、装置、计算机设备和存储介质
CN113254649A (zh) * 2021-06-22 2021-08-13 中国平安人寿保险股份有限公司 敏感内容识别模型的训练方法、文本识别方法及相关装置
CN113886885A (zh) * 2021-10-21 2022-01-04 平安科技（深圳）有限公司 数据脱敏方法、数据脱敏装置、设备及存储介质

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107368752B (zh) * 2017-07-25 2019-06-28 北京工商大学 一种基于生成式对抗网络的深度差分隐私保护方法
CN110263152B (zh) * 2019-05-07 2024-04-09 平安科技（深圳）有限公司 基于神经网络的文本分类方法、系统及计算机设备
CN110135193A (zh) * 2019-05-15 2019-08-16 广东工业大学 一种数据脱敏方法、装置、设备及计算机可读存储介质
CN110807207B (zh) * 2019-10-30 2021-10-08 腾讯科技（深圳）有限公司 数据处理方法、装置、电子设备及存储介质
CN111768325B (zh) * 2020-04-03 2023-07-25 南京信息工程大学 大数据隐私保护中基于生成对抗样本的安全性提升方法
CN111563275B (zh) * 2020-07-14 2020-10-20 中国人民解放军国防科技大学 一种基于生成对抗网络的数据脱敏方法
CN113297573B (zh) * 2021-06-11 2022-06-10 浙江工业大学 一种基于gan模拟数据生成的垂直联邦学习防御方法和装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110188202A (zh) * 2019-06-06 2019-08-30 北京百度网讯科技有限公司 语义关系识别模型的训练方法、装置及终端
WO2021027533A1 (zh) * 2019-08-13 2021-02-18 平安国际智慧城市科技股份有限公司 文本语义识别方法、装置、计算机设备和存储介质
CN111143884A (zh) * 2019-12-31 2020-05-12 北京懿医云科技有限公司 数据脱敏方法与装置、电子设备及存储介质
CN113254649A (zh) * 2021-06-22 2021-08-13 中国平安人寿保险股份有限公司 敏感内容识别模型的训练方法、文本识别方法及相关装置
CN113886885A (zh) * 2021-10-21 2022-01-04 平安科技（深圳）有限公司 数据脱敏方法、数据脱敏装置、设备及存储介质

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116757748A (zh) * 2023-08-14 2023-09-15 广州钛动科技股份有限公司 基于随机梯度攻击的广告点击预测方法
CN116757748B (zh) * 2023-08-14 2023-12-19 广州钛动科技股份有限公司 基于随机梯度攻击的广告点击预测方法
CN117290888A (zh) * 2023-11-23 2023-12-26 江苏风云科技服务有限公司 一种大数据的信息脱敏方法、存储介质及服务器
CN117290888B (zh) * 2023-11-23 2024-02-09 江苏风云科技服务有限公司 一种大数据的信息脱敏方法、存储介质及服务器
CN117932676A (zh) * 2024-01-26 2024-04-26 湖北消费金融股份有限公司 基于网络接口访问控制的数据脱敏方法及系统
CN117744127A (zh) * 2024-02-20 2024-03-22 北京佳芯信息科技有限公司 基于数据信息保护的数据加密认证方法及系统
CN117744127B (zh) * 2024-02-20 2024-05-07 北京佳芯信息科技有限公司 基于数据信息保护的数据加密认证方法及系统
CN117912624A (zh) * 2024-03-15 2024-04-19 江西曼荼罗软件有限公司 一种电子病历共享方法及系统
CN118278051A (zh) * 2024-06-03 2024-07-02 广州青莲网络科技有限公司 一种基于人工智能的数据脱敏方法及系统
CN118840269A (zh) * 2024-06-24 2024-10-25 大卓智能科技有限公司 图像和视频的脱敏模型的部署方法、装置、设备及介质
CN118840269B (zh) * 2024-06-24 2026-01-06 奇瑞汽车股份有限公司 图像和视频的脱敏模型的部署方法、装置、设备及介质
CN118748614A (zh) * 2024-07-12 2024-10-08 青岛海高设计制造有限公司 数据处理方法、装置、设备和存储介质
CN119808135A (zh) * 2024-11-12 2025-04-11 北京航空航天大学 面向离散时间动态图的隐私保护方法、装置和电子设备

Also Published As

Publication number Publication date
CN113886885A (zh) 2022-01-04
CN113886885B (zh) 2025-01-17

Similar Documents

Publication Publication Date Title
WO2023065632A1 (zh) 数据脱敏方法、数据脱敏装置、设备及存储介质
US12153693B2 (en) Sensitive data classification
US20250103746A1 (en) System and method for providing a privacy-aware prompt engineering system
CN113726784B (zh) 一种网络数据的安全监控方法、装置、设备及存储介质
US10430610B2 (en) Adaptive data obfuscation
CN113761577B (zh) 一种大数据脱敏的方法、装置、计算机设备及存储介质
US20240061952A1 (en) Identifying sensitive data using redacted data
US12111933B2 (en) System and method for dynamically updating existing threat models based on newly identified active threats
WO2021151358A1 (zh) 基于解释模型的分诊信息推荐方法、装置、设备及介质
WO2021189975A1 (zh) 机器行为识别方法、装置、设备及计算机可读存储介质
US12387007B2 (en) Personally identifiable information scrubber with language models
WO2020234515A1 (en) Compatible anonymization of data sets of different sources
CN113901821B (zh) 一种实体命名识别方法、装置、设备及存储介质
CN113326699B (zh) 数据检测方法、电子设备及存储介质
CN117009832A (zh) 异常命令的检测方法、装置、电子设备及存储介质
CN121029922A (zh) 问答场景的检索增强方法、装置、设备和存储介质
CN120632935A (zh) 一种文本隐私脱敏方法、装置、电子设备及存储介质
US20220351210A1 (en) Method and system for detection of abnormal transactional behavior
CN118277504B (zh) 一种基于关联性分析的文档关联存储方法及相关装置
CN113706174A (zh) 基于机器学习算法的违约预测方法、装置、设备及介质
KR20210130048A (ko) 분석 모델 학습 방법 및 장치
HK40063339A (zh) 数据脱敏方法、数据脱敏装置、设备及存储介质
HK40063339B (zh) 数据脱敏方法、数据脱敏装置、设备及存储介质
CN115687979A (zh) 威胁情报中指定技术的识别方法及装置、电子设备、存储介质
CN121502824B (zh) 一种提示词隐私保护方法、装置、设备、产品及介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22882257

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22882257

Country of ref document: EP

Kind code of ref document: A1