Self-adaptive sensitive information intelligent identification method and device, equipment, storage medium and product

Abstract

The invention discloses a self-adaptive sensitive information intelligent identification method, device and equipment, a storage medium and a product, relates to the technical field of sensitive information detection.The self-adaptive sensitive information intelligent identification method, device and equipment, the storage medium and the product comprise the steps that enterprise sensitive information flow batch data is subjected to block processing, and data blocks are obtained; carrying out feature extraction and vectorization processing on the data blocks to obtain feature vectors; constructing a retrieval binary tree based on enterprise sensitive information flow batch data; performing fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set; respectively performing feature vector retrieval and semantic structure information retrieval on feature vectors of the sensitive entity feature set to respectively obtain related contexts of each preset number, and reordering the related contexts of each preset number to obtain a target context; and inputting the target context into the large language model to obtain an enterprise sensitive information identification result, thereby realizing accurate identification of the enterprise sensitive information.

Description

Technical Field

[0001] The present application relates to the field of sensitive information detection technology, and in particular to adaptive sensitive information intelligent identification methods, devices, equipment, storage media and products. Background Art

[0002] Sensitive information identification is an important part of network security, aiming to quickly and accurately identify sensitive content involving privacy, confidentiality, malice, etc. from massive information. It is crucial to improve the accuracy and efficiency of sensitive information identification. Traditional sensitive information identification methods are mainly based on keyword and regular expression matching, and sensitive information is screened out by matching the response content with sensitive keywords or regular expressions specified empirically. However, due to the diversity and dynamism of network language, traditional keyword and regular expression matching methods rely on empirical knowledge, incomplete rules, and are prone to misjudgment and missed judgments. The type and level analysis of sensitive information is inaccurate and cannot be effectively classified. Machine learning methods automatically extract content features by training and learning a large amount of known sensitive and non-sensitive data, and construct classifiers for classification and identification. Common machine learning algorithms include naive Bayes, support vector machines, neural networks, etc. Machine learning methods show high accuracy and efficiency when processing large-scale data sets, but they have high requirements for training data annotation, high difficulty and cost. If sensitive information is not within the distribution range of the annotated data, the identification of sensitive information will be greatly reduced.

[0003] 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

[0004] The main purpose of this application is to provide an adaptive sensitive information intelligent identification method, device, equipment, storage medium and product, aiming to solve the technical problems that the regular expression matching method relies on empirical knowledge, has incomplete rules, is prone to misjudgment and missed judgment, cannot be adaptively updated, and has weak generalization.

[0005] To achieve the above objectives, the present application proposes an adaptive sensitive information intelligent identification method, the method comprising:

[0006] The batch data of the enterprise sensitive information flow is processed into blocks to obtain data blocks, and the data blocks are subjected to feature extraction and vectorization processing to obtain feature vectors;

[0007] Constructing a retrieval binary tree based on the batch data of the enterprise sensitive information flow;

[0008] Perform fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set;

[0009] Performing feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reordering the preset number of related contexts to obtain a target context;

[0010] The target context is input into a large language model to obtain enterprise sensitive information recognition results.

[0011] In one embodiment, the step of constructing a retrieval binary tree based on the batch data of the enterprise sensitive information flow includes:

[0012] Cleaning the sensitive batch data of the enterprise, and annotating and extracting the cleaned data to obtain sensitive entities and sensitive information categories;

[0013] Performing cluster analysis according to the sensitive entities and the sensitive information categories to obtain multiple clustering results of the sensitive entities;

[0014] The corresponding feature vector sets in the plurality of clustering results are respectively merged to obtain a feature vector index;

[0015] Determine a division point of the feature vector index in the first dimension, and create a node with the index vector corresponding to the division point, wherein the node includes a current index value, a node pointer pointing to a left child node and a right child node, and a feature set pointer of a feature set corresponding to the node;

[0016] Divide the index vectors of the partition points smaller than the first dimension into the left subtree of the partition points, and divide the index vectors of the partition points larger than the first dimension into the right subtree of the partition points, to obtain a binary tree of the first dimension;

[0017] Based on the dimension order, the root node of the binary tree of the second dimension is used as the left child node and the right child node of the binary tree of the first dimension to obtain a retrieval binary tree, and the dimension order of the first dimension is located before the dimension order of the second dimension.

[0018] In one embodiment, the step of performing cluster analysis according to the sensitive entities and the sensitive information categories to obtain a plurality of clustering results of the sensitive entities includes:

[0019] Obtaining sensitive data features according to the sensitive entity and the sensitive information category;

[0020] Determine an initial cluster center, determine an initial cluster center vector according to the initial cluster center, and obtain a sensitive data feature vector according to the sensitive data feature;

[0021] Obtaining a vector distance according to the initial cluster center vector and the sensitive data feature vector;

[0022] Determine a neighborhood set according to the vector distance, and determine the number of data points in the neighborhood set and the median of the data points in the feature dimension, obtain a new cluster center according to the neighborhood set, the number of data points and the median, and obtain multiple clustering results according to the new cluster center;

[0023] When the stability of the plurality of clustering results is greater than a preset stability, the plurality of clustering results are output.

[0024] In one embodiment, the step of performing fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set includes:

[0025] Converting the feature vector into a query vector, and comparing the query vector with the dimension value of the index vector of the tree node of the retrieval binary tree to obtain a comparison result;

[0026] When the comparison result is that the characteristic value of the query vector is less than the dimension value of the index vector of the tree node of the search binary tree, recursively search the left subtree of the search binary tree; when the comparison result is that the characteristic value of the query vector is greater than the dimension value of the index vector of the tree node of the search binary tree, recursively search the right subtree of the search binary tree;

[0027] When a leaf node is found recursively, recursive backtracking is performed to obtain a list of candidate nodes;

[0028] Calculate the Manhattan distance between the candidate nodes in the candidate node list and the query vector to determine the target node;

[0029] A sensitive entity feature set is determined according to the target node.

[0030] In one embodiment, the step of performing feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reordering the preset number of related contexts to obtain the target context includes:

[0031] Performing feature vector retrieval on the feature vector of the sensitive entity feature set, and calculating feature similarity between the feature vector and a first vector, where the first vector is a feature retrieval vector;

[0032] Obtaining a preset number of first relevant contexts according to the feature similarity;

[0033] Performing semantic structure information retrieval on the feature vector of the sensitive entity feature set, and obtaining a semantic retrieval score according to the feature vector and a second vector, wherein the second vector is a semantic structure vector;

[0034] Obtaining a preset number of second relevant contexts according to the semantic retrieval score;

[0035] Reordering scores are calculated for the first relevant context and the second relevant context, the reordering scores are arranged from high to low, and relevant contexts corresponding to a target number of reordering scores are selected to obtain a target context.

[0036] In one embodiment, after the step of inputting the target context into the large language model to obtain the enterprise sensitive information recognition result, the method further includes:

[0037] Determine a verification task, perform entity extraction on a sentence in the verification task, and receive a verification result based on the entity;

[0038] When the verification result is a misjudgment, the large language model is updated based on the verification result, and enterprise sensitive information is identified based on the updated large language model to obtain an enterprise sensitive information identification result.

[0039] In addition, to achieve the above-mentioned purpose, the present application also proposes an adaptive sensitive information intelligent identification device, the adaptive sensitive information intelligent identification device comprising:

[0040] The data processing module is used to process the batch data of the enterprise sensitive information flow in blocks to obtain data blocks, and perform feature extraction and vectorization processing on the data blocks to obtain feature vectors;

[0041] A binary tree construction module, used to construct a retrieval binary tree based on the batch data of the enterprise sensitive information flow;

[0042] A fuzzy retrieval module, used for performing fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set;

[0043] An accurate retrieval module, used to perform feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set, respectively, to obtain a preset number of related contexts, and to reorder the preset number of related contexts to obtain a target context;

[0044] The information identification module is used to input the target context into the large language model to obtain the enterprise sensitive information identification result.

[0045] In addition, to achieve the above-mentioned purpose, the present application also proposes an adaptive sensitive information intelligent identification device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, and the computer program is configured to implement the steps of the adaptive sensitive information intelligent identification method as described above.

[0046] 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 adaptive sensitive information intelligent identification method described above are implemented.

[0047] 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 adaptive sensitive information intelligent identification method as described above.

[0048] One or more technical solutions proposed in the present application have at least the following technical effects: performing block processing on batch data of enterprise sensitive information flow to obtain data blocks, and performing feature extraction and vectorization processing on the data blocks to obtain feature vectors; constructing a retrieval binary tree based on the batch data of enterprise sensitive information flow; performing fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set; performing feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reordering the preset number of related contexts to obtain a target context; inputting the target context into a large language model to obtain an enterprise sensitive information recognition result, thereby achieving accurate recognition of the enterprise sensitive information. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] 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.

[0050] 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.

[0051] Figure 1 A flowchart of the first embodiment of the adaptive sensitive information intelligent identification method of the present application is provided;

[0052] Figure 2 The main flow chart of knowledge retrieval enhancement provided for the first embodiment of the adaptive sensitive information intelligent identification method of the present application;

[0053] Figure 3 A schematic diagram of a feature fuzzy retrieval process provided for an embodiment of the adaptive sensitive information intelligent identification method of the present application;

[0054] Figure 4The main flow chart of hybrid retrieval and re-ranking provided for the first embodiment of the adaptive sensitive information intelligent identification method of the present application;

[0055] Figure 5 A flowchart of the second embodiment of the adaptive sensitive information intelligent identification method of the present application is provided;

[0056] Figure 6 The process of building a knowledge vector database of sensitive information of an enterprise provided in the first embodiment of the adaptive sensitive information intelligent identification method of this application;

[0057] Figure 7 A schematic diagram of a retrieval binary tree construction process provided for an embodiment of the adaptive sensitive information intelligent identification method of the present application;

[0058] Figure 8 A brief flowchart of the adaptive sensitive information intelligent identification method provided in Example 2 of the present application;

[0059] Fig. 9 This is a schematic diagram of the module structure of the adaptive sensitive information intelligent identification device according to the embodiment of the present application;

[0060] Fig.10 This is a schematic diagram of the device structure of the hardware operating environment involved in the adaptive sensitive information intelligent identification method in the embodiment of the present application.

[0061] 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

[0062] 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.

[0063] 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.

[0064] The main solution of the embodiment of the present application is: block processing of batch data of enterprise sensitive information flow to obtain data blocks, and feature extraction and vectorization processing of the data blocks to obtain feature vectors; constructing a retrieval binary tree based on the batch data of enterprise sensitive information flow; performing fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set; performing feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reordering the preset number of related contexts to obtain a target context; inputting the target context into a large language model to obtain an enterprise sensitive information recognition result.

[0065] In this embodiment, for ease of description, the following description is made by taking the adaptive sensitive information intelligent identification device as the execution subject.

[0066] Since the traditional sensitive information identification method in the prior art is mainly based on keyword and regular expression matching, sensitive keywords or regular expressions are empirically specified to match the response content, thereby screening out sensitive information. However, due to the diversity and dynamism of network language, the traditional keyword and regular expression matching method relies on empirical knowledge, the rules are incomplete, and it is easy to produce misjudgments and missed judgments. The type and level analysis of sensitive information is inaccurate and cannot be effectively classified. The machine learning method automatically extracts the features of the content by training and learning a large amount of known sensitive and non-sensitive data, and constructs a classifier for classification and identification. Common machine learning algorithms include naive Bayes, support vector machines, neural networks, etc. Machine learning methods show high accuracy and efficiency when processing large-scale data sets, but they have high requirements for training data annotation, and are difficult and costly. If the sensitive information is not within the distribution range of the annotated data, the identification of sensitive information will be greatly reduced.

[0067] The present application provides a solution, which adopts a retrieval binary tree construction based on clustering feature fusion to form an enterprise sensitive information database, and then performs fuzzy query based on the retrieval binary tree to improve recognition efficiency. For the fuzzy query results, a hybrid retrieval and re-ranking method is used to screen out the precise context and send it to the large model. The large model has a deeper understanding of the context in the generation task, and can better capture the semantics and logical structure of the text. The recognition task of enterprise sensitive information is cleverly converted into a language generation task of the large model, and special symbols are used to mark the sensitive information recognition results. Compared with the existing solutions, sensitive information can be identified more quickly, more effectively and more accurately.

[0068] It can be seen from the above embodiments that the present application obtains data blocks by performing block processing on batch data of enterprise sensitive information flow, and performs feature extraction and vectorization processing on the data blocks to obtain feature vectors; constructs a retrieval binary tree based on the batch data of enterprise sensitive information flow; performs fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set; performs feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reorders the preset number of related contexts to obtain a target context; inputs the target context into a large language model to obtain an enterprise sensitive information recognition result, thereby achieving accurate recognition of enterprise sensitive information.

[0069] It should be noted that 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, etc., or an electronic device capable of realizing the above functions, an adaptive sensitive information intelligent recognition device, etc. The following takes the adaptive sensitive information intelligent recognition device as an example to illustrate this embodiment and the following embodiments.

[0070] Based on this, the present application embodiment provides an adaptive sensitive information intelligent identification method, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the adaptive sensitive information intelligent identification method of the present application.

[0071] In this embodiment, the adaptive sensitive information intelligent identification method includes steps S10 to S50:

[0072] Step S10, performing block processing on the batch data of the enterprise sensitive information flow to obtain data blocks, and performing feature extraction and vectorization processing on the data blocks to obtain feature vectors.

[0073] It should be noted that enterprise sensitive information flow and batch data include enterprise sensitive information and enterprise flow and batch data information. Enterprise sensitive information includes personal identity information, such as employee or customer name, address, ID number, telephone number, email address, etc., financial data, business secrets, customer data and other information. Enterprise flow and batch data information is enterprise data information obtained by stream processing and batch processing of enterprise data. Feature vector refers to the numerical representation extracted from enterprise sensitive information that can represent the characteristics of the data.

[0074] In the specific implementation, when processing batch data of sensitive enterprise information, this embodiment adopts a block strategy to split the huge text data into smaller fragments. Through the intent-based block method, the information is cut in the form of intent packages using the natural language processing tool NLTK. This process ensures that the sentences after block division maintain both integrity and accuracy of expression. After the text is segmented, feature extraction can be performed on each text segment, including part-of-speech tagging, named entity recognition, stem extraction and other processing methods to obtain the feature value of the text, and then the feature value of the text is vectorized, for example, the text is converted into a vector based on TF-IDF to obtain the feature vector corresponding to the batch data of the enterprise sensitive information flow.

[0075] Step S20, constructing a retrieval binary tree based on the batch data of the enterprise sensitive information flow.

[0076] It should be noted that the retrieval binary tree is a special binary tree structure that can be used to organize and retrieve feature vectors in enterprise sensitive information. Based on the properties of the binary search tree, the enterprise sensitive information can be distributed in a certain pattern, thereby speeding up the retrieval speed.

[0077] In a specific implementation, when constructing a retrieval binary tree based on the batch data of the enterprise's sensitive information flow, it is necessary to first perform a certain degree of data cleaning and data labeling on the batch data of the enterprise's sensitive information flow, so that the batch data of the enterprise's sensitive information flow can be divided into different levels according to their importance and the characteristics of the enterprise itself. For example, in a financial institution, the customer's identity information, financial records, etc. are high-level sensitive information, and the account type, account opening date, etc. are secondary sensitive information. When performing a sensitive information identification task, the enterprise can specify the sensitivity of this task. A sensitive information identification task of a certain level only includes the retrieval of sensitive information with a sensitivity greater than or equal to the task level. In this embodiment, the data level can be divided into three levels, namely high level, medium level and low level, and a retrieval binary tree is established based on these three levels. When constructing a retrieval binary tree, it is possible to determine the corresponding index vector based on the batch data of the enterprise sensitive information flow, and determine the partition point based on the dimensional information in the index vector and the value in the dimensional information. Based on the partition point, the points smaller than the index vector corresponding to the partition point are assigned to the left subtree of the node, and the points larger than the index vector corresponding to the partition point are assigned to the right subtree of the node. The retrieval binary tree is constructed based on this method.

[0078] Step S30, performing fuzzy search based on the search binary tree and the feature vector to obtain a sensitive entity feature set.

[0079] It should be noted that the sensitive entity feature set refers to the set of candidate sensitive entities obtained after fuzzy retrieval, and the data in this set is relatively rough compared with the target data.

[0080] In the specific implementation, when searching for sensitive information, refer to Figure 2, input the real-time streaming data, batch data, or mixed text data combining real-time and batch characteristics of the enterprise. After receiving the data, the system processes it in blocks, and uses a block model to segment the data by paragraphs. After block division, each query block can be matched with the examples in the vector knowledge base in a more fine-grained manner. Subsequently, each data block is processed one by one: first, the text encoder is used to extract text features to obtain the corresponding Chunk features; then, the Chunk features are used to perform information retrieval in the previously constructed enterprise sensitive information knowledge vector library; finally, the retrieved information is sent to the instruction engineering construction of the large model as the context in the instruction, and the model is assisted in sensitive information identification to improve the accuracy of identification. Block division, feature extraction, etc., this embodiment focuses on the design of the information retrieval process. The information retrieval process includes category feature fuzzy retrieval and precise retrieval, and the precise retrieval stage includes hybrid retrieval and re-ranking. When performing fuzzy retrieval, the feature vector is converted into a query vector, and the query vector is compared with the dimension value of the index vector of the tree node of the retrieval binary tree to obtain a comparison result; when the comparison result is that the feature value of the query vector is less than the dimension value of the index vector of the tree node of the retrieval binary tree, a recursive search is performed in the left subtree of the retrieval binary tree, and when the comparison result is that the feature value of the query vector is greater than the dimension value of the index vector of the tree node of the retrieval binary tree, a recursive search is performed in the right subtree of the retrieval binary tree; when a leaf node is recursively found, recursive backtracking is performed to obtain a candidate node list; the Manhattan distance is calculated based on the candidate nodes in the candidate node list and the query vector to determine the target node; and a sensitive entity feature set is determined based on the target node.

[0081] When a retrieval binary tree is constructed in the knowledge vector library based on category fusion features, category fuzzy retrieval based on the retrieval binary tree is prioritized during retrieval. The specific retrieval method is: compare the dimension values ​​of the query vector and the index vector of the tree node in turn, decide whether to recursively search the left subtree or the right subtree, and after reaching the leaf node, backtrack to obtain a list of candidate nodes. Compare the Manhattan distance between the query vector and the node vector in the list, find the nearest index node, obtain the sensitive entity feature set corresponding to the index node, and then perform further precise retrieval in the feature set. If the level of sensitive information identification is specified during retrieval, exclude candidate nodes whose corresponding sensitive feature levels are lower than that level from the candidate nodes. For ease of understanding, refer to Figure 3, assuming that the feature vector of an information block is (7,3) at this time, we can get the candidate index nodes {(5,4), (7,2), (9,6)} by comparing the dimensions in sequence, and get the closest node (7,2) by comparing their Manhattan distances. Then, we can find the feature set of the corresponding category based on the pointer stored in the node. If the sensitive information level specified in this retrieval is medium, then remove the index vector (9,6) corresponding to the low-level sensitive feature from the candidate node list, retain the index vectors (5,4) and (7,2) corresponding to the high-level sensitive feature, and then search for the closest node. The fuzzy retrieval process quickly filters out a large amount of irrelevant information from the feature category dimension, thereby greatly speeding up the context retrieval speed.

[0082] Step S40, performing feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set, respectively, to obtain a preset number of related contexts, and reordering the preset number of related contexts to obtain a target context.

[0083] It should be noted that both the relevant context and the target context are semantic text information. The relevant context specifically refers to the semantic text information before and after the semantic text corresponding to the feature vector of the sensitive entity feature set, and the target context information is a preset number of relevant contexts in the relevant context.

[0084] In a specific implementation, a feature vector retrieval is performed on the feature vector of the sensitive entity feature set, and a feature similarity between the feature vector and a first vector is calculated, where the first vector is a feature retrieval vector; a preset number of first related contexts are obtained based on the feature similarity; a semantic structure information retrieval is performed on the feature vector of the sensitive entity feature set, and a semantic retrieval score is obtained based on the feature vector and a second vector, where the second vector is a semantic structure vector; a preset number of second related contexts are obtained based on the semantic retrieval score; a reordering score is calculated for the first related context and the second related context, the reordering scores are arranged from high to low, and related contexts corresponding to the target number of reordering scores are selected to obtain a target context.

[0085] Reference Figure 4 , the feature vector retrieval algorithm and the semantic structure information retrieval algorithm are used to retrieve respectively, and 15 relevant contexts are obtained for each retrieval algorithm. Finally, the 5 most relevant contexts are obtained by re-ranking and sent to the large model. Among them, the feature vector retrieval process is to calculate the similarity between vectors, and finally return the information block corresponding to the vector with higher similarity. The length of the feature vector may vary depending on the size of the information block. This method uses cosine similarity to measure vector similarity. Cosine similarity calculates the cosine value of the angle between two vectors. The larger the cosine similarity value, the more similar the vectors are.

[0086] Assume that the two vectors used for comparison are A = [a1, a2, ..., a n ] and B=[b1,b2,...,b n ], then the feature vector similarity score is:

[0087]

[0088] Where ||A|| and ||B|| are the norms of the two vectors, and A·B represents the dot product of the two vectors.

[0089] The cosine distance distinguishes differences in direction and is insensitive to absolute values, thus shielding the differences caused by text length and focusing the search results on text content. Then, the 15 results with the highest scores are taken out and their feature vector search scores are used for re-ranking calculation.

[0090] Text semantic retrieval combines existing word vector models for feature extraction, such as the CLIP model, which can capture deeper semantic relationships between texts. Then, the similarity between the semantic features of the text to be detected and the information block in the knowledge base is compared, and the semantic relevance between the two text feature vectors is evaluated to retrieve the corresponding relevant context results. The calculation formula for text semantic retrieval score is as follows:

[0091]

[0092] Among them, Q represents the text block to be detected, D is the sensitive information sentence sample in the corresponding knowledge vector library, q is the sensitive entity word in the sensitive information sentence, and the f function calculates the number of occurrences of a word in the sentence sample. k1 and b are constants. k1 is used to adjust the influence of word frequency, b is used to normalize the document length and control the influence of document length on frequency. avgdl is the average length of all documents in the document collection. IDF is the inverse document frequency, which is used to reduce the weight of words that appear frequently in all documents. Then, the 15 results with the highest scores are taken out, and their text semantic retrieval scores are used for the next calculation.

[0093] After obtaining the results retrieved by the two retrieval methods, the re-ranking scores of the retrieved sensitive data samples are obtained by mixing the scores of the two retrieval results in a certain proportion, and then the top 5 sample data are selected as the final result. The re-ranking scheme is as follows:

[0094] First, in order to eliminate the numerical differences between the two search methods, the search scores of the two search results are standardized. The search scores can be scaled to the same range by using minimum-maximum standardization or Z-score standardization.

[0095] Next, according to the distribution of the retrieval results of the two methods, the weights of the retrieval results of the two methods are dynamically adjusted. Considering that the data distribution shows the stability of the retrieval results to a certain extent, the standard deviation is the main consideration in the weight calculation.

[0096] Let σ v is the standard deviation of the original score of the feature vector retrieval result, σ s is the standard deviation of the original score of the text semantic retrieval result. For a piece of information in the retrieval result, its feature vector retrieval standardized score is score v , the standardized score of text semantic retrieval is score s , then its re-ranking score score r The calculation method is:

[0097]

[0098] According to this formula, if the standard deviation of the feature vector retrieval result score is large, the weight coefficient of the feature vector retrieval result is The weight coefficient of text semantic retrieval is reduced accordingly A small standard deviation means that the retrieval results have a high correlation with the samples to be tested in this dimension. At this time, the weight of the retrieval results of this dimension can be appropriately increased to filter out information with a higher correlation with the samples to be tested from the retrieval results.

[0099] Step S50: input the target context into a large language model to obtain a result of identifying enterprise sensitive information.

[0100] It should be noted that enterprise sensitive information identification refers to identifying and extracting text containing sensitive information from input information, and marking and outputting the sensitive information according to the sensitivity level.

[0101] In the specific implementation, the search results are arranged in descending order according to the re-ranking scores, and the five pieces of information with the highest scores are selected as context examples and sent to the large model. Through instruction engineering, the search context results are combined with user questions and instructions to form the instruction engineering data of the large language model. Sensitive information is identified and parsed based on the instruction engineering data and the large language model to obtain the specific information output of the sensitive information entity.

[0102] In a feasible implementation manner, after the step of inputting the target context into the large language model to obtain the enterprise sensitive information recognition result, the step further includes:

[0103] Determine a verification task, perform entity extraction on a sentence in the verification task, and receive a verification result based on the entity;

[0104] When the verification result is a misjudgment, the large language model is updated based on the verification result, and enterprise sensitive information is identified based on the updated large language model to obtain an enterprise sensitive information identification result.

[0105] In the specific implementation, the purpose of the enterprise sensitive information entity recognition project is to build a reasonable instruction system so that the large model can accurately understand the task requirements. At the same time, the context data is obtained through the knowledge retrieval enhancement module, so that the model can more accurately identify the enterprise sensitive information, thereby obtaining the final recognition result.

[0106] The enterprise sensitive information big model instruction of this embodiment mainly consists of three parts: the task description of sensitive information identification, sensitive information context examples and the detected data provided by the user. Task description of sensitive information identification: describes the specific situation of the task, including: role settings, entity categories to be extracted, etc. Sensitive information context examples: the context samples retrieved in the previous section are attached to the prompt as a few sample demonstrations. On the one hand, it can standardize the output style, and on the other hand, it can provide direct reference examples for sensitive information identification tasks to help them make predictions. The detected data provided by the user: the detected enterprise sensitive information data. The enterprise sensitive information big model instruction of this embodiment mainly consists of three parts:

[0107]

[0108]

[0109] If the input sequence does not contain any entities, the output will simply copy the input X; for sensitive entities in the input sequence, special marking symbols such as "@@" and "##" are used to surround them. This converts the sensitive entity task into generating text containing special label information, which reduces the difficulty of the task. You only need to mark the location of the enterprise sensitive entity and copy it for the rest.

[0110] In order to overcome the hallucination problem of the large language model of enterprise sensitive information and improve the accuracy of the model in sensitive information identification tasks, a sensitive data identification result verification module is designed in a targeted manner. The large model uses its general understanding ability to make a secondary judgment on the type of sensitive data identified, and screen out sensitive data with obvious category errors in the results.

[0111] The sensitive data category determination instruction engineering design of this embodiment is as follows:

[0112]

[0113] The sensitive information verification task description is as follows: the purpose is to specify the verification task, such as "Please verify whether the word extracted from the given sentence is a financial data sensitive entity, answer yes or no?". This tells the big model to verify whether the words extracted from the given sentence belong to sensitive entities of a specific sensitive category; sensitive information: mainly lists the detection results of sensitive information entities in a list for word-by-word judgment; detected data: mainly provides contextual information of the source of sensitive words for sensitive information entities to improve the accuracy of secondary type judgment verification.

[0114] This constitutes a secondary verification instruction project, which aims to determine whether the extracted sensitive entity belongs to a specific sensitive entity type through the secondary sensitive information type verification task. If the extracted sensitive entity belongs to the sensitive information type of the query, the answer is "yes", otherwise the answer is "no". For example, if in the previous recognition results, a sensitive information "ID card" is mistakenly labeled as financial information, then the model will use its general knowledge ability to find that the ID card does not belong to financial information, and thus correct the category of the sensitive information or remove the sensitive information as needed.

[0115] The present embodiment provides an adaptive sensitive information intelligent identification method, which obtains data blocks by performing block processing on batch data of enterprise sensitive information flow, and performs feature extraction and vectorization processing on the data blocks to obtain feature vectors; constructs a retrieval binary tree based on the batch data of enterprise sensitive information flow; performs fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set; performs feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reorders the preset number of related contexts to obtain a target context; inputs the target context into a large language model to obtain an enterprise sensitive information identification result, thereby achieving accurate identification of enterprise sensitive information.

[0116] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 5 , step S20 also includes steps S201 to S206:

[0117] S201: Clean the sensitive batch data of the enterprise, and annotate and extract the cleaned data to obtain sensitive entities and sensitive information categories.

[0118] It should be noted that the sensitive entity refers to the data content clearly stated in the data information, and the sensitive information category is the classification of the data content. When the data content is "the mobile phone number is 135XXXXXXXX", it can be determined that the sensitive entity is "135XXXXXXXX" and the sensitive information category is "mobile phone number / communication method".

[0119] S202: Perform cluster analysis according to the sensitive entities and the sensitive information categories to obtain multiple clustering results of the sensitive entities.

[0120] It should be understood that cluster analysis can be performed based on sensitive entities and sensitive information categories to determine the distribution of sensitive entities in sensitive data blocks in sensitive paragraphs, and the corresponding distribution of sensitive entities is the corresponding multiple clustering results.

[0121] S203: Fusing the corresponding feature vector sets in the plurality of clustering results respectively to obtain feature vector indexes.

[0122] It should be understood that the feature vector sets in the clustering results can be fused and the weight parameters of each vector can be determined. Specifically, the weight parameters can be reasonably assigned according to the actual situation of each enterprise, and the feature vectors can be weighted averaged to generate a feature vector index.

[0123] S204: Determine the division point of the feature vector index in the first dimension, and create a node with the index vector corresponding to the division point, wherein the node includes a current index value, node pointers pointing to the left child node and the right child node, and a feature set pointer of the feature set corresponding to the node.

[0124] It can be understood that the first dimension is the currently selected dimension. There can be n dimensions in each index set, and the dimensions can be sorted in sequence to obtain a list of dimensions arranged in sequence. If the current dimension is 2, then the current first dimension is the dimension of dimension 2, and the importance of the data in the first dimension and the median of the data are used as the dividing point. The dividing point includes the current index value, node pointers pointing to the left child node and the right child node, and the feature set pointer of the feature set corresponding to the node.

[0125] S205: Divide the index vectors of the division points smaller than the first dimension into the left subtree of the division points, and divide the index vectors of the division points larger than the first dimension into the right subtree of the division points, to obtain a binary tree of the first dimension.

[0126] It can be understood that when dividing the left and right subtrees, the index vector can be compared with the division point to determine the vector score between the index vector and the division point, and the index vector of the division point that is smaller than the first dimension is divided into the left subtree of the division point, and the index vector of the division point that is larger than the first dimension is divided into the right subtree of the division point, thereby obtaining a binary tree of the first dimension.

[0127] S206: Based on the dimension order, the root node of the binary tree of the second dimension is used as the left child node and the right child node of the binary tree of the first dimension to obtain a retrieval binary tree, and the dimension order of the first dimension is located before the dimension order of the second dimension.

[0128] It can be understood that the first dimension at this time is dimension 2. According to the dimension order, the second dimension can be determined as dimension 3 and other dimensional data thereafter. Based on the same construction method as the binary tree of the first dimension, multiple binary trees of the second dimension are constructed, and the root node of the binary tree of the second dimension is used as a child node of the binary tree of the first dimension to form a retrieval binary tree.

[0129] It should be understood that the partition point refers to the point used to distinguish the index vector of the current dimension. The partition point is a real point in the index vector of the current dimension. The partition point can divide the index vector into two left and right subtrees, and the left subtree or the right subtree has the same or similar characteristics. The first dimension can be determined as the currently selected dimension, and the second dimension is the other dimension after the current dimension is sorted according to the dimension.

[0130] In the specific implementation, when constructing the retrieval binary tree, the enterprise sensitive information knowledge vector library is also being constructed. When constructing the binary tree and the enterprise sensitive information knowledge vector library, the following can be used to determine the sensitive information of the enterprise: Figure 6 The process of constructing an enterprise sensitive information knowledge vector database is as follows: first, the sensitive data is cleaned and labeled, and the cleaned and labeled data is chunked to obtain data chunks, and then feature fusion and clustering are performed on the data to vectorize the data chunks to obtain vectorized indexes, and the vectorized indexes are stored in a value vector database. This embodiment organizes, cleans and labels the relevant sensitive information and confidentiality-related knowledge of a specific enterprise, and then analyzes the structure of the sensitive feature information, and then performs feature vectorization, thereby constructing an enterprise sensitive information knowledge vector database. This proposal can provide an enterprise with a private sensitive information library for knowledge storage and management, and also provides contextual example support for the identification of enterprise sensitive information.

[0131] First, data cleaning and manual data annotation are performed, and then sensitive text is segmented by text intent segmentation. Next, the classification categories of sensitive entities and the central features of each category are determined through iterative clustering. Then, the feature index of each category is obtained through feature fusion. Finally, a retrieval binary tree is constructed based on the category index, which together with the sensitive entity features constitutes a knowledge vector library. By clustering, sensitive information feature vectors with similar features are aggregated together, and then a binary tree is constructed using the category feature index after feature fusion, which can effectively improve the retrieval efficiency of the knowledge vector library.

[0132] First, data extraction is required. Data extraction operations are performed through data cleaning, data processing, metadata extraction, and metadata annotation. During data cleaning, relevant documents such as PDF, word, markdown, and enterprise sensitive information data descriptions are extracted and cleaned to select information related to sensitive information examples; data processing: including data format processing, removal of unrecognizable content, compression and formatting, etc.; metadata extraction: extracting relevant information such as enterprise sensitive information categories and classifications in the text; metadata annotation: various types of special identification annotations are performed on the description of enterprise sensitive information, such as "The bank president's ID number is 877862187607190018" is converted into a generated text sequence "The bank president's @@ID number** is @@877862187607190018##", and the special marking symbol @@## is used to mark the sensitive entities and sensitive information categories that need to be extracted. Then, data segmentation is performed. When segmentation is performed, large blocks of text are decomposed into small segments. This proposal adopts an intent-based segmentation method, which is segmented through professional intent packages. This embodiment uses the commonly used intent package and NLP-based NLTK tool to complete the segmentation, thereby ensuring the integrity of the segmented sentences and the accuracy of the expression. Then, through cluster analysis, the distribution of sensitive entities in sensitive data blocks in all sensitive paragraphs is understood, preparing for the next step of generating feature indexes for each category.

[0133] When performing feature fusion, a feature vector index can be formed. For each feature vector set {s1, s2, ...s n} to build a feature index for each category. When searching, fuzzy search can be performed on the category features first, and then precise search can be performed in the corresponding category. The feature index calculation method is as follows:

[0134]

[0135] where ω j is the weight parameter of each vector. Different enterprises can assign different weights to different feature entities according to their actual conditions to reflect the importance of a feature word in the entire sensitive entity set. j satisfy Used to achieve weighted average of feature vectors. W is the projection matrix obtained by linear discriminant analysis of the feature vector set, which is used to reduce the dimensionality of the feature vector. Through dimensionality reduction, redundant noise in some feature vectors can be eliminated and the most important features can be retained. At the same time, dimensionality reduction makes the generated feature vector index smaller and more concise, improves the efficiency of sensitive information retrieval, and also reduces storage space.

[0136] When constructing a knowledge vector library, for an index set with n dimensions {(x1,...,xn),(y1,...,yn),...}, starting from the first dimension, compare the values ​​of the index vectors in the dimension {x1,y1,...} in turn, find the division point m1 on the dimension, and then create a node M1 for the index vector corresponding to the division point of the dimension. The node contains the current index value, pointers to the left and right child nodes, and pointers to the feature set corresponding to the index node. All index vectors that are smaller than the division point m1 in the first dimension are assigned to the left subtree of node M1, and the remaining nodes are assigned to the right subtree of node M1. In the left and right subtrees of node M1, compare the values ​​of the second dimension {x2,y2,...} of each index set, and create the root nodes of the left and right subtrees with the index vectors corresponding to the division values ​​on the left and right sides as the left and right child nodes of M1. And so on, cyclically compare the values ​​of each index dimension in the index set until all index vectors have corresponding index tree nodes.

[0137] In the comparison of each dimension, the selection method of the dividing point comprehensively considers the importance of the sensitive information category and the numerical centrality of the dimension, compares the scores of each index vector on the dimension, and selects the index vector with the highest score as the dividing point. The score of the kth index vector on the i-th dimension is:

[0138] m k =α level k -β·|num k -median i |

[0139] Among them, level k is the importance score of the sensitive feature category corresponding to the kth index vector. The higher the importance of the sensitive feature, the higher the score. k is the value of the kth index vector in the i-th dimension, median i is the median of each index vector in the i-th dimension, and α and β are weight adjustment coefficients.

[0140] In the knowledge vector library, high-level sensitive information is more likely to be used. Therefore, in this embodiment, the index vector corresponding to the high-level sensitive information is more likely to be selected as the partition point, so that the index nodes that are accessed more frequently have a greater probability of being located in the upper layer of the tree. In subsequent queries, low-level information that is not related to the query can be screened out earlier, improving retrieval efficiency. At the same time, screening out high-level sensitive information at an earlier stage can reduce the probability of false detection or mismatching. Considering the data centering degree on the dimension is to better balance the number of nodes in the left and right subtrees. The more balanced the left and right subtrees are, the lower the height of the tree can be. Therefore, the index vector with a value closer to the median in a certain dimension is more likely to be selected.

[0141] For ease of understanding, assume that there is currently the following two-dimensional index vector set: {high level {(5,4), (7,2)}, medium level {(4,7)}, low level {(2,3), (8,1), (9,6)}}, and this two-dimensional data set contains the x-axis and the y-axis. First, calculate the vector score on the x-axis, and calculate that the vector (5,4) has the highest score. Take the vector index (5,4) as the root node, and divide the sub-index set {(2,3), (4,7)} with an x-axis less than 5 into the left subtree, and divide the remaining index set {(9,6), (8,1), (7,2)} with an x-axis greater than 5 into the right subtree. Then, in the sub-index set, calculate the scores of each vector on the y-axis to construct nodes to further divide the index set. Then, in the further generated sub-index set, divide the index set based on the x-axis, and so on. The binary tree construction process is as follows: Figure 7 As shown in the figure, by using feature vector indexes of different categories to build a knowledge vector library, it is used to provide context examples for sensitive information retrieval in the future to improve the accuracy of sensitive information identification. In the subsequent sensitive information retrieval process, the corresponding category can be matched according to the feature vector index, and then the specific sensitive entity can be matched in the category feature, thereby improving the efficiency of context retrieval.

[0142] In a feasible implementation manner, the step of performing cluster analysis according to the sensitive entities and the sensitive information categories to obtain multiple clustering results of the sensitive entities includes:

[0143] Obtaining sensitive data features according to the sensitive entity and the sensitive information category;

[0144] Determine an initial cluster center, determine an initial cluster center vector according to the initial cluster center, and obtain a sensitive data feature vector according to the sensitive data feature;

[0145] Obtaining a vector distance according to the initial cluster center vector and the sensitive data feature vector;

[0146] Determine a neighborhood set according to the vector distance, and determine the number of data points in the neighborhood set and the median of the data points in the feature dimension, obtain a new cluster center according to the neighborhood set, the number of data points and the median, and obtain multiple clustering results according to the new cluster center;

[0147] When the stability of the plurality of clustering results is greater than a preset stability, the plurality of clustering results are output.

[0148] In the specific implementation, first, the distance between the data and the center of each category is obtained by calculating the similarity between the features of each sensitive data. Then the data is classified into the category with the closest distance. If the distance to two categories is equal, one category can be selected. The initial cluster center can be randomly selected or manually selected.

[0149] For two vectors P = (p1, p2, ..., p n ) and Q=(q1,q2,...,q n ), the distance calculation function is defined as:

[0150]

[0151] This function can be used to measure the distance between enterprise sensitive feature arrays and analyze the feature differences of sensitive feature entities.

[0152] After all data are divided into corresponding categories, the new cluster center of each category is calculated, and then the old cluster center is updated with the new cluster center. The central eigenvalue is calculated as follows:

[0153]

[0154] represents the cluster center value of the jth feature after the m+1th iteration, Ν(i) is the neighborhood set of the ith data point, |Ν(i)| is the number of data points adjacent to it, is the median value of the kth data point in the jth feature dimension.

[0155] Each time the cluster center is updated, the median information of the data points in the neighborhood is weighted and summed. The clustering effect is more stable and the clustering results are not easily affected by extreme values, thereby improving the robustness and accuracy of the clustering.

[0156] At the same time, all data are re-clustered according to the updated cluster centers, and the above steps are repeated and iterated until all categories are stable. In this way, the data can be accurately assigned to the corresponding categories.

[0157] This embodiment calculates the similarity between sensitive data features, determines the distance between the data and the center of each category, and classifies the data into the nearest category. If the distance is equal, any category is selected. The initial cluster center can be randomly or manually selected. After the data is classified, the new cluster center of each category is calculated, and the old cluster center is updated with the new cluster center. The calculation of the new cluster center takes into account the median information of the data points in the field. The robustness and accuracy of clustering are improved by weighted summation. The updated cluster center is then used to reclassify the data, and the iterative process is repeated until the category is stable, ensuring that the data is accurately assigned to the corresponding category and improving the accuracy of data clustering assignment.

[0158] For example, in order to help understand the implementation process of the adaptive sensitive information intelligent identification method obtained by combining this embodiment with the above-mentioned embodiment 1, please refer to Figure 8 , Figure 8 A brief flow chart of an adaptive sensitive information intelligent identification method is provided, specifically: a knowledge vector database of enterprise sensitive information is constructed through processes such as data cleaning and manual sorting. The knowledge vector data will be used in two places. One is to use the labeled data to fine-tune the model, and the other is to serve as the basis for knowledge retrieval, provide reliable context examples for the model, and assist the model in identifying sensitive information to fine-tune the model and obtain a lightweight enterprise sensitive information large model used in this embodiment. The model will be used in two places. One is to perform sensitive information entity recognition, and the other is to perform self-verification of the recognition results, knowledge retrieval and sensitive information recognition. Extract the block features of the enterprise flow data, retrieve the most relevant context examples from the vector database through knowledge retrieval enhancement, combine the context examples with sensitive information recognition instructions to build an instruction project, assist the model in identifying sensitive information, and use the general knowledge ability of the large model to verify the sensitive information recognition results, exclude some obviously unreasonable results, and output the final recognition results.

[0159] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the adaptive sensitive information intelligent identification method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0160] This application also provides an adaptive sensitive information intelligent identification device, please refer to Fig. 9 , the adaptive sensitive information intelligent identification device comprises:

[0161] The data processing module 10 is used to process the batch data of the enterprise sensitive information flow into blocks to obtain data blocks, and perform feature extraction and vectorization processing on the data blocks to obtain feature vectors;

[0162] A binary tree construction module 20, used to construct a retrieval binary tree based on the batch data of the enterprise sensitive information flow;

[0163] A fuzzy retrieval module 30, configured to perform fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set;

[0164] The precise retrieval module 40 is used to perform feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set, respectively, to obtain a preset number of related contexts, and to reorder the preset number of related contexts to obtain a target context;

[0165] The information identification module 50 is used to input the target context into the large language model to obtain the enterprise sensitive information identification result.

[0166] In one embodiment, the binary tree construction module 20 is also used to clean the sensitive batch data of the enterprise, and to annotate and extract the cleaned data to obtain sensitive entities and sensitive information categories; to perform cluster analysis based on the sensitive entities and the sensitive information categories to obtain multiple clustering results of the sensitive entities; to fuse the corresponding feature vector sets in the multiple clustering results respectively to obtain a feature vector index; to determine the division point of the feature vector index in the first dimension, and to create a node with the index vector corresponding to the division point, the node including the current index value, node pointers pointing to the left child node and the right child node, and the feature set pointer of the feature set corresponding to the node; to divide the index vector of the division point smaller than the first dimension into the left subtree of the division point, and to divide the index vector of the division point larger than the first dimension into the right subtree of the division point, to obtain a binary tree of the first dimension; based on the dimensional order, the root node of the binary tree of the second dimension is used as the left child node and the right child node of the binary tree of the first dimension to obtain a retrieval binary tree, and the dimensional order of the first dimension is located before the dimensional order of the second dimension.

[0167] In one embodiment, the binary tree construction module 20 is also used to obtain sensitive data features based on the sensitive entity and the sensitive information category; determine an initial cluster center, determine an initial cluster center vector based on the initial cluster center, and obtain a sensitive data feature vector based on the sensitive data features; obtain a vector distance based on the initial cluster center vector and the sensitive data feature vector; determine a neighborhood set based on the vector distance, and determine the number of data points in the neighborhood set and the median of the data points in the feature dimension, obtain a new cluster center based on the neighborhood set, the number of data points and the median, and obtain multiple clustering results based on the new cluster center; when the stability of multiple clustering results is greater than a preset stability, output multiple clustering results.

[0168] In one embodiment, the fuzzy retrieval module 30 is further used to convert the feature vector into a query vector, compare the query vector with the dimension value of the index vector of the tree node of the retrieval binary tree, and obtain a comparison result; when the comparison result is that the feature value of the query vector is less than the dimension value of the index vector of the tree node of the retrieval binary tree, recursively search the left subtree of the retrieval binary tree; when the comparison result is that the feature value of the query vector is greater than the dimension value of the index vector of the tree node of the retrieval binary tree, recursively search the right subtree of the retrieval binary tree; when a leaf node is recursively found, recursive backtracking is performed to obtain a candidate node list; the Manhattan distance is calculated based on the candidate nodes in the candidate node list and the query vector to determine the target node; and a sensitive entity feature set is determined based on the target node.

[0169] In one embodiment, the precise retrieval module 40 is also used to perform feature vector retrieval on the feature vector of the sensitive entity feature set, calculate the feature similarity between the feature vector and the first vector, the first vector is the feature retrieval vector; obtain a preset number of first related contexts based on the feature similarity; perform semantic structure information retrieval on the feature vector of the sensitive entity feature set, obtain a semantic retrieval score based on the feature vector and the second vector, the second vector is the semantic structure vector; obtain a preset number of second related contexts based on the semantic retrieval score; calculate reordering scores for the first related context and the second related context, arrange the reordering scores from high to low, select related contexts corresponding to the target number of reordering scores, and obtain the target context.

[0170] In one embodiment, the information identification module 50 is also used to determine a verification task, perform entity extraction on sentences in the verification task, and receive a verification result based on the entity; when the verification result is a misjudgment, the large language model is updated based on the verification result, and enterprise sensitive information is identified based on the updated large language model to obtain an enterprise sensitive information identification result.

[0171] The adaptive sensitive information intelligent identification device provided by the present application adopts the adaptive sensitive information intelligent identification method in the above embodiment, which can solve the technical problems that the regular expression matching method relies on empirical knowledge, has incomplete rules, is prone to misjudgment and missed judgment, cannot be adaptively updated, and has weak generalization. Compared with the prior art, the beneficial effects of the adaptive sensitive information intelligent identification device provided by the present application are the same as the beneficial effects of the adaptive sensitive information intelligent identification method provided by the above embodiment, and the other technical features in the adaptive sensitive information intelligent identification device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0172] The present application provides an adaptive intelligent identification device for sensitive information, 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 adaptive intelligent identification method for sensitive information in the above-mentioned embodiment one.

[0173] Reference below Fig.10 , which shows a schematic diagram of the structure of an adaptive sensitive information intelligent identification device suitable for implementing the embodiment of the present application. The adaptive sensitive information intelligent identification device in the embodiment of the present application may include but is 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. Fig.10 The adaptive sensitive information intelligent identification 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.

[0174] like Fig.10As shown, the adaptive sensitive information intelligent recognition device may include a processing device 1001 (such as a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM: Read Only Memory) 1002 or the program loaded from the storage device 1003 to the random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the xxx device are also stored. The processing device 1001, ROM1002 and RAM1004 are connected to each other through a bus 1005. The 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 adaptive sensitive information intelligent identification device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows an adaptive sensitive information intelligent identification 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.

[0175] 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.

[0176] The adaptive sensitive information intelligent identification device provided by the present application adopts the adaptive sensitive information intelligent identification method in the above embodiment, which can solve the technical problems that the regular expression matching method relies on empirical knowledge, has incomplete rules, is prone to misjudgment and missed judgment, cannot be adaptively updated, and has weak generalization. Compared with the prior art, the beneficial effects of the adaptive sensitive information intelligent identification device provided by the present application are the same as the beneficial effects of the adaptive sensitive information intelligent identification method provided by the above embodiment, and the other technical features in the adaptive sensitive information intelligent identification device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0177] 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.

[0178] 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.

[0179] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the adaptive sensitive information intelligent identification method in the above-mentioned embodiment.

[0180] 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.

[0181] The computer-readable storage medium may be included in the adaptive sensitive information intelligent identification device; or may exist independently without being assembled into the adaptive sensitive information intelligent identification device.

[0182] The computer-readable storage medium carries one or more programs. When the one or more programs are executed by the adaptive sensitive information intelligent identification device, the adaptive sensitive information intelligent identification device:

[0183] The batch data of the enterprise sensitive information flow is processed into blocks to obtain data blocks, and the data blocks are subjected to feature extraction and vectorization processing to obtain feature vectors;

[0184] Constructing a retrieval binary tree based on the batch data of the enterprise sensitive information flow;

[0185] Perform fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set;

[0186] Performing feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reordering the preset number of related contexts to obtain a target context;

[0187] The target context is input into a large language model to obtain enterprise sensitive information recognition results.

[0188] 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).

[0189] 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.

[0190] 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.

[0191] The readable storage medium provided by the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned adaptive sensitive information intelligent identification method, which can solve the technical problems that the regular expression matching method relies on empirical knowledge, has incomplete rules, is prone to misjudgment and missed judgment, cannot be adaptively updated, and has weak generalization. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as the beneficial effects of the adaptive sensitive information intelligent identification method provided by the above-mentioned embodiment, and will not be repeated here.

[0192] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned adaptive sensitive information intelligent identification method.

[0193] The computer program product provided by the present application can solve the technical problems that the regular expression matching method relies on empirical knowledge, has incomplete rules, is prone to misjudgment and missed judgment, cannot be adaptively updated, and has weak generalization. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as the beneficial effects of the adaptive sensitive information intelligent identification method provided by the above embodiment, and will not be repeated here.

[0194] 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. An adaptive sensitive information intelligent identification method, characterized in that: The method includes: The batch data of the enterprise sensitive information flow is processed into blocks to obtain data blocks, and the data blocks are subjected to feature extraction and vectorization processing to obtain feature vectors; Constructing a retrieval binary tree based on the batch data of the enterprise sensitive information flow; Perform fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set; Performing feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reordering the preset number of related contexts to obtain a target context; The target context is input into a large language model to obtain enterprise sensitive information recognition results.

2. The method according to claim 1, characterized in that The step of constructing a retrieval binary tree based on the batch data of the enterprise sensitive information flow includes: Cleaning the sensitive batch data of the enterprise, and annotating and extracting the cleaned data to obtain sensitive entities and sensitive information categories; Performing cluster analysis according to the sensitive entities and the sensitive information categories to obtain multiple clustering results of the sensitive entities; The corresponding feature vector sets in the plurality of clustering results are respectively merged to obtain a feature vector index; Determine a division point of the feature vector index in the first dimension, and create a node with the index vector corresponding to the division point, wherein the node includes a current index value, a node pointer pointing to a left child node and a right child node, and a feature set pointer of a feature set corresponding to the node; Divide the index vectors of the partition points smaller than the first dimension into the left subtree of the partition points, and divide the index vectors of the partition points larger than the first dimension into the right subtree of the partition points, to obtain a binary tree of the first dimension; Based on the dimension order, the root node of the binary tree of the second dimension is used as the left child node and the right child node of the binary tree of the first dimension to obtain a retrieval binary tree, and the dimension order of the first dimension is located before the dimension order of the second dimension.

3. The method according to claim 2, characterized in that The step of performing cluster analysis according to the sensitive entities and the sensitive information categories to obtain multiple clustering results of the sensitive entities includes: Obtaining sensitive data features according to the sensitive entity and the sensitive information category; Determine an initial cluster center, determine an initial cluster center vector according to the initial cluster center, and obtain a sensitive data feature vector according to the sensitive data feature; Obtaining a vector distance according to the initial cluster center vector and the sensitive data feature vector; Determine a neighborhood set according to the vector distance, and determine the number of data points in the neighborhood set and the median of the data points in the feature dimension, obtain a new cluster center according to the neighborhood set, the number of data points and the median, and obtain multiple clustering results according to the new cluster center; When the stability of the plurality of clustering results is greater than a preset stability, the plurality of clustering results are output.

4. The method according to claim 1, characterized in that The step of performing fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set comprises: Converting the feature vector into a query vector, and comparing the query vector with the dimension value of the index vector of the tree node of the retrieval binary tree to obtain a comparison result; When the comparison result is that the characteristic value of the query vector is less than the dimension value of the index vector of the tree node of the search binary tree, recursively search the left subtree of the search binary tree; when the comparison result is that the characteristic value of the query vector is greater than the dimension value of the index vector of the tree node of the search binary tree, recursively search the right subtree of the search binary tree; When a leaf node is found recursively, recursive backtracking is performed to obtain a list of candidate nodes; Calculate the Manhattan distance between the candidate nodes in the candidate node list and the query vector to determine the target node; A sensitive entity feature set is determined according to the target node.

5. The method according to claim 1, characterized in that The step of performing feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set to obtain a preset number of related contexts, and reordering the preset number of related contexts to obtain the target context comprises: Performing feature vector retrieval on the feature vector of the sensitive entity feature set, and calculating feature similarity between the feature vector and a first vector, where the first vector is a feature retrieval vector; Obtaining a preset number of first relevant contexts according to the feature similarity; Performing semantic structure information retrieval on the feature vector of the sensitive entity feature set, and obtaining a semantic retrieval score according to the feature vector and a second vector, wherein the second vector is a semantic structure vector; Obtaining a preset number of second relevant contexts according to the semantic retrieval score; Reordering scores are calculated for the first relevant context and the second relevant context, the reordering scores are arranged from high to low, and relevant contexts corresponding to a target number of reordering scores are selected to obtain a target context.

6. The method according to any one of claims 1 to 5, characterized in that After the step of inputting the target context into the large language model to obtain the enterprise sensitive information identification result, the method further includes: Determine a verification task, perform entity extraction on a sentence in the verification task, and receive a verification result based on the entity; When the verification result is a misjudgment, the large language model is updated based on the verification result, and enterprise sensitive information is identified based on the updated large language model to obtain an enterprise sensitive information identification result.

7. An adaptive sensitive information intelligent identification device, characterized in that: The device comprises: The data processing module is used to process the batch data of the enterprise sensitive information flow in blocks to obtain data blocks, and perform feature extraction and vectorization processing on the data blocks to obtain feature vectors; A binary tree construction module, used to construct a retrieval binary tree based on the batch data of the enterprise sensitive information flow; A fuzzy retrieval module, used for performing fuzzy retrieval based on the retrieval binary tree and the feature vector to obtain a sensitive entity feature set; An accurate retrieval module, used to perform feature vector retrieval and semantic structure information retrieval on the feature vectors of the sensitive entity feature set, respectively, to obtain a preset number of related contexts, and to reorder the preset number of related contexts to obtain a target context; The information identification module is used to input the target context into the large language model to obtain the enterprise sensitive information identification result.

8. An adaptive sensitive information intelligent recognition 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 adaptive sensitive information intelligent identification method as described in 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 adaptive sensitive information intelligent identification method as described in 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 adaptive sensitive information intelligent identification method as described in any one of claims 1 to 6 are implemented.

Images