A multi-source heterogeneous geological data desensitization method and device

Abstract

The application discloses a multi-source heterogeneous geological data desensitization method and device, including the following steps: according to the input multi-source heterogeneous geological data, performing named entity recognition, and extracting geological entity data in the multi-source heterogeneous geological data; linking the geological entity data and the pre-input geological knowledge graph to obtain reasoning data; evaluating the reasoning data through a dynamic sensitivity grading evaluation model to obtain the sensitivity grading of the reasoning data with different sensitivities, and determining a sensitive target; performing geometric remodeling on the sensitive target, simulating the tectonic movement of the sensitive target, performing constraint and topological anti-flipping through a Jacobian determinant, constructing a Lagrange lock of key points, and obtaining spatial deformation data; and performing attribute desensitization and picture-text collaborative desensitization on the spatial deformation data to obtain desensitization data. The technical scheme can keep the real geological data hidden while maintaining a high degree of logical consistency with the original data.

Description

Technical Field

[0001] This application relates to the field of geological data processing technology, and in particular to a method and apparatus for desensitizing multi-source heterogeneous geological data. Background Technology

[0002] With the deepening of the "Digital Earth" strategy and the advancement of the "Deep Earth Exploration" major science and technology project, the geological exploration industry is undergoing a paradigm shift from experience-based mineral exploration to intelligent mineral exploration. In this process, geological big data has become a core production factor. In particular, for the typical barrier-type fold structure and low-temperature hydrothermal mineralization system (mainly fluorite, barite, and calcite) in southeastern Chongqing, high-dimensional, multi-source, and heterogeneous exploration data has irreplaceable value for training high-precision artificial intelligence (AI) mineral exploration models. Direct sharing or external transfer for research, bidding, or other purposes faces a serious risk of leakage.

[0003] Traditional general-purpose data anonymization techniques, such as simple coordinate offsetting, k-anonymity, differential privacy, or simple text anonymization, exhibit significant limitations when dealing with geological data that possesses high spatial autocorrelation, strict physical constraints, and complex genetic logic. This is particularly true in southeastern Chongqing, where mineral deposit formation is jointly constrained by tectonic stress fields, stratigraphic attributes, and weak geochemical anomalies. If these geological logics are severed during the anonymization process, the data becomes meaningless noise, preventing AI models from learning the true mineralization patterns. Therefore, developing a complete data anonymization process that both ensures privacy and security while maintaining the fidelity of geological logic has become an urgent need in the current field of geological big data applications. Summary of the Invention

[0004] This application provides a method and apparatus for desensitizing multi-source heterogeneous geological data, which can conceal the real geological data while maintaining a high degree of logical consistency with the original data.

[0005] Firstly, this application provides a method for desensitizing multi-source heterogeneous geological data, employing the following technical solution: A method for desensitizing multi-source heterogeneous geological data includes the following steps: Named entity recognition is performed on the input multi-source heterogeneous geological data to extract geological entity data from the multi-source heterogeneous geological data; By linking the geological entity data with the pre-input geological knowledge graph, inference data is obtained; The inference data is evaluated using a dynamic sensitivity grading assessment model to obtain sensitivity gradings for the inference data with different sensitivities, thereby identifying sensitive targets. The sensitive target is geometrically reshaped to simulate its construction motion. Constraints and topological anti-flipping are performed using Jacobian determinants to construct Lagrange locking of key points and obtain spatial deformation data. The spatial deformation data is subjected to attribute desensitization and image-text co-desensitization to obtain desensitized data.

[0006] Furthermore, named entity recognition is performed based on the input multi-source heterogeneous geological data to extract geological entity data from the multi-source heterogeneous geological data. Specifically, this includes the following steps: The pre-trained layer of the deep learning architecture is adaptively fine-tuned based on the multi-source heterogeneous geological data, and the multi-source heterogeneous geological data is converted into fine-tuned data with high-dimensional contextual semantic representation and output to the encoding layer of the deep learning architecture. The encoding layer captures the long-distance semantic dependencies of the fine-tuning data through a bidirectional long short-term memory network, and outputs the long-distance semantic dependencies and the fine-tuning data to the decoding layer of the deep learning architecture; The decoding layer generates the label sequence dependency relationship of the fine-tuned data through a conditional random field, thereby obtaining geological entity data that conforms to geological grammar logic.

[0007] Furthermore, linking the geological entity data and the pre-input geological knowledge graph includes the following steps: The geological entity data is described using the Web ontology language standard, and semantic reasoning is performed on the geological entity data to obtain the reasoning data. The reasoning data includes the geological entity names, as well as the sequence relationships, lithological characteristics, and ore-controlling attributes between the geological entity names.

[0008] Furthermore, the dynamic sensitivity grading assessment model includes a sensitivity scoring function, which is used to evaluate the inherent attributes of geological entity data in the inference data, as well as the connectivity between the inherent attributes and the geological knowledge graph. The inference data is graded by the sensitivity scoring function to obtain several levels of sensitive targets, and targeted processing strategies are adopted for different levels of sensitive targets.

[0009] Furthermore, the simulated construction motion of the sensitive target specifically includes the following steps: The sensitive target is discretized into several tiny units using tetrahedral or triangular meshes; The mesh is given a fiber-reinforced hyperelastic constitutive model, the construction direction vector is set as the construction principal axis, and the strain energy density function is implanted into each micro-unit. The strain energy density function is the sum of the matrix term and the anisotropy term. Anisotropic constraints are applied to the micro-units to lock their structural pattern; Boundary constraints and virtual body forces are applied to the mesh, and the nonlinear equilibrium equations of the micro-units are solved to obtain the simulation results of the construction motion; The constraint and topology anti-flipping method using Jacobian determinant includes: In the simulation of the constructed motion, a penalty term based on the Jacobian determinant is introduced. When the Jacobian determinant approaches 0, the penalty term approaches infinity. The Lagrange locking is used to apply hard constraints to the structural key points and preserve the ore-controlling structural relationships of the structural key points.

[0010] Furthermore, the attribute desensitization of the spatial deformation data includes the following steps: The spatial deformation data is characterized by generative adversarial network, virtual stratigraphic identifiers are synthesized based on the spatial deformation data, and the virtual stratigraphic identifiers are re-encapsulated into standardized geological metadata and injected into the desensitized database. The spatial deformation data is subjected to covariance noise addition by transforming the space by equal interval logarithmic ratio, and a gradient protection operator is introduced for weak anomaly features in the spatial deformation data to obtain attribute-desensitized spatial deformation data.

[0011] Furthermore, the image-text collaborative desensitization of the spatial deformation data includes the following steps: Scan the multi-source heterogeneous geological data and extract spatial relationship triples, which include the orientation, distance and target of the multi-source heterogeneous geological data; The spatial relationship triples are updated based on the spatial deformation data, and the updated spatial relationship triples are backfilled into the desensitized data using a natural language generation model.

[0012] Furthermore, the above method also includes performing Jacobi matrix-driven vector orientation correction and raster data inverse mapping alignment on the desensitized data to verify the holographic spatial alignment and consistency of the desensitized data.

[0013] Furthermore, the above method also includes evaluating the de-identified data and establishing quantitative evaluation indicators, which include the topological consistency, causal logic, construction style, and usability of the de-identified data; Determine whether the de-identified data meets all the criteria in the quantitative evaluation indicators; If all items in the quantitative evaluation indicators are satisfied, the de-identified data will be output; If any of the quantitative evaluation indicators are not met, the stiffness matrix of the mesh and / or the noise level are adjusted iteratively until all of the quantitative evaluation indicators are met, and the desensitized data is output.

[0014] Secondly, this application provides a multi-source heterogeneous geological data desensitization device, which adopts the following technical solution: A multi-source heterogeneous geological data desensitization device, employing the multi-source heterogeneous geological data desensitization method described above, includes: The data acquisition module is used to collect multi-source heterogeneous geological data; The data recognition module is used to perform named entity recognition based on the input multi-source heterogeneous geological data and extract geological entity data from the multi-source heterogeneous geological data. The data reasoning module is used to link the geological entity data and the pre-input geological knowledge graph to obtain reasoning data; The data sensitivity grading and evaluation module is used to evaluate the inference data through a dynamic sensitivity grading and evaluation model, obtain the sensitivity grading of the inference data with different sensitivities, and determine the sensitive targets. The data reshaping module is used to geometrically reshape the sensitive target, simulate the construction motion of the sensitive target, constrain and prevent topological flipping through Jacobian determinant, construct Lagrange locking of key points, and obtain spatial deformation data. The data desensitization module performs attribute desensitization and image-text collaborative desensitization on the spatial deformation data to obtain desensitized data.

[0015] In summary, this application includes at least one of the following beneficial technical effects: This application provides a method and apparatus for desensitizing multi-source heterogeneous geological data. By performing named entity recognition on the multi-source heterogeneous geological data, geological entity data is extracted. Then, by linking the geological entity data with a pre-input geological knowledge graph, inference data can be obtained. This not only extracts coordinates but also deeply analyzes the physical properties of strata and structures, providing knowledge anchors for subsequent semantic replacement. A dynamic sensitivity grading evaluation model is used to grade the sensitivity of the inference data, identifying sensitive targets and achieving automated classification and grading of sensitive entities. A geological knowledge graph is then constructed to guide subsequent processing. The sensitive targets are geometrically reshaped, and a complete finite element simulation is used to solve for a... The local displacement field is used to obtain the generating function for all subsequent heterogeneous data deformations. In order to ensure that the anonymized data can still guide the artificial intelligence model to identify core mineral exploration features while concealing the true identity of the data, this application generates a virtual stratigraphic package by performing attribute anonymization and image-text co-anonymization on the spatial deformation data, and establishes a real-time synchronization mechanism between text description and geometric map. This achieves complete concealment of real geographical locations and sensitive place names, while ensuring that the anonymized virtual dataset maintains a high degree of logical consistency with the original data in four dimensions: "construction topology, stratigraphic semantics, element distribution, and three-dimensional occurrence". This supports external institutions to directly use the anonymized data for high-quality artificial intelligence mineral exploration prediction and scientific research applications. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the multi-source heterogeneous geological data desensitization method in the embodiments of this application.

[0017] Figure 2 This is a schematic diagram of the data desensitization process for low-temperature hydrothermal mineral resources in southeastern Chongqing in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] This application provides a method for desensitizing multi-source heterogeneous geological data, aiming to solve the problem that traditional general desensitization techniques cannot be directly applied to geological data desensitization. Taking the low-temperature hydrothermal mineral deposits in southeastern Chongqing as an example, the southeastern Chongqing region is located in the southeastern Sichuan fold belt on the southeastern edge of the Yangtze Platform. Its most prominent feature is the development of typical "barrier-like" folds. This structural morphology is characterized by a compact anticline axis and a wide and open syncline, arranged in parallel in a north-northeast (NNE) direction. This unique "comb-like" structural pattern not only determines the topography but is also the core controlling factor for low-temperature hydrothermal mineralization.

[0020] Within this geological framework, the distribution of mineral resources (mainly fluorite and barite) exhibits a strong regularity. Firstly, it is strongly tectonically controlled, with ore bodies primarily occurring in the detached spaces of anticline axes, compressional-shear fault zones, or interlayer breccia zones. For example, in areas like Pengshui, the distribution of fluorite deposits is clearly controlled by the shear fields of northwest-trending (NW) or north-northeast-trending (NNE) faults. Secondly, it exhibits stratigraphic specificity, with ore-forming materials mostly concentrated in specific carbonate strata from the Cambrian to the Ordovician periods, such as the Honghuayuan Formation (O1h). Finally, it displays weak anomaly characteristics; these deposits have relatively low mineralization temperatures, resulting in low contrast of surface geochemical indicator elements (such as Hg, Sb, and As), making prospecting highly dependent on trace element ratios and spatial gradient characteristics.

[0021] Based on the aforementioned geological characteristics, existing desensitization technologies face at least the following unavoidable technical challenges when applied to multi-source heterogeneous data in southeastern Chongqing: Semantic-location decoupling fails: simple code replacement (such as replacing "Honghuayuan Group" with "L_05") cuts off the bridge between the AI ​​model and the external geological knowledge graph, making it unable to infer lithological properties.

[0022] Structural topological distortion: Random coordinate jitter or nonlinear distortion can easily lead to fault knotting, strata interpenetration, and even cause local stress field reversal (from tensile to compressive), which undermines the prediction logic of filling-type deposits. Geochemical “weak anomaly” annihilation: The noise added by differential privacy often exceeds the contrast of weak geological anomalies and disrupts the covariance between elements (such as the F / Ca ratio). Text "relative positioning" is difficult to clean: retaining relative descriptions such as "200 meters on the northwest wing of the anticline" makes it easy to deduce the absolute position in reverse, while deleting them results in the loss of data value. Asynchronous multi-source heterogeneous data: If borehole (1D), geological map (2D), and geochemical data (2D) use different transformation parameters, it will result in dirty data such as "the borehole falls on the wrong stratum".

[0023] In summary, there is an urgent need for a method that can understand geological semantics, preserve tectonic topology, maintain geochemical logic, and synchronize multiple sources in order to declassify geological data in southeastern Chongqing.

[0024] To address the above problems, this application discloses a method for desensitizing multi-source heterogeneous geological data, employing the following technical solution: Reference Figure 1 A method for desensitizing multi-source heterogeneous geological data includes the following steps: S101: Perform named entity recognition based on the input multi-source heterogeneous geological data, and extract geological entity data from the multi-source heterogeneous geological data; In this embodiment, multi-source heterogeneous geological data may include geological exploration reports, academic papers, and geological records. The sensitivity of multi-source heterogeneous geological data is often hidden in unstructured text descriptions and is highly specialized and nested. For example, "Ordovician Honghuayuan Formation Upper Bioclastic Limestone" is not just a stratigraphic name, but also implies multiple attributes such as age (Ordovician), lithology (limestone), and genesis (biogenic sedimentation). To accurately capture this information, this application employs a deep learning architecture based on BERT-BiLSTM-CRF for automatic extraction of geological entities.

[0025] Specifically, it includes the following steps: The pre-trained layer of the deep learning architecture is adaptively fine-tuned based on the multi-source heterogeneous geological data, and the multi-source heterogeneous geological data is converted into fine-tuned data with high-dimensional contextual semantic representation and output to the encoding layer of the deep learning architecture. The encoding layer captures the long-distance semantic dependencies of the fine-tuning data through a bidirectional long short-term memory network, and outputs the long-distance semantic dependencies and the fine-tuning data to the decoding layer of the deep learning architecture; The decoding layer generates the label sequence dependency relationship of the fine-tuned data through a conditional random field, thereby obtaining geological entity data that conforms to geological grammar logic.

[0026] In this implementation, the pre-training layer can use geological exploration reports, academic papers, and geological records as a corpus to perform domain-adaptive fine-tuning of the BERT model. This layer can capture high-dimensional contextual semantic representations of professional terms such as "anticline," "fracture," "alteration zone," and "vacuum space," solving the problem that general NLP models cannot understand geological context. The encoding layer uses a bidirectional long short-term memory network (BiLSTM) to capture long-distance semantic dependencies. For example, in the sentence "This ore body is located on... (describes a large number of lithological features)... the footwall of the fault zone," the model can identify the spatial dependency between "ore body" and "footwall of the fault zone" across long texts. The decoding layer uses conditional random fields (CRF) to solve the dependency problem of label sequences, ensuring that the annotation results conform to geological grammatical logic. For example, the CRF layer uses transition probability matrix constraints to ensure that the identified "group" level units (such as the Honghuayuan Formation) must belong to "system" level units (such as the Ordovician), preventing labeling errors such as reversed stratigraphic sequences.

[0027] The extracted geological entity data includes, but is not limited to: Geological units: strata (such as the "Honghuayuan Formation" and "Sangzhi Formation"), rock masses, ore bodies, etc.

[0028] Structural elements: faults (such as the "F1 fault"), folds (such as the "Qiyaoshan anticline"), joints, ductile shear zones, etc.

[0029] Spatial orientation: direction, inclination, distance, azimuth, coordinate string, etc.

[0030] Geochemical indicators: element name, grade value (e.g., CaF2%), ratio characteristics, etc.

[0031] S102: Link the geological entity data and the pre-input geological knowledge graph to obtain inference data; In one embodiment of this application, simple entity recognition is insufficient to support logically accurate desensitization; the system must "understand" the geological meaning behind the entities. Therefore, this invention uses the Web Ontology Language (OWL) standard to describe geological domain ontology, constructing a geological knowledge graph (KG). The geological knowledge graph and geological domain ontology not only contain entity names but also stratigraphic relationships between entities, lithological characteristics, and ore-controlling attributes, providing a logical basis for subsequent "semantic substitution," for example: Core ontology relation definition: 1. Subordination: e.g., “Honghuayuan Group ∈ Lower Ordovician”.

[0032] 2. Ore-hosting relationship: such as "carbonate rock → fluorite ore".

[0033] 3. Ore-controlling structures: such as "NNE-trending fault → hydrothermal channel".

[0034] 4. Spatial correlation: such as "skarnification ↔ contact zone".

[0035] The core classes and attribute definitions of the geological ontology are shown in Table 1 below: Table 1 By linking identified entities to this geological knowledge graph, semantic reasoning can be performed. For example, when the system identifies the "Honghuayuan Formation," even if it is renamed to "Virtual Stratum A," the system can automatically generate a virtual description that conforms to the characteristics of limestone based on the "lithology: limestone" attribute in the KG, instead of mistakenly describing it as sandstone, thus maintaining the consistency of geological logic.

[0036] S103: The inference data is evaluated using a dynamic sensitivity grading evaluation model to obtain sensitivity gradings of the inference data with different sensitivities, and sensitive targets are determined. In this embodiment, since not all geological data have the same sensitivity, this application introduces a dynamic sensitivity assessment mechanism based on inference potential to balance security and data utility, through a sensitivity scoring function. Taking into account the inherent properties of an entity and its connectivity in the KG: Here, Relations(e) represents the set of all relationships that entity e possesses (such as geological relationships like "assigned to" or "located in"). It is the inherent sensitivity of an entity, without considering its relationship with other entities; It is the association sensitivity, which is the sensitivity propagated from the associated entity r, and takes the inherent sensitivity of the associated entity; w1 and w2 represent weight coefficients (usually w1+w2=1), and their function is to adjust the importance ratio of "inherent attributes" and "associated background" in the final score; This represents the sum of the association sensitivities of all related relationships.

[0037] In one embodiment of this application, the inference data can be divided into the following levels: Red level ( Examples include precise coordinates, strategic mineral names, and high-precision borehole data. Subsequent processing strategies include: fully nonlinear deformation + pseudonymization + strong noise reduction.

[0038] Yellow level ( Examples include regional geological boundaries and indicative element ratios. Subsequent processing strategy: topology-preserving deformation + covariance-preserving noise reduction.

[0039] Green level ( Examples include general background geological knowledge and non-mineralized strata. Subsequent processing strategy: Preserve statistical characteristics.

[0040] For red-level sensitive targets, all processing strategies from steps S104 to S106 below need to be executed to perform a completely non-linear deformation of the sensitive targets, while pseudonymizing and adding strong noise to make this part of the data obtain the highest level of protection. For sensitive targets at the yellow level, you can choose to perform only the topology preservation deformation in step S104 and the covariance preservation noise addition in step S105, and then perform step S106 for verification, so that this part of the data can obtain a higher protection strength. For sensitive targets at the green level, only the topology preservation deformation in step S104 can be performed, and then step S106 can be performed for verification, so that the attribute values ​​and names of this part of the data remain unchanged, reducing the computational intensity. By preset threshold and Ensure that data with different sensitivities receives differentiated protection levels, balancing security and usability.

[0041] S104: Perform geometric reshaping on the sensitive target, simulate the construction motion of the sensitive target, constrain and prevent topological flipping through Jacobian determinant, construct Lagrange locking of key points, and obtain spatial deformation data; Reference Figure 1 and Figure 2 In this embodiment, after identifying sensitive targets, different strategies are selected based on the level of sensitivity. A global displacement field is solved using finite element method (FEM) simulation through physical constraints such as anisotropic hyperelastic material constraints, Jacobian determinant constraints, and Lagrange keypoint locking. As the generating function for all subsequent heterogeneous data deformation, the spatial anonymization requirements for geological maps are extremely high: they must prevent attackers from reconstructing the true location through graphic matching (such as river bends and terrain features), while ensuring that geologists can read the correct tectonic evolution history from the deformed map. Simple translations and rotations (rigid transformations) are easily restored; however, random nonlinear distortions (such as image liquefaction) can easily lead to topological errors such as fault knots and stratigraphic overlaps, destroying key mineral exploration clues such as "tectonic clamping".

[0042] In one embodiment of this application, the global displacement field The generation process is equivalent to simulating a controlled "virtual construction movement" in a computer. An exemplary implementation step is as follows: S401: Spatial discretization (mesh generation); The two-dimensional or three-dimensional geological space (domain) of southeastern Chongqing The data is discretized into countless tiny units. A tetrahedral (3D) or triangular (2D) mesh is used, with each vertex of the mesh designated as a node. .

[0043] Step 402: Assign anisotropic material properties (define "physical gene") To maintain the NNE-oriented distribution characteristics of the "partitioned folds", a fiber-reinforced hyperelastic constitutive model is applied to the mesh.

[0044] ① Define the principal structural axis: Based on the geological background of southeastern Chongqing, define the structural direction vector. Pointing to North-Northeast (NNE30°).

[0045] ② Energy injection: The function injects the strain energy density function into each element.

[0046] Strain energy density function Defined as: (Strain energy density function): Describes the "virtual elastic energy" stored per unit volume of a geological body during deformation. The system seeks the most natural deformation state by minimizing this energy.

[0047] (Deformation Gradient Tensor): This is a A matrix, mathematically, describes the geometric mapping relationship (including rotation, scaling, and shearing) from the "original coordinate point" to the "desensitized coordinate point" in space.

[0048] Matrix item The Neo-Hookean model is used to describe isotropic deformation, controlling for volume changes (Jacobi determinant). ).

[0049] (Shear modulus / Lame constant): Represents the rock's ability to resist shape changes (shear). The larger the rock, the harder it is, and the more difficult it is to deform.

[0050] (Lame's first constant): controls the volumetric compressibility of rocks. The larger the geological body, the more difficult it is to compress its volume.

[0051] (First strain invariant): , which measures the overall stretching degree, where C is the right Cauchy-Green deformation tensor, obtained by multiplying the deformation gradient tensor F itself with its transpose matrix, describing the pure deformation (stretching or compression) state of the object (here, the geological mesh) after removing rigid body rotation, and trace is the trace of the computation matrix, which is defined as the sum of the diagonal elements.

[0052] (Jacobi determinant): , representing the rate of volume change. This indicates that the volume remains constant; Indicates compression; Indicates expansion. In the formula... The item is to prevent volume collapse (i.e., to prevent...) (Change to 0 or a negative number to ensure the formation does not disappear).

[0053] Anisotropic terms : (Fourth pseudo-invariant): Elongation after square. Measures the elongation of a geological body along its tectonic axis ( The degree of stretching or compression. When When , it indicates no deformation along the axis.

[0054] (Structural Direction Vector): A unit vector pointing to the main structural trend (northeast, NNE) in southeastern Chongqing, defining the direction of the "reinforcing steel".

[0055] (Fiber stiffness coefficient): The unit is stress (MPa). It determines the "stiffness" of the structural axis. The larger the value, the more the deformation field tends to slide along the fault and fold axis, making it difficult to cut them off. This protects the continuity of the fault.

[0056] (Nonlinear hardening coefficient): A dimensionless constant. It determines the "sensitivity" to deformation. The larger the fold axis, the more exponentially the resistance increases when attempting to stretch it. This forcibly locks in the length and shape of the fold, preventing an antislope from being stretched too long or flattened during desensitization.

[0057] The purpose of this is to increase the stiffness along the structural axis, so that the deformation field tends to occur parallel to the structural line, thus avoiding unnatural distortion of the transverse structure.

[0058] ③ Anisotropy constraint: where The term contains an exponential hardening function. This means that when the grid is along... When stretched in the direction (fold axis), the resistance is small; but if an attempt is made to stretch perpendicular to the fold axis... When directional compression occurs (i.e., when the fold morphology is disrupted), the resistance increases exponentially. This is equivalent to implanting "virtual reinforcing bars" into the geological body, locking in the structural pattern.

[0059] Step 403: Apply virtual boundary conditions and physical forces (apply "external forces") To drive the mesh to deform, a set of virtual force fields needs to be applied: Boundary constraints: Fix the four corner points or edge nodes of the map (Dirichlet boundary conditions) to prevent the map from flying off as a whole.

[0060] Apply virtual stamina Apply a non-uniform force field inside the grid, such as applying a clockwise torque at the center of the map or a westward squeezing force in the southeast. This force is completely virtual and its purpose is solely to make the grid "move" without disrupting the topology.

[0061] Step 404: Solve the nonlinear equilibrium equations (calculate displacements) Solve the static equilibrium equations using the Newton-Raphson iterative method: in The divergence operator is derived by differentiating and summing the stress tensor σ. After receiving the stress tensor σ, the sum of the rates of change (partial derivatives) of σ in the x, y, and z directions is calculated. The result is a vector that represents the resultant force (internal force) per unit volume at that point due to uneven stress distribution. For the first Piola-Kirchhoff stress tensor, ; Virtual forces are artificially applied non-uniform force fields (such as torque and compressive force) used to drive mesh deformation.

[0062] in First Stress tensor. By solving this system of nonlinear equations, the global displacement field is obtained. Global displacement field This is a vector field defined over the entire geological body, describing the displacement of each material point under virtual tectonic motion. Here, "global displacement field" refers to a displacement field that is continuously distributed across the entire domain and satisfies physical uniqueness. Based on finite element discretization, the continuous equilibrium equations are transformed into nonlinear residual equations concerning the nodal displacements u. These are then solved using Newton-Raphson iterative linearization to obtain the displacement increment Δu. The iteration of Δu continues until convergence, ultimately yielding the global displacement field that satisfies both the equilibrium equations and boundary conditions. Thus, the transformed coordinates are obtained. This process simulates a "virtual tectonic movement," making the deformed geological body geomechanically plausible.

[0063] In one embodiment of this application, during large-amplitude nonlinear deformation, the mesh is prone to flipping, i.e., the local volume becomes negative. This causes the strata to "pass through" itself, disrupting the topology.

[0064] Therefore, we introduce a penalty term based on the Jacobian determinant into the optimization objective: in, The goal is to optimize the total energy output. It is elastic energy; These are the weighting coefficients for the topological constraints; For Jacobi determinant; Parameters used to control the steepness of the potential barrier; The total area of ​​two-dimensional or three-dimensional geological space that needs to be desensitized.

[0065] when At this point, the penalty term tends to infinity. This forms an infinitely high energy barrier, forcing the solver to strictly guarantee the correctness of the displacement field during optimization. .

[0066] This ensures the continuity of faults, prevents the reversal of stratigraphic sequence, and keeps closed structures (such as domes) closed, thus fundamentally eliminating topological errors.

[0067] In one embodiment of this application, for critical geological relationships (e.g., fault F1 must cut the anticline axis A), the system employs the Lagrange multiplier method to apply hard constraints: Alternatively, the relative positions can remain unchanged: in, For constraint functions; As a global displacement transformation function, for any point P in the domain, its new position Φ(P) after deformation can be obtained in the following way: , among them For displacement, if P is not a node, u(P) is calculated by interpolating the displacements of the surrounding nodes using the shape function (interpolation function) of the element in which it is located. These are the original key structural intersections (such as the intersection of a fault and an anticline axis). For the target location (or to maintain a relative relationship); For rotation / transformation matrices, preserve relative geometry. In the original, undeformed geological model, this is the coordinate vector of a specific point on a fault (e.g., a corner or center point of the fault plane). In the original, undeformed geological model, this is the coordinate vector of a specific point (e.g., a boundary point of the ore body). It is a fixed geometric vector that is directly read and calculated from the initial geological model. It serves as a "hard constraint" benchmark, ensuring that key geological elements (faults and ore bodies) maintain the correct spatial configuration after deformation.

[0068] The above methods ensure that even in the event of severe deformation across the entire region, the core ore-controlling structural relationships (such as "fault-controlled deposits") are accurately preserved for AI models to learn from.

[0069] S105: Perform attribute desensitization and image-text collaborative desensitization on the spatial deformation data to obtain desensitized data; Reference Figure 1 and Figure 2 In this embodiment, in order to enable the desensitized data to still guide the AI ​​model to identify core mineral exploration features while concealing the real identity, this application generates a virtual stratigraphic package by performing cross-modal attribute desensitization on the geological logic in the hidden space. At the same time, a real-time synchronization mechanism between text description and geometric map is established to block the vulnerability of reverse decryption by taking advantage of the inconsistency between text and graphics.

[0070] In one embodiment of this application, the attribute desensitization of the spatial deformation data includes the following steps: The spatial deformation data is characterized by generative adversarial network, virtual stratigraphic identifiers are synthesized based on the spatial deformation data, and the virtual stratigraphic identifiers are re-encapsulated into standardized geological metadata and injected into the desensitized database. The spatial deformation data is subjected to covariance noise addition by transforming the space by equal interval logarithmic ratio, and a gradient protection operator is introduced for weak anomaly features in the spatial deformation data to obtain attribute-desensitized spatial deformation data.

[0071] In this embodiment, to address the semantic loss caused by simple code replacement, this application designs a "semantic surrogate" generation mechanism, utilizing Generative Adversarial Networks (GANs) for representation learning within a geological knowledge graph (KG). For example, for sensitive strata such as the "Honghuayuan Formation," core logical genes closely related to low-temperature hydrothermal mineralization in southeastern Chongqing are extracted, including specific porosity-permeability combination vectors, lithological chemical reactivity characteristics, and brittle physical parameters of the strata. Then, instead of retrieving existing place names from external databases, a completely virtual stratum ID (e.g., Formation_S_YDN_001) is synthesized based on the aforementioned "logical genes" within a non-sensitive virtual feature space. The synthesized logical gene parameters are then repackaged into standardized geological metadata. By injecting this set of "virtual metadata" into the desensitized database, it is ensured that subsequent AI-powered mineral exploration models capture the physical role of the stratum in hydrothermal circulation and fluid infilling (e.g., "reservoir" or "caprock") during inference, rather than simple administrative labels, thereby achieving deep fidelity of prior mineralization knowledge.

[0072] In this embodiment, the constituent data in the spatial deformation data also need to be subjected to structured noise addition. For example, geochemical data (such as stream sediment measurement data) is typical constituent data, subject to "sum constraint" (Sum=100%). If standard difference privacy (Standard DP) is directly applied for independent noise addition, it will lead to spurious correlation and destroy the covariant relationship between elements (such as the symbiotic relationship between Zn and Pb).

[0073] This application introduces CoDA theory to perform privacy operations in the Euclidean space after isometric log-ratio (ilr) transformation.

[0074] First, construct an orthogonal basis. . Component vector Mapped to Real space of dimension : in The geometric mean is in In space, data is no longer subject to fixed rules and constraints; Transformed A vector in a real space. The original component vector (subject to fixed sum constraints, such as Sum=100%). It is an orthogonal basis matrix. component vector The geometric mean; T represents the transpose of the matrix.

[0075] To preserve the symbiotic relationships (causal logic) between elements, noise generation must adhere to the inherent geostatistical characteristics of the data. Calculation Covariance matrix of spatial data : in, This is the covariance matrix of the ilr spatial data, used to capture the co-occurrence relationships between elements (such as element Zn and element Pb). For spatial vectors The mean of; T represents the transpose.

[0076] ; The randomized values ​​have a mean of 0 and a covariance of . The noise vector; Noise intensity coefficient; It follows a multivariate Gaussian distribution. Specifically, the generated noise vector... Follows a multivariate Gaussian distribution And superimposed on .

[0077] In this way, the data with added noise, after being restored back to the ppm space, still retains the key indicator element ratio logic reflecting low-temperature hydrothermal mineralization, such as F / Ca and Sr / Ba. This ensures that the geochemical zoning conclusions derived by the AI ​​model are consistent with the original data. The ppm space is the original chemical composition concentration space in parts per million, which is subject to the mathematical constraints of the composition data. The noise is generated in the unconstrained ILR space according to the covariance structure of the original data and then restored back to the ppm space. This increases the diversity of the data while preserving the geochemical logic reflecting mineralization.

[0078] To address the weak anomaly characteristics of elements such as Hg and Sb in southeastern Chongqing, the system introduces a gradient protection operator. At each spatial point, the spatial gradient vector of the key indicator element is calculated. When injecting noise, the perturbation of the noise component in the gradient direction is minimized. , : Noise component; (The spatial gradient vector of the key indicator element). By suppressing noise components along the gradient direction, the system perfectly preserves the shape and diffusion trend of the abnormal halo while significantly perturbing the background value, ensuring that the core logic of "mineral exploration direction" is not lost. In one embodiment of this application, the image-text collaborative desensitization of spatial deformation data includes the following steps: Scan the multi-source heterogeneous geological data and extract spatial relationship triples, which include the orientation, distance and target of the multi-source heterogeneous geological data; The spatial relationship triples are updated based on the spatial deformation data, and the updated spatial relationship triples are backfilled into the desensitized data using a natural language generation model.

[0079] In this implementation, the geological report and geological map are tightly coupled. The report is filled with numerous descriptions of "relative location," such as: "The ore body is located on the northwest wing of the Qiyaoshan anticline, approximately 200 meters from the axis." If the place name is simply deleted, leaving only the distance description, an attacker can easily reconstruct the specific location using the un-anonymized distance information combined with a publicly available geological map. If the map is deformed but the text remains unchanged, a logical contradiction will arise: "The map shows a distance of 2km, while the report shows 200m," exposing the anonymization process. This application addresses this by designing a central synchronization engine with a global displacement field. This is the core mechanism that drives the real-time rewriting of text parameters.

[0080] In this embodiment, the central synchronization engine includes: Using NLP technology to automatically scan geological reports and extract the "azimuth-distance-target" triple, such as... and location description .

[0081] Real-time invocation of the displacement field generated in stage two Based on the new position after deformation Recalculate the geometric distance between the two points. and azimuth ,in The coordinates of a point on the anticline axis after deformation; in, The new distance after deformation; The new azimuth angle after deformation; Displacement field The transformed anticline axis and the new coordinates of the ore body.

[0082] Because PC-GSTN is anisotropic deformation, It may no longer be 200m, but 238m; "northwest wing" may become "20 degrees north of west".

[0083] Using a Natural Language Generation (NLG) model, the updated numerical values ​​are... and direction The data is then backfilled into the text template to generate a logically consistent new report. This WYSIWYG real-time correction ensures a perfect closed loop in the internal logic of the data.

[0084] S106: Perform Jacobi matrix-driven vector orientation correction and raster data inverse mapping alignment on the desensitized data and evaluate it.

[0085] Reference Figure 1 and Figure 2 In this implementation, as the final output stage of the process, it is necessary to ensure the physical consistency of all data in three-dimensional space and to conduct rigorous quality verification. Geological data contains a large amount of vector direction data, most typically borehole trajectories and stratigraphic attitudes. When nonlinear distortions occur in space (such as rotation + shear), if the borehole trajectory maintains its original angle, its contact relationship with the surrounding strata will be incorrect (for example, originally perpendicular to the strata, it becomes oblique after deformation, leading to an error in the calculated apparent thickness).

[0086] This system not only translates the orifice coordinates, but also utilizes the displacement field. Jacobian matrix at each depth point ( The inclination angle of the borehole and azimuth Perform local rotation correction.

[0087] Let the tangent vector of the borehole at a certain depth be... The corrected vector for: in, Corrected borehole or formation attitude vector (unit vector); : Original vector; Displacement field At point The Jacobian matrix at the location (describes local rotation and shearing). This operator ensures that vector elements rotate and deform along with the spatial medium, thus maintaining the relative angle between the borehole and the formation interface. This is crucial for whether the anonymized data can be directly imported into 3D geological modeling software for modeling.

[0088] For geophysical images or raster geological maps, the system adopts a method based on... The inverse mapping resampling algorithm ensures holographic visual alignment between the deformed raster pixels and the vector geological boundary. First, a blank new image (desensitized image) is created, and then for each pixel in the new image... Calculate its corresponding position in the original image: in, : Coordinates in the original image; Inverse transformation of the displacement field; : Pixel coordinates of the new image after desensitization.

[0089] Pick The color value at the specified location is filled with a new pixel. This ensures that the geological map and geophysical anomaly map remain clear and continuous after deformation, and completely overlap with the vector layer.

[0090] For mineral points (points), faults (lines), and stratigraphic extents (polygons) on geological maps, directly substitute the original coordinates. Function (global displacement transformation function). For any point ,calculate: : Represents the coordinate vector of any point in the geological body before deformation (original state).

[0091] : Represents the new coordinate vector of the same point after deformation, which is obtained by substituting the original coordinates into the displacement transformation function.

[0092] The fault lines bend along with the grid, but because It is a continuous function, the fault line will not break, and the topological adjacency relationship between strata remains unchanged.

[0093] To verify the desensitization effect, the system established a set of quantitative evaluation indicators (KPIs): Topological consistency: Calculate the Euler number characteristics, compare the number of holes and the number of connected components in the binarized constructed graph, and require... (Strict consistency), where ΔE is the difference in Euler number between the original geological map and the desensitized geological map. A difference in Euler number of 0 indicates topological consistency, with no new holes or fractures.

[0094] Causal logic: Calculate the Kullback-Leibler divergence of the distribution of element ratios (e.g., Sr / Ba), requiring... To ensure that geochemical data (geochemical data), even after noise has been added, still reflects the original mineralization patterns. Among these, The probability distribution of the original data (e.g., the histogram distribution of the original Sr / Ba ratio). : Probability distribution of data after anonymization.

[0095] Construction style: Calculate the rate of change of the aspect ratio of the folds to ensure it is within the allowable range.

[0096] Usability: Train a standard CNN model for mineral exploration prediction on both the original and anonymized images, requiring an accuracy loss. If the accuracy loss exceeds 3%, it indicates that too much key information was destroyed during the desensitization process (such as the spatial relationship between ore bodies and faults), which may cause artificial intelligence to fail to obtain key information.

[0097] The accuracy obtained by training and testing on the original data; The accuracy obtained after training and testing on anonymized data using the same model structure; If any indicator exceeds the limit, the system will automatically trigger parameter backtracking, adjust the stiffness matrix parameters of the finite element mesh or the noise level, and perform iterative optimization until the logic fidelity requirement is met.

[0098] This application also discloses a multi-source heterogeneous geological data desensitization device, which adopts the following technical solution: A multi-source heterogeneous geological data desensitization device, employing the multi-source heterogeneous geological data desensitization method described above, includes: The data acquisition module is used to collect multi-source heterogeneous geological data; The data recognition module is used to perform named entity recognition based on the input multi-source heterogeneous geological data and extract geological entity data from the multi-source heterogeneous geological data. The data reasoning module is used to link the geological entity data and the pre-input geological knowledge graph to obtain reasoning data; The data sensitivity grading and evaluation module is used to evaluate the inference data through a dynamic sensitivity grading and evaluation model, obtain the sensitivity grading of the inference data with different sensitivities, and determine the sensitive targets. The data reshaping module is used to geometrically reshape the sensitive target, simulate the construction motion of the sensitive target, constrain and prevent topological flipping through Jacobian determinant, construct Lagrange locking of key points, and obtain spatial deformation data. The data desensitization module performs attribute desensitization and image-text collaborative desensitization on the spatial deformation data to obtain desensitized data.

[0099] The multi-source heterogeneous geological data desensitization device of this application embodiment can realize any of the above-mentioned methods for desensitizing multi-source heterogeneous geological data, and the specific working process of each module in the multi-source heterogeneous geological data desensitization device can be referred to the corresponding process in the above-mentioned method embodiment.

[0100] In the several embodiments provided in this application, it should be understood that the provided methods and apparatus can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a component is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple components can be combined or integrated into another system, or some features can be ignored or not executed.

[0101] This application also discloses a computer device.

[0102] Computer equipment, including memory, processor, and computer program stored in memory and executable on the processor, wherein the processor executes the computer program to implement the multi-source heterogeneous geological data desensitization method described above.

[0103] This application also discloses a computer-readable storage medium.

[0104] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as any of the above-described methods for desensitizing multi-source heterogeneous geological data.

[0105] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0106] In summary, the multi-source heterogeneous geological data desensitization method and apparatus provided in this application differs from conventional keyword filtering. As shown in steps S101-S103, Named Entity Recognition (NER) technology, combined with geological ontology, can perform multi-level dynamic sensitivity rating of exploration data. Through Knowledge Graph (KG) technology, not only are coordinates extracted, but the physical properties of strata and structures are also deeply analyzed, providing "knowledge anchors" for subsequent semantic replacement. As shown in step S104, the geological space is treated as a non-uniform anisotropic elastic grid, and a complex nonlinear displacement field is solved using finite element method (FEM) simulation technology. While significantly deviating from the true location, this application forcibly introduces geological invariant constraints (such as fault intersection relationships and fold axis parallelism) to ensure that prospecting patterns such as anticlines and synclines still conform to the logical common sense of structural geology after geometric deformation. As shown in step S105, to address the "closure effect" of geochemical data, this application performs covariance-preserving noise addition in the equidistant logarithmic ratio (ilr) transformation space. In particular, a weak anomaly protection operator is introduced in the gradient domain, preserving the key anomaly halo morphology and element ratio logic of low-temperature hydrothermal deposits by perturbing the background value rather than the peak value. For the relative positioning description in the text report, this application establishes a closed loop linking the text and the spatial displacement field; by extracting spatial triples, the displacement field can be dynamically invoked. By recalculating the deformed distance and azimuth and automatically updating the text parameters, the risk of reverse positioning using logical contradictions between the text and graphics is fundamentally eliminated. For one-dimensional borehole data, this application not only implements coordinate translation but also uses the Jacobian matrix of the transformation field at various points in space to perform local rotation correction on the trajectory vector. This technology ensures that the relative cutting angle between the borehole and the inclined strata remains unchanged after large-scale spatial distortion, effectively supporting the geological modeling needs in complex three-dimensional scenes. Through the methods and apparatus in the embodiments of this application, the real geographical location and place names can be completely hidden, while ensuring that the anonymized virtual dataset maintains a high degree of logical consistency with the original data in four dimensions: "structural topology, stratigraphic semantics, element distribution, and three-dimensional attitude." This supports external institutions in directly using the anonymized data for high-quality artificial intelligence mineral exploration prediction and scientific research applications.

[0107] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0108] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for desensitizing multi-source heterogeneous geological data, characterized in that, Includes the following steps: Named entity recognition is performed on the input multi-source heterogeneous geological data to extract geological entity data from the multi-source heterogeneous geological data; By linking the geological entity data with the pre-input geological knowledge graph, inference data is obtained; The inference data is evaluated using a dynamic sensitivity grading assessment model to obtain sensitivity gradings for the inference data with different sensitivities, thereby identifying sensitive targets. The sensitive target is geometrically reshaped to simulate its construction motion. Constraints and topological anti-flipping are performed using Jacobian determinants to construct Lagrange locking of key points and obtain spatial deformation data. The spatial deformation data is subjected to attribute desensitization and image-text co-desensitization to obtain desensitized data.

2. The method for desensitizing multi-source heterogeneous geological data according to claim 1, characterized in that, Named entity recognition is performed on the input multi-source heterogeneous geological data, and the extraction of geological entity data from the multi-source heterogeneous geological data includes the following steps: The pre-trained layer of the deep learning architecture is adaptively fine-tuned based on the multi-source heterogeneous geological data, and the multi-source heterogeneous geological data is converted into fine-tuned data with high-dimensional contextual semantic representation and output to the encoding layer of the deep learning architecture. The encoding layer captures the long-distance semantic dependencies of the fine-tuning data through a bidirectional long short-term memory network, and outputs the long-distance semantic dependencies and the fine-tuning data to the decoding layer of the deep learning architecture; The decoding layer generates the label sequence dependency relationship of the fine-tuned data through a conditional random field, thereby obtaining geological entity data that conforms to geological grammar logic.

3. The method for desensitizing multi-source heterogeneous geological data according to claim 1, characterized in that, Linking the geological entity data and the pre-input geological knowledge graph includes the following steps: The geological entity data is described using the Web ontology language standard, and semantic reasoning is performed on the geological entity data to obtain the reasoning data. The reasoning data includes the geological entity names, as well as the sequence relationships, lithological characteristics, and ore-controlling attributes between the geological entity names.

4. The method for desensitizing multi-source heterogeneous geological data according to claim 1, characterized in that, The dynamic sensitivity grading assessment model includes a sensitivity scoring function, which is used to evaluate the inherent attributes of geological entity data in the inference data, as well as the connectivity between the inherent attributes and the geological knowledge graph. The inference data is graded by the sensitivity scoring function to obtain several levels of sensitive targets.

5. The method for desensitizing multi-source heterogeneous geological data according to any one of claims 1-4, characterized in that, The simulated motion of the sensitive target specifically includes the following steps: The sensitive target is discretized into several tiny units using tetrahedral or triangular meshes; The mesh is given a fiber-reinforced hyperelastic constitutive model, the construction direction vector is set as the construction principal axis, and the strain energy density function is implanted into each micro-unit. The strain energy density function is the sum of the matrix term and the anisotropy term. Anisotropic constraints are applied to the micro-units to lock their structural pattern; Boundary constraints and virtual body forces are applied to the mesh, and the nonlinear equilibrium equations of the micro-units are solved to obtain the simulation results of the construction motion; The constraint and topology anti-flipping method using Jacobian determinant includes: In the simulation of the constructed motion, a penalty term based on the Jacobian determinant is introduced. When the Jacobian determinant approaches 0, the penalty term approaches infinity. The Lagrange locking is used to apply hard constraints to the structural key points and preserve the ore-controlling structural relationships of the structural key points.

6. The method for desensitizing multi-source heterogeneous geological data according to claim 5, characterized in that, The attribute desensitization of the spatial deformation data includes the following steps: The spatial deformation data is characterized by generative adversarial network, virtual stratigraphic identifiers are synthesized based on the spatial deformation data, and the virtual stratigraphic identifiers are re-encapsulated into standardized geological metadata and injected into the desensitized database. The spatial deformation data is subjected to covariance noise addition by transforming the space by equal interval logarithmic ratio, and a gradient protection operator is introduced for weak anomaly features in the spatial deformation data to obtain attribute-desensitized spatial deformation data.

7. The method for desensitizing multi-source heterogeneous geological data according to claim 6, characterized in that, The method of image-text collaborative desensitization of the spatial deformation data includes the following steps: Scan the multi-source heterogeneous geological data and extract spatial relationship triples, which include the orientation, distance and target of the multi-source heterogeneous geological data; The spatial relationship triples are updated based on the spatial deformation data, and the updated spatial relationship triples are backfilled into the desensitized data using a natural language generation model.

8. The method for desensitizing multi-source heterogeneous geological data according to claim 7, characterized in that, It also includes performing Jacobi matrix-driven vector orientation correction and raster data inverse mapping alignment on the desensitized data to verify the holographic spatial alignment and consistency of the desensitized data.

9. The method for desensitizing multi-source heterogeneous geological data according to claim 8, characterized in that, It also includes evaluating the de-identified data and establishing quantitative evaluation indicators, which include the topological consistency, causal logic, construction style, and usability of the de-identified data; Determine whether the de-identified data meets all the criteria in the quantitative evaluation indicators; If all items in the quantitative evaluation indicators are satisfied, the de-identified data will be output; If any of the quantitative evaluation indicators are not met, the stiffness matrix of the mesh and / or the noise level are adjusted iteratively until all of the quantitative evaluation indicators are met, and the desensitized data is output.

10. A desensitization device for multi-source heterogeneous geological data, employing the desensitization method for multi-source heterogeneous geological data as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous geological data; The data recognition module is used to perform named entity recognition based on the input multi-source heterogeneous geological data and extract geological entity data from the multi-source heterogeneous geological data. The data reasoning module is used to link the geological entity data and the pre-input geological knowledge graph to obtain reasoning data; The data sensitivity grading and evaluation module is used to evaluate the inference data through a dynamic sensitivity grading and evaluation model, obtain the sensitivity grading of the inference data with different sensitivities, and determine the sensitive targets. The data reshaping module is used to geometrically reshape the sensitive target, simulate the construction motion of the sensitive target, constrain and prevent topological flipping through Jacobian determinant, construct Lagrange locking of key points, and obtain spatial deformation data. The data desensitization module performs attribute desensitization and image-text collaborative desensitization on the spatial deformation data to obtain desensitized data.

Images