Federal learning CIM system information security protection method based on block chain and TEE

Abstract

The invention relates to the field of smart city management and information security, and provides a federated learning CIM system information security protection method based on a block chain and a TEE, and the method achieves the security sharing and cooperative computing of multi-party data through combining the data tamper resistance of the block chain technology and the security computing capability of the TEE, and improves the security of the multi-party data. And the data privacy and the security of the model training process are ensured. According to the core technical scheme, the method comprises the steps of deploying an intelligent contract in a block chain network, and managing data access permission; local model training and security aggregation are carried out in a trusted execution environment, and global model parameters are protected through a differential privacy technology. The method can be used for traffic management, environment monitoring, energy optimization and other scenes in a smart city, and provides efficient, safe and intelligent decision support for a city management system.

Description

Technical Field

[0001] The present invention relates to the fields of smart city management and information security. Specifically, a method for information security protection of a federated learning urban information model system based on blockchain and trusted execution environment is proposed. Background Art

[0002] With the deepening of smart city construction, the City Information Model (CIM) has gradually become the core framework of smart city management. CIM integrates multi-source heterogeneous urban data (such as traffic, energy, environment, and public safety data) to support real-time monitoring, predictive analysis, and intelligent decision-making. However, the traditional CIM data processing methods have the following main problems: 1. Data privacy and sharing issues: Centralized data storage and analysis methods face the risk of data privacy leakage, especially when sensitive data (such as citizen locations, behavior patterns, etc.) are stored and processed through a single central point.

[0003] 2. Trust issues in data storage and management: Centralized data storage architectures have problems such as data tampering, loss, and non-traceability, making it difficult to guarantee the authenticity and integrity of data sources.

[0004] 3. Computational resource pressure and data silo phenomenon: Urban management systems in different departments or regions usually cannot share data due to reasons such as laws, regulations, and data sovereignty, resulting in serious data silo problems and restricting the intelligent level of smart city systems.

[0005] 4. Model security and generalization ability deficiencies: Existing centralized model training methods are difficult to handle data differences between different regions and are vulnerable to threats from malicious attackers, reducing the security and generalization ability of the model. Summary of the Invention

[0006] In view of the deficiencies of the prior art, the present invention provides a method for information security protection of a federated learning CIM system based on blockchain and TEE, aiming to solve the problems of insufficient data privacy protection, poor computational security, and data silo problems in existing smart city management systems. This method combines the distributed ledger of blockchain technology, the isolated computing of trusted execution environment, and the decentralized model training of federated learning to achieve trusted sharing of data, secure training and updating of models, and efficient privacy protection in the urban information model system.

[0007] The method of the present invention mainly includes the following three core contents: 1. Data sharing and trusted management mechanism based on blockchain To address the privacy risks and data tampering issues brought about by centralized data storage, the present invention applies blockchain technology to the data sharing and management of the City Information Model system. Through the distributed ledger, consensus mechanism, and smart contracts of the blockchain, trustworthy records of data sources and the traceability and immutability of data sharing operations are achieved.

[0008] 1.1 Data Encryption and Upload Each data holder (such as urban management departments, public service units) first encrypts the local data to generate corresponding data digests, and uploads the data digests and smart contracts to the blockchain network together. The smart contract defines the access rights and usage rules of the data, ensuring the transparency and compliance of the data sharing process. Specifically, it includes the following steps: (1) Local encryption processing. Each data holder executes a symmetric or asymmetric encryption algorithm to obtain encrypted data and the corresponding digest information .

[0009] (2) Upload to the blockchain. Write the data digest and related access policies (expressed through smart contracts) into the blockchain ledger, use the smart contract to verify the access rights, and verify the integrity and authenticity of the data source through zero-knowledge proof technology.

[0010] 1.2 Smart Contract-Driven Data Transaction and Verification Through smart contracts, cross-departmental data transaction and verification processes can be realized. If a certain department needs to access the data of other departments, a data request transaction is triggered through the smart contract. Nodes in the blockchain network perform consensus verification on this request to confirm the legality of the transaction. After the verification passes, the data can be transmitted or authorized to access in a point-to-point secure channel. In addition, the smart contract can also automatically execute data usage rules (such as charging policies, access time limits, etc.) to ensure the compliance and security of the data interaction process.

[0011] 1.3 Traceability of Data Operations The blockchain records every data upload, sharing, access, or update operation in the distributed ledger. Due to the immutable and distributed characteristics of the blockchain, the operation history of each piece of data can be traced based on the transaction records later, thus effectively guaranteeing the integrity and authenticity of the data.

[0012] 2. Secure Federated Learning Model Training and Update within a Trusted Execution Environment To avoid the data leakage risk and computing security issues brought by centralized model training, the present invention proposes a multi-level federated learning strategy based on a trusted execution environment. Each data holder executes model training and update in its local TEE environment, maximizing the privacy and security of data during the training process.

[0013] 2.1 Local model initialization and training Each data holder initializes the model parameters using local data in its TEE environment and performs local training according to the loss function. Common gradient descent methods (such as SGD) or their variants (such as Adam) can be used to iteratively update the model parameters.

[0014] 2.2 Encrypted gradient calculation and upload After completing local training, each holder calculates the model gradient for the current round and encrypts the gradient using homomorphic encryption technology to obtain the encrypted gradient . Subsequently, the holder uploads the encrypted gradient to the central coordination node, which is also in a trusted execution environment and can perform decryption and aggregation operations in a secure environment.

[0015] 2.3 Global model update based on weighted average After collecting and decrypting all the encrypted gradients uploaded by the holders in its TEE, the central coordination server performs a secure aggregation process to generate global model parameters. The previous round of global model parameters is saved by the central coordination node after completing the previous round of aggregation, stored on the blockchain as part of the system state, and read each time the model is updated.

[0016] 2.4 Multi-level federated learning strategy According to the data characteristics of different regions or departments, a multi-level federated learning hierarchical structure is designed. For example: Regional-level federated learning. First, perform multiple iterations of model training and aggregation within the same region or department to obtain a preliminary "regional model".

[0017] Global-level federated learning. Upload the regional models to the central coordination server for secondary aggregation on a wider scale to obtain the global model.

[0018] This hierarchical aggregation method can not only reduce the cross-regional data communication volume but also improve the generalization ability of the model while ensuring privacy.

[0019] 3. Model optimization mechanism based on differential privacy and optimization algorithms To further protect the privacy of the holders' data and improve the performance of the model, the present invention introduces differential privacy and multiple optimization strategies in the aggregation and update process of the global model to ensure that the model can converge quickly without leaking sensitive information.

[0020] 3.1 Differential Privacy Protection Mechanism After each round of global model aggregation is completed, the central coordination node injects noise into the global model parameters within its TEE environment to meet the requirements of differential privacy. Since TEE can protect the generation and addition process of noise from being monitored by the outside world, it further enhances the security of differential privacy protection.

[0021] 3.2 Dynamically Optimized Adaptive Learning Rate Adjustment In federated learning, different rounds may require different learning rates or regularization parameters to adapt to the dynamic changes in data distribution and model convergence speed. The central coordination server can adopt an adaptive strategy to and perform fine-tuning within the TEE according to the loss change trend of each round of training.

[0022] 3.3 Model Pruning and Quantization Optimization In order to reduce the resource consumption of the model during the transmission, deployment, and inference phases, the present invention introduces model pruning and quantization techniques after the global model training is completed: Model Pruning, removing neurons or weights that have little impact on prediction to reduce the scale of the network structure.

[0023] Model Quantization, converting floating-point weights to low-precision representations (such as 8-bit integers) to reduce the model storage requirements and computational overhead without significantly sacrificing accuracy.

[0024] Since the above operations can be executed or managed in the TEE environment, it can effectively achieve the comprehensive optimization of model scale and accuracy while ensuring security.

[0025] The specific technical solution of the present invention is as follows: An information security protection method for a federated learning CIM system based on blockchain and TEE, the method includes the following: Each data holder collects various data including real-time traffic data, energy utilization efficiency data, environmental monitoring data, and urban infrastructure operation data and preprocesses the data. The preprocessed data is stored in the local trusted execution environment in encrypted form and a data digest is generated, and the initial urban information model is initialized and trained using the local data; The consensus mechanism and smart contract of the blockchain network are used to verify and record the encrypted local model gradients uploaded by all parties through zero-knowledge proof technology. After the triggering conditions are met, the smart contract performs secure multi-party computation aggregation on the encrypted gradients in the trusted execution environment of the central coordination node to generate global model parameters. The smart contract triggering conditions include reaching a preset threshold of the number of participating parties, time interval, or training rounds; Each data holder receives the global model parameters through the smart contract and performs further model training in the local trusted execution environment in combination with the latest local data to calculate new local model gradients; Each data holder encrypts the locally computed model gradients through a secure multi-party computation protocol and uploads the encrypted gradient information through a secure channel in the blockchain. The central coordination node decrypts it within the trusted execution environment and aggregates the updated encrypted gradients of all parties and combines them with the previous round of global model to generate a new global model; The updated global model is distributed to each data holder through the smart contract, and the processes of gradient calculation, upload, aggregation, and model update are repeated until the global model parameters converge within the set accuracy range, completing the optimization and security protection of the city information model; The optimized city information model system is applied to multiple smart city scenarios including traffic flow optimization, smart grid dispatching, environmental quality monitoring, and urban emergency management to achieve real-time monitoring and decision support.

[0026] Furthermore, gradient descent method is used for local training, and the update formula is: ; Where: is the learning rate, is the gradient of the loss function with respect to the model parameters, is the regularization coefficient, is the updated local model parameters, is the local model parameters of the previous round.

[0027] Even further, the global model parameter update formula is as follows: ; Where: is the updated global model parameters, is the number of holders participating in federated learning; is the holder 's data weight; is the learning rate; is the regularization coefficient; is the global model parameters of the previous round.

[0028] Furthermore, during the training process, according to the loss change trend of each round of training, an adaptive strategy is adopted to and for adjustment: ; Among them: and are the adjustment factors of the learning rate and the regularization coefficient respectively, used to dynamically adjust the learning rate according to the model convergence speed, is the number of iteration rounds.

[0029] Furthermore, after the global model training is completed, the model pruning and quantization technology is introduced to remove neurons or weights with less impact on prediction, so as to reduce the scale of the network structure.

[0030] Furthermore, after the global model training is completed, the model quantization technology is introduced to convert the floating-point weights into low-precision representations, reducing the model storage requirements and computational overhead without significantly sacrificing accuracy.

[0031] Furthermore, the model parameter gradient aggregation calculation adopts a multi-level encryption strategy based on homomorphic encryption, combined with a secure multi-party computation protocol, to ensure that the data of the participating parties cannot be reverse-inferred during the aggregation process.

[0032] Furthermore, during the process of uploading the model parameter gradients, data transmission is carried out through the encrypted communication channel in the blockchain, combined with digital signatures, to ensure the integrity and source verifiability of the gradient information transmission.

[0033] Furthermore, the aggregation of the global model parameters adopts a hybrid strategy of multiple optimization algorithms including weighted average, stochastic gradient descent, and gradient boosting trees, so as to improve the generalization ability and robustness of the model.

[0034] Furthermore, a provably secure differential privacy algorithm is used to perturb the data during the model training and parameter aggregation processes, and then the processed model parameters are uploaded through the secure channel on the blockchain; the added differential privacy noise comes from a Gaussian distribution or a Laplace distribution to prevent reverse-inference of the original data held by the holder from the uploaded parameters.

[0035] Compared with the prior art, the beneficial effects of the present invention are: (1) Equal emphasis on privacy protection and computational security. Sensitive data is left in the local TEE environment for training, and the model update process is protected through homomorphic encryption, differential privacy, etc., minimizing the risk of data leakage.

[0036] (2) Trusted traceability and compliant sharing. Through the distributed ledger and smart contracts of the blockchain, strict identity authentication and log retention are carried out for data operations and access, realizing traceable and tamper-proof trusted data sharing.

[0037] (3) A flexible and scalable federated learning framework that supports multi-level federated learning modes (regional level, global level), and can reduce the resource overhead of the model during actual deployment through techniques such as pruning and quantization.

[0038] (4) Applicable to a variety of smart city applications, such as traffic flow prediction, environmental monitoring, energy consumption optimization, etc. Under the premise of ensuring data sovereignty and security, it collaboratively mines more urban data value.

[0039] In summary, the information security protection method for the urban information model system proposed by the present invention, which combines blockchain technology, trusted execution environment (TEE) and federated learning method, can effectively address the contradiction between smart city data sharing and privacy protection. This method combines the distributed ledger characteristics of blockchain and the secure computing ability of TEE to ensure the privacy, security and data integrity of the urban information model, and is applicable to smart city application scenarios such as urban traffic optimization, environmental monitoring, energy management and public safety, providing a reliable and efficient technical approach for the intelligent decision-making of urban management. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a data exchange flow chart of the present invention based on blockchain and trusted execution environment.

[0041] Figure 2 It is an architecture diagram of the federated learning urban information model system of the present invention based on blockchain and trusted execution environment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] Hereinafter, the exemplary embodiments of the present invention will be described in more detail with reference to the drawings. Although the exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be fully conveyed to those skilled in the art. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other. Hereinafter, the present invention will be described in detail with reference to the drawings and in combination with the embodiments.

[0043] The information security protection method for the federated learning urban information model system of the present invention based on blockchain and trusted execution environment aims to solve key problems such as data privacy protection, model training security and cross-departmental data sharing in the urban information model system. The following specific implementation steps can be implemented and applied in the multi-department collaborative environment of smart cities.

[0044] Figure 1A process for secure data exchange using blockchain and trusted execution environment is described. The user requests data, and both parties sign a smart contract through the blockchain and encrypt the data; the sharing party uploads the encrypted data, the system processes the data in the TEE, records the operation logs through the blockchain, and finally returns the result to the user for decryption and use. The whole process ensures the security, privacy and credibility of data exchange.

[0045] Figure 2 A federated learning architecture combining blockchain and trusted execution environment is shown, which is used to guarantee data privacy and security in the smart city information model system. The architecture includes: data collection and preprocessing, encrypted distributed storage, realizing data access control and sharing through blockchain smart contracts; performing local model training and secure encrypted parameter aggregation in the TEE, and protecting the global model through differential privacy; finally applying the optimized model to smart city scenarios such as urban traffic optimization, environmental monitoring, energy management, etc., to achieve secure data sharing and efficient computing.

[0046] The technical solution of the present invention includes the following content: 1. Collection and distributed storage Each urban management department (such as transportation, energy, environmental monitoring, etc.) conducts local collection and storage of massive urban data within its respective region to ensure data sovereignty and privacy security. Specifically, it includes: (1) Data collection, obtaining multi-dimensional data from multiple data sources (such as traffic flow sensors, energy consumption meters, air quality monitoring stations, etc.) to form a local dataset. Let represent the dataset held by the holder collected.

[0047] (2) Data preprocessing, performing data cleaning (such as removing noise and outliers), feature engineering (such as normalization, feature encoding, etc.) and formatting processing on to ensure data quality and format consistency.

[0048] (3) Distributed storage and encryption, each holder stores the preprocessed dataset in an encrypted form in a local trusted environment (such as a local database or private cloud), and generates a data digest for integrity verification. A symmetric encryption algorithm can be used to protect the data. Let

[0049] where is the symmetric encryption key, is the encrypted data. At the same time, a digest is generated through a hash function., and stores the summary in a smart contract on the blockchain. In the subsequent data sharing and verification process, the smart contract can automatically verify the integrity and source credibility of the shared data, which is used for data exchange request approval and data usage permission control.

[0050] 2. Data sharing and access control on blockchain platforms Under the premise of ensuring data privacy, the present invention uses blockchain technology to build a cross-departmental data sharing and access control platform. The specific process is as follows: (1) Blockchain network deployment: building a consortium blockchain network with the participation of multiple city management departments. Each department node can participate in the blockchain consensus and ledger storage. The distributed ledger of the blockchain can ensure the immutability and traceability of data operations.

[0051] (2) Definition and execution of smart contracts: deploy smart contracts on the blockchain to record and execute data access rights and sharing rules. For example, in a smart contract, it can be pre-defined which department nodes have the right to access a certain type of data, and what conditions must be met (such as authorization, payment, or other policies) to obtain the data.

[0052] (3) Data access request and verification: If a management department node needs to access data from other departments, it first initiates a request to the smart contract through the blockchain, and the blockchain nodes collaborate to complete the request legitimacy verification. After the verification is passed, the smart contract triggers peer-to-peer data sharing or access authorization, and uses end-to-end encryption to transmit data, thereby preventing the risk of data leakage during transmission.

[0053] 3. Local model training and parameter protection in a trusted execution environment In order to reduce the privacy and security issues that may arise from external data circulation, each holder completes model training and protection of sensitive parameters in a local trusted execution environment. The main steps are as follows: (1) Local model initialization, each data holder Initialize model parameters in its TEE , appropriate machine learning models (such as neural networks, decision trees, clustering models, etc.) can be selected for different tasks (such as traffic flow prediction, energy consumption analysis, etc.).

[0054] (2) Local training and parameter update for local datasets Use gradient descent or other optimization algorithms to update model parameters. is the learning rate, is the regularization coefficient, and local training is performed at The iteration formula during round update can be expressed as:

[0055] where is the learning rate; is the local loss function of the holder ; is the gradient of the loss function with respect to the parameter ; is the regularization coefficient, which is used to prevent the model from overfitting.

[0056] (3) Gradient Encryption and Upload: After local training is completed, each holder encrypts the model gradient calculated locally based on secure multi-party computation (MPC) to obtain the encrypted gradient , and uploads it to the corresponding central coordinator node in the blockchain network. Since these operations are performed in the TEE, it can effectively prevent external eavesdropping or malicious tampering.

[0057] 4. Blockchain-Driven Secure Aggregation and Global Model Update With the support of the blockchain, the central coordinator decrypts the encrypted model parameter encrypted gradients uploaded by each holder in the TEE environment, and then uses the weighted aggregation algorithm to calculate the global model parameter . If what each holder uploads is the locally updated model parameter, then it can be set as:

[0058] Here is the total number of participants, is the data weight of the holder , which is determined by the data volume or data quality.

[0059] Each data holder receives the global model parameter through the smart contract, performs further model training by combining the latest local data in the local trusted execution environment, and calculates the new local model gradient; each data holder encrypts the locally calculated model gradient through the secure multi-party computation protocol and uploads the encrypted gradient information through the secure channel in the blockchain. The central coordinator completes the decryption in the trusted execution environment, aggregates the encrypted gradients updated by each party using the secure multi-party computation protocol, and generates a new global model by combining the previous round of global model. The global model parameter update formula is as follows After the central coordination server collects and decrypts all the encrypted gradients uploaded by the holders in its TEE, it executes the secure aggregation process to generate the global model parameter . The aggregation formula is as follows:

[0060] Where: is the updated global model parameter, is the number of holders participating in federated learning; For the holder is the data weight (usually proportional to the data volume or data quality); is the learning rate; is the regularization coefficient; is the global model parameter of the previous round.

[0061] The global model parameter of the previous round is saved by the central coordinator after completing the aggregation of the previous round, stored on the blockchain as part of the system state, and read every time the model is updated.

[0062] To further prevent the original data of any holder from being inferred from the global model, differential privacy noise can be injected into Let be (Gaussian distribution) or (Laplace distribution), then the global model after adding differential privacy noise can be expressed as:

[0063] The addition of differential privacy noise is also completed in the TEE to prevent external knowledge of the noise distribution or numerical information, thus maximizing data privacy.

[0064] 5. Dynamic Parameter Optimization and Local Model Update After the holder obtains the latest global model parameters from the blockchain, it can perform dynamic optimization and update in combination with local conditions: (1) Adaptive learning rate and regularization adjustment. To adapt to the model convergence speed and data distribution changes in different rounds, the central coordinator can dynamically adjust the learning rate and the regularization coefficient in the TEE for the next round of federated learning. For example:

[0065] where is the learning rate adjustment factor, is the regularization coefficient adjustment factor (the adjustment factor can be a decay coefficient less than 1 or an amplification coefficient greater than 1), is the iteration round. Through this dynamic adjustment, the convergence speed can be accelerated in the early stage of training, and the learning rate can be reduced in the later stage to prevent overfitting or oscillation.

[0066] (2) Model pruning and quantization. Each holder can perform model pruning and model quantization operations: pruning parameters or nodes with less influence, and converting the model weights from floating-point numbers to low-precision representations (such as 8-bit integers), significantly reducing storage and computational overhead.

[0067] 6. Model Deployment and Iterative Optimization After multiple rounds of federated learning and dynamic optimization, the finally formed global model can be deployed in the actual business scenarios of urban management departments, such as traffic congestion prediction, environmental pollution warning, energy consumption analysis, etc. After deployment, as time goes by and the environment changes, each holder can continue to collect new data and continuously iterate and update the parameters and structure of the model according to real-time feedback: (1) Scenario-based deployment, applying the global model obtained from federated learning or the model fine-tuned by each local department to actual urban management tasks, and outputting prediction or decision results (such as traffic flow scheduling, environmental protection monitoring and warning, etc.).

[0068] (2) Continuous iterative optimization, the newly collected data during the actual operation process can be put back into local training again, and periodically conduct federated learning collaboration and update with other departments, so as to always maintain the accuracy and adaptability of the model in a dynamic environment.

[0069] 7. Security Monitoring and Auditing To ensure the transparency and security of the data sharing, model training, and parameter aggregation processes, the present invention also provides a monitoring and auditing mechanism based on blockchain: (1) Comprehensive logging and traceability, all key operations such as data upload, access authorization, model training, and aggregation are recorded in the blockchain distributed ledger. Due to the immutable feature of the blockchain, accurate traceability of the entire process can be achieved.

[0070] (2) Dynamic adjustment of security policies, managers can continuously improve or update the access policies of smart contracts and technical parameters such as differential privacy and encryption algorithms according to the on-chain audit results, combined with actual business requirements and security compliance requirements, to continuously ensure the overall security and privacy of the urban information model system.

[0071] The present invention has been described in detail through embodiments above, but the content described is only an exemplary embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. The protection scope of the present invention is defined by the claims. Those who utilize the technical solutions described in the present invention, or those skilled in the art inspired by the technical solutions of the present invention, within the essence and protection scope of the present invention, design similar technical solutions to achieve the above technical effects, or make equivalent changes and improvements to the application scope, etc., should still fall within the patent coverage protection scope of the present invention. It should be noted that for the sake of clear description, the description of the present invention omits the description of some components and processes that are not directly and obviously related to the protection scope of the present invention but are known to those skilled in the art.

Claims

1. A method for information security protection of a federated learning CIM system based on blockchain and TEE, characterized in that The method includes the following: Each data holder collects various data including real-time traffic data, energy utilization efficiency data, environmental monitoring data, and urban infrastructure operation data, and preprocesses the data. The preprocessed data is stored in an encrypted form in a local trusted execution environment and a data digest is generated. The initial urban information model is initialized and trained using the local data; The consensus mechanism and smart contract of the blockchain network are used to verify and record the encrypted local model gradients uploaded by each party through zero-knowledge proof technology. After the triggering conditions are met, the smart contract performs secure multi-party computation aggregation on the encrypted gradients in the trusted execution environment of the central coordination node to generate global model parameters. The smart contract triggering conditions include reaching a preset number threshold of participants, time interval, or number of training rounds; Each data holder receives the global model parameters through the smart contract and performs further model training in the local trusted execution environment in combination with the latest local data to calculate new local model gradients; Each data holder encrypts the locally calculated model gradients through a secure multi-party computation protocol and uploads the encrypted gradient information through a secure channel in the blockchain. The central coordination node decrypts the information in the trusted execution environment and aggregates the encrypted gradients updated by each party using the secure multi-party computation protocol and combines them with the previous round of global model to generate a new global model; The updated global model is distributed to each data holder through the smart contract, and the processes of gradient calculation, upload, aggregation, and model update are repeated until the global model parameters converge within the set accuracy range, completing the optimization and security protection of the urban information model; The optimized urban information model system is applied to multiple smart city scenarios including traffic flow optimization, smart grid scheduling, environmental quality monitoring, and urban emergency management; 2. The method according to claim 1, wherein Local training uses the gradient descent method, and the update formula is: ； Wherein: is the learning rate, is the gradient of the loss function with respect to the model parameters, is the regularization coefficient, is the updated local model parameters, is the local model parameters of the previous round.

3. The method according to claim 2, wherein The global model parameter update formula is as follows: ； Wherein: is the updated global model parameter, is the number of holders participating in federated learning, is the holder data weight, is the learning rate, is the regularization coefficient, is the global model parameter of the previous round.

4. The method according to claim 2 or 3, characterized in that, During the training process, according to the loss change trend of each round of training, an adaptive strategy is adopted to and make adjustments: ； Wherein: and are the adjustment factors of the learning rate and the regularization coefficient respectively, used to dynamically adjust the learning rate according to the model convergence speed, is the number of iterations.

5. The method according to claim 1, characterized in that, After the global model training is completed, model pruning and quantization techniques are introduced; 6. The method according to claim 1, characterized in that After the global model training is completed, model quantization techniques are introduced to convert floating-point weights to low-precision representations; 7. The method according to claim 1, wherein The aggregation calculation of model parameter gradients adopts a multi-level encryption strategy based on homomorphic encryption; 8. The method according to claim 1, wherein During the process of uploading model parameter gradients, data is transmitted through the encrypted communication channel of the blockchain; 9. The method according to claim 1, wherein The aggregation of global model parameters adopts a hybrid strategy of multiple optimization algorithms including weighted average, stochastic gradient descent, and gradient boosting tree; 10. The method according to claim 1, characterized in that, The differential privacy algorithm is used to perturb the data during the model training and parameter aggregation processes, and then the processed model parameters are uploaded through the secure channel on the blockchain. The added differential privacy noise comes from a Gaussian distribution or a Laplace distribution.

Images