Federal learning method based on training quality evaluation and on-chain reward payment

Abstract

The invention relates to a federated learning method based on training quality evaluation and on-chain reward payment, and belongs to the technical field of federated learning. According to the method, a block chain technology is utilized, a central server and a client are regarded as nodes on a block chain, data transmission between devices is realized through a block chain network, and payment of training quality is realized through an intelligent contract. Through a zero-knowledge certification technology, the certification of the training quality by the central server is realized on the premise of protecting user data privacy.

Description

Technical Field

[0001] The present invention relates to a federated learning method based on training quality evaluation and on-chain remuneration payment, belonging to the technical field of federated learning. Background Art

[0002] In recent years, with the development of technologies such as big data, the Internet of Things, and deep learning, machine learning has developed rapidly and made significant progress. Due to the adaptability, efficiency, diversity, and predictiveness of machine learning, machine learning technology has been widely used in finance, healthcare, natural language processing, computer vision, big data analysis, and other fields, greatly facilitating people's lives and work.

[0003] Effective machine learning models often require a large amount of high-quality data support. However, in most cases, the data available to trainers is very limited. At the same time, due to constraints in industry competition, legal supervision, and user privacy and security, the phenomenon of "data islands" has been created. Federated Learning technology can effectively solve the "data island" phenomenon while protecting user data privacy.

[0004] Federated learning technology is a distributed machine learning method that allows multiple participants (such as devices, organizations, or nodes) to jointly train a model without sharing the original data. Each participant retains the data locally and trains the model on its own data, and then sends the model update (not the data itself) to the central server. The central server aggregates the model updates of all participants, generates a global model, and distributes it back to each participant for the next round of training. This process is repeated many times until the model converges. Therefore, federated learning technology can achieve data sharing and model training while protecting user privacy and data security.

[0005] However, in the process of federated learning, the quality of data trained locally by each client varies greatly, which requires the central server to verify and pay for the training quality of each client. To solve this problem, it is necessary to find a federated learning privacy protection solution that can prove the training quality of the client and pay for it. Summary of the invention

[0006] The purpose of the present invention is to solve the technical problem of how to evaluate the quality of client training data and pay remuneration during the process of federated learning, and creatively propose a federated learning method based on training quality evaluation and on-chain remuneration payment.

[0007] The method of the present invention uses blockchain technology to regard the central server and the client as nodes on the blockchain. The data transmission between devices is realized through the blockchain network, and the payment for the training quality is realized through smart contracts. Through zero-knowledge proof technology, the central server can prove the training quality under the premise of protecting the privacy of user data.

[0008] The present invention is implemented by adopting the following technical solutions.

[0009] A federated learning method based on training quality evaluation and on-chain reward payment. The participating devices include a central server S, clients C1, C2, ..., C n , a trusted parameter generator A. The central server S is responsible for distributing models and keys for proof, as well as deploying related smart contracts. Clients C1, C2, …, C n Responsible for model training and generating proof of training quality. Trusted parameter generator A is responsible for distributing the parameters required for secure aggregation.

[0010] Step 1: Device initialization.

[0011] First, the central server S generates a verification key (VK) and a proving key (PK) based on the training task. Afterwards, S distributes PK to each client for generating proof of the training process, and VK is used by S to verify the correctness of the proof later. S deploys smart contracts on the blockchain network for subsequent automatic verification, payment, and aggregation.

[0012] The central server S deposits money into the smart contract in advance, and then automatically verifies the proofs submitted by each client through the smart contract. The smart contract fits the losses submitted by each client, and then automatically pays and aggregates according to the fitting results. The trusted parameter generator A generates random numbers r1, r2, ...r for all clients. n , used for subsequent security aggregation,

[0013] Step 2: Model training.

[0014] Federated learning sets up several rounds of training. For the kth round of training, the central server S will aggregate the global model m after the k-1th round of training. k Distribute to each client C1, C2, ..., C n This step ensures that all participating devices start training from the same starting point, thus ensuring the effectiveness of subsequent training.

[0015] Forward propagation, loss calculation and back propagation are the core processes of model training. n Receive the global model mk Then, using the dataset Train the model. By a number of forms such as (X j ,y j ) sample composition, X j is the characteristic vector of the sample, y j is the label corresponding to the sample. For each sample (X j ,y j ), the forward propagation will be X j The process of passing the input data through the neural network layer by layer to calculate the output result. The result of forward propagation is the network's calculation of the input data X j The predicted value of . Output result is a probability distribution vector, which indicates the probability of the sample belonging to each category. After obtaining the predicted value, the cross entropy loss function is used to calculate the predicted value and the true value y j The loss between As follows:

[0016]

[0017] Among them, K is the number of categories.

[0018] After that, back propagation is performed. The gradient is calculated according to the loss function and the network weights and biases are adjusted to minimize the loss. The gradient update δ of the sample is recorded during the back propagation process, as follows:

[0019]

[0020] The above are the gradient update formulas for the weight and bias of the lth layer respectively. Among them, Z (l) is the error term at level l. represents partial derivative, L represents loss function, W (l) represents the weight matrix of the lth layer, b (l) represents the bias of the lth layer. (X (l-1) ) T It is the transpose of the activation value of the previous layer, and T represents transpose.

[0021] Each client records the loss set of all samples during training and gradient update δ.

[0022] Step 3: Proof of training quality.

[0023] For client C i , after the kth round of training, a loss set is generated and the gradient update δ i Client C iGenerate r distributed by the central A through parameters i , perform mask generation on the gradient update δ i Client C i Submit and the loss set to the smart contract in its entirety.

[0024] To verify the training quality of each client without accessing the training process and training data, the client generates a concise non-interactive zero-knowledge proof for the training process (which can be implemented using zk-SNARKs). The client non-interactively generates random numbers (which can be achieved through the Fiat-Shamir transformation), randomly selects t samples from all sample sets to obtain Utilize the calculation process of the samples in to generate the proof proof.

[0025] Subsequently, the client submits the generated proof to the central server S for verification. If the verification passes, the central server S believes that the client has not fabricated. For example, among 10,000 proofs containing 1,000 forged proofs, randomly selecting 50, the probability of drawing a forged proof and failing the verification is 99.49%. Therefore, the verifier has a high confidence level to confirm the contribution of the client.

[0026] Step 4: Training quality payment.

[0027] The smart contract first verifies the proofs proof submitted by each client. Then, it models all the sample loss data submitted by the clients that pass the verification and selects a normal distribution for fitting. After that, according to the fitted loss distribution function, two threshold values t1 and t2 are obtained, and t1 < t2, which are used to judge the range of the sample loss l, thereby determining the category of the sample. If l < t1, it is considered a valid sample; if l > t2, it is considered an invalid sample.

[0028] The smart contract makes the above judgments on the losses uploaded by each client. For client C i , if there are invalid samples in the samples, that is then it is considered that the training quality of this client is unqualified, and the central server S will not use the gradient update of this client for aggregation, nor will it make a payment to this client. If there are no invalid samples in the small part of the samples drawn, it is considered that the training quality of this client is qualified, and the smart contract realizes automatic payment to this client. At the same time, the more samples in in the loss set, the more remuneration will be paid to this client.

[0029] ​Afterwards, the model parameters submitted by clients with qualified training quality are averaged and aggregated through the smart contract according to the number of valid samples, and the aggregated results are distributed to each client for the next round of training.

[0030] Beneficial Effects

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] The method of the present invention combines zero-knowledge proof and federated learning technology, heuristically selects samples by calculating cross-entropy loss, realizes the evaluation of training quality and verification of the training process, and automatically pays rewards through blockchain. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the model of the method of the present invention.

[0034] Figure 2 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0035] The method of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0036] like Figure 1 The figure shows a typical federated learning application scenario. The central server and each participating device are regarded as nodes on the blockchain, and the distribution and upload of model parameters are realized through the blockchain network. The central server has a global model, and each client has training data of different quality. The central server distributes the global model to the client, and the client uses its own local data for training. Then, the loss generated in the forward propagation is used to generate a proof and uploaded to the central server for verification, thereby realizing the proof of training quality. Finally, the central server pays the client and performs security aggregation based on the training quality.

[0037] like Figure 2 As shown in FIG. 1 , a federated learning method based on training quality evaluation and on-chain reward payment includes the following steps:

[0038] Step 1: Device initialization.

[0039] First, the central server S generates a verification key VK and a proof key PK.

[0040] Afterwards, S distributes PK to each participant to generate relevant proofs, and VK is used by S to verify the correctness of the proofs. S deploys the smart contracts required by the blockchain network for subsequent automatic verification, payment, and aggregation.

[0041] S constructs a circuit representing the training process, including forward propagation, back propagation, and loss calculation. The above circuit is converted into R1CS form (Rank-1 Constraint System), expressed as a set of polynomial forms. Then, the Lagrange interpolation formula is used to convert R1CS into an equivalent QAP form (Quadratic Arithmetic Program), thereby obtaining VK and PK.

[0042] The central server S deposits money into the smart contract in advance, then automatically verifies the proof submitted by the client through the smart contract, and automatically pays and aggregates according to the fitting results.

[0043] The trusted parameter generator A generates random numbers r1, r2, ...r for n clients n , used for subsequent security aggregation, A distributes the random number r i Generate a promise:

[0044] C(r i )=H(r i )

[0045] Where H is a hash function used to generate a commitment for data.

[0046] Step 2: Model training.

[0047] Federated learning sets up several rounds of training. For the kth round of training, the central server S will aggregate the global model m after the k-1th round of training. k Distributed to each client C1, C2, ..., C n This step ensures that all participating devices start training from the same starting point, thus ensuring the effectiveness of subsequent training.

[0048] Clients C1, C2, …, C n Receive the global model m k Then, using the dataset Train the model. By a number of forms such as (X j ,y j ) sample composition, where X j is the characteristic vector of the sample, y j is the label corresponding to the sample.

[0049] Forward propagation is the process of passing input data through the neural network layer by layer to calculate the output result. The result of forward propagation is the network's predicted value for the input data. Output result Is a probability distribution vector, indicating the probability of the sample belonging to each category. The details are as follows:

[0050] Y [l] =W [l] X [l-1] +b [l]

[0051]

[0052] Among them, Y [l] represents the output of the lth layer; Indicates the final output result. [l] and b [l] are the weights and biases of the lth layer; the softmax activation function is used to convert the linear output into a probability distribution.

[0053] After getting the predicted value, use the cross entropy loss function to calculate the predicted value and the true value y j The error between Suppose there are K categories and the cross entropy loss is The calculation method is as follows:

[0054]

[0055] After that, back propagation is performed to calculate the gradient according to the loss function and adjust the network weights and biases to minimize the loss. The gradient update δ of the sample is recorded during the back propagation process. The details are as follows:

[0056]

[0057] Among them, Z (l) is the error term of the lth layer, (X (l-1) ) T is the transpose of the activation values ​​of the previous layer.

[0058] During the training process, each client records the error set of all samples and gradient update δ.

[0059] Step 3: Proof of training quality.

[0060] For client C i , after the kth round of training, a loss set is generated and the gradient update δ i . C i The random number r distributed by the parameter generation center A i δ i Perform mask generation C i Will And the loss set All submitted to the smart contract.

[0061] in:

[0062]

[0063] in, Represents the masked gradient update value.

[0064] In order to verify the training quality of each client without accessing the training process and training data, the client can use zk-SNARKs to generate a concise non-interactive zero-knowledge proof of the training process. Each client uses the error set of all samples And gradient update δ generates challenges, according to which all samples are collected from the challenge Randomly select t samples from use The calculation process of the sample in is used to generate the proof:

[0065] proof: j ,δ j =train(X j ,y j ,m k ,PK)|j∈C}

[0066] Among them, l j represents the error of the jth client; m k represents the global model after the k-1th round of training aggregation; C represents the client; train represents the training process.

[0067] zk-SNARKs (Zero-Knowledge Succinct Non-Interactive Argument of Knowledge) is an efficient zero-knowledge proof technology that allows the prover to prove the correctness of a proposition to the verifier without leaking any information about the proposition.

[0068] For client C i :

[0069] 1.C i For its private data D i Generate a commitment as follows:

[0070] C(D i )=H(X i ,y i )

[0071] Where H is a hash function used to generate a commitment for data.

[0072] 2. Each client is based on the error set of all samples Generate a challenge along with the gradient update δ, and randomly select t samples from the entire sample set according to the challenge to obtain Use The calculation process of the samples in is used to generate a proof. For client C i ,

[0073] 3.C i Generate a witness for all the constraints in the circuit generated by the central server, including inputs, intermediate values, and outputs. Then, use the proof key PK and The calculation process of the samples in to generate a proof containing polynomials. The values of these polynomials at specific points can prove the correctness of the calculation process without revealing the specific input data.

[0074] Step 4: Training quality payment.

[0075] The smart contract uses the verification key VK, the proofs submitted by each client, and the corresponding losses and gradient updates for verification, thereby verifying the legality of the training process.

[0076] After that, evaluate the training quality. Model all the sample loss data submitted by the clients that pass the verification, and choose a normal distribution for fitting. Obtain two threshold values t1, t2, t1 < t2, according to the fitted loss distribution function, to determine the range of the sample loss l and thus determine the category of the sample. If l < t1, it is considered a valid sample; if l > t2, it is considered an invalid sample. Since the predicted value of a valid sample after one round of training has a small gap with the true value, the loss l will also be small. If there are invalid data samples in the training set, the predicted value of the invalid sample has a large gap with the true value, so its loss l will be large.

[0077] In summary, there will be an obvious difference in the loss l between normal samples and clean samples, so they can be well distinguished by setting thresholds.

[0078] Let the set of sample losses calculated during the training process of the i-th client be be N is the total number of samples in the local dataset of the client.

[0079] The specific steps for fitting the normal distribution include:

[0080] First, calculate the sample statistics. To fit the normal distribution, it is necessary to calculate the mean and standard deviation of the sample losses. Specifically as follows:

[0081] Sample mean μ i :

[0082]

[0083] in, Represents the error loss of the jth sample in the sample set of the i-th client.

[0084] Sample standard deviation σ i :

[0085]

[0086] Among them, μ i Represents the mean of the sample loss of the i-th client.

[0087] Then, select the distribution model. In this method, select the normal distribution. Fitting. The probability density function of the normal distribution is as follows:

[0088]

[0089] Among them, μ represents the sample mean, σ represents the sample standard deviation, π represents pi, and x represents the value of the random variable.

[0090] After that, we fit the distribution. We use the maximum likelihood estimation method to estimate the distribution parameters. For a normal distribution, the sample mean and sample standard deviation are the parameter estimates.

[0091] Let t1=μ-σ, t2=μ+3σ.

[0092] The smart contract makes the above judgment on the losses uploaded by each client. i , if there are invalid samples in the sample, that is The training quality of the client is considered unqualified, and the central server will not use the client's gradient update for aggregation, nor will it pay the client. If there are no invalid samples in the extracted samples, the training quality of the client is considered qualified, and the client will be automatically paid through the smart contract. middle, The more samples there are, the more remuneration will be paid to the client. The remuneration is calculated as follows:

[0093]

[0094] Among them, P(x) is the final remuneration paid, W is the initial remuneration, and x is middle The number of samples, D i C i The size of the dataset.

[0095] Afterwards, the model parameters submitted by the clients with qualified training quality are averaged and aggregated through the smart contract. therefore Calculate the security aggregation result After that, it is distributed to each client for the next round of training.

[0096] Example

[0097] In order to make the purpose, technical solution, applicability, innovation and advantages of the present invention more clearly expressed, the present invention is further described in detail by proposing an application example of the method of the present invention.

[0098] At present, federated learning has a certain degree of application in the field of Internet of Things. In daily life, more and more people choose to wear smart bracelets. Among them, each smart bracelet can be regarded as an Internet of Things device. Everyone collects their own health data through the application platform and stores it on the local device. The research institution designs the initial health analysis model and sends it to all participating devices. Each participant uses his own health data to train the model on the local device and generates a local model update. After that, the participant sends his local model to the research institution, and the research structure performs secure aggregation to form a new model without accessing any individual's original data. However, during the training process, the quality of the data trained by various financial institutions varies, which requires the research structure to verify and pay for the training quality of each participant.

[0099] Therefore, through the federated learning method based on training quality evaluation and on-chain remuneration payment proposed in the present invention, the research institution is regarded as the central server and each participant is regarded as the client. The training quality can be evaluated and automatically paid through the blockchain according to the training quality of each client.

[0100] The specific process is as follows:

[0101] Step 1: The research institution generates the key for proof and deploys the smart contract on the blockchain. The trusted parameter generator generates random numbers and distributes them to each participant.

[0102] Step 2: Participants train based on the distributed model and their own local data and submit the training results to the research institution.

[0103] Step 3: In order to verify the training quality of each participant without accessing the training process and training data, each participant uses zk-SNARKs to generate a zero-knowledge proof and submit it to the smart contract.

[0104] Step 4: The research institution automatically verifies the proof through smart contracts, then models all sample losses submitted by the verified participants, selects the normal distribution for fitting, and conducts training quality assessment and automatic payment through smart contracts. The smart contract averages and aggregates the model parameters submitted by participants with qualified training quality, and distributes the aggregated results to each participant for the next round of training.

[0105] At this point, the process is complete.

Claims

1. A federated learning method based on training quality evaluation and on-chain reward payment, in which the participating devices include a central server S, clients C1, C2, ..., C n , a trusted parameter generator A; the central server S is responsible for distributing models and keys used for proof, as well as deploying related smart contracts; clients C1, C2, …, C n Responsible for model training and generating proof of training quality; trusted parameter generator A is responsible for distributing the parameters required for secure aggregation; It is characterized in that It includes the following steps: Step 1: Device initialization; First, the central server S generates a verification key VK and a proof key PK according to the training task; After that, the central server S distributes the PK to each client for generating proofs during the training process, and the VK is used by S to verify the correctness of the proofs later; S deploys a smart contract on the blockchain network for subsequent automatic verification, payment, and aggregation; The central server S deposits money into the smart contract in advance, and then automatically verifies the proofs submitted by each client through the smart contract. The smart contract fits the losses submitted by each client, and then automatically pays and aggregates according to the fitting results; the trusted parameter generator A generates random numbers r1, r2, ...r for all clients. n , used for subsequent security aggregation, Step 2: Model training; Federated learning sets up several rounds of training; for the kth round of training, the central server S aggregates the global model m after the k-1th round of training. k Distribute to each client C1, C2, ..., C n ; Clients C1, C2, …, C n Receive the global model m k Then, using the dataset Train the model; By a number of forms such as (X j ,y j ) sample composition, X j is the characteristic vector of the sample, y j is the label corresponding to the sample; for each sample (X j ,y j ), the forward propagation will be X j The process of passing the input data through the neural network layer by layer to calculate the output result; the result of forward propagation is the network's calculation of the input data X j The predicted value of is a probability distribution vector, which indicates the probability of the sample belonging to each category. After obtaining the predicted value, the cross entropy loss function is used to calculate the predicted value. and the true value y j The loss between As follows: Where K is the number of categories; After that, backpropagation is performed; the gradient is calculated according to the loss function and the network weights and biases are adjusted to minimize the loss; during the backpropagation process, the gradient update δ of the sample is recorded as follows: The above are the gradient update formulas for the weight and bias of the lth layer respectively; where Z (l) is the error term of the lth layer; represents partial derivative, L represents loss function, W (l) represents the weight matrix of the lth layer, b (l) represents the bias of the lth layer; (X (l-1) ) T It is the transpose of the activation value of the previous layer, and T represents transpose; Each client records the loss set of all samples during training and gradient update δ; Step 3: Proof of training quality; For client C i , after the kth round of training, a loss set is generated and the gradient update δ i ; Client C i The r distributed by center A is generated by parameters i , update the gradient δ i Perform mask generation Client C i Will And the loss set All submitted to the smart contract; The client generates a concise non-interactive zero-knowledge proof of the training process; the client generates random numbers non-interactively from all sample sets Randomly select t samples from use The calculation process of the sample is used to generate proof; The client subsequently submits the generated proof to the central server S for verification; if the verification passes, the central server S believes that the client has not faked; Step 4: Payment for training quality; The smart contract first verifies the proofs proof submitted by each client; then, it models all the sample loss data submitted by the clients that pass the verification and selects a normal distribution for fitting; after that, two threshold values t1, t2 are obtained according to the fitted loss distribution function, and t1 < t2, which are used to judge the range of the sample loss l, so as to determine the category of the sample; if l < t1, it is considered a valid sample; if l > t2, it is considered an invalid sample; The smart contract makes the above judgment on the losses uploaded by each client; i , if there are invalid samples in the sample, that is The training quality of the client is considered unqualified, and the central server S will not use the client's gradient update for aggregation, nor will it pay the client. If there are no invalid samples in the small number of samples extracted, the training quality of the client is considered qualified, and the client is automatically paid through the smart contract. At the same time, the loss set middle The more samples a client has, the more remuneration it will pay to the client. After that, through the smart contract, the model parameters submitted by the clients with qualified training quality are averaged and aggregated according to the number of valid samples, and the aggregated result is distributed to each client for the next round of training.

2. A federated learning method based on training quality evaluation and on-chain remuneration payment as claimed in claim 1, characterized in that: In step 1, A distributes the random number r i Generate a promise: C(r i )=H(r i ) Where H is a hash function used to generate a commitment of the data.

3. A federated learning method based on training quality evaluation and on-chain remuneration payment as claimed in claim 1, characterized in that: In step 2, forward propagation is the process of passing the input data through the neural network layer by layer to calculate the output result. The result of forward propagation is the network's predicted value of the input data, and the output result is Is a probability distribution vector, indicating the probability of the sample belonging to each category, as follows: Y [l] =W [l] X [l-1] +b [l] Among them, Y [l] represents the output of the lth layer; Indicates the final output result; W [l] and b [l] are the weights and biases of the lth layer; the softmax activation function is used to convert the linear output into a probability distribution.

4. A federated learning method based on training quality evaluation and on-chain remuneration payment as claimed in claim 1, characterized in that: In step 3, for client C i , after the kth round of training, a loss set is generated and the gradient update δ i ; C i The random number r distributed by the parameter generation center A i δ i Perform mask generation C i Will And the loss set All submitted to the smart contract; among them: in, Represents the gradient update value after masking; The client uses zk-SNARKs to generate concise non-interactive zero-knowledge proofs for the training process; each client generates a set of errors based on all samples. And gradient update δ generates challenges, according to which all samples are collected from the challenge Randomly select t samples from use The calculation process of the sample in is used to generate the proof: proof:{l j ,δ j =train(X j ,y j ,m k ,PK)|j∈C} Among them, l j represents the error of the jth client; m k represents the global model after the k-1th round of training aggregation; C represents the client; train represents the training process; For client C i : C i For its private data D i Generate a commitment as follows: C(D i )=H(X i ,y i ) Where H is a hash function used to generate a commitment of the data; Each client is based on the error set of all samples And the gradient update δ generates a challenge challenge, according to the challenge from all sample sets Randomly select t samples from use The calculation process of the sample in is used to generate proof; for client C i , C i Generate a witness for all constraints in the circuit generated by the central server, including inputs, intermediate values, and outputs; then, use the proof key PK and The computation process of the sample in is a proof containing a polynomial.

5. A federated learning method based on training quality evaluation and on-chain remuneration payment as claimed in claim 1, characterized in that: In step 4, let the sample loss set calculated during the training process of the i-th client be for N is the total number of samples in the client's local dataset; The specific steps for fitting the normal distribution include: First, calculate the sample statistics; to fit the normal distribution, the mean and standard deviation of the sample loss need to be calculated; specifically as follows: Sample mean μ i : in, represents the error loss of the jth sample in the sample set of the i-th client; Sample standard deviation σ i : Among them, μ i represents the mean of the sample loss of the i-th client; Then, select the distribution model; select Normal distribution Fitting; the probability density function of the normal distribution is as follows: Where μ represents the sample mean, σ represents the sample standard deviation, π represents pi, and x represents the value of the random variable; After that, perform the fitting distribution; use the maximum likelihood estimation method to estimate the distribution parameters; for the normal distribution, the sample mean and the sample standard deviation are the parameter estimation values; Let t1 = μ - σ, t2 = μ + 3σ; The smart contract makes the above judgment on the losses uploaded by each client; i , if there are invalid samples in the sample, that is If the training quality of the client is not up to standard, the central server will not use the client's gradient update for aggregation, nor will it pay the client. If there are no invalid samples in the extracted samples, the training quality of the client is considered to be qualified, and the client will be automatically paid through the smart contract. At the same time, the loss set middle, The more samples there are, the more remuneration will be paid to the client; the remuneration calculation formula is as follows: Among them, P(x) is the final remuneration paid, W is the initial remuneration, and x is middle The number of samples, D i C i The size of the dataset; Afterwards, the model parameters submitted by the clients with qualified training quality are averaged and aggregated through the smart contract; since the random number meets therefore Calculate the security aggregation result After that, it is distributed to each client for the next round of training.

Images