Federated learning method and system based on anchor aggregation

Abstract

The present invention discloses a method and system for federated learning based on anchor aggregation, including multiple clients and a central server, each client having a communication channel with the central server; the client is used to guide the local data set to train the current local model according to the global anchor point of the current round, and after updating the local anchor point, upload the new local anchor point to the central server, wherein the local anchor point is a feature point that can represent the classification category determined according to the feature vector corresponding to the local data; the central server is used to aggregate and update the new local anchor points uploaded by all clients to obtain the global anchor points corresponding to each classification category, and send the global anchor points to each client to guide the next round of local training. The method and system effectively reduce the communication level of federated training on the basis of ensuring the accuracy of model training, and support federated learning of heterogeneous models.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and information security, and in particular to a federated learning method and system based on anchor point aggregation. Background Art

[0002] With the development of artificial intelligence, massive user data has become an important basis for model training. However, in reality, data between different departments and industries is difficult to communicate with each other due to economic interests, user privacy, and permission restrictions. How to connect data fragments between different data units to fully utilize data and improve the effect of artificial intelligence training has become one of the hot topics in current artificial intelligence research.

[0003] In this context, federated learning was born. Federated learning can be seen as an extended form of distributed learning. In federated learning, different data owners can join the data federation and jointly train a task through a centralized parameter server. During the entire training process, the data of different data owners are retained locally, and they mainly communicate data information indirectly through parameter information such as models, thereby achieving a training effect similar to that of using all client data for joint training, such as a federated learning method disclosed in patent document CN113516151A, and another example is a federated prediction method based on federated learning disclosed in patent document CN112364943A.

[0004] The most widely used federated learning algorithm at present is the federated average algorithm. However, most algorithms represented by the federated average algorithm transmit local training information through the client's local training model. Since deep learning training models often have complex model structures, federated learning algorithms that use model parameters or training gradients as information media require higher communication costs. At the same time, since the central server mainly performs federated averaging on the models or gradients uploaded by different clients, the model structure used by each client participating in the federated training must be the same, which greatly restricts the hardware conditions and training purposes of the participating devices. Summary of the invention

[0005] In view of the above, in order to solve the problems of model heterogeneity and high communication costs, the purpose of the present invention is to provide a federated learning method and system based on anchor point aggregation, which effectively reduces the communication level of federated training while ensuring the accuracy of model training, and supports federated learning of heterogeneous models.

[0006] To achieve the above-mentioned purpose of the invention, the embodiment provides a federated learning system based on anchor aggregation, including multiple clients and a central server, each client having a communication channel with the central server;

[0007] The client is used to guide the local data set to train the current local model according to the global anchor point of the current round, and after updating the local anchor point, upload the new local anchor point to the central server, where the local anchor point is a feature point that can represent the classification category determined according to the feature vector corresponding to the local data;

[0008] The central server is used to aggregate and update the new local anchor points uploaded by all clients to obtain the global anchor points corresponding to each classification category, and send the global anchor points to each client to guide the next round of local training.

[0009] In one embodiment, when the client guides the local data set to train the local model based on the current local anchor point, the loss function loss used is:

[0010] loss = αloss sup +(1-α)loss an

[0011] Among them, α is the weight, loss sup Constructing a task loss for the classification prediction task based on the classification prediction results output by the local model;

[0012] loss an The guidance loss for guiding local model training based on the global anchor point is used to measure the relationship between the feature vector extracted from the input sample data and the global anchor point guiding local training in the feature projection space, which is expressed as:

[0013]

[0014] in, It represents the real feature vector of the i-th input data belonging to the j-class data after being projected by the local model in the feature projection space. It represents the vector representation of the j-th global anchor point in the feature projection space. relation(·) represents the relationship function, which is used to measure the distance or similarity. T is a hyperparameter with a value of 0-1. n is the category index and N is the total number of categories.

[0015] In one embodiment, when training a local model, the client uses a momentum optimization algorithm to update the local anchor point, including:

[0016] During the training of the local model, the momentum optimization algorithm is used to update the current local anchor point based on the real feature vector of the sample data in the feature projection space after being projected by the local model. After iterating through multiple batches of data, the feature points in the real feature vector that can represent the classification category are obtained as new local anchor points.

[0017] In one embodiment, when training a local model, the client uses a central loss function to update the local anchor point, including:

[0018] The local anchor point is regarded as a training parameter. During the training process, the difference between the real feature vector of the sample data in the feature projection space after being projected by the local model and the local anchor point is made to construct a central loss function. The central loss function is used to update the position of the local anchor point to reduce the average distance between the real feature vector and the local anchor point. After multiple batches of data iterations, the new local anchor point can balance the distance between the corresponding feature vectors in each category to approach the center of each feature vector and achieve stability, which is sufficient to represent the data characteristics of each classification category.

[0019] In one embodiment, the central server aggregates and updates the new local anchor points uploaded by all clients, including: weighted summing and averaging all the new local anchor points representing each classification category to obtain a global anchor point for each classification category.

[0020] In one embodiment, before federated learning, the central server is further used to construct a federated learning task and send it to each client, wherein the federated learning task includes an initial local model, an initial global anchor point, and federated learning hyperparameters, wherein the federated learning hyperparameters include a gradient descent algorithm optimizer, a local training learning rate, a total number of clients participating in federated learning, a number of rounds of local training of the client per round, a total number of communication rounds, a spatial dimension of a local anchor point, a weight decay rate, and a momentum value or a learning rate required during the update of the local anchor point;

[0021] The client and the central server perform federated learning based on the federated learning task.

[0022] In one embodiment, the local model includes a feature extraction network, a projection network and an output layer, wherein the feature extraction network is used to extract the feature map of the input sample data, the projection network is used to project the feature map into the feature projection space to obtain the feature vector, and the output layer is used to perform prediction calculations based on the feature vector and classify the prediction results.

[0023] In one embodiment, the client is also used to upload local model parameters to the central server; the central server is also used to aggregate the local model parameters uploaded by all clients to obtain a global model, and send the global model to each client as the client's current local model.

[0024] After federated learning is completed, each client uses its current local model as the classification model for classification prediction tasks.

[0025] To achieve the above-mentioned purpose of the invention, the embodiment further provides a federated learning method based on anchor point aggregation, wherein the federated learning method adopts the above-mentioned federated learning system, and the federated learning method includes:

[0026] The central server establishes a communication channel with the client, creates a federated learning task, and sends it to each client;

[0027] The client trains the current local model based on the local data set guided by the global anchor point of the current round, and after updating the local anchor point, uploads the new local anchor point to the central server, where the local anchor point is a feature point that can represent the classification category determined according to the feature vector corresponding to the local data;

[0028] The central server aggregates and updates the new local anchor points uploaded by all clients to obtain the global anchor points corresponding to each classification category, and sends the global anchor points to each client to guide the next round of local training.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] Each client trains the current local model based on the local data set guided by the global anchor point of the current round, updates the local anchor point and uploads it to the central server. Since the local anchor point is a feature point that can represent the classification category extracted from the feature map corresponding to the local data, the global information obtained by all local anchor points can integrate the category information of each client and has global representativeness. The guidance and constraint training of the global anchor point can improve the efficiency of federated learning. At the same time, the central server aggregates the local anchor points uploaded by the client, and then sends the aggregated anchor point to each client for the next round of federated learning. In this way, the aggregated anchor point realizes the guidance and constraint of the global data distribution.

[0031] In the process of federated learning, local anchor points that condense local data knowledge are uploaded to the central server instead of local models for global information exchange. Since the number of local anchor point parameters is very small compared to the local model parameters, this effectively reduces the communication level of federated training while ensuring the accuracy of model training, thereby improving communication efficiency.

[0032] When federated learning is performed only through anchor point aggregation without aggregating model parameters, the local models trained by each client can be heterogeneous and suitable for their respective application needs, breaking the unified restrictions on model structure and improving the flexibility of federated training.

[0033] When performing federated learning of anchor point aggregation and model parameter aggregation at the same time, the client uploads the local anchor point and local model parameters at the same time. The anchor point aggregation and parameter aggregation can be organically combined. Since the anchor point communication level differs from the parameter communication level by at least 1000 orders of magnitude, the anchor point communication transmission can be ignored relative to the parameter communication transmission. Training with the anchor point as a constraint guide can improve the federated performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 is a structural diagram of a federated learning system based on anchor point aggregation provided in an embodiment;

[0036] Figure 2 is a schematic diagram of federated learning provided by an embodiment;

[0037] Figure 3 is a flow chart of a federated learning method based on anchor point aggregation provided by an embodiment;

[0038] Figure 4 It is a flowchart of a federated learning method based on anchor point aggregation provided by another embodiment. DETAILED DESCRIPTION

[0039] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0040] Figure 1 It is a structural diagram of a federated learning system based on anchor point aggregation provided in an embodiment. Figure 2 Schematic diagram of federated learning provided by the embodiment. Figure 1 and Figure 2 As shown, the federated learning system provided in the embodiment includes multiple clients and a central server, wherein the client can be any electronic device with computing power, including but not limited to smart watches, smart phones, various computers, etc. The central server can be a terminal server located in the terminal, or a cloud server located in the cloud. The central server has a communication channel with each client participating in the federated learning to achieve secure communication of data and messages.

[0041] In the embodiment, the central server serves as the instructor of the federated learning system, and is used to construct a federated learning task, and sends it to each client to organize the federated learning. The federated learning task includes an initial local model, an initial global anchor point, and federated learning hyperparameters, wherein the federated learning hyperparameters include a gradient descent algorithm optimizer, a local training learning rate lr, the total number of clients participating in the federated learning, the number of local training rounds of the client (local epoch), the total communication rounds (global epoch), the spatial dimension of the local anchor point, the weight decay rate (weight_decay), and the momentum value or learning rate required during the local anchor point update process.

[0042] It should be noted that the spatial dimension of the local anchor point constrains the spatial dimension of each local anchor point, which can be 64, 128 or 256, that is, 64, 128 or 256 data are used to represent a local anchor point.

[0043] Each client has a different data set, which is used for local training based on the received federated learning tasks. Before local training, it is also used to preprocess the local data to obtain high-quality training data.

[0044] The local model includes a feature extraction network, a projection network and an output layer, wherein the feature extraction network is used to extract the feature map of the input sample data, and the projection network is used to project the feature map into a feature projection space of limited dimension to obtain the corresponding feature vector. The output layer is used to perform prediction calculations based on the feature vector and classify the prediction results. Preferably, the feature extraction network can be a feature extraction network of a Simple-CNN network, a feature extraction network of Vgg-11, a feature extraction network of ResNet-18, and a feature extraction network of ResNet-54. The projection network can adopt an MLP network. The output layer can adopt a fully connected neural network.

[0045] The local training of each client includes: training the current local model based on the local data set guided by the global anchor point of the current round, updating the local anchor point, and uploading the new local anchor point to the central server.

[0046] In the embodiment, when the client performs the initial round of federated learning, the current global anchor point is the initial global anchor point issued by the central server; when performing other rounds of federated learning other than the initial round, the global anchor point of the current round is the global anchor point issued by the central server. It should be noted that during the training process, feature extraction condenses the feature extraction of local data, and takes the feature points in the real feature vector corresponding to the feature projection that can represent the overall features of the classification category (feature distribution knowledge of local data) as local anchor points. The local anchor points contain the feature distribution knowledge of local data of each type of data, which plays a key role in data classification.

[0047] In an embodiment, when the client guides the local data set to train the local model based on the current global anchor point, the training algorithm adopted is mainly an algorithm that combines the loss function from the true value label based on general supervised learning and the loss function from the relationship measurement function based on contrastive learning.

[0048] Specifically, the loss function used is:

[0049] loss = αloss sup +(1-α)loss an

[0050] Among them, α is the weight, loss sup Construct a task loss for the classification prediction task based on the classification prediction result output by the local model, which can be the classification prediction result out and the true value label The cross entropy loss function is expressed as

[0051] loss an The guidance loss for guiding local model training based on the global anchor point is used to measure the relationship between the feature vector extracted from the input sample data and the global anchor point guiding local training in the feature projection space, which is expressed as:

[0052]

[0053] in, It represents the real feature vector of the i-th input data belonging to the j-class data after being projected by the local model in the feature projection space. It represents the vector representation of the j-th global anchor point in the feature projection space. relation(·) represents the relationship function, which is used to measure the distance or similarity. T is a hyperparameter with a value of 0-1. n is the category index and N is the total number of categories.

[0054] In the embodiment, the relation function relation(·) can be used to measure the Euclidean distance or cosine similarity. an The local feature projection of the client is calibrated mainly based on the global anchor point, so that the feature projection of a certain category after feature extraction in the feature projection space is close to the global one.

[0055] In the embodiment, when the client trains the local model, it is also necessary to update the local anchor point. The core idea is that the updated new local anchor point is the center of the local data category feature projection. Specifically, two methods can be used to update the local anchor point. Method 1: Use momentum optimization algorithm to update the local anchor point, including:

[0056] During the training of the local model, the real feature vector of the sample data in the feature projection space after being projected by the local model is used, and the momentum optimization algorithm is used to update the current local anchor point. After multiple batches of data iterations, the feature points that can represent the classification category in the real feature vector are obtained as new local anchor points. It should be noted that when the momentum optimization algorithm is used to update the local anchor point, the required momentum value is configured by the central server and sent to the client through the federated learning task.

[0057] Method 2: Use the central loss function to update the local anchor point, including:

[0058] The local anchor point is regarded as a training parameter. During the training process, the real feature vector of the sample data in the feature projection space after being projected by the local model is subtracted from the local anchor point to construct a central loss function. The central loss function is used to update the position of the local anchor point to reduce the average distance between the real feature vector and the local anchor point. After multiple batches of data iterations, the new local anchor point can balance the distance between the corresponding feature vectors in each category to approach the center of each feature vector and achieve stability, so as to be sufficient to represent the data characteristics of each classification category. It should be noted that when the central loss function is used to update the local anchor point, the required learning rate is set by the central server configuration and sent to the client through the federated learning task.

[0059] The new local anchor points updated by the above two methods are uploaded to the central server for anchor aggregation. The central server, as the aggregation center of federated learning, is used to aggregate and update the new local anchor points uploaded by all clients to obtain the global anchor points corresponding to each classification category, and send the global anchor points to each client as the global anchor point of the client's current round.

[0060] In the embodiment, when the central server aggregates and updates the new local anchor points uploaded by all clients, the method adopted includes: weighted summing and averaging all the new local anchor points representing each classification category to obtain a global anchor point for each classification category.

[0061] In the embodiment, the number of classification categories for each client during local training is agreed upon, and the number of local anchor points is configured to be equal to the number of classification categories, that is, a representative feature point is extracted from each type of data as a local anchor point. In this way, when performing anchor aggregation, all local anchor points representing each classification category are weighted summed and averaged to obtain a global anchor point representing the classification category. The number of global anchor points obtained in this way is equal to the number of classification categories, which ensures that the dimensions of the local anchor points uploaded by the client are equal to the dimensions of the global anchor points.

[0062] In the embodiment, data of the same category will be more concentrated in the feature space after being extracted and processed by this model. The projection center of the projection of data of the same category is used as the local anchor point to express the properties of the local data feature projection. The central server integrates all local anchor points, condenses and extracts global data information, and obtains a global anchor point that can represent the global data projection information. This method of adjusting the training of the local model based on the global anchor point that guides local training makes the local data features approach the global unified standard, so that the data features of the same category are more concentrated in the feature space, thereby improving the effect of federated training.

[0063] In this federated learning process that uses local anchors for guidance and constraints, since there is no need to upload local model parameters, only local anchors need to be uploaded, the communication level of local anchors is greatly reduced, which significantly improves communication efficiency. Furthermore, when federated learning is performed only through anchor aggregation, model parameters are not aggregated. Therefore, the local models trained by each client are heterogeneous and suitable for their respective application needs. The unified restrictions on the model structure improve the flexibility of federated training.

[0064] In another embodiment, when uploading a new local anchor point, the client is also used to upload local model parameters to the central server; at this time, the central server is also used to aggregate the local model parameters uploaded by all clients to obtain a global model, and send the global model to each client for the next round of local training.

[0065] When this type of federated learning performs anchor point aggregation and model parameter aggregation simultaneously, the client uploads the local anchor point and local model parameters at the same time. The anchor point aggregation and parameter aggregation can be organically combined. Since the anchor point communication level differs from the parameter communication level by at least 1000 orders of magnitude, the anchor point communication transmission can be ignored compared to the parameter communication transmission, which can improve the federated performance of the model.

[0066] Figure 3 is a flow chart of a federated learning method based on anchor point aggregation provided in an embodiment. Figure 3 As shown, the anchor point aggregation-based federated learning method provided in the embodiment adopts the anchor point aggregation-based federated learning system provided in the above embodiment. Specifically, the federated learning method includes the following steps:

[0067] Step 1: The center establishes a communication channel with the client.

[0068] Step 2: The central server negotiates with the client and creates a federated learning task, and sends the federated learning task to each client.

[0069] Among them, the federated learning task includes the initial local model, the initial global anchor point, and the federated learning hyperparameters.

[0070] Step 3: The client preprocesses the local data to improve the quality of the local data.

[0071] Step 4: The client guides the local data set to train the current local model based on the current global anchor point, and after updating the local anchor point, uploads the new local anchor point to the central server.

[0072] It should be noted that when the client is performing the initial round of federated learning, the current global anchor point is the initial global anchor point issued by the central server; when performing other rounds of federated learning other than the initial round, the current global anchor point is the global anchor point issued by the central server.

[0073] In step 5, the central server aggregates and updates the new local anchor points uploaded by all clients to obtain the global anchor points corresponding to each classification category, and sends the global anchor points to each client to guide the next round of federated learning.

[0074] Step 6: Iterate steps 4 and 5 until the training termination condition is reached.

[0075] In the embodiment, the training termination condition is the total number of communication rounds.

[0076] Figure 4 is a flow chart of a federated learning method based on anchor point aggregation provided in an embodiment. Figure 4 As shown, the anchor point aggregation-based federated learning method provided in the embodiment adopts the anchor point aggregation-based federated learning system provided in the above embodiment. Specifically, the federated learning method includes the following steps:

[0077] Step 1: The center establishes a communication channel with the client.

[0078] Step 2: The central server negotiates with the client and creates a federated learning task, and sends the federated learning task to each client.

[0079] Among them, the federated learning task includes the initial local model, the initial local anchor point, and the federated learning hyperparameters.

[0080] Step 3: The client preprocesses the local data to improve the quality of the local data.

[0081] Step 4: The client guides the local data set to train the current local model based on the current global anchor point, updates the local anchor point and local model parameters, and then uploads the new local anchor point and local model parameters to the central server.

[0082] It should be noted that when the client is performing the initial round of federated learning, the current global anchor point is the initial global anchor point issued by the central server, and the current local model is the initial local model issued by the central server; when performing other rounds of federated learning other than the initial round, the current global anchor point is the global anchor point issued by the central server, and the current local model is the global model issued by the central server.

[0083] In step 5, the central server aggregates and updates the new local anchor points and local model parameters uploaded by all clients to obtain the global anchor points and global models corresponding to each classification category, and sends the global anchor points and global models to each client as the client's current global anchor point and current local model for the next round of federated learning.

[0084] Step 6: Iterate steps 4 and 5 until the training termination condition is reached.

[0085] In the embodiment, the training termination condition is the total number of communication rounds.

[0086] In the above-mentioned federated learning method, different clients extract features of various categories of local data through feature extraction networks and projection networks, and then process them to form local anchor points representing the distribution characteristics of local data. After that, in the global communication process, only local anchor points that use condensed knowledge are selected for information exchange according to actual needs, or both models and local anchor points are used for information exchange. After obtaining the local anchor points of each participating client, the central parameter server performs decision and processing to obtain a global anchor point that can represent global data information. In subsequent training, the global anchor point or / and the global model that can represent global information are sent to each client participating in the training to guide subsequent federated learning training. This can effectively reduce the communication level of federated training on the basis of ensuring the accuracy of model training, while supporting federated learning of heterogeneous models.

[0087] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A federated learning system based on anchor point aggregation, characterized in that: It includes multiple clients and a central server, and each client has a communication channel with the central server; The client is used to guide the local data set to train the current local model according to the global anchor point of the current round, and after updating the local anchor point, upload the new local anchor point to the central server, where the local anchor point is a feature point that can represent the classification category determined according to the feature vector corresponding to the local data; Among them, the first method: using the momentum optimization algorithm to update the local anchor point, including: in the process of training the local model, according to the real feature vector of the sample data after being projected by the local model in the feature projection space, using the momentum optimization algorithm to update the current local anchor point, and after iterating multiple batches of data, obtaining the feature points that can represent the classification category in the real feature vector as the new local anchor point; Method 2: Using the central loss function to update the local anchor point, including: considering the local anchor point as a training parameter, in the training process, taking the difference between the real feature vector of the sample data in the feature projection space after being projected by the local model and the local anchor point to construct the central loss function, and using the central loss function to update the position of the local anchor point to reduce the average distance between the real feature vector and the local anchor point. After multiple batches of data iterations, the new local anchor point can balance the distance between the corresponding feature vectors in each category, so as to approach the center of each feature vector and achieve stability, so as to be sufficient to represent the data characteristics of each classification category; The central server is used to aggregate and update the new local anchor points uploaded by all clients to obtain the global anchor points corresponding to each classification category, and send the global anchor points to each client to guide the next round of local training.

2. The anchor point aggregation-based federated learning system according to claim 1, characterized in that: When the client guides the local data set to train the local model based on the current local anchor point, the loss function used is: loss=αloss sup +(1-α)loss an Among them, α is the weight, loss sup Constructing a task loss for the classification prediction task based on the classification prediction results output by the local model; loss an It is the guidance loss for guiding local model training based on the global anchor point, which is used to measure the relationship between the feature vector extracted from the input sample data and the global anchor point guiding local training in the feature projection space, and is expressed as: in, It represents the real feature vector of the i-th input data belonging to the j-class data after being projected by the local model in the feature projection space. It represents the vector representation of the j-th global anchor point in the feature projection space. relation(·) represents the relationship function, which is used to measure the distance or similarity. T is a hyperparameter with a value of 0-1. n is the category index and N is the total number of categories.

3. The anchor point aggregation-based federated learning system according to claim 1, characterized in that: The central server aggregates and updates the new local anchor points uploaded by all clients, including: weighted summing and averaging all new local anchor points representing each classification category to obtain a global anchor point for each classification category.

4. The anchor point aggregation-based federated learning system according to claim 1, characterized in that: Before federated learning, the central server is also used to build a federated learning task and send it to each client. The federated learning task includes the initial local model, the initial global anchor point, and the federated learning hyperparameters. The federated learning hyperparameters include the gradient descent algorithm optimizer, the local training learning rate, the total number of clients participating in federated learning, the number of local training rounds of the client, the total number of communication rounds, the spatial dimension of the local anchor point, the weight decay rate, and the momentum value or learning rate required during the local anchor point update process. The client and the central server perform federated learning based on the federated learning task.

5. The anchor point aggregation-based federated learning system according to claim 1, characterized in that: The local model includes a feature extraction network, a projection network and an output layer, wherein the feature extraction network is used to extract the feature map of the input sample data, the projection network is used to project the feature map into the feature projection space to obtain the feature vector, and the output layer is used to perform prediction calculations based on the feature vector and classify the data to predict the results.

6. The anchor point aggregation-based federated learning system according to any one of claims 1 to 5, characterized in that: The client is also used to upload local model parameters to the central server; the central server is also used to aggregate the local model parameters uploaded by all clients to obtain a global model, and send the global model to each client as the client's current local model.

7. The anchor point aggregation-based federated learning system according to any one of claims 1 to 5, characterized in that: After federated learning is completed, each client uses its current local model as the classification model for classification prediction tasks.

8. The anchor point aggregation-based federated learning system according to claim 6, characterized in that: After federated learning is completed, each client uses its current local model as the classification model for classification prediction tasks.

9. A federated learning method based on anchor point aggregation, characterized in that: The federated learning method adopts the federated learning system according to any one of claims 1 to 8, and the federated learning method includes: The central server establishes a communication channel with the client, creates a federated learning task, and sends it to each client; The client guides the local data set to train the current local model based on the global anchor point of the current round, and after updating the local anchor point, uploads the new local anchor point to the central server, where the local anchor point is a feature point that can represent the classification category determined according to the feature vector corresponding to the local data; Among them, the first method: using the momentum optimization algorithm to update the local anchor point, including: in the process of training the local model, according to the real feature vector of the sample data after being projected by the local model in the feature projection space, using the momentum optimization algorithm to update the current local anchor point, and after iterating multiple batches of data, obtaining the feature points that can represent the classification category in the real feature vector as the new local anchor point; Method 2: Using the central loss function to update the local anchor point, including: considering the local anchor point as a training parameter, in the training process, taking the difference between the real feature vector of the sample data in the feature projection space after being projected by the local model and the local anchor point to construct the central loss function, and using the central loss function to update the position of the local anchor point to reduce the average distance between the real feature vector and the local anchor point. After multiple batches of data iterations, the new local anchor point can balance the distance between the corresponding feature vectors in each category, so as to approach the center of each feature vector and achieve stability, so as to be sufficient to represent the data characteristics of each classification category; The central server aggregates and updates the new local anchor points uploaded by all clients to obtain the global anchor points corresponding to each classification category, and sends the global anchor points to each client to guide the next round of local training.

Images