Method for learning local model of federated learning framework using classification of training data

Abstract

The present invention relates to a method for training a local model of a federated learning framework using training data classification, characterized by comprising the steps of: a client classifying training data into two classifications; a client generating a mini-batch for training, wherein the client generates the mini-batch by adjusting the ratio of samples classified into the two classifications included in the mini-batch to a set ratio; and a client performing training of a training model using the mini-batch with the adjusted ratio of samples.

Description

Technology Field

[0001] The present invention relates to federated learning, and more specifically, to a method for training a local model of federated learning. Background Technology

[0003] Federated learning is a deep learning method developed for the purpose of protecting sensitive information, including personal information. It utilizes a method of creating a global model that encompasses the whole based on the local models of each client (also referred to as participant, client, party, edge device, node, user, etc.) participating in the federated learning network.

[0004] More specifically, each client collects data to train a local model and uploads the information of the trained local model to the server. The server aggregates the local model information to update the global model and transmits the updated global model information to the clients. The clients update their local models using the global model information received from the server.

[0005] In this process, since the server and client only exchange information from the global and local models, there is no concern that data collected by the client will be transmitted externally, which offers advantages in terms of information protection.

[0006] Meanwhile, the background technology of the present invention is disclosed in Korean Patent Publication No. 10-2019-0096872 (August 20, 2019). The problem to be solved

[0008] The purpose of the present invention is to provide a local model training method for a federated learning framework using training data classification, which classifies training data to enable the learning of more data that has a greater influence on training performance, thereby improving the performance (accuracy) of federated learning. means of solving the problem

[0010] A method for training a local model of a federated learning framework using training data classification according to the present invention is characterized by comprising: a step in which a client classifies training data into two classifications; a step in which the client generates a mini-batch for training, wherein the client generates the mini-batch by adjusting the ratio of samples classified into the two classifications included in the mini-batch to a set ratio; and a step in which the client performs training of a training model using the mini-batch in which the ratio of samples has been adjusted.

[0011] In the step of classifying the training data of the present invention into two categories, the client is characterized by classifying the training data into a forgettable sample and an unforgettable sample.

[0012] In the present invention, the client is characterized by classifying the training data into the forgettable samples and the unforgettable samples based on catastrophic forgetting.

[0013] In the present invention, the client is characterized by comparing the result obtained by training the training data on the training model with the result obtained by training the training data again on the training model on which the training data was trained, thereby classifying the training data into the forgettable samples and the unforgettable samples.

[0014] In the present invention, the client is characterized by classifying samples as the forgettable samples in which the result obtained by training the learning data on the learning model and the result obtained by training the learning data again are different.

[0015] In the present invention, the client is characterized by classifying samples for which the result obtained by re-training is an incorrect answer among samples for which the result obtained by training the learning data to the learning model is a correct answer as the forgettable samples.

[0016] The present invention is characterized by further including the step of the client receiving the set ratio from the server.

[0017] The present invention is characterized by further including the step of the client receiving information about the learning model from the server before the step of classifying the learning data into two categories.

[0018] The present invention is characterized by further including, after the step of performing learning of the learning model, the step of the client transmitting information about the learning model that has been learned to the server.

[0019] In the present invention, the information regarding the learning model and the information regarding the learning model on which the learning was performed are characterized as being weights for the learning model. Effects of the invention

[0021] The local model training method of a federated learning framework using training data classification according to the present invention has the effect of improving the performance of a training model by classifying training data and adjusting the ratio of samples included in a mini-batch, thereby allowing more data that has a greater influence on training performance to be trained. Brief explanation of the drawing

[0023] FIG. 1 is an exemplary diagram illustrating a system in which federated learning is performed according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating a method for performing federated learning according to an embodiment of the present invention. FIG. 3 is a flowchart illustrating a local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention. FIG. 4 is an example diagram illustrating the operation method of a local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention. Figure 5 is an example diagram showing the results of comparing the performance of a learning model using a local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention with the performance of a learning model using general FedAvg. Figure 6 is an example diagram showing performance according to the ratio of samples in a local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention. Specific details for implementing the invention

[0024] Hereinafter, an embodiment of a local model training method of a federated learning framework using training data classification according to the present invention will be described with reference to the attached drawings. In this process, the thickness of lines or the size of components depicted in the drawings may be exaggerated for clarity and convenience of explanation. Furthermore, the terms described below are defined considering their functions in the present invention, and these may vary depending on the intention or convention of the user or operator. Therefore, the definitions of these terms should be based on the content throughout this specification.

[0025] FIG. 1 is an exemplary diagram for explaining a system in which federated learning is performed according to an embodiment of the present invention, FIG. 2 is a flowchart for explaining a method for performing federated learning according to an embodiment of the present invention, FIG. 3 is a flowchart for explaining a local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention, and FIG. 4 is an exemplary diagram for explaining the operation method of a local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention. Referring to the above, the local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention is described as follows.

[0026] As illustrated in FIG. 1, federated learning is performed in a system including a server (100) and clients (200).

[0027] The server (100) and the client (200) may be configured to include a processor for deep learning computation, a learning model, a storage device for storing learning data, a communication device for communication with each other, and a plurality of clients (200) may be configured to connect with the server (100) to perform federated learning.

[0028] For example, a user's smart device, personal computer, etc. can be used as a client (200).

[0029] As illustrated in Fig. 2, associative learning can be performed through the following process.

[0030] First, the server (100) creates an initial learning model (global model) using its own learning data and distributes it to clients (200) (S100). Subsequently, each client (200) performs learning using the learning data it possesses (S200). Once learning is performed, each client (200) possesses an individually updated learning model (local model), and each client (200) transmits information about the updated learning model to the server (100) (S300). The server (100) updates the initial learning model (global model) based on the information collected from the clients (200) (S400) and redistributes the updated learning model to the clients (200) (S500).

[0031] Federated learning is performed by repeatedly training the local model and updating the global model in this manner.

[0032] At this time, the information transmitted by each client (200) to the server (100) is the weights of the learning model, and the information transmitted when the server (100) redistributes the updated learning model to the clients (200) may also be weights. Here, weights may refer to a set of variables that can be learned by a deep learning neural network, and the server (100) can improve the global model using the collected weights.

[0033] In addition, for such federated learning, the FedAvg algorithm (HB McMahan et al., Communication-Efficient Learning of Deep Networks from Decentralized Data, 2016.), the FedMA algorithm (H. Wang et al., Federated Learning with Matched Averaging, 2019.), and the FedProx algorithm (T. Li et al, Federated Optimization in Heterogeneous Networks, 2018.) can be used, and various other known algorithms can also be applied.

[0034] However, since the operation method of federated learning using weights and the application of the aforementioned algorithm are matters widely known to those skilled in the art of the present invention, a further detailed explanation is omitted in the present invention.

[0036] Meanwhile, as can be seen in FIGS. 3 and 4, the method for training a local model of a federated learning framework using training data classification according to the present invention can be performed as follows in training a local model of a client (200).

[0037] First, the client (200) can classify the learning data by distinguishing the degree of significance of the learning.

[0038] Specifically, the client (200) can classify the training data into two categories: data that is more useful for training and general data.

[0039] For example, the client (200) can classify data that is meaningful for learning based on the catastrophic forgetting phenomenon. Catastrophic forgetting refers to the phenomenon where a learning model forgets the data it has learned. That is, when there is a model f(x) that correctly predicts data x1, there exists data that cannot be correctly predicted after learning x2 because the value of f(x1) for x1 has changed, and such data may correspond to a forgettable sample.

[0040] On the other hand, there are unforgettable samples that are not forgotten after being trained on the learning model, and the client (200) can classify the training data into forgettable samples and unforgettable samples.

[0041] The distinction between these forgettable and unforgettable samples can be made by testing whether the data is forgotten or not through training the learning model with the training data.

[0042] That is, the client (200) learns the training data and distinguishes it into forgettable samples and unforgettable samples (S210).

[0043] Specifically, the results obtained while training the training data on a training model (local model) are recorded first, and samples where the results obtained differ when the training data is trained again can be determined to be forgettable samples.

[0044] Alternatively, when training with training data, if the model is retrained on samples that get the correct answer, samples that do not get the correct answer can be determined to be forgettable samples.

[0045] In addition, when training with training data, samples that fail to guess the correct answer from the beginning can also be considered as forgettable samples.

[0046] The client (200) can be configured to separate the training data into forgettable samples and unforgettable samples and add a flag for this to the training data.

[0047] Afterwards, the client (200) organizes the training data into a mini-batch for training, and adjusts the ratio of forgettable samples and unforgettable samples included in the mini-batch to a set ratio (S220).

[0048] The aforementioned forgettable and unforgettable samples can be viewed as representing data that is difficult for a model to learn and data that is easy for a model to learn. Therefore, while both forgettable and unforgettable samples are necessary for training, forgettable samples are more important than unforgettable samples in determining the performance of a deep learning model.

[0049] Therefore, the performance of a deep learning model can be further improved by repeatedly exposing it to forgettable samples during the learning process.

[0050] Furthermore, samples that fail to identify the correct answer may possess characteristics that should be ignored, which make learning difficult. For example, in a learning model that distinguishes dog images, if the dog is captured too small or if the arm of a person holding the dog is also visible, these instances can be repeatedly exposed to the model to assign higher weight to the parts that need to be prioritized and lower the values ​​for the parts to be ignored.

[0051] However, since forgettable samples can make model training difficult, it is necessary to set this ratio to an appropriate level.

[0052] At this time, the ratio of forgettable samples and unforgettable samples can be transmitted from the server (100) to the client (200). The server (100) may be configured to select this ratio by considering the type of deep learning model, etc., or may be configured so that the user inputs the ratio to the server (100) to set it.

[0053] Meanwhile, as described above, in federated learning, when each local model is trained on the clients (200), the trained information is aggregated to the server (100), the server (100) updates the global model through this information, and redistributes it to the clients (200). Therefore, changes and losses of information occur in terms of each local model. Consequently, there is a higher possibility of forgetting previously trained samples compared to a general deep learning environment.

[0054] That is, since the likelihood of catastrophic forgetting events occurring in a federated learning environment is higher than in a general deep learning environment, the present invention allows for the performance of a learning model to be improved by adjusting the ratio of forgettable samples to unforgettable samples when configuring a mini-batch.

[0055] Subsequently, the client (200) performs training using a mini-batch with an adjusted sample ratio (S230). That is, the client (200) can perform training by configuring the mini-batch randomly, but including a certain proportion of permissible samples.

[0056] When the learning is completed, the client (200) can proceed to step S300 of FIG. 2 described above and transmit the information of the updated learning model to the server (100).

[0058] Figure 5 is an example diagram showing the results of comparing the performance of a learning model using a local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention with the performance of a learning model using general FedAvg.

[0059] In the test of Figure 5, the CIFAR-10 image dataset was used. Virtual clients were created and image data was randomly distributed, and the performance indicator shown in the figure is accuracy. Hereinafter, the method according to the present invention is referred to as B-Fed (Sample Boosted Federated Learning).

[0060] Figures 5(a) and 5(b) show the LeNet-5 model applied to 28 and 16 virtual clients, respectively. Figures 5(c) and 5(d) show the VGG-9 model applied to 16 virtual clients, but with 10 and 20 epochs, respectively. The weighted proportion of forgettable samples is 30%. It can be observed that there is a difference of 1–3% between FedAvg and B-Fed, and that B-Fed consistently outperforms FedAvg by more than 1% at any given time point.

[0061] Figure 6 is an example diagram showing performance according to the ratio of samples in a local model learning method of a federated learning framework using learning data classification according to an embodiment of the present invention.

[0062] Figure 6 shows the results when only the weighting ratio of the forgettable samples is changed under the same conditions as Figure 5. Baseline represents FedAvg, as the weighting ratio is 0%, and Boost All represents 100%. As can be seen in Figure 6, performance decreases inversely when the weighting ratio of the forgettable samples becomes excessively large. It can be observed that from 40% onwards, even the average performance and final performance fall compared to FedAvg. However, below 40%, the average, final, peak performance, and convergence point are all higher than FedAvg, showing a difference of approximately 1–3%.

[0063] Although the present invention has been described with reference to the embodiments illustrated in the drawings, this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the technical scope of protection of the present invention should be determined by the following claims. Explanation of the symbols

[0065] 100: Server 200: Client

Claims

Claim 1 A method for training a local model of a federated learning framework using training data classification, comprising: a step in which a client classifies training data into two categories, namely forgettable samples and unforgettable samples; a step in which the client generates a mini-batch for training, wherein the client generates the mini-batch by adjusting the ratio of the forgettable samples and the unforgettable samples included in the mini-batch to a set ratio; and a step in which the client performs training of a training model using the mini-batch in which the ratio of samples has been adjusted. Claim 2 delete Claim 3 A method for training a local model of a federated learning framework using training data classification, wherein, in claim 1, the client classifies the training data into the forgettable samples and the unforgettable samples based on catastrophic forgetting. Claim 4 A local model learning method of a federated learning framework using learning data classification, wherein, in claim 1, the client distinguishes the learning data into the forgettable samples and the unforgettable samples by comparing the result obtained by training the learning data on the learning model with the result obtained by training the learning data again on the learning model on which the learning data was trained. Claim 5 A local model learning method of a federated learning framework using training data classification, characterized in that, in claim 4, the client classifies samples in which the result obtained by training the training data on the training model and the result obtained by training the training data again are different as the forgettable samples. Claim 6 A local model learning method of a federated learning framework using training data classification, characterized in that, in claim 4, the client classifies samples for which the result obtained by retraining is incorrect among samples for which the result obtained by training the training data to the training model is correct as the forgettable samples. Claim 7 A method for training a local model of a federated learning framework using training data classification, characterized in that, in claim 1, the client further includes the step of receiving the set ratio from the server. Claim 8 A method for training a local model of a federated learning framework using training data classification, characterized in that, in claim 1, prior to the step of classifying the training data into two categories, the client further includes the step of receiving information about the training model from the server. Claim 9 A method for training a local model of a federated learning framework using training data classification, characterized in that, in claim 8, after the step of performing training of the training model, the client further includes the step of transmitting information about the training model to the server. Claim 10 A method for training a local model of a federated learning framework using training data classification, characterized in that, in claim 9, the information about the training model and the information about the training model on which the training was performed are weights for the training model.

Images