WO2020073492A1 - 数据安全处理方法、装置、计算机设备及存储介质 - Google Patents

数据安全处理方法、装置、计算机设备及存储介质 Download PDF

Info

Publication number
WO2020073492A1
WO2020073492A1 PCT/CN2018/122734 CN2018122734W WO2020073492A1 WO 2020073492 A1 WO2020073492 A1 WO 2020073492A1 CN 2018122734 W CN2018122734 W CN 2018122734W WO 2020073492 A1 WO2020073492 A1 WO 2020073492A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
original
user data
target
training
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/CN2018/122734
Other languages
English (en)
French (fr)
Inventor
史光辉
王涵
王建明
肖京
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 WO2020073492A1 publication Critical patent/WO2020073492A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/60Protecting data
    • G06F21/62Protecting access to data via a platform, e.g. using keys or access control rules
    • G06F21/6218Protecting 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
    • G06F21/6245Protecting personal data, e.g. for financial or medical purposes
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/243Classification techniques relating to the number of classes
    • G06F18/24323Tree-organised classifiers

Definitions

  • This application relates to the field of data security technology, and in particular, to a data security processing method, device, computer equipment, and storage medium.
  • Embodiments of the present application provide a data security processing method, device, computer equipment, and storage medium, to solve the problem of large data volume for user data desensitization processing, inconvenient operation, and time and energy consumption.
  • a data security processing method including:
  • the data processing request includes a target combination feature
  • the valid user data is regarded as safe user data.
  • a data security processing device including:
  • a data processing request obtaining module configured to obtain a data processing request, the data processing request including target combination characteristics
  • a user data acquisition module to be tested configured to acquire user data to be tested from a user database according to the target combination characteristics
  • An effective user data acquisition module configured to determine a target noise extraction range according to a data security processing model, process the user data to be tested based on the target noise extraction range, and obtain effective user data;
  • a data security processing module configured to input the effective user data into the data security processing model, and obtain an output value of a target combination characteristic corresponding to the effective user data
  • the safe user data acquisition module is configured to use the effective user data as safe user data when the output value of the target combination feature is within a preset monitoring range.
  • a computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor.
  • the processor executes the computer-readable instructions, the following steps are implemented:
  • the data processing request includes a target combination feature
  • the valid user data is regarded as safe user data.
  • One or more non-volatile readable storage media storing computer-readable instructions, which when executed by one or more processors, cause the one or more processors to implement the following steps:
  • the data processing request includes a target combination feature
  • the valid user data is regarded as safe user data.
  • FIG. 1 is an application scenario diagram of a data security processing method in an embodiment of the present application
  • FIG. 2 is a flowchart of a data security processing method in an embodiment of the present application
  • FIG. 3 is a specific flowchart of step S30 in FIG. 2;
  • FIG. 4 is a flowchart before step S30 in FIG. 2;
  • FIG. 5 is a specific flowchart of step S301 in FIG. 4;
  • FIG. 6 is a schematic diagram of a data security processing device in an embodiment of the present application.
  • FIG. 7 is a schematic diagram of a computer device in an embodiment of the present application.
  • the data security processing method provided by this application can be applied in the application environment as shown in FIG. 1, in which the client communicates with the server through the network.
  • the client may be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server can be implemented by an independent server or a server cluster composed of multiple servers.
  • a data security processing method is provided.
  • the method is applied to the server in FIG. 1 as an example for illustration, and includes the following steps:
  • S10 Acquire a data processing request, and the data processing request includes a target combination feature.
  • the data processing request in this embodiment refers to a request sent by a client to a server for processing user data, and the data processing request includes a target combination feature.
  • the target combination feature refers to the feature about the user data that the user fills in the client as needed, including but not limited to the user's age, gender, whether there is a car, whether to purchase insurance (including but not limited to car insurance, property insurance) stored in the user database And life insurance). Understandably, the target combination feature includes at least one feature of user data.
  • the user database refers to a database for storing user data.
  • the user data stored in the user database in this embodiment includes but is not limited to the user's basic information, purchase behavior and transaction data.
  • basic information includes but is not limited to name, age (date of birth), gender, birthplace, ethnicity, educational background, and work experience.
  • Purchase behavior refers to the behavior of users purchasing products on the platform, including but not limited to financial products and insurance products purchased by users.
  • Transaction data refers to the recorded data of the user's investment, consumption and transfer.
  • the server After the user sets the target combination feature on the client, it is sent to the server, and the server obtains user data corresponding to the target combination feature from the user database as the user data to be tested according to the obtained target combination feature.
  • the user data to be tested refers to user data that needs to be tested.
  • S30 Determine the target noise extraction range according to the data security processing model, process the user data to be tested based on the target noise extraction range, and obtain valid user data.
  • this embodiment adds user data corresponding to the non-target combination feature to the user data to be tested, so that the acquired effective user data is not all user data corresponding to the target combination feature, which reduces the accuracy of basic information user data. Ensure the security of user data.
  • the data security processing model refers to a pre-trained model for determining the target noise extraction range.
  • the target noise extraction range refers to the range used to extract user data corresponding to non-target combination features.
  • the data security processing model in this embodiment includes a target gradient lifting decision tree model and a target logistic regression model.
  • the target gradient lifting decision tree model refers to the pre-trained gradient lifting decision tree that meets user requirements.
  • Gradient lifting decision tree (Gradient Boosting Decision Tree, GBDT) is an iterative decision tree algorithm, which consists of multiple decision trees. The conclusions of all trees are added up as the final prediction result.
  • the target logistic regression model refers to a pre-trained logistic regression model used to obtain the target noise extraction range.
  • Logistic regression model (Logistic Regression, LR) is a regression model that reduces the prediction range and limits the predicted value to [0,1].
  • the target noise extraction range is determined according to the security processing model, and user data corresponding to a specific number of non-target combination features is extracted from the user database and added to the user data to be tested to generate valid user data.
  • Effective user data refers to data obtained by adding user data corresponding to non-target combination features to the user data to be tested. It should be noted that the specific number in this embodiment refers to the number within the target noise extraction range.
  • the server extracts the user data corresponding to the non-target combination feature according to the target noise extraction range and adds it to the user data to be tested, so that the secure user data obtained in the subsequent steps can not only ensure the security of the user data, but also meet the demand for the value of the user data.
  • the effective user data after obtaining the effective user data, it is necessary to further verify the effective user data to determine whether the effective user data meets user requirements. Therefore, it is also necessary to input the acquired effective user data into a pre-trained data security processing model for processing, and obtain the output value of the effective user data after being processed by the data security processing model.
  • the output value is a probability value used to detect whether valid user data meets the requirements set by the user.
  • the preset monitoring range refers to a range set in advance for detecting whether valid user data meets user requirements.
  • the preset monitoring range set in this implementation is 70% -90%, which can not only ensure the security of user data, but also meet the value requirements for user data.
  • the calculated value of the target combination feature is within the preset monitoring range through the calculation of the data security processing model, it means that the valid user data meets the requirements, and the valid user data is regarded as the safe user data.
  • Steps S10-S50 Obtain the user data to be tested from the user database through the target combination feature, and then extract the user data corresponding to a specific number of non-target combination features according to the target noise extraction range determined by the data security processing model and add it to the user to be tested In the data, to obtain effective user data, reduce the accuracy of effective user data to ensure the safety of user data.
  • the effective user data After obtaining effective user data, in order to further determine whether the effective user data meets the requirements, it is also necessary to input the effective user data into the data security processing model to obtain the output value of the target combination feature corresponding to the effective user data, when the output value is in the preset
  • the monitoring scope indicates that the effective user data meets the requirements, and the effective user data is sent as safe user data to the client that needs to send the data processing request, which not only ensures that the user of the data cannot easily infer the user's personal privacy information, but also maximizes The value of the target data to the fullest extent.
  • step S30 processing the user data to be tested based on the target noise extraction range to obtain valid user data specifically includes the following steps:
  • the server selects user data corresponding to a specific number of non-target combination features from the user database as target noise data according to the target noise extraction range, and provides a data source for subsequent steps.
  • the server selects any number from 5000-25000 from the user database according to the target noise extraction range 5000-25000 The user data corresponding to the non-target combination feature is used as the target noise data.
  • the target noise data is added to the user data to be tested, and the effective user data is acquired, so that the incomplete effective user data is the user data corresponding to the target combination feature, reducing the accuracy of the effective user data, making Effective user data can not only satisfy users' needs for user data, but also ensure the security of user data.
  • Steps S31-S32 select the user data corresponding to the non-target combination feature from the user database as the target noise data through the target noise extraction range, and add the target noise data to the user data to be tested to obtain valid user data to ensure the user data Security.
  • the data security processing method before step S30, that is, before the step of determining the target noise extraction range according to the data security processing model, the data security processing method further includes the following steps:
  • S301 Obtain data to be trained, and divide the data to be trained into a training set and a test set.
  • the data to be trained is obtained from the sample database, and the data to be trained is divided into a training set and a test set for training models and test models.
  • the model to be trained in this embodiment refers to the original gradient lifting decision tree model.
  • the sample database refers to a database used to store data to be trained.
  • the data to be trained refers to the data used to train the original gradient lifting decision tree model.
  • a training set is a set of data to be trained for training the original gradient boosting decision tree model.
  • a test set is a set of data to be trained for testing whether the trained original gradient lifting decision tree model is accurate.
  • the model parameters include the maximum depth and the maximum number of iterations of the gradient lifting decision tree.
  • the model parameters in the original gradient lifting decision tree model need to be initially set.
  • the model parameters in this embodiment include the maximum depth and the maximum number of iterations of the original gradient lifting decision tree. According to experimental data, the maximum depth of the original gradient lifting decision tree is set to 3, and the maximum number of iterations is set to 50, which has the best effect. In this embodiment, the maximum depth is set to 3 and the maximum number of iterations is 50.
  • Initializing the model parameters of the original gradient lifting decision tree model can shorten the training time and improve the recognition accuracy when training the original gradient lifting decision tree model.
  • the data to be trained in the training set is input into the original gradient lifting decision tree model, and the original gradient lifting decision tree model will be included in the target combination feature corresponding to the training data Select a feature as the first bifurcation point, obtain the residual of the training data at the bifurcation point in the training set, and then bifurcate the remaining features in the target combination feature again, and the residual corresponding to the first bifurcation point
  • the difference is used as the input of the second decision tree and iterates continuously.
  • the training of the original gradient lifting decision tree model is stopped and the original gradient lifting decision tree is obtained.
  • the original combined feature refers to the combined feature corresponding to the training set input to the original gradient lifting decision tree model, and corresponding to each decision tree path.
  • the original combination feature is input into the original logistic regression model, and the probability corresponding to the original combination feature is calculated by the Sigmoid function, that is, the output value corresponding to the original combination feature.
  • the Sigmoid function in this embodiment can be expressed as: Among them, g (x) is the probability of occurrence of the original combination feature, g (x) ⁇ (0,1), x is the original combination feature, T is the parameter that the user changes according to the actual setting ⁇ , ⁇ is the original gradient
  • the output value corresponding to the original combined feature is within the preset monitoring range, it means that the original gradient lifting decision tree model and the original logistic regression model are successfully trained, and the original gradient lifting decision tree model and the original logistic regression model are used as the target gradient lifting Decision tree model and objective logistic regression model.
  • S306 The target gradient lifting decision tree model and the target logistic regression model are tested using the data to be trained corresponding to the test set. If the obtained output value corresponding to each data to be trained is within the preset monitoring range, the target gradient Improve the decision tree model and the target logistic regression model as the data security processing model.
  • the original gradient lifting decision tree model and the original logistic regression model are used as data security processing models.
  • Steps S301 to S306 the original gradient lifting decision tree model is trained from the training data to be trained, and the original combined features corresponding to each decision tree path in the original gradient lifting decision tree model are input into the original logistic regression model to obtain the original The output value corresponding to the combined feature. If the output value corresponding to the original combination feature is within the preset monitoring range, it means that the original gradient lifting decision tree model and the original logistic regression model are successfully trained. In order to prevent overfitting, it is also necessary to input the training data corresponding to the test set into the trained target gradient lifting decision tree model and target logistic regression model for testing.
  • the output value corresponding to each training data is within the preset monitoring range, it means that the trained target gradient lifting decision tree model and the target logistic regression model meet the requirements, and can be used as a data security processing model to finally determine the target noise extraction range and use it for subsequent steps to obtain valid user data.
  • step S301 the data to be trained is acquired, and the data to be trained is divided into a training set and a test set, specifically including the following steps:
  • the model training request includes training combination features.
  • the model training request refers to a request sent by the client for model training.
  • the training combination feature refers to the feature of user data for training the model set by the user on the client as needed. Specifically, the user sets the training combination feature on the client as required, and then clicks the send operation, and the server will obtain a model training request carrying the training combination feature.
  • S3012 Select training user data matching the training combination feature and non-training user data not matching the training combination feature from the user database according to the training combination feature.
  • the server selects training user data that matches the training combination feature and non-training user data that does not match the training combination feature from the user database according to the training combination feature included in the model training request.
  • the training user data refers to user data matching the characteristics of the training combination.
  • Non-training user data refers to user data that does not match the characteristics of the training combination. If the combined training features are ages 20-30 and purchase of property insurance, the user data of ages 20-30 and purchase of property insurance will be obtained as training user data according to the training features. Risk user data as non-training user data. Obtaining training user data and non-training user data according to the characteristics of the training combination provides a data source for subsequent acquisition of original positive samples and original negative samples.
  • the training combination feature is used as the user tag of the training user data
  • the non-target combination feature is used as the user tag of the non-training user data. For example, if the training feature is 20-30 years old and purchases P & C insurance as user tags for training user data, the age is not 20-30 years old, and P & C insurance is not purchased as non-training user data tags.
  • the number of positive samples refers to a preset number for acquiring original positive samples.
  • the original positive samples refer to the training user data extracted from the training user data according to the preset number of positive samples. According to the preset number of positive samples, the original positive samples are obtained from the training user data, which provides a data source for obtaining effective positive samples and effective negative samples in the subsequent steps.
  • S3014 According to the proportion of positive and negative samples, select the corresponding non-training user data from the non-training user data as the original negative sample.
  • the server selects the corresponding non-training user data from the non-training user data according to the preset positive and negative sample ratios as the original negative samples.
  • the positive and negative sample ratio refers to a preset ratio used to determine the number of original negative samples according to the number of positive samples.
  • the number of original negative samples in this embodiment is determined according to the number of positive samples, so that the number of original negative samples and the number of positive samples are proportional to meet the requirements for subsequent model training.
  • S3015 Extract negative noise data from the original positive samples according to the first noise extraction range, and add the negative noise data to the original negative samples to generate valid negative samples.
  • the first decimation range refers to a preset range of proportions that use original positive samples as negative noise data.
  • Negative noise data refers to artificially modifying the user tags corresponding to part of the user data in the original positive sample, so that the user tags carried by the part of the user data become user tags corresponding to non-target user features.
  • the user tags carried in the user data in the original positive sample are ages 20-30 years and purchase of property insurance, and some user tags are artificially modified to user tags corresponding to non-target combination features (such as modifying the user tags to 35-45 years old and (Purchase property insurance, or modify the user label to 20-30 years old and purchase auto insurance, or modify the user label corresponding to any non-target combination features such as 30-40 years old and having a car).
  • the server randomly extracts a specific amount of user data from the original positive sample according to the first noise extraction range, and artificially modifies the user tags corresponding to the part of the user data to a non-target combination
  • the user tag corresponding to the feature, the user data of which part of the user tag is modified is regarded as negative noise data.
  • the specific number refers to any number within the first noise extraction range.
  • the number of original positive samples is 100,000
  • the first extraction range is set to 0.1-0.3.
  • 10,000-30,000 user data are randomly selected from 100,000 original positive samples.
  • the user tag carried in the user data is modified to the user tag corresponding to the non-target combination feature, and the part of the user data is regarded as negative noise data.
  • S3016 Extract positive noise data from the original negative samples according to the second noise extraction range, and add the positive noise data to the original positive samples to generate valid positive samples.
  • the second decimation range refers to a preset range of proportions that use original negative samples as positive noise data.
  • Positive noise data refers to changing the user label corresponding to some user data in the original negative sample into the user label corresponding to the target combination feature. It should be noted that since the number of positive samples and the number of negative samples are selected according to the preset ratio of positive and negative samples, the first extraction range and the second extraction range are also inconsistent, such as setting the first extraction range If it is 0.1-0.3, then the second extraction range is set to 0.05-0.2.
  • a specific amount of user data is randomly selected from the original negative sample, the user label corresponding to the part of the user data is artificially modified, the user label is modified to the user label corresponding to the target combination feature, and the part is modified
  • the original negative samples of user tags are used as positive noise data. Understandably, the specific number refers to any number within the second noise extraction range.
  • the positive noise data is added to the original positive samples to form valid positive samples. Understandably, a valid positive sample refers to a positive sample formed by adding negative noise data to the original positive sample.
  • the number of original negative samples is 1 million
  • the second extraction range is set to 0.05-0.2. According to the second extraction range, any number in the range of 50,000-200,000 is randomly selected from 1 million original negative samples.
  • User data and then artificially modify the user tags carried in the extracted user data to the user tags corresponding to the target combination feature. After the user label is modified, the original negative sample of the modified user label is added as positive noise data to the original positive sample to generate a valid positive sample.
  • S3017 Use valid positive samples and valid negative samples as the data to be trained, and store them in the sample database.
  • the effective positive samples and effective negative samples are used as training data Stored in the sample database.
  • Steps S3011 to S3017 by adding positive noise data to the original positive samples to obtain valid positive samples, and adding negative noise data to the original negative samples to obtain valid negative samples, forcibly changing the accuracy of the original positive samples and the original negative samples to make the generated effective
  • the user data in the positive sample and the effective negative sample are not all user data with target combination characteristics or non-target combination characteristics, to avoid output of real user data, and there is no need to desensitize some sensitive data in the user data one by one. , To ensure the security of user data.
  • step S304 after inputting the original combined feature into the original logistic regression model and obtaining the output value corresponding to the original combined feature, the data security processing method further includes: when the output value corresponding to the original combined feature Below the preset monitoring range, the first noise extraction range and the second noise extraction range are reduced, and the ratio of positive and negative samples is increased; when the output value corresponding to the original combined feature is higher than the preset monitoring range, the first noise is increased Decimation range and second noise extraction range, and reduce the proportion of positive and negative samples.
  • the resulting output value corresponding to the original combined feature is not within the preset monitoring range, which means the first noise extraction Inaccurate setting of the range, second noise extraction range, and positive and negative sample ratios requires dynamic adjustment of the first noise extraction range, second noise extraction range, and positive and negative sample ratios.
  • the output value corresponding to the original combined feature is lower than the preset monitoring range, it indicates that the number of negative samples is too large, and the ratio of positive and negative samples should be increased to reduce the number of negative samples.
  • the settings of the first noise extraction range and the second noise extraction range are too high, reducing the settings of the first noise extraction range and the second noise extraction range.
  • the output value corresponding to the original combination feature is higher than the preset monitoring range, it means that the number of negative samples is too small, and the ratio of positive and negative samples should be reduced and the number of negative samples should be increased.
  • the settings of the first noise extraction range and the second noise extraction range are too low, and the settings of the first noise extraction range and the second noise extraction range should be increased.
  • the data security processing method provided in this application obtains the user data to be tested from the user database through the target combination feature, and then extracts the user data corresponding to a specific number of non-target combination features according to the target noise extraction range determined by the data security processing model. In the user data to be tested, the accuracy of the user data to be tested is reduced to ensure the security of the user data.
  • the target noise extraction range is determined according to the trained data security processing model, so that the effective user data obtained according to the target noise extraction range is more accurate and meets user requirements, which not only ensures the safety of user data but also enables user data to play the value of data .
  • the range means that the effective user data meets the requirements, and the effective user data is sent to the user as safe user data.
  • the output value corresponding to the original combination feature is lower than the preset monitoring range, the first noise extraction range and the second noise extraction range are reduced, and the ratio of positive and negative samples is increased; when the output value corresponding to the original combination feature is higher than the preset monitoring range Range, the first noise extraction range and the second noise extraction range are increased, and the ratio of positive and negative samples is reduced, so that obtaining safe user data is closer to user requirements.
  • a data security processing device is provided, and the data security processing device corresponds to the data security processing method in the above embodiment in one-to-one correspondence.
  • the data security processing device includes a data processing request acquisition module 10, a user data acquisition module to be tested 20, a valid user data acquisition module 30, a data security processing module 40 and a security user data acquisition module 50.
  • the detailed description of each functional module is as follows:
  • the data processing request obtaining module 10 is used to obtain a data processing request, and the data processing request includes a target combination characteristic.
  • the user data to be tested acquiring module 20 is used to acquire user data to be tested from the user database according to the target combination characteristics.
  • the effective user data obtaining module 30 is configured to determine a target noise extraction range according to the data security processing model, process the user data to be tested based on the target noise extraction range, and obtain effective user data.
  • the data security processing module 40 is configured to input valid user data into the data security processing model and obtain the output value of the target combination feature corresponding to the valid user data.
  • the safe user data acquisition module 50 is configured to use valid user data as safe user data when the output value of the target combination feature is within a preset monitoring range.
  • the effective user data acquisition module 30 includes a target noise data acquisition unit and an effective user data acquisition unit.
  • the target noise data acquisition unit is used to select user data corresponding to non-target combination features from the user database as target noise data based on the target noise extraction range.
  • the effective user data acquisition unit is used to add target noise data to the user data to be tested to acquire effective user data.
  • the data security processing device further includes:
  • the data processing unit to be trained is used to obtain data to be trained and divide the data to be trained into a training set and a test set.
  • the model parameter initialization unit is used to initialize the model parameters of the original gradient lifting decision tree model.
  • the model parameters include the maximum depth and the maximum number of iterations of the gradient lifting decision tree.
  • the model training unit is used to input the training data corresponding to the training set into the original gradient lifting decision tree model.
  • the training depth in the original gradient lifting decision tree model reaches the maximum depth and the number of iterations reaches the maximum number of iterations, the training of the original training stops
  • the gradient lifting decision tree model obtains the original combined features corresponding to each decision tree path in the original gradient lifting decision tree model.
  • the output value calculation unit is used to input the original combined features into the original logistic regression model to obtain the output values corresponding to the original combined features.
  • the model determination unit is used to use the original gradient lifting decision tree model and the original logistic regression model as the target gradient lifting decision tree model and the target logistic regression model when the output value corresponding to the original combined feature is within the preset monitoring range.
  • the model test unit is used to test the target gradient lifting decision tree model and the target logistic regression model using the training data corresponding to the test set. If the obtained output value corresponding to each training data is within the preset monitoring range, Then the target gradient lifting decision tree model and target logistic regression model are used as the data security processing model.
  • the data processing unit to be trained includes a model training request acquisition subunit, user data selection subunit, original positive sample acquisition subunit, original negative sample acquisition subunit, effective negative sample acquisition subunit, effective positive sample acquisition subunit and The data to be trained generates a subunit.
  • the model training request obtaining subunit is used to obtain a model training request, and the model training request includes training combination features.
  • the user data selection subunit is used to select training user data that matches the training combination feature and non-training user data that does not match the training combination feature from the user database based on the training combination feature.
  • the original positive sample acquisition subunit is used to select the corresponding training user data from the training user data as the original positive sample according to the preset number of positive samples.
  • the original negative sample acquisition subunit is used to select the corresponding non-training user data from the non-training user data according to the proportion of positive and negative samples as the original negative sample.
  • the effective negative sample acquisition subunit is used to extract negative noise data from the original positive sample according to the first noise extraction range, and add the negative noise data to the original negative sample to generate an effective negative sample.
  • the effective positive sample acquisition subunit is used to extract positive noise data from the original negative sample according to the second noise extraction range, and add the positive noise data to the original positive sample to generate an effective positive sample.
  • the to-be-trained data generating subunit is used to store valid positive samples and valid negative samples as to-be-trained data in the sample database.
  • the data security processing method is also used to reduce the first noise extraction range and the second noise extraction range and increase the ratio of positive and negative samples when the output value corresponding to the original combination feature is lower than the preset monitoring range; The output value corresponding to the feature is higher than the preset monitoring range, then the first noise extraction range and the second noise extraction range are increased, and the ratio of positive and negative samples is reduced.
  • Each module in the above data security processing device may be implemented in whole or in part by software, hardware, or a combination thereof.
  • the above modules may be embedded in the hardware or independent of the processor in the computer device, or may be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 7.
  • the computer device includes a processor, memory, network interface, and database connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer-readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer-readable instructions in the non-volatile storage medium.
  • the database of the computer device is used to store data involved in the data security processing method.
  • the network interface of the computer device is used to communicate with external terminals through a network connection.
  • the computer readable instructions are executed by the processor to implement a data security processing method.
  • a computer device which includes a memory, a processor, and computer-readable instructions stored on the memory and executable on the processor.
  • the processor implements the computer-readable instructions to implement the data security processing method described above
  • the steps shown in FIG. 2 are steps S10-S50. Or the steps shown in Figures 3 to 5. To avoid repetition, I will not repeat them here.
  • the processor implements the computer-readable instructions to implement the steps of the data security processing device, as shown in FIG. 6, the data processing request acquisition module 10, the user data acquisition module to be tested 20, the effective user data acquisition module 30, and data security processing Module 40 and secure user data acquisition module 50. To avoid repetition, I will not repeat them here.
  • one or more non-volatile readable storage media storing computer-readable instructions are provided.
  • the computer-readable instructions are executed by one or more processors, the one or more processors are implemented
  • the steps of the data security processing method in the above embodiment are shown as step S10-step S50 shown in FIG. 2. Or the steps shown in Figures 3 to 5. To avoid repetition, I will not repeat them here.
  • the one or more processors implement the steps of the data security processing apparatus in the foregoing embodiment, as shown in FIG.
  • the user data acquisition module 20, the effective user data acquisition module 30, the data security processing module 40, and the secure user data acquisition module 50 To avoid repetition, I will not repeat them here.
  • Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM random access memory
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • DDRSDRAM double data rate SDRAM
  • ESDRAM enhanced SDRAM
  • SLDRAM synchronous chain (Synchlink) DRAM
  • RDRAM direct RAM
  • DRAM direct memory bus dynamic RAM
  • RDRAM memory bus dynamic RAM

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Bioethics (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Security & Cryptography (AREA)
  • Computer Hardware Design (AREA)
  • Medical Informatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本申请公开了一种数据安全处理方法、装置、计算机设备及存储介质,该方法包括获取数据处理请求,数据处理请求包括目标组合特征;根据目标组合特征,从用户数据库中获取待测用户数据;根据数据安全处理模型确定目标噪声抽取范围,基于目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;将有效用户数据输入到数据安全处理模型中,获取有效用户数据对应的目标组合特征的输出值;当目标组合特征的输出值在预设监测范围内,则将有效用户数据作为安全用户数据。该方法既保证数据的使用者不能轻易推断出用户的个人隐私信息,又能最大化程度地发挥目标数据的价值。

Description

数据安全处理方法、装置、计算机设备及存储介质
本申请以2018年10月12日提交的申请号为201811187262.8,名称为“数据安全处理方法、装置、计算机设备及存储介质”的中国发明专利申请为基础,并要求其优先权。
技术领域
本申请涉及数据安全技术领域,尤其涉及一种数据安全处理方法、装置、计算机设备及存储介质。
背景技术
近年来,随着信息技术的迅猛发展,大数据的应用越来越广泛,数据共享已成为社会发展的一种趋势。但数据在共享时需要符合特定的条件,保证不能包含个人身份的信息以及可轻易推断出个人隐私信息。如金融保险行业进行数据共享时,根据中国银行业监督管理委员会(简称银监会)和中国保险监督管理委员会(简称中国保监会)等规定,对表示用户身份的信息以及可轻易推断出用户身份的信息需要进行脱敏处理,既保证数据的使用者不能轻易推断出用户的个人隐私信息,又能最大化发挥数据的价值。当前按照规定对用户数据逐一进行脱敏处理,脱敏处理数据量大,不方便操作并且耗时耗力。
发明内容
本申请实施例提供一种数据安全处理方法、装置、计算机设备及存储介质,以解决用户数据脱敏处理数据量大,不方便操作并且耗时耗力的问题。
一种数据安全处理方法,包括:
获取数据处理请求,所述数据处理请求包括目标组合特征;
根据所述目标组合特征,从用户数据库中获取待测用户数据;
根据数据安全处理模型确定目标噪声抽取范围,基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;
将所述有效用户数据输入到所述数据安全处理模型中,获取所述有效用户数据对应的目标组合特征的输出值;
当所述目标组合特征的输出值在预设监测范围内,则将所述有效用户数据作为安全用户数据。
一种数据安全处理装置,包括:
数据处理请求获取模块,用于获取数据处理请求,所述数据处理请求包括目标组合特征;
待测用户数据获取模块,用于根据所述目标组合特征,从用户数据库中获取待测用户数据;
有效用户数据获取模块,用于根据数据安全处理模型确定目标噪声抽取范围,基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;
数据安全处理模块,用于将所述有效用户数据输入到所述数据安全处理模型中,获取所述有效用户数据对应的目标组合特征的输出值;
安全用户数据获取模块,用于当所述目标组合特征的输出值在预设监测范围内,则将所述有效用户数据作为安全用户数据。
一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,所述处理器执行所述计算机可读指令时实现如下步骤:
获取数据处理请求,所述数据处理请求包括目标组合特征;
根据所述目标组合特征,从用户数据库中获取待测用户数据;
根据数据安全处理模型确定目标噪声抽取范围,基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;
将所述有效用户数据输入到所述数据安全处理模型中,获取所述有效用户数据对应的目标组合特征的输出值;
当所述目标组合特征的输出值在预设监测范围内,则将所述有效用户数据作为安全用户数据。
一个或多个存储有计算机可读指令的非易失性可读存储介质,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器实现如下步骤:
获取数据处理请求,所述数据处理请求包括目标组合特征;
根据所述目标组合特征,从用户数据库中获取待测用户数据;
根据数据安全处理模型确定目标噪声抽取范围,基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;
将所述有效用户数据输入到所述数据安全处理模型中,获取所述有效用户数据对应的目标组合特征的输出值;
当所述目标组合特征的输出值在预设监测范围内,则将所述有效用户数据作为安全用户数据。
本申请的一个或多个实施例的细节在下面的附图及描述中提出。本申请的其他特征和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请实施例的技术方案,下面将对本申请实施例的描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅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性的前提下,还可以根据这些附图获得其他的附图。
图1是本申请一实施例中数据安全处理方法的一应用场景图;
图2是本申请一实施例中数据安全处理方法的一流程图;
图3是图2中步骤S30的一具体流程图;
图4是图2中步骤S30之前的一流程图;
图5是图4中步骤S301的一具体流程图;
图6是本申请一实施例中数据安全处理装置的一示意图;
图7是本申请一实施例中计算机设备的一示意图。
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作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。
本申请提供的数据安全处理方法,可应用在如图1的应用环境中,其中,客户端通过网络与服务器进行通信。其中,客户端可以但不限于各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一实施例中,如图2所示,提供一种数据安全处理方法,以该方法应用在图1中的服务器为例进行说明,包括如下步骤:
S10:获取数据处理请求,数据处理请求包括目标组合特征。
本实施例中的数据处理请求指客户端发送给服务器用于处理用户数据的请求,该数据处理请求包括目标组合特征。该目标组合特征指用户根据需要在客户端填写的关于用户数据的特征,包括但不限于用户数据库中存储的用户的年龄、性别、是否有车、是否购买保险(包括但不限于车险、产险和寿险)等特征。可以理解地,该目标组合特征包括至少一个用户数据的特征。
S20:根据目标组合特征,从用户数据库中获取待测用户数据。
其中,用户数据库指用于存储用户数据的数据库。本实施例中的用户数据库中存储的用户数据包括但不限于用户的基本信息、购买行为和交易数据。其中,基本信息包括但不限于姓名、年龄(出生年月)、性别、籍贯、民族、教育背景和工作经历等。购买行为指用户在平台购买产品的行为,包括但不限于用户购买的理财产品和保险产品。交易数据指用户的投资、消费和转账的记录数据。
具体地,当用户在客户端设置好目标组合特征后,发送给服务器,服务器根据获取到的目标组合特征,从用户数据库中获取与目标组合特征对应的用户数据作为待测用户数据。其中,待测用户数据指需要进行测试的用户数据。
S30:根据数据安全处理模型确定目标噪声抽取范围,基于目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据。
为了保护用户数据的安全性,本实施例对待测用户数据添加非目标组合特征对应的用户数据,使得获取的有效用户数据不全为目标组合特征对应的用户数据,降低基本信息用户数据的准确性来保证用户数据的安全性。
其中,数据安全处理模型指预先训练好的用于确定目标噪声抽取范围的模型。目标噪 声抽取范围指用于抽取非目标组合特征对应的用户数据的范围。本实施例中的数据安全处理模型包括目标梯度提升决策树模型和目标逻辑回归模型。目标梯度提升决策树模型指预先训练好的满足用户要求的梯度提升决策树,梯度提升决策树(Gradient Boosting Decision Tree,GBDT)是一种迭代的决策树算法,该算法由多棵决策树组成,所有树的结论累加起来作为最终预测结果。目标逻辑回归模型指预先训练好的用于获取目标噪声抽取范围的逻辑回归模型。逻辑回归模型(Logistic Regression,LR)是一种减小预测范围,将预测值限定为[0,1]间的一种回归模型。
在获取待测用户数据后,根据安全处理模型确定目标噪声抽取范围,从用户数据库中抽取特定数量的非目标组合特征对应的用户数据,加入待测用户数据中,生成有效用户数据。有效用户数据指在待测用户数据中添加了非目标组合特征对应的用户数据后得到的数据。需要说明的是,本实施例中的特定数量指在目标噪声抽取范围内的数量。服务器根据目标噪声抽取范围抽取非目标组合特征对应的用户数据加入到待测用户数据中,使得后续步骤获取的安全用户数据既能保证用户数据的安全性,又满足对用户数据的价值的需求。
S40:将有效用户数据输入到数据安全处理模型中,获取有效用户数据对应的目标组合特征的输出值。
具体地,在获取有效用户数据后,还需要对有效用户数据进行进一步地验证处理,确定有效用户数据是否满足用户要求。因此,还需将获取的有效用户数据输入到预先训练好的数据安全处理模型中进行处理,获取有效用户数据在经过数据安全处理模型处理得到的输出值。该输出值是一个概率值,用于检测有效用户数据是否满足用户设定的要求。
S50:当目标组合特征的输出值在预设监测范围内,则将有效用户数据作为安全用户数据。
具体地,预设监测范围指预先设置的用于检测有效用户数据是否满足用户要求的范围。本实施中设置的预设监测范围为70%-90%,既能保证用户数据的安全性,又满足对用户数据的价值需求。当通过数据安全处理模型的计算,获取的目标组合特征的输出值在预设监测范围内,则表示有效用户数据满足要求,将有效用户数据作为安全用户数据。
步骤S10-步骤S50,通过目标组合特征从用户数据库中获取待测用户数据,然后根据数据安全处理模型确定的目标噪声抽取范围,抽取特定数量的非目标组合特征对应的用户数据加入到待测用户数据中,以获取有效用户数据,降低有效用户数据的准确性来保证用户数据的安全性。在获取有效用户数据后,为了进一步确定有效用户数据是否满足要求,还需要将有效用户数据输入到数据安全处理模型中,获取有效用户数据对应的目标组合特征的输出值,当输出值在预设监测范围,则表示有效用户数据满足要求,将有效用户数据作为安全用户数据发送给需要的发送数据处理请求的客户端,既保证数据的使用者不能轻易推断出用户的个人隐私信息,又能最大化程度地发挥目标数据的价值。
在一实施例中,如图3所示,步骤S30中,基于目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据,具体包括如下步骤:
S31:基于目标噪声抽取范围,从用户数据库选取非目标组合特征对应的用户数据作为目标噪声数据。
具体地,在数据安全处理模型确定目标噪声抽取范围后,服务器根据目标噪声抽取范围,从用户数据库选取特定数量的非目标组合特征对应的用户数据作为目标噪声数据,为后续步骤提供数据来源。
如根据数据安全处理模型确定目标噪声抽取范围为0.05-0.25,非目标组合特征对应的用户数据有10万个,服务器根据目标噪声抽取范围5000-25000,从用户数据库选取5000-25000中任意一个数字的非目标组合特征对应的用户数据作为目标噪声数据。
S32:将目标噪声数据加入到待测用户数据中,获取有效用户数据。
具体地,在获取目标噪声数据后,将目标噪声数据加入到待测用户数据中,获取有效用户数据,使得有效用户数据不全为目标组合特征对应的用户数据,降低有效用户数据的准确性,使得有效用户数据既可以满足用户对用户数据的需求,又能保证用户数据的安全性。
步骤S31-步骤S32,通过目标噪声抽取范围,从用户数据库选取非目标组合特征对应的用户数据作为目标噪声数据,并将目标噪声数据加入到待测用户数据中,获取有效用户数据来保证用户数据的安全性。
在一实施例中,如图4所示,在步骤S30之前,即根据数据安全处理模型确定目标噪声抽取范围的步骤之前,数据安全处理方法还包括如下步骤:
S301:获取待训练数据,将待训练数据划分为训练集和测试集。
具体地,从样本数据库中获取待训练数据,并将待训练数据划分为训练集和测试集用于训练模型和测试模型。本实施例中需要训练的模型指原始梯度提升决策树模型。其中,样本数据库指用于存储待训练数据的数据库。待训练数据指用于训练原始梯度提升决策树模型的数据。训练集(training set)是用于训练原始梯度提升决策树模型的待训练数据的集合。测试集(test set)是用于测试训练好的原始梯度提升决策树模型是否准确的待训练数据的集合。
S302:初始化原始梯度提升决策树模型的模型参数,模型参数包括梯度提升决策树的最大深度和最大迭代次数。
具体地,在对原始梯度提升决策树模型进行训练之前,首先需要对原始梯度提升决策树模型中的模型参数进行初始化设置。本实施例中的模型参数包括原始梯度提升决策树的最大深度和最大迭代次数。根据实验数据表示,原始梯度提升决策树的最大深度设置为3,最大迭代次数设置为50,效果最好,本实施例中初始化设置其最大深度为3,最大迭代次数为50。对原始梯度提升决策树模型的模型参数进行初始化设置可以在后续训练原始梯度提升决策树模型时,缩短训练时间,提高识别准确率。
S303:将训练集对应的待训练数据输入到原始梯度提升决策树模型中,当原始梯度提升决策树模型中的训练深度达到最大深度且迭代次数达到最大迭代次数,则停止训练原始 梯度提升决策树模型,获取原始梯度提升决策树模型中每个决策树路径对应的原始组合特征。
具体地,在对原始梯度提升决策树模型进行初始化设置后,将训练集中的待训练数据输入到原始梯度提升决策树模型中,原始梯度提升决策树模型会在待训练数据对应的目标组合特征中选择一个特征作为第一个分叉点,获取训练集中待训练数据在该分叉点的残差,然后将目标组合特征中剩余的特征进行再次分叉,将第一个分叉点对应的残差作为第二个决策树输入,不断迭代,当原始梯度提升决策树模型中的训练深度达到最大深度且迭代次数达到最大迭代次数,则停止训练原始梯度提升决策树模型,获取原始梯度提升决策树模型中每个决策树路径对应的原始组合特征。其中,原始组合特征指训练集对应的待训练数据输入到原始梯度提升决策树模型中,每个决策树路径对应的组合特征。
S304:将原始组合特征输入到原始逻辑回归模型中,获取原始组合特征对应的输出值。
具体地,在获取原始组合特征之后,将原始组合特征输入到原始逻辑回归模型中,通过Sigmoid函数计算原始组合特征对应的概率,即原始组合特征对应的输出值。本实施例中的Sigmoid函数可表示为:
Figure PCTCN2018122734-appb-000001
其中,g (x)为原始组合特征发生的概率,g (x)∈(0,1),x为原始组合特征,T为用户根据实际设定的θ变化快慢的参数,θ为原始梯度提升决策树模型中每个用户数据的特征对应的权重的组合。q=1表示原始组合特征为目标组合特征成立,q=0表示原始组合特征为目标组合特征不成立,即原始组合特征为非目标组合特征。
S305:当原始组合特征对应的输出值在预设监测范围内,则将原始梯度提升决策树模型和原始逻辑回归模型作为目标梯度提升决策树模型和目标逻辑回归模型。
具体地,当原始组合特征对应的输出值在预设监测范围内,则表示原始梯度提升决策树模型和原始逻辑回归模型训练成功,将原始梯度提升决策树模型和原始逻辑回归模型作为目标梯度提升决策树模型和目标逻辑回归模型。
S306:采用测试集对应的待训练数据对目标梯度提升决策树模型和目标逻辑回归模型进行测试,若获取到的每一待训练数据对应的输出值均在预设监测范围内,则将目标梯度提升决策树模型和目标逻辑回归模型作为数据安全处理模型。
在确定原始梯度提升决策树模型和原始逻辑回归模型为目标梯度提升决策树模型和目标逻辑回归模型之后,为了防止过拟合,需要采用测试集中的待训练数据对目标梯度提升决策树模型和目标逻辑回归模型进行测试,验证目标梯度提升决策树模型和目标逻辑回归模型的准确性,若获取到的测试集中每一待训练数据对应的输出值均在预设监测范围 内,则标识将目标梯度提升决策树模型和目标逻辑回归模型训练的比较成功,满足要求,则将目标梯度提升决策树模型和目标逻辑回归模型作为数据安全处理模型。
步骤S301-步骤S306,通过训练集中的待训练数据训练原始梯度提升决策树模型,并将原始梯度提升决策树模型中每个决策树路径对应的原始组合特征输入到原始逻辑回归模型中,获取原始组合特征对应的输出值。若原始组合特征对应的输出值在预设监测范围内,则表示原始梯度提升决策树模型和原始逻辑回归模型训练成功。为了防止出现过拟合的情况,还需要将测试集对应的待训练数据输入到训练好的目标梯度提升决策树模型和目标逻辑回归模型中进行测试,若每一待训练数据对应的输出值均在预设监测范围内,则表示训练后的目标梯度提升决策树模型和目标逻辑回归模型符合需求,可以作为最终确定目标噪声抽取范围的数据安全处理模型,为后续步骤获取有效用户数据使用。
在一实施例中,如图5所示,步骤S301,获取待训练数据,将待训练数据划分为训练集和测试集,具体包括如下步骤:
S3011:获取模型训练请求,模型训练请求包括训练组合特征。
其中,模型训练请求指客户端发送的用于进行模型训练的请求。训练组合特征指用户在客户端根据需要设置的用于训练模型的用户数据的特征。具体地,用户按照要求在客户端设置训练组合特征,然后点击发送的操作,服务器会获取到携带有训练组合特征的模型训练请求。
S3012:根据训练组合特征,从用户数据库中选取与训练组合特征匹配的训练用户数据和与训练组合特征不匹配的非训练用户数据。
具体地,在获取到模型训练请求后,服务器根据模型训练请求中包括的训练组合特征,从用户数据库中选取与训练组合特征匹配的训练用户数据和与训练组合特征不匹配的非训练用户数据。其中,训练用户数据指与训练组合特征匹配的用户数据。非训练用户数据指与训练组合特征不匹配的用户数据。如训练合特征为年龄20-30岁和购买产险,则根据训练特征获取年龄20-30岁和购买产险的用户数据作为训练用户数据,选取不同时包括年龄在20-30岁和购买产险的用户数据作为非训练用户数据。根据训练组合特征获取训练用户数据和非训练用户数据为后续获取原始正样本和原始负样本提供数据来源。
进一步地,为方便后续步骤对用户数据进行处理,在获取训练用户数据后,将训练组合特征作为训练用户数据的用户标签,非目标组合特征作为非训练用户数据的用户标签。如将训练合特征为年龄20-30岁和购买产险作为训练用户数据的用户标签,将年龄不在20-30岁,没有购买产险作为非训练用户数据的标签。
S3013:按照预设的正样本数量,从训练用户数据中选取对应的训练用户数据作为原始正样本。
本实施例中正样本数量指预先设置的用于获取原始正样本的数量。原始正样本指从训练用户数据中根据预设的正样本数量抽取的训练用户数据。按照预设的正样本数量,从训练用户数据中获取原始正样本,为后续步骤获取有效正样本和有效负样本提供数据来源。
S3014:按照正负样本比例,从非训练用户数据中选取对应的非训练用户数据作为原始负样本。
具体地,在获取原始正样本后,服务器根据预先设置的正负样本比例,从非训练用户数据中选取对应的非训练用户数据作为原始负样本。其中,正负样本比例指预先设置的用于根据正样本数量确定原始负样本的数量的比列。本实施例中的原始负样本的数量是根据正样本数量决定的,使得原始负样本的数量和正样本数量成一定比例,满足后续进行模型训练的要求。
S3015:根据第一噪声抽取范围从原始正样本中抽取负噪声数据,并将负噪声数据加入到原始负样本中,生成有效负样本。
其中,第一抽取范围指预先设置好的将原始正样本作为负噪声数据的比例范围。负噪声数据指将原始正样本中部分用户数据对应的用户标签进行人为进行修改,使得该部分用户数据携带的用户标签变为非目标用户特征对应的用户标签。如原始正样本中的用户数据携带的用户标签为年龄20-30岁和购买产险,人为将部分用户标签修改为非目标组合特征对应的用户标签(如将用户标签修改为35-45岁和购买产险,或者将用户标签修改为20-30岁和购买车险,或者修改为30-40岁和有车等任意非目标组合特征对应的用户标签)。
具体地,在获取原始正样本和原始负样本之后,服务器根据第一噪声抽取范围,从原始正样本中随机抽取特定数量的用户数据,人为将该部分用户数据对应的用户标签修改为非目标组合特征对应的用户标签,将该部分修改了用户标签的用户数据作为负噪声数据。可以理解地,该特定数量指第一噪声抽取范围内的任意数字。在获取负噪声数据后,将负噪声数据加入到原始负样本中,生成有效负样本,降低原始负样本的准确性,使得原始负样本中的用户数据不全为非目标组合特征对应的用户数据,达到降低原始负样本中用户数据的敏感性的目的。
例如,原始正样本的数量为10万个,第一抽取范围设置为0.1-0.3,根据第一抽取范围,从10万个原始正样本中随机抽取1万-3万的用户数据,将该部分用户数据携带的用户标签修改为非目标组合特征对应的用户标签,将该部分用户数据作为负噪声数据。
S3016:根据第二噪声抽取范围从原始负样本中抽取正噪声数据,并将正噪声数据加入到原始正样本中,生成有效正样本。
其中,第二抽取范围指预先设置好的将原始负样本作为正噪声数据的比例范围。正噪声数据指将原始负样本中部分用户数据对应的用户标签变为目标组合特征对应的用户标签。需要说明的是,由于正样本的数量与负样本的数量是按照预先设置好的正负样本比例选取的,因此,第一抽取范围和第二抽取范围也是不一致的,如将第一抽取范围设置为0.1-0.3,则第二抽取范围设置为0.05-0.2。
具体地,根据第二抽取范围,从原始负样本中随机抽取特定数量的用户数据,人为修改该部分用户数据对应的用户标签,将用户标签修改为目标组合特征对应的用户标签,将该部分修改了用户标签的原始负样本作为正噪声数据。可以理解地,该特定数量指第二噪 声抽取范围内的任意数字。获取正噪声数据后,将正噪声数据添加到原始正样本中,形成有效正样本。可以理解地,有效正样本指在原始正样本中加入负噪声数据后形成的正样本。
例如,原始负样本的数量为100万个,第二抽取范围设置为0.05-0.2,根据第二抽取范围,从100万个原始负样本中随机抽取5万-20万范围内任何一个数字对应的用户数据,然后将抽取的用户数据携带的用户标签人为修改为目标组合特征对应的用户标签。修改用户标签后,将修改用户标签的原始负样本作为正噪声数据,加入到原始正样本中,生成有效正样本。
在原始正样本中加入正噪声数据,使得原始正样本不只是仅包括真正的目标组合特征对应的用户数据,还包括一些非目标组合特征的正噪声数据,使得原始正样本中的用户数据不是100%真实的训练用户数据,可以避免输出真实的用户数据,也不用对用户数据中的一些敏感数据进行逐一脱敏处理,保证了用户数据的安全性。
S3017:将有效正样本和有效负样本作为待训练数据,并存储在样本数据库中。
具体地,在获取有效正样本和有效负样本后,为了方便后续训练原始梯度提升决策树模型和验证训练好的目标梯度提升决策树模型是否准确,将有效正样本和有效负样本作为待训练数据存储在样本数据库中。
步骤S3011-步骤S3017,通过对原始正样本加入正噪声数据获取有效正样本,对原始负样本加入负噪声数据获取有效负样本,强行改变原始正样本和原始负样本的准确性,使得生成的有效正样本和有效负样本中的用户数据不全为含有目标组合特征的用户数据或者非目标组合特征的用户数据,避免输出真实的用户数据,也不用对用户数据中的一些敏感数据进行逐一脱敏处理,保证了用户数据的安全性。
在一实施例中,在步骤S304,将原始组合特征输入到原始逻辑回归模型中,获取原始组合特征对应的输出值的步骤之后,该数据安全处理方法还包括:当原始组合特征对应的输出值低于预设监测范围,则减小第一噪声抽取范围和第二噪声抽取范围,并提高正负样本比例;当原始组合特征对应的输出值高于预设监测范围,则增大第一噪声抽取范围和第二噪声抽取范围,并降低正负样本比例。
具体地,当训练集对应的待训练数据经过原始梯度提升决策树模型和原始逻辑回归模型的训练与计算,得到的原始组合特征对应的输出值不在预设监测范围内,则表示第一噪声抽取范围、第二噪声抽取范围和正负样本比例设置的不准确,需要动态调整第一噪声抽取范围、第二噪声抽取范围和正负样本比例。
进一步地,当原始组合特征对应的输出值低于预设监测范围,则表示负样本数量过多,应该提高正负样本比例,减少负样本数量。此时,第一噪声抽取范围和第二噪声抽取范围设置的偏高,减小第一噪声抽取范围和第二噪声抽取范围的设置。当原始组合特征对应的输出值高于预设监测范围,则表示负样本数量过少,应该降低正负样本比例,增加负样本数量。此时,第一噪声抽取范围和第二噪声抽取范围设置的偏低,应该提高第一噪声抽取范围和第二噪声抽取范围的设置。
本申请提供的数据安全处理方法,通过目标组合特征从用户数据库中获取待测用户数据,然后根据数据安全处理模型确定的目标噪声抽取范围,抽取特定数量的非目标组合特征对应的用户数据加入到待测用户数据中,降低待测用户数据的准确性来保证用户数据的安全性。该目标噪声抽取范围是根据训练好的数据安全处理模型确定的,使得根据目标噪声抽取范围获取的有效用户数据更加准确,满足用户要求,既保证用户数据的安全又能使用户数据发挥数据的价值。获取有效用户数据后,为了进一步确定有效用户数据是否满足要求,还需要将有效用户数据输入到数据安全处理模型中,获取有效用户数据对应的目标组合特征的输出值,当输出值在预设监测范围,则表示有效用户数据满足要求,将有效用户数据作为安全用户数据发送给需要的用户。当原始组合特征对应的输出值低于预设监测范围,则减小第一噪声抽取范围和第二噪声抽取范围,并提高正负样本比例;当原始组合特征对应的输出值高于预设监测范围,则增大第一噪声抽取范围和第二噪声抽取范围,并降低正负样本比例,使得获取安全用户数据更加接近用户的要求。
应理解,上述实施例中各步骤的序号的大小并不意味着执行顺序的先后,各过程的执行顺序应以其功能和内在逻辑确定,而不应对本申请实施例的实施过程构成任何限定。
在一实施例中,提供一种数据安全处理装置,该数据安全处理装置与上述实施例中数据安全处理方法一一对应。如图6所示,该数据安全处理装置包括数据处理请求获取模块10、待测用户数据获取模块20、有效用户数据获取模块30、数据安全处理模块40和安全用户数据获取模块50。各功能模块详细说明如下:
数据处理请求获取模块10,用于获取数据处理请求,数据处理请求包括目标组合特征。
待测用户数据获取模块20,用于根据目标组合特征,从用户数据库中获取待测用户数据。
有效用户数据获取模块30,用于根据数据安全处理模型确定目标噪声抽取范围,基于目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据。
数据安全处理模块40,用于将有效用户数据输入到数据安全处理模型中,获取有效用户数据对应的目标组合特征的输出值。
安全用户数据获取模块50,用于当目标组合特征的输出值在预设监测范围内,则将有效用户数据作为安全用户数据。
进一步地,有效用户数据获取模块30包括目标噪声数据获取单元和有效用户数据获取单元。
目标噪声数据获取单元,用于基于目标噪声抽取范围,从用户数据库选取非目标组合特征对应的用户数据作为目标噪声数据。
有效用户数据获取单元,用于将目标噪声数据加入到待测用户数据中,获取有效用户数据。
进一步地,在有效用户数据获取模块之前,数据安全处理装置还包括:
待训练数据处理单元,用于获取待训练数据,将待训练数据划分为训练集和测试集。
模型参数初始化单元,用于初始化原始梯度提升决策树模型的模型参数,模型参数包括梯度提升决策树的最大深度和最大迭代次数。
模型训练单元,用于将训练集对应的待训练数据输入到原始梯度提升决策树模型中,当原始梯度提升决策树模型中的训练深度达到最大深度且迭代次数达到最大迭代次数,则停止训练原始梯度提升决策树模型,获取原始梯度提升决策树模型中每个决策树路径对应的原始组合特征。
输出值计算单元,用于将原始组合特征输入到原始逻辑回归模型中,获取原始组合特征对应的输出值。
模型确定单元,用于当原始组合特征对应的输出值在预设监测范围内,则将原始梯度提升决策树模型和原始逻辑回归模型作为目标梯度提升决策树模型和目标逻辑回归模型。
模型测试单元,用于采用测试集对应的待训练数据对目标梯度提升决策树模型和目标逻辑回归模型进行测试,若获取到的每一待训练数据对应的输出值均在预设监测范围内,则将目标梯度提升决策树模型和目标逻辑回归模型作为数据安全处理模型。
进一步地,待训练数据处理单元包括模型训练请求获取子单元、用户数据选取子单元、原始正样本获取子单元、原始负样本获取子单元、有效负样本获取子单元、有效正样本获取子单元和待训练数据生成子单元。
模型训练请求获取子单元,用于获取模型训练请求,模型训练请求包括训练组合特征。
用户数据选取子单元,用于根据训练组合特征,从用户数据库中选取与训练组合特征匹配的训练用户数据和与训练组合特征不匹配的非训练用户数据。
原始正样本获取子单元,用于按照预设的正样本数量,从训练用户数据中选取对应的训练用户数据作为原始正样本。
原始负样本获取子单元,用于按照正负样本比例,从非训练用户数据中选取对应的非训练用户数据作为原始负样本。
有效负样本获取子单元,用于根据第一噪声抽取范围从原始正样本中抽取负噪声数据,并将负噪声数据加入到原始负样本中,生成有效负样本。
有效正样本获取子单元,用于根据第二噪声抽取范围从原始负样本中抽取正噪声数据,并将正噪声数据加入到原始正样本中,生成有效正样本。
待训练数据生成子单元,用于将有效正样本和有效负样本作为待训练数据存储在样本数据库中。
进一步地,数据安全处理方法还用于当原始组合特征对应的输出值低于预设监测范围,则减小第一噪声抽取范围和第二噪声抽取范围,并提高正负样本比例;当原始组合特征对应的输出值高于预设监测范围,则增大第一噪声抽取范围和第二噪声抽取范围,并降低正负样本比例。
关于数据安全处理装置的具体限定可以参见上文中对于数据安全处理方法的限定,在 此不再赘述。上述数据安全处理装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中,也可以以软件形式存储于计算机设备中的存储器中,以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中,提供了一种计算机设备,该计算机设备可以是服务器,其内部结构图可以如图7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、网络接口和数据库。其中,该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储数据安全处理方法涉及的数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种数据安全处理方法。
在一个实施例中,提供了一种计算机设备,包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机可读指令,处理器执行计算机可读指令时实现上述数据安全处理方法的步骤,如图2所示的步骤S10-步骤S50。或者图3至图5所示的步骤。为避免重复,这里不再赘述。或者,处理器执行计算机可读指令时实现上述数据安全处理装置的步骤,如图6所示的数据处理请求获取模块10、待测用户数据获取模块20、有效用户数据获取模块30、数据安全处理模块40和安全用户数据获取模块50。为避免重复,这里不再赘述。
在一个实施例中,提供了一个或多个存储有计算机可读指令的非易失性可读存储介质,计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器实现上述实施例中数据安全处理方法的步骤,如图2所示的步骤S10-步骤S50。或者图3至图5所示的步骤。为避免重复,这里不再赘述。或者,计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器实现上述实施例中数据安全处理装置的步骤,如图6所示的数据处理请求获取模块10、待测用户数据获取模块20、有效用户数据获取模块30、数据安全处理模块40和安全用户数据获取模块50。为避免重复,这里不再赘述。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机可读指令来指令相关的硬件来完成,所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中,该计算机可读指令在执行时,可包括如上述各方法的实施例的流程。其中,本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM以多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM (SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
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,仅以上述各功能单元、模块的划分进行举例说明,实际应用中,可以根据需要而将上述功能分配由不同的功能单元、模块完成,即将所述装置的内部结构划分成不同的功能单元或模块,以完成以上描述的全部或者部分功能。
以上所述实施例仅用以说明本申请的技术方案,而非对其限制;尽管参照前述实施例对本申请进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围,均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种数据安全处理方法,其特征在于,包括:
    获取数据处理请求,所述数据处理请求包括目标组合特征;
    根据所述目标组合特征,从用户数据库中获取待测用户数据;
    根据数据安全处理模型确定目标噪声抽取范围,基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;
    将所述有效用户数据输入到所述数据安全处理模型中,获取所述有效用户数据对应的目标组合特征的输出值;
    当所述目标组合特征的输出值在预设监测范围内,则将所述有效用户数据作为安全用户数据。
  2. 如权利要求1所述的数据安全处理方法,其特征在于,所述基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据,包括:
    基于所述目标噪声抽取范围,从用户数据库选取非目标组合特征对应的用户数据作为目标噪声数据;
    将所述目标噪声数据加入到待测用户数据中,获取有效用户数据。
  3. 如权利要求1所述的数据安全处理方法,其特征在于,在所述根据数据安全处理模型确定目标噪声抽取范围的步骤之前,所述数据安全处理方法还包括:
    获取待训练数据,将所述待训练数据划分为训练集和测试集;
    初始化原始梯度提升决策树模型的模型参数,所述模型参数包括梯度提升决策树的最大深度和最大迭代次数;
    将训练集对应的待训练数据输入到所述原始梯度提升决策树模型中,当所述原始梯度提升决策树模型中的训练深度达到所述最大深度且迭代次数达到最大迭代次数,则停止训练所述原始梯度提升决策树模型,获取所述原始梯度提升决策树模型中每个决策树路径对应的原始组合特征;
    将所述原始组合特征输入到原始逻辑回归模型中,获取所述原始组合特征对应的输出值;
    当所述原始组合特征对应的输出值在所述预设监测范围内,则将所述原始梯度提升决策树模型和所述原始逻辑回归模型作为目标梯度提升决策树模型和目标逻辑回归模型;
    采用测试集对应的待训练数据对所述目标梯度提升决策树模型和所述目标逻辑回归模型进行测试,若获取到的每一所述待训练数据对应的输出值均在所述预设监测范围内,则将所述目标梯度提升决策树模型和目标逻辑回归模型作为数据安全处理模型。
  4. 如权利要求3所述的数据安全处理方法,其特征在于,所述获取待训练数据,将所述待训练数据划分为训练集和测试集,包括:
    获取模型训练请求,所述模型训练请求包括训练组合特征;
    根据所述训练组合特征,从用户数据库中选取与所述训练组合特征匹配的训练用户数据和与所述训练组合特征不匹配的非训练用户数据;
    按照预设的正样本数量,从所述训练用户数据中选取对应的训练用户数据作为原始正样本;
    按照正负样本比例,从所述非训练用户数据中选取对应的非训练用户数据作为原始负样本;
    根据第一噪声抽取范围从原始正样本中抽取负噪声数据,并将所述负噪声数据加入到原始负样本中,生成有效负样本;
    根据第二噪声抽取范围从原始负样本中抽取正噪声数据,并将所述正噪声数据加入到原始正样本中,生成有效正样本;
    将所述有效正样本和所述有效负样本作为待训练数据存储在样本数据库中。
  5. 如权利要求4所述的数据安全处理方法,其特征在于,在所述将所述原始组合特征输入到原始逻辑回归模型中,获取所述原始组合特征对应的输出值的步骤之后,所述数据安全处理方法还包括:
    当所述原始组合特征对应的输出值低于所述预设监测范围,则减小所述第一噪声抽取范围和所述第二噪声抽取范围,并提高所述正负样本比例;当所述原始组合特征对应的输出值高于所述预设监测范围,则增大所述第一噪声抽取范围和所述第二噪声抽取范围,并降低所述正负样本比例。
  6. 一种数据安全处理装置,其特征在于,包括:
    数据处理请求获取模块,用于获取数据处理请求,所述数据处理请求包括目标组合特征;
    待测用户数据获取模块,用于根据所述目标组合特征,从用户数据库中获取待测用户数据;
    有效用户数据获取模块,用于根据数据安全处理模型确定目标噪声抽取范围,基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;
    数据安全处理模块,用于将所述有效用户数据输入到所述数据安全处理模型中,获取所述有效用户数据对应的目标组合特征的输出值;
    安全用户数据获取模块,用于当所述目标组合特征的输出值在预设监测范围内,则将所述有效用户数据作为安全用户数据。
  7. 如权利要求6所述的数据安全处理装置,其特征在于,在有效用户数据获取模块之前,所述数据安全处理装置还包括:
    待训练数据处理单元,用于获取待训练数据,将所述待训练数据划分为训练集和测试集;
    模型参数初始化单元,用于初始化原始梯度提升决策树模型的模型参数,所述模型参数包括梯度提升决策树的最大深度和最大迭代次数;
    模型训练单元,用于将训练集对应的待训练数据输入到所述原始梯度提升决策树模型中,当所述原始梯度提升决策树模型中的训练深度达到所述最大深度且迭代次数达到最大迭代次数,则停止训练所述原始梯度提升决策树模型,获取所述所述原始梯度提升决策树模型中每个决策树路径对应的原始组合特征;
    输出值计算单元,用于将所述原始组合特征输入到原始逻辑回归模型中,获取所述原始组合特征对应的输出值;
    模型确定单元,用于当所述原始组合特征对应的输出值在所述预设监测范围内,则将所述原始梯度提升决策树模型和所述原始逻辑回归模型作为目标梯度提升决策树模型和目标逻辑回归模型;
    模型测试单元,用于采用测试集对应的待训练数据对所述目标梯度提升决策树模型和所述目标逻辑回归模型进行测试,若获取到的每一所述待训练数据对应的输出值均在所述预设监测范围内,则将所述目标梯度提升决策树模型和目标逻辑回归模型作为数据安全处理模型。
  8. 如权利要求7所述的数据安全处理装置,其特征在于,所述待训练数据处理单元,包括:
    模型训练请求获取子单元,用于获取模型训练请求,所述模型训练请求包括训练组合特征;
    用户数据选取子单元,用于根据所述训练组合特征,从用户数据库中选取与所述训练组合特征匹配的安全用户数据和与所述训练组合特征不匹配的非安全用户数据;
    原始正样本获取子单元,用于按照预设的正样本数量,从所述安全用户数据中选取对应的安全用户数据作为原始正样本;
    原始负样本获取子单元,用于按照正负样本比例,从所述非安全用户数据中选取对应的非安全用户数据作为原始负样本;
    有效负样本获取子单元,用于根据第一噪声抽取范围从原始正样本中抽取负噪声数据,并将所述负噪声数据加入到原始负样本中,生成有效负样本;
    有效正样本获取子单元,用于根据第二噪声抽取范围从原始负样本中抽取正噪声数据,并将所述正噪声数据加入到原始正样本中,生成有效正样本;
    待训练数据生成子单元,用于将所述有效正样本和所述有效负样本作为待训练数据存储在样本数据库中。
  9. 如权利要求6所述的数据安全处理装置,其特征在于,所述有效用户数据获取模块,包括:
    目标噪声数据获取单元,用于基于所述目标噪声抽取范围,从用户数据库选取非目标组合特征对应的用户数据作为目标噪声数据;
    有效用户数据获取单元,用于将所述目标噪声数据加入到待测用户数据中,获取有效用户数据。
  10. 如权利要求8所述的数据安全处理装置,其特征在于,所述数据安全处理装置还包括:
    当所述原始组合特征对应的输出值低于所述预设监测范围,则减小所述第一噪声抽取范围和所述第二噪声抽取范围,并提高所述正负样本比例;当所述原始组合特征对应的输出值高于所述预设监测范围,则增大所述第一噪声抽取范围和所述第二噪声抽取范围,并降低所述正负样本比例。
  11. 一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,其特征在于,所述处理器执行所述计算机可读指令时实现如下步骤:
    获取数据处理请求,所述数据处理请求包括目标组合特征;
    根据所述目标组合特征,从用户数据库中获取待测用户数据;
    根据数据安全处理模型确定目标噪声抽取范围,基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;
    将所述有效用户数据输入到所述数据安全处理模型中,获取所述有效用户数据对应的目标组合特征的输出值;
    当所述目标组合特征的输出值在预设监测范围内,则将所述有效用户数据作为安全用户数据。
  12. 如权利要求11所述的计算机设备,其特征在于,所述基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据,包括:
    基于所述目标噪声抽取范围,从用户数据库选取非目标组合特征对应的用户数据作为目标噪声数据;
    将所述目标噪声数据加入到待测用户数据中,获取有效用户数据。
  13. 如权利要求11所述的计算机设备,其特征在于,在所述根据数据安全处理模型确定目标噪声抽取范围的步骤之前,所述处理器执行所述计算机可读指令时还实现如下步骤:
    获取待训练数据,将所述待训练数据划分为训练集和测试集;
    初始化原始梯度提升决策树模型的模型参数,所述模型参数包括梯度提升决策树的最大深度和最大迭代次数;
    将训练集对应的待训练数据输入到所述原始梯度提升决策树模型中,当所述原始梯度提升决策树模型中的训练深度达到所述最大深度且迭代次数达到最大迭代次数,则停止训练所述原始梯度提升决策树模型,获取所述原始梯度提升决策树模型中每个决策树路径对应的原始组合特征;
    将所述原始组合特征输入到原始逻辑回归模型中,获取所述原始组合特征对应的输出值;
    当所述原始组合特征对应的输出值在所述预设监测范围内,则将所述原始梯度提升决 策树模型和所述原始逻辑回归模型作为目标梯度提升决策树模型和目标逻辑回归模型;
    采用测试集对应的待训练数据对所述目标梯度提升决策树模型和所述目标逻辑回归模型进行测试,若获取到的每一所述待训练数据对应的输出值均在所述预设监测范围内,则将所述目标梯度提升决策树模型和目标逻辑回归模型作为数据安全处理模型。
  14. 如权利要求13所述的计算机设备,其特征在于,所述获取待训练数据,将所述待训练数据划分为训练集和测试集,包括:
    获取模型训练请求,所述模型训练请求包括训练组合特征;
    根据所述训练组合特征,从用户数据库中选取与所述训练组合特征匹配的训练用户数据和与所述训练组合特征不匹配的非训练用户数据;
    按照预设的正样本数量,从所述训练用户数据中选取对应的训练用户数据作为原始正样本;
    按照正负样本比例,从所述非训练用户数据中选取对应的非训练用户数据作为原始负样本;
    根据第一噪声抽取范围从原始正样本中抽取负噪声数据,并将所述负噪声数据加入到原始负样本中,生成有效负样本;
    根据第二噪声抽取范围从原始负样本中抽取正噪声数据,并将所述正噪声数据加入到原始正样本中,生成有效正样本;
    将所述有效正样本和所述有效负样本作为待训练数据存储在样本数据库中。
  15. 如权利要求14所述的计算机设备,其特征在于,在所述将所述原始组合特征输入到原始逻辑回归模型中,获取所述原始组合特征对应的输出值的步骤之后,所述处理器执行所述计算机可读指令时还实现如下步骤:
    当所述原始组合特征对应的输出值低于所述预设监测范围,则减小所述第一噪声抽取范围和所述第二噪声抽取范围,并提高所述正负样本比例;当所述原始组合特征对应的输出值高于所述预设监测范围,则增大所述第一噪声抽取范围和所述第二噪声抽取范围,并降低所述正负样本比例。
  16. 一个或多个存储有计算机可读指令的非易失性可读存储介质,其特征在于,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器实现如下步骤:
    获取数据处理请求,所述数据处理请求包括目标组合特征;
    根据所述目标组合特征,从用户数据库中获取待测用户数据;
    根据数据安全处理模型确定目标噪声抽取范围,基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据;
    将所述有效用户数据输入到所述数据安全处理模型中,获取所述有效用户数据对应的目标组合特征的输出值;
    当所述目标组合特征的输出值在预设监测范围内,则将所述有效用户数据作为安全用户数据。
  17. 如权利要求16所述的非易失性可读存储介质,其特征在于,所述基于所述目标噪声抽取范围对待测用户数据进行处理,获取有效用户数据,包括:
    基于所述目标噪声抽取范围,从用户数据库选取非目标组合特征对应的用户数据作为目标噪声数据;
    将所述目标噪声数据加入到待测用户数据中,获取有效用户数据。
  18. 如权利要求16所述的非易失性可读存储介质,其特征在于,在所述根据数据安全处理模型确定目标噪声抽取范围的步骤之前,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器还执行如下步骤:
    获取待训练数据,将所述待训练数据划分为训练集和测试集;
    初始化原始梯度提升决策树模型的模型参数,所述模型参数包括梯度提升决策树的最大深度和最大迭代次数;
    将训练集对应的待训练数据输入到所述原始梯度提升决策树模型中,当所述原始梯度提升决策树模型中的训练深度达到所述最大深度且迭代次数达到最大迭代次数,则停止训练所述原始梯度提升决策树模型,获取所述原始梯度提升决策树模型中每个决策树路径对应的原始组合特征;
    将所述原始组合特征输入到原始逻辑回归模型中,获取所述原始组合特征对应的输出值;
    当所述原始组合特征对应的输出值在所述预设监测范围内,则将所述原始梯度提升决策树模型和所述原始逻辑回归模型作为目标梯度提升决策树模型和目标逻辑回归模型;
    采用测试集对应的待训练数据对所述目标梯度提升决策树模型和所述目标逻辑回归模型进行测试,若获取到的每一所述待训练数据对应的输出值均在所述预设监测范围内,则将所述目标梯度提升决策树模型和目标逻辑回归模型作为数据安全处理模型。
  19. 如权利要求18所述的非易失性可读存储介质,其特征在于,所述获取待训练数据,将所述待训练数据划分为训练集和测试集,包括:
    获取模型训练请求,所述模型训练请求包括训练组合特征;
    根据所述训练组合特征,从用户数据库中选取与所述训练组合特征匹配的训练用户数据和与所述训练组合特征不匹配的非训练用户数据;
    按照预设的正样本数量,从所述训练用户数据中选取对应的训练用户数据作为原始正样本;
    按照正负样本比例,从所述非训练用户数据中选取对应的非训练用户数据作为原始负样本;
    根据第一噪声抽取范围从原始正样本中抽取负噪声数据,并将所述负噪声数据加入到原始负样本中,生成有效负样本;
    根据第二噪声抽取范围从原始负样本中抽取正噪声数据,并将所述正噪声数据加入到原始正样本中,生成有效正样本;
    将所述有效正样本和所述有效负样本作为待训练数据存储在样本数据库中。
  20. 如权利要求19所述的非易失性可读存储介质,其特征在于,在所述将所述原始组合特征输入到原始逻辑回归模型中,获取所述原始组合特征对应的输出值的步骤之后,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器还执行如下步骤:
    当所述原始组合特征对应的输出值低于所述预设监测范围,则减小所述第一噪声抽取范围和所述第二噪声抽取范围,并提高所述正负样本比例;当所述原始组合特征对应的输出值高于所述预设监测范围,则增大所述第一噪声抽取范围和所述第二噪声抽取范围,并降低所述正负样本比例。
PCT/CN2018/122734 2018-10-12 2018-12-21 数据安全处理方法、装置、计算机设备及存储介质 Ceased WO2020073492A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811187262.8 2018-10-12
CN201811187262.8A CN109543442B (zh) 2018-10-12 2018-10-12 数据安全处理方法、装置、计算机设备及存储介质

Publications (1)

Publication Number Publication Date
WO2020073492A1 true WO2020073492A1 (zh) 2020-04-16

Family

ID=65843885

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/122734 Ceased WO2020073492A1 (zh) 2018-10-12 2018-12-21 数据安全处理方法、装置、计算机设备及存储介质

Country Status (2)

Country Link
CN (1) CN109543442B (zh)
WO (1) WO2020073492A1 (zh)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111753914A (zh) * 2020-06-29 2020-10-09 北京百度网讯科技有限公司 模型优化方法和装置、电子设备及存储介质
CN112101570A (zh) * 2020-09-22 2020-12-18 北京百度网讯科技有限公司 一种模型训练和使用方法、装置、设备及存储介质
CN112632607A (zh) * 2020-12-22 2021-04-09 中国建设银行股份有限公司 一种数据处理方法、装置及设备
CN113780365A (zh) * 2021-08-19 2021-12-10 支付宝(杭州)信息技术有限公司 样本生成方法和装置
CN114418095A (zh) * 2022-01-21 2022-04-29 腾讯科技(深圳)有限公司 联邦神经网络的训练方法、装置及电子设备
CN116167460A (zh) * 2023-03-13 2023-05-26 抖音视界有限公司 用于拆分学习的方法、装置、电子设备和介质
CN117216684A (zh) * 2023-07-27 2023-12-12 北京奇艺世纪科技有限公司 数据资源的分配方法、装置、电子设备及存储介质
CN112347476B (zh) * 2020-11-13 2024-02-02 脸萌有限公司 数据保护方法、装置、介质及设备

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112070310A (zh) * 2020-09-10 2020-12-11 腾讯科技(深圳)有限公司 基于人工智能的流失用户预测方法、装置及电子设备
CN116842371A (zh) * 2022-03-21 2023-10-03 北京转转精神科技有限责任公司 用于构建训练样本的方法、装置、电子设备及存储介质
CN115292597B (zh) * 2022-08-11 2025-09-23 成都安恒信息技术有限公司 一种反用户画像的方法及系统

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102549614A (zh) * 2009-10-07 2012-07-04 微软公司 用于维护用户简档的隐私的隐私库
CN106339714A (zh) * 2016-08-10 2017-01-18 上海交通大学 多层嵌入差分隐私到决策树模型的隐私风险控制方法
CN108537055A (zh) * 2018-03-06 2018-09-14 南京邮电大学 一种数据查询隐私保护的隐私预算分配和数据发布方法及其系统

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8627483B2 (en) * 2008-12-18 2014-01-07 Accenture Global Services Limited Data anonymization based on guessing anonymity
US20150095017A1 (en) * 2013-09-27 2015-04-02 Google Inc. System and method for learning word embeddings using neural language models
CN104133903B (zh) * 2014-08-04 2017-05-17 华中科技大学 一种基于隐私保护的数据索引方法
CN106656981B (zh) * 2016-10-21 2020-04-28 东软集团股份有限公司 网络入侵检测方法和装置
CN107491557A (zh) * 2017-09-06 2017-12-19 徐州医科大学 一种基于差分隐私的TopN协同过滤推荐方法
CN107895277A (zh) * 2017-09-30 2018-04-10 平安科技(深圳)有限公司 在应用程序中推送贷款广告的方法、电子装置及介质
CN107782442A (zh) * 2017-10-24 2018-03-09 华北电力大学(保定) 基于大数据和随机森林的变压器多特征参量选择方法
CN108389125B (zh) * 2018-02-27 2022-06-07 挖财网络技术有限公司 信贷申请的逾期风险预测方法及装置
CN108520181B (zh) * 2018-03-26 2022-04-22 联想(北京)有限公司 数据模型训练方法和装置

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102549614A (zh) * 2009-10-07 2012-07-04 微软公司 用于维护用户简档的隐私的隐私库
CN106339714A (zh) * 2016-08-10 2017-01-18 上海交通大学 多层嵌入差分隐私到决策树模型的隐私风险控制方法
CN108537055A (zh) * 2018-03-06 2018-09-14 南京邮电大学 一种数据查询隐私保护的隐私预算分配和数据发布方法及其系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LI, RUIQI: "Non-official translation: Churn Prediction Models for Anonymous Telecom Customer Dataset", ECONOMICS AND MANAGEMENT, CHINA MASTER'S THESES FULL-TEXT DATABASE, no. 1, 15 January 2018 (2018-01-15), ISSN: 1674-0246 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111753914A (zh) * 2020-06-29 2020-10-09 北京百度网讯科技有限公司 模型优化方法和装置、电子设备及存储介质
CN111753914B (zh) * 2020-06-29 2024-04-16 北京百度网讯科技有限公司 模型优化方法和装置、电子设备及存储介质
CN112101570A (zh) * 2020-09-22 2020-12-18 北京百度网讯科技有限公司 一种模型训练和使用方法、装置、设备及存储介质
CN112347476B (zh) * 2020-11-13 2024-02-02 脸萌有限公司 数据保护方法、装置、介质及设备
CN112632607A (zh) * 2020-12-22 2021-04-09 中国建设银行股份有限公司 一种数据处理方法、装置及设备
CN112632607B (zh) * 2020-12-22 2024-04-26 中国建设银行股份有限公司 一种数据处理方法、装置及设备
CN113780365A (zh) * 2021-08-19 2021-12-10 支付宝(杭州)信息技术有限公司 样本生成方法和装置
CN114418095A (zh) * 2022-01-21 2022-04-29 腾讯科技(深圳)有限公司 联邦神经网络的训练方法、装置及电子设备
CN114418095B (zh) * 2022-01-21 2025-08-19 腾讯科技(深圳)有限公司 联邦神经网络的训练方法、装置及电子设备
CN116167460A (zh) * 2023-03-13 2023-05-26 抖音视界有限公司 用于拆分学习的方法、装置、电子设备和介质
CN117216684A (zh) * 2023-07-27 2023-12-12 北京奇艺世纪科技有限公司 数据资源的分配方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
CN109543442B (zh) 2024-11-26
CN109543442A (zh) 2019-03-29

Similar Documents

Publication Publication Date Title
WO2020073492A1 (zh) 数据安全处理方法、装置、计算机设备及存储介质
US20200097329A1 (en) Method and device for virtual resource allocation, modeling, and data prediction
KR102351085B1 (ko) 처리될 트랜잭션에 관련된 위험 식별 방법, 장치 및 전자 디바이스
RU2713574C1 (ru) Системы и устройства для оценки архитектуры и реализации стратегий в области обеспечения безопасности
CN110222992A (zh) 一种基于受骗群体画像的网络诈骗预警方法及装置
US11232182B2 (en) Open data biometric identity validation
CN111542833A (zh) 用于文档认证的分类筛选引擎
CN110060068A (zh) 商户评估方法、装置、电子设备及可读存储介质
US8838430B1 (en) Detection of memory access violation in simulations
US20250200193A1 (en) Detecting synthetic user accounts using synthetic patterns learned via machine learning
CN110163655A (zh) 基于梯度提升树的坐席分配方法、装置、设备及存储介质
CN113886260A (zh) 自动化测试方法、系统、计算机设备及存储介质
CN112634017A (zh) 远程开卡激活方法、装置、电子设备及计算机存储介质
CN116720186A (zh) 一种恶意代码识别方法、装置、电子设备及存储介质
CN117807929A (zh) 一种基于深度学习的模型仿真预测方法及系统
WO2021043023A1 (zh) 图像处理方法及装置、分类器训练方法以及可读存储介质
EP4621739A1 (en) Model training method and related device
CN113283388B (zh) 活体人脸检测模型的训练方法、装置、设备及存储介质
CN111046380B (zh) 一种基于对抗样本增强模型抗攻击能力的方法和系统
CN114743592B (zh) 一种药物临床试验结果预测方法、装置、设备及介质
US12387459B2 (en) System and method for image de-identification to humans while remaining recognizable by machines
CN114372862A (zh) 数据处理方法、装置、计算机设备和存储介质
WO2025102839A1 (zh) 模型训练方法、生物特征图像生成方法及装置
US12591774B2 (en) Ordering infrastructure using application terms
CN113420841B (zh) 有毒样本数据生成方法及装置

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: 18936811

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 06.08.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18936811

Country of ref document: EP

Kind code of ref document: A1