A spatio-temporal fine-grained privacy protection method and system for industrial manufacturing scene distributed learning

Abstract

The application provides a kind of space-time fine-grained privacy protection method and system for industrial manufacturing scene distributed learning.The method comprises: industrial manufacturing scene distributed learning construction;Space hierarchical modeling and privacy level division;Differential privacy, homomorphic encryption and data desensitization line node side configuration and execution;Time dynamic privacy control mechanism design;Space-time joint privacy parameter optimization and aggregation;Global aggregation and privacy budget recovery;Model updating and stage feedback adjustment.The application comprehensively considers the spatial safety level difference and time phase change characteristics in the industrial manufacturing process, by introducing hierarchical differential privacy, homomorphic encryption and data desensitization mechanism in the spatial dimension, introducing dynamic privacy budget and encryption intensity control mechanism in the time dimension, forming a space-time joint fine-grained privacy protection system.

Description

Technical Field

[0001] This invention belongs to the field of privacy protection technology in distributed learning in industrial manufacturing scenarios, specifically involving a spatiotemporal fine-grained privacy protection method and system for distributed learning in industrial manufacturing scenarios. Background Technology

[0002] With the digitalization and intelligentization of industrial manufacturing, distributed learning is widely used for collaborative model training across different production lines to avoid privacy risks associated with centralized data. However, existing privacy protection mechanisms are mostly statically configured, making it difficult to meet the multi-level security requirements and dynamic production characteristics in industrial scenarios.

[0003] On the one hand, different production lines exhibit significant differences in security levels and data sensitivity. Core production lines involve critical process and equipment information, while ordinary production lines only contain general production data. Using differential privacy or homomorphic encryption of uniform strength could lead to insufficient protection for highly sensitive production lines or excessive protection for low-sensitivity production lines, impacting system performance. On the other hand, industrial production exhibits distinct phases, such as different privacy protection requirements during R&D, trial production, process adjustments, and mass production. However, existing methods cannot dynamically adjust privacy budgets or encryption strength according to these phases. Furthermore, data also varies in timeliness; historical data differs from newly added data in sensitivity. Uniform desensitization could easily lead to a waste of privacy resources or a decrease in utilization.

[0004] To overcome the above challenges, designing a technical solution that can simultaneously achieve dynamic fine-grained privacy protection in both spatial and temporal dimensions to meet the security, flexibility, and practicality requirements of distributed learning in industrial manufacturing scenarios is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a spatiotemporal fine-grained privacy protection method and system for distributed learning in industrial manufacturing scenarios. It aims to solve the static, single-level, and coarse-grained problems in privacy protection of distributed learning in industrial manufacturing in the prior art, and to achieve multi-level and dynamic privacy protection across production lines and production stages.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A spatiotemporal fine-grained privacy protection method for distributed learning in industrial manufacturing scenarios, the method comprising:

[0008] A distributed learning framework consisting of a central server and production line nodes is constructed. Each production line node independently conducts training using its private data in its local production line environment and uploads the parameter updates to the central server.

[0009] Based on a distributed learning framework, spatial hierarchical modeling and privacy level classification are performed according to the security level and data sensitivity of each production line node to determine the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node.

[0010] Based on the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node, differential privacy, homomorphic encryption and data anonymization operations are performed on each production line node configuration on the production line node side, respectively injecting noise into the parameter update amount, performing ciphertext encryption and feature blurring processing;

[0011] A dynamic privacy control mechanism is constructed in the time dimension, and the differential privacy budget, encryption strength and data desensitization coefficient are adaptively adjusted through stage identification and time decay model;

[0012] A spatiotemporal joint privacy parameter optimization model is constructed with the goal of minimizing the global privacy cost function and constraining the model training loss to not exceed a preset threshold, thereby achieving joint optimization of privacy protection strength and model performance.

[0013] The central server performs weighted aggregation on the encrypted parameter update amount without decryption to generate new global model parameters, and reclaims and updates the privacy budget of each production line node;

[0014] The new global model parameters are distributed to each production line node. Based on the privacy consumption and model performance feedback at each stage, the privacy parameters for the next round of training are dynamically adjusted, forming a closed-loop adaptive privacy control process.

[0015] Preferably, based on the differential privacy budget, homomorphic encryption method, and data anonymization coefficient of each production line node, the method of configuring and performing differential privacy, homomorphic encryption, and data anonymization operations on each production line node at the production line node side, and injecting noise into the parameter update amount, performing ciphertext encryption, and feature blurring processing respectively includes:

[0016] Differential privacy mechanism: Before each round of parameter upload, noise is injected into the locally generated parameter update amount to achieve differential privacy protection.

[0017] ;

[0018] in, This represents the parameter update amount after noise injection. For locally generated parameter update amount, Indicates noise intensity. It is a Gaussian function;

[0019] Homomorphic encryption mechanism: Encrypted computation is performed after differential privacy processing is completed.

[0020] ;

[0021] in, This refers to the amount of time the encrypted parameters are updated. Indicates the encryption function selected by the production line node;

[0022] Data anonymization mechanism: Before entering local training, attribute anonymization and feature blurring operations are performed on highly sensitive data.

[0023] ;

[0024] in, This is the anonymized data. This is highly sensitive data. This is the data anonymization coefficient.

[0025] Preferably, a dynamic privacy control mechanism is constructed in the time dimension, and the differential privacy budget, encryption strength, and data anonymization coefficient are adaptively adjusted through stage identification and time decay models.

[0026] The differential privacy budget as a function of time is expressed as:

[0027] ;

[0028] in, For the current differential privacy budget, For the initial differential privacy budget, This is a stage adjustment coefficient;

[0029] To address the timeliness differences between historical and newly added data, a data anonymization coefficient decay model over time is introduced:

[0030] ;

[0031] in, The anonymization coefficient for the current data. The initial data anonymization coefficient is used. This indicates the interval between the data generation time and the current time. This is the decay rate parameter;

[0032] While the differential privacy budget and data anonymization coefficient change over time, the homomorphic encryption parameters and aggregation strategy are adjusted simultaneously.

[0033] Preferably, the method for constructing a spatiotemporal joint privacy parameter optimization model, aiming to minimize the global privacy cost function and constraining the model training loss to not exceed a preset threshold, to achieve joint optimization of privacy protection strength and model performance includes:

[0034] The global privacy cost function is:

[0035] ;

[0036] in, , , These are the weighting coefficients for the spatial, temporal, and desensitization dimensions, respectively. For differential privacy budgeting;

[0037] The constraints are:

[0038] ;

[0039] in, For global training loss, For baseline model performance, This is the performance deviation threshold.

[0040] Preferred methods for reclaiming and updating the privacy budget at each production line node include:

[0041] ;

[0042] in, For the first Differential privacy budget for round training, For the first Differential privacy budget for round training, For adjustment coefficients, This represents the change in the model's convergence performance.

[0043] This invention also provides a spatiotemporal fine-grained privacy protection system for distributed learning in industrial manufacturing scenarios. The system is used to implement the aforementioned method and includes: a framework construction module, a spatial module, a configuration execution module, a temporal module, a spatiotemporal joint module, an aggregation and recycling module, and an update feedback module.

[0044] The framework building module is used to build a distributed learning framework that includes a central server and production line nodes. Each production line node independently conducts training using its private data in its local production line environment and uploads the parameter updates to the central server.

[0045] The spatial module is used to perform spatial hierarchical modeling and privacy level classification based on the security level and data sensitivity of each production line node, and to determine the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node, based on the distributed learning framework.

[0046] The configuration execution module is used to configure and execute differential privacy, homomorphic encryption and data desensitization operations on each production line node based on the differential privacy budget, homomorphic encryption method and data desensitization coefficient of each production line node, respectively injecting noise into the parameter update amount, performing ciphertext encryption and feature blurring processing;

[0047] The time module is used to build a dynamic privacy control mechanism in the time dimension. Through stage identification and time decay model, it adaptively adjusts the differential privacy budget, encryption strength and data desensitization coefficient.

[0048] The spatiotemporal joint module is used to construct a spatiotemporal joint privacy parameter optimization model. With the goal of minimizing the global privacy cost function, it constrains the model training loss to not exceed a preset threshold, thereby achieving joint optimization of privacy protection strength and model performance.

[0049] The aggregation and recycling module is used by the central server to perform weighted aggregation on the encrypted parameter update amount to generate new global model parameters without decryption, and to reclaim and update the privacy budget of each production line node.

[0050] The update feedback module is used to distribute new global model parameters to each production line node. Based on the privacy consumption and model performance feedback at each stage, the privacy parameters for the next round of training are dynamically adjusted to form a closed-loop adaptive privacy control process.

[0051] Preferably, the configuration execution module includes:

[0052] Differential privacy mechanism: Before each round of parameter upload, noise is injected into the locally generated parameter update amount to achieve differential privacy protection.

[0053] ;

[0054] in, This represents the parameter update amount after noise injection. For locally generated parameter update amount, Indicates noise intensity. It is a Gaussian function;

[0055] Homomorphic encryption mechanism: Encrypted computation is performed after differential privacy processing is completed.

[0056] ;

[0057] in, This refers to the amount of time the encrypted parameters are updated. Indicates the encryption function selected by the production line node;

[0058] Data anonymization mechanism: Before entering local training, attribute anonymization and feature blurring operations are performed on highly sensitive data.

[0059] ;

[0060] in, This is the anonymized data. This is highly sensitive data. This is the data anonymization coefficient.

[0061] Preferably, the time module includes:

[0062] The differential privacy budget as a function of time is expressed as:

[0063] ;

[0064] in, For the current differential privacy budget, For the initial differential privacy budget, This is a stage adjustment coefficient;

[0065] To address the timeliness differences between historical and newly added data, a data anonymization coefficient decay model over time is introduced:

[0066] ;

[0067] in, The anonymization coefficient for the current data. The initial data anonymization coefficient is used. This indicates the interval between the data generation time and the current time. This is the decay rate parameter;

[0068] While the differential privacy budget and data anonymization coefficient change over time, the homomorphic encryption parameters and aggregation strategy are adjusted simultaneously.

[0069] Preferably, the spatiotemporal joint module includes:

[0070] The global privacy cost function is:

[0071] ;

[0072] in, , , These are the weighting coefficients for the spatial, temporal, and desensitization dimensions, respectively. For differential privacy budgeting;

[0073] The constraints are:

[0074] ;

[0075] in, For global training loss, For baseline model performance, This is the performance deviation threshold.

[0076] Preferably, the aggregation and recycling module includes:

[0077] ;

[0078] in, For the first Differential privacy budget for round training, For the first Differential privacy budget for round training, For adjustment coefficients, This represents the change in the model's convergence performance.

[0079] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0080] 1. Spatial hierarchical privacy protection is highly effective.

[0081] By implementing multi-level differential privacy, homomorphic encryption, and data anonymization strategies based on production line security levels and data sensitivity in the spatial dimension, differentiated configuration of privacy protection is achieved. This mechanism effectively prevents the risk of data leakage in highly sensitive production lines while avoiding excessive computation and performance degradation for low-sensitivity production line nodes.

[0082] 2. Time-based dynamic privacy control enhances system adaptability.

[0083] By introducing a time-dynamic privacy budget and a desensitization coefficient decay mechanism, the system can automatically adjust the strength of privacy protection according to changes in production stage and data timeliness. This method achieves adaptive privacy control throughout the entire lifecycle of R&D, prototyping, and mass production, improving the system's ability to balance privacy and performance in long-term operation.

[0084] 3. Spatiotemporal joint optimization achieves a synergistic balance between privacy and performance.

[0085] By constructing a spatiotemporal joint privacy parameter optimization model and a budget recovery mechanism, global collaborative optimization of differential privacy budget, encryption strength, and desensitization coefficient is achieved. This mechanism ensures that while improving model convergence efficiency and accuracy, the overall security, stability, and optimal resource utilization of the distributed system are maintained. Attached Figure Description

[0086] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0087] Figure 1 This is a schematic diagram of the spatiotemporal fine-grained privacy protection method for distributed learning in industrial manufacturing scenarios according to an embodiment of the present invention. Detailed Implementation

[0088] 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.

[0089] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0090] Example 1

[0091] This invention provides a spatiotemporal fine-grained privacy protection method for distributed learning in industrial manufacturing scenarios, suitable for the need to balance privacy protection and training performance in multi-production line collaborative modeling, including the following steps:

[0092] A distributed learning framework consisting of a central server and production line nodes is constructed. Each production line node independently conducts training using its private data in its local production line environment and uploads the parameter updates to the central server.

[0093] Based on a distributed learning framework, spatial hierarchical modeling and privacy level classification are performed according to the security level and data sensitivity of each production line node to determine the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node.

[0094] Based on the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node, differential privacy, homomorphic encryption and data anonymization operations are performed on each production line node configuration on the production line node side, respectively injecting noise into the parameter update amount, performing ciphertext encryption and feature blurring processing;

[0095] A dynamic privacy control mechanism is constructed in the time dimension, and the differential privacy budget, encryption strength and data desensitization coefficient are adaptively adjusted through stage identification and time decay model;

[0096] A spatiotemporal joint privacy parameter optimization model is constructed with the goal of minimizing the global privacy cost function and constraining the model training loss to not exceed a preset threshold, thereby achieving joint optimization of privacy protection strength and model performance.

[0097] The central server performs weighted aggregation on the encrypted parameter update amount without decryption to generate new global model parameters, and reclaims and updates the privacy budget of each production line node;

[0098] The new global model parameters are distributed to each production line node. Based on the privacy consumption and model performance feedback at each stage, the privacy parameters for the next round of training are dynamically adjusted, forming a closed-loop adaptive privacy control process.

[0099] like Figure 1 As shown, the specific implementation process of the present invention is as follows:

[0100] Step 1: Constructing Distributed Learning for Industrial Manufacturing Scenarios

[0101] The spatiotemporal fine-grained privacy protection method for distributed learning in this invention is implemented based on a distributed learning system. This system includes a central server (Server) and multiple production line nodes (Clients). Each production line node independently trains its model using its private data within its local production line environment and uploads the parameter update results of its local model to the central server. The central server receives and merges the parameter update information from each production line node to complete the global model update, and then distributes the updated model parameters to each production line node, achieving cross-production line joint optimization. This process keeps data locally within the factory, ensuring that production and modeling tasks are executed in parallel.

[0102] Step 1.1 Overall Architecture

[0103] In industrial manufacturing scenarios, the system adopts a distributed learning framework of "central server - multiple production line nodes". Specifically:

[0104] Central Server: Responsible for the aggregation, updating and distribution of global model parameters; maintaining functions such as production line node registration, communication scheduling, and model version management; receiving parameter update results from each production line node and performing global aggregation in each training iteration.

[0105] Production line node (Client): Deployed in edge computing terminals of each independent production line or workshop, it uses local production line sensor data, visual inspection data, process logs and other private data for model training; after completing local training, each production line node only uploads the parameter gradient or weight difference, and does not upload the original sample data.

[0106] This architecture enables the decoupling and parallelization of computational tasks in a multi-production-line environment, ensuring that the model training process does not depend on centralized data storage, thereby effectively avoiding the risk of industrial data leakage.

[0107] Step 1.2 Distributed Training Process

[0108] The distributed learning process consists of multiple synchronous iterative rounds, each round including the following steps:

[0109] 1. Global Initialization: The central server loads the initial model parameters. And distribute it to all production line nodes;

[0110] 2. Local Training: Each production line node performs several rounds of local updates using local production line data to obtain the [number]th [item / process]. iThe production line node at the first Local model parameters of the wheel ;

[0111] 3. Parameter Upload: Difference in calculated parameters for production line nodes The difference results are perturbed and encrypted. For the first i The production line node at the first Local model parameters of the wheel;

[0112] 4. Global Aggregation: The server aggregates data based on the weight of production line nodes. Execute weighted fusion: ,in It can be dynamically adjusted based on the scale, stability, or latency of production line node data. N This refers to the number of production line nodes.

[0113] 5. Model Distribution: The server will distribute the aggregated global model parameters. The data is distributed to each production line node for the next round of training, enabling continuous convergence and synchronous updates.

[0114] This distributed learning framework enables joint modeling and knowledge sharing across production lines while maintaining data locality. It effectively improves the generalization ability and stability of the model, ensuring that the production process and modeling tasks are carried out synchronously, and providing basic support for hierarchical aggregation and feature optimization in subsequent steps.

[0115] Step 2: Spatial hierarchical modeling and privacy level classification

[0116] To address the differences in data sensitivity, security level, and leakage risk among different industrial production lines, this invention designs a hierarchical modeling and privacy level classification mechanism in the spatial dimension. This mechanism is used to establish configurable differential privacy parameters, encryption methods, and desensitization strengths, providing a foundation for subsequent dynamic control.

[0117] Step 2.1 Production Line Safety Level Classification

[0118] First, the spatial security levels of each industrial production line participating in distributed learning are categorized into three levels: core production lines, important production lines, and general production lines. Core production lines, involving critical process parameters, core components, or high-value visual data, belong to the highest security level; important production lines, involving routine process and equipment operation data, have a medium security level; and general production lines, containing only routine operation or environmental data, have the lowest security level. The system generates a security level identifier for each production line based on enterprise security standards, production line task attributes, and data usage scenarios. This serves as the basis for configuring subsequent privacy policies.

[0119] Step 2.2 Data Sensitivity Assessment and Classification

[0120] After determining the security level, the system further performs sensitivity assessments and classifications on the local data of each production line. Assessment indicators include the identifiability of the data content, its business relevance, and the impact of potential leakage. The system divides the data into three categories:

[0121] Highly sensitive data: such as unpublished product design drawings, experimental quality test samples, and equipment control instructions;

[0122] Zhongmin data: such as routine production process parameters and equipment operation logs;

[0123] Low-sensitivity data: such as environmental monitoring values ​​and general process statistics.

[0124] The data sensitivity of each production line is based on This is indicated and stored in the local policy configuration file, along with the security level identifier ( These together constitute the basic information for privacy classification.

[0125] Step 2.3 Privacy Level Policy Generation

[0126] After determining the security level identifier and data sensitivity, these two are combined and mapped to a privacy level policy table for each production line node. This policy table is used to determine the differential privacy budget, homomorphic encryption method, and data anonymization strength for each production line node. The system uses a mapping function... The privacy configuration parameters for each production line node are calculated and defined as follows: .in: A differential privacy budget is used to control the intensity of noise injection; The homomorphic encryption method determines the encryption level of the uploaded parameters (fully homomorphic, partially homomorphic, or lightweight encryption). is the data anonymization coefficient, representing the proportion of data obfuscated before local training.

[0127] The mapping results can be presented in tabular form in the system, used to assign an initial privacy policy to each production line, and invoked and dynamically updated during subsequent training.

[0128] Step 2.4 Strategy Management and Version Tracking

[0129] To ensure the traceability and stability of the policy, the system establishes a privacy policy version management mechanism. When production line tasks, data sources, or security requirements change, the system automatically reassesses the production line level and data sensitivity, and generates a new version of the policy file. Historical versions will be archived with timestamps to allow for quick restoration to the previous effective version in case of anomalies or backtracking analysis, enabling the strategy to be rollbackable and continuously optimized. Through the aforementioned spatial hierarchical modeling and privacy level classification mechanism, this invention enables the ability to configure privacy policies differently based on production line security risks and data sensitivity in a distributed learning system, providing fundamental support for subsequent dynamic time-based control and global aggregation optimization.

[0130] Step 3: Node-side privacy protection configuration and execution

[0131] Based on the aforementioned spatial hierarchical modeling and privacy level classification results, this invention configures a corresponding privacy protection strategy for each production line node, and implements differential privacy, homomorphic encryption, and data anonymization operations during model upload and local training. This step ensures that each production line node meets the preset privacy strength requirements during parameter updates and data processing, achieving multi-layered protection before upload and local leakage prevention control.

[0132] Step 3.1 Differential Privacy Mechanism

[0133] Before uploading parameters in each round, noise is injected into the locally generated model update or gradient vector to achieve differential privacy protection. The operation process is as follows:

[0134] ;

[0135] in, This represents the parameter update amount after noise injection. For locally generated parameter update amount, The noise intensity is represented by the differential privacy budget. Decide, It is a Gaussian function.

[0136] The core production line uses a smaller To enhance protection strength, while general production lines can choose larger ones. To minimize performance loss, the system performs norm clipping on the gradient vector before injecting noise, ensuring consistent data distribution across all production line nodes. This mechanism effectively prevents the risk of inferring local training data features from model parameters.

[0137] Step 3.2 Homomorphic Encryption Mechanism

[0138] After differential privacy processing, the system encrypts the uploaded parameters or gradient results to prevent information leakage during intermediate transmission. The calculation format is as follows:

[0139] ;

[0140] in, This refers to the amount of time the encrypted parameters are updated. This indicates the encryption function selected by the production line node.

[0141] The core production line employs Fully Homomorphic Encryption (FHE), enabling direct addition and multiplication operations within the ciphertext field. Important production lines utilize Partially Homomorphic Encryption (PHE), retaining only additivity or multiplication characteristics. General production lines employ lightweight encryption combined with Transport Layer Security (TLS) for protection. This encryption mechanism ensures that data uploaded during parameter processing remains unreadable and immutable without affecting aggregation calculations.

[0142] Step 3.3 Data anonymization mechanism

[0143] Before the data enters the local training module, the system performs attribute desensitization and feature blurring operations on highly sensitive data. Its calculation form can be expressed as:

[0144] ;

[0145] in, This is the anonymized data. This is highly sensitive data. This is the data anonymization coefficient, used to control the proportion of visual information or salient features retained.

[0146] For visual data, the system blurs or randomly perturbs the background and non-defect regions; for time-series or parametric data, it masks some information through intervalization, hashing, and frequency domain attenuation. This process retains only task-related features, thereby reducing potential privacy risks without affecting model performance.

[0147] Through the configuration and execution of differential privacy, homomorphic encryption, and data anonymization on the production line node side, this invention establishes a multi-layered security protection structure before and after model uploading, realizing end-to-end privacy protection from data source → local training → parameter uploading → aggregation calculation, ensuring the balance and controllability of cross-production line distributed learning between data security and modeling performance.

[0148] Step 4: Design of a Time-Dynamic Privacy Control Mechanism

[0149] In terms of time, the industrial manufacturing process typically goes through different stages, such as R&D and prototyping, process changes, and stable mass production. The privacy protection requirements differ significantly at each stage. The R&D stage often involves highly sensitive design and experimental data, requiring strong privacy protection; while in the mass production stage, the data structure is stable, and the protection level can be appropriately reduced to improve model training efficiency.

[0150] To this end, the present invention constructs a dynamic privacy control mechanism in the time dimension, which adaptively adjusts the differential privacy budget, encryption strength and desensitization coefficient through stage identification and time decay model.

[0151] Step 4.1 Dynamically Adjusting the Privacy Budget

[0152] Define stage variables , respectively representing the research and development stage Pilot-scale stage and mass production stage .

[0153] The system automatically loads the corresponding privacy template based on the current stage and adjusts the current differential privacy budget accordingly. .

[0154] The differential privacy budget as a function of time is expressed as:

[0155] ;

[0156] in, For the current differential privacy budget, For the initial differential privacy budget, Adjustment coefficient for each stage: R&D stage This indicates strong protection and weak learning; mass production stage. This indicates weak protection and strong learning; pilot stage To achieve a balance between the two.

[0157] This mechanism allows for dynamic adjustment of differential privacy strength based on the production stage, enabling a smooth transition from high security to high performance.

[0158] Step 4.2 Time decay of data anonymization coefficient

[0159] To address the timeliness differences between historical and newly added data, a data anonymization coefficient decay model over time is introduced:

[0160] ;

[0161] in, The anonymization coefficient for the current data. The initial data anonymization coefficient is used. This indicates the interval between the data generation time and the current time. This is the decay rate parameter.

[0162] This mechanism ensures that as data ages, its sensitivity and protection strength decrease in tandem, thereby improving the efficiency of using historical data in retraining and model calibration and avoiding the ineffective consumption of privacy budgets.

[0163] Step 4.3 Dynamic Encryption and Policy Linkage

[0164] While the differential privacy budget and data anonymization coefficient change over time, the homomorphic encryption parameters and aggregation strategy are adjusted simultaneously.

[0165] When the system detects a decrease in the privacy budget or a decay of the desensitization coefficient to a preset threshold, it automatically triggers an encryption strategy switch: downgrading from fully homomorphic encryption (FHE) to partially homomorphic encryption (PHE), or switching to lightweight encryption mode during mass production to reduce communication overhead and encryption computation burden.

[0166] This dynamic linkage mechanism is implemented through time stamping and policy control interfaces, ensuring consistency between privacy protection strength, encryption policies, and data lifecycle.

[0167] Step 4.4 Adaptive Optimization in the Time Dimension

[0168] After each training round, the system evaluates the privacy protection effectiveness in real time based on changes in model performance and the rate of privacy consumption. It automatically corrects for fluctuations in training performance or excessively rapid depletion of the privacy budget. or Parameters enable adaptive balancing between stages.

[0169] Through this mechanism, the system can dynamically adjust privacy protection parameters at different production stages, balancing training stability and privacy security, and forming a continuous self-optimization mechanism in the time dimension.

[0170] Step 5: Spatiotemporal Joint Privacy Parameter Optimization and Aggregation

[0171] To ensure that the spatial and temporal privacy strategies work synergistically, this invention constructs a spatiotemporal joint privacy parameter optimization model to achieve a global balance between privacy protection strength and model performance. The system dynamically and jointly optimizes the spatial differential privacy parameters, temporal stage adjustment coefficients, and data anonymization parameters at the central endpoint, and synchronously updates the global weights during the aggregation process.

[0172] Step 5.1 Design of Global Privacy Cost Function

[0173] To measure the contribution of privacy parameters of different dimensions to the overall security of the system, this invention defines a global privacy cost function as follows:

[0174] ;

[0175] in, , , These are the weighting coefficients for the spatial, temporal, and desensitization dimensions, respectively. For differential privacy budgeting.

[0176] This cost function is used to quantify the cost of privacy protection in multi-production line nodes and multi-stage scenarios, so that trade-offs can be adjusted in subsequent optimizations.

[0177] Step 5.2 Joint Optimization Constraints and Solution

[0178] The system minimizes the cost to global privacy. The goal is to constrain training performance to remain within an acceptable range.

[0179] The constraints are:

[0180] ;

[0181] in, For global training loss, For baseline model performance, This is the performance deviation threshold.

[0182] This constraint ensures that while reducing the cost of privacy, the model's accuracy and convergence performance do not suffer significant degradation.

[0183] Step 5.3 Aggregation and Dynamic Weight Adjustment

[0184] After the encrypted upload is completed at each production line node, the central server performs a weighted aggregation of the encrypted parameters from each production line node. The aggregation formula is as follows:

[0185] ;

[0186] in, These are the parameters after weighted aggregation. The aggregation weights are dynamically adjusted based on the security level, remaining privacy budget, and network latency of the production line nodes. For production line nodes using fully homomorphic encryption, the system performs computable aggregation directly in the ciphertext domain; for lightweight encrypted production line nodes, a secure channel aggregation strategy is adopted. After aggregation, the system updates the global model parameters and recalculates the privacy cost for the next round of optimization.

[0187] Step 5.4 Spatiotemporal parameter collaborative update

[0188] After each round of optimization, the system dynamically adjusts the spatial dimension based on privacy budget consumption and changes in model performance. With the time dimension .

[0189] If a performance degradation exceeding the performance deviation threshold is detected... The system will automatically increase the number of production line nodes. or weaken To restore training efficiency; conversely, when performance is stable but security risks increase, the privacy budget is automatically tightened or the desensitization intensity is increased.

[0190] This two-way control mechanism enables a continuous balance between privacy and performance across time and space.

[0191] Step 6: Global Aggregation and Privacy Budget Reclamation

[0192] Step 6.1 Global Aggregation Mechanism

[0193] After completing local encryption and data anonymization, each production line node uploads the encrypted model updates to the central server. Without decrypting any production line node data, the server performs a weighted aggregation operation to generate new global model parameters. The aggregation process is defined as follows:

[0194] ;

[0195] in, The amount of encrypted parameter updates uploaded by the production line nodes. This is the aggregate weight.

[0196] The aggregation weights are dynamically determined based on factors such as the reliability of the production line nodes, their training contribution, the rate of privacy budget consumption, and communication latency. For production line nodes using fully homomorphic encryption (FHE), the server performs aggregation operations directly in the encrypted domain; for lightweight encrypted production line nodes, weighted aggregation is performed within a secure channel. This mechanism ensures that the aggregation process achieves global model updates without exposing any original parameters, effectively preventing the risk of intermediate data leakage.

[0197] Step 6.2 Privacy Budget Recovery and Adaptive Adjustment

[0198] After global aggregation is complete, the system reclaims and dynamically updates the privacy budget for each production line node to prevent budget exhaustion or security shifts during long-term training. The system calculates the privacy budget adjustment based on the model's convergence and performance changes during this training round. The update formula is as follows:

[0199] ;

[0200] in, For the first Differential privacy budget for round training, For the first Differential privacy budget for round training, For adjustment coefficients, This represents the change in the model's convergence performance.

[0201] When model performance improves When the model performance deteriorates or fluctuations intensify, the system appropriately relaxes the privacy budget to improve training efficiency; when the model performance deteriorates or fluctuations intensify, the budget is automatically tightened to enhance protection. Through this budget recovery mechanism, the system can achieve an adaptive balance between privacy budget and model performance in multiple iterations, forming a closed-loop control process of "budget allocation - training consumption - dynamic recovery".

[0202] Step 7: Model Update and Stage Feedback Adjustment

[0203] After completing global aggregation, the central server distributes the updated global model to each production line node. Upon receiving the new model, each production line node restarts its local training process based on its own production line status and production stage, thus achieving synchronous updates and continuous optimization of the distributed system.

[0204] During model deployment, the system records in real-time metrics such as privacy budget consumption, model performance trends, and production planning cycles for the current training round, using these as feedback for the next stage of privacy configuration parameters. Based on this feedback, the system adjusts the differential privacy budget for the next stage. Homomorphic encryption method Data anonymization coefficient Reassessment and parameter reallocation: If the privacy budget consumption of a production line node is too high or the security risk increases in the previous stage, the system will automatically increase the encryption level or increase the desensitization coefficient; if the training performance of a production line node declines or the model converges insufficiently, the system will appropriately relax the privacy budget to restore training efficiency; if the production cycle changes or the stage switches (such as from trial production to mass production), the system will automatically load the corresponding stage template and reset the privacy parameter baseline.

[0205] Through this feedback and reconfiguration mechanism, the system forms a closed-loop privacy control process of "aggregation-distribution-evaluation-adjustment", which enables model updates and phased privacy strategies to continuously and adaptively evolve, ensuring the privacy security stability and model performance continuity of industrial distributed learning in long-term operation.

[0206] In summary, this invention provides a spatiotemporal fine-grained privacy protection method for distributed learning in industrial manufacturing scenarios. It comprehensively considers the differences in spatial security levels and the characteristics of temporal phase changes in the industrial manufacturing process. By introducing hierarchical differential privacy, homomorphic encryption and data desensitization mechanisms in the spatial dimension, and introducing dynamic privacy budget and encryption strength control mechanisms in the temporal dimension, a spatiotemporal joint fine-grained privacy protection system is formed.

[0207] Example 2

[0208] This invention also provides a spatiotemporal fine-grained privacy protection system for distributed learning in industrial manufacturing scenarios, used to implement the methods described in the foregoing embodiments, including: a framework construction module, a spatial module, a configuration execution module, a temporal module, a spatiotemporal joint module, an aggregation and recycling module, and an update feedback module;

[0209] The framework building module is used to build a distributed learning framework that includes a central server and production line nodes. Each production line node independently conducts training using its private data in its local production line environment and uploads the parameter updates to the central server.

[0210] The spatial module is used to perform spatial hierarchical modeling and privacy level classification based on the security level and data sensitivity of each production line node, and to determine the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node, based on the distributed learning framework.

[0211] The configuration execution module is used to configure and execute differential privacy, homomorphic encryption and data desensitization operations on each production line node based on the differential privacy budget, homomorphic encryption method and data desensitization coefficient of each production line node, respectively injecting noise into the parameter update amount, performing ciphertext encryption and feature blurring processing;

[0212] The time module is used to build a dynamic privacy control mechanism in the time dimension. Through stage identification and time decay model, it adaptively adjusts the differential privacy budget, encryption strength and data desensitization coefficient.

[0213] The spatiotemporal joint module is used to construct a spatiotemporal joint privacy parameter optimization model. With the goal of minimizing the global privacy cost function, it constrains the model training loss to not exceed a preset threshold, thereby achieving joint optimization of privacy protection strength and model performance.

[0214] The aggregation and recycling module is used by the central server to perform weighted aggregation on the encrypted parameter update amount to generate new global model parameters without decryption, and to reclaim and update the privacy budget of each production line node.

[0215] The update feedback module is used to distribute new global model parameters to each production line node. Based on the privacy consumption and model performance feedback at each stage, the privacy parameters for the next round of training are dynamically adjusted to form a closed-loop adaptive privacy control process.

[0216] Furthermore, the configuration execution module includes:

[0217] Differential privacy mechanism: Before each round of parameter upload, noise is injected into the locally generated parameter update amount to achieve differential privacy protection.

[0218] ;

[0219] in, This represents the parameter update amount after noise injection. For locally generated parameter update amount, Indicates noise intensity. It is a Gaussian function;

[0220] Homomorphic encryption mechanism: Encrypted computation is performed after differential privacy processing is completed.

[0221] ;

[0222] in, This refers to the amount of time the encrypted parameters are updated. Indicates the encryption function selected by the production line node;

[0223] Data anonymization mechanism: Before entering local training, attribute anonymization and feature blurring operations are performed on highly sensitive data.

[0224] ;

[0225] in, This is the anonymized data. This is highly sensitive data. This is the data anonymization coefficient.

[0226] Furthermore, the time module includes:

[0227] The differential privacy budget as a function of time is expressed as:

[0228] ;

[0229] in, For the current differential privacy budget, For the initial differential privacy budget, This is a stage adjustment coefficient;

[0230] To address the timeliness differences between historical and newly added data, a data anonymization coefficient decay model over time is introduced:

[0231] ;

[0232] in, The anonymization coefficient for the current data. The initial data anonymization coefficient is used. This indicates the interval between the data generation time and the current time. This is the decay rate parameter;

[0233] While the differential privacy budget and data anonymization coefficient change over time, the homomorphic encryption parameters and aggregation strategy are adjusted simultaneously.

[0234] Furthermore, the spatiotemporal joint module includes:

[0235] The global privacy cost function is:

[0236] ;

[0237] in, , , These are the weighting coefficients for the spatial, temporal, and desensitization dimensions, respectively. For differential privacy budgeting;

[0238] The constraints are:

[0239] ;

[0240] in, For global training loss, For baseline model performance, This is the performance deviation threshold.

[0241] Furthermore, the aggregation and recycling module includes:

[0242] ;

[0243] in, For the first Differential privacy budget for round training, For the first Differential privacy budget for round training, For adjustment coefficients, This represents the change in the model's convergence performance.

[0244] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A spatiotemporal fine-grained privacy protection method for distributed learning in industrial manufacturing scenarios, characterized in that, The method includes: A distributed learning framework consisting of a central server and production line nodes is constructed. Each production line node independently conducts training using its private data in its local production line environment and uploads the parameter updates to the central server. Based on a distributed learning framework, spatial hierarchical modeling and privacy level classification are performed according to the security level and data sensitivity of each production line node to determine the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node. Based on the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node, differential privacy, homomorphic encryption and data anonymization operations are performed on each production line node configuration on the production line node side, respectively injecting noise into the parameter update amount, performing ciphertext encryption and feature blurring processing; A dynamic privacy control mechanism is constructed in the time dimension, and the differential privacy budget, encryption strength and data desensitization coefficient are adaptively adjusted through stage identification and time decay model. A spatiotemporal joint privacy parameter optimization model is constructed with the goal of minimizing the global privacy cost function and constraining the model training loss to not exceed a preset threshold, thereby achieving joint optimization of privacy protection strength and model performance. The central server performs weighted aggregation on the encrypted parameter update amount without decryption to generate new global model parameters, and reclaims and updates the privacy budget of each production line node; The new global model parameters are distributed to each production line node. Based on the privacy consumption and model performance feedback at each stage, the privacy parameters for the next round of training are dynamically adjusted, forming a closed-loop adaptive privacy control process.

2. The method of claim 1, wherein, Based on the differential privacy budget, homomorphic encryption method, and data anonymization coefficient of each production line node, differential privacy, homomorphic encryption, and data anonymization operations are configured and performed on each production line node at the production line node side. The methods for injecting noise into parameter update amounts, performing ciphertext encryption, and feature obfuscation processing include: Differential privacy mechanism: Before each round of parameter upload, noise is injected into the locally generated parameter update amount to achieve differential privacy protection. ； wherein is the parameter update after injection of noise, is the locally generated parameter update, denotes the noise strength, is a Gaussian function; Homomorphic encryption mechanism: Encrypted computation is performed after differential privacy processing. ； wherein, is the updated parameter after encryption, denotes the encryption function used by the line node; Data anonymization mechanism: Before entering local training, attribute anonymization and feature blurring operations are performed on highly sensitive data. ； wherein, is the data after desensitization, is the high sensitive data, is the data desensitization coefficient.

3. The method of claim 2, wherein, Methods for constructing a dynamic privacy control mechanism over time, which adaptively adjust differential privacy budget, encryption strength, and data anonymization coefficient through stage identification and time decay models, include: The differential privacy budget as a function of time is expressed as: ； in, For the current differential privacy budget, For the initial differential privacy budget, This is a stage adjustment coefficient; To address the timeliness differences between historical and newly added data, a data anonymization coefficient decay model over time is introduced: ； in, The anonymization coefficient for the current data. The initial data anonymization coefficient is used. This indicates the interval between the data generation time and the current time. This is the decay rate parameter; While the differential privacy budget and data anonymization coefficient change over time, the homomorphic encryption parameters and aggregation strategy are adjusted simultaneously.

4. The method of claim 3, wherein, Methods for jointly optimizing privacy protection strength and model performance by constructing a spatiotemporal joint privacy parameter optimization model, aiming to minimize the global privacy cost function, and constraining the model training loss to not exceed a preset threshold, include: The global privacy cost function is: ； in, , , These are the weighting coefficients for the spatial, temporal, and desensitization dimensions, respectively. For differential privacy budgeting; The constraints are: ； wherein, is the global training loss, is the baseline model performance, is the performance bias threshold.

5. The method of claim 4, wherein, Methods for reclaiming and updating the privacy budget at each production line node include: ； in, For the first Differential privacy budget for round training, For the first Differential privacy budget for round training, For adjustment coefficients, This represents the change in the model's convergence performance.

6. A spatiotemporal fine-grained privacy protection system for distributed learning in industrial manufacturing scenarios, the system being used to implement the method described in any one of claims 1-5, characterized in that, The system includes: a framework construction module, a spatial module, a configuration execution module, a time module, a spatiotemporal joint module, an aggregation and recycling module, and an update feedback module; The framework building module is used to build a distributed learning framework that includes a central server and production line nodes. Each production line node independently conducts training using its private data in its local production line environment and uploads the parameter updates to the central server. The spatial module is used to perform spatial hierarchical modeling and privacy level classification based on the security level and data sensitivity of each production line node, and to determine the differential privacy budget, homomorphic encryption method and data anonymization coefficient of each production line node, based on the distributed learning framework. The configuration execution module is used to configure and execute differential privacy, homomorphic encryption and data desensitization operations on each production line node based on the differential privacy budget, homomorphic encryption method and data desensitization coefficient of each production line node, respectively injecting noise into the parameter update amount, performing ciphertext encryption and feature blurring processing; The time module is used to build a dynamic privacy control mechanism in the time dimension. Through stage identification and time decay model, it adaptively adjusts the differential privacy budget, encryption strength and data desensitization coefficient. The spatiotemporal joint module is used to construct a spatiotemporal joint privacy parameter optimization model. With the goal of minimizing the global privacy cost function, it constrains the model training loss to not exceed a preset threshold, thereby achieving joint optimization of privacy protection strength and model performance. The aggregation and recycling module is used by the central server to perform weighted aggregation on the encrypted parameter update amount to generate new global model parameters without decryption, and to reclaim and update the privacy budget of each production line node. The update feedback module is used to distribute new global model parameters to each production line node. Based on the privacy consumption and model performance feedback at each stage, the privacy parameters for the next round of training are dynamically adjusted to form a closed-loop adaptive privacy control process.

7. The system of claim 6, wherein, The configuration execution module includes: Differential privacy mechanism: Before each round of parameter upload, noise is injected into the locally generated parameter update amount to achieve differential privacy protection. ； wherein is the parameter update after injection of noise, is the locally generated parameter update, denotes the noise strength, is a Gaussian function; Homomorphic encryption mechanism: Encrypted computation is performed after differential privacy processing. ； wherein, is the updated parameter after encryption, represents the encryption function used by the line node; Data anonymization mechanism: Before entering local training, attribute anonymization and feature blurring operations are performed on highly sensitive data. ； wherein, is the data after desensitization, is the high sensitive data, is the data desensitization coefficient.

8. The system of claim 7, wherein, The time module includes: The differential privacy budget as a function of time is expressed as: ； wherein, is the current differential privacy budget, is the initial differential privacy budget, is a phase adjustment coefficient; To address the timeliness differences between historical and newly added data, a data anonymization coefficient decay model over time is introduced: ； in, The anonymization coefficient for the current data. The initial data anonymization coefficient is used. This indicates the interval between the data generation time and the current time. This is the decay rate parameter; While the differential privacy budget and data anonymization coefficient change over time, the homomorphic encryption parameters and aggregation strategy are adjusted simultaneously.

9. The system of claim 8, wherein, The spatiotemporal co-operation module includes: The global privacy cost function is: ； in, , , These are the weighting coefficients for the spatial, temporal, and desensitization dimensions, respectively. For differential privacy budgeting; The constraints are: ； in, For global training loss, For baseline model performance, This is the performance deviation threshold.

10. The system of claim 9, wherein, The aggregation and recycling module includes: ； wherein, is the differential privacy budget for the first round of training, is the differential privacy budget for the first round of training, is the adjustment coefficient, is the change in model convergence performance.

Images