KR102646338B1 - System, method, computer-readable storage medium and computer program for personalized federated learning based on individual data of client - Google Patents

Abstract

The federated learning system according to an embodiment of the present invention transmits the first parameter of the extractor among the federated learning models including the extractor and the classifier to each client device, receives the learned first parameter from each client device, and receives the first parameter learned from each client device. A central server that updates federated learning models; And a plurality of people learning the first parameter of the federated learning model using data they own while maintaining the value of the second parameter of the classifier among the federated learning models and transmitting the learned first parameter to the central server. Can include client devices.

Description

Federated learning system, method, computer-readable recording medium and computer program tailored to the client's individual data

The present invention relates to a federated learning system, method, computer-readable recording medium, and computer program tailored to a client's individual data.

With recent developments in cloud and big data technology, artificial intelligence (AI) technology is being applied to various services. In order to apply such artificial intelligence technology to services, the process of learning an artificial intelligence model based on a large amount of data must be preceded.

Training artificial intelligence models requires significant computer resources to perform large-scale calculations. For example, a cloud computing service is a service that provides cloud computing infrastructure to learn artificial intelligence models without installing complex hardware and software. Because cloud computing is based on centralization of resources, all necessary data must be stored in cloud memory and used for model training. Data centralization offers many advantages in terms of maximizing efficiency, but it also carries the risk of leakage of user personal data and incurs significant network costs as data transfer is involved.

Recently, federated learning has been actively researched to overcome these problems. Federated learning is a learning method that collects models directly learned by each client device based on individual data held by multiple client devices into a central location, rather than collecting user personal data in a central location as before. Since this type of federated learning does not centrally collect user personal data, there is less risk of privacy infringement, and network costs can be reduced because only the parameters of the updated model can be transmitted.

However, since the datasets of each client device actually used in federated learning are different in number, type, distribution, etc., the imbalance of data used in the federated learning process causes forgetting (ex. catastrophic forgetting), in which the direction of learning is lost. Problems may arise.

Furthermore, although the model on which federated learning is finally completed generally shows high performance, it may show insufficient performance depending on the case when applied to an individual client device using a data set of a specific distribution.

Republic of Korea Patent Publication No. 10-2021-0132500 (published on November 4, 2021): Joint learning system and method

The problem that the present invention aims to solve is that various problems can be caused by data imbalance of each client device used in federated learning, so each client device first learns the extract part of the federated learning model and sends it to the central server. A federated learning system, method, computer-readable recording medium, and computer that allows each client device to update some parameters of the federated learning model, and secondarily, each client device individually learns the classification of the federated learning model according to its own data set. To provide a program.

However, the problem to be solved by the present invention is not limited to the above-mentioned, and includes purposes that are not mentioned but can be clearly understood by those skilled in the art from the description below. can do.

The federated learning system according to an embodiment of the present invention transmits the first parameter of the extractor among the federated learning models including the extractor and the classifier to each client device, receives the learned first parameter from each client device, and receives the first parameter learned from each client device. A central server that updates federated learning models; And a plurality of people learning the first parameter of the federated learning model using data they own while maintaining the value of the second parameter of the classifier among the federated learning models and transmitting the learned first parameter to the central server. Can include client devices.

Additionally, when the plurality of client devices receive a federated learning model on which federated learning has been completed from the central server, they can update the second parameter of the federated learning model using data they own.

Additionally, the second parameter may maintain a preset value according to a predetermined weight initialization algorithm during the learning process of the first parameter of each client device.

Additionally, the second parameter may maintain a preset value according to an orthogonal initialization algorithm during the learning process of the first parameter of each client device.

Additionally, the classification unit may include the last layer in contact with the output layer among the layers constituting the joint learning model.

Additionally, the extractor may include layers from the frontmost layer in contact with the input layer to immediately before the last layer among the layers constituting the federated learning model.

A federated learning method performed by a central server and a plurality of client devices according to an embodiment includes the steps of the central server transmitting a first parameter of the extraction unit among the federated learning models including an extraction unit and a classification unit to each client device; Each of a plurality of client devices learns the first parameter of the federated learning model using its own data while maintaining the value of the second parameter of the classifier among the federated learning models, and transmits the learned first parameter to the central server. Transferring to; And it may include the central server receiving a first parameter learned from each client device to update the joint learning model.

In addition, after the step of updating the federated learning model, each of the plurality of client devices receives a federated learning model on which federated learning has been completed from the central server, and updates the second parameter of the federated learning model using data held by the plurality of client devices. may further include.

Additionally, the step of transmitting to the central server may include controlling the second parameter to maintain a preset value according to a predetermined weight initialization algorithm during the learning process of the first parameter.

Additionally, the second parameter may include controlling the second parameter to maintain a preset value according to an orthogonal initialization algorithm during the learning process of the first parameter.

In addition, the classification unit includes the last layer in contact with the output layer among the layers constituting the federated learning model, and the extraction unit starts from the last layer in contact with the input layer among the layers constituting the federated learning model. It can include the layer immediately preceding the layer.

A computer-readable recording medium storing a computer program according to an embodiment, the method comprising: acquiring a first parameter of an extraction unit among a federated learning model including an extraction unit and a classification unit; Learning a first parameter of the joint learning model using data owned by the joint learning model while maintaining the value of the second parameter of the classifier among the joint learning models; And it may include instructions for causing the processor to perform a federated learning method of transmitting the learned first parameter to the central server.

A computer program stored in a computer-readable recording medium according to an embodiment, comprising: acquiring a first parameter of the extraction unit among a federated learning model including an extraction unit and a classification unit; Learning a first parameter of the joint learning model using data owned by the joint learning model while maintaining the value of the second parameter of the classifier among the joint learning models; And it may include instructions for causing the processor to perform a federated learning method of transmitting the learned first parameter to the central server.

According to an embodiment of the present invention, the federated learning model is divided into an extraction unit and a classification unit, and in the first learning process of learning a global model, the federated learning speed is increased by intensively learning the classification unit of the federated learning model from data held by each client device. It can be improved.

In addition, in the second learning process, in which the local model is individually trained after completing the training of the global model, the extractor is individually learned using the data held by each client device, so that each client device makes customized decisions based on the data set it holds. It has a federated learning model with boundaries, so each client device can use a model with greatly improved accuracy for the data it mainly uses.

The effects that can be obtained from the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by those skilled in the art from the description below. will be.

1 is a configuration diagram of a federated learning system according to an embodiment of the present invention.
Figure 2 is an exemplary diagram showing the structure of a federated learning model used in a federated learning system according to an embodiment of the present invention.
Figure 3 is an example diagram showing an operation in which a central server transmits a first parameter to each client device and each client device learns the first parameter according to an embodiment of the present invention.
Figure 4 shows an operation in which each client device transmits the updated first parameter to the central server and the central server collects each first parameter to update the first parameter of the federated learning model according to an embodiment of the present invention. This is an example diagram.
Figure 5 is an example diagram illustrating an operation in which a central server transmits a fully trained federated learning model to each client device and each client device learns a second parameter of the federated learning model, according to an embodiment of the present invention.
Figure 6 is an example diagram showing the appearance of a second parameter individualized according to the data set held by each client device according to an embodiment of the present invention.
Figure 7 is a flowchart of a federated learning method according to an embodiment of the present invention.
Figure 8 shows a federated learning model (FedBABU) generated by a federated learning system according to an embodiment of the present invention and a federated learning model generated by an existing federated learning algorithm measured by applying the same dataset to a plurality of client devices. This is a comparison table of accuracy.

The advantages and features of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and can be implemented in various forms. The present embodiments are merely provided to ensure that the disclosure of the present invention is complete and to those skilled in the art to which the present invention pertains. It is provided to fully inform the person of the scope of the invention, and the scope of the invention is only defined by the claims.

In describing embodiments of the present invention, detailed descriptions of well-known functions or configurations will be omitted unless actually necessary. The terms described below are terms defined in consideration of functions in the embodiments of the present invention, and may vary depending on the intention or custom of the user or operator. Therefore, the definition should be made based on the contents throughout this specification.

The functional blocks shown in the drawings and described below are only examples of possible implementations. Other functional blocks may be used in other implementations without departing from the spirit and scope of the detailed description. Additionally, although one or more functional blocks of the present invention are shown as individual blocks, one or more of the functional blocks of the present invention may be a combination of various hardware and software configurations that perform the same function.

Additionally, the expression that it includes certain components is an open expression and simply refers to the existence of the components, and should not be understood as excluding additional components.

Furthermore, when it is mentioned that a component is connected or connected to another component, it should be understood that although it may be directly connected or connected to the other component, other components may exist in between.

In addition, expressions such as 'first, second', etc. are expressions used only to distinguish a plurality of configurations, and do not limit the order or other characteristics between the configurations.

Hereinafter used ‘…’ wealth', '… Terms such as 'unit' refer to a unit that processes at least one function or operation, and may be implemented as hardware, software, or a combination of hardware and software.

Figure 1 is a configuration diagram of a federated learning system 10 according to an embodiment of the present invention.

Referring to FIG. 1, the federated learning system 10 according to one embodiment may include a central server 100 and a plurality of client devices 200.

The central server 100 and the client device 200 are computing devices that include memory and a processor, and overall operations can be performed by instructions stored in the memory and operations of the processor.

The central server 100 and the client device 200 may store an artificial intelligence neural network model designed with the same structure to perform federated learning.

Hereinafter, the artificial intelligence neural network model used for federated learning according to the embodiments of this document will be referred to as the 'federated learning model'. In addition, if explanation is needed to distinguish the device where the 'federated learning model' is stored, the federated learning model stored in the central server 100 is referred to as a 'global model' and the federated learning model stored in the client device 200 is referred to as a 'local model'. We will refer to them separately.

The general operation of the central server 100 and the client device 200 constituting the federated learning system 10 to train a federated learning model is as follows.

First, the central server 100 may transmit parameter values set in the global model to each client device 200.

Next, each client device 200 can train a local model using its own data and transmit the parameters of the learned local model to the central server 100.

Thereafter, the central server 100 may update the parameters of the global model by collecting the parameters of the local model learned by each client device 200.

In this way, a series of processes in which the central server 100 transmits parameters to the client device 200, collects newly learned parameters, and then updates the model can be understood as one round of federated learning. A round of federated learning may be conducted in multiple rounds depending on the design, and after the final round is conducted, the parameters of the updated global model may be determined as the parameters of the final federated learning model.

At this time, the central server 100 may select some client devices 200 among the plurality of client devices 200 and transmit parameters for each round of federated learning according to a predetermined method (ex. FedAvg, FedSGD, FedMA, etc.). .

At this time, the central server 100 may update the parameters of the global model by combining parameters collected from the client device 200 according to a predetermined method (ex. FedAvg, FedSGD, FedMA, etc.).

Meanwhile, in federated learning, if the number, type, distribution, etc. of data held by each client device 200 are different from each other, forgetting (ex. catastrophic forgetting) problems may occur, and individual clients using data sets of a specific distribution may occur. When applying a federated learning model to the device 200, there are cases where insufficient performance is shown. To solve this problem, the federated learning system 10 according to the embodiment of this document uses a method of learning by dividing parameters to be commonly used by each client device 200 and parameters to be used individually among the parameters of the federated learning model. present.

Figure 2 is an exemplary diagram showing the structure of a federated learning model used in the federated learning system 10 according to an embodiment of the present invention.

Referring to FIG. 2, a federated learning model according to one embodiment may include a neural network consisting of an input layer, a hidden layer, and an output layer. The federated learning system 10 according to one embodiment further divides the hidden layer of the federated learning model into an extractor and a classifier.

The extraction unit may include layers from the frontmost layer that contacts the input layer among the layers that make up the federated learning model to the layer immediately in front of the layer corresponding to the lastest layer of the hidden layer. For example, the extractor may include a network layer including parameters for performing a convolution operation by applying weights and biases to predetermined feature values or feature vectors. Hereinafter, the parameters learned in the extraction unit are collectively referred to as 'first parameters'.

The classification unit may include the last layer in contact with the output layer among the layers constituting the federated learning model. For example, the classification unit may include a network layer that includes parameters to distinguish a decision boundary for distinguishing classes for the output layer. Hereinafter, the parameters learned in the classification unit are collectively referred to as 'second parameters'.

Figure 3 is an example diagram showing an operation in which the central server 100 transmits a first parameter to each client device 200 and each client device 200 learns the first parameter according to an embodiment of the present invention. .

The central server 100 may transmit the first parameter set in the global model to each client device 200. At this time, the central server 100 selects some client devices 200 among all client devices 200 for each round of federated learning according to a predetermined method (ex. FedAvg, FedSGD, FedMA, etc.), transmits and updates parameters. You can do it.

Each client device (200-1, 200-2, ..., 200-n) sets the first parameter value transmitted by the central server 100 to the extractor of the local model stored respectively as the initial value for learning. And, the first parameter can be individually learned using individually held data (D1, D2, D3) using a predefined learning algorithm (ex. CNN, RNN, MLP, etc.). At this time, each client device (200-1, 200-2, ..., 200-n) can learn only the first parameter by ensuring that the second parameter value set in the classification unit of the local model is not updated and maintains the same value. there is.

As an example, the central server 100 and all client devices 200 are set to have the same second parameter in the federated learning model they each store, and the value may be stipulated to remain the same without being updated.

As an example, the second parameter value may be stipulated to have a preset value according to a predetermined weight initialization algorithm (ex. orthogonal initialization). For example, the central server 100 determines a second parameter value to be applied to the classifier according to a weight initialization algorithm and spreads the second parameter value to all client devices 200 before a round of federated learning begins, During the learning process of the first parameter, it can be stipulated that the value of the second parameter remains unchanged during the learning process.

Figure 4 shows that, according to an embodiment of the present invention, each client device 200 transmits the updated first parameter to the central server 100, and the central server 100 collects each first parameter to form a federated learning model. This is an example diagram showing the operation of updating the first parameter.

Referring to FIG. 4, each client device 200 may generate a newly learned first parameter value using individual data D1, D2, and D3 held by each client device 200. Each client device 200 may transmit the updated first parameter value to the central server 100. At this time, the central server 100 may update the first parameter value of the global model by combining the first parameter values collected from the client device 200 according to a predetermined method (e.g. FedAvg, FedSGD, FedMA, etc.). .

According to the embodiment of this document, the process of FIGS. 3 and 4 described above can be understood as one of the learning rounds of a global model in which the central server 100 and the client device 200 participate together. The global model learning round of FIGS. 3 and 4 may be performed a preset number of rounds depending on the designer's selection.

Figure 5 shows that according to an embodiment of the present invention, the central server 100 transmits a fully trained federated learning model to each client device 200, and each client device 200 learns the second parameter of the federated learning model. This is an example showing the operation.

Referring to FIG. 5, the central server 100 may transmit a federated learning model (ex. final first parameter value) on which global learning has been completed according to the process of FIGS. 3 and 4 to each client device 200.

Each client device (200-1, 200-2, ..., 200-n) sets the final first parameter value transmitted by the central server 100 to the extractor of the local model each stored, and individually The second parameter can be individually learned using the existing data (D1, D2, D3) using a pre-defined learning algorithm (ex. CNN, RNN, MLP, etc.). At this time, each client device (200-1, 200-2, ..., 200-n) controls the first parameter value set in the classification unit of the local model not to be updated and maintains the same value to learn only the second parameter. You can.

Figure 6 is an example diagram showing the appearance of a second parameter individualized according to the data set held by each client device 200 according to an embodiment of the present invention.

Referring to Figure 6, each client device (200-1, 200-2, ..., 200-n) individually learned the second parameter using the data (D1, D2, D3) stored respectively. Therefore, they may have different second parameter values. That is, each client device (200-1, 200-2, ..., 200-n) has an individually customized decision boundary classification unit suitable for classifying data mainly used by the user of the client device (200). It can be learned to have.

As such, the embodiment of this document primarily allows each client device 200 to learn the extractor of the federated learning model so that the central server 100 updates the extractor of the federated learning model, and secondarily, each client device 200 ), when using a federated learning model, individually learns the classification part of the federated learning model according to the data set it owns, allowing it to use a federated learning model customized to the individual data distribution for each client device 200.

Figure 7 is a flowchart of a federated learning method according to an embodiment of the present invention.

Each step of the federated learning method according to FIG. 7 can be performed by the central server 100 and the client device 200 of the federated learning system 10 described with reference to FIGS. 1 to 6, and each step is described as As follows.

In step S1010, the central server 100 may transmit the first parameter of the extractor of the joint learning model to each client device 200.

In step S1020, each client device 200 learns the first parameter of the federated learning model using its own data while maintaining the value of the second parameter of the classifier in the federated learning model, and places the learned first parameter in the center. It can be transmitted to the server 100.

In step S1030, the central server 100 may receive the learned first parameter from each client device 200 and update the joint learning model.

In step S1040, each of the plurality of client devices 200 may receive a federated learning model on which federated learning has been completed from the central server 100 and update the second parameter of the federated learning model using data it possesses.

Meanwhile, in addition to the steps shown in FIG. 7, as various embodiments are configured to perform the contents described with FIGS. 1 to 6, a new step for performing applicable operations in addition to the steps shown in FIG. 7 can be added. Meanwhile, the configuration of additional steps and the operations of the components that are the subject of each step to perform the corresponding steps have been described in FIGS. 1 to 6, so duplicate descriptions will be omitted.

Figure 8 shows a federated learning model (FedBABU) generated by the federated learning system 10 according to an embodiment of the present invention and a federated learning model generated by an existing federated learning algorithm in the same manner in a plurality of client devices 200. This is a comparison table of accuracy measured by applying the dataset.

Referring to FIG. 8, s is a variable that sets the degree of imbalance of data held by each client device 200, and f and τ are variables that set the learning environment for federated learning. It can be seen that the accuracy of the embodiment of the present invention (FedBABU) in various environments set according to the variables shown in FIG. 8 is greatly improved compared to the accuracy of the federated learning model generated by the existing algorithm.

According to the above-described embodiment, the federated learning model is divided into an extraction unit and a classification unit, and in the first learning process of learning a global model, federated learning is performed by intensively learning the classification unit of the federated learning model from data held by each client device 200. Speed can be improved. In addition, in the secondary learning process of individually learning the local model after completing training of the global model, the extractor is individually learned using the data held by each client device 200, so that each client device 200 learns the data held by it. It has a federated learning model with a decision boundary customized by the set, and thus each client device 200 can use a model with greatly improved accuracy for the data it mainly uses.

Embodiments of the present invention described above can be implemented through various means. For example, embodiments of the present invention may be implemented by hardware, firmware, software, or a combination thereof.

In the case of hardware implementation, the method according to embodiments of the present invention uses one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), and Programmable Logic Devices (PLDs). , can be implemented by FPGAs (Field Programmable Gate Arrays), processors, controllers, microcontrollers, microprocessors, etc.

In the case of implementation by firmware or software, the method according to embodiments of the present invention may be implemented in the form of a module, procedure, or function that performs the function or operation described above. A computer program in which software codes, etc. are recorded may be stored in a computer-readable recording medium or memory unit and driven by a processor. The memory unit is located inside or outside the processor and can exchange data with the processor through various known means.

Additionally, combinations of each block of the block diagram and each step of the flow diagram attached to the present invention may be performed by computer program instructions. Since these computer program instructions can be mounted on the encoding processor of a general-purpose computer, special-purpose computer, or other programmable data processing equipment, the instructions performed through the encoding processor of the computer or other programmable data processing equipment are included in each block or block of the block diagram. Each step of the flowchart creates a means to perform the functions described. These computer program instructions may also be stored in computer-usable or computer-readable memory that can be directed to a computer or other programmable data processing equipment to implement a function in a particular way, so that the computer-usable or computer-readable memory The instructions stored in can also produce manufactured items containing instruction means that perform the functions described in each block of the block diagram or each step of the flow diagram. Computer program instructions can also be mounted on a computer or other programmable data processing equipment, so that a series of operational steps are performed on the computer or other programmable data processing equipment to create a process that is executed by the computer, thereby generating a process that is executed by the computer or other programmable data processing equipment. Instructions that perform processing equipment may also provide steps for executing functions described in each block of the block diagram and each step of the flow diagram.

In addition, each block or each step may represent a module, segment, or portion of code that includes one or more executable instructions for executing a specified logical function. Additionally, it should be noted that in some alternative embodiments it is possible for the functions mentioned in blocks or steps to occur out of order. For example, two blocks or steps shown in succession may in fact be performed substantially simultaneously, or the blocks or steps may sometimes be performed in reverse order depending on the corresponding function.

As such, a person skilled in the art to which the present invention pertains will understand that the present invention can be implemented in other specific forms without changing its technical idea or essential features. Therefore, the embodiments described above should be understood in all respects as illustrative and not restrictive. The scope of the present invention is indicated by the claims described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. .

10: Federated learning system
100: Central server
200: Client device

Claims (12)

A central server that transmits the first parameter of the extraction unit among the joint learning models including an extraction unit and a classification unit to each client device, and updates the joint learning model by receiving the first parameter learned from each client device; and
Maintaining the value of the second parameter of the classifier among the federated learning models, the first parameter of the federated learning model is learned using the data owned by the joint learning model, and the learned first parameter is transmitted to the central server, and the comprising a plurality of client devices that update a second parameter of the federated learning model using the data it has,
Federated learning system. According to paragraph 1,
The plurality of client devices are:
Upon receiving a federated learning model for which federated learning has been completed from the central server, updating the second parameter of the received federated learning model,
Federated learning system. According to paragraph 1,
The second parameter is,
During the learning process of the first parameter of each client device, a preset value is maintained according to a predetermined weight initialization algorithm,
Federated learning system. According to paragraph 3,
The second parameter is,
During the learning process of the first parameter of each client device, the preset value is maintained according to an orthogonal initialization algorithm,
Federated learning system. According to paragraph 1,
The classification department,
Among the layers constituting the federated learning model, it includes the last layer in contact with the output layer,
The extraction unit,
Among the layers constituting the federated learning model, including the layer from the frontmost layer in contact with the input layer to the layer immediately before the last layer,
Federated learning system. In a federated learning method performed by a central server and a plurality of client devices,
A central server transmitting the first parameter of the extraction unit among the joint learning models including the extraction unit and the classification unit to each client device;
Each of a plurality of client devices learns the first parameter of the federated learning model using its own data while maintaining the value of the second parameter of the classifier among the federated learning models, and transmits the learned first parameter to the central server. Transferring to;
A central server receiving learned first parameters from each client device to update the joint learning model; and
Each of the plurality of client devices updating a second parameter of the federated learning model using data it holds;
Including,
Federated learning methods. According to clause 6,
The step of updating the second parameter is,
After updating the federated learning model,
Further comprising the step of each of a plurality of client devices receiving a federated learning model on which federated learning has been completed from the central server, and updating a second parameter of the received federated learning model,
Federated learning methods. According to clause 6,
The step of transmitting to the central server is,
Including controlling the second parameter to maintain a preset value according to a predetermined weight initialization algorithm during the learning process of the first parameter,
Federated learning methods. According to clause 8,
The second parameter is,
In the process of learning the first parameter, controlling the second parameter to maintain a preset value according to an orthogonal initialization algorithm,
Federated learning methods. According to clause 6,
The classification department,
Among the layers constituting the federated learning model, it includes the last layer in contact with the output layer,
The extraction unit,
Among the layers constituting the federated learning model, including the layer from the frontmost layer in contact with the input layer to the layer immediately before the last layer,
Federated learning methods. A computer-readable recording medium storing a computer program,
Obtaining a first parameter of the extraction unit among the joint learning model including the extraction unit and the classification unit;
Learning a first parameter of the joint learning model using data owned by the joint learning model while maintaining the value of the second parameter of the classifier among the joint learning models;
transmitting the learned first parameter to a central server; and
Containing instructions for causing the processor to perform a federated learning method of updating a second parameter of the federated learning model using data held by the processor,
A computer-readable recording medium. A computer program stored on a computer-readable recording medium,
Obtaining a first parameter of the extraction unit among the joint learning model including the extraction unit and the classification unit;
learning a first parameter of the joint learning model using data owned by the joint learning model while maintaining the value of the second parameter of the classifier among the joint learning models;
transmitting the learned first parameter to a central server; and
Containing instructions for causing the processor to perform a federated learning method of updating a second parameter of the federated learning model using data held by the processor,
computer program.

Images