Automobile insurance data security processing method and system based on intelligent algorithm

Abstract

The present application relates to the technical field of data security processing, and more particularly to a car insurance data security processing method and system based on intelligent algorithm. The car insurance business data containing sensitive fields is acquired, the field dependency network is constructed based on the historical access records, the target field is desensitized and converted according to the network, the connected subgraph of the historical access field sequence in the field dependency network is identified and the combination completeness index is calculated, when the index exceeds the early warning threshold, the desensitization intensity level of the key field in the candidate field is improved, and then the desensitization intensity benchmark is adjusted according to the risk accumulation value feedback. The method realizes dynamic and differentiated desensitization of sensitive fields, and effectively improves the security of car insurance data.

Description

Technical Field

[0001] This invention relates to the field of data security processing technology, and in particular to a method and system for processing automotive insurance data security based on intelligent algorithms. Background Technology

[0002] In the field of auto insurance data security processing, existing practices mainly rely on static data masking rules and role-based access control mechanisms. These methods are simple to implement and have low management costs, and are therefore widely used in current insurance business systems.

[0003] However, the inherent flaw of static data masking lies in its neglect of the dependencies between sensitive fields. In auto insurance data, a single non-sensitive field (such as vehicle identification number, number of claims, or premium amount) poses low risk when exposed individually, but the combination of multiple fields can reconstruct highly sensitive information (such as precise vehicle owner identity or claims details). Attackers can obtain different fields multiple times and piece together a complete profile through co-occurrence analysis. Static rules cannot detect this cross-field combination risk, resulting in insufficient masking strength to resist association inference attacks.

[0004] While existing dynamic data masking methods attempt to adjust intensity based on access frequency or time windows, they generally lack structured analysis of the historical behavioral sequences of the accessing entity. When an accessing entity continuously requests a set of logically related fields, they cannot identify this anomalous pattern of progressively approaching sensitive information. For example, a user might first access the vehicle brand, then accident records, and finally the vehicle owner's identification number. This sequential access behavior may indicate malicious data theft, but traditional methods can only assess the sensitivity of each field individually, failing to provide early warnings and strengthen the protection of subsequent fields based on the temporal co-occurrence patterns between fields. The shortcomings are concentrated in two dimensions: first, the dependencies between fields are not modeled for data masking decisions; second, there is a lack of dynamic monitoring and prediction capabilities for access sequence combinations. These two deficiencies make existing solutions weak in protecting against progressive attacks targeting field combinations. Summary of the Invention

[0005] This invention provides a method and system for secure processing of automotive insurance data based on intelligent algorithms, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for secure processing of automotive insurance data based on intelligent algorithms, comprising:

[0007] Retrieve auto insurance business data containing sensitive fields;

[0008] Based on the co-occurrence frequency and temporal association patterns of different sensitive fields in historical business access records, a dependency strength coefficient is calculated for each pair of sensitive fields, and a field dependency network is constructed with sensitive fields as nodes and dependency strength coefficients as edge weights;

[0009] Based on the field dependency network, the target field set involved in the current data access request is de-identified and transformed.

[0010] For the historical access field sequence of the current access subject, identify the connected subgraph formed by the historical access field sequence in the field dependency network, calculate the structural matching degree between the connected subgraph and the field combinations in the abnormal combination pattern library to obtain the combination completeness index; when the combination completeness index exceeds the warning threshold, predict the candidate field set for the access subject's subsequent access, and increase the desensitization intensity level of the key fields in the candidate field set that belong to the field combination;

[0011] The cumulative risk value is calculated based on the combined integrity index, and the desensitization intensity benchmark is adjusted based on the cumulative risk value.

[0012] Based on the co-occurrence frequency and temporal association patterns of different sensitive fields in historical business access records, a dependency strength coefficient is calculated for each pair of sensitive fields. A field dependency network is constructed with sensitive fields as nodes and dependency strength coefficients as edge weights, including:

[0013] Iterate through the historical business access records, extract the set of sensitive fields appearing in each access session, count the co-occurrence frequency of each pair of sensitive fields in different sessions, and normalize to obtain the co-occurrence frequency value;

[0014] For each pair of co-occurring sensitive fields, their access timestamps within each session are extracted, the time interval between the access times of the two fields is calculated, and a decay function is applied to the time interval to generate a time-series decay weight, wherein the smaller the time interval, the larger the time-series decay weight;

[0015] The co-occurrence frequency value is multiplied by the temporal decay weight to obtain the composite correlation degree, which is then used as the dependency strength coefficient.

[0016] A weighted directed graph is constructed using sensitive fields as nodes and the aforementioned dependency strength coefficients as edge weights to serve as a field dependency network;

[0017] The community detection algorithm is applied to the field-dependent network to identify connected components, and sensitive fields within the same connected component are marked as high-dependency field clusters.

[0018] Based on the field dependency network, the target field set involved in the current data access request is de-identified and transformed, including:

[0019] The dependency strength coefficients of field pairs within the target field set are extracted from the field dependency network and compared with a preset association threshold to divide them into a set of high-dependency field pairs and a set of low-dependency field combinations.

[0020] Obtain the desensitization strength benchmark of the current access subject and determine the transformation strength level of the current desensitization conversion;

[0021] For the set of highly dependent field pairs, a family of mapping functions is selected based on the transformation intensity level. A mapping function is assigned to each field pair, and the same mapping function is applied to the original values ​​of the two fields in the field pair to obtain the desensitized value. The family of mapping functions contains mapping functions with different degrees of reversibility; the higher the transformation intensity level, the lower the degree of reversibility of the mapping function family is selected.

[0022] For the set of low-dependency field combinations, the obfuscation perturbation amplitude is determined according to the transformation intensity level. An obfuscation function based on the obfuscation perturbation amplitude is generated independently for each field, and a desensitization value is obtained for each field. The higher the transformation intensity level, the greater the obfuscation perturbation amplitude.

[0023] The de-identified values ​​of the set of highly dependent fields and the set of low-dependent fields are combined and returned to the current data access request as a response result.

[0024] For the set of highly dependent field pairs, a family of mapping functions is selected based on the transformation intensity level. A mapping function is assigned to each field pair, and the same mapping function is applied to the original values ​​of the two fields in the field pair to obtain the desensitized value, including:

[0025] For each field pair in the set of highly dependent field pairs, a set of numerical samples for that field pair is extracted from historical business records. The Pearson correlation coefficient, Spearman rank correlation coefficient, and mutual information value between the two field values ​​are calculated. The magnitudes of the Pearson correlation coefficient, Spearman rank correlation coefficient, and mutual information value are compared, and the relationship type corresponding to the indicator with the largest value is selected as the association relationship type of that field pair. The largest Pearson correlation coefficient is marked as a linear relationship, the largest Spearman rank correlation coefficient is marked as an ordinal relationship, and the largest mutual information value is marked as a non-linear dependency relationship.

[0026] Homomorphic mapping functions are constructed based on the association type and the transformation intensity level of the field pair, where linear relationships correspond to a family of linear transformation functions, ordered relationships correspond to a family of monotonically increasing functions, and nonlinear dependencies correspond to a family of polynomial functions. The higher the transformation intensity level, the greater the degree to which the parameters of the function family deviate from the identity mapping.

[0027] Apply the homomorphic mapping function to the original values ​​of the two fields in the field pair to calculate the desensitized values, and verify that the numerical pair still maintains the association type after desensitization.

[0028] For the historical access field sequence of the current access subject, identify the connected subgraph formed by the historical access field sequence in the field dependency network, and calculate the structural matching degree between the connected subgraph and the field combinations in the abnormal combination pattern library to obtain the combination completeness index, including:

[0029] The sensitive fields in the historical access field sequence are mapped to the corresponding nodes in the field-dependent network, and the connected subgraphs formed by the edges connecting the corresponding nodes are extracted;

[0030] Extract the target topology corresponding to each field combination from the abnormal combination pattern library, calculate the structural matching degree between the connected subgraph and each target topology, and select the target topology with the highest structural matching degree as the target abnormal combination;

[0031] Identify the set of missing nodes in the target abnormal combination that do not appear in the connected subgraph, calculate the node centrality metric of each missing node in the field dependency network, and mark the missing nodes with node centrality metric values ​​higher than a preset centrality threshold as key bridging nodes;

[0032] The ratio of the number of nodes already included in the connected subgraph to the total number of nodes in the target abnormal combination is used as the basic completeness. The ratio of the number of key bridging nodes to the total number of missing nodes is used as the key missing ratio. The basic completeness and the key missing ratio are weighted and combined to obtain the combination completeness index.

[0033] When the completeness index of the combination exceeds the warning threshold, predict the candidate field set for subsequent visits by the access subject, and increase the desensitization intensity level of key fields belonging to the field combination in the candidate field set, including:

[0034] Extract the access timestamps from the historical access field sequence, calculate the number of nodes that grow in the connected subgraph within a preset time window, and use the ratio of the number of nodes that grow to the duration of the preset time window as the subgraph expansion rate; dynamically adjust the warning threshold according to the subgraph expansion rate, wherein the higher the subgraph expansion rate, the lower the warning threshold;

[0035] When the combined completeness index exceeds the warning threshold, the candidate field prediction process is triggered.

[0036] Extract the set of boundary nodes of the connected subgraph from the field dependency network, calculate the shortest path length between each boundary node and each node in the set of missing nodes, and select the missing nodes whose shortest path length is lower than a preset path threshold as the candidate field set;

[0037] Identify the fields in the candidate field set that belong to the target anomaly combination and are marked as key bridging nodes as key fields, obtain the current desensitization strength level of the current access subject, and increase the desensitization strength level of the key fields by a preset number based on the current desensitization strength level.

[0038] Calculate the cumulative risk value based on the combined integrity index, and adjust the desensitization intensity benchmark based on the cumulative risk value, including:

[0039] Extract the access timestamps of each sensitive field in the historical access field sequence, calculate the time interval sequence of adjacent field accesses, the sequence of field access order, and the sequence of field access frequency changes within a unit time window, and combine them into a time-series feature vector;

[0040] Extract standard field access sequences corresponding to each business scenario from a pre-set business process template library, calculate the vector similarity between the time-series feature vector and the standard time-series feature vector corresponding to each standard field access sequence, and use the vector similarity as a measure of the rationality of the access motivation;

[0041] The risk accumulation value is calculated by weighting and summing the visit motivation rationality measure and the portfolio completeness index.

[0042] A time decay function is applied to the accumulated risk value to obtain a decayed accumulated risk value, and the decayed accumulated risk value is fed back to the desensitization conversion step as a desensitization intensity benchmark;

[0043] Identify abnormal access records in the historical access field sequence whose access motivation rationality metric value is lower than a preset rationality threshold, extract the co-occurring sensitive field pairs in the abnormal access records, and reduce the dependency strength coefficient of the corresponding edge of the co-occurring sensitive field pairs in the field dependency network.

[0044] A second aspect of this invention provides a vehicle insurance data security processing system based on intelligent algorithms, comprising:

[0045] The data acquisition unit is used to acquire auto insurance business data containing sensitive fields;

[0046] The network construction unit is used to calculate the dependency strength coefficient for each pair of sensitive fields based on the co-occurrence frequency and temporal association pattern of different sensitive fields in historical business access records, and to construct a field dependency network with sensitive fields as nodes and dependency strength coefficients as edge weights.

[0047] The de-identification and conversion unit is used to perform de-identification and conversion on the target field set involved in the current data access request according to the field dependency network;

[0048] The complete unit is used to identify the connected subgraph formed by the historical access field sequence of the current access subject in the field dependency network, and calculate the structural matching degree between the connected subgraph and the field combination in the abnormal combination pattern library to obtain the completeness index.

[0049] The intensity enhancement unit is used to predict the set of candidate fields to be accessed by the accessing subject in the future when the completeness index of the combination exceeds the warning threshold, and to increase the desensitization intensity level of the key fields in the candidate field set that belong to the field combination.

[0050] The feedback adjustment unit is used to calculate the cumulative risk value based on the combined integrity index, and adjust the desensitization intensity benchmark based on the cumulative risk value.

[0051] A third aspect of the present invention provides an electronic device, comprising:

[0052] processor;

[0053] Memory used to store processor-executable instructions;

[0054] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0055] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0056] This invention implements dynamic data masking for auto insurance business data based on field dependencies. Sensitive fields are no longer processed in isolation, but rather protected in a coordinated manner based on a dependency network constructed from co-occurrence frequency and temporal correlation. For the target field set of the current access request, the data masking transformation directly refers to the dependency strength between fields, effectively blocking the path of deriving high-sensitivity fields from low-sensitivity fields through relational queries, significantly improving the data leakage protection effect.

[0057] By combining the connected subgraph features of historical access field sequences in the dependency network, a combinatorial integrity index is introduced to accurately capture abnormal access patterns. When the index exceeds the warning threshold, candidate fields for subsequent access are proactively predicted and the anonymization strength level is increased accordingly. Enhanced protection is initiated before sensitive combinations are fully acquired, significantly reducing the risk of sensitive field combination leakage, while avoiding performance loss caused by blindly upgrading anonymization for all fields.

[0058] The system calculates the cumulative risk value based on the integrity index of the combined data and dynamically adjusts the desensitization intensity benchmark based on feedback. It can automatically adjust the protection threshold according to the real-time behavioral risks of the accessing entity, automatically tightening the desensitization strategy in high-risk scenarios and maintaining normal access efficiency in low-risk scenarios, achieving a dynamic balance between security and availability. The overall solution upgrades data security protection from static rules to an intelligent adaptive system through field network modeling and behavioral pattern recognition. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating a method for secure processing of automotive insurance data based on intelligent algorithms.

[0060] Figure 2 This is a flowchart of the adaptive desensitization transformation process for the target field set based on field dependency networks. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0063] Figure 1 This is a flowchart illustrating the vehicle insurance data security processing method based on intelligent algorithms according to an embodiment of the present invention. The vehicle insurance data security processing method based on intelligent algorithms includes:

[0064] Retrieve auto insurance business data containing sensitive fields;

[0065] Based on the co-occurrence frequency and temporal association patterns of different sensitive fields in historical business access records, a dependency strength coefficient is calculated for each pair of sensitive fields, and a field dependency network is constructed with sensitive fields as nodes and dependency strength coefficients as edge weights;

[0066] Based on the field dependency network, the target field set involved in the current data access request is de-identified and transformed.

[0067] For the historical access field sequence of the current access subject, identify the connected subgraph formed by the historical access field sequence in the field dependency network, calculate the structural matching degree between the connected subgraph and the field combinations in the abnormal combination pattern library to obtain the combination completeness index; when the combination completeness index exceeds the warning threshold, predict the candidate field set for the access subject's subsequent access, and increase the desensitization intensity level of the key fields in the candidate field set that belong to the field combination;

[0068] The cumulative risk value is calculated based on the combined integrity index, and the desensitization intensity benchmark is adjusted based on the cumulative risk value.

[0069] In one optional implementation, based on the co-occurrence frequency and temporal association pattern of different sensitive fields in historical business access records, a dependency strength coefficient is calculated for each pair of sensitive fields, and a field dependency network is constructed with sensitive fields as nodes and dependency strength coefficients as edge weights, including:

[0070] Iterate through the historical business access records, extract the set of sensitive fields appearing in each access session, count the co-occurrence frequency of each pair of sensitive fields in different sessions, and normalize to obtain the co-occurrence frequency value;

[0071] For each pair of co-occurring sensitive fields, their access timestamps within each session are extracted, the time interval between the access times of the two fields is calculated, and a decay function is applied to the time interval to generate a time-series decay weight, wherein the smaller the time interval, the larger the time-series decay weight;

[0072] The co-occurrence frequency value is multiplied by the temporal decay weight to obtain the composite correlation degree, which is then used as the dependency strength coefficient.

[0073] A weighted directed graph is constructed using sensitive fields as nodes and the aforementioned dependency strength coefficients as edge weights to serve as a field dependency network;

[0074] The community detection algorithm is applied to the field-dependent network to identify connected components, and sensitive fields within the same connected component are marked as high-dependency field clusters.

[0075] For example, when traversing historical business access records, the access session is used as the basic processing unit. Each session corresponds to a complete business operation process, such as insurance claim processing, policy inquiry, or underwriting approval. Within a single session, the set of all accessed sensitive fields is extracted. Sensitive fields include, but are not limited to, fields involving personal privacy or core business information, such as vehicle owner's name, ID number, license plate number, accident location, bank account number, and damage assessment amount. For any two sensitive fields... and The number of sessions in which both occur simultaneously across all historical sessions is counted, and this number is denoted as the co-occurrence count. and based on the total number of sessions Normalize the values ​​to obtain the co-occurrence frequency values. The calculation method is as follows The normalized co-occurrence frequency values ​​fall between 0 and 1. The larger the value, the higher the probability that the two fields are requested simultaneously in business access, reflecting a strong correlation between them at the business logic level.

[0076] After obtaining the co-occurrence frequency values, a temporal association pattern is further introduced to refine the characterization of the dependencies between fields. This involves analyzing the sensitive fields for each co-occurrence pair. and Extract their access timestamps within each session and record them as follows: and Calculate the time interval between when two fields are accessed in the same session. The shorter the time interval, the higher the probability that the two fields are accessed consecutively, and the stronger their temporal correlation; conversely, if the access interval is long, the correlation between the two is relatively weak. To transform the time interval into a quantifiable weight value, a decay function is applied to the time interval, using an exponential decay form to generate temporally decaying weights. The calculation formula is: ,in For decay rate control parameters, Configured according to business timeliness requirements, for time-sensitive business scenarios. Larger values ​​(e.g.) =0.1), which causes the temporal correlation to decay rapidly; for scenarios requiring long-term monitoring, a smaller value for λ (such as...) =0.01), maintaining the continued influence of historical information, used to adjust the sensitivity of the time interval to the weighting effect. When When it approaches 0, A value close to 1 indicates that the two fields are accessed almost simultaneously, showing the strongest temporal correlation; as Increase Monotonically decreasing, which aligns with the design goal of "the smaller the time interval, the greater the time decay weight". If a pair of fields has co-occurring records in multiple sessions, the average of the time decay weights calculated across all sessions is taken as the final time decay weight for that pair of fields.

[0077] Co-occurrence frequency value With time decay weight Perform a product operation to obtain the composite correlation degree. ,Right now Composite correlation combines the frequency of field co-occurrence with the temporal tightness, providing a more comprehensive reflection of the business dependency between two fields. Field pairs with high co-occurrence frequency but dispersed access intervals will have a relatively lower composite correlation due to the reduced weighting of temporal decay; conversely, field pairs with moderate co-occurrence frequency but highly concentrated access intervals may maintain a higher composite correlation. Composite correlation... Directly used as a field and The dependency strength coefficient between the two fields objectively quantifies the strength of the association between the two fields in historical access behavior.

[0078] Using all sensitive fields as the node set and the dependency strength coefficient between each field pair as the edge weight, a weighted directed graph is constructed as the field dependency network. The directionality of the directed graph is reflected in the temporal order of access: if in most co-occurring sessions, the field... The access timestamp is earlier than the field Then establish a diagram from point to For directed edges, the edge weights are assigned the corresponding dependency strength coefficients. If the access order of two fields shows no significant pattern across different sessions, bidirectional edges can be established and weighted accordingly. For field pairs with dependency strength coefficients below a preset minimum threshold, no edge connections are established in the graph to avoid sparse noise interfering with subsequent analysis of the network structure. Once the field dependency network is constructed, it intuitively reflects the association topology between sensitive fields in the auto insurance business, providing a structured basis for the formulation of subsequent de-identification strategies.

[0079] After the field dependency network is constructed, a community detection algorithm is executed to identify connected components. The core objective of the community detection algorithm is to group closely connected nodes in the network into the same community. Common methods include the Louvain algorithm based on modularity optimization or fast community identification methods based on label propagation. Taking the Louvain algorithm as an example, the algorithm iteratively optimizes the network modularity. To determine the community division results, modularity This reflects the degree of difference between the sum of edge weights within a community in a real-world network and the expected value of the random network. A higher value indicates a more significant community structure. In field dependency networks, the algorithm aggregates sensitive fields with high dependency strength coefficients, frequent co-occurrence, and close temporal order into the same connected component. For example, fields such as vehicle owner's ID number, name, contact number, and home address are often requested in the same session during claims processing, and their co-occurrence frequency and temporal decay weights are both high, thus they are identified as the same connected component.

[0080] Marking sensitive fields within the same connected component as highly dependent field clusters signifies that these fields are strongly correlated in business access; accessing any one field within a cluster increases the probability of accessing the others simultaneously. The marking results of highly dependent field clusters play a crucial role in subsequent anomaly detection: as an access subject gradually accumulates access records for fields within a cluster, it can more accurately determine whether they are systematically piecing together a complete set of sensitive information. For different highly dependent field clusters, hierarchical labeling can be implemented based on the business sensitivity of the fields within the cluster. For example, field clusters involving financial account information and those containing only basic vehicle information may differ in their subsequent de-identification intensity allocation, thereby achieving differentiated data security protection strategies.

[0081] In one optional implementation, based on the field dependency network, the target field set involved in the current data access request is de-identified and transformed, including:

[0082] The dependency strength coefficients of field pairs within the target field set are extracted from the field dependency network and compared with a preset association threshold to divide them into a set of high-dependency field pairs and a set of low-dependency field combinations.

[0083] Obtain the desensitization strength benchmark of the current access subject and determine the transformation strength level of the current desensitization conversion;

[0084] For the set of highly dependent field pairs, a family of mapping functions is selected based on the transformation intensity level. A mapping function is assigned to each field pair, and the same mapping function is applied to the original values ​​of the two fields in the field pair to obtain the desensitized value. The family of mapping functions contains mapping functions with different degrees of reversibility; the higher the transformation intensity level, the lower the degree of reversibility of the mapping function family is selected.

[0085] For the set of low-dependency field combinations, the obfuscation perturbation amplitude is determined according to the transformation intensity level. An obfuscation function based on the obfuscation perturbation amplitude is generated independently for each field, and a desensitization value is obtained for each field. The higher the transformation intensity level, the greater the obfuscation perturbation amplitude.

[0086] The de-identified values ​​of the set of highly dependent fields and the set of low-dependent fields are combined and returned to the current data access request as a response result.

[0087] Combination Figure 2 This document describes the adaptive data masking process based on a field dependency network for the target field set. After constructing the field dependency network, a differentiated data masking strategy needs to be adopted for each target field set involved in the current data access request, taking into account the dependency strength relationships between fields. The process involves extracting the dependency strength coefficients for all field pairs within the target field set from the field dependency network. Compare it with a preset association threshold Perform a comparison one by one. When the dependency strength coefficient of a certain field pair... At that time, this field is classified into the set of highly dependent field pairs. ;when At that time, the corresponding field is categorized into the low-dependency field combination set. Preset association threshold The value can be configured according to the business scenario. In the car insurance business, fields such as license plate number and vehicle identification number, and insured's name and ID number often have high dependency strength and will be classified into the high dependency field pair set, while fields with weak correlation, such as insurance amount and accident location, will be classified into the low dependency field combination set.

[0088] After determining the field grouping, obtain the de-identification strength benchmark corresponding to the current access subject. Desensitization intensity benchmark It is dynamically adjusted from historical risk accumulation values, reflecting the risk assessment results accumulated by the current user in the system. The range of values ​​is mapped to discrete transformation intensity levels. For example, it can be divided into four levels: low, medium, high, and very high, each corresponding to different levels of desensitization treatment. (Change intensity level) The higher the value, the stricter the data protection requirements for the accessing entity, and the stronger the subsequent mapping function and the magnitude of the obfuscation disturbance.

[0089] For collections of highly dependent fields The core principle of field pairs in this approach is to apply the same mapping function to both fields within the same pair. This ensures that the consistency of the association between the two fields is not disrupted after anonymization, thus preventing attackers from inferring the correspondence between the original data by comparing the anonymized values ​​of the two fields. This is based on the transformation strength level. From the family of mapping functions In this context, a specific mapping function is selected for each field pair. Family of mapping functions It contains a series of mapping functions with different degrees of reversibility. Mapping functions with high reversibility (such as deterministic encryption and format-preserving encryption) can restore the original value in authorized scenarios and are suitable for situations with lower transformation strength levels. Mapping functions with low reversibility (such as one-way hash mapping and tokenization replacement) cannot be directly restored after de-identification and are suitable for situations with higher transformation strength levels. Transformation Strength Level The higher the value, the more likely the mapping function will have lower reversibility. For field pairs... , select the mapping function Simultaneously acting on raw value and raw value Desensitization values ​​were obtained. and Because the two fields use the same mapping function, if the original values ​​are the same, the values ​​after de-identification will also be the same; if the original values ​​are different, the values ​​after de-identification will also be different. This ensures that the relative relationship between the fields is maintained after de-identification, meeting the data consistency requirements in some business scenarios, while also shielding the real content from external attackers.

[0090] Taking the license plate number and vehicle identification number (VIN) in car insurance as an example, the two appear with extremely high frequency in historical access records, indicating a high degree of dependence. Typically far exceeds the association threshold. Therefore, it is classified into the set of highly dependent field pairs. When changing the intensity level At a medium level, a format-preserving encryption function can be selected as the mapping function. The same format-preserving encryption is applied to both the license plate number and VIN code, ensuring that the de-identified license plate number still conforms to the license plate format specifications and the VIN code retains its original number of digits and character structure. However, the actual content has been replaced, making it impossible to reconstruct the true vehicle information. When the strength level is changed... When upgraded to a higher level, a one-way hash mapping function is used, and the de-identified values ​​are completely irreversible, further enhancing data security.

[0091] For low-dependency field combination sets The fields in the dataset, due to their weak interrelationships, can be desensitized independently without maintaining consistent mapping between them. This is based on the transformation intensity level. Determine the magnitude of the confusion disturbance The higher the transformation intensity level, the better. The larger the value, the wider the range of random perturbations applied to the original value of the field. For each field... Independently generate obfuscation functions Confusion function Based on the amplitude of the confusion disturbance For numerical fields, random noise following a uniform or Gaussian distribution can be superimposed on the original value, with the noise amplitude controlled by... Control; for categorical fields, it can be done according to A predetermined replacement probability is used to replace the original category value with another value within the same category space; the higher the replacement probability, the stronger the obfuscation. Each field generates its own obfuscation function and applies it separately, resulting in its own desensitized value, without interference between fields.

[0092] Taking the location of the accident and the insured amount in car insurance as an example, the two have a low frequency of occurrence in historical visits, indicating a high degree of dependence. Below the association threshold It is classified into the low-dependency field combination set. When changing the intensity level At low levels, the amplitude of the confusion disturbance The magnitude is relatively small, and the location of the accident is only generalized at the provincial level, with only a small proportion of random noise added to the insured amount; when the intensity level changes... When promoted to a higher level, The increased size and generalization of the accident location to a large regional level significantly amplify the noise amplitude of the insured amount, making it more difficult to infer the original information of low-dependency fields.

[0093] Complete the collection of highly dependent fields Combination set with low-dependency fields After the data anonymization process, the anonymized values ​​from both parts are combined according to the field order and structure of the target field set in the original data access request to form a complete anonymized response, which is then returned to the current data access request as a response. The entire anonymization process is transparent to the accessing entity. The data obtained by the accessing entity maintains the same format and structure as the original data, but key sensitive content has been differentiated. This satisfies the basic requirements for business data availability while effectively preventing the direct leakage of sensitive information. This differentiated anonymization mechanism, which combines field dependencies with transformation intensity levels, can protect data security while avoiding excessive anonymization in low-risk access scenarios, thus balancing data security and business efficiency.

[0094] In one optional implementation, for the set of highly dependent field pairs, a family of mapping functions is selected according to the transformation intensity level, a mapping function is assigned to each field pair, and the same mapping function is applied to the original values ​​of the two fields in the field pair to obtain a desensitized value, including:

[0095] For each field pair in the set of highly dependent field pairs, a set of numerical samples for that field pair is extracted from historical business records. The Pearson correlation coefficient, Spearman rank correlation coefficient, and mutual information value between the two field values ​​are calculated. The magnitudes of the Pearson correlation coefficient, Spearman rank correlation coefficient, and mutual information value are compared, and the relationship type corresponding to the indicator with the largest value is selected as the association relationship type of that field pair. The largest Pearson correlation coefficient is marked as a linear relationship, the largest Spearman rank correlation coefficient is marked as an ordinal relationship, and the largest mutual information value is marked as a non-linear dependency relationship.

[0096] Homomorphic mapping functions are constructed based on the association type and the transformation intensity level of the field pair, where linear relationships correspond to a family of linear transformation functions, ordered relationships correspond to a family of monotonically increasing functions, and nonlinear dependencies correspond to a family of polynomial functions. The higher the transformation intensity level, the greater the degree to which the parameters of the function family deviate from the identity mapping.

[0097] Apply the homomorphic mapping function to the original values ​​of the two fields in the field pair to calculate the desensitized values, and verify that the numerical pair still maintains the association type after desensitization.

[0098] For example, for each field pair in the set of highly dependent field pairs, before applying the desensitization transformation, it is necessary to first extract a set of numerical samples of that field pair from historical business records to quantify the statistical correlation characteristics between the two fields. Specifically, record entries containing the complete values ​​of that field pair are extracted from historical business records, forming a set of records of length [length missing]. The paired sample sequences, where This indicates the number of valid samples. A small sample size can lead to unstable statistical indicators, therefore it is required that... The sample size should not be lower than the preset minimum sample size threshold. When the sample size is insufficient, additional samples should be taken from the extended history window.

[0099] After obtaining the sample set, three types of statistical correlation indicators were calculated. The Pearson correlation coefficient was used. This measures the degree of linear correlation between the values ​​of two fields, and its value range is... The closer the absolute value is to 1, the stronger the linear correlation. Spearman rank correlation coefficient. Numerical rank ranking calculations are more robust to non-normal distributions and outliers, and can capture monotonic but non-linear correlations. The value range is also [missing information]. Mutual information value Measuring the generalized dependency between two fields can capture not only linear relationships but also nonlinear dependencies of any form. Its value is a non-negative real number; a larger value indicates a stronger dependency. Since the three types of indicators have different dimensions and ranges, they need to be normalized before comparison. The absolute values ​​of the Pearson correlation coefficient and Spearman rank correlation coefficient are then compared with the normalized mutual information values ​​on a unified scale.

[0100] Compare the normalized values ​​of the three categories of indicators, and select the relationship type corresponding to the indicator with the largest value as the association type for that field pair. If If the absolute value of is the largest, then the field pair is marked as having a linear relationship, indicating a significant linear covariance between the two fields; if If the absolute value of is the largest, it is marked as an ordinal relation, indicating a stable monotonically increasing or decreasing trend between the two fields, but not strictly linear; if If the normalized value is the largest, it is marked as a non-linear dependency, indicating a complex non-linear coupling structure between the two fields that cannot be described by a simple linear or monotonic function. Accurate identification of the relationship type is fundamental to the subsequent construction of homomorphic mapping functions; different types correspond to different function family selection strategies.

[0101] After determining the type of association, based on the type of association and the level of transformation strength... Construct homomorphic mapping functions for field pairs. The core design principle of homomorphic mapping functions is that after applying the same mapping function to the original values ​​of the two fields in a field pair, the desensitized numerical pair still retains the original association type. That is, the desensitization operation protects numerical privacy without destroying the statistical structure between fields.

[0102] For field pairs marked as linear relationships, a family of linear transformation functions is used for desensitization. The general form of a linear transformation function is: ,in This is the scaling factor. This represents the translation amount. Since linear transformations are closed under both multiplication and addition, applying the same linear transformation to two fields ensures that the Pearson correlation coefficient between the desensitized values ​​of the two fields remains unchanged from the original value, thus satisfying the homomorphism requirement. Transformation intensity level. The higher, The magnitude of deviation from 1 and The larger the absolute value, the more distorted the original value, and the stronger the privacy protection. In actual construction, and The range of values ​​is determined by The corresponding parameter range is determined, and random sampling is performed within this range to increase unpredictability.

[0103] For field pairs marked as ordinal, a family of monotonically increasing functions is used for desensitization. Monotonically increasing functions maintain the numerical order, so applying the same monotonically increasing function to two fields ensures that the Spearman rank correlation coefficient between the desensitized values ​​remains consistent with the original rank correlation coefficient. Optional monotonically increasing functions include exponential, logarithmic, and power functions; the specific choice is determined by [the relevant authority / organization]. The decision is made. The higher the transformation intensity level, the more the selected function deviates from the identity mapping. The greater the degree of influence, for example, at high intensity levels, prioritizing power functions or exponential functions with larger curvatures significantly reshapes the original numerical distribution while preserving the sorting structure. It is important to note that the domain of the selected monotonically increasing function must cover the actual value range of the field to avoid range overflow or mapping failure.

[0104] For field pairs marked as having non-linear dependencies, a family of polynomial functions is used for desensitization. (Polynomial functions) It has sufficient expressive power to fully obfuscate the original numerical values ​​while preserving the complex non-linear dependency structure between fields. Let be the order of the polynomial. These are the coefficients of each order. Transformation intensity level. The higher the polynomial order, the better. The larger the value, the greater the deviation of the coefficients of each order from the coefficients corresponding to the identity mapping, and the stronger the desensitization intensity. After applying the same polynomial function to two fields, the mutual information value between the desensitized values ​​of the two fields remains approximately consistent with the original mutual information value, satisfying the homomorphism preservation requirement of nonlinear dependencies.

[0105] After constructing the homomorphic mapping function, the constructed function is applied to the original values ​​of the two fields in the field pair to calculate their respective desensitized values. After the desensitization calculation is complete, a verification step of the association type needs to be performed to confirm that the desensitized numerical pairs still retain the original association type. The verification method is consistent with the initial association identification step: the Pearson correlation coefficient, Spearman rank correlation coefficient, and mutual information value are recalculated for the desensitized numerical pairs. The dominant index is judged according to the same comparison rules. If the dominant index type is consistent with the association type of the original label, the verification passes; if the verification fails, the parameters of the mapping function are adjusted and the desensitization calculation is re-executed until the verification passes or the maximum number of retries is reached. This verification mechanism ensures that the desensitization process protects privacy without destroying the semantic association structure between fields, allowing statistical analysis and model training based on the desensitized data to still provide effective business insights, while preventing attackers from reverse-engineering the original sensitive values ​​through the association relationships between fields.

[0106] In one optional implementation, for the historical access field sequence of the current accessing subject, a connected subgraph formed by the historical access field sequence in the field dependency network is identified, and the structural matching degree between the connected subgraph and the field combinations in the abnormal combination pattern library is calculated to obtain a combination completeness index, including:

[0107] The sensitive fields in the historical access field sequence are mapped to the corresponding nodes in the field-dependent network, and the connected subgraphs formed by the edges connecting the corresponding nodes are extracted;

[0108] Extract the target topology corresponding to each field combination from the abnormal combination pattern library, calculate the structural matching degree between the connected subgraph and each target topology, and select the target topology with the highest structural matching degree as the target abnormal combination;

[0109] Identify the set of missing nodes in the target abnormal combination that do not appear in the connected subgraph, calculate the node centrality metric of each missing node in the field dependency network, and mark the missing nodes with node centrality metric values ​​higher than a preset centrality threshold as key bridging nodes;

[0110] The ratio of the number of nodes already included in the connected subgraph to the total number of nodes in the target abnormal combination is used as the basic completeness. The ratio of the number of key bridging nodes to the total number of missing nodes is used as the key missing ratio. The basic completeness and the key missing ratio are weighted and combined to obtain the combination completeness index.

[0111] For example, for the historical access field sequence of the current user, each sensitive field involved is mapped to a corresponding node in the field dependency network. The historical access field sequence records the field identifiers accessed sequentially by the user over a period of time. By traversing this sequence, fields belonging to the sensitive field set are extracted, and their corresponding nodes are located in the field dependency network. After node location, the structure of direct or indirect connections between these nodes in the field dependency network is extracted, forming a connected subgraph. The construction of the connected subgraph is based on edges with a non-zero dependency strength coefficient; that is, as long as there is a valid dependency relationship between two nodes, the corresponding edge is included in the connected subgraph. If there is no direct edge between some accessed field nodes, but they can be connected through other accessed field nodes, they are also considered as part of the same connected subgraph. In this way, a connected subgraph that reflects the structural distribution of the current user's actual access behavior in the field dependency network is obtained.

[0112] An abnormal combination pattern library pre-stores several known abnormal field combinations. Each field combination corresponds to a target topology, describing the connection relationships and edge weight distribution patterns between the field nodes in the combination. The target topology corresponding to each field combination is extracted one by one from the abnormal combination pattern library, and the structural matching degree between the connected subgraph and each target topology is calculated. The structural matching degree is calculated using graph isomorphism or subgraph matching methods, comprehensively considering the degree of overlap of node sets and the similarity of edge connection patterns. Specifically, the node coverage rate is the ratio of the number of nodes shared by the connected subgraph and the target topology to the total number of nodes in the target topology; the edge coverage rate is the ratio of the number of edges shared by both to the total number of edges in the target topology. The weighted average of the node coverage rate and the edge coverage rate yields the structural matching degree. The calculation method is as follows ,in For a connected subgraph, For the first A set of nodes for a target topology. Let the set of edges be a connected subgraph. For the first A set of edges for a target topological structure. and The weight coefficients for node coverage and edge coverage are respectively, and satisfy the following conditions: After traversing all target topologies in the abnormal combination pattern library, the structural matching degree is selected. The highest target topology is taken as the target anomaly combination, and subsequent analysis revolves around this target anomaly combination.

[0113] Identify the set of missing nodes in the target anomalous combination that do not appear in the connected subgraph. Missing nodes refer to the set of nodes in the target anomalous combination. The set of nodes in a connected subgraph that exists in the graph. The set of nodes not included in the list, i.e., the set of missing nodes. For each node in the set of missing nodes, calculate its node centrality metric within the field dependency network. The node centrality metric combines degree centrality and betweenness centrality. Degree centrality reflects the number of edges directly connecting a node to other nodes in the field dependency network, while betweenness centrality reflects the frequency with which a node serves as an intermediate node in the shortest path. For each node in the set of missing nodes... Its comprehensive centrality measure Degree centrality With betweenness centrality The weighted combination yields, i.e. ,in and The corresponding weight coefficients and satisfying The comprehensive centrality measure Above the preset centrality threshold Missing nodes are marked as critical bridging nodes. These nodes occupy a structural hub position in the field dependency network, and their access will significantly affect the overall integrity of the field combination. A preset centrality threshold is used. Based on the distribution of node centrality metrics in the field-dependent network, the 80th percentile of all node centrality metrics is typically used.

[0114] After obtaining the number of nodes already contained in the connected subgraph, the total number of nodes in the target abnormal combination, the number of key bridging nodes, and the total number of missing node sets, the combination completeness index is calculated. Let the number of nodes already contained in the connected subgraph be... The total number of nodes for the target abnormal combination is The number of critical bridging nodes is The total number of missing nodes is Then the basic completeness Defined as Key missing percentage Defined as (when season Combinatorial completeness index It is obtained by weighting the basic completeness and the proportion of critical missing items, i.e. ,in and The corresponding weight coefficients and satisfying . The range of values ​​is A higher value indicates a higher degree of fit between the current user's historical access behavior and the target anomaly combination, and a greater potential risk of data fabrication.

[0115] Combinatorial completeness index It balances the coverage of visited fields with the impact of unvisited key bridging nodes on the overall integrity of anomaly combinations. Basic Integrity Directly reflects the percentage of fields accessed by the accessing subject in the target anomaly combination, and the key missing percentage. The assessment should then be supplemented by considering the structural importance of the remaining missing fields to avoid underestimating high-risk scenarios based solely on quantity. For example, when the accessing entity has accessed most of the peripheral fields in the target anomaly combination but has not yet accessed a few key bridging nodes located at network hubs, higher and The weighted average of the two is also not low. This accurately reflects the degree to which the access behavior is close to completing the assembly of abnormal fields, providing a reliable basis for subsequent decisions on increasing the level of desensitization. When When the warning threshold is exceeded, the anonymization strength level of key fields in the candidate field set is increased, thereby intervening in advance before the accessing subject actually completes the access to the complete field combination and reducing the risk of data leakage.

[0116] In one optional implementation, when the completeness index of the combination exceeds the warning threshold, the candidate field set for subsequent visits by the accessing subject is predicted, and the desensitization intensity level of the key fields belonging to the field combination in the candidate field set is increased, including:

[0117] Extract the access timestamps from the historical access field sequence, calculate the number of nodes that grow in the connected subgraph within a preset time window, and use the ratio of the number of nodes that grow to the duration of the preset time window as the subgraph expansion rate;

[0118] The warning threshold is dynamically adjusted based on the subgraph expansion rate, wherein the higher the subgraph expansion rate, the lower the warning threshold.

[0119] When the combined completeness index exceeds the warning threshold, the candidate field prediction process is triggered.

[0120] Extract the set of boundary nodes of the connected subgraph from the field dependency network, calculate the shortest path length between each boundary node and each node in the set of missing nodes, and select the missing nodes whose shortest path length is lower than a preset path threshold as the candidate field set;

[0121] Identify the fields in the candidate field set that belong to the target anomaly combination and are marked as key bridging nodes as key fields, obtain the current desensitization strength level of the current access subject, and increase the desensitization strength level of the key fields by a preset number based on the current desensitization strength level.

[0122] For example, after calculating the completeness index, it is necessary to further integrate the time dimension information of access behavior to dynamically calibrate the early warning triggering mechanism. The access timestamp of each field is extracted from the historical access field sequence. Using a preset time window (e.g., set to 30 minutes or 60 minutes, configurable according to business scenarios) as the statistical unit, the number of newly added nodes in the connected subgraph within that time window is counted. Dividing this number of node increases by the duration of the preset time window yields the subgraph expansion rate. The unit is "nodes per minute," reflecting the speed at which an accessing entity explores combinations of sensitive fields within a unit of time. A higher subgraph expansion rate indicates that the accessing entity is rapidly accumulating field access coverage in a short period of time, and its behavior has stronger proactive detection characteristics. Therefore, it is necessary to lower the warning threshold to respond to potential risks in advance.

[0123] The dynamic adjustment of the warning threshold adopts a segmented mapping strategy. Let the initial baseline value of the warning threshold be... Subgraph expansion rate The corresponding threshold adjustment amount is Then the dynamic early warning threshold satisfy ,in It is about The threshold is a monotonically increasing function, meaning the higher the expansion rate, the greater the decrease in threshold. To prevent the threshold from being too low and causing an excessively high false alarm rate, a lower bound is set for the threshold. When the calculation result is lower than At that time, with This is the final warning threshold. When the combined completeness index... Exceeding the dynamic warning threshold When this happens, the candidate field prediction process is triggered, leading to subsequent field prediction and desensitization strength enhancement stages.

[0124] The core idea of ​​candidate field prediction starts from the "boundaries" of the connected subgraph and searches for the closest field node in the graph structure of the field dependency network that is most likely to be visited by the next accessing entity. A boundary node is a node in the connected subgraph that has at least one edge connecting it to a node outside the subgraph; that is, an interface node between the subgraph and the external network. All boundary nodes in the connected subgraph are extracted to form a boundary node set. For the set of missing nodes Each node in ,calculate and Find the shortest path length between all boundary nodes in the array, and take the minimum value as the node. Shortest distance to a connected subgraph The shortest path length is calculated based on the graph structure of the field dependency network, and the path weights can be expressed using the dependency strength coefficient. The reciprocal of the path cost (the higher the dependency strength, the lower the path cost) makes highly related fields closer in terms of path distance.

[0125] Let the preset path threshold be For satisfying Missing nodes Incorporate it into the candidate field set The preset path threshold is... The selection mechanism is set based on the average path length of the network dependent on the field, typically 0.5 to 1 times the average path length. This filtering mechanism ensures that the candidate field set only includes nodes that are structurally closely connected to the currently visited field, excluding distant fields with low correlation to the current access path, thereby reducing noise in the candidate set and improving prediction accuracy. In practical engineering implementation, The value can be adaptively set according to the average path length of the network that the field depends on, for example, it can be set to a range of 0.5 times to 1 times the average path length to balance recall and precision.

[0126] In the candidate field set Once identified, the key fields requiring enhanced desensitization strength need to be determined. Two criteria are used to determine key fields: first, the field belongs to the currently matched target field combination in the abnormal combination pattern library; second, the field is marked as a key bridging node in the field dependency network. The marking of key bridging nodes comes from the result of the comprehensive evaluation based on betweenness centrality and degree centrality in claim 5, i.e., the comprehensive centrality metric. Exceeding the preset centrality threshold The nodes. Candidate fields that meet the above two conditions are identified as key fields, forming a set of key fields. .

[0127] Obtain the anonymization strength level of the current access subject at the current moment, denoted as This level reflects the strength of the desensitization measures applied to the accessing subject under the current risk assessment status. For the key field set... For each field in the dataset, assign its desensitization strength level to... Increase the preset level based on The desensitization intensity level was improved. . The default value can be set to 1 or 2 levels, or it can be set according to the combination completeness index. With warning threshold The difference is dynamically determined, and the larger the difference, the more levels are raised, so as to achieve a positive proportional match between the intensity of risk response and the degree of risk.

[0128] The increased anonymization strength only applies to critical fields. Fields in the candidate field set that do not belong to the target field combination or are not marked as critical bridging nodes maintain their original anonymization strength, preventing excessive anonymization from impacting data availability in normal business processes. In practical applications, critical fields in auto insurance business data typically include highly sensitive fields such as vehicle owner ID numbers, vehicle identification numbers, and claims history. If these fields are systematically associated and collected by the accessing entity, they will pose a serious threat to the privacy and security of the data subject. By increasing the anonymization strength of candidate critical fields in advance, even if the accessing entity subsequently initiates access requests for these fields, the obtained data has already undergone a higher level of anonymization processing, thereby effectively reducing the risk of data leakage without blocking access.

[0129] The entire candidate field prediction and desensitization intensity enhancement process forms a closed loop: the subgraph expansion rate drives the dynamic adjustment of the warning threshold, making risk warnings more sensitive; the shortest path filtering from boundary nodes to missing nodes ensures the accuracy of candidate field predictions; dual-condition filtering based on key bridging node marking and target field combination attribution ensures the targeting of desensitization enhancement measures; and the dynamic configuration of preset levels ensures that the response intensity matches the actual risk level. These mechanisms work together to enable desensitization processing to apply stricter protection to potentially high-risk fields before the access behavior actually completes the combination and splicing of sensitive fields, achieving proactive defense against data security threats.

[0130] In one optional implementation, a cumulative risk value is calculated based on the combined integrity index, and the desensitization intensity benchmark is adjusted based on the cumulative risk value, including:

[0131] Extract the access timestamps of each sensitive field in the historical access field sequence, calculate the time interval sequence of adjacent field accesses, the sequence of field access order, and the sequence of field access frequency changes within a unit time window, and combine them into a time-series feature vector;

[0132] Extract standard field access sequences corresponding to each business scenario from a pre-set business process template library, calculate the vector similarity between the time-series feature vector and the standard time-series feature vector corresponding to each standard field access sequence, and use the vector similarity as a measure of the rationality of the access motivation;

[0133] The risk accumulation value is calculated by weighting and summing the visit motivation rationality measure and the portfolio completeness index.

[0134] A time decay function is applied to the accumulated risk value to obtain a decayed accumulated risk value, and the decayed accumulated risk value is fed back to the desensitization conversion step as a desensitization intensity benchmark;

[0135] Identify abnormal access records in the historical access field sequence whose access motivation rationality metric value is lower than a preset rationality threshold, extract the co-occurring sensitive field pairs in the abnormal access records, and reduce the dependency strength coefficient of the corresponding edge of the co-occurring sensitive field pairs in the field dependency network.

[0136] For example, the access timestamps of each sensitive field in the historical access field sequence are extracted, and three types of temporal feature components are constructed based on these. The time interval sequence of adjacent field access reflects the rhythmic pattern of the visitor switching between different sensitive fields. If the intervals are too uniform or exhibit mechanical periodicity, it often indicates abnormal automated batch access behavior. The sequence of field access order records the order in which sensitive fields are accessed, which is used for comparison with the field access order in the standard business process. The sequence of changes in field access frequency within a unit time window captures the fluctuations in access intensity over time, which can reflect whether there are sudden concentrated outbreaks of access behavior. The above three types of components are concatenated into a unified temporal feature vector, which serves as the input representation for subsequent assessment of the rationality of access motivation.

[0137] Standard field access sequences corresponding to various business scenarios are extracted from a pre-set business process template library. These scenarios cover typical processes such as auto insurance claims, renewal review, underwriting assessment, and subrogation. The business process template library is constructed from historical normal business operation data. For each business scenario type, access records completed by verified normal business personnel under standard business processes are selected from historical access logs. Field access order, access time interval distribution, and access frequency features are extracted. After cluster analysis and expert review, standard templates are formed, with 3 to 5 typical standard templates corresponding to each business scenario to cover different operation paths. For each standard field access sequence, a corresponding standard time-series feature vector is generated using the same construction method as the time-series feature vector. The cosine similarity between the current time-series feature vector and each standard time-series feature vector is calculated. The maximum similarity across all business scenarios is taken as the measure of the rationality of the access motivation, denoted as . . The closer it is to 1, the more closely the current access behavior matches a legitimate business process, and the stronger the rationality of the access motivation. The closer the value is to 0, the greater the deviation from all known standard procedures, indicating a higher suspicion of anomaly.

[0138] Measure the rationality of the visit motivation With portfolio completeness index Perform a weighted summation to calculate the cumulative risk value. .because The lower the value, the higher the risk. The higher the value, the higher the risk; therefore, when applying a weighted sum, [the following applies]. The complement value is calculated as follows: ;in, The weighting coefficients for the completeness index are as follows: The weighting coefficients are used to compensate for the rationality of the visit motivation; both are positive real numbers and satisfy the following conditions: This weighted approach allows the cumulative risk value to simultaneously consider both the abnormal completeness of field combinations and the degree of deviation in the motivation behind the access behavior, avoiding false negatives or missed positives caused by a single-dimensional assessment. In actual deployment, the risk accumulation value can be adjusted based on the difference in sensitivity of the two types of risk signals to the business scenario. and Perform dynamic configuration.

[0139] Accumulated risk value By applying a time decay function, the cumulative risk value after decay is obtained. The introduction of time decay is to reflect the time-sensitivity of risk—the further back in time historical access behavior is from the current moment, the weaker its impact on the current security state should be. The decay function adopts an exponential decay form: ;in, The historical time span corresponding to the cumulative risk value from the current moment. The time decay rate parameter controls how quickly the risk value diminishes over time. When the value is large (e.g., 0.05), the risk value decays rapidly, and the system is more sensitive to recent behavior; When the value is small (e.g., 0.005), the historical risk signal has a longer duration of influence on the current decision. Accumulated risk value after decay. Feedback is sent to the desensitization process and used as a benchmark for the desensitization strength of the current user. The basis for updating. Specifically, will The value range of the desensitization intensity benchmark is mapped to the range of values, so that the desensitization benchmark is automatically adjusted upward under high-risk conditions and restored to the normal level under low-risk conditions, thereby realizing the adaptive dynamic adjustment of the desensitization strategy.

[0140] Identify the rationality measure of access motivation in the historical access field sequence. Below the preset reasonableness threshold Access records that violate this rule are marked as abnormal access records. A preset reasonableness threshold is set. The threshold can be configured based on the security sensitivity of the business scenario, typically set between 0.3 and 0.5. Access records below this threshold are considered to deviate significantly from known legitimate business processes. Co-occurring sensitive field pairs are extracted from all abnormal access records; that is, field pairs consisting of any two sensitive fields accessed simultaneously in the same abnormal access session.

[0141] For the extracted co-occurrence sensitive field pairs, reduce their dependency strength coefficient in the corresponding edges of the field dependency network. Specifically, for each anomalous co-occurrence field pair... Its current dependency strength coefficient Multiply by a reduction factor less than 1 The dependency strength coefficient will be set to [value] in the upcoming update. ,in To reduce the ratio parameter, The value is typically set between 0.7 and 0.9, and is dynamically adjusted based on the severity of the abnormal access records, with lower values ​​corresponding to more severe anomalies. This is to achieve stronger edge weight suppression. If the same field pair appears repeatedly in multiple abnormal access records, its dependency strength coefficient is repeatedly reduced. The number of reductions is positively correlated with the frequency of the field pair in the abnormal records, ensuring that the network edge weights of high-frequency abnormal co-occurring field pairs are suppressed more significantly.

[0142] The core logic of this feedback mechanism is as follows: if a pair of fields frequently appears in abnormal access behavior, it indicates that their correlation in normal business scenarios may be overestimated, or the combination of the field pair itself carries a high risk of privacy leakage. Therefore, it is necessary to proactively weaken the correlation strength at the field dependency network level. This reduces the probability of the field pair being identified as a reasonable association during subsequent de-identification and anomaly detection, further enhancing overall data security protection capabilities. By linking the calculation results of the risk accumulation value with the dynamic updates of the field dependency network, the entire data security processing flow forms a closed-loop adaptive mechanism, continuously accumulating security knowledge from historical access behavior and gradually optimizing de-identification strategies and anomaly detection accuracy.

[0143] A second aspect of this invention provides a vehicle insurance data security processing system based on intelligent algorithms, comprising:

[0144] The data acquisition unit is used to acquire auto insurance business data containing sensitive fields;

[0145] The network construction unit is used to calculate the dependency strength coefficient for each pair of sensitive fields based on the co-occurrence frequency and temporal association pattern of different sensitive fields in historical business access records, and to construct a field dependency network with sensitive fields as nodes and dependency strength coefficients as edge weights.

[0146] The de-identification and conversion unit is used to perform de-identification and conversion on the target field set involved in the current data access request according to the field dependency network;

[0147] The complete unit is used to identify the connected subgraph formed by the historical access field sequence of the current access subject in the field dependency network, and calculate the structural matching degree between the connected subgraph and the field combination in the abnormal combination pattern library to obtain the completeness index.

[0148] The intensity enhancement unit is used to predict the set of candidate fields to be accessed by the accessing subject in the future when the completeness index of the combination exceeds the warning threshold, and to increase the desensitization intensity level of the key fields in the candidate field set that belong to the field combination.

[0149] The feedback adjustment unit is used to calculate the cumulative risk value based on the combined integrity index, and adjust the desensitization intensity benchmark based on the cumulative risk value.

[0150] A third aspect of the present invention provides an electronic device, comprising:

[0151] processor;

[0152] Memory used to store processor-executable instructions;

[0153] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0154] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0155] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

Claims

1. A method for secure processing of automobile insurance data based on intelligent algorithms, characterized in that, include: Retrieve auto insurance business data containing sensitive fields; Based on the co-occurrence frequency and temporal association patterns of different sensitive fields in historical business access records, a dependency strength coefficient is calculated for each pair of sensitive fields, and a field dependency network is constructed with sensitive fields as nodes and dependency strength coefficients as edge weights; Based on the field dependency network, the target field set involved in the current data access request is de-identified and transformed. For the historical access field sequence of the current access subject, identify the connected subgraph formed by the historical access field sequence in the field dependency network, calculate the structural matching degree between the connected subgraph and the field combinations in the abnormal combination pattern library to obtain the combination completeness index; when the combination completeness index exceeds the warning threshold, predict the candidate field set for the access subject's subsequent access, and increase the desensitization intensity level of the key fields in the candidate field set that belong to the field combination; The cumulative risk value is calculated based on the combined integrity index, and the desensitization intensity benchmark is adjusted based on the cumulative risk value.

2. The method according to claim 1, characterized in that, Based on the co-occurrence frequency and temporal association patterns of different sensitive fields in historical business access records, a dependency strength coefficient is calculated for each pair of sensitive fields. A field dependency network is constructed with sensitive fields as nodes and dependency strength coefficients as edge weights, including: Iterate through the historical business access records, extract the set of sensitive fields appearing in each access session, count the co-occurrence frequency of each pair of sensitive fields in different sessions, and normalize to obtain the co-occurrence frequency value; For each pair of co-occurring sensitive fields, their access timestamps within each session are extracted, the time interval between the access times of the two fields is calculated, and a decay function is applied to the time interval to generate a time-series decay weight, wherein the smaller the time interval, the larger the time-series decay weight; The co-occurrence frequency value is multiplied by the temporal decay weight to obtain the composite correlation degree, which is then used as the dependency strength coefficient. A weighted directed graph is constructed using sensitive fields as nodes and the aforementioned dependency strength coefficients as edge weights to serve as a field dependency network.

3. The method according to claim 1, characterized in that, Based on the field dependency network, the target field set involved in the current data access request is de-identified and transformed, including: The dependency strength coefficients of field pairs within the target field set are extracted from the field dependency network and compared with a preset association threshold to divide them into a set of high-dependency field pairs and a set of low-dependency field combinations. Obtain the desensitization strength benchmark of the current access subject and determine the transformation strength level of the current desensitization conversion; For the set of highly dependent field pairs, a family of mapping functions is selected based on the transformation intensity level. A mapping function is assigned to each field pair, and the same mapping function is applied to the original values ​​of the two fields in the field pair to obtain the desensitized value. The family of mapping functions contains mapping functions with different degrees of reversibility; the higher the transformation intensity level, the lower the degree of reversibility of the mapping function family is selected. For the set of low-dependency field combinations, the obfuscation perturbation amplitude is determined according to the transformation intensity level. An obfuscation function based on the obfuscation perturbation amplitude is generated independently for each field, and a desensitization value is obtained for each field. The higher the transformation intensity level, the greater the obfuscation perturbation amplitude.

4. The method according to claim 3, characterized in that, For the set of highly dependent field pairs, a family of mapping functions is selected based on the transformation intensity level. A mapping function is assigned to each field pair, and the same mapping function is applied to the original values ​​of the two fields in the field pair to obtain the desensitized value, including: For each field pair in the set of highly dependent field pairs, a set of numerical samples for that field pair is extracted from historical business records. The Pearson correlation coefficient, Spearman rank correlation coefficient, and mutual information value between the two field values ​​are calculated. The magnitudes of the Pearson correlation coefficient, Spearman rank correlation coefficient, and mutual information value are compared, and the relationship type corresponding to the indicator with the largest value is selected as the association relationship type of that field pair. The largest Pearson correlation coefficient is marked as a linear relationship, the largest Spearman rank correlation coefficient is marked as an ordinal relationship, and the largest mutual information value is marked as a non-linear dependency relationship. Homomorphic mapping functions are constructed based on the association type and the transformation intensity level of the field pair, where linear relationships correspond to a family of linear transformation functions, ordered relationships correspond to a family of monotonically increasing functions, and nonlinear dependencies correspond to a family of polynomial functions. The higher the transformation intensity level, the greater the degree to which the parameters of the function family deviate from the identity mapping. The desensitized value is calculated by applying the homomorphic mapping function to the original values ​​of the two fields in the field pair.

5. The method according to claim 1, characterized in that, For the historical access field sequence of the current access subject, identify the connected subgraph formed by the historical access field sequence in the field dependency network, and calculate the structural matching degree between the connected subgraph and the field combinations in the abnormal combination pattern library to obtain the combination completeness index, including: The sensitive fields in the historical access field sequence are mapped to the corresponding nodes in the field-dependent network, and the connected subgraphs formed by the edges connecting the corresponding nodes are extracted; Extract the target topology corresponding to each field combination from the abnormal combination pattern library, calculate the structural matching degree between the connected subgraph and each target topology, and select the target topology with the highest structural matching degree as the target abnormal combination; Identify the set of missing nodes in the target abnormal combination that do not appear in the connected subgraph, calculate the node centrality metric of each missing node in the field dependency network, and mark the missing nodes with node centrality metric values ​​higher than a preset centrality threshold as key bridging nodes; The ratio of the number of nodes already included in the connected subgraph to the total number of nodes in the target abnormal combination is used as the basic completeness. The ratio of the number of key bridging nodes to the total number of missing nodes is used as the key missing ratio. The basic completeness and the key missing ratio are weighted and combined to obtain the combination completeness index.

6. The method according to claim 5, characterized in that, When the completeness index of the combination exceeds the warning threshold, predict the candidate field set for subsequent visits by the access subject, and increase the desensitization intensity level of key fields belonging to the field combination in the candidate field set, including: Extract the access timestamps from the historical access field sequence, calculate the number of nodes that grow in the connected subgraph within a preset time window, and use the ratio of the number of nodes that grow to the duration of the preset time window as the subgraph expansion rate; dynamically adjust the warning threshold according to the subgraph expansion rate, wherein the higher the subgraph expansion rate, the lower the warning threshold; When the combined completeness index exceeds the warning threshold, the candidate field prediction process is triggered. Extract the set of boundary nodes of the connected subgraph from the field dependency network, calculate the shortest path length between each boundary node and each node in the set of missing nodes, and select the missing nodes whose shortest path length is lower than a preset path threshold as the candidate field set; Identify the fields in the candidate field set that belong to the target anomaly combination and are marked as key bridging nodes as key fields, obtain the current desensitization strength level of the current access subject, and increase the desensitization strength level of the key fields by a preset number based on the current desensitization strength level.

7. The method according to claim 1, characterized in that, Calculate the cumulative risk value based on the combined integrity index, and adjust the desensitization intensity benchmark based on the cumulative risk value, including: Extract the access timestamps of each sensitive field in the historical access field sequence, calculate the time interval sequence of adjacent field accesses, the sequence of field access order, and the sequence of field access frequency changes within a unit time window, and combine them into a time-series feature vector; Extract standard field access sequences corresponding to each business scenario from a pre-set business process template library, calculate the vector similarity between the time-series feature vector and the standard time-series feature vector corresponding to each standard field access sequence, and use the vector similarity as a measure of the rationality of the access motivation; The risk accumulation value is calculated by weighting and summing the visit motivation rationality measure and the portfolio completeness index. A time decay function is applied to the accumulated risk value to obtain a decayed accumulated risk value, and the decayed accumulated risk value is fed back to the desensitization conversion step as a desensitization intensity benchmark; Identify abnormal access records in the historical access field sequence whose access motivation rationality metric value is lower than a preset rationality threshold, extract the co-occurring sensitive field pairs in the abnormal access records, and reduce the dependency strength coefficient of the corresponding edge of the co-occurring sensitive field pairs in the field dependency network.

8. A vehicle insurance data security processing system based on intelligent algorithms, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire auto insurance business data containing sensitive fields; The network construction unit is used to calculate the dependency strength coefficient for each pair of sensitive fields based on the co-occurrence frequency and temporal association pattern of different sensitive fields in historical business access records, and to construct a field dependency network with sensitive fields as nodes and dependency strength coefficients as edge weights. The de-identification and conversion unit is used to perform de-identification and conversion on the target field set involved in the current data access request according to the field dependency network; The complete unit is used to identify the connected subgraph formed by the historical access field sequence of the current access subject in the field dependency network, and calculate the structural matching degree between the connected subgraph and the field combination in the abnormal combination pattern library to obtain the completeness index. The intensity enhancement unit is used to predict the set of candidate fields to be accessed by the accessing subject in the future when the completeness index of the combination exceeds the warning threshold, and to increase the desensitization intensity level of the key fields in the candidate field set that belong to the field combination. The feedback adjustment unit is used to calculate the cumulative risk value based on the combined integrity index, and adjust the desensitization intensity benchmark based on the cumulative risk value.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images