A Distributed Data Compliance Review Method Based on Data Security
By employing distributed local review and an improved TransE model, combined with a dual-array trie algorithm and risk source attribution analysis, the accuracy and security issues of data compliance review in distributed data environments are addressed, enabling efficient risk identification and location.
Patent Information
- Application Number
- CN202611359895.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-17
AI Technical Summary
Existing data compliance review methods are difficult to accurately identify sensitive fields and their variant expressions, determine the true degree of violation, identify potential violations and locate the source of risk in a distributed environment, and there are risks of data leakage during cross-domain transfer and centralized storage.
By employing distributed local review, a dual-array trie algorithm, an improved TransE model, and risk source attribution analysis, and through preprocessing, rule trigger strength generation, local matching, review risk distribution analysis, and relationship reasoning, a comprehensive judgment is made on compliance rule hits, cross-node risk propagation, and hidden violation relationships.
It improves the automation and security of compliance review of multi-source heterogeneous data, avoids the risk of data leakage, improves the efficiency of compliance rule hitting and the processing stability in complex scenarios, and enables accurate analysis of hidden violation relationships and risk sources.
Smart Images

Figure CN123116302A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data security technology, and in particular to a distributed data compliance review method based on data security. Background Technology
[0002] With the continuous flow of data resources across government, finance, healthcare, industrial internet, and enterprise operations, data is typically distributed across different business systems, databases, interface platforms, and audit nodes, forming a multi-node, multi-entity, and multi-path data processing environment. Data compliance review requires simultaneous attention to multiple elements, including data objects, field attributes, access subjects, operational behaviors, permission status, flow paths, and audit records, continuously assessing the access, sharing, retrieval, transmission, and auditing processes. Existing data compliance review methods often rely on centralized collection, manual rule verification, or single-field matching, concentrating raw data from distributed nodes onto a unified platform for review. This easily increases the risks of cross-domain data transfer, centralized storage leaks, and unauthorized copying, failing to meet the security requirements of data remaining within its domain, local review, and unified aggregation.
[0003] Existing compliance review methods typically focus on static rule matching, manual verification, or simple log statistics, lacking effective analytical capabilities for co-occurrence relationships of fields among multi-source heterogeneous data, changes in access behavior, cross-node flow paths, and hidden risk sources. In actual review processes, the same data object may have different naming conventions, different interface encapsulation forms, and different business call paths. Relying solely on fixed-word matching is insufficient to accurately identify sensitive fields and their variant expressions. The same access behavior may also generate different compliance risks due to differences in permission status, flow direction, operation frequency, and audit retention. It is difficult to determine the true degree of violation based solely on a single hit result. Traditional string matching methods are prone to low efficiency in status retrieval, inaccurate path positioning, and isolated hit results when faced with long fields, complex interface parameters, and large amounts of log text. Traditional relational reasoning models are also insufficient to fully express the complex relationships between fields, behaviors, permissions, paths, and risks in data compliance scenarios, resulting in insufficient identification of potential violations, inaccurate positioning of risk sources, and a lack of traceable evidence in review reports.
[0004] Therefore, how to provide a distributed data compliance review method based on data security is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a distributed data compliance review method based on data security. This invention comprehensively utilizes distributed local review, a dual-array trie algorithm, an improved TransE model, and a risk source attribution analysis data security processing method. It details the entire process from collecting data compliance review elements from various distributed data nodes, through preprocessing, rule trigger strength generation, local term matching, review risk distribution analysis, relationship reasoning, risk source location, to the generation of a final report. Structurally, it innovatively incorporates ternary interval conversion, segmented jump distance rearrangement, component residual recursive correction, vector internal differential perturbation, and component interleaved projection, ensuring that the original data remains within the local nodes while comprehensively identifying compliance rule hits, cross-node risk propagation, and implicit violation relationships. Compared with existing technologies, this invention offers advantages such as data remaining within the data domain, high review efficiency, accurate risk identification, clear source location, and applicability to multi-source heterogeneous distributed data environments.
[0006] A distributed data compliance review method based on data security according to an embodiment of the present invention includes: Data compliance review elements are collected from each distributed data node, and these elements are preprocessed to generate the data objects to be reviewed and node review summaries. Based on the node review summary, the cross-field co-occurrence frequency is discretized, and a nonlinear interactive accumulation operation is performed to generate a rule trigger intensity sequence; The algorithm of double array trie is used to perform local matching of the data objects to be reviewed, the state is filtered by ternary interval conversion, and the segment order addressing is performed by segment jump distance rearrangement to obtain the term hit result; Align the node review summary with the term hit results by field behavior trajectory, perform item-by-item deviation decomposition and local conflict expansion, and generate a review risk distribution sequence through cross-node propagation path calibration and path stability recursive analysis. The review risk distribution sequence is processed by ternary extraction to generate review relationship triples. An improved TransE model is constructed, and relationship reasoning is performed on the review relationship triples. Error backfilling is performed based on component residual recursive correction. Vector internal difference perturbation is introduced to process adjacent term differences. Component staggered projection is used to arrange components in an alternating manner to obtain the relationship discrimination result. Based on the term hit results and relationship discrimination results, a hierarchical structure verification is performed. The rule trigger intensity sequence is used as the hierarchical recursive input. Contribution accumulation path analysis and risk source attribution ranking are performed to generate risk source location results. The risk distribution sequence and risk source location results are hierarchically compressed and encoded, and cross-node consistency verification and redundancy resolution are performed on the structured review summary to generate a distributed data compliance review report.
[0007] Optionally, the data compliance review elements include data objects, field attributes, access subjects, operation behaviors, permission status, flow paths, and audit records.
[0008] Optionally, the preprocessing of data compliance review elements includes data cleaning, format standardization, local numbering, field classification, missing data completion, duplicate removal, and counting statistics.
[0009] Optionally, generating the review summary of the data object and node to be reviewed includes: Establish local collection channels on each distributed data node, and read the original review entries from the node resource directory, interface call chain, permission credential index, data flow credential and local audit index according to the collection cycle. Configure the node serial number, source location number and time slice number for each original review entry. The original review items are processed by time slice merging, field path concatenation, subject identity desensitization, flow link compression, item hash digest generation and abnormal format removal, and the processed original review items are converted into data objects to be reviewed with a unified index structure. The node review summary is generated by statistically analyzing the object size, call frequency, permission change frequency, cross-domain transfer frequency, and audit retention frequency of the data objects to be reviewed based on the node serial number. The statistical results are then combined with the source location number, time slice number, and item hash digest to generate the node review summary.
[0010] Optionally, the generation rule triggers an intensity sequence, including: Read the node review summary according to the node identifier and time slice number, and arrange the object size, call frequency, number of permission changes, number of cross-domain transfers and number of audit retentions within the same time slice in order to generate a field behavior sequence; The co-occurrence frequency of the field behavior sequence is discretized across fields. The number of times any two types of fields appear simultaneously in the current time slice is taken as the co-occurrence count. The co-occurrence count is divided by the sum of the occurrence counts of all fields in the current time slice to obtain the co-occurrence frequency. The co-occurrence frequencies are divided into discrete levels in ascending order and the first structure tensor is generated. Arrange the access subject dimension, data object dimension, and operation behavior dimension in the first structural tensor into a three-axis sequence to obtain the structural tensor transition sequence. Subtract the structural tensor transition sequences of adjacent time slices one by one to obtain the time series fluctuation term. A nonlinear interactive accumulation operation is performed on the structure tensor transition sequence. The structure tensor transition value of the current time slice is multiplied by the structure tensor transition value of the previous time slice, and then the absolute value of the time series fluctuation term is added to obtain the rule trigger intensity value of the current time slice. The rule trigger intensity values are arranged in order according to the time slice number to generate the rule trigger intensity sequence.
[0011] Optionally, obtaining the term matching result includes: Based on the rules, a local search term is generated by triggering the intensity sequence. A double-array trie structure containing a base address array, a check array, and a termination marker array is constructed based on the local search term. Read the data object to be reviewed, convert the current matching status, current character encoding, and the associated position of the check array into the base address interval, character interval, and check interval, respectively, and perform ternary interval conversion on the base address interval, character interval, and check interval to obtain the status filtering result; The long-field text and log text in the data object to be reviewed are divided into character segments. A jump distance sequence is generated based on the first and last character segment encoding and the difference in the encoding of adjacent characters within the segment. The segment jump distance rearrangement is used to determine the access position within the segment and the starting state of the next character segment, thus obtaining the segment addressing result. The state filtering results and segment addressing results are used to confirm the hit. For the position where the term termination mark exists in the termination mark array, the hit term number, hit object number, hit field position and hit time segment number are recorded to generate the term hit result.
[0012] Optionally, generating the review risk distribution sequence includes: Read the node review summary and term matching results, align the field behavior trajectory according to the node identifier, time slice number, object number and field position, and generate a multi-source review result sequence; The multi-source review result sequence is decomposed item by item to obtain the call deviation by subtracting the average of the current call frequency from the average of the historical call frequency; the permission deviation by subtracting the average of the current number of permission changes from the average of the historical number of permission changes; and the flow deviation by subtracting the average of the current number of cross-domain transfers from the average of the historical number of cross-domain transfers. The review conflict chain set is formed according to the time slice order. The review conflict chain set is expanded locally, and conflict chains with different access subjects, different permission states or different flow paths under the same time slice and the same object are expanded adjacently, and cross-node propagation paths are generated according to node identifiers and flow paths; The path is labeled and stability is recursively analyzed for cross-node propagation paths. The path stability is obtained by dividing the number of times the same path appears in consecutive time slices by the total number of time slices to which the path belongs. The path stability is multiplied by the number of corresponding conflict chains to obtain the path risk distribution value. The review risk distribution sequence is generated according to the node identifier, time slice number and path order.
[0013] Optionally, obtaining the relationship determination result includes: The review risk distribution sequence is processed by ternary extraction to generate review relation triples. After the translation operation of the original TransE model, component residual recursive correction is set. Before the translation operation, vector internal difference perturbation is set. Before relation discrimination, component staggered projection is set to construct an improved TransE model. Read the head entity, relation, and tail entity from the examine relation triplet, and look up the corresponding vector representations in the entity vector table and relation vector table respectively to obtain the head entity vector, relation vector, and tail entity vector. Add the head entity vector and relation vector to obtain the predicted tail entity vector. Perform component-wise difference processing on the tail entity vector and the predicted tail entity vector to obtain the component residuals. Fill the component residuals back into the head entity vector, relation vector, and predicted tail entity vector to obtain the corrected predicted tail entity vector. Introducing vector internal difference perturbation for adjacent component difference processing, the head entity vector, relation vector, and tail entity vector are processed for adjacent component difference processing respectively to obtain head entity difference sequence, relation difference sequence, and tail entity difference sequence. A difference deviation sequence is generated based on the head entity difference sequence, relation difference sequence, and tail entity difference sequence. The component interleaving projection is used to perform component interleaving arrangement. The head entity vector, relation vector and tail entity vector are interleaved according to the main component, object component, permission component, behavior component, rule component and risk component. The interleaved head entity vector, relation vector and tail entity vector are then processed by component difference to obtain the interleaved deviation sequence. Based on the corrected predicted tail entity vector, the difference deviation sequence, and the staggered deviation sequence, a relation discrimination value is generated. The relation discrimination value is then associated with the examined relation triplet to generate a relation discrimination result. Using real review relation triples as positive samples and triples obtained by replacing any entity or relation in the review relation triples as negative samples, calculate the relationship discriminant values for positive and negative samples respectively. Update and improve the TransE model according to the training direction where the relationship discriminant value of positive samples is less than that of negative samples, and obtain the improved TransE model after training.
[0014] Optionally, the generation of risk source location results includes: Read the term hit results, relationship discrimination results, and rule trigger intensity sequence, and perform hierarchical merging according to node identifier, time slice number, object number, and field position to generate a hierarchical verification sequence; The hierarchical verification sequence is structurally verified. The absolute value of the difference between the number of term hits and the number of relation discriminations is obtained. The hit difference, relation discrimination, and rule trigger intensity of the corresponding time slice are added together to obtain the risk contribution value of the current level. The risk contribution value of the previous level is added together with the risk contribution value of the current level to obtain the recursive risk contribution value. The risk contribution values are concatenated in the order of field position, operation behavior, permission status and flow path to generate a contribution accumulation path. The total path contribution value is obtained by adding up the recursive risk contribution values within the same path. The risk source is then sorted from largest to smallest according to the total path contribution value to generate the risk source location result.
[0015] Optionally, generating the distributed data compliance review master report includes: Read the risk distribution sequence and risk source location results, and trim the summary field according to the node identifier, time slice number, risk source, path risk distribution value, path contribution total value and risk source sorting result. Convert the trimmed summary field into a fixed-length coded segment to generate a structured review summary. Cross-node consistency verification is performed on structured review summaries. Structured review summaries with the same risk source location, the same path identifier, and the same time segment number are compared, and consistent summaries and summaries to be reviewed are generated according to the difference in risk level and the difference in risk type. Redundancy is eliminated in the consistent summary and the summary to be reviewed. Summaries with the same node identifier, the same time slot number, the same risk type, and the same risk source location are merged into report records. The report records are arranged hierarchically according to node identifier, risk type, risk level, risk source location, and time slot number to generate a distributed data compliance review master report.
[0016] The beneficial effects of this invention are: This invention enhances the automation and security of compliance review for multi-source heterogeneous data through a distributed local review and data security processing workflow. Compared to existing centralized collection and review methods, this invention completes the collection, preprocessing, local matching, and risk analysis of data compliance review elements on each distributed data node, avoiding the risks of data leakage, unauthorized copying, and cross-domain transfer caused by the centralized movement of original data. Through the collaborative processing of rule trigger strength sequences, dual-array trie algorithms, and review risk distribution sequences, it can quickly identify abnormal features in fields, behaviors, permissions, and transfer paths, improving the efficiency of compliance rule hitting and the processing stability in complex review scenarios.
[0017] This invention improves the TransE model for relational reasoning using triples of review relationships, and combines this with a risk source localization and hierarchical compressed report generation process to achieve accurate analysis of implicit violation relationships, cross-node risk propagation, and risk sources. Component residual recursive correction, vector internal differential perturbation, and component staggered projection enhance relational discrimination capabilities, enabling the identification of abnormal access, unauthorized sharing, abnormal permissions, and transfer risks that are difficult to detect with simple term matching. Compared to existing methods relying on manual verification or static rule matching, this invention offers the technical advantages of data not leaving the domain, high review efficiency, accurate risk identification, clear source localization, standardized report generation, and applicability to distributed data environments. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a distributed data compliance review method based on data security proposed in this invention; Figure 2 This is a structural block diagram of the double-array trie algorithm for a distributed data compliance review method based on data security proposed in this invention; Figure 3 This is a functional diagram of the improved TransE model of the distributed data compliance review method based on data security proposed in this invention. Detailed Implementation
[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0020] refer to Figure 1 , Figure 2 and Figure 3 A distributed data compliance review method based on data security includes: Data compliance review elements are collected from each distributed data node, and these elements are preprocessed to generate the data objects to be reviewed and node review summaries. Based on the node review summary, the cross-field co-occurrence frequency is discretized, and a nonlinear interactive accumulation operation is performed to generate a rule trigger intensity sequence; The algorithm of double array trie is used to perform local matching of the data objects to be reviewed, the state is filtered by ternary interval conversion, and the segment order addressing is performed by segment jump distance rearrangement to obtain the term hit result; Align the node review summary with the term hit results by field behavior trajectory, perform item-by-item deviation decomposition and local conflict expansion, and generate a review risk distribution sequence through cross-node propagation path calibration and path stability recursive analysis. The review risk distribution sequence is processed by ternary extraction to generate review relationship triples. An improved TransE model is constructed, and relationship reasoning is performed on the review relationship triples. Error backfilling is performed based on component residual recursive correction. Vector internal difference perturbation is introduced to process adjacent term differences. Component staggered projection is used to arrange components in an alternating manner to obtain the relationship discrimination result. Based on the term hit results and relationship discrimination results, a hierarchical structure verification is performed. The rule trigger intensity sequence is used as the hierarchical recursive input. Contribution accumulation path analysis and risk source attribution ranking are performed to generate risk source location results. The risk distribution sequence and risk source location results are hierarchically compressed and encoded, and cross-node consistency verification and redundancy resolution are performed on the structured review summary to generate a distributed data compliance review report.
[0021] In this embodiment, the data compliance review elements include data objects, field attributes, access subjects, operation behaviors, permission status, flow paths, and audit records.
[0022] In this embodiment, the preprocessing of data compliance review elements includes data cleaning, format standardization, local numbering, field classification, missing data completion, duplicate removal, and counting statistics.
[0023] In this embodiment, generating the review summary of the data object to be reviewed and the node includes: Establish local collection channels on each distributed data node, and read the original review entries from the node resource directory, interface call chain, permission credential index, data flow credential and local audit index according to the collection cycle. Configure the node serial number, source location number and time slice number for each original review entry. The original review entries undergo time-slice merging, field path concatenation, subject identity desensitization, flow link compression, entry hash digest generation, and abnormal format removal. The processed original review entries are then converted into data objects to be reviewed with a unified index structure. Specifically, the conversion of the processed original review entries into data objects to be reviewed with a unified index structure involves: The node serial number is used as the first-level index, the source location number as the second-level index, and the time slice number as the third-level index. The time slice number is obtained by dividing the time difference between the collection time of the review item and the preset baseline time by 10 minutes and then rounding down. The field path concatenation result is converted into field path encoding, the subject identity desensitization result is converted into subject anonymization encoding, and the circulation link compression result is converted into circulation link encoding. Together with the item hash digest, they form the object index key. Review entries with the same node serial number, source location number, time slice number, and object index key are merged. During merging, the number of entries, the number of calls, the number of permission changes, and the number of transfers are accumulated, and the latest audit time is retained. If any review entry is missing any of the node serial number, time slice number, field path code, or entry hash digest, it is determined to be an incomplete index entry and is removed. If the number of duplicate entries under the same object index key exceeds one, only one master record is retained and the number of duplicates is written to the statistics field, forming a data object to be reviewed that includes node index, source index, time index, object index, main index, transfer index, and digest index. The node review summary is generated by statistically analyzing the object size, call frequency, permission change frequency, cross-domain transfer frequency, and audit retention frequency of the data objects to be reviewed based on the node serial number. The statistical results are then combined with the source location number, time slice number, and item hash digest to generate the node review summary.
[0024] In this embodiment, the generation of the rule-triggered intensity sequence includes: Read the node review summary according to the node identifier and time slice number, and arrange the object size, call frequency, number of permission changes, number of cross-domain transfers and number of audit retentions within the same time slice in order to generate a field behavior sequence; The cross-field co-occurrence frequency of the field behavior sequence is discretized. The number of times any two types of fields appear simultaneously within the current time slice is taken as the co-occurrence count. The co-occurrence count is divided by the sum of the occurrence counts of all fields within the current time slice to obtain the co-occurrence frequency. The co-occurrence frequencies are divided into discrete levels in ascending order and the first structure tensor is generated. Specifically, the cross-field co-occurrence frequency discretization of the field behavior sequence is as follows: The field behavior sequence is split into multiple time slice sequences according to the time slice number. The occurrence frequency of each field category is counted in each time slice sequence. The co-occurrence frequency of any two fields in the same data object to be reviewed or the same audit record is taken as the co-occurrence frequency of the field category pair. The occurrence frequency of all field categories in the current time slice is added together to obtain the total number of field occurrences. The co-occurrence frequency of each field category pair is divided by the total number of field occurrences to obtain the corresponding co-occurrence frequency. When the total number of field occurrences is zero, the co-occurrence frequency is recorded as zero. The co-occurrence frequencies of all field category pairs in the current time slice are arranged in ascending order and divided into 5 discrete levels according to the number of arrangements. The lowest group is recorded as level 1 and the highest group is recorded as level 5. Using the time slice number, field category pair number and discrete level as three index directions, the discrete level of each field category pair in the corresponding time slice is written into the tensor position. The position where no co-occurrence relationship occurs is written as zero value, generating the first structure tensor. The access subject dimension, data object dimension, and operation behavior dimension in the first structural tensor are arranged along three axes to obtain the structural tensor transition sequence. The time series fluctuation term is obtained by subtracting the structural tensor transition sequences of adjacent time slices, where the three-axis arrangement is as follows: Read the discrete level values corresponding to each time slice from the first structure tensor, and establish the access subject axis, data object axis, and operation behavior axis respectively. The access subject axis is arranged in ascending order according to the subject anonymous code, the data object axis is arranged in ascending order according to the object index key, and the operation behavior axis is arranged in ascending order according to the behavior type code. Write the discrete level values of the same access subject, the same data object, and the same operation behavior in the same time slice to the three-axis coordinate position. If there are multiple discrete level values at the same three-axis coordinate position, add the discrete level values and divide by the quantity to get the average discrete level value of the current position. If there is no corresponding record, write zero value to the current position. Arrange the three-axis tensors generated in each time slice in order from early to late according to the time slice number to obtain the structure tensor transition sequence. Subtract the tensor values of the same three-axis coordinate position in two adjacent time slices one by one to obtain the time series fluctuation term of the corresponding coordinate position, which is used to represent the change amplitude of the combination of access subject, data object, and operation behavior in continuous time slices. A nonlinear interactive accumulation operation is performed on the structure tensor transition sequence. The structure tensor transition value of the current time slice is multiplied by the structure tensor transition value of the previous time slice, and then the absolute value of the time series fluctuation term is added to obtain the rule trigger intensity value of the current time slice. The rule trigger intensity values are arranged in order according to the time slice number to generate the rule trigger intensity sequence.
[0025] In this embodiment, obtaining the term matching result includes: Based on the rules, a local search terminology is generated by triggering the intensity sequence. A double-array trie structure containing a base address array, a check array, and a termination marker array is then constructed based on the local search terminology. The local search terminology is generated based on the rule trigger intensity sequence. Specifically, the rule trigger intensity sequence is read according to the node identifier and time slice number. Each rule trigger intensity value is associated with the corresponding review rule number. The rule trigger intensity values within the same node are normalized. The review rule numbers are arranged from largest to smallest according to the normalization value. The review rule number with a normalization value of not less than 0.60 is selected as the local search rule. If there are fewer than 20 review rule numbers that meet the conditions, they are padded to 20 in the order of arrangement. The search fields, field aliases, interface parameter names, path fragments and review keywords corresponding to the local search rule are read. The same text content is hashed to remove duplicates. Each retained text is configured with a term number, corresponding rule number and time slice number, and the local search terminology is generated. A double-array trie structure containing a base address array, a check array, and a termination marker array is constructed based on the local search terminology. Specifically, each term in the local search terminology is converted into a character encoding sequence and arranged in lexicographical order. The root state number 0 is used as the initial state. The character encoding of each term is read sequentially. For all sub-character encodings in the current state, the minimum available base address value is found so that the array positions obtained by adding each sub-character encoding to the base address value are all unoccupied. The minimum available base address value is written to the position corresponding to the current state in the base address array. The array position corresponding to each sub-character is written to the check array. The value of the current position in the check array is set to the current state number. When the character encoding sequence reaches the end of the term, the corresponding state position is written to the termination marker array. The termination marker array records the term number, the corresponding rule number, and the time slice number. This process is repeated until all terms in the local search terminology are written, resulting in a double-array trie structure for local matching. Read the data object to be reviewed, and convert the current matching status, current character encoding, and associated position of the check array into base address interval, character interval, and check interval, respectively. Perform ternary interval conversion on the base address interval, character interval, and check interval to obtain the status filtering result. Specifically, the status filtering result is as follows: Read the current character, the previous character, and the next character from the data object to be reviewed. Take the value of the current matching state in the base address array as the center base address value. Take the three consecutive available array positions before and after the center base address value to form a base address interval. Take the current character encoding as the center character value. Combine the difference between the previous character encoding and the current character encoding, and the difference between the next character encoding and the current character encoding to form a character interval. Read all positions in the verification array whose parent state number is equal to the current matching state and merge them into verification intervals according to consecutive subscripts. Each base address value in the base address interval is added to each character value in the character interval to obtain the candidate state position. The positions where the candidate state position falls into the verification interval and the corresponding value in the verification array is equal to the current matching state are retained, and the remaining candidate state positions are discarded. If the same character generates multiple retained positions, they are sorted in ascending order according to the distance between the candidate state position and the position obtained by adding the center base address value to the current character encoding. The position with the highest sorting is retained as the valid state of the current character. The valid states corresponding to each character are arranged in the matching order to form the state filtering result. Long-field text and log text in the data object to be reviewed are segmented into character segments. A jump distance sequence is generated based on the first and last character segment codes and the difference in codes between adjacent characters within a segment. Segmented jump distance rearrangement is used to determine the access position within a segment and the starting state of the next character segment, resulting in a segment addressing result. Specifically, the segment addressing result is as follows: Long text fields and log text are divided into character segments according to field separators, path separators, operation separators, and a fixed length. The fixed length is set to 16 characters. Continuous text exceeding 16 characters and without a separator is further divided into the next character segment. The character encoding of the first character and the character encoding of the last character of each character segment are read. The absolute value of the difference between two adjacent character encodings in the segment is calculated in turn. The first character encoding of the segment is added to the first difference to obtain the first jump distance value. The previous jump distance value is added to the next adjacent difference to obtain the subsequent jump distance values, forming the jump distance sequence of the character segments. The termination state of the previous character segment is used as the starting state of the current character segment. The first character segment uses the root state as its starting state. The starting state is added to the jump values in the jump distance sequence to obtain the access position within the segment. The access position within the segment is verified in the verification array to see if it belongs to the current starting state. If the verification passes, the access position within the segment is retained. If the verification fails, the positions are rearranged in ascending order of jump distance values and addressed again. The last verified position is used as the starting state of the next character segment. The access positions within each segment and the starting state of the next character segment are recorded in the character segment order to obtain the segment order addressing result. The state filtering results and segment addressing results are used to confirm the hit. For positions in the termination marker array where a term termination marker exists, the hit term number, hit object number, hit field position, and hit time segment number are recorded. This generates the term hit result, specifically as follows: Align the valid state sequence in the state filtering results with the intra-segment access positions in the segment addressing results according to the hit object number, field position, and time slice number. Compare the array positions corresponding to the two types of results bit by bit. When the valid state corresponding to the same character position is the same as the intra-segment access position, or both point to the same termination marker array position, it is recorded as a one-time fatal hit. Divide the number of one-time fatal hits by the total number of character positions participating in the comparison to obtain the hit consistency rate. The hit consistency rate is not less than 0.85. If the last valid state or the last intra-segment access position has a term termination marker in the termination marker array, it is confirmed that the field or log segment hits the corresponding term. Read the term ID and rule ID stored in the termination marker array, and combine them with the object ID, field position, time slice number and node identifier in the data object to be reviewed to generate a hit record. If there are multiple hit records for the same field position, retain the hit record with the highest hit consistency rate. If the hit consistency rates are the same, retain the hit record with the longest character length. Finally, arrange the hit records according to the node identifier, time slice number and hit object ID to generate the term hit result.
[0026] In this embodiment, generating the review risk distribution sequence includes: Read the node review summary and term matching results, align the field behavior trajectory according to the node identifier, time slice number, object number and field position, and generate a multi-source review result sequence; The multi-source review result sequence is decomposed item by item to obtain the call deviation by subtracting the average of the current call frequency from the average of the historical call frequency; the permission deviation by subtracting the average of the current number of permission changes from the average of the historical number of permission changes; and the flow deviation by subtracting the average of the current number of cross-domain transfers from the average of the historical number of cross-domain transfers. The review conflict chain set is formed according to the time slice order. The review conflict chain set undergoes local conflict expansion. Conflict chains with different access subjects, different permission states, or different flow paths under the same time slice and the same object are extended adjacently. Cross-node propagation paths are generated according to node identifiers and flow paths. Specifically, the local conflict expansion of the review conflict chain set involves: The set of conflict chains is grouped according to node identifier, time slice number, and hit object number. Within each group, the access subject, permission status, and flow path corresponding to each conflict chain are read. When any two conflict chains have the same hit object number and time slice number, and at least one of the access subject, permission status, or flow path is different, the two conflict chains are marked as adjacent conflict chains. The conflict expansion value between adjacent conflict chains is calculated. The conflict expansion value is the sum of the difference value of access subject, permission status, and flow path. The same value is recorded as 0, and the different value is recorded as 1. When the conflict expansion value is not less than 1, the adjacent conflict chains are added to the same local expansion group, and the conflict chains in the local expansion group are connected according to the order of the flow path. If adjacent conflict chains are distributed across different distributed data nodes, then the node identifier is used as the path node, and the flow path is used as the path edge. The corresponding nodes are connected in sequence to form a cross-node propagation path. Each cross-node propagation path is recorded as the starting node, ending node, nodes passed through, flow path number, conflict expansion value, and time slice number to complete the local conflict expansion processing of the conflict chain set. Path identification and stability recursive analysis are performed on cross-node propagation paths. The path stability is obtained by dividing the number of times the same path appears within consecutive time slices by the total number of time slices to which the path belongs. The path stability is then multiplied by the number of corresponding conflict chains to obtain the path risk distribution value. A review risk distribution sequence is generated according to node identifier, time slice number, and path order. Specifically, the path identification and stability recursive analysis for cross-node propagation paths involves: Read the starting node, ending node, nodes passed through, path number, and time slice number of each cross-node propagation path. Concatenate the paths in the order of starting node-passing node-ending node-path number to generate a path identifier. Group records with the same path identifier into the same path. Arrange the occurrence records of the same path from earliest to latest according to the time slice number. Set the consecutive occurrence count and the total number of time slices to which the path belongs for each path. When the path appears in the current time slice and also appears in the previous time slice, the consecutive occurrence count is incremented by 1. When the path appears in the current time slice but does not appear in the previous time slice, the consecutive occurrence count is recorded as 1. The total number of time slices to which the path belongs is the number of time slices from the time slice when the path first appears to the current time slice. The path stability is obtained by dividing the maximum number of consecutive occurrences of the same path by the total number of time slices to which the path belongs. The number of conflict chains associated with the path in the current time slice is counted. The path stability is multiplied by the number of conflict chains to obtain the path risk distribution value. The path records are sorted according to the node identifier, time slice number, path identifier, and path risk distribution value to form a review risk distribution sequence.
[0027] In this embodiment, obtaining the relationship discrimination result includes: The review risk distribution sequence is subjected to ternary extraction to generate review relation triples. After the translation operation of the original TransE model, component residual recursive correction is set. Before the translation operation, vector internal difference perturbation is set. Before relation discrimination, component staggered projection is set to construct an improved TransE model. Specifically, the review risk distribution sequence is subjected to ternary extraction to generate review relation triples. Read the node identifier, time slice number, path identifier, risk source location, path risk distribution value and associated conflict chain information in the risk distribution sequence. Parse the starting node, passing nodes, ending nodes and flow path number from the path identifier. Take the corresponding access subject or starting node in the associated conflict chain as the head entity, and take the operation behavior, flow path number or risk trigger type as the relation, and take the field object, risk source location or ending node as the tail entity. Records with the same head entity, relation, and tail entity within the same time slice are merged. During merging, the path risk distribution values are added to obtain the triplet risk value. The occurrence frequency, node identifier, and time slice number are recorded. When there are multiple relations between the same head entity and the same tail entity, the triplets formed by different relations are retained. When the triplet risk value is zero, it is not written into the training sample. Finally, the review relation triplet is generated, consisting of head entity, relation, tail entity, triplet risk value, and time slice number. Read the head entity, relation, and tail entity from the examine relation triplet, and look up the corresponding vector representations in the entity vector table and relation vector table respectively to obtain the head entity vector, relation vector, and tail entity vector. Add the head entity vector and relation vector to obtain the predicted tail entity vector. Perform component-wise difference processing on the tail entity vector and the predicted tail entity vector to obtain the component residuals. Fill the component residuals back into the head entity vector, relation vector, and predicted tail entity vector to obtain the corrected predicted tail entity vector. An internal vector difference perturbation is introduced for adjacent-term difference processing. Adjacent-component difference processing is performed on the head entity vector, relation vector, and tail entity vector respectively, resulting in head entity difference sequences, relation difference sequences, and tail entity difference sequences. A difference deviation sequence is then generated based on these sequences. Specifically, the internal vector difference perturbation for adjacent-term difference processing is as follows: Read the head entity vector, relation vector, and tail entity vector corresponding to the review relation triplet. The three vectors have the same dimension, which is 120 dimensions. Process them sequentially from dimension 1 to dimension 120 according to the component index. For the head entity vector, subtract the previous component from the next component to obtain 119 head entity difference values. Apply the same method to the relation vector and tail entity vector to obtain 119 relation difference values and 119 tail entity difference values respectively. Add the head entity difference value and relation difference value at the same index position to obtain the predicted difference value. Subtract the tail entity difference value at the corresponding index position from the predicted difference value to obtain the difference deviation value at the index position. Arrange the 119 difference deviation values in component order to generate a difference deviation sequence, which is used to represent the degree of deviation of the review relation triplet in the changing trend of adjacent vector components. The component interleaving projection is used to perform component interleaving arrangement. The head entity vector, relation vector and tail entity vector are interleaved according to the main component, object component, permission component, behavior component, rule component and risk component. The interleaved head entity vector, relation vector and tail entity vector are then processed by component difference to obtain the interleaved deviation sequence. Based on the corrected predicted tail entity vector, the differenced deviation sequence, and the staggered deviation sequence, a relation discrimination value is generated. This relation discrimination value is then correlated with the examined relation triplet to generate a relation discrimination result. Specifically, the generation of the relation discrimination result is as follows: Read the true tail entity vector, the corrected predicted tail entity vector, the difference deviation sequence, and the staggered deviation sequence corresponding to the triplet of the review relation. Normalize each vector component to a value range of -1 to 1. Calculate the basic deviation value, which is the sum of the absolute values of the differences between the true tail entity vector and the corrected predicted tail entity vector at 120 identical component positions, divided by 120. Calculate the difference deviation value, which is the sum of the absolute values of the 119 difference deviation values in the difference deviation sequence, divided by 119. Calculate the staggered deviation value, which is the sum of the absolute values of 120 staggered deviation values in the staggered deviation sequence divided by 120. Add the base deviation value, the differential deviation value, and the staggered deviation value to obtain the relationship discrimination value. When the relationship discrimination value is less than 0.50, it is marked as a low-risk relationship; when the relationship discrimination value is greater than or equal to 0.50 and less than 1.20, it is marked as a medium-risk relationship; and when the relationship discrimination value is greater than or equal to 1.20, it is marked as a high-risk relationship. Associate the relationship discrimination value, risk relationship label, review relationship triplet, node identifier, time slice number, and path identifier into the same result record to generate the relationship discrimination result. Using real review relation triples as positive samples and triples obtained by replacing any entity or relation in the review relation triples as negative samples, the relation discriminant values for positive and negative samples are calculated respectively. The TransE model is updated and improved according to the training direction where the relation discriminant value of positive samples is less than that of negative samples, resulting in the trained improved TransE model. Specifically, the trained improved TransE model is as follows: The triples in the review relationship triples that have been confirmed to have compliance or risk relationships are used as positive samples. For each positive sample, the head entity, relationship or tail entity are replaced to generate 3 negative samples. The replacement objects are selected from the entity set or relationship set of the same distributed data node, and the replaced negative samples must not be duplicates of the existing positive samples. The relationship discrimination values of the positive samples and the negative samples are calculated separately. The relationship discrimination value is obtained by adding the basic deviation value, the differential deviation value and the staggered deviation value. Calculate the training loss value, which is the result of subtracting the negative sample relationship discriminant value from the positive sample relationship discriminant value and adding the margin constant. The margin constant is set to 1. When the result is less than 0, the training loss value is recorded as 0. When the result is greater than 0, the entity vector table and relationship vector table are updated according to the result. After each round of training, each component of the entity vector and relationship vector is normalized to a value range of -1 to 1. Training is stopped after 100 rounds of continuous training or when the change in the training loss value is less than 0.001 for 10 consecutive rounds. The entity vector table, relationship vector table, component residual recursive correction parameters, vector internal difference perturbation parameters, and component staggered projection arrangement rules at the time of stopping training are retained to obtain the improved TransE model after training.
[0028] In this embodiment, generating the risk source location result includes: The algorithm reads the term hit results, relation determination results, and rule trigger strength sequence, and performs hierarchical merging based on node identifier, time slice number, object number, and field position to generate a hierarchical verification sequence. Specifically, the hierarchical merging process for generating the hierarchical verification sequence is as follows: Read the node identifier, time slice number, hit object number, and hit field position from the term hit results, read the node identifier, time slice number, examined relationship triplet, and relationship judgment value from the relationship judgment results, and read the node identifier, time slice number, and rule trigger strength value from the rule trigger strength sequence. Use the combination of node identifier, time slice number, object number, and field position as the merge key to group term hit records, relationship judgment records, and rule trigger strength records with the same merge key into the same level position. Within each level, the number of records where terms are hit is counted as the term hit count, and the number of records where relationships are identified is counted as the relationship identification count. The relationship identification values of all records within the level are summed and then divided by the number of records where relationships are identified to obtain the relationship identification mean. When the number of records where relationships are identified is zero, the relationship identification mean is recorded as 0. The rule trigger intensity value within the same node and the same time slice is used as the rule trigger input for the level. The level positions are arranged in the order of node identifier, time slice number, object number, and field position. The term hit count, relationship identification count, relationship identification mean, and rule trigger input are written into the corresponding positions to generate a hierarchical verification sequence. The hierarchical verification sequence is structurally verified. The absolute value of the difference between the number of term hits and the number of relation discriminations is obtained. The hit difference, relation discrimination, and rule trigger intensity of the corresponding time slice are added together to obtain the risk contribution value of the current level. The risk contribution value of the previous level is added together with the risk contribution value of the current level to obtain the recursive risk contribution value. The recursive risk contribution values are concatenated according to the order of field position, operation behavior, permission status, and flow path to generate a contribution accumulation path. The recursive risk contribution values within the same path are summed to obtain the total path contribution value. The risk source attribution results are then generated by sorting the total path contribution values from largest to smallest. The specific steps for attributing risk sources are as follows: Read the recursive risk contribution value corresponding to each level in the contribution accumulation path, and summarize the contribution value according to field position, operation behavior, permission status and flow path respectively. The field position contribution value is the sum of all recursive risk contribution values associated with the same field position, the operation behavior contribution value is the sum of all recursive risk contribution values associated with the same operation behavior, the permission status contribution value is the sum of all recursive risk contribution values associated with the same permission status, and the flow path contribution value is the sum of all recursive risk contribution values associated with the same flow path. Divide the contribution value of each risk source by the sum of the contribution values of all risk sources to obtain the risk source percentage. When the sum of the contribution values of all risk sources is zero, the percentage of each risk source is recorded as zero. Sort the cumulative contribution paths in descending order of total path contribution value. If the total path contribution values are the same, sort them in descending order of risk source percentage. If the risk source percentages are the same, sort them in ascending order of time slice number. Select the field position, operation behavior, permission status and flow path at the top of the sorting results as the main risk sources. Combine the risk source name, total path contribution value, risk source percentage, node identifier, time slice number and object number to generate the risk source location result.
[0029] In this embodiment, generating the distributed data compliance review summary report includes: Read the risk distribution sequence and risk source location results, and trim the summary field according to node identifier, time slice number, risk source, path risk distribution value, path contribution total value, and risk source sorting result. Convert the trimmed summary field into a fixed-length coded segment to generate a structured review summary. The specific steps of summary field trimming are as follows: Read the node identifier, time slice number, path identifier, path risk distribution value, and associated object number from the risk distribution sequence. Read the risk source, total path contribution value, and risk source sorting result from the risk source location result. Using the node identifier, time slice number, and path identifier as the summary pruning key, merge records with the same pruning key in both types of results into the same summary candidate record. In each summary candidate record, only the node identifier, time slice number, risk source, path risk distribution value, total path contribution value, risk source sorting result, path identifier, and object number are retained. Delete the original field values, complete log content, main plaintext information, and original circulation vouchers. When there are multiple risk source records under the same pruning key, the top 3 risk sources are retained according to the risk source sorting result from smallest to largest. When there are multiple path risk distribution values for the same risk source, the maximum value is taken as the path risk distribution value of that risk source. When there are multiple path contribution total values for the same risk source, the sum of the path contribution total values is taken as the path contribution total value of the risk source. The retained fields are arranged in the order of node identifier, time slice number, risk source, path risk distribution value, path contribution total value and sorting result to form the pruned summary field. Cross-node consistency checks are performed on structured review summaries. Structured review summaries with the same risk source location, the same path identifier, and the same time slot number are compared. Consistent summaries and summaries to be reviewed are generated according to the difference in risk level and risk type. The cross-node consistency check is specifically performed as follows: Read the structured review digests uploaded by each distributed data node. Use the risk source location, path identifier, and time slice number as the consistency verification key. Group structured review digests with the same verification key but different node identifiers into the same verification group. Within each verification group, convert the risk level into a numerical level, with low risk recorded as 1, medium risk as 2, and high risk as 3. Calculate the risk level difference between any two digests. The risk level difference is the absolute value of the difference between the numerical levels of the two digests. Compare the risk type codes of the two digests. If the risk type codes are the same, record 0; if they are different, record 1. When the maximum risk level difference of all digests in the same verification group is 0 and the risk type difference value is 0, mark the digests in the verification group as consistent digests. When the maximum risk level difference is greater than 0 or there is a risk type difference value of 1, mark the digests in the verification group as digests to be reviewed. Record the maximum risk level difference, the difference risk type, and the corresponding node identifier to complete the cross-node consistency verification. Redundancy is eliminated in the consistent summary and the summary pending review. Summaries with the same node identifier, time slice number, risk type, and risk source location are merged into a single report record. These report records are then hierarchically arranged according to node identifier, risk type, risk level, risk source location, and time slice number to generate a distributed data compliance review master report. The redundancy elimination of the consistent summary and the summary pending review is specifically as follows: Read the node identifier, time slice number, risk type, risk level, risk source location, path identifier, path risk distribution value, total path contribution value, and review mark from the consistent summary and the summary to be reviewed. Use the node identifier, time slice number, risk type, and risk source location as redundancy resolution keys. Group the summaries with the same redundancy resolution keys into the same report candidate group. Within each report candidate group, count the number of summaries as the number of merges. Convert the risk level to a numerical level and take the largest numerical level as the report risk level. Add the risk distribution values of each path to obtain the total report risk distribution value. Add the total contribution values of each path to obtain the total report contribution value. Retain the path identifier with the largest total report contribution value as the main path identifier. If the report candidate group contains a summary to be reviewed, the report record is marked as pending review. If it only contains a consistent summary, it is marked as consistent. Duplicate summaries are deleted, and only report records containing node identifier, time slice number, risk type, report risk level, risk source location, main path identifier, total report risk distribution value, total report contribution value, number of merges, and review status are retained. The records are then arranged hierarchically according to node identifier, risk type, report risk level, risk source location, and time slice number to generate a distributed data compliance review master report.
[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to an enterprise-level distributed data compliance review cycle, connecting six distributed data nodes, including business data nodes, interface call nodes, permission credential nodes, data flow nodes, and audit archiving nodes. Each node recorded a total of 1862 original review entries, 2463 data objects, 518 field attribute records, 348 access subjects, 706 operation behaviors, 1765 permission status change records, 268 flow paths, and 1923 audit records. Traditional centralized review methods require uploading original fields, interface logs, and flow credentials, simulating a data transmission volume of 8.6GB; this invention completes processing locally on each node, uploading only a structured review summary, with an uploaded data volume of 2.4GB.
[0031] After the data enters the processing flow, each node performs data cleaning, format standardization, local numbering, field classification, missing data completion, duplicate removal, and counting statistics on the data compliance review elements. This process removes 184 abnormal format entries, completes 317 missing field indexes, and removes 425 duplicate audit records, resulting in 2389 data objects to be reviewed and 3680 node review summaries. Traditional methods use direct log aggregation, resulting in a duplicate record rate of 8.3%; after processing by this invention, the duplicate record rate is reduced to 1.1%.
[0032] A rule trigger strength sequence is generated based on the node review summary. Taking a node review window as an example, the object size is 1860, the call frequency is 7420, the number of permission changes is 543, the number of cross-domain transfers is 876, and the number of audit retentions is 1690. Cross-field co-occurrence frequency discretization is performed, resulting in 1248 field co-occurrence combinations. After triaxial arrangement and nonlinear interactive cumulative calculation, a rule trigger strength sequence of 1192 items is generated, with an average trigger strength of 0.64. 438 rule items have a trigger strength of not less than 0.60.
[0033] During the local matching phase, the system generates a local search term list based on the strength sequence triggered by the rules, reducing the term list from 920 candidate terms to 276. A base address array, a check array, and a termination marker array are constructed, and state filtering is performed using ternary interval conversion. Segment addressing is performed using segmented jump distance rearrangement. Taking a 184-character interface log as an example, it is divided into 13 character segments and 13 jump distance sequences are generated. Traditional methods require 184 state accesses, while this invention only requires 96. Throughout the entire cycle, this invention completes 4126 text matches, obtaining 864 term matching results.
[0034] In the risk distribution analysis phase, the node review summaries and term hit results were aligned with field behavior trajectories, generating 1827 multi-source review result sequences. Call deviation, permission deviation, and flow deviation were calculated, forming a set of 984 review conflict chains. The data object currently has 126 calls, with a historical average of 43, resulting in a call deviation of 83. The current number of permission changes is 9, with a historical average of 2, resulting in a permission deviation of 7. The current number of cross-domain flows is 17, with a historical average of 5, resulting in a flow deviation of 12. After local conflict expansion, 5 cross-node propagation paths were formed, one of which appeared 6 times within 8 consecutive time slices, resulting in a path stability of 0.75, 11 associated conflict chains, and a path risk distribution value of 8.25.
[0035] In the relational reasoning phase, triplet extraction was performed on the censorship risk distribution sequence, generating 5260 censorship relation triplets. The training samples included 3600 positive samples and 1080 negative samples. Negative samples were obtained by replacing the access subject, operation behavior, or field object. The positive sample was subject anonymization code A023-batch export-field object F118, with a triplet risk value of 0.82. The negative samples included subject anonymization code A411-batch export-field object F118, subject anonymization code A023-normal query-field object F118, and subject anonymization code A023-batch export-field object F702. The improved TransE model used a 120-dimensional vector, and training stopped when the loss change was below 0.001 after 78 rounds. In the test set of 2900 triplets, this invention identified 1684 high-risk relations, 326 medium-risk relations, and 990 low-risk relations.
[0036] During the risk source identification phase, a hierarchical structure verification was performed based on the term hit results, relationship identification results, and rule trigger strength sequence to generate risk source identification results. Statistics show that the contribution of the flow path was 38.7%, permission status contributed 27.4%, operational behavior contributed 21.6%, and field position contributed 12.3%. The path with the highest contribution had a total contribution value of 36.8, corresponding to cross-node flow path anomalies; the second highest contribution source was frequent changes in permission status, with a total contribution value of 27.1. 4100 structured review summaries were generated. After cross-node consistency verification and redundancy resolution, 1126 report records were formed, 2864 consistent summaries were generated, and 262 summaries are pending review.
[0037] In the comparative experiment, the method of this invention and the traditional centralized rule review method used the same training samples, both consisting of 3600 positive samples and 2080 negative samples, and the test samples were both 1200 review relationship triplets. On the same batch of test data, the accuracy rate of sensitive word hit of the traditional method was 81.6%, while that of this invention was 92.4%; the recognition rate of implicit violation relationships of the traditional method was 63.8%, while that of this invention was 84.1%; the recognition rate of cross-node risk paths of the traditional method was 58.7%, while that of this invention was 86.3%; the accuracy rate of risk source location of the traditional method was 61.5%, while that of this invention was 88.9%; the traditional method generated 940 records awaiting manual review, while that of this invention generated 262; the total review time of the traditional method was 126 minutes, while that of this invention was 54 minutes; the amount of original data transmitted outside the traditional method was 428.6 GB, while that of this invention was 0 GB. It can be seen that this invention can improve local matching efficiency, implicit risk identification capability, cross-node propagation analysis capability, and risk source location accuracy without the original data leaving the domain.
[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A distributed data compliance review method based on data security, characterized in that, include: Data compliance review elements are collected from each distributed data node, and these elements are preprocessed to generate the data objects to be reviewed and node review summaries. Based on the node review summary, the cross-field co-occurrence frequency is discretized, and a nonlinear interactive accumulation operation is performed to generate a rule trigger intensity sequence; The algorithm of double array trie is used to perform local matching of the data objects to be reviewed, the state is filtered by ternary interval conversion, and the segment order addressing is performed by segment jump distance rearrangement to obtain the term hit result; Align the node review summary with the term hit results by field behavior trajectory, perform item-by-item deviation decomposition and local conflict expansion, and generate a review risk distribution sequence through cross-node propagation path calibration and path stability recursive analysis. The review risk distribution sequence is processed by ternary extraction to generate review relationship triples. An improved TransE model is constructed, and relationship reasoning is performed on the review relationship triples. Error backfilling is performed based on component residual recursive correction. Vector internal difference perturbation is introduced to process adjacent term differences. Component staggered projection is used to arrange components in an alternating manner to obtain the relationship discrimination result. Based on the term hit results and relationship discrimination results, a hierarchical structure verification is performed. The rule trigger intensity sequence is used as the hierarchical recursive input. Contribution accumulation path analysis and risk source attribution ranking are performed to generate risk source location results. The risk distribution sequence and risk source location results are hierarchically compressed and encoded, and cross-node consistency verification and redundancy resolution are performed on the structured review summary to generate a distributed data compliance review report.
2. The distributed data compliance review method based on data security according to claim 1, characterized in that, The data compliance review elements include data objects, field attributes, access subjects, operation behaviors, permission status, flow paths, and audit records.
3. The distributed data compliance review method based on data security according to claim 1, characterized in that, The preprocessing of data compliance review elements includes data cleaning, format standardization, local numbering, field classification, missing data completion, duplicate removal, and counting statistics.
4. The distributed data compliance review method based on data security according to claim 1, characterized in that, The generation of the data object to be reviewed and the node review summary includes: Establish local collection channels on each distributed data node, and read the original review entries from the node resource directory, interface call chain, permission credential index, data flow credential and local audit index according to the collection cycle. Configure the node serial number, source location number and time slice number for each original review entry. The original review items are processed by time slice merging, field path concatenation, subject identity desensitization, flow link compression, item hash digest generation and abnormal format removal, and the processed original review items are converted into data objects to be reviewed with a unified index structure. The node review summary is generated by statistically analyzing the object size, call frequency, permission change frequency, cross-domain transfer frequency, and audit retention frequency of the data objects to be reviewed based on the node serial number. The statistical results are then combined with the source location number, time slice number, and item hash digest to generate the node review summary.
5. The distributed data compliance review method based on data security according to claim 1, characterized in that, The generation rule triggers an intensity sequence, including: Read the node review summary according to the node identifier and time slice number, and arrange the object size, call frequency, number of permission changes, number of cross-domain transfers and number of audit retentions within the same time slice in order to generate a field behavior sequence; The co-occurrence frequency of the field behavior sequence is discretized across fields. The number of times any two types of fields appear simultaneously in the current time slice is taken as the co-occurrence count. The co-occurrence count is divided by the sum of the occurrence counts of all fields in the current time slice to obtain the co-occurrence frequency. The co-occurrence frequencies are divided into discrete levels in ascending order and the first structure tensor is generated. Arrange the access subject dimension, data object dimension, and operation behavior dimension in the first structural tensor into a three-axis sequence to obtain the structural tensor transition sequence. Subtract the structural tensor transition sequences of adjacent time slices one by one to obtain the time series fluctuation term. A nonlinear interactive accumulation operation is performed on the structure tensor transition sequence. The structure tensor transition value of the current time slice is multiplied by the structure tensor transition value of the previous time slice, and then the absolute value of the time series fluctuation term is added to obtain the rule trigger intensity value of the current time slice. The rule trigger intensity values are arranged in order according to the time slice number to generate the rule trigger intensity sequence.
6. The distributed data compliance review method based on data security according to claim 1, characterized in that, The obtained term matching results include: Based on the rules, a local search term is generated by triggering the intensity sequence. A double-array trie structure containing a base address array, a check array, and a termination marker array is constructed based on the local search term. Read the data object to be reviewed, convert the current matching status, current character encoding, and the associated position of the check array into the base address interval, character interval, and check interval, respectively, and perform ternary interval conversion on the base address interval, character interval, and check interval to obtain the status filtering result; The long-field text and log text in the data object to be reviewed are divided into character segments. A jump distance sequence is generated based on the first and last character segment encoding and the difference in the encoding of adjacent characters within the segment. The segment jump distance rearrangement is used to determine the access position within the segment and the starting state of the next character segment, thus obtaining the segment addressing result. The state filtering results and segment addressing results are used to confirm the hit. For the position where the term termination mark exists in the termination mark array, the hit term number, hit object number, hit field position and hit time segment number are recorded to generate the term hit result.
7. The distributed data compliance review method based on data security according to claim 1, characterized in that, The generation of the review risk distribution sequence includes: Read the node review summary and term matching results, align the field behavior trajectory according to the node identifier, time slice number, object number and field position, and generate a multi-source review result sequence; The multi-source review result sequence is decomposed item by item to obtain the call deviation by subtracting the average of the current call frequency from the average of the historical call frequency; the permission deviation by subtracting the average of the current number of permission changes from the average of the historical number of permission changes; and the flow deviation by subtracting the average of the current number of cross-domain transfers from the average of the historical number of cross-domain transfers. The review conflict chain set is formed according to the time slice order. The review conflict chain set is expanded locally, and conflict chains with different access subjects, different permission states or different flow paths under the same time slice and the same object are expanded adjacently, and cross-node propagation paths are generated according to node identifiers and flow paths; The path is labeled and stability is recursively analyzed for cross-node propagation paths. The path stability is obtained by dividing the number of times the same path appears in consecutive time slices by the total number of time slices to which the path belongs. The path stability is multiplied by the number of corresponding conflict chains to obtain the path risk distribution value. The review risk distribution sequence is generated according to the node identifier, time slice number and path order.
8. The distributed data compliance review method based on data security according to claim 1, characterized in that, The obtained relationship discrimination result includes: The review risk distribution sequence is processed by ternary extraction to generate review relation triples. After the translation operation of the original TransE model, component residual recursive correction is set. Before the translation operation, vector internal difference perturbation is set. Before relation discrimination, component staggered projection is set to construct an improved TransE model. Read the head entity, relation, and tail entity from the examine relation triplet, and look up the corresponding vector representations in the entity vector table and relation vector table respectively to obtain the head entity vector, relation vector, and tail entity vector. Add the head entity vector and relation vector to obtain the predicted tail entity vector. Perform component-wise difference processing on the tail entity vector and the predicted tail entity vector to obtain the component residuals. Fill the component residuals back into the head entity vector, relation vector, and predicted tail entity vector to obtain the corrected predicted tail entity vector. Introducing vector internal difference perturbation for adjacent component difference processing, the head entity vector, relation vector, and tail entity vector are processed for adjacent component difference processing respectively to obtain head entity difference sequence, relation difference sequence, and tail entity difference sequence. A difference deviation sequence is generated based on the head entity difference sequence, relation difference sequence, and tail entity difference sequence. The component interleaving projection is used to perform component interleaving arrangement. The head entity vector, relation vector and tail entity vector are interleaved according to the main component, object component, permission component, behavior component, rule component and risk component. The interleaved head entity vector, relation vector and tail entity vector are then processed by component difference to obtain the interleaved deviation sequence. Based on the corrected predicted tail entity vector, the difference deviation sequence, and the staggered deviation sequence, a relation discrimination value is generated. The relation discrimination value is then associated with the examined relation triplet to generate a relation discrimination result. Using real review relation triples as positive samples and triples obtained by replacing any entity or relation in the review relation triples as negative samples, calculate the relationship discriminant values for positive and negative samples respectively. Update and improve the TransE model according to the training direction where the relationship discriminant value of positive samples is less than that of negative samples, and obtain the improved TransE model after training.
9. The distributed data compliance review method based on data security according to claim 1, characterized in that, The generated risk source location results include: Read the term hit results, relationship discrimination results, and rule trigger intensity sequence, and perform hierarchical merging according to node identifier, time slice number, object number, and field position to generate a hierarchical verification sequence; The hierarchical verification sequence is structurally verified. The absolute value of the difference between the number of term hits and the number of relation discriminations is obtained. The hit difference, relation discrimination, and rule trigger intensity of the corresponding time slice are added together to obtain the risk contribution value of the current level. The risk contribution value of the previous level is added together with the risk contribution value of the current level to obtain the recursive risk contribution value. The risk contribution values are concatenated in the order of field position, operation behavior, permission status and flow path to generate a contribution accumulation path. The total path contribution value is obtained by adding up the recursive risk contribution values within the same path. The risk source is then sorted from largest to smallest according to the total path contribution value to generate the risk source location result.
10. A distributed data compliance review method based on data security according to claim 1, characterized in that, The generation of the distributed data compliance review master report includes: Read the risk distribution sequence and risk source location results, and trim the summary field according to the node identifier, time slice number, risk source, path risk distribution value, path contribution total value and risk source sorting result. Convert the trimmed summary field into a fixed-length coded segment to generate a structured review summary. Cross-node consistency verification is performed on structured review summaries. Structured review summaries with the same risk source location, the same path identifier, and the same time segment number are compared, and consistent summaries and summaries to be reviewed are generated according to the difference in risk level and the difference in risk type. Redundancy is eliminated in the consistent summary and the summary to be reviewed. Summaries with the same node identifier, the same time slot number, the same risk type, and the same risk source location are merged into report records. The report records are arranged hierarchically according to node identifier, risk type, risk level, risk source location, and time slot number to generate a distributed data compliance review master report.
