Multi-party joint training method of performance prediction model based on material data privacy protection

Abstract

The application discloses a multi-party joint training method of a performance prediction model based on material data privacy protection. In the joint training process, a plurality of participants transmits the parameters of a local graph learning model to other participants after noise disturbance, so that the privacy and safety of the local material data are ensured, and the parameters of the local graph learning model are disturbed by noise before transmission, so that reasoning attacks by an enemy are avoided. In addition, the application obtains a global graph node relationship matrix through secure calculation, so that the local graph node relationship matrix of each participant is not leaked, and the performance of the local graph learning model can be improved. The application avoids leakage of material data and model parameters in the joint training process of the plurality of participants.

Description

Technical Field

[0001] This invention relates to the field of materials performance prediction technology, and in particular to a multi-party joint training method for performance prediction models based on materials data privacy protection. Background Technology

[0002] Currently, in the field of materials property prediction, using machine learning to analyze materials data and guide the preparation of new materials can significantly shorten the entire preparation cycle. However, due to the insufficient sample size of a single institution and the fact that the entire materials production process cannot be completed in a single institution, model training in a single institution is insufficient to obtain a high-performance model. Furthermore, given the significant value of materials data, uploading it to a server for centralized training incurs substantial communication overhead and increases the risk of privacy breaches. In practical applications, the server setup makes the entire multi-party collaborative training process more vulnerable to malicious attacks. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-party joint training method for performance prediction models based on material data privacy protection, so as to achieve multi-party collaborative training while avoiding the leakage of material data and thus achieving material data privacy protection.

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

[0005] A multi-party joint training method for a performance prediction model based on materials data privacy protection is disclosed. This method is applied to a decentralized distributed system comprising multiple participants, each of whom can only access local process parameters. The training method includes the following steps:

[0006] Each participant will share the step index of its local process with other participants in the decentralized distributed system;

[0007] Each participant generates a local graph node relationship matrix based on all step indexes and the step indexes of local process steps;

[0008] Each participating party jointly calculates the local graph node relationship matrix to obtain the global graph node relationship matrix;

[0009] Each participant trains the local graph learning model based on the global graph node relationship matrix, local process parameters, and material performance parameters, and then shares the parameters of the trained local graph learning model with other participants in the decentralized distributed system after noise perturbation.

[0010] Each participant aggregates and averages the parameters of their local graph learning model to obtain the parameters of the global graph learning model, and sets the parameters of the global graph learning model as the initial parameters of the local graph learning model.

[0011] The process returns to the step "Each participant trains the local graph learning model based on the local graph node relationship matrix, local process parameters, and material performance parameters, and shares the parameters of the trained local graph learning model with other participants in the decentralized distributed system", until the iteration termination condition is met.

[0012] Optionally, each participant generates a local graph node relationship matrix based on all step indices and the step index of the local process step, specifically including:

[0013] Each participant generates an initial local graph node relationship matrix based on the step index of its local process steps;

[0014] Each participant determines the maximum index value based on all step indexes;

[0015] Each participant aligns the initial local graph node relationship matrix according to the maximum index value to generate an aligned local graph node relationship matrix.

[0016] Optionally, each participant may jointly calculate the local graph node relationship matrix to obtain the global graph node relationship matrix, which specifically includes:

[0017] Each participant randomly generates a numerical matrix with values ​​in the interval (0, 1]. Among them, X i Let m be the numerical matrix of the i-th participant, i = 1, 2, ..., N, where N is the number of participants. max The maximum value at index. This represents the size of the local graph node relationship matrix;

[0018] Active participants generate the first perturbation matrix Second perturbation matrix The third perturbation matrix is ​​D = BΘC; where Θ is the dot product between matrix elements, and the active participant is the initiator of multi-party joint training.

[0019] The active participants generate their first sub-secret B based on the number of participants in the decentralized distributed system and the first, second, and third perturbation matrices. i =SHR(B), Second Sub-Secret C i =SHR(C) and the third sub-secret D i =SHR(D), and send the first, second, and third subsecrets of each passive participant to each passive participant; where B i C i and D iThese are the first sub-secret, second sub-secret, and third sub-secret of the i-th participant, respectively. The passive participant is a participant in the decentralized distributed system other than the active participant.

[0020] Each participant calculates its third sub-secret E based on its first, second, and third sub-secrets and the local graph node relationship matrix. i =A i -B i And the fourth son's secret F i =X i -C i Among them, E i and F i For the third and fourth sub-secrets of the i-th participant; A i Let be the local graph node relationship matrix of the i-th participant;

[0021] Each passive participant sends its third and fourth sub-secrets to the active participant;

[0022] The active participants calculate the first master secret based on the third and fourth sub-secrets of each participant. Second Secret And send the first master secret and the second master secret to each passive participant;

[0023] Each participating party calculates its own total secret matrix based on the first total secret and the second total secret, and shares the total secret matrix with other participating parties; wherein, the total secret matrix of the active participating party r is calculated. The formula is: Calculate the total secret matrix of passive participant j The formula is:

[0024] Each participant calculates the global relation matrix based on the total secret matrix shared by all participants.

[0025] Each participant calculates the global graph node relationship matrix based on the aforementioned global relationship matrix. Among them, finalAdj xy This represents the element in the x-th row and y-th column of the global graph node relationship matrix, a. xy This represents the element in the x-th row and y-th column of the global relation matrix.

[0026] Optionally, the parameters of the trained local graph learning model can be perturbed by noise, specifically including:

[0027] Each participant extracts important parameters from the parameters of the local graph learning model; the important parameters are those other than the important parameters that have a greater impact on the output of the local graph learning model than the parameters of the local graph learning model method.

[0028] Each of the aforementioned participants used noise to perturb key parameters.

[0029] Optionally, each participant extracts important parameters from the parameters of the local graph learning model, specifically including:

[0030] The influence matrix of the sample set on the neurons of the local graph learning model is calculated using the influence degree calculation method.

[0031] Based on the influence matrix of the sample set on the neurons of the local graph learning model, the influence matrix of the weight parameters between the neurons of the local graph learning model is calculated as follows:

[0032] weightImpactRate =

[0033] {weightImpactRate 1 weightImpactRate 2 ..., weightImpactRate K-1};

[0034]

[0035] Where weightImpactRate is the influence matrix of the weight parameters between neurons in the local graph learning model, and K represents the number of hidden layers in the local graph learning model; weightImpactRate 1 This is the influence matrix of each neuron in the first hidden layer of the local graph learning model on each neuron in the second hidden layer, where weightImpactRate is the weight. 2 This is the influence matrix of each neuron in the second hidden layer of the local graph learning model on each neuron in the third hidden layer, where weightImpactRate is the weight. K-1 This is the influence matrix of each neuron in the (K-1)th hidden layer of the local graph learning model on each neuron in the Kth hidden layer; Let be the influence degree of the 0th neuron in the kth hidden layer on the pth neuron in the (k+1)th hidden layer, and let represent the influence matrix of each neuron in the kth hidden layer on each neuron in the (k+1)th hidden layer: weightImpactRate. k The element in row o and column p of the array;

[0036] Based on the influence matrix of the weight parameters between the neurons in the local graph learning model, the weight parameters of each neuron in the local graph learning model are sorted to obtain the first weight parameter sequence.

[0037] Based on the magnitude of the weight parameters of each neuron in the local graph learning model, the weight parameters of each neuron in the local graph learning model are sorted to obtain the second weight parameter sequence.

[0038] Obtain the first preset number of weight parameters from the first weight parameter sequence to form a first weight parameter set; obtain the first preset number of weight parameters from the second weight parameter sequence to form a second weight parameter set.

[0039] Calculate the intersection of the first weight parameter and the second weight parameter, and use it as the set of important parameters.

[0040] Optionally, the method for calculating the influence degree specifically includes:

[0041] The influence matrix of each sample in the sample set on the neurons of the local graph learning model is calculated as follows:

[0042] neuronImpact m ={output(x m )>α}={O m，1 O m，2 , ...O m，K};

[0043] Among them, neuronImpact m Let m be the influence matrix of the m-th sample on the neurons of the local graph learning model, and output(x) m ) represents the m-th sample x m The output of each neuron in the local graph learning model after inputting the input, where α is a preset threshold, and O m，k Let $m$ be the influence vector of the $m$-th sample on the $k$-th hidden layer of the local graph learning model, where $k = 1, 2, ..., K$, and $K$ represents the number of hidden layers in the local graph learning model. m，k ={neu m，k，1 , neu m，k，2 , ...neu m，k，N}, neu m，k，n Let S be the influence of the m-th sample on the n-th neuron of the k-th hidden layer in the local graph learning model, where n = 1, 2, ..., S, and S represents the number of neurons in the k-th hidden layer of the local graph learning model. h m，k，n For the m-th sample x m The output value of the nth neuron in the kth hidden layer of the corresponding local graph learning model;

[0044] Based on the influence matrix of each sample on the neurons of the local graph learning model, the influence matrix of the sample set on the neurons of the local graph learning model is calculated as follows:

[0045]

[0046] Where M represents the number of samples in the sample set.

[0047] Optionally, based on the influence matrix of the sample set on the weight parameters of each neuron in the local graph learning model, the weight parameters of each neuron in the local graph learning model are sorted, and the formula for obtaining the first weight parameter sequence is as follows:

[0048] weightImportIndex=sort(weightImpactRate);

[0049] Here, weightImportIndex is the first weight parameter sequence, and sort() is the sorting function.

[0050] Optionally, based on the magnitude of the weight parameters of each neuron in the local graph learning model, the weight parameters of each neuron are sorted to obtain the second weight parameter sequence. The formula is as follows:

[0051] magIndex = sort(weight);

[0052] Where magIndex is the second weight parameter sequence, weight is the weight parameter magnitude matrix composed of the magnitudes of the weight parameters of each neuron in the local graph learning model, and sort() is the sorting function.

[0053] A multi-party joint training device for a performance prediction model based on materials data privacy protection, wherein the training device is applied to the above-mentioned training method, comprising:

[0054] The training device includes multiple computers deployed on each participant's computer. Each computer is equipped with a step-by-step index sharing module, a local graph node relationship matrix generation module, a global graph node relationship matrix calculation module, a training module, an aggregation module, and a return module.

[0055] The step index sharing module among participants is used to share the step index of local process steps with other participants in the decentralized distributed system;

[0056] The local graph node relationship matrix generation module for each participant is used to generate a local graph node relationship matrix based on all step indices and the step index of the local process steps;

[0057] The global graph node relationship matrix calculation module for each participant is used to jointly calculate the local graph node relationship matrix to obtain the global graph node relationship matrix.

[0058] Each participant's training module is used to train the local graph learning model based on the global graph node relationship matrix, local process parameters, and material performance parameters, and to share the parameters of the trained local graph learning model with other participants in the decentralized distributed system after noise perturbation.

[0059] The aggregation module of each participant is used to aggregate and average the parameters of the local graph learning model sent by each participant to obtain the parameters of the global graph learning model, and set the parameters of the global graph learning model as the initial parameters of the local graph learning model.

[0060] Each participant's return module is used to return the training module until the iteration termination condition is met.

[0061] An electronic device includes a memory disposed in each participant, a processor disposed in each participant, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the training method.

[0062] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0063] This invention discloses a multi-party joint training method for performance prediction models based on material data privacy protection. During joint training, the parameters of the local graph learning model are perturbed with noise before being transmitted to other participants, ensuring the privacy and security of the local material data. Furthermore, perturbing the parameters of the local graph learning model with noise before transmission prevents adversary inference attacks. In addition, this invention securely computes the global graph node relationship matrix, ensuring that the relationship matrices of each participating local graph node are not leaked and improving the performance of the local graph learning model. This invention avoids the leakage of material data and model parameters during multi-party joint training. Attached Figure Description

[0064] 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.

[0065] Figure 1 The flowchart illustrates a multi-party joint training method for a performance prediction model based on material data privacy protection, as provided in this embodiment of the invention. Detailed Implementation

[0066] 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.

[0067] The purpose of this invention is to provide a multi-party joint training method for performance prediction models based on material data privacy protection, so as to achieve multi-party collaborative training while avoiding the leakage of material data and thus achieving material data privacy protection.

[0068] 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.

[0069] like Figure 1 As shown, this embodiment of the invention provides a multi-party joint training method for a performance prediction model based on material data privacy protection. The training method is applied to a decentralized distributed system, which includes multiple participants, each of whom can only obtain local process parameters. The training method includes the following steps:

[0070] Step 101: Each participant shares the step index of its local process steps with other participants in the decentralized distributed system.

[0071] Step 102: Each participant generates a local graph node relationship matrix based on all step indices and the step index of the local process step.

[0072] Step 103: Each participant jointly calculates the local graph node relationship matrix to obtain the global graph node relationship matrix;

[0073] Step 104: Each participant trains the local graph learning model based on the global graph node relationship matrix, local process parameters, and material performance parameters, and then shares the parameters of the trained local graph learning model with other participants in the decentralized distributed system after noise perturbation.

[0074] Step 105: Each participant aggregates the parameters of their local graph learning model and averages them to obtain the parameters of the global graph learning model, and sets the parameters of the global graph learning model as the initial parameters of the local graph learning model.

[0075] Step 106, return to step "Each participant trains the local graph learning model based on the local graph node relationship matrix, local process parameters and material performance parameters, and shares the parameters of the trained local graph learning model with other participants in the decentralized distributed system", until the iteration termination condition is met.

[0076] The training method of the present invention will be described below based on the following specific embodiments. Each participating party includes material mechanism 0, material mechanism 1, and material mechanism 2. Based on this specific embodiment, the specific implementation steps of the training method of the present invention are as follows.

[0077] Step 101 specifically includes:

[0078] Material mechanism 1 sends the step indexes {1,2,3} of the preparation process steps (ingot casting, heating, hot rolling) to material mechanism 2 and material structure 3. Material mechanism 2 sends the step indexes {4,5,6,7} of the preparation process steps (coiling, milling, initial rolling, annealing) to material mechanism 1 and material structure 3. Material mechanism 3 sends the step indexes {8,7,9,10} of the preparation process steps (cooling, annealing, final rolling, finished product) to material structure 1 and material structure 2. 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.

[0079] Step 102 specifically includes:

[0080] Each participant constructs dependencies between preparation processes based on its local preparation process and sends the maximum and minimum indices of the preparation processes to other neighboring participants. In the graph learning method, graph nodes represent preparation process steps, and a graph node relationship matrix characterizes the spatial dependencies between material preparation processes.

[0081] Each participant obtains the maximum and minimum values ​​of the graph nodes based on the accepted graph node indices, aligns its local node adjacency matrix according to the maximum index, and securely jointly computes the global graph node relationships. Based on the above example, the minimum index is... min =1 and the maximum index max =10.

[0082] The local graph node relationship matrix is ​​aligned based on the index of the local graph node relationship moments and the maximum value of the index to obtain the aligned local graph node relationship matrix; the aligned local graph node relationship matrix is ​​m. mmax ×m mmax Matrix; where m max Indicates the maximum value of the index. max It is 10.

[0083] Based on the above example, the local graph node relationship matrix is ​​completed, retaining the node relationships of the current participants and representing nodes that do not exist in the current participants with 0. The aligned local graph node relationship matrix corresponding to Material Organization 1 is as follows:

[0084]

[0085] The aligned local graph node relationship matrix corresponding to Material Mechanism 2 is as follows:

[0086]

[0087] The aligned local graph node relationship matrix corresponding to Material Mechanism 3 is as follows:

[0088]

[0089] Step 103 specifically includes:

[0090] Each participant randomly generates a numerical matrix with values ​​in the interval (0, 1]. Among them, X i Let m be the numerical matrix of the i-th participant, i = 1, 2, ..., N, where N is the number of participants. max The maximum value at index. This represents the size of the local graph node relationship matrix;

[0091] Active participants generate the first perturbation matrix Second perturbation matrix The third perturbation matrix is ​​D = BΘC; where Θ is the dot product between matrix elements, and the active participant is the initiator of multi-party joint training.

[0092] The active participants generate their first sub-secret B based on the number of participants in the decentralized distributed system and the first, second, and third perturbation matrices. i =SHR(B), Second Sub-Secret C i =SHR(C) and the third sub-secret D i =SHR(D), and send the first, second, and third subsecrets of each passive participant to each passive participant; where B i C i and D i These are the first sub-secret, second sub-secret, and third sub-secret of the i-th participant, respectively. The passive participant is a participant in the decentralized distributed system other than the active participant.

[0093] Each participant calculates its third sub-secret E based on its first, second, and third sub-secrets and the local graph node relationship matrix. i =A i -B i And the fourth son's secret F i =X i -C i Among them, E i and F i For the third and fourth sub-secrets of the i-th participant; A i Let be the local graph node relationship matrix of the i-th participant;

[0094] Each passive participant sends its third and fourth sub-secrets to the active participant;

[0095] The active participants calculate the first master secret based on the third and fourth sub-secrets of each participant. Second Secret And send the first master secret and the second master secret to each passive participant;

[0096] Each participating party calculates its own total secret matrix based on the first total secret and the second total secret, and shares the total secret matrix with other participating parties; wherein, the total secret matrix of the active participating party r is calculated. The formula is: Calculate the total secret matrix of passive participant j The formula is:

[0097] Each participant calculates the global relation matrix based on the total secret matrix shared by all participants.

[0098] Each participant calculates the global graph node relationship matrix based on the aforementioned global relationship matrix. Among them, finalAdj xy This represents the element in the x-th row and y-th column of the global graph node relationship matrix, a. xy This represents the element in the x-th row and y-th column of the global relation matrix.

[0099] in,

[0100] Each of the aforementioned participants views the aligned local graph node relationship matrix as a sub-secret. And randomly generate a matrix of non-zero values ​​between (0, 1]. The numerical matrix of material mechanism 1 is as follows:

[0101]

[0102] The numerical matrix of material mechanism 2 is as follows:

[0103]

[0104] The numerical matrix of material mechanism 3 is as follows:

[0105]

[0106] Participant P r Treating other participants as active participants, and considering them as active participants. i≠0 Viewed as a passive participant, participant P r Generates a perturbation matrix with values ​​of [0, 1]. D = BΘC, where Θ is the dot product of matrix elements.

[0107]

[0108]

[0109]

[0110] The active participant Pr generates the first sub-secret based on the number of participants N. Second son's secret The secret of the third child SHR is a method of secret sharing. And B i≠r C i≠r D i≠r Send to other participants P i≠r The secret sharing among the participants, specifically B, is as follows:

[0111]

[0112]

[0113]

[0114] C's secret sharing is:

[0115]

[0116]

[0117]

[0118] D's secret sharing is:

[0119]

[0120]

[0121]

[0122] Each participant P i Calculate E i =A i -B i F i =X i -C i Among them, the passive participant will E i and F i Send to the active participants. Each participant's E i and F i They are respectively:

[0123]

[0124]

[0125]

[0126]

[0127]

[0128]

[0129] Active participant P r calculate Then send E and F to the other passive participants.

[0130]

[0131]

[0132] Active participant P r calculate and will Send to other passive participants P i≠r For example, material mechanism 1 is taken as the active participant, wherein...

[0133]

[0134] Passive participant P i≠r calculate And send it to other participants.

[0135]

[0136]

[0137] Each participating party calculates

[0138]

[0139] Each participating party according to

[0140] Step 104 specifically includes:

[0141] Each participant inputs the global graph node relationships, local process parameters, and initial global model parameters into the local graph learning model for model training, thereby obtaining the local graph learning model parameters for this iteration. After perturbing the local graph model parameters with noise, the parameters are sent to other neighboring participants.

[0142] During the training process, each participant inputs the global adjacency matrix and the process parameters of the preparation process into the local graph model, outputs the performance parameters of the material, and continuously iterates the training until the difference between the output performance and the actual performance is less than the set threshold.

[0143] The trained local graph model can be used for testing material properties and for adjusting process parameters. For adjusting process parameters, an enumeration method can be used to determine some process parameters first. Then, the global adjacency matrix and some process parameters are input into the local graph model for performance prediction. This process is repeated until the optimal process parameters are found to achieve the desired material preparation. Alternatively, particle swarm optimization or genetic algorithms can be used to optimize different process parameters.

[0144] The parameters of the local graph learning model are subjected to noise perturbation. The specific steps include:

[0145] Each participant extracts the location of important model parameters in the local graph learning model parameters based on the influence of the input on the neurons;

[0146] Each of the aforementioned participants uses noise to perturb the corresponding model parameters based on the location of the important model parameters;

[0147] Each participant extracts the location of important model parameters in the local graph learning model parameters based on the influence of the input on the neurons. The specific steps include:

[0148] The influence matrix of each sample in the sample set on the neurons of the local graph learning model is calculated as follows:

[0149] neuronImpact m ={output(x m )>α}={O m，1 O m，2 , ...O m，K};

[0150] Among them, neuronImpact m Let m be the influence matrix of the m-th sample on the neurons of the local graph learning model, and output(x) m ) represents the m-th sample x m The output of each neuron in the local graph learning model after inputting the input, where α is a preset threshold, and O m，k Let $m$ be the influence vector of the $m$-th sample on the $k$-th hidden layer of the local graph learning model, where $k = 1, 2, ..., K$, and $K$ represents the number of hidden layers in the local graph learning model. m，k ={neu m，k，1 , neu m，k，2 , ...neu m，k，N}, neu m，k，n Let S be the influence of the m-th sample on the n-th neuron of the k-th hidden layer in the local graph learning model, where n = 1, 2, ..., S, and S represents the number of neurons in the k-th hidden layer of the local graph learning model. h m，k，n For the m-th sample x m The output value of the nth neuron in the kth hidden layer of the corresponding local graph learning model;

[0151] Based on the influence matrix of each sample on the neurons of the local graph learning model, the influence matrix of the sample set on the neurons of the local graph learning model is calculated as follows:

[0152]

[0153] Where M represents the number of samples in the sample set.

[0154] Based on the influence matrix of the sample set on the neurons of the local graph learning model, the influence matrix of the weight parameters between the neurons of the local graph learning model is calculated as follows:

[0155] weightImpactRate =

[0156] {weightImpactRate 1 weightImpactRate 2 ..., weightImpactRate K-1};

[0157]

[0158] Where weightImpactRate is the influence matrix of the weight parameters between neurons in the local graph learning model, and K represents the number of hidden layers in the local graph learning model; weightImpactRate1 This is the influence matrix of each neuron in the first hidden layer of the local graph learning model on each neuron in the second hidden layer, where weightImpactRate is the weight. 2 This is the influence matrix of each neuron in the second hidden layer of the local graph learning model on each neuron in the third hidden layer, where weightImpactRate is the weight. K-1 This is the influence matrix of each neuron in the (K-1)th hidden layer of the local graph learning model on each neuron in the Kth hidden layer; Let be the influence degree of the 0th neuron in the kth hidden layer on the pth neuron in the (k+1)th hidden layer, and let represent the influence matrix of each neuron in the kth hidden layer on each neuron in the (k+1)th hidden layer: weightImpactRate. k The element in row o and column p.

[0159] Each participating party ranks the influence of the sample set on the weight parameters of each neuron in the local graph learning model according to the formula weightImportIndex = sort(weightImpactRate);

[0160] Each participating party sorts the magnitude of the weight parameter to obtain magIndex = sort(weight);

[0161] Each participating party simultaneously considers the impact of the input sample on the weight and the weight magnitude to obtain the important weight parameter ImportIndex = weightImportIndex∩magIndex.

[0162] Example 2

[0163] Embodiment 2 of the present invention provides a multi-party joint training device for a performance prediction model based on material data privacy protection. The training device is applied to the above-mentioned training method and includes:

[0164] The training device includes multiple computers deployed on each participant's computer. Each computer is equipped with a step-by-step index sharing module, a local graph node relationship matrix generation module, a global graph node relationship matrix calculation module, a training module, an aggregation module, and a return module.

[0165] The step index sharing module among participants is used to share the step index of local process steps with other participants in the decentralized distributed system;

[0166] The local graph node relationship matrix generation module for each participant is used to generate a local graph node relationship matrix based on all step indices and the step index of the local process steps;

[0167] The global graph node relationship matrix calculation module for each participant is used to jointly calculate the local graph node relationship matrix to obtain the global graph node relationship matrix.

[0168] Each participant's training module is used to train the local graph learning model based on the global graph node relationship matrix, local process parameters, and material performance parameters, and to share the parameters of the trained local graph learning model with other participants in the decentralized distributed system after noise perturbation.

[0169] The aggregation module of each participant is used to aggregate and average the parameters of the local graph learning model sent by each participant to obtain the parameters of the global graph learning model, and set the parameters of the global graph learning model as the initial parameters of the local graph learning model.

[0170] Each participant's return module is used to return the training module until the iteration termination condition is met.

[0171] Example 3

[0172] Embodiment 3 of the present invention provides an electronic device, including a memory disposed in each participating party, a processor disposed in each participating party, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the training method of Embodiment 1.

[0173] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0174] This invention discloses a multi-party joint training method for performance prediction models based on material data privacy protection. During joint training, the parameters of the local graph learning model are perturbed with noise before being transmitted to other participants, ensuring the privacy and security of the local material data. Furthermore, perturbing the parameters of the local graph learning model with noise before transmission prevents adversary inference attacks. In addition, this invention securely computes the global graph node relationship matrix, ensuring that the relationship matrices of each participating local graph node are not leaked and improving the performance of the local graph learning model. This invention avoids the leakage of material data and model parameters during multi-party joint training.

[0175] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0176] 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 joint training method for performance prediction models based on materials data privacy protection, characterized in that, The training method is applied to a decentralized distributed system, which includes multiple participants, each of which can only obtain local process parameters; the training method includes the following steps: Each participant will share the step index of its local process with other participants in the decentralized distributed system; Each participant generates a local graph node relationship matrix based on all step indexes and the step indexes of local process steps; Each participating party jointly calculates the local graph node relationship matrix to obtain the global graph node relationship matrix; Each participant trains the local graph learning model based on the global graph node relationship matrix, local process parameters, and material performance parameters, and then shares the parameters of the trained local graph learning model with other participants in the decentralized distributed system after noise perturbation. Each participant aggregates and averages the parameters of their local graph learning model to obtain the parameters of the global graph learning model, and sets the parameters of the global graph learning model as the initial parameters of the local graph learning model. Return to step "Each participant trains the local graph learning model based on the local graph node relationship matrix, local process parameters, and material performance parameters, and shares the parameters of the trained local graph learning model with other participants in the decentralized distributed system", until the iteration termination condition is met; The parameters of the trained local graph learning model are subjected to noise perturbation, specifically including: Each participant extracts important parameters from the parameters of the local graph learning model; the important parameters are those other than the important parameters that have a greater impact on the output of the local graph learning model than the parameters of the local graph learning model method. Each of the aforementioned participants used noise to perturb key parameters; The key parameters extracted by each participant from the local graph learning model include: The influence matrix of the sample set on the neurons of the local graph learning model is calculated using the influence degree calculation method.

2. The multi-party joint training method for performance prediction models based on material data privacy protection according to claim 1, characterized in that, Each participating party generates a local graph node relationship matrix based on all step indices and the step index of the local process step, specifically including: Each participant generates an initial local graph node relationship matrix based on the step index of its local process steps; Each participant determines the maximum index value based on all step indexes; Each participant aligns the initial local graph node relationship matrix according to the maximum index value to generate an aligned local graph node relationship matrix.

3. The multi-party joint training method for performance prediction models based on material data privacy protection according to claim 1, characterized in that, Each participating party jointly calculates the local graph node relationship matrix to obtain the global graph node relationship matrix, which specifically includes: Each participant randomly generates a numerical matrix with values ​​in the interval (0,1]. ;in, For the first i Numerical matrix of each participant i= 1, 2, ..., N , N The number of participants. The maximum value at index. This represents the size of the local graph node relationship matrix; Active participants generate the first perturbation matrix The second perturbation matrix and the third perturbation matrix ;in, It is the dot product between matrix elements, and the active participant is the initiator of multi-party joint training; The active participants generate their first sub-secret based on the number of participants in the decentralized distributed system and the first, second, and third perturbation matrices. The Secret of the Second Son And the secret of the third son And send the first, second, and third sub-secrets of each passive participant to each passive participant; among which, , and The first i The passive participants are those other than the active participants in the decentralized distributed system, and the passive participants are those other than the active participants. Each participant calculates its third sub-secret based on its first, second, and third sub-secrets and the local graph node relationship matrix. And the secret of the fourth son ;in, and Let the third and fourth sub-secrets of the i-th participant be used; For the first i Local graph node relationship matrix of each participant; Each passive participant sends its third and fourth sub-secrets to the active participant; The active participants calculate the first master secret based on the third and fourth sub-secrets of each participant. Second Secret The first and second general secrets are then sent to each passive participant. Each participating party calculates its own total secret matrix based on the first total secret and the second total secret, and shares the total secret matrix with other participating parties; wherein, the calculation of the active participating party... r Total Secret Matrix The formula is: ,in, , and Given the first, second, and third subsecrets of the active participant r, calculate the subsecrets of the passive participant r. j Total Secret Matrix The formula is: , j= 1, 2, ..., N , ; Each participant calculates the global relation matrix based on the total secret matrix shared by all participants. ; Each participant calculates the global graph node relationship matrix based on the global relationship matrix. ;in, Represents the first node in the global graph node relationship matrix. x Line 1 y Column elements, Represents the first in the global relation matrix x Line 1 y The elements of the column.

4. The multi-party joint training method for performance prediction models based on material data privacy protection according to claim 1, characterized in that, The important parameters extracted by each participant from the local graph learning model also include: Based on the influence matrix of the sample set on the neurons of the local graph learning model, the influence matrix of the weight parameters between the neurons of the local graph learning model is calculated as follows: ； ； in, This is the influence matrix of the weight parameters between neurons in the local graph learning model. This indicates the number of hidden layers in the local graph learning model; This is the influence matrix of each neuron in the first hidden layer of the local graph learning model on each neuron in the second hidden layer. This is the influence matrix of each neuron in the second hidden layer of the local graph learning model on each neuron in the third hidden layer. For the local graph learning model K -1 The neurons in each hidden layer are paired with the first K The influence matrix of each neuron in each hidden layer; For the first k The o-th neuron in the n-th hidden layer pairs with the n-th... k +1 hidden layer p The influence degree of the k-th neuron represents the influence of each neuron in the k-th hidden layer on the first hidden layer. k Influence matrix of each neuron in +1 hidden layer The element in the o-th row and p-th column of the array; and These represent the influence of the sample set on the o-th neuron of the k-th hidden layer and the p-th neuron of the (k+1)-th hidden layer of the local graph learning model, respectively. Based on the influence matrix of the weight parameters between the neurons in the local graph learning model, the weight parameters of each neuron in the local graph learning model are sorted to obtain the first weight parameter sequence. Based on the magnitude of the weight parameters of each neuron in the local graph learning model, the weight parameters of each neuron in the local graph learning model are sorted to obtain the second weight parameter sequence. Obtain the first preset number of weight parameters from the first weight parameter sequence to form a first weight parameter set; obtain the first preset number of weight parameters from the second weight parameter sequence to form a second weight parameter set. Calculate the intersection of the first weight parameter and the second weight parameter, and use it as the set of important parameters.

5. The multi-party joint training method for performance prediction models based on material data privacy protection according to claim 4, characterized in that, The method for calculating the influence specifically includes: The influence matrix of each sample in the sample set on the neurons of the local graph learning model is calculated as follows: ； in, For the first m The influence matrix of each sample on the neurons of the local graph learning model Represents the m-th sample The output of each neuron in the local graph learning model after inputting the input. For the preset threshold, For the first m The nth sample is the first local graph learning model k The influence vector of each hidden layer k= 1, 2, ..., K , K This indicates the number of hidden layers in the local graph learning model. , For the first m The nth sample is the first local graph learning model k The first hidden layer n The degree of influence of each neuron n= 1, 2, ..., S , S In the local graph learning model, the first... k The number of neurons in each hidden layer , For the first m Sample The corresponding local graph learning model's first In the hidden layer, the th The output value of each neuron; Based on the influence matrix of each sample on the neurons of the local graph learning model, the influence matrix of the sample set on the neurons of the local graph learning model is calculated as follows: ； in, M This indicates the number of samples in the sample set.

6. The multi-party joint training method for performance prediction models based on material data privacy protection according to claim 4, characterized in that, Based on the influence matrix of the sample set on the weight parameters of each neuron in the local graph learning model, the weight parameters of each neuron in the local graph learning model are sorted, and the formula for obtaining the first weight parameter sequence is as follows: ； in, This is the first weight parameter sequence. ( ) is the sorting function.

7. The multi-party joint training method for performance prediction models based on material data privacy protection according to claim 4, characterized in that, Based on the magnitude of the weight parameters of each neuron in the local graph learning model, the formula for sorting the weight parameters of each neuron to obtain the second weight parameter sequence is as follows: ； in, For the second weight parameter sequence, This is a weight parameter magnitude matrix composed of the magnitudes of the weight parameters of each neuron in the local graph learning model. ( ) is the sorting function.

8. A multi-party joint training device for performance prediction models based on material data privacy protection, characterized in that, The training device is applied to the training method according to any one of claims 1-7, comprising: The training device includes multiple computers deployed on each participant's computer. Each computer is equipped with a step index sharing module, a local graph node relationship matrix generation module, a global graph node relationship matrix calculation module, a training module, an aggregation module, and a return module. The step index sharing module among participants is used to share the step index of local process steps with other participants in the decentralized distributed system; The local graph node relationship matrix generation module for each participant is used to generate a local graph node relationship matrix based on all step indices and the step index of the local process steps; The global graph node relationship matrix calculation module for each participant is used to jointly calculate the local graph node relationship matrix to obtain the global graph node relationship matrix. The training modules of each participant are used to train the local graph learning model based on the global graph node relationship matrix, local process parameters and material performance parameters, and to share the parameters of the trained local graph learning model with other participants in the decentralized distributed system after noise perturbation. The aggregation module of each participant is used to aggregate and average the parameters of the local graph learning model sent by each participant to obtain the parameters of the global graph learning model, and set the parameters of the global graph learning model as the initial parameters of the local graph learning model. Each participant's return module is used to return the training module until the iteration termination condition is met.

9. An electronic device, characterized in that, It includes a memory deployed in each participant, a processor deployed in each participant, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the training method as described in any one of claims 1 to 7.

Images