A federated learning method, apparatus and system

Abstract

A federated learning method, device and system, relating to the field of artificial intelligence, the method comprising: a first node obtaining data distribution information of a plurality of second nodes according to a target data feature required by a training task, the data distribution information of any second node being used to indicate a data category to which business data conforming to the target data feature and stored locally by the second node belongs; the first node selecting at least two target second nodes from the plurality of second nodes according to a target data category required by the training task and the data distribution information of the plurality of second nodes, any target second node locally storing target business data conforming to the target data feature and belonging to the target data category; and the first node instructing the at least two target second nodes to perform federated learning to obtain a federated learning model corresponding to the target data category in the training task. In the case that the participants have multiple data distributions, the method can avoid the trained model being affected by data poisoning.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a federated learning method, apparatus and system. Background Technology

[0002] Federated learning is an emerging foundational technology for artificial intelligence (AI). Its design goal is to enable efficient machine learning (ML) among multiple participants or computing nodes, while ensuring information security during big data exchange, protecting terminal and personal data privacy, and guaranteeing legal compliance. The machine learning algorithms used in federated learning are not limited to important algorithms such as neural networks and random forests, and it is expected to become the foundation for next-generation AI collaborative algorithms and networks.

[0003] Horizontal federated learning is a key branch of federated learning. The system architecture of horizontal federated learning includes a coordinating node and several participating nodes. The coordinating node sends an initial AI model to each participating node. Each participating node trains its own AI model using its own dataset and sends the updated model parameters / gradient values ​​back to the coordinating node. The coordinating node then aggregates the updated model parameters / gradient values ​​received from the participating nodes (e.g., using a federated averaging algorithm) and sends the aggregated updated model back to the participating nodes. This process is repeated until the model converges or a preset iteration stopping condition is met. In this architecture, the original datasets of the participating nodes never leave their local machines, protecting user privacy and data security while reducing the communication overhead of sending the original datasets.

[0004] However, in reality, the datasets of each participating node may be unevenly sized. Furthermore, the distribution of datasets across different participating nodes often varies significantly. For example, some participating nodes may have a relatively uniform dataset distribution, while others may have multiple sub-datasets with substantial differences between them. Therefore, when different participating nodes use datasets with different distributions to train the same AI model, the AI ​​model is susceptible to data poisoning, leading to a decrease in the accuracy of the updated model after aggregation.

[0005] Therefore, ensuring the accuracy of federated learning models remains a pressing issue that needs to be addressed. Summary of the Invention

[0006] This application provides a federated learning method, apparatus, and system that helps ensure the accuracy of federated learning models.

[0007] In a first aspect, embodiments of this application provide a federated learning method that can be applied to a federated learning system including a first node and multiple second nodes, and can be implemented by the first node acting as a coordinator.

[0008] In this method, the first node obtains data distribution information of multiple second nodes based on the target data features required for the training task. The data distribution information of any one second node is used to indicate the data category to which the business data that conforms to the target data features and is locally stored in the second node belongs. Based on the target data category required for the training task and the data distribution information of the multiple second nodes, the first node selects at least two target second nodes from the multiple second nodes. Any one of the target second nodes locally stores target business data that conforms to the target data features and belongs to the target data category. The first node instructs the at least two target second nodes to perform federated learning to obtain a federated learning model corresponding to the target data category in the training task.

[0009] This scheme allows the first node to selectively instruct at least two target second nodes to perform federated learning based on the data distribution information of each of the multiple second nodes, thereby obtaining a federated learning model corresponding to the target data features. This approach allows for the acquisition of corresponding federated learning models for different data distributions, minimizing the poisoning effect of different data distributions on the model and ensuring the accuracy of the obtained federated learning model.

[0010] It is understood that in this embodiment, the data category is a division of business data stored locally on the second node side. The target data feature can be a set of data features (or a data feature group, including multiple data features). After the second node identifies and classifies the business data that conforms to the target data feature, it can obtain sub-data sets belonging to at least one data category. Each data category corresponds to a data distribution, and the sub-data set corresponding to each data category is also the data set corresponding to that data distribution, which can be used to test and evaluate the AI ​​model of the corresponding data category. For example, the target data feature may include a combination of multiple data features such as height, weight, chest circumference, and hip circumference. After identifying and classifying the business data according to the target data feature, the data category obtained may be a division of body types such as thin, slightly overweight, obese, and overweight, where "thin," "slightly overweight," "obese," and "overweight" are data categories, and the distribution corresponds to a data distribution.

[0011] In one possible design, each second node deploys at least one data analysis model, each data analysis model corresponding to a data feature group and used to identify the data category of business data conforming to the corresponding data feature group; the first node obtains data distribution information of multiple second nodes according to the target data features required by the training task, including: the first node sending a first query message to each of the multiple second nodes according to the target data features; wherein the first query message sent to any second node includes the identifier of the target data feature and the identifier of the target data analysis model, the target data analysis model corresponding to the target data feature; the first node receives corresponding data distribution information from the multiple second nodes, the data distribution information of any second node being used to indicate the identifier of at least one data category, and the data information of business data belonging to the at least one data category stored by the second node.

[0012] This scheme allows for the deployment of at least one data analysis model on multiple second nodes. This model enables the identification and categorization of business data stored locally on any of the second nodes that conforms to specific data characteristics, thereby obtaining the data distribution information of the second nodes. It is understood that this embodiment only illustrates how data analysis models can be used to analyze local data sets on the second node side, and does not limit the specific implementation method for obtaining the data distribution information of the second nodes. In other embodiments, the first node can obtain the data distribution information of any one of the multiple second nodes through any suitable method, and this application does not limit this.

[0013] In one possible design, the first query message sent by the first node to any second node also includes the identifier of the target data category, wherein the data distribution information returned by the second node includes: the identifier of the target data category, and the data information of the target business data belonging to the target data category stored by the second node.

[0014] This scheme allows the first node to indicate the target data category to any second node, enabling the second node to provide feedback on the target business data belonging to that category, stored locally. This allows the first node to select a suitable AI model for that target data category from the feedback from multiple second nodes, thereby obtaining the corresponding federated learning model.

[0015] In one possible design, before the first node obtains the data distribution information of multiple second nodes based on the target data features required for the training task, the method further includes: the first node sending a data analysis model deployment message to each of the multiple second nodes, wherein the data analysis model deployment message sent to any second node includes: an identifier of the at least one data analysis model, and a model file of the at least one data analysis model.

[0016] With this scheme, the first node can deploy at least one data analysis model on each of the multiple second nodes by adding data analysis model deployment messages between the first node and the second node. Then, the data distribution information of each second node can be obtained by using the data analysis model deployed on the second node.

[0017] In one possible design, the first node instructs the at least two target second nodes to perform federated learning to obtain a federated learning model corresponding to the target data category in the training task. This includes: the first node sending a model training message to each of the at least two target second nodes, wherein the model training message sent to any target second node includes an identifier of a target artificial intelligence (AI) model, which corresponds to the target data category; and the first node obtaining the federated learning model corresponding to the target data category in the training task based on the updated AI models received from the at least two target second nodes.

[0018] This scheme allows the first node to select at least two target second nodes from among multiple second nodes that can participate in the federated learning process of the federated learning model corresponding to the target data category, based on the data distribution information of multiple second nodes. The first node then includes the identifier of the target AI model in the model training messages sent to these two target second nodes, instructing them to train the target AI model using stored target business data to obtain an updated AI model. This process is repeated until the model converges or reaches a preset iteration stopping condition, thereby obtaining the federated learning model corresponding to the target data category. Because this federated learning model is trained using a dataset with the same data distribution, the effects of data poisoning are avoided, ensuring the accuracy of the federated learning model.

[0019] In one possible design, the model training message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

[0020] This scheme allows the first node to include the identifier of the target data category and the identifier of the target data analysis model in the model training message sent to any target second node. This instructs the target second node to use the training data from the stored target business data for federated learning to obtain an updated AI model. This ensures the use of correct training data for model training, avoids the effects of data poisoning, and ultimately guarantees the accuracy of the aggregated federated learning model.

[0021] In one possible design, the first node instructs the at least two target second nodes to perform federated learning to obtain a federated learning model corresponding to the target data category in the training task. The design further includes: the first node sending a model evaluation message to each of the at least two target second nodes, wherein the model evaluation message sent to any target second node includes an identifier and evaluation metric for the target evaluation model, and the target evaluation model corresponds to the target data category; and the first node receiving corresponding model evaluation results from the at least two target second nodes.

[0022] This scheme allows the first node to include the identifier of the target evaluation model in the model evaluation message sent to the target second node. This enables the target second node to perform model evaluation using test data from the stored target business data. This ensures that the correct test data is used for model evaluation, avoiding inaccurate model evaluations.

[0023] In one possible design, the model evaluation message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

[0024] This scheme allows the first node to include the identifier of the target data category and the identifier of the target data analysis model in the model evaluation message sent to the target second node. This enables the target second node to use test data from the stored target business data for model evaluation, ensuring the use of correct test data and avoiding inaccurate model evaluations.

[0025] In one possible design, the federated learning system is a wireless AI model-driven network system; the first node includes a model management function (MMF) module; any second node includes a model training function (MTF) module, a data management function (DMF) module, and a model evaluation function (MEF) module; wherein, at least one data analysis model is deployed in the DMF module or the MTF module; the first node sends a first query message to each second node, including: the MMF module sending the first query message to the DMF module or the MTF module of each second node.

[0026] With this scheme, when the federated learning system is implemented as a wireless AI model-driven network system, the MMF module of the first node can communicate with the corresponding functional module of the second node, and carry relevant indication information in each message sent. This allows the corresponding functional model of the second node to use the correct data when performing its function to complete the federated learning, thereby avoiding the impact of data poisoning and ensuring the accuracy of the federated learning model corresponding to each data category.

[0027] In one possible design, the method further includes: the first node sending a mapping table to each of the plurality of second nodes, wherein the mapping table sent to any second node is used to record the mapping relationship between the identifier of data features, the identifier of AI models, the identifier of data analysis models, and the identifier of data categories.

[0028] This scheme allows the first node to send a mapping table to the second node. When instructing the corresponding second node to perform federated learning, the target data category identifier or target data analysis model identifier no longer needs to be specified in the relevant messages sent. When frequent model training messages need to be sent for model iteration, the amount of message transmission between communication interfaces can be effectively reduced, thereby reducing signaling overhead.

[0029] Secondly, embodiments of this application provide a federated learning method that can be applied to any second node in a federated learning system that includes a first node and multiple second nodes.

[0030] In this method, the second node receives a first query message from the first node, the first query message indicating the target data features required for the training task; the second node sends data distribution information to the first node based on the target data features, the data distribution information indicating the data category to which the business data that conforms to the target data features and is locally stored by the second node belongs; the second node, according to the instruction of the first node, uses the stored target business data belonging to the target data category to train the target artificial intelligence (AI) model corresponding to the target data category, thereby obtaining an updated AI model; the second node sends the updated AI model to the first node, so that the first node obtains the federated learning model corresponding to the target data category in the training task.

[0031] In one possible design, the second node is deployed with at least one data analysis model, each data analysis model corresponding to a data feature group and used to identify the data category of business data that conforms to the corresponding data feature group; the first query message includes an identifier of the target data feature and an identifier of the target data analysis model, the target data analysis model corresponding to the target data feature; the second node sends data distribution information to the first node according to the target data feature, including: the second node using the target data analysis model to identify the data category of stored business data that conforms to the target data feature, and obtaining data information of business data belonging to at least one data category; the second node sends the data distribution information to the first node, the data distribution information being used to indicate the identifier of at least one data category, and the data information of business data belonging to at least one data category.

[0032] In one possible design, the first query message may also include an identifier of the target data category, wherein the data distribution information includes: the identifier of the target data category, and data information of the target business data belonging to the target data category stored by the second node.

[0033] In one possible design, before the second node receives the first query message from the first node, the method further includes: the second node receiving a data analysis model deployment message from the first node; wherein the data analysis model deployment message includes: an identifier of the at least one data analysis model, and a model file of the at least one data analysis model.

[0034] In one possible design, the second node, according to the instructions of the first node, uses stored target business data belonging to the target data category to train a target artificial intelligence (AI) model corresponding to the target data category to obtain an updated AI model. This includes: the second node receiving a model training message from the first node, wherein the model training message includes an identifier for the target AI model, and the target AI model corresponds to the target data category; the second node obtaining stored target business data that conforms to the target data characteristics and belongs to the target data category based on the AI ​​model identifier; and the second node training the AI ​​model based on the target business data to obtain an updated AI model.

[0035] In one possible design, the model training message may also include: an identifier of the target data category and an identifier of the target data analysis model.

[0036] In one possible design, the method further includes: the second node receiving a model evaluation message from the first node and evaluating the target evaluation model using the target business data, wherein the target evaluation model message includes an identifier and evaluation indicators for the target evaluation model, and the target evaluation model corresponds to the target data category; the second node sending the model evaluation result to the first node.

[0037] In one possible design, the model evaluation message may also include: an identifier of the target data category and an identifier of the target data analysis model.

[0038] In one possible design, the federated learning system is a wireless AI model-driven network system; the first node includes a model management function (MMF) module; any second node includes a model training function (MTF) module, a data management function (DMF) module, and a model evaluation function (MEF) module; wherein, at least one data analysis model is deployed in the DMF module or the MTF module; the second node receives a first query message from the first node, including: the DMF module or the MTF module receiving the first query message from the MMF module.

[0039] In one possible design, where the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the MTF module, after the DMF module receives the first query message from the MTF module, the method further includes: the DMF module sending a data analysis message to the MTF module, the data analysis message including a full dataset of data that conforms to the target data characteristics stored by the DMF module, an identifier of the target data category, and an identifier of the data analysis model, and the data analysis message being used to instruct the MTF module to use the target data analysis model to identify the data category of the full dataset.

[0040] In one possible design, where the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the DMF module, the second node receives model training messages from the first node, including: the MTF module receiving the model training messages from the MMF module; after the MTF module receives the model training messages from the MMF module, the method further includes: the MTF module sending a second query message to the DMF module, the second query message being used to instruct the DMF module to provide feedback to the MTF module on the training data set in the target business data; wherein the second query message includes: an identifier of the target data feature, an identifier of the target AI model, and first data type indication information; or, an identifier of the target data feature, an identifier of the target data category, an identifier of the target data analysis model, and first data type indication information.

[0041] In one possible design, where the MEF module and the MTF module are located in different entities, after the MEF module receives the model evaluation message from the MTF module, the method further includes: the MEF module sending a third query message to the DMF module, the third query message being used to instruct the DMF module to provide feedback to the MEF module on the test data set in the target business data; wherein the third query message includes: an identifier of the target data feature, an identifier of the target AI model, and second data type indication information; or, an identifier of the target data feature, an identifier of the target data category, an identifier of the target data analysis model, and second data type indication information.

[0042] Thirdly, embodiments of this application provide a federated learning apparatus applied to a federated learning system including a first node and multiple second nodes. The apparatus includes: a communication unit, configured to acquire data distribution information of multiple second nodes based on target data features required for a training task, wherein the data distribution information of any one second node is used to indicate the data category to which business data conforming to the target data features stored locally on the second node belongs; and a processing unit, configured to select at least two target second nodes from the multiple second nodes based on the target data category required for the training task and the data distribution information of the multiple second nodes, and instruct the at least two target second nodes to perform federated learning, so as to obtain a federated learning model in the training task corresponding to the target data category in which any one target second node locally stores target business data that conforms to the target data features and belongs to the target data category.

[0043] In one possible design, each second node deploys at least one data analysis model, each data analysis model corresponding to a data feature group and used to identify the data category of business data conforming to the corresponding data feature group; the communication unit is used to: send a first query message to each of the plurality of second nodes according to the target data feature; wherein the first query message sent to any second node includes an identifier of the target data feature and an identifier of the target data analysis model, the target data analysis model corresponding to the target data feature; and receive corresponding data distribution information from the plurality of second nodes, wherein the data distribution information of any second node is used to indicate the identifier of at least one data category, and the data information of business data belonging to the at least one data category stored by the second node.

[0044] In one possible design, the first query message sent by the first node to any second node also includes the identifier of the target data category, wherein the data distribution information returned by the second node includes: the identifier of the target data category, and the data information of the target business data belonging to the target data category stored by the second node.

[0045] In one possible design, before the first node obtains the data distribution information of multiple second nodes based on the target data features required for the training task, the communication unit is further configured to: send a data analysis model deployment message to each of the multiple second nodes, wherein the data analysis model deployment message sent to any second node includes: the identifier of the at least one data analysis model, and the model file of the at least one data analysis model.

[0046] In one possible design, the processing unit is configured to: send a model training message to each of the at least two target second nodes, wherein the model training message sent to any target second node includes an identifier of a target artificial intelligence (AI) model, the target AI model corresponding to the target data category; and obtain a federated learning model corresponding to the target data category in the training task based on the updated AI model received from the at least two target second nodes.

[0047] In one possible design, the model training message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

[0048] In one possible design, the communication unit is further configured to: send a model evaluation message to each of the at least two target second nodes, wherein the model evaluation message sent to any target second node includes an identifier and an evaluation index of the target evaluation model, the target evaluation model corresponding to the target data category; and receive corresponding model evaluation results from the at least two target second nodes.

[0049] In one possible design, the model evaluation message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

[0050] In one possible design, the federated learning system is a wireless AI model-driven network system; the first node includes a model management function (MMF) module; any second node includes a model training function (MTF) module, a data management function (DMF) module, and a model evaluation function (MEF) module; wherein, at least one data analysis model is deployed in the DMF module or the MTF module; the communication unit sends a first query message to each second node, including: the communication unit of the MMF module sends the first query message to the DMF module or the MTF module of each second node.

[0051] In one possible design, the communication unit is further configured to: send a mapping table to each of the plurality of second nodes, wherein the mapping table sent to any one of the second nodes is used to record the mapping relationship between the identifiers of data features, the identifiers of AI models, the identifiers of data analysis models, and the identifiers of data categories.

[0052] Fourthly, embodiments of this application provide a federated learning apparatus, applied to any second node in a federated learning system including a first node and multiple second nodes. The apparatus includes: a communication unit, configured to receive a first query message from the first node, the first query message indicating target data features required for a training task; and to send data distribution information to the first node based on the target data features, the data distribution information indicating the data category to which business data conforming to the target data features, stored locally on the second node, belongs; a processing unit, configured to train a target artificial intelligence (AI) model corresponding to the target data category using stored target business data belonging to the target data category, according to the instructions of the first node, to obtain an updated AI model; the communication unit is further configured to send the updated AI model to the first node, so that the first node obtains the federated learning model corresponding to the target data category in the training task.

[0053] In one possible design, the second node is deployed with at least one data analysis model, each data analysis model corresponding to a data feature group and used to identify the data category of business data that conforms to the corresponding data feature group; the first query message includes the identifier of the target data feature and the identifier of the target data analysis model, the target data analysis model corresponding to the target data feature; the processing unit is used to use the target data analysis model to identify the data category of the stored business data that conforms to the target data feature, and obtain data information of business data belonging to at least one data category; the communication unit is also used to send the data distribution information to the first node, the data distribution information being used to indicate the identifier of at least one data category, and the data information of business data belonging to at least one data category.

[0054] In one possible design, the first query message may also include an identifier of the target data category, wherein the data distribution information includes: the identifier of the target data category, and data information of the target business data belonging to the target data category stored by the second node.

[0055] In one possible design, the communication unit is further configured to: receive a data analysis model deployment message from the first node before the first node receives the first query message; wherein the data analysis model deployment message includes: an identifier of the at least one data analysis model, and a model file of the at least one data analysis model.

[0056] In one possible design, the communication unit is used to receive a model training message from the first node, wherein the model training message includes an identifier of the target AI model, and the target AI model corresponds to the target data category; the processing unit is used to obtain, based on the identifier of the AI ​​model, stored target business data that conforms to the characteristics of the target data and belongs to the target data category, and train the AI ​​model based on the target business data to obtain an updated AI model.

[0057] In one possible design, the model training message may also include: an identifier of the target data category and an identifier of the target data analysis model.

[0058] In one possible design, the communication unit is further configured to: receive a model evaluation message from the first node, and evaluate the target evaluation model using the target service data, wherein the target evaluation model message includes an identifier and evaluation index of the target evaluation model, and the target evaluation model corresponds to the target data category; and send the model evaluation result to the first node.

[0059] In one possible design, the model evaluation message may also include: an identifier of the target data category and an identifier of the target data analysis model.

[0060] In one possible design, the federated learning system is a wireless AI model-driven network system; the first node includes a model management function (MMF) module; any second node includes a model training function (MTF) module, a data management function (DMF) module, and a model evaluation function (MEF) module; wherein, at least one data analysis model is deployed in the DMF module or the MTF module; the communication unit receives a first query message from the first node, including: the communication unit of the DMF module or the communication unit of the MTF module receives the first query message from the MMF module.

[0061] In one possible design, where the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the MTF module, after the communication unit of the DMF module receives the first query message from the MTF module, the communication unit of the DMF module is further configured to: send a data analysis message to the MTF module, the data analysis message including the full dataset of data that conforms to the target data characteristics stored by the DMF module, the identifier of the target data category, and the identifier of the data analysis model, and the data analysis message is used to instruct the MTF module to use the target data analysis model to identify the data category of the full dataset.

[0062] In one possible design, where the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the DMF module, the second node receives model training messages from the first node, including: the communication unit of the MTF module receiving the model training messages from the MMF module; after receiving the model training messages from the MMF module, the communication unit of the MTF module is further configured to: send a second query message to the DMF module, the second query message instructing the DMF module to provide feedback to the MTF module on the training data set in the target business data; wherein the second query message includes: the identifier of the target data feature, the identifier of the target AI model, and first data type indication information; or, the identifier of the target data feature, the identifier of the target data category, the identifier of the target data analysis model, and first data type indication information.

[0063] In one possible design, where the MEF module and the MTF module are located in different entities, the communication unit is configured to: receive the model evaluation message from the MMF module via the communication unit of the MEF module; after receiving the model evaluation message from the MMF module via the communication unit of the MEF module, the communication unit of the MEF module is further configured to: send a third query message to the DMF module, the third query message being used to instruct the DMF module to provide feedback to the MEF module on the test data set in the target business data; wherein, the third query message includes: an identifier of the target data feature, an identifier of the target AI model, and second data type indication information; or, an identifier of the target data feature, an identifier of the target data category, an identifier of the target data analysis model, and second data type indication information.

[0064] Fifthly, embodiments of this application provide a federated learning system, including the federated learning apparatus described in any one of the third aspects above; and the federated learning apparatus described in any one of the fourth aspects.

[0065] Sixthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the method described in any one of the first or second aspects above.

[0066] In a seventh aspect, embodiments of this application provide a computer program product that, when run on a computer, causes the computer to perform the method described in any one of the first or second aspects above.

[0067] Eighthly, embodiments of this application provide a chip, the chip including a processor and a data interface, the processor reading instructions stored in a memory through the data interface to execute the method described in any one of the first or second aspects above.

[0068] In one possible design, the chip may further include a memory storing instructions, and the processor is configured to execute the instructions stored in the memory. When the instructions are executed, the processor is configured to perform the method described in any one of the first or second aspects above. Attached Figure Description

[0069] Figure 1 A schematic diagram of a federated learning scenario;

[0070] Figure 2 A diagram illustrating the general process of federated learning;

[0071] Figures 3A-3BA schematic diagram of a wireless AI model-driven network system;

[0072] Figure 4 This is a schematic diagram of federated learning in a network system driven by a wireless AI model.

[0073] Figure 5 A schematic diagram illustrating the training principle of the federated learning model provided in this application embodiment;

[0074] Figures 6A-6C A schematic diagram of the system architecture applicable to the embodiments of this application;

[0075] Figure 7 A flowchart illustrating the federated learning method provided in this application embodiment;

[0076] Figure 8 A flowchart illustrating the federated learning method provided in this application embodiment;

[0077] Figure 9 A flowchart illustrating the federated learning method provided in this application embodiment;

[0078] Figure 10 A flowchart illustrating the federated learning method provided in this application embodiment;

[0079] Figure 11 A flowchart illustrating the federated learning method provided in this application embodiment;

[0080] Figure 12 A flowchart illustrating the federated learning method provided in this application embodiment;

[0081] Figure 13 A flowchart illustrating the federated learning method provided in this application embodiment;

[0082] Figure 14 A flowchart illustrating the federated learning method provided in this application embodiment;

[0083] Figure 15 A schematic diagram of the federated learning device provided in the embodiments of this application;

[0084] Figure 16 A schematic diagram of a federated learning device provided in an embodiment of this application. Detailed Implementation

[0085] To make it easier to understand, let's first combine... Figure 1 and Figure 2 This provides an illustrative example of the scenarios and processes of federated learning.

[0086] See Figure 1In a federated learning scenario, there may be a coordinating node and multiple participating nodes. The coordinating node is the coordinator in the federated learning process, while the participating nodes are the participants in the federated learning process and also the owners of the dataset. For ease of understanding and distinction, in this embodiment, the coordinating node is referred to as the first node 110, and the participating nodes are referred to as the second node 120.

[0087] The first node 110 and the second node 120 can be any node that supports data transmission (such as a network node). For example, the first node can be a server, or a parameter server, or an aggregation server. The second node can be a client, such as a mobile terminal or a personal computer.

[0088] The first node 110 can be used to maintain the federated learning model. The second node 120 can obtain the federated learning model from the first node 110 and train it locally using its local training dataset to obtain a local model. After training the local model, the second node 120 can send the local model back to the first node 110 so that the first node 110 can update or optimize the federated learning model. This process is repeated for multiple iterations until the federated learning model converges or reaches a preset iteration stopping condition (e.g., reaching the maximum number of iterations or the longest training time).

[0089] The following is combined Figure 2 This section introduces the general process of federated learning.

[0090] S210: The first node 110 constructs the federated learning model. The first node can construct a general machine learning model or a specific machine learning model according to requirements. Taking image recognition as an example, the first node can construct a convolutional neural network (CNN) as the federated learning model.

[0091] S220: First node 110 selects second node 120. The second node 120 selected by first node 110 will receive the federated learning model issued by first node 110. First node 110 can randomly select second node 120, or it can select second node 120 according to a certain strategy. For example, first node 110 can choose second node 120 with a larger amount of training data required by the federated learning model.

[0092] S230: The second node 120 obtains or receives the federated learning model from the first node 110. For example, in one implementation, the second node 120 may actively request the first node 110 to distribute the federated learning model. Alternatively, in another implementation, the first node 110 may actively distribute the federated learning model to the second node. Taking the second node 120 as the client and the first node 110 as the server as an example, the client can download the federated learning model from the server.

[0093] S240: The second node 120 trains the federated learning model using local training data to obtain a local model. The second node 120 can use the federated learning model as the initial model for the local model, and then use the local training data to train the initial model in one or more steps to obtain the local model.

[0094] In S250, the first node 110 aggregates the local models trained by the second node 120 to obtain an updated federated learning model. For example, in one implementation, the first node 110 can perform a weighted summation of the parameters of multiple local models of the second nodes 120 and use the result of the weighted summation as the updated federated learning model.

[0095] The process described in S220-S250 can be viewed as an iteration in the federated learning process. The first node 110 and the second node 120 can repeatedly execute steps S220-S250 until the federated learning model converges or reaches the preset effect.

[0096] With the rise of Artificial Intelligence (AI) technology, AI has helped various industries solve problems that traditional algorithms could not. Currently, the wireless network field is also attempting to introduce AI technology to improve wireless network performance. Wireless AI model-driven network systems primarily address the issues of distributing, updating, and coordinating AI algorithm models within wireless networks. The following section will combine... Figure 3A , Figure 3B and Figure 4 This paper provides an exemplary illustration of the scenario and process of federated learning in a wireless AI model-driven network system.

[0097] Figure 3A An example of the functional modules that a wireless AI model-driven network system may contain is shown.

[0098] See Figure 3BThe main functional modules of a wireless AI model-driven network system can include: a model management function (MMF module), a model training function (MTF module), a model evaluation function (MEF module), and a data management function (DMF module). The MMF module manages the model lifecycle and can trigger model training and evaluation. The MTF module trains local models and outputs model files after training. The MEF module evaluates the performance of the trained model using a test set. The DMF module subscribes to and stores the datasets needed by the model and provides data query and retrieval services. The MTF and MEF modules can each send data requests to the DMF module to obtain datasets (including training and test sets).

[0099] The MMF module acts as the coordinator, while the MTF, MEF, and DMF modules are the participants. When a training task is triggered, the MMF module selects several MTF modules as participants in the model training process. Upon receiving the model training message, the MTF module queries the DMF module for data based on the data feature names. In addition to the data feature names, the data query message includes the data type (type), indicating whether the queried data is training or testing data. The DMF module queries the entire dataset based on the data feature names and then returns the training dataset from the entire dataset to the MTF module. The MTF module uses the training data returned by the DMF module to train the model and sends a training completion notification message to the MMF module upon completion. After receiving training completion messages from all participating MTF modules, the MMF module performs aggregation processing on the model. The MMF module repeats this process until the conditions for stopping training are met.

[0100] It is understood that the wireless AI model-driven network system to which the embodiments of this application are applicable may also include Figure 3B Other modules shown besides the main functional modules are only those relevant to this application and are not intended to limit the wireless AI model-driven network architecture and its functional implementation.

[0101] See Figure 4 In the process of federated learning based on wireless AI model-driven network architecture, Figure 3B The interaction process between the various functional modules is shown below:

[0102] S401: When the training task is triggered, the MMF module obtains the initial training model.

[0103] S402: The MMF module sends a training model deployment message to the registered MTF modules. The training model deployment message includes the training model name and training model file, so that each MTF module can deploy the training model locally according to the training model deployment message.

[0104] S403: The MMF module randomly selects several participants from the registered participants (i.e., the MTF module) to participate in model training.

[0105] S404: The MMF module sends a model training message to the MTF module that needs to participate in model training. The model training message includes the training model name and training model file to trigger the MTF module to start the model training process.

[0106] S405: The MTF module sends a data query message to the DMF module. This data query message includes information such as the data feature name and data type indication of the data required for model training.

[0107] S406: The DMF module sends a data query confirmation message to the MTF module, which carries the dataset that meets the data query information requirements.

[0108] S407: The MTF module uses the dataset returned by the DMF module to train the model. After several iterations, the model training is completed and the training model file is updated.

[0109] S408: The MTF module sends a training completion notification message to the MMF module. This training completion notification message includes the training model name, training model file, and amount of training data used in the training of the local model trained by the MTF module.

[0110] S409: After the MMF module collects the training completion notification messages returned by all the MTF modules participating in the current training round, it aggregates the locally trained models uploaded by the MTF modules using an aggregation algorithm (such as federated averaging) to obtain the updated model parameters.

[0111] S410: The MMF module determines whether the conditions for stopping training are met. If not, it repeats S403-S409 to proceed with the next round of participant selection and model training until the conditions for stopping training are met, at which point the current process ends. The conditions for stopping training could be, for example, whether the maximum number of training iterations or the maximum training time has been reached.

[0112] To ensure the performance of horizontal federated learning when the datasets of the participants are uneven in size, the industry has targeted... Figure 4The federated learning process shown has been improved, specifically including:

[0113] (1) In the S403 participant selection phase, the MMF module sends the data feature name to the DMF module to query the data information of each participant and selects a suitable participant to participate in the current model training process based on the data information fed back by each participant.

[0114] (2) During the local model training phase, in S408, the MMF module instructs the MTF module on the amount of data used for local model training.

[0115] Since deep learning typically assumes that data is independently distributed, if the same AI model is trained using data from different distributions, the model is susceptible to data poisoning, which can lead to a decrease in the accuracy of the aggregated model.

[0116] In practical applications, the data distribution among different participants often varies significantly. Figure 4 In the federated learning process shown, during the participant selection phase of S403, the MMF module randomly selects participants for this round of model training, making it impossible to select participants with the same data distribution. Therefore, the trained model is susceptible to poisoning effects from different participants' varying data distributions. During the local model training phase, the MTF module only obtains data from the DMF module through data feature names. Thus, the MTF module can only obtain the full dataset with those features and uses it for model training. However, when multiple distributions exist in the full dataset, the model is affected by data poisoning, resulting in slower training convergence and decreased model performance.

[0117] And based on Figure 4 In the improved scheme, the MMF module only queries the DMF module for data information by data feature name. Therefore, it can only retrieve the full dataset with that data feature and return the full dataset information (e.g., data length). However, when the data has multiple distributions, the DMF module cannot distinguish between different distributions based on the data name. Consequently, the MMF module cannot obtain information for each data distribution, and the problem of the MMF module selecting participants with different data distributions to train the same model still exists. Similarly, during the local model training phase, the MTF module can only obtain data from the DMF module by data feature name. Therefore, it can only obtain the full dataset with that data feature and uses the full dataset for model training. When multiple distributions exist in the full dataset, the problems of slower training convergence and decreased model performance caused by data poisoning cannot be avoided.

[0118] In view of this, embodiments of this application provide a federated learning scheme that helps ensure the accuracy of federated learning models. In this scheme, the method and apparatus are based on the same technical concept. Since the methods and apparatus solve problems based on similar principles, their implementations can refer to each other, and repeated details will not be elaborated further.

[0119] This scheme can be applied to a federated learning system comprising a first node and multiple second nodes. When a training task is triggered, the first node, acting as the coordinator, determines the target data features and target data category based on the training task. Based on the target data features, the first node obtains data distribution information from multiple second nodes. The data distribution information of any one second node indicates the data category to which the business data stored locally on that second node that conforms to the target data features belongs. Then, based on the target data category and the data distribution information of the multiple second nodes, the first node selects at least two target second nodes and instructs these at least two target second nodes to perform federated learning to obtain a federated learning model corresponding to the target data category in the training task.

[0120] The data distribution information fed back by the second nodes can serve as one of the decision-making criteria. Based on the data distribution information from multiple second nodes, the first node can select at least two target second nodes that store business data that matches the target data characteristics and belongs to the target data category. These target second nodes will then participate in the model training process of the federated learning model corresponding to the target data category, thereby obtaining the corresponding federated learning model for that target data category. Furthermore, during the process of coordinating the training of the federated learning model corresponding to the target data category by at least two target second nodes, the first node can include relevant instruction information in the messages sent to any of the target second nodes. This instruction instructs each target second node to select target business data belonging to that target data category to train or evaluate the corresponding AI model, thus obtaining the federated learning model corresponding to the target data category.

[0121] This scheme identifies the data category to which business data that matches the characteristics of the target data belongs. Each data category corresponds to a data distribution. For each data category, the corresponding target business data is used to complete model training or model evaluation. Thus, when each participant has multiple data distributions, corresponding federated learning models are obtained for different data distributions. This minimizes the poisoning effect of different data distributions of different participant nodes on the federated learning model, thereby ensuring the accuracy of the obtained federated learning model.

[0122] To facilitate understanding, the following will be combined with... Figure 5 This application introduces the principles of the federal learning program.

[0123] See Figure 5 When a training task is triggered, the first node, acting as the coordinator, can distribute at least one data analysis model to multiple second nodes. Each data analysis model can be a clustering or classification model, and each model corresponds to a set of data features and can be used to identify data categories of business data that conform to the corresponding data features. This divides the full dataset conforming to the corresponding data features into sub-data sets belonging to at least one data category. Each sub-data set of a data category can be used for model training or evaluation of the AI ​​model corresponding to that data category. This at least one data analysis model can be provided by a model provider, stored on the first node, or obtained by the first node from the model provider's server (or other storage node); this application does not limit this.

[0124] The second node can deploy at least one data analysis model locally, based on the instructions of the first node. Then, using this at least one data analysis model, the local raw data is identified, analyzed, and categorized to identify different data categories of business data that conform to corresponding data characteristics, resulting in sub-data sets corresponding to different data categories. Each data category corresponds to a data distribution.

[0125] For each data category, the first node can lead at least two second nodes that store business data belonging to that data category as participating nodes, and instruct each participating node to use the business data of the corresponding data category stored locally to train and evaluate the AI ​​model corresponding to that data category in order to obtain the federated learning model corresponding to that data category.

[0126] The first node can add parameter information, such as the identifier of the target data category and the identifier of the target data analysis model, to the data information query message sent to each of the multiple second nodes. This is to instruct each second node to identify and classify the locally stored data according to the instructions and to provide feedback on the corresponding data distribution information. This allows the first node to select at least two target second nodes from the multiple second nodes based on the data distribution information provided by the multiple second nodes, and to participate in the training process of the AI ​​model corresponding to the target data category.

[0127] Furthermore, when instructing the at least two target second nodes to train the AI ​​model corresponding to the target data category, the first node can add parameter information, such as the identifier of the target data category and the identifier of the target data analysis model, to the model training message or model evaluation message sent to any one of the at least two target second nodes. This instructs each target second node to use the target business data corresponding to that target data category for model training or model evaluation, so as to obtain the federated learning model corresponding to that target data category. This solves the problem of decreased model accuracy after aggregation caused by different participants using data from different distributions to train the same AI model.

[0128] It is understood that the embodiments in this application are merely illustrative of how data identification and analysis can be achieved through data analysis models, and do not constitute any limitation on their specific implementation. In other embodiments, the first node may learn about the data distribution on the second node side through other means and lead the training and evaluation process of the various federated learning models corresponding to different data categories, which will not be elaborated here.

[0129] Before providing a detailed explanation of the embodiments of this application, the system structure involved in the embodiments of this application will be introduced first.

[0130] In one alternative implementation, the embodiments of this application can be applied to Figure 1 The federated learning system shown.

[0131] In one alternative implementation, the embodiments of this application can be applied to Figures 3A-3B The wireless AI model-driven network system shown herein can be implemented differently depending on the specific implementation of the functional modules included in the system when the embodiments of this application are applied to different scenarios. The following is a combination of... Figures 6A-6C Provide an example.

[0132] Example 1:

[0133] In one example, see Figure 6AThis application's embodiments can be applied to horizontal federated learning scenarios within an Enabler of Network Automation (eNA) architecture. eNA is a novel intelligent network architecture based on network data analytics function (NWDAF). In this scenario, horizontal federated learning can be performed between multiple local NWDAFs. Specifically, the central NWDAF implements the functionality of the MMF module; the data collection coordination function (DCCF) implements the functionality of the DMF module, collecting data from the corresponding network function (NF) module; and the local NWDAF implements the functionality of the MTF and MEF modules. It is understood that the functionality of the MTF and MEF modules can be implemented by the same local NWDAF or by different local NWDAF instances; this application does not limit this.

[0134] Example 2:

[0135] In one example, see Figure 6B This application's embodiments can be applied to a horizontal federated learning scenario within the user equipment (UE)-radio access network (RAN) context of the 3rd Generation Partnership Project (3GPP). In this scenario, horizontal federated learning is performed among multiple UEs. The RAN can implement the MMF module's functionality, while the UEs can implement the MTF, MEF, and DMF modules' functionality.

[0136] Example 3:

[0137] In one example, see Figure 6C This application's embodiments can be applied to federated learning scenarios within radio access networks (RANs) – element management / network management environments. In this scenario, horizontal federated learning can be performed across multiple RANs. Specifically, the element management system (EMS) / network management system (NMS) can implement the MMF module's functionality, while the RANs can implement the MTF, MEF, and DMF modules' functionality.

[0138] It is understandable that the examples above are only used in conjunction with specific application scenarios. Figures 3A-3B The illustrated functional modules are examples of specific implementations and not limitations on the relevant entities. In other embodiments, the relevant entities may also implement other functions, which will not be elaborated here.

[0139] based on Figure 1 , Figures 3A-3B , Figures 6A-6C The system architecture shown illustrates the federated learning scheme of this application, where the first node (or MMF module) and the second node (or the various functional modules of the second node) communicate with each other. However, due to differences in the information exchanged between the various functional modules, this scheme can have different implementations. The following section, in conjunction with the appendix... Figure 7-14 This paper provides a detailed description of the federated learning method provided in the embodiments of this application. It is understood that... Figure 7-14 In the flowchart shown, on the second node side, functional modules such as the MTF module, MEF module, and DMF module can be deployed in different (or not completely identical) entities (e.g.) Figure 6A The scenario shown can also be that the same entity is divided into functional modules such as MTF module, MEF module, and DMF module based on its functional logic (e.g., Figure 6B , Figure 6C (Scenario shown). Accordingly, communication between the MMF module on the first node side and the various functional modules on the second node side can be achieved through the communication interface with the entity where each functional module is located.

[0140] It should be noted that, in this embodiment, during the model training phase, an updated AI model is obtained by training the target AI model. The updated AI models from each second node are aggregated at the first node to obtain a federated learning model. In the next iteration, the first node distributes this federated learning model to each target second node as the target AI model to be trained. This process is repeated until the training task ends, resulting in a federated learning model corresponding to the target data category. During the model evaluation phase, after the first node triggers the evaluation task, it can distribute the aggregated federated learning model to each target second node as the AI ​​model to be evaluated. This allows each target second node to evaluate the federated learning model using the target business data corresponding to the respective target data category.

[0141] It should be noted that the AI ​​model identifiers in the mapping table maintained by the MMF module correspond to different AI models to be trained or evaluated at different stages.

[0142] Example 1

[0143] In this embodiment, the MMF module can add a data analysis model deployment message to the communication interface with the DMF module. This data analysis model deployment message can be used to deploy a data analysis model in the DMF module, so that the DMF module can use the deployed data analysis model to identify, analyze and classify the locally stored data to distinguish the data categories of business data that meet the corresponding data characteristics.

[0144] During the participant selection phase, when requesting data from multiple DMF modules, the MMF module can include the identifiers of the target data features and the target data analysis model required for the training task in the data query message (i.e., the first query message) sent to any DMF module. This allows the multiple DMF modules to use their respective target data analysis models to identify and analyze the full dataset that matches the target data features, dividing it into sub-data sets belonging to at least one data category. The MMF modules then include the identifier of that at least one data category and the data information of the sub-data sets belonging to that category in the data distribution information fed back to the MMF module. Furthermore, based on the data distribution information fed back by different DMF modules, the MMF module can select at least two suitable target DMF modules as target participants (since DMF modules correspond to MTF modules, selecting target DMF modules is equivalent to selecting the target second node and other functional modules), participating in the training and evaluation process of the federated learning model corresponding to the target data category.

[0145] During the local model training phase, when the MMF module requests local model training from the MTF module of the target participant, it can also include the identifiers of the target data features and the target data analysis model required for the training task in the model training message sent to the target MTF module. This allows the MTF module to select target business data belonging to the target data category according to the instructions, and use the target business data to train the AI ​​model corresponding to the target data category locally. This avoids the final model being affected by data poisoning due to differences in data distribution among different participants.

[0146] See Figure 7 When a training task is triggered, the steps of a federated learning method may include:

[0147] S700: Determine the target data features and target data categories based on the training task.

[0148] It is understood that in the embodiments of this application, the target data category can be determined based on the training task, or it can be determined by combining the training task and the data distribution information of multiple DMF modules after receiving the data distribution information of multiple DMF modules. This application does not limit this. Furthermore, the target data feature can refer to a single data feature or a group of data features (including multiple data features). In the case of a data feature group, the multiple data features included in the data feature group can have corresponding relationships or they can be unrelated to each other. This application does not limit this either.

[0149] S701: The MMF module acquires the data analysis model and the initial AI model based on the training task, and establishes a mapping relationship table.

[0150] The MMF module can acquire at least one data analysis model and / or at least one initial AI model, depending on the requirements of the training task. Both the data analysis model and / or the initial AI model can be provided by the model provider. The MMF module can store the corresponding model files locally, or it can obtain the corresponding models from the model provider's server or other storage devices.

[0151] Data analysis models are clustering or classification models. Each data analysis model corresponds to a set of data features and is used to identify data categories of business data that conform to the corresponding set of data features, obtaining at least one data category. Each data category corresponds to a data distribution and an initial AI model. The initial AI model corresponding to each data category can be used to learn and predict the data set of that data category.

[0152] Each data analysis model can be used to divide the full dataset that meets the corresponding data characteristics into sub-data sets belonging to different data categories, and different data categories can be distinguished by different identifiers. This mapping table can be used to record the mapping relationship between the identifiers of data features, AI models, data analysis models, and data categories. The identifier of the AI ​​model can be the name of the AI ​​model (including the name of the model used for training or the name of the model used for testing), and the identifier of the data category can be a data class index, for example, the identifier of the data category can be 1, 2, 3, etc., to indicate which data category it is.

[0153] S702: The MMF module sends a data analysis model deployment message to each of the multiple DMF modules and receives feedback from the DMF modules.

[0154] The data analysis model deployment message may include: an identifier for at least one data analysis model, and a model file for the at least one data analysis model, to deploy the at least one data analysis model in the corresponding DMF module. Feedback from the DMF module may include a notification message indicating that the at least one data analysis model has been successfully deployed locally.

[0155] S703: The MMF module sends training model deployment messages to each of the multiple MTF modules and receives feedback from the MTF modules.

[0156] The training model deployment message includes: an identifier for at least one initial AI model, and a model file for the at least one initial AI model. It is understood that the implementation steps of S703 are not limited to this; for example, they can be completed at any stage between S701 and S707.

[0157] The participant selection process includes:

[0158] The participant selection process can be implemented in two ways:

[0159] In one optional implementation, S704a: The MMF module sends a data information query message (i.e., the first query message) to each of the multiple DMF modules.

[0160] The data information query message may include: the identifier of the target data feature, the identifier of the target data category, the identifier of the target data analysis model, and query indicator (object) information, etc.

[0161] The identifier of the target data feature is used to indicate the data features that the target business data required for the training task conforms to, so that the DMF module can obtain the full set of data that conforms to the target data feature from the original data set stored locally.

[0162] The identifier of the target data category is used to indicate the data information of the target business data belonging to the target data category in the full dataset that meets the characteristics of the target data.

[0163] The identifier of the target data analysis model is used to instruct the DMF module to use the corresponding target data analysis model to identify, analyze and classify the full dataset that meets the target data characteristics, so as to divide the full dataset into sub-data sets belonging to different data categories.

[0164] The query indicator information is used to indicate the data information that needs to be queried for each data category, such as the size of the sub-data set, the data generation period, etc.

[0165] After receiving the data information query message, the DMF module uses the corresponding target data analysis model to identify, analyze, and classify the corresponding full dataset based on the relevant indication information contained in the query message. This results in the data distribution information that needs to be fed back to the MMF module. The data distribution information may include the identifier of the target data category, as well as the data information of the target business data that conforms to the characteristics of the target data and belongs to the target data category stored by the DMF module, including the size of the sub-data set and the data generation period.

[0166] S705a: Each DMF module returns a data information query confirmation message to the MMF module.

[0167] The data information query confirmation message includes the corresponding data distribution information found, including the identifier of the target data category, and the data information of the target business data that conforms to the characteristics of the target data and belongs to the target data category stored in the DMF module, including the size of the sub-data set, the data generation period, etc.

[0168] In one optional implementation, S704b: The MMF module sends a data information query message (i.e., the first query message) to each of the multiple DMF modules.

[0169] The data information query message may include: the identifier of the target data feature (data name), the identifier of the target data analysis model, and query indicator (object) information, etc.

[0170] Since the data information query message does not indicate the identifier of the target data category, the DMF module can query the data information of business data belonging to at least one data category obtained after identifying the full set of data that meets the characteristics of the target data based on the target data analysis model, including the size of the sub-data set, the data generation period, etc.

[0171] S705b: Each DMF module returns a data information query confirmation message to the MMF module.

[0172] The data information query confirmation message carries the data distribution information of the DMF module. This data distribution information can be implemented as a query result list, which includes: an identifier of at least one data category, and data information of business data belonging to the at least one data category respectively.

[0173] It is understood that, in specific implementation, the MMF module can choose to obtain the corresponding data distribution information of any DMF module in any of the implementation methods of S704a-S705a or S704b-S705b, and this application does not limit it in this regard.

[0174] S706: The MMF module, based on the data distribution information fed back by multiple DMF modules, uses its internal algorithm to select at least two target participants (including at least two target second nodes and their respective corresponding functional modules). That is, it selects which participants will participate in the model training process of the federated learning model corresponding to the target data category.

[0175] Understandably, the data category identifiers and corresponding data information included in the data distribution information returned by multiple DMF modules can serve as one of the decision-making criteria. By combining the existence of the target business data required for the training task in the business data locally stored by each DMF module, and the data volume of the corresponding sub-dataset, at least two target participants corresponding to the target data category are selected. These at least two target participants can participate in the current model training process to obtain the federated learning model corresponding to the target data category.

[0176] For the target data category, after selecting at least two target participants, the local model training process includes:

[0177] S707: The MMF module sends model training messages to any target MTF module.

[0178] The model training message includes: the identifier of the target AI model, which corresponds to the target data category; the model file of the target AI model; the identifier of the target data category; the identifier of the target data analysis model; and training data volume indication information, etc.

[0179] The identifier of the target data category can be used to indicate the data category to which the sub-data set used by the target MTF module for model training belongs.

[0180] The identifier of the target data analysis model can be used to indicate the target data analysis model required by the target MTF module to obtain the sub-data set belonging to the target data category.

[0181] Training data quantity indication information is used to indicate the amount of data required to train the model.

[0182] S708: Any target MTF module sends a data query message (i.e., the second query message) to the corresponding target DMF module.

[0183] The data query message includes: identifiers of target data features, data type (type) indications, identifiers of target data categories, training data volume indications, and identifiers of target data analysis models.

[0184] The data type indication information can include two types: indication for training (train) or indication for testing (test), used to obtain training data sets and test data sets, respectively. In this embodiment, for ease of distinction, the data type indication information used to indicate training is referred to as the first data type indication information, and the data type indication information used to indicate testing is referred to as the second data type indication information. In S708, this data type is the first data type indication information, used to indicate training (train) in order to obtain the training data set.

[0185] S709: The target DMF module sends a data query confirmation message to the corresponding target MTF module.

[0186] The data query confirmation message includes the target business data that meets the target data characteristics and belongs to the target data category, which is the sub-data set corresponding to the target data category and is the training data set.

[0187] S710: Any target MTF module uses the sub-data set corresponding to the target data category returned by the corresponding target DMF module to train the target AI model and obtain the updated AI model.

[0188] S711: After the target model training is completed, any target MTF module sends a training completion notification message to the MMF module.

[0189] The training completion notification message includes the identifier of the trained AI model, the model file of the updated AI model, and the amount of data in the training dataset used to train the AI ​​model.

[0190] S712: After collecting the training completion notification messages returned by at least two target MTF modules participating in this round of model training, the MMF module aggregates the updated AI models returned by each target MTF module using an aggregation algorithm (such as federated averaging) and updates the parameters of the federated learning model.

[0191] For example, in this aggregation algorithm, the size of the training dataset used by each target MTF module to train the corresponding AI model can be used to determine the weight factor of the updated AI model trained by the corresponding target MTF module. This weight factor is used to represent the weight of the corresponding AI model when performing aggregation processing.

[0192] S713: The MMF module determines whether the conditions for stopping training are met. If not, it returns to repeat S704-S712 to iterate through the next round of participant selection, model training, and model aggregation until the conditions for stopping training are met, at which point the current process ends, i.e., the current model training process ends, completing the training task for the target data category and obtaining the federated learning model corresponding to that target data category.

[0193] Example 2

[0194] This embodiment is an improvement based on Embodiment 1. For similarities, please refer to the combined... Figure 7 The relevant descriptions will not be repeated below.

[0195] In this embodiment, the MMF module can send a mapping table to each of the multiple MTF modules. This mapping table represents the relationship between four identifiers: data feature identifier, AI model identifier, data analysis model identifier, and data category identifier. Therefore, in the model training message sent by the MMF module to each target MTF module, it is no longer necessary to specify the identifiers of the target data category and the target data analysis model required for the current model training. Thus, when the MMF module needs to frequently send model training messages to each target MTF module, the amount of message transmission between communication interfaces can be effectively reduced, thereby reducing signaling overhead.

[0196] See Figure 8 When a training task is triggered, the steps of a federated learning method may include:

[0197] S800: Based on the training task, determine the target data features and target data categories. For a detailed explanation, please refer to S700 above; it will not be repeated here.

[0198] S801: The MMF module acquires the data analysis model and the initial AI model based on the training task, and establishes a mapping table. For a detailed explanation, please refer to S701 above; it will not be repeated here.

[0199] S802: The MMF module sends a data analysis model deployment message to each of the multiple DMF modules and receives feedback from the DMF modules. For a detailed explanation, please refer to S702 above; it will not be repeated here.

[0200] S803: The MMF module sends training model deployment messages to each of the multiple MTF modules and receives feedback from the MTF modules.

[0201] The training model deployment message includes: an identifier for at least one initial AI model; a model file for the at least one initial AI model; and a mapping table.

[0202] This mapping table can record the identifiers of the data categories and the data analysis models corresponding to at least one initial AI model, making it convenient to find the identifiers of the target data categories and the target data analysis models corresponding to the identifiers of the target AI model in subsequent processes.

[0203] It is understood that the implementation steps of S803 are not limited to this; for example, it can be completed at any stage between S801 and S807. It is also understood that, in specific implementations, depending on the actual method steps, in S803, the MMF module may send the mapping table to each DMF module, or it may only send the mapping relationship between the identifier of the target AI model, the identifier of the target data category, and the identifier of the target data analysis model to each DMF module; this application does not limit this.

[0204] The participant selection process includes:

[0205] The participant selection process can be implemented in two ways:

[0206] In one optional implementation: S804a: The MMF module sends a data information query message (i.e., the first query message) to each of the multiple DMF modules.

[0207] The data information query message may include: the identifier of the target data feature, the identifier of the target data category, the identifier of the target data analysis model, and query indicator information, etc.

[0208] S805a: Each DMF module returns a data information query confirmation message to the MMF module.

[0209] The data information query confirmation message includes the corresponding data distribution information found, including the identifier of the target data category, and the data information of the target business data that conforms to the characteristics of the target data and belongs to the target data category stored in the DMF module, including the size of the sub-data set, the data generation period, etc.

[0210] In one alternative implementation:

[0211] S804b: The MMF module sends a data information query message (i.e., the first query message) to each of the multiple DMF modules.

[0212] The data information query message may include: the identifier of the target data features, the identifier of the target data analysis model, and query indicator information, etc.

[0213] S805b: Each DMF module returns a data information query confirmation message to the MMF module.

[0214] The data information query confirmation message carries the data distribution information of the DMF module. This data distribution information can be implemented as a query result list, which includes: an identifier of at least one data category, and data information of business data belonging to the at least one data category stored in the DMF module.

[0215] It is understood that, in specific implementations, the MMF module can choose to obtain the corresponding data distribution information of any DMF module using any of the implementation methods S804a-S805a or S804b-S805b, and this application does not limit this. For a detailed explanation of S804a-S805a or S804b-S805b, please refer to the above S704a-S705a or S704b-S705b, which will not be repeated here.

[0216] S806: The MMF module, based on the data distribution information fed back by multiple DMF modules, uses its internal algorithm to select at least two target participants (including at least two target second nodes and their respective corresponding functional modules). That is, it selects which participants will participate in the model training process of the federated learning model corresponding to the target data category.

[0217] Understandably, the data category identifiers and corresponding data information included in the data distribution information returned by multiple DMF modules can serve as one of the decision-making criteria. By combining the existence of the target business data required for the training task in the business data stored locally by each DMF, and the data volume of the corresponding sub-dataset, at least two target participants corresponding to the target data category are selected. These at least two target participants can participate in the current model training process to obtain the federated learning model corresponding to the target data category.

[0218] For the target data category, after selecting at least two target participants, the local model training process includes:

[0219] S807: The MMF module sends model training messages to any target MTF module.

[0220] The model training message includes: the identifier of the target AI model, which corresponds to the target data category; the model file of the target AI model; and training data volume indication information, etc. For a detailed explanation, please refer to S707 above, which will not be repeated here.

[0221] S808: The target MTF module uses the mapping table to find the identifier of the target data category and the identifier of the target data analysis model that correspond to the identifier of the target AI model contained in the model training message.

[0222] S809: Any target MTF module sends a data query message (i.e., the second query message) to the corresponding target DMF module.

[0223] The data query message includes: identifiers of target data features, data type information, identifiers of target data categories, training data volume information, and identifiers of the target data analysis model. For a detailed explanation, please refer to section S708 above; it will not be repeated here.

[0224] S810: Any target DMF module sends a data query confirmation message to the corresponding target MTF module.

[0225] The data query confirmation message includes target business data that meets the target data characteristics and belongs to the target data category, obtained through the target data analysis model. This is the sub-data set corresponding to the target data category, which serves as the training data set. For a detailed explanation, please refer to S709 above; it will not be repeated here.

[0226] S811: The target MTF module uses the sub-data set corresponding to the target data category returned by the target DMF module to train the target AI model, obtaining the updated AI model. For a detailed explanation, please refer to S710 above; it will not be repeated here.

[0227] S812: After the model training is completed, any target MTF module sends a training completion notification message to the MMF module.

[0228] The training completion notification message includes the identifier of the trained AI model, the updated AI model file, and the amount of data in the training dataset used to train the AI ​​model. For a detailed explanation, please refer to section S711 above; it will not be repeated here.

[0229] S813: After collecting training completion notification messages from at least two target MTF modules participating in this round of model training, the MMF module aggregates the updated AI models returned by each target MTF module using an aggregation algorithm (e.g., federated averaging) to update the parameters of the federated learning model. For a detailed explanation, please refer to S712 above; it will not be repeated here.

[0230] S814: The MMF module determines whether the conditions for stopping training are met. If not, it returns to repeat S804-S813 to proceed with the next round of participant selection, model training, and model aggregation until the conditions for stopping training are met, at which point the current process ends, i.e., the current model training process ends, completing the training task for the target data category and obtaining the federated learning model corresponding to the target data category.

[0231] Compared to Embodiment 1, since the MMF module has already sent a mapping table to each MTF module in S803, when any MTF module is selected as a participant and receives the model training message in S808, it can use the mapping table to find the identifier of the target data category and the identifier of the target data analysis model corresponding to the identifier of the target AI model contained in the model training message. It then carries these identifiers in the data query message sent to the DMF module, allowing the DMF module to obtain and feed back the target business data (i.e., the training data set) required for training the corresponding target AI model, based on the instructions in the received data query message. Therefore, when the MMF module needs to frequently send model training messages to each target MTF module, the amount of message transmission between communication interfaces can be effectively reduced, thereby reducing signaling overhead.

[0232] Example 3

[0233] This embodiment is an improvement based on Embodiment 1. For similarities, please refer to the combined... Figure 7 The relevant descriptions will not be repeated below.

[0234] In this embodiment, the MMF module can send a mapping table to each of the multiple DMF modules. This mapping table represents the relationship between four identifiers: data feature identifier, AI model identifier, data analysis model identifier, and data category identifier. Therefore, in the model training messages sent by the MMF module to each MTF module, and in the data query messages sent by each MTF module to the corresponding DMF module, it is no longer necessary to specify the identifiers of the target data category and the target data analysis model required for the current model training. Thus, when the MMF module needs to frequently send model training messages to each target MTF module, the amount of message transmission between communication interfaces can be effectively reduced, thereby reducing signaling overhead.

[0235] See Figure 9 When a training task is triggered, the steps of a federated learning method may include:

[0236] S900: Based on the training task, determine the target data features and target data categories. For a detailed explanation, please refer to S700 above; it will not be repeated here.

[0237] S901: The MMF module acquires the data analysis model and the initial AI model based on the training task, and establishes a mapping table. For a detailed explanation, please refer to S701 above; it will not be repeated here.

[0238] S902: The MMF module sends a data analysis model deployment message to each of the multiple DMF modules and receives feedback from the DMF modules.

[0239] The data analysis model deployment message may include: an identifier for at least one data analysis model; a model file for the at least one data analysis model to be deployed in the DMF module; and a mapping table. Feedback from the DMF module may include a notification message indicating that the at least one data analysis model has been successfully deployed locally.

[0240] This mapping table can record the identifiers of the data categories and the data analysis models corresponding to at least one initial AI model, making it convenient to find the identifiers of the target data categories and the target data analysis models corresponding to the identifiers of the target AI model in subsequent processes.

[0241] S903: The MMF module sends the training model deployment message to each of the multiple MTF modules and receives feedback from the MTF modules. For a detailed explanation, please refer to S703 above, which will not be repeated here. It is understood that the implementation steps of S903 are not limited to this; for example, it can be completed at any stage between S901 and S907.

[0242] The participant selection process includes:

[0243] The participant selection process can be implemented in two ways:

[0244] In one optional implementation: S904a: The MMF module sends a data information query message (i.e., the first query message) to each of the multiple DMF modules.

[0245] The data information query message may include: the identifier of the target data feature, the identifier of the target data category, the identifier of the target data analysis model, and query indicator information, etc.

[0246] S905a: Each DMF module returns a data information query confirmation message to the MMF module.

[0247] The data information query confirmation message includes the corresponding data distribution information found, including the identifier of the target data category, and the data information of the target business data that conforms to the characteristics of the target data and belongs to the target data category stored in the DMF module, including the size of the sub-data set, the data generation period, etc.

[0248] In one alternative implementation:

[0249] S904b: The MMF module sends a data information query message (i.e., the first query message) to each of the multiple DMF modules.

[0250] The data information query message may include: the identifier of the target data feature (data name), the identifier of the target data analysis model, and query indicator (object) information, etc.

[0251] S905b: The DMF module returns a data information query confirmation message to the MMF module.

[0252] The data information query confirmation message carries the data distribution information of the DMF module. This data distribution information can be implemented as a query result list, which includes: an identifier of at least one data category, and data information of business data belonging to the at least one data category stored in the DMF module.

[0253] It is understood that, in specific implementations, the MMF module can choose to obtain the corresponding data distribution information of any DMF module using any of the implementation methods S904a-S905a or S904b-S905b, and this application does not impose any restrictions on this. For detailed explanations, please refer to S704a-S705a or S704b-S705b above, which will not be repeated here.

[0254] S906: The MMF module, based on the data distribution information fed back by multiple DMF modules, uses its internal algorithm to select at least two target participants (including at least two target second nodes and their respective corresponding functional modules). That is, it selects which participants will participate in the model training process of the federated learning model corresponding to the target data category.

[0255] Understandably, the data category identifiers and corresponding data information included in the data distribution information returned by multiple DMF modules can serve as one of the decision-making criteria. By combining the existence of the target business data required for the training task in the business data stored locally by each DMF, and the data volume of the corresponding sub-dataset, at least two target participants corresponding to the target data category are selected. These at least two target participants can participate in the current model training process to obtain the federated learning model corresponding to the target data category.

[0256] For the target data category, after selecting at least two target participants, the local model training process includes:

[0257] S907: The MMF module sends a model training message to any target MTF module.

[0258] The model training message includes: the identifier of the target AI model, which corresponds to the target data category; the model file of the target AI model; and training data volume indication information, etc. For a detailed explanation, please refer to S707 above, which will not be repeated here.

[0259] S908: The target MTF module sends a data query message (i.e., the second query message) to the corresponding target DMF module. This data query message includes: the identifier of the target data features, the identifier of the target AI model, and data type indication information. In S908, this data type is the first data type indication information, used to indicate training in order to obtain the training dataset.

[0260] S909: After receiving the data query message, the target DMF module, according to the mapping table, looks up the identifier of the target data category and the identifier of the target data analysis model corresponding to the identifier of the target AI model contained in the model training message. Based on the identifiers of the target data category and the target data analysis model, it obtains the target business data belonging to the target data category.

[0261] S910: The target DMF module sends a data query confirmation message to the corresponding target MTF module.

[0262] The data query confirmation message includes the target business data that was retrieved and belongs to the target data category, i.e., the sub-data set corresponding to the target data category, which is the training data set. For a detailed explanation, please refer to S709 above, which will not be repeated here.

[0263] S911: Any target MTF module uses the sub-data set corresponding to the target data category returned by the target DMF module to train the target AI model, obtaining the updated AI model. For a detailed explanation, please refer to S710 above; it will not be repeated here.

[0264] S912: After model training is complete, any target MTF module sends a training completion notification message to the MMF module. This notification message includes the identifier of the trained AI model, the updated AI model file, and the amount of data in the training dataset used to train the AI ​​model. For a detailed explanation, please refer to S711 above, which will not be repeated here.

[0265] S913: After collecting training completion notification messages from at least two target MTF modules participating in this round of model training, the MMF module aggregates the updated AI models returned by each target MTF module using an aggregation algorithm (e.g., federated averaging) to update the parameters of the federated learning model. For a detailed explanation, please refer to S712 above; it will not be repeated here.

[0266] S914: The MMF module determines whether the conditions for stopping training are met. If not, it returns to repeat S904-S913 to proceed with the next round of participant selection, model training, and model aggregation until the conditions for stopping training are met, at which point the current process ends, i.e., the current model training process ends, completing the training task for the target data category and obtaining the federated learning model corresponding to the target data category.

[0267] Compared to Embodiment 1, since the MMF module has already sent a mapping table to each DMF module in S902, when any MTF module is selected as a participant and receives the model training message in S907, it can include the identifier of the target AI model in the data query message sent to the DMF module. This allows the DMF module to look up the identifier of the target data category and the identifier of the target data analysis model corresponding to the identifier of the target AI model based on the mapping table. Then, based on the identifier of the target data category and the identifier of the target data analysis model, it feeds back the target business data (i.e., the training dataset) required for training the target AI model to the corresponding MTF module, thereby obtaining the federated learning model corresponding to the target data category. Therefore, when the MMF module needs to frequently send model training messages to each target MTF module, the amount of message transmission between communication interfaces can be effectively reduced, thus reducing signaling overhead.

[0268] Example 4

[0269] This embodiment illustrates a method for training a horizontal federated learning model when the data analysis model and the AI ​​model to be trained are jointly deployed on an MTF module, provided they cannot be separated. It is understood that in this embodiment, the data analysis function can be implemented by the AI ​​model itself, or the AI ​​model can be bound to the data analysis model; this application does not limit this approach.

[0270] In this embodiment, the AI ​​model to be trained is deployed in each MTF module. Since the data analysis model and the AI ​​model to be trained cannot be separated, a data analysis request message can be added between the DMF module and the MTF module. This allows the DMF module to utilize the data analysis model deployed in the corresponding MTF module to identify, analyze, and classify the full dataset that meets the corresponding data characteristics, obtain the data distribution information of the business data stored locally in the DMF, and then feed this data distribution information back to the MMF module. Furthermore, the MMF module can select appropriate target participants to participate in the model training process of the federated learning model for the target data category based on the data distribution information of different DMF modules. During the local model training phase, when instructing the MTF module to perform local model training, the MMF module can also include the identifier of the target data category and the identifier of the target data analysis model in the model training message sent to the MTF module. This allows the MTF module to select target business data belonging to the target data category for local model training of the target AI model based on the identifier of the target data category and the identifier of the target data analysis model, thereby avoiding the data poisoning effect on the final federated learning model due to differences in data distribution among different participants.

[0271] See Figure 10 When a training task is triggered, the steps of a federated learning method may include:

[0272] S1000: Based on the training task, determine the target data features and target data categories. For a detailed explanation, please refer to S700 above; it will not be repeated here.

[0273] S1001: The MMF module obtains the model package required for the training task and maintains the mapping table.

[0274] The model package includes at least one data analysis model, an initial AI model corresponding to each data category, and a mapping relationship between data category identifiers and AI model identifiers. The mapping table maintained by the MMF module records the mapping relationships between data feature identifiers, AI model identifiers, data analysis model identifiers, and data category identifiers. This includes the data category identifiers and data analysis model identifiers corresponding to at least one AI model, facilitating the subsequent lookup of the target data category identifier and target data analysis model identifier corresponding to the target AI model identifier using this mapping table.

[0275] S1002: The MMF module sends a training model deployment message to each of the multiple MTF modules and receives feedback from the DMF module.

[0276] The training model deployment message includes: an identifier for at least one data analysis model and a model file for at least one data analysis model, an identifier for the initial AI model corresponding to each data category and the corresponding model file, and a mapping table.

[0277] The participant selection process includes:

[0278] The participant selection process can be implemented in two ways:

[0279] In one optional implementation, S1003a: The MMF module sends a data information query message (i.e., the first query message) to each of the multiple DMF modules.

[0280] The data information query message may include: the identifier of the target data feature, the identifier of the target data category, the identifier of the target data analysis model, and query indicator information, etc.

[0281] S1004a: Each DMF module sends a data analysis request message to the corresponding MTF module.

[0282] The data analysis request includes the identifier of the target data category, the identifier of the target data analysis model, and the full set of business data containing the symbolic target data characteristics stored in DMF.

[0283] S1005a: Based on the identifier of the target data analysis model, the MTF module uses the corresponding target data analysis model to identify, analyze, and classify the full dataset, obtains a set of sub-data of business data belonging to the target data category, and sends a data analysis request confirmation message back to the corresponding DMF module.

[0284] The data analysis request confirmation message includes a set of sub-data points corresponding to the respective data categories obtained after dividing the full dataset according to the data analysis model.

[0285] S1006a: After analyzing each sub-data set, the DMF module sends a data information query confirmation message to the MMF module.

[0286] The data information query confirmation message includes the corresponding data distribution information found, including the identifier of the target data category, and the data information of the target business data that conforms to the characteristics of the target data and belongs to the target data category stored in the DMF module, including the size of the sub-data set, the data generation period, etc.

[0287] In one optional implementation, S1003b: The MMF module sends a data information query message (i.e., the first query message) to each of the multiple DMF modules.

[0288] The data information query message may include: the identifier of the target data features, the identifier of the target data analysis model, and query indicator (object) information, etc.

[0289] S1004b: Each DMF module sends a data analysis request message to the corresponding MTF module.

[0290] The data analysis request includes the identifier of the target data analysis model and a full set of business data containing symbolic target data characteristics stored in DMF.

[0291] S1005b: Based on the identifier of the target data analysis model, the MTF module uses the corresponding target data analysis model to identify, analyze, and classify the full dataset, obtaining a set of sub-data belonging to the target data category, and then sends a data analysis request confirmation message back to the corresponding DMF module. This data analysis request confirmation message includes the sub-data set corresponding to the relevant data category obtained after dividing the full dataset according to the target data analysis model.

[0292] S1006b: After analyzing each sub-data set, the DMF module sends a data information query confirmation message to the MMF module.

[0293] The data information query confirmation message carries the data distribution information of the DMF module. This data distribution information can be implemented as a query result list, which includes: an identifier of at least one data category, and data information of business data belonging to the at least one data category stored in the DMF module.

[0294] It is understood that, in specific implementation, the MMF module can choose to obtain the corresponding data distribution information of any DMF module in any of the implementation methods of S1003a-S1006a or S1003b-S1006b, and this application does not limit it in this regard.

[0295] S1007: The MMF module, based on the data distribution information fed back by multiple DMF modules, uses its internal algorithm to select at least two target participants (including at least two target second nodes and their respective corresponding functional modules). That is, it selects which participants will participate in the model training process of the federated learning model corresponding to the target data category.

[0296] Understandably, the data category identifiers and corresponding data information included in the data distribution information returned by multiple DMF modules can serve as one of the decision-making criteria. By combining the existence of the target business data required for the training task in the business data stored locally by each DMF, and the data volume of the corresponding sub-dataset, at least two target participants corresponding to the target data category are selected. These at least two target participants can participate in the current model training process to obtain the federated learning model corresponding to the target data category.

[0297] For the target data category, after selecting at least two target participants, the local model training process includes:

[0298] S1008: The MMF module sends a model training message to any target MTF module.

[0299] The model training message includes: the identifier of the target AI model, which corresponds to the target data category; the model file of the target AI model; and training data volume indication information, etc. For a detailed explanation, please refer to S707 above, which will not be repeated here.

[0300] S1009: The target MTF module sends a data query message (i.e., the second query message) to the corresponding target DMF module.

[0301] The data query message includes: identifiers of target data features, data type, target data category, target data analysis model, and training data volume information. For a detailed explanation, please refer to section S708 above; it will not be repeated here.

[0302] S1010: The target DMF module sends a data query confirmation message back to the corresponding MTF module.

[0303] The data query confirmation message includes the target business data that meets the target data characteristics and belongs to the target data category, which is the sub-data set corresponding to the target data category and is the training data set.

[0304] S1011: The target MTF module uses the sub-data set corresponding to the target data category returned by the target DMF module to train the target AI model and obtain the updated AI model.

[0305] S1012: After the target model training is completed, any target MTF module sends a training completion notification message to the corresponding target MMF module. The training completion notification message includes the identifier of the trained AI model, the model file of the updated AI model, and the amount of data in the training dataset used to train the AI ​​model.

[0306] S1013: After collecting the training completion notification messages returned by at least two target MTF modules participating in this round of model training, the MMF module aggregates the updated AI models returned by each target MTF module using an aggregation algorithm (such as federated averaging) and updates the parameters of the federated learning model.

[0307] For example, in this aggregation algorithm, the size of the training dataset used by each target MTF module to train the corresponding AI model can be used to determine the weight factor of the updated AI model trained by the corresponding target MTF module. This weight factor is used to represent the weight of the corresponding AI model when performing aggregation processing.

[0308] S1014: The MMF module determines whether the conditions for stopping training are met. If not, it returns to repeat S704-S712 to iterate through the next round of participant selection, model training, and model aggregation until the conditions for stopping training are met, at which point the current process ends, i.e., the current model training process ends, completing the training task for the target data category and obtaining the federated learning model corresponding to the target data category.

[0309] Example 5

[0310] This embodiment is another implementation scheme with the same scenario as Embodiment 4. The difference from Embodiment 4 is that in Embodiment 5, the MMF module can directly request data information from the MTF module.

[0311] In this embodiment, the AI ​​model to be trained is deployed in each MTF module. Since the data analysis model and the AI ​​model to be trained cannot be separated, a data information query message is added between the MMF module and the MTF module. The MTF module can then provide the MMF module with data distribution information of the business data stored in the corresponding DMF module. Furthermore, the MMF module can select appropriate target participants to join the model training process of the federated learning model for the target data category based on the data distribution information of different DMF modules. During the local model training phase, when requesting local model training from the MTF module, the MMF module can also include the identifier of the target data category and the identifier of the target data analysis model in the model training message sent to the MTF module. This allows the MTF module to select a sub-data set belonging to the corresponding target data category for local model training based on the identifier of the target data category and the identifier of the target data analysis model, thereby avoiding data poisoning of the final model due to differences in data distribution among different participants.

[0312] See Figure 11 When a training task is triggered, the steps of a federated learning method may include:

[0313] S1100: Based on the training task, determine the target data features and target data categories. For a detailed explanation, please refer to S700 above; it will not be repeated here.

[0314] S1101: The MMF module acquires the model packages required for the training task and maintains a mapping table. For a detailed explanation, please refer to S1001 above; it will not be repeated here.

[0315] S1102: The MMF module sends a training model deployment message to each of the multiple MTF modules and receives feedback from the DMF module.

[0316] The training model deployment message includes: an identifier for at least one data analysis model and a model file for at least one data analysis model, an identifier for the initial AI model corresponding to each data category and the corresponding model file, and a mapping table.

[0317] The participant selection process includes:

[0318] The participant selection process can be implemented in two ways:

[0319] In one optional implementation, S1103a: The MMF module sends a data information query message (i.e., the first query message) to any target MTF module.

[0320] The data information query message may include: the identifier of the target data feature, the identifier of the target data category, the identifier of the target data analysis model, and query indicator information, etc.

[0321] S1104a: Each target MTF module sends a data query message to the corresponding target DMF module.

[0322] The data query message includes an identifier of the target data characteristics.

[0323] S1105a: The target DMF module sends a data query confirmation message to the corresponding target MTF module.

[0324] The data query confirmation message includes the full set of data that matches the characteristics of the target data, stored in the target DMF module.

[0325] S1106a: Based on the identifier of the target data analysis model, the target MTF module uses the corresponding target data analysis model to identify, analyze, and classify the full dataset, obtains a set of sub-data of business data belonging to the target data category, and sends a data information query confirmation message back to the MTF module.

[0326] The data information query confirmation message includes the corresponding data distribution information retrieved, including the identifier of the target data category, and the data information of the target business data that conforms to the characteristics of the target data and belongs to the target data category stored in the DMF module, including the size of the sub-data set, the data generation period, etc.

[0327] In one optional implementation, S1103b: The MMF module sends a data information query message (i.e., the first query message) to the target MTF module. This data information query message may include: the identifier of the target data feature, the identifier of the target data analysis model, and query indicator (object) information, etc.

[0328] S1104b: The target MTF module sends a data query message to the corresponding target DMF module.

[0329] The data query message includes an identifier of the target data characteristics.

[0330] S1105b: The target DMF module sends a data query confirmation message to the corresponding target MTF module. This data query confirmation message includes the full set of data stored by the target DMF module that conforms to the characteristics of the target data.

[0331] S1106b: Based on the identifier of the target data analysis model, the target MTF module uses the corresponding target data analysis model to identify, analyze and classify the full dataset, obtains a set of sub-data belonging to at least one data category, and sends a data information query confirmation message back to the MTF module.

[0332] The data information query confirmation message carries the data distribution information of the DMF module. This data distribution information can be implemented as a query result list, which includes: an identifier of at least one data category, and data information of business data belonging to the at least one data category respectively.

[0333] It is understandable that, in specific implementation, the MMF module can choose to obtain the corresponding data distribution information of any DMF using any of the implementation methods S1103a-S1106a or S1103b-S1106b, and this application does not limit this.

[0334] S1107: The MMF module, based on the data distribution information fed back by multiple DMF modules, uses its internal algorithm to make decisions to select at least two target participants (including at least two target second nodes and their respective corresponding functional modules). That is, it selects which participants will participate in this model training process.

[0335] Understandably, the data category identifiers and corresponding data information included in the data distribution information returned by the MTF module can serve as one of the decision-making criteria. By combining factors such as whether the target business data required for the training task exists in the business data locally stored in each DMF, and the data volume of the sub-dataset, at least two target participants can be selected. These at least two target participants can participate in the current model training process to obtain the federated learning model corresponding to the target data category.

[0336] For the target data category, after selecting at least two target participants, the local model training process includes:

[0337] S1108: The MMF module sends a model training message to any target MTF module.

[0338] The model training message includes: the identifier of the target AI model, which corresponds to the target data category; the model file of the target AI model; and training data volume indication information, etc. For a detailed explanation, please refer to S707 above, which will not be repeated here.

[0339] S1109: Each target MTF module sends a data query message to the corresponding target DMF module.

[0340] The data query message includes: an identifier of the target data feature.

[0341] S1110: The target DMF module sends a data query confirmation message back to the corresponding MTF module.

[0342] The data query confirmation message includes a complete set of business data that matches the characteristics of the target data.

[0343] S1111: Each target MTF module, based on the mapping table, looks up the identifier of the target data category and the identifier of the target data analysis model corresponding to the identifier of the target data feature. Based on the identifier of the target data analysis model, it uses the corresponding target data analysis model to identify, analyze, and classify the full dataset, thereby obtaining a sub-data set of business data belonging to the target data category. The obtained sub-data set is then used to train the target AI model, resulting in an updated AI model.

[0344] S1112: After the target model training is completed, any target MTF module sends a training completion notification message to the corresponding target MMF module. The training completion notification message includes the identifier of the trained AI model, the model file of the updated AI model, and the amount of data in the training dataset used to train the AI ​​model.

[0345] S1113: After collecting the training completion notification messages returned by at least two target MTF modules participating in this round of model training, the MMF module aggregates the updated AI models returned by each target MTF module using an aggregation algorithm (such as federated averaging) and updates the parameters of the federated learning model.

[0346] For example, in this aggregation algorithm, the size of the training dataset used by each target MTF module to train the corresponding AI model can be used to determine the weight factor of the updated AI model trained by the corresponding target MTF module. This weight factor is used to represent the weight of the corresponding AI model when performing aggregation processing.

[0347] S1114: The MMF module determines whether the conditions for stopping training are met. If not, it returns to repeat S1103-S1113 to iterate through the next round of participant selection, model training, and model aggregation until the conditions for stopping training are met, at which point the current process ends, i.e., the current model training process ends, completing the training task for the target data category and obtaining the federated learning model corresponding to the target data category.

[0348] Therefore, based on the above embodiments one through five, Figure 1 , Figures 3A-3B , Figures 6A-6C The system architecture shown allows the MMF module, acting as the first node in the coordinator, to interact with the MTF and DMF modules, acting as second nodes in the participant group. This allows the MMF module to obtain data distribution information from multiple candidate participants. Based on this data distribution information, the MMF module selects at least two suitable target participants to participate in the training of the federated learning model for the target data category. Furthermore, during the local model training phase, the coordinator guides each target participant to select a sub-dataset of the corresponding data category to train the AI ​​model for that data category locally. This ensures an updated AI model for the corresponding data category, preventing data poisoning of the final model due to differences in data distribution among participants.

[0349] It should be noted that in the above embodiments, the MEF module and the MTF module, which are used to implement the model evaluation function, are considered to be located in the same entity. After the MTF module completes local model training, during the model evaluation phase, the MEF module obtains a test data set belonging to the target data category to evaluate the updated AI model corresponding to the trained target data category, or to evaluate the federated learning model corresponding to the target data category issued by the MMF. Specifically, the way the MEF module obtains the test data set from the DMF during the model evaluation phase is basically the same as the process by which the MTF module obtains the training data set in Embodiments 1-5. Detailed implementation details can be found in the relevant descriptions above. It is understood that since the MEF module and the MTF module implement different functions, the names of the messages exchanged between the MEF module and other modules to obtain the corresponding test data set may be different from the names of the relevant messages in the MTF module.

[0350] Example 6

[0351] In some embodiments, the MTF module and the MEF module can be deployed separately, i.e., located in different entities. In this case, to facilitate the implementation of the model evaluation function, the MEF module needs to interact with other modules to obtain the evaluation task and the test data set required to complete the corresponding evaluation task. Therefore, based on embodiments one through five above, a model evaluation task can be added during model training to evaluate the obtained model. It is understood that the evaluation task can be triggered by the MTF module as needed, and can be sent to the MEF module at any time as required; this application does not limit this.

[0352] See Figure 12 When an evaluation task is triggered, the steps of a federated learning method may include:

[0353] S1200: Evaluation task triggered.

[0354] S1201: The MMF module sends an evaluation model deployment message to any target MEF module.

[0355] The deployment message for the evaluation model includes the identifier of the target evaluation model and the corresponding model file. This target evaluation model can be a federated learning model corresponding to the specific target data category, or an updated AI model obtained by training the target MTF corresponding to the target MEF module.

[0356] S1202: The MMF module sends a model evaluation message to any target MEF module.

[0357] The model evaluation message includes the identifier of the target evaluation model, evaluation indicator information, the identifier of the target data category corresponding to the target evaluation model, and the identifier of the target data analysis model.

[0358] S1203: The target MEF module sends a data query message to the corresponding target DMF module.

[0359] The data query message includes: the identifier of the target data features, the data type, the identifier of the target data category, the training data volume indication information, and the identifier of the target data analysis model.

[0360] The data type can include two types: training (train) and test (test), used to obtain the training data set and the test data set, respectively. In S1203, the data type is test (test) to obtain the test data set.

[0361] S1204: The target DMF module sends a data query confirmation message to the corresponding target MEF module.

[0362] The data query confirmation message includes the target business data that meets the target data characteristics and belongs to the target data category, which is the sub-data set corresponding to the target data category and is the test data set.

[0363] S1205: The target MEF module uses the sub-data set corresponding to the target data category returned by the target DMF module to evaluate the target evaluation model and sends a model evaluation confirmation message to the MMF module.

[0364] The model evaluation confirmation message includes the evaluation results, such as the accuracy, precision, and prediction error of the target evaluation model.

[0365] Therefore, by including the identifier of the target data category and the identifier of the target data analysis model in the model evaluation message sent from the MMF module to the MEF module, the MEF module can obtain the correct test dataset required for the evaluation task to evaluate the target evaluation model, thus avoiding inaccurate model evaluation.

[0366] Example 7

[0367] In an alternative implementation, the MMF module can also send the mapping table mentioned in the above embodiments to the MEF. Furthermore, the model evaluation message sent by the MMF module to the target MEF module no longer needs to specify the identifier of the target data category and the identifier of the target analysis model required for the current model evaluation. The MEF module can use the mapping table, the identifier of the target evaluation model, and based on the identifier of the target data category and the identifier of the target analysis model, request the corresponding sub-data set (for test data set) from the corresponding DMF module to evaluate the target evaluation model.

[0368] See Figure 13 When an evaluation task is triggered, the steps of a federated learning method may include:

[0369] S1300: Evaluation task triggered.

[0370] S1301: The MMF module sends an evaluation model deployment message to any target MEF module.

[0371] The deployment message for the evaluation model includes the identifier of the target evaluation model, the model file corresponding to the evaluation model, and a mapping table. The target evaluation model can be a federated learning model corresponding to the relevant target data category, or an updated AI model obtained by training the target MTF corresponding to the target MEF module.

[0372] S1302: The MMF module sends a model evaluation message to any target MEF module.

[0373] The model evaluation message includes the identifier of the target evaluation model and evaluation indicator information.

[0374] S1303: The target MEF module uses the mapping table to find the identifier of the target data category and the identifier of the target data analysis model that correspond to the identifier of the target evaluation model contained in the model evaluation message.

[0375] S1304: The target MEF module sends a data query message to the corresponding target DMF module.

[0376] The data query message includes: identifiers of target data features, data type, target data category, and target data analysis model. The data type can include either "train" or "test," used to obtain the training and test datasets, respectively. In S1203, the data type is set to "test" to obtain the test dataset.

[0377] S1305: The target DMF module sends a data query confirmation message to the corresponding target MEF module.

[0378] The data query confirmation message includes the target business data that meets the target data characteristics and belongs to the target data category, which is the sub-data set corresponding to the target data category and is the test data set.

[0379] S1306: The target MEF module uses the sub-data set corresponding to the target data category returned by the target DMF module to evaluate the target evaluation model and sends a model evaluation confirmation message to the MMF module.

[0380] The model evaluation confirmation message includes the evaluation results, such as the accuracy, precision, and prediction error of the target evaluation model.

[0381] Example 8

[0382] In an optional implementation, the MMF module can also send the mapping table mentioned in the above embodiments to the DMF module. Therefore, the model evaluation message sent by the MMF module to the target MEF module no longer needs to specify the identifiers of the target data category and the target analysis model required for the current model evaluation. Instead, the data query message sent by the MEF module to the DMF module can include the identifier of the target evaluation model. The DMF module can then use the mapping table to find the identifier of the target data category and the identifier of the target analysis model corresponding to that identifier, and return the corresponding sub-data set (the test data set) to the relevant MEF module to evaluate the target evaluation model.

[0383] See Figure 14 When an evaluation task is triggered, the steps of a federated learning method may include:

[0384] S1400: Evaluation task triggered.

[0385] S1401: The MMF module sends a data analysis model deployment message to any target DMF module.

[0386] The data analysis model deployment message includes the data analysis model's identifier, model file, and mapping table. It's understandable that if the DMF module has already deployed the data analysis model, then only the mapping table needs to be sent in step S1401. If the DMF module already stores the mapping table, then step S1401 can be ignored.

[0387] S1402: The MMF module sends an evaluation model deployment message to any target MEF module.

[0388] The deployment message for the evaluation model includes the identifier of the target evaluation model and the corresponding model file. This target evaluation model can be a federated learning model corresponding to the specific target data category, or an updated AI model obtained by training the target MTF corresponding to the target MEF module.

[0389] S1403: The MMF module sends a model evaluation message to the target MEF module.

[0390] The model evaluation message includes the identifier of the target evaluation model and evaluation indicator information.

[0391] S1404: The target MEF module sends a data query message to the corresponding target DMF module.

[0392] The data query message includes: the identifier of the target data feature, the data type, and the identifier of the target evaluation model. The data type can include either "train" or "test," used to obtain the training dataset and the test dataset, respectively. In S1203, the data type is set to "test" to obtain the test dataset.

[0393] S1405: The target DMF module uses the mapping table to find the identifier of the target data category and the identifier of the target data analysis model that correspond to the identifier of the target evaluation model. Based on the identifier of the target data category and the identifier of the target data analysis model, it obtains the corresponding sub-data set, which is the test data set.

[0394] S1406: The target DMF module sends a data query confirmation message to the corresponding target MEF module.

[0395] The data query confirmation message includes target business data obtained from the target DMF query that conforms to the target data characteristics and belongs to the target data category, i.e., the sub-data set corresponding to the target data category, which is the test data set.

[0396] S1407: The target MEF module uses the sub-data set corresponding to the target data category returned by the target DMF module to evaluate the target evaluation model and sends a model evaluation confirmation message to the MMF module.

[0397] The model evaluation confirmation message includes the evaluation results, such as the accuracy, precision, and prediction error of the target evaluation model.

[0398] Therefore, through the above embodiments six to eight, the first node acting as the coordinator can, according to the evaluation task, instruct the MEF modules of at least two target second nodes to obtain the test data set corresponding to the target data category, and evaluate the target evaluation model corresponding to the target data category, thereby avoiding inaccurate evaluation of the target evaluation model corresponding to the target data category.

[0399] At this point, the attached document has been completed. Figures 7-14 The specific implementation of the federated learning scheme of this application is described in various embodiments.

[0400] In this scheme, the first node, acting as the coordinator, can select at least two target second nodes containing business data of the target data category to participate in the model training process based on the data distribution information of multiple second nodes, thereby obtaining a federated learning model for the corresponding target data category. Furthermore, during the process of coordinating the training of the federated learning model corresponding to the target data category by at least two target second nodes, the first node can include relevant information in messages sent to the target second nodes to instruct each target second node to select target business data belonging to that target data category to train or evaluate the corresponding AI model, thus obtaining the federated learning model corresponding to the target data category. This scheme identifies the data category to which business data conforming to the target data characteristics belongs; each data category corresponds to a data distribution; and for each data category, the corresponding business data is used to complete model training or evaluation. Therefore, even when participating parties have multiple data distributions, corresponding federated learning models can be obtained for different data distributions, minimizing the poisoning effect of different data distributions on the federated learning model and ensuring the accuracy of the obtained federated learning model.

[0401] Based on the same technical concept, this application also provides a federated learning device, see [link to relevant documentation]. Figure 15 The federated learning device 1500 may include a communication unit 1510 and a processing unit 1520. The communication unit 1510 and the processing unit 1520 may be used to implement the methods described in the above embodiments or the embodiments provided.

[0402] When the federated learning device is implemented as the first node mentioned above, the federated learning device can achieve:

[0403] The communication unit 1510 is used to acquire data distribution information of multiple second nodes according to the target data features required for the training task. The data distribution information of any second node is used to indicate the data category to which the business data that conforms to the target data features stored locally by the second node belongs. The processing unit 1520 is used to select at least two target second nodes from the multiple second nodes according to the target data category required for the training task and the data distribution information of the multiple second nodes, and instruct the at least two target second nodes to perform federated learning, so as to obtain that any target second node in the federated learning model corresponding to the target data category in the training task has target business data that conforms to the target data features and belongs to the target data category stored locally.

[0404] In one example, each second node deploys at least one data analysis model, each data analysis model corresponding to a data feature group and used to identify the data category of business data that conforms to the corresponding data feature group; the communication unit 1510 is used to: send a first query message to each of the plurality of second nodes according to the target data feature; wherein the first query message sent to any second node includes the identifier of the target data feature and the identifier of the target data analysis model, the target data analysis model corresponding to the target data feature; and receive corresponding data distribution information from the plurality of second nodes, wherein the data distribution information of any second node is used to indicate the identifier of at least one data category, and the data information of business data belonging to the at least one data category stored by the second node.

[0405] In one example, the first query message sent by the first node to any second node also includes the identifier of the target data category, wherein the data distribution information returned by the second node includes: the identifier of the target data category, and the data information of the target business data belonging to the target data category stored by the second node.

[0406] In one example, before the first node obtains the data distribution information of multiple second nodes based on the target data features required for the training task, the communication unit 1510 is further configured to: send a data analysis model deployment message to each of the multiple second nodes, wherein the data analysis model deployment message sent to any second node includes: the identifier of the at least one data analysis model, and the model file of the at least one data analysis model.

[0407] In one example, the processing unit 1520 is configured to: send a model training message to each of the at least two target second nodes, wherein the model training message sent to any target second node includes an identifier of a target artificial intelligence (AI) model, the target AI model corresponding to the target data category; and obtain a federated learning model corresponding to the target data category in the training task based on the updated AI model received from the at least two target second nodes.

[0408] In one example, the model training message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

[0409] In one example, the communication unit 1510 is further configured to: send a model evaluation message to each of the at least two target second nodes, wherein the model evaluation message sent to any target second node includes an identifier of the target evaluation model and an evaluation index, the target evaluation model corresponding to the target data category; and receive corresponding model evaluation results from the at least two target second nodes.

[0410] In one example, the model evaluation message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

[0411] In one example, the federated learning system is a wireless AI model-driven network system; the first node includes a model management function (MMF) module; any second node includes a model training function (MTF) module, a data management function (DMF) module, and a model evaluation function (MEF) module; wherein, at least one data analysis model is deployed in the DMF module or the MTF module; the communication unit 1510 sends a first query message to each second node, including: the communication unit of the MMF module sends the first query message to the DMF module or the MTF module of each second node.

[0412] In one example, the communication unit 1510 is further configured to: send a mapping relationship table to each of the plurality of second nodes respectively, wherein the mapping relationship table sent to any one of the second nodes is used to record the mapping relationship between the identifier of data feature, the identifier of AI model, the identifier of data analysis model, and the identifier of data category.

[0413] When the federated learning device is implemented as the first node mentioned above, the federated learning device can achieve:

[0414] Communication unit 1510 is configured to receive a first query message from the first node, the first query message indicating the target data features required for the training task; and send data distribution information to the first node according to the target data features, the data distribution information indicating the data category to which the business data that conforms to the target data features stored locally by the second node belongs; processing unit 1520 is configured to train a target artificial intelligence (AI) model corresponding to the target data category using the stored target business data belonging to the target data category, according to the instructions of the first node, to obtain an updated AI model; the communication unit is further configured to send the updated AI model to the first node so that the first node obtains the federated learning model corresponding to the target data category in the training task.

[0415] In one example, the second node is deployed with at least one data analysis model, each data analysis model corresponding to a data feature group and used to identify the data category of business data that conforms to the corresponding data feature group; the first query message includes the identifier of the target data feature and the identifier of the target data analysis model, the target data analysis model corresponding to the target data feature; the processing unit is used to use the target data analysis model to identify the data category of the stored business data that conforms to the target data feature, and obtain data information of business data that belongs to at least one data category; the communication unit is also used to send the data distribution information to the first node, the data distribution information being used to indicate the identifier of at least one data category, and the data information of business data that belongs to at least one data category.

[0416] In one example, the first query message also includes an identifier of the target data category, wherein the data distribution information includes: the identifier of the target data category, and data information of the target business data belonging to the target data category stored by the second node.

[0417] In one example, the communication unit 1510 is further configured to: receive a data analysis model deployment message from the first node before the first node receives the first query message; wherein the data analysis model deployment message includes: an identifier of the at least one data analysis model, and a model file of the at least one data analysis model.

[0418] In one example, the communication unit 1510 is used to receive a model training message from the first node, wherein the model training message includes an identifier of the target AI model, and the target AI model corresponds to the target data category; the processing unit is used to obtain, based on the identifier of the AI ​​model, stored target business data that conforms to the characteristics of the target data and belongs to the target data category, and train the AI ​​model based on the target business data to obtain an updated AI model.

[0419] In one example, the model training message also includes: an identifier for the target data category and an identifier for the target data analysis model.

[0420] In one example, the communication unit 1510 is further configured to: receive a model evaluation message from the first node, and evaluate the target evaluation model using the target service data, wherein the target evaluation model message includes an identifier and evaluation index of the target evaluation model, and the target evaluation model corresponds to the target data category; and send the model evaluation result to the first node.

[0421] In one example, the model evaluation message also includes: an identifier of the target data category and an identifier of the target data analysis model.

[0422] In one example, the federated learning system is a wireless AI model-driven network system; the first node includes a model management function (MMF) module; any second node includes a model training function (MTF) module, a data management function (DMF) module, and a model evaluation function (MEF) module; wherein, at least one data analysis model is deployed in the DMF module or the MTF module; the communication unit receives a first query message from the first node, including: the communication unit of the DMF module or the communication unit of the MTF module receives the first query message from the MMF module.

[0423] In one example, where the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the MTF module, after the communication unit of the DMF module receives the first query message from the MTF module, the communication unit of the DMF module is further configured to: send a data analysis message to the MTF module, the data analysis message including the full dataset of data that conforms to the target data characteristics stored by the DMF module, the identifier of the target data category, and the identifier of the data analysis model, and the data analysis message is used to instruct the MTF module to use the target data analysis model to identify the data category of the full dataset.

[0424] In one example, where the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the DMF module, the second node receives a model training message from the first node, including: the communication unit of the MTF module receiving the model training message from the MMF module; after the communication unit of the MTF module receives the model training message from the MMF module, the communication unit of the MTF module is further configured to: send a second query message to the DMF module, the second query message being used to instruct the DMF module to provide feedback to the MTF module on the training data set in the target business data; wherein, the second query message includes: the identifier of the target data feature, the identifier of the target AI model, and first data type indication information; or, the identifier of the target data feature, the identifier of the target data category, the identifier of the target data analysis model, and first data type indication information.

[0425] In one example, when the MEF module and the MTF module are located in different entities, the communication unit is configured to: receive the model evaluation message from the MMF module through the communication unit of the MEF module; after receiving the model evaluation message from the MMF module through the communication unit of the MEF module, the communication unit of the MEF module is further configured to: send a third query message to the DMF module, the third query message being used to instruct the DMF module to provide feedback to the MEF module on the test data set in the target business data; wherein, the third query message includes: the identifier of the target data feature, the identifier of the target AI model, and second data type indication information; or, the identifier of the target data feature, the identifier of the target data category, the identifier of the target data analysis model, and second data type indication information.

[0426] Based on the same technical concept, this application also provides a federated learning device, which can apply the first node or the second node described above and can implement the methods provided in the above embodiments. (See also...) Figure 16 As shown, the federated learning device 1600 includes a memory 1601, a processor 1602, and a transceiver 1603, wherein the memory 1601, the processor 1602, and the transceiver 1603 are interconnected.

[0427] Optionally, the memory 1601, the processor 1602, and the transceiver 1603 are interconnected via a bus 1604. The memory 1601 stores program code, and the processor 1602 can retrieve the program code from the memory 1601 and execute corresponding processing. The bus 1604 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 16 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0428] It is understood that the memory 1601 is used to store program instructions and data, etc. Specifically, the program instructions may include program code, which includes computer operation instructions. The memory 1601 may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. The processor 1602 executes the program instructions stored in the memory 1601 and uses the data stored in the memory 1601 to implement the above functions, thereby implementing the federated learning method provided in the above embodiments.

[0429] It is understood that this application Figure 16The memory 1601 can be volatile memory or non-volatile memory, or may include both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DR RAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0430] Based on the above embodiments, this application also provides a computer program that, when run on a computer, causes the computer to execute the methods provided in the above embodiments.

[0431] Based on the above embodiments, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a computer, causes the computer to perform the methods provided in the above embodiments.

[0432] The storage medium can be any available medium that a computer can access. For example, but not limited to, a computer-readable medium can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer.

[0433] Based on the above embodiments, this application also provides a chip for reading a computer program stored in a memory to implement the method provided in the above embodiments.

[0434] Based on the above embodiments, this application provides a chip system including a processor for supporting a computer device in implementing the functions involved in the service equipment, forwarding equipment, or site equipment in the above embodiments. In one possible design, the chip system further includes a memory for storing necessary programs and data of the computer device. This chip system may be composed of chips or may include chips and other discrete components.

[0435] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0436] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0437] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0438] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0439] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope of protection of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A federated learning method, characterized in that, The method, applied to a federated learning system comprising a first node and multiple second nodes, includes: The first node obtains data distribution information of multiple second nodes based on the target data features required for the training task. The data distribution information of any second node is used to indicate the data category to which the business data that conforms to the target data features stored locally by the second node belongs. The first node selects at least two target second nodes from the plurality of second nodes based on the target data category required by the training task and the data distribution information of the plurality of second nodes. Each target second node locally stores target business data that conforms to the target data characteristics and belongs to the target data category. The first node instructs the at least two target second nodes to perform federated learning in order to obtain a federated learning model corresponding to the target data category in the training task. The federated learning system is a wireless AI model-driven network system; the first node includes the model management function (MMF) module of the wireless AI model-driven network system; any second node includes the model training function (MTF) module, the data management function (DMF) module, and the model evaluation function (MEF) module of the wireless AI model-driven network system. In the federated learning scenario under the enabled network automation (eNA) architecture, the central network data analysis function (NWDAF) is used to implement the function of the MMF module, the data collection coordination function (DCCF) is used to implement the function of the DMF module, and the local NWDAF is used to implement the functions of the MTF and MEF modules. In the federated learning scenario of UE-RAN, the RAN implements the function of the MMF module, and the UE implements the functions of the MTF module, the MEF module and the DMF module. In the federated learning scenario under the RAN-Network Element Management EMS / Network Management NMS scenario, the EMS / NMS implements the function of the MMF module, and the RAN implements the functions of the MTF module, the MEF module and the DMF module.

2. The method according to claim 1, characterized in that, Each second node is equipped with at least one data analysis model, and each data analysis model corresponds to a data feature group and is used to identify the data category of business data that conforms to the corresponding data feature group. The first node obtains data distribution information of multiple second nodes based on the target data features required for the training task, including: The first node sends a first query message to each of the plurality of second nodes according to the target data features; wherein the first query message sent to any second node includes an identifier of the target data features and an identifier of the target data analysis model, and the target data analysis model corresponds to the target data features; The first node receives corresponding data distribution information from the plurality of second nodes respectively. The data distribution information of any second node is used to indicate the identifier of at least one data category, as well as the data information of business data belonging to the at least one data category stored by the second node.

3. The method according to claim 2, characterized in that, The first query message sent by the first node to any second node also includes the identifier of the target data category, wherein the data distribution information fed back by the second node includes: the identifier of the target data category, and the data information of the target business data belonging to the target data category stored by the second node.

4. The method according to claim 2 or 3, characterized in that, Before the first node obtains the data distribution information of multiple second nodes based on the target data features required for the training task, the method further includes: The first node sends a data analysis model deployment message to each of the plurality of second nodes, wherein the data analysis model deployment message sent to any second node includes: the identifier of the at least one data analysis model, and the model file of the at least one data analysis model.

5. The method according to claim 2 or 3, characterized in that, The first node instructs the at least two target second nodes to perform federated learning to obtain a federated learning model corresponding to the target data category in the training task, including: The first node sends a model training message to each of the at least two target second nodes, wherein the model training message sent to any target second node includes an identifier of the target artificial intelligence (AI) model, and the target AI model corresponds to the target data category; The first node obtains the federated learning model corresponding to the target data category in the training task based on the updated AI model received from the at least two target second nodes.

6. The method according to claim 5, characterized in that, The model training message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

7. The method according to claim 5, characterized in that, The first node instructs the at least two target second nodes to perform federated learning to obtain a federated learning model corresponding to the target data category in the training task, and further includes: The first node sends a model evaluation message to each of the at least two target second nodes, wherein the model evaluation message sent to any target second node includes the identifier of the target evaluation model and the evaluation index, and the target evaluation model corresponds to the target data category; The first node receives the corresponding model evaluation results from the at least two target second nodes.

8. The method according to claim 7, characterized in that, The model evaluation message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

9. The method according to claim 2 or 3, characterized in that, The at least one data analysis model is deployed in the DMF module or the MTF module; The first node sends a first query message to each second node, including: The MMF module sends the first query message to the DMF module or the MTF module of each second node.

10. The method according to claim 2 or 3, characterized in that, The method further includes: The first node sends a mapping table to each of the plurality of second nodes. The mapping table sent to any second node is used to record the mapping relationship between the identifier of data features, the identifier of AI models, the identifier of data analysis models, and the identifier of data categories.

11. A federated learning method, characterized in that, The method, applied to a federated learning system comprising a first node and multiple second nodes, includes: The second node receives a first query message from the first node, the first query message being used to indicate the target data features required for the training task; The second node sends data distribution information to the first node based on the target data characteristics. The data distribution information is used to indicate the data category to which the business data that conforms to the target data characteristics and is stored locally by the second node belongs. The first node is used to select at least two target second nodes from the plurality of second nodes based on the target data category required by the training task and the data distribution information of the plurality of second nodes. According to the instructions of the first node, the second target node uses the stored target business data belonging to the target data category to train the target artificial intelligence (AI) model corresponding to the target data category, and obtains the updated AI model. The target second node sends the updated AI model to the first node, so that the first node obtains the federated learning model corresponding to the target data category in the training task; The federated learning system is a wireless AI model-driven network system; the first node includes the model management function (MMF) module of the wireless AI model-driven network system; any second node includes the model training function (MTF) module, the data management function (DMF) module, and the model evaluation function (MEF) module of the wireless AI model-driven network system. In the federated learning scenario under the enabled network automation (eNA) architecture, the central network data analysis function (NWDAF) is used to implement the function of the MMF module, the data collection coordination function (DCCF) is used to implement the function of the DMF module, and the local NWDAF is used to implement the functions of the MTF and MEF modules. In the federated learning scenario of UE-RAN, the RAN implements the function of the MMF module, and the UE implements the functions of the MTF module, the MEF module and the DMF module. In the federated learning scenario under the RAN-Network Element Management EMS / Network Management NMS scenario, the EMS / NMS implements the function of the MMF module, and the RAN implements the functions of the MTF module, the MEF module and the DMF module.

12. The method according to claim 11, characterized in that, The second node is equipped with at least one data analysis model, each data analysis model is associated with a data feature group and is used to identify the data category of business data that conforms to the corresponding data feature group; the first query message includes the identifier of the target data feature and the identifier of the target data analysis model, the target data analysis model being associated with the target data feature; The second node sends data distribution information to the first node based on the target data characteristics, including: The second node uses the target data analysis model to identify the data categories of stored business data that conform to the target data characteristics, and obtains data information of business data that belong to at least one data category. The second node sends the data distribution information to the first node. The data distribution information is used to indicate the identifier of at least one data category, and the data information of business data belonging to the at least one data category respectively.

13. The method according to claim 12, characterized in that, The first query message also includes the identifier of the target data category, wherein the data distribution information includes: the identifier of the target data category, and the data information of the target business data belonging to the target data category stored by the second node.

14. The method according to claim 12 or 13, characterized in that, Before the second node receives the first query message from the first node, the method further includes: The second node receives a data analysis model deployment message from the first node; wherein the data analysis model deployment message includes: the identifier of the at least one data analysis model, and the model file of the at least one data analysis model.

15. The method according to claim 12 or 13, characterized in that, The second node, according to the instructions of the first node, uses stored target business data belonging to the target data category to train the target artificial intelligence (AI) model corresponding to the target data category, thereby obtaining an updated AI model, including: The second node receives a model training message from the first node, wherein the model training message includes an identifier of the target AI model, and the target AI model corresponds to the target data category; The second node obtains target business data that conforms to the target data characteristics and belongs to the target data category, based on the identifier of the target AI model; The second node trains the target AI model based on the target business data to obtain an updated AI model.

16. The method according to claim 15, characterized in that, The model training message also includes: the identifier of the target data category and the identifier of the target data analysis model.

17. The method according to claim 15, characterized in that, The method further includes: The second node receives a model evaluation message from the first node and evaluates the target evaluation model using the target business data. The target evaluation model message includes the identifier and evaluation index of the target evaluation model, and the target evaluation model corresponds to the target data category. The second node sends the model evaluation results to the first node.

18. The method according to claim 17, characterized in that, The model evaluation message also includes: the identifier of the target data category and the identifier of the target data analysis model.

19. The method according to claim 12 or 13, characterized in that, The at least one data analysis model is deployed in the DMF module or the MTF module; The second node receives a first query message from the first node, including: The DMF module or the MTF module receives the first query message from the MMF module.

20. The method according to claim 19, characterized in that, When the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the MTF module, after the DMF module receives the first query message from the MTF module, the method further includes: The DMF module sends a data analysis message to the MTF module. The data analysis message includes the full dataset that conforms to the target data characteristics stored by the DMF module, the identifier of the target data category, and the identifier of the data analysis model. The data analysis message is used to instruct the MTF module to use the target data analysis model to identify the data category of the full dataset.

21. The method according to claim 19, characterized in that, When the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the DMF module, the second node receives model training messages from the first node, including: The MTF module receives the model training message from the MMF module; After the MTF module receives the model training message from the MMF module, the method further includes: The MTF module sends a second query message to the DMF module, the second query message being used to instruct the DMF module to return the training data set in the target service data to the MTF module; The second query message includes: The identifiers of the target data features, the identifiers of the target AI model, and the first data type indication information; or The identifiers of the target data features, the target data category, the target data analysis model, and the first data type indication information.

22. The method according to claim 19, characterized in that, When the MEF module and the MTF module are located in different entities, after the MEF module receives the model evaluation message from the MTF module, the method further includes: The MEF module sends a third query message to the DMF module, the third query message being used to instruct the DMF module to return the test data set in the target service data to the MEF module; The third query message includes: The identifiers of the target data features, the identifiers of the target AI model, and the second data type indication information; or The identifiers of the target data features, the target data category, the target data analysis model, and the second data type indication information.

23. A federated learning device, characterized in that, For use in a federated learning system comprising a first node and multiple second nodes, the apparatus includes: The communication unit is used to obtain data distribution information of multiple second nodes according to the target data features required for the training task. The data distribution information of any second node is used to indicate the data category to which the business data that conforms to the target data features stored locally on the second node belongs. The processing unit is configured to select at least two target second nodes from the plurality of second nodes according to the target data category required by the training task and the data distribution information of the plurality of second nodes, and instruct the at least two target second nodes to perform federated learning, so as to obtain that any target second node of the federated learning model corresponding to the target data category in the training task has target business data that conforms to the target data characteristics and belongs to the target data category locally stored. The federated learning system is a wireless AI model-driven network system; the first node includes the model management function (MMF) module of the wireless AI model-driven network system; any second node includes the model training function (MTF) module, the data management function (DMF) module, and the model evaluation function (MEF) module of the wireless AI model-driven network system. In the federated learning scenario under the enabled network automation (eNA) architecture, the central network data analysis function (NWDAF) is used to implement the function of the MMF module, the data collection coordination function (DCCF) is used to implement the function of the DMF module, and the local NWDAF is used to implement the functions of the MTF and MEF modules. In the federated learning scenario of UE-RAN, the RAN implements the function of the MMF module, and the UE implements the functions of the MTF module, the MEF module and the DMF module. In the federated learning scenario under the RAN-Network Element Management EMS / Network Management NMS scenario, the EMS / NMS implements the function of the MMF module, and the RAN implements the functions of the MTF module, the MEF module and the DMF module.

24. The apparatus according to claim 23, characterized in that, Each second node deploys at least one data analysis model, each data analysis model corresponding to a data feature group and used to identify the data category of business data that conforms to the corresponding data feature group; the communication unit is used for: Based on the target data features, a first query message is sent to each of the plurality of second nodes; wherein the first query message sent to any second node includes an identifier of the target data features and an identifier of the target data analysis model, and the target data analysis model corresponds to the target data features; The system receives corresponding data distribution information from the plurality of second nodes, wherein the data distribution information of any one second node is used to indicate the identifier of at least one data category, and the data information of the business data belonging to the at least one data category stored by the second node.

25. The apparatus according to claim 24, characterized in that, The first query message sent by the first node to any second node also includes the identifier of the target data category, wherein the data distribution information fed back by the second node includes: the identifier of the target data category, and the data information of the target business data belonging to the target data category stored by the second node.

26. The apparatus according to claim 24 or 25, characterized in that, Before the first node obtains the data distribution information of multiple second nodes based on the target data features required for the training task, the communication unit is further configured to: A data analysis model deployment message is sent to each of the plurality of second nodes, wherein the data analysis model deployment message sent to any one of the second nodes includes: the identifier of the at least one data analysis model, and the model file of the at least one data analysis model.

27. The apparatus according to claim 24 or 25, characterized in that, The processing unit is used for: A model training message is sent to each of the at least two target second nodes, wherein the model training message sent to any target second node includes an identifier of the target artificial intelligence (AI) model, and the target AI model corresponds to the target data category; Based on the updated AI models received from the at least two target second nodes, a federated learning model corresponding to the target data category in the training task is obtained.

28. The apparatus according to claim 27, characterized in that, The model training message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

29. The apparatus according to claim 27, characterized in that, The communication unit is also used for: A model evaluation message is sent to each of the at least two target second nodes, wherein the model evaluation message sent to any target second node includes the identifier of the target evaluation model and the evaluation index, and the target evaluation model corresponds to the target data category; The corresponding model evaluation results are received from the second nodes of the at least two targets respectively.

30. The apparatus according to claim 29, characterized in that, The model evaluation message sent to any target second node also includes: the identifier of the target data category and the identifier of the target data analysis model.

31. The apparatus according to claim 24 or 25, characterized in that, The federated learning system is a wireless AI model-driven network system; the first node includes a model management function (MMF) module; any second node includes a model training function (MTF) module, a data management function (DMF) module, and a model evaluation function (MEF) module. The at least one data analysis model is deployed in the DMF module or the MTF module; The communication unit sends a first query message to each second node, including: The communication unit of the MMF module sends the first query message to the DMF module or the MTF module of each second node.

32. The apparatus according to claim 24 or 25, characterized in that, The communication unit is also used for: A mapping table is sent to each of the plurality of second nodes respectively. The mapping table sent to any one of the second nodes is used to record the mapping relationship between the identifier of data feature, the identifier of AI model, the identifier of data analysis model, and the identifier of data category.

33. A federated learning device, characterized in that, For use in a federated learning system comprising a first node and multiple second nodes, the apparatus includes: A communication unit is configured to receive a first query message from a first node, the first query message indicating the target data features required for a training task; and send data distribution information to the first node according to the target data features, the data distribution information indicating the data category to which business data that conforms to the target data features and is locally stored on a second node belongs; and the first node is configured to select at least two target second nodes from the plurality of second nodes according to the target data category required for the training task and the data distribution information of the plurality of second nodes. The processing unit is configured to, according to the instructions of the first node, use stored target business data belonging to the target data category to train the target artificial intelligence (AI) model corresponding to the target data category, and obtain an updated AI model. The communication unit is also used to send the updated AI model to the first node, so that the first node obtains the federated learning model corresponding to the target data category in the training task; The federated learning system is a wireless AI model-driven network system; the first node includes the model management function (MMF) module of the wireless AI model-driven network system; any second node includes the model training function (MTF) module, the data management function (DMF) module, and the model evaluation function (MEF) module of the wireless AI model-driven network system. In the federated learning scenario under the enabled network automation (eNA) architecture, the central network data analysis function (NWDAF) is used to implement the function of the MMF module, the data collection coordination function (DCCF) is used to implement the function of the DMF module, and the local NWDAF is used to implement the functions of the MTF and MEF modules. In the federated learning scenario of UE-RAN, the RAN implements the function of the MMF module, and the UE implements the functions of the MTF module, the MEF module and the DMF module. In the federated learning scenario under the RAN-Network Element Management EMS / Network Management NMS scenario, the EMS / NMS implements the function of the MMF module, and the RAN implements the functions of the MTF module, the MEF module and the DMF module.

34. The apparatus according to claim 33, characterized in that, The second node is equipped with at least one data analysis model, each data analysis model is associated with a data feature group and is used to identify the data category of business data that conforms to the corresponding data feature group; the first query message includes the identifier of the target data feature and the identifier of the target data analysis model, the target data analysis model being associated with the target data feature; The processing unit is used to use the target data analysis model to identify the data categories of stored business data that conform to the target data characteristics, and to obtain data information of business data that belong to at least one data category. The communication unit is further configured to send the data distribution information to the first node, the data distribution information being used to indicate the identifier of at least one data category, and the data information of business data belonging to the at least one data category respectively.

35. The apparatus according to claim 34, characterized in that, The first query message also includes the identifier of the target data category, wherein the data distribution information includes: the identifier of the target data category, and the data information of the target business data belonging to the target data category stored by the second node.

36. The apparatus according to claim 34 or 35, characterized in that, The communication unit is also used for: Before the first node receives the first query message, a data analysis model deployment message is received from the first node; wherein the data analysis model deployment message includes: the identifier of the at least one data analysis model, and the model file of the at least one data analysis model.

37. The apparatus according to claim 34 or 35, characterized in that, The communication unit is used to receive model training messages from the first node, wherein the model training messages include the identifier of the target AI model, and the target AI model corresponds to the target data category; The processing unit is used to obtain target business data that conforms to the target data characteristics and belongs to the target data category, based on the identifier of the target AI model, and to train the target AI model based on the target business data to obtain an updated AI model.

38. The apparatus according to claim 37, characterized in that, The model training message also includes: the identifier of the target data category and the identifier of the target data analysis model.

39. The apparatus according to claim 37, characterized in that, The communication unit is also used for: The system receives a model evaluation message from the first node and evaluates the target evaluation model using the target business data. The target evaluation model message includes the identifier and evaluation index of the target evaluation model, and the target evaluation model corresponds to the target data category. Send the model evaluation results to the first node.

40. The apparatus according to claim 39, characterized in that, The model evaluation message also includes: the identifier of the target data category and the identifier of the target data analysis model.

41. The apparatus according to claim 34 or 35, characterized in that, The federated learning system is a wireless AI model-driven network system; the first node includes a model management function (MMF) module; any second node includes a model training function (MTF) module, a data management function (DMF) module, and a model evaluation function (MEF) module. The at least one data analysis model is deployed in the DMF module or the MTF module; The communication unit receives a first query message from the first node, including: The communication unit of the DMF module or the communication unit of the MTF module receives the first query message from the MMF module.

42. The apparatus according to claim 41, characterized in that, When the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the MTF module, after the communication unit of the DMF module receives the first query message from the MTF module, the communication unit of the DMF module is further configured to: send a data analysis message to the MTF module, the data analysis message including the full set of data that conforms to the target data characteristics stored by the DMF module, the identifier of the target data category, and the identifier of the data analysis model, and the data analysis message is used to instruct the MTF module to use the target data analysis model to identify the data category of the full set of data.

43. The apparatus according to claim 41, characterized in that, When the DMF module and the MTF module are located in different entities, and the at least one data analysis model is deployed in the DMF module, the second node receives model training messages from the first node, including: The communication unit of the MTF module receives the model training message from the MMF module; After the communication unit of the MTF module receives the model training message from the MMF module, the communication unit of the MTF module is further configured to: send a second query message to the DMF module, wherein the second query message is configured to instruct the DMF module to return the training data set in the target service data to the MTF module; The second query message includes: The identifiers of the target data features, the identifiers of the target AI model, and the first data type indication information; or The identifiers of the target data features, the target data category, the target data analysis model, and the first data type indication information.

44. The apparatus according to claim 41, characterized in that, When the MEF module and the MTF module are located in different entities, the communication unit is used for: The model evaluation message is received from the MMF module through the communication unit of the MEF module; After receiving the model evaluation message from the MMF module through the communication unit of the MEF module, the communication unit of the MEF module is further configured to: send a third query message to the DMF module, wherein the third query message is configured to instruct the DMF module to provide the MEF module with the test data set in the target service data; The third query message includes: The identifiers of the target data features, the identifiers of the target AI model, and the second data type indication information; or The identifiers of the target data features, the target data category, the target data analysis model, and the second data type indication information.

45. A federated learning system, characterized in that, include: The federated learning device as described in any one of claims 23-32; as well as The federated learning device as described in any one of claims 33-44.

46. ​​A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the method of any one of claims 1-10, or the method of any one of claims 11-22.

47. A computer program product, characterized in that, When the computer program product is run on a computer, it causes the computer to perform the method of any one of claims 1-10, or the method of any one of claims 11-22.

Images