Financial data desensitization method based on edge computing

Abstract

The invention discloses a financial data desensitization method based on edge computing. The method comprises the following steps that remote certification is completed at an edge node, a desensitization strategy packet is loaded, and a trusted execution environment is initialized; financial data streams are collected and subjected to structured segmentation, and a field list is generated; inputting the field list into a double-layer hypersphere embedding network for coding, and outputting a spherical embedding vector and a hierarchical mapping relation; applying adaptive rotation disturbance in the hypersphere space according to the sensitivity difference to obtain a disturbed embedded vector and a disturbance parameter; the disturbance vector is input into a decoding unit to generate desensitization data, and a desensitization result is output through verification of a risk threshold value and a utility threshold value; generating a transformation spectrum and recording an audit log based on the chained hash; and transmitting a desensitization result according to a minimum necessary principle, and carrying out sealed storage or destruction on the original data. According to the method, data effectiveness is considered while data security is guaranteed, and credible, auditable and controllable domain outbound of high-sensitivity financial data is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of privacy computing and edge computing, and particularly relates to a financial data desensitization method based on edge computing. BACKGROUND

[0002] With the comprehensive online of financial business, a large amount of sensitive data related to personal identity information, account data and transaction behavior is continuously generated by terminal devices and business systems. The existing common techniques for financial data desensitization include field masking, hash operation, format-preserving encryption and differential privacy, which mostly rely on a central server for unified processing. Although such methods can reduce the risk of direct exposure of sensitive information to a certain extent, they still have the problems of link leakage risk and high processing delay due to the centralized transmission of data to the central node.

[0003] The existing methods generally use static rules for masking or noise addition, which is difficult to perform hierarchical processing on different sensitivity fields, resulting in excessive desensitization reducing data usability or insufficient desensitization causing privacy leakage. At the same time, the methods lack the ability of structured modeling of field relationships, making it difficult to maintain data distribution characteristics after desensitization. Most methods do not establish a complete process record and traceability mechanism, making it difficult to meet the requirements of compliance audit and responsibility definition of the financial industry.

[0004] Therefore, how to provide a financial data desensitization method based on edge computing is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a financial data desensitization method based on edge computing. The present application provides a technical solution that can complete data processing on site in an edge computing environment and achieve fine desensitization based on sensitivity grading, thereby reducing the risk of privacy leakage while maintaining data usability and compliance.

[0006] According to the financial data desensitization method based on edge computing, the method comprises the following steps: Completing remote attestation in the edge computing node, loading the desensitization policy package and initializing the trusted execution environment; Collecting the financial data stream and performing structured segmentation to obtain a field list; Inputting the field list into a double-layer hypersphere embedding network for encoding to output a spherical embedding vector and a mapping relationship between the field and the spherical level; Applying adaptive rotation perturbation based on sensitivity difference in the hypersphere space to obtain a perturbed embedding vector and corresponding perturbation parameters; Inputting the perturbed embedding vector into a decoding unit to generate desensitized data and output a desensitized result that passes the preset risk threshold and utility threshold check; The perturbation parameter is mapped to generate a transformation spectrum, and an audit log is formed based on a chain hash; The desensitization result is transmitted to the out-domain target according to the principle of minimum necessity, and the original financial data is stored or destroyed.

[0007] Further, the desensitization strategy package includes: field sensitivity grading rules, mapping boundaries of inner and outer spherical surfaces, inner and outer spherical surface radius configurations and minimum intervals, value intervals of rotation perturbation angle and rotation axis, risk threshold and utility threshold, and format preservation constraint templates, accompanied by version numbers and effective ranges, which take effect after signature verification and validity period verification in the trusted execution environment.

[0008] Further, the collected financial data stream includes transaction messages, operation logs and free texts, and the collection process is cached and uniformly formatted according to the timestamp sequence; The structured segmentation is based on transaction message templates, log field mapping tables and text parsing rules, and the financial data stream is parsed field by field to generate a field list containing field identification, field original value, context position information and sensitivity level.

[0009] Further, the generation process of the spherical embedding vector and the mapping relationship between the field and the spherical level includes: The initial vector is obtained by the type sub-coding unit and the numerical sub-coding unit according to the field category, wherein the type sub-coding unit is used to process character type or enumeration type fields, including account number, certificate category, and region code, and is mapped to the initial vector by one-hot encoding or embedded dictionary method; The numerical sub-coding unit is used to process numerical type fields, including amount, timestamp, and transaction frequency, and is converted to the initial vector by normalization or segmentation mapping method; The initial vector is subjected to amplitude normalization and mean removal processing to obtain a unit direction vector; According to the sensitivity grading rules in the desensitization strategy package, the fields with sensitivity level of the first type are designated as the outer spherical level, the fields with sensitivity level of the second type are designated as the inner spherical level, and the fields with sensitivity level between the first type and the second type are determined according to the preset boundary rules; Fixed or adaptive radius is set for the inner sphere and the outer sphere, and the outer sphere radius is greater than the inner sphere radius and a minimum interval is set; The unit direction vector is projected on the spherical surface according to the corresponding radius and the level to which it belongs to obtain a spherical embedding vector; In the projection process, the intra-cluster compactness constraint and the inter-cluster angle interval constraint are applied; The field identification, the level to which it belongs, the radius configuration, the coding unit identification and the projection coordinates are recorded to form the mapping relationship between the field and the spherical level.

[0010] Further, the ratio of the inner spherical radius and the outer spherical radius of the double-layer hyperspherical embedding network is dynamically set during operation, and the determination process of the ratio includes: According to the sensitivity grading threshold in the desensitization strategy package, the fields are initially hierarchically divided; After the embedding vector is generated, clustering operation is performed on the vectors in the same level, and the clustering operation adopts a hierarchical clustering method based on spherical cosine similarity, taking the included angle between the vectors as the distance measure; When the average included angle of any cluster in the clustering result is greater than a preset threshold, it is determined that the cluster is too dispersed, triggering automatic expansion of the spherical radius of the level; After the desensitization result is verified by the risk threshold and the utility threshold, if the utility index decreases by more than a set range, the radius interval between the inner sphere and the outer sphere is reduced accordingly; The adjusted radius parameters and field level mapping relationship are output, and the radius parameters are written into the mapping record.

[0011] Further, the generation process of the adaptive rotation disturbance includes: For the field embedding vector mapped to the outer sphere, the disturbance angle range is set to a first preset interval; For the field embedding vector mapped to the inner sphere, the disturbance angle range is set to a second preset interval, and the upper limit of the first preset interval is greater than the upper limit of the second preset interval; According to the random seed derived from the session key combined with the field identifier, the direction vector of the rotation axis is determined, and the direction vector maintains unit length in the embedding space; the session key is generated by the edge computing node when establishing a data processing session, and the session key derives the random seed through a pseudo-random function as a random source of rotation disturbance; The field identifier is generated in the structured splitting stage and contains field position and category information, and the direction vector maintains unit length in the embedding space.

[0012] The rotation angle is sampled in the corresponding preset interval, and the rotation angle and the rotation axis are applied to the embedding vector to obtain the disturbed embedding vector; When the desensitization result generated by the disturbed embedding vector does not meet the risk threshold and the utility threshold, the sampling interval of the rotation angle or the direction of the rotation axis is adjusted, and the disturbance is re-executed until the verification requirement is met; The rotation angle, rotation axis direction and field identifier are recorded as disturbance parameters as input for generating a transformation spectrum.

[0013] Further, the decoding unit executes different decoding rules according to the field category after receiving the disturbed embedding vector: For account number type and certificate number type fields, according to the format preservation constraint template in the strategy package, format preservation data consistent with the original length and check digit is outputted; For address type fields, only position information to a preset administrative district level is outputted; For time type fields, time values are mapped to a preset time interval; For amount type fields, rounding is performed according to the precision set in the desensitization strategy package, and additional noise is introduced in the decoding stage; After the decoding result is generated, risk and utility verification is performed through the following steps: A disclosure risk index is calculated, which is determined based on the uniqueness of the field in the embedding space, the distribution characteristics of the original value of the field, the context position information and the possible re-identification probability of the external knowledge base; A task utility index is calculated, which is determined based on the distribution preservation degree of the desensitized data, the prediction accuracy retention rate of the downstream risk control model and the deviation of the statistical report index; When the disclosure risk index is not greater than the risk threshold, and the task utility index is not less than the utility threshold, it is determined that the desensitization result is effective and is outputted; If the conditions are not met, the decoding parameters or the disturbance amplitude are adjusted, and the decoding and verification are re-executed until the threshold requirements are met.

[0014] Further, the generation process of the transformation spectrum includes: The mapping relationship between the field and the spherical level and the corresponding disturbance parameter are combined to form a record containing the field identifier, the belonging level, the radius configuration, the projection coordinate, the rotation axis direction, the rotation angle value and the context position information; The record is hashed after adding the version number, the effective range and the timestamp of the strategy package to generate a digest, and is linked with the previous digest to form a chain hash structure to obtain an audit log.

[0015] Further, the transmission of the minimum necessary principle includes: Only the fields and statistical characteristics directly related to the business purpose in the desensitization result that pass the threshold verification are outputted, and additional information unrelated to the current task is shielded; The original financial data is stored in an encrypted form for a short period in the edge computing node, and a destruction instruction is triggered after the storage period expires or the business purpose is completed, the data block is irreversibly deleted and a destruction record is generated, which is included in the audit log.

[0016] The beneficial effects of the present application are: The application avoids the link exposure and centralized storage risk caused by uploading sensitive financial data to the central server, meets the compliance requirements of 'data local processing and minimum necessary transmission', and effectively improves the data security and processing real-time performance. Based on the structured modeling mode of double-layer hypersphere embedding network, hierarchical mapping and vectorization representation of different sensitivity fields are realized, the distribution characteristics and correlation between fields are maintained while ensuring the desensitization intensity, and the usability of desensitized data in risk control, statistical analysis and modeling prediction is improved.

[0017] Further, the adaptive rotation perturbation mechanism proposed by the application can dynamically adjust the perturbation amplitude according to the field sensitivity difference, and balance the privacy protection and data utility; the double check of risk threshold and utility threshold ensures that the desensitization result meets the compliance requirements while maintaining the business operability. The transformation spectrum of the perturbation parameter and the mapping relationship is formed into an unforgeable audit log through chain hashing, realizing the whole-process traceability and non-repudiation of the desensitization process, and providing reliable support for the responsibility definition and compliance audit of financial data processing. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the embodiments of the application are used to explain the application, and do not constitute a limitation on the application. In the drawings:

[0019] Fig. 1 A flowchart of a financial data desensitization method based on edge computing is proposed for the application; Fig. 2 A field embedding and hierarchical mapping flowchart of a financial data desensitization method based on edge computing is proposed for the application; Fig. 3 An adaptive perturbation and audit log generation flowchart of a financial data desensitization method based on edge computing is proposed for the application. DETAILED DESCRIPTION

[0020] The application will now be further described in detail in conjunction with the drawings. These drawings are all simplified schematic diagrams, and only schematically show the basic structure of the application, and therefore only show the components related to the application.

[0021] Reference Figs. 1-3 A financial data desensitization method based on edge computing, comprising the following steps: Complete remote attestation in the edge computing node, load the desensitization strategy package and initialize the trusted execution environment; Collect financial data streams and perform structured segmentation to obtain a field list; The field list is input into the double-layer hyperspherical embedding network for encoding, and a spherical embedding vector and a mapping relationship between the field and the spherical level are output. An adaptive rotation perturbation is applied in the hyperspherical space based on the sensitivity difference, and a perturbed embedding vector and a corresponding perturbation parameter are obtained. The perturbed embedding vector is input into a decoding unit to generate de-sensitized data, and a de-sensitized result passing a preset risk threshold and a utility threshold is output. The perturbation parameter and the mapping relationship are used to generate a transformation spectrum, and an audit log is formed based on a chain hash. The de-sensitized result is transmitted to an out-domain target according to the principle of minimum necessity, and the original financial data is sealed and stored or destroyed.

[0022] In an embodiment, the edge computing node first completes remote attestation in the startup phase. Remote attestation is initiated by a security module built-in a trusted execution environment (TEE), and a node hardware fingerprint, a firmware version, and a trusted execution environment measurement value are sent to a remote verification service. The verification service returns a signed credential, and after the credential passes verification, the node enters a trusted state.

[0023] In this embodiment, the de-sensitization strategy package includes field sensitivity grading rules, mapping boundaries of inner and outer spherical surfaces, inner and outer spherical surface radius configurations and minimum intervals, value intervals of rotation perturbation angles and rotation axes, risk thresholds and utility thresholds, and format preservation constraint templates, with version numbers and effective ranges, which take effect after signature verification and validity period verification in the trusted execution environment.

[0024] After loading is completed, the node initializes memory space and security registers inside the TEE, establishes a data cache area, an embedding vector calculation area, and a perturbation parameter storage area, ensures that subsequent processing flows run in the trusted execution environment, and avoids external attacks and data leakage.

[0025] In this embodiment, the edge computing node accesses multiple types of financial data sources in real time through a preset collection agent module, including transaction messages generated by a transaction processing system, operation logs generated by an operation and maintenance system, and free text data in a business system, and the collection process is cached in timestamp order and uniformly formatted. For transaction messages, field values are parsed according to a common message structure in the financial industry (including a message header, a message body, and a verification section), and the order of appearance and original values are recorded. For operation logs, log keywords and values are aligned to standardized fields using a log field mapping table, such as “login user”, “IP address”, and “operation time”. For free text, text parsing rules are called to divide the text into candidate field units, and sensitive entities including account numbers, ID numbers, phone numbers, and amount values are extracted.

[0026] The structured segmentation is based on transaction message templates, log field mapping tables, and text analysis rules, and each field of the financial data stream is parsed to generate a field list containing field identification, field original value, context position information, and sensitivity level: Field identification: a unique code composed of field position, category, and data source type; Field original value: the actual data content obtained during collection; Context position information: the relative position of the field in the message, log, or text and the relationship with adjacent fields; Sensitivity level: determined according to the sensitivity grading rules in the de-sensitization strategy package, for example, an account field is labeled as a first-level sensitivity, and a timestamp field is labeled as a second-level sensitivity.

[0027] The collection module sends the field list to the embedded network input buffer, preparing for the subsequent vectorization coding of the double-layer hyperspherical embedded network.

[0028] In this embodiment, the generation process of the spherical embedded vector and the mapping relationship of the field and the spherical level includes: The type sub-coding unit is used to process character or enumeration type fields, including account number, certificate type, and region code, and the mapping steps include: Based on the de-sensitization strategy package or external standard library, a complete symbol set is constructed; Each symbol in the set is assigned a unique index number to form a symbol index table; An embedding matrix is initialized, with the number of rows corresponding to the size of the symbol set and the number of columns being the preset embedding dimension, and each row vector being a dense vector corresponding to the symbol; When an input field is entered, first find its index number according to the symbol, and then extract the dense vector of the corresponding row from the embedding matrix as the initial vector of the field; The embedding matrix can be selected to remain fixed during operation, or it can be adaptively updated in the spherical projection stage combined with the clustering results, so that the dense vectors of the same type of symbols are more compactly distributed in the spherical space; The numerical sub-coding unit is used to process numerical type fields, including amount, timestamp, and transaction frequency, and the mapping steps include: The original numerical field is normalized according to the preset value range, or mapped to a segmented value according to the interval division rule; Convert the normalized value or segmented value to a dense vector as the initial vector of the field; Perform amplitude normalization and mean removal on the initial vector to obtain a unit direction vector, wherein: Calculate the mean vector for the initial vectors of the same batch, and subtract the mean vector from each initial vector to eliminate the overall distribution bias; After de-meaning, the Euclidean norm of each vector is calculated, and each component of the vector is divided by the norm to normalize the amplitude of the resulting vector to 1; The normalized vector is located on the unit sphere, ensuring the scale consistency and comparability of different fields in the embedding space; The unit direction vector is hierarchically divided according to the sensitivity grading rules in the desensitization strategy package, and the fields with a sensitivity level of the first type are designated as the outer spherical surface level, the fields with a sensitivity level of the second type are designated as the inner spherical surface level, and the fields with a sensitivity level between the two are determined according to a preset boundary function to determine their level; Fixed radii or adaptive radii are set for the inner and outer spheres, and the outer sphere radius is greater than the inner sphere radius and maintains a minimum interval, ensuring spatial isolation between different sensitivity levels; The unit direction vector is projected onto the spherical surface according to the corresponding radius of the level to obtain a spherical embedding vector, including: When the field is determined to be an outer spherical surface level, the length of the unit direction vector is scaled to the outer spherical surface radius value; When the field is determined to be an inner spherical surface level, the length of the unit direction vector is scaled to the inner spherical surface radius value; If the field sensitivity is in the boundary interval, the target radius is determined by interpolation between the inner and outer radii according to the boundary rule, and scaling is performed; The projected embedding vector is located on the spherical surface of the double-layer hyperspherical structure, ensuring that different sensitivity fields form hierarchical isolation in spatial distribution.

[0029] In the spherical projection process, intra-cluster compactness constraints and inter-cluster angle separation constraints are applied, wherein: The intra-cluster compactness constraint is: For a set of embedding vectors formed by fields of the same category, the spherical cosine similarity between any two vectors is calculated; A compactness threshold is set, and when the similarity between any vector in the cluster and the cluster center is lower than the threshold, the vector is triggered for fine tuning; The fine tuning process gradually moves the vector closer to the cluster center by applying a small rotation to the vector on the spherical tangent plane until the compactness requirement is met; The intra-cluster compactness constraint ensures the convergence of the distribution of embedded fields of the same category and reduces the intra-cluster dispersion.

[0030] The inter-cluster angle separation constraint is: For a set of embedding vectors formed by fields of different categories, the angle between the cluster center vectors is calculated; An angle separation threshold is set, and when the angle between any two clusters is less than the threshold, it is determined that the clusters are too close; When the distance is determined to be too close, the angle between the clusters is restored to above the threshold value by increasing the radius value of one of the clusters or rotating the overall direction of the clusters; The angle interval constraint between clusters ensures that different categories of fields maintain sufficient angular separation in the embedding space, avoiding the weakening of sensitivity differences.

[0031] The mapping relationship between the field and the spherical level is formed by recording the field identifier, the corresponding level, the radius configuration, the encoding unit identifier, and the projection coordinates.

[0032] In this embodiment, the ratio of the inner spherical radius to the outer spherical radius of the double-layer hyperspherical embedding network is dynamically set during operation, and the determination process of the ratio includes: According to the sensitivity grading threshold in the desensitization strategy package, the field is initially divided into levels; After the embedding vector is generated, clustering operation is performed on the vectors in the same level, and the clustering operation adopts a hierarchical clustering method based on spherical cosine similarity: The cosine similarity between vectors is used as the distance metric, and the lower the cosine similarity, the farther the distance; The clustering process merges the vector clusters with the highest similarity from bottom to top until the similarity of all clusters is below the preset merging threshold; The cluster set and the corresponding cluster center are obtained.

[0033] When the average angle of any cluster in the clustering result is greater than the preset threshold, it is determined that the cluster is too dispersed, and the automatic expansion of the spherical radius of the level is triggered; The radius expansion mechanism includes: The radius value of the level to which the cluster belongs is gradually enlarged by a proportional factor, for example, by 5% to 10% each time; While expanding the radius, the direction of the embedding vector within the cluster is kept unchanged, and only the module length is adjusted; The expansion and verification are repeated until the average angle within the cluster is less than the threshold value, and the compactness requirement is met.

[0034] Through this step, it is ensured that the fields within the cluster will not be excessively dispersed in space, and the controllability of the same type of fields is improved.

[0035] After the desensitization result is verified by the risk threshold and the utility threshold, if the utility index decreases by more than a set range, the radius interval between the inner sphere and the outer sphere is reduced accordingly; The adjusted radius parameters and the field level mapping relationship are output, and the radius parameters are written into the mapping record.

[0036] In this embodiment, the generation process of the adaptive rotation disturbance includes: For the field embedding vector mapped to the outer sphere, the disturbance angle range is set to a first preset interval; For the field embedding vector mapped to the inner spherical surface, the perturbation angle range is set as a second preset interval, and the upper limit of the first preset interval is greater than the upper limit of the second preset interval, so that the outer spherical surface obtains stronger perturbation protection, and the inner spherical surface maintains higher data information utility; According to the random seed derived from the session key and the field identifier, the direction vector of the rotation axis is determined, and the direction vector maintains a unit length in the embedding space. The rotation axis ensures the randomness and field correlation of the perturbation, and avoids pattern leakage caused by long-term fixed perturbation of similar fields. The session key is generated by the edge computing node when establishing a data processing session. The session key derives a random seed through a pseudo-random function as a random source for rotation perturbation. The field identifier is generated in the structured splitting stage and contains field position and category information. The direction vector maintains a unit length in the embedding space.

[0037] The rotation angle is sampled in the corresponding preset interval, and the rotation angle and the rotation axis are applied to the embedding vector to obtain the perturbed embedding vector. When the de-sensitization result generated by the perturbed embedding vector does not meet the risk threshold and the utility threshold, the sampling interval of the rotation angle or the direction of the rotation axis is adjusted, and the perturbation is re-executed until the verification requirement is met: If the risk index is greater than the threshold, it means that the perturbation is insufficient, and the angle needs to be increased or the direction of the rotation axis needs to be changed; If the utility index is lower than the threshold, it means that the perturbation is excessive, and the angle needs to be reduced or an alternative rotation axis needs to be used; The rotation angle, rotation axis direction and field identifier are recorded as perturbation parameters as input for generating a transformation spectrum.

[0038] In this embodiment, the decoding unit executes different decoding rules according to the field category after receiving the perturbed embedding vector: For account number and ID number fields, the format preserving data consistent with the original length and check digit is output according to the format preserving constraint template in the policy package: The decoding unit first retrieves the original format template of the corresponding field (including length, segmentation structure and check digit rule); When the perturbation vector is mapped back to the character space, the length consistent with the original template is maintained; The corresponding check digit is recalculated by the check algorithm in the decoding unit to ensure that the de-sensitized data still meets the check logic of the business system; For example: the de-sensitized ID number still maintains the 18-bit format, and the Luhn algorithm or specific region code rule is used to ensure the validity of the check digit.

[0039] For address type fields, only output location information to the preset administrative district level: The decoding unit maps the perturbation vector to the address level dictionary; Only keep to the preset administrative district level (such as province or city) when output, and shield more fine-grained street and house number information; For example: the original address is "No. 27, Zhongguancun Avenue, Haidian District, Beijing", and only "Haidian District, Beijing" is kept after desensitization.

[0040] For time type fields, map the time value to the preset time interval: Map the perturbed vector to the time interval; Desensitize the specific time point to a certain predefined interval (such as hour, day, or month); For example: the original transaction timestamp "2025-08-18 14:32:45" is desensitized to "afternoon, August 18, 2025".

[0041] For amount type fields, round according to the precision set in the desensitization strategy package, and introduce additional noise in the decoding stage: The decoding unit rounds the amount according to the precision requirement in the strategy package; Add small noise (such as ±0.5%) when output to ensure stable statistical distribution and avoid single transaction being re-identified; For example: the original amount "5321.78" can be output as "5300 ± Δ" after desensitization.

[0042] After the decoding result is generated, the risk and utility verification is performed through the following steps: Calculate the disclosure risk index, which is based on the uniqueness measure of the field in the embedding space, the distribution characteristics of the original value of the field, the possible re-identification probability of the external knowledge base, and the context location information: ; Where, is the disclosure risk index; is the weight parameter, configured by the desensitization strategy package, and the sum of the four is 1; represents the embedding space uniqueness measure; represents the original value distribution characteristic risk; represents the context location information contribution; represents the external knowledge base re-identification risk.

[0043] ; Where, represents a set of fields within a certain spherical level; represents the field spherical embedding vector of the field; representing field angular distance to nearest neighbor in the same hierarchy level; representing minimum proximity threshold, configured by strategy package.

[0044] ; wherein representing total number of fields; representing original value of the th field; representing probability of the field value in the original distribution; representing lower bound of distribution smoothing (to avoid probability of 0), configured by strategy package; representing probability after lower bound correction.

[0045] ; wherein, representing context field set of the field (from location information in field list); representing conditional probability of field value given the context; representing context boost ratio, representing the recognition improvement brought by the context; representing normalized upper limit of context boost ratio, configured by strategy package.

[0046] ; wherein, is the number of candidate entries retrieved from external knowledge base for the field , when , field uniquely matches → high risk (value 1); when , take inverse ratio decay; when , no match → risk is 0.

[0047] Calculate task utility index, which is determined based on the distribution preservation degree of the de-identified data, the prediction accuracy preservation rate of the downstream risk control model, and the deviation of statistical report indicators: ; wherein, is a weight parameter, configured by strategy package, and is 1; is the distribution preservation degree; is the accuracy preservation rate of the downstream model;

[0048] ; wherein, is the discretization space of the field value; is the original data distribution; is the desensitized data distribution.

[0049] ; wherein, is the accuracy of the original data on the downstream risk control model; is the accuracy of the desensitized data on the same model; is a smoothing constant, taking a value of 10 -8 .

[0050] ; wherein, is the number of key report indicators defined by the strategy package; is the value of the th indicator in the original data; : the value of the th indicator in the desensitized data; is a smoothing constant, taking a value of 10 -4 .

[0051] When the disclosure risk index is not greater than the risk threshold, and the task utility index is not less than the utility threshold, it is determined as an effective desensitization result and output; If the conditions are not met, adjust the decoding parameters or the perturbation amplitude, and re-execute the decoding and verification until the threshold requirements are met.

[0052] In the embodiment, the generation process of the transformed spectrum includes: The mapping relationship between the field and the spherical level and the corresponding perturbation parameters are combined to form a record containing field identification, belonging level, radius configuration, projection coordinates, rotation axis direction, rotation angle value, and context position information; The record is hashed after adding the version number, effective range, and timestamp of the strategy package to generate a digest, and linked with the previous digest to form a chain hash structure to obtain an audit log.

[0053] In the embodiment, the transmission of the minimum necessary principle includes: Only the fields and statistical features directly related to the business purpose in the desensitized result after threshold verification are output, and additional information unrelated to the current task is shielded; The original financial data is stored in encrypted form for a short period in the edge computing node, and after the storage period expires or the business purpose is completed, a destruction instruction is triggered to irreversibly delete the data block and generate a destruction record, which is included in the audit log.

[0054] Example 1: In order to verify the feasibility of the present application in implementation, the present application is applied to the cross-border payment business scene of a certain commercial bank, and the test data is derived from 5000 real transaction messages, operation logs and free text notes. The experiment is carried out on an edge computing node, which is configured as an Intel SGX trusted execution environment, and a desensitization policy package that has passed signature verification is loaded.

[0055] In the collected original transaction message, the main fields include: account number, certificate number, amount, time, address.

[0056] Account number type: fixed length of 16 bits, taken from the account number field of the message; Certificate number type: length of 18 bits, taken from the log user information field; Amount type: numerical precision of 0.01 yuan; Time type: accurate to seconds; Address type: including province, city, district and street information.

[0057] Structured segmentation is based on transaction message templates, log mapping tables and text parsing rules to obtain a field list, and each list record contains a field identifier, an original value, a context position and a sensitivity level. For example:

[0058] Field identifier: acct_no_001; Original value: 6222123456789012; Context position: message field 05; Sensitivity level: first level (most sensitive).

[0059] Encode the segmented fields: Type sub-coding unit: character type fields such as account number, certificate number and address use embedded dictionary method, and the dictionary size is 10 (digits), 35 (letters and symbols) and 5000 (common place name table) respectively. Each entry is mapped to a 64-dimensional initial vector.

[0060] Numerical sub-coding unit: the amount field is normalized to the [0, 1] interval; the time field is mapped to a periodic angle according to 86400 seconds in a day, and then converted to a two-dimensional sine / cosine vector.

[0061] All initial vectors are normalized and de-meaned to obtain unit vectors, and are mapped to a double-layer hypersphere according to the sensitivity grading rules: First level sensitive field (account number, certificate number) is mapped to the outer sphere surface, with a radius r2=1.0; Second level sensitive field (amount, time) is mapped to the inner sphere surface, with a radius r1=0.8; The address type field is classified into the outer sphere according to the policy boundary.

[0062] Apply perturbation to the sphere embedding vector: Outer spherical field perturbation angle interval [5°, 15°]; Inner spherical field perturbation angle interval [1°, 5°]; The perturbation axis is determined by a random seed generated by a session key + field identification hash, ensuring reproducibility within the same session.

[0063] For example, the account number field 6222123456789012 is rotated by 12° on the outer sphere after perturbation, obtaining new embedding coordinates.

[0064] Perturbation vector input decoding unit: Account number, ID number → format preserving template generates a string of the same length; Amount → rounded to integer bits and added with ±1 yuan noise; Time → mapped to a 30-minute granularity interval; Address → retained to the "district and county" level.

[0065] Then calculate the disclosure risk index RDI and the task utility index TUI: RDI mainly depends on the uniqueness of the embedding space + field distribution characteristics; TUI mainly measures the distribution preservation degree and the accuracy preservation degree of the downstream risk control model.

[0066] Table 1 experimental results table

[0067] As shown in Table 1, the account number / ID number type data: large angle perturbation combined with format preserving template, effectively hides the original features, RDI is controlled at about 0.2; TUI>0.89, ensuring that risk control modeling is still available.

[0068] Amount / time type data: numerical fields in the inner sphere, small perturbation angle, limited precision loss, TUI up to 0.95; indicating that business indicators have little impact.

[0069] Address type data: only retained to "district and county", RDI reduced to 0.19, avoiding precise positioning; TUI 0.87, statistical models can still run normally.

[0070] From the above experiments, it can be seen that the present application can effectively reduce the disclosure risk while maintaining data utility through the whole process of embedding mapping, perturbation and decoding in real business scenarios, and has reproducibility and landing nature.

[0071] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An edge computing-based financial data desensitization method, characterized in that, The method comprises the following steps: Completing remote attestation, loading a de-identification policy package and initializing a trusted execution environment in an edge computing node; Collecting financial data streams and performing structured segmentation to obtain a field list; Inputting the field list into a double-layer hypersphere embedding network for coding to output a spherical embedding vector and a mapping relationship between the field and the spherical level; Applying adaptive rotational disturbance based on sensitivity difference in the hypersphere space to obtain a disturbed embedding vector and a corresponding disturbance parameter; Inputting the disturbed embedding vector into a decoding unit to generate de-identified data and output a de-identified result that passes a preset risk threshold and utility threshold check; Generating a transformation spectrum from the disturbance parameter and the mapping relationship and forming an audit log based on a chain hash; Transferring the de-identified result to an out-domain target according to the principle of minimum necessity and performing sealed storage or destruction on the original financial data.

2. The financial data desensitization method based on edge computing according to claim 1, characterized in that, The de-identification policy package comprises: field sensitivity grading rules, mapping boundaries of the inner sphere and the outer sphere, inner and outer sphere radius configurations and minimum intervals, value intervals of the rotation disturbance angle and the rotation axis, risk thresholds and utility thresholds and format preservation constraint templates, accompanied by version numbers and effective ranges, which take effect after signature verification and validity period verification in the trusted execution environment.

3. The financial data desensitization method based on edge computing according to claim 1, characterized in that, The collected financial data streams comprise transaction messages, operation logs and free texts, and the collection process caches and uniformly formats the data according to the timestamp sequence; The structured segmentation performs field-by-field parsing on the financial data streams based on transaction message templates, log field mapping tables and text parsing rules to generate a field list containing field identifiers, field original values, context position information and sensitivity levels.

4. The financial data desensitization method based on edge computing according to claim 1, characterized in that, The generation process of the spherical embedding vector and the mapping relationship between the field and the spherical level comprises: Obtaining initial vectors from type sub-coding units and numerical value sub-coding units according to field categories, wherein the type sub-coding units are used to process character type or enumeration type fields, including account numbers, certificate categories and region codes, and map them into initial vectors in the form of one-hot coding or embedding dictionary; The numerical value sub-coding unit is used to process numerical value type fields, including amounts, timestamps and transaction frequencies, and converts them into initial vectors in the form of normalization or segmentation mapping; Performing amplitude normalization and mean removal processing on the initial vectors to obtain unit direction vectors; According to the sensitivity grading rules in the de-identification policy package, the fields with a sensitivity level of the first type are designated as the outer sphere level, the fields with a sensitivity level of the second type are designated as the inner sphere level, and the fields with a sensitivity level between the first type and the second type are determined to belong to a level according to a preset boundary rule; Fixed or adaptive radii are set for the inner sphere and the outer sphere, and the outer sphere radius is greater than the inner sphere radius and a minimum interval is set; Performing spherical projection on the unit direction vectors according to the belonging level and the corresponding radius to obtain a spherical embedding vector; In the projection process, cluster tightness constraints and cluster angle interval constraints are applied; The mapping relationship between the field and the spherical level is formed by recording the field identifier, the belonging level, the radius configuration, the coding unit identifier and the projection coordinates.

5. The financial data desensitization method based on edge computing according to claim 1, characterized in that, The ratio of the inner sphere radius and the outer sphere radius of the double-layer hypersphere embedding network is dynamically set during operation, and the determination process of the ratio comprises: According to the sensitivity grading threshold in the desensitization strategy package, the fields are initially hierarchically divided; After the embedding vector is generated, a clustering operation is performed on the vectors in the same level, and the clustering operation adopts a hierarchical clustering method based on spherical cosine similarity, taking the included angle between the vectors as the distance measure; When the average included angle of any cluster in the clustering result is greater than a preset threshold, it is determined that the cluster is too dispersed, and the automatic expansion of the spherical radius of the level is triggered; After the desensitization result is verified by the risk threshold and the utility threshold, if the utility index decreases by more than a set range, the radius interval of the inner sphere and the outer sphere is correspondingly reduced; The adjusted radius parameters and the field level mapping relationship are output, and the radius parameters are written into the mapping record.

6. The financial data desensitization method based on edge computing according to claim 1, characterized in that, The generation process of the adaptive rotation perturbation includes: For the field embedding vector mapped to the outer sphere, the perturbation angle range is set to a first preset interval; For the field embedding vector mapped to the inner sphere, the perturbation angle range is set to a second preset interval, and the upper limit of the first preset interval is greater than the upper limit of the second preset interval; According to the random seed derived from the session key and the field identifier, the direction vector of the rotation axis is determined, and the direction vector maintains unit length in the embedding space; the session key is generated by the edge computing node when establishing a data processing session, and the session key derives the random seed through a pseudo-random function as a random source of the rotation perturbation; The field identifier is generated in the structured splitting stage and contains field position and category information, and the direction vector maintains unit length in the embedding space; The rotation angle is sampled in the corresponding preset interval, and the rotation angle and the rotation axis are applied to the embedding vector to obtain the perturbed embedding vector; When the desensitization result generated by the perturbed embedding vector does not meet the risk threshold and the utility threshold, the sampling interval of the rotation angle or the direction of the rotation axis is adjusted, and the perturbation is re-executed until the verification requirement is met; The rotation angle, rotation axis direction and field identifier are recorded as perturbation parameters as input for generating a transformation spectrum.

7. The financial data desensitization method based on edge computing according to claim 1, characterized in that, The decoding unit, after receiving the perturbed embedding vector, executes different decoding rules according to the field category: For account number and ID number fields, according to the format preservation constraint template in the strategy package, the format preservation data consistent with the original length and check digit is output; For address type fields, only the position information of the preset administrative district level is output; For time type fields, the time value is mapped to a preset time interval; For the amount type field, according to the precision set in the desensitization strategy package, rounding is performed, and additional noise is introduced in the decoding stage; After the decoding result is generated, the risk and utility verification is performed by the following steps: Calculate the disclosure risk index, which is determined based on the uniqueness of the field in the embedding space, the distribution characteristics of the original value of the field, the context position information and the possible recognition probability of the external knowledge base; Calculate the task utility index, which is determined based on the distribution preservation degree of the desensitized data, the prediction accuracy preservation rate of the downstream risk control model and the deviation of the statistical report index. When the risk index is not greater than the risk threshold value, and the task utility index is not less than the utility threshold value, it is determined that the effective desensitization result is obtained and outputted; If the condition is not met, the decoding parameters or the perturbation amplitude are adjusted, and the decoding and checking are re-executed until the threshold requirement is reached.

8. The financial data desensitization method based on edge computing according to claim 1, characterized in that, The generation process of the transformation spectrum includes: The mapping relationship between the field and the spherical level and the corresponding perturbation parameters are combined to form a record containing the field identification, the belonging level, the radius configuration, the projection coordinates, the rotation axis direction, the rotation angle value and the context position information; The summary is generated by hash operation after the version number, the effective range and the timestamp of the joining policy package are recorded, and the chain structure of the hash is formed with the last summary link to obtain the audit log.

9. The financial data desensitization method based on edge computing according to claim 1, characterized in that, The transmission of the minimum necessary principle includes: Only the fields and statistical characteristics directly related to the business purpose in the desensitization result after the threshold checking are outputted, and the additional information unrelated to the current task is shielded; The original financial data are stored in the form of encryption in the edge computing node for short-term sealed storage, and after the storage period expires or the business purpose is completed, a destruction instruction is triggered, the data block is irreversibly deleted, and a destruction record is generated, which is included in the audit log.

Images