Cross-industry data sharing method and system based on dynamic grading and privacy calculation

Abstract

This invention discloses a cross-industry data sharing method and system based on dynamic hierarchical classification and privacy computing, relating to the fields of blockchain and privacy computing technologies. The method includes: acquiring raw data from multiple industry entities, extracting feature information from their evaluation dimensions, determining sensitivity scores, and classifying sensitivity levels; generating protected data using corresponding privacy protection technologies based on the sensitivity levels; generating digital fingerprints for the protected data and storing them on the blockchain; simultaneously performing joint analysis using the protected data based on a federated learning framework and storing the results on the blockchain; monitoring contextual changes in the protected data, reassessing the sensitivity, and updating the protection strategy and on-chain storage. The system includes a data hierarchical module, a privacy protection module, a trusted analysis module, and a dynamic adjustment module. This invention achieves secure sharing and trusted collaborative analysis of cross-industry data.

Description

Technical Field

[0001] This invention relates to the fields of blockchain and privacy computing technology, specifically to a cross-industry data sharing method and system based on dynamic hierarchical and privacy computing. Background Technology

[0002] Currently, traditional data sharing and privacy protection solutions typically employ a uniform privacy protection strategy to process all data. This leads to overprotection of low-sensitivity data, resulting in a loss of data value, while providing insufficient protection for highly sensitive data, posing risks of privacy leaks and trade secrets.

[0003] While existing privacy-preserving computation frameworks can achieve "data usable but not visible," their computation processes and results lack effective third-party verification mechanisms. Furthermore, blockchain-based evidence storage technology itself cannot address the privacy protection of raw data before it is uploaded to the blockchain, resulting in a failure to organically integrate the two. In addition, many attempts at data grading employ static rules, failing to dynamically adjust protection strategies based on data flow, changes in usage scenarios, or policy and regulatory updates. Moreover, existing technologies often incur significant computational overhead, making efficient deployment on resource-constrained edge devices difficult.

[0004] Therefore, there is a need to provide a cross-industry data sharing method and system that can achieve dynamic data sensitivity classification, deep integration of privacy protection and blockchain evidence storage, and adaptive adjustment capabilities, in order to resolve the contradiction between data security and value mining in industrial collaboration and improve the comprehensive utilization level of data resources. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-industry data sharing method and system based on dynamic hierarchical and privacy-preserving computation to solve the problems raised in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a cross-industry data sharing method and system based on dynamic hierarchical and privacy-preserving computation. The methods include: Step S1: Obtain raw data from multiple industry entities, extract feature information of the raw data in several evaluation dimensions, determine the sensitivity score of the raw data based on the feature information, and classify the raw data into high sensitivity level, medium sensitivity level and low sensitivity level according to the sensitivity score; Step S2: Based on the sensitivity level of the original data, apply corresponding privacy protection technologies to each sensitivity level to generate corresponding protected data; specifically, for high-sensitivity data, homomorphic encryption is used to generate high-sensitivity protected data; for medium-sensitivity data, differential privacy technology is used to generate medium-sensitivity protected data; and for low-sensitivity data, hashing is used to generate low-sensitivity protected data. The protected data includes the high-sensitivity protected data, the medium-sensitivity protected data, and the low-sensitivity protected data. Step S3: Generate a digital fingerprint for the protected data, construct the digital fingerprint into a hash tree structure, and store the root hash value of the hash tree structure on the blockchain for evidence storage; based on the federated learning framework, perform multi-party joint analysis using the protected data forms corresponding to different sensitivity levels to obtain joint analysis results; perform hash processing on the joint analysis results, store the result hash value on the blockchain for evidence storage, and associate it with the protected data evidence storage record to form a trusted chain; Step S4: Monitor the contextual environment change events of the data during its lifecycle. When a trigger event is detected, extract the contextual feature information of the data and re-evaluate the sensitivity of the data. If the sensitivity change obtained from the re-evaluation meets the threshold comparison condition, update the protection strategy of the data and update the on-chain evidence accordingly.

[0007] Furthermore, step S1 includes: Step S1-1: Obtain raw data from several industry entities, extract feature information of the raw data in several evaluation dimensions, and construct a feature vector X=(x1,x2,…,x…). m ); Step S1-2: Standardize the feature vector to obtain standardized feature values. y ij =[x ij -min(x j )] / [max(x j )-min(x j )], where x ij Let x be the original feature value of the i-th sample in the j-th dimension, min(x) j ) represents the minimum value of all samples in the j-th dimension, max(x) j ) represents the maximum value of all samples in the j-th dimension, y ij These are the standardized eigenvalues; Step S1-3: Calculate the objective weights of each dimension ,in , k = 1 / [ln(n)], where n is the total number of data samples participating in the evaluation; obtain the subjective weights w corresponding to each dimension. j The objective weight V j With the subjective weight w j The weighted fusion is performed to obtain the comprehensive weight. , where λ j =α·V j +(1-α)·w j α is the harmonic coefficient; Step S1-4: Based on the comprehensive weight W j and the standardized eigenvalue y ij Calculate the sensitivity score of the data. , where C(r i The value range of ) is [0,1], and the lower the score, the more sensitive the data; Step S1-5: Divide the original data into high sensitivity level, medium sensitivity level, and low sensitivity level according to the sensitivity score; wherein, when 0≤C(r i When C(r) < 0.3, it is classified as high sensitivity; when 0.3 ≤ C(r) i When 0.7 < C(r), it is classified as medium sensitivity; when 0.7 ≤ C(r) < C(r), it is classified as medium sensitivity. i If the value is less than or equal to 1, it is classified as a low sensitivity level.

[0008] Furthermore, step S2 includes: Step S2-1: When the sensitivity score C(r) of the original data i ) satisfies 0≤C(r) i When the value is less than 0.3, the data is classified as high-sensitivity. High-sensitivity data is then encrypted using the SM4 symmetric encryption algorithm to generate SM4 ciphertext C. sm4 =SM4_Encrypt(K sm4 D original (IV), where K sm4 For the encryption key, D original The original data is represented by IV, which is the initialization vector. Then, the Paillier homomorphic encryption algorithm is used to encrypt the SM4 ciphertext or its key values, generating Paillier ciphertext C. paillier =g m ·r n mod n 2 Where g and n are public key parameters, m is the plaintext value, and r is a random number; the Paillier ciphertext is combined with metadata to generate highly sensitive protected data. high =(C paillier Metadata, which includes data identifiers, hierarchical tags, key indexes, and access policies; Step S2-2: When the sensitivity score C(r) of the original data i ) satisfies 0.3≤C(r) i When the sensitivity is less than 0.7, it is classified as medium sensitivity. For data of medium sensitivity, the global sensitivity ΔS of the query function f is determined. The scaling parameter b of the Laplace distribution is calculated based on the privacy budget ε, where b = ΔS / ε. Noise values ​​η∼Laplace(b) are extracted from the Laplace distribution, where the probability density function of the Laplace distribution is f(η) = (1 / 2b) × e^(-ε / 2). [-(∣η∣) / b] The noise value η is added to the query results of the original data to generate sensitive protected data. medium =f(D original )+η; Step S2-3: When the sensitivity score C(r) of the original data i ) satisfies 0.7≤C(r) i When the value is less than or equal to 1, the data is classified as low-sensitivity. The SM3 hash algorithm is used to process the low-sensitivity data, generating a hash value that serves as the protected low-sensitivity data. low =SM3(D original ).

[0009] Furthermore, step S3 includes: Step S3-1: Concatenate the protected data with the current timestamp, and use the SM3 hash algorithm to calculate the hash value as the leaf node hash. Leaf =SM3(Data type ∥Timestamp), where Data type Data high Data medium or Data low ∥ represents a concatenation operation; a dynamic Merkle tree is constructed by pairing several leaf node hashes together and merging them upwards, with the parent node hash being H. parent =SM3(H left ∥H right This continues until the root hash (Root) is calculated. Hash The data provider uses the SM2 private key to sign the root hash, obtaining the signature (r,s)=SM2_Sign(d A ,Root Hash The root hash and the signature are sent to the blockchain network, and the smart contract verifies the validity of the signature and records the root hash on the chain. Step S3-2: Based on the federated learning framework, conduct multi-party joint analysis using protected data formats corresponding to different sensitivity levels, including: for highly sensitive data, using homomorphic encryption and horizontal federated learning in a collaborative manner, each participant uses Paillier homomorphically encrypted ciphertext data C locally. paillier Calculate gradient g i The encryption gradient Enc(g) will be used. i The data is sent to the aggregation server, which uses Paillier's homomorphic addition property, Enc(m1)·Enc(m2)=Enc(m1+m2), to aggregate the gradient as Enc(g)=∏Enc(g). i The aggregated ciphertext gradient is sent back to each party, which then decrypts it using its private key and updates its local model. For moderately sensitive data, a combination of differential privacy and longitudinal federated learning is used to protect the moderately sensitive data. medium The data is sent to federated learning nodes, where each node uses the noisy features for secure sample alignment and model training, achieving joint modeling across samples or features. For low-sensitivity data, the low-sensitivity protected data is... low Used for source domain or auxiliary analysis in federated transfer learning to achieve cross-industry knowledge reuse; Step S3-3: Hash the joint analysis results and store the hash value of the results on the blockchain, associating it with the protected data storage record to form a trusted chain.

[0010] Furthermore, step S4 includes: Step S4-1: Continuously monitor context environment change events during the data lifecycle. These context environment change events include at least one of the following: changes in data relationships, updates to data security policies, and abnormal data access frequency. Format the monitored context environment change events into a standard event object Event=(type, data...) id The event is evaluated using the severity and timestamp of the monitored event, and prioritized according to the severity and type of the event. When the severity of a monitored context change event reaches a trigger threshold, context information related to the data corresponding to the monitored context change event is extracted from the data graph and compliance graph to generate an enhanced feature vector X. new The enhanced feature vector adds dynamic features on the basis of the original features; Step S4-2: Using a federated learning framework, the security events or audit results subsequently triggered by the historical data hierarchical records of the aforementioned industry entities are used as training data to jointly train a federated learning sensitivity prediction model M without sharing the original data. fedThe enhanced feature vector X new Input to the federated learning sensitivity prediction model M fed The sensitivity score C after reassessment was obtained. new (r i )=M fed (X new W global ), where W global C represents the global model parameters after federated aggregation. new (r i This indicates the probability of data privacy breaches in the current context. Step S4-3: Calculate the change in sensitivity ΔC = |C new (r i )-C(r i ) |, using a hysteresis threshold for comparison, if ΔC≤γ upgrade Then execute the upgrade protection process if ΔC ≥ γ downgrade Then the downgrade protection process is executed, where γ upgrade To upgrade the threshold, γ downgrade This is the downgrade threshold; Step S4-4: When executing the upgrade protection process, locate the original data copy based on the data association record, and re-encrypt the original data using a new encryption key using a high-sensitivity protection process to generate new high-sensitivity protected data. high_new ; Calculate the hash value of the new highly sensitive protected data. high_new The blockchain smart contract is invoked to update the Merkle leaf nodes, replacing the old leaf hash with the new hash. high_new The new root hash is calculated and added to the chain. When the downgrade protection process is executed, authorization and approval are triggered. After approval, the original highly sensitive data is decrypted, and differential privacy noise is applied to the decrypted original data to generate new moderately sensitive protected data. medium_new ; Calculate the hash value of the new sensitive protected data. medium_new The blockchain smart contract is invoked to update the Merkle leaf nodes, replacing the old leaf hash with the new hash. medium_new And calculate the new root hash and put it on the chain.

[0011] The system includes: a data classification module, a privacy protection module, a trusted analysis module, and a dynamic adjustment module; The data classification module is used to acquire raw data from multiple industry entities, extract feature information of the raw data in several evaluation dimensions, determine the sensitivity score of the raw data, and divide the raw data into several sensitivity levels according to the sensitivity score. The privacy protection module is used to process data with corresponding privacy protection technologies according to the sensitivity level of the original data, and generate protected data. The trusted analysis module is used to generate digital fingerprints for the protected data and store them on the blockchain. At the same time, based on the federated learning framework, it performs joint analysis using the protected data forms corresponding to different sensitivity levels to obtain joint analysis results, and stores the joint analysis results on the blockchain to form a trusted chain. The dynamic adjustment module is used to monitor the contextual changes of the protected data during its lifecycle. When a triggering event is detected, the sensitivity of the protected data is reassessed, and the protection strategy and on-chain evidence storage of the protected data are updated based on the reassessment results.

[0012] The data classification module includes: The feature extraction unit is used to extract feature information of the raw data in several evaluation dimensions after obtaining the raw data from several industry entities. The weighting calculation unit is used to calculate the sensitivity score of the original data by weighted fusion of objective weights and subjective weights. The grading unit is used to divide the original data into high sensitivity level, medium sensitivity level and low sensitivity level according to the sensitivity score.

[0013] The privacy protection module includes: The highly sensitive protected unit is used to generate highly sensitive protected data by using a dual encryption method that combines symmetric encryption and homomorphic encryption when the data is determined to be of a high sensitivity level. The medium-sensitive protected unit is used to generate medium-sensitive protected data by adding noise using differential privacy when the data is determined to be of medium sensitivity level. The low-sensitivity protected unit is used to generate low-sensitivity protected data by hashing when the data is determined to be of a low sensitivity level.

[0014] The trust analysis module includes: The evidence storage unit is used to concatenate the protected data with the timestamp, calculate the hash value as the leaf node, construct a hash tree structure, and store the root hash value on the blockchain for evidence storage. The joint analysis unit is used to perform joint analysis based on the federated learning framework. For high-sensitivity data, it adopts a combination of homomorphic encryption and horizontal federated learning. For medium-sensitivity data, it adopts a combination of differential privacy and vertical federated learning. For low-sensitivity data, it adopts federated transfer learning for knowledge reuse. The result storage unit is used to perform hash processing on the joint analysis results and store the result hash value on the blockchain.

[0015] The dynamic adjustment module includes: An event monitoring unit is used to monitor contextual environment change events of the protected data during its lifecycle, including at least one of data lineage change, external policy and regulation change, and abnormal data access behavior. The re-evaluation unit is used to extract the context information of the protected data and re-evaluate the sensitivity of the protected data when the monitored context environment change event reaches the trigger condition. The strategy update unit is used to execute an upgrade protection process or a downgrade protection process based on the comparison between the reassessment result and the original sensitivity, and update the on-chain evidence accordingly.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a dynamic data sensitivity grading model, calculates sensitivity scores by weighting and fusing objective and subjective weights, and classifies data into three sensitivity levels: high, medium, and low based on the scores. Different privacy protection technologies such as homomorphic encryption, differential privacy, and hash processing are matched for different levels, thus resolving the contradiction between data value loss and insufficient privacy protection caused by a unified protection strategy, and achieving a balance between strong protection of highly sensitive data and efficient utilization of low-sensitivity data.

[0017] 2. This invention deeply integrates privacy computing with blockchain notarization. By generating digital fingerprints from protected data and constructing a hash tree structure, the root hash value is stored on the blockchain. Simultaneously, it utilizes multiple modes such as homomorphic encryption and horizontal federated learning, differential privacy and vertical federated learning, and federated transfer learning for joint analysis. The hash values ​​of the analysis results are then stored on the blockchain to form a trusted chain. This solves the problems of the lack of third-party verification mechanisms for the privacy computing process and results, and the inability of blockchain notarization to protect the privacy of the original data before it is uploaded to the blockchain. It achieves the unity of "data usable but invisible" and the verification of the computing process and the immutability of the results.

[0018] 3. This invention uses an event-driven dynamic hierarchical adjustment mechanism to monitor in real time contextual changes such as changes in data relationships, updates to data security policies, and abnormal data access frequencies. When the triggering conditions are met, contextual information is extracted to reassess the sensitivity, and an upgrade or downgrade protection process is executed based on the reassessment results, while simultaneously updating the on-chain evidence. This solves the problem that static hierarchical strategies cannot adapt to changes in data flow, usage scenarios, and policy and regulatory updates, enabling data protection strategies to be dynamically adjusted with environmental changes, thus improving the system's adaptability and continuous compliance. Attached Figure Description

[0019] Figure 1This is a schematic diagram illustrating the data classification, privacy protection, and federated analysis process of the cross-industry data sharing method based on dynamic classification and privacy computing of this invention. Figure 2 This is a schematic diagram illustrating the dynamic reassessment and adaptive adjustment process of the protection strategy in the cross-industry data sharing method based on dynamic hierarchical and privacy computing of the present invention. Figure 3 This is a schematic diagram of the cross-industry data sharing system based on dynamic hierarchical and privacy computing of the present invention. Detailed Implementation

[0020] 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.

[0021] Example: Figures 1-3 As shown, this invention provides a technical solution: a cross-industry data sharing method and system based on dynamic hierarchical and privacy-preserving computation. The methods include: Step S1: Obtain raw data from multiple industry entities, extract feature information of the raw data in several evaluation dimensions, determine the sensitivity score of the raw data based on the feature information, and classify the raw data into high sensitivity level, medium sensitivity level and low sensitivity level according to the sensitivity score; Step S1-1: The system connects to the agricultural smart planting system, industrial MES system, service industry transaction system, and logistics management system through the data acquisition layer to obtain raw data from multiple industry entities. Raw data includes: in agriculture, soil moisture, meteorological data, crop growth data, agricultural input data, and farmer identity information; in processing, equipment operating parameters, process formulas, raw material data, quality testing data, energy consumption, and environmental protection data; in the service industry, consumer purchase records, consumer behavior data, and consumer profiles; and in logistics, cargo information, transportation trajectory, logistics node data, and delivery data. For unstructured data, it is first parsed and key information extracted, then converted into structured or semi-structured data. The system performs multi-dimensional feature mapping based on three predefined evaluation dimensions V={V type V risk V compliance}, to conduct in-depth analysis of the data content; Among them, the data type dimension V typeBased on the industry data dictionary, data is mapped to specific categories, such as mapping soil pH to agricultural environmental data, and equipment rotation speed to industrial operating parameters, and a basic sensitivity value v is assigned to each category. t This value reflects the inherent sensitivity of this type of data; Leakage risk dimension V risk A lightweight risk assessment model is used for quantification. This model analyzes the identifiability (whether it contains direct identifiers), inferability (e.g., high spending may infer a high-income group), and correlation (whether the data will generate new sensitive information when correlated with other data), and outputs a comprehensive risk score v. r ; Legal compliance dimension V compliance The system uses a built-in compliance knowledge graph for detection. This graph is linked to laws and regulations such as the Data Security Law and the Personal Information Protection Law. If data contains personal biometric features, it is automatically flagged and assigned a very high compliance risk score. c ; For the i-th sample (i=1,2,…,n), its feature values ​​across the above m evaluation dimensions are combined into a feature vector x. i =(x i1 ,x i2 ,…,x im ), where x ij This represents the feature value of the i-th sample in the j-th dimension; this feature vector comprehensively reflects the inherent, static, and sensitive attributes of the i-th sample. Step S1-2: Standardize the feature vector to obtain standardized feature values. y ij =[x ij -min(x j )] / [max(x j )-min(x j )], where x ij Let x be the original feature value of the i-th sample in the j-th dimension, min(x) j ) represents the minimum value of all samples in the j-th dimension, max(x) j ) represents the maximum value of all samples in the j-th dimension, y ij These are the standardized eigenvalues; Step S1-3: Calculate the objective weights of each dimension ,in , k=1 / [ln(n)], where n is the total number of data samples participating in the evaluation; the system administrator presets the subjective weight vector w based on the current regulatory focus. j During the period of stringent regulation under the Personal Information Protection Law, the weight of legal compliance dimensions was increased; the aforementioned objective weight V was adjusted.j With the subjective weight w j The weighted fusion is performed to obtain the comprehensive weight. , where λ j =α·V j +(1-α)·w j α is the harmonic coefficient; Step S1-4: Based on the comprehensive weight W j and the standardized eigenvalue y ij Calculate the sensitivity score of the data. , where C(r i The value range of ) is [0,1], and the lower the score, the more sensitive the data; Step S1-5: Classify levels according to sensitivity scores: when 0 ≤ C(r) i When C(r) < 0.3, it is classified as high sensitivity; when 0.3 ≤ C(r) i When 0.7 < C(r), it is classified as medium sensitivity; when 0.7 ≤ C(r) < C(r), it is classified as medium sensitivity. i If the value is less than or equal to 1, it is classified as a low sensitivity level.

[0022] Step S2: Based on the sensitivity level of the original data, apply corresponding privacy protection technologies to each sensitivity level to generate corresponding protected data; specifically, for high-sensitivity data, homomorphic encryption is used to generate high-sensitivity protected data; for medium-sensitivity data, differential privacy technology is used to generate medium-sensitivity protected data; and for low-sensitivity data, hashing is used to generate low-sensitivity protected data. The protected data includes the high-sensitivity protected data, the medium-sensitivity protected data, and the low-sensitivity protected data. Step S2-1: When the sensitivity score C(r) of the original data i ) satisfies 0≤C(r) i When the value is less than 0.3, the data is classified as high-sensitivity. High-sensitivity data is then encrypted using the SM4 symmetric encryption algorithm to generate SM4 ciphertext C. sm4 =SM4_Encrypt(K sm4 D original (IV), where K sm4 For the encryption key, D original The original data is represented by IV, which is the initialization vector. Then, the Paillier homomorphic encryption algorithm is used to encrypt the SM4 ciphertext or its key values, generating Paillier ciphertext C. paillier =g m ·r n mod n 2Where g and n are public key parameters, m is the plaintext value, and r is a random number; the Paillier ciphertext is combined with metadata to generate highly sensitive protected data. high =(C paillier Metadata, which includes data identifiers, hierarchical tags, key indexes, and access policies; Step S2-2: When the sensitivity score C(r) of the original data i ) satisfies 0.3≤C(r) i When the sensitivity is less than 0.7, it is classified as medium sensitivity. For medium sensitivity data, such as regional sales statistics and average equipment energy consumption, the global sensitivity ΔS of the query function f is determined. The scaling parameter b of the Laplace distribution is calculated based on the privacy budget ε (initial value set to 0.5). Noise values ​​η∼Laplace(b) are extracted from the Laplace distribution, where the probability density function of the Laplace distribution is f(η)=(1 / 2b)×e [-(∣η∣) / b] The noise value η is added to the query results of the original data to generate sensitive protected data. medium =f(D original )+η; Step S2-3: When the sensitivity score C(r) of the original data i ) satisfies 0.7≤C(r) i When the value is less than or equal to 1, the data is classified as low-sensitivity. For low-sensitivity data, such as anonymized ambient temperature, the SM3 hash algorithm is used to perform hash processing, generating a hash value as the low-sensitivity protected data. low =SM3(D original ).

[0023] Step S3: Generate a digital fingerprint for the protected data, construct the digital fingerprint into a hash tree structure, and store the root hash value of the hash tree structure on the blockchain for evidence storage; based on the federated learning framework, perform multi-party joint analysis using the protected data forms corresponding to different sensitivity levels to obtain joint analysis results; perform hash processing on the joint analysis results, store the result hash value on the blockchain for evidence storage, and associate it with the protected data evidence storage record to form a trusted chain; Step S3-1: Concatenate the protected data with the current timestamp, and use the SM3 hash algorithm to calculate the hash value as the leaf node hash. Leaf =SM3(Data type ∥Timestamp), where Data type Data high Data medium or Data low∥ represents a concatenation operation; a dynamic Merkle tree is constructed by pairing several leaf node hashes together and merging them upwards, with the parent node hash being H. parent =SM3(H left ∥H right This continues until the root hash (Root) is calculated. Hash The data provider uses the SM2 private key to sign the root hash, obtaining the signature (r,s)=SM2_Sign(d A ,Root Hash The root hash and the signature are sent to the blockchain network, and the smart contract verifies the validity of the signature and records the root hash on the chain. Step S3-2: Based on the federated learning framework, conduct multi-party joint analysis using protected data formats corresponding to different sensitivity levels; for high-sensitivity data, a collaborative approach of homomorphic encryption and horizontal federated learning is adopted, with each participant using Paillier homomorphically encrypted ciphertext data C locally. paillier Calculate gradient g i The encryption gradient Enc(g) will be used. i The data is sent to the aggregation server, which uses Paillier's homomorphic addition property, Enc(m1)·Enc(m2)=Enc(m1+m2), to aggregate the gradient as Enc(g)=∏Enc(g). i The aggregated ciphertext gradient is sent back to each party, which then decrypts it using its private key and updates its local model. For moderately sensitive data, a combination of differential privacy and longitudinal federated learning is used to protect the moderately sensitive data. medium The data is sent to federated learning nodes, where each node uses the noisy features for secure sample alignment and model training, achieving joint modeling across samples or features. For low-sensitivity data, the low-sensitivity protected data is... low Used for source domain or auxiliary analysis in federated transfer learning to achieve cross-industry knowledge reuse; Step S3-3: Hash the joint analysis results and store the hash value of the results on the blockchain, associating it with the protected data storage record to form a trusted chain.

[0024] Step S4: Monitor the context environment change events of the data during its lifecycle. When a trigger event is detected, extract the context feature information of the data and re-evaluate the sensitivity of the data. If the sensitivity change obtained from the re-evaluation meets the threshold comparison condition, update the protection strategy of the data and update the on-chain evidence accordingly. Step S4-1: Continuously monitor contextual change events throughout the data lifecycle. These contextual change events include at least one of the following: changes in data relationships, updates to data security policies, and abnormal data access frequencies. The monitored content includes changes in data relationships (e.g., data being used for joint analysis with logistics data), updates to data security policies (e.g., regulatory agencies issuing new data classification guidelines), and abnormal data access frequencies (e.g., low-sensitivity datasets being accessed frequently, potentially indicating a related attack). Format the monitored contextual change events into a standard event object Event=(type, data). id The event is evaluated using the severity and timestamp of the monitored event, and prioritized according to the severity and type of the event. When the severity of a monitored context change event reaches a trigger threshold, contextual information related to the data corresponding to the monitored context change event is extracted from the data graph and compliance graph, generating an enhanced feature vector X. new The enhanced feature vector adds dynamic features on the basis of the original features, such as the average sensitivity of associated data and the credit rating of the joint analysis parties.

[0025] Step S4-2: Using a federated learning framework, the security events or audit results subsequently triggered by the historical data hierarchical records of the aforementioned industry entities are used as training data to jointly train a federated learning sensitivity prediction model M without sharing the original data. fed The enhanced feature vector X new Input to the federated learning sensitivity prediction model M fed The sensitivity score C after reassessment was obtained. new (r i )=M fed (X new W global ), where W global C represents the global model parameters after federated aggregation. new (r i This indicates the probability of data privacy breaches in the current context. Step S4-3: Calculate the change in sensitivity ΔC = |C new (r i )-C(r i ) |, using a hysteresis threshold for comparison, if ΔC≤γ upgrade Then execute the upgrade protection process if ΔC ≥ γ downgrade Then the downgrade protection process is executed, where γ upgrade To upgrade the threshold, γ downgrade This is the downgrade threshold; Step S4-4: When executing the upgrade protection process, locate the original data copy based on the data association record, and re-encrypt the original data using a new encryption key using a high-sensitivity protection process to generate new high-sensitivity protected data. high_new ; Calculate the hash value of the new highly sensitive protected data. high_new The blockchain smart contract is invoked to update the Merkle leaf nodes, replacing the old leaf hash with the new hash. high_new The new root hash is calculated and added to the chain. When the downgrade protection process is executed, authorization and approval are triggered. After approval, the original highly sensitive data is decrypted, and differential privacy noise is applied to the decrypted original data to generate new moderately sensitive protected data. medium_new ; Calculate the hash value of the new sensitive protected data. medium_new The blockchain smart contract is invoked to update the Merkle leaf nodes, replacing the old leaf hash with the new hash. medium_new And calculate the new root hash and put it on the chain.

[0026] The system includes: a data classification module, a privacy protection module, a trusted analysis module, and a dynamic adjustment module; The data classification module is used to acquire raw data from multiple industry entities, extract feature information of the raw data in several evaluation dimensions, determine the sensitivity score of the raw data, and divide the raw data into several sensitivity levels according to the sensitivity score. The privacy protection module is used to process data with corresponding privacy protection technologies according to the sensitivity level of the original data, and generate protected data. The trusted analysis module is used to generate digital fingerprints for the protected data and store them on the blockchain. At the same time, based on the federated learning framework, it performs joint analysis using the protected data forms corresponding to different sensitivity levels to obtain joint analysis results, and stores the joint analysis results on the blockchain to form a trusted chain. The dynamic adjustment module is used to monitor the contextual changes of the protected data during its lifecycle. When a triggering event is detected, the sensitivity of the protected data is reassessed, and the protection strategy and on-chain evidence storage of the protected data are updated based on the reassessment results.

[0027] The data classification module includes: The feature extraction unit is used to extract feature information of the raw data in several evaluation dimensions after obtaining the raw data from several industry entities. The weighting calculation unit is used to calculate the sensitivity score of the original data by weighted fusion of objective weights and subjective weights. The grading unit is used to divide the original data into high sensitivity level, medium sensitivity level and low sensitivity level according to the sensitivity score.

[0028] The privacy protection module includes: The highly sensitive protected unit is used to generate highly sensitive protected data by using a dual encryption method that combines symmetric encryption and homomorphic encryption when the data is determined to be of a high sensitivity level. The medium-sensitive protected unit is used to generate medium-sensitive protected data by adding noise using differential privacy when the data is determined to be of medium sensitivity level. The low-sensitivity protected unit is used to generate low-sensitivity protected data by hashing when the data is determined to be of a low sensitivity level.

[0029] The trust analysis module includes: The evidence storage unit is used to concatenate the protected data with the timestamp, calculate the hash value as the leaf node, construct a hash tree structure, and store the root hash value on the blockchain for evidence storage. The joint analysis unit is used to perform joint analysis based on the federated learning framework. For high-sensitivity data, it adopts a combination of homomorphic encryption and horizontal federated learning. For medium-sensitivity data, it adopts a combination of differential privacy and vertical federated learning. For low-sensitivity data, it adopts federated transfer learning for knowledge reuse. The result storage unit is used to perform hash processing on the joint analysis results and store the result hash value on the blockchain.

[0030] The dynamic adjustment module includes: An event monitoring unit is used to monitor contextual environment change events of the protected data during its lifecycle, including at least one of data lineage change, external policy and regulation change, and abnormal data access behavior. The re-evaluation unit is used to extract the context information of the protected data and re-evaluate the sensitivity of the protected data when the monitored context environment change event reaches the trigger condition. The strategy update unit is used to execute an upgrade protection process or a downgrade protection process based on the comparison between the reassessment result and the original sensitivity, and update the on-chain evidence accordingly.

[0031] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A cross-industry data sharing method based on dynamic grading and privacy calculation, characterized in that: Includes the following steps: The process involves acquiring raw data from several industry entities, extracting feature information from the raw data across several evaluation dimensions, determining the sensitivity score of the raw data based on the feature information, and dividing the raw data into several sensitivity levels according to the sensitivity score. Based on the sensitivity level of the original data, corresponding privacy protection technologies are applied to data with different sensitivity levels to generate protected data. A digital fingerprint is generated for the protected data and stored on the blockchain for evidence. At the same time, based on the federated learning framework, joint analysis is performed using the protected data forms corresponding to different sensitivity levels to obtain joint analysis results. The joint analysis results are then stored on the blockchain for evidence, forming a trusted chain. The system monitors contextual change events of the protected data throughout its lifecycle. When a monitored contextual change event meets the triggering condition, the sensitivity of the protected data is reassessed, and the protection strategy and on-chain evidence storage of the protected data are updated based on the reassessment results.

2. The cross-industry data sharing method based on dynamic hierarchical and privacy-preserving computation according to claim 1, characterized in that: The classification into several sensitivity levels includes: After acquiring raw data from several industry entities, feature information of the raw data on several evaluation dimensions is extracted. Objective weights of each evaluation dimension are calculated based on the information entropy of the feature information on each evaluation dimension, and subjective weights of each evaluation dimension are set according to business needs. The objective weights and subjective weights are weighted and fused to obtain a comprehensive weight. The sensitivity score of the raw data is calculated based on the comprehensive weight, and the raw data is divided into high sensitivity level, medium sensitivity level, and low sensitivity level according to the sensitivity score. Specifically, when the sensitivity score is less than a first threshold, it is classified as high sensitivity level; when the sensitivity score is between the first threshold and a second threshold, it is classified as medium sensitivity level; and when the sensitivity score is greater than or equal to the second threshold, it is classified as low sensitivity level.

3. The cross-industry data sharing method based on dynamic hierarchical and privacy-preserving computation according to claim 1, characterized in that: Different privacy protection technologies are used to process data with different sensitivity levels, including: When a data is determined to be highly sensitive, a dual encryption method combining symmetric encryption and homomorphic encryption is used to generate highly sensitive protected data; when a data is determined to be moderately sensitive, a differential privacy method with added noise is used to generate moderately sensitive protected data; when a data is determined to be low sensitive, hash processing is used to generate low-sensitivity protected data; the highly sensitive protected data, the moderately sensitive protected data, and the low-sensitivity protected data constitute protected data.

4. The cross-industry data sharing method based on dynamic hierarchical and privacy-preserving computation according to claim 1, characterized in that: The formation of a trusted chain includes: The protected data is concatenated with a timestamp, and the hash value is calculated as the leaf node. A hash tree structure is constructed, and the root hash value is stored on the blockchain for evidence. Based on the federated learning framework, homomorphic encryption and horizontal federated learning are used for joint analysis of highly sensitive data. Differential privacy and vertical federated learning are used for joint analysis of medium-sensitive data. Federated transfer learning is used for knowledge reuse of low-sensitive data. The joint analysis results are hashed, and the hash value of the result is stored on the blockchain for evidence.

5. The cross-industry data sharing method based on dynamic hierarchical and privacy-preserving computation according to claim 1, characterized in that: The protection strategy and on-chain evidence storage for updating the protected data based on the reassessment results include: The system monitors contextual change events of the protected data throughout its lifecycle. These contextual change events include at least one of the following: changes in data relationships, updates to data security policies, and abnormal data access frequencies. When a monitored contextual change event meets the reassessment conditions, the system extracts the contextual information of the protected data and reassesses the sensitivity of the protected data. Based on the comparison between the reassessment result and the original sensitivity, an upgrade protection process is executed when the comparison result meets preset upgrade conditions, and a downgrade protection process is executed when the comparison result meets preset downgrade conditions. Specifically, during the upgrade protection process, the system locates a copy of the original data based on the data relationship record, re-encrypts the original data using a high-sensitivity protection process, generates new high-sensitivity protected data, and updates the on-chain notarization. During the downgrade protection process, authorization and approval are triggered. After approval, the original high-sensitivity data is decrypted and differential privacy noise is applied to generate new medium-sensitivity protected data, and the on-chain notarization is updated.

6. A cross-industry data sharing system based on dynamic hierarchical and privacy-preserving computation, used to execute the cross-industry data sharing method based on dynamic hierarchical and privacy-preserving computation as described in any one of claims 1-5, characterized in that: The system includes: Data classification module, privacy protection module, trusted analysis module, and dynamic adjustment module; The data classification module is used to acquire raw data from multiple industry entities, extract feature information of the raw data in several evaluation dimensions, determine the sensitivity score of the raw data, and divide the raw data into several sensitivity levels according to the sensitivity score. The privacy protection module is used to process data with corresponding privacy protection technologies according to the sensitivity level of the original data, and generate protected data. The trusted analysis module is used to generate digital fingerprints for the protected data and store them on the blockchain. At the same time, based on the federated learning framework, it performs joint analysis using the protected data forms corresponding to different sensitivity levels to obtain joint analysis results, and stores the joint analysis results on the blockchain to form a trusted chain. The dynamic adjustment module is used to monitor the contextual changes of the protected data during its lifecycle. When a triggering event is detected, the sensitivity of the protected data is reassessed, and the protection strategy and on-chain evidence storage of the protected data are updated based on the reassessment results.

7. The cross-industry data sharing system based on dynamic hierarchical and privacy-preserving computation according to claim 6, characterized in that: The data classification module includes: The feature extraction unit is used to extract feature information of the raw data in several evaluation dimensions after obtaining the raw data from several industry entities. The weighting calculation unit is used to calculate the sensitivity score of the original data by weighted fusion of objective weights and subjective weights. The grading unit is used to divide the original data into high sensitivity level, medium sensitivity level and low sensitivity level according to the sensitivity score.

8. The cross-industry data sharing system based on dynamic hierarchical and privacy-preserving computation according to claim 6, characterized in that: The privacy protection module includes: The highly sensitive protected unit is used to generate highly sensitive protected data by using a dual encryption method that combines symmetric encryption and homomorphic encryption when the data is determined to be of a high sensitivity level. The medium-sensitive protected unit is used to generate medium-sensitive protected data by adding noise using differential privacy when the data is determined to be of medium sensitivity level. The low-sensitivity protected unit is used to generate low-sensitivity protected data by hashing when the data is determined to be of a low sensitivity level.

9. The cross-industry data sharing system based on dynamic hierarchical and privacy-preserving computation according to claim 6, characterized in that: The reliability analysis module includes: The evidence storage unit is used to concatenate the protected data with the timestamp, calculate the hash value as the leaf node, construct a hash tree structure, and store the root hash value on the blockchain for evidence storage. The joint analysis unit is used to perform joint analysis based on the federated learning framework. For high-sensitivity data, it adopts a combination of homomorphic encryption and horizontal federated learning. For medium-sensitivity data, it adopts a combination of differential privacy and vertical federated learning. For low-sensitivity data, it adopts federated transfer learning for knowledge reuse. The result storage unit is used to perform hash processing on the joint analysis results and store the result hash value on the blockchain.

10. The cross-industry data sharing system based on dynamic hierarchical and privacy-preserving computation according to claim 6, characterized in that: The dynamic adjustment module includes: An event monitoring unit is used to monitor contextual environment change events of the protected data during its lifecycle, including at least one of data lineage change, external policy and regulation change, and abnormal data access behavior. The re-evaluation unit is used to extract the context information of the protected data and re-evaluate the sensitivity of the protected data when the monitored context environment change event reaches the trigger condition. The strategy update unit is used to execute an upgrade protection process or a downgrade protection process based on the comparison between the reassessment result and the original sensitivity, and update the on-chain evidence accordingly.

Images