Financial integrated data security storage method and system based on cloud computing

Abstract

The application provides a financial integrated data security storage method and system based on cloud computing, relates to the technical field of cloud computing, and comprises the following steps: obtaining financial data and analyzing the semantics thereof, extracting business attributes and correlation to construct a multi-dimensional sensitivity evaluation model, and performing sensitivity scoring on data fields; performing differential encryption according to the score to generate layered encrypted data; generating a dynamic desensitization rule set based on access request information, and returning after selectively decrypting and desensitizing target fields according to the rule set. The application realizes fine security management and control and dynamic self-adaptive desensitization of financial data storage, and improves the flexibility of data security and access control.

Description

Technical Field

[0001] This invention relates to cloud computing technology, and more particularly to a cloud-based method and system for secure storage of integrated financial data. Background Technology

[0002] In the field of secure financial data storage in cloud computing environments, existing technologies typically employ standardized data encryption and access control mechanisms to ensure information security. The common practice is that enterprises encrypt all financial data before uploading it to cloud storage, generally using a unified encryption algorithm and key. Access control relies on preset static permission policies, granting access to specific data files or database tables based on user roles or departments. This model aims to prevent data theft during transmission and static storage through encryption, and to restrict unauthorized access through permission management.

[0003] The sensitivity of different fields within financial data varies significantly; for example, bank account numbers, transaction amounts, and publicly disclosed company names are drastically different in importance. A one-size-fits-all, uniform encryption strategy fails to provide granular protection based on the intrinsic value of the data, resulting in insufficient protection for highly sensitive information or unnecessary performance overhead for less sensitive information. Furthermore, static, role-based access control policies lack scenario adaptability. They cannot make real-time judgments based on dynamic context information such as access purpose, time, and location, making it difficult to address risks such as unauthorized access or abuse of privileges. This rigidity is particularly evident in data sharing and collaboration scenarios, either excessively restricting access and impacting business efficiency or leaving security vulnerabilities. Summary of the Invention

[0004] This invention provides a cloud computing-based integrated financial data security storage method and system, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a cloud-based method for secure storage of integrated financial data, comprising:

[0006] Obtain the financial data to be stored, perform semantic parsing on the financial data to extract the business attributes and relationships of the data fields; based on the business attributes and relationships, construct a multi-dimensional sensitivity assessment model, and perform sensitivity quantification scoring on each field in the financial data to obtain structured financial data;

[0007] The structured financial data is subjected to differentiated encryption processing based on sensitivity scores. Data fields with different sensitivity score ranges are encrypted using different key systems to generate layered encrypted data.

[0008] When an access request is received, the identity of the accessing subject, the purpose of access, and the spatiotemporal context information of access are extracted from the request; based on the access information and the output of the multi-dimensional sensitivity assessment model, a dynamic desensitization rule set for the current access scenario is generated.

[0009] Based on the dynamic desensitization rule set and key index relationship, target data fields are selectively decrypted from the cloud storage resource pool, and the decrypted data fields are processed according to the desensitization transformation method defined in the rule set to generate desensitized financial data;

[0010] The anonymized financial data is returned to the accessing entity, and the access frequency and abnormal access patterns of the data fields during this access process are recorded. The access records are then used to optimize and update the multi-dimensional sensitivity assessment model.

[0011] Based on the aforementioned business attributes and relationships, a multi-dimensional sensitivity assessment model is constructed, and sensitivity metrics are quantitatively scored for each field in the financial data, resulting in structured financial data including:

[0012] The business attributes are semantically vectorized and encoded, and a field association graph is constructed. The association strength weight is marked on each directed edge of the field association graph.

[0013] Based on the semantic vectorized encoding and the field association graph with association strength weights, a multi-dimensional sensitivity assessment model is established, including a single-field sensitivity assessment sub-model and an association sensitivity propagation sub-model.

[0014] The initial sensitivity score is calculated by using the single-field sensitivity assessment sub-model to perform semantic vectorization encoding of each data field, and the basic sensitivity score of each field is obtained. The basic sensitivity score is then assigned as the initial sensitivity attribute of the corresponding node in the field association graph.

[0015] The sensitivity propagation sub-model performs sensitivity propagation calculation on the field association graph. Based on the association strength weight of the directed edge and the basic sensitivity score of the connected node, the sensitivity attribute of each node is iteratively updated until the sensitivity attribute converges, and a comprehensive sensitivity score considering the association effect between fields is obtained.

[0016] The comprehensive sensitivity score is bound to the corresponding data fields to generate structured financial data carrying sensitivity score information.

[0017] The structured financial data is subjected to differentiated encryption processing based on sensitivity scores. Data fields with different sensitivity score ranges are encrypted using different key systems to generate layered encrypted data, including:

[0018] Statistical analysis is performed on the sensitivity scores to determine the set of quantiles in the score distribution. Based on the set of quantiles, multiple sensitivity score intervals are divided, and a corresponding key system is constructed for each sensitivity score interval.

[0019] For each of the aforementioned sensitivity scoring intervals, a corresponding key derivation topology is established. The key derivation topology defines the derivation path and derivation level from the root key to the target encryption key, and the key derivation topology for different sensitivity scoring intervals has a derivation level depth adapted to its sensitivity.

[0020] Based on the sensitivity score of each data field, determine the sensitivity score interval to which it belongs, obtain the key derivation topology corresponding to the sensitivity score interval, and establish a mapping relationship between the data field and the corresponding key derivation topology.

[0021] Based on the mapping relationship, multi-level key derivation operations are performed along the derivation path defined by the corresponding key derivation topology to generate target encryption keys corresponding to their sensitivity score ranges for each data field.

[0022] The target encryption key is used to encrypt the data fields belonging to the corresponding sensitivity score range to obtain encrypted data fields. All encrypted data fields are then associated with and stored with the identification information of the key-derived topology to form hierarchical encrypted data.

[0023] Based on the mapping relationship, multi-level key derivation operations are performed along the derivation path defined by the corresponding key derivation topology to generate target encryption keys for each data field corresponding to its sensitivity score range, including:

[0024] Based on the mapping relationship, determine the corresponding key derivation topology for each data field, and extract the root key identifier and derivation path information from the key derivation topology.

[0025] The derivation path information is structured and parsed to extract the derivation transformation rules of each derivation level and the key material transfer relationship that defines the key material transfer method between levels;

[0026] The root key is obtained based on the root key identifier, and the root key is input as the initial key material into the first derivation level of the key derivation topology to start the key derivation process;

[0027] Following the derivation hierarchy order defined by the derivation path information, the derivation transformation rule corresponding to the current derivation hierarchy is obtained, and the key material received at the current derivation hierarchy is transformed according to the derivation transformation rule to generate output key material.

[0028] According to the key material transfer relationship, the output key material of the current derivation level is transferred to the next derivation level as its input key material. The derivation process is repeated until all derivation levels are processed, and the key material output by the final derivation level is determined as the target encryption key.

[0029] Based on the access information and the output of the multi-dimensional sensitivity assessment model, a dynamic desensitization rule set is generated for the current access scenario, including:

[0030] The access subject characteristics and access behavior characteristics are extracted from the access information, and the sensitivity metrics of each data field are extracted from the output of the multi-dimensional sensitivity assessment model. The access subject characteristics and access behavior characteristics are fused to generate an access risk assessment vector for assessing the current access risk.

[0031] Based on the access risk assessment vector and the sensitivity quantification index, a desensitization intensity mapping function is constructed to determine the degree of desensitization processing. The desensitization intensity mapping function will dynamically calculate the desensitization intensity according to the combined characteristics of access risk and data sensitivity.

[0032] The sensitivity metrics of each data field and the access risk assessment vector are input as input parameters to the desensitization strength mapping function, and the target desensitization strength that each data field should adopt in the current access scenario is calculated.

[0033] Based on the target desensitization intensity, determine the desensitization operation type that is appropriate to the desensitization intensity for each data field, and set the corresponding desensitization parameter configuration based on the desensitization operation type to ensure that the desensitization effect meets the security requirements;

[0034] The de-identification operation type and the de-identification parameter configuration are integrated into the system to generate a dynamic de-identification rule set for the current access scenario.

[0035] Based on the dynamic de-identification rule set and key index relationship, target data fields are selectively decrypted from the cloud storage resource pool, and the decrypted data fields are processed according to the de-identification transformation method defined in the rule set to generate de-identified financial data, including:

[0036] Extract the desensitization transformation method corresponding to each data field from the dynamic desensitization rule set, construct the association mapping table between the data field and the desensitization transformation method, and establish a distributed key retrieval mechanism based on intelligent load balancing.

[0037] Based on the distributed key retrieval mechanism and key index relationship, the storage location of each target data field in the cloud storage resource pool and the corresponding key acquisition path are determined, and the key derivation level sequence required for decryption key is obtained from the key management node.

[0038] Access the key management node along the key acquisition path, and use an adaptive key recombination algorithm to dynamically combine and calculate the acquired key materials at each level to generate a decryption key for decrypting the target data field;

[0039] The decryption key is used to decrypt the target data field at the corresponding storage location in the cloud storage resource pool to obtain the decrypted data field.

[0040] Based on the association mapping table, the desensitization processing parameters corresponding to each of the decrypted data fields are obtained, and the desensitization processing parameters are applied to the desensitization transformation method to perform desensitization transformation operation on each of the decrypted data fields to generate desensitized financial data.

[0041] Accessing the key management node along the key acquisition path, and using an adaptive key recombination algorithm to dynamically combine and calculate the acquired key materials at each level, a decryption key for decrypting the target data field is generated, including:

[0042] Following the key derivation hierarchy sequence defined by the key acquisition path, the key management node is accessed sequentially, the corresponding key material is obtained from each key derivation hierarchy, and the obtained key material is preliminarily structured through an adaptive key reorganization algorithm to form a hierarchical key material set adapted to the current system load characteristics.

[0043] Deep structural feature extraction is performed on each level of key material in the hierarchical key material set to obtain the entropy distribution characteristics of each level of key material and the real-time load status of the system. Based on the entropy distribution characteristics and the real-time load status of the system, the contribution weight of each level of key material in the subsequent dynamic combination operation is calculated.

[0044] Based on the contribution weight, the key materials at each level in the hierarchical key material set are intelligently weighted to generate a weighted key material sequence that reflects the importance of each level of key material.

[0045] Based on the adaptive key reorganization algorithm, the resource optimization configuration of each weighted key material in the weighted key material sequence is performed, and the layer-by-layer fusion operation is performed according to the hierarchical order defined by the key derivation hierarchical sequence. The weighted key material of the current level is adaptively mixed and transformed with the fusion result of the previous level to generate the fusion output of the current level.

[0046] Repeat the layer-by-layer fusion operation until all levels of weighted key material have been processed, and determine the final level fusion output as the dynamically adaptive decryption key.

[0047] A second aspect of this invention provides a cloud-based integrated financial data security storage system, comprising:

[0048] The data parsing unit is used to acquire financial data to be stored, perform semantic parsing on the financial data to extract the business attributes and relationships of the data fields; based on the business attributes and relationships, construct a multi-dimensional sensitivity assessment model, and perform sensitivity quantification scoring on each field in the financial data to obtain structured financial data.

[0049] The layered encryption unit is used to perform differentiated encryption processing on the structured financial data according to the sensitivity score, and to encrypt data fields with different key systems for different sensitivity score ranges to generate layered encrypted data;

[0050] The access information unit is used to extract the access subject's identity, access purpose, and access spatiotemporal context information from the access request when an access request is received; and to generate a dynamic desensitization rule set for the current access scenario based on the access information and the output of the multi-dimensional sensitivity assessment model.

[0051] The selective decryption unit is used to selectively decrypt target data fields from the cloud storage resource pool according to the dynamic desensitization rule set and key index relationship, and process the decrypted data fields according to the desensitization transformation method defined in the rule set to generate desensitized financial data;

[0052] The model optimization unit is used to return the anonymized financial data to the accessing subject, and at the same time record the access frequency and abnormal access patterns of the data fields during this access process, and use the access records to optimize and update the multi-dimensional sensitivity assessment model.

[0053] A third aspect of the present invention provides an electronic device, comprising:

[0054] processor;

[0055] Memory used to store processor-executable instructions;

[0056] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0057] 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.

[0058] The beneficial effects of this application are as follows:

[0059] This method significantly enhances the security of financial data storage. By semantically parsing financial data and constructing a multi-dimensional sensitivity assessment model, it achieves a deep understanding of the business attributes and relationships of data fields. This makes sensitivity quantification and scoring more accurate, avoiding the one-sidedness of traditional methods that rely on single dimensions or static rules for sensitivity judgment. This lays a scientific foundation for subsequent differentiated security strategies. Differential encryption processing is applied based on sensitivity scores, encrypting data fields with different sensitivity levels using independent key systems, forming a layered encrypted data structure. This design effectively isolates data at different security levels; even if the key at one level is leaked, the security of the entire dataset will not be compromised, greatly enhancing the data's resistance to attacks in a cloud storage environment.

[0060] In the data access phase, a set of desensitization rules matching the current scenario is dynamically generated by comprehensively considering the identity of the accessing entity, the purpose of access, and the access context. This method achieves fine-grained access control and data desensitization, maximizing the concealment of core sensitive information while ensuring normal business operations, effectively preventing the risk of data leakage during access and use. Selective decryption and desensitization based on dynamic rules and key indexes avoids the security exposure window caused by full decryption in traditional solutions, improving data processing efficiency and security. The system automatically records access frequency and abnormal patterns, and uses this feedback information to continuously optimize the sensitivity assessment model. This allows the model to adapt to business changes and new security threats, and the model's assessment accuracy and the effectiveness of security policies evolve over time, forming a closed-loop security enhancement mechanism. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the cloud computing-based secure storage method for financial data according to an embodiment of the present invention.

[0062] Figure 2 This is a flowchart illustrating key derivation and encrypted data storage based on sensitivity partitioning, as described in an embodiment of the present invention. Detailed Implementation

[0063] 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.

[0064] 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.

[0065] Figure 1 This is a flowchart illustrating the cloud-based integrated financial data security storage method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] Obtain the financial data to be stored, perform semantic parsing on the financial data to extract the business attributes and relationships of the data fields; based on the business attributes and relationships, construct a multi-dimensional sensitivity assessment model, and perform sensitivity quantification scoring on each field in the financial data to obtain structured financial data;

[0067] The structured financial data is subjected to differentiated encryption processing based on sensitivity scores. Data fields with different sensitivity score ranges are encrypted using different key systems to generate layered encrypted data.

[0068] When an access request is received, the identity of the accessing subject, the purpose of access, and the spatiotemporal context information of access are extracted from the request; based on the access information and the output of the multi-dimensional sensitivity assessment model, a dynamic desensitization rule set for the current access scenario is generated.

[0069] Based on the dynamic desensitization rule set and key index relationship, target data fields are selectively decrypted from the cloud storage resource pool, and the decrypted data fields are processed according to the desensitization transformation method defined in the rule set to generate desensitized financial data;

[0070] The anonymized financial data is returned to the accessing entity, and the access frequency and abnormal access patterns of the data fields during this access process are recorded. The access records are then used to optimize and update the multi-dimensional sensitivity assessment model.

[0071] In one optional implementation, a multi-dimensional sensitivity assessment model is constructed based on the business attributes and the relationship, and sensitivity metrics are quantitatively scored for each field in the financial data to obtain structured financial data including:

[0072] The business attributes are semantically vectorized and encoded, and a field association graph is constructed. The association strength weight is marked on each directed edge of the field association graph.

[0073] Based on the semantic vectorized encoding and the field association graph with association strength weights, a multi-dimensional sensitivity assessment model is established, including a single-field sensitivity assessment sub-model and an association sensitivity propagation sub-model.

[0074] The initial sensitivity score is calculated by using the single-field sensitivity assessment sub-model to perform semantic vectorization encoding of each data field, and the basic sensitivity score of each field is obtained. The basic sensitivity score is then assigned as the initial sensitivity attribute of the corresponding node in the field association graph.

[0075] The sensitivity propagation sub-model performs sensitivity propagation calculation on the field association graph. Based on the association strength weight of the directed edge and the basic sensitivity score of the connected node, the sensitivity attribute of each node is iteratively updated until the sensitivity attribute converges, and a comprehensive sensitivity score considering the association effect between fields is obtained.

[0076] The comprehensive sensitivity score is bound to the corresponding data fields to generate structured financial data carrying sensitivity score information.

[0077] To address the varying sensitivity levels of different fields in financial data, a multi-dimensional sensitivity assessment model is constructed to achieve refined sensitivity measurement scoring. First, semantic vectorization encoding is performed on business attributes, converting attributes such as "customer name," "transaction amount," and "bank account number" into high-dimensional semantic vector representations. A pre-trained domain language model is then used to encode the business attribute description text for each field, generating dense vectors with dimensions of 512 or 768. These semantic vectors capture the semantic features and business meaning of the fields. For example, "customer contact number" is encoded into a vector. This makes fields with similar semantics closer together in the vector space.

[0078] After semantic vectorization encoding, a field relationship graph is constructed to represent the dependencies between fields. Each field in the financial data is used as a node in the graph. By analyzing foreign key constraints, computational dependencies, and reference relationships in the business logic, directed edges are established between related fields. For example, the "Customer ID" field in the "Order Details Table" has a foreign key relationship with the "Customer ID" field in the "Customer Information Table," and a directed edge is established in the graph pointing from the former to the latter. For computational fields, such as "Amount Payable," which is calculated from "Unit Price" and "Quantity," directed edges are established from the "Unit Price" and "Quantity" nodes respectively to the "Amount Payable" node.

[0079] For each directed edge in the field association graph, annotate the association strength weight to quantify the degree of association. Association Strength Weight The value ranges from 0 to 1, representing the strength of the association influence from node i to node j. For foreign key associations, the weight is set to 0.9 for mandatory foreign key constraints and 0.6 for optional foreign keys. For computational dependencies, the association strength is determined based on the weight coefficient of the field in the calculation formula. For example, in a weighted summation formula, if the weight coefficient of a certain field accounts for 30%, the association strength weight of the corresponding directed edge is set to 0.3. For reference relationships, the weight value is determined by statistically analyzing the actual reference frequency in historical data. A reference frequency exceeding 80% is set to a weight of 0.8, and a reference frequency between 50% and 80% is set to a weight of 0.5. For field associations with a reference frequency of less than 50%, a piecewise linear mapping strategy is adopted to determine the weight: For associations with a reference frequency between 20% and 50%, a linear interpolation method is used to calculate the weight value. The specific calculation formula is: weight = 0.2 + (reference frequency - 0.2). For example, the weight is 0.35 when the reference frequency is 35% and 0.42 when the reference frequency is 42%. For weak associations with a reference frequency of less than 20%, a minimum weight strategy is adopted, and the lower limit of the weight is set to 0.1. That is, associations with a reference frequency between 0 and 20% are uniformly assigned a base weight of 0.1. For edges with extremely low reference frequencies (below 5%) and determined by business experts to be accidental associations, they can be directly removed during the graph construction stage without establishing directed edge connections.

[0080] A multi-dimensional sensitivity assessment model is established based on semantic vectorized encoding and weighted field association graphs. This model comprises two core components: a single-field sensitivity assessment sub-model and an association sensitivity propagation sub-model. The single-field sensitivity assessment sub-model employs a multilayer perceptron structure, with the semantic vector of the field as input. Feature transformation is performed through a three-layer fully connected network, and the basic sensitivity score of the field is finally output. The hidden layer dimensions of the network are 256, 128, and 64, respectively. The ReLU activation function is used, and the output layer uses the Sigmoid function to normalize the scores to the range of 0 to 1. This sub-model is trained using supervised learning, and the training data includes various financial fields and their corresponding expert-annotated sensitivity levels.

[0081] The sensitivity propagation sub-model is designed based on the message passing mechanism of graph neural networks, performing sensitivity propagation calculations on the field association graph. This sub-model treats the field association graph as a directed weighted graph G=(V, E, W), where V is the set of nodes, E is the set of directed edges, and W is the set of edge weights. The propagation process adopts an iterative update method. In the k-th iteration, node i receives information from all neighboring nodes pointing to it. The sensitivity information is aggregated using a weighted summation function, and the updated sensitivity is obtained. The aggregation weight is determined by considering both the node's inherent sensitivity and the sensitivity influenced by neighbor propagation. It is based on the strength of the directed edge association. Current sensitivity of the source node The decision is made jointly. The aggregate weight is calculated using a product-fusion method, and the specific formula is: Aggregate Weight ,in This represents the weight of the directed edge association strength from node i to node j. This represents the current sensitivity value of source node i in the k-th iteration. This calculation method embodies a dual-effect mechanism: the association strength weight reflects the structural correlation between fields at the business logic level, and the current sensitivity value reflects the amount of sensitive information actually carried by the source node. Through multiplication, a significant propagation effect only occurs when both the edge association strength and the source node sensitivity are high. For example, if... and Then aggregate weight This indicates that node i has a strong sensitivity propagation effect on node j; if and Then aggregate weight The propagation impact is relatively weak. This product fusion mechanism ensures that sensitivity propagation both follows the constraints of the graph structure and dynamically responds to the real-time sensitivity status of nodes.

[0082] A single-field sensitivity assessment sub-model is used to calculate initial sensitivity scores for each data field. The semantic vectors of the fields are sequentially input into the sub-model to obtain the basic sensitivity scores for each field. For example, the "Bank Account Number" field has a calculated basic sensitivity score of 0.92, the "Customer Name" field has a score of 0.85, and the "Order Date" field has a score of 0.35. These basic sensitivity scores are then assigned as initial sensitivity attributes to the corresponding nodes in the field association graph, forming a graph structure with initial sensitivity information.

[0083] The sensitivity propagation sub-model performs multi-round iterative sensitivity propagation calculations on the field association graph. In each iteration, all nodes in the graph are traversed. For node i, the sensitivity information of all directed edges pointing to that node and their source nodes is collected. A weighted aggregation value is calculated by multiplying the sensitivity of the source node by the association strength weight of the corresponding edge and then summing the results. The aggregation value is then fused with the node's own basic sensitivity. The fusion ratio is controlled by the hyperparameter α, typically set to 0.7, indicating that 70% of the node's own sensitivity features are retained while absorbing 30% of the influence of neighbor propagation. The updated sensitivity is then assigned back to the node attributes, and the next iteration continues.

[0084] A convergence criterion is set to terminate the iteration process. After each iteration, the maximum value of the change in sensitivity of all nodes is calculated. When this maximum change is less than a preset threshold of 0.001, the sensitivity attribute is considered to have converged to a stable state. Convergence is usually achieved after 5 to 10 iterations. After convergence, the sensitivity attribute of each node is the comprehensive sensitivity score that considers the influence of inter-field relationships. For example, the basic sensitivity of the "Order Number" field is 0.4, but due to its strong correlation with the highly sensitive "Customer Number" field, the comprehensive sensitivity increases to 0.68 after propagation calculation.

[0085] The overall sensitivity score is bound to the corresponding data field to generate structured financial data carrying sensitivity score information. A sensitivity score attribute field is added to the metadata layer of the data table to store the overall sensitivity score value corresponding to each business field.

[0086] In one optional implementation, the structured financial data is subjected to differentiated encryption processing based on sensitivity scores. Data fields with different sensitivity score ranges are encrypted using different key systems to generate layered encrypted data, including:

[0087] Statistical analysis is performed on the sensitivity scores to determine the set of quantiles in the score distribution. Based on the set of quantiles, multiple sensitivity score intervals are divided, and a corresponding key system is constructed for each sensitivity score interval.

[0088] For each of the aforementioned sensitivity scoring intervals, a corresponding key derivation topology is established. The key derivation topology defines the derivation path and derivation level from the root key to the target encryption key, and the key derivation topology for different sensitivity scoring intervals has a derivation level depth adapted to its sensitivity.

[0089] Based on the sensitivity score of each data field, determine the sensitivity score interval to which it belongs, obtain the key derivation topology corresponding to the sensitivity score interval, and establish a mapping relationship between the data field and the corresponding key derivation topology.

[0090] Based on the mapping relationship, multi-level key derivation operations are performed along the derivation path defined by the corresponding key derivation topology to generate target encryption keys corresponding to their sensitivity score ranges for each data field.

[0091] The target encryption key is used to encrypt the data fields belonging to the corresponding sensitivity score range to obtain encrypted data fields. All encrypted data fields are then associated with and stored with the identification information of the key-derived topology to form hierarchical encrypted data.

[0092] like Figure 2 As shown, the method includes:

[0093] After completing the sensitivity scoring of structured financial data, differentiated encryption processing needs to be implemented for data fields with different sensitivity levels based on the scoring results. First, sensitivity scores for all data fields are collected to form a scoring sample set. Statistical analysis is then performed on this sample set to obtain the distribution characteristics of the scores. By calculating quartiles, deciles, or percentiles, key nodes in the score distribution are determined. For example, in a financial system containing 500 data fields, the sensitivity score range is 0 to 100. Statistical analysis yields the 25th percentile as 32, the 50th percentile as 58, the 75th percentile as 82, and the 90th percentile as 91. Based on these quantiles, sensitivity score intervals are divided into five ranges: low sensitivity [0, 32], low-to-medium sensitivity [32, 58], medium-to-high sensitivity [58, 82], high sensitivity [82, 91], and extremely high sensitivity [91, 100]. Each sensitivity score interval is configured with an independent key system to ensure that data with different sensitivity levels are encrypted using mutually isolated mechanisms.

[0094] A corresponding key derivation topology is established for each sensitivity score range. This topology adopts a tree-like hierarchical architecture, defining the derivation path from the root key to the final encryption key. For the low sensitivity range, a two-layer derivation structure is constructed. The root key directly generates the target encryption key through a single derivation function. The derivation function uses a hash-based key derivation function (HKDF), taking the root key and field identifier as input and outputting a 256-bit target encryption key. For the medium-low sensitivity range, a three-layer derivation structure is constructed. The root key first derives an intermediate key, and then the intermediate key, combined with a timestamp parameter, derives the target encryption key, increasing the derivation complexity. For the medium-high sensitivity range, a four-layer derivation structure is constructed, introducing a domain key layer and a business key layer. The derivation path is root key to domain key, then to business key, and finally to the target encryption key. For the high sensitivity range, a five-layer derivation structure is constructed, adding a scenario key layer on top of the four layers, bringing the derivation depth to five levels. For extremely sensitive regions, a six-layer derivation structure is constructed. At the highest layer, a seed key generated by a hardware security module is introduced as the root key source. The derivation process involves six layers: seed key, master key, domain key, block key, business key, and finally the target encryption key. Different derivation parameters and salt values ​​are used between each layer to ensure that even if a key at one layer is leaked, the security of other layers will not be affected.

[0095] Iterate through all structured financial data fields and determine the sensitivity score range to which each field belongs based on its sensitivity score. For example, the field "Employee Basic Salary" has a score of 45, belonging to the low-to-medium sensitivity range [32, 58); the field "Bank Account Balance" has a score of 88, belonging to the high sensitivity range [82, 91); and the field "Company Registered Address" has a score of 18, belonging to the low sensitivity range [0, 32]. Query the key-derived topology identifier corresponding to each sensitivity score range and establish a mapping table from fields to topologies. This mapping table records the field's unique identifier, field name, sensitivity score, range number, and corresponding topology identifier. The mapping table is stored using a hash index structure, supporting fast retrieval of the corresponding key-derived topology based on the field identifier, with a constant retrieval time complexity.

[0096] Based on the established mapping relationship, key derivation operations are performed on each data field to generate the target encryption key. For fields belonging to the low-sensitivity range, their two-layer derivation topology is obtained, the root key is read from the secure storage area, and the field identifier is input as the derivation parameter. HKDF is then called to perform a derivation operation to obtain the target encryption key. For fields belonging to the medium-high sensitivity range, their four-layer derivation topology is obtained, and derivation from the root key to the domain key, from the domain key to the business key, and from the business key to the target encryption key are performed sequentially. Each derivation layer incorporates contextual information such as the current timestamp, field attribute type, and data owner identifier as derivation parameters to ensure the uniqueness and unpredictability of the derivation results. For fields in the extremely high sensitivity range, the seed key is obtained from the hardware security module, and the derivation is performed sequentially through six layers, with an interval of no less than 10 milliseconds between each layer to avoid timing attacks, ultimately generating the target encryption key. All derivation processes record the derivation path and the hash digest of each layer's key, stored in the audit log, but the plaintext of the key is not recorded.

[0097] The generated target encryption key is used to encrypt the corresponding data fields. For structured numerical fields, an order-preserving encryption algorithm is used to complete encryption while ensuring the comparability of ciphertext, supporting range queries in the encrypted state. For string fields, AES-GCM authentication encryption mode is used with a key length of 256 bits, generating a 128-bit authentication tag to ensure the confidentiality and integrity of the ciphertext. For date and time fields, a format-preserving encryption algorithm is used to ensure that the encrypted data retains the date and time format, facilitating application layer processing. After encryption, the encrypted data fields, the corresponding key derivation topology identifier, the hash digest of the derivation path, the encryption algorithm identifier, initialization vector, and other information are assembled into an encrypted data packet. All encrypted data packets are stored according to sensitivity score intervals, forming a hierarchical encrypted data structure. This structure logically isolates data of different sensitivity levels. During physical storage, different security levels of storage media can be selected according to the sensitivity level. High-sensitivity data is stored on encrypted hard drives or hardware security modules, while low-sensitivity data is stored in a regular database, achieving layered protection at the storage level.

[0098] In one optional implementation, based on the mapping relationship, multi-level key derivation operations are performed along the derivation path defined by the corresponding key derivation topology to generate target encryption keys for each data field corresponding to its sensitivity score range, including:

[0099] Based on the mapping relationship, determine the corresponding key derivation topology for each data field, and extract the root key identifier and derivation path information from the key derivation topology.

[0100] The derivation path information is structured and parsed to extract the derivation transformation rules of each derivation level and the key material transfer relationship that defines the key material transfer method between levels;

[0101] The root key is obtained based on the root key identifier, and the root key is input as the initial key material into the first derivation level of the key derivation topology to start the key derivation process;

[0102] Following the derivation hierarchy order defined by the derivation path information, the derivation transformation rule corresponding to the current derivation hierarchy is obtained, and the key material received at the current derivation hierarchy is transformed according to the derivation transformation rule to generate output key material.

[0103] According to the key material transfer relationship, the output key material of the current derivation level is transferred to the next derivation level as its input key material. The derivation process is repeated until all derivation levels are processed, and the key material output by the final derivation level is determined as the target encryption key.

[0104] After establishing the mapping relationship between sensitivity score intervals and key derivation topologies, specific key derivation operations need to be performed to generate encryption keys corresponding to each data field. For each data field to be processed, its corresponding key derivation topology is first determined based on the established mapping relationship. This mapping relationship can be represented as a lookup table or a hash mapping. By inputting the sensitivity score of the data field or its corresponding score interval, the corresponding topology identifier can be quickly located.

[0105] Once the target topology is determined, key information needs to be extracted from its definition to guide subsequent derivation operations. The key derivation topology is typically stored in a structured data format, containing a root key identifier field and a derivation path information field. The root key identifier is used to locate the root key stored in the key management system; it can be a unique key number, a hash digest, or a reference path to the key's storage location. The derivation path information describes the complete derivation chain from the root key to the final target key, usually represented by a tree or directed graph structure, where each node represents a derivation level, and edges represent the derivation relationships between levels.

[0106] Structured parsing of the derivation path information is the core preparatory step in the key derivation process. This parsing process requires traversing all nodes in the derivation path and extracting the derivation transformation rules for each derivation level. These rules define how to transform input key material into output key material. Common transformation rules include derivation algorithms based on cryptographic hash functions, derivation algorithms based on key derivation functions such as HKDF, or derivation algorithms based on cryptographic message authentication codes such as HMAC. Specifically, the transformation rule for a particular derivation level specifies using the SHA-256 hash function to combine the input key material with specific derivation parameters, or using the HKDF algorithm combined with context information strings for key expansion.

[0107] While extracting the derivation transformation rules, it is necessary to parse and define the key material transfer relationships that define the key material transfer methods between levels. These relationships specify how the output key material generated at the current level is transferred to the next level, including the data format, whether encoding conversion is required, and whether additional derivation context information is needed. In some scenarios, the output of a derivation level needs to be transferred to multiple downstream levels simultaneously. In this case, the transfer relationships will clearly define the number of key materials received by each downstream level and their format requirements.

[0108] After parsing the path information, the root key is retrieved from the key management system using the root key identifier. The key management system can be a hardware security module, a key vault, or a distributed key management service; the corresponding root key data can be obtained by providing the root key identifier as a query parameter. The obtained root key, as the initial key material, is input into the first derivation level defined in the key derivation topology, marking the formal start of the key derivation process. The format of the initial key material is typically a fixed-length byte sequence, such as 256-bit or 512-bit binary data.

[0109] The key derivation process is executed hierarchically according to the derivation path information. For the currently processed derivation level, the corresponding derivation transformation rule is first obtained. This rule specifies the type of transformation algorithm and related parameter configurations. The received key material is processed according to the transformation rule. For example, if the transformation rule specifies the use of the HMAC-SHA256 algorithm, the current key material is used as the key input, combined with a predefined derivation tag or counter value as the message input, and the calculated HMAC output is the output key material for that level. If the transformation rule requires the use of the HKDF algorithm, the salt value used in the extraction step and the information parameters used in the expansion step need to be specified. The required length of output key material is generated through a two-stage process.

[0110] In certain derivation levels, transformation rules involve binding operations with data field attributes, that is, introducing metadata information of the data fields, such as field name, field type, or field path, into the derivation calculation. This binding mechanism allows different data fields within the same sensitivity range to generate independent encryption keys even when using the same derivation topology, thereby improving key isolation. In practice, the field metadata information can be encoded into a byte sequence and used as an additional input parameter to the derivation transformation function for computation.

[0111] After the current derivation level completes the transformation, the generated output key material is passed to the next derivation level according to the key material transfer relationship. This transfer operation involves data format conversion, such as encoding the binary key material into Base64 or hexadecimal format to adapt to the input requirements of downstream levels. If the transformation rules of the next derivation level require additional context information, the transfer process also needs to prepare the corresponding context data and transfer it along with the key material. Upon receiving the transferred key material, the next derivation level uses it as its input key material to continue performing the derivation transformation.

[0112] The key derivation process is executed sequentially according to the hierarchical order of the derivation path. After each level completes its derivation transformation, its output is passed downstream, forming the flow of key material within the derivation topology. For derivation topologies with multiple branches, each level needs to be processed in the order of topological sorting, ensuring that each level only begins execution after all its upstream levels have been processed. The derivation process is repeated until all derivation levels are processed. At this point, the key material output by the final derivation level, i.e., the endpoint of the derivation path, is the target encryption key for the target data field. This target encryption key corresponds to the sensitivity score range of the data field, and its derivation depth is proportional to the algorithm complexity and sensitivity level, thereby achieving a differentiated key strength protection strategy.

[0113] In one optional implementation, based on the access information and the output of the multi-dimensional sensitivity assessment model, a dynamic desensitization rule set for the current access scenario is generated, including:

[0114] The access subject characteristics and access behavior characteristics are extracted from the access information, and the sensitivity metrics of each data field are extracted from the output of the multi-dimensional sensitivity assessment model. The access subject characteristics and access behavior characteristics are fused to generate an access risk assessment vector for assessing the current access risk.

[0115] Based on the access risk assessment vector and the sensitivity quantification index, a desensitization intensity mapping function is constructed to determine the degree of desensitization processing. The desensitization intensity mapping function will dynamically calculate the desensitization intensity according to the combined characteristics of access risk and data sensitivity.

[0116] The sensitivity metrics of each data field and the access risk assessment vector are input as input parameters to the desensitization strength mapping function, and the target desensitization strength that each data field should adopt in the current access scenario is calculated.

[0117] Based on the target desensitization intensity, determine the desensitization operation type that is appropriate to the desensitization intensity for each data field, and set the corresponding desensitization parameter configuration based on the desensitization operation type to ensure that the desensitization effect meets the security requirements;

[0118] The de-identification operation type and the de-identification parameter configuration are integrated into the system to generate a dynamic de-identification rule set for the current access scenario.

[0119] When generating a dynamic de-identification rule set for the current access scenario, key feature elements are extracted from the access information. The extraction of access subject features covers dimensions such as user identity, role and permission level, department type, and historical access reputation score. For example, for access subjects in a medical system, information such as job category (doctor, nurse, administrative staff), professional qualification level, and department affiliation can be extracted; for enterprise data access scenarios, attributes such as employee ID, job level, and business line affiliation are extracted. The extraction of access behavior features focuses on elements such as access time period (working or non-working hours), access frequency, data request scale, access source address, and type of terminal device used. These features are processed through feature vectorization, converting textual categorical data into numerical representations, providing standardized input for subsequent risk assessment.

[0120] When extracting sensitivity metrics for each data field from the output of a multi-dimensional sensitivity assessment model, a complete metric system is needed, including business sensitivity scores, privacy sensitivity scores, compliance sensitivity scores, and overall sensitivity weights. Each data field corresponds to a set of sensitivity metrics, which are typically normalized to a range of 0 to 1. For example, the ID number field has quantitative metrics of 0.6 for business sensitivity, 0.95 for privacy sensitivity, and 0.9 for compliance sensitivity, resulting in a weighted overall sensitivity of 0.85.

[0121] The fusion of visitor characteristics and visit behavior characteristics employs a multi-level fusion strategy. The first level standardizes the features to eliminate the influence of different feature dimensions. The second level assigns differentiated weight coefficients based on the contribution of each feature to risk assessment. This weighting is typically determined through statistical analysis of historical security incident data; for example, the weight coefficient for visits outside of working hours can be set to 1.5, while the weight coefficient for visits from abnormal geographical locations can be set to 2.0. The third level concatenates or sums the weighted feature vectors to generate a comprehensive visit risk assessment vector. The dimensional design of this vector needs to balance expressive power and computational efficiency; in practice, 8 to 16-dimensional vector representations are commonly used.

[0122] When constructing a desensitization intensity mapping function, a mathematical model reflecting the synergistic effect of access risk and data sensitivity needs to be designed. This function takes an access risk assessment vector and a data field sensitivity metric as input, and the output is typically defined as a discrete intensity level from 0 to 5, where 0 indicates no desensitization is needed and 5 represents the highest intensity of desensitization. The mapping function can be designed using a piecewise linear function, a non-linear sigmoid function, or a query method based on a rule table. For example, when the overall risk value of the access risk assessment vector is below the threshold of 0.3 and the data sensitivity is below 0.4, the mapping output intensity level is 1; when the risk value is between 0.3 and 0.6 and the sensitivity is between 0.4 and 0.7, the mapping output intensity level is 3; and when the risk value exceeds 0.8 or the sensitivity exceeds 0.9, the mapping output intensity level is 5. To enhance the function's adaptability, a dynamic adjustment mechanism can be introduced to fine-tune the mapping relationship based on real-time monitoring of data breach events.

[0123] When calculating the sensitivity metrics and access risk assessment vectors of each data field as input parameters, a field-by-field traversal approach is adopted. For each field in the dataset, its corresponding sensitivity metric is extracted, combined with the risk assessment vector of the current access scenario, and substituted into the desensitization strength mapping function for calculation. For example, for data records containing five fields—name, ID number, mobile phone number, home address, and consumption amount—the target desensitization strengths are calculated to be 1, 5, 4, 3, and 2, respectively. The calculation of the target desensitization strength uses a weighted product normalization method, and the specific calculation process consists of three steps: First, extract the comprehensive sensitivity of the field. (Value range 0 to 1) and the comprehensive risk assessment value of the current access scenario. (Value range: 0 to 1); Second step: Calculate the desensitization demand intensity index. The first step involves converting the desensitization requirement intensity index I into discrete intensity levels through segmented mapping. The mapping rules are as follows: intensity level 0 is I < 0.1, intensity level 1 is I < 0.25, intensity level 2 is I < 0.45, intensity level 3 is I < 0.45, intensity level 4 is I < 0.65, and intensity level 5 is I ≥ 0.85. For example, if the overall sensitivity of the name field is 0.6, the access risk assessment value is 0.4, and the scenario adjustment coefficient is 1.3, then I = 0.6 × 0.4 × 1.3 = 0.312, which maps to intensity level 2. However, if the name field is marked as a non-critical display field in the rule configuration, it can be downgraded to intensity level 1. The overall sensitivity of the ID number field is 0.95, and the access risk assessment value is 0.7. Therefore, I = 0.95 × 0.7 × 1.3 = 0.8645, which is mapped to strength level 5, and adopts full masking or encryption processing.

[0124] The calculation process needs to consider the correlation between fields. When multiple field combinations pose a risk of identity re-identification, the anonymization strength of the associated fields should be appropriately increased to ensure that the anonymity of the overall dataset meets privacy protection standards such as k-anonymity or l-diversity. The mechanism for increasing the anonymization strength of associated fields adopts a quasi-identifier detection and strength increment compensation strategy, which is implemented in four stages: In the first stage, by analyzing the semantic and statistical characteristics of the dataset's fields, the set of fields that may constitute quasi-identifier combinations is identified, such as combinations of {age, gender, postal code} and {date of birth, occupation, workplace}. These fields may have low anonymization strength individually, but when combined, they can uniquely identify an individual. In the second stage, a k-anonymity verification test is performed on the identified quasi-identifier combinations. The equivalence class size distribution of the dataset under the current anonymization strength configuration is calculated. If there are equivalence class sizes smaller than the preset k value (usually k≥5), the anonymization strength is increased. If the situation is deemed unfavorable, a risk of re-identification is identified. In the third stage, for quasi-identifier combinations that do not meet k-anonymity, an incremental strength compensation algorithm is initiated. This algorithm prioritizes increasing the strength of fields with the least impact on data availability. The specific selection strategy is based on the field's information entropy and business value score, prioritizing fields with lower information entropy and lower business value scores. The strength increment is typically set to 1 to 2 levels. In the fourth stage, after the strength increase, k-anonymity and l-diversity indicators are re-verified. If the requirements are still not met, the strength of other related fields is iteratively increased until the privacy protection standard is met or all related fields reach the highest strength level of 5. For example, the initial anonymization strengths of the {age, gender, zip code} combination in the dataset are 2, 1, and 2 respectively. Verification reveals that the minimum equivalence class size is 3, which does not meet the k=5 requirement. The algorithm calculates that the zip code field has the lowest information entropy, increasing its strength from 2 to 4 (generalizing to the city level). After re-verification, the minimum equivalence class size reaches 6, meeting the k-anonymity requirement. If this is still not met, the strength of the age field is further increased from 2 to 3 (generalizing to the 10-year-old age group) to ensure that the privacy protection standard is met.

[0125] When determining the specific de-identification operation type based on the calculated target de-identification strength, a mapping relationship between strength level and operation type is established. Strength level 1 typically corresponds to partial masking operations, such as keeping the surname but hiding the given name; strength level 2 corresponds to randomization operations that preserve formatting, such as replacing the middle four digits of a phone number with random numbers; strength level 3 corresponds to generalization processing, such as generalizing the precise birth date to the birth year; strength level 4 corresponds to complete masking or hash encryption; and strength level 5 corresponds to data deletion or replacement with fixed placeholders. Each operation type requires configuration of corresponding de-identification parameters, including the selection of masking characters (asterisk, hash symbol, or other symbols), the number of digits to retain, the generalization granularity, the hash algorithm type, and the salting strategy. Parameter configuration should follow the principle of consistency to ensure that the same strength level produces equivalent security protection effects on different fields.

[0126] The process of configuring anonymization parameters requires a comprehensive consideration of the balance between data availability and security. For scenarios where the statistical analysis value of the data needs to be preserved, priority should be given to methods that preserve the format and use encryption or differential privacy by adding noise. For scenarios where only display is required and no further calculation is needed, stronger masking or generalization strategies can be adopted. For example, if the age field in medical data is used for disease statistical analysis, parameters that generalize to the 5-year or 10-year age range can be set; if it is only used for displaying patient information, it can be directly masked or deleted.

[0127] The system integration of de-identification operation types and parameter configurations is achieved by generating a structured rule description file. This file, in JSON or XML format, contains complete information such as field identifiers, sensitivity indicators, target de-identification strength, operation type, parameter configuration, and execution priority. The rule set is organized in a hierarchical structure: the top layer is the access scenario identifier, the middle layer is the data table or data category, and the bottom layer is the de-identification rule entries for specific fields. The integration process also requires establishing a rule version management mechanism to record metadata such as rule generation time, triggering conditions, and validity period, supporting rule auditing, traceability, and dynamic updates. The generated dynamic de-identification rule set is ultimately provided to the data access control module in the form of an interface to achieve automated execution of real-time de-identification processing.

[0128] In one optional implementation, based on the dynamic de-identification rule set and key index relationship, target data fields are selectively decrypted from the cloud storage resource pool, and the decrypted data fields are processed according to the de-identification transformation method defined in the rule set to generate de-identified financial data, including:

[0129] Extract the desensitization transformation method corresponding to each data field from the dynamic desensitization rule set, construct the association mapping table between the data field and the desensitization transformation method, and establish a distributed key retrieval mechanism based on intelligent load balancing.

[0130] Based on the distributed key retrieval mechanism and key index relationship, the storage location of each target data field in the cloud storage resource pool and the corresponding key acquisition path are determined, and the key derivation level sequence required for decryption key is obtained from the key management node.

[0131] Access the key management node along the key acquisition path, and use an adaptive key recombination algorithm to dynamically combine and calculate the acquired key materials at each level to generate a decryption key for decrypting the target data field;

[0132] The decryption key is used to decrypt the target data field at the corresponding storage location in the cloud storage resource pool to obtain the decrypted data field.

[0133] Based on the association mapping table, the desensitization processing parameters corresponding to each of the decrypted data fields are obtained, and the desensitization processing parameters are applied to the desensitization transformation method to perform desensitization transformation operation on each of the decrypted data fields to generate desensitized financial data.

[0134] Based on the generated dynamic desensitization rule set and key index relationship, it is necessary to selectively decrypt target data fields from the cloud storage resource pool and complete subsequent desensitization processing. First, the desensitization transformation methods corresponding to each data field are extracted from the dynamic desensitization rule set, establishing refined desensitization strategies for fields with different levels of sensitivity in financial data. For amount fields, a numerical range transformation method can be used to convert precise amounts into range representations; for account number fields, a partial character masking method can be used, retaining the first four and last three digits while replacing the middle part with a specific symbol; for personnel name fields, a surname retention method can be used, replacing the given name with an asterisk. Based on these extracted desensitization transformation methods, a mapping table is constructed between data fields and desensitization transformation methods. This mapping table uses a key-value pair structure, where the key is the data field identifier, and the value is a structured object containing information such as desensitization transformation type, parameter configuration, and priority. Simultaneously, a distributed key retrieval mechanism based on intelligent load balancing is established. This mechanism monitors the real-time load status of each key management node, including indicators such as processor utilization, memory usage, and network bandwidth utilization, and calculates the comprehensive load score for each node. The distribution strategy for key retrieval requests is dynamically adjusted based on load scores, prioritizing the routing of key acquisition requests to nodes with lower loads to avoid retrieval delays caused by single-point overload.

[0135] Based on the established distributed key retrieval mechanism and key index relationship, the specific storage location of each target data field in the cloud storage resource pool is determined. By parsing the storage location identifier recorded in the key index relationship, the specific storage block number and offset address in the cloud storage resource pool are located. For data fields stored on different storage nodes, a storage location descriptor containing information such as node address, storage block identifier, data start offset, and data length is generated. Simultaneously, based on the key hierarchy information in the key index relationship, the corresponding key acquisition path is determined. This path describes the complete traversal sequence from the top-level key management node down to the leaf node. The key derivation hierarchy sequence required to obtain the decryption key from the key management node is obtained. This sequence contains a complete hierarchy of master key layer, intermediate derivation layer, and leaf key layer, with each layer recording the storage identifier and derivation parameters of the corresponding key material.

[0136] Following the established key acquisition path, key management nodes are accessed sequentially according to hierarchy. An authentication request is initiated to the top-level master key management node, and the master key material is obtained upon successful authentication. This master key material is not directly used for decryption but serves as the root key for deriving other levels of keys. Subsequently, derived key materials are requested from intermediate derivation layer nodes, carrying the hash digest of the master key material and the identifier information of the current data field in the request. After verifying the validity of the request, the intermediate derivation layer nodes use key derivation functions to generate intermediate-level key materials based on the master key material. Finally, leaf key layer nodes are accessed to obtain the leaf key material directly associated with the target data field. An adaptive key reassembly algorithm is used to dynamically combine the key materials obtained from each level. This algorithm adaptively selects the combination strategy based on the length and type of the key materials. For fixed-length key materials, bitwise XOR operations are used for combination; for variable-length key materials, a hash chain combination method is used, using the upper-level key material as the input parameter for lower-level key derivation, iteratively generating the final decryption key layer by layer. During the combination operation, a data field identifier is introduced as a salt value to ensure that even with the same key level material, a unique decryption key can be generated for different data fields.

[0137] The generated decryption key is used to decrypt the target data field at the corresponding storage location in the cloud storage resource pool. Based on the information in the storage location descriptor, a data read request is initiated to the cloud storage resource pool to retrieve the encrypted target data field. The corresponding decryption algorithm is selected based on the encryption algorithm used during encryption; typically, AES-256 symmetric encryption is used for data decryption. The decryption key is used as the algorithm input, along with the initialization vector used during encryption, to perform decryption operations on the read ciphertext data, outputting the plaintext data field. After decryption, the plaintext data is verified for integrity by comparing the checksum attached to the decrypted data with the recalculated checksum value to verify the correctness of the decryption process and that the data has not been tampered with.

[0138] Based on the previously constructed association mapping table, the anonymization parameters for each decrypted data field are queried. For the amount field, the anonymization parameters include the interval granularity value and the number of decimal places to retain; for the account number field, the parameters include the prefix length to retain, the suffix length to retain, and the masking symbol character; for the name field, the parameters include the character position to retain and the type of symbol to replace. These anonymization parameters are applied to the corresponding anonymization transformation methods, and the anonymization transformation operation is performed on each decrypted data field. For amount anonymization, the range to which the amount value belongs is determined, and the precise amount is replaced with a range representation; for account number anonymization, the first four and last three characters are extracted, and the middle part is replaced with an asterisk; for name anonymization, the surname is retained, and the given name is replaced with an asterisk. After completing the anonymization transformation of all fields, the anonymized data fields are assembled according to the structure of the original financial data to generate anonymized financial data. This data retains the necessary business analysis value while effectively protecting sensitive information from direct exposure.

[0139] In one optional implementation, the key management node is accessed along the key acquisition path, and an adaptive key reassembly algorithm is used to dynamically combine and calculate the acquired key materials at each level to generate a decryption key for decrypting the target data field, including:

[0140] Following the key derivation hierarchy sequence defined by the key acquisition path, the key management node is accessed sequentially, the corresponding key material is obtained from each key derivation hierarchy, and the obtained key material is preliminarily structured through an adaptive key reorganization algorithm to form a hierarchical key material set adapted to the current system load characteristics.

[0141] Deep structural feature extraction is performed on each level of key material in the hierarchical key material set to obtain the entropy distribution characteristics of each level of key material and the real-time load status of the system. Based on the entropy distribution characteristics and the real-time load status of the system, the contribution weight of each level of key material in the subsequent dynamic combination operation is calculated.

[0142] Based on the contribution weight, the key materials at each level in the hierarchical key material set are intelligently weighted to generate a weighted key material sequence that reflects the importance of each level of key material.

[0143] Based on the adaptive key reorganization algorithm, the resource optimization configuration of each weighted key material in the weighted key material sequence is performed, and the layer-by-layer fusion operation is performed according to the hierarchical order defined by the key derivation hierarchical sequence. The weighted key material of the current level is adaptively mixed and transformed with the fusion result of the previous level to generate the fusion output of the current level.

[0144] Repeat the layer-by-layer fusion operation until all levels of weighted key material have been processed, and determine the final level fusion output as the dynamically adaptive decryption key.

[0145] During key reassembly, the corresponding key management nodes need to be accessed sequentially according to the key derivation hierarchy sequence defined by the key acquisition path. Specifically, the key derivation hierarchy sequence contains a three- to five-level hierarchical structure, with each level corresponding to the storage address of one or more key management nodes. When accessing the first-level key management node, a secure connection channel is established through the node identifier, and the first-level key material is obtained from that node. This key material is typically a 128- to 256-bit binary data block. After obtaining the first-level key material, the pointer information of the next-level node contained in the key material is used to locate and access the second-level key management node. This process continues until the key material of all levels is obtained. During the acquisition process, the adaptive key reassembly algorithm performs preliminary structuring processing on the obtained key material. This processing includes format standardization, length normalization, and metadata extraction of the key material. Format standardization converts the key material returned by different nodes into a standard binary block format, and length normalization ensures that the effective number of bits in each key material meets the requirements of subsequent operations. After initial structuring, a hierarchical key material set is formed. Each element in this set corresponds to a key derivation level and carries metadata information for that level, including the level number, timestamp, and node load status. This hierarchical organization allows subsequent key reassembly operations to adapt to the current system's load characteristics, avoiding performance bottlenecks under high load scenarios.

[0146] When performing deep structural feature extraction on the key materials at each level of the hierarchical key material set, entropy is calculated for each key material. Entropy calculation involves statistically analyzing the frequency of occurrence of each byte in the key material and calculating a randomness measure based on information entropy theory. The entropy distribution feature includes not only the overall entropy value but also the local entropy distribution across different byte segments. For example, a 256-bit key material is divided into eight 32-bit segments, and the local entropy value of each segment is calculated, forming an eight-dimensional entropy distribution vector. The entropy distribution feature reflects the quality of the key material; a high entropy value indicates higher randomness and security strength. Simultaneously, the real-time system load status is obtained through a monitoring interface. This status includes current CPU utilization, memory usage, network bandwidth utilization, and the response latency of the key management node. The real-time system load status is represented by standardized numerical values ​​ranging from 0 to 1, where 0 represents idle and 1 represents full load. Based on the entropy distribution feature and the real-time system load status, the contribution weight of each level of key material in subsequent dynamic combination operations is calculated. The calculation of contribution weights comprehensively considers the security strength of the key materials and the current computing power of the system. This ensures that when the system load is high, key materials with lower processing complexity but still meeting security requirements are prioritized, while when the system load is low, the security advantages of high-entropy key materials are fully utilized. Specifically, the contribution weights are calculated by weighted summation of the principal components of the entropy distribution characteristics and the reciprocal of the system's real-time load state. The weighting coefficients are dynamically determined based on historical computation data through an adaptive adjustment mechanism.

[0147] Based on the calculated contribution weights, intelligent weighting is applied to the key materials at each level of the hierarchical key material set. Intelligent weighting generates weighted key materials by numerically mapping the original binary data of each key material to its corresponding contribution weight. In one specific embodiment, the weighting process employs bit-modulation, adjusting the influence of key bits in the key material according to the magnitude of the contribution weight. Key materials with higher contribution weights have their key bits given greater weight in subsequent fusion operations, while key materials with lower contribution weights have their bit influence reduced accordingly. Through this weighting process, a weighted key material sequence reflecting the importance of each level of key material is generated. This sequence maintains the order of the original key derivation hierarchy sequence, but each element carries data content that has undergone intelligent weighting.

[0148] The adaptive key reassembly algorithm optimizes resource allocation for each weighted key material in the weighted key material sequence. Resource optimization allocates appropriate computational resources to each weighted key material based on the available computing resources and real-time load of the system. When the system load is low, more computational resources are allocated to high-weighted key materials, employing a more complex hybrid transformation algorithm. When the system load is high, the computational complexity is automatically reduced, using a lightweight hybrid transformation algorithm. After resource allocation, layer-by-layer fusion operations are performed according to the hierarchical order defined in the key derivation hierarchy sequence. The layer-by-layer fusion operation starts from the first level, using the weighted key materials of the first level as the initial fusion input. When processing the second level, the weighted key materials of the second level are adaptively hybridized with the fusion result of the first level.

[0149] The adaptive hybrid transformation employs a combination of hash mixing, XOR combination, and shift diffusion. Specifically, the fusion result of the preceding level is used as the seed input and XORed with the weighted key material of the current level. Then, a cyclic shift operation is performed on the XOR result, with the shift number dynamically determined based on the current level's sequence number. Finally, hash mixing is performed on the shifted result to generate the fusion output of the current level. This fusion output serves as the fusion result of the preceding level for the next level's fusion operation. This layer-by-layer fusion operation is repeated until all levels of weighted key material have been processed. The final level's fusion output possesses the comprehensive characteristics of the key material from each level and achieves dynamic response to system load through an adaptive adjustment mechanism. The final level's fusion output is used as the decryption key, with a length identical to the encryption key length of the target data field, for subsequent data decryption operations. The entire key reassembly process achieves a dynamic balance between security and performance through an adaptive mechanism, ensuring efficient generation of high-quality decryption keys under different load scenarios.

[0150] A second aspect of this invention provides a cloud-based integrated financial data security storage system, comprising:

[0151] The data parsing unit is used to acquire financial data to be stored, perform semantic parsing on the financial data to extract the business attributes and relationships of the data fields; based on the business attributes and relationships, construct a multi-dimensional sensitivity assessment model, and perform sensitivity quantification scoring on each field in the financial data to obtain structured financial data.

[0152] The layered encryption unit is used to perform differentiated encryption processing on the structured financial data according to the sensitivity score, and to encrypt data fields with different key systems for different sensitivity score ranges to generate layered encrypted data;

[0153] The access information unit is used to extract the access subject's identity, access purpose, and access spatiotemporal context information from the access request when an access request is received; and to generate a dynamic desensitization rule set for the current access scenario based on the access information and the output of the multi-dimensional sensitivity assessment model.

[0154] The selective decryption unit is used to selectively decrypt target data fields from the cloud storage resource pool according to the dynamic desensitization rule set and key index relationship, and process the decrypted data fields according to the desensitization transformation method defined in the rule set to generate desensitized financial data;

[0155] The model optimization unit is used to return the anonymized financial data to the accessing subject, and at the same time record the access frequency and abnormal access patterns of the data fields during this access process, and use the access records to optimize and update the multi-dimensional sensitivity assessment model.

[0156] A third aspect of the present invention provides an electronic device, comprising:

[0157] processor;

[0158] Memory used to store processor-executable instructions;

[0159] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0160] 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.

[0161] 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.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cloud computing-based method for secure storage of integrated financial data, characterized in that: include: Obtain the financial data to be stored, and perform semantic parsing on the financial data to extract the business attributes and relationships of the data fields; Based on the business attributes and the relationship, a multi-dimensional sensitivity assessment model is constructed, and the sensitivity of each field in the financial data is quantitatively scored to obtain structured financial data. The structured financial data is subjected to differentiated encryption processing based on sensitivity scores. Data fields with different sensitivity score ranges are encrypted using different key systems to generate layered encrypted data. When an access request is received, extract the access subject's identity, access purpose, and access spatiotemporal context information from the request. Based on the access information and the output of the multi-dimensional sensitivity assessment model, a dynamic desensitization rule set is generated for the current access scenario; Based on the dynamic desensitization rule set and key index relationship, target data fields are selectively decrypted from the cloud storage resource pool, and the decrypted data fields are processed according to the desensitization transformation method defined in the rule set to generate desensitized financial data; The anonymized financial data is returned to the accessing entity, and the access frequency and abnormal access patterns of the data fields during this access process are recorded. The access records are then used to optimize and update the multi-dimensional sensitivity assessment model.

2. The method according to claim 1, characterized in that, Based on the aforementioned business attributes and relationships, a multi-dimensional sensitivity assessment model is constructed, and sensitivity metrics are quantitatively scored for each field in the financial data, resulting in structured financial data including: The business attributes are semantically vectorized and encoded, and a field association graph is constructed. The association strength weight is marked on each directed edge of the field association graph. Based on the semantic vectorized encoding and the field association graph with association strength weights, a multi-dimensional sensitivity assessment model is established, including a single-field sensitivity assessment sub-model and an association sensitivity propagation sub-model. The initial sensitivity score is calculated by using the single-field sensitivity assessment sub-model to perform semantic vectorization encoding of each data field, and the basic sensitivity score of each field is obtained. The basic sensitivity score is then assigned as the initial sensitivity attribute of the corresponding node in the field association graph. The sensitivity propagation sub-model performs sensitivity propagation calculation on the field association graph. Based on the association strength weight of the directed edge and the basic sensitivity score of the connected node, the sensitivity attribute of each node is iteratively updated until the sensitivity attribute converges, and a comprehensive sensitivity score considering the association effect between fields is obtained. The comprehensive sensitivity score is bound to the corresponding data fields to generate structured financial data carrying sensitivity score information.

3. The method according to claim 1, characterized in that, The structured financial data is subjected to differentiated encryption processing based on sensitivity scores. Data fields with different sensitivity score ranges are encrypted using different key systems to generate layered encrypted data, including: Statistical analysis is performed on the sensitivity scores to determine the set of quantiles in the score distribution. Based on the set of quantiles, multiple sensitivity score intervals are divided, and a corresponding key system is constructed for each sensitivity score interval. For each of the aforementioned sensitivity scoring intervals, a corresponding key derivation topology is established. The key derivation topology defines the derivation path and derivation level from the root key to the target encryption key, and the key derivation topology for different sensitivity scoring intervals has a derivation level depth adapted to its sensitivity. Based on the sensitivity score of each data field, determine the sensitivity score interval to which it belongs, obtain the key derivation topology corresponding to the sensitivity score interval, and establish a mapping relationship between the data field and the corresponding key derivation topology. Based on the mapping relationship, multi-level key derivation operations are performed along the derivation path defined by the corresponding key derivation topology to generate target encryption keys corresponding to their sensitivity score ranges for each data field. The target encryption key is used to encrypt the data fields belonging to the corresponding sensitivity score range to obtain encrypted data fields. All encrypted data fields are then associated with and stored with the identification information of the key-derived topology to form hierarchical encrypted data.

4. The method according to claim 3, characterized in that, Based on the mapping relationship, multi-level key derivation operations are performed along the derivation path defined by the corresponding key derivation topology to generate target encryption keys for each data field corresponding to its sensitivity score range, including: Based on the mapping relationship, determine the corresponding key derivation topology for each data field, and extract the root key identifier and derivation path information from the key derivation topology. The derivation path information is structured and parsed to extract the derivation transformation rules of each derivation level and the key material transfer relationship that defines the key material transfer method between levels; The root key is obtained based on the root key identifier, and the root key is input as the initial key material into the first derivation level of the key derivation topology to start the key derivation process; Following the derivation hierarchy order defined by the derivation path information, the derivation transformation rule corresponding to the current derivation hierarchy is obtained, and the key material received at the current derivation hierarchy is transformed according to the derivation transformation rule to generate output key material. According to the key material transfer relationship, the output key material of the current derivation level is transferred to the next derivation level as its input key material. The derivation process is repeated until all derivation levels are processed, and the key material output by the final derivation level is determined as the target encryption key.

5. The method according to claim 1, characterized in that, Based on the access information and the output of the multi-dimensional sensitivity assessment model, a dynamic desensitization rule set is generated for the current access scenario, including: The access subject characteristics and access behavior characteristics are extracted from the access information, and the sensitivity metrics of each data field are extracted from the output of the multi-dimensional sensitivity assessment model. The access subject characteristics and access behavior characteristics are fused to generate an access risk assessment vector for assessing the current access risk. Based on the access risk assessment vector and the sensitivity quantification index, a desensitization intensity mapping function is constructed to determine the degree of desensitization processing. The desensitization intensity mapping function will dynamically calculate the desensitization intensity according to the combined characteristics of access risk and data sensitivity. The sensitivity metrics of each data field and the access risk assessment vector are input as input parameters to the desensitization strength mapping function, and the target desensitization strength that each data field should adopt in the current access scenario is calculated. Based on the target desensitization intensity, determine the desensitization operation type that is appropriate to the desensitization intensity for each data field, and set the corresponding desensitization parameter configuration based on the desensitization operation type to ensure that the desensitization effect meets the security requirements; The de-identification operation type and the de-identification parameter configuration are integrated into the system to generate a dynamic de-identification rule set for the current access scenario.

6. The method according to claim 1, characterized in that, Based on the dynamic de-identification rule set and key index relationship, target data fields are selectively decrypted from the cloud storage resource pool, and the decrypted data fields are processed according to the de-identification transformation method defined in the rule set to generate de-identified financial data, including: Extract the desensitization transformation method corresponding to each data field from the dynamic desensitization rule set, construct the association mapping table between the data field and the desensitization transformation method, and establish a distributed key retrieval mechanism based on intelligent load balancing. Based on the distributed key retrieval mechanism and key index relationship, the storage location of each target data field in the cloud storage resource pool and the corresponding key acquisition path are determined, and the key derivation level sequence required for decryption key is obtained from the key management node. Access the key management node along the key acquisition path, and use an adaptive key recombination algorithm to dynamically combine and calculate the acquired key materials at each level to generate a decryption key for decrypting the target data field; The decryption key is used to decrypt the target data field at the corresponding storage location in the cloud storage resource pool to obtain the decrypted data field. Based on the association mapping table, the desensitization processing parameters corresponding to each of the decrypted data fields are obtained, and the desensitization processing parameters are applied to the desensitization transformation method to perform desensitization transformation operation on each of the decrypted data fields to generate desensitized financial data.

7. The method according to claim 6, characterized in that, Accessing the key management node along the key acquisition path, and using an adaptive key recombination algorithm to dynamically combine and calculate the acquired key materials at each level, a decryption key for decrypting the target data field is generated, including: Following the key derivation hierarchy sequence defined by the key acquisition path, the key management node is accessed sequentially, the corresponding key material is obtained from each key derivation hierarchy, and the obtained key material is preliminarily structured through an adaptive key reorganization algorithm to form a hierarchical key material set adapted to the current system load characteristics. Deep structural feature extraction is performed on each level of key material in the hierarchical key material set to obtain the entropy distribution characteristics of each level of key material and the real-time load status of the system. Based on the entropy distribution characteristics and the real-time load status of the system, the contribution weight of each level of key material in the subsequent dynamic combination operation is calculated. Based on the contribution weight, the key materials at each level in the hierarchical key material set are intelligently weighted to generate a weighted key material sequence that reflects the importance of each level of key material. Based on the adaptive key reorganization algorithm, the resource optimization configuration of each weighted key material in the weighted key material sequence is performed, and the layer-by-layer fusion operation is performed according to the hierarchical order defined by the key derivation hierarchical sequence. The weighted key material of the current level is adaptively mixed and transformed with the fusion result of the previous level to generate the fusion output of the current level. Repeat the layer-by-layer fusion operation until all levels of weighted key material have been processed, and determine the final level fusion output as the dynamically adaptive decryption key.

8. A cloud-based integrated financial data security storage system, used to implement the method of any one of claims 1-7, characterized in that, include: The data parsing unit is used to acquire the financial data to be stored and to perform semantic parsing on the financial data to extract the business attributes and relationships of the data fields; Based on the business attributes and the relationship, a multi-dimensional sensitivity assessment model is constructed, and the sensitivity of each field in the financial data is quantitatively scored to obtain structured financial data. The layered encryption unit is used to perform differentiated encryption processing on the structured financial data according to the sensitivity score, and to encrypt data fields with different key systems for different sensitivity score ranges to generate layered encrypted data; The access information unit is used to extract the access subject's identity, access purpose, and access spatiotemporal context information from the access request when an access request is received. Based on the access information and the output of the multi-dimensional sensitivity assessment model, a dynamic desensitization rule set is generated for the current access scenario; The selective decryption unit is used to selectively decrypt target data fields from the cloud storage resource pool according to the dynamic desensitization rule set and key index relationship, and process the decrypted data fields according to the desensitization transformation method defined in the rule set to generate desensitized financial data; The model optimization unit is used to return the anonymized financial data to the accessing subject, and at the same time record the access frequency and abnormal access patterns of the data fields during this access process, and use the access records to optimize and update the multi-dimensional sensitivity assessment model.

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