A multi-party collaborative graph learning method and system based on material data privacy protection

Abstract

The application relates to a multi-party cooperation graph learning method and system based on material data privacy protection. In the learning method, multiple participants independently perform machine learning, and model parameters are uploaded to a server for joint training, so that material data is prevented from being leaked in the training process. Noise disturbance is performed on the model parameter transmission process, the leakage of the model parameters is avoided, and the graph node relationship matrix is set to participate in the training of the local graph learning model, the dependence relationship information of a material preparation process is fully considered, and the performance of the obtained model can be improved.

Description

Technical Field

[0001] This invention relates to the field of material property prediction technology, and in particular to a multi-party collaborative graph learning method and system based on material data privacy protection. Background Technology

[0002] With the rapid development of artificial intelligence, applying machine learning to the field of materials performance prediction, especially in guiding the preparation of new materials, can significantly shorten the entire preparation cycle. However, a single institution has a limited sample size of materials, and it cannot encompass all processes in the entire material production process, thus insufficient to train a high-performance model. Furthermore, the value of materials data, the substantial communication overhead, and the potential for privacy breaches during transmission render centralized learning by uploading local materials data to a server no longer suitable. Summary of the Invention

[0003] In view of this, the present invention provides a multi-party collaborative graph learning method and system based on material data privacy protection, so as to avoid the leakage of material data during the collaborative model training process of multiple participants.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A multi-party collaborative graph learning method based on material data privacy protection is disclosed. This method is applied to a distributed system comprising a server and multiple participants. Each participant is able to obtain partial preparation processes, process parameters of these processes, and performance parameters of the material after production. The learning method includes the following steps:

[0006] Each participant constructs a local graph node relationship matrix based on the known partial preparation process, and sends the index of the local graph node relationship matrix to the server; the local graph node relationship matrix is ​​used to characterize the dependency relationship between each preparation process in the production process.

[0007] The server determines the maximum index value based on the index of the local graph node relationship matrix of each participant, counts the number of participants, and sends the maximum index value and the number of participants to each participant;

[0008] Each participant aligns and encodes the local graph node relationship matrix according to the maximum index value and the number of participants to obtain the encoded local graph node relationship matrix, and then encrypts the encoded local graph node relationship matrix and sends it to the server.

[0009] The server aggregates the encoded local graph node relationship matrix sent by each participant to obtain a global graph node relationship matrix, and then sends the global graph node relationship matrix to each participant.

[0010] Each participant uses the global graph node relationship matrix, the known process parameters of some processes, and the performance parameters of the materials after production to train the local graph learning model, obtain local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation;

[0011] The server averages the local graph learning model parameters sent by each participant to obtain the global graph learning model parameters, and then sends the global graph learning model parameters to each participant.

[0012] Each participant sets the global graph learning parameters as the initial parameters of the local graph learning model, and returns to the step "each participant uses the global graph node relationship matrix and the known process parameters of some processes, as well as the performance parameters of the materials after production, to train the local graph learning model, obtain the local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation", until the iteration termination condition is met.

[0013] Optionally, each participant aligns and encodes the local graph node relationship matrix according to the maximum index value and the number of participants to obtain the encoded local graph node relationship matrix, specifically including:

[0014] The local graph node relationship matrix is ​​aligned based on its index and the maximum value of the index to obtain an aligned local graph node relationship matrix; the aligned local graph node relationship matrix is ​​n. max ×n max Matrix; where n max Indicates the maximum index value;

[0015] Based on the number of participants, the number of bits occupied by each element in the aligned local graph node relationship matrix is ​​determined using the formula m = log2(nq) + 1; where m represents the number of bits occupied by each element in the aligned local graph node relationship matrix, n represents the number of participants, and q represents the remainder parameter, q = n mod 2.

[0016] Based on the number of bits occupied by each element in the aligned local graph node relationship matrix, the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data is determined using the formula t = r / m; where t represents the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data, and r represents the number of bits of the data type of the operating system of the computer used by the participant.

[0017] Encode each t element in the aligned local graph node relationship matrix into an r-bit data, resulting in an encoded local graph node relationship matrix composed of r-bit data.

[0018] Optionally, the algorithm used to encrypt the encoded local graph node relationship matrix is ​​a homomorphic encryption algorithm.

[0019] Optionally, the specific steps for noise perturbation of the local graph learning model parameters are as follows:

[0020] The parameter data of each column of the local graph learning model parameters are sorted in ascending order of value to obtain the parameter sequence of each column;

[0021] Candidate sets are formed by selecting l feature cut points from the parameter data of each column based on the quantiles of each column; s k ={s k1 ,s k2 ,…,s kl}, where s k Let s represent the candidate set of the k-th column. k1 s k2 and s kl These are the 1st, 2nd, and 1st feature cut points of the parameter data in the kth column, respectively;

[0022] Construct quantile histograms for the candidate sets in each column;

[0023] Based on the quantile histogram of the candidate set for each column, the maximum statistical frequency region for each column is determined, and noise perturbation is added to the parameter data within the maximum statistical frequency region for each column.

[0024] Optionally, quantiles can be calculated as follows: M represents the quantile of the k-th column. kmax and M kmin represents the maximum and minimum values ​​of the parameter data in column k, and l represents the number of candidate sets for partitioning.

[0025] A multi-party collaborative graph learning system based on material data privacy protection is proposed. This system is applied to a distributed system comprising a server and multiple participants. Each participant can obtain partial preparation processes, process parameters of these processes, and performance parameters of the material after production. The learning system includes:

[0026] This map node relationship matrix construction module is used to enable each participant to construct a local map node relationship matrix based on the known partial manufacturing processes, and send the index of the local map node relationship matrix to the server; the local map node relationship matrix is ​​used to characterize the dependency relationship between each manufacturing process in the production process.

[0027] The index maximum value and parameter number determination module is used to determine the index maximum value based on the index of the local graph node relationship matrix of each participant through the server, count the number of participants, and send the index maximum value and the number of participants to each participant;

[0028] The alignment and encoding module is used to align and encode the local graph node relationship matrix by each participant according to the maximum index value and the number of participants, to obtain the encoded local graph node relationship matrix, and to encrypt the encoded local graph node relationship matrix and send it to the server.

[0029] The relationship matrix aggregation module is used to aggregate the encoded local graph node relationship matrix sent by each participant through the server to obtain a global graph node relationship matrix, and then send the global graph node relationship matrix to each participant.

[0030] The local training module is used to train the local graph learning model by each participant using the global graph node relationship matrix and the process parameters of some processes and the performance parameters of the materials after production, to obtain the local graph learning model parameters, and then send the local graph learning model parameters to the server after noise perturbation.

[0031] The average calculation module is used to calculate the average of the local graph learning model parameters sent by each participant through the server, obtain the global graph learning model parameters, and send the global graph learning model parameters to each participant.

[0032] The return module is used to set the global graph learning parameters as the initial parameters of the local graph learning model by each participant, and return to the step "each participant uses the global graph node relationship matrix and the known process parameters of some processes and the performance parameters of the materials after production to train the local graph learning model, obtain the local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation", until the iteration termination condition is met.

[0033] Optionally, the alignment and encoding module specifically includes:

[0034] The alignment submodule is used to align the local graph node relationship matrix according to the index of the local graph node relationship matrix and the maximum value of the index, to obtain an aligned local graph node relationship matrix; the aligned local graph node relationship matrix is ​​n. max ×n max Matrix; where n max Indicates the maximum index value;

[0035] The first calculation submodule is used to determine the number of bits occupied by each element in the aligned local graph node relationship matrix based on the number of participants, using the formula m = log2(nq) + 1; where m represents the number of bits occupied by each element in the aligned local graph node relationship matrix, n represents the number of participants, and q represents the remainder parameter, q = n mod 2.

[0036] The second calculation submodule is used to determine the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data, based on the number of bits occupied by each element in the aligned local graph node relationship matrix, using the formula t = r / m; where t represents the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data, and r represents the number of bits of the data type of the operating system of the computer used by the participant.

[0037] The encoding submodule is used to encode each t element in the aligned local graph node relationship matrix into an r-bit data, resulting in an encoded local graph node relationship matrix composed of r-bit data.

[0038] Optionally, the algorithm used to encrypt the encoded local graph node relationship matrix is ​​a homomorphic encryption algorithm.

[0039] Optionally, the specific steps for noise perturbation of the local graph learning model parameters are as follows:

[0040] The parameter data of each column of the local graph learning model parameters are sorted in ascending order of value to obtain the parameter sequence of each column;

[0041] Candidate sets are formed by selecting l feature cut points from the parameter data of each column based on the quantiles of each column; s k ={s k1 ,s k2 ,…,s kl}, where s k Let s represent the candidate set of the k-th column. k1 s k2 and s kl These are the 1st, 2nd, and 1st feature cut points of the parameter data in the kth column, respectively;

[0042] Construct quantile histograms for the candidate sets in each column;

[0043] Based on the quantile histogram of the candidate set for each column, the maximum statistical frequency region for each column is determined, and noise perturbation is added to the parameter data within the maximum statistical frequency region for each column.

[0044] Optionally, quantiles can be calculated as follows: M represents the quantile of the k-th column. kmax and M kmin represents the maximum and minimum values ​​of the parameter data in column k, and l represents the number of candidate sets for partitioning.

[0045] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0046] This invention discloses a multi-party collaborative graph learning method and system based on material data privacy protection. The learning method includes the following steps: Each participant constructs a local graph node relationship matrix based on the learned partial preparation process, and sends the index of the local graph node relationship matrix to the server; the server determines the maximum index value based on the index of each participant's local graph node relationship matrix, counts the number of participants, and sends the maximum index value and the number of participants to each participant; each participant aligns and encodes the local graph node relationship matrix according to the maximum index value and the number of participants to obtain an encoded local graph node relationship matrix, and encrypts the encoded local graph node relationship matrix before sending it to the server; the server processes the data sent by each participant... The encoded local graph node relationship matrix is ​​aggregated to obtain a global graph node relationship matrix, which is then sent to each participant. Each participant uses the global graph node relationship matrix, known process parameters of some processes, and performance parameters of the materials after production to train the local graph learning model, obtaining local graph learning model parameters. These parameters are then noise-perturbed and sent to the server. The server averages the local graph learning model parameters sent by each participant to obtain global graph learning model parameters, which are then sent to each participant. Each participant sets the global graph learning parameters as the initial parameters for their local graph learning model and continues training. This invention allows multiple participants to independently perform machine learning, uploading model parameters to the server for joint processing. This avoids leakage of material data during training. The invention also incorporates noise perturbation during parameter transmission to prevent leakage. Furthermore, the inclusion of a graph node relationship matrix in the local graph learning model training fully considers the dependencies in material preparation processes, thus improving the performance of the trained model. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be 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.

[0048] Figure 1 This is a flowchart illustrating a multi-party collaborative graph learning method based on material data privacy protection, provided as an embodiment of the present invention. Detailed Implementation

[0049] 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.

[0050] The purpose of this invention is to provide a multi-party collaborative graph learning method and system based on material data privacy protection, so as to avoid the leakage of material data during the collaborative model training process of multiple participants.

[0051] 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.

[0052] Multiple participants independently perform machine learning, and uploading model parameters to a server for joint training is an effective way to solve the problem of material parameter leakage. Furthermore, current machine learning methods rarely consider the dependencies between material data. Incorporating the dependencies (graph node adjacency relationships) and material property features into the graph learning model can significantly improve model performance. In joint learning based on model parameter aggregation, although the data of each material organization is protected, model parameters can still cause privacy leaks in the local graph model. Therefore, this invention designs an efficient and privacy-preserving multi-party collaborative graph learning method for material data. This method not only ensures the secure transmission of graph model parameters and material data dependency information but also reduces communication overhead and improves the performance of joint learning. Based on this, this invention provides a multi-party collaborative graph learning method and system based on material data privacy protection. It protects the material preparation process data, where each preparation process is a graph node. Different material organizations are responsible for different preparation processes, and the directed connections between preparation processes are node adjacency relationships. Different preparation processes include different process parameters. The process steps with the greatest impact on material preparation are identified based on the prediction results of the graph neural network. The details are as follows:

[0053] Example 1

[0054] This invention provides a multi-party collaborative graph learning method based on material data privacy protection. The implementation of embodiment 1 is based on the following example:

[0055] Material organization 1 sends the preparation process steps (ingot casting, heating, hot rolling) index {1,2,3} to the server; Material organization 2 sends the preparation process steps (coiling, milling, initial rolling, annealing) index {4,5,6,7} to the server; Material organization 3 sends the preparation process steps (cooling, annealing, final rolling, finished product) index {8,7,9,10} to the server. Different preparation process steps include different process parameters. For example, ingot casting includes process parameters such as temperature and rotation speed; annealing includes process parameters such as annealing temperature, aging temperature, and aging time.

[0056] like Figure 1 As shown, the learning method includes the following steps:

[0057] Step 101: Each participant constructs a local graph node relationship matrix based on the known partial preparation process, and sends the index of the local graph node relationship matrix to the server; the local graph node relationship matrix is ​​used to characterize the dependency relationship between each preparation process in the production process.

[0058] Step 102: The server determines the maximum index value based on the index of each participant's local graph node relationship matrix, counts the number of participants, and sends the maximum index value and the number of participants to each participant. According to the above example, the minimum index value is ind. min =1 and the maximum index ind max =10.

[0059] Step 103: Each participant aligns and encodes the local graph node relationship matrix according to the maximum index value and the number of participants to obtain the encoded local graph node relationship matrix, and then encrypts the encoded local graph node relationship matrix and sends it to the server.

[0060] Step 103: Each participant aligns and encodes the local graph node relationship matrix according to the maximum index value and the number of participants to obtain the encoded local graph node relationship matrix, specifically including:

[0061] The local graph node relationship matrix is ​​aligned based on its index and the maximum value of the index to obtain an aligned local graph node relationship matrix; the aligned local graph node relationship matrix is ​​n. max ×n max Matrix; where n max This indicates the maximum index value.

[0062] Based on the above example, the local graph node relationship matrix is ​​completed, retaining the node relationships of the current participant and representing non-existent nodes in the current participant with 0. The aligned local graph node relationship matrix corresponding to Material Mechanism 1 is as follows:

[0063]

[0064] The aligned local graph node relationship matrix corresponding to Material Mechanism 2 is as follows:

[0065]

[0066] The aligned local graph node relationship matrix corresponding to Material Mechanism 2 is as follows:

[0067]

[0068] Based on the number of participants, the number of bits occupied by each element in the aligned local graph node relationship matrix is ​​determined using the formula m = log2(nq) + 1; where m represents the number of bits occupied by each element in the aligned local graph node relationship matrix, n represents the number of participants, and q represents the remainder parameter, q = n mod 2; according to the above example, each participant receives the number of participants n = 3 from the server, and the number of bits occupied by each element in the aligned local graph node relationship matrix m = 2.

[0069] Based on the number of bits occupied by each element in the aligned local graph node relationship matrix, the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data is determined using the formula t = r / m. Here, t represents the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data, and r represents the number of bits of the data type of the operating system of the computer used by the participant. According to the above example, based on the data type in the current computer operating system, the number of bits occupied by the selected encoded data type is r = 32. Inputting the number of bits r and the number of bits m occupied by each element into the formula t = r / m, the number of elements in the aligned local graph node relationship matrix that can be encoded by an r-bit data type is t = 16.

[0070] Encode each t element in the aligned local graph node relationship matrix into an r-bit data, resulting in an encoded local graph node relationship matrix composed of r-bit data.

[0071] Encode the t elements of the aligned local graph node relation matrix into new data d. Then, each row of the aligned local graph node relation matrix is ​​processed... The encoding process yields a local graph node relation matrix of size (p, b). Based on the above example, for the aligned local graph node relation matrix A∈R 10×10 Encoding is performed to obtain the encoded local graph node relationship matrix.

[0072] For example, converting the aligned local graph node relationship matrix corresponding to material mechanism 1 into a binary bit matrix is... Then, converted to a 32-bit decimal matrix (i.e., the encoded local graph node relationship matrix) is:

[0073] Convert the aligned local map node relationship matrix corresponding to Material Mechanism 2 into a binary bit matrix. Then, converted to a 32-bit decimal matrix, it is:

[0074] Convert the aligned local map node relationship matrix corresponding to material mechanism 3 into a binary bit matrix. Then, converted to a 32-bit decimal matrix, it is:

[0075] The encoded local graph node relationship matrix The public key is then input into a homomorphic encryption algorithm (such as the Paillier algorithm) for encryption, resulting in...

[0076] Step 104: The server aggregates the encoded local graph node relationship matrix sent by each participant to obtain the global graph node relationship matrix, and sends it to each participant. That is, it... Send to the server for aggregation, and get The aggregated encAggedratedA is sent to all participants.

[0077] Step 105: Each participant uses the global graph node relationship matrix, the known process parameters of some processes, and the performance parameters of the materials after production to train the local graph learning model, obtain local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation; each participant inputs encAggedratedA and their private key into a homomorphic encryption algorithm for decryption, obtaining the global graph node relationship matrix aggedratedA = Dec(encAggedratedA, sk). The training step is then executed.

[0078] During the training process, for each participant, the global graph node relationship matrix and some process parameters are used as inputs to the local graph learning model, and the performance parameters of the material after production are used as outputs to train the local graph learning model parameters until the difference between the performance parameters output by the local graph learning model and the actual performance parameters is less than a certain threshold.

[0079] The trained local graph learning model can be used to predict material properties and adjust process parameters. For process parameter adjustment, an enumeration method can be used to determine several process parameters. These determined process parameters and the global graph node relationship matrix are then input into the trained local graph learning model for performance parameter prediction. Process parameters with superior performance are selected to guide the fabrication process. Alternatively, particle swarm optimization, genetic algorithms, etc., can be combined to optimize performance parameters based on different process parameters. The enumeration method is suitable for scenarios where key fabrication processes and key process parameters are known.

[0080] Each participant adds noise perturbation to some parameters of the graph learning model and sends the perturbated parameters to the server for secure aggregation. The server then sends the aggregated global graph learning model parameters back to each participant, and the learning process continues iteratively until the iteration termination condition is met. The specific steps are as follows:

[0081] Participants from all parties, based on quantiles Find the local model parameter M k The candidate set s of l feature cut points of the column data distribution k ={s k1 ,s k2 ,…,s kl},in M kmax M is the maximum value of the model parameters in the k-th column. kmin It is the minimum value of the model parameter in the k-th column;

[0082] Each participant, based on the candidate set s of each column k Construct a quantile histogram fGrad = {fGrad} k1 fGrad k2 ,…fGrad kl};

[0083] The participants compared the quantile histograms of each column to obtain the maximum statistical frequency region of each column, maxfGrad = max(fGrad);

[0084] Each participant adds noise perturbation to the region of maximum statistical frequency in each column to obtain the perturbed local graph learning model parameters, and then sends the perturbed local graph learning model parameters to the server.

[0085] Step 106: The server averages the local graph learning model parameters sent by each participant to obtain the global graph learning model parameters, and then sends the global graph learning model parameters to each participant.

[0086] Step 107: Each participant sets the global graph learning parameters as the initial parameters of their local graph learning model, returning to step "Each participant uses the global graph node relationship matrix, the known process parameters of some processes, and the performance parameters of the materials after production to train the local graph learning model, obtain local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation," until the iteration termination condition is met. The server sends the aggregated global graph learning model parameters to each participant to continue learning and iterating until the iteration termination condition is met. The iteration termination condition is typically model convergence or the number of iterations reaching a certain threshold.

[0087] Example 2

[0088] Embodiment 2 of the present invention provides a multi-party collaborative graph learning system based on material data privacy protection. The learning system is applied to a distributed system, which includes a server and multiple participants. Each participant can obtain partial preparation processes, process parameters of partial preparation processes, and performance parameters of the material after production. The learning system includes:

[0089] This map node relationship matrix construction module is used to enable each participant to construct a local map node relationship matrix based on the known partial manufacturing processes, and send the index of the local map node relationship matrix to the server; the local map node relationship matrix is ​​used to characterize the dependency relationship between each manufacturing process in the production process.

[0090] The index maximum value and parameter number determination module is used to determine the index maximum value based on the index of the local graph node relationship matrix of each participant through the server, count the number of participants, and send the index maximum value and the number of participants to each participant;

[0091] An alignment and encoding module is used to align and encode the local graph node relationship matrix by having each participant align and encode it according to the maximum index value and the number of participants, thereby obtaining an encoded local graph node relationship matrix. The encoded local graph node relationship matrix is ​​then encrypted and sent to the server. Specifically, the alignment and encoding module includes an alignment submodule, used to align the local graph node relationship matrix according to the index and the maximum index value, thereby obtaining an aligned local graph node relationship matrix. The aligned local graph node relationship matrix is ​​n. max ×n max Matrix; where n maxThe system consists of three submodules: a first calculation submodule, a second calculation submodule, and an encoding submodule. The first submodule encodes each t element in the aligned local graph node relationship matrix using the formula m = log2(nq) + 1, where m represents the number of bits per element, n represents the number of participants, and q represents the remainder parameter, q = n mod 2. The second submodule encodes each t element in the aligned local graph node relationship matrix using the formula t = r / m, where t represents the number of elements that can be represented by the same r-bit data, and r represents the number of bits in the data type of the computer operating system used by the participants. The encoding submodule encodes each t element in the aligned local graph node relationship matrix into an r-bit data unit, resulting in an encoded local graph node relationship matrix composed of r-bit data. The encryption algorithm for the encoded local graph node relationship matrix is ​​a homomorphic encryption algorithm.

[0092] The relationship matrix aggregation module is used to aggregate the encoded local graph node relationship matrix sent by each participant through the server to obtain a global graph node relationship matrix, and then send the global graph node relationship matrix to each participant.

[0093] The local training module is used to train the local graph learning model using the global graph node relationship matrix, the known process parameters of some processes, and the performance parameters of the materials after production, respectively, by each participant. This trains the local graph learning model parameters, which are then noise-perturbed before being sent to the server. The specific steps for noise perturbing the local graph learning model parameters are as follows: The parameter data of each column of the local graph learning model parameters are sorted in ascending order of value to obtain the parameter sequence of each column; A candidate set consisting of l feature cut points of the parameter data in each column is formed based on the quantile of each column. k ={s k1 ,s k2 ,…,s kl}, where s k Let s represent the candidate set of the k-th column. k1 s k2 and s kl Let the first, second, and l feature cut points of the parameter data in column k be respectively; construct the quantile histogram of the candidate set for each column; based on the quantile histogram of the candidate set for each column, determine the maximum statistical frequency region of each column, and add noise perturbation to the parameter data in the maximum statistical frequency region of each column.

[0094] The quantiles are calculated as follows: M represents the quantile of the k-th column. kmax and M kmin represents the maximum and minimum values ​​of the parameter data in column k, and l represents the number of candidate sets for partitioning.

[0095] The average calculation module is used to calculate the average of the local graph learning model parameters sent by each participant through the server, obtain the global graph learning model parameters, and send the global graph learning model parameters to each participant.

[0096] The return module is used to set the global graph learning parameters as the initial parameters of the local graph learning model by each participant, and return to the step "each participant uses the global graph node relationship matrix and the known process parameters of some processes and the performance parameters of the materials after production to train the local graph learning model, obtain the local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation", until the iteration termination condition is met.

[0097] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0098] This invention discloses a multi-party collaborative graph learning method and system based on material data privacy protection. The learning method includes the following steps: Each participant constructs a local graph node relationship matrix based on the learned partial preparation process, and sends the index of the local graph node relationship matrix to the server; the server determines the maximum index value based on the index of each participant's local graph node relationship matrix, counts the number of participants, and sends the maximum index value and the number of participants to each participant; each participant aligns and encodes the local graph node relationship matrix according to the maximum index value and the number of participants to obtain an encoded local graph node relationship matrix, and encrypts the encoded local graph node relationship matrix before sending it to the server; the server processes the data sent by each participant... The encoded local graph node relationship matrix is ​​aggregated to obtain a global graph node relationship matrix, which is then sent to each participant. Each participant uses the global graph node relationship matrix, known process parameters of some processes, and performance parameters of the materials after production to train the local graph learning model, obtaining local graph learning model parameters. These parameters are then noise-perturbed and sent to the server. The server averages the local graph learning model parameters sent by each participant to obtain global graph learning model parameters, which are then sent to each participant. Each participant sets the global graph learning parameters as the initial parameters for their local graph learning model and continues training. This invention allows multiple participants to independently perform machine learning, uploading model parameters to the server for joint processing. This avoids leakage of material data during training. The invention also incorporates noise perturbation during parameter transmission to prevent leakage. Furthermore, the inclusion of a graph node relationship matrix in the local graph learning model training fully considers the dependencies in material preparation processes, thus improving the performance of the trained model.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0100] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-party collaborative graph learning method based on material data privacy protection, characterized in that, The learning method is applied to a distributed system, which includes a server and multiple participants. Each participant can learn about some of the preparation processes in the material production process, the process parameters of some of the preparation processes, and the performance parameters of the material after production. The learning method includes the following steps: Each participant constructs a local graph node relationship matrix based on the learned partial preparation process, and sends the index of the local graph node relationship matrix to the server; The local graph node relationship matrix is ​​used to characterize the dependency relationship between each preparation process in the production process of the partial preparation process; The server determines the maximum index value based on the index of the local graph node relationship matrix of each participant, counts the number of participants, and sends the maximum index value and the number of participants to each participant; Each participant aligns and encodes the local graph node relationship matrix according to the maximum index value and the number of participants to obtain the encoded local graph node relationship matrix, and then encrypts the encoded local graph node relationship matrix and sends it to the server. The server aggregates the encoded local graph node relationship matrix sent by each participant to obtain a global graph node relationship matrix, and then sends the global graph node relationship matrix to each participant. Each participant uses the global graph node relationship matrix, the known process parameters of some processes, and the performance parameters of the materials after production to train the local graph learning model, obtain local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation; The server averages the local graph learning model parameters sent by each participant to obtain the global graph learning model parameters, and then sends the global graph learning model parameters to each participant. Each participant sets the global graph learning parameters as the initial parameters of the local graph learning model, and returns to the step "Each participant uses the global graph node relationship matrix and the known process parameters of some processes, as well as the performance parameters of the materials after production, to train the local graph learning model, obtain the local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation", until the iteration termination condition is met; Each participant aligns and encodes the local graph node relationship matrix according to the maximum index value and the number of participants, obtaining the encoded local graph node relationship matrix, specifically including: The local graph node relationship matrix is ​​aligned based on its index and the maximum value of the index to obtain an aligned local graph node relationship matrix; the aligned local graph node relationship matrix is ​​n. max ×n max Matrix; where n max Indicates the maximum index value; Based on the number of participants, the number of bits occupied by each element in the aligned local graph node relationship matrix is ​​determined using the formula m = log2(nq) + 1; where m represents the number of bits occupied by each element in the aligned local graph node relationship matrix, n represents the number of participants, and q represents the remainder parameter, q = n mod 2. Based on the number of bits occupied by each element in the aligned local graph node relationship matrix, the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data is determined using the formula t = r / m; where t represents the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data, and r represents the number of bits of the data type of the operating system of the computer used by the participant. Encode each t element in the aligned local graph node relationship matrix into an r-bit data, resulting in an encoded local graph node relationship matrix composed of r-bit data.

2. The multi-party collaborative graph learning method based on material data privacy protection according to claim 1, characterized in that, The algorithm used to encrypt the encoded local graph node relationship matrix is ​​a homomorphic encryption algorithm.

3. The multi-party collaborative graph learning method based on material data privacy protection according to claim 1, characterized in that, The specific steps for noise perturbation of the parameters of the local graph learning model are as follows: The parameter data of each column of the local graph learning model parameters are sorted in ascending order of value to obtain the parameter sequence of each column; Candidate sets are formed by selecting l feature cut points from the parameter data of each column based on the quantiles of each column; s k ={s k1 ,s k2 ,…,s kl }, where s k Let s represent the candidate set of the k-th column. k1 s k2 and s kl These are the 1st, 2nd, and 1st feature cut points of the parameter data in the kth column, respectively; Construct quantile histograms for the candidate sets in each column; Based on the quantile histogram of the candidate set for each column, the maximum statistical frequency region for each column is determined, and noise perturbation is added to the parameter data within the maximum statistical frequency region for each column.

4. The multi-party collaborative graph learning method based on material data privacy protection according to claim 3, characterized in that, The method for calculating quantiles is as follows: M represents the quantile of the k-th column. kmax and M kmin represents the maximum and minimum values ​​of the parameter data in column k, and l represents the number of candidate sets for partitioning.

5. A multi-party collaborative graph learning system based on material data privacy protection, characterized in that, The learning system is applied to a distributed system, which includes a server and multiple participants. Each participant can learn about some of the preparation processes in the material production process, the process parameters of some of the preparation processes, and the performance parameters of the material after production. The learning system includes: This map node relationship matrix construction module is used to enable each participant to construct a local map node relationship matrix based on the known partial manufacturing processes, and send the index of the local map node relationship matrix to the server; the local map node relationship matrix is ​​used to characterize the dependency relationship between each manufacturing process in the production process. The index maximum value and parameter number determination module is used to determine the index maximum value based on the index of the local graph node relationship matrix of each participant through the server, count the number of participants, and send the index maximum value and the number of participants to each participant; The alignment and encoding module is used to align and encode the local graph node relationship matrix by each participant according to the maximum index value and the number of participants, to obtain the encoded local graph node relationship matrix, and to encrypt the encoded local graph node relationship matrix and send it to the server. The relationship matrix aggregation module is used to aggregate the encoded local graph node relationship matrix sent by each participant through the server to obtain a global graph node relationship matrix, and then send the global graph node relationship matrix to each participant. The local training module is used to train the local graph learning model by each participant using the global graph node relationship matrix and the process parameters of some processes and the performance parameters of the materials after production, to obtain the local graph learning model parameters, and then send the local graph learning model parameters to the server after noise perturbation. The average calculation module is used to calculate the average of the local graph learning model parameters sent by each participant through the server, obtain the global graph learning model parameters, and send the global graph learning model parameters to each participant. The return module is used to set the global graph learning parameters as the initial parameters of the local graph learning model by each participant, and return to the step "each participant uses the global graph node relationship matrix and the known process parameters of some processes, as well as the performance parameters of the materials after production, to train the local graph learning model, obtain the local graph learning model parameters, and send the local graph learning model parameters to the server after noise perturbation", until the iteration end condition is met; The alignment and encoding module specifically includes: The alignment submodule is used to align the local graph node relationship matrix according to the index of the local graph node relationship matrix and the maximum value of the index, to obtain an aligned local graph node relationship matrix; the aligned local graph node relationship matrix is ​​n. max ×n max Matrix; where n max Indicates the maximum index value; The first calculation submodule is used to determine the number of bits occupied by each element in the aligned local graph node relationship matrix based on the number of participants, using the formula m = log2(nq) + 1; where m represents the number of bits occupied by each element in the aligned local graph node relationship matrix, n represents the number of participants, and q represents the remainder parameter, q = n mod 2. The second calculation submodule is used to determine the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data, based on the number of bits occupied by each element in the aligned local graph node relationship matrix, using the formula t = r / m; where t represents the number of elements in the aligned local graph node relationship matrix that can be represented by the same r-bit data, and r represents the number of bits of the data type of the operating system of the computer used by the participant. The encoding submodule is used to encode each t element in the aligned local graph node relationship matrix into an r-bit data, resulting in an encoded local graph node relationship matrix composed of r-bit data.

6. The multi-party collaborative graph learning system based on material data privacy protection according to claim 5, characterized in that, The algorithm used to encrypt the encoded local graph node relationship matrix is ​​a homomorphic encryption algorithm.

7. The multi-party collaborative graph learning system based on material data privacy protection according to claim 5, characterized in that, The specific steps for noise perturbation of the parameters of the local graph learning model are as follows: The parameter data of each column of the local graph learning model parameters are sorted in ascending order of value to obtain the parameter sequence of each column; Candidate sets are formed by selecting l feature cut points from the parameter data of each column based on the quantiles of each column; s k ={s k1 ,s k2 ,…,s kl }, where s k Let s represent the candidate set of the k-th column. k1 s k2 and s kl These are the 1st, 2nd, and 1st feature cut points of the parameter data in the kth column, respectively; Construct quantile histograms for the candidate sets in each column; Based on the quantile histogram of the candidate set for each column, the maximum statistical frequency region for each column is determined, and noise perturbation is added to the parameter data within the maximum statistical frequency region for each column.

8. The multi-party collaborative graph learning system based on material data privacy protection according to claim 7, characterized in that, The method for calculating quantiles is as follows: M represents the quantile of the k-th column. kmax and M kmin represents the maximum and minimum values ​​of the parameter data in column k, and l represents the number of candidate sets for partitioning.

Images