Data desensitization processing method and device

Abstract

According to the data desensitization processing method and device, to-be-desensitized data are obtained, logical coherence desensitization is carried out on the to-be-desensitized data through a pre-constructed logical coherence desensitization model, desensitization corpus is obtained, logical coherence desensitization is desensitization retaining the logical relation between data items, and the desensitization corpus is used for desensitization of the logical relation between the data items. The logical relationship at least comprises a calculation relationship, a regional relationship, a time relationship and a comparison relationship. By means of the scheme, due to the fact that the context logic relation exists among the multiple data items, in order to keep the logic relation of desensitization during desensitization, logic coherence desensitization is conducted on the data to be desensitized through the pre-constructed logic coherence desensitization model, and the desensitization efficiency is improved. The logical coherence desensitization refers to desensitization of logical relationships such as calculation relationships, regional relationships, time relationships and comparison relationships among the data items, and the purpose of obtaining desensitization corpora with the logical relationships reserved in the data desensitization process is achieved.

Description

Technical Field

[0001] The present application relates to the technical field of data desensitization processing, and more specifically, to a data desensitization processing method and device. Background Art

[0002] Data desensitization refers to the deformation of certain sensitive information through desensitization rules to achieve reliable protection of sensitive privacy data. In this way, the desensitized real data sets can be safely used in development, testing and other non-production environments as well as outsourcing environments.

[0003] Existing data desensitization technology is usually simulated desensitization. Simulated desensitization achieves a balance between privacy protection and data availability by generating synthetic data similar to the data item, and is an effective method for desensitizing sensitive data items.

[0004] However, simulation masking only satisfies the simulation problem after masking of a single data item, and has no ability to maintain logical coherence, that is, it has no ability to perform masking in the presence of contextual logical relationships.

[0005] Therefore, how to retain logical relationships in the process of desensitizing data is an urgent problem to be solved in this application. Summary of the invention

[0006] In view of this, the present application discloses a data desensitization processing method and device, which aims to perform logical coherence desensitization on the data to be desensitized through a pre-constructed logical coherence desensitization model, so as to achieve the purpose of obtaining desensitized corpus that retains logical relationships in the process of desensitizing the data.

[0007] In order to achieve the above purpose, the disclosed technical solution is as follows:

[0008] The first aspect of the present application discloses a data desensitization processing method, the method comprising:

[0009] Obtain the data to be desensitized;

[0010] Performing logical coherence desensitization on the data to be desensitized by using a pre-built logical coherence desensitization model to obtain desensitized corpus;

[0011] Among them, the logical coherence desensitization is a desensitization that retains the logical relationship between each data item; the logical relationship at least includes calculation relationship, geographical relationship, time relationship and comparison relationship.

[0012] Preferably, the method of performing logical coherence desensitization on the data to be desensitized by using a pre-constructed logical coherence desensitization model to obtain desensitized corpus includes:

[0013] Obtain sensitive data items in the data to be desensitized through a pre-built logical coherence desensitization model;

[0014] Determine the logical relationship to be matched of the sensitive data item in the context through the prompt word requirements in the logical coherence desensitization model; wherein the prompt word requirements include desensitization requirements and output format requirements;

[0015] According to the to-be-matched logical relationship, the logical relationship is matched to obtain a matching result; the matching result is a matching result that the to-be-matched logical relationship is consistent with the logical relationship;

[0016] According to the matching results, the sensitive data items are desensitized for logical coherence to obtain desensitized corpus.

[0017] Preferably, it also includes:

[0018] When receiving an output requirement, parsing the requirement type of the output requirement; wherein the requirement type at least includes a requirement type that conforms to a calculation relationship, a requirement type that conforms to a geographical relationship, a requirement type that conforms to a time relationship, and / or a requirement type that conforms to a comparison relationship;

[0019] According to the logical consistency desensitization model and the requirement type, the data to be desensitized is subjected to logical consistency desensitization.

[0020] Preferably, after performing logical coherence desensitization on the data to be desensitized by using a pre-built logical coherence desensitization model to obtain desensitized corpus, the method further includes:

[0021] The desensitized corpus is used to train the vertical domain big model to avoid the situation where the vertical domain big model leaks sensitive data items.

[0022] Preferably, it also includes:

[0023] In the process of performing logical coherence desensitization on the data to be desensitized, when all data items related to logical relationships in the data to be desensitized are desensitized correctly, the correctly desensitized logical relationships are marked.

[0024] The second aspect of the present application discloses a data desensitization processing device, the device comprising:

[0025] An acquisition unit, used to acquire the data to be desensitized;

[0026] The first desensitization unit is used to perform logical coherence desensitization on the data to be desensitized by using a pre-constructed logical coherence desensitization model to obtain a desensitized corpus; wherein the logical coherence desensitization is a desensitization that retains the logical relationship between each data item; the logical relationship includes at least a calculation relationship, a geographical relationship, a time relationship and a comparison relationship.

[0027] Preferably, the first desensitization unit comprises:

[0028] An acquisition module is used to acquire sensitive data items in the data to be desensitized through a pre-built logical coherence desensitization model;

[0029] A determination module, used to determine the logical relationship to be matched of the sensitive data item in the context through the prompt word requirements in the logical coherence desensitization model; wherein the prompt word requirements include desensitization requirements and output format requirements;

[0030] A matching module, used for matching the to-be-matched logical relationship with the logical relationship to obtain a matching result; the matching result is a matching result that the to-be-matched logical relationship is consistent with the logical relationship;

[0031] The desensitization module is used to perform logical coherence desensitization on the sensitive data items according to the matching results to obtain desensitized corpus.

[0032] Preferably, it also includes:

[0033] A parsing unit, configured to parse the requirement type of the output requirement upon receiving the output requirement; wherein the requirement type at least includes a requirement type that conforms to a calculation relationship, a requirement type that conforms to a geographical relationship, a requirement type that conforms to a time relationship, and / or a requirement type that conforms to a comparison relationship;

[0034] The second desensitizing unit is used to perform logical consistency desensitization on the data to be desensitized according to the logical consistency desensitization model and the requirement type.

[0035] Preferably, it also includes:

[0036] The training unit is used to train the vertical domain big model with the desensitized corpus to avoid the vertical domain big model from leaking sensitive data items.

[0037] Preferably, it also includes:

[0038] The marking unit is used to mark the correctly desensitized logical relationship when all data items related to the logical relationship in the data to be desensitized are desensitized correctly during the process of performing logical coherence desensitization on the data to be desensitized.

[0039] Through the above technical scheme, it can be known that the present application discloses a data desensitization processing method and device, which obtains the data to be desensitized, and performs logical coherence desensitization on the data to be desensitized through a pre-constructed logical coherence desensitization model to obtain a desensitized corpus, wherein the logical coherence desensitization is a desensitization that retains the logical relationship between each data item, and the logical relationship includes at least a calculation relationship, a geographical relationship, a time relationship, and a comparative relationship. Through the above scheme, since there are contextual logical relationships between multiple data items, in order to maintain the desensitized logical relationship during desensitization, a pre-constructed logical coherence desensitization model is used to perform logical coherence desensitization on the data to be desensitized, and the logical coherence desensitization is a desensitization that retains the calculation relationship, geographical relationship, time relationship, comparative relationship, and other logical relationships between each data item, so as to achieve the purpose of obtaining a desensitized corpus that retains the logical relationship in the process of desensitizing the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] 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, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0041] Figure 1 A flowchart of a data desensitization processing method disclosed in an embodiment of the present application;

[0042] Figure 2 A schematic diagram comparing the logical coherence effects of simulation desensitization and model desensitization disclosed in the embodiments of the present application;

[0043] Figure 3 A schematic diagram of data desensitization processing disclosed in an embodiment of the present application;

[0044] Figure 4 A schematic diagram of the structure of a data desensitization processing device disclosed in an embodiment of the present application;

[0045] Figure 5 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present application. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0047] In this application, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element.

[0048] As can be seen from the background technology, however, simulation desensitization only satisfies the simulation problem after desensitization of a single data item, and has no ability to maintain logical coherence, that is, it has no ability to desensitize in the presence of contextual logical relationships. Therefore, how to retain logical relationships in the process of desensitizing data is a problem that needs to be solved urgently in this application. Therefore, how to retain logical relationships in the process of desensitizing data is a problem that needs to be solved urgently in this application.

[0049] In order to solve the above problems, the present application discloses a data desensitization processing method and device, obtains the data to be desensitized, and performs logical coherence desensitization on the data to be desensitized through a pre-constructed logical coherence desensitization model to obtain a desensitized corpus, wherein the logical coherence desensitization is a desensitization that retains the logical relationship between each data item, and the logical relationship includes at least a calculation relationship, a regional relationship, a time relationship, and a comparative relationship. Through the above scheme, since there is a contextual logical relationship between multiple data items, in order to maintain the desensitized logical relationship during desensitization, a pre-constructed logical coherence desensitization model is used to perform logical coherence desensitization on the data to be desensitized, and the logical coherence desensitization is a desensitization that retains the calculation relationship, regional relationship, time relationship, comparative relationship and other logical relationships between each data item, so as to achieve the purpose of obtaining a desensitized corpus that retains the logical relationship in the process of desensitizing the data. The specific implementation method is specifically described by the following embodiments.

[0050] refer to Figure 1 As shown, a data desensitization processing method disclosed in an embodiment of the present application is provided, and the data desensitization processing method mainly includes the following steps:

[0051] S101: Obtain data to be desensitized.

[0052] The data to be desensitized are data that have not been desensitized by the logical coherence desensitization model, that is, the original corpus.

[0053] S102: Perform logical consistency desensitization on the data to be desensitized by using a pre-built logical consistency desensitization model; wherein, logical consistency desensitization is desensitization that retains the logical relationship (i.e., logical coherence) between each data item; the logical relationship includes at least calculation relationship, geographical relationship, time relationship, and comparison relationship.

[0054] Due to the existence of contextual logical relationships between multiple data items in the corpus, it is particularly necessary to maintain the logical coherence of desensitization during desensitization. Only in this way can these logics be learned when training the vertical domain big model, so as to better serve the usage scenarios in specific fields, such as intelligent writing of judicial documents.

[0055] When training a vertical domain big model, it is essential to use corpus. Many of the corpus are not disclosed within the industry / organization and contain a large amount of sensitive information. Once leaked, it will cause very serious consequences. Therefore, we need to desensitize the corpus to prevent sensitive data items such as personal information from being leaked through the vertical domain big model. The most common hallucination problem of big models (the hallucination problem of big models is mainly manifested in the inconsistency between the generated content and real-world facts or user input) may be caused by errors in the logical relationship of the corpus. The corpus used for vertical domain big model training must first be desensitized, and ordinary simulation desensitization will destroy the logical relationship of the corpus. Therefore, this solution proposes to implement logical coherence desensitization of the corpus based on the big model, which can allow the vertical domain big model to learn as much of the internal logical relationship of the corpus as possible while avoiding the leakage of sensitive information, thereby reducing the hallucination problem of the vertical domain big model.

[0056] There are many types of logical coherence. This solution summarizes and defines four types of logical coherence (i.e., computational relationship, geographical relationship, temporal relationship, and comparative relationship). In the following experiments, we focus on whether these logical coherences are maintained before and after desensitization. We mainly compare two methods: simulation desensitization (existing technology) and logical coherence desensitization model desensitization (this solution).

[0057] Among them, the calculation relationship is the logical relationship of calculation between data items; the regional relationship indicates the logical relationship of inclusion between regions; the time relationship indicates the logical relationship of time sequence between data items; and the comparison relationship indicates the logical relationship of comparison between data items.

[0058] Specifically, the process of performing logical consistency desensitization on the data to be desensitized is performed through the pre-built logical consistency desensitization model, as shown in A1-A4.

[0059] A1: Obtain sensitive data items from the data to be desensitized through the pre-built logical coherence desensitization model.

[0060] Among them, sensitive data items include time, region, data comparison, etc.

[0061] A2: Determine the logical relationship to be matched in the context of sensitive data items through the prompt word requirements in the logical coherence desensitization model; the prompt word requirements include desensitization requirements and output format requirements.

[0062] For example, the sensitive data items in the data to be desensitized are Hong Kong, Yuen Long, and graduated from a public school in 1999. The logical relationships to be matched in the context are geographical relationship (Hong Kong, Yuen Long, a public school) and time relationship (1999).

[0063] A3: Match the logical relationship to be matched with the logical relationship to obtain a matching result; the matching result is a matching result in which the logical relationship to be matched is consistent with the logical relationship.

[0064] For example, according to the logical relationship to be matched being a geographical relationship and a time relationship, the matching result of the geographical relationship of the logical relationship to be matched matches the geographical relationship in the logical relationship, and the matching result of the time relationship of the logical relationship to be matched matches the time relationship in the logical relationship.

[0065] A4: Based on the matching results, perform logical coherence desensitization on sensitive data items.

[0066] For example, based on the matching results of the geographical relationship of the logical relationship to be matched and the geographical relationship in the logical relationship, and the matching results of the temporal relationship of the logical relationship to be matched and the temporal relationship in the logical relationship, the sensitive data items are desensitized for logical coherence to obtain desensitized corpus that retains the geographical relationship and the temporal relationship.

[0067] When an output requirement is received during the process of performing logical consistency masking on the data to be masked, the requirement type of the output requirement is parsed, and logical consistency masking is performed on the data to be masked according to the logical consistency masking model and the requirement type.

[0068] The requirement types include at least a requirement type that complies with a calculation relationship, a requirement type that complies with a geographical relationship, a requirement type that complies with a time relationship, and / or a requirement type that complies with a comparison relationship.

[0069] For other logical relationships outside of these four categories, the large model desensitization method only needs to add relevant requirements to the prompt output requirements to perform logical coherence desensitization for new logical relationships, such as: {requires compliance with **calculation relationships**, such as clear calculation relationships such as addition, subtraction, multiplication and division between data items, and maintains data item logic} requirement type. Large model desensitization has low threshold and high-efficiency scalability for new logical relationships.

[0070] An example of simulated desensitization (existing technology) is shown in Table 1:

[0071]

[0072] Table 1

[0073] It can be seen from the content in Table 1 that the simulation desensitization of the prior art only meets the simulation problem after desensitization of a single data item, and is unable to maintain logical coherence.

[0074] In order to facilitate understanding of the logical coherence of this solution, an example is given in conjunction with Table 2.

[0075]

[0076]

[0077] Table 2

[0078] As can be seen from Table 2, the model desensitization of this solution retains the logical relationship of calculation relationship, regional relationship, time relationship and comparison relationship. In addition, there is an inclusion relationship between regional relationships, thus avoiding the confusion of the relationship between regional relationships during the desensitization process.

[0079] A total of 67 samples were collected from the public network, which contain one or more logical relationship types. In the same sample, each logical relationship type may appear multiple times. The number of samples corresponding to the specific logical relationship type is shown in Table 3. The same logical relationship in the same sample is uniformly recorded as 1. In the process of logical coherence desensitization of the desensitized data, when all data items related to the logical relationship in the desensitized data are desensitized correctly, the correctly desensitized logical relationship is marked. For example, in a certain sample, if and only if all data items related to the calculation relationship are desensitized correctly, the logical relationship is marked as successfully maintained, otherwise it is marked as failed. The total number of all logical relationships is 106.

[0080] Logical relationship type The number of samples Total number of logical relationships Calculation relationship 23 23 Regional Relationship 20 20 Time relationship 42 42 Contrast 21 21

[0081] Table 3

[0082] In the experiment in Table 3, a logical coherence desensitization experiment was conducted on 67 samples containing logical relationships. The logical relationships mainly include calculation relationships, regional relationships, time relationships, and comparison relationships. These four types of logical relationships will serve as a comparison benchmark for the desensitization capabilities of the simulation desensitization method and the large model desensitization method.

[0083] The logical coherence desensitization model makes clear prompt word requirements for four logical relationships in the prompt, namely, desensitization requirements and output format requirements. The specific prompt word requirements are shown in Table 4.

[0084]

[0085] Table 4

[0086] The desensitization result of a specific sample. The original content of the sample is as follows:

[0087] Qiu was born in Shantou, Guangdong, China. His father Qiu was a professor of philosophy and a well-known educator. When Qiu was only a few months old, his family moved to Hong Kong and lived in Yuen Long. When Qiu was 14 years old, his father Qiu, who taught philosophy at the university, passed away, and his mother raised him alone. He studied in the first grade of Yuen Long Wing On School, and in the second grade he entered Shatin Public School and graduated in 1960. In high school, he studied at Hong Kong Pui Ching Middle School. Later, in 1966, he entered the Department of Mathematics of Chung Chi College, Chinese University of Hong Kong. In his third year of university, he was recommended by Professor Stephen Salaff to study at the University of California, Berkeley, under the tutelage of Chen, and graduated early (Yau completed a four-year course in three years in 1969, and Chung Chi College allowed him to graduate, but he did not meet the annual requirements of the Chinese University of Hong Kong at that time and was unable to obtain a degree.) Two years later (1971), he obtained his doctorate degree, and then did a one-year postdoctoral research at the Institute of Advanced Mathematics, and then worked as an assistant professor at Stony Brook University, State University of New York for two years. In 1974, he became an associate professor at Stanford University. In 1979, he returned to the Institute for Advanced Mathematics at Princeton as a professor. From 1984 to 1987, he was a professor at the University of California, San Diego. From 1987 to 2022, he taught at Harvard University.

[0088] According to the original content of the above samples, the results of simulation desensitization and guided desensitization are shown in Table 5.

[0089]

[0090]

[0091] Table 5

[0092] In Table 5, the true labels of the sample with the following three logical relationships are as follows:

[0093] 1. Geographical relationship: Hong Kong, Yuen Long, Yuen Long Wing On School, Shatin Public School, Hong Kong Pui Ching Middle School;

[0094] 2. Computational relations: 1966, completed in 1969 in three years, two years later (1971), one year, two years, 1974;

[0095] 3. Time relationship: 1960, 1966, 1969, 1971, 1974, 1979, 1984 to 1987, 1987 to 2022.

[0096] In the specific sample desensitization effect in Table 5, the three logical relationships of the large model desensitization are all retained, achieving logical coherence desensitization. However, the simulation desensitization loses these logical relationships, such as the confusion of regional relationships, the appearance of multiple provinces and cities; the university graduation time is earlier than the enrollment time. Therefore, it can be seen from Table 5 that the logical coherence desensitization of this scheme is better than the simulation desensitization of the existing technology.

[0097] According to Tables 1, 2, 3, 4 and 5 above, the logical coherence effects of simulation desensitization (existing technology) and this solution are compared. The logical coherence effects of simulation desensitization and large model desensitization are compared. Figure 2 shown.

[0098] Figure 2 In the figure, it can be seen that the large model desensitization (this solution) has the best effect, and the four logical relationships are basically all retained, while the simulation desensitization (existing technology) is basically unable to retain these logical relationships in terms of calculation relationships and regional relationships. It has a certain ability to retain these logical relationships in terms of time relationships and comparison relationships. This is because simulation desensitization randomly generates similar data, and in terms of time relationships and comparison relationships, the logical redundancy between each data item is relatively large. The result of random desensitization may just keep these logical relationships. However, for calculation relationships and regional relationships, the logical redundancy between each data item is very small, and the data items after desensitization have clear requirements, so simulation desensitization is basically unable to retain logical relationships for this. Therefore, the large model desensitization of this solution is better than simulation desensitization.

[0099] This solution is the first to pay attention to the problem of corpus desensitization in the training process of large vertical models in the industry, and pioneered the first method to implement logical coherence desensitization of corpus based on large models, which greatly reduces the information leakage risk of large vertical models. In the task of logical coherence desensitization, for 4 types of logical relationships, the desensitization effect of the large model is better than simulation desensitization. And for the desensitization of new logical relationships, only minor changes to the model prompt are needed to achieve the expected effect.

[0100] In order to facilitate the understanding of the data desensitization process, combined with Figure 3 Provide explanation.

[0101] Figure 3 In the process, first, the original corpus is obtained; the original corpus is the data to be desensitized that has not been desensitized for logical coherence.

[0102] Then, the original corpus is desensitized for logical coherence through a preset language model to obtain desensitized corpus; the desensitized corpus is data that meets the preset sensitivity specifications.

[0103] Among them, the preset language model can use the GPT-4o model, or the privately deployed large model such as Tongyi Qianwen. The "o" in the name of GPT-4o stands for Omni, which means all-powerful, highlighting its multifunctional characteristics. GPT-4o can reason about audio, vision and text in real time. GPT-4o can handle a variety of different languages ​​while improving speed and quality.

[0104] It should be noted that the preset sensitive specifications include but are not limited to "GB / T 35273-2020 Information Security Technology Personal Information Security Specification" and so on.

[0105] The corpus includes multiple data items and their contextual information. Some data items are sensitive data items. In the corpus training scenario, the original corpus needs to be desensitized before it can be used as desensitized corpus to train the vertical domain large model. The premise of desensitization is to correctly identify these data items, especially sensitive data items. There are many sensitive data items in personal information. Common sensitive data items can refer to "GB / T 35273-2020 Information Security Technology Personal Information Security Specification", such as personal property information including identification information (passwords), license plate numbers, etc., personal health and physiological information including fertility information, medication records, etc., personal biometric information including personal genes, fingerprints, etc., and personal identity information including ID cards, passports, etc.

[0106] The collection, use, storage, sharing and transfer of user personal information involved in this plan are in compliance with the relevant laws and regulations, and require the notification of users and the acquisition of their consent or authorization. When applicable, user personal information is de-identified and / or anonymized and / or encrypted.

[0107] Logical coherence desensitization refers to desensitization that retains logical relationships. Due to the existence of contextual logical relationships between multiple data items in the corpus, it is particularly necessary to maintain the logical coherence of desensitization during desensitization, so that these logics can be learned when training the vertical domain large model, so as to better serve the use scenarios in specific fields, such as intelligent writing of judicial documents, etc.

[0108] Finally, the vertical domain large model is trained through desensitized corpus to avoid the situation where the vertical domain large model leaks sensitive data items.

[0109] The desensitized corpus of this solution is used to train vertical domain big models in other industries, thereby greatly reducing the risk of information leakage of vertical domain big models.

[0110] In an embodiment of the present application, since there are contextual logical relationships between multiple data items, in order to maintain the desensitized logical relationships during desensitization, a pre-constructed logical consistency desensitization model is used to perform logical consistency desensitization on the desensitized data. Logical consistency desensitization is to retain the computational relationships, geographical relationships, temporal relationships, comparative relationships and other logical relationships between the various data items, thereby achieving the purpose of obtaining desensitized corpus that retains logical relationships in the process of desensitizing the data.

[0111] Based on the above embodiments Figure 1 A data desensitization processing method is disclosed, and the embodiment of the present application also discloses a data desensitization processing device, such as Figure 4As shown, the data desensitization processing device includes:

[0112] An acquisition unit 401 is used to acquire data to be desensitized;

[0113] The first desensitizing unit 402 is used to perform logical coherence desensitization on the data to be desensitized by using a pre-built logical coherence desensitization model to obtain a desensitized corpus; wherein the logical coherence desensitization is a desensitization that retains the logical relationship between each data item; the logical relationship includes at least a calculation relationship, a geographical relationship, a time relationship and a comparison relationship.

[0114] Furthermore, the first desensitization unit 402 includes:

[0115] An acquisition module is used to acquire sensitive data items in the data to be desensitized through a pre-built logical coherence desensitization model;

[0116] A determination module, used to determine the logical relationship to be matched of the sensitive data item in the context through the prompt word requirements in the logical coherence desensitization model; wherein the prompt word requirements include desensitization requirements and output format requirements;

[0117] A matching module is used to match the logical relationship to be matched with the logical relationship to obtain a matching result; the matching result is a matching result in which the logical relationship to be matched is consistent with the logical relationship;

[0118] The desensitization module is used to perform logical coherence desensitization on sensitive data items according to the matching results to obtain desensitized corpus.

[0119] Furthermore, the data desensitization processing device also includes:

[0120] A parsing unit, configured to parse the requirement type of the output requirement upon receiving the output requirement; wherein the requirement type at least includes a requirement type that conforms to a calculation relationship, a requirement type that conforms to a geographical relationship, a requirement type that conforms to a time relationship, and / or a requirement type that conforms to a comparison relationship;

[0121] The second desensitizing unit is used to perform logical consistency desensitization on the data to be desensitized according to the logical consistency desensitization model and the requirement type.

[0122] Furthermore, the data desensitization processing device also includes:

[0123] The training unit is used to train the vertical domain big model with the desensitized corpus to avoid the vertical domain big model from leaking sensitive data items.

[0124] Furthermore, the data desensitization processing device also includes:

[0125] The marking unit is used to mark the correctly desensitized logical relationship when all data items related to the logical relationship in the data to be desensitized are desensitized correctly during the process of performing logical coherence desensitization on the data to be desensitized.

[0126] In an embodiment of the present application, since there are contextual logical relationships between multiple data items, in order to maintain the desensitized logical relationships during desensitization, a pre-constructed logical consistency desensitization model is used to perform logical consistency desensitization on the desensitized data. Logical consistency desensitization is to retain the computational relationships, geographical relationships, temporal relationships, comparative relationships and other logical relationships between the various data items, thereby achieving the purpose of obtaining desensitized corpus that retains logical relationships in the process of desensitizing the data.

[0127] An embodiment of the present application further provides a storage medium, which includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the above-mentioned data desensitization processing method.

[0128] The present application also provides an electronic device, whose structural diagram is as follows: Figure 5 As shown, it specifically includes a memory 501 and one or more instructions 502, wherein the one or more instructions 502 are stored in the memory 501 and are configured to be executed by one or more processors 503 to execute the one or more instructions 502 to perform the above-mentioned data desensitization processing method.

[0129] For the aforementioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0130] It should be noted that each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referred to each other. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0131] The steps in the methods of the embodiments of the present application can be adjusted in order, combined, and deleted according to actual needs.

[0132] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0133] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0134] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A data desensitization processing method, characterized in that: The method comprises: Obtain the data to be desensitized; Performing logical coherence desensitization on the data to be desensitized by using a pre-built logical coherence desensitization model to obtain desensitized corpus; Among them, the logical coherence desensitization is a desensitization that retains the logical relationship between each data item; the logical relationship at least includes calculation relationship, geographical relationship, time relationship and comparison relationship.

2. The method according to claim 1, characterized in that The pre-built logical coherence desensitization model is used to perform logical coherence desensitization on the data to be desensitized to obtain desensitized corpus, including: Obtain sensitive data items in the data to be desensitized through a pre-built logical coherence desensitization model; Determine the logical relationship to be matched of the sensitive data item in the context through the prompt word requirements in the logical coherence desensitization model; wherein the prompt word requirements include desensitization requirements and output format requirements; According to the to-be-matched logical relationship, the logical relationship is matched to obtain a matching result; the matching result is a matching result that the to-be-matched logical relationship is consistent with the logical relationship; According to the matching results, the sensitive data items are desensitized for logical coherence to obtain desensitized corpus.

3. The method according to claim 1, characterized in that: Also includes: When receiving an output requirement, parsing the requirement type of the output requirement; wherein the requirement type at least includes a requirement type that conforms to a calculation relationship, a requirement type that conforms to a geographical relationship, a requirement type that conforms to a time relationship, and / or a requirement type that conforms to a comparison relationship; According to the logical consistency desensitization model and the requirement type, the data to be desensitized is subjected to logical consistency desensitization.

4. The method according to claim 1, characterized in that: After performing logical coherence desensitization on the data to be desensitized by using a pre-built logical coherence desensitization model to obtain desensitized corpus, the method further includes: The desensitized corpus is used to train the vertical domain big model to avoid the situation where the vertical domain big model leaks sensitive data items.

5. The method according to claim 1, characterized in that Also includes: In the process of performing logical coherence desensitization on the data to be desensitized, when all data items related to logical relationships in the data to be desensitized are desensitized correctly, the correctly desensitized logical relationships are marked.

6. A data desensitization processing device, characterized in that: The device comprises: An acquisition unit, used to acquire the data to be desensitized; The first desensitization unit is used to perform logical coherence desensitization on the data to be desensitized by using a pre-constructed logical coherence desensitization model to obtain a desensitized corpus; wherein the logical coherence desensitization is a desensitization that retains the logical relationship between each data item; the logical relationship includes at least a calculation relationship, a geographical relationship, a time relationship and a comparison relationship.

7. The device according to claim 6, characterized in that The first desensitization unit comprises: An acquisition module is used to acquire sensitive data items in the data to be desensitized through a pre-built logical coherence desensitization model; A determination module, used to determine the logical relationship to be matched of the sensitive data item in the context through the prompt word requirements in the logical coherence desensitization model; wherein the prompt word requirements include desensitization requirements and output format requirements; A matching module, used for matching the to-be-matched logical relationship with the logical relationship to obtain a matching result; the matching result is a matching result that the to-be-matched logical relationship is consistent with the logical relationship; The desensitization module is used to perform logical coherence desensitization on the sensitive data items according to the matching results to obtain desensitized corpus.

8. The device according to claim 6, characterized in that Also includes: A parsing unit, configured to parse the requirement type of the output requirement upon receiving the output requirement; wherein the requirement type at least includes a requirement type that conforms to a calculation relationship, a requirement type that conforms to a geographical relationship, a requirement type that conforms to a time relationship, and / or a requirement type that conforms to a comparison relationship; The second desensitizing unit is used to perform logical consistency desensitization on the data to be desensitized according to the logical consistency desensitization model and the requirement type.

9. The device according to claim 6, characterized in that Also includes: The training unit is used to train the vertical domain big model with the desensitized corpus to avoid the vertical domain big model from leaking sensitive data items.

10. The device according to claim 6, characterized in that Also includes: The marking unit is used to mark the correctly desensitized logical relationship when all data items related to the logical relationship in the data to be desensitized are desensitized correctly during the process of performing logical coherence desensitization on the data to be desensitized.

Images