Identity verification method for trusted platform to support block chain

Abstract

The invention relates to an identity verification method for a trusted platform to support a block chain, which comprises the following steps: firstly, each participant registers on a block chain network through a node of the block chain, then a management node sets the number of consensus committee as A in advance, and broadcasts a training program and global model parameters to all nodes; locally constructing an initial model, training to obtain local model parameters, independently performing local model verification on the obtained local model parameters by using MAE, then broadcasting to the management node, performing common model verification on the local model parameters by the management node, then performing mixed verification, and finally performing local model verification. And the local model parameters of the mixed verification requirement are packaged into a block B'again and stored, the management node downloads the block B ', aggregates the local model parameters in the block B', takes an aggregation result as a new global model parameter, and finally performs A times of communication and selects A consensus committee. According to the invention, the security is high while data tracing is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of blockchain and relates to an identity authentication method for a trusted platform supporting blockchain. Background Art

[0002] Although using digital communication to transmit data can achieve traceability, real-time and cost savings, digital communication is more vulnerable to cyber attacks and poses security and privacy risks. Limited IoT computing resources make it difficult to apply complex security policies, and there is also the risk of data manipulation. In cloud platforms, there is also the additional risk of denial of service (DoS) attacks or IP spoofing.

[0003] Data leakage may occur during data storage, data transmission, and data sharing, which may cause serious problems for owners and suppliers. In this regard, existing works mainly focus on utilizing aggregate information about the data without compromising the privacy of participants. They address this issue by making some modifications to the key contributions of the original data, such as k-anonymity algorithms, L-diversity algorithms. But most of the methods assume that the attacker has only limited background knowledge, where the data is still vulnerable to algorithm-based attacks or background knowledge attacks.

[0004] In order to protect the data privacy of users while maintaining a useful AI model, Google proposed the concept of federated learning in 2017. The core idea of ​​federated learning is to establish a global model based on virtual fusion data by training a distributed model among multiple data sources with local data, without sharing local training data, only by exchanging model parameters or intermediate results to achieve a balance between data privacy protection and data sharing. Due to its privacy characteristics, federated learning has broad application prospects in smart cities, e-medicine, wireless communications, mobile edge networks and other fields. Ideally, the shared model obtained by federated learning is similar to or better than the model obtained by training data on the central server. Therefore, enterprises can legally and effectively integrate data to extract information, while individuals or other institutions holding data can still retain control over the data while enjoying the AI ​​services provided by the enterprise.

[0005] According to the partitioning method of data, federated learning can be divided into horizontal federated learning (HFL) and vertical federated learning (VFL). Although HFL protects data privacy to a certain extent, the HFL protocol still has the potential risk of leaking local private data information. Secure multi-party computation, homomorphic encryption (HE), and differential privacy (DP) are the three most common privacy protection mechanisms, which have been theoretically and empirically proven to be effective for HFL. Compared with HFL, VFL is more likely to appear in practical applications. The training data of participating clients in VFL have the same sample identity (ID) space, but different feature spaces. Current privacy-preserving VFL systems are built on the assumption that all data labels are stored on only one visitor or activity party, which is unrealistic in many practical applications.

[0006] Differential privacy is a key concept that protects individual privacy while allowing meaningful data analysis. It involves adding controlled noise to computations to prevent sensitive information from leaking. In federated learning, multiple parties collaborate to build a shared machine learning model without sharing the original data, and ensuring privacy is critical. Differential privacy techniques such as secure aggregation and local differential privacy play a key role in protecting privacy in federated learning environments. Differential privacy (DP) proposed by Dwork et al. in 2006 shows that even if an attacker has all the remaining tuples of sensitive data, the provable privacy guarantees published by database records do not cause a significant loss in query accuracy. Secure aggregation encrypts model updates before sharing them with a central server, preventing direct access to individual updates. Techniques such as homomorphic encryption and secure multi-party computation enable efficient privacy-preserving computations across distributed datasets. Practical applications of differential privacy in federated learning, such as in healthcare and finance, have demonstrated its importance in balancing data-driven insights with individual privacy rights. Geyer et al. proposed a client-based differential privacy federated learning framework that combines differential privacy and federated learning to protect the privacy of client data and enable model training and updating.

[0007] Node identity authentication is an important part of the federated learning model. The authentication protocol is a communication protocol implemented through cryptography and communication technology. It confirms the network identity of the communication object by securely transmitting identity information between two or more communication objects. Through the authentication protocol, data transmission subjects can confirm each other's network identity before sending private data. The key negotiated by the two communicating parties can ensure the security of data during transmission, greatly reducing the risk of private data being stolen and cracked by attackers. Authentication is a prerequisite for data access and is widely used as a key technology for data transmission and information security in scenarios such as the Industrial Internet of Things, Internet of Vehicles, smart healthcare, and cloud databases.

[0008] Blockchain forms a decentralized distributed ledger through cryptography and peer-to-peer communication, using smart contracts to implement programmable, automated, and trusted transactions. Its purpose is to ensure that data cannot be tampered with while achieving decentralized data sharing. The main direction of combining federated learning with blockchain research is to introduce blockchain to build a distributed trusted computing framework for federated learning, thereby improving the mutual trust between nodes and the credibility of the aggregation model. The reasons for introducing blockchain in federated learning can be summarized into two points. On the one hand, the participating nodes and communication processes in federated learning may be subject to different types of attacks. On the other hand, maliciously participating nodes or servers may actively launch attacks to undermine the security and fairness of federated learning.

[0009] Recently, some scholars have combined blockchain and federated learning to achieve security and privacy protection in federated learning. Traditional federated learning aggregates updates in a centralized manner. Once the central server fails or is attacked, the process of federated learning may not continue. To solve this problem, Reference 1 (A blockchain for auditable federated learning with trust and incentive, in: 2019 5th International Conference on Big Data Computing and Communications, BIGCOM, IEEE, 2019, pp.151-159.) uses blockchain to implement a local model parameter update scheme that cannot be tampered with. In more detail, it designs a federated learning architecture based on a blockchain peer-to-peer network, in which the MPT (Merkle Patricia Tree) of the structure globally stored in the blockchain. Reference 2 (A reliable and accountable privacy-preserving federated learning framework using the blockchain, in: Proceedings of the2019 ACM SIGSAC Conference on Computer and Communications Security, 2019, pp.2561–2563.) proposed a privacy-preserving framework based on blockchain, which eliminates the semi-honesty assumption of participants and adopts encryption technology to protect data privacy, but the parameter server still has the risk of single point failure. Reference 3 (Blockchained on-device federated learning, IEEE Commun. Lett. 24 (6) (2019)1279–1283.) introduced blockchain into federated learning, solved the single point failure problem, and analyzed the optimal block generation rate. However, it ignored the privacy protection problem of intermediate parameters. Abali et al. suggested adding Laplace noise to the gradient of the training model. In addition, a collaborative learning system based on additive homomorphic encryption was proposed. However, due to the inability to achieve a balance between security and efficiency, the expected results were minimal.

[0010] The issue of privacy protection in federated learning remains to be studied. Currently, mainstream research relies on technologies such as DP (Reference 4: Shuffled model of differential privacy in federated learning. In: International conference on artificial intelligence and statistics. PMLR; 2021, pp. 2521–9.), HE (Reference 5: Privacy-preserving federated learning based on multi-key homomorphic encryption. Int J Intell Syst 2022.), and SMC (Reference 6: Averifiable federated learning scheme based on secure multi-party computation. In: International conference on wireless algorithms, systems, and applications. Springer; 2021, pp. 198–209.). All of these technologies can be used to protect privacy security in federated learning. However, it should be noted that the use of these technologies increases the complexity of computation and communication, so a balance needs to be struck between privacy security and computational efficiency.

[0011] Reference 7 (Privacy-preserving deep learning. In: Proceedings of the 22ndACM SIGSAC conference on computer and communications security.2015, pp. 1310–21. ) implements a privacy-preserving federated learning framework by adaptively sharing a small portion of the shared model parameters and using different privacy protection mechanisms. However, this approach requires a trade-off between model accuracy and data privacy. Further trade-offs are made between model convergence performance and privacy protection level, and it is believed that the better the convergence performance, the lower the protection level. Using secret sharing technology and encryption schemes to protect the security of private data is a representative technology for implementing SMC. PFMLP, a privacy-preserving machine learning framework based on partially homomorphic encryption and federated learning, significantly improves model accuracy compared to differential privacy. Although the model training efficiency has been improved, the model calculation efficiency under actual massive data remains to be verified. Reference 8 ("Blockchain and federated learning for privacy-preserved data sharing in industrial IoT", IEEE Transactions on Industrial Informatics, vol. 16, no. 6, pp. 4177-4186, 2019.) and Reference 9 ("Privacy-preserving blockchain-based federated learning for traffic flow prediction", Future Generation Computer Systems, vol. 117, pp. 328-337, 2021.) apply local differential sensitivity technology to blockchain federated learning. By perturbing the original data with noise, the data diversity of industrial Internet and intelligent transportation is protected. Reference 10 ("Decentralized privacy using blockchain-enabled federated learning in fogcomputing", IEEE Internet of Things Journal, vol. 7, no. 6, pp. 5171-5183, 2020.) designs a hybrid identity mechanism based on signature and encryption protocols to prevent attackers from stealing blockchain data information stored in the model.However, this mechanism consumes a lot of computational overhead and is difficult to deploy in local devices when the number of federated learning iterations is too large.

[0012] Therefore, it is of great significance to study an identity authentication method for trusted platforms supporting blockchain to solve the above problems. Summary of the invention

[0013] The purpose of the present invention is to solve the problems existing in the prior art and provide an identity authentication method for a trusted platform supporting blockchain.

[0014] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0015] An identity authentication method for a trusted platform supporting blockchain, comprising the following steps:

[0016] (1) Each participant registers on the blockchain network through a blockchain node and obtains a unique user ID;

[0017] (2) The management node pre-sets the number of consensus committees to A;

[0018] (3) The management node broadcasts the training program and global model parameters to all nodes used by participants;

[0019] (4) After receiving the model parameters and training program, the participant uses the global model parameters and training program to construct an initial model locally, and uses local data to train the initial model to obtain a local model (i.e., a local model), thereby obtaining local model parameters; differential privacy technology is used during the training process;

[0020] (5) All participants will independently use MAE (mean absolute error) to verify the local model parameters, upload the local model parameters and local model verification results to the blockchain network through the node and broadcast them to the management node;

[0021] (6) The management node performs public model verification on the local model parameters, and combines the public verification results with the local model verification results to obtain a mixed verification result. The quality of the local model parameters is then judged based on the mixed verification result, and the members of the consensus committee are selected. The selected members of the consensus committee then form a consensus committee and assign a unique identification ID to the consensus committee. The identity information of the selected members of the consensus committee is verified by the management node through the zero-knowledge proof protocol (FFS zero-knowledge proof protocol) and the DH algorithm.

[0022] (7) The members of the consensus committee compare the results of local model verification and select the one with the best model quality to become the leader of the committee; the leader collects the local model parameters obtained by all participants in this round of communication and generates block B, which is sent to other members of the committee for mixed verification of the local model parameters in the block, and the local model parameters that pass the mixed verification requirements are sent to the leader. The leader packages all the local model parameters that pass the mixed verification requirements into block B' again, sends it to the blockchain network for storage, and the management node downloads block B' and aggregates the local model parameters in block B', and uses the aggregation result as the new global model parameter;

[0023] (8) Conduct A rounds of communication and select A consensus committees; each round of communication process includes steps (3) to (7);

[0024] Among them, during the aggregation process, the local model parameters in block II are hierarchically encrypted on the smart contract of the blockchain.

[0025] As the preferred technical solution:

[0026] As described above, an identity authentication method for a trusted platform supporting blockchain, during the training process, the specific process of using differential privacy technology is as follows:

[0027] Assume that in each round of communication, the number of participants is N, the number of nodes used by the participants is N, and N nodes are defined as: ;node For any node in P, node Local dataset , x i is the sample input, y i is the basic truth; the dataset of all participating nodes in federated learning is defined as D; participants pass through the node The global model parameters received are: , the purpose of federated learning is to train a global model , so that the loss function on D Minimize, where the global model is the model obtained by aggregating the parameters of all participants; and the loss function on D Minimize, is to make Minimize the loss function on ;

[0028] node exist The loss function on is as follows:

[0029] ;

[0030] in, Based on the global model The data loss function ;

[0031] Use the stochastic gradient descent algorithm (SGD) to update the parameters in the opposite direction of the gradient of the loss function relative to the model parameters to minimize the loss function on D;

[0032] The loss function on the parameter in the opposite direction of the gradient of the model parameters is expressed as:

[0033] ;

[0034] The model parameter formula after updating iteration is:

[0035] ;

[0036] in, Represents the learning rate, which is 10 -6 ~10 -1 , t is the number of update iterations; the model parameters after iterative update through local data are the local model parameters;

[0037] When t=1, = ;

[0038] In order to achieve privacy protection of parameters in the opposite direction of the gradient of the loss function relative to the model parameters, noise is added to the updated model parameter formula based on the Laplace mechanism of LDP;

[0039] The updated model parameter formula after adding noise is:

[0040] ;

[0041] in, is the Laplace noise, s is the local sensitivity, represents the cross entropy of model quality evaluation, is the dynamic privacy noise adjustment coefficient, which controls the noise size based on cross entropy. , Budget for privacy, , is the maximum number of update iterations, is the total budget constraint, ;

[0042] Differential privacy protection based on the Laplace mechanism reduces the risk of local data privacy leakage caused by inference attacks.

[0043] As described above, a method for authentication of trusted platforms supporting blockchains, local model validation using MAE refers to evaluating the quality of the model through MAE:

[0044] The formula for MAE to evaluate the quality of the model is:

[0045] ;

[0046] in, For Model The mean absolute error of the model means is the local model of the model parameters, For basic facts, For Model The predicted value of , b is the bias, is the sample input, The lower the model The higher the accuracy;

[0047] The process of public model verification is as follows: before the consensus committee is formed, a public verification set is defined , participants at the node Calculate the MAE of the local model on the public validation set:

[0048] The calculation formula of MAE on the public validation set is:

[0049]

[0050] in, The number of public validation sets is j, For Model The predicted value of is the ground truth in the public validation set;

[0051] The results of the hybrid verification are: ;in, is a weight parameter used to balance local verification results and public verification results.

[0052] As described above, an authentication method for a trusted platform supporting blockchain, in order to ensure the reliability of the identity of the participating nodes, we adopted an authentication scheme based on the zero-knowledge proof protocol and the DH algorithm; zero-knowledge proof is a method in which one party (the prover) can prove to another party (the verifier) ​​that they know the value x without passing any information other than that they know the value x. The essence of zero-knowledge proof is that it is trivial to prove that someone knows certain information by simply revealing certain information. The Feige Fiat Shamir (FFS) scheme uses a public key cryptosystem that requires only a small number of modular operations and is therefore faster than other public key algorithms. In addition, FFS is suitable for scenarios where the computing resources of all participants are limited, such as federated learning. Therefore, this application selects the FFS zero-knowledge proof protocol, which provides a certain guarantee for not providing the leakage of user identity privacy information. In addition, the DH key exchange algorithm is used as the key agreement mechanism, which involves modular power operations and has high computational complexity, and it can provide more reliable guarantees. All nodes participating in federated learning need to register before joining the blockchain to obtain the public parameter information required by the above authentication method. The registration process is as follows:

[0053] (1) Set a system security parameter .

[0054] (2) Calculate the interaction parameters of the DH algorithm , where a and q are system preset parameters.

[0055] (3) Generate a series of local parameters, including a random number r, a random symbol number s (s is −1 or 1), .

[0056] (4) Set a large integer m and calculate , used to form the public parameter information of zero-knowledge proof.

[0057] (5) Component registration information registration = Broadcast to the blockchain network, id is the node identification information.

[0058] The authentication process of user P and the consensus committee is as follows:

[0059] (1) User P sends a request message to DTC to query and obtain the public parameter information of DTC (distributed trusted consensus) member registration. This information is sent by the node Zero-knowledge proof of public parameters and DH algorithm parameters of Then, P will generate the user ID , Timestamp and random sequence values Send to .

[0060] (2) Node After receiving the message from user P, a binary string is randomly generated , whose value is 0 or 1, and sends the message to user P, including the digital string and node id , Timestamp and serial number .

[0061] (3) Node Request to obtain the registration information of user P.

[0062] (4) User P obtains the node The binary string sent After that, according to the random number r and Calculate the zero-knowledge proof parameter information and set the current timestamp and sequence value N3 is sent to the node .

[0063] (5) User P calculates the public Then encrypt the information with K and send it to .

[0064] (6) Node After receiving the information sent by P, the identity of P is verified based on the zero-knowledge proof protocol. If the verification is successful, it means that the identity of P is correct. Calculate the public key based on the registration information of user P obtained from the blockchain Then, decrypt with K The message is transmitted and encrypted, and finally, the encrypted information is sent to user P. After the authentication is completed, K is the session key used for data exchange.

[0065] As described above, an authentication method for a trusted platform supporting a blockchain, during the aggregation process, the model parameters are hierarchically encrypted on the smart contract of the blockchain, including Ciphertext aggregation of layers and bias layers;

[0066] for The ciphertext aggregation method of the layer is calculated as follows:

[0067] ;

[0068] in, is the total percentage of each participant participating in the model training set; represents the aggregation parameter of the weight layer, represents the encryption weight of each user;

[0069] For the bias layer, homomorphic calculation is required, and the calculation formula is as follows:

[0070] ;

[0071] in, represents the aggregation parameter of the bias layer, , represents homomorphic operation, represents the encrypted bias weight of each user. Since homomorphic algorithms can take up a lot of computing resources, we use layer to reduce computational overhead, and use Layers use different privacy algorithms to protect privacy;

[0072] This application implements the aggregation process of the parameters of the hierarchical encryption model of the smart contract in the blockchain network. This process is safe and is not affected by any individual user. This is the result of all users maintaining the entire blockchain network. The results are stored in the blockchain network. For specific implementation, see the algorithm:

[0073]

[0074] As described above, in an authentication method for a trusted platform supporting a blockchain, the members of the consensus committee compare the results of local model verification and select the one with the best model quality to become the committee leader. The specific process is as follows:

[0075] Voting is performed among committee members. Each member can vote for other members' models based on the results of pairwise comparison. Assuming that each committee member The voting weight is determined by the following formula:

[0076] ;

[0077] ;

[0078] Indicates committee members The total number of votes cast, For committee members Local verification results on For committee members The local verification result on . The smaller the value, the better.

[0079] Beneficial effects:

[0080] (1) The present invention provides an identity authentication method for a trusted platform supporting blockchain, which applies the authentication mechanism of a federated learning system supporting blockchain to fiber product traceability. While achieving data traceability, it also restricts access to data to protect data sharing security.

[0081] (2) The present invention provides an identity authentication method for a trusted platform supporting blockchain, which introduces a federated learning authorized consensus-PoQ protocol to reduce computing costs and communication resource costs and achieve efficient data sharing.

[0082] (3) The present invention provides an identity authentication method for a trusted platform supporting blockchain, which combines a blockchain network to establish a decentralized FL training model and uses a secure aggregation method to improve the security of the model training environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 A flowchart of an identity authentication method for a trusted platform supporting blockchain;

[0084] Figure 2 This is the schematic diagram of the consensus-PoQ protocol. DETAILED DESCRIPTION

[0085] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0086] An authentication method for a trusted platform supporting blockchain, such as Figure 1 As shown, the following steps are included:

[0087] (1) Each participant registers on the blockchain network through a blockchain node and obtains a unique user ID;

[0088] (2) The management node pre-sets the number of consensus committees to A;

[0089] (3) The management node broadcasts the training program and global model parameters to all nodes used by participants;

[0090] (4) After receiving the model parameters and training program, the participant uses the global model parameters and training program to construct an initial model locally, and uses local data to train the initial model to obtain a local model (i.e., local model), thereby obtaining local model parameters. In the training process, differential privacy technology is used, and the specific process is as follows;

[0091] Assume that in each round of communication, the number of participants is N, the number of nodes used by the participants is N, and N nodes are defined as: ;node For any node in P, node Local dataset , x i is the sample input, y i is the basic truth; the dataset of all participating nodes in federated learning is defined as D; participants pass through the node The global model parameters received are: , the purpose of federated learning is to train a global model , so that the loss function on D Minimize, where the global model is the model obtained by aggregating the parameters of all participants; and the loss function on D Minimize, is to make Minimize the loss function on ;

[0092] node exist The loss function on is as follows:

[0093] ;

[0094] in, Based on the global model The data loss function ;

[0095] Use the stochastic gradient descent algorithm (SGD) to update the parameters in the opposite direction of the gradient of the loss function relative to the model parameters to minimize the loss function on D;

[0096] The loss function on the parameter in the opposite direction of the gradient of the model parameters is expressed as:

[0097] ;

[0098] The model parameter formula after updating iteration is:

[0099] ;

[0100] in, Represents the learning rate, which is 10 -6 ~10 -1 , t is the number of update iterations;

[0101] When t=1, = ;

[0102] Based on the Laplace mechanism of LDP, noise is added to the updated model parameter formula;

[0103] The updated model parameter formula after adding noise is:

[0104] ;

[0105] in, is the Laplace noise, s is the local sensitivity, represents the cross entropy of model quality evaluation, is the dynamic privacy noise adjustment coefficient, , Budget for privacy, , is the maximum number of update iterations, is the total budget constraint, ;

[0106] (5) All participants will independently use MAE (mean absolute error) to verify the local model parameters, upload the local model parameters and local model verification results to the blockchain network through the node and broadcast them to the management node;

[0107] Local model validation using MAE means evaluating the quality of the model through MAE:

[0108] The formula for MAE to evaluate the quality of the model is:

[0109] ;

[0110] in, For Model The mean absolute error of the model means is the local model of the model parameters, For basic facts, For Model The predicted value of , b is the bias, is the sample input, The lower the model The higher the accuracy;

[0111] The process of public model verification is as follows: before the consensus committee is formed, a public verification set is defined , participants at the node Calculate the MAE of the local model on the public validation set;

[0112] The calculation formula of MAE on the public validation set is:

[0113] ;

[0114] in, The number of public validation sets is j, For Model The predicted value of is the ground truth in the public validation set;

[0115] The results of the hybrid verification are: ;in, is the weight parameter;

[0116] (6) The management node performs public model verification on the local model parameters, and combines the public verification results with the local model verification results to obtain a mixed verification result. The quality of the local model parameters is then judged based on the mixed verification result, and the members of the consensus committee are selected. The selected members of the consensus committee then form a consensus committee and assign a unique identification ID to the consensus committee. The identity information of the selected members of the consensus committee is verified by the management node through the zero-knowledge proof protocol (FFS zero-knowledge proof protocol) and the DH algorithm.

[0117] (7) Figure 2 As shown, the members of the consensus committee compare the results of local model verification and select the one with the best model quality to become the committee leader; the leader collects the local model parameters obtained by all participants in this round of communication and generates block B, and sends block B to other members of the committee to perform mixed verification on the local model parameters in the block, and sends the local model parameters that pass the mixed verification requirements to the leader. The leader packages all the local model parameters that pass the mixed verification requirements into block B' again, sends it to the blockchain network for storage, and the management node downloads block B', aggregates the local model parameters in block B', and uses the aggregation results as the new global model parameters;

[0118] (8) Conduct A rounds of communication and select A consensus committees; each round of communication process includes steps (3) to (7);

[0119] In the aggregation process, the local model parameters in Block II are encrypted hierarchically on the smart contract of the blockchain, including Ciphertext aggregation of layers and bias layers;

[0120] for The ciphertext aggregation method of the layer is calculated as follows:

[0121] ;

[0122] in, is the total percentage of each participant participating in the model training set; represents the aggregation parameter of the weight layer, represents the encryption weight of each user;

[0123] For the ciphertext aggregation method of the bias layer, homomorphic calculation is required, and the calculation formula is as follows:

[0124] ;

[0125] in, represents the aggregation parameter of the bias layer, , represents homomorphic operation, Represents the encrypted bias weight of each user.

Claims

1. An identity authentication method for a trusted platform supporting blockchain, characterized in that: The steps include: (1) Each participant registers on the blockchain network through a blockchain node and obtains a unique user ID; (2) The management node pre-sets the number of consensus committees to A; (3) The management node broadcasts the training program and global model parameters to all nodes used by participants; (4) After receiving the model parameters and training program, the participant uses the global model parameters and training program to construct an initial model locally, and uses local data to train the initial model to obtain a local model, thereby obtaining local model parameters; differential privacy technology is used during the training process; (5) All participants will independently use MAE to verify the local model parameters obtained, upload the local model parameters and local model verification results to the blockchain network through the node and broadcast them to the management node; (6) The management node performs public model verification on the local model parameters, and combines the public verification results with the local model verification results to obtain a mixed verification result. The quality of the local model parameters is then judged based on the mixed verification result, and the members of the consensus committee are selected. The selected members of the consensus committee then form a consensus committee and assign a unique identification ID to the consensus committee. The identity information of the selected members of the consensus committee is verified by the management node through the zero-knowledge proof protocol and the DH algorithm. (7) The members of the consensus committee compare the results of local model verification and select the one with the best model quality to become the leader of the committee; the leader collects the local model parameters obtained by all participants in this round of communication and generates block B, which is sent to other members of the committee for mixed verification of the local model parameters in the block, and the local model parameters that pass the mixed verification requirements are sent to the leader. The leader packages all the local model parameters that pass the mixed verification requirements into block B' again, sends it to the blockchain network for storage, and the management node downloads block B' and aggregates the local model parameters in block B', and uses the aggregation result as the new global model parameter; (8) Conduct A rounds of communication and select A consensus committees; each round of communication process includes steps (3) to (7); Among them, during the aggregation process, the local model parameters in block II are hierarchically encrypted on the smart contract of the blockchain.

2. The identity authentication method for a trusted platform supporting blockchain according to claim 1, characterized in that: During the training process, the specific process of using differential privacy technology is as follows: Assume that in each round of communication, the number of participants is N, the number of nodes used by the participants is N, and N nodes are defined as: ;node For any node in P, node Local dataset , x i is the sample input, y i It is a basic fact; The dataset of all participating nodes in federated learning is defined as D; participants pass through nodes The model parameters received are: ; node exist The loss function on is as follows: ； in, It is based on the model The data loss function ; The loss function on the parameter in the opposite direction of the gradient of the model parameters is expressed as: ； The updated model parameter formula is: ； in, Represents the learning rate, which is 10 -6 ~10 -1 , t is the number of updates; When t=1, = ; Based on the Laplace mechanism of LDP, noise is added to the updated model parameter formula; The updated model parameter formula after adding noise is: ； in, is the Laplace noise, s is the local sensitivity, represents the cross entropy of model quality evaluation, is the dynamic privacy noise adjustment coefficient, , Budget for privacy, , is the maximum number of updates, is the total budget constraint, .

3. The identity authentication method for a trusted platform supporting blockchain according to claim 1, characterized in that: Local model validation using MAE means evaluating the quality of the model through MAE: The formula for MAE to evaluate the quality of the model is: ； in, For Model The mean absolute error of the model means is the local model of the model parameters, For basic facts, For Model The predicted value of , b is the bias, is the sample input; The process of public model verification is as follows: before the consensus committee is formed, a public verification set is defined , participants at the node Calculate the MAE of the local model on the public validation set: The calculation formula of MAE on the public validation set is: in, The number of public validation sets is j, For Model The predicted value of is the ground truth in the public validation set; The results of the hybrid verification are: ;in, is the weight parameter.

4. The identity authentication method for a trusted platform supporting blockchain according to claim 1, characterized in that: During the aggregation process, the model parameters are encrypted in layers on the blockchain smart contract, including Ciphertext aggregation of layers and bias layers; For the ciphertext aggregation method of the weight layer, the calculation formula is: ； in, is the total percentage of each participant participating in the model training set; represents the aggregation parameter of the weight layer, represents the encryption weight of each user; For the ciphertext aggregation method of the bias layer, homomorphic calculation is required, and the calculation formula is as follows: ； in, represents the aggregation parameter of the bias layer, , represents homomorphic operation, Represents the encrypted bias weight of each user.

Images