Supply chain data security management method and system

Abstract

The invention provides a supply chain data security management method and system, and the method comprises the steps: carrying out the hierarchical division and distributed storage processing of supply chain data according to production links, and generating a process correlation hash value; constructing a feature extractor containing a process constraint relationship based on the supply chain data and the process association hash value, and extracting by adopting domain adversarial features to obtain supply chain data features meeting physical constraints; performing multi-stage differential disturbance according to the supply chain data characteristics and the process parameter association map, and correcting the disturbed data based on the process constraint relationship; and according to the disturbed process parameter data and the process association hash value, carrying out dynamic on-chain confrontation verification on the model access behavior, and when abnormal access is detected, carrying out differential disturbance intensity adjustment and access limitation. According to the method, process constraints and multi-stage differential disturbance are introduced in the transfer learning process, competitors are prevented from obtaining commercial confidential information through model reverse reasoning, and the safety and effectiveness of data sharing are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of data security management, and in particular to a supply chain data security management method and system. Background Art

[0002] In the current supply chain collaborative manufacturing environment, upstream and downstream companies usually use intelligent prediction systems based on transfer learning for production planning and quality management. The typical system architecture includes data acquisition layer, feature extraction layer, model training layer and prediction output layer. Among them, the data acquisition layer is responsible for collecting various production data, such as equipment operating parameters, product quality data, material consumption data, etc.; the feature extraction layer performs dimensionality reduction and feature engineering on the original data; the model training layer trains and optimizes the model based on the transfer learning algorithm; the prediction output layer is responsible for generating various prediction results and decision recommendations.

[0003] However, in the process of model migration, since the production data of different companies often contain commercial secrets such as core process parameters and formula data, although the existing management methods can ensure the security of data transmission and storage, they cannot effectively prevent competitors from extracting sensitive information from the original training data from the optimized model through reverse reasoning of the model. This hidden data leakage may cause damage to the core competitiveness of the enterprise. Summary of the invention

[0004] The main purpose of the present invention is to solve the technical problem that the intelligent prediction system based on transfer learning is easy to be used by competitors to obtain core process parameters through reverse reasoning during the model migration process in the existing supply chain collaborative manufacturing environment; A first aspect of the present invention provides a supply chain data security management method, the supply chain data security management method comprising: The supply chain data is divided into layers and distributedly stored according to the production links of the supply chain, and the process-related hash value is generated based on the process constraint information and the distributedly stored supply chain data; Building a feature extractor containing a process constraint relationship based on the supply chain data and process-related hash values, and performing domain adversarial feature extraction on the supply chain data based on the feature extractor to obtain supply chain data features that meet physical constraints; Performing multi-level differential disturbance on the supply chain data according to the supply chain data characteristics and the preset process parameter association map, and correcting the disturbed data according to the process constraint relationship to obtain disturbed process parameter data; The model access behavior is dynamically verified on the chain based on the disturbed process parameter data and the process-associated hash value. When abnormal model access is detected, differentiated disturbance intensity adjustments and access restrictions are performed on the distributed stored supply chain data.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the step of performing hierarchical division and distributed storage processing on the supply chain data according to the supply chain production link, and generating a process-related hash value based on the process constraint information and the distributedly stored supply chain data includes: Classify the supply chain data into production links to obtain raw material link data, manufacturing link data and product link data; Use IPFS distributed storage to store the classified supply chain data in blocks to obtain data blocks and data block indexes; Constructing a Merkle tree for the data block based on the process constraint information to obtain a Merkle tree with process constraints; A process-associated hash value is generated according to the Merkle tree with process constraints and process constraint information.

[0006] Optionally, in a second implementation of the first aspect of the present invention, constructing a feature extractor containing a process constraint relationship based on the supply chain data and the process-associated hash value, performing domain adversarial feature extraction on the supply chain data based on the feature extractor, and obtaining supply chain data features that satisfy physical constraints includes: Perform data domain labeling on process parameter data in the supply chain data to obtain data samples with source identifiers, group process process-related hash values ​​according to the source identifiers, and obtain feature constraint conditions for each data domain; Performing source domain and target domain division processing on the data sample to obtain an initial domain division result, performing consistency verification processing on the initial domain division result according to the feature constraint condition to obtain a domain division strategy that meets the process constraint; Performing feature space mapping processing on the process parameter data according to the domain partitioning strategy to obtain an initial feature representation, and performing gradient reversal layer embedding processing on the initial feature representation to obtain an adversarial feature extractor; The adversarial feature extractor is subjected to process constraint injection processing to obtain a feature extraction network with process constraints, and the supply chain data is subjected to feature extraction processing based on the feature extraction network to obtain supply chain data features that meet physical constraints.

[0007] Optionally, in a third implementation of the first aspect of the present invention, performing feature space mapping processing on the process parameter data according to the domain partitioning strategy to obtain an initial feature representation, and performing gradient reversal layer embedding processing on the initial feature representation to obtain an adversarial feature extractor includes: Performing feature dimension analysis processing on the domain partitioning strategy to obtain a dimension weight vector, and performing weighted mapping processing on the process parameter data according to the dimension weight vector to obtain an initial feature space; Performing nonlinear transformation processing on the initial feature space to obtain a hidden layer feature representation, and performing principal component projection processing on the hidden layer feature representation to obtain a reduced dimension feature matrix; Performing reverse gradient calculation processing on the dimension reduction feature matrix to obtain gradient reversal parameters, and performing back propagation processing according to the gradient reversal parameters to obtain an adversarial gradient matrix; The adversarial gradient matrix is ​​subjected to feature reconstruction processing to obtain an adversarial feature representation, and a feature extraction network is constructed based on the adversarial feature representation to obtain an adversarial feature extractor.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, performing multi-level differential perturbation on the supply chain data according to the supply chain data characteristics and a preset process parameter association map, and correcting the disturbed data according to the process constraint relationship, to obtain the disturbed process parameter data includes: Quantifying the process sensitivity of the supply chain data features to obtain a parameter sensitivity matrix, and performing correlation analysis based on the parameter sensitivity matrix and the process parameter correlation map to obtain a multi-level sensitive data set; Perform differential privacy budget allocation processing on the multi-level sensitive data set to obtain disturbance budget values ​​at each level, and perform parameter disturbance amount calculation processing on the disturbance budget values ​​to obtain initial disturbance data; Performing constraint verification processing on the initial disturbance data and the process constraint relationship to obtain a constraint violation data identifier, and performing constraint verification processing on the initial disturbance data according to the constraint violation data identifier to obtain a constraint violation degree; The initial disturbance data is subjected to process constraint correction processing according to the constraint violation degree to obtain corrected disturbance data, and the corrected disturbance data is subjected to process parameter reconstruction processing to obtain disturbed process parameter data.

[0009] Optionally, in a fifth implementation manner of the first aspect of the present invention, performing differential privacy budget allocation processing on the multi-level sensitive data set to obtain disturbance budget values ​​at each level, performing parameter disturbance amount calculation processing on the disturbance budget value to obtain initial disturbance data includes: Performing privacy sensitivity calculation processing on the multi-level sensitive data set to obtain a sensitivity scoring matrix, and performing budget weight allocation processing according to the sensitivity scoring matrix to obtain an initial budget allocation plan; Performing constraint optimization processing on the initial budget allocation scheme to obtain an optimized budget value, and performing Laplace noise generation processing according to the optimized budget value to obtain a noise parameter set; Performing a step-by-step disturbance amount calculation process on the noise parameter set to obtain a disturbance amount matrix, and performing a parameter disturbance superposition process according to the disturbance amount matrix to obtain a disturbance superposition result; The disturbance superposition result is numerically normalized to obtain normalized disturbance data, and parameter reconstruction is performed according to the normalized disturbance data to obtain initial disturbance data.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the performing dynamic on-chain adversarial verification on the model access behavior according to the disturbed process parameter data and the process-associated hash value, and performing differentiated disturbance intensity adjustment and access restriction on the distributed stored supply chain data when abnormal model access is detected, includes: Perform smart contract deployment processing on the disturbed process parameter data to obtain an adversarial verification contract, generate verification rules based on the adversarial verification contract and the process-related hash value, and obtain dynamic on-chain verification rules; Perform feature extraction on the model access behavior to obtain an access feature vector, perform adversarial verification based on the access feature vector and dynamic on-chain verification rules, and obtain a verification score and anomaly identification; Performing credibility quantification processing on the model access behavior according to the verification score and the abnormal identification to obtain access credibility, and performing dynamic threshold determination processing on the access credibility to obtain an access control decision; According to the access control decision, a differentiated disturbance intensity calculation process is performed on the distributedly stored supply chain data to obtain a disturbance adjustment plan, and the disturbance adjustment plan is deployed on-chain to obtain an updated access restriction strategy.

[0011] A second aspect of the present invention provides a supply chain data security management system, the supply chain data security management system comprising: A hierarchical storage module is used to perform hierarchical division and distributed storage processing on the supply chain data according to the production links of the supply chain, and to generate a process-related hash value based on the process constraint information and the distributedly stored supply chain data; A feature adversarial module, used to construct a feature extractor containing a process constraint relationship based on the supply chain data and the process-related hash value, and to perform domain adversarial feature extraction on the supply chain data based on the feature extractor to obtain supply chain data features that meet physical constraints; A differential perturbation module, used to perform multi-level differential perturbation on the supply chain data according to the supply chain data characteristics and a preset process parameter association map, and to correct the disturbed data according to the process constraint relationship to obtain disturbed process parameter data; The access verification module is used to perform dynamic on-chain adversarial verification on the model access behavior based on the disturbed process parameter data and the process-associated hash value, and to perform differentiated disturbance intensity adjustment and access restriction on the distributedly stored supply chain data when abnormal model access is detected.

[0012] The above-mentioned supply chain data security management method and system generates process-related hash values ​​by dividing the supply chain data into layers and distributing the storage processing according to the production links; constructs a feature extractor containing process constraint relationships based on the supply chain data and process-related hash values, and uses domain adversarial features to extract supply chain data features that meet physical constraints; performs multi-level differential perturbations based on the supply chain data features and process parameter association maps, and corrects the perturbed data based on the process constraint relationships; performs dynamic on-chain adversarial verification of model access behavior based on the perturbed process parameter data and process-related hash values, and performs differentiated perturbation intensity adjustment and access restrictions when abnormal access is detected. The present invention prevents competitors from obtaining commercial confidential information through model reverse reasoning and ensures the security and effectiveness of data sharing by introducing process constraints and multi-level differential perturbations in the process of transfer learning.

[0013] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0014] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of a first embodiment of a supply chain data security management method in an embodiment of the present invention; Figure 2 It is a schematic diagram of an embodiment of a supply chain data security management system in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0018] To facilitate understanding of this embodiment, a supply chain data security management method disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps: 101. Perform hierarchical division and distributed storage processing on the supply chain data according to the production links of the supply chain, and generate process-related hash values ​​based on process constraint information and distributedly stored supply chain data; In one embodiment of the present invention, the supply chain data is hierarchically divided and distributedly stored according to the production links of the supply chain, and a process-associated hash value is generated based on the process constraint information and the distributedly stored supply chain data, including: classifying the supply chain data according to the production links to obtain raw material link data, manufacturing link data and product link data; using IPFS distributed storage to store the classified supply chain data in blocks to obtain data blocks and data block indexes; constructing a Merkle tree for the data blocks based on the process constraint information to obtain a Merkle tree with process constraints; and generating a process-associated hash value based on the Merkle tree with process constraints and the process constraint information.

[0019] Specifically, when classifying the data in the production process, first establish a data attribute mapping table, which contains multiple dimensional information such as data type identifier, data source identifier, data timestamp, data business attributes, etc. For raw material link data, the system extracts supplier qualification information (such as quality system certification, environmental certification, etc.), basic information of raw materials (such as physical properties, chemical composition, specifications, etc.), raw material inspection data (such as incoming material inspection report, physical performance test data, chemical composition analysis data, etc.), inventory information (such as batch number, storage location, storage conditions, etc.) and upstream supply chain traceability information (such as raw material origin, transportation information, transit warehousing information, etc.). For manufacturing link data, the focus is on production equipment operating parameters (such as equipment temperature, pressure, speed, power, etc.), process control parameters (such as processing temperature, insulation time, cooling rate, additive ratio, etc.), online detection data (such as real-time quality detection data, process parameter monitoring data, equipment status monitoring data, etc.), production environment parameters (such as workshop temperature and humidity, cleanliness, air quality, etc.) and production process traceability information (such as operators, production batches, material input records, etc.). For product link data, the system collects finished product inspection data (such as appearance inspection, performance testing, reliability verification, etc.), packaging information (such as packaging specifications, protection requirements, identification information, etc.), warehousing and logistics data (such as storage conditions, transportation requirements, distribution information, etc.), customer usage feedback (such as product usage data, performance evaluation, fault information, etc.) and after-sales service records (such as maintenance records, upgrade and transformation information, life prediction, etc.).

[0020] Specifically, when using the IPFS distributed storage system for data storage, the data in each link is first preprocessed, including data format unification, data encoding conversion, data compression and other operations. Then, the optimal data slicing strategy is determined based on the time correlation and spatial locality of the data. The system adopts a dynamic slicing algorithm and uses different slice sizes for different types of data: for structured data (such as production parameters, test data, etc.), a fixed-size slice (256KB) is used; for unstructured data (such as images, videos, etc.), the slice size (512KB-1MB) is dynamically adjusted according to the complexity of the data content. The SHA-256 algorithm is used to calculate the content-addressed hash value of each data slice, and the association between data slices is recorded. The system also establishes a multi-level index structure: the first-level index records the basic information of the data block (such as hash value, size, creation time, etc.); the second-level index records the storage location and access path of the data block; the third-level index records the association and organizational structure between data blocks. At the same time, for frequently accessed data blocks, the system will retain the cache in the local node and establish a cache update mechanism to ensure data consistency and availability.

[0021] Specifically, when constructing a Merkle tree with process constraints, it is first necessary to formally describe and mathematically model the process constraint information. The process constraint information includes quantitative relationships between parameters (such as thermodynamic equations, material mechanics equations, etc.), qualitative relationships (such as process parameter change trends, quality characteristics correlation, etc.) and normative requirements (such as safety limits, process procedures, etc.). The system converts these constraint relationships into constraint verification functions, the input of the function is the relevant process parameters, and the output is the constraint satisfaction score. In the construction of the Merkle tree, the data blocks are first organized into a tree structure according to the process flow sequence. The leaf nodes store the hash values ​​of the data blocks. In addition to storing the hash values ​​of the child nodes, the non-leaf nodes also embed the corresponding process constraint verification functions. The system constructs the tree structure in a bottom-up manner: first, the hash values ​​of the leaf nodes are calculated, then the connection relationships and weights between the nodes are determined according to the process constraint relationships, and finally the hash values ​​and constraint verification functions of the non-leaf nodes are calculated layer by layer.

[0022] Specifically, the process of generating process-related hash values ​​combines the two dimensions of data integrity verification and process constraint verification. First, the system performs a process constraint verification function on each node in the Merkle tree to obtain a series of constraint satisfaction scores. These scores reflect whether the data complies with process rules and physical laws. Then, the system performs bit operations (including XOR operations and circular shift operations) on the constraint satisfaction scores and the original hash values ​​of the nodes to obtain intermediate hash values ​​that incorporate process constraint information. Next, the system uses the improved SHA-256 algorithm to perform multiple rounds of calculations on these intermediate hash values. The algorithm adds the encoding process of process constraint information on the basis of the standard SHA-256 to ensure that the hash value generated in the end can reflect the process relevance of the data. The process-related hash value generated in the end is represented in hexadecimal, where the upper 128 bits are used for data integrity verification and the lower 128 bits are used for process constraint verification. The hash value is recorded in the blockchain network together with the data block index as the basis for data access and verification.

[0023] 102. Construct a feature extractor containing process constraint relationships based on supply chain data and process-related hash values, perform domain adversarial feature extraction on the supply chain data based on the feature extractor, and obtain supply chain data features that meet physical constraints; In one embodiment of the present invention, the feature extractor containing the process constraint relationship is constructed based on the supply chain data and the process-associated hash value, and the domain adversarial feature extraction is performed on the supply chain data based on the feature extractor to obtain the supply chain data features that meet the physical constraints, including: performing data domain labeling processing on the process parameter data in the supply chain data to obtain data samples with source identification, grouping the process-associated hash values ​​according to the source identification to obtain feature constraints of each data domain; performing source domain and target domain partitioning processing on the data samples to obtain initial domain partitioning results, performing consistency verification processing on the initial domain partitioning results according to the feature constraints to obtain a domain partitioning strategy that meets the process constraints; performing feature space mapping processing on the process parameter data according to the domain partitioning strategy to obtain an initial feature representation, performing gradient reversal layer embedding processing on the initial feature representation to obtain an adversarial feature extractor; performing process constraint injection processing on the adversarial feature extractor to obtain a feature extraction network with process constraints, performing feature extraction processing on the supply chain data according to the feature extraction network to obtain supply chain data features that meet the physical constraints.

[0024] Specifically, when processing the data domain labeling of the supply chain data, the data source feature vector is first constructed. The vector contains multi-dimensional identifiers such as the source enterprise information, production line information, equipment identification, and collection time of the data. At the same time, semantic analysis is performed on the process parameter data to extract the process characteristics. For example, for the injection molding process, key parameters such as mold temperature, injection pressure, and holding time are extracted; for the heat treatment process, process parameters such as heating temperature, holding time, and cooling rate are extracted. Based on these characteristics, the system generates a unique source identification code for each data sample. The identification code adopts a hierarchical coding method: the first layer represents the enterprise code, the second layer represents the production line code, the third layer represents the process type code, and the fourth layer represents the parameter type code. Then, the process-related hash values ​​are grouped according to the source identification, and a constraint relationship matrix between parameters is established in each group. The matrix describes the physical constraints, process limitations, and quality requirements between process parameters, thereby obtaining the characteristic constraints of each data domain.

[0025] Specifically, the source domain and the target domain are divided. The system first constructs a feature space based on the statistical characteristics of the data samples (such as mean, variance, distribution form, etc.) and process characteristics (such as parameter range, change law, correlation, etc.). In this feature space, the clustering center of the data sample is determined by calculating the Mahalanobis distance and process similarity between samples. The initial division of the source domain and the target domain is based on these clustering centers, while considering the time series characteristics of the data to ensure that the data distribution of adjacent time periods has a certain continuity. For the initial domain division results, the system verifies the consistency according to the feature constraints. The verification process includes parameter range verification, physical constraint verification, and process rule verification. If the division result that does not meet the constraints is found, it is optimized by adjusting the clustering parameters and boundary conditions, and finally a domain division strategy that meets the process constraints is obtained.

[0026] Specifically, according to the domain partitioning strategy, the system constructs a feature space mapping network. The network adopts a multi-layer perceptron structure. The number of nodes in the input layer is the same as the dimension of the original process parameters. The hidden layer uses a nonlinear activation function for feature transformation, and the output layer dimension is determined according to the principal component analysis results of the process characteristics. Process constraint verification units are embedded in each layer of the network. These units monitor the satisfaction of physical constraints during the feature transformation process. When the gradient reversal layer is embedded in the initial feature representation, the feature distribution difference between the source domain and the target domain is first calculated, and then the gradient reversal function is designed. This function maximizes the difference between domains while keeping the process characteristics unchanged. The network parameters are optimized through the back-propagation algorithm, so that the feature extractor can extract shared features while maintaining domain adversarial properties.

[0027] Specifically, finally, the process constraints are embedded into the adversarial feature extractor to form a complete feature extraction network. The constraint injection process first converts the physical laws and process rules into loss functions. For example, for heat treatment processes, the laws of thermodynamics are converted into constraint functions of temperature-time curves; for machining processes, the principles of material mechanics are converted into constraint functions of cutting parameters. These constraint functions are combined with the original loss function of the network to form a multi-objective optimization problem. Through the alternating optimization strategy, the satisfaction of process constraints is guaranteed while domain adversarial learning. When the network training converges, the feature extraction network is used to process the supply chain data. The generated features not only maintain the essential characteristics of the data, but also satisfy the physical constraints and have strong domain adaptability.

[0028] Furthermore, the process parameter data is subjected to feature space mapping processing according to the domain partitioning strategy to obtain an initial feature representation, and the initial feature representation is subjected to gradient reversal layer embedding processing to obtain an adversarial feature extractor, including: feature dimension analysis processing is performed on the domain partitioning strategy to obtain a dimension weight vector, and the process parameter data is subjected to weighted mapping processing according to the dimension weight vector to obtain an initial feature space; nonlinear transformation processing is performed on the initial feature space to obtain a hidden layer feature representation, and principal component projection processing is performed according to the hidden layer feature representation to obtain a reduced dimension feature matrix; reverse gradient calculation processing is performed on the reduced dimension feature matrix to obtain gradient reversal parameters, and back propagation processing is performed according to the gradient reversal parameters to obtain an adversarial gradient matrix; feature reconstruction processing is performed on the adversarial gradient matrix to obtain an adversarial feature representation, and feature extraction network construction processing is performed according to the adversarial feature representation to obtain an adversarial feature extractor.

[0029] Specifically, when performing feature dimension analysis on the domain partitioning strategy, the contribution of each dimension of the process parameters to the overall feature space is first calculated. The physical meaning and process importance of the parameters are considered during the analysis. For example, in the injection molding process, parameters such as mold temperature and injection pressure have higher weights, while auxiliary parameters such as mold opening and closing time have lower weights. By calculating the information entropy and mutual information of each dimensional feature, the independence and information content of the features are evaluated, and the weights are corrected in combination with the knowledge of process experts. The system adopts an adaptive weight allocation algorithm, which considers the physical coupling relationship between parameters, such as the relationship between temperature and pressure, speed and force, and finally generates a dimensional weight vector. Based on the weight vector, the process parameter data is weighted mapped to convert the original parameter space into the initial feature space, which retains the physical meaning and correlation of the process parameters.

[0030] Specifically, when the initial feature space is transformed nonlinearly, a deep neural network structure is used, which includes multiple hidden layers. Each hidden layer uses a different nonlinear activation function. For example, the first layer uses the ReLU function to capture the positive activation features of the parameters, the second layer uses the tanh function to process the bidirectional change features of the parameters, and the third layer uses the sigmoid function to achieve feature normalization. During the transformation process, multiple residual connections are set to ensure that the original process feature information will not be lost in the deep network. At the same time, batch normalization is performed after each layer of transformation to reduce the offset of the feature distribution. After nonlinear transformation, the hidden layer feature representation is obtained, which contains the high-order correlation of the process parameters. Then, the hidden layer features are subjected to principal component analysis, the covariance matrix and eigenvalues ​​of the features are calculated, and the principal components with a cumulative contribution rate of more than 95% are selected to construct a projection matrix, and the high-dimensional features are projected into a low-dimensional space to obtain a reduced-dimensional feature matrix.

[0031] Specifically, when performing reverse gradient calculation, a discriminator network is first constructed, which is used to distinguish whether the feature comes from the source domain or the target domain. The discriminator adopts a multi-layer feedforward network structure and is trained by minimizing the binary cross entropy loss. During the training process, the gradient of the discriminator to the feature is calculated. These gradients reflect the contribution of the feature to the domain classification. Through the gradient reversal layer, the direction of these gradients is reversed to generate gradient reversal parameters. These parameters are used for back propagation to update the parameters of the feature extraction network so that the extracted features can confuse the discriminator. This process is implemented through an alternating optimization algorithm. In each iteration, the feature extractor is fixed to update the discriminator, and then the discriminator is fixed to update the feature extractor, and finally the adversarial gradient matrix is ​​obtained.

[0032] Specifically, finally, feature reconstruction is performed based on the adversarial gradient matrix. The reconstruction process adopts an encoder-decoder structure. The encoder encodes the adversarial gradient information into the feature space, and the decoder restores the encoded features to the original parameter space. In the encoding process, the attention mechanism is used to highlight important process features while maintaining the adversarial nature of the features. The reconstructed adversarial feature representation contains both the original process features and has strong domain migration capabilities. Based on these feature representations, a multi-layer feature extraction network is constructed, which includes a feature extraction layer, an adversarial learning layer, and a feature fusion layer. The feature extraction layer uses a convolutional structure to capture local features, the adversarial learning layer realizes domain adaptation, and the feature fusion layer integrates features at different levels to ultimately form an adversarial feature extractor.

[0033] 103. Perform multi-level differential disturbance on the supply chain data according to the supply chain data characteristics and the preset process parameter association map, and correct the disturbed data according to the process constraint relationship to obtain disturbed process parameter data; In one embodiment of the present invention, the supply chain data is subjected to multi-level differential perturbation according to the supply chain data characteristics and the preset process parameter association map, and the perturbed data is corrected according to the process constraint relationship to obtain the perturbed process parameter data, including: quantifying the process sensitivity of the supply chain data characteristics to obtain a parameter sensitivity matrix, performing correlation analysis according to the parameter sensitivity matrix and the process parameter association map to obtain a multi-level sensitive data set; performing differential privacy budget allocation on the multi-level sensitive data set to obtain perturbation budget values ​​at each level, performing parameter perturbation amount calculation on the perturbation budget values ​​to obtain initial perturbation data; performing constraint verification on the initial perturbation data and the process constraint relationship to obtain a constraint violation data identifier, performing constraint verification on the initial perturbation data according to the constraint violation data identifier to obtain a constraint violation degree; performing process constraint correction on the initial perturbation data according to the constraint violation degree to obtain corrected perturbation data, and performing process parameter reconstruction on the corrected perturbation data to obtain perturbed process parameter data.

[0034] Specifically, when quantifying the process sensitivity of supply chain data features, we first establish a parameter sensitivity assessment system, which includes three dimensions: process impact, commercial value, and leakage risk. For process impact, we score the degree of direct impact of parameters on product quality by analyzing them. For example, in the steel smelting process, parameters such as temperature and chemical composition have a high process impact. For commercial value, we evaluate the degree of correlation between parameters and the core competitiveness of the enterprise. For example, formula ratios, special process parameters, etc. have a high commercial value. For leakage risk, we quantify the degree of loss that may be caused by the acquisition of parameters by competitors. Through the comprehensive scoring of these three dimensions, we construct a parameter sensitivity matrix. Then, we cross-analyze the matrix with the process parameter association map to identify the correlation and impact path between parameters, and finally divide the data into three levels of data sets: high sensitivity, medium sensitivity, and low sensitivity.

[0035] Specifically, an adaptive budget allocation strategy is adopted when allocating differential privacy budgets. First, a basic privacy budget is set for each sensitivity level. A larger privacy budget is allocated to high-sensitivity data sets to provide stronger protection, while a relatively smaller budget is allocated to medium-sensitivity and low-sensitivity data sets. Budget allocation also needs to consider the frequency of data use and access scenarios. For data with high frequency access, the budget value should be appropriately increased, while for data with low frequency access, the budget value can be reduced. Based on the allocated budget value, the Laplace mechanism is used to calculate the perturbation of each parameter. The size of the perturbation is inversely proportional to the budget value and proportional to the sensitivity of the parameter. The initial perturbation data is obtained by adding the calculated noise.

[0036] Specifically, when performing constraint verification on the initial perturbation data, first extract the quantitative constraints and qualitative constraints in the process constraint relationship. Quantitative constraints include the physical limits of parameters (such as temperature range, pressure range, etc.) and the numerical relationships between parameters (such as proportional relationships, additive relationships, etc.); qualitative constraints include the changing trend of parameters, the direction of correlation, etc. The perturbed data is verified by these constraints. If a parameter value exceeds the physical limit or violates the relationship between parameters, it is marked as constraint violation data. For the marked constraint violation data, calculate the degree of deviation from the constraint conditions, including the degree of exceeding the limit, the degree of relationship violation, etc. These indicators together constitute the quantitative indicators of the degree of constraint violation.

[0037] Specifically, finally, the disturbance data is corrected for process constraints based on the degree of constraint violation. The correction process adopts an iterative optimization method. Each iteration first corrects the parameters with a larger degree of violation, and the original disturbance amount and constraint conditions are comprehensively considered during the correction. If there is a strong correlation between the parameters, the relevant parameters need to be corrected at the same time to maintain the process constraint relationship. The corrected data needs to be re-constrained and verified until all parameters meet the process constraints. Then the process parameters of the corrected disturbance data are reconstructed. The reconstruction process includes parameter normalization, dimensional recovery and numerical modification to ensure that the reconstructed parameters not only maintain the disturbance effect but also meet the process requirements, and finally obtain the perturbed process parameter data that meets the actual process requirements.

[0038] Furthermore, the differential privacy budget allocation processing is performed on the multi-level sensitive data set to obtain disturbance budget values ​​at each level, and the parameter disturbance amount calculation processing is performed on the disturbance budget value to obtain initial disturbance data, which includes: performing privacy sensitivity calculation processing on the multi-level sensitive data set to obtain a sensitivity scoring matrix, performing budget weight allocation processing according to the sensitivity scoring matrix to obtain an initial budget allocation scheme; performing constraint optimization processing on the initial budget allocation scheme to obtain an optimized budget value, performing Laplace noise generation processing according to the optimized budget value to obtain a noise parameter set; performing step-by-step disturbance amount calculation processing on the noise parameter set to obtain a disturbance amount matrix, performing parameter disturbance superposition processing according to the disturbance amount matrix to obtain a disturbance superposition result; performing numerical normalization processing on the disturbance superposition result to obtain normalized disturbance data, and performing parameter reconstruction processing according to the normalized disturbance data to obtain initial disturbance data.

[0039] Specifically, when calculating the privacy sensitivity of multi-level sensitive data sets, a multi-dimensional sensitivity evaluation index system is first established. The system includes data value dimensions (such as the commercial value, technical value, and strategic value of data), security risk dimensions (such as the impact of data leakage, the risk of data abuse, etc.), and usage scenario dimensions (such as data access frequency, data sharing scope, etc.). For each dimension, a detailed scoring standard is set. For example, in the data value dimension, core process parameters and key quality indicators have higher value scores; in the security risk dimension, formula data and process control parameters have higher risk scores; in the usage scenario dimension, real-time monitoring data and quality inspection data have higher usage scores. The analytic hierarchy process is used to perform weighted calculations on each dimension, and finally a sensitivity scoring matrix containing all evaluation indicators is obtained. Then, a budget weight allocation strategy is designed based on the matrix, and a nonlinear mapping function is used to convert the sensitivity score into a budget weight. The higher the sensitivity of the data, the higher the budget weight is, forming an initial budget allocation plan.

[0040] Specifically, when constraining the optimization of the initial budget allocation scheme, three main constraints need to be considered: the overall privacy budget constraint, the process quality constraint, and the data availability constraint. First, based on the total privacy budget limit set by the system, the budget allocation value of each data item is normalized to ensure that the total budget does not exceed the limit. Secondly, according to the process quality requirements, the minimum interference threshold of the key process parameters is set to ensure that the data after adding noise can still meet the process requirements. Finally, based on the data availability requirements, the tolerance range of data perturbations is set to ensure that the data still has practical application value after perturbations. The Lagrange multiplier method is used to solve these constraints to obtain the optimized budget value that meets multiple constraints. According to the optimized budget value, random noise is generated using the Laplace distribution, and the scale parameter of the noise is inversely proportional to the budget value, thereby obtaining a noise parameter set.

[0041] Specifically, a hierarchical perturbation strategy is adopted when calculating the perturbation amount step by step. First, the perturbation amount is calculated for the data at the high sensitivity level, and the size of the perturbation amount is proportional to the budget value and the sensitivity of the data. Then the perturbation amount is calculated for the data at the medium sensitivity level, while considering the correlation with the highly sensitive data to avoid data inconsistency caused by the perturbation. Finally, the perturbation amount of the low sensitivity level data is calculated to ensure that the perturbed data still maintains basic statistical characteristics. The perturbations of all levels are organized into a perturbation amount matrix, which reflects the degree of perturbation at different levels and different parameters. Then, the original data is perturbed and superimposed based on the matrix. The superposition process adopts an additive perturbation mechanism while maintaining the relative relationship between the parameters.

[0042] Specifically, finally, the disturbance superposition results are numerically normalized to ensure the consistency of data distribution. The normalization process first calculates the statistical characteristics of the disturbance data (such as mean, variance, etc.), and then normalizes it through methods such as min-max normalization or z-score normalization. Different normalization strategies are used for different types of parameters, such as linear normalization for continuous parameters and interval mapping normalization for discrete parameters. The normalized data needs to be reconstructed. The reconstruction process includes dimensional recovery, physical meaning conversion and numerical modification to ensure that the reconstructed data not only maintains the privacy protection effect, but also conforms to the value characteristics of the actual process parameters, and finally obtains the initial disturbance data.

[0043] 104. Perform dynamic on-chain adversarial verification on model access behavior based on the disturbed process parameter data and process-related hash values. When abnormal model access is detected, differentiated disturbance intensity adjustments and access restrictions are performed on distributed storage supply chain data.

[0044] In one embodiment of the present invention, the dynamic on-chain adversarial verification of the model access behavior is performed according to the disturbed process parameter data and the process-associated hash value, and the distributed storage supply chain data is subjected to differentiated disturbance intensity adjustment and access restriction when abnormal model access is detected, including: performing smart contract deployment processing on the disturbed process parameter data to obtain an adversarial verification contract, generating verification rule processing according to the adversarial verification contract and the process-associated hash value to obtain dynamic on-chain verification rules; performing feature extraction processing on the model access behavior to obtain an access feature vector, performing adversarial verification processing according to the access feature vector and the dynamic on-chain verification rules to obtain a verification score and anomaly identification; performing credibility quantification processing on the model access behavior according to the verification score and the anomaly identification to obtain access credibility, performing dynamic threshold determination processing on the access credibility to obtain an access control decision; performing differentiated disturbance intensity calculation processing on the distributed storage supply chain data according to the access control decision to obtain a disturbance adjustment plan, performing on-chain deployment processing on the disturbance adjustment plan to obtain an updated access restriction strategy.

[0045] Specifically, when deploying smart contracts, it is first necessary to encode the disturbed process parameter data and convert it into a data structure that can be processed by smart contracts. The encoding process adopts a hierarchical structure. The first layer is the basic data layer, which contains the numerical value, unit and timestamp of the parameter; the second layer is the disturbance information layer, which records the disturbance type, disturbance intensity and disturbance range of the parameter; the third layer is the constraint information layer, which contains the physical constraints, process constraints and quality constraints of the parameter. Based on these encoded data, an adversarial verification contract is constructed. The contract structure includes a data verification module, an access control module and a disturbance adjustment module. By deploying smart contracts, the system embeds the process-related hash value into the contract and establishes a data integrity verification mechanism. In the process of generating verification rules, the process constraint information is converted into a rule expression, including parameter threshold rules, parameter relationship rules and timing constraint rules. These rules together constitute a dynamic on-chain verification rule set.

[0046] Specifically, a multi-dimensional analysis method is used to extract the features of model access behavior. First, the frequency features, period features, and duration features of access are extracted from the time dimension; the geographic location features, network topology features, and device identification features of access are extracted from the spatial dimension; and the operation type features, data volume features, and access pattern features of access are extracted from the behavior dimension. These features are standardized to form an access feature vector, and the vector dimension is determined by the number of features. The adversarial verification process matches the access feature vector with the dynamic chain verification rules, calculates the conformity of the features with the rules, and analyzes the degree of abnormality of the access behavior. The verification score reflects the compliance of the access behavior, and the score ranges from 0 to 100, where a score above 80 indicates normal access, 60 to 80 indicates that it needs to be paid special attention, and a score below 60 is marked as abnormal access. The abnormal identification records the specific types of abnormalities in detail, such as frequent access abnormalities, unauthorized access abnormalities, or pattern abnormalities.

[0047] Specifically, in the access credibility quantification stage, the system comprehensively considers the verification score and anomaly identification, and constructs a credibility assessment model based on multiple factors. The evaluation factors include historical access records, current access characteristics, degree of abnormal behavior, and process knowledge relevance. Each factor has a corresponding weight coefficient, and the access credibility score is obtained through weighted calculation. The dynamic threshold judgment adopts an adaptive method to dynamically adjust the threshold parameters according to the system operation status and security level. When the access credibility is lower than the threshold, the system generates corresponding control decisions, and the decision types include temporarily restricting access, increasing the disturbance intensity, or completely denying access. These decisions are then converted into specific access control policies.

[0048] Specifically, in the process of differentiated calculation of disturbance intensity, the basic principle of disturbance adjustment is first determined according to the access control decision. For low-credibility access behaviors, the system increases the intensity of data disturbance accordingly, and the disturbance intensity is negatively correlated with the credibility. The disturbance adjustment scheme contains three key elements: disturbance parameter selection, disturbance intensity calculation, and disturbance implementation strategy. The specific calculation of disturbance intensity takes into account the sensitivity of parameters, the abnormality of access, and the tolerance of the process. After the calculated disturbance adjustment scheme is verified for security, it is deployed to the blockchain network through smart contracts. The contract execution process will automatically trigger corresponding access restriction measures, including increasing data noise, reducing data accuracy, or limiting access frequency. The transaction records generated during the entire deployment process are permanently stored in the blockchain, ensuring the transparency and immutability of the access control strategy, and ultimately achieving security protection for supply chain data.

[0049] In this embodiment, a real-time updated environmental digital twin model is constructed by collecting and fusion processing of multi-dimensional environmental data of the construction site. According to the model and the individual physiological data of the workers, the safety status of each worker is adaptively evaluated to obtain an individualized dynamic safety threshold. Combining the environmental digital twin model, individualized dynamic safety threshold and real-time status data, the graph neural network reasoning technology is used to comprehensively evaluate the safety situation and predict potential risk trends. According to the safety situation assessment results and risk prediction trends, a situation-aware hierarchical processing method is used to generate adaptive early warning and intervention plans for different risk levels, providing accurate safety prompts. This method monitors the safety status of workers in real time and predicts early warning according to the dynamic changes of the construction site, improves the intelligence level and accuracy of construction safety management, and reduces the risk of accidents.

[0050] The above describes the supply chain data security management method in the embodiment of the present invention. The following describes the supply chain data security management system in the embodiment of the present invention. Figure 2 , an embodiment of the supply chain data security management system in an embodiment of the present invention includes: The hierarchical storage module 201 is used to perform hierarchical division and distributed storage processing on the supply chain data according to the production links of the supply chain, and generate a process-related hash value based on the process constraint information and the distributedly stored supply chain data; A feature confrontation module 202 is used to construct a feature extractor containing a process constraint relationship based on the supply chain data and the process-related hash value, and to perform domain confrontation feature extraction on the supply chain data based on the feature extractor to obtain supply chain data features that meet physical constraints; The differential perturbation module 203 is used to perform multi-level differential perturbation on the supply chain data according to the supply chain data characteristics and the preset process parameter association map, and to correct the disturbed data according to the process constraint relationship to obtain disturbed process parameter data; The access verification module 204 is used to perform dynamic on-chain adversarial verification on the model access behavior based on the disturbed process parameter data and the process-associated hash value, and to perform differentiated disturbance intensity adjustment and access restriction on the distributed stored supply chain data when abnormal model access is detected.

[0051] In an embodiment of the present invention, the supply chain data security management system runs the above supply chain data security management method, and the supply chain data security management system generates process-related hash values ​​by layering and distributed storage processing of the supply chain data according to the production links; constructs a feature extractor containing process constraint relationships based on the supply chain data and process-related hash values, and uses domain adversarial features to extract supply chain data features that meet physical constraints; performs multi-level differential perturbations based on the supply chain data features and process parameter association maps, and corrects the perturbed data based on the process constraint relationships; performs dynamic on-chain adversarial verification of model access behavior based on the perturbed process parameter data and process-related hash values, and performs differentiated perturbation intensity adjustment and access restrictions when abnormal access is detected. The present invention prevents competitors from obtaining commercial confidential information through model reverse reasoning and ensures the security and effectiveness of data sharing by introducing process constraints and multi-level differential perturbations in the process of transfer learning.

[0052] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device, or unit can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0053] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0054] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A supply chain data security management method, characterized in that: The supply chain data security management method comprises: The supply chain data is divided into layers and distributedly stored according to the production links of the supply chain, and the process-related hash value is generated based on the process constraint information and the distributedly stored supply chain data; Building a feature extractor containing a process constraint relationship based on the supply chain data and process-related hash values, and performing domain adversarial feature extraction on the supply chain data based on the feature extractor to obtain supply chain data features that meet physical constraints; Performing multi-level differential disturbance on the supply chain data according to the supply chain data characteristics and the preset process parameter association map, and correcting the disturbed data according to the process constraint relationship to obtain disturbed process parameter data; The model access behavior is dynamically verified on the chain based on the disturbed process parameter data and the process-associated hash value. When abnormal model access is detected, differentiated disturbance intensity adjustments and access restrictions are performed on the distributed stored supply chain data.

2. The supply chain data security management method according to claim 1, characterized in that: The step of performing hierarchical division and distributed storage processing on the supply chain data according to the production links of the supply chain, and generating a process-related hash value based on the process constraint information and the distributedly stored supply chain data includes: Classify the supply chain data into production links to obtain raw material link data, manufacturing link data and product link data; Use IPFS distributed storage to store the classified supply chain data in blocks to obtain data blocks and data block indexes; Constructing a Merkle tree for the data block based on the process constraint information to obtain a Merkle tree with process constraints; A process-associated hash value is generated according to the Merkle tree with process constraints and process constraint information.

3. The supply chain data security management method according to claim 1, characterized in that: The step of constructing a feature extractor containing a process constraint relationship based on the supply chain data and the process-related hash value, and performing domain adversarial feature extraction on the supply chain data based on the feature extractor to obtain supply chain data features that satisfy physical constraints includes: Perform data domain labeling on process parameter data in the supply chain data to obtain data samples with source identifiers, group process process-related hash values ​​according to the source identifiers, and obtain feature constraint conditions for each data domain; Performing source domain and target domain division processing on the data sample to obtain an initial domain division result, performing consistency verification processing on the initial domain division result according to the feature constraint condition to obtain a domain division strategy that meets the process constraint; Performing feature space mapping processing on the process parameter data according to the domain partitioning strategy to obtain an initial feature representation, and performing gradient reversal layer embedding processing on the initial feature representation to obtain an adversarial feature extractor; The adversarial feature extractor is subjected to process constraint injection processing to obtain a feature extraction network with process constraints, and the supply chain data is subjected to feature extraction processing based on the feature extraction network to obtain supply chain data features that meet physical constraints.

4. The supply chain data security management method according to claim 3 is characterized in that: The process parameter data is subjected to feature space mapping processing according to the domain partitioning strategy to obtain an initial feature representation, and the initial feature representation is subjected to gradient reversal layer embedding processing to obtain an adversarial feature extractor, which includes: Performing feature dimension analysis processing on the domain partitioning strategy to obtain a dimension weight vector, and performing weighted mapping processing on the process parameter data according to the dimension weight vector to obtain an initial feature space; Performing nonlinear transformation processing on the initial feature space to obtain a hidden layer feature representation, and performing principal component projection processing on the hidden layer feature representation to obtain a reduced dimension feature matrix; Performing reverse gradient calculation processing on the dimension reduction feature matrix to obtain gradient reversal parameters, and performing back propagation processing according to the gradient reversal parameters to obtain an adversarial gradient matrix; The adversarial gradient matrix is ​​subjected to feature reconstruction processing to obtain an adversarial feature representation, and a feature extraction network is constructed based on the adversarial feature representation to obtain an adversarial feature extractor.

5. The supply chain data security management method according to claim 1, characterized in that: The multi-level differential perturbation of the supply chain data is performed according to the supply chain data characteristics and the preset process parameter association map, and the perturbed data is corrected according to the process constraint relationship to obtain the perturbed process parameter data including: Quantifying the process sensitivity of the supply chain data features to obtain a parameter sensitivity matrix, and performing correlation analysis based on the parameter sensitivity matrix and the process parameter correlation map to obtain a multi-level sensitive data set; Perform differential privacy budget allocation processing on the multi-level sensitive data set to obtain disturbance budget values ​​at each level, and perform parameter disturbance amount calculation processing on the disturbance budget values ​​to obtain initial disturbance data; Performing constraint verification processing on the initial disturbance data and the process constraint relationship to obtain a constraint violation data identifier, and performing constraint verification processing on the initial disturbance data according to the constraint violation data identifier to obtain a constraint violation degree; The initial disturbance data is subjected to process constraint correction processing according to the constraint violation degree to obtain corrected disturbance data, and the corrected disturbance data is subjected to process parameter reconstruction processing to obtain disturbed process parameter data.

6. The supply chain data security management method according to claim 5, characterized in that: The performing differential privacy budget allocation processing on the multi-level sensitive data set to obtain disturbance budget values ​​at each level, and performing parameter disturbance amount calculation processing on the disturbance budget values ​​to obtain initial disturbance data includes: Performing privacy sensitivity calculation processing on the multi-level sensitive data set to obtain a sensitivity scoring matrix, and performing budget weight allocation processing according to the sensitivity scoring matrix to obtain an initial budget allocation plan; Performing constraint optimization processing on the initial budget allocation scheme to obtain an optimized budget value, and performing Laplace noise generation processing according to the optimized budget value to obtain a noise parameter set; Performing a step-by-step disturbance amount calculation process on the noise parameter set to obtain a disturbance amount matrix, and performing a parameter disturbance superposition process according to the disturbance amount matrix to obtain a disturbance superposition result; The disturbance superposition result is numerically normalized to obtain normalized disturbance data, and parameter reconstruction is performed according to the normalized disturbance data to obtain initial disturbance data.

7. The supply chain data security management method according to claim 1, characterized in that: The dynamic on-chain adversarial verification of the model access behavior is performed according to the disturbed process parameter data and the process-related hash value, and the differentiated disturbance intensity adjustment and access restriction are performed on the distributed storage supply chain data when abnormal model access is detected, including: Perform smart contract deployment processing on the disturbed process parameter data to obtain an adversarial verification contract, generate verification rules based on the adversarial verification contract and the process-related hash value, and obtain dynamic on-chain verification rules; Perform feature extraction on the model access behavior to obtain an access feature vector, perform adversarial verification based on the access feature vector and dynamic on-chain verification rules, and obtain a verification score and anomaly identification; Performing credibility quantification processing on the model access behavior according to the verification score and the abnormal identification to obtain access credibility, and performing dynamic threshold determination processing on the access credibility to obtain an access control decision; According to the access control decision, a differentiated disturbance intensity calculation process is performed on the distributedly stored supply chain data to obtain a disturbance adjustment plan, and the disturbance adjustment plan is deployed on-chain to obtain an updated access restriction strategy.

8. A supply chain data security management system, characterized in that: The supply chain data security management system includes: A hierarchical storage module is used to perform hierarchical division and distributed storage processing on the supply chain data according to the production links of the supply chain, and to generate a process-related hash value based on the process constraint information and the distributedly stored supply chain data; A feature adversarial module, used to construct a feature extractor containing a process constraint relationship based on the supply chain data and the process-related hash value, and to perform domain adversarial feature extraction on the supply chain data based on the feature extractor to obtain supply chain data features that meet physical constraints; A differential perturbation module, used to perform multi-level differential perturbation on the supply chain data according to the supply chain data characteristics and a preset process parameter association map, and to correct the disturbed data according to the process constraint relationship to obtain disturbed process parameter data; The access verification module is used to perform dynamic on-chain adversarial verification on the model access behavior based on the disturbed process parameter data and the process-associated hash value, and to perform differentiated disturbance intensity adjustment and access restriction on the distributedly stored supply chain data when abnormal model access is detected.

Images