Intelligent data desensitization management method for sensitive data

Abstract

The invention discloses an intelligent data desensitization management method for sensitive data. The method comprises the steps of obtaining target interaction data in a financial loan platform; inputting the target interaction data into a data identification model to obtain sensitive data output by the data identification model; the data identification model is obtained by training based on label results of the sample interaction data and the corresponding sensitive data; determining a desensitization mode of the sensitive data based on the data type of the sensitive data; and desensitizing the sensitive data based on the desensitization mode to obtain target data. According to the method, the sensitive data in the target interaction data are identified, and then based on the desensitization mode of the sensitive data, different types of sensitive data can be subjected to independent desensitization processing through different desensitization modes, so that the security degree of desensitization of the sensitive data in the target interaction data is improved, and the situation that the sensitive data in the target interaction data cannot be desensitized after a single type of data is cracked is avoided. And other types of data are easy to crack, so that the data security risk of the financial loan platform is reduced.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data intelligent desensitization management method for sensitive data. Background Art

[0002] With the rapid development of Internet finance, more and more financial services are provided through online platforms, and users conduct various financial activities on these platforms, such as applying for loans and managing accounts. In this process, users will generate a large amount of interactive data, which may contain sensitive information, such as personal identity information, financial account information, financial information, and credit information. Protecting the security of this sensitive information is one of the important tasks of the financial industry and a key area of ​​concern for regulators. In recent years, data leakage incidents have occurred frequently, posing a serious threat to the privacy and security of users. Therefore, desensitization of sensitive data is an issue that needs to be addressed urgently.

[0003] At present, financial loan platforms usually use data recognition models to automatically identify sensitive data. After identification, the sensitive data is desensitized using desensitization technology. Although this has improved the security of financial data to a certain extent, financial loan platforms currently generally use the same or similar desensitization algorithms to desensitize the data during the processing of the identified sensitive data, resulting in low security of the desensitized data. In particular, when one type of data is cracked, other data with similar processing algorithms are also easily cracked, which brings security risks to the data desensitization of financial loan platforms. Summary of the invention

[0004] The purpose of the present invention is to solve the problems in the prior art and to propose a data intelligent desensitization management method for sensitive data.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A data intelligent desensitization management method for sensitive data, comprising:

[0007] Obtain target interaction data in the financial loan platform;

[0008] Input the target interaction data into a data recognition model to obtain sensitive data output by the data recognition model; the sensitive data includes personal identity information, financial account information, financial information and credit information; the data recognition model is trained based on the label results of the sample interaction data and its corresponding sensitive data;

[0009] Determining a desensitization mode for the sensitive data based on the data type of the sensitive data;

[0010] The sensitive data is desensitized based on the desensitization mode to obtain target data.

[0011] According to the present invention, a data intelligent desensitization management method for sensitive data is provided, wherein the personal identity information includes identity card number information, name information and date of birth information; the sensitive data is desensitized based on the desensitization mode to obtain target data, including:

[0012] If the sensitive data is identity card number information, the identity card number information is split to obtain multiple segments of number information; at least one first target number information in the multiple segments of number information is converted into its corresponding English letter information; the English letter information and the second target number information are spliced ​​to obtain the target data; the second target number information is the number information in the multiple segments of number information except the first target number information;

[0013] If the sensitive data is the name information, convert at least one first target character in the name information into a target coded character by using a first preset formula; concatenate the target coded character and the second target character to obtain the target data; the second target character is a character in the name information other than the first target character;

[0014] If the sensitive data is date of birth information, the date of birth information is split to obtain multiple segments of time information; at least one first target time information in the multiple segments of time information is offset by a preset length to obtain offset time information; the offset time information and the second target time information are spliced ​​to obtain the target data; the second target time information is the time information in the multiple segments of time information except the first target time information.

[0015] According to the present invention, a data intelligent desensitization management method for sensitive data is provided, wherein the sensitive data is desensitized based on the desensitization mode to obtain target data, including:

[0016] If the sensitive data is financial account information, the financial account information is split into multiple sub-segment information;

[0017] For at least one first target sub-segment information among the plurality of sub-segment information, converting a first target number in the first target sub-segment information into a first digital code value; the first target number is at least one digit in the first target sub-segment information;

[0018] Converting the first digital code value using a second preset formula to obtain a converted second digital code value;

[0019] Converting the second digital code value into a corresponding target value;

[0020] The target data is obtained by splicing the target numerical value, the second target sub-segment, and the second target number; the second target sub-segment is the sub-segment information of the multiple sub-segment information except the first target sub-segment information; the second target number is the number in the first target sub-segment except the first target number.

[0021] According to the present invention, a data intelligent desensitization management method for sensitive data is provided, wherein the sensitive data is desensitized based on the desensitization mode to obtain target data, including:

[0022] If the sensitive data is the financial information, the financial information is processed by a third preset formula to obtain a target mapping value; the target mapping value is mapped within a preset interval;

[0023] Performing nonlinear transformation on the target mapping value through a fourth preset formula to obtain blurred financial information;

[0024] A random noise value is added to the blurred financial information to obtain the target data.

[0025] According to the present invention, a data intelligent desensitization management method for sensitive data is provided, wherein the credit information includes credit score information and overdue record text description information, and the sensitive data is desensitized based on the desensitization mode to obtain target data, including:

[0026] If the sensitive data is credit score information, a first score value and a second score value of the platform credit score are obtained; the first score value is less than the credit score information; the second score value is greater than the credit score information; the first score value and the second score value are grouped into intervals to obtain a plurality of score sub-intervals; based on the credit score information and the plurality of score sub-intervals, an interval number of the score sub-interval where the credit score information is located is obtained; the credit score information and the interval number are processed by a fifth preset formula to obtain the target data;

[0027] If the sensitive data is text description information of overdue records, the text description information of overdue records is subjected to sensitive word recognition to determine a plurality of sensitive keywords; the plurality of sensitive keywords are subjected to fuzzy processing to obtain the target data.

[0028] According to the present invention, a data intelligent desensitization management method for sensitive data is provided, and the training process of the data recognition model includes:

[0029] Acquire historical data of the financial loan platform; the historical data includes historical sensitive data and historical general data;

[0030] Classify the historical sensitive data to obtain first classification data, second classification data, third classification data and fourth classification data; the first classification data is sample data in the user application process data, the second classification data is sample data in the user repayment data, the third classification data is sample data in the communication data between the user and the platform customers, and the fourth classification data is sample data in the user operation behavior data;

[0031] Based on the first classified data, the second classified data, the third classified data, the fourth classified data and the historical common data, a pre-trained deep neural network is trained to obtain a trained data recognition model.

[0032] According to the present invention, a data intelligent desensitization management method for sensitive data is provided, wherein the cross entropy loss function of the trained data recognition model is:

[0033]

[0034] Where N is the total number of samples in the training data set; x i Represents the sample data after cleaning, encoding and vectorization; y i Represented as sample x i The true label of f(x i ; θ) represents the data recognition model for the input sample x i The predicted output of ; θ represents the parameters of the data recognition model.

[0035] A data intelligent desensitization management device for sensitive data, comprising:

[0036] An acquisition unit, used for acquiring target interaction data in a financial loan platform;

[0037] A computing unit, configured to input the target interaction data into a data recognition model to obtain sensitive data output by the data recognition model; the sensitive data includes personal identity information, financial account information, financial information, and credit information; the data recognition model is trained based on sample interaction data and its corresponding sensitive data label results;

[0038] A matching unit, configured to determine a desensitization mode of the sensitive data based on a data type of the sensitive data;

[0039] An output unit is used to desensitize the sensitive data based on the desensitization mode to obtain target data.

[0040] The present invention provides an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned method for intelligent data desensitization management of sensitive data when executing the program.

[0041] The present invention provides a non-transitory readable storage medium on which a program is stored. When the program is executed by a processor, the steps of the above-mentioned data intelligent desensitization management method for sensitive data are implemented.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] The present invention provides a data intelligent desensitization management method for sensitive data, which obtains target interaction data in a financial loan platform, inputs the target interaction data into a data recognition model, identifies sensitive data in the target interaction data, and then determines a desensitization mode for the sensitive data based on the data type of the sensitive data, so that different types of sensitive data are individually desensitized by different desensitization modes, thereby improving the security level of desensitizing sensitive data in the target interaction data, avoiding the risk that other types of data are easily decrypted after a single type of data is decrypted, and thus reducing the data security risk of the financial loan platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 A schematic diagram of a flow chart of a method for intelligent desensitization management of sensitive data provided by an embodiment of the present invention;

[0046] Figure 2 A schematic diagram of the structure of a data intelligent desensitization management device for sensitive data provided by an embodiment of the present invention;

[0047] Figure 3 This is a schematic structural diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Figure 1 The following is a flow chart of a method for intelligent desensitization management of sensitive data provided by an embodiment of the present invention. Figure 1 As shown, the method comprises:

[0050] Step 101, obtaining target interaction data in a financial loan platform;

[0051] Specifically, in the data desensitization process, since the sources of the target interaction data are diverse, systematic acquisition of the target interaction data can avoid missing data sources that may contain sensitive information.

[0052] Including: obtaining the first sub-data, second sub-data, third sub-data and fourth sub-data in the financial loan platform; the first sub-data is the application process data on the financial loan platform; the second sub-data is the repayment data on the financial loan platform; the third sub-data is the communication data with platform customers on the financial loan platform; the fourth sub-data is the operation behavior data on the financial loan platform.

[0053] Among them, the application process data includes the loan application form filled out by the user on the financial loan platform, covering personal identity information (such as name, ID number, date of birth, etc.) and loan-related information (such as loan amount, loan user, repayment period, etc.), and personal identity information and loan-related information are usually stored in the platform's database in a structured form, which can be obtained through the database interface with the loan application system.

[0054] Repayment data includes the user's repayment history (such as repayment date, repayment amount, repayment method), whether it is overdue and the number of overdue days, etc. This information can be obtained from the repayment management system of the financial loan platform. Its storage form may be a repayment record table in the database, which is associated with the user's loan account.

[0055] The communication data with platform customers includes the communication records between users and platform customer service personnel through online customer service systems, customer service hotlines and other channels, including consultations on loan products, feedback on issues during the application process, explanations of repayment difficulties, etc. The above data are usually stored in the customer service system database in text form and can be obtained through specific interfaces.

[0056] Operational behavior data includes the time and frequency of users logging into the platform, records of browsing different loan product pages on the platform, adjustments to parameters such as loan amounts and terms, and clicks and inputs during the application or repayment process. The above data is obtained through the platform's log system, and during the acquisition process, the system automatically converts it into an analyzable format to facilitate subsequent operations.

[0057] The first sub-data, the second sub-data, the third sub-data and the fourth sub-data are combined to obtain target interaction data.

[0058] Furthermore, when effectively combining the four sets of data together, it is necessary to first find the association key between them. In financial loan platforms, the user's ID is usually used as the association key, so that whether it is application process data, repayment data, communication data with platform customers or operation behavior data, they are all generated around the user's activities on the platform. Therefore, each user has a unique user ID, and different types of data are associated through this ID.

[0059] Step 102, input the target interaction data into the data recognition model to obtain sensitive data output by the data recognition model; sensitive data includes personal identity information, financial account information, financial information and credit information; the data recognition model is trained based on the label results of the sample interaction data and its corresponding sensitive data.

[0060] Specifically, by inputting the target interaction data into the data recognition model and obtaining the output sensitive data, intelligence can be achieved in the process of sensitive data recognition. At the same time, the type of sensitive data can be intelligently identified. Therefore, through mechanized training of the data recognition model, the efficiency and accuracy of recognition can be obtained.

[0061] Furthermore, in the process of training the data recognition model, the following steps are specifically included:

[0062] Obtain historical data of the financial loan platform; historical data includes historical sensitive data and historical general data;

[0063] Classify historical sensitive data to obtain first-classified data, second-classified data, third-classified data, and fourth-classified data; the first-classified data is sensitive data in user application process data, the second-classified data is sensitive data in user repayment data, the third-classified data is sensitive data in communication data between users and platform customers, and the fourth-classified data is sensitive data in user operation behavior data;

[0064] Based on the first classification data, the second classification data, the third classification data, the fourth classification data and the historical general data, the pre-trained deep neural network model is trained to obtain a data recognition model. In the process of training the model with the deep neural network, sensitive data can be effectively judged from the target interaction data, providing support for data security and privacy protection of the financial loan platform.

[0065] Step 103: Determine a desensitization mode for the sensitive data based on the data type of the sensitive data.

[0066] Specifically, since sensitive data includes four types: personal identity information, financial account information, financial information and credit information, personal identity information includes ID number information, name information and date of birth information; credit information includes credit score information and credit report information; credit report information includes overdue record text description information and detailed record information.

[0067] Therefore, when the sensitive data is personal identity information, the desensitization mode is the first desensitization mode;

[0068] The first desensitization mode includes three sub-modes, namely sub-mode A, sub-mode B and sub-mode C, which correspond to the ID card number information, name information and date of birth information respectively.

[0069] When the sensitive data is financial account information, the desensitization mode is the second desensitization mode;

[0070] When the sensitive data is financial information, the desensitization mode is the third desensitization mode;

[0071] When the sensitive data is credit information, the desensitization mode is the fourth desensitization mode.

[0072] The fourth desensitization mode also includes two sub-modes, namely sub-mode D and sub-mode E, which correspond to credit score information and overdue record text description information respectively.

[0073] Through four different desensitization modes and their corresponding sub-modes, more and more flexible options are provided for the desensitization of sensitive data to meet the desensitization needs of different types of sensitive data.

[0074] Step 104: desensitize the sensitive data based on the desensitization mode to obtain target data.

[0075] Specifically, when the sensitive data is personal identity information, the sensitive data is desensitized based on the first desensitization mode, including:

[0076] If the sensitive data is ID card number information, use sub-mode A to split the ID card number information to obtain multiple segments of number information; convert at least one first target number information in the multiple segments of number information into its corresponding English letter information; splice the English letter information and the second target number information to obtain the target data; the second target number information is the number information in the multiple segments of number information except the first target number information.

[0077] If the sensitive data is name information, sub-mode B is used to convert at least one first target character in the name information into a target coded character through a first preset formula; the target coded character and the second target character are concatenated to obtain target data; the second target character is the character in the name information other than the first target character.

[0078] If the sensitive data is date of birth information, use sub-mode C to split the date of birth information to obtain multiple time information segments; offset at least one first target time information in the multiple time information segments by a preset length to obtain offset time information; concatenate the offset time information and the second target time information to obtain target data; the second target time information is the time information in the multiple time information segments except the first target time information.

[0079] When the sensitive data is financial account information, the sensitive data is desensitized based on the second desensitization mode, including: if the sensitive data is financial account information, the financial account information is split to obtain multiple sub-segment information; for at least one first target sub-segment information in the multiple sub-segment information, the first target number in the first target sub-segment information is converted into a first digital code value; the first target number is at least one digit in the first target sub-segment information; the first digital code value is converted by a second preset formula to obtain a converted second digital code value; the second digital code value is converted into a corresponding target numerical value; the target numerical value, the second target sub-segment, and the second target number are spliced ​​to obtain the target data; the second target sub-segment is the sub-segment information in the multiple sub-segment information except the first target sub-segment information; the second target number is the number in the first target sub-segment except the first target number.

[0080] When the sensitive data is financial information, the sensitive data is desensitized based on the third desensitizing mode, including: if the sensitive data is financial information, the financial information is processed through the third preset formula to obtain a target mapping value; the target mapping value is mapped within a preset interval; the target mapping value is nonlinearly transformed through a fourth preset formula to obtain blurred financial information; and a random noise value is added to the blurred financial information to obtain target data.

[0081] When the sensitive data is credit information, the sensitive data is desensitized based on the fourth desensitization mode, including:

[0082] If the sensitive data is credit information, sub-mode D is used. If the sensitive data is credit score information, the first score value and the second score value of the platform credit score are obtained; the first score value is less than the credit score information; the second score value is greater than the credit score information; the first score value and the second score value are grouped into intervals to obtain multiple score sub-intervals; based on the credit score information and the multiple score sub-intervals, the interval number of the score sub-interval where the credit score information is located is obtained; the credit score information and the interval number are processed by the fifth preset formula to obtain the target data.

[0083] If the sensitive data is the text description information of overdue records, sub-mode E is adopted to perform sensitive word recognition on the text description information of overdue records to determine multiple sensitive keywords; the multiple sensitive keywords are fuzzy processed to obtain the target data.

[0084] The present invention relates to the field of data processing, and proposes a data intelligent desensitization management method for sensitive data. In the present invention, the proposed data intelligent desensitization management method obtains target interaction data in a financial loan platform, inputs the target interaction data into a data recognition model, identifies sensitive data in the target interaction data, and then determines the desensitization mode of the sensitive data based on the data type of the sensitive data, so that different types of sensitive data are individually desensitized by different desensitization modes, improves the security of desensitizing sensitive data in the target interaction data, avoids the risk that other types of data are easily cracked after a single type of data is cracked, and thus reduces the data security risk of the financial loan platform.

[0085] In addition, in step 101, the first sub-data, the second sub-data, the third sub-data and the fourth sub-data are combined to obtain target interaction data, including:

[0086] The first sub-data, the second sub-data, the third sub-data and the fourth sub-data are respectively subjected to original format recognition and unified processing to obtain first sub-structure data, second sub-structure data, third sub-structure data and fourth sub-structure data in a unified format; in the original format recognition and unified processing process, since the first sub-data, the second sub-data, the third sub-data and the fourth sub-data come from different data sources and their formats are different, the source and format of each sub-data are firstly identified and determined, and then unified formatting processing is performed on the data in different formats.

[0087] Specifically, in one embodiment, for structured data (such as application process data and partial repayment data), which are stored in different database systems (such as MySQL, Oracle, etc.), they can be extracted through data extraction tools such as ETL tools, and then converted into a unified structured format, such as uniformly converted into a data table form represented by rows and columns to ensure that the definition and data type of each field are consistent. For communication data in text form, its encoding can be unified into common encoding formats such as UTF-8 to facilitate subsequent processing. For operational behavior data, if it is in JSON format, the key information can be parsed and extracted, and converted into a format compatible with other data. If it is also converted into a data table form, different operational behaviors will be recorded as different fields.

[0088] The first substructure data, the second substructure data, the third substructure data and the fourth substructure data are denoised at the same time to obtain the first data to be fused, the second data to be fused, the third data to be fused and the fourth data to be fused respectively. In the process of denoising the data, each substructure data is first cleaned to remove the noise data and invalid data therein. Specifically, in one embodiment, for the application process data, there may be information filled in by the user by mistake (such as the ID card number filled in does not meet the format requirements), and these invalid records can be identified and deleted by setting data validation rules (such as regular expression validation of the ID card number format). For the repayment data, there may be test data or abnormal records caused by system failure (such as the repayment amount is negative or the date format is wrong), and these abnormal data are removed. For the communication data with the platform customers, it may contain some irrelevant greetings, emoticons, etc., which can be deleted through text processing technology (such as regular expression matching, lexical analysis of natural language processing, etc.). For the operation behavior data, there may be test operation records automatically generated by the system or incomplete records caused by network problems, and complete and valid operation behavior records are screened out. Then, the first data to be fused, the second data to be fused, the third data to be fused and the fourth data to be fused are obtained.

[0089] Based on the user ID, the first data to be fused, the second data to be fused, the third data to be fused, and the fourth data to be fused are associated and fused to obtain the target interaction data. During the specific operation in the data fusion process, each data to be fused can be first stored in a temporary data table or data structure, and then associated and integrated according to the user ID, and finally a target interaction data set containing all the data information to be fused is formed, in which each record corresponds to the complete interaction data of a user. In addition, after being combined into the target interaction data, the data can also be sorted according to business needs. For example, in one embodiment, if you want to sort the target interaction data according to the time sequence of the user's loan application, you can perform a sorting operation according to the application time field in the application process data.

[0090] Through the above steps, the four types of sub-data obtained from the financial loan platform (i.e., application process data, repayment data, communication data, and operational behavior data) can be effectively combined into target interaction data, providing a basis for subsequent data analysis, data desensitization, and other processing.

[0091] In step 102, the deep neural network adopts a multi-layer perceptron architecture. In one embodiment, for the input layer, taking the first classified data as an example, the data value features of the first classified data, such as the application amount, application age, etc., are directly used as input values ​​of the input neurons. For text features, the word vectors after the text features are processed, if the word vector dimension is 100 each time, a description of the application purpose has 20 valid words after word segmentation, these 20 word vectors are concatenated in sequence to form a 2000-dimensional vector, which is input into the corresponding neurons of the input layer to complete the input of the first classified data in the input layer.

[0092] After that, in the first hidden layer, after the input layer data is passed into the first hidden layer, each neuron receives the weighted sum input from all neurons in the input layer, and the calculation formula is: l =∑ l ω lμ x l +b l , where x l is the output of the lth neuron in the input layer, ω lμ is the weight connecting the lth neuron in the input layer and the μth neuron in the first hidden layer, b l is the bias of the lth neuron in the first hidden layer, and then passes through the ReLU function y l =max(0, z l ) is activated to obtain the output of the first hidden layer 1 neuron; and for the text feature part, the 2000-dimensional text vector is fused with the numerical feature through a specific weight matrix, so that the first hidden layer begins to learn the correlation between the text and the numerical feature, maps it to the 128-dimensional space, and represents it together with the results obtained after calculation of other numerical features in the 128-dimensional space.

[0093] Next, in the second hidden layer, hidden layer 2 has 64 neurons, and the output of 128 neurons in hidden layer 1 is used as the input of hidden layer 2. The above weighted summation and activation process is repeated to further abstract and compress the features, and learn more advanced feature combinations, such as learning the association pattern between certain specific income ranges and specific application purposes for sensitive data judgment, and further mapping the 128-dimensional data to 64 dimensions. Repeat the second hidden layer operation in subsequent hidden layers until a more accurate state is reached.

[0094] Specifically, the cross entropy loss function of the data recognition model is obtained as:

[0095]

[0096] In the financial loan platform data, there may also be noise data, which may affect the judgment of the model, so we define z iRepresented as data x i The noise level can be measured by some data quality evaluation indicators, such as data integrity and consistency, with a value range of [0,1]), and a noise penalty term is introduced to make the model more cautious in predicting high-noise data, so that the loss function of the noise processing term is obtained as:

[0097]

[0098] When the data noise is large (close to 1), this term will produce a large penalty if the model predicts incorrectly.

[0099] Where N is the total number of samples in the training data set; x i Represented as interaction data samples after cleaning, encoding and vectorization; y i Represented as sample x i The true label, y i The value of y is 0 or 1. i When it is 1, it means sample x i It is sensitive data. i When it is 0, it means sample x i belongs to ordinary data; f(x i ; θ) represents the data recognition model for the input sample x i The predicted output of ; θ represents the parameters of the data recognition model.

[0100] The final comprehensive loss function is:

[0101] L=L CE +γL noise ;

[0102] Among them, γ represents the weight parameter of the noise processing loss function, which can be adjusted according to the specific business scenario and data characteristics. By adjusting the above loss function, a more robust and adaptable data recognition model can be trained.

[0103] In step 104, in the first desensitization mode:

[0104] When the sensitive data is ID card number information, use sub-mode A to split the ID card number information to obtain multiple segments of number information; convert at least one first target number information in the multiple segments of number information into its corresponding English letter information; splice the English letter information and the second target number information to obtain the target data; the second target number information is the number information in the multiple segments of number information except the first target number information.

[0105] Specifically, the first target number information can be either one segment of number information or multiple segments of number information. This embodiment takes one segment of number information as an example. In one embodiment, if the ID number information is 34******4***5*254, after splitting it, four segments of number information, 34***, ***4, ***5, and *254, are obtained; the target number information 34*** in the first segment of information can be converted into its corresponding English letter information ty***; then the English letters ty*** and the second target number information are spliced ​​to obtain the target data, i.e. ty******4***5*254; the second target number information is the number information in the multiple segments of number information except the first target number information.

[0106] When the sensitive data is name information, sub-mode B is used to convert at least one first target character in the name information into a target coded character through a first preset formula; the target coded character and the second target character are concatenated to obtain the target data; the second target character is the character in the name information other than the first target character.

[0107] Specifically, each Chinese character is a character information, and the first target character can be a single character or a combination of multiple consecutive characters. This embodiment takes the first target character as a single character as an example. In one embodiment, if the name is Zhang San Wanzi, the character three is converted into the target coded character The desensitized name is Zhang ball.

[0108] The first preset formula is:

[0109] c ^ =chr((ord(c i )*w i )mod q+k1), get the target encoding character c converted by each character ci- ^ i;

[0110] Among them, ord(c i ) represents the Unicode code value of character ci; wi represents the randomly generated weight value of character ci; q represents the base of the modulo operation (usually a sufficiently large value to ensure that the transformed character is still a valid displayable character); k1 represents the offset used to adjust the character range (usually set to a constant). Through the above-mentioned weighted character transformation and modulo operation, a desensitization method is provided that can retain some features of the name and effectively hide sensitive information. According to actual needs, various parameters can be flexibly adjusted to achieve different levels of privacy protection effects.

[0111] When the sensitive data is date of birth information, sub - pattern C is used to split the date of birth information to obtain multiple segments of time information; at least one first target time information in the multiple segments of time information is offset by a preset length to obtain offset time information; the offset time information and the second target time information are concatenated to obtain the target data; the second target time information is the time information other than the first target time information in the multiple segments of time information. Here, the first target time information is any one or more combinations of the multiple segments of time information.

[0112] Specifically, in one embodiment, if the date of birth is October 9, 1994, when it is divided into three segments according to year, month, and day, the first segment is 1994, the second segment is 10, and the third segment is 9. At this time, the third segment is selected as the first target time information, and 30 days is set as the offset preset length. Then the offset third segment is 39. After concatenating it with the first segment 1994 and the second segment 9, the finally desensitized date of birth is November 8, 1994, and the date is automatically converted.

[0113] When in the second desensitization mode:

[0114] In one embodiment, the financial account information is split to obtain multiple sub - fragment information. For example, the original financial account data A=a1a2…a n is split into three segments, including the head H=a1a2...a m1 , the middle M=a m+1 a m+2 …a m2 , and the tail T=a m2+ 1a m2+2 …a n ; where m1 and m2 are split points determined according to the account type, and m1 and m2 can be automatically adjusted according to actual needs to adapt to the characteristics of different account lengths and structures, improving the generality of the algorithm.

[0115] For at least one first target sub - fragment information in the multiple sub - fragment information, the first target number in the first target sub - fragment information is converted into a first digital coding value. For example, ASCII code value processing is performed on the middle M, so that each digit a j , m1<j≤m2 in the middle M is converted into the corresponding ASCII code value d j , that is, the first digital coding value. Then, the converted ASCII code value d i is converted through a second preset formula to obtain the converted second digital coding value. The second preset formula is:

[0116] d ^ i =(d i +k2)mod256;

[0117] Among them, k2 is the offset value based on the account generation time;

[0118] The changed ASCII code value d ^ i Convert back to target value M - Finally, the target value, the second target sub-segment, and the second target number are spliced ​​together to obtain the target data A - =HM - T.

[0119] Through the above algorithm processing, the algorithm has a certain degree of reversibility. For authorized internal personnel, if they know the account generation time or platform-specific key, they can restore the original account through reverse calculation. In addition, the first target sub-segment information can be multiple sub-segments. When using the second preset formula for conversion, the characteristics of the ASCII code value and the dynamic offset value k2 are utilized, so that even if the account is partially leaked during storage and transmission, it is difficult for the attacker to restore the complete account information. This makes the desensitization method unique and increases the difficulty of cracking.

[0120] When in the third desensitization mode:

[0121] That is, if the sensitive data is financial information, the financial information is processed by the third preset formula to obtain a target mapping value; the target mapping value is mapped within a preset interval; the target mapping value is nonlinearly transformed by the fourth preset formula to obtain blurred financial information; and a random noise value is added to the blurred financial information to obtain target data.

[0122] Specifically, the third preset formula is as follows:

[0123] The preset interval is [0,100);

[0124] The target mapping value F after being processed by the third preset formula STD Within [0,100).

[0125] Where F represents the initial financial information. For different types of initial financial information (such as income, assets, and liabilities), the units are all in monetary units (such as yuan). min and F max They represent the minimum and maximum values ​​of the financial data of this type determined according to the business scenario. For example, in one embodiment, for the income situation, the income range of platform users is roughly between 2,000 yuan and 50,000 yuan, then F min =2000, F max =5000.

[0126] The fourth preset formula is as follows:

[0127]

[0128] The nonlinear characteristics of the sine function are used to further blur the original value of the data and obtain the blurred financial information F trans , making the differences between adjacent data less obvious after desensitization.

[0129] Finally, the blurred financial information F trans Add random noise values ​​to obtain target data. The formula for adding random noise values ​​is as follows:

[0130] F ^ =F trans +R;

[0131] Where R represents random noise, and its value range is [-∈, ∈], and ∈ represents the noise intensity parameter determined according to data sensitivity and business requirements.

[0132] Through the steps of standardization, nonlinear transformation and adding random noise, the original financial data is effectively blurred. Even if the data is leaked to a certain extent, it is difficult for attackers to accurately infer the user's true financial status, which protects the user's privacy. In addition, at the macro level, the processed data still retains a certain relative size relationship to facilitate subsequent use, and the parameter F min 、F max ,∈ can be adjusted according to different financial data types (income, assets, liabilities) and business scenarios to meet usage requirements.

[0133] If the sensitive data is credit information, it is in the fourth desensitization mode:

[0134] At this time, according to the fact that the credit information is credit score information, sub-mode D is used, that is, if the sensitive data is credit score information, the first score value and the second score value of the platform credit score are obtained; the first score value is less than the credit score information; the second score value is greater than the credit score information; the first score value and the second score value are grouped into intervals to obtain multiple score sub-intervals; based on the credit score information and the multiple score sub-intervals, the interval number of the score sub-interval where the credit score information is located is obtained; the credit score information and the interval number are processed by the fifth preset formula to obtain the target data. The target data thus obtained is the score information after the information score information is desensitized.

[0135] The fifth preset formula is:

[0136] And satisfy S∈[S k ,S k+1 );

[0137] Among them, S -represents the target data after credit score information is desensitized, k represents the interval number of the score sub-interval where the credit score information is located; R represents the time-based random factor value, which is the number of milliseconds of the current timestamp, R∈[0,9); S represents the credit score information; Expressed as an adjustment coefficient. (Set according to the strictness of data desensitization and data characteristics)

[0138] Specifically, in one embodiment, if the credit score information is 498 points, the first score value can be 400 points, the second score value can be 600 points, and the score between 400 points and 600 points is divided into a score sub-interval every 20 points. The score sub-interval where the score information is 500 points is numbered 5, that is, it is in the score sub-interval of 480-500, and R is selected as 5. Select 0.5 and substitute it into the fifth preset formula. That is, the de-sensitized credit score information is 679.

[0139] By implementing grouping and time-based random factors, the relative position information of credit scores is retained to a certain extent (credit scores in different intervals still have distinction after desensitization), while increasing the randomness and uncertainty of the desensitized results, effectively protecting the privacy of credit scores. In addition, the system still has a certain degree of usability in macro analysis (such as statistical distribution of credit status by interval).

[0140] If the credit information is the text description information of the overdue record, sub-mode E is adopted, that is, the text description information of the overdue record is subjected to sensitive word recognition to determine multiple sensitive keywords; multiple sensitive keywords are subjected to fuzzy processing to obtain target data.

[0141] Specifically, in the process of identifying sensitive words, a sensitive word library is first constructed. The sensitive word library is classified and constructed according to the common vocabulary types in overdue records, and stored in the database of the financial loan platform. When used, it is directly retrieved from the database. The fuzzification processing of sensitive keywords is to convert sensitive keywords into fuzzy words. While constructing the sensitive word library, its corresponding fuzzy word library is constructed synchronously, so that each sensitive keyword in the database corresponds to a fuzzy word.

[0142] Furthermore, in one embodiment, words related to overdue time, such as precise time words such as three months, half a year, and one year, words related to overdue amount, such as precise monetary words such as five thousand yuan, ten thousand yuan, and one hundred thousand yuan, and words describing the reasons for overdue, such as unemployment, business failure, and difficulty in capital turnover, are all sensitive keywords, and their corresponding fuzzy words are: precise time words can be converted into fuzzy words such as several months, a period of time, and a long time; precise monetary words can be converted into fuzzy words such as a certain amount of money, a certain amount of money, and a large amount of funds; precise reason words can be converted into fuzzy words such as changes in personal financial status, fluctuations in economic conditions, and capital flow problems.

[0143] Figure 2 A schematic diagram of the structure of a data intelligent desensitization management device for sensitive data provided by an embodiment of the present invention, such as Figure 2 As shown, the device includes: an acquisition unit 10, used to acquire target interaction data in a financial loan platform; a calculation unit 20, used to input the target interaction data into a data recognition model to obtain sensitive data output by the data recognition model; the sensitive data includes personal identity information, financial account information, financial information and credit information; the data recognition model is trained based on sample interaction data and its corresponding sensitive data label results; a matching unit 30, used to determine a desensitization mode of the sensitive data based on the data type of the sensitive data; an output unit 40, used to desensitize the sensitive data based on the desensitization mode to obtain target data.

[0144] Figure 3 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 3 As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330 and a communication bus 340, wherein the processor 310, the communication interface 320 and the memory 330 communicate with each other through the communication bus 340. The processor 310 may call the logic instructions in the memory 330 to execute a data intelligent desensitization management method for sensitive data, the method comprising: obtaining target interaction data in a financial loan platform; inputting the target interaction data into a data recognition model to obtain sensitive data output by the data recognition model; the sensitive data includes personal identity information, financial account information, financial information and credit information; the data recognition model is trained based on the label results of the sample interaction data and its corresponding sensitive data; and determining the desensitization mode of the sensitive data based on the data type of the sensitive data.

[0145] In addition, the logic instructions in the above-mentioned memory 330 can be implemented in the form of a software functional unit and can be stored in a readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a device (which can be a person, a server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0146] On the other hand, the present invention also provides a program product, which includes a program stored on a non-transitory readable storage medium, and the program includes program instructions. When the program instructions are executed, a data intelligent desensitization management method for sensitive data provided by the above methods can be executed, and the method includes: obtaining target interaction data in a financial loan platform; inputting the target interaction data into a data recognition model to obtain sensitive data output by the data recognition model; the sensitive data includes personal identity information, financial account information, financial information and credit information; the data recognition model is trained based on the label results of sample interaction data and its corresponding sensitive data; and determining the desensitization mode of the sensitive data based on the data type of the sensitive data. On the other hand, the present invention also provides a non-transitory readable storage medium having a program stored thereon, which, when executed by a processor, is implemented to execute the above-mentioned data intelligent desensitization management method for sensitive data, the method comprising: obtaining target interaction data in a financial loan platform; inputting the target interaction data into a data recognition model to obtain sensitive data output by the data recognition model; the sensitive data includes personal identity information, financial account information, financial information and credit information; the data recognition model is trained based on the label results of sample interaction data and its corresponding sensitive data; and determining the desensitization mode of the sensitive data based on the data type of the sensitive data.

[0147] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A data intelligent desensitization management method for sensitive data, characterized in that: include: Obtain target interaction data in the financial loan platform; Inputting the target interaction data into a data recognition model to obtain sensitive data output by the data recognition model; The sensitive data includes personal identity information, financial account information, financial information and credit information; The data recognition model is trained based on the label results of sample interaction data and its corresponding sensitive data; Determining a desensitization mode for the sensitive data based on the data type of the sensitive data; The sensitive data is desensitized based on the desensitization mode to obtain target data.

2. The intelligent desensitization management method for sensitive data according to claim 1, characterized in that: The personal identity information includes ID number information, name information and date of birth information; Desensitizing the sensitive data based on the desensitization mode to obtain target data includes: If the sensitive data is identity card number information, the identity card number information is split to obtain multiple segments of number information; Convert at least one first target number information in the multiple number information segments into its corresponding English letter information; concatenate the English letter information and the second target number information to obtain the target data; the second target number information is the number information in the multiple number information segments except the first target number information; If the sensitive data is the name information, converting at least one first target character in the name information into a target coded character by using a first preset formula; concatenating the target coded character and a second target character to obtain the target data; the second target character is a character in the name information other than the first target character; If the sensitive data is date of birth information, the date of birth information is split to obtain multiple segments of time information; at least one first target time information in the multiple segments of time information is offset by a preset length to obtain offset time information; the offset time information and the second target time information are spliced ​​to obtain the target data; the second target time information is the time information in the multiple segments of time information except the first target time information.

3. The intelligent desensitization management method for sensitive data according to claim 1, characterized in that: Desensitizing the sensitive data based on the desensitization mode to obtain target data includes: If the sensitive data is financial account information, the financial account information is split into multiple sub-segment information; For at least one first target sub-segment information among the plurality of sub-segment information, converting a first target number in the first target sub-segment information into a first digital code value; the first target number is at least one digit in the first target sub-segment information; Converting the first digital code value using a second preset formula to obtain a converted second digital code value; Converting the second digital code value into a corresponding target value; The target data is obtained by splicing the target numerical value, the second target sub-segment, and the second target number; the second target sub-segment is the sub-segment information of the multiple sub-segment information except the first target sub-segment information; the second target number is the number in the first target sub-segment except the first target number.

4. The method for intelligent desensitization management of sensitive data according to claim 1, characterized in that: Desensitizing the sensitive data based on the desensitization mode to obtain target data includes: If the sensitive data is the financial information, the financial information is processed by a third preset formula to obtain a target mapping value; the target mapping value is mapped within a preset interval; Performing nonlinear transformation on the target mapping value through a fourth preset formula to obtain blurred financial information; A random noise value is added to the blurred financial information to obtain the target data.

5. The method for intelligent desensitization management of sensitive data according to claim 1, characterized in that: The credit information includes credit score information and overdue record text description information, and the sensitive data is desensitized based on the desensitization mode to obtain target data, including: If the sensitive data is credit score information, a first score value and a second score value of the platform credit score are obtained; the first score value is less than the credit score information; the second score value is greater than the credit score information; the first score value and the second score value are grouped into intervals to obtain a plurality of score sub-intervals; based on the credit score information and the plurality of score sub-intervals, an interval number of the score sub-interval where the credit score information is located is obtained; the credit score information and the interval number are processed by a fifth preset formula to obtain the target data; If the sensitive data is text description information of overdue records, the text description information of overdue records is subjected to sensitive word recognition to determine a plurality of sensitive keywords; the plurality of sensitive keywords are subjected to fuzzy processing to obtain the target data.

6. The method for intelligent desensitization management of sensitive data according to claim 1, characterized in that: The training process of the data recognition model includes: Acquire historical data of the financial loan platform; the historical data includes historical sensitive data and historical general data; Classify the historical sensitive data to obtain first classification data, second classification data, third classification data and fourth classification data; the first classification data is sample data in the user application process data, the second classification data is sample data in the user repayment data, the third classification data is sample data in the communication data between the user and the platform customers, and the fourth classification data is sample data in the user operation behavior data; Based on the first classified data, the second classified data, the third classified data, the fourth classified data and the historical common data, a pre-trained deep neural network is trained to obtain a trained data recognition model.

7. The method for intelligent desensitization management of sensitive data according to claim 6, characterized in that: The cross entropy loss function of the trained data recognition model is: Where N represents the total number of samples in the training data set; x i Represents the sample data after cleaning, encoding and vectorization; y i Represented as sample x i The true label of f(x i ; θ) represents the data recognition model for the input sample x i The predicted output of ; θ represents the parameters of the data recognition model.

8. A data intelligent desensitization management device for sensitive data, characterized in that: include: An acquisition unit, used for acquiring target interaction data in a financial loan platform; A computing unit, used for inputting the target interaction data into a data recognition model to obtain sensitive data output by the data recognition model; The sensitive data includes personal identity information, financial account information, financial information and credit information; The data recognition model is trained based on sample interaction data and its corresponding sensitive data label results; A matching unit, configured to determine a desensitization mode of the sensitive data based on a data type of the sensitive data; An output unit is used to desensitize the sensitive data based on the desensitization mode to obtain target data.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the data intelligent desensitization management method for sensitive data as described in any one of claims 1 to 7 are implemented.

10. A non-transitory readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the data intelligent desensitization management method for sensitive data as described in any one of claims 1 to 7 are implemented.

Images