A data collaboration method based on a verifiable mechanism

Abstract

This invention relates to a data collaboration method based on a verifiable mechanism, relating to the fields of machine learning and data security. The method includes: dividing multiple participants in a set data collaboration task into verifiers and multiple computers; training a local model using local data from each computer; after each computer's local model completes a set number of training rounds, sending proof information to the verifier; verifying the decrypted proof information and counting the number of parameter merging signals issued by computers whose decrypted proof information is valid; if the number of computers issuing parameter merging signals is greater than or equal to a set threshold, then the verifier merges the model parameters of the computers who issued parameter merging signals to obtain the current aggregated model; if the current aggregated model meets a set standard, then the current aggregated model is used as the final aggregated model. This invention improves the security of privacy data and the credibility of shared data.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and data security technology, and in particular to a data collaboration method based on a verifiable mechanism. Background Technology

[0002] In today's era, with the rapid development of artificial intelligence, data has gradually become a crucial productive force. Enterprises across various industries collect vast amounts of data, but the existence of "data silos" can severely impact their development. Privacy protection during data collaboration has also become a major challenge. For example, in the medical field, medical data is often highly sensitive and private, containing various sensitive patient information such as names, addresses, contact information, examination records, disease diagnoses, and even family genetic history. If this sensitive data is tampered with or leaked, individuals may suffer negative social judgment, even discrimination or insults, affecting their lives and mental well-being. Therefore, the security and reliability of data privacy during multi-party data collaboration still need improvement. Summary of the Invention

[0003] The purpose of this invention is to provide a data collaboration method based on a verifiable mechanism, which improves the security of privacy data and the credibility of shared data.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A data collaboration method based on a verifiable mechanism includes:

[0006] The multiple participants in setting up the data collaboration task are divided into one verifier and multiple computers; each participant includes a local model with the same local data and network structure; the local model of the verifier is denoted as the aggregate model.

[0007] The verifier generates a proof key and a verification key, and sends the proof key to multiple computers.

[0008] The local data of each of the aforementioned computers is used to train the current local model of each corresponding computer for a set number of rounds;

[0009] After each of the aforementioned computers completes a set number of rounds of model training, its current local model sends proof information to the verifier.

[0010] The verifier decrypts each of the proof information according to the verification key, verifies the decrypted proof information, and counts the number of parameter merging signals issued by the computer that are valid after decryption of the proof information;

[0011] Determine whether the number of the calculators that issued the parameter merging signal is greater than or equal to a set threshold.

[0012] If the number of the computers that issue parameter merging signals is greater than or equal to a set threshold, then the verifier merges the model parameters of the computers that issue parameter merging signals to obtain the current aggregated model.

[0013] If the number of the computers that issue parameter merging signals is less than the set threshold, return to the step "use the local data of each computer to train the current local model of each corresponding computer for a set number of rounds";

[0014] Determine whether the prediction accuracy of the current aggregation model meets the set standard;

[0015] If the set standard is not met, return to step "Use the local data of each of the computers to train the current local model of each corresponding computer for a set number of rounds";

[0016] If the set criteria are met, the current aggregation model will be used as the final aggregation model to complete the set data collaboration task.

[0017] Optionally, after each of the computers completes a set number of rounds of model training, its current local model sends proof information to the verifier, specifically including:

[0018] Each of the aforementioned computers employs a zero-knowledge, concise, non-interactive knowledge proof technique to convert the computational task of training the current local model for a set number of rounds into an arithmetic circuit. The arithmetic circuit is then encrypted using the proof key to generate the proof information, which is a quadratic arithmetic polynomial.

[0019] Optionally, the verifier decrypts each of the proof information according to the verification key, verifies the decrypted proof information, and counts the number of parameter merging signals issued by the computer whose decrypted proof information is valid, specifically including:

[0020] The decrypted proof information includes multiple polynomials. Pre-defined mathematical operations are performed on each of the polynomials to generate a result polynomial.

[0021] Determine whether the resulting polynomial is the same as the quadratic arithmetic polynomial;

[0022] If they are the same, the decrypted proof information is valid, causing the j-th computer to issue a parameter merging signal.

[0023] If they are different, the decrypted proof information is invalid. Then, the local data of the j-th computer is used to train the current local model of the j-th computer for a set number of rounds, and the model parameters after each round of training are uploaded to the verifier.

[0024] Optionally, the step of using the local data of each of the computational users to train the current local model of each corresponding computational user for a set number of rounds specifically includes:

[0025] When the verifier receives the model parameters after the k-th round of training of the j-th computer, it uses the verifier's local data and the current aggregated model to verify the model parameters after the k-th round of training of the j-th computer. When the local model corresponding to the model parameters after the k-th round of training of the j-th computer is greater than the prediction accuracy of the aggregated model, a positive interaction is recorded and the number of positive interactions of the j-th computer is updated; otherwise, a negative interaction is recorded and the number of negative interactions of the j-th computer is updated.

[0026] After the local model of the j-th computer completes the training of the set number of rounds, the proof information generated by the j-th computer using the proof key is sent to the verifier.

[0027] Optionally, after setting the current aggregation model as the final aggregation model, the method further includes:

[0028] The reputation value of each computational is determined based on the number of positive interactions, the number of negative interactions, and the communication quality of each computational; the communication quality of each computational is the probability of successful data parameter transmission between each computational and the verifier.

[0029] The verifier's reputation value is updated based on the reputation value of each calculator.

[0030] Optionally, dividing the multiple participants in setting up the data collaboration task into one verifier and multiple computers specifically includes:

[0031] The participant with the highest reputation value is designated as the verifier, and the other participants besides the verifier are designated as calculators.

[0032] Optionally, the calculator's reputation value is:

[0033] T i→j =b i→j +au i→j ;

[0034] Among them, T i→j Let b represent the reputation value of the j-th calculator. i→j , where u represents the trust ratio of the j-th calculator. i→j Let represent the uncertainty ratio of the j-th calculator, and a represent the coefficient of the uncertainty ratio, a∈[0,1];

[0035]

[0036] u i→j =1-q i→j ;

[0037] Where α represents the number of positive interactions for the j-th calculator, β represents the number of negative interactions for the j-th calculator, κ represents the weighting coefficient for the number of positive interactions, η represents the weighting coefficient for the number of negative interactions, and q i→j This represents the communication quality of the j-th computer.

[0038] Optionally, the verifier's reputation value is:

[0039]

[0040] Where T represents the reputation value of the verifier, and n represents the number of the calculators.

[0041] Optionally, if the number of computers issuing parameter merging signals is greater than or equal to a set threshold, then the verifier merges the model parameters of the computers issuing the parameter merging signals to obtain the current aggregated model, specifically including:

[0042] The average value of each parameter in the model parameters of the calculator that issued the parameter merging signal is used to replace the corresponding parameters in the current aggregate model to obtain the aggregate model with updated parameters.

[0043] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0044] The present invention enables the computer to train the model using local data and only sends the parameters obtained from the model training to the verifier, thereby avoiding sharing local data with other participants and improving the security of privacy data. In addition, the verification of the proof information is realized through the proof key and the verification key, which improves the credibility of the shared data. Attached Figure Description

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 This is a flowchart illustrating a data collaboration method based on a verifiable mechanism according to the present invention. Figure 1 ;

[0047] Figure 2 This is a flowchart illustrating a data collaboration method based on a verifiable mechanism according to the present invention. Figure 2;

[0048] Figure 3 This is a schematic diagram of the computational model for the verifiable mechanism of this 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 data collaboration method based on a verifiable mechanism, which improves the security of privacy data and the credibility of shared data.

[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] Figure 1 This is a flowchart illustrating a data collaboration method based on a verifiable mechanism according to the present invention. Figure 1 , Figure 2 This is a flowchart illustrating a data collaboration method based on a verifiable mechanism according to the present invention. Figure 2 ,like Figure 1-Figure 2 As shown, a data collaboration method based on a verifiable mechanism includes:

[0053] Step 101: Divide the multiple participants in the data collaboration task into one verifier and multiple computers; each participant includes a local model with the same local data and network structure; the local model of the verifier is denoted as the aggregate model.

[0054] Step 101 specifically includes:

[0055] The participant with the highest reputation value is designated as the verifier, and the other participants besides the verifier are designated as calculators.

[0056] If each participant is setting up a data collaboration task for the first time, their reputation value will be the default initial value.

[0057] This invention proposes a data collaboration method based on a verifiable mechanism and discloses a computational model based on a verifiable mechanism, such as... Figure 3As shown, verifying the calculation results of each participant effectively reduces the probability of malicious falsification and improves the credibility of the calculation results. While protecting the privacy and security of each participant's local data, it enables mutual verification and collaborative data sharing among all parties. Furthermore, by electing a leader based on the participants' reputation scores to be responsible for verifying calculation results, aggregating model parameters, and calculating and updating reputation scores, data collaboration among the participants becomes more fair and efficient.

[0058] In the verifiable mechanism-based computational model, participants are divided into verifiers and computers based on their functions. Verifiers are the leaders, selected based on the reputation scores of all participants. Computers are the other participants, and the model is based on a zero-knowledge proof mechanism. The entire algorithm mainly includes preparation, generating proof information, and information verification, such as... Figure 3 As shown.

[0059] In the computational model based on the verifiable mechanism, three polynomial functions, Keygen(), Compute(), and Verify(), are defined as follows.

[0060] Keygen(): This function generates a random key. During the preparation phase, the verifier generates two keys: a proof key and a PK key. F and verification key VK F The function is represented as follows:

[0061] (PK F VK F )←Keygen(y,1 λ )

[0062] Where y is the task function and λ is the security parameter.

[0063] Compute(): During the computation phase, each participant completes its computation task and uses the PK. F A verification message is generated, which is a quadratic arithmetic polynomial. This quadratic arithmetic polynomial is obtained by the arithmetic circuit after the computation task is converted by the computer. Finally, each participant transmits polynomials V(x), W(x), Y(x), and y(x) to the verifier. Here, V(x), W(x), and Y(x) represent the polynomials of the left, right, and output lines of the arithmetic circuit, respectively, and y(x) is the final result of the verification message. The function is represented as follows:

[0064] (y(x),π y )←Compute(PK F ,x)

[0065] Where x is the input, π yThe proof of the correctness of y(x) is derived from V(x), W(x), and Y(x).

[0066] The left, right, and output lines of an arithmetic circuit refer to the polynomials generated during the process of converting an arithmetic circuit into a quadratic arithmetic program, which are used to generate quadratic arithmetic polynomials for verification.

[0067] Arithmetic circuits contain multiplication gates and addition gates. A multiplication gate is considered to have exactly two input lines, called the left and right lines, used for data input. The output line is used for data output.

[0068] Verify(): During the verification phase, the verifier uses the verification key VK. F The decryption process involves using the obtained V(x), W(x), and Y(x) to compute a polynomial h(x). If the generated polynomial h(x) is consistent with y(x) and the computation process is correct, the computation task is proven to be valid; otherwise, the computation task is proven to be invalid. The function is represented as follows:

[0069] {0,1}←Verify(VK F ,x,y(x),π y )

[0070] The function outputs 1 when the polynomial h(x) = y(x), which proves that the computation task is valid; otherwise, it outputs 0, which proves that the computation task is invalid.

[0071] Specifically, step 101 further includes: the verifier uses the Keygen() function to generate a proof key and a verification key, and the proof key is distributed to each computer.

[0072] Step 102: The verifier generates a proof key and a verification key, and sends the proof key to multiple computers.

[0073] Step 103: Use the local data of each of the aforementioned computers to train the current local model of each corresponding computer for a set number of rounds.

[0074] Step 103 specifically includes:

[0075] When the verifier receives the model parameters after the k-th round of training of the j-th computer, it uses the verifier's local data and the current aggregated model to verify the model parameters after the k-th round of training of the j-th computer. When the local model corresponding to the model parameters after the k-th round of training of the j-th computer is greater than the prediction accuracy of the aggregated model, a positive interaction is recorded and the number of positive interactions of the j-th computer is updated; otherwise, a negative interaction is recorded and the number of negative interactions of the j-th computer is updated.

[0076] After the local model of the j-th computer completes the training of the set number of rounds, the proof information generated by the j-th computer using the proof key is sent to the verifier.

[0077] Step 104: After completing the set number of rounds of model training, each of the computers' current local models sends proof information to the verifier.

[0078] Step 104 specifically includes:

[0079] Each of the aforementioned computers employs a zero-knowledge concise non-interactive knowledge proof technique (zk-SNARK) to convert the computational task of training the current local model for a set number of rounds into an arithmetic circuit. The arithmetic circuit is then encrypted using the proof key to generate the proof information, which is a quadratic arithmetic polynomial.

[0080] The computational tasks also include loading data and data preprocessing.

[0081] Step 105: The verifier decrypts each of the proof information according to the verification key, verifies the decrypted proof information, and counts the number of parameter merging signals issued by the computer that are valid after decryption of the proof information.

[0082] Step 105 specifically includes:

[0083] The decrypted proof information includes multiple polynomials. Pre-defined mathematical operations are performed on each polynomial, including addition, subtraction, multiplication, and division, to generate a result polynomial.

[0084] Here, after decryption, the verifier obtains polynomials including V(x), W(x), Y(x), and y(x). The verifier performs a preset mathematical operation on V(x), W(x), Y(x), and y(x) to obtain a result polynomial that can be V(x)*W(x)-Y(x). If y(x) = V(x)*W(x)-Y(x), the information is valid. If no result polynomial with the same result as y(x) can be calculated, the information is invalid.

[0085] Determine whether the resulting polynomial is the same as the quadratic arithmetic polynomial;

[0086] If they are the same, the decrypted proof information is valid, causing the j-th computer to issue a parameter merging signal.

[0087] If they are different, the decrypted proof information is invalid. Then, the local data of the j-th computer is used to train the current local model of the j-th computer for a set number of rounds, and the model parameters after each round of training are uploaded to the verifier.

[0088] The verifier accumulates the number of times the decrypted proof information corresponding to each computer is invalid. If the accumulated number exceeds the preset number of verifications, the corresponding computer is refused further participation in model training.

[0089] The leader uses the RONI (Reject On Negative Impact) scheme for evaluation. This scheme evaluates the training performance of each participant's model by comparing the leader's performance when updating the model locally with local data versus not updating it locally. If a participant's local model update improves the model's prediction performance, the leader considers the training iteration a positive interaction with the participant, and vice versa. If the performance degradation of a local model update at the leader's level exceeds a pre-defined threshold, that local model update is rejected when integrating all local model updates.

[0090] Step 106: Determine whether the number of the calculators that issued the parameter merging signal is greater than or equal to a set threshold.

[0091] If the number of the calculators that issue the parameter merging signal is greater than or equal to a set threshold, then step 107 is executed.

[0092] Step 107: The verifier merges the model parameters of the computer that issued the parameter merging signal to obtain the current aggregated model.

[0093] Step 107 specifically includes:

[0094] The average value of each parameter in the model parameters of the calculator that issued the parameter merging signal is used to replace the corresponding parameters in the current aggregate model to obtain the aggregate model with updated parameters.

[0095] If the number of the calculators that issue the parameter merging signal is less than the set threshold, return to step 103.

[0096] Step 108: Determine whether the prediction accuracy of the current aggregation model meets the set standard.

[0097] If the set standard is not met, return to step 103.

[0098] If the set standard is met, proceed to step 109.

[0099] Step 109: Use the current aggregation model as the final aggregation model to complete the set data collaboration task.

[0100] For model training with a specified number of rounds, the tuple vector γ is used. i→j ={b i→j ,d i→j ,u i→j} is used to express the reputation value of participant j in processing task i for model training.

[0101]

[0102] The reputation vector expression reveals that positive interactions increase a participating node's reputation value, while negative interactions decrease it. To increase the cost of malicious behavior and reduce negative interaction events, the weight of negative interactions in reputation calculation should be higher than that of positive interactions.

[0103] After completing the task, the leader updates the reputation values ​​of each participant for task i to each participant's records. Subsequent tasks can then use these reputation values ​​to elect participants with high reputation values ​​to undertake tasks such as verifying calculation results, aggregating model parameters, and calculating and updating the reputation values ​​of each participant.

[0104] After setting the current aggregation model as the final aggregation model, the process also includes:

[0105] The reputation value of each computational is determined based on the number of positive interactions, the number of negative interactions, and the communication quality of each computational; the communication quality of each computational is the probability of successful transmission of data parameters between each computational and the verifier.

[0106] Leaders calculate average communication quality n is the number of calculators, q k This represents the communication quality of the k-th computer.

[0107] The verifier's reputation value is updated based on the reputation value of each calculator.

[0108] The calculator's reputation value is:

[0109] T i→j =b i→j +au i→j ;

[0110] Among them, T i→j Let b represent the reputation value of the j-th calculator. i→j , where u represents the trust ratio of the j-th calculator. i→j Let represent the uncertainty ratio of the j-th calculator, and a represent the coefficient of the uncertainty ratio, a∈[0,1];

[0111]

[0112] u i→j =1-q i→j ;

[0113] Where α represents the number of positive interactions for the j-th calculator, β represents the number of negative interactions for the j-th calculator, κ represents the weighting coefficient for the number of positive interactions, η represents the weighting coefficient for the number of negative interactions, and q i→j This represents the communication quality of the j-th computer.

[0114] The verifier's reputation value is:

[0115]

[0116] Where T represents the reputation value of the verifier, and n represents the number of the calculators.

[0117] The final aggregation model is used for text classification or image recognition.

[0118] In a specific implementation, the final aggregation model is used for vehicle identification. Each participating party does not want to disclose its local vehicle data. The final aggregation model is obtained through a data collaboration method based on a verifiable mechanism according to the present invention. The collected vehicle images are input into the final aggregation model, and the vehicle type and vehicle location are output.

[0119] As a specific implementation method, taking the MNIST dataset classification task as an example, the aggregation model here refers to the aggregation model obtained by different data collaborators using their own MNIST data for machine learning training. Throughout the process, there is no direct sharing of raw data; only parameters are shared, thereby achieving the goal of protecting the data privacy of the participants.

[0120] This invention, based on verifiable computation and zero-knowledge proofs, discloses a key verification mechanism to verify the credibility of computation results from each participant in data collaboration, reducing the likelihood of malicious transmission of false or erroneous information that could affect the final outcome. Furthermore, it employs a reputation-based election method to select a leader from among the participants to be responsible for verifying computation results, aggregating model parameters, and calculating and updating reputation values.

[0121] In verifiable computation, participants are divided into computers and verifiers based on their functions. The computer first converts the computational task into an arithmetic circuit, then uses cryptographic techniques to create public and quickly verifiable mathematical relations and their values. Finally, the computer sends these values ​​to the verifier. The verifier can use these values ​​to check whether the public verification relations are satisfied. Verifiable computation aims not only to obtain the correct result but also to verify that result with far less computation than the result itself requires.

[0122] Zero-knowledge proofs refer to the ability of a prover to convince a verifier that a statement is true without revealing any useful information to the verifier. Zero-knowledge proofs do not involve large amounts of data transmission or computation during use; compared to other privacy-preserving computation schemes, the verification process is simple and computationally insignificant. Furthermore, zero-knowledge proof algorithms are non-interactive and do not require the participation of a trusted third party.

[0123] The relevant terms in this invention are explained below.

[0124] Data collaboration: Data owners share data information to utilize data information they do not possess, thereby further exploring and enhancing the value of their data.

[0125] Reputation score: Used to assess the credibility of each task participant, and is affected by the trust ratio, skepticism ratio, and uncertainty ratio.

[0126] Leader: Elected by all task participants based on reputation, responsible for aggregating model parameters and calculating the reputation value of each participating node, and updating the reputation value.

[0127] Verifiable computation: Verifiable computation refers to the ability of a task initiator to delegate computation to a potentially untrusted computing power provider while retaining the ability to verify the results. The (untrusted) computing power provider is required to submit proof of the correctness of the computation results upon completing the computation task.

[0128] Zero-knowledge proof: A zero-knowledge proof refers to a proof in which the prover can convince the verifier that a statement is true without revealing any useful information to the verifier. It typically requires meeting the following three conditions: completeness, reliability, and zero knowledge.

[0129] Completeness: If the prover's statements are all true, then an honest prover can convince a verifier who strictly enforces the agreement of a certain assertion;

[0130] Reliability: If the prover's statement is false, i.e., he may not know the secret information, then the probability that the prover will successfully deceive the verifier who strictly enforces the protocol is very small;

[0131] Zero-knowledge: During the verification process, the verifier can obtain no additional information beyond believing that the prover possesses certain secret information. Therefore, the verifier cannot prove to a third party that he possesses the secret.

[0132] 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.

[0133] 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 data collaboration method based on a verifiable mechanism, characterized in that, include: The multiple participants in setting up the data collaboration task are divided into one verifier and multiple computers; each participant includes a local model with the same local data and network structure; the local model of the verifier is denoted as the aggregate model. The verifier generates a proof key and a verification key, and sends the proof key to multiple computers. The local data of each of the aforementioned computers is used to train the current local model of each corresponding computer for a set number of rounds; After each of the aforementioned computers completes a set number of rounds of model training, its current local model sends proof information to the verifier. The verifier decrypts each of the proof information according to the verification key, verifies the decrypted proof information, and counts the number of parameter merging signals issued by the computer that are valid after decryption of the proof information. Determine whether the number of the calculators that issued the parameter merging signal is greater than or equal to a set threshold. If the number of the computers that issue parameter merging signals is greater than or equal to a set threshold, then the verifier merges the model parameters of the computers that issue parameter merging signals to obtain the current aggregated model. If the number of the computers that issue parameter merging signals is less than the set threshold, return to the step "use the local data of each computer to train the current local model of each corresponding computer for a set number of rounds"; Determine whether the prediction accuracy of the current aggregation model meets the set standard; If the set standard is not met, return to step "use the local data of each of the computers to train the current local model of each corresponding computer for a set number of rounds"; If the set criteria are met, the current aggregation model will be used as the final aggregation model to complete the set data collaboration task. After each of the aforementioned computers completes a set number of rounds of model training, its current local model sends proof information to the verifier, specifically including: Each of the aforementioned computers employs a zero-knowledge concise non-interactive knowledge proof technique to convert the computational task of training the current local model for a set number of rounds into an arithmetic circuit. The arithmetic circuit is then encrypted using the proof key to generate the proof information, which is a quadratic arithmetic polynomial. The verifier decrypts each of the proof information according to the verification key, verifies the decrypted proof information, and counts the number of parameter merging signals issued by the computer whose decrypted proof information is valid, specifically including: The decrypted proof information includes multiple polynomials. Pre-defined mathematical operations are performed on each of the polynomials to generate a result polynomial. Determine whether the resulting polynomial is the same as the quadratic arithmetic polynomial; If they are the same, the decrypted proof information is valid, causing the j-th computer to issue a parameter merging signal. If they are different, the decrypted proof information is invalid. Then, the local data of the j-th computer is used to train the current local model of the j-th computer for a set number of rounds, and the model parameters after each round of training are uploaded to the verifier.

2. The data collaboration method based on a verifiable mechanism according to claim 1, characterized in that, The step of training the current local model of each corresponding computer for a set number of rounds using the local data of each computer specifically includes: When the verifier receives the model parameters after the k-th round of training of the j-th computer, it uses the verifier's local data and the current aggregated model to verify the model parameters after the k-th round of training of the j-th computer. When the local model corresponding to the model parameters after the k-th round of training of the j-th computer is greater than the prediction accuracy of the aggregated model, a positive interaction is recorded and the number of positive interactions of the j-th computer is updated; otherwise, a negative interaction is recorded and the number of negative interactions of the j-th computer is updated. After the local model of the j-th computer completes the training of the set number of rounds, the proof information generated by the j-th computer using the proof key is sent to the verifier.

3. The data collaboration method based on a verifiable mechanism according to claim 2, characterized in that, After setting the current aggregation model as the final aggregation model, the process also includes: The reputation value of each computational is determined based on the number of positive interactions, the number of negative interactions, and the communication quality of each computational; the communication quality of each computational is the probability of successful data parameter transmission between each computational and the verifier. The verifier's reputation value is updated based on the reputation value of each calculator.

4. The data collaboration method based on a verifiable mechanism according to claim 3, characterized in that, The division of multiple participants in setting up the data collaboration task into one verifier and multiple computers specifically includes: The participant with the highest reputation value is designated as the verifier, and the other participants besides the verifier are designated as calculators.

5. The data collaboration method based on a verifiable mechanism according to claim 3, characterized in that, The calculator's reputation value is: T i→j =b i→j +au i→j ; Among them, T i→j Let b represent the reputation value of the j-th calculator. i→j , where represents the trust ratio of the j-th calculator, and u i→j Let represent the uncertainty ratio of the j-th calculator, and a represent the coefficient of the uncertainty ratio, a∈[0,1]; u i→j =1-q i→j ; Where α represents the number of positive interactions for the j-th calculator, β represents the number of negative interactions for the j-th calculator, κ represents the weighting coefficient for the number of positive interactions, η represents the weighting coefficient for the number of negative interactions, and q i→j This represents the communication quality of the j-th computer.

6. The data collaboration method based on a verifiable mechanism according to claim 5, characterized in that, The verifier's reputation value is: Where T represents the reputation value of the verifier, and n represents the number of the calculators.

7. The data collaboration method based on a verifiable mechanism according to claim 1, characterized in that, If the number of computers issuing parameter merging signals is greater than or equal to a set threshold, then the verifier merges the model parameters of the computers issuing the parameter merging signals to obtain the current aggregated model, specifically including: The average value of each parameter in the model parameters of the calculator that issued the parameter merging signal is used to replace the corresponding parameters in the current aggregate model to obtain the aggregate model with updated parameters.

Images