Sensitive data risk auditing method and device, electronic equipment, storage medium and program product

Abstract

The invention discloses a sensitive data risk auditing method, electronic equipment, a sensitive data risk auditing device, a storage medium and a program product, and relates to the technical field of data processing.The sensitive data risk auditing method comprises the steps that target text information corresponding to original data is input into a preset target sensitive information recognition model, at least one piece of sensitive information in the target text information and sensitive categories corresponding to the sensitive information are obtained through prediction of the target sensitive information recognition model; according to the sensitive category corresponding to each piece of sensitive information and a preset sensitive level rule, determining a sensitive level corresponding to each piece of sensitive information; and determining the risk level of the original data based on the sensitivity level of each piece of sensitive information and the frequency of each sensitivity level. According to the technical scheme of the invention, the problems of low efficiency and low accuracy during auditing of sensitive information are solved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a sensitive data risk audit method, device, electronic device, computer-readable storage medium, and computer program product. Background Art

[0002] In order to protect the sensitive privacy data of users, it is usually necessary to audit and identify the sensitive data on the data-bearing platform, and accurately analyze the sensitive data included in the original data, as well as the type and risk level of the sensitive data. The commonly used sensitive data identification solution in the big data environment based on traffic analysis is to identify the interactive traffic between the upper-level application and the big data platform by using protocol parsing, so as to obtain the metadata information of the big database, make a preliminary judgment on sensitive information through simple character feature matching, and extract the "semantic content" behind the corresponding table field based on the SQL-like (Structured Query Language) statements generated by the actual interaction between the upper-level application and the big data platform, and identify sensitive information based on the semantic results.

[0003] However, this sensitive information identification strategy is based only on simple character feature matching and SQL statement analysis, which makes it difficult to fully understand complex contextual associations or semantic ambiguity, and is prone to misidentification of sensitive information or failure to discover potential sensitive data. Moreover, processing real-time big data traffic and performing in-depth analysis may consume a lot of computing resources and time, especially in the face of high-speed real-time traffic. This method may not be able to achieve real-time or near real-time monitoring of sensitive information, resulting in low audit efficiency of sensitive data.

[0004] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The main purpose of this application is to provide a sensitive data risk audit method, device, electronic device, computer-readable storage medium and computer program product, aiming to solve the technical problems of low accuracy and efficiency of current sensitive information identification strategies.

[0006] To achieve the above objectives, this application proposes a sensitive data risk audit method, which includes:

[0007] Inputting the target text information corresponding to the original data into a preset target sensitive information recognition model, and predicting at least one sensitive information in the target text information and the sensitive categories corresponding to each sensitive information through the target sensitive information recognition model;

[0008] Determine the sensitivity level corresponding to each of the sensitive information according to the sensitivity category corresponding to each of the sensitive information and the preset sensitivity level rule, wherein the sensitivity level rule is used to characterize the correspondence between the sensitive category and the sensitivity level;

[0009] Based on the sensitivity level of each sensitive information and the frequency of each sensitivity level, the risk level of the original data is determined.

[0010] In one embodiment, the step of determining the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level includes:

[0011] Traversing the sensitivity levels of each of the sensitive information, and counting the frequencies corresponding to each of the sensitivity levels;

[0012] Calculate the product of each sensitivity level and the corresponding frequency to obtain the level number corresponding to each sensitivity level;

[0013] Calculate the sum of the level numbers corresponding to each of the sensitivity levels to determine the level number of the original data;

[0014] The risk level corresponding to the original data is determined according to the corresponding relationship between the preset level number and the risk level and the level number of the original data.

[0015] In one embodiment, before the step of inputting the target text information corresponding to the original data into a preset target sensitive information recognition model, the method further includes:

[0016] Collect training data within a preset range;

[0017] Preprocessing the training data to obtain processed training text information, wherein the preprocessing process at least includes data cleaning, standardization, part-of-speech tagging, and deletion of stop words;

[0018] Obtaining sensitive categories of sensitive information in the training text information to obtain sensitive category label information;

[0019] Training a pre-built initial sensitive information recognition model according to the training text information and the sensitive category label information;

[0020] When the harmonic mean of the precision and recall of the initial sensitive information recognition model is higher than a preset threshold, the model training is stopped and the initial sensitive information recognition model is set as the target sensitive information recognition model.

[0021] In one embodiment, the step of training the pre-built initial sensitive information recognition model according to the training text information and the sensitive category label information includes:

[0022] Input the training text information into the initial sensitive information recognition model, and the initial sensitive information recognition model outputs a corresponding recognition result, wherein the recognition result includes the sensitive information and sensitive categories in the training text information;

[0023] Iteratively optimizing model parameters of the initial sensitive information recognition model based on the sensitive category label information corresponding to the training text information and the sensitive information and sensitive category corresponding to the recognition result;

[0024] Determine the number of correct instances, the number of misjudged instances, and the number of missed instances according to the sensitive category label information corresponding to the training text information and the sensitive information and sensitive category corresponding to the recognition result;

[0025] Calculating the accuracy of the initial sensitive information recognition model according to the number of correct instances and the number of misjudged instances;

[0026] The recall rate of the initial sensitive information identification model is calculated based on the number of correct instances and the number of missed instances.

[0027] In one embodiment, before the step of inputting the target text information corresponding to the original data into a preset target sensitive information recognition model, the method further includes:

[0028] When the original data is text data, deleting non-text content in the original data to obtain original text information, wherein the non-text content includes at least one of a hypertext markup language tag, a non-text character, and a blank character;

[0029] When the original data is non-text data, parsing the original data to obtain original text information;

[0030] Based on a predefined dictionary, the original text information is segmented to obtain each word included in the original text information, wherein the predefined dictionary includes predefined sensitive words and proprietary words;

[0031] By using a preset word segmentation tool or a pre-trained part-of-speech tagging model, a part-of-speech tag is added to each word in the original text information to obtain annotated text information, wherein the part-of-speech tag includes at least a name, a verb, and an adjective;

[0032] Based on a preset stop word list, the stop words in the annotated text information are deleted to obtain the target text information.

[0033] In one embodiment, after the step of determining the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level, the method further includes:

[0034] A risk audit report is generated based on the sensitive information corresponding to the original data, the sensitivity levels corresponding to the sensitive information, the frequency of the sensitivity levels, the risk levels, and the frequency of the sensitivity levels corresponding to the multiple sensitive categories in the original data.

[0035] In addition, to achieve the above purpose, the present application also proposes a sensitive data risk audit device, the sensitive data risk audit device comprising:

[0036] A sensitive category identification module, used to input the target text information corresponding to the original data into a preset target sensitive information identification model, and predict at least one sensitive information in the target text information and the sensitive category corresponding to each sensitive information through the target sensitive information identification model;

[0037] a sensitivity level determination module, which determines the sensitivity level corresponding to each of the sensitive information according to the sensitivity category corresponding to each of the sensitive information and a preset sensitivity level rule, wherein the sensitivity level rule is used to characterize the correspondence between the sensitivity category and the sensitivity level;

[0038] The risk level determination module determines the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level.

[0039] In addition, to achieve the above-mentioned purpose, the present application also proposes an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the sensitive data risk audit method as described above.

[0040] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by the processor, the steps of the sensitive data risk audit method described above are implemented.

[0041] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the sensitive data risk audit method described above.

[0042] The present application proposes a sensitive data risk audit method, which first inputs the target text information corresponding to the original data into a preset target sensitive information recognition model, and predicts at least one sensitive information in the target text information and the sensitive category corresponding to each of the sensitive information through the target sensitive information recognition model, and then determines the sensitivity level corresponding to each of the sensitive information according to the sensitive category corresponding to each of the sensitive information and the preset sensitivity level rule, wherein the sensitivity level rule is used to characterize the correspondence between the sensitive category and the sensitivity level, and finally determines the risk level of the original data based on the sensitivity level of each of the sensitive information and the frequency of each of the sensitive levels. The technical solution of the present application adopts a target sensitive information recognition model based on deep learning to identify and process the pre-processed target text information, which can efficiently and quickly identify the sensitive information and sensitive categories therein, facilitate risk rating of sensitive data existing in the original data, and finally determine the risk level of the original data by combining the sensitivity level of each sensitive information and the frequency of each sensitivity level. Compared with the traditional sensitive data recognition method based on real-time traffic analysis, it can more comprehensively check all audited data, and through the sensitive information recognition model to identify the sensitive category of the input data, and further determine the corresponding sensitivity level and risk level, it can efficiently audit the sensitive information existing in a large amount of data, overcoming the defect of the traditional solution that real-time processing and analysis of real-time large-volume data requires a lot of computing resources and time, which not only increases the accuracy, but also improves the audit efficiency, and facilitates better protection and identification of sensitive data. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0044] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0045] Figure 1 A flowchart of the first embodiment of the sensitive data risk audit method of this application is provided;

[0046] Figure 2 A flowchart of a sensitive data risk audit method in an embodiment of the present application;

[0047] Figure 3 This is a flowchart showing an example of the implementation steps of the sensitive data risk audit method in the embodiment of the present application;

[0048] Figure 4 This is a schematic diagram of the structure of a sensitive data risk audit device in an embodiment of the present application;

[0049] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the sensitive data risk audit method in the embodiment of the present application.

[0050] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0051] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0052] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0053] The execution subject of this embodiment may be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, a server, etc., or an electronic device or a control device capable of realizing the above functions. The following takes the server as an execution subject as an example to illustrate this embodiment and the following embodiments.

[0054] The current sensitive data identification scheme in the big data environment based on traffic analysis has certain defects: first, it emphasizes independence from the big data platform, but still needs to parse the interactive traffic between the upper-level application and the big data platform, which means that if the communication protocol is changed or the encryption is enhanced, the parsing may fail, affecting the accuracy of sensitive data identification; second, there are false positives and negatives: based on simple character feature matching and SQL statement analysis, it may not be able to fully understand complex contextual associations or semantic ambiguity, resulting in misidentification of sensitive information or failure to effectively discover all potential sensitive data; third, processing big data traffic and performing in-depth analysis may consume a lot of computing resources and time, especially when facing high-speed real-time traffic. This method may not be able to achieve real-time or near real-time monitoring of sensitive information; fourth, for sensitive information that is deliberately transmitted in secret, such as after obfuscation, encoding or other forms of deformation, this method may be difficult to effectively identify, especially in scenarios of high-level attacks or evasion of detection.

[0055] In order to overcome the above-mentioned defects, the technical solution of the embodiment of the present application is to achieve the following objectives: based on the user data security management specifications, security detection can be performed on the user sensitive information and data risks currently existing in the access system. When the security detection finds problems, the user can be reminded in real time, and the risk level of the overall data security of the platform can be assessed.

[0056] In order to achieve the above objectives, the present application embodiment provides a sensitive data risk audit method, referring to Figure 1 , Figure 1 This is a flowchart of the first embodiment of the sensitive data risk audit method of the present application, and the sensitive data risk audit method includes:

[0057] Step S10, inputting the target text information corresponding to the original data into a preset target sensitive information recognition model, and predicting at least one sensitive information in the target text information and the sensitive categories corresponding to each sensitive information through the target sensitive information recognition model;

[0058] The raw data may include various types of data, such as text data, audio data, image data, etc. First, the raw data must be converted into text form to facilitate processing by the pre-trained target information recognition model. In addition, in order to improve the efficiency of the recognition process, the extracted raw text information needs to be further pre-processed, such as data cleaning and standardization.

[0059] Before step S10, it is necessary to collect original data. Specifically, the target task-related corpus can be obtained through existing services, network data can be crawled, and open source data sets can be obtained. The scope of collection includes personal privacy information, corporate confidential information, government sensitive information, network information security information, etc. The data collection standards include: the data format uniformly adopts structured data storage to ensure that the data is easy to parse and process; the data quality ensures that the data is complete, accurate, consistent, and has no missing values ​​and errors; data security requires the use of encryption and decryption technology to ensure the security of data during collection, storage, and transmission; the frequency of data updates is determined according to actual needs to ensure the timeliness of the data; before starting data collection, a compliance review is first conducted to ensure that all data collection activities comply with relevant laws and regulations, and obtain necessary licenses and authorizations. In addition, it is necessary to clarify the source of data, determine the source of data, such as public data sets, partner sharing, user authorization, etc., and ensure the legitimacy and reliability of the data source. For data that needs to be provided by users, the user agreement and privacy policy should clearly inform the user of the purpose, scope and method of using the data, and obtain the user's explicit consent. In addition, legal data collection technologies are used, such as API (Application Programming Interface) calls, web crawlers, etc., but they must be carried out within the scope permitted by the data provider. For sensitive data, encrypted transmission and secure storage technologies are used to prevent data from being intercepted or tampered with during transmission.

[0060] Among them, the target sensitive information recognition model is a pre-trained and optimized machine learning model that can identify sensitive information in input text, such as the UIE (Universal Information Extraction) model, which is a natural language processing model based on deep learning and aims to solve information extraction tasks, including entity recognition, relationship extraction, event extraction, etc. The UIE model is usually based on large-scale pre-trained language models, such as BERT (Bidirectional Encoder Representations from Transformers), RoBERTa (Robustly Optimized BERT Pretraining Approach), ALBERT (ALite BERT), etc. These models are pre-trained on a large amount of text data, learn rich language representations, and can accurately identify sensitive information in target text information and determine the corresponding sensitive categories.

[0061] The UIE model is characterized by its versatility and can adapt to a variety of different information extraction tasks without the need for a large number of model structure adjustments for specific tasks, which makes the UIE model very flexible and convenient in practical applications. Therefore, the UIE model is usually trained on specific information extraction tasks through fine-tuning to adapt to specific data sets and task requirements.

[0062] It is understandable that the target text information includes sensitive information of various sensitive categories, where the sensitive categories can be predefined based on the application field. As an example, in a feasible sensitive information management rule, the specific definition, characteristics and information elements of each category of data such as Class A, Class B, Class C, and Class D involved in the system are clearly defined (refer to Table 1 and Table 2 below). These categories may include but are not limited to personal identity information (such as ID number, mobile phone number), enterprise core business data, etc.

[0063] Table 1: Customer information classification table

[0064]

[0065]

[0066]

[0067] Table 2: Classification of enterprise operation and management data

[0068]

[0069]

[0070]

[0071]

[0072] Step S20, determining the sensitivity level corresponding to each sensitive information according to the sensitivity category corresponding to each sensitive information and the preset sensitivity level rule, wherein the sensitivity level rule is used to characterize the correspondence between the sensitivity category and the sensitivity level;

[0073] It is understandable that different sensitive categories have different corresponding sensitivity levels. In order to further analyze the sensitivity level of sensitive information in the original data, it is necessary to convert each sensitive category output by the model into a corresponding sensitivity level. The sensitivity level rule can be in the form of a two-dimensional table. For example, according to a feasible sensitive information management rule, the sensitivity level can be divided into 4 levels: 1, 2, 3, and 4. The specific sensitive information risk level is shown in Table 3 below.

[0074] Table 3: Data classification table

[0075]

[0076]

[0077] Based on the above Table 3, the sensitivity level of each sensitive information in the original data can be determined.

[0078] Step S30, determining the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level.

[0079] In order to comprehensively evaluate the sensitivity risk of raw data from an overall perspective, it is necessary to comprehensively consider the sensitivity level of each sensitive information. In a set of raw data, there are sensitive information of multiple sensitivity levels. For the raw data, the higher the sensitivity level of the sensitive information, the higher the final risk level. The higher the frequency of sensitive information, the higher the final risk level. Both the sensitivity level and the frequency are positively correlated with the risk level.

[0080] In the embodiment of the present application, a risk level identification model can be pre-constructed, and the determined sensitivity level of each sensitive information and the frequency of sensitive information of each sensitivity level are input into the risk level identification model, and the risk level of the original data is identified by the risk level identification model.

[0081] Furthermore, the risk level may be expressed in digital form (such as a positive integer from 1 to 1000, or even above 1000), or may be expressed in terms such as “extremely high”, “high”, “medium”, and “low”.

[0082] Furthermore, in a feasible embodiment, after the step of determining the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level, the method further includes:

[0083] Step S40, generating a risk audit report according to each sensitive information corresponding to the original data, the sensitivity level corresponding to each sensitive information, the frequency of each sensitivity level, the risk level, and the frequency of the sensitivity level corresponding to multiple sensitive categories in the original data;

[0084] Among them, the risk audit report also includes.

[0085] In order to more clearly, completely and comprehensively display the risk of sensitive information of the original data, a completed risk audit report is generated after determining the sensitive information corresponding to the original data, the sensitivity level corresponding to each sensitive information, the frequency of each sensitive level and the risk level. The risk audit report clearly points out the sensitive information in the original data, the number of occurrences, the risk level and other detailed information, so that the staff can intuitively understand the risk of the original data, and can further make corresponding rectification suggestions and risk control measures.

[0086] On the basis of the aforementioned application embodiments, the above-mentioned sensitive information identification model and risk level identification model are connected to the audited system through an API (Application Programming Interface) interface, thereby realizing automated and intelligent auditing of massive data in the system and improving the audit efficiency of sensitive information.

[0087] For example, refer to Figure 2 ,The process of a feasible sensitive data risk audit method may include: model building, machine learning model optimization, data source scanning, data auditing, and output of audit report.

[0088] Furthermore, in a feasible embodiment, the step of extracting original text information corresponding to the original data, preprocessing the original text information, and obtaining the target text information includes:

[0089] Step A10, when the original data is text data, delete non-text content in the original data to obtain original text information, wherein the non-text content includes at least one of a hypertext markup language tag, a non-text character, and a blank character;

[0090] For text data, format standardization is performed to remove irrelevant HTML (Hypertext Markup Language) tags, special characters, blank characters and other non-text content to obtain purer original text information.

[0091] Step A20, when the original data is non-text data, the original data is parsed to obtain original text information;

[0092] For non-text data (such as pictures, audio, and video), special tools or services are used to parse and extract the text information (for example, using OCR (Optical Character Recognition) technology to recognize text in pictures, convert speech to text, etc.) for subsequent text analysis.

[0093] In addition, the original text information must be standardized, specifically converting line breaks and tabs in the text into spaces or other symbols to ensure the consistency of the text structure and facilitate subsequent word segmentation. For data in the form of tables and lists that may contain sensitive information, structured extraction can be performed and converted into an easy-to-process text format.

[0094] After step A20, it is necessary to perform part-of-speech tagging on each word in the original text information. Specifically, for numeric strings related to sensitive information in the original text information, such as mobile phone numbers, ID card numbers, order numbers, etc., they need to be retained and specially marked so that the subsequent sensitive information recognition model can accurately capture them.

[0095] In addition, before the words in the original text information are tagged with parts of speech, the text is first segmented to separate the words in the continuous text, so that each word can be tagged with parts of speech separately to ensure the accurate segmentation of sensitive words (such as name, address, product model, etc.). Then, the part-of-speech tagging function of the word segmentation tool or the pre-trained part-of-speech tagging model is used to add a part-of-speech tag to each segmented word. It should be noted that in the sensitive information recognition scenarios of various sensitive categories, the focus is on nouns, verbs, adjectives and other content words, especially nouns (such as names of people, groups, product names, etc.), verbs (such as purchase, payment, disclosure, etc.) and adjectives (such as important, confidential, sensitive, etc.) related to sensitive information.

[0096] Specifically, part-of-speech tagging includes the following steps:

[0097] Step A30, segmenting the original text information based on a predefined dictionary to obtain the words included in the original text information, wherein the predefined dictionary includes predefined sensitive words and proprietary words;

[0098] It should be noted that for various types of sensitive information, it is necessary to customize the configuration of the word segmentation tool, such as adding predefined dictionaries in specific fields (such as enterprise operation management data, user service-related information, user data and service content information, user identity and authentication information, etc.) to improve the recognition accuracy of the model for specific field vocabulary.

[0099] Step A40, add part-of-speech tags to each word in the original text information through a preset word segmentation tool or a pre-trained part-of-speech tagging model, where the part-of-speech tags at least include nouns, verbs, and adjectives.

[0100] In the embodiments of the present application, word segmentation tools such as jieba (a third-party Chinese word segmentation library) and HanLP (a natural language processing tool) can be used to accurately segment Chinese texts to ensure the accurate segmentation of sensitive words (such as names, addresses, product models, etc.).

[0101] Exemplarily, in addition to including nouns, verbs, and adjectives, the part-of-speech tags can also include personal names, punctuation marks, numbers, place names, pronouns, brand names, etc. to conduct a comprehensive and in-depth part-of-speech analysis of the words in the original information.

[0102] Step A50, based on a preset stop word list, delete the stop words in the annotated text information to obtain the target text information.

[0103] Among them, the stop word list is customized based on the word frequency statistics of the corpus and combined with various sensitive information recognition scenarios. During the process of deleting stop words, common words such as "de" (of) and "le" (了) that do not affect the judgment of sensitive information are mainly removed to reduce the noise in the text information and improve the recognition and processing efficiency of the model.

[0104] In addition, while deleting the general stop words in the text information, it is necessary to retain high-frequency words that may be related to sensitive information, such as "leak", "violate regulations", "complaint", etc. Although these words may belong to stop words in general texts, they may have important prompting effects in sensitive information recognition scenarios.

[0105] Exemplarily, after word segmentation and part-of-speech tagging, remove the words that are marked as stop words and are irrelevant to sensitive information according to the part-of-speech information and the customized stop word list, and retain the keywords closely related to various types of sensitive information, such as "customer", "order", "contract", "price", "channel", "partner", etc. These words help the model better understand the text context and improve the accuracy of sensitive information recognition.

[0106] Exemplarily, when the original data is text data, the steps of a feasible preprocessing method for the original text information may include: first, remove the non-text content in the data to clean the data, and the original text includes: Zhang San (138****5678) purchased a C-brand mobile phone in a building in District B, City A, with the order number #1234**., the corresponding cleaned text includes: Zhang San (138****5678) bought a mobile phone of brand C in a building in District Y, City A. Order number: #1234**. Then unify the format: The original text includes: Zhang San: 138****5678\tCity A, District Y, a certain building Purchase: Mobile phone of brand C Order number: #1234** After unifying the format, the text includes: Zhang San: 138****5678City A, District Y, a certain buildingPurchase: Mobile phone of brand COrder number: #1234**. Then perform Chinese word segmentation. The original text includes: Zhang San (138****5678) bought a mobile phone of brand C in a building in District Y, City A. Order number: #1234**. The word segmentation results include: Zhang San / ( / 138****678 / ) / City A / District Y / a certain / building / buy / mobile phone of brand C / , / order number / # / 1234**. Then perform part-of-speech tagging. The tagged text information includes: "Zhang San (person's name)", "((punctuation mark)", "138****5678 (number)", ") (punctuation mark)", "City A (place name)", "District Y (place name)", "a certain (pronoun)", "building (noun)", "buy (verb)", "brand C (brand name)", ", (punctuation mark)", "order number (noun)", "# (special symbol)", "1234** (number)", ". (punctuation mark)". Then remove stop words and retain keywords. Determine the stop word set. For example, some examples of stop words are: of, already, in, and, for, is, have, at, etc., just. Perform the steps of removing stop words and retaining keywords. The original text (already cleaned and segmented) includes: Zhang San (138****5678) bought a mobile phone of brand C in a building in District Y, City A. Order number: #1234**. After removing stop words, the text includes: Zhang San (138****5678) bought a mobile phone of brand C in a building in District Y, City A. Order number: #1234**. Among them, the retained keywords (such as "buy", "order number") are not removed because they are related to sensitive information identification. In actual operation, the stop word list may be adjusted according to specific scenarios to ensure that key context cue words are not accidentally deleted.

[0107] After the above text preprocessing steps for various sensitive information identification scenarios, the original text has been transformed into a feature vector suitable for machine learning models, and the words related to sensitive information have been specifically retained, providing effective data support for the subsequent sensitive information identification process.

[0108] Through the above text preprocessing steps for various sensitive information identification scenarios, not only irrelevant non-text content has been removed, Chinese word segmentation and part-of-speech tagging have been completed, but also special attention has been paid to the words related to various sensitive information, general stop words have been reasonably removed, and at the same time, key context cue words have been retained, laying a solid data foundation for the subsequent model learning and sensitive information identification steps.

[0109] In a feasible embodiment, the step of determining the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level includes:

[0110] Step S31, traversing the sensitivity levels of each sensitive information, and counting the frequencies corresponding to each sensitivity level;

[0111] In the embodiment of the present application, for each sensitive information of a sensitive category, all the sensitive information instances initially identified are traversed, and the number of times each type appears in the data set is recorded. Specifically, a hash table or a database index can be used to quickly count the frequency of each sensitive information type.

[0112] For example, if 5 ID numbers, 3 telephone numbers, and 10 email addresses are identified, the number of these different categories of sensitive information is recorded respectively, and then the frequencies corresponding to these sensitivity levels are further updated according to the sensitivity levels corresponding to these categories of sensitive information.

[0113] Exemplarily, an empty hash table (dictionary) is initialized, the key represents the sensitive information type, and the value represents the number of times the sensitive information of this type appears. The steps corresponding to the above statistical method may include: assuming that there is the following sensitive information list, each element is a tuple containing the sensitive information type and its occurrence, including: ID number "12345*****12345678", telephone number "138****5678", ID number "98765*****12345678", email "*****@example.com", then traverse the sensitive information list, use the hash table to count the number of times each type of information appears, and finally output the statistical results.

[0114] In the above example, when a new sensitive information type is encountered, if the type is already in the hash table, the count is increased; if not, a new key is added to the hash table and the count is initialized to 1. Finally, the hash table sensitive_info_counts stores the frequency of occurrence of various types of sensitive information.

[0115] Then, according to the aforementioned Table 3, the sensitivity level corresponding to each sensitive category is queried to obtain the frequency corresponding to each sensitivity level.

[0116] Step S32, calculating the product of each sensitivity level and the corresponding frequency, and obtaining the level number corresponding to each sensitivity level;

[0117] Step S33, calculating the sum of the level numbers corresponding to each sensitivity level, and determining the level number of the original data;

[0118] The calculation method of the level number of each sensitivity level may be: level number = sensitivity level * corresponding frequency.

[0119] For example, if the frequency of occurrence of sensitive information with a sensitivity level of 1 is 5, the corresponding level number is 1*5=5; if the frequency of occurrence of sensitive information with a sensitivity level of 2 is 6, the corresponding level number is 2*6=12, and so on.

[0120] When calculating the level number of the original data, the sum of the level numbers of each sensitivity level is calculated. Taking four sensitivity levels as an example, if the level number of sensitivity level 1 is 5, the level number of sensitivity level 2 is 12, the level number of sensitivity level 3 is 18, and the level number of sensitivity level 4 is 16, then the level number of the original data is 5+12+18+16=51.

[0121] Step S34, determining the risk level corresponding to the original data according to the correspondence between the preset level number and the risk level and the level number of the original data.

[0122] In the embodiment of the present application, the risk level is defined as 4 levels, and the corresponding relationship between the risk level and the definition description is described in Table 4 below.

[0123] Table 4: Risk level table

[0124] Risk Level Definition or description Very high The most serious risk high Less serious risk middle Serious risks Low Mild impact

[0125] In addition, there is a positive correlation between the level number and the risk level. For example, the corresponding relationship between the level number and the risk level includes: a level number between 0-10 belongs to low risk, a level number between 11-100 belongs to medium risk, a level number between 101-1000 belongs to high risk, and a level number above 1000 belongs to extremely high risk.

[0126] In another feasible embodiment, before the step of inputting the target text information corresponding to the original data into a preset target sensitive information recognition model, the method further includes:

[0127] Step B10, collecting training data within a preset range;

[0128] Step B20, preprocessing the training data to obtain processed training text information, the preprocessing process at least includes data cleaning, standardization, part-of-speech tagging, and deletion of stop words;

[0129] Step B30, obtaining the sensitive category of the sensitive information in the training text information, and obtaining sensitive category label information;

[0130] Step B40, training the pre-built initial sensitive information recognition model according to the training text information and the sensitive category label information;

[0131] Step B50, when the harmonic mean of the precision and recall of the initial sensitive information recognition model is higher than a preset threshold, stop model training and set the initial sensitive information recognition model as the target sensitive information recognition model.

[0132] The embodiment of the present application also provides a method for pre-training a sensitive information recognition model, which first requires the collection of training data for model training, wherein the preset range is data involving various sensitive information, and the authorization and permission of the data source or data provider are obtained. Similarly, after the training data is collected, the training data needs to be pre-processed, and the pre-processing process is the same as the pre-processing process described in step A10 to step A50, and can be referred to the aforementioned content, and will not be repeated here.

[0133] In addition, in the step of obtaining the sensitive categories of sensitive information in the training text information, the data labeling standards can be specified through expert experience, and professional labelers can label the data to obtain the sensitive categories corresponding to each sensitive information in the training text information. In the data labeling process, it can be performed by a tool specially designed to identify and label potential sensitive information (such as an initial sensitive information recognition model to be trained). Through their experience and professional knowledge, the staff can train and adjust the model to ensure that it can accurately identify various types of sensitive information. When using this model, the staff can verify and correct the output of the model according to the pre-set standards to further improve the accuracy of risk detection. Through the risk detection model, sensitive information in a large amount of data can be quickly identified, which greatly improves the efficiency of data processing. With the assistance of the model, professional labelers can focus on verifying and refining the recognition results of the model instead of labeling from scratch, thereby effectively saving time and resources. The obtained sensitive category label information is used as the true label of the training text information, which is convenient for evaluating the loss value, precision rate and other performance of the model during the training process, so as to obtain the target sensitive information recognition model with the final performance meeting the requirements.

[0134] In another feasible embodiment, the sensitive category label information obtained after annotation can be audited, such as formulating clear audit standards and processes, including but not limited to data accuracy, consistency, completeness and compliance. The audit standards should be consistent with the data annotation standards and be able to adapt to the ever-changing legal and regulatory requirements. The audit can be conducted in the form of manual review or automated tools, such as using regular expressions to check the data format, or using existing risk detection models to screen out possible incorrect annotations. This can improve audit efficiency and reduce human errors. Based on the audit results and feedback, the audit process and annotation standards are continuously optimized. Through continuous improvement, the quality of data annotation and audit can be improved to ensure the accuracy and reliability of the data.

[0135] Furthermore, after obtaining the training text information and its corresponding sensitive category label information, the pre-built initial sensitive information recognition model can be trained. Specifically, in the embodiment of the present application, the sensitive information recognition task is converted into a named entity recognition task in the NLP (Natural Language Processing) task, thereby forming a recognition modeling solution for sensitive information.

[0136] During the training process, the training text information can be divided into a training set, a validation set (for model parameter adjustment), a test set (for testing model performance), etc. Multiple text information in the training set are sequentially input into the initial sensitive information recognition model. After the initial sensitive information recognition model outputs the corresponding prediction results, the corresponding loss value is calculated. The model parameters of the model are iteratively optimized with the minimization of the loss value as the optimization goal. Finally, the harmonic mean of the precision and recall rate is higher than the preset threshold as the condition for stopping the model training, and the target sensitive information recognition model with performance that meets the requirements is obtained. Among them, the harmonic mean of the precision and recall rate is the F1 score, which is jointly determined by the precision and recall rate, and the precision and recall rate are both positively correlated with the F1 score.

[0137] Specifically, precision indicates the proportion of entities recognized by the model that are truly positive examples. It focuses on the reliability of the model's prediction results, that is, how many of the instances predicted by the model to be a certain entity type are actually correct instances of that type. Recall, also known as sensitivity, indicates the proportion of entities that actually exist that are successfully recognized by the model. It measures the comprehensiveness of the model's recognition ability, that is, how many of all actually existing entities are correctly found by the model. The F1 score is the harmonic mean of the precision and recall rates. It takes into account both the accuracy and comprehensiveness of the model and aims to provide a single comprehensive evaluation indicator. The F1 score reaches its maximum value when the precision and recall rates are equal, and when either of them approaches 0, the F1 score will also approach 0.

[0138] A high precision rate means that the vast majority of sensitive information (such as name, mobile phone number, ID number, email address, and home address) identified by the model is real sensitive information, and there are few cases of misidentification. A high recall rate means that most of the sensitive information actually existing in the text is successfully identified by the model, and there are few omissions.

[0139] In summary, the model shows excellent accuracy and comprehensiveness in identifying sensitive information, and can efficiently find various types of sensitive information from texts with relatively few false positives and false negatives. This level of performance has high effectiveness and reliability in practical applications.

[0140] In another feasible embodiment, a small batch of data can be used for testing first to iterate the parameters of the initial sensitive information recognition model. At the same time, a control group Bert+CRF model can be added for control experiments. For performance comparison and evaluation, the UIE model is used as the experimental group and the Bert+CRF model is used as the control group. The advantages and disadvantages of the two models are evaluated by comparing their performance on the same data set.

[0141] Among them, BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained model based on Transformer, which can effectively extract and process semantic information in text. In traditional natural language processing tasks, it is usually necessary to retrain the model for specific tasks. However, BERT can make the model have broad application prospects through pre-training. In addition, CRF (Conditional RandomField) is a model based on conditional random fields, which can consider the features in the text in an orderly manner, thereby realizing accurate recognition of new entities. In the field of Internet finance, new entities usually appear in specific contexts, and the CRF model can identify and classify new entities based on contextual information.

[0142] In the BERT+CRF model proposed in the embodiment of the present application, the BERT model is first used to extract features from the text data, and then the extracted features are input into the CRF model to identify new entities. Specifically, the BERT model is used to regularize the text and extract context features that can reflect the semantic information in the text. These features are then input into the CRF model, and the CRF model is used to identify and classify new entities.

[0143] For example, after using 10,000 pieces of finely annotated data to fine-tune the pre-trained model, the performance of the model has been significantly improved, so that its F1 score on the entity recognition task has reached 90 points. Referring to the comparison of the F1 scores of various models and the UIE model shown in Table 5 below, it can be seen that UIE is more sensitive to sensitive information recognition than the BERT+CRF model, and the final F1 score (F1_entity is 96.57) is higher than the F1 score of the BERT+CRF model (96.05, 96.08 and 96.13), and the effect is better. It has high effectiveness and reliability in practical applications.

[0144] Table 5:

[0145]

[0146] Among them, solution is the solution, epoch is the training round, f1_token refers to the F1 score calculation based on the token level, f1_entity refers to the F1 score calculation based on the entity level, and comment refers to the comment content. Bert+crf+init is used to represent the model after the CRF weights are initialized with training data, and Bert+crf+freeze is used to represent the model after the CRF weights are generated by the training data.

[0147] Furthermore, in a feasible embodiment, the step of training the pre-built initial sensitive information recognition model according to the training text information and the sensitive category label information includes:

[0148] Step B41, inputting the trained text information into an initial sensitive information recognition model, and the initial sensitive information recognition model outputs a corresponding recognition result, wherein the recognition result includes sensitive information and sensitive categories in the training text information;

[0149] Step B42, iteratively optimizing the model parameters of the initial sensitive information recognition model based on the sensitive category label information corresponding to the training text information and the sensitive information and sensitive category corresponding to the recognition result;

[0150] Step B43, determining the number of correct instances, the number of misjudged instances, and the number of missed instances based on the sensitive category label information corresponding to the training text information and the sensitive information and sensitive categories corresponding to the recognition result;

[0151] Step B44, calculating the accuracy of the initial sensitive information recognition model according to the number of correct instances and the number of misjudged instances;

[0152] Step B45, calculating the recall rate of the initial sensitive information recognition model based on the number of correct instances and the number of missed instances.

[0153] In the embodiment of the present application, the Word2Vec (word to vector, a related model for generating word vectors) model can be used to embed the preprocessed text data to generate a high-dimensional feature vector at the word level to capture the semantic relationship between words. Specifically, word2vec is a process of mapping a sparse word vector in the form of one-hot into an n-dimensional (n is generally several hundred) dense vector using a one-layer neural network (i.e., CBOW, continuous bag of words model). In order to speed up the model training speed, tricks (optimization techniques) include the use of Hierarchical softmax (a Huffman tree structure), negative sampling (positive and negative sample construction), Huffman Tree, etc.

[0154] The Word2Vec model uses a layer of neural network to map one-hot vectors to dense word vectors. In order to improve training efficiency and model performance, the following techniques and strategies are used: Since the traditional softmax function has high computational complexity, because for each word, it is necessary to calculate its similarity with all other words. So Hierarchical Softmax reduces the amount of calculation by building a binary tree. Each internal node represents a word, and the leaf node represents the output of the word. In this way, for each word, it is only necessary to calculate its similarity with its parent node and two child nodes. In addition, during the training process, in order to avoid calculating all possible word pairs, negative sampling technology can be used. Specifically, negative sampling randomly selects some words that are not related to the target word as negative samples, and only predicts these negative samples. This method can not only reduce the amount of calculation, but also improve the generalization ability of the model. In addition, Huffman Tree is similar to Hierarchical Softmax and is also a technology for reducing computational complexity. By building a Huffman tree, words can be mapped to a binary tree, thereby reducing the amount of calculation when calculating probabilities.

[0155] Furthermore, in the field of natural language processing (NLP), words constitute the basic unit of processing. When embarking on the task of part-of-speech tagging, a series of samples are usually used for learning, which consist of words (x) and their part-of-speech tags (y). The learning goal is to establish a mapping from words to parts of speech, that is, x->y (see Table 1 and Table 2 for specific mapping relationships). It can be understood that traditional mapping establishment methods cover algorithms such as Bayesian classifiers and support vector machines (SVM). However, these algorithms usually require the input to be of numerical type, while words in NLP are essentially a symbolic representation of human language, whether in Chinese, English or Latin. Therefore, it is necessary to convert these symbols into numerical values, or more precisely, embed them into a mathematical space. This conversion process is called word embedding, and the Word2vec model mentioned above is a feasible method of word embedding technology.

[0156] In the specific application of NLP, x can be regarded as a specific word in a sentence, and y is other words in the context of the word. Then the mapping function f here corresponds to the common "language model" in NLP. The core task of this model is to evaluate the sample (x, y) to determine whether it conforms to the rules of natural language. In other words, the language model aims to determine whether the combination of words x and y constitutes a smooth and reasonable language expression.

[0157] Furthermore, in the process of model selection and initialization, the UIE model can be selected as the basic model. The deep learning-based natural language processing framework, such as BERT, RoBERTa, ALBERT and other pre-trained language models, has strong language understanding capabilities and rich context information capture capabilities, and is suitable for various information extraction tasks. Then fine-tuning the selected UIE model to adapt to the sensitive information identification task, load the pre-processed text data (i.e., training text information) and the corresponding sensitive information labels into the model, and ensure that the data format is consistent with the model input requirements.

[0158] After receiving the input training text information, the model outputs the corresponding sensitive information recognition results. Specifically, for the input text data, the model outputs the sensitive information category label corresponding to each word (or word fragment), such as "name", "mobile phone number", "ID number", "email address", "home address", etc. These labels identify the potential location and category of sensitive information in the text, which facilitates the subsequent sensitive information management and implementation of protection measures.

[0159] Exemplarily, data is transferred through standardized data formats (such as CSV, JSON, etc.) to ensure that the preprocessed text data and annotated data can be smoothly input into the UIE model for training and reasoning. The output of the model training phase (such as the model weight file) is used to save the trained model state for subsequent verification, evaluation, and deployment. In actual applications, after the model is deployed, it receives raw text input without preprocessing, automatically executes the preprocessing process internally, and then uses the trained UIE model to identify sensitive information, and finally returns the recognition result.

[0160] After determining the recognition result, we can compare the various entities recognized by the model to see which ones are positive examples that are consistent with the sensitive category label information (i.e., the correct sensitive category is assigned to the entity), which ones are misidentified (i.e., the wrong sensitive category is assigned to the entity), and which ones are omitted (i.e., sensitive information that should have been assigned a sensitive category is not assigned a sensitive category), so as to determine the TP (True Positives, the number of instances correctly recognized as a certain entity type by the model, i.e., the number of correct instances), FP (False Positives, the number of instances of a certain entity type that are not of this type that are mistakenly judged by the model, i.e., the number of misjudged instances), and FN (False Negatives, the number of instances that are actually a certain entity type but are omitted (i.e., not recognized) by the model, i.e., the number of omitted instances) in this recognition process of the model.

[0161] For example, the calculation formula of precision is as follows:

[0162]

[0163] The calculation formula of recall is as follows:

[0164]

[0165] Furthermore, the calculation formula of F1 score is as follows:

[0166]

[0167] According to Table 4, after fine-tuning the initial sensitive information identification model (UIE) with 10,000 finely labeled data, the F1 score of 90 points was obtained in the entity recognition task (i.e., the task of sensitive information identification of training text information). This means that the precision and recall of the model are both very high and close, and the F1 score reaches the maximum value when the two are balanced, and the model training process can be ended to obtain the target sensitive information recognition model.

[0168] For ease of understanding, based on the aforementioned application embodiments, a feasible implementation step of the sensitive data risk audit method is as follows: Figure 3 As shown, it specifically includes: first input data such as {data: "Name: Zhang San, Gender: Male, ID Number 11010*****08088888"}, then perform data cleaning and word segmentation, and the processing result is: ['Name' 'Zhang San', 'Gender' 'Male', 'ID Number' '11010*****08088888'], and then perform vectorization processing to obtain a 36*36 matrix, which is input into the sensitive information recognition model for model prediction. If the data has been desensitized, the process ends; if the data has not been desensitized, the output The non-massified type ['Name', 'Zhang San', 'Gender', 'Male', 'ID Number', '11010*****08088888'][0,0,A11,A11,A11,0,0,A12,0,0,0,0,A13,13,13,13,13,13,13] contains sensitive data types A11, A12, and A13. The risk level of the masks is further determined to be low risk, and the process ends if it is medium, high, or extremely high risk. A prompt or warning is issued to remind the staff to perform masks on this batch of data in a timely manner.

[0169] In the embodiment of the present application, the Word2Vev model is used to pre-process data through word embedding technology combined with pre-set rules. Word embedding is a way to convert a single word into a digital representation (vector) of a word. Each word is mapped to a vector. Then, based on the UIE model, a sensitive information recognition model is constructed and trained to identify sensitive information of pre-processed text information, which solves the problem of low efficiency and low accuracy of traditional solutions in auditing and identifying sensitive information.

[0170] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the sensitive data risk audit method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0171] This application also provides a sensitive data risk audit device, referring to Figure 4 , the sensitive data risk audit device comprises:

[0172] The sensitive category identification module 10 is used to input the target text information corresponding to the original data into a preset target sensitive information identification model, and predict at least one sensitive information in the target text information and the sensitive category corresponding to each sensitive information through the target sensitive information identification model;

[0173] A sensitivity level determination module 20, which determines the sensitivity level corresponding to each of the sensitive information according to the sensitivity category corresponding to each of the sensitive information and a preset sensitivity level rule, wherein the sensitivity level rule is used to characterize the correspondence between the sensitivity category and the sensitivity level;

[0174] The risk level determination module 30 determines the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level.

[0175] In one embodiment, the risk level determination module 30 is further configured to:

[0176] Traversing the sensitivity levels of each of the sensitive information, and counting the frequencies corresponding to each of the sensitivity levels;

[0177] Calculate the product of each sensitivity level and the corresponding frequency to obtain the level number corresponding to each sensitivity level;

[0178] Calculate the sum of the level numbers corresponding to each of the sensitivity levels to determine the level number of the original data;

[0179] The risk level corresponding to the original data is determined according to the corresponding relationship between the preset level number and the risk level and the level number of the original data.

[0180] In one embodiment, the sensitive data risk audit device further includes a model training module, wherein the model training module is used to:

[0181] Collect training data within a preset range;

[0182] Preprocessing the training data to obtain processed training text information, wherein the preprocessing process at least includes data cleaning, standardization, part-of-speech tagging, and deletion of stop words;

[0183] Obtaining sensitive categories of sensitive information in the training text information to obtain sensitive category label information;

[0184] Training a pre-built initial sensitive information recognition model according to the training text information and the sensitive category label information;

[0185] When the harmonic mean of the precision and recall of the initial sensitive information recognition model is higher than a preset threshold, the model training is stopped and the initial sensitive information recognition model is set as the target sensitive information recognition model.

[0186] In one embodiment, the model training module is further used to:

[0187] Input the training text information into the initial sensitive information recognition model, and the initial sensitive information recognition model outputs a corresponding recognition result, wherein the recognition result includes the sensitive information and sensitive categories in the training text information;

[0188] Iteratively optimizing model parameters of the initial sensitive information recognition model based on the sensitive category label information corresponding to the training text information and the sensitive information and sensitive category corresponding to the recognition result;

[0189] Determine the number of correct instances, the number of misjudged instances, and the number of missed instances according to the sensitive category label information corresponding to the training text information and the sensitive information and sensitive category corresponding to the recognition result;

[0190] Calculating the accuracy of the initial sensitive information recognition model according to the number of correct instances and the number of misjudged instances;

[0191] The recall rate of the initial sensitive information identification model is calculated based on the number of correct instances and the number of missed instances.

[0192] In one embodiment, the sensitive data risk audit device further includes a preprocessing module, and the preprocessing module is used to:

[0193] When the original data is text data, deleting non-text content in the original data to obtain original text information, wherein the non-text content includes at least one of a hypertext markup language tag, a non-text character, and a blank character;

[0194] When the original data is non-text data, parsing the original data to obtain original text information;

[0195] Based on a predefined dictionary, the original text information is segmented to obtain each word included in the original text information, wherein the predefined dictionary includes predefined sensitive words and proprietary words;

[0196] By using a preset word segmentation tool or a pre-trained part-of-speech tagging model, a part-of-speech tag is added to each word in the original text information to obtain annotated text information, wherein the part-of-speech tag includes at least a name, a verb, and an adjective;

[0197] Based on a preset stop word list, the stop words in the annotated text information are deleted to obtain the target text information.

[0198] In one embodiment, the risk level determination module 30 is further configured to:

[0199] A risk audit report is generated based on the sensitive information corresponding to the original data, the sensitivity levels corresponding to the sensitive information, the frequency of the sensitivity levels, the risk levels, and the frequency of the sensitivity levels corresponding to the multiple sensitive categories in the original data.

[0200] The sensitive data risk audit device provided by the present application adopts the sensitive data risk audit method in the above embodiment, which can solve the technical problem of low accuracy and efficiency of the current sensitive information identification strategy. Compared with the prior art, the beneficial effects of the sensitive data risk audit device provided by the present application are the same as the beneficial effects of the sensitive data risk audit method provided by the above embodiment, and the other technical features in the sensitive data risk audit device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0201] The present application provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the sensitive data risk audit method in the above-mentioned embodiment one.

[0202] Reference below Figure 5 , which shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present application. The electronic devices in the embodiments of the present application may include but are not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The electronic device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0203] like Figure 5As shown, the electronic device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the electronic device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows an electronic device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.

[0204] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0205] The electronic device provided by the present application adopts the sensitive data risk audit method in the above embodiment, which can solve the technical problem of low accuracy and efficiency of the current sensitive information identification strategy. Compared with the prior art, the beneficial effects of the electronic device provided by the present application are the same as the beneficial effects of the sensitive data risk audit method provided by the above embodiment, and the other technical features in the electronic device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0206] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0207] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0208] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, wherein the computer-readable program instructions are used to execute the sensitive data risk audit method in the above-mentioned embodiment.

[0209] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0210] The computer-readable storage medium may be included in the electronic device, or may exist independently without being installed in the electronic device.

[0211] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by an electronic device, the electronic device: inputs the target text information corresponding to the original data into a preset target sensitive information recognition model, and predicts at least one sensitive information in the target text information and the sensitive category corresponding to each of the sensitive information through the target sensitive information recognition model; determines the sensitivity level corresponding to each of the sensitive information according to the sensitive category corresponding to each of the sensitive information and the preset sensitivity level rule, wherein the sensitivity level rule is used to characterize the correspondence between the sensitive category and the sensitivity level; determines the risk level of the original data based on the sensitivity level of each of the sensitive information and the frequency of each of the sensitive levels.

[0212] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0213] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0214] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0215] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned sensitive data risk audit method, and can solve the technical problem of low accuracy and efficiency of the current sensitive information identification strategy. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the sensitive data risk audit method provided in the above-mentioned embodiment, and will not be repeated here.

[0216] The present application also provides a computer program product, including a computer program, which implements the steps of the sensitive data risk audit method as described above when executed by a processor.

[0217] The computer program product provided by this application can solve the technical problem of low accuracy and efficiency of the current sensitive information identification strategy. Compared with the prior art, the beneficial effects of the computer program product provided by this application are the same as the beneficial effects of the sensitive data risk audit method provided by the above embodiment, which will not be repeated here.

[0218] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A sensitive data risk audit method, characterized in that: The sensitive data risk audit method includes: Inputting the target text information corresponding to the original data into a preset target sensitive information recognition model, and predicting at least one sensitive information in the target text information and the sensitive categories corresponding to each sensitive information through the target sensitive information recognition model; Determine the sensitivity level corresponding to each of the sensitive information according to the sensitivity category corresponding to each of the sensitive information and the preset sensitivity level rule, wherein the sensitivity level rule is used to characterize the correspondence between the sensitive category and the sensitivity level; Based on the sensitivity level of each sensitive information and the frequency of each sensitivity level, the risk level of the original data is determined.

2. The sensitive data risk audit method according to claim 1, characterized in that: The step of determining the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level comprises: Traversing the sensitivity levels of each of the sensitive information, and counting the frequencies corresponding to each of the sensitivity levels; Calculate the product of each sensitivity level and the corresponding frequency to obtain the level number corresponding to each sensitivity level; Calculate the sum of the level numbers corresponding to each of the sensitivity levels to determine the level number of the original data; The risk level corresponding to the original data is determined according to the corresponding relationship between the preset level number and the risk level and the level number of the original data.

3. The sensitive data risk audit method according to claim 1, characterized in that: Before the step of inputting the target text information corresponding to the original data into a preset target sensitive information recognition model, the method further includes: Collect training data within a preset range; Preprocessing the training data to obtain processed training text information, wherein the preprocessing process at least includes data cleaning, standardization, part-of-speech tagging, and deletion of stop words; Obtaining sensitive categories of sensitive information in the training text information to obtain sensitive category label information; Training a pre-built initial sensitive information recognition model according to the training text information and the sensitive category label information; When the harmonic mean of the precision and recall of the initial sensitive information recognition model is higher than a preset threshold, the model training is stopped and the initial sensitive information recognition model is set as the target sensitive information recognition model.

4. The sensitive data risk audit method according to claim 3, characterized in that: The step of training the pre-built initial sensitive information recognition model according to the training text information and the sensitive category label information includes: Input the training text information into the initial sensitive information recognition model, and the initial sensitive information recognition model outputs a corresponding recognition result, wherein the recognition result includes the sensitive information and sensitive categories in the training text information; Iteratively optimizing model parameters of the initial sensitive information recognition model based on the sensitive category label information corresponding to the training text information and the sensitive information and sensitive category corresponding to the recognition result; Determine the number of correct instances, the number of misjudged instances, and the number of missed instances according to the sensitive category label information corresponding to the training text information and the sensitive information and sensitive category corresponding to the recognition result; Calculating the accuracy of the initial sensitive information recognition model according to the number of correct instances and the number of misjudged instances; The recall rate of the initial sensitive information identification model is calculated based on the number of correct instances and the number of missed instances.

5. The sensitive data risk audit method according to claim 1, characterized in that: Before the step of inputting the target text information corresponding to the original data into a preset target sensitive information recognition model, the method further includes: When the original data is text data, deleting non-text content in the original data to obtain original text information, wherein the non-text content includes at least one of a hypertext markup language tag, a non-text character, and a blank character; When the original data is non-text data, parsing the original data to obtain original text information; Based on a predefined dictionary, the original text information is segmented to obtain each word included in the original text information, wherein the predefined dictionary includes predefined sensitive words and proprietary words; By using a preset word segmentation tool or a pre-trained part-of-speech tagging model, a part-of-speech tag is added to each word in the original text information to obtain annotated text information, wherein the part-of-speech tag includes at least a name, a verb, and an adjective; Based on a preset stop word list, the stop words in the annotated text information are deleted to obtain the target text information.

6. The sensitive data risk audit method according to any one of claims 1 to 5, characterized in that: After the step of determining the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level, the method further includes: A risk audit report is generated based on the sensitive information corresponding to the original data, the sensitivity levels corresponding to the sensitive information, the frequencies of the sensitivity levels, the risk levels, and the frequencies of the sensitivity levels corresponding to the multiple sensitive categories in the original data.

7. A sensitive data risk audit device, characterized in that: The sensitive data risk audit device comprises: A sensitive category identification module, used to input the target text information corresponding to the original data into a preset target sensitive information identification model, and predict at least one sensitive information in the target text information and the sensitive category corresponding to each sensitive information through the target sensitive information identification model; a sensitivity level determination module, which determines the sensitivity level corresponding to each of the sensitive information according to the sensitivity category corresponding to each of the sensitive information and a preset sensitivity level rule, wherein the sensitivity level rule is used to characterize the correspondence between the sensitivity category and the sensitivity level; The risk level determination module determines the risk level of the original data based on the sensitivity level of each sensitive information and the frequency of each sensitivity level.

8. An electronic device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the sensitive data risk audit method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the sensitive data risk audit method according to any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the sensitive data risk audit method according to any one of claims 1 to 6 are implemented.

Images