Federal category incremental learning method, user side, server and storage medium

Abstract

The invention discloses a federal category incremental learning method, a user side, a server and a storage medium, and the method comprises the steps: enabling a conditional generative adversarial network model sent by the server to serve as a local model of the user side, then training the local model in two stages: the first stage is to train the local model according to the current task data and the simulated historical task data, and the second stage is to train the local model according to the target task data and the simulated historical task data after balanced sampling; and the user side sends the trained local models to a server, so that the server aggregates the local models to generate a target global model. According to the method, the user side uses the conditional generative adversarial network model to generate the simulation task data in a targeted manner, catastrophic forgetting caused by incremental learning is effectively relieved, classification discrimination caused by unbalanced data is reduced by using two-stage training, and the recognition accuracy of the finally generated target global model can be improved.

Description

Technical Field

[0001] The present application relates to the field of federated incremental learning, and in particular to a federated category incremental learning method, a user terminal, a server and a storage medium. Background Art

[0002] In the era of artificial intelligence that relies on massive data, data is undoubtedly the most valuable resource. However, data is usually stored in a dispersed and fragmented manner in various departments and institutions, forming data islands, which makes it extremely difficult to extract effective information from massive data. Many obstacles need to be overcome to achieve data integration and effective use. In addition, the public's attention to data security has increased significantly, which also brings challenges to data privacy protection.

[0003] Federated learning (Federated Learning / Federated machine learning) technology can solve the data island problem while protecting data privacy. In federated learning, it can usually be divided into client and server. The client refers to each local device or entity participating in federated learning, such as mobile devices (such as smartphones, tablets), sensor nodes or local servers. These devices usually hold local data, such as user's personal data or sensor data. The server refers to the centralized entity responsible for coordinating and organizing the federated learning process. The server is usually responsible for managing and controlling the global model update of federated learning, coordinating the participation of the client, and processing the aggregated updated model. During the federated learning process, the data of the client does not need to be transmitted to the server, but the data is trained locally on the client, thereby reducing the risk of data leakage. After the client uses the local data to independently train the model, it sends the model update to the server. The server is responsible for collecting the model updates of all clients and using the corresponding algorithm to aggregate these updates to form a global model. The aggregated global model updates are sent back to the client, and the client uses these updates to improve its own model, thereby solving the data island problem.

[0004] However, in real-world applications, in addition to considering the constraints of privacy protection, solving the data island problem also requires dealing with the problem of limited storage space. Due to the limited storage space of user-side devices, it is impossible to save a large amount of training data for a long time, which makes the training data available only for a period of time. Therefore, historical data cannot be used because it cannot be saved for a long time, but relying only on currently available data for updates will cause many machine learning models to suffer catastrophic forgetting. For this reason, the concept of incremental learning was proposed to develop intelligent systems that can learn and accumulate knowledge from new tasks while avoiding forgetting learned knowledge in the absence of previous data. Incremental learning scenarios are mainly divided into three categories: task increment, domain increment, and category increment. In task increment, the task label is known, and only predictions need to be made within the label space of a specific task; in domain increment, the problem structure remains unchanged, but the input distribution changes; and in category increment, the algorithm must gradually learn to distinguish the increasing categories. Each incremental task contains different classes, and the algorithm needs to identify all encountered categories, so federated category incremental learning is also more realistic for building robust classifiers in the real world. Furthermore, Qi et al. (Better Generative Replay for Continual Federated Learning [C], The Eleventh International Conference on Learning Representations (ICLR), Kigali Rwanda, 2023, 1-17.) proposed the FedCIL (Federated Class-Incremental Learning) method in 2023, which uses replay technology and knowledge distillation to compensate for the lack of old data and overcome the forgetting problem. However, although the FedCIL method effectively addresses the problem of data forgetting by using replay technology and knowledge distillation for data enhancement, when faced with unbalanced training data, this method will further aggravate the bias caused by imbalance, leading to discrimination against more categories, resulting in more serious catastrophic forgetting. Therefore, it is urgent to propose a new federated class incremental learning method to solve the challenges faced by federated class incremental learning in imbalanced data scenarios, that is, how to build a unified global model that can recognize and classify all learned categories without relying on the storage of old data.

[0005] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0006] The embodiments of the present application provide a federated category incremental learning method, a user terminal, a server, and a storage medium, aiming to solve the technical problem of how to build a unified global model that can identify and classify all learned categories without relying on the storage of old data.

[0007] To achieve the above object, an embodiment of the present application provides a federated category incremental learning method, which is applied to a user end. The federated category incremental learning method includes the following:

[0008] Receiving the conditional generative adversarial network model sent by the server, and using the conditional generative adversarial network model as a local model of the user end;

[0009] According to current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task, performing a first stage training of a first preset round on the local model;

[0010] Performing balanced sampling on the current task data corresponding to the current task to obtain target task data;

[0011] According to the target task data, or according to the target task data and the simulated historical task data, performing a second-stage training of a second preset round on the local model;

[0012] The local model is sent to the server, so that the server generates a target global model based on the local model.

[0013] Optionally, after the step of sending the local model to the server, the method further includes:

[0014] When receiving the target global model sent by the server, updating the target global model to the conditional generative adversarial network model;

[0015] Continue to execute the step of using the conditional generative adversarial network model as a local model of the user end.

[0016] Optionally, the method further includes:

[0017] Upon receiving the target global model sent by the server, determining the number of communications between the client and the server;

[0018] If the number of communications reaches the preset number of communications, task data of the next task is obtained, and the task data is updated to the current task data of the current task.

[0019] Optionally, the step of performing the first stage training of the first preset round on the local model based on the current task data corresponding to the current task, or based on the current task data and simulated historical task data corresponding to the current task, includes: if the current task is the first task, using the current task data as the input of the local model; if the current task is not the first task, using the current task data and the simulated historical task data as the input of the local model, wherein the simulated historical task data is generated based on the historical local model of the previous task.

[0020] Optionally, the step of performing a first-stage training of a first preset round on the local model according to current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task, includes:

[0021] If the current task is the first task, determine the total loss of the local model based on the current task data;

[0022] Updating model parameters of the local model according to the total loss;

[0023] If the current task is not the first task, determining the total loss of the local model based on the current task data and the simulated historical task data;

[0024] Model parameters of the local model are updated according to the total loss.

[0025] Optionally, the step of performing a second preset round of second-stage training on the local model based on the target task data and the simulated historical task data includes: if the current task is the first task, using the target task data as input to the local model; if the current task is not the first task, using the target task data and the simulated historical task data as input to the local model, wherein the simulated historical task data is generated based on the historical local model of the previous task.

[0026] Optionally, before the step of performing a second preset round of second stage training on the local model according to the target task data and the simulated historical task data, the step includes:

[0027] Freeze model parameters of the feature extraction layer of the local model.

[0028] Optionally, the step of performing a second-stage training of a second preset round on the local model according to the target task data and the simulated historical task data includes:

[0029] If the current task is the first task, determining the total loss of the local model based on the target task data;

[0030] Updating model parameters of the local model according to the total loss;

[0031] If the current task is not the first task, determining the total loss of the local model based on the target task data and the simulated historical task data;

[0032] Model parameters of the local model are updated according to the total loss.

[0033] To achieve the above objectives, an embodiment of the present application provides a federated category incremental learning method, which is applied to a server. The federated category incremental learning method includes the following:

[0034] Send the initialized conditional generative adversarial network model to the user end;

[0035] When receiving the local model sent by the user terminal, aggregating the local model to generate a global model;

[0036] The global model is trained for preset rounds according to the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain a target global model.

[0037] Optionally, after the step of training the global model for a preset round according to the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain the target global model, the method further comprises:

[0038] The target global model is sent to the user end.

[0039] Optionally, before the step of sending the target global model to the user terminal, the method further includes:

[0040] Determining the number of communications between the server and the client;

[0041] If the number of communications does not reach the communication number threshold, the target global model is sent to the user terminal.

[0042] In addition, to achieve the above-mentioned purpose, the present application also provides a user terminal, which includes: a memory, a processor, and a federated category incremental learning program stored in the memory and executable on the processor, and the federated category incremental learning program implements the steps of the federated category incremental learning method as described above when executed by the processor.

[0043] In addition, to achieve the above-mentioned purpose, the present application also provides a server, which includes: a memory, a processor, and a federated category incremental learning program stored on the memory and executable on the processor, and the federated category incremental learning program implements the steps of the federated category incremental learning method as described above when executed by the processor.

[0044] In addition, to achieve the above-mentioned purpose, the present application also provides a computer-readable storage medium, on which a federated category incremental learning program is stored. When the federated category incremental learning program is executed by a processor, the steps of the federated category incremental learning method as described above are implemented.

[0045] A method for incremental learning of federated categories proposed in the first embodiment of the present application, a user terminal, a server and a storage medium. In the technical solution provided in this embodiment, the user terminal receives the conditional generative adversarial network model sent by the server, and uses the conditional generative adversarial network model as the local model of the user terminal. According to the current task data corresponding to the current task, or according to the current task data corresponding to the current task and the simulated historical task data, the local model is trained for the first stage of the first preset round, and the current task data corresponding to the current task is balanced sampled to obtain the target task data, and then according to the target task data, or according to the target task data and the simulated historical task data, the local model is trained for the second stage of the second preset round. After completing the two-stage training, the local model is sent to the server so that the server generates a target global model based on the local model. In the first stage training and the second stage training of the local model, the simulated historical task data is used as the input of the local model for training, thereby effectively alleviating the catastrophic forgetting problem caused by incremental learning. The classification discrimination caused by unbalanced data is solved through two-stage replay training, and the recognition accuracy of the target global model generated by the server can be improved.

[0046] This application has at least the following beneficial effects:

[0047] (1) Since the existing federated incremental learning methods cannot effectively handle unbalanced data, the unbalanced data seriously deepens the catastrophic forgetting of the model, causing classification discrimination against majority category samples, which seriously affects the final model recognition accuracy. Based on the ACGAN network and combined with the playback method, this application proposes a method for federated category incremental learning in an unbalanced data environment. The method generates simulated balanced historical data and current data on the user side and the server through the ACGAN network, and uses these simulated task data to train the model, thereby alleviating the catastrophic forgetting phenomenon caused by incremental learning.

[0048] (2) For the user side that stores unbalanced data, this application designs a two-stage retraining method. The first stage is to train the local model based on the current task data and simulated historical task data, and the second stage is to train the local model based on the target task data after balanced sampling and simulated historical task data. It can reduce the negative impact of unbalanced data on the local model and alleviate the classification discrimination problem, thereby improving the anti-catastrophic forgetting ability and final accuracy of the target global model generated by the server. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flowchart of the first embodiment of the federated category incremental learning method of the present application;

[0050] Figure 2 This is a training diagram of the user end of this application;

[0051] Figure 3 A flowchart of a second embodiment of the federated category incremental learning method of the present application;

[0052] Figure 4 This is a training diagram of the server;

[0053] Figure 5 It is a schematic diagram of the terminal structure of the hardware operating environment involved in the embodiment of the present application.

[0054] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0055] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0056] In order to solve the technical problem of how to build a unified global model that can identify and classify all learned categories without relying on old data storage, this application proposes a federated category incremental learning method, whose main solution steps include the following:

[0057] In the technical solution provided in this embodiment, by receiving the conditional generative adversarial network model sent by the server, the conditional generative adversarial network model is used as the local model of the user end, and the local model is trained for the first stage of the first preset round according to the current task data corresponding to the current task, or according to the current task data corresponding to the current task and the simulated historical task data, and the current task data corresponding to the current task is balanced sampled to obtain the target task data, and then the local model is trained for the second stage of the second preset round according to the target task data, or according to the target task data and the simulated historical task data. After completing the two-stage training, the local model is sent to the server so that the server generates a target global model based on the local model. In the present application, the simulated historical task data is used as the input of the local model for training in both the first stage training and the second stage training of the local model, thereby effectively alleviating the catastrophic forgetting problem caused by incremental learning. The classification discrimination caused by unbalanced data is solved through two-stage replay training, and the recognition accuracy of the target global model generated by the server can be improved.

[0058] In order to better understand the above technical solution, exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0059] Reference Figure 1 In the first embodiment of the federated category incremental learning method of the present application, the federated category incremental learning method is applied to the user end, including steps S10 to S50:

[0060] Step S10: receiving the conditional generative adversarial network model sent by the server, and using the conditional generative adversarial network model as a local model of the user end;

[0061] In this embodiment, after the server initializes the conditional generative adversarial network model, it is uniformly sent to each user terminal. The user terminal uses the conditional generative adversarial network model as a local model and trains the local model to learn the newly added categories or category sets of the user terminal, so that it has the ability to recognize and classify new data. The local model includes a generator and a discriminator.

[0062] The conditional generative adversarial network model of the present application is an ACGAN (Conditional Image Synthesis With Auxiliary Classifier GANs) network model, wherein the ACGAN network model can not only generate realistic simulation task data, but also control the category and quantity of generated simulation task data, so that when the local model is trained in two stages, class-balanced simulation task data can be generated, thereby reducing the classification discrimination caused by unbalanced data.

[0063] Step S20: performing a first-stage training of a first preset round on the local model according to current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task;

[0064] In this embodiment, in the first stage of training, the local model needs to undergo a first preset round of training, and each round is to train the local model based on the current task data corresponding to the current task, or based on the current task data and simulated task data corresponding to the current task.

[0065] Optionally, the first preset round may be set to 160.

[0066] Specifically, the step of performing the first stage training of the first preset round on the local model according to the current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task, includes: if the current task is the first task, using the current task data as the input of the local model. If the current task is not the first task, using the current task data and simulated historical task data as the input of the local model, wherein the simulated historical task data is generated based on the historical local model of the previous task.

[0067] In this embodiment, the local model is trained and learned for different types of tasks in turn to continuously increase the recognition ability of different types of data. The current task data corresponding to the current task is stored on the user side and can be directly obtained. The simulated historical task data is generated based on the historical local model of the previous task.

[0068] It should be noted that the historical local model of the previous task refers to the local model obtained after the last two-stage training in the federated category incremental learning of the previous task. This historical local model has the ability to recognize all categories of data learned in the historical tasks. Therefore, the category labels of all categories learned in the historical tasks and the corresponding generation quantity of each category label can be input into the historical local model of the previous task to generate simulated historical task data corresponding to all categories learned in the historical tasks.

[0069] Optionally, the number of generated data corresponding to each category label input into the historical local model can be set to be the same, so that the number of generated simulated historical task data for each category is the same, that is, class-balanced simulated historical task data is generated, thereby reducing classification discrimination caused by unbalanced data.

[0070] It should also be noted that when the application is conducting the first stage of training, if the current task is not the first task, the application will use the current task data and simulated historical task data corresponding to the current task as the input of the local model for training, so that the local model will not forget the recognition ability of all previously learned data corresponding to the categories in the incremental learning of the federated categories, thereby solving the catastrophic forgetting problem caused by the incremental learning of the federated categories. And the simulated historical task data is generated using the historical local model of the previous task, so the user end does not need to rely on the learned historical task data, and then the user end does not need to store the historical task data, which can save the user end's memory resources.

[0071] Optionally, since the local model is an ACGAN network model, when generating simulated task data, the input of the local model generator includes a category label and a noise vector. The noise vector can be randomly generated based on a Gaussian distribution. Therefore, when the user terminal generates simulated historical task data based on the historical local model, it only needs to input the category label and the number of generation corresponding to the category label.

[0072] For example, y = [A, A, A, B, B, B, C, C, C] can be input into the historical local model, and the historical local model will generate 3 simulated historical task data for each of categories A, B, and C. It can be understood that each letter (A, B, C) in y = [A, A, A, B, B, B, C, C, C] represents a category label, and the number of repetitions represents the number of samples that need to be generated for each category.

[0073] Furthermore, the step of performing the first-stage training of the local model for the first preset round according to the current task data corresponding to the current task, or according to the current task data corresponding to the current task and the simulated historical task data, includes: if the current task is the first task, determining the total loss of the local model based on the current task data, and then updating the model parameters of the local model according to the total loss. If the current task is not the first task, determining the total loss of the local model based on the current task data and the simulated historical task data, and then updating the model parameters of the local model according to the total loss.

[0074] In this embodiment, when the local model is training in the first stage, if the current task is the first task, the input of the local model is the current task data, and then the local model performs category learning based on the input current task data. If the current task is not the first task, the input of the local model is the current task data and the simulated historical task data, and then the local model performs category learning based on the input current task data and the simulated historical task data. When the categories corresponding to the current task data and the simulated historical task data are learned, the generator generates corresponding simulated task data based on the categories.

[0075] Simulated historical mission data generated by the historical local model of the previous mission It is expressed as:

[0076]

[0077] in, represents the generator of the historical local model of the previous task, z represents the noise vector, y represents the category label, Represents all the categories currently learned by the historical local model of the previous task of user u.

[0078] The total loss L of the local model u It is expressed as:

[0079]

[0080] in, is the loss of the discriminator D, is the loss of the generator. Represents the total loss calculated by the training data currently input by the local model. The training data currently input to the local model. Represents the generator of the historical local model through the previous task The total loss calculated from the generated simulated historical task data, Represents the generator of the historical local model through the previous task The generated simulated historical task data. u Represents the total loss of the historical local model for the previous task The total loss with the current local model The weight hyperparameter between u ∈[0,1].

[0081] It is understandable that the total loss L of the local model u is the total loss based on the local model itself and the total loss of the historical local model of the previous task Obtained by calculation.

[0082] The loss L of the discriminator D D It is expressed as:

[0083]

[0084] L D =L D,adv +L D,aux

[0085] Among them, L D,adv represents the adversarial loss of the discriminator D, X real The training data for the model input, X fake The simulated task data generated by the model's generator D is expected by the model to be able to discriminate the input X real and the simulated task data X generated by the generator G fake Correctly distinguish. D,aux L is the classification loss of the discriminator D. It is expected that the discriminator D can correctly classify the real or simulated task data. D is the loss of the discriminator D. It should be noted that the loss of the discriminator D

[0086] The loss L of the generator G G It is expressed as:

[0087]

[0088] L G =L G,adv +L G,aux

[0089] Among them, L G,adv is the adversarial loss of the generator G, X real The training data for the model input, X fake The simulated task data generated by the model's generator D is the simulated task data X generated by the model's generator G. fake It can fool the discriminator D, making it think that the simulated task data is real. G,aux is the classification loss of generator G, and the model expects the simulated task data X generated by generator G fake It can also be correctly classified by the discriminator D. G is the loss of generator G. It should be noted that the loss of generator G

[0090] Optionally, the step of updating the model parameters of the local model according to the total loss includes: updating the model parameters of the local model according to the total loss of the local model and using a gradient descent method.

[0091] Specifically, the model parameters D after the local model discriminator D is updated u,k+1 It is expressed as:

[0092] X one =X ins +X his

[0093] D u,k+1 =D u,k -η×▽L(D u,k ;X one )

[0094] Where η is the learning rate, ▽L(D u,k ;X now ) is the total loss at time step k for all parameters of the local model D u,k The gradient of X ins is the current task data, X his The simulated historical task data, X one The training data currently input to the local model in the first stage.

[0095] Step S30: performing balanced sampling on the current task data corresponding to the current task to obtain target task data;

[0096] In this embodiment, before conducting the second stage of training, the current task data of each category is balanced and sampled so that the number of task data of each category in the sampled target task data is the same, avoiding the bias of the model due to the imbalance of the categories of the task data input into the local model, thereby reducing the classification discrimination caused by the imbalanced data.

[0097] Step S40: performing a second-stage training of a second preset round on the local model according to the target task data, or according to the target task data and the simulated historical task data;

[0098] In this embodiment, in the second stage of training, the local model needs to undergo a second preset round of training, and each round is to train the local model based on the target task data, or based on the target task data and the simulated task data.

[0099] It should be noted that during the first stage of training, the current task data is directly used as the input of the local model without balanced sampling of the current task data. This is to enable the local model to be exposed to the characteristics and distribution of all real data during the training process, so that the generator of the local model can learn a wider range of data features, including samples in rare or extreme cases, thereby improving the diversity and quality of the data generated by the local model generator. During the second stage of training, balanced sampling of the current task data is performed to reduce the classification discrimination caused by unbalanced data.

[0100] Optionally, the second preset round may be set to 40.

[0101] Specifically, the step of performing a second-stage training of the local model for a second preset round according to the target task data and the simulated historical task data includes: if the current task is the first task, using the target task data as the input of the local model. If the current task is not the first task, using the target task data and the simulated historical task data as the input of the local model, wherein the simulated historical task data is generated based on the historical local model of the previous task.

[0102] In this embodiment, the generation of simulated historical task data can refer to the generation scheme given in the first stage of training above, which will not be repeated here.

[0103] Similarly, when the application is conducting the second stage of training, if the current task is not the first task, the application will use the target task data and simulated historical task data as the input of the local model for training, so that the local model will not forget the recognition ability of all previously learned data corresponding to the categories in the incremental learning of the federated categories, thereby solving the catastrophic forgetting problem caused by the incremental learning of the federated categories. And the simulated historical task data is generated using the historical local model of the previous task, so the user end does not need to rely on the learned historical task data, and thus the user end does not need to store the historical task data, which can save memory resources.

[0104] It should be noted that, before the step of performing a second preset round of second stage training on the local model according to the target task data and the simulated historical task data, it includes: freezing the model parameters of the feature extraction layer of the local model.

[0105] In this embodiment, after the first stage of training of the local model is completed and before the second stage of training of the local model, it is necessary to freeze the model parameters of the feature extraction layer of the local model, that is, fix the model parameters of the feature extraction layer of the local model so that the local model will not update the model parameters of the feature extraction layer during the second stage of training. In this way, the feature extraction layer of the local model is not trained in the second stage of training, but other layers of the local model are trained. This can retain all the features and distribution of the real data learned by the local model in the first stage of training, and solve the classification discrimination caused by unbalanced data.

[0106] Furthermore, the step of performing a second-stage training of the local model for a second preset round according to the target task data and the simulated historical task data includes: if the current task is the first task, determining the total loss of the local model based on the target task data, and then updating the model parameters of the local model according to the total loss. If the current task is not the first task, determining the total loss of the local model based on the target task data and the simulated historical task data, and then updating the model parameters of the local model according to the total loss.

[0107] In this embodiment, since the model parameters of the feature extraction layer of the local model are frozen, when the model parameters of the local model are updated according to the total loss, the model parameters of the feature extraction layer of the local model will not be updated.

[0108] It should be noted that the local model generates simulated task data based on the target task data and simulated historical task data, and the total loss of the local model has been determined in detail in the above-mentioned first stage training, which will not be repeated here.

[0109] Optionally, the step of updating the model parameters of the local model according to the total loss includes: updating the model parameters of the local model according to the total loss of the local model and using a gradient descent method.

[0110] It should be noted that since the feature extraction layer of the local model is frozen in the second stage of training, the model parameter update of the local model discriminator is different from that of the local model discriminator after the first stage of training. In the second stage of training, the updated model parameters of the local model discriminator D are expressed as:

[0111] X two =X bal +X his

[0112]

[0113] Where η is the learning rate, is the total loss at time step k for the classification layer parameters The gradient of X two Represents the current training data input to the local model in the first stage, X bal represents the target task data obtained after balanced sampling, X his For the simulated historical data, represents the classification layer model parameters of the discriminator D, represents the model parameters of the feature extraction layer of the discriminator D, and u represents the user end.

[0114] It can be known that the model parameters of the local model discriminator feature extraction layer will not change. Therefore, in the second stage of training, the other layers of the local model except the feature extraction layer are trained.

[0115] Step S50: Send the local model to the server, so that the server generates a target global model based on the local model.

[0116] In this embodiment, after the local model completes the first stage training of the first preset round and the second stage training of the second preset round, the training is completed, and then the trained local model is sent to the server, so that the server aggregates the local models sent by each user terminal to generate a global model, and optimizes the generated global model to generate a target global model. This makes it unnecessary for the server to share the task data of each user terminal, and protects the data privacy of the user terminal.

[0117] Furthermore, after the step of sending the local model to the server, it also includes: when receiving the target global model sent by the server, updating the target global model to a conditional generative adversarial network model, and continuing to execute the step of using the conditional generative adversarial network model as the local model of the user end.

[0118] In this embodiment, the server aggregates the local models sent by each user terminal to generate a target global model, and then sends the target global model to each user terminal. The user terminal uses the target global model as the local model of the user terminal and continues to train the local model, thereby solving the data island problem between different user terminals. The training process of the user terminal can refer to Figure 2 , Figure 2 This is a training diagram of the user end of this application.

[0119] Optionally, the method also includes: when the user terminal receives the target global model sent by the server, determining the number of communications between the user terminal and the server, if the number of communications reaches a preset number of communications, obtaining task data for the next task, and updating the task data to the current task data of the current task, thereby performing federated category incremental learning for the next task.

[0120] In this embodiment, the training of each task requires multiple communications with the server. When the user receives the target global model sent by the server, it is necessary to determine the current number of communications with the server. If the number of communications reaches the preset number of communications, it indicates that the federated category incremental learning of the current task is completed and enters the federated category incremental learning of the next task.

[0121] It should be noted that when the user end is performing incremental learning of the federated category of the current task, it is necessary to send the local model after each two-stage training to the server so that the server generates a target global model based on the local model. After the server generates the target global model, it sends it back to the user end. Then the user end receives the target global model sent by the server and updates it to a conditional generative adversarial network model, thereby continuing to execute the step of using the conditional generative adversarial network model as the local model of the user end, so as to continue the two-stage training of the local model, so as to continue the incremental learning of the federated category for the current task. When the number of communications between the user end and the server reaches the preset number of communications, it indicates that the incremental learning of the federated category for the current task is completed, and the incremental learning of the federated category for the next task is entered.

[0122] Optionally, the preset number of communications can be determined based on the communication number threshold and the number of tasks. For example, if the communication number threshold is 200 and the number of tasks is 5, then 200 / 5=40, so the preset number of communications can be 40, 80, 120, 160, and 200 respectively. Or directly set the preset number of communications to 40, but when the number of communications reaches the preset number of communications, the counted number of communications needs to be cleared. That is, the number of communications for each task needs to be counted from zero. For example, when performing federated category incremental learning for the first task, if the number of communications reaches 40, then enter the federated category incremental learning for the next task, and clear the number of communications 40 to zero, and restart the counting. When the number of communications reaches 40 again, continue to enter the federated category incremental learning for the next task.

[0123] In the technical solution provided in this embodiment, the user terminal receives the conditional generative adversarial network model sent by the server, uses the conditional generative adversarial network model as the local model of the user terminal, performs the first stage training of the first preset round on the local model according to the current task data corresponding to the current task, or according to the current task data corresponding to the current task and the simulated historical task data, performs balanced sampling on the current task data corresponding to the current task, obtains the target task data, and then performs the second stage training of the second preset round on the local model according to the target task data, or according to the target task data and the simulated historical task data, after completing the two-stage training, sends the local model to the server, so that the server generates the target global model based on the local model. In the present application, the simulated historical task data is used as the input of the local model for training in both the first stage training and the second stage training of the local model, thereby effectively alleviating the catastrophic forgetting problem caused by incremental learning. The classification discrimination caused by unbalanced data is solved through two-stage replay training, and the recognition accuracy of the target global model generated by the server can be improved in the end.

[0124] Reference Figure 3 In the second embodiment of the federated category incremental learning method of the present application, the federated category incremental learning method is applied to the server, including steps S60 to S80:

[0125] Step S60: sending the initialized conditional generative adversarial network model to the user end, wherein the conditional generative adversarial network model is an ACGAN network model;

[0126] In this embodiment, the conditional generative adversarial network model is an ACGAN network model. The server sends the conditional generative adversarial network model to each user terminal, so that each user terminal performs federated category incremental learning on the task data of the user terminal based on the conditional generative adversarial network model.

[0127] Specifically, the step of performing federated category incremental learning on the task data of the user terminal based on the conditional generative adversarial network model includes: receiving the conditional generative adversarial network model sent by the server, using the conditional generative adversarial network model as the local model of the user terminal, performing a first-stage training of a first preset round on the local model according to the current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task, performing balanced sampling on the current task data corresponding to the current task to obtain target task data, and then performing a second-stage training of a second preset round on the local model according to the target task data, or according to the target task data and simulated historical task data, and after completing the two-stage training, sending the local model to the server so that the server generates a target global model based on the local model.

[0128] The simulated task data generated by the generator G of the conditional generative adversarial network model is:

[0129] X fake =G(z,y|y∈Y)

[0130] Among them, z is the noise vector, y is the category label, and Y is all the categories currently learned. The discriminator D distinguishes the data category and the truth of the data as follows:

[0131] Class,Real=D(x|x∈X fake ∪X real )

[0132] Among them, the generator G needs to generate the corresponding category of simulation task dataset X according to the specified category. fake , the discriminator D is used to distinguish the real data X real With the simulation task data X fake , and predict the data category at the same time. Real is the probability that the model prediction is real data, and Class is the probability vector predicted by the model, where each component represents the probability P(y|x y ). Although the discriminator D requires two different classification layers to make the above two different predictions on the data, they both share the feature extraction layer D feat , only at the classification layer D class There is a difference.

[0133] The total loss of the discriminator D of the conditional generative adversarial network model is expressed as:

[0134]

[0135] L D =L D,adv +L D,aux

[0136] Among them, L D,adv Represents the adversarial loss of the discriminator D. The model expects the discriminator D to input X real and the simulated task data X generated by the generator G fake Correctly distinguish. D,aux L is the classification loss of the discriminator D, and it is expected that the discriminator D can correctly classify it regardless of whether it is real or simulated data. D is the total loss of the discriminator D.

[0137] The total loss of the generator G of the conditional generative adversarial network model is expressed as:

[0138]

[0139] L G =LG,adv +L G,aux

[0140] Among them, L G,adv is the adversarial loss of generator G, and the model expects the simulated task data X generated by generator G fake It can fool the discriminator D, making it think that the simulated task data is real. G,aux is the classification loss of generator G, and the model expects the simulated task data X generated by generator G fake It can also be correctly classified by the discriminator D. G is the total loss of the generator G.

[0141] Therefore, the total loss L of the conditional generative adversarial network model is W for:

[0142] L W =L D +L G

[0143] Step S70: upon receiving the local model sent by the user terminal, aggregating the local model to generate a global model;

[0144] In this embodiment, after the user terminal performs two-stage training on the local model, the local model is sent to the server. When the server receives the local model sent by the user terminal, it aggregates the local model to generate a global model.

[0145] Specifically, the server aggregates the model parameters of the local models sent by each user terminal according to a preset aggregation formula to obtain the aggregation parameters of the global model, and then updates the model parameters of the initialized conditional generative adversarial network model to the aggregation parameters of the global model, thereby generating a global model.

[0146] The preset aggregation formula is expressed as:

[0147]

[0148] Among them, W s represents the model parameters of the global model, {W u |u∈U} represents the model parameters of the local model sent by the user end, n u W u represents the contribution of each user u, n u represents the number of task data owned by user u, and |X| represents the total number of task data owned by all users {u∈U}.

[0149] Step S80: training the global model for a preset number of rounds based on the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain a target global model.

[0150] In this embodiment, the category labels and generation quantities currently learned by the local model are input into the local model to generate simulation task data corresponding to all categories currently learned by the local model, and all category labels and generation quantities learned by the historical local model of the previous task are input into the historical local model to generate simulation historical task data corresponding to all categories learned by the historical local model. Therefore, the generated simulation task data and simulation historical task data include task data of all categories that have been learned by each user terminal. The simulation task data and simulation historical task data are used as inputs of the global model for preset rounds of training to finally obtain the target global model. Please refer to Figure 4 , Figure 4 Schematic diagram of server training.

[0151] It should be noted that the historical local model of the previous task refers to the local model obtained after the last two-stage training in the federated category incremental learning of the previous task. This historical local model has the ability to recognize all categories of data learned in the historical tasks. Therefore, the category labels of all categories learned in the historical tasks and the corresponding generation quantity of each category label can be input into the historical local model of the previous task to generate simulated historical task data corresponding to all categories learned in the historical tasks.

[0152] Optionally, the number of generated data corresponding to each category label input into the historical local model can be set to be the same, so that the number of generated simulated historical task data for each category is the same, that is, class-balanced simulated historical task data is generated, thereby reducing classification discrimination caused by unbalanced data.

[0153] It should also be noted that the target global model obtained by training the simulated task data and the simulated historical task data as the input of the global model can alleviate the catastrophic forgetting caused by incremental learning, thereby improving the recognition accuracy of the target global model.

[0154] Optionally, the preset number of rounds may be set to 200.

[0155] Specifically, the simulated historical task data generated by the historical local model of the previous task Simulation task data generated by local models Respectively expressed as:

[0156]

[0157]

[0158] in, represents simulated historical task data, z represents noise vector, y represents category label, Represents all categories learned by the historical local model of the previous task, The generator of the historical local model of the previous task, u represents the user end, and U represents the set of all user ends. Represents simulation task data, represents the generator of the local model, Indicates the category currently learned by the local model.

[0159] Furthermore, the global model updates model parameters according to the calculated total loss during training.

[0160] Global Model W s Total loss L s It is expressed as:

[0161]

[0162] in, represents the discriminator loss of the global model, represents the generator loss of the global model. Represents the total loss calculated from the simulated task data of the current task generated by the local model. Represents the total loss calculated from the simulated historical task data generated by the historical local model of the previous task. s represents the weight hyperparameter between the total loss of the historical local model of the previous task and the total loss of the local model, δ s ∈[0,1].

[0163] It is understandable that the total loss L of the global model s is the total loss based on the local model itself and the total loss of the historical local model of the previous task Obtained by calculation.

[0164] Optionally, after the step of training the global model for preset rounds based on the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain the target global model, the step includes: sending the target global model to the user end.

[0165] In this embodiment, after obtaining the target global model, the server sends the target global model to the user end, so that when the user end receives the target global model sent by the server, the target global model is updated to a conditional generative adversarial network model, and then the step of using the conditional generative adversarial network model as the local model of the user end is continued, thereby continuing to perform federated category incremental learning on the local model. Then the server can continue to receive the local model of the user end, and continue to perform the step of aggregating the local model to generate a global model.

[0166] Optionally, before the step of sending the target global model to the user terminal, the step further includes: determining the number of communications between the server and the user terminal, and if the number of communications does not reach a communication number threshold, sending the target global model to the user terminal.

[0167] In this embodiment, if the number of communications reaches the communication number threshold, it indicates that the currently obtained target global model is the optimized target global model, the federated category incremental learning of the user end is completed, and the learning of all categories has been completed.

[0168] Optionally, the communication number threshold may be set to 200.

[0169] In the technical solution provided in this embodiment, the server sends the initialized conditional generative adversarial network model to the user terminal, so that the user terminal performs federated category incremental learning based on the target conditional generative adversarial network. When the local model sent by the user terminal is received, the local model is aggregated to generate a global model, and then the global model is trained for preset rounds according to the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain the target global model, thereby improving the recognition accuracy of the target global model under the premise of protecting data privacy.

[0170] Refer to Table 1, which is a comparison of the accuracy between the federated category incremental learning method of the present application and other methods of the prior art. In order to verify the technical effect of the federated category incremental learning method proposed in the present application, the image classification accuracy of the final global model was tested using the Fashion-MNIST dataset and the EMnist-Letters dataset at different imbalance ratios. Compared with FedAvg, a method widely used in federated learning, the accuracy of the federated category incremental learning method proposed in the present application is increased by 2 to 3 times. Compared with the latest federated incremental learning method FedCIL, the accuracy is increased by 5.7% to 22.38% at different imbalance ratios.

[0171] Table 1

[0172]

[0173] Reference Figure 5 , Figure 5 This is a schematic diagram of the terminal structure of the hardware operating environment involved in the embodiment of the present application.

[0174] The terminal in the embodiment of the present application can be a user terminal or a server.

[0175] like Figure 5 As shown, the terminal may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), a mouse, etc., and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0176] Those skilled in the art will understand that Figure 5 The terminal structure shown in the figure does not constitute a limitation on the terminal, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0177] like Figure 5 As shown, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a federated category incremental learning program.

[0178] When the terminal is a user terminal, Figure 5 In the terminal shown, the network interface 1004 is mainly used to connect to the backend server and perform data communication with the backend server; the processor 1001 can be used to call the federated category incremental learning program stored in the memory 1005 and perform the following operations:

[0179] Receiving the conditional generative adversarial network model sent by the server, and using the conditional generative adversarial network model as a local model of the user end;

[0180] According to current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task, performing a first stage training of a first preset round on the local model;

[0181] Performing balanced sampling on the current task data corresponding to the current task to obtain target task data;

[0182] According to the target task data, or according to the target task data and the simulated historical task data, performing a second-stage training of a second preset round on the local model;

[0183] The local model is sent to the server, so that the server generates a target global model based on the local model.

[0184] Further, the processor 1001 may call the federated category incremental learning program stored in the memory 1005, and further perform the following operations:

[0185] When receiving the target global model sent by the server, updating the target global model to the conditional generative adversarial network model;

[0186] Continue to execute the step of using the conditional generative adversarial network model as a local model of the user end.

[0187] Upon receiving the target global model sent by the server, determining the number of communications between the client and the server;

[0188] If the number of communications reaches the preset number of communications, task data of the next task is obtained, and the task data is updated to the current task data of the current task.

[0189] Further, the processor 1001 may call the federated category incremental learning program stored in the memory 1005, and further perform the following operations:

[0190] If the current task is the first task, the current task data is used as the input of the local model; if the current task is not the first task, the current task data and the simulated historical task data are used as the input of the local model, wherein the simulated historical task data is generated based on the historical local model of the previous task.

[0191] Further, the processor 1001 may call the federated category incremental learning program stored in the memory 1005, and further perform the following operations:

[0192] If the current task is the first task, determine the total loss of the local model based on the current task data;

[0193] Updating model parameters of the local model according to the total loss;

[0194] If the current task is not the first task, determining the total loss of the local model based on the current task data and the simulated historical task data;

[0195] Model parameters of the local model are updated according to the total loss.

[0196] Further, the processor 1001 may call the federated category incremental learning program stored in the memory 1005, and further perform the following operations:

[0197] If the current task is the first task, the target task data is used as the input of the local model; if the current task is not the first task, the target task data and the simulated historical task data are used as the input of the local model, wherein the simulated historical task data is generated based on the historical local model of the previous task.

[0198] Freeze model parameters of the feature extraction layer of the local model.

[0199] Further, the processor 1001 may call the federated category incremental learning program stored in the memory 1005, and further perform the following operations:

[0200] If the current task is the first task, determining the total loss of the local model based on the target task data;

[0201] Updating model parameters of the local model according to the total loss;

[0202] If the current task is not the first task, determining the total loss of the local model based on the target task data and the simulated historical task data;

[0203] Model parameters of the local model are updated according to the total loss.

[0204] When the terminal is a server, Figure 5 In the terminal shown, the network interface 1004 is mainly used to connect to the backend server and perform data communication with the backend server; the processor 1001 can be used to call the federated category incremental learning program stored in the memory 1005 and perform the following operations:

[0205] Send the initialized conditional generative adversarial network model to the user end;

[0206] When receiving the local model sent by the user terminal, aggregating the local model to generate a global model;

[0207] The global model is trained for preset rounds according to the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain a target global model.

[0208] Further, the processor 1001 may call the federated category incremental learning program stored in the memory 1005, and further perform the following operations:

[0209] The target global model is sent to the user end.

[0210] Determining the number of communications between the server and the client;

[0211] If the number of communications does not reach the communication number threshold, the target global model is sent to the user terminal.

[0212] Further, the processor 1001 may call the federated category incremental learning program stored in the memory 1005, and further perform the following operations:

[0213] If the current task is the first task, the simulated task data generated by the local model is used as the input of the global model; if the current task is not the first task, the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task are used as the input of the global model.

[0214] In addition, to achieve the above-mentioned purpose, the present application also provides a user terminal, which includes: a memory, a processor, and a federated category incremental learning program stored in the memory and executable on the processor, and the federated category incremental learning program implements the steps of the federated category incremental learning method as described above when executed by the processor.

[0215] In addition, to achieve the above-mentioned purpose, the present application also provides a server, which includes: a memory, a processor, and a federated category incremental learning program stored on the memory and executable on the processor, and the federated category incremental learning program implements the steps of the federated category incremental learning method as described above when executed by the processor.

[0216] In addition, to achieve the above-mentioned purpose, the present application also provides a computer-readable storage medium, on which a federated category incremental learning program is stored. When the federated category incremental learning program is executed by a processor, the steps of the federated category incremental learning method as described above are implemented.

[0217] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0218] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0219] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD) as described above, and includes a number of instructions for a terminal device to execute the methods described in each embodiment of the present application.

[0220] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A federated category incremental learning method, characterized in that: Applied to the user side, the federated category incremental learning method includes: Receiving the conditional generative adversarial network model sent by the server, and using the conditional generative adversarial network model as a local model of the user end; According to current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task, performing a first stage training of a first preset round on the local model; Performing balanced sampling on the current task data corresponding to the current task to obtain target task data; According to the target task data, or according to the target task data and the simulated historical task data, performing a second-stage training of a second preset round on the local model; The local model is sent to the server, so that the server generates a target global model based on the local model.

2. The method according to claim 1, characterized in that After the step of sending the local model to the server, the method further includes: When receiving the target global model sent by the server, updating the target global model to the conditional generative adversarial network model; Continue to execute the step of using the conditional generative adversarial network model as a local model of the user end.

3. The method according to claim 2, characterized in that The method further comprises: Upon receiving the target global model sent by the server, determining the number of communications between the client and the server; If the number of communications reaches the preset number of communications, task data of the next task is obtained, and the task data is updated to the current task data of the current task.

4. The method according to claim 1, characterized in that The step of performing the first stage training of the first preset round on the local model according to the current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task, includes: if the current task is the first task, using the current task data as the input of the local model; if the current task is not the first task, using the current task data and the simulated historical task data as the input of the local model, wherein the simulated historical task data is generated based on the historical local model of the previous task.

5. The method according to claim 4, characterized in that The step of performing a first-stage training of a first preset round on the local model according to the current task data corresponding to the current task, or according to the current task data and simulated historical task data corresponding to the current task, comprises: If the current task is the first task, determine the total loss of the local model based on the current task data; Updating model parameters of the local model according to the total loss; If the current task is not the first task, determining the total loss of the local model based on the current task data and the simulated historical task data; Model parameters of the local model are updated according to the total loss.

6. The method according to claim 1, characterized in that The step of performing a second-stage training of the local model for a second preset round based on the target task data and the simulated historical task data includes: if the current task is the first task, using the target task data as the input of the local model; if the current task is not the first task, using the target task data and the simulated historical task data as the input of the local model, wherein the simulated historical task data is generated based on the historical local model of the previous task.

7. The method according to claim 6, characterized in that Before the step of performing a second-stage training of a second preset round on the local model according to the target task data and the simulated historical task data, the method includes: Freeze model parameters of the feature extraction layer of the local model.

8. The method according to claim 7, characterized in that The step of performing a second-stage training of a second preset round on the local model according to the target task data and the simulated historical task data comprises: If the current task is the first task, determining the total loss of the local model based on the target task data; Updating model parameters of the local model according to the total loss; If the current task is not the first task, determining the total loss of the local model based on the target task data and the simulated historical task data; Model parameters of the local model are updated according to the total loss.

9. A federated category incremental learning method, characterized in that: Applied to the server, the federated category incremental learning method includes: Send the initialized conditional generative adversarial network model to the user end; When receiving the local model sent by the user terminal, aggregating the local model to generate a global model; The global model is trained for preset rounds according to the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain a target global model.

10. The method according to claim 9, characterized in that After the step of training the global model for a preset round according to the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain the target global model, the method includes: The target global model is sent to the user end.

11. The method according to claim 10, characterized in that Before the step of sending the target global model to the user terminal, the method further includes: Determining the number of communications between the server and the client; If the number of communications does not reach the communication number threshold, the target global model is sent to the user terminal.

12. The method according to claim 9, characterized in that The step of training the global model for a preset round according to the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task to obtain the target global model includes: If the current task is the first task, the simulated task data generated by the local model is used as the input of the global model; if the current task is not the first task, the simulated task data generated by the local model and the simulated historical task data generated by the historical local model of the previous task are used as the input of the global model.

13. A user terminal, characterized in that: The user terminal comprises: A memory, a processor, and a federated category incremental learning program stored in the memory and executable on the processor, wherein the federated category incremental learning program, when executed by the processor, implements the steps of the federated category incremental learning method as described in any one of claims 1 to 8.

14. A server, characterized in that: The server comprises: A memory, a processor, and a federated category incremental learning program stored in the memory and executable on the processor, wherein the federated category incremental learning program, when executed by the processor, implements the steps of the federated category incremental learning method as described in any one of claims 9 to 12.

15. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, on which a federated category incremental learning program is stored. When the federated category incremental learning program is executed by a processor, the steps of the federated category incremental learning method as described in any one of claims 1 to 12 are implemented.

Images