Hot rolled strip crown prediction method and system based on privacy protection and federated learning

Abstract

This invention provides a method and system for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning, belonging to the field of hot-rolled strip steel crown prediction and privacy protection technology. The method includes: generating a system key and performing capability adaptation partitioning for each distributed client; each distributed client training a local hot-rolled strip steel crown prediction model using local rolling training data, and quantizing, compressing, and encrypting the model weights; an aggregator, based on RAMCFE and AoNE, securely aggregating the encrypted weight data from multiple clients to calculate the weight update amount of the global model; each distributed client receiving the weight update amount of the global model and updating its local hot-rolled strip steel crown prediction model; and each distributed client using the updated local hot-rolled strip steel crown prediction model to predict the crown of the local hot-rolled strip steel based on local rolling data, and outputting the prediction result. This invention enables the prediction of hot-rolled strip steel crown.

Description

Technical Field

[0001] This invention belongs to the field of hot-rolled strip crown prediction and privacy protection technology, specifically referring to a method and system for predicting hot-rolled strip crown based on privacy protection and federated learning. Background Technology

[0002] Hot-rolled strip steel, as a fundamental material of modern industry, is widely used in construction, automobile manufacturing, and key national projects. However, the crown of hot-rolled strip steel directly determines product safety, service life, and production qualification rate. Traditional methods for predicting the crown of hot-rolled strip steel rely on empirical experiments or physical metallurgical models, which suffer from high costs, long cycles, and poor adaptability. In contrast, data-driven artificial neural network models, with their powerful nonlinear fitting capabilities, have become a core technology for optimizing rolling processes, reducing defect rates (such as center waviness and edge waviness), and achieving real-time quality control.

[0003] However, existing artificial neural network prediction models face critical privacy and security challenges in industrial applications. The weights of these models, as core parameters, are easily stolen and copied if transmitted or stored in plaintext. Furthermore, the parameters can be used to infer sensitive training data (such as rolling force, inlet temperature, and proprietary alloy composition), severely damaging a company's intellectual property and trade secrets.

[0004] To address this privacy issue, existing technologies primarily employ three schemes: homomorphic encryption, differential privacy, and secure multi-party computation. However, none of these can meet the specific constraints of hot-rolled strip steel production. Homomorphic encryption supports arbitrary computation of encrypted data, but incurs high computational overhead, resulting in prediction latency far exceeding industrial real-time requirements and making it unsuitable for real-time control scenarios in the rolling process. Differential privacy achieves privacy protection by introducing noise, but noise reduces prediction accuracy, making it difficult to meet the micron-level error requirements for scenarios such as hot-rolled strip steel crown prediction. Secure multi-party computation requires frequent interactions in distributed workshops or factories, increasing network latency and failure risks, and cannot adapt to the dynamic start / stop requirements of clients in industrial environments. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method and system for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning. The technical solution is as follows: On the one hand, a method for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning is provided, which includes: S1. Key Initialization and Dynamic Client Partitioning: Generate system keys and perform capability adaptation partitioning for various distributed clients, laying the foundation for subsequent encryption and aggregation; S2. Local Training and Security Encryption: Each distributed client uses local rolling training data to train a local hot-rolled strip crown prediction model, and quantizes, compresses, and encrypts the model weights. S3. Secure Aggregation: The aggregator is based on the robust adaptive multi-client function encryption algorithm RAMCFE and the all-or-nothing encapsulation mechanism AoNE to securely aggregate encrypted weight data from multiple clients and calculate the weight update amount of the global model. S4. Model Update and Termination Check: Each distributed client receives the weight update amount of the global model and updates the local hot-rolled strip crown prediction model, and performs iterative iteration and termination judgment. S5. Prediction Output: Each distributed client uses the updated local hot-rolled strip crown prediction model to predict the crown of the hot-rolled strip to be predicted based on local rolling data and outputs the prediction results.

[0006] On the other hand, a hot-rolled strip crown prediction system based on privacy protection and federated learning is provided, the system comprising: The key initialization and client dynamic partitioning module is used to generate system keys and perform partitioning for various distributed client capabilities, laying the foundation for subsequent encryption and aggregation. The local training and security encryption module is used by each distributed client to train a local hot-rolled strip crown prediction model using local rolling training data, and to quantize, compress, and encrypt the model weights. The secure aggregation module is used by the aggregator to securely aggregate encrypted weight data from multiple clients based on the robust adaptive multi-client function encryption algorithm RAMCFE and the all-or-nothing encapsulation mechanism AoNE, and calculate the weight update amount of the global model. The model update and termination check module is used by each distributed client to receive the weight update amount of the global model and update the local hot-rolled strip crown prediction model, and to perform iterative looping and termination judgment. The prediction output module is used by each distributed client to use the updated local hot-rolled strip crown prediction model to predict the crown of the hot-rolled strip to be predicted based on local rolling data and output the prediction results.

[0007] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described method for predicting the crown of hot-rolled strip based on privacy protection and federated learning.

[0008] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement the above-described method for predicting the crown of hot-rolled strip based on privacy protection and federated learning.

[0009] The beneficial effects of the technical solution provided by this invention include at least the following: This invention strikes a reasonable balance between privacy protection, prediction accuracy, and computational efficiency. By integrating the robust adaptive multi-client function encryption algorithm RAMCFE with the all-or-nothing encapsulation mechanism AoNE for privacy protection, and combining vector compression algorithm with dynamic load balancing of adaptive client partitioning strategy, it significantly reduces the computation and transmission overhead of encrypted aggregation and improves the global model iteration convergence efficiency while ensuring that model weights and sensitive process data are not leaked. Ultimately, it obtains the optimal model and meets the high-precision requirements for hot-rolled strip crown prediction.

[0010] When solving the problem of privacy protection of artificial neural network model parameters in the distributed hot-rolled strip steel production scenario, this invention can resist model theft and sensitive process data back-inference attacks, and can achieve full encryption protection of single client parameters without the need for a trusted third party. While ensuring privacy and security, this invention can meet the industrial-grade real-time and accuracy requirements of hot-rolled strip steel production, and is suitable for complex industrial environments with heterogeneous workshop computing capabilities and dynamic start-stop of clients.

[0011] This invention proposes a fault isolation architecture that integrates adaptive sharding based on computing power and full packet loss tolerance, through... The scoring mechanism dynamically senses the computing power and network status of heterogeneous workshops, and uses the AoNE mechanism to limit the impact of single-point disconnection to a local subset. This "sharding isolation-local discarding-global convergence" strategy solves the defect of traditional multi-client function encryption that causes global aggregation failure when nodes are offline. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0013] Figure 1 This is a flowchart of a method for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning, provided by an embodiment of the present invention. Figure 2This is a convergence trajectory diagram of a hot-rolled strip crown prediction method based on privacy protection and federated learning provided in an embodiment of the present invention under a high random drop rate of 20%. Figure 3 This is a prediction accuracy diagram of a hot-rolled strip crown prediction method based on privacy protection and federated learning provided in an embodiment of the present invention, under a high random drop rate of 20%. Figure 4 This is a block diagram of a hot-rolled strip crown prediction system based on privacy protection and federated learning provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0014] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0015] This invention provides a method for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning. This method can be implemented by an electronic device, which can be a terminal or a server. Figure 1 The diagram shown is a flowchart of the method. The processing flow may include the following steps: S1. Key Initialization and Dynamic Client Partitioning: Generate system keys and perform capability adaptation partitioning for various distributed clients, laying the foundation for subsequent encryption and aggregation; Optionally, S1 specifically includes: S1.1. Common Parameter Generation: The aggregator or the first participating client performs the initialization process to generate bilinear group parameters. and hash function To obtain an all-or-nothing encapsulation setup and public parameters Distribute the training round counter to all potential clients via a secure channel and initialize the training round counter. ; S1.2. Client Capability Assessment and Ranking: Each client Encrypted reporting of its real-time status information, including: CPU load rate. and network communication delay The aggregator uses a weighted comprehensive evaluation function to calculate the capability score for each client. To quantify the node processing capacity in heterogeneous industrial environments, nodes with sufficient computing resources and stable communication are selected. The calculation formula is as follows:

[0016] in, Indicates the client Current CPU load rate, Represents the available computing power reserve of a node; Indicates the client Network communication latency with the aggregator; The maximum tolerable delay threshold is set and used to normalize communication metrics; To prevent tiny constants with a denominator of zero; and These are the weighting coefficients for computing power and communication power, respectively, and they satisfy... And accordingly to all Sort the active clients in descending order to obtain an ordered set. ; S1.3. Dynamic Partitioning and Label Allocation: Set the size of the security target subset to... One client, will Divided into A subset, before Each subset contains 3 clients, and the last subset contains the remaining clients. Assign a unique session label .

[0017] For example, if there are 10 clients, the first 3 subsets each contain 3 clients, and the last subset contains 1 client.

[0018] S2. Local Training and Security Encryption: Each distributed client uses local rolling training data to train a local hot-rolled strip crown prediction model, and quantizes, compresses, and encrypts the model weights. Optionally, S2 specifically includes: S2.1. Local model training: per client Using its proprietary hot-rolling process dataset, a local hot-rolled strip crown prediction model is constructed. This model is a 34-12-1 topological BP neural network model. The 34-dimensional input layer receives 34 types of process feature data, including: inlet thickness, temperature, and rolling force, bending roll force, and roll shifting amount for each stand of the finishing mill. The 12-dimensional hidden layer uses activation functions to extract nonlinear metallurgical features. The single-neuron output layer outputs the final predicted hot-rolled strip crown value. Through forward and backward propagation, the client updates its local weight vector. To minimize the prediction error of strip crown; S2.2. Weight Quantization and Finite Field Mapping: For the client In the The weight update amount from the 12-dimensional hidden layer to the output layer obtained from local training. First, a 16-bit symmetric uniform quantization strategy is adopted, and the scaling factor is set. Map floating-point weights to symmetric integer intervals To retain 13 decimal places of precision, the quantization formula is:

[0019] Perform an offset mapping to accommodate the function encryption algorithm's requirement for non-negative integer input, transferring signed integers... Convert to positive integer :

[0020] At this point, the weighted components are mapped to the positive integer field. The client directly outputs the positive integers of these 12 dimensions. Construct the weight vector to be encrypted. :

[0021] S2.3. Function Encryption: Client Use public parameters Generated local private key and the session tag of the current fragment The weight vector constructed in step S2.2 Execute robust adaptive multi-client function encryption RAMCFE, with clients utilizing subsets public key set of all members Generate vector ciphertext :

[0022] Simultaneously, the client integrates the All-or-No Encapsulation (AoNE) mechanism to generate partial functional decryption key shares, including a random mask. The decryption key share of the aforementioned functions Functional components and package components It consists of two parts. The calculation formula is as follows:

[0023] in, For the client The private key; Aggregate function vector Corresponding client The weighted component (set to 1 in this embodiment of the invention); For the client A random mask matrix for sampling is used to achieve full packet loss tolerance; Function vector The hash mapping value;

[0024] Using the All-or-No-Encapsulation Algorithm (AoNE), a private key is encapsulated. For random mask matrix Encrypt it and combine it with the current subset. and conversation tags Binding, the client will ciphertext Decryption key share of the secret part function Send to the aggregator.

[0025] S3. Secure Aggregation: The aggregator is based on the robust adaptive multi-client function encryption algorithm RAMCFE and the all-or-nothing encapsulation mechanism AoNE to securely aggregate encrypted weight data from multiple clients and calculate the weight update amount of the global model. Optionally, S3 specifically includes: S3.1. Subset Integrity Verification: The aggregator receives the vector ciphertext from each client. and partial decryption key share Then, first check each subset. Whether the data is complete can only be determined if data packets from all clients within the subset have been delivered and the session tag is valid. Only when there is consistency is the subset marked as "complete"; otherwise, it is considered an "invalid subset" and discarded directly (triggering AoNE's full packet loss tolerance mechanism). S3.2. Complete decryption key synthesis: For the verified complete subset The aggregator utilizes an all-or-nothing encapsulation mechanism to encapsulate the key shares uploaded by all clients within the subset. By combining the subsets, if the subset is complete, the aggregator will successfully remove the random mask and synthesize the complete functional decryption key corresponding to that subset. ; S3.3. Vector Decompression and Global Model Calculation: The aggregator determines the function weight vector for this round based on the current federated learning aggregation strategy. The functional weight vector The contribution weights of each client's data in the aggregation process are defined, and the aggregator will... Set as an all-one vector corresponding to the size of the subset, so as to achieve subset pairing in the ciphertext field. The aggregator sums all client model updates using the synthesized key. RAMCFE decryption is performed on the set of ciphertext vectors within the subset. Based on the encryption properties of the inner product function, the decryption operation actually calculates the client weight vector set. With function vector Inner product: Decryption directly yields the exponential form of the subset aggregation vector:

[0026] The aggregator then employs the small-step-big-step algorithm (BSGS) to solve for the discrete logarithm within the polynomial, from... Recover plaintext aggregate vector (Since the aggregated weight update values ​​usually fall within a small integer range, the Small Step Big Step algorithm (BSGS) can be used), including: First, the boundaries of the search space are determined and a pre-computation table is constructed; Set the maximum range of values ​​for the aggregation result. Its calculation formula is ,in For the number of subset clients, The maximum quantization value for a single weight; then the step size parameter is calculated. To reduce the time complexity to And construct a "small step" hash table by calculating a series of bases. power value ,in Traversing the interval The generated key-value pairs Stored in a table for quick lookup; Secondly, perform a search match to recover the plaintext values; Calculate the "big step" baseline value and let the variable Incrementing from 0 Calculate the current search value After each calculation, check if the Val exists in the hash table. Once a match is found, it is considered to exist. Make This means that the equation has been found. The solution is found, at which point the search stops and the formula is applied. It directly recovers the algebraic sum of the encryption weights of all clients within the subset; Finally, the aggregator summarizes all valid subsets marked as "complete". These are denoted as Valid, and dequantization and offset removal operations are performed to calculate the global model's weight update. :

[0027] The Then it is broadcast to all clients.

[0028] S4. Model Update and Termination Check: Each distributed client receives the weight update amount of the global model and updates the local hot-rolled strip crown prediction model, and performs iterative iteration and termination judgment. Optionally, S4 specifically includes: S4.1. Model Update: Each client receives the weight update amount of the global model. And use it to replace the local model parameters; S4.2. Termination condition judgment: If the maximum number of training rounds is reached. Or, the change in the global model parameters, i.e., the change in the weight update amount of the global model, is less than a threshold: The training process ends when this time is reached; S4.3. Iteration or Termination: If not terminated, let If the condition is met, return to step S2; otherwise, output the output. This serves as the weight vector for the final trained local hot-rolled strip crown prediction model.

[0029] S5. Prediction Output: Each distributed client uses the updated local hot-rolled strip crown prediction model to predict the crown of the hot-rolled strip to be predicted based on local rolling data and outputs the prediction results.

[0030] Optionally, to ensure the prediction accuracy of the artificial neural network model after the function encryption algorithm, this embodiment of the invention uses the mean square error as the local objective function to minimize the prediction error of the strip convexity:

[0031] in, It is the actual convexity value. These are model predictions. That is the sample size.

[0032] To demonstrate the effectiveness of the method proposed in this embodiment, the method was trained on a proprietary hot-rolled strip dataset containing 34-dimensional process features and consisting of 10,000 samples to simulate processes with... A federated environment with 20 clients was used. The data was divided into an 80% training set and a 20% test set. To verify the fault tolerance of the proposed scheme in a dynamic industrial environment, 100 rounds of robustness testing were conducted, simulating a 20% high random disconnection rate. Figure 2 As shown, the mean squared error loss of the artificial neural network prediction model under the method of the present invention continues to decrease, and the accuracy eventually approaches 100%, indicating that the method proposed in the present invention effectively maintains the continuity and convergence of federated learning. Figure 3 The predictive performance of the final model is further demonstrated. The scatter plot shows that the predicted values ​​are closely aligned with the measured ground values ​​on the ideal line, indicating that the deviation of most samples is minimal, meeting the high-precision requirements of industrial rolling control.

[0033] like Figure 4 As shown, this embodiment of the invention also provides a hot-rolled strip crown prediction system based on privacy protection and federated learning, the system comprising: The key initialization and client dynamic partitioning module 410 is used to generate system keys and perform capability adaptation partitioning for various distributed clients, laying the foundation for subsequent encryption and aggregation. The local training and security encryption module 420 is used by each distributed client to train a local hot-rolled strip crown prediction model using local rolling training data, and to quantize, compress, and encrypt the model weights. The secure aggregation module 430 is used by the aggregator to securely aggregate encrypted weight data from multiple clients based on the robust adaptive multi-client function encryption algorithm RAMCFE and the all-or-nothing encapsulation mechanism AoNE, and to calculate the weight update amount of the global model. The model update and termination check module 440 is used for each distributed client to receive the weight update amount of the global model and update the local hot-rolled strip crown prediction model, and to perform iterative looping and termination judgment. The prediction output module 450 is used by each distributed client to use the updated local hot-rolled strip crown prediction model to predict the crown of the local hot-rolled strip based on the local rolling data and output the prediction results.

[0034] Optionally, the key initialization and client dynamic partitioning module is specifically used for: S1.1. Common Parameter Generation: The aggregator or the first participating client performs the initialization process to generate bilinear group parameters. and hash function To obtain an all-or-nothing encapsulation setup and public parameters Distribute the training round counter to all potential clients via a secure channel and initialize the training round counter. ; S1.2. Client Capability Assessment and Ranking: Each client Encrypted reporting of its real-time status information, including: CPU load rate. and network communication delay The aggregator uses a weighted comprehensive evaluation function to calculate the capability score for each client. To quantify the node processing capacity in heterogeneous industrial environments, nodes with sufficient computing resources and stable communication are selected. The calculation formula is as follows:

[0035] in, Indicates the client Current CPU load rate, Represents the available computing power reserve of a node; Indicates the client Network communication latency with the aggregator; The maximum tolerable delay threshold is set and used to normalize communication metrics; To prevent tiny constants with a denominator of zero; and These are the weighting coefficients for computing power and communication power, respectively, and they satisfy... And accordingly to all Sort the active clients in descending order to obtain an ordered set. ; S1.3. Dynamic Partitioning and Label Allocation: Set the size of the security target subset to... One client, will Divided into A subset, before Each subset contains 3 clients, and the last subset contains the remaining clients. Assign a unique session label .

[0036] Optionally, the local training and security encryption module is specifically used for: S2.1. Local model training: per client Using its proprietary hot-rolling process dataset, a local hot-rolled strip crown prediction model is constructed. This model is a 34-12-1 topological BP neural network model. The 34-dimensional input layer receives 34 types of process feature data, including: inlet thickness, temperature, and rolling force, bending roll force, and roll shifting amount for each stand of the finishing mill. The 12-dimensional hidden layer uses activation functions to extract nonlinear metallurgical features. The single-neuron output layer outputs the final predicted hot-rolled strip crown value. Through forward and backward propagation, the client updates its local weight vector. To minimize the prediction error of strip crown; S2.2. Weight Quantization and Finite Field Mapping: For the client In the The weight update amount from the 12-dimensional hidden layer to the output layer obtained from local training. First, a 16-bit symmetric uniform quantization strategy is adopted, and the scaling factor is set. Map floating-point weights to symmetric integer intervals To retain 13 decimal places of precision, the quantization formula is:

[0037] Perform an offset mapping to accommodate the function encryption algorithm's requirement for non-negative integer input, transferring signed integers... Convert to positive integer :

[0038] At this point, the weighted components are mapped to the positive integer field. The client directly outputs the positive integers of these 12 dimensions. Construct the weight vector to be encrypted. :

[0039] S2.3. Function Encryption: Client Use public parameters Generated local private key and the session tag of the current fragment The weight vector constructed in step S2.2 Execute robust adaptive multi-client function encryption RAMCFE, with clients utilizing subsets public key set of all members Generate vector ciphertext :

[0040] Simultaneously, the client integrates the All-or-No Encapsulation (AoNE) mechanism to generate partial functional decryption key shares, including a random mask. The decryption key share of the aforementioned functions Functional components and package components It consists of two parts. The calculation formula is as follows:

[0041] in, For the client The private key; Aggregate function vector Corresponding client The weighted components; For the client A random mask matrix for sampling is used to achieve full packet loss tolerance; Function vector The hash mapping value;

[0042] Using the All-or-No-Encapsulation Algorithm (AoNE), a private key is encapsulated. For random mask matrix Encrypt it and combine it with the current subset. and conversation tags Binding, the client will ciphertext Decryption key share of the secret part function Send to the aggregator.

[0043] Optionally, the security aggregation module is specifically used for: S3.1. Subset Integrity Verification: The aggregator receives the vector ciphertext from each client. and partial decryption key share Then, first check each subset. Whether the data is complete can only be determined if data packets from all clients within the subset have been delivered and the session tag is valid. Only when there is consistency is the subset marked as "complete"; otherwise, it is considered an "invalid subset" and discarded directly. S3.2. Complete decryption key synthesis: For the verified complete subset The aggregator utilizes an all-or-nothing encapsulation mechanism to encapsulate the key shares uploaded by all clients within the subset. By combining the subsets, if the subset is complete, the aggregator will successfully remove the random mask and synthesize the complete functional decryption key corresponding to that subset. ; S3.3. Vector Decompression and Global Model Calculation: The aggregator determines the function weight vector for this round based on the current federated learning aggregation strategy. The functional weight vector The contribution weights of each client's data in the aggregation process are defined, and the aggregator will... Set as an all-one vector corresponding to the size of the subset, so as to achieve subset pairing in the ciphertext field. The aggregator sums all client model updates using the synthesized key. RAMCFE decryption is performed on the set of ciphertext vectors within the subset. Based on the encryption properties of the inner product function, the decryption operation actually calculates the client weight vector set. With function vector Inner product: Decryption directly yields the exponential form of the subset aggregation vector:

[0044] The aggregator then employs the small-step-big-step algorithm (BSGS) to solve for the discrete logarithm within the polynomial, from... Recover plaintext aggregate vector ,include: First, the boundaries of the search space are determined and a pre-computation table is constructed; Set the maximum range of values ​​for the aggregation result. Its calculation formula is ,in For the number of subset clients, The maximum quantization value for a single weight; then the step size parameter is calculated. To reduce the time complexity to And construct a "small step" hash table by calculating a series of bases. power value ,in Traversing the interval The generated key-value pairs Stored in a table for quick lookup; Secondly, perform a search match to recover the plaintext values; Calculate the "big step" baseline value and let the variable Incrementing from 0 Calculate the current search value After each calculation, check if the Val exists in the hash table. Once a match is found, it is considered to exist. Make This means that the equation has been found. The solution is found, at which point the search stops and the formula is applied. It directly recovers the algebraic sum of the encryption weights of all clients within the subset; Finally, the aggregator summarizes all valid subsets marked as "complete". These are denoted as Valid, and dequantization and offset removal operations are performed to calculate the global model's weight update. :

[0045] The Then it is broadcast to all clients.

[0046] Optionally, the model update and termination check module is specifically used for: S4.1. Model Update: Each client receives the weight update amount of the global model. And use it to replace the local model parameters; S4.2. Termination condition judgment: If the maximum number of training rounds is reached. Or, the change in the global model parameters, i.e., the change in the weight update amount of the global model, is less than a threshold: The training process ends when this time is reached; S4.3. Iteration or Termination: If not terminated, let If the condition is met, return to step S2; otherwise, output the output. This serves as the weight vector for the final trained local hot-rolled strip crown prediction model.

[0047] The hot-rolled strip crown prediction system based on privacy protection and federated learning provided in this embodiment of the invention has a functional structure that corresponds to the hot-rolled strip crown prediction method based on privacy protection and federated learning provided in this embodiment of the invention, and will not be described again here.

[0048] Figure 5 This is a schematic diagram of the structure of an electronic device 500 provided in an embodiment of the present invention. The electronic device 500 may vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 501 and one or more memories 502. The memory 502 stores at least one instruction, which is loaded and executed by the processor 501 to implement the steps of the above-described method for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning.

[0049] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to perform the above-described method for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0050] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning, characterized in that, The method includes: S1. Key Initialization and Dynamic Client Partitioning: Generate system keys and perform capability adaptation partitioning for various distributed clients, laying the foundation for subsequent encryption and aggregation; S2. Local Training and Security Encryption: Each distributed client uses local rolling training data to train a local hot-rolled strip crown prediction model, and quantizes, compresses, and encrypts the model weights. S3. Secure Aggregation: The aggregator is based on the robust adaptive multi-client function encryption algorithm RAMCFE and the all-or-nothing encapsulation mechanism AoNE to securely aggregate encrypted weight data from multiple clients and calculate the weight update amount of the global model. S4. Model Update and Termination Check: Each distributed client receives the weight update amount of the global model and updates the local hot-rolled strip crown prediction model, and performs iterative iteration and termination judgment. S5. Prediction Output: Each distributed client uses the updated local hot-rolled strip crown prediction model to predict the crown of the hot-rolled strip to be predicted based on local rolling data and outputs the prediction results.

2. The method according to claim 1, characterized in that, S1 specifically includes: S1.

1. Common Parameter Generation: The aggregator or the first participating client performs the initialization process to generate bilinear group parameters. and hash function To obtain an all-or-nothing encapsulation setup and public parameters Distribute the training round counter to all potential clients via a secure channel and initialize the training round counter. ; S1.

2. Client Capability Assessment and Ranking: Each client Encrypted reporting of its real-time status information, including: CPU load rate. and network communication delay The aggregator uses a weighted comprehensive evaluation function to calculate the capability score for each client. To quantify the node processing capacity in heterogeneous industrial environments, nodes with sufficient computing resources and stable communication are selected. The calculation formula is as follows: in, Indicates the client Current CPU load rate, Represents the available computing power reserve of a node; Indicates the client Network communication latency with the aggregator; The maximum tolerable delay threshold is set and used to normalize communication metrics; To prevent tiny constants with a denominator of zero; and These are the weighting coefficients for computing power and communication power, respectively, and they satisfy... And accordingly to all Sort the active clients in descending order to obtain an ordered set. ; S1.

3. Dynamic Partitioning and Label Allocation: Set the size of the security target subset to... One client, will Divided into A subset, before Each subset contains 3 clients, and the last subset contains the remaining clients. Assign a unique session label .

3. The method according to claim 2, characterized in that, S2 specifically includes: S2.

1. Local model training: per client Using its proprietary hot-rolling process dataset, a local hot-rolled strip crown prediction model is constructed. This model is a 34-12-1 topological BP neural network model. The 34-dimensional input layer receives 34 types of process feature data, including: inlet thickness, temperature, and rolling force, bending roll force, and roll shifting amount for each stand of the finishing mill. The 12-dimensional hidden layer uses activation functions to extract nonlinear metallurgical features. The single-neuron output layer outputs the final predicted hot-rolled strip crown value. Through forward and backward propagation, the client updates its local weight vector. To minimize the prediction error of strip crown; S2.

2. Weight Quantization and Finite Field Mapping: For the client In the The weight update amount from the 12-dimensional hidden layer to the output layer obtained from local training. First, a 16-bit symmetric uniform quantization strategy is adopted, and the scaling factor is set. Map floating-point weights to symmetric integer intervals To retain 13 decimal places of precision, the quantization formula is: Perform an offset mapping to accommodate the function encryption algorithm's requirement for non-negative integer input, transferring signed integers... Convert to positive integer : At this point, the weighted components are mapped to the positive integer field. The client directly outputs the positive integers of these 12 dimensions. Construct the weight vector to be encrypted. : S2.

3. Function Encryption: Client Use public parameters Generated local private key and the session tag of the current fragment The weight vector constructed in step S2.2 Execute robust adaptive multi-client function encryption RAMCFE, with clients utilizing subsets public key set of all members Generate vector ciphertext : Simultaneously, the client integrates the All-or-No Encapsulation (AoNE) mechanism to generate partial functional decryption key shares, including a random mask. The decryption key share of the aforementioned functions Functional components and package components It consists of two parts. The calculation formula is as follows: in, For the client The private key; Aggregate function vector Corresponding client The weighted components; For the client A random mask matrix for sampling is used to achieve full packet loss tolerance; Function vector The hash mapping value; Using the All-or-No-Encapsulation Algorithm (AoNE), a private key is encapsulated. For random mask matrix Encrypt it and combine it with the current subset. and conversation tags Binding, the client will ciphertext Decryption key share of the secret part function Send to the aggregator.

4. The method according to claim 3, characterized in that, S3 specifically includes: S3.

1. Subset Integrity Verification: The aggregator receives the vector ciphertext from each client. and partial decryption key share Then, first check each subset. Whether the data is complete can only be determined if data packets from all clients within the subset have been delivered and the session tag is valid. Only when there is consistency is the subset marked as "complete"; otherwise, it is considered an "invalid subset" and discarded directly. S3.

2. Complete decryption key synthesis: For the verified complete subset The aggregator utilizes an all-or-nothing encapsulation mechanism to encapsulate the key shares uploaded by all clients within the subset. By combining the subsets, if the subset is complete, the aggregator will successfully remove the random mask and synthesize the complete functional decryption key corresponding to that subset. ; S3.

3. Vector Decompression and Global Model Calculation: The aggregator determines the function weight vector for this round based on the current federated learning aggregation strategy. The functional weight vector The contribution weights of each client's data in the aggregation process are defined, and the aggregator will... Set as an all-one vector corresponding to the size of the subset, so as to achieve subset pairing in the ciphertext field. The aggregator sums all client model updates using the synthesized key. RAMCFE decryption is performed on the set of ciphertext vectors within the subset. Based on the encryption properties of the inner product function, the decryption operation actually calculates the client weight vector set. With function vector Inner product: Decryption directly yields the exponential form of the subset aggregation vector: The aggregator then employs the small-step-big-step algorithm (BSGS) to solve for the discrete logarithm within the polynomial, from... Recover plaintext aggregate vector ,include: First, the boundaries of the search space are determined and a pre-computation table is constructed; Set the maximum range of values ​​for the aggregation result. Its calculation formula is ,in For the number of subset clients, The maximum quantization value for a single weight; then the step size parameter is calculated. To reduce the time complexity to And construct a "small step" hash table by calculating a series of bases. power value ,in Traversing the interval The generated key-value pairs Stored in a table for quick lookup; Secondly, perform a search match to recover the plaintext values; Calculate the "large step" baseline value and let the variable Incrementing from 0 Calculate the current search value After each calculation, check if the Val exists in the hash table. Once a match is found, it is considered to exist. Make This means that the equation has been found. The solution is found, at which point the search stops and the formula is applied. It directly recovers the algebraic sum of the encryption weights of all clients within the subset; Finally, the aggregator summarizes all valid subsets marked as "complete," denoted as Valid, performs dequantization and offset removal operations, and calculates the weight update amount of the global model. : The Then it is broadcast to all clients.

5. The method according to claim 4, characterized in that, S4 specifically includes: S4.

1. Model Update: Each client receives the weight update amount of the global model. And use it to replace the local model parameters; S4.

2. Termination condition judgment: If the maximum number of training rounds is reached. Or, the change in the global model parameters, i.e., the change in the weight update amount of the global model, is less than a threshold: The training process ends when this time is reached; S4.

3. Iteration or Termination: If not terminated, let If the condition is met, return to step S2; otherwise, output the output. This serves as the weight vector for the final trained local hot-rolled strip crown prediction model.

6. A system for predicting the crown of hot-rolled strip steel based on privacy protection and federated learning, characterized in that, The system includes: The key initialization and client dynamic partitioning module is used to generate system keys and perform partitioning for various distributed client capabilities, laying the foundation for subsequent encryption and aggregation. The local training and security encryption module is used by each distributed client to train a local hot-rolled strip crown prediction model using local rolling training data, and to quantize, compress, and encrypt the model weights. The secure aggregation module is used by the aggregator to securely aggregate encrypted weight data from multiple clients based on the robust adaptive multi-client function encryption algorithm RAMCFE and the all-or-nothing encapsulation mechanism AoNE, and calculate the weight update amount of the global model. The model update and termination check module is used by each distributed client to receive the weight update amount of the global model and update the local hot-rolled strip crown prediction model, and to perform iterative looping and termination judgment. The prediction output module is used by each distributed client to use the updated local hot-rolled strip crown prediction model to predict the crown of the hot-rolled strip to be predicted based on local rolling data and output the prediction results.

7. The system according to claim 6, characterized in that, The key initialization and client dynamic partitioning module is specifically used for: S1.

1. Common Parameter Generation: The aggregator or the first participating client performs the initialization process to generate bilinear group parameters. and hash function To obtain an all-or-nothing encapsulation setup and public parameters Distribute the training round counter to all potential clients via a secure channel and initialize the training round counter. ; S1.

2. Client Capability Assessment and Ranking: Each client Encrypted reporting of its real-time status information, including: CPU load rate. and network communication delay The aggregator uses a weighted comprehensive evaluation function to calculate the capability score for each client. To quantify the node processing capacity in heterogeneous industrial environments, nodes with sufficient computing resources and stable communication are selected. The calculation formula is as follows: in, Indicates the client Current CPU load rate, Represents the available computing power reserve of a node; Indicates the client Network communication latency with the aggregator; The maximum tolerable delay threshold is set and used to normalize communication metrics; To prevent tiny constants with a denominator of zero; and These are the weighting coefficients for computing power and communication power, respectively, and they satisfy... And accordingly to all Sort the active clients in descending order to obtain an ordered set. ; S1.

3. Dynamic Partitioning and Label Allocation: Set the size of the security target subset to... One client, will Divided into A subset, before Each subset contains 3 clients, and the last subset contains the remaining clients. Assign a unique session label .

8. The system according to claim 7, characterized in that, The local training and security encryption module is specifically used for: S2.

1. Local model training: per client Using its proprietary hot-rolling process dataset, a local hot-rolled strip crown prediction model is constructed. This model is a 34-12-1 topological BP neural network model. The 34-dimensional input layer receives 34 types of process feature data, including: inlet thickness, temperature, and rolling force, bending roll force, and roll shifting amount for each stand of the finishing mill. The 12-dimensional hidden layer uses activation functions to extract nonlinear metallurgical features. The single-neuron output layer outputs the final predicted hot-rolled strip crown value. Through forward and backward propagation, the client updates its local weight vector. To minimize the prediction error of strip crown; S2.

2. Weight Quantization and Finite Field Mapping: For the client In the The weight update amount from the 12-dimensional hidden layer to the output layer obtained from local training. First, a 16-bit symmetric uniform quantization strategy is adopted, and the scaling factor is set. Map floating-point weights to symmetric integer intervals To retain 13 decimal places of precision, the quantization formula is: Perform an offset mapping to accommodate the function encryption algorithm's requirement for non-negative integer input, transferring signed integers... Convert to positive integer : At this point, the weighted components are mapped to the positive integer field. The client directly outputs the positive integers of these 12 dimensions. Construct the weight vector to be encrypted. : S2.

3. Function Encryption: Client Use public parameters Generated local private key and the session tag of the current fragment The weight vector constructed in step S2.2 Execute robust adaptive multi-client function encryption RAMCFE, with clients utilizing subsets public key set of all members Generate vector ciphertext : Simultaneously, the client integrates the All-or-No Encapsulation (AoNE) mechanism to generate partial functional decryption key shares, including a random mask. The decryption key share of the aforementioned functions Functional components and package components It consists of two parts. The calculation formula is as follows: in, For the client The private key; Aggregate function vector Corresponding client The weighted components; For the client A random mask matrix for sampling is used to achieve full packet loss tolerance; Function vector The hash mapping value; Using the All-or-No-Encapsulation Algorithm (AoNE), a private key is encapsulated. For random mask matrix Encrypt it and combine it with the current subset. and conversation tags Binding, the client will ciphertext Decryption key share of the secret part function Send to the aggregator.

9. The system according to claim 8, characterized in that, The security aggregation module is specifically used for: S3.

1. Subset Integrity Verification: The aggregator receives the vector ciphertext from each client. and partial decryption key share Then, first check each subset. Whether the data is complete can only be determined if data packets from all clients within the subset have been delivered and the session tag is valid. Only when there is consistency is the subset marked as "complete"; otherwise, it is considered an "invalid subset" and discarded directly. S3.

2. Complete decryption key synthesis: For the verified complete subset The aggregator utilizes an all-or-nothing encapsulation mechanism to encapsulate the key shares uploaded by all clients within the subset. By combining the subsets, if the subset is complete, the aggregator will successfully remove the random mask and synthesize the complete functional decryption key corresponding to that subset. ; S3.

3. Vector Decompression and Global Model Calculation: The aggregator determines the function weight vector for this round based on the current federated learning aggregation strategy. The functional weight vector The contribution weights of each client's data in the aggregation process are defined, and the aggregator will... Set as an all-one vector corresponding to the size of the subset, so as to achieve subset pairing in the ciphertext field. The aggregator sums all client model updates using the synthesized key. RAMCFE decryption is performed on the set of ciphertext vectors within the subset. Based on the encryption properties of the inner product function, the decryption operation actually calculates the client weight vector set. With function vector Inner product: Decryption directly yields the exponential form of the subset aggregation vector: The aggregator then employs the small-step-big-step algorithm (BSGS) to solve for the discrete logarithm within the polynomial, from... Recover plaintext aggregate vector ,include: First, the boundaries of the search space are determined and a pre-computation table is constructed; Set the maximum range of values ​​for the aggregation result. Its calculation formula is ,in For the number of subset clients, The maximum quantization value for a single weight; then the step size parameter is calculated. To reduce the time complexity to And construct a "small step" hash table by calculating a series of bases. power value ,in Traversing the interval The generated key-value pairs Stored in a table for quick lookup; Secondly, perform a search match to recover the plaintext values; Calculate the "large step" baseline value and let the variable Incrementing from 0 Calculate the current search value After each calculation, check if the Val exists in the hash table. Once a match is found, it is considered to exist. Make This means that the equation has been found. The solution is found, at which point the search stops and the formula is applied. It directly recovers the algebraic sum of the encryption weights of all clients within the subset; Finally, the aggregator summarizes all valid subsets marked as "complete," denoted as Valid, performs dequantization and offset removal operations, and calculates the weight update amount of the global model. : The Then it is broadcast to all clients.

10. The system according to claim 9, characterized in that, The model update and termination check module is specifically used for: S4.

1. Model Update: Each client receives the weight update amount of the global model. And use it to replace the local model parameters; S4.

2. Termination condition judgment: If the maximum number of training rounds is reached. Or, the change in the global model parameters, i.e., the change in the weight update amount of the global model, is less than a threshold: The training process ends when this time is reached; S4.

3. Iteration or Termination: If not terminated, let If the condition is met, return to step S2; otherwise, output the output. This serves as the weight vector for the final trained local hot-rolled strip crown prediction model.

Images