Data sharing security decision-making methods and systems using artificial intelligence

Abstract

This invention provides a data sharing security decision-making method and system applying artificial intelligence. It receives data sharing task description information and associated raw talent data sets as input into a context-aware mapping network to generate a task data association strength mapping table. This table is then matched with a pre-set set of basic security labels to resolve conflicts between data sharing requirements and security constraints, generating a fine-grained data protection guidance document. The corresponding data units in the raw talent data set undergo structured de-identification transformation to obtain intermediate data forms. Based on access permission rules and transmission encryption rules, a set of corresponding access verification rules and a set of transmission encryption rules are generated. These three sets are deployed to a controlled collaborative computing environment for joint data sharing analysis. Data access behavior undergoes permission verification and encryption compliance checks. After the joint analysis is completed, the sharing analysis results are output. This invention ensures the compliance and reliability of the sharing analysis result report.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a data sharing security decision-making method and system that applies artificial intelligence. Background Technology

[0002] With the deepening application of artificial intelligence technology in the field of data security, analyzing and controlling security risks during the data sharing process can protect data privacy and security while meeting data sharing needs. Currently, the common approach is to pre-configure a fixed set of security rules, perform data anonymization on the original data before sharing, and control the scope of data access according to a preset access permission list. Some methods also introduce simple rule matching algorithms to adjust the security strategy during the sharing process. However, these existing methods often suffer from insufficient adaptability between security strategies and specific sharing needs in practical applications. This leads to either over-protection that restricts the effective use of data or insufficient protection that introduces data security risks when facing complex data sharing scenarios. Furthermore, data anonymization and security rule formulation are often carried out independently, which can easily result in mismatches between the anonymized data format and the security rules. These problems collectively affect the accuracy and reliability of data sharing security decisions. Summary of the Invention

[0003] This invention provides a data sharing security decision-making method and system that applies artificial intelligence.

[0004] In a first aspect, embodiments of the present invention provide a data sharing security decision-making method applying artificial intelligence, comprising: receiving data sharing task description information and an associated set of original talent data; inputting the data sharing task description information and the set of original talent data into a preset context-aware mapping network, performing intent recognition on the task description information through a semantic parsing layer, extracting entity relationships from data units in the original talent data set through an entity association layer, and generating a task data association strength mapping table by combining the intent recognition result and the entity relationship extraction result; invoking a pre-configured data protection policy generator, performing policy matching between the task data association strength mapping table and a preset set of basic security labels, coordinating the conflict between data sharing requirements and security constraints through a multi-objective trade-off layer of the policy generator, and generating fine-grained numbers for each data unit in the original talent data set. According to the protection guidelines document, this document contains data anonymization rules, access permission rules, and transmission encryption rules. Based on the data anonymization rules in the fine-grained data protection guidelines document, the corresponding data units in the original talent data set are subjected to structured anonymization transformation to obtain an intermediate data form that meets security requirements. Then, a matching set of access verification rules and a set of transmission encryption rules are generated according to the access permission rules and transmission encryption rules. The intermediate data form, the set of access verification rules, and the set of transmission encryption rules are deployed to a pre-defined controlled collaborative computing environment for data sharing and joint analysis. The built-in rule verification engine in the environment calls the access verification rule set in real time to verify the permissions of data access behavior, and calls the transmission encryption rule set to check the encryption compliance of the data transmission process. After the joint analysis operation is completed, the shared analysis results that meet the requirements of the data sharing task description information are output.

[0005] Secondly, embodiments of the present invention provide a solution.

[0006] The embodiments of this application have the following beneficial effects: This invention receives data sharing task description information and a set of raw talent data. It utilizes a context-aware mapping network to perform intent recognition processing on the task description information and entity relationship extraction processing on data units to generate a task data association strength mapping table. This enables semantic-level dynamic association between data sharing tasks and data units, upgrading the matching between task requirements and data entities from traditional static keyword matching to deep association based on intent and entity relationships. This lays the foundation for accurate generation of subsequent security strategies. The invention then calls a data protection strategy generator to match the task data association strength mapping table with a basic security tag set. A multi-objective trade-off layer handles conflict coordination between data sharing requirements and security constraints to generate fine-grained data protection guidance documents. This dynamically balances multi-dimensional objective conflicts in the data sharing process, transforming security strategies from a single security-oriented or requirement-oriented approach to a collaborative optimization of security and requirements. This avoids the over-protection or under-protection that may result from traditional static rules. This approach enhances the applicability and flexibility of data protection strategies. By performing structured de-identification transformation on the original talent data set according to fine-grained data protection guidelines to obtain intermediate data forms, and simultaneously generating corresponding access verification rule sets and transmission encryption rule sets, it achieves integrated linkage between data de-identification processing and security rule formulation. This ensures a precise correspondence between the de-identified data form and the access and transmission rules, avoiding security vulnerabilities caused by a disconnect between data processing and rule configuration, and enhancing the security and controllability of the entire data lifecycle. Deploying the intermediate data form, access verification rule sets, and transmission encryption rule sets to a controlled collaborative computing environment, and using a rule verification engine and encryption compliance check module for real-time permission verification and encryption compliance checks, enables dynamic security monitoring of the data sharing and analysis process. This extends security verification from traditional pre-configuration to real-time monitoring throughout the entire process, effectively preventing unauthorized access and transmission risks during analysis and ensuring the compliance and reliability of shared analysis results reports. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the structure of the computer system provided in the embodiments of this application; Figure 2 This is a flowchart illustrating the data sharing security decision-making method using artificial intelligence provided in this application embodiment. Detailed Implementation

[0008] The data sharing security decision-making method using artificial intelligence provided in this invention is applied to a computer system. In some embodiments, the computer system can be a laptop, a desktop computer, an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0009] See Figure 1 This is a schematic diagram of the structure of the computer system provided in the embodiments of this application. Figure 1 The computer system shown includes at least one processor 210, memory 250, at least one network interface 220, and an external interface 230. The various components in the computer system 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 1 The general labeled all buses as Bus System 240.

[0010] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0011] External interface 230 may include, for example, one or more speakers and / or one or more visual displays. External interface 230 may also include one or more input devices 432, such as a keyboard, mouse, microphone, touch screen display, camera, etc.

[0012] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0013] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this application embodiment is intended to include any suitable type of memory.

[0014] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0015] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; network communication module 252 is used to reach other communication devices, such as shared devices, via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 include: Bluetooth, WiFi, and Universal Serial Bus (USB), etc.; presentation module 253 is used to enable the display of information (e.g., external interfaces for operating peripheral devices and displaying content and information) via one or more output devices 231 associated with external interface 230 (e.g., display screen, speaker, etc.); input processing module 254 is used to detect and translate one or more user inputs or interactions from one or more input devices 232.

[0016] See Figure 2 This is a flowchart illustrating the data sharing security decision-making method using artificial intelligence provided in this application embodiment. Next, we will combine... Figure 2 The steps shown are explained.

[0017] Step S100: Receive data sharing task description information and associated raw talent data set, wherein the data sharing task description information includes a description of data usage scenarios, a description of data requirement types, and a description of the scope of data sharing, and the raw talent data set includes multi-dimensional data units.

[0018] The data usage scenario description details how the data will be applied in actual business operations. For example, in a talent recruitment scenario, it might be used to screen candidates who meet specific job requirements; in a talent development scenario, it might be used to develop personalized training plans. The data requirement type description clarifies the specific type of data required, such as the candidate's educational background, work experience, and skill level. The data sharing scope description defines the boundaries of data sharing, including the data source range (e.g., which departments or databases within the company), time range (e.g., data from the past year or a specific time period), and data granularity (e.g., detailed data or aggregated data).

[0019] The raw talent dataset is a collection of various data related to talent. Multi-dimensional data units mean that these data describe talent from multiple different angles and levels. For example, basic personal information includes name, gender, and age; educational background includes graduating institution, major, and degree; work experience includes previous companies, positions, and length of employment; and skills include professional skills and language proficiency. These multi-dimensional data units are interconnected and together constitute a comprehensive description of talent.

[0020] In practical applications, when an enterprise needs to share talent data across departments to support strategic decision-making, the system receives a data sharing task description. Assuming the data usage scenario is described as "screening talent with specific skills and experience for the company's strategic expansion projects," the data requirement type is described as "information on talent with relevant industry project experience and specific technical skills," and the data sharing scope is described as "data on all current employees within the company." Simultaneously, the system collects related raw talent data sets from various data sources (such as human resource management systems, project management systems, etc.), integrating this data to form a complete raw talent data set.

[0021] Step S200: Input the data sharing task description information and the original talent data set into a preset context-aware mapping network. Perform intent recognition on the task description information through the semantic parsing layer, extract entity relationships from the data units in the original talent data set through the entity association layer, and generate a task data association strength mapping table by combining the intent recognition results and the entity relationship extraction results.

[0022] In one implementation, step S200 may specifically include the following steps S210 to S260: Step S210: Input the data sharing task description information into the semantic parsing layer of the context-aware mapping network, segment the data in the task description information using scene description to obtain multiple scene description word units, perform part-of-speech tagging on the scene description word units, and select noun word units and verb word units as key scene elements.

[0023] Word segmentation is a fundamental operation in natural language processing, which divides continuous text into individual words according to certain rules. When segmenting data usage scenario descriptions, such as "evaluating the innovation capabilities of talents in projects," the segmented data can be divided into scenario description units such as "evaluate," "talents," "in," "project," "in," and "innovation capabilities."

[0024] Part-of-speech tagging assigns the part of speech (PSS) to each word, such as noun, verb, adjective, adverb, etc. In the example above, "talent," "project," and "innovation capability" are noun-type words, while "evaluation" is a verb-type word. Noun-type and verb-type words are selected as key scenario elements because these words typically express the core content of the data usage scenario more directly. Noun-type words clarify the subject and object in the scenario, while verb-type words indicate specific operations or behaviors.

[0025] For example, natural language processing tools, such as open-source NLP libraries, can be used, which include dedicated word segmenters and part-of-speech taggers. First, the data, using a scenario description as input, is passed to the word segmenter, which segments the text into scenario description word units according to preset rules. Then, these word units are fed into the part-of-speech tagger, which tags each word unit with a part of speech based on the grammatical and semantic features of the words. Finally, noun word units and verb word units are selected as key scenario elements based on their parts of speech.

[0026] Step S220: Perform dependency parsing on the data requirement type description, extract core requirement entity words and requirement modifiers, and construct requirement entity-modifier association pairs.

[0027] Dependency parsing is a method used to analyze the dependency relationships between words in a sentence, revealing its grammatical structure and semantic relationships. In data requirement type descriptions, core requirement entities are words that express the main requirements and are the core objects of the requirements. Requirement modifiers are words that further describe and limit the core requirement entities, clarifying the specific characteristics and conditions of the requirements.

[0028] For example, in the data requirement type description "software engineer with advanced programming skills," "software engineer" is the core requirement entity, and "with advanced programming skills" is the requirement modifier. The purpose of constructing requirement entity-modifier association pairs is to link the core requirement entity and the requirement modifier to form a clear requirement expression for subsequent processing and analysis. In practical applications, a deep learning-based dependency parsing model can be used. This model typically consists of an input layer, a hidden layer, and an output layer. The input layer receives the text describing the data requirement type, the hidden layer extracts and transforms the text's features, and the output layer outputs the dependency relationships between the words. Based on the output dependency relationships, the core requirement entity and requirement modifier are extracted, and requirement entity-modifier association pairs are constructed.

[0029] Step S230: Associate key scene elements and demand entities-modifiers with the intent classifier of the input semantic parsing layer, and perform sequence labeling through the bidirectional LSTM network therein to generate a task intent label sequence. The task intent label sequence includes at least one of data query intent label, data statistics intent label and data modeling intent label.

[0030] The intent classifier is a module used to classify task intents. It receives key scene elements and demand entity-modifier association pairs as input and determines the type of task intent through internal algorithms and models. The Bidirectional Long Short-Term Memory Network (LSTM) is a deep learning network with the ability to process sequence data bidirectionally, considering both context and semantic information to more accurately capture semantic information. Sequence labeling is the process of assigning a label to each element in a sequence. In this step, the Bidirectional LSTM network processes the sequence composed of key scene elements and demand entity-modifier association pairs, assigning a corresponding intent label to each element based on its semantic information. Data query intent labels indicate that the task intent is to query specific data, such as querying information on talents with certain skills; data statistics intent labels indicate that the task intent is to perform statistical analysis on data, such as calculating the average years of service of talents in different departments; data modeling intent labels indicate that the task intent is to build a data model, such as building a talent performance evaluation model.

[0031] In one implementation, step S230 may specifically include the following steps S231 to S235: Step S231: Perform word vector conversion on the noun and verb units in the key scene elements to generate a scene element word vector sequence. Concatenate the word vectors of the entity words and modifiers in the demand entity-modifier association pair to generate a demand association pair word vector sequence.

[0032] Word vectorization is the process of converting words into vector representations. This maps words in natural language to a vector space, enabling computers to better process and understand them. When performing word vectorization on noun and verb units within key scene elements, pre-trained word vector models can be used, such as those trained on large-scale corpora. These models learn word co-occurrence relationships from a large amount of text, converting each word into a fixed-length vector. The vector's dimension is typically determined by the model's training settings.

[0033] For example, for key scenario elements "talent" and "assessment," a pre-trained word vector model is used to convert "talent" into a vector and "assessment" into a vector. These vectors form a sequence of scenario element word vectors. For demand entity-modifier association pairs, such as (software engineer, with advanced programming skills), "software engineer" and "with advanced programming skills" are converted into word vectors respectively. Then, these two word vectors are concatenated along their vector dimensions to obtain a longer vector. Multiple such vectors form a sequence of demand association pair word vectors.

[0034] Step S232: Associate the scene element word vector sequence with the demand. Input the word vector sequence into the input layer of the bidirectional LSTM network. The forward LSTM unit extracts the temporal features of the word vector sequence from left to right to obtain the forward temporal feature vector. The backward LSTM unit extracts the temporal features of the word vector sequence from right to left to obtain the backward temporal feature vector.

[0035] A bidirectional LSTM network consists of forward LSTM units and backward LSTM units. The forward LSTM unit starts from the first element of the sequence and processes each element sequentially, extracting temporal features from left to right. The backward LSTM unit starts from the last element of the sequence and processes each element in reverse, extracting temporal features from right to left.

[0036] In the input layer, the sequence of scene element word vectors and the sequence of demand association word vectors are fed into the bidirectional LSTM network. The forward LSTM unit updates the hidden state at the current time step based on the current input and the hidden state at the previous time step, thereby extracting the temporal features from left to right. For example, for the scene element word vector sequence "[talent vector, evaluation vector]", the forward LSTM unit first processes the "talent vector" to obtain a hidden state, and then uses this hidden state and the "evaluation vector" as input to update the hidden state at the current time step, finally obtaining the forward temporal feature vector.

[0037] The backward LSTM unit works similarly, but in the reverse order: it processes the last element of the sequence first, then proceeds forward sequentially to obtain the backward temporal feature vector. By combining forward and backward LSTM units, the bidirectional LSTM network can simultaneously consider the context of the sequence, capturing the temporal features more comprehensively.

[0038] Step S233: Concatenate the forward temporal feature vector and the backward temporal feature vector along the feature dimension to generate a bidirectional fused feature vector. Input the bidirectional fused feature vector into the fully connected layer of the classifier and perform a nonlinear transformation through the activation function to obtain the intent classification feature vector.

[0039] Concatenating the forward and backward temporal feature vectors along the feature dimension fuses temporal feature information from left to right and right to left, resulting in a more comprehensive and richer feature representation. For example, if the forward and backward temporal feature vectors are both vectors, concatenating them along the vector dimension yields a longer bidirectional fused feature vector. Fully connected layers in neural networks multiply the input feature vectors by a set of learnable weight matrices, add bias terms, and then perform a non-linear transformation using an activation function. The activation function introduces non-linearity, enabling the network to learn more complex patterns and features. Common activation functions include ReLU (Rectified Linear Unit) and Sigmoid.

[0040] In this step, the bidirectional fused feature vector is input into a fully connected layer. The fully connected layer performs a linear transformation on the input based on the weight matrix and bias terms, and then uses an activation function to perform a non-linear mapping on the result of the linear transformation. The final intent classification feature vector contains key information for intent classification and can more accurately represent the intent features of the task.

[0041] Step S234: Call the CRF layer of the intent classifier to perform sequence labeling and decoding on the intent classification feature vector, calculate the label transition probability matrix, find the label path with the highest probability based on the Viterbi algorithm, and generate the initial intent label sequence.

[0042] The CRF (Conditional Random Fields) layer is a probabilistic graphical model used for sequence labeling, which can consider the dependencies between labels in a sequence. When decoding sequence labels from intent-classification feature vectors, the CRF layer calculates a label transition probability matrix, which represents the probability of transitioning from one label to another. This probability matrix is ​​learned based on training data and reflects the transition patterns and probabilities between different labels.

[0043] The Viterbi algorithm is a dynamic programming algorithm used to find the label path with the highest probability given a model and observation sequence. In this step, based on the calculated label transition probability matrix and intent classification feature vector, the Viterbi algorithm finds the path with the highest probability from all possible label paths. The label sequence corresponding to this path is the initial intent label sequence. For example, assuming possible intent labels are "data query intent label," "data statistics intent label," and "data modeling intent label," the Viterbi algorithm will determine the most probable intent label at each time step based on probability calculations, thus generating the initial intent label sequence.

[0044] Step S235: Perform post-processing on the initial intent label sequence to remove redundant and conflicting labels, and merge semantically similar intent labels to obtain an optimized task intent label sequence.

[0045] Redundant tags are tags that appear repeatedly in the initial intent tag sequence or do not substantially contribute to the expression of the intent. Conflicting tags are tags that contradict each other logically. For example, if a sequence contains both a "data query intent tag" and a "data delete intent tag," but the current task obviously cannot perform query and delete operations simultaneously, these two tags are conflicting tags. Merging semantically similar intent tags simplifies the tag sequence and improves the efficiency of tag expression. For example, the "detailed data query intent tag" and the "basic data query intent tag" are semantically similar and can be merged into a single "data query intent tag."

[0046] For example, the post-processing operation can be completed by formulating a series of rules and strategies. First, based on preset rules, it is determined which tags are redundant and conflicting, and these are removed from the initial intent tag sequence. Then, semantic similarity calculation methods, such as word vector-based similarity calculation methods, are used to determine which tags are semantically similar, and they are merged into a single tag. Finally, an optimized task intent tag sequence is obtained, which more concisely and accurately expresses the task's intent.

[0047] Step S240: Input the original talent data set into the entity association layer of the context-aware mapping network, perform named entity recognition on the text data in the data unit, extract the data entity name and entity type, extract the entity attributes on the numerical data, and obtain the entity attribute key-value pair set.

[0048] Named entity recognition (NER) identifies entities with specific meanings within text, such as names of people, places, organizations, and times. In a raw talent dataset, textual data may contain information such as names of individuals, companies, schools, and projects. NER can extract these entity names and determine their types. For example, "Zhang San" is a person's name, "XX Technology Co., Ltd." is an organization name, and "XX University" is a school name.

[0049] Entity attribute extraction for numerical data involves identifying the entity attributes represented by these values ​​and representing them as key-value pairs. For example, in work experience data for talent, "years of service" is numerical data, and entity attribute extraction can yield key-value pairs (years of service, corresponding numerical value).

[0050] In practical applications, deep learning-based named entity recognition models, such as those based on BERT (Bidirectional Encoder Representations from Transformers), can be used. These models learn rich linguistic knowledge and semantic information through large-scale pre-training, enabling them to accurately identify named entities in text. Inputting textual data into the model will output the names and types of the identified data entities. For numerical data, entity attribute key-value pairs can be determined using a data dictionary or metadata. The data dictionary records the meaning and attributes of each data field, allowing numerical data to be converted into key-value pair format.

[0051] Step S250: Extract entity relation triples from the entity name, entity type and entity attribute key-value pair set using the relation extraction model of the entity association layer, and generate an entity relation triple set containing the head entity, relation type and tail entity.

[0052] Relation extraction models are used to extract relationships between entities from data. They can analyze information from entity names, entity types, and key-value pairs of entity attributes to identify the inherent connections between different entities. An entity-relation triple is a structure representing a relationship between entities, consisting of a head entity, a relation type, and a tail entity. For example, in talent data, (Zhang San, employed at, XX Technology Co., Ltd.) is an entity-relation triple, where "Zhang San" is the head entity, "employed at" is the relation type, and "XX Technology Co., Ltd." is the tail entity.

[0053] In one implementation, step S250 may specifically include the following steps S251 to S255: Step S251: Input the entity name, entity type, and the set of entity attribute key-value pairs into the feature construction layer of the relation extraction model. Perform character-level feature extraction on the entity name to generate an entity name character feature vector, perform one-hot encoding on the entity type to generate an entity type encoding vector, and perform word vector transformation on the keys and values in the set of entity attribute key-value pairs respectively to generate an attribute key word vector and an attribute value word vector.

[0054] Character-level feature extraction is to split the entity name into characters, then extract the features of each character, and finally combine them into an entity name character feature vector. For example, for the entity name "Zhang San", it is split into "Zhang" and "San", and the features of each character are extracted, such as the encoding information of the character, its position in the character table, the stroke features of the character, etc., and then these features are combined into a vector.

[0055] One-hot encoding is to convert the entity type into a binary vector, where only one position is 1 and the rest are 0. For example, the entity types are "person name", "organization name", "place name". If the entity type is "person name", the corresponding one-hot encoding vector may be [1, 0, 0].

[0056] Perform word vector transformation on the keys and values in the set of entity attribute key-value pairs respectively, and use a pre-trained word vector model to convert the keys and values into vector representations. For example, for the key-value pair ("working years", corresponding numerical value), convert "working years" into an attribute key word vector and convert "corresponding numerical value" into an attribute value word vector.

[0057] Step S252: Fuse the entity name character feature vector, entity type encoding vector, attribute key word vector, and attribute value word vector to generate an entity comprehensive feature vector. Input the entity comprehensive feature vector into the Bi-GRU layer of the relation extraction model, and perform temporal dependence modeling on the feature vector through a gated recurrent unit to obtain an entity temporal feature vector.

[0058] Feature fusion is to combine different types of feature vectors into a more comprehensive feature vector. The entity name character feature vector, entity type encoding vector, attribute key word vector, and attribute value word vector can be concatenated in the vector dimension to obtain an entity comprehensive feature vector. The Bi-GRU (Bidirectional Gated Recurrent Unit) layer is a bidirectional gated recurrent unit network, which consists of a forward GRU unit and a backward GRU unit. The gated recurrent unit controls the flow of information through a gating mechanism and can effectively capture long-term dependencies in sequence data.

[0059] The entity composite feature vector is input into the Bi-GRU layer. The forward GRU unit processes the sequence starting from the first element, while the backward GRU unit processes the sequence starting from the last element. The combination of the two can better capture the temporal dependencies between entities. During processing, the GRU unit updates the hidden state at the current time step based on the input at the current time step and the hidden state at the previous time step, ultimately obtaining the entity temporal feature vector.

[0060] Step S253: Call the attention mechanism layer of the relation extraction model to calculate the attention of the entity temporal feature vector, generate the entity attribute attention weight vector, and perform weighted summation on the entity temporal feature vector based on the weight vector to obtain the entity feature vector with enhanced attribute association.

[0061] The attention mechanism layer is designed to make the model focus more on the important parts of the sequence. In this step, the attention mechanism layer calculates the attention level of entity temporal feature vectors and assigns an attention weight to each feature vector. The entity attribute attention weight vector represents the importance of each entity attribute in the overall sequence; the larger the weight value, the more important the attribute.

[0062] Weighted summation of entity temporal feature vectors based on weight vectors involves multiplying each entity temporal feature vector by its corresponding attention weight, and then summing all the results to obtain the entity feature vector that strengthens attribute associations. This highlights important entity attributes and improves the model's ability to identify entity relationships. For example, assuming there are multiple entity temporal feature vectors, namely vector A, vector B, vector C, etc., and corresponding attention weights A, B, C, etc., then the entity feature vector that strengthens attribute associations is weight A × vector A + weight B × vector B + weight C × vector C + ... Step S254: Input the entity feature vector with enhanced attribute association into the convolutional layer of the relationship extraction model, perform local feature extraction on the feature vector through multi-scale convolution kernel to obtain multi-scale entity relationship feature map, and perform global max pooling on the multi-scale entity relationship feature map to generate entity relationship feature vector.

[0063] A convolutional layer is a type of layer in deep learning that uses convolution kernels to perform convolution operations on the input feature vector, extracting local features. Multi-scale convolution kernels refer to using convolution operations with kernels of different sizes, allowing features to be extracted at different scales, resulting in richer local feature information. Different sized convolution kernels can capture features of different granularities; for example, small convolution kernels can capture local, detailed features, while large convolution kernels can capture more macroscopic features.

[0064] The multi-scale entity relationship feature map is the output of the convolutional layer, containing entity relationship features extracted from different scales. Global max pooling maximizes each channel in the multi-scale entity relationship feature map, compressing the feature map into a single vector. Global max pooling reduces the dimensionality of the features while preserving the most important feature information.

[0065] For example, convolutional layers can perform convolution operations in parallel using multiple convolutional kernels of different sizes, such as small, medium, and large kernels. The convolutional results are then concatenated to obtain a multi-scale entity relationship feature map. Finally, global max pooling is performed on this feature map to generate entity relationship feature vectors.

[0066] Step S255: Input the entity relationship feature vector into the classification layer of the relationship extraction model, calculate the probability value of each relationship type through the softmax function, select the relationship type with the highest probability value as the target relationship type between entities, construct an entity relationship triplet containing the head entity name, the target relationship type and the tail entity name, and combine multiple entity relationship triplets to form an entity relationship triplet set.

[0067] The classification layer is used to classify entity relationships. The softmax function is a commonly used activation function to transform the input vector into a probability distribution. In this step, the entity relationship feature vector is input into the classification layer, and after processing by the softmax function, a probability value for each relationship type is obtained. These probability values ​​represent the likelihood that entities belong to different relationship types.

[0068] The relationship type with the highest probability value is selected as the target relationship type between entities. For example, assuming the possible relationship types are "employed at", "graduated from", and "cooperate with", the softmax function calculates that "employed at" has the highest probability value, so "employed at" is the target relationship type.

[0069] Construct an entity relation triple that includes a head entity name, a target relation type, and a tail entity name, for example (Zhang San, works at XX Technology Co., Ltd.). Combining multiple such entity relation triples together forms an entity relation triple set.

[0070] Step S260: Input the task intent label sequence and entity relation triple set into the fusion computing layer of the context-aware mapping network, calculate the semantic similarity value between each intent label and entity relation triple, and construct a task data association strength mapping table based on the similarity value. The row dimension of the mapping table is the intent label in the intent label sequence, the column dimension is the triple identifier in the entity relation triple set, and the element is the corresponding semantic similarity value.

[0071] The fusion computing layer is a layer in the context-aware mapping network used to fuse different information, enabling comprehensive analysis and processing of task intent label sequences and entity relation triples. The semantic similarity value between each intent label and entity relation triple can be calculated using semantic similarity calculation methods, such as cosine similarity calculation based on word vectors.

[0072] Cosine similarity measures the similarity between two vectors by calculating the cosine of the angle between them; a value closer to 1 indicates a higher similarity. In this step, the intent label and entity relation triples are converted into vector representations, and then their cosine similarity is calculated. Specifically, the words in the intent label and entity relation triples are converted into word vectors, and these word vectors are then combined or averaged to obtain the intent label vector and entity relation triple vector. Finally, the cosine similarity between these two vectors is calculated.

[0073] A task data association strength mapping table is constructed based on similarity values. The row dimension represents the intent tags in the intent tag sequence, the column dimension represents the triple identifiers in the entity relationship triple set, and the element is the corresponding semantic similarity value. This mapping table can intuitively show which data entity relationships are strongly associated with each task intent, helping decision-makers better understand the relevance of the data, thus providing an important reference for the formulation of subsequent data protection strategies.

[0074] Step S300: Invoke the pre-configured data protection policy generator, match the task data association strength mapping table with the preset basic security label set, and coordinate the conflict between data sharing requirements and security constraints through the multi-objective trade-off layer of the policy generator to generate a fine-grained data protection guidance document for each data unit in the talent original data set. This document includes data desensitization rules, access permission rules and transmission encryption rules.

[0075] In one implementation, step S300 may specifically include the following steps S310 to S360: Step S310: Invoke the policy input layer of the data protection policy generator to receive the task data association strength mapping table and the preset basic security label set, wherein the basic security label set includes data sensitivity labels, data access level labels and data transmission security labels.

[0076] The policy input layer serves as the input interface for the data protection policy generator, receiving a task data association strength mapping table and a basic security label set. The task data association strength mapping table provides information on the association strength between task intents and data entity relationships, while the basic security label set provides security standards and a basis for subsequent policy matching and conflict coordination.

[0077] In practical applications, when the data protection policy generator starts, it calls the policy input layer, passing the task data association strength mapping table and the basic security label set as input. The policy input layer performs format checks and validations on this input data to ensure its integrity and accuracy before passing it to subsequent processing modules.

[0078] Step S320: Normalize the semantic similarity values ​​in the task data association strength mapping table to generate association strength normalization values; based on the relationship between the entity relation triples and the data units in the original talent data set, aggregate the association strength normalization values ​​corresponding to each triple to its corresponding data unit to generate a comprehensive association strength value for each data unit; based on the comprehensive association strength value, divide the data units in the original talent data set into high-association data units, medium-association data units, and low-association data units.

[0079] Normalization transforms data into a specific range, such as the [0, 1] interval, which eliminates differences in the dimensions and scales of the data, making different data comparable. To normalize the semantic similarity values ​​in the task data association strength mapping table, common normalization methods can be used, such as linear normalization. This method obtains the normalized association strength value by subtracting the minimum value from the original semantic similarity value and then dividing by the difference between the maximum and minimum values.

[0080] Based on the relationship between entity relation triples and data units in the original talent dataset, the normalized association strength value corresponding to each triple is aggregated to its corresponding data unit. For example, if an entity relation triple (Zhang San, works at, XX Technology Co., Ltd.) is associated with the talent data unit "Zhang San", then the normalized association strength value of that triple is added to the "Zhang San" data unit. Through this aggregation method, a comprehensive association strength value for each data unit can be obtained.

[0081] Based on the comprehensive correlation strength value, data units in the original talent dataset are divided into high-correlation data units, medium-correlation data units, and low-correlation data units. This division can be based on preset thresholds; for example, data units with a comprehensive correlation strength value greater than a certain high threshold are classified as high-correlation data units, those with a comprehensive correlation strength value between the high and low thresholds are classified as medium-correlation data units, and those with a comprehensive correlation strength value less than the low threshold are classified as low-correlation data units.

[0082] Step S330: Match the high-association data units, medium-association data units, and low-association data units with the data sensitivity tags in the basic security tag set to generate the initial security tag combination for each data unit.

[0083] Tag matching compares data units with varying degrees of relevance to data sensitivity tags in a basic set of security tags to identify the most suitable data sensitivity tag. Generally, highly relevant data units are matched with higher-sensitivity data sensitivity tags because these data units are more closely related to the task intent and may contain more sensitive information; moderately relevant data units are matched with moderately sensitive data sensitivity tags; and low-relevance data units are matched with lower-sensitivity data sensitivity tags.

[0084] For example, a mapping table can be established to record the correspondence between data units with different levels of correlation and data sensitivity labels. Based on this mapping table, high-correlation data units, medium-correlation data units, and low-correlation data units are matched with their corresponding labels to generate an initial security label combination for each data unit. In addition to the data sensitivity label, the initial security label combination may also include initial settings for data access level labels and data transmission security labels.

[0085] Step S340: Input the initial security label combination into the multi-objective trade-off layer of the data protection strategy generator, and extract the data usage frequency parameter, data processing timeliness parameter, and data accuracy requirement parameter from the data sharing requirements through the constraint condition parsing module within the layer, and extract the data desensitization level parameter, access control parameter, and transmission encryption strength parameter from the security constraints.

[0086] The multi-objective trade-off layer is the core module in the data protection policy generator, used to coordinate data sharing needs and security constraints. The constraint parsing module, a submodule within the multi-objective trade-off layer, is responsible for extracting key parameters from the input information. The data usage frequency parameter indicates how frequently the data is used in the data sharing task; for example, some data may need to be used daily, while some data may only need to be used occasionally. The data processing timeliness parameter indicates the time requirements for data processing, such as how long the data needs to be processed. The data precision requirement parameter indicates the required level of accuracy of the data, such as the number of decimal places to retain.

[0087] The data anonymization level parameter in the security constraints indicates the degree of data anonymization, such as light anonymization, medium anonymization, and heavy anonymization. The access control parameter specifies the scope and restrictions of access permissions for different users. The transmission encryption strength parameter indicates the encryption strength used during data transmission, such as high-strength encryption, medium-strength encryption, and low-strength encryption. In practical applications, the constraint parsing module parses the initial security label combination, extracting parameters related to data sharing requirements and security constraints. These parameters serve as the basis for subsequent conflict detection and coordination.

[0088] Step S350: Detect conflicts between the data usage frequency parameter and the data desensitization level parameter, the data processing timeliness parameter and the access control parameter, and the data accuracy requirement parameter and the transmission encryption strength parameter through a conflict coordination algorithm of the multi-objective trade-off layer, and generate conflict coordination results.

[0089] In one implementation, step S350 may specifically include the following steps S351 to S355: Step S351: For the divided high-association data units, medium-association data units, and low-association data units, respectively, input their data usage frequency parameters into the first conflict detection module of the conflict coordination algorithm, obtain the desensitization frequency threshold range corresponding to the association strength level of the data unit, and determine whether the data usage frequency parameter exceeds the upper limit or is lower than the lower limit of the corresponding desensitization frequency threshold range. If so, it is determined that there is a first conflict between the data usage frequency parameter and the data desensitization level parameter of the data unit, and a first conflict type identifier and conflict degree quantification value are generated.

[0090] The first conflict detection module is a submodule of the conflict coordination algorithm, used to detect conflicts between the data usage frequency parameter and the data anonymization level parameter. Different anonymization frequency threshold ranges are preset for data units with different association strength levels. These threshold ranges are determined based on extensive experimental and empirical data, reflecting a reasonable matching range between data usage frequency and data anonymization level under different association strengths.

[0091] The data usage frequency parameters for highly correlated, moderately correlated, and low-correlation data units are input into the first conflict detection module. The module obtains the corresponding desensitization frequency threshold range based on the correlation strength level of the data units. Then, the data usage frequency parameter is compared with this threshold range. If the data usage frequency parameter exceeds the upper limit of the threshold range, it indicates that the data is used too frequently, and the current data desensitization level may be too high, resulting in excessive restrictions on data use. If the data usage frequency parameter is below the lower limit of the threshold range, it indicates that the data is used too infrequently, and the current data desensitization level may be too low, posing a risk of data leakage.

[0092] When a first conflict is detected, the first conflict detection module generates a first conflict type identifier to identify the specific type of conflict. It also generates a conflict severity quantification value, which indicates the severity of the conflict. The conflict severity quantification value is calculated based on the deviation of the data usage frequency parameter from a threshold range; the greater the deviation, the higher the conflict severity quantification value.

[0093] Step S352: Input the data processing timeliness parameter into the second conflict detection module of the conflict coordination algorithm, obtain the preset permission verification timeliness threshold, calculate the absolute value of the difference between the data processing timeliness parameter and the permission verification timeliness threshold, if the absolute value of the difference exceeds the preset timeliness tolerance range, it is determined that there is a second conflict between the data processing timeliness parameter and the access control parameter, and a second conflict type identifier and conflict degree quantification value are generated.

[0094] The second conflict detection module is a submodule in the conflict coordination algorithm used to detect conflicts between data processing timeliness parameters and access control parameters. The preset access verification timeliness threshold is a time threshold set according to the system's performance and security requirements, representing the time required to complete access verification under normal circumstances.

[0095] After the data processing timeliness parameter is input into the second conflict detection module, the module obtains the preset permission verification timeliness threshold and calculates the absolute value of the difference between the data processing timeliness parameter and the permission verification timeliness threshold. If this absolute value exceeds the preset timeliness tolerance range, it indicates that there is a significant difference between the time requirements for data processing and the time requirements for permission verification, which may lead to data not being processed in a timely manner or permission verification being too cumbersome, thus causing a conflict.

[0096] When a second conflict is detected, the second conflict detection module generates a second conflict type identifier to identify the specific type of conflict, and also generates a conflict severity quantification value. The conflict severity quantification value can be calculated based on the degree of deviation between the absolute value of the difference and the timeliness tolerance range; the greater the deviation, the higher the conflict severity quantification value.

[0097] Step S353: Input the data precision requirement parameter into the third conflict detection module of the conflict coordination algorithm, obtain the preset encryption precision loss threshold, calculate the theoretical precision loss value after transmission encryption based on the data precision requirement parameter, if the theoretical precision loss value exceeds the encryption precision loss threshold, it is determined that there is a third conflict between the data precision requirement parameter and the transmission encryption strength parameter, and generate a third conflict type identifier and conflict degree quantification value.

[0098] The third conflict detection module is a submodule in the conflict coordination algorithm used to detect conflicts between the data precision requirement parameter and the transmission encryption strength parameter. The preset encryption precision loss threshold is a threshold set according to the importance of the data and the application scenario, representing the maximum degree of data precision loss allowed during transmission encryption.

[0099] After the data precision requirement parameters are input into the third collision detection module, the module obtains the preset encryption precision loss threshold and calculates the theoretical precision loss value after transmission encryption based on the data precision requirement parameters. The theoretical precision loss value is calculated based on the characteristics of the encryption algorithm and the data precision requirement parameters, and represents the degree of precision loss that the data may lose under the current transmission encryption strength.

[0100] If the theoretical precision loss value exceeds the encryption precision loss threshold, it indicates that the current transmission encryption strength may be too high, causing the data precision to fail to meet requirements, thus resulting in a collision. When a third collision is detected, the third collision detection module generates a third collision type identifier to identify the specific type of collision, and also generates a collision severity quantification value. The collision severity quantification value can be calculated based on the deviation between the theoretical precision loss value and the encryption precision loss threshold; the greater the deviation, the higher the collision severity quantification value.

[0101] Step S354: Input the first conflict type identifier, the second conflict type identifier, the third conflict type identifier and the corresponding conflict degree quantification value into the priority sorting module of the conflict coordination algorithm, sort the conflict types according to the preset conflict priority rules, and generate a conflict priority sequence.

[0102] The priority sorting module is a submodule in the conflict coordination algorithm used to determine the priority of conflict handling. The preset conflict priority rules are formulated based on factors such as the importance of the data sharing task, security requirements, and business needs, and specify the processing order of different types of conflicts.

[0103] After inputting the first conflict type identifier, the second conflict type identifier, the third conflict type identifier, and the corresponding conflict severity quantification value into the priority sorting module, the module will sort each conflict type according to the preset conflict priority rules. For example, the conflict between data usage frequency and data anonymization level may be specified as having the highest priority, because this conflict may directly affect the availability of data; the conflict between data processing timeliness and access control has the next highest priority; and the conflict between data accuracy requirements and transmission encryption strength has the lowest priority.

[0104] Finally, the priority sorting module generates a conflict priority sequence, which represents the processing order of each conflict type, providing guidance for subsequent conflict resolution.

[0105] Step S355: Based on the conflict priority sequence and the conflict degree quantification value, call the conflict resolution strategy library of the conflict coordination algorithm, match the corresponding conflict mitigation rules, and generate a conflict coordination result containing the conflict type, conflict priority, conflict mitigation rules, and rule application order.

[0106] The conflict resolution strategy library is a database storing various conflict mitigation rules within the conflict coordination algorithm, containing multiple solutions for different types of conflicts. These conflict mitigation rules are summarized based on extensive experiments and experience, and can effectively resolve conflicts between data sharing needs and security constraints.

[0107] Based on the conflict priority sequence and the quantified conflict severity, corresponding conflict mitigation rules are matched from the conflict resolution strategy library. For example, for a conflict between data usage frequency and data anonymization level, if the conflict severity is high, the rule for lowering the data anonymization level or adjusting the data usage frequency may be chosen; for a conflict between data processing timeliness and access control, the rule for optimizing the access verification process or adjusting the data processing timeliness requirements may be chosen.

[0108] Finally, a conflict coordination result is generated, which includes conflict type, conflict priority, conflict mitigation rules, and the order in which the rules are applied. The conflict coordination result provides specific guidance and basis for subsequent adjustments to the initial security label combination.

[0109] Step S360: Adjust the tag parameters in the initial security tag combination according to the conflict coordination results, generate an optimized target security tag combination, match the corresponding protection policy from the preset policy rule base based on the target security tag combination, and generate a fine-grained data protection guidance document containing data desensitization rules, access permission rules and transmission encryption rules.

[0110] In one implementation, step S360 may specifically include the following steps S361 to S366: Step S361: Analyze the conflict mitigation rules and the order of rule application in the conflict coordination results, and extract the tag adjustment parameters corresponding to each conflict type in the order of rule application. The first conflict type corresponds to the data desensitization level adjustment parameter, the second conflict type corresponds to the access control adjustment parameter, and the third conflict type corresponds to the transmission encryption strength adjustment parameter.

[0111] First, the conflict resolution results are analyzed. These results include conflict mitigation rules and the order in which they are applied. The conflict mitigation rules specify how to resolve different types of conflicts, while the order in which they are applied specifies the execution order of these rules. Following the order of rule application, the label adjustment parameters corresponding to each conflict type are extracted sequentially. For the first conflict type (conflict between data usage frequency and data anonymization level), the corresponding label adjustment parameter is the data anonymization level adjustment parameter, used to adjust the data anonymization level in the initial security label combination. For the second conflict type (conflict between data processing timeliness and access control), the corresponding label adjustment parameter is the access control adjustment parameter, used to adjust the data access level label and access control parameters. For the third conflict type (conflict between data precision requirements and transmission encryption strength), the corresponding label adjustment parameter is the transmission encryption strength adjustment parameter, used to adjust the data transmission security label and transmission encryption strength parameters.

[0112] Step S362: For each conflicting data unit, add its corresponding data anonymization level adjustment parameter to the data anonymization level parameter in the initial security label combination of the data unit to generate the adjusted target data anonymization level parameter; add its corresponding access control adjustment parameter to the access control parameter in the initial security label combination of the data unit to generate the target access control parameter; add its corresponding transmission encryption strength adjustment parameter to the transmission encryption strength parameter in the initial security label combination of the data unit to generate the target transmission encryption strength parameter.

[0113] For conflicting data units, adjustments are made according to the extracted tag adjustment parameters. The data anonymization level adjustment parameter is superimposed on the data anonymization level parameter in the initial security tag combination for that data unit. For example, if the initial data anonymization level parameter is light anonymization, and the data anonymization level adjustment parameter is to reduce anonymization by one level, then the adjusted target data anonymization level parameter will be a lower level of anonymization. Access control adjustment parameters are superimposed on the access control parameters in the initial security tag combination for that data unit. For example, if the initial access control parameters stipulate that certain users can only perform read-only operations, and the access control adjustment parameter adds partial write operation permissions, then the adjusted target access control parameters will allow these users to perform partial write operations. Transmission encryption strength adjustment parameters are superimposed on the transmission encryption strength parameter in the initial security tag combination for that data unit. For example, if the initial transmission encryption strength parameter is medium-strength encryption, and the transmission encryption strength adjustment parameter is to reduce encryption strength, then the adjusted target transmission encryption strength parameter will be low-strength encryption.

[0114] Step S363: Combine the target data desensitization level parameters, target access control parameters, and target transmission encryption strength parameters to form an optimized target security label combination, and associate each data unit with a unique target security label combination.

[0115] The adjusted target data anonymization level parameters, target access control parameters, and target transmission encryption strength parameters are combined to form an optimized target security label combination. This target security label combination better balances data sharing needs and security constraints, and can better protect data security and privacy. A unique target security label combination is associated with each data unit, ensuring that each data unit has a corresponding and appropriate protection strategy. This enables fine-grained protection of each data unit in the original talent dataset.

[0116] Step S364: Load the protection policy template that matches the target security label combination from the preset policy rule base. The policy rule base contains multiple templates corresponding to different security label combinations. Each template contains a preset data desensitization rule template, access permission rule template, and transmission encryption rule template.

[0117] The preset policy rule base is a database that stores various protection policy templates, categorized and stored according to different security tag combinations. Each security tag combination corresponds to a protection policy template, which contains preset data anonymization rule templates, access permission rule templates, and transmission encryption rule templates.

[0118] Based on the target security label combination for each data unit, a matching protection policy template is loaded from a pre-defined policy rule base. For example, if the target security label combination is high-sensitivity data, strict access permissions, and high-strength encrypted transmission, then the corresponding protection policy template will be loaded. The data anonymization rule template in this template may specify stricter anonymization algorithms and a wider range of anonymized fields; the access permission rule template may specify stricter role permissions and shorter permission durations; and the transmission encryption rule template may specify more advanced encryption protocols and more complex key management methods.

[0119] Step S365: Adjust the rule parameters in the protection strategy template according to the specific parameter values ​​in the target security label combination. For example, determine the desensitization algorithm type and desensitization field range in the data desensitization rule according to the target data desensitization level parameter, determine the role permission list and permission validity period in the access permission rule according to the target access permission control parameter, and determine the encryption protocol and key management method in the transmission encryption rule according to the target transmission encryption strength parameter.

[0120] Based on the specific parameter values ​​in the target security label combination, the rule parameters in the loaded protection policy template are adjusted. For data anonymization rules, the anonymization algorithm type and the range of anonymized fields are determined according to the target data anonymization level parameter. If the target data anonymization level parameter is severe anonymization, a more complex anonymization algorithm may be selected, such as replacement anonymization, encryption anonymization, etc., and the range of anonymized fields may be expanded.

[0121] For access permission rules, the role permission list and permission validity period are determined based on the target access permission control parameters. If the target access permission control parameter is strict access control, the list of allowed operations may be reduced, and the validity period of permissions may be shortened.

[0122] For transmission encryption rules, the encryption protocol and key management method are determined based on the target transmission encryption strength parameter. If the target transmission encryption strength parameter is high-strength encryption, then a more advanced encryption protocol, such as AES (Advanced Encryption Standard), may be selected, and a more secure key management method, such as key escrow or periodic key replacement, may be adopted.

[0123] Step S366: Organize the adjusted rule parameters into structured rule entries, where data desensitization rule entries include desensitization algorithm identifiers, indexes of fields to be desensitized, and desensitization replacement rules; access permission rule entries include role identifiers, a list of allowed operations, and the effective period of permissions; and transmission encryption rule entries include encryption algorithm identifiers, key generation methods, and transmission verification rules. After sorting multiple rule entries according to the normalized value of the association strength of data units, integrate them into a fine-grained data protection guidance document.

[0124] The adjusted rule parameters are organized into structured rule entries, each containing specific information. Data masking rule entries include a masking algorithm identifier, specifying the type of masking algorithm used; an index of the field to be masked, identifying the field requiring masking; and a masking replacement rule, specifying the exact method of masking.

[0125] Access permission rule entries include a role identifier, used to specify different user roles; a list of allowed operations, used to list the operations that the role is allowed to perform; and a permission effective period, used to specify the time range within which the permissions for the role are effective.

[0126] The transmission encryption rule entries include an encryption algorithm identifier, which specifies the encryption algorithm used; a key generation method, which describes how the key is generated; and transmission verification rules, which specify the verification methods during data transmission.

[0127] Multiple rule entries are sorted according to the normalized value of the association strength of the data units, with the rule entries corresponding to data units with higher association strength values ​​appearing first. This sorting aims to prioritize the protection of data units highly relevant to the task. Finally, the sorted rule entries are integrated into a fine-grained data protection guidance document, which provides detailed protection guidance for subsequent data processing and sharing.

[0128] Step S400: Based on the data desensitization rules in the fine-grained data protection guidelines document, perform structured desensitization transformation on the corresponding data units in the original talent data set to obtain an intermediate data form that meets security requirements. Then, generate a set of corresponding access verification rules and a set of transmission encryption rules according to the access permission rules and transmission encryption rules. The access verification rule set includes data access subject authentication rules and data operation permission verification rules, and the transmission encryption rule set includes data transmission channel encryption protocol rules and data fragmentation encryption rules.

[0129] Based on the data anonymization rules in the fine-grained data protection guidelines, structured anonymization transformation is performed on the data units in the original talent dataset that require anonymization. Structured anonymization transformation processes data according to specific rules and procedures to protect data privacy and security. For example, if the data anonymization rules stipulate that the ID card numbers of talents should be anonymized using a replacement anonymization algorithm to replace some digits of the ID card number with specific characters, then the ID card number data units in the original talent dataset will be processed according to this rule.

[0130] Through de-identification and transformation, an intermediate data format that meets security requirements is obtained. This intermediate data format, which is data that has undergone de-identification processing, satisfies the basic needs of data sharing while protecting data privacy and sensitive information.

[0131] Based on the access permission rules in the fine-grained data protection guidelines document, a corresponding set of access verification rules is generated. This set includes data access subject authentication rules and data operation permission verification rules. Data access subject authentication rules verify the authenticity and legitimacy of the subject accessing the data, using methods such as username and password verification, fingerprint verification, and digital certificate verification. Data operation permission verification rules check whether the access subject has the permission to perform specific operations, such as whether querying, modifying, or deleting data is permitted.

[0132] Based on the transmission encryption rules in the fine-grained data protection guidelines document, a corresponding set of transmission encryption rules is generated. This set includes data transmission channel encryption protocol rules and data fragmentation encryption rules. The data transmission channel encryption protocol rules specify the encryption protocols used during data transmission, such as SSL / TLS (Secure Sockets Layer / Transport Layer Security), to ensure data security during transmission. The data fragmentation encryption rules are used to fragment data and encrypt each fragment, improving the security and reliability of data transmission.

[0133] In one implementation, step S400 may specifically include the following steps S410 to S450: Step S410: Parse the data desensitization rule entries in the fine-grained data protection guidelines document, extract the desensitization algorithm type, desensitization field location, and desensitization replacement character, and locate the data unit to be desensitized and the target field within the unit in the original talent data set based on the desensitization field location.

[0134] This section analyzes the data masking rule entries in the fine-grained data protection guidelines document. These entries contain detailed information about the masking process. It extracts the masking algorithm type, such as replacement masking, masking masking, and encryption masking, as different algorithms are suitable for different data types and security requirements. It also extracts the masking field position, specifying the location of the data field to be masked within the data unit, such as which field within the unit. Finally, it extracts the masking replacement characters, used to replace sensitive characters in the replacement masking algorithm. Based on the masking field position, it locates the data unit to be masked and the target field within that unit in the original talent dataset. For example, if the masking field position is specified as the 3rd field in the data unit, then the 3rd field of all data units in the original talent dataset will be found as the target field.

[0135] Step S420: Call the desensitization processing function that matches the desensitization algorithm type to perform desensitization conversion on the original data value of the target field. If the desensitization algorithm type is replacement desensitization, then use desensitization replacement characters to replace some characters in the target field; if it is masking desensitization, then replace the sensitive characters in the target field with preset masking characters; if it is encryption desensitization, then use a preset encryption function to reversibly encrypt the target field to generate the desensitized data field value.

[0136] Based on the extracted de-identification algorithm type, the matching de-identification processing function is invoked. If the de-identification algorithm type is replacement de-identification, the de-identification processing function will replace some characters in the target field with de-identification replacement characters. For example, for a talent's mobile phone number, the de-identification replacement character is "*", which might replace the middle few digits of the mobile phone number with "*".

[0137] If the de-identification algorithm type is masked de-identification, the de-identification processing function will replace sensitive characters in the target field with preset masked characters. Preset masked characters are usually uniform characters, such as "#". For example, for a talent's ID card number, the digits in the birth date portion will be replaced with "#".

[0138] If the de-identification algorithm type is encryption de-identification, the de-identification processing function will reversibly encrypt the target field using a preset encryption function. Reversible encryption means that the encrypted data can be restored to the original data when needed. The preset encryption function can be a symmetric encryption algorithm, such as AES, or an asymmetric encryption algorithm, such as RSA. During encryption, the original data value of the target field is encrypted using the corresponding key to generate the de-identified data field value.

[0139] In one implementation, step S420 may specifically include the following steps S421 to S423: Step S421: When the desensitization algorithm type is replacement desensitization, obtain the character length of the target field, calculate the number of characters to be replaced according to the preset replacement ratio, select the corresponding number of character positions from the target field by random sampling, replace the characters at the selected positions with desensitized replacement characters, keep the remaining characters unchanged, and generate the field value after replacement desensitization.

[0140] Once the data masking algorithm type is determined to be replacement masking, the first step is to obtain the character length of the target field, which can be achieved using string processing functions. The preset replacement ratio is pre-set based on the data sensitivity and security requirements, representing the proportion of characters to be replaced in the target field. Based on the replacement ratio and the character length of the target field, the number of characters to be replaced is calculated. A corresponding number of character positions are selected from the target field using random sampling. Random sampling ensures that the replacement character positions are random, increasing the randomness and security of the masking process. A random number generator can be used to generate random character position indices. Then, the masked replacement characters are used to replace the characters at the selected positions, while leaving the remaining characters unchanged, ultimately generating the masked field value.

[0141] Step S422: When the desensitization algorithm type is masking desensitization, the preset sensitive character recognition function is called to identify the set of sensitive characters in the target field. The set of sensitive characters includes preset personal identification characters, contact information characters, and privacy data characters. Each character in the identified set of sensitive characters is replaced with a preset masking character to generate the masked desensitized field value.

[0142] The pre-defined sensitive character recognition function is specifically designed to identify sensitive characters in a target field, based on pre-defined sensitive character rules. The sensitive character set includes common personal identification characters, such as numbers in ID card numbers and specific characters in names; contact information characters, such as characters in mobile phone numbers and email addresses; and private data characters, such as numbers in bank card numbers. The sensitive character recognition function scans the target field to identify the sensitive character set. Then, each character in the identified sensitive character set is replaced with a pre-defined masking character, such as "#". In this way, sensitive information in the target field is masked, generating a masked and anonymized field value.

[0143] Step S423: When the desensitization algorithm type is encryption desensitization, obtain the preset encryption function and the corresponding initial key, convert the original data value of the target field into a byte array, call the encryption function to perform encryption operation on the byte array using the initial key to generate an encrypted byte array, convert the encrypted byte array into a hexadecimal string to generate the encrypted desensitized field value, and store the initial key in the preset key management library and associate it with the desensitized field identifier.

[0144] When the de-identification algorithm type is encryption de-identification, the first step is to obtain the preset encryption function and its corresponding initial key. The preset encryption function can be a symmetric or asymmetric encryption algorithm, and the initial key is a crucial parameter for performing the encryption operation. The original data value of the target field is converted into a byte array, because encryption functions typically process byte data. Then, the encryption function is called to perform an encryption operation on the byte array using the initial key, generating an encrypted byte array. The encryption operation converts the original data into ciphertext, ensuring data security.

[0145] The encrypted byte array is converted into a hexadecimal string for easy storage and transmission. This generates the encrypted and anonymized field values. Simultaneously, to enable decryption of the encrypted data when needed, the initial key is stored in a pre-defined key management library and associated with the anonymized field identifier. This allows the initial key to be retrieved from the key management library based on the field identifier when decryption is required later.

[0146] Step S430: Replace the target field value in the original data unit with the de-identified data field value to obtain the preliminary de-identified data unit. Perform data integrity verification on the preliminary de-identified data unit and check whether the format of the data unit after de-identification conforms to the preset data structure specification. If it does not conform, repeat the de-identification conversion operation until the format conforms to the specification and obtain an intermediate data form that meets the security requirements.

[0147] In one implementation, step S430 may specifically include the following steps S431 to S436: Step S431: Obtain the preset data structure specification document, and parse the data unit format requirements in the document, including the requirements for the number of fields, the data type of fields, the range of field lengths, and the logical relationships between fields.

[0148] First, obtain the pre-defined data structure specification document from the system's configuration file or database. This document details the format requirements that data units should follow. Parse the document and extract key information. The field quantity requirement specifies the number of fields that a data unit should contain, serving as a fundamental constraint on the data unit format. The field data type requirement clarifies the data type that each field should have, such as integer, character, or date, ensuring data accuracy and consistency. The field length range requirement specifies the allowed length range for each field, preventing field data from being too long or too short. The logical relationship requirement describes the logical connections between different fields, such as the association between primary keys and foreign keys, and the calculation relationships between fields, ensuring the logical rationality of the data.

[0149] Step S432: Compare the actual number of fields in the preliminary de-identified data unit with the required number of fields. If the actual number of fields is less than the required number of fields, the format is determined to be non-compliant and a field missing error flag is generated. If the actual number of fields is more than the required number of fields, the format is determined to be non-compliant and a field redundancy error flag is generated.

[0150] The actual number of fields in the initially anonymized data units is counted, and then compared with the field count requirements parsed from the data structure specification document.

[0151] If the actual number of fields is less than the required number of fields, it means that the data unit is missing necessary fields, which may lead to incomplete data information. In this case, the format is determined to be non-compliant, and a field missing error flag is generated to mark that the data unit has a problem with missing fields.

[0152] If the actual number of fields exceeds the field requirement, it means that there are redundant fields in the data unit. This may affect data processing and analysis. The format is judged to be non-compliant, and a field redundancy error mark is generated to indicate that there are redundant fields in the data unit.

[0153] Step S433: Check whether the actual data type of each field in the preliminary desensitized data unit is consistent with the field data type requirements. If there are fields with inconsistent data types, the format is determined to be non-compliant and a data type error flag is generated.

[0154] Iterate through each field in the initially anonymized data unit to obtain its actual data type. Then, compare the actual data type of each field with the data type requirements specified in the data structure specification document.

[0155] If a field with inconsistent data types is found, it means that the data in that field does not meet the format requirements, which may lead to errors or inaccurate results in data processing. In this case, the format is determined to be non-compliant, and a data type error flag is generated to indicate that there is a data type mismatch problem.

[0156] Step S434: Check whether the actual length of each field in the preliminary desensitized data unit is within the field length range requirement. If there is a field length exceeding the range, it is determined that the format does not conform to the specification and a field length error mark is generated.

[0157] For each field in the initially anonymized data unit, obtain its actual length. Compare the actual length of each field with the field length range specified in the data structure specification document. If any field length exceeds the range, whether too long or too short, it indicates that the data in that field does not meet the format requirements and may affect data storage and use. A field length error flag is generated to indicate that the field has an abnormal length.

[0158] Step S435: Verify whether the actual logical relationship between fields in the preliminary de-identified data unit is consistent with the requirements for logical relationship between fields. For example, whether the primary key field is unique and whether the foreign key field has a corresponding record in the associated table. If the logical relationship is inconsistent, the format is determined to be non-compliant and a logical relationship error flag is generated.

[0159] Based on the logical relationship requirements between fields specified in the data structure specification document, the actual logical relationships between fields in the initially anonymized data units are verified. For primary key fields, their uniqueness is checked, i.e., whether there are duplicate primary key values ​​in the data unit. Duplicate primary key values ​​will compromise data uniqueness and integrity, leading to data management chaos. For foreign key fields, the existence of corresponding records in the associated tables is checked. Foreign keys are used to establish relationships between different tables; if a foreign key field value does not have a corresponding record in the associated table, the data association will fail. If the actual logical relationship between fields is inconsistent with the requirements, the format is deemed non-compliant, and a logical relationship error flag is generated to indicate that the data unit has a logical relationship anomaly.

[0160] Step S436: Summarize the error flags for missing fields, redundant fields, data types, field lengths, and logical relationships. If no error flags are found, the initial de-identified data unit format is determined to be compliant with the specifications, and it is identified as the intermediate data form. If error flags are found, adjust the parameters in the de-identification conversion operation according to the error flag type, such as adjusting the position of the de-identified field or the number of characters replaced, and repeat the de-identification conversion operation until the generated de-identified data unit passes all format checks and obtains an intermediate data form that meets security requirements.

[0161] The system will summarize the error flags generated during the data integrity verification process, including those for missing fields, redundant fields, data types, field lengths, and logical relationships.

[0162] If the summary results show no error indicators, it means that the format of the initially desensitized data unit fully conforms to the preset data structure specifications, and it is determined as an intermediate data form, which can be used for subsequent data sharing and analysis.

[0163] If an error flag is present, it indicates a formatting issue with the initial de-identified data unit, requiring adjustment. Adjust the parameters in the de-identification conversion operation accordingly, based on the type of error flag. For example, if the error flag indicates a field length error, it may be necessary to adjust the position of the de-identified field or the number of characters replaced to change the field length. Perform the de-identification conversion operation again, and perform data integrity verification on the generated de-identified data unit once more. Repeat this process until the generated de-identified data unit passes all format checks, ultimately yielding an intermediate data format that meets security requirements.

[0164] Step S440: Parse the access permission rule entries in the fine-grained data protection guidance document, extract the access subject role type, data operation permission list, and permission validity period, match the corresponding authentication method from the preset role permission mapping library according to the access subject role type, determine the set of allowed operation types according to the data operation permission list, determine the permission effective time interval according to the permission validity period, and combine the authentication method, operation type set, and permission effective time interval to generate access verification rule entries. Multiple access verification rule entries constitute the access verification rule set.

[0165] The access permission rules in the fine-grained data protection guidelines document are parsed to extract key information. The access subject role type distinguishes different user roles, such as administrators, regular users, and auditors, each with different permissions and scopes of operation. The data operation permission list details the types of data operations permitted for this role, such as querying, modifying, and deleting. The permission validity period specifies the time frame within which the role holds the corresponding permissions, ensuring the timeliness and security of permissions.

[0166] Based on the extracted access subject role type, the corresponding authentication method is retrieved from the pre-defined role and permission mapping database. The role and permission mapping database stores the correspondence between different roles and authentication methods, and is pre-configured according to security policies and business needs. For example, an administrator role may require multi-factor authentication, including password, SMS verification code, and fingerprint recognition; a regular user role may only require password verification.

[0167] The set of permitted operation types is determined based on the data operation permission list. This set specifies the specific operations that the access subject can perform within the permission validity period. The effective time range of the permissions is determined based on the permission validity period, including the start and end times. The authentication method, the set of operation types, and the effective time range are combined to generate access verification rule entries. Multiple such access verification rule entries constitute an access verification rule set, used for effective control and management of data access behavior.

[0168] Step S450: Parse the transmission encryption rule entries in the fine-grained data protection guidance document, extract the encryption algorithm type, key length parameter and encryption mode, determine the corresponding key generation method according to the encryption algorithm type, determine the key byte length according to the key length parameter, determine the data block method and padding method according to the encryption mode, combine the key generation method, key byte length, data block method and padding method to generate transmission encryption rule entries, and multiple transmission encryption rule entries constitute a transmission encryption rule set.

[0169] The key parameters are extracted by analyzing the transmission encryption rules entries in the fine-grained data protection guidelines document. The encryption algorithm type specifies the encryption algorithm used during data transmission, such as AES and DES. Different encryption algorithms have different security and performance characteristics. The key length parameter determines the length of the encryption key; a longer key length increases encryption security but also increases the computational complexity of encryption and decryption. The encryption mode specifies how the encryption algorithm processes data. Common encryption modes include ECB, CBC, and CFB, each with different data block and padding methods. The corresponding key generation method is determined based on the encryption algorithm type. Different encryption algorithms require different key generation methods; for example, the AES algorithm can use a random number generator to generate the key. The byte length of the key is determined based on the key length parameter to ensure that the generated key meets the specified length requirements. The data block and padding methods are determined based on the encryption mode. The data block method divides the data to be encrypted into fixed-size blocks, and the padding method handles cases where the data length is not an integer multiple of the block size, ensuring that the data can be encrypted completely. The key generation method, key byte length, data block method, and padding method are combined to generate transmission encryption rule entries. Multiple transmission encryption rule entries constitute a transmission encryption rule set, which is used to encrypt and protect the data transmission process, ensuring the security and integrity of the data during transmission.

[0170] Step S500: Deploy the intermediate data format, access verification rule set, and transmission encryption rule set to the preset controlled collaborative computing environment, perform data sharing and joint analysis, and use the built-in rule verification engine of the environment to call the access verification rule set in real time to verify the data access behavior and call the transmission encryption rule set to check the encryption compliance of the data transmission process. After the joint analysis operation is completed, output the sharing analysis results that meet the requirements of the data sharing task description information.

[0171] In one implementation, step S500 may specifically include the following steps S510-S560: Step S510: Transmit intermediate data to the distributed storage node of the controlled collaborative computing environment through the preset environment deployment interface, load the access verification rule set into the rule storage area of ​​the environment's rule verification engine, and load the transmission encryption rule set into the rule configuration area of ​​the environment's encryption compliance check module.

[0172] The pre-defined environment deployment interface serves as a bridge connecting external systems and the controlled collaborative computing environment. It provides standardized interfaces and protocols for securely and efficiently transmitting intermediate data formats, access verification rule sets, and transmission encryption rule sets to the controlled collaborative computing environment.

[0173] This interface transmits intermediate data to distributed storage nodes in a controlled collaborative computing environment. Distributed storage nodes are a storage architecture that distributes data across multiple nodes, offering advantages such as high availability, scalability, and fault tolerance. Storing intermediate data on distributed storage nodes facilitates subsequent data access and processing.

[0174] The set of access verification rules is loaded into the rule store of the environment's rule validation engine. The rule store is the area in the rule validation engine used to store access verification rules. The rule validation engine can quickly read rules from this area and perform real-time verification of data access behavior.

[0175] The set of transmission encryption rules is loaded into the rule configuration area of ​​the environment's encryption compliance check module. This module is responsible for checking the encryption compliance of data transmission processes. The rule configuration area stores the transmission encryption rules, allowing the module to check transmitted data packets according to these rules, ensuring the security and compliance of data transmission.

[0176] Step S520: Start the joint analysis task scheduler of the controlled collaborative computing environment, and generate an analysis task execution plan based on the data requirement type description in the data sharing task description information. The plan includes the task execution order, task resource allocation scheme and task result aggregation method.

[0177] Launching the joint analysis task scheduler in a controlled collaborative computing environment is a core component responsible for managing and scheduling joint analysis tasks that share data.

[0178] Based on the data requirement type description in the initial data sharing task description information, the joint analysis task scheduler generates a detailed analysis task execution plan. The task execution order specifies the order in which each analysis task is executed, ensuring that tasks are performed in a logical sequence, thereby improving analysis efficiency and accuracy.

[0179] The task resource allocation scheme determines the computing and storage resources required for each analysis task, allocating resources rationally based on the task's complexity and importance to ensure smooth task execution. The task result aggregation method defines how to integrate and summarize the results of various analysis tasks to obtain the final shared analysis results.

[0180] Step S530: According to the analysis task execution plan, the computing nodes of the environment are invoked to perform joint analysis operations on the intermediate data in the distributed storage nodes. During the analysis process, when the computing node initiates a data access request, the rule verification engine captures the access subject's identity information, the request operation type, and the request timestamp in the access request in real time.

[0181] Based on the generated analysis task execution plan, the joint analysis task scheduler invokes compute nodes in the controlled collaborative computing environment to perform joint analysis operations on intermediate data formats in the distributed storage nodes. Compute nodes are devices or processes with computing capabilities, responsible for executing specific data analysis tasks.

[0182] During the analysis process, compute nodes may initiate data access requests as needed for the task. When a data access request is generated, the environment's built-in rule validation engine will capture key information in the request in real time, including the identity information of the accessing subject to identify the visitor; the type of request operation to clarify the operation the accessing subject wants to perform, such as querying or statistics; and the request timestamp to record the time when the request occurred for subsequent time-related validation.

[0183] Step S540: The rule verification engine matches the access verification rule entries corresponding to the access subject's identity information from the access verification rule set, checks whether the access subject's identity information has passed the authentication method verification, checks whether the request operation type belongs to the operation type set, and checks whether the request timestamp is within the permission effective time interval. If all checks pass, the data access operation is allowed; otherwise, the data access operation is rejected and an access exception log is recorded.

[0184] In one implementation, step S540 may specifically include the following steps S541 to S546: Step S541: The rule validation engine extracts the access subject's identity information from the access request, standardizes the format of the identity information, removes redundant characters and spaces, and generates a standard identity identifier.

[0185] The rule validation engine first extracts the access subject's identity information from the captured access requests. This information may include the username, ID number, or national ID card number entered by the user. Since the user-entered identity information may have irregular formats, such as containing extra spaces or special characters, it needs to be standardized.

[0186] Redundant characters and spaces in identity information are removed using specific string processing functions, converting it to a standardized format. For example, leading and trailing spaces are removed from usernames, and separators are removed from ID card numbers. After processing, a standard identity identifier is generated, which can be used more accurately for subsequent identity matching and verification.

[0187] Step S542: Traverse each access verification rule entry from the access verification rule set, compare whether the access subject role type in the entry is consistent with the role type corresponding to the standard identity identifier, and determine the rule entries with consistent role types as candidate verification rule entries.

[0188] The rule validation engine iterates through every access validation rule entry in the access validation rule set. For each entry, it extracts the access subject role type, such as administrator, regular user, visitor, etc. The extracted access subject role type is then compared with the role type corresponding to the standard identity identifier. If they match, the rule entry is designated as a candidate validation rule entry. Candidate validation rule entries are rules that may be applicable to the current access subject, and further authentication and permission checks will be performed based on these entries.

[0189] Step S543: Extract the authentication method from the candidate verification rule entries, call the corresponding authentication function according to the authentication method, input the standard identity identifier into the function for authentication, generate an authentication success identifier if the authentication is successful, otherwise generate an authentication failure identifier.

[0190] The authentication method is extracted from the defined candidate authentication rule entries. Different roles may correspond to different authentication methods, such as single-factor authentication (password authentication) and multi-factor authentication (password + SMS verification code). Based on the extracted authentication method, the corresponding authentication function is called. These authentication functions are pre-developed and configured to implement different authentication logics. A standard identity identifier is passed as input to the authentication function, which verifies the identity according to the specific authentication method. If the authentication function returns a successful verification result, it indicates that the accessing entity's identity is legitimate, and an authentication success flag is generated to mark successful authentication. If the verification fails, an authentication failure flag is generated, indicating that there is a problem with the accessing entity's identity.

[0191] Step S544: Extract the operation type set from the candidate verification rule entries, check whether the request operation type is an element in the operation type set, if so, generate an operation type matching identifier, otherwise generate an operation type mismatch identifier.

[0192] Extract the set of allowed operation types from the candidate verification rule entries. This set lists the types of data operations that the role is allowed to perform.

[0193] The request operation type in the access request is compared with the set of operation types to check if the request operation type is an element in the set. If it is, it means that the operation requested by the access subject is allowed, and an operation type matching identifier is generated. If the request operation type is not in the set of operation types, it means that the access subject does not have permission to perform the operation, and an operation type mismatch identifier is generated.

[0194] Step S545: Extract the permission effective time interval from the candidate verification rule entries, check whether the request timestamp is greater than the start time of the interval and less than the end time of the interval. If so, generate a time validity identifier; otherwise, generate a time invalid identifier.

[0195] Extract the permission effective time range from the candidate verification rule entries. This range specifies the time range within which the access subject has the corresponding permissions, including the start and end times of the permissions.

[0196] The request timestamp in the access request is compared with the permission effective time range to check if the request timestamp falls between the start and end times of the range. If the request timestamp is greater than the start time and less than the end time of the range, it indicates that the request was made within the permission validity period, and a time validity flag is generated. If the request timestamp is not within the range, it indicates that the accessing subject made the request without permission, and a time invalid flag is generated.

[0197] Step S546: Summarize the authentication pass flag, operation type matching flag, and time validity flag. If all three flags are in a pass state, the data access operation is deemed to meet the permission requirements and is allowed. If any flag is in a failure state, the data access operation is deemed to not meet the permission requirements, the data access operation is refused, and the access subject's identity information, request operation type, request timestamp, and failure flag type are recorded in the access exception log.

[0198] The generated authentication credentials are aggregated using an identifier, an operation type matching identifier, and a time validity identifier. Only when all three identifiers are in a passed state is it considered that the accessing subject's identity is legitimate, the requested operation is permitted, and the request time is within the permission validity period. Therefore, the data access operation is deemed to meet the permission requirements, and the compute node is allowed to perform the data access operation.

[0199] If any of the following conditions are met (e.g., authentication failure, operation type mismatch, or invalid time), the data access operation is deemed to be invalid and access is denied. Simultaneously, detailed information such as the access subject's identity, request operation type, request timestamp, and failure flag type is recorded in the access exception log for subsequent security audits and problem analysis to identify the cause of the access exception and take appropriate measures.

[0200] Step S550: During the transmission of intermediate data between computing nodes, the encryption compliance check module captures the transmission data packets in real time, matches the corresponding transmission encryption rule entries from the transmission encryption rule set, checks whether the encryption algorithm type of the data packet is consistent with the rule entry, checks whether the length of the encryption key meets the key byte length requirement, and checks whether the data grouping method and padding method are consistent with the rule entry. If all checks pass, the transmission operation is allowed; otherwise, the transmission operation is interrupted and an encryption exception log is recorded.

[0201] During the transmission of intermediate data between different computing nodes, the encryption compliance check module in the environment monitors and captures the transmitted data packets in real time.

[0202] The encryption compliance check module searches the transmission encryption rule set for the transmission encryption rule entry corresponding to the currently transmitted data packet. This may be done by matching some identifying information in the data packet, such as the source node, destination node, and data type.

[0203] After locating the corresponding transmission encryption rule entry, a detailed inspection of the transmitted data packets is performed. First, check if the encryption algorithm type of the data packet matches the encryption algorithm type specified in the rule entry. Different encryption algorithms have different security and performance characteristics; if the encryption algorithm types do not match, the security of data transmission may not be guaranteed.

[0204] Next, the encryption key length is checked to ensure it meets the key byte length requirements specified in the rule entries. Key length is a crucial factor affecting encryption security; if the key length does not meet the requirements, the encrypted data may be easily cracked. The data block and padding methods are also checked to ensure they are consistent with the rule entries. These methods determine how data is processed during encryption; inconsistencies may lead to errors or inability to properly decrypt the data. If all checks pass, the encryption method of the transmitted data packet complies with the transmission encryption rules, and the transmission operation is allowed to continue. If any check fails, the transmission operation is interrupted to prevent insecure data transmission. Simultaneously, an encryption exception log is recorded, containing information about the transmitted data packet and the specific items that failed the checks, for subsequent analysis and processing to identify and fix the encryption compliance issues.

[0205] Step S560: After the joint analysis operation is completed, the task scheduler aggregates the analysis results of each computing node according to the task result aggregation method, and generates a shared analysis result that includes data analysis conclusions, data statistical charts and data trend predictions, ensuring that the content meets the requirements of the data usage scenario description, data requirement type description and data sharing scope description in the data sharing task description information.

[0206] In one implementation, step S560 may specifically include the following steps S561 to S566: Step S561: The task scheduler receives the local analysis results uploaded by each computing node. Each local analysis result includes the analysis dimension identifier, analysis index value, and analysis confidence parameter.

[0207] After the joint analysis operation is completed, the joint analysis task scheduler waits for each computing node to upload its partial analysis results. Each computing node, after completing its assigned analysis task, organizes and encapsulates the results to form partial analysis results containing key information. The analysis dimension identifier in the partial analysis results clarifies which dimension the analysis was conducted from, such as time, geographic, or business type. The analysis metric values ​​are the specific analysis results data under that dimension, such as sales revenue, number of users, and growth rate. The analysis confidence parameter indicates the reliability of the analysis results, reflecting the uncertainty and error range in the analysis process.

[0208] Step S562: According to the aggregation rules in the task result aggregation method, perform a weighted average of the index values ​​of the local analysis results with the same analysis dimension identifier, and use the weight value as the analysis confidence parameter of each local analysis result to generate the global analysis index value.

[0209] The task scheduler processes the local analysis results uploaded by each computing node according to the aggregation rules in the previously defined task result aggregation method. For local analysis results with the same analysis dimension identifier, their analysis index values ​​are calculated using a weighted average.

[0210] The weights in the weighted average are the confidence parameters of each local analysis result. A higher confidence parameter indicates a more reliable local analysis result, and thus carries greater weight in the weighted average calculation. This approach allows for a comprehensive consideration of the reliability of each local analysis result, generating more accurate global analysis index values.

[0211] Step S563: Perform statistical analysis on the global analysis indicator values ​​and generate data statistical charts, including trend line charts, comparison bar charts, and percentage pie charts. The trend line chart is used to show the trend of indicator values ​​over time, the comparison bar chart is used to show the comparison of indicator values ​​in different analysis dimensions, and the percentage pie chart is used to show the proportion of each sub-dimensional indicator value to the total indicator value.

[0212] After obtaining the global analysis indicator values, statistical analysis is performed on them. The purpose of statistical analysis is to present the data in a more intuitive and understandable way, helping users better understand the characteristics and relationships of the data. Data statistical charts are generated, including trend line charts, comparison bar charts, and percentage pie charts. The trend line chart uses time as the horizontal axis and the global analysis indicator values ​​as the vertical axis. By connecting the indicator values ​​at various time points, it shows the trend of indicator values ​​over time, allowing users to clearly see the development trend of the data. The comparison bar chart displays the global analysis indicator values ​​of different analysis dimensions in the form of a bar chart. The height of the bars compares the size of the indicator values ​​of different dimensions, making it easy for users to compare the differences between different dimensions. The percentage pie chart presents the proportion of the global analysis indicator values ​​of each sub-dimension to the total indicator values ​​in a pie chart format. The size of each sector represents the proportion of the sub-dimension indicator value in the total, allowing users to intuitively understand the relative importance of each sub-dimension in the overall picture.

[0213] Step S564: Based on the global analysis index values ​​and data statistical charts, and combined with the data usage scenario description in the data sharing task description information, generate data analysis conclusions. The conclusions include a summary of data characteristics, explanations of data anomalies, and suggestions for data application.

[0214] Based on the generated global analysis indicator values ​​and data statistical charts, and combined with the data usage scenario description in the initially received data sharing task description information, a comprehensive analysis and judgment are performed to generate data analysis conclusions. The data analysis conclusions include a summary of data characteristics, summarizing and generalizing the data characteristics reflected by the global analysis indicator values ​​and data statistical charts, such as data distribution patterns and trend characteristics. The data outlier description analyzes and explains any anomalies that occur in the data, identifying the causes of the anomalies, such as data entry errors or changes in the external environment.

[0215] Step S565: Based on the historical trend of global analysis index values, call the preset trend prediction model to predict the index values ​​for future analysis periods, and generate data trend prediction results, including the predicted value sequence and prediction confidence interval.

[0216] Based on the historical trends of global analysis indicators, a pre-defined trend prediction model is used to predict the indicator values ​​for future analysis periods. The trend prediction model can be based on time series analysis, such as the ARIMA model or exponential smoothing model, or it can be based on machine learning, such as neural network models or support vector machine models.

[0217] Historical data of global analytical indicators are input into the trend prediction model. The model learns and trains based on the data's changing patterns and characteristics, then predicts the indicator values ​​for future analysis periods. The generated trend prediction results include a sequence of predicted values, i.e., the predicted indicator values ​​for each future analysis period. Simultaneously, prediction confidence intervals are generated, representing the reliable range of the prediction results and reflecting the uncertainties and errors in the prediction process. Prediction confidence intervals help users assess the reliability of the prediction results and consider the possible range of error when making decisions.

[0218] Step S566: Organize the data analysis conclusions, data statistical charts, and data trend prediction results in a structured manner according to the preset template to generate preliminary shared analysis results. Check whether the data range meets the data sharing range description requirements in the data sharing task description information. If not, adjust the display range of the data statistical charts and the description range of the data analysis conclusions until the content fully meets all the requirements of the data sharing task description information to obtain the final shared analysis results.

[0219] The generated data analysis conclusions, statistical charts, and trend prediction results are organized in a structured manner according to a pre-set template. This template specifies the format and layout of the shared analysis results, making them more standardized and consistent. After organization, preliminary shared analysis results are generated. These results are then checked, with a focus on ensuring the data range conforms to the data sharing scope description requirements in the data sharing task description. The data sharing scope description specifies the data source range, time range, and data granularity, ensuring the shared analysis results do not exceed these limits. If the data range in the preliminary shared analysis results is found to be inconsistent with the requirements, the display range of the statistical charts needs to be adjusted, for example, by narrowing or widening the data interval displayed in the charts, to conform to the data sharing scope. Simultaneously, the scope of the data analysis conclusions is adjusted accordingly to avoid including data information outside the specified range. This process of checking and adjusting is repeated until the content of the shared analysis results fully meets all the requirements of the data sharing task description, including the data usage scenario description, data requirement type description, and data sharing scope description, ultimately yielding final shared analysis results that can be used for decision-making and sharing.

Claims

1. A data sharing security decision-making method applying artificial intelligence, characterized in that, The method includes: Receive data sharing task description information and associated raw talent data set; The task description information and the original talent data set are input into a preset context-aware mapping network. The semantic parsing layer performs intent recognition on the task description information, and the entity association layer extracts entity relationships from the data units in the original talent data set. The task data association strength mapping table is generated by combining the intent recognition results and the entity relationship extraction results. The pre-configured data protection policy generator is invoked to match the task data association strength mapping table with the preset basic security label set. The multi-objective trade-off layer of the policy generator coordinates the conflict between data sharing requirements and security constraints, and generates a fine-grained data protection guidance document for each data unit in the talent original data set. This document includes data desensitization rules, access permission rules and transmission encryption rules. Based on the data desensitization rules in the fine-grained data protection guidelines, the corresponding data units in the original talent data set are subjected to structured desensitization transformation to obtain intermediate data forms that meet security requirements. A set of access verification rules and a set of transmission encryption rules are generated according to access permission rules and transmission encryption rules. The intermediate data format, access verification rule set, and transmission encryption rule set are deployed to a pre-defined controlled collaborative computing environment for data sharing and joint analysis. The built-in rule verification engine of the environment calls the access verification rule set in real time to verify the data access behavior and calls the transmission encryption rule set to check the encryption compliance of the data transmission process. After the joint analysis operation is completed, the shared analysis results that meet the requirements of the data sharing task description information are output.

2. The method according to claim 1, characterized in that, The process involves inputting the data sharing task description information and the original talent data set into a preset context-aware mapping network. The semantic parsing layer performs intent recognition on the task description information, and the entity association layer extracts entity relationships from the data units in the original talent data set. Finally, a task data association strength mapping table is generated by combining the intent recognition results and the entity relationship extraction results. The task description information of data sharing is input into the semantic parsing layer of the context-aware mapping network. The data in the task description information is segmented using scene description to obtain multiple scene description word units. The scene description word units are then labeled with part of speech, and noun word units and verb word units are selected as key scene elements. Dependency parsing is performed on the data requirement type descriptions to extract core requirement entity words and requirement modifiers, and requirement entity-modifier association pairs are constructed. The key scene elements and demand entities-modifiers are associated with the intent classifier of the input semantic parsing layer, and the sequence is labeled through the bidirectional LSTM network to generate a task intent label sequence. The raw talent data set is input into the entity association layer of the context-aware mapping network. Named entity recognition is performed on the text data in the data unit to extract the data entity name and entity type. Entity attributes are extracted from the numerical data to obtain a set of entity attribute key-value pairs. The entity relationship extraction model of the entity association layer extracts entity relationship triples from the entity name, entity type and entity attribute key-value pair set, and generates a set of entity relationship triples containing head entity, relationship type and tail entity. The task intent label sequence and entity relation triple set are input into the fusion computing layer of the context-aware mapping network. The semantic similarity value between each intent label and entity relation triple is calculated. Based on the similarity value, a task data association strength mapping table is constructed. The row dimension of the mapping table is the intent label in the intent label sequence, the column dimension is the triple identifier in the entity relation triple set, and the element is the corresponding semantic similarity value.

3. The method according to claim 2, characterized in that, The intention classifier, which associates key scene elements and demand entities with modifiers, with the input semantic parsing layer, performs sequence labeling through its bidirectional LSTM network to generate a task intent label sequence, including: Word vector transformation is performed on noun and verb units in key scene elements to generate scene element word vector sequence. Word vectors of entity words and modifiers in demand entity-modifier association pairs are concatenated to generate demand association pair word vector sequence. The word vector sequence of scene elements is associated with the demand. The word vector sequence is input into the input layer of the bidirectional LSTM network. The forward LSTM unit extracts the temporal features of the word vector sequence from left to right to obtain the forward temporal feature vector. The backward LSTM unit extracts the temporal features of the word vector sequence from right to left to obtain the backward temporal feature vector. The forward temporal feature vector and the backward temporal feature vector are concatenated along the feature dimension to generate a bidirectional fused feature vector. The bidirectional fused feature vector is then input into the fully connected layer of the classifier and subjected to a nonlinear transformation through an activation function to obtain the intent classification feature vector. The CRF layer of the intent classifier is invoked to perform sequence labeling and decoding on the intent classification feature vector, calculate the label transition probability matrix, find the label path with the highest probability based on the Viterbi algorithm, and generate the initial intent label sequence. Post-processing is performed on the initial intent label sequence to remove redundant and conflicting labels, resulting in an optimized task intent label sequence.

4. The method according to claim 2, characterized in that, The entity relationship extraction model, through the entity association layer, extracts entity relationship triples from the entity name, entity type, and entity attribute key-value pair set, generating an entity relationship triple set containing the head entity, relationship type, and tail entity, including: The feature construction layer of the model is constructed by inputting entity name, entity type and entity attribute key-value pair set into the relation. Character-level feature extraction is performed on entity name to generate entity name character feature vector. One-hot encoding is performed on entity type to generate entity type encoding vector. Keys and values ​​in entity attribute key-value pair set are transformed into word vectors to generate attribute key word vectors and attribute value word vectors. The entity name character feature vector, entity type encoding vector, attribute key word vector, and attribute value word vector are fused to generate a comprehensive entity feature vector. The comprehensive entity feature vector is then input into the Bi-GRU layer of the relation extraction model. Temporal dependency modeling is performed on the feature vector through a gated recurrent unit to obtain the entity temporal feature vector. The attention mechanism layer of the relation extraction model is invoked to calculate the attention of the entity temporal feature vector, generate the entity attribute attention weight vector, and perform weighted summation on the entity temporal feature vector based on the weight vector to obtain the entity feature vector that strengthens the attribute association. The entity feature vectors that enhance attribute associations are input into the convolutional layer of the relationship extraction model. Local feature extraction is performed on the feature vectors through multi-scale convolutional kernels to obtain multi-scale entity relationship feature maps. Global max pooling is then performed on the multi-scale entity relationship feature maps to generate entity relationship feature vectors. The entity relationship feature vector is input into the classification layer of the relationship extraction model. The probability value of each relationship type is calculated by the softmax function. The relationship type with the highest probability value is selected as the target relationship type between entities. Entity relationship triples containing the head entity name, the target relationship type, and the tail entity name are constructed. Multiple entity relationship triples are combined to form an entity relationship triple set.

5. The method according to claim 1, characterized in that, The process involves invoking a pre-configured data protection policy generator to match the task data association strength mapping table with a preset set of basic security labels. The policy generator's multi-objective trade-off layer coordinates conflicts between data sharing needs and security constraints, generating fine-grained data protection guidelines for each data unit in the original talent dataset. These guidelines include: The policy input layer of the data protection policy generator is invoked to receive a task data association strength mapping table and a preset set of basic security labels, which includes data sensitivity labels, data access level labels and data transmission security labels. The semantic similarity values ​​in the task data association strength mapping table are normalized to generate association strength normalized values. Based on the relationship between entity relation triples and data units in the original talent data set, the association strength normalized values ​​corresponding to each triple are aggregated to their respective data units to generate a comprehensive association strength value for each data unit. Based on the comprehensive association strength value, the data units in the original talent data set are divided into high-association data units, medium-association data units, and low-association data units. The high-association data units, medium-association data units, and low-association data units are respectively matched with the data sensitivity tags in the basic security tag set to generate the initial security tag combination for each data unit; The initial security label combination is input into the multi-objective trade-off layer of the data protection strategy generator. The constraint parsing module within the layer extracts the data usage frequency parameter, data processing timeliness parameter, and data accuracy requirement parameter from the data sharing requirements, and extracts the data desensitization level parameter, access control parameter, and transmission encryption strength parameter from the security constraints. The conflict coordination algorithm of the multi-objective trade-off layer is used to detect conflicts between data usage frequency parameters and data desensitization level parameters, data processing timeliness parameters and access control parameters, and data accuracy requirement parameters and transmission encryption strength parameters, and generate conflict coordination results. Based on the conflict coordination results, the tag parameters in the initial security tag combination are adjusted to generate an optimized target security tag combination. Based on the target security tag combination, the corresponding protection policy is matched from the preset policy rule base to generate a fine-grained data protection guidance document containing data desensitization rules, access permission rules, and transmission encryption rules.

6. The method according to claim 5, characterized in that, The conflict coordination algorithm using a multi-objective trade-off layer performs conflict detection on data usage frequency parameters and data desensitization level parameters, data processing timeliness parameters and access control parameters, and data accuracy requirement parameters and transmission encryption strength parameters, generating conflict coordination results, including: For the high-association data units, medium-association data units, and low-association data units after division, the data usage frequency parameter is input into the first conflict detection module of the conflict coordination algorithm to obtain the desensitization frequency threshold range corresponding to the association strength level of the data unit. It is determined whether the data usage frequency parameter exceeds the upper limit or is lower than the lower limit of the corresponding desensitization frequency threshold range. If so, it is determined that there is a first conflict between the data usage frequency parameter and the data desensitization level parameter of the data unit, and a first conflict type identifier and conflict degree quantification value are generated. The data processing timeliness parameter is input into the second conflict detection module of the conflict coordination algorithm to obtain the preset permission verification timeliness threshold. The absolute value of the difference between the data processing timeliness parameter and the permission verification timeliness threshold is calculated. If the absolute value of the difference exceeds the preset timeliness tolerance range, it is determined that there is a second conflict between the data processing timeliness parameter and the access control parameter, and a second conflict type identifier and conflict degree quantification value are generated. The data precision requirement parameter is input into the third conflict detection module of the conflict coordination algorithm to obtain the preset encryption precision loss threshold. The theoretical precision loss value after transmission encryption is calculated based on the data precision requirement parameter. If the theoretical precision loss value exceeds the encryption precision loss threshold, it is determined that there is a third conflict between the data precision requirement parameter and the transmission encryption strength parameter, and a third conflict type identifier and conflict degree quantification value are generated. The first conflict type identifier, the second conflict type identifier, the third conflict type identifier, and the corresponding conflict degree quantification value are input into the priority sorting module of the conflict coordination algorithm. The conflict types are sorted according to the preset conflict priority rules to generate a conflict priority sequence. Based on the conflict priority sequence and the quantification value of the conflict degree, the conflict resolution strategy library of the conflict coordination algorithm is invoked to match the corresponding conflict mitigation rules and generate a conflict coordination result that includes the conflict type, conflict priority, conflict mitigation rules and the order in which the rules are applied.

7. The method according to claim 5, characterized in that, The process involves adjusting the tag parameters in the initial security tag combination based on the conflict coordination results to generate an optimized target security tag combination. Then, based on this target security tag combination, corresponding protection policies are matched from a preset policy rule base to generate a fine-grained data protection guidance document containing data anonymization rules, access permission rules, and transmission encryption rules. The conflict resolution results are analyzed to determine the conflict mitigation rules and their application order. The tag adjustment parameters corresponding to each conflict type are extracted in the order of rule application. The first conflict type corresponds to the data desensitization level adjustment parameter, the second conflict type corresponds to the access control adjustment parameter, and the third conflict type corresponds to the transmission encryption strength adjustment parameter. For each conflicting data unit, its corresponding data anonymization level adjustment parameter is superimposed on the data anonymization level parameter in the initial security label combination of the data unit to generate the adjusted target data anonymization level parameter; its corresponding access control adjustment parameter is superimposed on the access control parameter in the initial security label combination of the data unit to generate the target access control parameter; its corresponding transmission encryption strength adjustment parameter is superimposed on the transmission encryption strength parameter in the initial security label combination of the data unit to generate the target transmission encryption strength parameter. The target data desensitization level parameters, target access control parameters, and target transmission encryption strength parameters are combined to form an optimized target security label combination, and each data unit is associated with a unique target security label combination. Load protection policy templates that match the target security label combination from the preset policy rule base. The policy rule base contains multiple templates corresponding to different security label combinations. Each template contains preset data desensitization rule templates, access permission rule templates, and transmission encryption rule templates. Adjust the rule parameters in the protection strategy template according to the specific parameter values ​​in the target security label combination; determine the desensitization algorithm type and desensitization field range in the data desensitization rule according to the target data desensitization level parameter; determine the role permission list and permission validity period in the access permission rule according to the target access permission control parameter; and determine the encryption protocol and key management method in the transmission encryption rule according to the target transmission encryption strength parameter. The adjusted rule parameters are organized into structured rule entries. The data desensitization rule entries include desensitization algorithm identifiers, indexes of fields to be desensitized, and desensitization replacement rules. The access permission rule entries include role identifiers, a list of allowed operations, and the effective period of permissions. The transmission encryption rule entries include encryption algorithm identifiers, key generation methods, and transmission verification rules. After sorting the multiple rule entries according to the normalized value of the association strength of the data units, they are integrated into a fine-grained data protection guidance document.

8. The method according to claim 1, characterized in that, The process involves performing structured desensitization transformation on corresponding data units in the original talent dataset based on the data desensitization rules in the fine-grained data protection guidelines document, obtaining an intermediate data form that meets security requirements, and generating a set of corresponding access verification rules and a set of transmission encryption rules based on access permission rules and transmission encryption rules, including: The data anonymization rules in the fine-grained data protection guidelines document are analyzed to extract the anonymization algorithm type, anonymized field location, and anonymized replacement characters. Based on the anonymized field location, the data unit to be anonymized and the target field within the unit are located in the original talent data set. Call the desensitization processing function that matches the desensitization algorithm type to perform desensitization transformation on the original data value of the target field. If the desensitization algorithm type is replacement desensitization, then use desensitization replacement characters to replace some characters in the target field; if it is masking desensitization, then replace the sensitive characters in the target field with preset masking characters; if it is encryption desensitization, then use a preset encryption function to reversibly encrypt the target field to generate the desensitized data field value. Replace the target field value in the original data unit with the data field value after desensitization to obtain the preliminary desensitized data unit. Perform data integrity verification on the preliminary desensitized data unit to check whether the format of the data unit after desensitization conforms to the preset data structure specification. If it does not conform, repeat the desensitization conversion operation until the format conforms to the specification and obtain an intermediate data form that meets the security requirements. The access permission rule entries in the fine-grained data protection guidelines document are parsed to extract the access subject role type, data operation permission list, and permission validity period. The corresponding authentication method is matched from the preset role permission mapping library according to the access subject role type. The set of allowed operation types is determined according to the data operation permission list. The time interval for permission to take effect is determined according to the permission validity period. The authentication method, operation type set, and permission time interval are combined to generate access verification rule entries. Multiple access verification rule entries constitute an access verification rule set. The transmission encryption rule entries in the fine-grained data protection guidelines document are analyzed to extract the encryption algorithm type, key length parameter, and encryption mode. The corresponding key generation method is determined according to the encryption algorithm type, the key byte length is determined according to the key length parameter, and the data block method and padding method are determined according to the encryption mode. The key generation method, key byte length, data block method, and padding method are combined to generate transmission encryption rule entries. Multiple transmission encryption rule entries constitute a transmission encryption rule set.

9. The method according to claim 8, characterized in that, The process of performing data integrity verification on the initially de-identified data units involves checking whether the format of the de-identified data units conforms to the preset data structure specifications. If not, the de-identification conversion operation is repeated until the format conforms to the specifications, resulting in an intermediate data form that meets security requirements. This includes: Obtain the preset data structure specification document, and parse the data unit format requirements in the document, including the requirements for the number of fields, the data type of fields, the length range of fields, and the logical relationship requirements between fields; The actual number of fields in the preliminary de-identified data unit is compared with the required number of fields. If the actual number of fields is less than the required number of fields, the format is determined to be non-compliant and a field missing error is generated. If the actual number of fields is more than the required number of fields, the format is determined to be non-compliant and a field redundancy error is generated. Check whether the actual data type of each field in the preliminary de-identified data unit is consistent with the field data type requirements. If there are fields with inconsistent data types, the format is determined to be non-compliant and a data type error flag is generated. Check whether the actual length of each field in the preliminary de-identified data unit is within the required field length range. If there are fields whose lengths exceed the range, the format is determined to be non-compliant and a field length error flag is generated. Verify whether the actual logical relationship between fields in the preliminary de-identified data unit is consistent with the required logical relationship between fields. If the logical relationship is inconsistent, the format is determined to be non-compliant and a logical relationship error flag is generated. The system summarizes error flags for missing fields, redundant fields, data types, field lengths, and logical relationships. If no error flags are found, the initial de-identified data unit format is deemed to conform to the specifications, and it is designated as an intermediate data form. If error flags are found, the parameters in the de-identification conversion operation are adjusted according to the error flag type, such as adjusting the position of the de-identified fields or the number of characters replaced. The de-identification conversion operation is then repeated until the generated de-identified data unit passes all format checks, resulting in an intermediate data form that meets security requirements.

10. A computer system, characterized in that, include: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the data sharing security decision-making method applying artificial intelligence as described in any one of claims 1 to 9.

Images