TEE and federated learning-based data sharing method and system
By using trusted execution environment and symmetric encryption algorithms in federated learning, the problems of low computing efficiency and large communication overhead caused by homomorphic encryption are solved, efficient and secure data sharing is achieved, and the accuracy and security of learning results are ensured. .
Patent Information
- Application Number
- CN202411305780.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-19
- Publication Date
- 2025-01-03
AI Technical Summary
In the existing federated learning scenarios, homomorphic encryption leads to low computing efficiency, large communication overhead, and noise accumulation leads to reduced learning result accuracy.
By combining symmetric encryption algorithms in a trusted execution environment, multiple execution nodes are established for local intelligent model training, and model parameters are encrypted and transmitted to the central aggregator for aggregation, forming a global intelligent model to ensure data privacy and security while reducing the number of data. Computational complexity and communication overhead.
An efficient and secure data sharing solution is realized, which improves the computational efficiency of federated learning and the accuracy of learning results, reduces communication overhead, and ensures the security and integrity of data.
Smart Images
Figure CN119248878A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of federated learning, and in particular to a data sharing method and system based on TEE and federated learning. Background Art
[0002] TEE, or Trusted Execution Environment, is an isolated and secure execution environment built by hardware and software. It allows code to be executed in an isolated, protected area without interference from the operating system or other software components, thereby providing a higher level of data protection and security. Federated learning is a distributed machine learning framework with privacy protection. It allows multiple devices or computing nodes to train models without sharing original data. Specifically, each participant uses local data to train the model locally, and sends the updated parameters of the model to the central server for aggregation, thereby building a more accurate global model.
[0003] In the current mainstream federated learning algorithms, differential privacy, secure multi-party computing, homomorphic encryption and other technologies are usually used to solve the security computing problems at different stages. In order to ensure the security of the parameter interaction calculation process in the joint modeling process, the homomorphic encryption algorithm is usually used to achieve the joint establishment of the prediction model. Although homomorphic encryption has certain advantages in the process of federated learning, such as providing significant privacy protection effects, it still has the following problems:
[0004] 1. Homomorphic encryption operations are much more complicated than plaintext operations. This means that when processing large-scale data or performing complex calculations, homomorphic encryption will significantly increase the demand for computing resources and processing time, ultimately causing model training to become extremely slow.
[0005] 2. The data after homomorphic encryption is usually much larger than the original data, which means that in the process of federated learning, a larger amount of data needs to be transmitted between participants, thereby increasing the communication overhead.
[0006] 3. In some homomorphic encryption schemes, noise will accumulate during the calculation process. If the noise is not managed properly, excessive noise accumulation may lead to a decrease in the accuracy of the final learning result or decryption failure.
[0007] In summary, the existing federated learning scenarios have technical problems such as low computing efficiency caused by the use of homomorphic encryption, high communication overhead, and noise accumulation leading to reduced accuracy of learning results. Summary of the invention
[0008] The present application provides a data sharing method and system based on TEE and federated learning, which is used to solve the technical problems in the prior art of federated learning scenarios, such as low computing efficiency, high communication overhead, and decreased accuracy of learning results caused by the use of homomorphic encryption.
[0009] In view of the above problems, this application provides a data sharing method and system based on TEE and federated learning.
[0010] In a first aspect, the present application provides a data sharing method based on TEE and federated learning, the method comprising:
[0011] Acquire multiple data participants, and establish multiple trusted execution environments for the multiple data participants;
[0012] Creating a plurality of execution nodes in the plurality of trusted execution environments;
[0013] According to the predetermined target intelligent model characteristics, the multiple execution nodes are used to train the intelligent models according to the local data samples of the multiple data participants, so as to obtain multiple local intelligent models;
[0014] Identify data sharing objects based on the data sharing requests of the multiple data participants to obtain a sharing identification participant group;
[0015] Sending multiple model parameters of multiple local intelligent models belonging to the shared recognition participant group to a central aggregator through the multiple execution nodes for aggregation to obtain a global intelligent model;
[0016] The global intelligent model is returned to the multiple data participants.
[0017] In a second aspect, the present application provides a data sharing system based on TEE and federated learning, the system comprising:
[0018] A trusted execution environment establishment module, used to obtain multiple data participants and establish multiple trusted execution environments for the multiple data participants;
[0019] An execution node creation module, used to create multiple execution nodes in the multiple trusted execution environments;
[0020] A model training module, used to train the intelligent model according to the local data samples of the multiple data participants using the multiple execution nodes according to the predetermined target intelligent model characteristics, so as to obtain multiple local intelligent models;
[0021] A sharing object identification module, used to identify data sharing objects based on the data sharing requests of the multiple data participants, and obtain a sharing identification participant group;
[0022] A global intelligent model acquisition module, used for sending multiple model parameters of multiple local intelligent models belonging to the shared recognition participating group to a central aggregator through the multiple execution nodes for aggregation to obtain a global intelligent model;
[0023] A return module is used to return the global intelligent model to the multiple data participants.
[0024] One or more technical solutions provided in this application have at least the following technical effects or advantages:
[0025] The data sharing method based on TEE and federated learning provided in the present application obtains multiple data participants and establishes multiple trusted execution environments for the multiple data participants; creates multiple execution nodes in the multiple trusted execution environments; according to the predetermined target intelligent model characteristics, the multiple execution nodes respectively train the intelligent model according to the local data samples of the multiple data participants to obtain multiple local intelligent models; based on the data sharing requests of the multiple data participants, the data sharing object is identified to obtain a shared identification participant group; through the multiple execution nodes, the multiple model parameters of the multiple local intelligent models belonging to the shared identification participant group are sent to the central aggregator for aggregation to obtain a global intelligent model; the global intelligent model is returned to the multiple data participants, which solves the technical problems in the prior art of low computational efficiency, high communication overhead, and reduced accuracy of learning results caused by the use of homomorphic encryption in the federated learning scenario, and realizes an efficient and secure data sharing solution, achieving the technical effect of ensuring high efficiency and reliability of data sharing without sacrificing security, and has the characteristics of low communication overhead and no sacrifice of accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 A flowchart of a data sharing method based on TEE and federated learning is provided for this application;
[0027] Figure 2 A schematic diagram of the data sharing system structure based on TEE and federated learning is provided for this application.
[0028] Explanation of the accompanying drawings: trusted execution environment establishment module 11, execution node creation module 12, model training module 13, shared object identification module 14, global intelligent model acquisition module 15, return module 16. DETAILED DESCRIPTION
[0029] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0030] Embodiment 1
[0031] This application provides a data sharing method based on TEE and federated learning. The core is to combine a trusted execution environment with a symmetric encryption algorithm to provide an efficient federated learning data sharing method to solve the problems of low computing efficiency, high communication overhead, and noise accumulation caused by the use of homomorphic encryption in federated learning scenarios, and further improve the security and efficiency of data sharing.
[0032] Homomorphic encryption is a form of encryption that allows certain types of computation to be performed on ciphertext without decrypting it. This means that operations can be performed on encrypted data while maintaining data privacy, and the result obtained is in encrypted form. When this result is decrypted, it is the same as the result of performing the same operation directly on the plaintext.
[0033] like Figure 1 As shown, the present application provides a data sharing method based on TEE and federated learning, the method comprising:
[0034] Step S100: Acquire multiple data participants, and establish multiple trusted execution environments for the multiple data participants;
[0035] Specifically, first, multiple participants who need to update local models through federated learning are obtained. These participants refer to multiple devices, organizations or data centers with local models that provide data foundation for the global intelligent model of federated learning of the central aggregator, and their local models and global models belong to the same learning type category, that is, they are aimed at processing the same type of data or solving the same data problem. For example, the participants in the same federated learning can be the personal data of different users of mobile devices, financial institutions with private data under financial platforms, different medical institutions with disease research or diagnosis programs in medical databases, research institutes or universities that may have specific data sets and expertise that are inconvenient to be made public, or different companies and enterprises that have competitive relationships in the same field but need to conduct overall industry analysis.
[0036] Furthermore, multiple trusted execution environments are established through multiple data participants, that is, each participant participating in the joint learning model training is equipped with an execution environment isolated from local data processing, which is used for subsequent encrypted communication and transmission of local model parameter data with the central aggregator. This execution environment is jointly constructed by hardware and software, which allows the code to be executed in an isolated and protected area without interference from the operating system or other software components. Therefore, the establishment of this trusted execution environment serves as a data interaction bridge and establishes a data transmission channel, which fully guarantees the privacy and security of local data, and also lays the foundation for subsequent global model aggregation processing, so that the central aggregator receives the training effect equivalent to the complete original data.
[0037] Step S200: creating multiple execution nodes in the multiple trusted execution environments;
[0038] Step S300: According to the predetermined target intelligent model features, the plurality of execution nodes are used to train intelligent models according to the local data samples of the plurality of data participants, to obtain a plurality of local intelligent models;
[0039] Furthermore, step S300 of the present application also includes:
[0040] Step S310: The predetermined target intelligent model features include model input features, model output features and model network features.
[0041] Optionally, each participant creates multiple execution nodes in its own trusted execution environment, which are used to run multiple independent and secure execution units or instances within the TEE, such as those set up based on different tasks, data or code segments to be processed, to ensure that each execution node runs in a secure and isolated environment, thereby preventing unauthorized access and tampering. By dividing multiple execution nodes, TEE can better manage and isolate different resources, improving the security, reliability and efficiency of the system.
[0042] Further, according to the characteristics of the predetermined target intelligent model, based on multiple execution nodes, the local data samples of the participant are used to train the local intelligent model, wherein the predetermined target intelligent model refers to the target model that multiple participants in federated learning need to obtain, and the characteristics of the target model are extracted to guide the construction of the local intelligent model. The predetermined target intelligent model characteristics include model input characteristics, model output characteristics and model network characteristics, and the model network characteristics refer to the structural characteristics of the model. That is, it can be understood that multiple participants do not directly share sample data, but only share model parameters trained by sample data, so the model structure of each participant remains consistent. For example, several institutions with similar businesses cooperate, and the characteristics that their local models need to learn and predict are similar, but the users and samples are different, so the predetermined target intelligent model characteristics are needed to guide the construction structure of the local intelligent model. Finally, based on the guidance of the target intelligent model characteristics, the execution nodes are used as the basis for network transmission, and the local data samples are used as training data to train and obtain the local intelligent model.
[0043] By building a local intelligent model, basic data can be provided for the subsequent encrypted transmission of model parameters, laying the foundation for federated learning and ensuring the feasibility of the solution.
[0044] Step S400: Identify data sharing objects based on the data sharing requests of the multiple data participants to obtain a sharing identification participant group;
[0045] Furthermore, step S400 of the present application also includes:
[0046] Step S410: parsing the data sharing requests of the multiple data participants to obtain the data exchange object of each data participant;
[0047] Step S420: Based on the data exchange object, data participants with two-way exchange requests are combined to obtain the shared identification participant group.
[0048] Exemplarily, when conducting federated learning, multiple participants in federated learning need to send data sharing request instructions to the aggregator. The aggregator identifies the data sharing objects sent by the multiple participants. When the object is determined to be included in the model parameter range of the global model, the participant sending the shared data is added to the shared identification participant group, which refers to the set of participants who provide data contributions to the aggregator model training.
[0049] Specifically, the data sharing request of the participants is parsed, the data sharing needs of the participants to which the shared data belongs are identified, the sharing needs of multiple participants in the aggregator are matched, and the data participants with two-way exchange needs are combined and divided into the same sharing identification group, wherein there can be multiple participants for sharing demand matching, and there can be multiple participants in the sharing identification group.
[0050] By dividing the shared recognition participants into groups, we can clarify the learning tasks, ensure the participation and completeness of the sample data, and thus improve the accuracy of the model.
[0051] Step S500: sending multiple model parameters of multiple local intelligent models belonging to the shared recognition participant group to a central aggregator through the multiple execution nodes for aggregation to obtain a global intelligent model;
[0052] Specifically, multiple execution nodes may correspond to multiple local intelligent models, and each local intelligent model has multiple model parameters. Multiple execution nodes will encrypt the multiple model parameters of the participants in the shared identification participating group and send them to the central aggregator for federated learning to obtain a global intelligent model. The data integration of multiple model parameters is completed through the execution nodes, providing a data basis for the fusion training of the global intelligent model.
[0053] Furthermore, step S500 of the present application also includes:
[0054] Step S510-1: performing local model aggregation based on the multiple model parameters by the central aggregator to obtain a first aggregated model;
[0055] Step S520-1: returning the first aggregation model to the multiple execution nodes, and performing loss analysis on the first aggregation model by the multiple execution nodes using local data samples to obtain multiple model loss functions;
[0056] Furthermore, step S520-1 of the present application also includes:
[0057] Step S521-1: downloading the first aggregation model through the multiple execution nodes, and decrypting it to obtain a first decrypted model;
[0058] Step S522-1: Perform loss analysis based on the first decryption model and the local intelligent model to obtain the multiple model loss functions.
[0059] Step S530-1: Perform iterative training based on the multiple model loss functions to obtain the global intelligent model.
[0060] Exemplarily, the central aggregator performs local model aggregation on multiple model parameters participating in the joint learning to obtain a first aggregated model. The aggregation method can be a simple average (weighted average or arithmetic average) or a more complex aggregation algorithm, such as weighted aggregation based on trust. The goal of aggregation is to generate a global model update that represents the contributions of all participants. The central aggregator applies the aggregated model update to the initial global model, that is, the predetermined target intelligent model, to generate a new global model, that is, the first aggregated model. This new global model not only includes the local learning results of each participant, but also ensures data privacy and security.
[0061] Further, the first aggregation model is sent to the local according to multiple execution nodes, and the execution node uses the local data sample to perform loss analysis on the sent model to obtain multiple model loss functions. Each participant optimizes and iterates the aggregation model based on the loss function, thereby obtaining an aggregation function with optimal accuracy and good aggregation effect. Specifically, each participant downloads the first aggregation model to the local according to the execution node and decrypts the model. The specific decryption process is as follows: the participants negotiate keys in advance and share symmetric keys. Key negotiation and symmetric encryption are not limited to one or several specified algorithms, but a sufficiently secure algorithm strategy must be used. When the participants upload local model parameters, they use symmetric keys in TEE to symmetrically encrypt the local model parameters. Since the channel cannot be guaranteed to be secure during the transmission process, the symmetric encryption is performed. The aggregator first decrypts with the symmetric key and then aggregates. Next, the aggregator encrypts with the symmetric key and sends it to each participant. After each participant decrypts, the local model is updated to ensure the security of data transmission.
[0062] Then, the local sample data is input into the decrypted aggregation model, and the aggregation model is verified and analyzed in combination with the local intelligent model to obtain the model loss function, and the accuracy change of the global model relative to the local model is compared. Furthermore, iterative training is performed based on multiple model loss functions. When the training converges, the global intelligent model is obtained. The accuracy of the global aggregation model is guaranteed through loss analysis and iterative training.
[0063] Furthermore, step S530-1 of the present application also includes:
[0064] Step S531-1: if the loss functions of the multiple models do not meet the loss convergence condition, update the first aggregation model using the local data sample to obtain multiple second local models;
[0065] Step S532-1: Sending model parameters of the plurality of second local models to the central aggregator for local model aggregation to obtain a second aggregated model;
[0066] Step S533-1: performing loss analysis on the second aggregation model through the multiple execution nodes to obtain multiple updated loss functions;
[0067] Step S534-1: If the multiple updated loss functions do not meet the loss convergence condition, repeat the training until the updated loss function meets the loss convergence condition to obtain the global intelligent model.
[0068] Optionally, the specific aggregation model optimization iterative process is: when multiple model loss functions do not meet the loss convergence condition, that is, compare the accuracy of the global model relative to the local model, when the global model accuracy is less than a certain threshold of the local model accuracy, use local data samples to update the first aggregation model to obtain multiple second local models, and then send the updated local models to the central aggregator for re-aggregation, and so on, continuously update and iterate until the updated loss function meets the loss convergence condition, and obtain the global intelligent model, so as to obtain the final learning result. The accuracy threshold is set by technical personnel in this field and is not limited here.
[0069] Furthermore, step S500 of the present application also includes:
[0070] Step S510-2: embedding an encryption algorithm in the plurality of execution nodes;
[0071] Step S520-2: Encrypt the multiple model parameters using the encryption algorithm, and send the multiple encrypted model parameters to the central aggregator for federated learning.
[0072] Step S600: Return the global intelligent model to the multiple data participants.
[0073] Specifically, the encryption of model parameters is completed by executing the encryption algorithm embedded in the node, and exists in the form of ciphertext in data transmission. Even if it is illegally accessed, the attacker cannot obtain the original data content, which greatly reduces the risk of information leakage and improves the trust of the system. Furthermore, the final optimized global intelligent model is sent to multiple data participants, and the local models of the participants are updated, thus completing federated learning.
[0074] Through the technical solutions of the above embodiments, the data sharing method based on TEE and federated learning provided by this application has the following technical effects:
[0075] 1. The computing process within the TEE (Trusted Execution Environment) is not open to users, including the negotiated keys and data generated during the computing process. Users are only allowed to input data and obtain output results. Therefore, information other than the results will not be exposed to users, which meets the need to protect the privacy of all participants.
[0076] 2. Data is transmitted in the channel in the form of ciphertext, and the key is only known by each participating party in the TEE environment. When using sufficiently secure key negotiation and symmetric encryption algorithms, data transmission can be considered secure.
[0077] 3. The trusted execution environment provides a secure independent processing environment, ensuring the confidentiality and integrity of the calculations and storage therein, and since the data will not be leaked to the user, the data calculation and storage are considered safe.
[0078] 4. Compared with plaintext calculation, the data only adds symmetric encryption and decryption steps when entering and leaving the domain, the operation complexity is low, and the symmetric encryption algorithm significantly surpasses homomorphic encryption in efficiency. Therefore, this solution solves the problem of low computational efficiency caused by the use of homomorphic encryption in federated learning.
[0079] 5. In symmetric encryption algorithms, the lengths of plaintext and ciphertext are approximately equal, so compared to plaintext calculations, this scheme does not increase communication overhead.
[0080] 6. Symmetric encryption algorithms do not introduce additional noise, so the calculation accuracy is the same as the plaintext calculation.
[0081] After the above process description and analysis, this solution ensures that the privacy information of all parties will not be leaked during the federated learning process by integrating the trusted execution environment and the symmetric encryption algorithm. The environment in which all parties participate in the federated learning is located in the trusted execution environment, where the data is calculated in plain text. When data needs to be exchanged, the plain text data is transmitted after symmetric encryption, and the receiver decrypts it in the trusted execution environment. The trusted execution environment ensures the security of local computing, while the symmetric encryption ensures the security of transmission. Since the symmetric encryption algorithm significantly surpasses homomorphic encryption in efficiency, and its plaintext and ciphertext lengths are approximately equal, no additional noise is introduced, it has the characteristics of high efficiency, low communication overhead, and no sacrifice of accuracy compared to other solutions such as homomorphic encryption. At the same time, this solution is not limited to the federated learning scenario, and it is also widely applicable in other data sharing scenarios, thus realizing an efficient and secure data sharing solution, which ensures the high efficiency and reliability of data sharing without sacrificing security.
[0082] Embodiment 2
[0083] Based on the same inventive concept as the data sharing method based on TEE and federated learning in the aforementioned embodiment, Figure 2 As shown, the present application provides a data sharing system based on TEE and federated learning, and the system includes:
[0084] A trusted execution environment establishment module 11 is used to obtain multiple data participants and establish multiple trusted execution environments for the multiple data participants;
[0085] An execution node creation module 12, configured to create a plurality of execution nodes in the plurality of trusted execution environments;
[0086] The model training module 13 is used to train the intelligent model according to the local data samples of the multiple data participants using the multiple execution nodes according to the predetermined target intelligent model characteristics to obtain multiple local intelligent models;
[0087] A sharing object identification module 14, configured to identify data sharing objects based on the data sharing requests of the plurality of data participants, and obtain a sharing identification participant group;
[0088] A global intelligent model acquisition module 15, configured to send multiple model parameters of multiple local intelligent models belonging to the shared recognition participant group to a central aggregator through the multiple execution nodes for aggregation to obtain a global intelligent model;
[0089] The return module 16 is used to return the global intelligent model to the multiple data participants.
[0090] Furthermore, the shared object identification module 14 is further configured to perform the following steps:
[0091] Parsing the data sharing requests of the multiple data participants to obtain the data exchange object of each data participant;
[0092] Based on the data exchange object, data participants with two-way exchange requests are combined to obtain the shared identification participant group.
[0093] Furthermore, the global intelligent model obtaining module 15 is also used to perform the following steps:
[0094] Performing local model aggregation based on the multiple model parameters by the central aggregator to obtain a first aggregated model;
[0095] Returning the first aggregation model to the multiple execution nodes, and performing loss analysis on the first aggregation model by using local data samples through the multiple execution nodes to obtain multiple model loss functions;
[0096] Iterative training is performed based on the multiple model loss functions to obtain the global intelligent model.
[0097] Furthermore, the global intelligent model obtaining module 15 is also used to perform the following steps:
[0098] If the loss functions of the multiple models do not meet the loss convergence condition, updating the first aggregation model using the local data sample to obtain multiple second local models;
[0099] Sending the model parameters of the plurality of second local models to the central aggregator for local model aggregation to obtain a second aggregated model;
[0100] Performing loss analysis on the second aggregation model through the multiple execution nodes to obtain multiple updated loss functions;
[0101] If the multiple updated loss functions do not meet the loss convergence condition, the training is repeated until the updated loss function meets the loss convergence condition, thereby obtaining the global intelligent model.
[0102] Furthermore, the global intelligent model obtaining module 15 is also used to perform the following steps:
[0103] The encryption algorithm is embedded in the plurality of execution nodes;
[0104] The multiple model parameters are encrypted by the encryption algorithm, and the multiple encrypted model parameters are sent to the central aggregator for federated learning.
[0105] Furthermore, the global intelligent model obtaining module 15 is also used to perform the following steps:
[0106] Downloading the first aggregation model through the multiple execution nodes and decrypting it to obtain a first decrypted model;
[0107] Loss analysis is performed based on the first decryption model and the local intelligent model to obtain the multiple model loss functions.
[0108] Furthermore, the model training module 13 is also used to perform the following steps:
[0109] The predetermined target intelligent model features include model input features, model output features and model network features.
[0110] Through the above detailed description of the data sharing method based on TEE and federated learning in this specification, those skilled in the art can clearly understand the data sharing method and system based on TEE and federated learning in this embodiment. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.
[0111] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A data sharing method based on TEE and federated learning, characterized in that: include: Acquire multiple data participants, and establish multiple trusted execution environments for the multiple data participants; Creating a plurality of execution nodes in the plurality of trusted execution environments; According to the predetermined target intelligent model characteristics, the multiple execution nodes are used to train the intelligent models according to the local data samples of the multiple data participants, so as to obtain multiple local intelligent models; Identify data sharing objects based on the data sharing requests of the multiple data participants to obtain a sharing identification participant group; Sending multiple model parameters of multiple local intelligent models belonging to the shared recognition participant group to a central aggregator through the multiple execution nodes for aggregation to obtain a global intelligent model; The global intelligent model is returned to the multiple data participants.
2. The method according to claim 1, characterized in that The data sharing object identification is performed based on the data sharing requests of the multiple data participants to obtain a sharing identification participant group, including: Parsing the data sharing requests of the multiple data participants to obtain the data exchange object of each data participant; Based on the data exchange object, data participants with two-way exchange requests are combined to obtain the shared identification participant group.
3. The method according to claim 1, characterized in that The plurality of execution nodes send the plurality of model parameters of the plurality of local intelligent models belonging to the shared recognition participant group to the central aggregator for aggregation to obtain a global intelligent model, including: Performing local model aggregation based on the multiple model parameters by the central aggregator to obtain a first aggregated model; Returning the first aggregation model to the multiple execution nodes, and performing loss analysis on the first aggregation model by using local data samples through the multiple execution nodes to obtain multiple model loss functions; Iterative training is performed based on the multiple model loss functions to obtain the global intelligent model.
4. The method according to claim 3, characterized in that Performing iterative training based on the multiple model loss functions to obtain the global intelligent model includes: If the loss functions of the multiple models do not meet the loss convergence condition, updating the first aggregation model using the local data sample to obtain multiple second local models; Sending the model parameters of the plurality of second local models to the central aggregator for local model aggregation to obtain a second aggregated model; Performing loss analysis on the second aggregation model through the multiple execution nodes to obtain multiple updated loss functions; If the multiple updated loss functions do not meet the loss convergence condition, the training is repeated until the updated loss function meets the loss convergence condition, thereby obtaining the global intelligent model.
5. The method according to claim 3, characterized in that The plurality of execution nodes send a plurality of model parameters of a plurality of local intelligent models belonging to the shared recognition participant group to a central aggregator for federated learning, including: The encryption algorithm is embedded in the plurality of execution nodes; The multiple model parameters are encrypted by the encryption algorithm, and the multiple encrypted model parameters are sent to the central aggregator for federated learning.
6. The method according to claim 3, characterized in that The first aggregation model is returned to the multiple execution nodes, and the multiple execution nodes use local data samples to perform loss analysis on the first aggregation model to obtain multiple model loss functions, including: Downloading the first aggregation model through the multiple execution nodes and decrypting it to obtain a first decrypted model; Loss analysis is performed based on the first decryption model and the local intelligent model to obtain the multiple model loss functions.
7. The method according to claim 1, characterized in that The predetermined target intelligent model features include model input features, model output features and model network features.
8. Data sharing system based on TEE and federated learning, characterized by: For implementing the data sharing method based on TEE and federated learning according to any one of claims 1 to 7, the system comprises: A trusted execution environment establishment module, used to obtain multiple data participants and establish multiple trusted execution environments for the multiple data participants; An execution node creation module, used to create multiple execution nodes in the multiple trusted execution environments; A model training module, used to train the intelligent model according to the local data samples of the multiple data participants using the multiple execution nodes according to the predetermined target intelligent model characteristics, so as to obtain multiple local intelligent models; A sharing object identification module, used to identify data sharing objects based on the data sharing requests of the multiple data participants, and obtain a sharing identification participant group; A global intelligent model acquisition module, used for sending multiple model parameters of multiple local intelligent models belonging to the shared recognition participating group to a central aggregator through the multiple execution nodes for aggregation to obtain a global intelligent model; A return module is used to return the global intelligent model to the multiple data participants.