Control system for data sensitivity levels based on dynamic optimization

Abstract

The application relates to the technical field of data security control, and particularly discloses a control system for data sensitivity levels based on dynamic optimization, which comprises an original data stream acquisition module, a sensitive scoring and semantic understanding module, a semantic drift quantitative evaluation module, a scoring enhancement and mapping module and a permission matching module. The original data stream acquisition module is used for acquiring original data request packets, static rule dictionaries and external event streams. The sensitive scoring and semantic understanding module is used for matching and scoring to obtain a basic sensitivity score, and performing feature extraction and vectorization conversion to obtain a data content vector and a global event context vector. The semantic drift quantitative evaluation module is used for performing semantic drift quantitative evaluation to obtain a semantic drift coefficient. The scoring enhancement and mapping module is used for performing dynamic enhancement calculation to obtain an enhanced sensitivity score, and performing boundary truncation and discretization mapping to obtain a dynamic sensitivity level. The permission matching module is used for performing permission matching between the dynamic sensitivity level and a user attribute archive to obtain an access control decision result. The application realizes millisecond-level adaptive reconstruction and accurate permission ruling of the sensitivity level.

Description

Technical Field

[0001] This invention relates to the field of data security control technology, and more specifically, to a control system based on dynamically optimized data sensitivity levels. Background Technology

[0002] With the advent of the big data era, massive and heterogeneous data continuously generates value in its flow. Traditional security models centered on perimeter protection are no longer sufficient to cope with complex data security challenges. Therefore, building a fine-grained internal access control mechanism has become crucial for ensuring data privacy. Because the sensitivity of data varies significantly across different times, scenarios, and business contexts, static access control models cannot effectively perceive the evolution of the context and the inherent risks of the data itself. There is an urgent need to construct a data sensitivity level control scheme based on dynamic optimization to achieve accurate capture and adaptive management of changes in data value and risk.

[0003] Existing data sensitivity assessments largely rely on static classification and grading standards or fixed keyword libraries, assigning static labels upon data entry. This approach leads to rigid permission decision-making logic, making it difficult to adapt to real-time fluctuations in sensitivity throughout the data lifecycle. The core technical bottleneck lies in the failure of semantic drift detection. Existing models often separate data content from the external business context, causing the system to fail to perceive the surge in sensitivity of specific words during emergencies. For example, "mask" is normally just an ordinary commodity, but its semantic sensitivity increases dramatically during public health emergencies due to changes in the global context. This decoupling of semantics and context, along with the lack of time-varying feature representation, makes existing technologies prone to lagging or inaccurate sensitivity assessments when facing adversarial attacks or complex business environments.

[0004] Therefore, a control scheme based on dynamically optimized data sensitivity levels is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a control system based on dynamically optimized data sensitivity levels, which is achieved through the following technical solution: A control system based on dynamically optimized data sensitivity levels, comprising: The raw data stream acquisition module is used to acquire raw data request packets containing target data content and system time, a preset static rule dictionary, and external event streams within the current time window; The sensitivity scoring and semantic understanding module is used to match and score the target data content based on a static rule dictionary to obtain a basic sensitivity score, and to use a pre-trained word embedding model to extract features and vectorize the target data content and external event stream to obtain data content vectors and global event context vectors respectively. The semantic drift quantization evaluation module is used to perform semantic drift quantization evaluation on the data content vector and the global event context vector based on the semantic mapping matrix to obtain the semantic drift coefficient; The scoring enhancement and mapping module is used to dynamically enhance the sensitivity score by adding the semantic drift coefficient as an increment to the basic sensitivity score based on a preset drift gain factor. The enhanced sensitivity score is then truncated and discretized to obtain the dynamic sensitivity level for the final decision. The permission matching module is used to match dynamic sensitivity levels with user attribute profiles to obtain corresponding access control decision results.

[0006] Preferably, the sensitivity scoring and semantic understanding module includes: The basic sensitivity score determination unit is used to traverse the static rule dictionary to match sensitive keywords in the target data content and determine the basic sensitivity score based on the matching frequency. The word embedding processing unit is used to perform stop word removal on the target data content and input the processed text sequence into the pre-trained word embedding model for feature mapping to obtain the data content vector; The business context hotspot representation unit is used to aggregate hotspots in external event streams using the word frequency-inverse document frequency algorithm to extract a set of high-weight keywords, and then converts the set of high-weight keywords into a global event context vector that can represent the current business context hotspots through a pre-trained word embedding model.

[0007] Preferably, the basic sensitivity score determination unit is used to determine the basic sensitivity score using the following formula: ; in, Basic sensitivity score, For the first The preset weight coefficient for each sensitive keyword. To indicate the first The frequency of occurrence of each sensitive keyword in the target data content This is the default base security score. This represents the maximum allowed static score by the system. To find the minimum value function, N This represents the total number of sensitive keywords matched in the target data content.

[0008] Preferably, the semantic drift quantification evaluation module includes: The spatial alignment transformation unit is used to perform spatial alignment transformation on the global event context vector based on the semantic mapping matrix to obtain the transformed semantic context vector. The correlation estimation unit is used to estimate the correlation between the transformed semantic context vector and the data content vector to obtain the original drift correlation score. The activation unit is used to perform nonlinear activation and coefficient normalization on the original drift correlation score to obtain the semantic drift coefficient.

[0009] Preferably, the scoring enhancement and mapping module is used for: The dynamic increment ratio is obtained by multiplying the semantic drift coefficient by the drift gain factor, and then the dynamic increment ratio is superimposed on the basic sensitivity score through a multiplicative model to obtain the enhanced sensitivity score. Based on the preset global maximum score limit, the enhanced sensitivity score is numerically saturated and boundary-limited to obtain the final continuous score; The final continuous score is mapped to a predefined discrete security hierarchy using a piecewise ladder function to obtain the dynamic sensitivity level.

[0010] Preferably, the permission matching module is used for: Quantitatively assess user security capabilities by analyzing the feature values ​​in the user attribute profile to obtain a user capability score; Using the dynamic sensitivity level as the index key, retrieve the corresponding minimum capability requirement threshold in the policy threshold matrix, and calculate the difference between the user's capability score and the minimum capability requirement threshold to obtain the permission gap value. Based on the permission gap value, a preset hierarchical decision logic is matched to obtain access control decision results that include full permission, downgrade and desensitization, or blocking alarm instructions.

[0011] Preferably, the correlation estimation unit is used for: Obtain a pre-defined normal semantic basis matrix containing multiple orthogonal basis vectors; The transformed semantic context vector is projected onto the safe subspace constructed by the normal semantic basis matrix to obtain a pure risk context vector; Calculate the noise entanglement penalty coefficient of the data content vector on the normal semantic basis matrix; Based on the noise entanglement penalty coefficient, a penalty-based comparative drift score is performed on the data content vector and the pure risk context vector to obtain the original drift correlation score.

[0012] Compared with existing technologies, this invention provides a data sensitivity level control system based on dynamic optimization. While establishing a basic sensitivity score using a static rule dictionary, it constructs a two-dimensional vector representation of data and external events through a pre-trained model and introduces a semantic mapping matrix to calculate the semantic drift coefficient of data content relative to global hot events, quantifying the nonlinear driving force of environmental dynamics on data risk. Furthermore, to address background noise interference, the system constructs a security subspace using orthogonal basis matrices. By orthogonally decomposing semantic vectors and calculating noise entanglement penalties, it removes normal industry background features, retaining only pure risk semantics for comparative scoring. Finally, the denoised drift coefficient is dynamically superimposed onto the basic sensitivity score, achieving millisecond-level adaptive reconstruction of sensitivity levels and precise permission adjudication. Attached Figure Description

[0013] The above and other objects, features, and advantages of the present invention will become more apparent from the more detailed description of the embodiments of the invention in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same parts or steps.

[0014] Figure 1 This is a system block diagram of a control system based on dynamically optimized data sensitivity levels, provided for an embodiment of the present invention.

[0015] Figure 2 A flowchart of a control system based on dynamically optimized data sensitivity levels provided for embodiments of the present invention.

[0016] Figure 3 A block diagram of the sensitivity scoring and semantic understanding module in a control system based on dynamically optimized data sensitivity levels, provided in an embodiment of the present invention.

[0017] Figure 4 This is a block diagram of the semantic drift quantization evaluation module in a control system based on dynamically optimized data sensitivity levels, provided in an embodiment of the present invention. Detailed Implementation

[0018] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0019] To address the technical bottlenecks in existing technologies, such as the inability of static rules to perceive real-time semantic evolution and the inability of systems to distinguish between normal contexts and real risks due to overlapping common industry terms, leading to high false alarm rates, this invention proposes a data sensitivity level control system based on dynamic optimization. The system first represents the original data stream and external event hotspots using a two-dimensional vectorization method to establish a basic sensitivity score. Then, to eliminate interference from high-frequency non-sensitive words common in industries such as healthcare and finance, the system introduces a pre-set normal semantic basis matrix. Through orthogonal projection technology, background components are subtracted from the complex context, extracting a pure risk context vector unaffected by industry-wide semantic interference. Simultaneously, this invention sets a noise entanglement penalty by quantifying the projection modulus of data content within the security subspace, thereby nonlinearly suppressing data containing only mediocre words. Based on this, the semantic drift coefficient calculated using the penalty mechanism dynamically enhances the basic sensitivity score. After boundary truncation and piecewise ladder mapping, a final discrete dynamic sensitivity level is generated, driving the permission decision point to accurately adjudicate user access behavior. This achieves real-time level adjustment with zero human intervention while significantly improving the robustness of defense in complex contexts.

[0020] Figure 1 This is a system block diagram of a control system based on dynamically optimized data sensitivity levels, provided for an embodiment of the present invention. Figure 2 A flowchart of a data sensitivity level-based control system based on dynamic optimization provided in an embodiment of the present invention. Figure 1 and Figure 2 As shown, the data sensitivity level control system 100 based on dynamic optimization provided in this embodiment of the invention includes: a raw data stream acquisition module 110, used to acquire raw data request packets containing target data content and system time, a preset static rule dictionary, and external event streams within the current time window; a sensitivity scoring and semantic understanding module 120, used to match and score the target data content based on the static rule dictionary to obtain a basic sensitivity score, and to use a pre-trained word embedding model to extract features and vectorize the target data content and external event streams to obtain data content vectors and global event context vectors; a semantic drift quantization evaluation module 130, used to perform semantic drift quantization evaluation on the data content vectors and global event context vectors based on a semantic mapping matrix to obtain semantic drift coefficients; a score enhancement and mapping module 140, used to dynamically enhance the basic sensitivity score by adding the semantic drift coefficients as increments based on a preset drift gain factor, and to perform boundary truncation and discretization mapping on the enhanced score to obtain a dynamic sensitivity level for final decision-making; and a permission matching module 150, used to match the dynamic sensitivity level with user attribute profiles to obtain corresponding access control decision results.

[0021] It is understandable that during the flow and sharing of big data, the sensitivity of data is not constant but exhibits significant dynamic fluctuations due to changes in external social hotspots, industry emergencies, and business evolution cycles. Relying solely on a fixed sensitive word database for identification will not effectively address compliance risks caused by contextual drift. Therefore, in the technical solution of this invention, a multi-dimensional data perception foundation is constructed by acquiring the original data request packet containing the target data content and system time, a pre-set static rule dictionary, and the external event stream within the current time window. This foundation integrates static compliance benchmarks, time-series features, and macro-level business background. This provides subsequent stages with feature inputs possessing semantic depth and time-varying characteristics, ensuring the system's ability to perceive surges in data risks under specific circumstances.

[0022] Specifically, in one example of the present invention, the system captures data through a real-time monitoring module deployed at the application layer gateway or data exchange node. The raw data stream acquisition module 110 first uses deep packet inspection technology to intercept the currently flowing network access request, extracts the target data content from the packet's payload area using a protocol decoding algorithm, and synchronously reads the clock offset field in the packet header to determine the system time, then encapsulates it into a raw data request packet in a unified format. Subsequently, the raw data stream acquisition module 110 retrieves a static rule dictionary maintained by domain experts from persistent storage media through a pre-established database connection pool. This static rule dictionary contains various sensitive feature items defined by laws and regulations and their initial risk weights. Finally, the raw data stream acquisition module 110 subscribes to topic channels in a distributed message middleware to access heterogeneous data streams from external news monitoring platforms or internal enterprise alarm systems in real time. It uses a time window sliding algorithm to intercept external event streams corresponding to the current access time, thereby achieving synchronous collection and integration of information about the data object and its macro-environment.

[0023] It is understandable that semantic inference relying solely on deep learning models may overlook fundamental compliance red lines due to the black-box nature of the models, while static rule matching alone struggles to capture implicit semantic connections in complex contexts, resulting in a disconnect between stability and flexibility in sensitivity assessment. Therefore, in the technical solution of this invention, the sensitivity scoring and semantic understanding module 120 further matches and scores the target data content based on a static rule dictionary to obtain a basic sensitivity score. A pre-trained word embedding model is then used to extract and vectorize features from the target data content and the external event stream to obtain data content vectors and global event context vectors, thereby constructing a dual-modal feature foundation that combines rigid compliance bottom lines with flexible semantic representation. This ensures that the system can quickly identify explicit sensitive features while providing a high-dimensional mathematical representation for subsequent accurate quantification of implicit semantic drift between data and the external environment.

[0024] Figure 3 This is a block diagram of the sensitivity scoring and semantic understanding module in a control system based on dynamically optimized data sensitivity levels, provided in an embodiment of the present invention. Figure 3 As shown, in an embodiment of the present invention, the sensitivity scoring and semantic understanding module 120 includes: a basic sensitivity scoring determination unit 121, used to traverse a static rule dictionary to match sensitive keywords in the target data content, and determine a basic sensitivity score based on the matching frequency; a word embedding processing unit 122, used to perform stop word removal processing on the target data content, and input the processed text sequence into a pre-trained word embedding model for feature mapping to obtain a data content vector; and a business context hotspot representation unit 123, used to perform hotspot aggregation on external event streams using a word frequency-inverse document frequency algorithm to extract a set of high-weight keywords, and convert the set of high-weight keywords into a global event context vector that can represent the current business context hotspot through a pre-trained word embedding model.

[0025] It is understandable that in big data access control scenarios, certain types of sensitive information, such as personally identifiable information, financial transaction records, or confidential business terminology, have a constant baseline risk in terms of legal compliance and corporate security strategies. If dynamic semantic inference is relied upon without strict constraints on explicit sensitive features, the system's ability to identify known risky terms will decline. Therefore, in the technical solution of this invention, the basic sensitivity scoring determination unit 121 traverses the static rule dictionary to match sensitive keywords in the target data content and determines the basic sensitivity score based on the matching frequency. This constructs a compliance baseline for data security management, establishing an inherent risk quantification value unaffected by environmental fluctuations. This ensures that each original data request packet undergoes deterministic compliance review before entering the complex dynamic evolution assessment, providing a reliable calculation basis for the subsequent superposition of semantic drift coefficients.

[0026] More specifically, in a specific example of the present invention, a pre-set static rule dictionary is first loaded from non-volatile memory. This static rule dictionary stores a large number of sensitive keywords identified by domain experts and their corresponding preset weight coefficients through a hash mapping table structure. After receiving the original data request packet, the system uses a word segmentation engine to perform a full-text scan of the target data content and employs a multi-pattern matching algorithm to compare each text unit with the entries in the static rule dictionary. When a match is successful, the system not only identifies the type of the sensitive keyword but also simultaneously records its matching frequency in the target data content. By correlating the preset weight coefficients of each sensitive keyword with its occurrence probability, a basic sensitivity score is determined using the following formula: ; in, Basic sensitivity scoring is used to quantify the static, inherent risk of the target data content. For the first The preset weight coefficient for each sensitive keyword reflects the importance of that keyword in the static compliance system. To indicate the first The frequency of occurrence of each sensitive keyword in the target data content This is the default base security score, serving as the initial security floor of the system. This is the maximum allowed static score value by the system, used to prevent score values ​​from overflowing. N This represents the total number of sensitive keywords matched in the target data content. This is a function that takes the minimum value. In this mathematical model, the cumulative term in the formula... It implements linear integration of explicit risk elements in the target data content. By multiplying preset weight coefficients with matching frequencies, it can enable data requests containing high-value sensitive words or high-frequency sensitive features to obtain higher initial scores. (Introduction) The aim is to provide a unified risk assessment benchmark for all data flowing through the system, ensuring that the system can still output a controllable basic sensitivity score even when no static rules are hit. The formula as a whole adopts... Function truncation mechanisms have key engineering significance, as they achieve this through... Numerical saturation processing of the calculation results can effectively suppress abnormal surges in scores caused by excessively long target data content or malicious keyword stuffing, thereby limiting the basic sensitivity score within the preset legal range and reserving the necessary numerical space for subsequent dynamic enhancement calculations based on semantic drift coefficients.

[0027] Specifically, since the target data content in the original data packet usually contains a large number of stop words such as "de", "le", "zai" that lack core semantic information, these words not only occupy computing resources, but also dilute the weights of key sensitive features in the vector space, resulting in deviations in subsequent semantic drift evaluation. Therefore, in the technical solution of the present invention, the stop word removal process is further performed on the target data content by the word embedding processing unit 122, and the processed text sequence is input into the pre-trained word embedding model for feature mapping to obtain the data content vector, so as to filter out the semantic noise in the text and extract high-purity deep semantic features. In this way, the unstructured text can be converted into a dense vector representation with mathematical operation characteristics, laying an accurate feature foundation for quantifying the semantic association between the data content and external hot events in the subsequent steps.

[0028] More specifically, in a specific example of the present invention, the word embedding processing unit 122 first receives the target data content transmitted through from the basic sensitivity score determination unit, and uses the word segmentation engine to segment it into a discrete sequence of word items. Subsequently, the system loads the general stop word library stored locally, and eliminates the meaningless auxiliary words that match successfully in the sequence through the string comparison algorithm to generate the cleaned text sequence. Then, the word embedding processing unit 122 starts the word embedding model pre-loaded and trained with a large-scale corpus. The word embedding model can be a lightweight bidirectional encoder representation model or a distributed word vector model. By using the cleaned text sequence as the input for multi-layer mapping, the discrete symbols in the text are converted into high-dimensional real-space coordinates with the ability to express semantic similarity, and finally the data content vector is output. Through the parameter weights trained inside the word embedding model, the business core words (such as masks, shortages, control) contained in the text sequence are mapped into vector coordinates with specific directivity; the generated data content vector completely records the semantic component distribution of the target data content in the multi-dimensional space, thus eliminating the semantic gap brought by different expression methods. Through this feature mapping, the system realizes the deep understanding of business data at the context level, ensuring that the subsequent steps can accurately identify the degree to which the data content vector deviates from the normal semantic space under the drive of specific external events through vector dot product or projection operations.

[0029] It is understandable that, given the massive and unstructured nature of external event streams in a big data environment, directly introducing the raw text stream into the evaluation model would result in excessively high computational dimensionality and difficulty in focusing on the core business risks. This would prevent the system from perceiving the real-time driving effect of events such as sudden public health emergencies or drastic changes in industry policies on data sensitivity. Therefore, in the technical solution of this invention, the business context hotspot representation unit 123 further utilizes the word frequency-inverse document frequency algorithm to aggregate hotspots in the external event stream to extract a set of high-weight keywords. A pre-trained word embedding model then converts this set of high-weight keywords into a global event context vector that can represent current business context hotspots, thereby constructing a semantic benchmark that can reflect the macro-environmental situation in real time. This effectively removes background noise from external information, mapping scattered social hotspots into high-dimensional feature vectors with mathematical computational properties, thus providing accurate input support for subsequently quantifying the semantic correlation between specific data content and global hotspot events.

[0030] More specifically, in a specific example of the present invention, the business context hotspot representation unit 123 first receives the external event stream within the current time window in real time through a streaming data interface, and preprocesses the text content using a word segmentation engine to remove common auxiliary words and conjunctions. The system then initiates a word frequency-inverse document frequency (IF-IDF) algorithm to calculate the word frequency of each term in the current event stream and its inverse document frequency in the historical corpus, evaluating the representative score of the term within the time window through their product. Based on this, the system sorts the terms from highest to lowest score and selects the top K core terms to form a high-weight keyword set. Next, the system loads a word embedding model pre-trained on a large-scale industry corpus, mapping each keyword in the high-weight keyword set to a dense low-dimensional real vector. To obtain a comprehensive representation that balances the weights of each hotspot, the system uses the TF-IDF score corresponding to each keyword as a weight coefficient, performs weighted summation and normalization on all word vectors, and finally generates a global event context vector. The weighted aggregation calculation formula executed by the system is as follows: ; in, Represents the global event context vector, serving as the center coordinates of the current macro-level business context in the vector space; This represents the total number of keywords contained in the set of high-weight keywords extracted from the external event stream; Indicates the first The term frequency-inverse document frequency weight value of each keyword in the high-weight keyword set is used to represent the intensity of the term's popularity in the current context; This represents the mapping function of the pre-trained word embedding model; Indicates the first The text content of each keyword. This formula is obtained through... The system performs vector superposition of multiple risk signals from the external environment (such as high-weight terms like shortage, control, and blockade), causing the generated global event context vector to shift in the multi-dimensional semantic space towards the most sensitive risk; while the denominator term... This serves as a numerical normalization mechanism, ensuring that the magnitude of the global event context vector does not abnormally increase with fluctuations in external information, thereby maintaining the determinism and robustness of the system's evaluation of semantic drift coefficients across different time windows. In this way, the system can transform abstract changes in the social environment into standardized mathematical features, thus accurately capturing the surge in sensitivity of ordinary supplies such as masks due to changes in the global context.

[0031] It is understandable that in big data dynamic access control scenarios, the sensitivity of data does not depend in isolation on its static content, but is heavily influenced by the macro-business context and sudden external events. For example, medical supplies such as masks are considered low-sensitivity general supplies in normal circumstances, but their sensitivity can drift non-linearly in the context of public health emergencies due to the surge in social relevance. If the system only recognizes the semantics of the data itself, it will fail to capture the surge in risk under specific circumstances, leading to delays or inaccuracies in access control decisions. Therefore, in the technical solution of this invention, the semantic drift quantification assessment module 130 further performs semantic drift quantification assessment on the data content vector and the global event context vector based on the semantic mapping matrix to obtain the semantic drift coefficient, thereby accurately measuring the risk correlation strength of the target data content relative to the current hot events.

[0032] Figure 4 This is a block diagram of a semantic drift quantization evaluation module in a control system based on dynamically optimized data sensitivity levels, provided in an embodiment of the present invention. Figure 4 As shown, in an embodiment of the present invention, the semantic drift quantization evaluation module 130 includes: a spatial alignment transformation unit 131, used to perform spatial alignment transformation on the global event context vector based on the semantic mapping matrix to obtain the transformed semantic context vector; a correlation estimation unit 132, used to perform correlation estimation on the transformed semantic context vector and the data content vector to obtain the original drift correlation score; and an activation unit 133, used to perform nonlinear activation and coefficient normalization on the original drift correlation score to obtain the semantic drift coefficient.

[0033] Specifically, since external event streams typically originate from publicly available news or internet public opinion, while target data content originates from an enterprise's internal business database, although both are processed by the same word embedding model, due to the heterogeneity of their context sources, their vector representations may exhibit slight misalignments in reference coordinates or differences in the distribution of semantic features in multidimensional space. Directly performing dot product operations makes it difficult to accurately capture the true correlation between the two in a specific risk dimension. Therefore, in the technical solution of this invention, a spatial alignment transformation unit 131 further performs a spatial alignment transformation on the global event context vector based on the semantic mapping matrix to obtain a transformed semantic context vector. This performs linear calibration and feature remapping of the vector space, eliminating spatial heterogeneity between heterogeneous data sources. In this way, macroscopic external hotspot features can be accurately projected onto a semantic coordinate system consistent with the internal data, ensuring that subsequent drift assessment is based on a unified and aligned geometric benchmark, thereby significantly improving the mathematical accuracy and business relevance of semantic drift coefficient calculation.

[0034] More specifically, in a specific example of the present invention, the spatial alignment transformation unit 131 first loads a preset semantic mapping matrix from the system's model parameter library. This semantic mapping matrix is ​​a square matrix with learnable parameters, and its dimension is consistent with the vector dimension output by the word embedding model. After receiving the global event context vector output by the business context hotspot representation unit 123, the system performs matrix multiplication using a high-performance linear algebra computation library. This operation essentially rotates, scales, or shears the global event context vector in the feature space, aligning its feature distribution with the vector space distribution of the target data content. Through this mathematical transformation, the system can dynamically adjust the weights of external event features in various dimensions, strengthening semantic components highly relevant to the current business risk while suppressing irrelevant background noise components, ultimately generating a transformed semantic context vector calibrated in spatial geometry, providing standardized input for subsequent accurate correlation estimation with the data content vector. During the execution of the spatial alignment transformation unit 131, the system uses the following linear transformation formula for calculation: ; in, This represents the transformed semantic context vector, which is a spatially calibrated representation of external environment features. This represents a pre-defined semantic mapping matrix, which serves as a linear transformation operator connecting the external event space and the internal data space. The input global event context vector represents the original external hotspot semantics; This represents matrix multiplication. The formula has a clear spatial geometric meaning in practical business scenarios: through... In essence, the system performs a remapping of the feature space. Assume... It includes features from two dimensions: "epidemic" and "logistics," while internal data focuses more on the specific risk of material shortages. It can use its internal parameter weights to... The vector direction is rotated and aligned towards the core risk axis of resource shortage. The generated... It is no longer just the raw mathematical expression of external news, but a feature vector that has been filtered and adapted by business logic. This allows the calculation of the angle between this vector and the data content vector (i.e., correlation) in subsequent steps to truly reflect the degree to which external events drive the sensitivity of internal data, rather than just reflecting the literal co-occurrence of words.

[0035] It is understandable that, after spatial alignment transformation, although the data content vector representing internal data semantics and the transformed semantic context vector representing external macro-risks are mapped to a unified feature space, the specific semantic correlation between them has not been quantified digitally. This makes it impossible for the system to directly determine whether the data content truly hits the current risk hotspot. Therefore, in the technical solution of this invention, the correlation estimation unit 132 further performs correlation estimation on the transformed semantic context vector and data content vector to obtain the original drift correlation score, thereby measuring the projection strength or geometric approximation degree of the target data content on the current business risk dimension. In this way, abstract semantic relationships can be transformed into calculable scalar values, providing a direct mathematical basis for subsequently generating semantic drift coefficients through nonlinear activation, thereby determining whether the data has experienced sensitivity drift due to changes in the external environment.

[0036] More specifically, in a specific example of the present invention, the relevance estimation unit 132 processes the input feature vector using vector algebra computation logic. The system performs a dot product operation between the data content vector and the transformed semantic context vector. This operation is geometrically equivalent to projecting the data content vector onto the direction of the transformed semantic context vector. The system calculates the cumulative product of the feature components of the two vectors in each semantic dimension. Through this linear calculation method, the system aggregates the matching degree of multiple feature dimensions into a single scalar value, namely the original drift relevance score, which is calculated using the following formula: ; in, This represents the original drift correlation score; Represents a data content vector; This represents the transformed semantic context vector; This represents the vector transpose operation. This formula quantifies the similarity of two vectors in semantic space: when the semantic features of the data content are highly consistent with the feature direction of the external event (for example, both have high weights in dimensions such as shortage or urgency), the result of the dot product operation will produce a large positive value, thus indicating a high-risk correlation; conversely, if the two directions are perpendicular or divergent, the score will approach zero or be negative, indicating that the data content has not been affected by the current external hot events.

[0037] In particular, the bilinear feature projection and correlation calculation steps in the above embodiments have a fundamental logical flaw in handling the sensitive drift determination in the high-dimensional semantic space. This flaw stems from the fact that simple dot product operations cannot orthogonally decouple accompanying noise. In complex business contexts, globally sensitive events (such as sudden public health events) and target business data (such as routine medical supply records) often share a large number of non-sensitive general industry terms, which constitute a safe but high-frequency background noise subspace. For example, when analyzing whether masks should be upgraded to sensitive data, related terms such as hospitals, logistics, and inventory exist not only in the context of sensitive events but also widely in ordinary non-sensitive operation records. The first embodiment directly calculates the collinearity (dot product) of these two vectors, which causes the model to be unable to distinguish whether the correlation comes from the matching of core risk features (such as shortages and controls) or from the overlap of normal background features. This neglect of the normal semantic subspace leads to falsely high drift scores for security data containing a large number of common terms, as the projection of the event vector onto the security dimension overlaps with the data. This ultimately results in severe false alarms when the system faces highly overlapping contexts, failing to accurately capture genuine business semantic drift. Therefore, in the technical solution of this invention, a preferred embodiment is further employed to estimate the correlation between the transformed semantic context vector and the data content vector. This achieves precise decoupling of risk features and background features through orthogonal subspace projection and a noise penalty mechanism. This ensures that the system only generates high-score responses to semantic drifts that truly pose business risks, while maintaining low sensitivity evaluations for ordinary business data containing the same industry terms, significantly improving the robustness and accuracy of dynamic data sensitivity level control.

[0038] More specifically, in another specific example of the present invention, the correlation estimation unit 132 is used to: obtain a preset normal semantic basis matrix containing multiple orthogonal basis vectors; project the transformed semantic context vector onto the safe subspace constructed by the normal semantic basis matrix to obtain a pure risk context vector; calculate the noise entanglement penalty coefficient of the data content vector on the normal semantic basis matrix; and, based on the noise entanglement penalty coefficient, perform a penalty-based comparative drift score on the data content vector and the pure risk context vector to obtain the original drift correlation score.

[0039] It is understandable that, in complex business contexts, globally sensitive events and target business data often share a large number of non-sensitive, general industry terms. These terms constitute a safe but high-frequency background noise subspace. If simple dot product operations are directly used for correlation estimation, the model will be unable to distinguish whether the correlation originates from the matching of core risk features or from the overlap of normal background features. Therefore, in the technical solution of this invention, a pre-set normal semantic basis matrix containing multiple orthogonal basis vectors is further obtained, and the transformed semantic context vector is projected onto the safe subspace constructed by the normal semantic basis matrix to obtain a pure risk context vector. This purifies the semantic expression of external events from the source, and the transformed semantic context vector is spatially decomposed using the Schmitt orthogonalization method. In this way, the general medical industry background in epidemic events can be effectively stripped away, retaining only pure risk semantics such as emergency control, avoiding interference from general background noise in subsequent matching, and significantly reducing the false alarm rate in high-noise environments.

[0040] In a specific example of this invention, a preset normal semantic basis matrix is ​​first read from non-volatile memory. This normal semantic basis matrix is ​​a mathematical space representation constructed by orthogonalizing a set of word vectors covering everyday language, medical industry terminology, and various stop words. Subsequently, the system performs background orthogonal decomposition, calculating the projection component (background component) of the transformed semantic context vector onto the safe subspace spanned by the normal semantic basis matrix to quantify the distribution of neutral words such as hospitals, logistics, and inventory in the external event flow. Finally, the system directionally subtracts this background component from the transformed semantic context vector, eliminating normal semantic features that, while possessing high-frequency statistical characteristics, do not pose immediate security risks, thereby obtaining a pure risk context vector perpendicular to the normal space. The process of obtaining and projecting the pure risk context vector described above can be represented by the following formula: ; in, This represents the pure risk context vector, which characterizes the core risk semantics after removing background noise. This represents the transformed semantic context vector, which contains feature information from the original global event context. This represents a pre-defined normal semantic basis matrix, used to define the feature benchmarks for security contexts. This indicates the calculation of projection coefficients for extracting only background-related features. This formula is executed... The calculated projection coefficients can quantify the semantic intensity in the transformed semantic context vector that highly overlaps with the normal medical background. The right side of the formula... The product of the projection coefficients reconstructs a background component representing industry-wide commonality, and a subtraction operation is used to mathematically remove directed background information related to masks. The resulting pure risk context vector... This ensures that subsequent correlation estimations are no longer affected by overlapping terms such as hospitals and logistics, enabling the system to accurately distinguish between mask inventory registration in ordinary business operations and mask control risks in public health emergencies. This achieves high robustness and high accuracy in dynamic data sensitivity level control.

[0041] Correspondingly, although core risk characteristics have been extracted from the macro context after background separation, the target data content itself may still contain a large amount of industry-standard jargon, everyday greetings, or redundant information lacking actual security risks. This content is considered correct but meaningless. Without quantitative suppression of such mundane features, the system is highly susceptible to misjudging data containing only common vocabulary as highly sensitive, leading to overgeneralization of access control policies. Therefore, in the technical solution of this invention, the data content vector is further projected onto each security basis direction defined by the normal semantic basis matrix, and the square of its projection modulus is calculated to quantify the mundane nature of the target data itself, i.e., to determine whether the data content is mainly composed of common security terms. This allows for the accurate identification of low-value data that, while potentially having a weak random correlation with the risk vector, is essentially composed entirely of security terms such as inventory lists and everyday greetings, providing a deterministic quantitative basis for subsequent suppression of false drift.

[0042] In a specific example of this invention, the calculation of the noise entanglement penalty coefficient is implemented through the following steps. First, the system retrieves a preset normal semantic basis matrix from memory. This normal semantic basis matrix consists of multiple mutually orthogonal basis vectors, representing a protected, risk-free safety subspace in specific industries such as healthcare. Then, the system performs a vector space projection operation, performing a transpose dot product operation between the vectorized data content vector and the normal semantic basis matrix to extract the projection components of the data content on each safety basis. Finally, the system performs an L2 norm square operation on the obtained projection vectors, and by accumulating the energy distributions on each safety dimension, finally generates the noise entanglement penalty coefficient. During the calculation of the noise entanglement penalty coefficient, the system generates the noise entanglement penalty coefficient using the following formula: ; in, This represents the noise entanglement penalty coefficient, used to quantify the component strength of data in the safe subspace. The square operation represents the L2 norm. This represents a data content vector, which characterizes the semantic features of the data to be evaluated. This represents the transpose of a pre-defined normal semantic basis matrix, used to extract the projection of data onto a safe basis. In the formula... The system partially achieves accurate extraction of industry background features from the data being processed. For example, in the context of medical supply distribution, if words like "hospital," "logistics," and "inventory" frequently appear in the target data, the operation will generate a large projection component. By performing the L2 norm square operation as a whole, the system can non-linearly amplify the weights of these normal background signals, resulting in a noise entanglement penalty coefficient. This becomes a benchmark for measuring the mediocrity of data. In subsequent drift assessments, even if the data shows a weak statistical correlation with sudden and sensitive events such as the pandemic due to word overlap, it is still considered a mediocre indicator. Even with extremely high values, the system can still determine that the data is essentially a low-risk mask inventory registration operation, rather than emergency risk information under control. This penalty mechanism based on entanglement quantification ensures that the system can accurately remove general background noise and achieve adaptive and precise control over semantic drift in highly overlapping contexts.

[0043] It is understandable that, in complex business contexts, even after orthogonal decomposition removes industry background noise from the macro-context, the target data may still contain a large number of non-risk characteristic words such as inventory lists and daily greetings. If only simple linear dot product is used for correlation calculation, the accumulation of these mediocre features will still cause a false increase in the correlation score, making it impossible for the system to distinguish whether the correlation stems from the matching of core risk features or from the overlap of normal background features. Therefore, in the technical solution of this invention, a penalty-based comparative drift score is further performed on the data content vector and the pure risk context vector based on the noise entanglement penalty coefficient to obtain the original drift correlation score. This introduces a non-linear signal-to-noise ratio calculation paradigm, mathematically decoupling risk features from background features. In this way, automated contrast enhancement can be achieved, ensuring that positive incentives are generated only when the data features are highly consistent with the pure risk features, while forcibly lowering the data score containing a large number of common words, significantly reducing the false alarm rate in high-noise environments, and achieving high robustness in controlling the sensitivity level of dynamic data.

[0044] In a specific example of the present invention, the correlation estimation unit 132 performs a penalty-based contrast drift scoring process through the following steps: First, the system obtains the data content vector representing the deep semantics of the data, the clean risk context vector after removing normal background noise, and the noise entanglement penalty coefficient quantifying the mediocrity of the target data, all generated in the preceding steps. The system then performs a transpose dot product operation on the data content vector and the clean risk context vector to extract the collinearity strength of the two in the non-background risk dimension. Simultaneously, the system retrieves preset adjustment hyperparameters from the configuration library and performs a nonlinear mapping process with the noise entanglement penalty coefficient to construct the denominator of the suppression term. Finally, the system generates an original drift correlation score that accurately reflects the degree of business risk shift by calculating the ratio of the numerator to the denominator of the suppression term. This penalty-based contrast drift scoring process can be represented by the following formula: ; in, This represents the original drift correlation score, which serves as the dynamic incremental basis for indicating data sensitivity. This represents a data content vector that reflects the semantic distribution of the target data. This represents a pure risk context vector, characterizing the core risk semantics of the current external event after removing the general background. This represents the denoised dot product between the two, used to measure the correlation between data features and pure risk features; The noise entanglement penalty coefficient represents the intensity of the trivial component of the data in the safe subspace, and is used to nonlinearly map the noise penalty onto the denominator to adjust the suppression strength. This represents the system's preset hyperparameters, used to adjust the suppression strength of the noise penalty in the denominator. (Numerator term in the formula) This ensures that scores are only positively incentivized when the target data content truly touches upon pure risk semantics such as emergency control or material shortages. For example, in a public health event scenario, data containing emergency control instructions would receive a high score in this category. Meanwhile, the denominator of the formula... It acts as a sensitivity adjuster, adjusting the noise entanglement penalty coefficient when the target data contains a large amount of medical industry jargon or correct but meaningless information such as inventory records. This will rapidly increase and drive denominator expansion, thereby forcibly lowering the final original drift relevance score. This design ensures that the system can accurately capture semantic drifts that truly pose business risks, while maintaining a low sensitivity evaluation for ordinary business data that contains the same industry terms but is inherently safe, thus avoiding a surge in compliance costs due to rule generalization.

[0045] The preferred embodiments described above aim to address the sensitivity misjudgment problem caused by normal semantic overlap under traditional linear projection perspectives through a denoising mechanism based on orthogonal subspace projection. Through the combined effect of orthogonal decomposition and entanglement penalty, the system achieves precise mathematical decoupling of sudden risk characteristics and normal background characteristics, significantly reducing the false alarm rate in high-noise environments. The final result is that the system can accurately capture semantic drifts that truly pose business risks (such as mask shortages), while maintaining low sensitivity evaluation for ordinary business data containing the same industry terms (such as mask inventory registration). This achieves high robustness and high accuracy in dynamic data sensitivity level control, avoiding a surge in compliance costs or business disruptions caused by rule generalization.

[0046] It is understandable that, since the original drift correlation score is the initial association strength calculated in a multi-dimensional space by the data content vector and the transformed semantic context vector, its value range is within the entire real number space, including negative values ​​representing semantic irrelevance or negative correlation, as well as excessively large positive values ​​representing extreme risk association. If it is directly applied to sensitivity enhancement calculation, it will lead to mathematical instability or numerical overflow in the subsequent level assessment logic. Therefore, in the technical solution of this invention, the original drift correlation score is further nonlinearly activated and its coefficients are normalized by activation unit 133, thereby filtering out the interference of negative correlation and locking the range of positive gain through a nonlinear mapping mechanism. In this way, it can be ensured that the semantic drift coefficient is always within the predefined standard ratio range, realizing the accurate quantification of the dynamic increment of data sensitivity, while effectively suppressing the frequent jumps in sensitivity level caused by extreme fluctuations in the original score.

[0047] More specifically, in a specific example of the present invention, activation unit 133 first receives the original drift correlation score output by correlation estimation unit 132. The system loads a preset normalized function model and performs feature transformation on the original drift correlation score through deterministic mathematical activation logic. In specific implementation, the system inputs the original drift correlation score into the Sigmoid activation function, which uses the exponential operation property to map input values ​​in the infinite interval to an open interval of zero to one. In this way, scores representing weak correlation or no correlation are compressed to a numerical space approaching zero, while high scores representing significant risk association are smoothly mapped to a normalized scalar approaching one. Finally, the system determines the processed value as the semantic drift coefficient for subsequent sensitivity superposition calculation. It is worth mentioning that, through the processing of the Sigmoid activation function, when the original drift correlation score is a large positive value, the exponential term quickly approaches zero, causing the generated semantic drift coefficient to quickly approach one, thereby triggering an effective enhancement of the basic sensitivity score. When the score is negative, the output semantic drift coefficient will rapidly decay and approach zero, ensuring that the system does not generate erroneous drift gain when the data content does not match the current external event flow. Through this non-linear activation mechanism, the system mathematically constructs a risk judgment threshold with saturation characteristics, which can maintain the overall stability of the dynamic sensitivity level judgment logic while ensuring high sensitivity to sudden events.

[0048] It is understandable that, since the basic sensitivity score obtained based on static rules only reflects the inherent compliance risks of the data content, while the semantic drift coefficient quantifies the intensity of risk variation driven by the external macro environment, simply adding the two linearly may not effectively reflect the multiplier amplification effect of sudden emergency contexts on data sensitivity, and continuous floating-point score values ​​cannot be directly used for rigid permission threshold matching. Therefore, in the technical solution of this invention, the scoring enhancement and mapping module 140 further uses a preset drift gain factor to dynamically enhance the basic sensitivity score by adding the semantic drift coefficient as an increment to obtain an enhanced sensitivity score, and performs boundary truncation and discretization mapping on the enhanced sensitivity score to obtain a dynamic sensitivity level for the final decision, thereby constructing a scoring evolution logic that integrates static compliance benchmarks and dynamic environment drift gains. In this way, it can ensure that the data sensitivity rating shows a significant upward trend when driven by high-risk semantic drift coefficients, while converging the continuous mathematical calculation results into an executable discrete security level, providing a deterministic input basis for subsequent access control decisions.

[0049] Specifically, in this embodiment of the invention, the scoring enhancement and mapping module 140 is used to: multiply the semantic drift coefficient by the drift gain factor to obtain a dynamic increment ratio, and superimpose the dynamic increment ratio onto the basic sensitivity score through a multiplicative model to obtain an enhanced sensitivity score; based on a preset global maximum score limit, perform numerical saturation and boundary constraints on the enhanced sensitivity score to obtain a final continuous score; and map the final continuous score to a predefined discrete security hierarchy through a piecewise step function to obtain a dynamic sensitivity level.

[0050] More specifically, in a specific example of the present invention, the scoring enhancement and mapping module 140 first reads the drift gain factor configured by the system. This drift gain factor, as a sensitivity hyperparameter for adjusting the weight of external environmental influences, determines the system's response magnitude to semantic drift. The system multiplies this drift gain factor with the semantic drift coefficient output in the previous step to calculate the dynamic increment ratio. Subsequently, the system uses a multiplicative growth model to superimpose this dynamic increment ratio as a multiplier factor onto the basic sensitivity score to calculate the enhanced sensitivity score reflecting the current comprehensive risk. To prevent the score value from expanding infinitely due to extreme environmental drift and exceeding the system's processing range, the system introduces a preset global maximum score upper limit, performs numerical saturation processing on the enhanced sensitivity score, and takes the smaller of the calculated value and the upper limit value as the final continuous score. Finally, the system uses a piecewise ladder function or rounding logic to map the final continuous score in the continuous interval to a predefined discrete security level system (such as L1 to L5 levels), and outputs the final dynamic sensitivity level.

[0051] During the dynamic enhancement and boundary truncation process, the system executes the following calculation formula: ; in, This represents the final continuous score after numerical saturation processing; This indicates the preset global maximum score limit; This represents a function that takes the minimum value. Indicates the basic sensitivity score; Indicates the drift gain factor; This represents the semantic drift coefficient. This formula embodies the control principle of benchmark-based, drift-enhancing approach: This part constitutes the dynamic multiplier, which is determined when the external environment is stable. At that time, the continuous scoring was maintained at the static benchmark. Nearby; and when major events such as an epidemic outbreak occur, leading to When it rises, As an amplifier, it drives the overall score to increase exponentially, thus rapidly raising the score of ordinary supplies such as masks to the high-risk range. Meanwhile, The function ensures that even with extremely drastic drift, the final continuous score will not be exceeded. The defined system range (e.g., 100 points) ensures the stability of the numerical system.

[0052] During the discretization mapping process, the system executes the following calculation formula: ; in, Indicates the dynamic sensitivity level; This represents the floor function; Indicates the final continuous score; This represents the step size of the score for each level in a discrete security hierarchy. The formula represents the transformation of mathematical risk values ​​into management decision values: assuming the system defines 5 security levels with a total score of 100 points, then the step size... It is 20. If the calculated The score is 85 (the original static score may have been only 40 points, which is enhanced due to drift). Through this formula mapping, the system will output level 5 (L5), which can directly drive the permission matching module to execute the highest level "blocking alarm" strategy, realizing the transformation from numerical calculation to business risk control.

[0053] It is understandable that in big data access environments, data sensitivity fluctuates dynamically due to external context. Traditional role-based static permission granting mechanisms cannot detect real-time changes in data risk, leading to unauthorized access or risk exposure in special scenarios where sensitivity surges due to external events. Therefore, in the technical solution of this invention, the permission matching module 150 further matches the dynamic sensitivity level with the user attribute profile to obtain the corresponding access control decision result. This establishes a dynamic adjudication mechanism based on risk and capability equivalence, so that the granting of access permissions no longer depends solely on static identity, but on a quantitative comparison between the real-time data risk value and the user's current security capabilities. In this way, fine-grained and adaptive access control can be achieved. While ensuring that high-risk data is strictly controlled, the hierarchical decision logic reduces the rigidity of security policies and significantly improves the system's security protection efficiency under complex attack and defense situations.

[0054] Specifically, in this embodiment of the invention, the permission matching module 150 is used to: quantitatively evaluate the user's security capabilities by analyzing the feature values ​​in the user's attribute profile to obtain a user capability score; retrieve the corresponding minimum capability requirement threshold in the policy threshold matrix using the dynamic sensitivity level as the index key, and calculate the difference between the user capability score and the minimum capability requirement threshold to obtain a permission gap value; and match the permission gap value with a preset hierarchical decision logic to obtain an access control decision result that includes a full grant, downgrade desensitization, or blocking alarm instruction.

[0055] More specifically, in a big data access environment, an entity's access permissions should not be determined solely by its department or static role, but must comprehensively consider its real-time security status in the current access session, such as the compliance of the terminal device and the risk level of the network access environment, to address constantly evolving security threats. Therefore, in the technical solution of this invention, the feature values ​​in the user attribute profile are further quantitatively evaluated to obtain a user capability score, thereby establishing a dynamic credit measurement mechanism based on multi-factor fusion. This allows the multi-dimensional, fragmented features of an entity to be mapped into a unified mathematical scalar, providing a quantitative basis for subsequent precise comparison and alignment between data risk levels and user capabilities.

[0056] In a specific example of this invention, user attribute profiles are first synchronously obtained from the system's identity authentication module and then structuredly parsed to extract feature values ​​for multiple key dimensions. These feature values ​​include job level weights, terminal device trust scores, and network environment security indices. The system execution process includes: first, normalizing the extracted attribute feature values ​​to map them to a predefined numerical space; second, performing a weighted summation operation on the feature values ​​based on the security weights pre-set in the system configuration library; and finally, calculating a corrected score by combining the real-time background factors of the current access session and adding it to the weighted result, thereby producing a user capability score that represents the user's credit limit in the current context.

[0057] The following calculation formula is used during the execution of the quantitative assessment of user security capabilities: ; in, This represents the user's capability score, which quantifies the user's overall permission level in the current access session; M represents the total number of user attribute dimensions considered. This represents the preset weight coefficient of the m-th user attribute, used to reflect the contribution of different dimensions such as job level and equipment to security capabilities; This represents the normalized feature value of the m-th user attribute extracted from the user attribute profile; This represents the additional correction score imposed by the current environmental context, used to dynamically adjust the security assessment results based on the real-time network environment. The formula achieves global integration of subject attribute characteristics through a multi-factor linear combination mechanism. Specifically, in the formula... By assigning points to core factors such as job level weight and device trust level, a basic competency foundation was established for the user, allowing users with higher job levels or those using highly trusted terminals to obtain a higher base score; while The correction item acts as an environmental feedback mechanism. When a user initiates access from an untrusted network or a remote environment, the negative gain generated by this item will effectively lower the user's ability score, thereby mathematically restricting their access to highly dynamic and sensitive data, ensuring that permission granting is dynamically adapted to the real-time risk environment.

[0058] More specifically, in big data access environments, data sensitivity exhibits dynamic fluctuations driven by external context. Traditional role-based static permission granting mechanisms cannot perceive real-time changes in data risk. Without establishing a quantitative comparison mechanism for real-time risk levels, it is difficult to determine whether a user's current security background is sufficient to cover specific data access risks. Therefore, in the technical solution of this invention, the corresponding minimum capability requirement threshold is retrieved from the policy threshold matrix using the dynamic sensitivity level as the index key, and the difference between the user's capability score and the minimum capability requirement threshold is calculated to obtain the permission gap value. This establishes a bridge between data risk and the subject's capability through mathematical difference calculation. In this way, the compliance level of access behavior can be accurately measured, providing deterministic logical support for subsequent tiered handling actions.

[0059] In a specific example of this invention, the permission matching module first loads a policy threshold matrix pre-installed in the system configuration library. This policy threshold matrix defines the minimum security capability baseline required for access at each level within the security hierarchy. The system obtains the dynamic sensitivity level calculated by the preceding module and uses it as the primary index key to perform a location query on the policy threshold matrix, thereby retrieving the minimum capability requirement threshold required under the current data risk state within milliseconds. Subsequently, the system calls the arithmetic logic unit to perform a subtraction calculation, subtracting the retrieved minimum capability requirement threshold from the user capability score quantified from the user attribute profile, ultimately generating a permission gap value representing the permission matching depth.

[0060] The following formula is used in the calculation of the permission gap value: ; in, This represents the permission gap value, used to quantify the redundancy or inadequacy of a user's current capabilities relative to data risks. The user capability score is a comprehensive credit score based on a quantitative assessment of the user's attribute profile. The policy threshold matrix stores predefined security threshold values ​​for each level. It represents the dynamic sensitivity level and serves as the primary key of the retrieval matrix index; This represents the matrix retrieval operation function. The formula contains... This partially represents the dynamic protection benchmark on the object side, which changes in real time as risks drift due to external context. Through By subtracting from this baseline, the system converts the qualitative security hierarchy relationship into a quantitative numerical deviation. (Accessibility gap value) The positive or negative attribute of the difference directly determines the direction of the decision. When the difference is positive, it means that the user's security capabilities in the current context are sufficient to cover the data risks, and the system can execute the permission instruction. When the difference is negative, it intuitively reveals the permission gap between the subject's capabilities and the object's risks, thereby triggering the system to downgrade or desensitize the access behavior or directly block it, ensuring that the data remains under control during special periods of heightened sensitivity.

[0061] It is understandable that, in a dynamically changing business context, the sensitivity of data may fluctuate instantaneously. A simple allow or block decision logic is insufficient to balance data security and business continuity, especially when there is a slight discrepancy between the user's capability score and the current dynamic sensitivity level requirements; direct blocking would impair business availability. Therefore, in the technical solution of this invention, a preset hierarchical decision logic is further matched based on the permission gap value to obtain access control decision results that include full allow, downgraded desensitization, or blocking alarm instructions, thereby constructing an adaptive control strategy with a risk buffering mechanism. This allows for differentiated control actions based on the redundancy or gap in user capabilities relative to data risk. While strictly controlling high-risk leaks, desensitization processing is performed on medium-risk requests to maximize business accessibility.

[0062] In a specific example of this invention, after obtaining the permission gap value, the permission matching module executes segmented discrimination logic by calling a predefined decision mapping operator. The system first compares the permission gap value with zero to determine whether the user's security capabilities fully cover the current data risk; if the permission gap value is less than zero, the system further determines whether the gap is within a preset elasticity range. During execution, the system retrieves the threshold between degraded service buffers by searching configuration parameters, and extracts the corresponding control instructions from the decision instruction library based on the spatial position of the permission gap value relative to this threshold. The final access control decision result not only defines the admission status of the current session but also synchronously drives the underlying execution gateway to physically intercept, mask, or fully transmit the data stream. In this process, the mapping logic of the hierarchical decision follows the following segmented function calculation formula: ; in, It indicates the access control decision result, including the specific security instructions executed by the system in response to the current access request; This represents the permission gap value, which is the mathematical difference between the user's ability score and the demand threshold obtained from the policy threshold matrix; This represents the threshold between the downgraded service buffers, serving as a flexible boundary parameter for determining whether to perform data downgrade and desensitization processing. In practical applications, this formula achieves a precise conversion from mathematical deviations to physical execution logic. Specifically, in the formula... The conditional branch ensures that when a user has sufficient credit, the system issues a full grant command, guaranteeing the efficient execution of legitimate access. (Introducing...) The interval discrimination logic reflects the adaptability of this solution. When the target data content experiences an unexpected increase in dynamic sensitivity due to semantic drift, causing a slight gap in user permissions, the system does not take a rigid blocking action. Instead, it dynamically masks sensitive fields in the target data content by outputting a Masking command, thus maintaining business continuity while keeping risks under control. Finally, through... The system establishes a hard defense line against unauthorized access by using a discrimination branch, and forcibly issues blocking commands to prevent substantial leakage of core sensitive assets, thereby achieving a deep integration of security strength and business flexibility.

[0063] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A control system based on dynamically optimized data sensitivity levels, characterized in that, include: The raw data stream acquisition module is used to acquire raw data request packets containing target data content and system time, a preset static rule dictionary, and external event streams within the current time window; The sensitivity scoring and semantic understanding module is used to match and score the target data content based on a static rule dictionary to obtain a basic sensitivity score, and to use a pre-trained word embedding model to extract features and vectorize the target data content and external event stream to obtain data content vectors and global event context vectors respectively. The semantic drift quantization evaluation module is used to perform semantic drift quantization evaluation on the data content vector and the global event context vector based on the semantic mapping matrix to obtain the semantic drift coefficient; The scoring enhancement and mapping module is used to dynamically enhance the sensitivity score by adding the semantic drift coefficient as an increment to the basic sensitivity score based on a preset drift gain factor. The enhanced sensitivity score is then truncated and discretized to obtain the dynamic sensitivity level for the final decision. The permission matching module is used to match dynamic sensitivity levels with user attribute profiles to obtain corresponding access control decision results.

2. The control system based on dynamically optimized data sensitivity level according to claim 1, characterized in that, The sensitivity scoring and semantic understanding module includes: The basic sensitivity score determination unit is used to traverse the static rule dictionary to match sensitive keywords in the target data content and determine the basic sensitivity score based on the matching frequency. The word embedding processing unit is used to perform stop word removal on the target data content and input the processed text sequence into the pre-trained word embedding model for feature mapping to obtain the data content vector; The business context hotspot representation unit is used to aggregate hotspots in external event streams using the word frequency-inverse document frequency algorithm to extract a set of high-weight keywords, and then converts the set of high-weight keywords into a global event context vector that can represent the current business context hotspots through a pre-trained word embedding model.

3. The control system based on dynamically optimized data sensitivity level according to claim 2, characterized in that, The basic sensitivity score determination unit is used to determine the basic sensitivity score using the following formula: ； in, Basic sensitivity score, For the first The preset weight coefficient for each sensitive keyword. To indicate the first The frequency of occurrence of each sensitive keyword in the target data content This is the default base security score. This represents the maximum allowed static score by the system. To find the minimum value function, N This represents the total number of sensitive keywords matched in the target data content.

4. The control system based on dynamically optimized data sensitivity level according to claim 1, characterized in that, The semantic drift quantification evaluation module includes: The spatial alignment transformation unit is used to perform spatial alignment transformation on the global event context vector based on the semantic mapping matrix to obtain the transformed semantic context vector. The correlation estimation unit is used to estimate the correlation between the transformed semantic context vector and the data content vector to obtain the original drift correlation score. The activation unit is used to perform nonlinear activation and coefficient normalization on the original drift correlation score to obtain the semantic drift coefficient.

5. The control system based on dynamically optimized data sensitivity level according to claim 1, characterized in that, The scoring enhancement and mapping module is used for: The dynamic increment ratio is obtained by multiplying the semantic drift coefficient by the drift gain factor, and then the dynamic increment ratio is superimposed on the basic sensitivity score through a multiplicative model to obtain the enhanced sensitivity score. Based on the preset global maximum score limit, the enhanced sensitivity score is numerically saturated and boundary-limited to obtain the final continuous score; The final continuous score is mapped to a predefined discrete security hierarchy using a piecewise ladder function to obtain the dynamic sensitivity level.

6. The control system based on dynamically optimized data sensitivity level according to claim 1, characterized in that, The permission matching module is used for: Quantitatively assess user security capabilities by analyzing the feature values ​​in the user attribute profile to obtain a user capability score; Using the dynamic sensitivity level as the index key, retrieve the corresponding minimum capability requirement threshold in the policy threshold matrix, and calculate the difference between the user's capability score and the minimum capability requirement threshold to obtain the permission gap value. Based on the permission gap value, a preset hierarchical decision logic is matched to obtain access control decision results that include full permission, downgrade and desensitization, or blocking alarm instructions.

7. The control system based on dynamically optimized data sensitivity level according to claim 4, characterized in that, The correlation estimation unit is used for: Obtain a pre-defined normal semantic basis matrix containing multiple orthogonal basis vectors; The transformed semantic context vector is projected onto the safe subspace constructed by the normal semantic basis matrix to obtain a pure risk context vector; Calculate the noise entanglement penalty coefficient of the data content vector on the normal semantic basis matrix; Based on the noise entanglement penalty coefficient, a penalty-based comparative drift score is performed on the data content vector and the pure risk context vector to obtain the original drift correlation score.

Images