Differential privacy de-sensitization processing method for high-sensitive public data

Abstract

This application relates to the field of privacy desensitization technology, specifically a differential privacy desensitization method for highly sensitive public data. The method includes: parsing query requests to establish node indexes and real data; constructing a hierarchical summation constraint matrix; and deriving a privacy theoretical degree based on a permission-based privacy budget to generate a weight matrix; extracting historical desensitized data to calculate the cluster-level data centrality; dynamically deriving the noise truncation threshold for each node based on the privacy theoretical degree; generating initial floating-point noise using the inverse Laplace function; constructing a secure floating-point noise value after threshold pruning; and superimposing the noise with real data to generate preliminary desensitized data; constructing a constrained objective function based on the weight matrix; optimizing and reconstructing the preliminary desensitized data; and outputting the final desensitized data. This application solves the problems of "inconsistent summation" and "prone to negative numbers" in multi-level data differential privacy desensitization, improving the reliability and efficiency of differential privacy desensitization.

Description

Technical Field

[0001] This application relates to the field of privacy desensitization technology, specifically to a differential privacy desensitization method for highly sensitive public data. Background Technology

[0002] In smart city and government data sharing scenarios, government service platforms frequently need to issue multi-level statistical reports to authorized external operating agencies (e.g., simultaneously displaying nested data on the city's total population, the population of each district, and the population of each street). Differential privacy mechanisms, by injecting random error distributions into precise statistical results to mask the true information of individuals, are currently a common technical means to prevent reverse engineering of data.

[0003] Faced with the rigid rule that multi-level government data must meet the requirement of consistent summation between different levels, the independent noise addition mode of conventional differential privacy directly disrupts this equation, causing front-end verification errors. Existing technologies often use post-hoc proportional allocation to repair this conflict, but this method is extremely prone to causing devastating damage to low-base-number underlying nodes: when negative noise is superimposed on the underlying grid or when it is forcibly allocated excessive amounts, its value is very likely to drop below zero. This counterintuitive negative desensitization result will directly lead to the complete collapse of the logic on the data consumption side, thereby reducing the reliability and efficiency of differential privacy desensitization. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a differential privacy desensitization method for highly sensitive public data, thereby resolving the existing issues.

[0005] The differential privacy desensitization method for highly sensitive public data in this application adopts the following technical solution: One embodiment of this application provides a differential privacy de-identification method for highly sensitive public data, the method comprising the following steps: The system counts the de-identified query requests sent by the current caller to establish the node index and obtain the real business data of each node; it constructs a hierarchical summation constraint matrix based on the parent-child logical inclusion relationship between nodes; it extracts the caller's role identification code, matches it to obtain the global query privacy budget to determine the local privacy budget of each node, and combines it with the preset global sensitivity to determine the privacy theoretical degree in order to construct a weight matrix. Extract historical de-identified business data from each node and calculate the intra-cluster data centrality of each node using the hierarchical summation constraint matrix; determine the noise boundary truncation threshold for each node based on the privacy theory degree and the intra-cluster data centrality. For each node, a floating-point number is generated, and combined with a local privacy budget, the initial floating-point noise value of each node is generated using the Laplace inverse cumulative distribution function. The initial floating-point noise value is compared with the noise boundary truncation threshold to construct a secure floating-point noise value for each node, which is used to de-identify real business data to obtain preliminary de-identified data. An objective function is constructed using a weight matrix and secure floating-point noise values ​​to repair the initial de-identified data and obtain the final de-identified data.

[0006] Preferably, the process of constructing the hierarchical summation constraint matrix is ​​as follows: Establish an initial zero matrix of size M×N, where M is the total number of summation equations and N is the total number of nodes; For the j-th summary equation, extract the direct child nodes under its corresponding parent node p; In the j-th row of the initial zero matrix, set the element with column index p to 1, set the element at the direct child node corresponding to the parent node p to -1, and keep the rest of the elements as 0; traverse all summation equations and use the generated matrix as the hierarchical summation constraint matrix.

[0007] Preferably, the method for determining the local privacy budget of each node is as follows: The global query privacy budget, obtained by matching the caller role identifier, is evenly distributed among all nodes, and the distribution result is used as the local privacy budget.

[0008] Preferably, the process of constructing the weight matrix is ​​as follows: The reciprocal of the square of the privacy theoretical degree of each node is used as the data modification weight of each node; Construct an N×N dimensional diagonal matrix, where N represents the total number of nodes. Set the elements on the diagonal of the diagonal matrix to the data modification weights of the corresponding nodes, and denote the resulting matrix as the weight matrix.

[0009] Preferably, the calculation process for the cluster-wide data centrality is as follows: Obtain historical business count values ​​from the historical anonymized business data of each node; A node includes a parent node and its direct child nodes. Under any parent node, the sum of the historical business count value of the parent node and the preset factor is calculated, and the ratio of the historical business count value of each direct child node to the corresponding sum value is calculated and recorded as the historical proportion of each direct child node under any parent node. The preset factor is a preset constant greater than 0. Calculate the sum of the squared historical proportions of all direct child nodes under any parent node, and use it as the intra-cluster data centrality of any parent node and all its direct child nodes; traverse all parent nodes and their direct child nodes to obtain the intra-cluster data centrality of each node.

[0010] Preferably, the expression for the noise boundary truncation threshold of each node is: In the formula, The noise boundary truncation threshold for node i; This represents the privacy-theoretical degree of node i; This represents the cluster centrality of node i. represents the preset failure tolerance probability limit value; ln[ ] represents the logarithmic function with the natural constant as the base.

[0011] Preferably, the process for generating the initial floating-point noise values ​​for each node is as follows: Substitute the floating-point number and privacy theory degree of each node into the Laplace inverse cumulative distribution function, and denote the resulting function value as the initial floating-point noise value of each node.

[0012] Preferably, the process of constructing the secure floating-point noise value for each node is as follows: If the absolute value of the initial floating-point noise value of node i is greater than the noise boundary truncation threshold, then the sign direction of the initial floating-point noise value of node i is retained, and its absolute value is reset to the noise boundary truncation threshold as the safe floating-point noise value. Otherwise, the initial floating-point noise value of node i is used as the safe floating-point noise value.

[0013] Preferably, the preliminary de-identified data is the result of adding the real business data to the secure floating-point noise value.

[0014] Preferably, the objective function The expression is: In the formula, Represents the column vector to be determined; A vector representing the initial de-identified data of all nodes; Represents the weight matrix; represents the transpose function; min[ ] represents the minimum value function.

[0015] This application has at least the following beneficial effects: This application constructs a hierarchical constraint matrix representing the summation equation of business by analyzing the single-root tree topology, and derives the privacy theoretical degree of each node by combining the local privacy budget of permission authentication allocation. Then, a weight matrix is ​​constructed based on the reciprocal square of the theoretical degree. By using the mathematical reciprocal relationship, the modification cost of macroscopic large cardinality nodes is amplified exponentially. This serves as a penalty mechanism to predefine the "avoidance red line" for subsequent reconstruction optimization. Thus, without exposing real sensitive data, the algorithm's underlying logic forces the solver to transfer the correction error that must be consumed later to microscopic small cardinality nodes, thereby maximizing the credibility of the macroscopic statistical data. Furthermore, this application extracts secure and leak-free historical desensitized data from the previous cycle, combines it with a hierarchical summation constraint matrix to calculate the cluster data centrality that objectively reflects the vulnerability of each node's topological branch imbalance, and dynamically derives a specific noise boundary truncation threshold in conjunction with privacy theory. It also establishes a precise "risk countermeasure" mechanism using a logarithmic probability mapping formula, which can automatically narrow the error survival space of dangerous nodes in extremely unbalanced branches or high-noise states, directly eliminating extreme negative noise that may cause global distortion in subsequent reconstruction at the generation stage, effectively avoiding the phenomenon of robbing Peter to pay Paul caused by forcibly repairing huge negative values. Furthermore, this application avoids logarithmic overflow by adding underlying floating-point extremum clamp protection and accurately generates continuous floating-point noise conforming to the Laplace distribution using the inverse cumulative distribution function. Then, it performs absolute boundary pruning based on the pre-derived truncation threshold and superimposes it onto the real business data. This not only completely eliminates the system-level deadlock risk from the engineering execution level and ensures the absolute robustness of the noise addition process, but also provides a differentiable and derivative-based basic data environment for subsequent consistency reconstruction calculations of massive nodes by shifting the data form from the discrete integer domain to the continuous floating-point domain. This avoids the exponential computing power explosion problem caused by discrete combinatorial optimization and greatly improves the computation response time. Ultimately, this application embeds the weight matrix as a penalty coefficient into the objective function and constructs a quadratic programming model by combining hierarchical summation equality constraints and non-negative lower bound constraints. After performing low-latency gradient optimization in the continuous real number domain, it performs discretization and rounding output, cleverly transferring the cost of fixing account conflicts precisely to micro-nodes with extremely low weights. This restores the self-consistent logic of multi-level data summation at the algebraic level. At the same time, with the non-negative hard constraint, it eliminates the possibility of negative number anomalies displayed on the front end. Finally, without touching the bottom line of the credibility of macro data, it outputs a final de-identified report that balances absolute business compliance and high-strength privacy protection, improving the reliability and efficiency of differential privacy de-identification. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the steps of a differential privacy de-identification method for highly sensitive public data provided in one embodiment of this application; Figure 2 This is a schematic diagram of a secure floating-point noise value extraction process provided in one embodiment of this application. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by this application to achieve the intended purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the differential privacy desensitization processing method for highly sensitive public data proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] The following description, in conjunction with the accompanying drawings, details the specific scheme of the differential privacy desensitization method for highly sensitive public data provided in this application.

[0021] This application provides a differential privacy de-identification method for highly sensitive public data in one embodiment. Specifically, it provides the following differential privacy de-identification method for highly sensitive public data. Please refer to [link / reference]. Figure 1 The method includes the following steps: Step S1: Count the de-identified query requests sent by the current caller to establish the node index and obtain the real business data of each node; construct a hierarchical summation constraint matrix based on the parent-child logical inclusion relationship between nodes; extract the role identification code of the caller, match and obtain the global query privacy budget to determine the local privacy budget of each node, and combine it with the preset global sensitivity to determine the privacy theoretical degree in order to construct a weight matrix.

[0022] In multi-level government data sharing and statistical query scenarios, there is a core data contradiction: on the one hand, objective business logic requires that multi-level node data must strictly meet the bottom-up summation consistency (e.g., the sum of data from each street must be absolutely equal to the district-level total); on the other hand, in order to prevent the leakage of individual privacy, a differential privacy mechanism must be introduced to anonymize the query results, but the inevitable cost of adding random noise is to destroy the original accuracy of the data.

[0023] If a conventional "independent noise addition" strategy is adopted, the random errors introduced by each node at each level will accumulate and superimpose when aggregated upwards, resulting in a significant "discrepancy" between the city-level macro data and the district / county-level micro data, thus losing statistical usability. To resolve this contradiction, this embodiment abandons isolated noise addition to nodes and instead adopts a "global collaborative reconstruction" strategy. Before injecting noise, the de-identified query requests sent by the current caller are first counted to establish node indexes and obtain the real business data of each node; a hierarchical summation constraint matrix is ​​constructed based on the parent-child logical inclusion relationship between nodes; the caller's role identifier code is extracted, and the global query privacy budget is matched to determine the local privacy budget of each node. Combined with a preset global sensitivity, the privacy theoretical degree is determined to construct a weight matrix. This matrix guides subsequent optimization calculations on "which nodes should be modified first and which nodes should be modified less," thereby maximizing the credibility of macro data while meeting privacy and security red lines. The specific process is as follows: S101: Count the de-identified query requests sent by the current caller to establish the node index and obtain the real business data of each node.

[0024] In multi-level government affairs query scenarios, the business summary count value of a high-level node (e.g., city-level) must objectively be equal to the sum of the business count values ​​of all its directly subordinate low-level nodes (e.g., district-level nodes within its jurisdiction). Specifically, to prevent logical deadlocks caused by multiple inheritance when processing node inclusion relationships in subsequent algorithms, this embodiment first enforces a mandatory precondition: the business entities involved in the intercepted multi-level statistical query requests must conform to a single-rooted tree topology, ensuring that any subsequent low-level node belongs to at most one directly summarizing parent node.

[0025] After intercepting the de-identified query request from the caller, a regular abstract syntax parser is invoked to perform field parsing on the query statement of the de-identified query request. Through parsing, all independent business entities contained in the query statement at different spatial or temporal aggregation levels are extracted, and the total number of independent business entities is counted. Subsequently, consecutive node indexes are assigned to these independent business entities, and the actual business data of the nodes is obtained, where the value of the node index is... .

[0026] S102: Construct a hierarchical summation constraint matrix based on the parent-child logical inclusion relationship between nodes.

[0027] The system queries and extracts the direct hierarchical summary relationships between independent business entities in the pre-defined data asset hierarchy directory. To clarify the mathematical boundaries of the matrix construction, the tree-like directory is traversed according to the following pre-defined rules to obtain valid summary equations: For node i, it is checked whether it has at least one direct subordinate child node; if so, it is determined that node i and all its subordinate direct child nodes objectively constitute an independent summation constraint unit, and a valid summary equation of the form "parent node business value = ∑ direct child node business value" is generated accordingly; if node i is a terminal leaf node without subordinate child nodes, or its subordinate nodes are not included in all current nodes, it is skipped. After the traversal is completed, the total number M of summary equations actually contained in the current de-identified query request is counted. Subsequently, an equation index j is established, whose value increases from 1 to M, and a zero matrix of size M×N with all internal elements initially set to 0 is constructed as the initial hierarchical summation constraint matrix.

[0028] For the j-th summary equation, locate its corresponding parent node and all its direct child nodes. Based on the order of the nodes, directly locate the node index p of the parent node, and extract the node indices corresponding to all its direct child nodes to form an index set K.

[0029] In the initial hierarchical summation constraint matrix, the first... In the row, first label the column as The values ​​of the parent node matrix elements are set to Set the elements at the direct child nodes corresponding to the parent node p to -1, and keep the rest of the elements at 0; traverse all summation equations and use the generated matrix as the hierarchical summation constraint matrix.

[0030] S103: Extract the caller's role identifier code, match it to obtain the global query privacy budget to determine the local privacy budget of each node, and combine it with the preset global sensitivity to determine the privacy theoretical degree in order to construct a weight matrix.

[0031] The system parses the current caller's identity information from the authentication credentials (such as a session token) carried in the anonymized query request sent from the front end, and extracts the role identifier code bound to the caller based on a standard permission authentication mechanism. Then, using this role identifier code as the index key, it performs a lookup in a pre-defined permission-privacy mapping table to obtain the total global query privacy budget authorized for the caller. During the budget allocation phase, an equal-distribution strategy is adopted, dividing the total global query privacy budget directly by the total number of nodes N to allocate an equal local privacy budget to each node.

[0032] Furthermore, the generation scale of differential privacy noise needs to be determined in conjunction with a preset global sensitivity Δf. For the single-individual counting and statistics scenario in government affairs (such as querying the actual population of a jurisdiction) addressed in this embodiment, since the absolute value of the maximum impact of adding, deleting, or changing a single individual record on the final statistical count is 1, the preset global sensitivity Δf is directly set to 1.

[0033] Then, the preset global sensitivity will be used. Substituting the local privacy budget of each node into the basic calculation formula of the Laplace mechanism, the theoretical degree of privacy for each node is calculated. The specific formula is as follows: Privacy Theoretical Degree of Node i The expression is: In the formula, Represents the local privacy budget of node i; This indicates the preset global sensitivity.

[0034] The calculation logic of this formula reflects an inverse proportional relationship: due to the characteristics of the division operator, the local privacy budget is derived based on role-based access control. The smaller the value (representing a lower level of access control for the caller and less information that can be disclosed), the higher the calculated privacy theoretical level. The larger the value, the wider the range of random error fluctuations that need to be injected into the node in subsequent steps in order to conceal the true individual information.

[0035] Because the statistical base of business data varies by magnitude across different levels of nodes, if no differentiated restrictions are imposed on the modification magnitude of each node during subsequent data reconstruction based on the hierarchical summation constraint matrix to satisfy the equality constraints, the algorithm, in minimizing the overall error, will distribute the adjustment amount equally among all nodes without directionality. This indiscriminate error distribution, once applied to nodes with small base numbers, can easily lead to relative error distortion, causing complete distortion of local business data. To overcome this solution bias, a cost penalty mechanism is introduced. Based on privacy theory, an independent data modification weight is assigned to each node to construct a weight matrix. Specifically: the data modification weight of node i... The expression is: In the formula, This represents the privacy-theoretical degree of node i.

[0036] The engineering design logic of this formula lies in using the reciprocal relationship to establish the cost of numerical modification: when the privacy theoretical degree corresponding to a certain node... The smaller the value (typically representing a large cardinal number node requiring higher reliability of macroeconomic data), the more significantly the inverse of its squared term will be amplified, leading to a change in the weighting of the resulting data. Increase in absolute value.

[0037] Furthermore, construct an N×N dimensional diagonal matrix, where N represents the total number of nodes. Set the elements on the diagonal of the diagonal matrix as the data modification weights of the corresponding nodes, and denote the resulting matrix as the weight matrix.

[0038] Based on the weight matrix, it can be understood that it is essentially a set of penalty coefficients used to quantify the cost of data modification at each node, reflecting the anti-disturbance priority of nodes at different levels in the subsequent algebraic reconstruction optimization process. The calculation of this weight matrix is ​​only affected by the privacy theory degree: the smaller the privacy theory degree of a node (usually indicating that the node has a sufficient privacy budget and a large business base, belonging to the macro-statistical node), the larger its corresponding weight value in the weight matrix, which reflects that the corresponding node has a very high data credibility weight. Its direct impact is that in the iterative calculation of minimizing the objective function, any small modification deviation caused to this high-weight node will be amplified into a huge penalty increment by algebra, thus forcing the solution algorithm to avoid adjusting this type of node. Conversely, the larger the privacy theory degree of a node (usually indicating that the node has a small business base, belonging to the micro-level detail node), the smaller its corresponding weight value, which reflects that the corresponding node has a wider allowable range of change in the reconstruction process. Its impact is that when the solution algorithm must consume errors to satisfy the equality constraints, it will automatically follow the cost minimization principle and preferentially distribute the correction amount to these low-weight nodes.

[0039] Thus, this embodiment constructs a hierarchical constraint matrix representing the summation equation of business by analyzing the single-root tree topology, and derives the privacy theoretical degree of each node by combining the local privacy budget of permission authentication allocation. Then, a weight matrix is ​​constructed based on the reciprocal square of this theoretical degree. By using the mathematical reciprocal relationship, the modification cost of macroscopic large cardinality nodes is amplified exponentially, which serves as a penalty mechanism to predefine the "avoidance red line" for subsequent reconstruction optimization. In this way, without exposing real sensitive data, the algorithm's underlying logic forces the solver to transfer the correction error that must be consumed in the future to microscopic small cardinality nodes, thus maximizing the credibility of the macroscopic statistical data.

[0040] Step S2: Extract historical de-identified business data of each node, and calculate the intra-cluster data centrality of each node in combination with the hierarchical summation constraint matrix; determine the noise boundary truncation threshold of each node based on the privacy theory degree and the intra-cluster data centrality.

[0041] Although the objective function of the subsequent step S3 sets a non-negative hard constraint, if the initial noise is not truncated beforehand in the current noise addition stage, small-cardinality nodes are prone to generating extreme negative noise (e.g., a true value of 5, resulting in a random error of -100) that exceeds their own business cardinality under the Laplace distribution. If it is forcibly required to be no less than zero in subsequent planning, a huge compensation error will be generated, which will be forced to be distributed to the macroscopic parent nodes, resulting in serious distortion of the macroscopic statistical data. Therefore, this embodiment extracts the historical desensitized business data of each node and calculates the intra-cluster data concentration of each node in combination with the hierarchical summation constraint matrix; based on the privacy theory degree and the intra-cluster data concentration, the noise boundary truncation threshold of each node is determined, limiting the total amount of initial extreme error injection from the source and preventing global distortion caused by robbing Peter to pay Paul during subsequent algebraic optimization.

[0042] S201: Extract the historical de-identified business data of each node, and calculate the cluster data centrality of each node in combination with the hierarchical summation constraint matrix.

[0043] In differential privacy mechanisms, using current real business data to determine the noise cutoff boundary would cause the output boundary range to directly carry the distribution characteristics of the current real business data, thus mathematically violating the independence principle of differential privacy and causing privacy leaks. To completely eliminate this illegal mapping, historical anonymized business data from each node in the previous statistical period is extracted instead. Since this historical data is publicly available and does not consume the current privacy budget, its use to calculate the numerical concentration of the tree-like topology can objectively reflect the long-term distribution pattern of the business grid (e.g., the base number of remote streets is consistently low) while ensuring absolute security and compliance in the calculation process. The statistical period depends on the frequency of government business updates. For example, if it is a monthly report, the data for the previous month is released on the 1st of each month. If the current query is for real business data in March, then the previous statistical period is February.

[0044] Furthermore, since a node includes a parent node and its direct child nodes, under any parent node, the sum of the historical service count value of any parent node and the preset factor is calculated, and the ratio of the historical service count value of each direct child node to the corresponding sum is calculated and recorded as the historical proportion of each direct child node under any parent node. Here, the preset factor is a preset constant greater than 0. In this embodiment, the preset factor is set to 1, which is used to eliminate the division deadlock error caused by the program performing division operation when the historical service data of the parent node is zero. Under the premise of ensuring that the denominator is not 0 and does not excessively affect the calculation result, the implementer can also set it according to the specific situation. This embodiment does not impose any special restrictions. Furthermore, the sum of the historical percentage squares of all direct child nodes under any parent node is calculated as the intra-cluster data centrality of any parent node and all its direct child nodes; all parent nodes and their direct child nodes are traversed to obtain the intra-cluster data centrality of each node.

[0045] Specifically, after completing the initial calculations described above, conflict resolution needs to be performed on intermediate nodes with dual attributes in the multi-level tree structure. For intermediate nodes that are simultaneously a child node of a parent node and a parent node of a lower-level child node, the historical proportion inherited when the node is a parent node and the historical proportion calculated when the node is a lower-level parent node are extracted. The maximum value of these two values ​​is then selected as the final effective intra-cluster data centrality for the intermediate node. The engineering logic for selecting the maximum value is as follows: a higher centrality value indicates that the local topological branch in which the node is located is more unbalanced and more vulnerable. Using the maximum value ensures that the most stringent noise boundary constraints are applied from the most dangerous and vulnerable perspective when deriving the truncation threshold later, thereby completely eliminating the risk of negative overflow. For terminal leaf nodes without lower-level children or root nodes without upper-level children, the aforementioned calculation results are directly used. After the traversal and determination are completed, the final determined intra-cluster data centrality for each node is output.

[0046] Based on the intra-cluster data centrality, it can be understood that the calculated intra-cluster data centrality is strictly constrained numerically by... Within the dimensionless interval (the lower limit is 0 when all business volumes are 0, and the upper limit is 1 because a smoothing factor of 1 is added to the denominator), it essentially reflects the degree of balanced development of each branch in a "family". Its calculation is only affected by one factor, which is the sum of the squares of the proportion of each child node to the parent node. When this value is larger, it indicates that the data is more distorted, that is, the business distribution is extremely concentrated on a single child node, and the business base of the other child nodes is close to zero. This reflects that the business distribution in this area is extremely unbalanced. If any negative noise is introduced, their values ​​will immediately drop below zero and become negative. Conversely, when the data concentration within the cluster is smaller, it indicates that the business volume of each child node is relatively evenly distributed and the size of each node is similar. This reflects that the business development in this area is balanced. This means that these nodes are highly robust and have a strong tolerance for random noise. It is difficult for them to become negative due to the same degree of random noise.

[0047] Furthermore, since the random error generated by the standard Laplace distribution may produce infinitely large negative extrema, this embodiment determines the noise boundary truncation threshold for each node based on the privacy theory degree and the cluster data centrality. This noise boundary truncation threshold is used to apply a forced boundary clamp to the generated original Laplace noise, eliminating or forcibly truncating extreme noise values ​​that exceed the threshold. Specifically: In this embodiment, as one implementation method, the noise boundary truncation threshold of node i is... The expression is: In the formula, This represents the privacy-theoretical degree of node i; This represents the cluster centrality of node i. This represents the preset failure tolerance probability limit value. The value range is usually 100. to In this embodiment, the following settings are provided. In practical applications, as other implementation methods, implementers can also set their own methods according to specific circumstances. This embodiment does not impose any special restrictions; ln[] represents the logarithmic function with the natural constant as the base.

[0048] Based on the noise boundary truncation threshold, it can be understood that the algebraic terms... This describes a dynamic amplification regulator that characterizes the probability tolerance boundary. The engineering design principle of this term is reflected in a strict positive and negative feedback control logic: the more unbalanced the data distribution of the aggregation cluster to which a node belongs, and the more dangerous the small-cardinality, vulnerable nodes, the higher the concentration parameter... The larger the value, the more likely the combination term will be. The value increases synchronously. As the combined term increases, the argument part... The overall value increases; because the logarithmic function is in the interval The inner value is negative and monotonically increasing. As the argument increases, the negative result calculated by the logarithmic function moves closer to zero; that is, the absolute value of the negative number decreases. Finally, this decreasing absolute value is compared with the negative parameter at the beginning of the formula. Multiplying them together results in the final derived cutoff threshold. The positive constant becomes smaller in absolute value; through this coherent algebraic transmission chain, smaller boundary threshold constants are automatically calculated for those more vulnerable topological branches, thereby imposing stricter restrictions on extreme value generation and compressing the survival space for small cardinality nodes to generate extreme negative errors from the source.

[0049] Thus, this embodiment extracts secure, leak-free historical anonymized data from the previous cycle, calculates the cluster data centrality that objectively reflects the vulnerability of each node's topological branch imbalance by combining the hierarchical summation constraint matrix, and dynamically derives a specific noise boundary truncation threshold using privacy theory. A precise "risk countermeasure" mechanism is established using the logarithmic probability mapping formula, which can automatically narrow the error survival space of dangerous nodes in extremely unbalanced branches or high-noise states. This effectively eliminates extreme negative noise that could cause global distortion in subsequent reconstruction at the generation stage, avoiding the "robbing Peter to pay Paul" phenomenon caused by forcibly patching huge negative values.

[0050] Step S3: Generate floating-point numbers for each node, and combine them with the local privacy budget. Using the Laplace inverse cumulative distribution function, generate the initial floating-point noise value for each node. Compare the initial floating-point noise value with the noise boundary truncation threshold to construct the secure floating-point noise value for each node, which is used to de-identify real business data to obtain preliminary de-identified data.

[0051] When dealing with the consensus algebra reconstruction of massive government service nodes, performing combinatorial optimization in the discrete integer domain will inevitably lead to an exponential explosion of computing power, resulting in solution timeouts. To overcome this computational bottleneck, this embodiment generates floating-point numbers for each node and, combined with a local privacy budget, uses the Laplace inverse cumulative distribution function to generate initial floating-point noise values ​​for each node. The initial floating-point noise values ​​are compared with noise boundary truncation thresholds to construct secure floating-point noise values ​​for each node. These are used to de-identify real business data to obtain preliminary de-identified data, shifting the subsequent constraint solution process from the discrete domain to the continuous real number domain, thereby allowing for rapid convergence of the solution. The specific process is as follows: S301: Generate floating-point numbers for each node and, in conjunction with a local privacy budget, use the Laplace inverse cumulative distribution function to generate initial floating-point noise values ​​for each node.

[0052] For node i, an initial value range is generated using a pseudo-random number generator. Uniformly distributed random floating-point numbers within In standard mathematical mappings, this random number will be directly substituted into a function containing logarithms. In the expression. However, in actual computer programming environments, due to the extreme limitations of underlying floating-point precision, if the generated... Extremely close to the boundary value or This will result in the expression The calculation results are extremely close to At this point, the logarithmic function This will trigger an illegal mathematical overflow error, causing the entire noise-adding program to crash and deadlock.

[0053] To prevent logarithmic overflow caused by such extreme values, when generating random floating-point numbers... Subsequently, a boundary constraint protection logic was forcibly added. A very small floating-point truncation lower limit constant was preset. (In this embodiment, it is set) Then, all generated random floating-point numbers The value is forced to be clamped in Within the interval. Specifically, when generated Less than At that time, The value is reset to ;when Greater than At that time, Reset the value to After this clamping protection operation, a safe and compliant random floating-point number is output. This completely eliminates the system-level deadlock risks of division by zero and overflow from the source of the project.

[0054] Furthermore, receive the clamped random floating-point number And combined with privacy theory and noise boundary cutoff threshold This generates and trims floating-point noise with a lower bound on the absolute error boundary. This embodiment uses random floating-point numbers... Corresponding privacy theory degree Substituting all values ​​into the Laplace inverse cumulative distribution function, we can generate the initial floating-point noise value for node i. The specific calculation formula is as follows: ; In the formula, This represents the privacy-theoretical degree of node i; represents a floating-point number of a node; sgn() represents a sign function; ln() represents a logarithmic function with the natural constant as the base.

[0055] Based on the initial floating-point noise value, it can be understood that the combined term... It plays a dual role in controlling both direction and amplitude: its positive or negative attribute is extracted by the sign function, which determines the final initial floating-point noise value generated. Is it better to increase the positive error of the value or decrease the negative error of the value? And its absolute value? Substituted into the logarithmic argument combination term In China, due to The uniform randomness of the logarithmic combination ensures that the system can accurately simulate the peaked, heavy-tailed divergence characteristics of the Laplace probability distribution. The coefficients at the front of the formula... This serves as a scaling operator, ensuring that the overall spread of the error strictly adheres to privacy budget requirements.

[0056] S302: Compare the initial floating-point noise value with the noise boundary truncation threshold to construct the safe floating-point noise value of each node, which is used to desensitize real business data to obtain preliminary desensitized data.

[0057] Calculate the initial floating-point noise value Then, the initial floating-point noise value is compared with the noise boundary truncation threshold to construct the safe floating-point noise value for each node. If the absolute value of the initial floating-point noise value of node i is greater than the noise boundary truncation threshold, the sign direction of the initial floating-point noise value of node i is retained, and its absolute value is reset to the noise boundary truncation threshold as the safe floating-point noise value. Otherwise, the initial floating-point noise value of node i is used as the safe floating-point noise value.

[0058] Preferably, the schematic diagram of the secure floating-point noise value extraction process provided in this embodiment is as follows: Figure 2 As shown.

[0059] Furthermore, in this embodiment, the result of adding the real business data of each node to the secure floating-point noise value is used as the preliminary desensitized data. That is, through this addition operation, preliminary desensitized data that both masks the real information and limits the extreme negative amplitude is obtained.

[0060] Thus, this embodiment avoids logarithmic overflow by adding underlying floating-point extremum clamp protection and accurately generates continuous floating-point noise conforming to the Laplace distribution using the inverse cumulative distribution function. Then, after absolute boundary pruning based on the pre-derived truncation threshold, it is superimposed on the real business data. This not only completely eliminates the risk of system-level deadlock from the engineering execution level and ensures the absolute robustness of the noise addition process, but also provides a differentiable and derivative-based basic data environment for the subsequent consistency reconstruction calculation of massive nodes by shifting the data form from the discrete integer domain to the continuous floating-point domain. This avoids the exponential computing power explosion problem caused by discrete combinatorial optimization and greatly improves the computational response time.

[0061] Step S4: Construct an objective function using the weight matrix and secure floating-point noise values ​​to repair the initial de-identified data and obtain the final de-identified data.

[0062] Differential privacy noise addition inevitably disrupts the originally rigorous summation equation of government reports. Traditional desensitization methods typically use a post-hoc proportional allocation method to forcibly smooth out errors. This often forces the already small base of micro-nodes to bear the reduction, resulting in negative statistical results that defy common sense. Therefore, this embodiment uses a weight matrix and secure floating-point noise values ​​to construct an objective function to repair the initially desensitized data and obtain the final desensitized data. Specifically: In this embodiment, the objective function The expression is: In the formula, Represents the column vector to be determined; A vector representing the initial de-identified data of all nodes; Represents the weight matrix; represents the transpose function; min[ ] represents the minimum value function.

[0063] Based on the objective function, the weight matrix can be understood as a combination of terms within the objective function. Characterized the initial noisy data with privacy protection in order to recover the summation equation. The forced secondary offset modification, multiplied by its transpose, mathematically constitutes the sum of squares of the modifications at each node. Specifically, the formula multiplies the sum of squares by a weight matrix. As an amplification factor, this design exhibits a strong tendency towards directional optimization and a positive feedback pattern: weight matrix The values ​​on the main diagonal represent the modification cost penalty multiplier for the corresponding node; for macroscopically large cardinality nodes (i.e., nodes with large allocated privacy budgets and small privacy theoretical degrees), their weight matrix... The corresponding values ​​are relatively large, and when the algorithm attempts to modify these macroscopic nodes, the tiny differences... Once the result is squared and multiplied by a very large weight multiplier, it leads to an explosive increase in the overall objective function result. Since the underlying logic of the quadratic programming solver is to find the minimum of the objective function, this huge algebraic increase will force the optimization algorithm to exhibit risk-averse behavior—the algorithm will automatically avoid modifying the data of macroscopic nodes with extremely large weights, and instead prioritize distributing the necessary correction errors to the microscopic bottom-level nodes with smaller weights. Thus, mathematically, it is ensured that the statistical results of the large number of macroscopic nodes will not undergo serious data deviation during the reconstruction process.

[0064] Furthermore, this embodiment injects the hierarchical summation constraint matrix as a hard condition into the model and performs iterative optimization in the continuous real number domain. This is to prevent the solver from directly setting the constraint matrix in pursuit of minimizing the objective function. Therefore, insurmountable algebraic boundaries must be set for it. First, the hierarchical summation constraint matrix is ​​injected into the model as the first-level equality boundary condition, forcibly defining the column vector to be determined. The matrix product equation must be satisfied: In the formula, Represents the hierarchical summation constraint matrix; based on the hierarchical summation constraint matrix The construction rule, which expands to the following, is equivalent to: for every set of nodes containing parent-child relationships, the value of the parent node must be exactly equal to the sum of the values ​​of all its direct child nodes. This constraint completely repairs the structural damage caused by the independent addition of noise in step S3 at the algebraic level.

[0065] Furthermore, a second-level inequality lower bound boundary condition is injected to forcibly specify the unknown column vector. Each element to be searched contained within All must be greater than or equal to zero: This hard constraint of the inequality fundamentally eliminates any business anomalies that include negative population or negative traffic flow at the computational boundary.

[0066] After injecting the aforementioned dual-constraint boundary conditions, the well-known Interior Point Method is invoked as the core iterative solution tool. Within an absolutely continuous real number space, guided by the continuous reduction of the objective function result, multiple rounds of gradient descent optimization are performed. To prevent infinite loops, a threshold is set for the difference between the objective function calculation results generated by two consecutive iterations. The calculation is iterated continuously until the difference is less than a set stopping threshold, at which point the quadratic programming solution process is considered to have converged. After convergence, the unknown column vector that has reached its optimal stable state is extracted and output as the completed continuous floating-point reconstructed column vector. .

[0067] The continuous real number results after optimization are converted into integer format by the business side to generate compliant and anonymized reports that can be recognized and displayed on the front-end large screen.

[0068] Due to the continuous floating-point reconstruction column vector with floating-point decimal places It cannot be directly recognized and rendered by discrete integer chart components in the business front end (such as bar charts or tables that only support displaying integer numbers of people), and requires a well-known discretization alignment transformation to be performed on it.

[0069] Reconstructing the column vector for continuous floating-point points For each floating-point element internally, the built-in numerical conversion function is called sequentially to perform a rounding operation. Specifically, if the decimal part of the floating-point element is greater than or equal to... When the integer part is 1, add one to it and truncate the decimal part; when the decimal part is strictly less than 1, add one to it and truncate the decimal part. When the fractional part is discarded, the fractional part is discarded directly. Since the second constraint of the objective function has ensured that all values ​​are not less than zero, after the above carry-over and discard operations, all continuous floating-point values ​​are forcibly converted to non-negative integer types, and the extremely small truncation error caused by this conversion usually does not affect the judgment of macro business indicators.

[0070] After conversion, these non-negative integers are rearranged strictly according to the preset node index order to generate the final de-identified data vector. Finally, the de-identified data vector is encapsulated into a report according to the data communication protocol format requested by the front end (such as JSON or XML format) and sent to the data sharing terminal that initiated the query. Thus, the full-cycle de-identification process that takes into account individual privacy and anti-cracking, macro and micro data summation and self-consistency, and the requirement of absolute non-negativity of numerical values ​​is completed.

[0071] Thus, this embodiment cleverly transfers the cost of fixing accounting conflicts precisely to micro-nodes with extremely low weights by embedding the weight matrix as a penalty coefficient into the objective function and constructing a quadratic programming model in conjunction with hierarchical summation equality constraints and non-negative lower bound constraints. After performing low-latency gradient optimization in the continuous real number domain, discretization and rounding are performed to output the data. This restores the self-consistent logic of multi-level data summation at the algebraic level. At the same time, the non-negative hard constraint eliminates the possibility of negative number anomalies in the front-end display. Finally, without touching the bottom line of the credibility of macro data, a final de-identified report that balances absolute business compliance and high-strength privacy protection is output, improving the reliability and efficiency of differential privacy de-identification.

[0072] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this specification have been described above. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0073] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0074] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them; modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions of some of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for differential privacy de-sensitization processing of high-sensitive public data, characterized in that, The method includes the following steps: The system counts the de-identified query requests sent by the current caller to establish the node index and obtain the real business data of each node; it constructs a hierarchical summation constraint matrix based on the parent-child logical inclusion relationship between nodes; it extracts the caller's role identification code, matches it to obtain the global query privacy budget to determine the local privacy budget of each node, and combines it with the preset global sensitivity to determine the privacy theoretical degree in order to construct a weight matrix. Extract historical de-identified business data from each node and calculate the intra-cluster data centrality of each node using the hierarchical summation constraint matrix; determine the noise boundary truncation threshold for each node based on the privacy theory degree and the intra-cluster data centrality. For each node, a floating-point number is generated, and combined with a local privacy budget, the initial floating-point noise value of each node is generated using the Laplace inverse cumulative distribution function. The initial floating-point noise value is compared with the noise boundary truncation threshold to construct a secure floating-point noise value for each node, which is used to de-identify real business data to obtain preliminary de-identified data. A target function is constructed using a weight matrix and secure floating-point noise values ​​to repair the initial de-identified data and obtain the final de-identified data. The process of constructing the hierarchical summation constraint matrix is ​​as follows: Establish an initial zero matrix of size M×N, where M is the total number of summation equations and N is the total number of nodes; For the j-th summary equation, extract the direct child nodes under its corresponding parent node p; In the j-th row of the initial zero matrix, set the element with column index p to 1, set the element at the direct child node corresponding to the parent node p to -1, and keep the rest of the elements to 0; traverse all summation equations and use the generated matrix as the hierarchical summation constraint matrix; The calculation process for the centrality of the data within the cluster is as follows: Obtain historical business count values ​​from the historical anonymized business data of each node; A node includes a parent node and its direct child nodes. Under any parent node, the sum of the historical business count value of the parent node and the preset factor is calculated, and the ratio of the historical business count value of each direct child node to the corresponding sum value is calculated and recorded as the historical proportion of each direct child node under any parent node. The preset factor is a preset constant greater than 0. Calculate the sum of the squared historical proportions of all direct child nodes under any parent node, and use it as the intra-cluster data centrality of any parent node and all its direct child nodes; traverse all parent nodes and their direct child nodes to obtain the intra-cluster data centrality of each node.

2. The method of claim 1, wherein the method comprises: The method for determining the local privacy budget of each node is as follows: The global query privacy budget, obtained by matching the caller role identifier, is evenly distributed among all nodes, and the distribution result is used as the local privacy budget.

3. The method of claim 1, wherein the method comprises: The process of constructing the weight matrix is ​​as follows: The reciprocal of the square of the privacy theoretical degree of each node is used as the data modification weight of each node; Construct an N×N dimensional diagonal matrix, where N represents the total number of nodes. Set the elements on the diagonal of the diagonal matrix to the data modification weights of the corresponding nodes, and denote the resulting matrix as the weight matrix.

4. The differential privacy desensitization method for highly sensitive public data as described in claim 1, characterized in that, The expression for the noise boundary truncation threshold of each node is: In the formula, The noise boundary truncation threshold for node i; Represents the privacy-theoretical degree of node i; This represents the cluster centrality of node i. represents the preset failure tolerance probability limit value; ln[ ] represents the logarithmic function with the natural constant as the base.

5. The differential privacy desensitization method for highly sensitive public data as described in claim 1, characterized in that, The process for generating the initial floating-point noise values ​​for each node is as follows: Substitute the floating-point number and privacy theory degree of each node into the Laplace inverse cumulative distribution function, and denote the resulting function value as the initial floating-point noise value of each node.

6. The differential privacy desensitization method for highly sensitive public data as described in claim 1, characterized in that, The process of constructing the safe floating-point noise value for each node is as follows: If the absolute value of the initial floating-point noise value of node i is greater than the noise boundary truncation threshold, then the sign direction of the initial floating-point noise value of node i is retained, and its absolute value is reset to the noise boundary truncation threshold as the safe floating-point noise value. Otherwise, the initial floating-point noise value of node i is used as the safe floating-point noise value.

7. The differential privacy desensitization method for highly sensitive public data as described in claim 1, characterized in that, The initial de-identified data is the result of adding the real business data to the secure floating-point noise value.

8. The differential privacy desensitization method for highly sensitive public data as described in claim 1, characterized in that, objective function The expression is: In the formula, Represents the column vector to be determined; A vector representing the initial de-identified data of all nodes; Represents the weight matrix; Represents the transpose function; min[ ] represents the minimum value function.

Images