KR102390553B1 - Federated learning method and system - Google Patents

Abstract

Disclosed are a method and system for associated learning. The method for associated learning according to one embodiment comprises: a step of training a machine learning model; a step of compressing a first model parameter of the machine learning model; a step of transmitting the compressed first model parameter to a parameter server; a step of receiving a generated general model parameter based on the compressed model parameter in the parameter server; a step of obtaining a second model parameter by restoring the general model parameter; and a step of initializing the machine learning model based on the second model parameter. Therefore, the present invention is capable of providing a technology that protects personal information.

Description

FEDERATED LEARNING METHOD AND SYSTEM

The disclosure below relates to a federated learning method and system.

Cloud computing and edge computing are being used to satisfy a high amount of computation due to the explosive increase in smart devices that generate a large amount of data, such as mobile phones, wearable devices, and autonomous vehicles.

Cloud computing and edge computing methods, which are conventional technologies, raise concerns about personal information exposure caused by sending local data of a client to another client.

In the conventional cloud computing method, when transmitting data from a smart device as a parameter, a network communication bottleneck occurs due to a large amount of data.

In the conventional edge computing research, there is an edge computing method in which a computing source is shared by connecting an edge device capable of computing in real time according to the location of a client. However, this method also has a disadvantage in that the client's machine learning performance is inferior to that of cloud computing.

The background art described above is possessed or acquired by the inventor in the process of deriving the disclosure of the present application, and cannot necessarily be said to be a known technology disclosed to the general public prior to the present application.

The following embodiments may provide a technology capable of alleviating a network communication bottleneck and protecting personal information in federated learning.

However, the technical tasks are not limited to the above-described technical tasks, and other technical tasks may exist.

A learning method according to an embodiment includes the steps of: learning a machine learning model; compressing a first model parameter of the machine learning model; and transmitting the compressed first model parameter to a parameter server; Receiving a general model parameter generated based on the compressed model parameter from a parameter server, restoring the general model parameter to obtain a second model parameter, and the machine learning model based on the second model parameter Including the step of initializing.

The compressing may include compressing the first model parameter through a neural network.

The general model parameter may be generated based on an average for each component of a plurality of compressed model parameters received by the parameter server.

The average for each component may be an average of non-zero components among components of the same index.

The acquiring may include acquiring the second model parameter through a neural network.

The first model parameter and the second model parameter may be a one-dimensional vector.

The learning method according to an embodiment includes the steps of: receiving compressed model parameters from a plurality of clients; generating a general model parameter based on an average of each component of the compressed model parameters; and transmitting to a plurality of clients.

The average for each component may be an average of non-zero components among components of the same index.

The generating may include selecting some of the compressed model parameters and generating the general model parameters based on the some.

An electronic device for performing federated learning according to an embodiment includes a memory including instructions and a processor for executing the instructions, and when the instructions are executed by the processor, the processor generates a machine learning model learning, learning the machine learning model, compressing the first model parameters of the machine learning model, sending the compressed first model parameters to the parameter server, and generating based on the compressed model parameters in the parameter server Receive a general model parameter, restore the general model parameter to obtain a second model parameter, and initialize the machine learning model based on the second model parameter.

In the compressing, the processor may compress the first model parameter through a neural network.

The general model parameter may be generated based on an average for each component of a plurality of compressed model parameters received by the parameter server.

The average for each component may be an average of non-zero components among components of the same index.

The processor may acquire the second model parameter through a neural network.

The first model parameter and the second model parameter may be a one-dimensional vector.

A server for performing federated learning according to an embodiment includes a memory including instructions and a processor for executing the instructions, and when the instructions are executed by the processor, the processor is compressed from a plurality of clients Received model parameters, generate a general model parameter based on the component-by-component average of the compressed model parameters, and transmit the general model parameter to the plurality of clients.

The average for each component may be an average of non-zero components among components of the same index.

The processor may select some of the compressed model parameters and generate the general model parameters based on the portions.

1 is a diagram for conceptually explaining a federated learning method according to an embodiment.
2 is a diagram for explaining in detail a method of associative learning.
3A is a diagram for explaining an update operation of the federated learning method, and FIG. 3B is a flowchart of the update operation.
4A is a diagram for explaining an execution operation of the federated learning method, and FIG. 4B is a flowchart of the execution operation.
5A is a diagram for explaining a download operation of the federated learning method, and FIG. 5B is a flowchart of the download operation.
6 shows a federated learning system according to an embodiment.

Specific structural or functional descriptions of the embodiments are disclosed for purposes of illustration only, and may be changed and implemented in various forms. Accordingly, the actual implementation form is not limited to the specific embodiments disclosed, and the scope of the present specification includes changes, equivalents, or substitutes included in the technical spirit described in the embodiments.

Although terms such as first or second may be used to describe various elements, these terms should be interpreted only for the purpose of distinguishing one element from another. For example, a first component may be termed a second component, and similarly, a second component may also be termed a first component.

When a component is referred to as being “connected to” another component, it may be directly connected or connected to the other component, but it should be understood that another component may exist in between.

The singular expression includes the plural expression unless the context clearly dictates otherwise. In this specification, terms such as "comprise" or "have" are intended to designate that the described feature, number, step, operation, component, part, or combination thereof exists, and includes one or more other features or numbers, It should be understood that the possibility of the presence or addition of steps, operations, components, parts or combinations thereof is not precluded in advance.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present specification. does not

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. In the description with reference to the accompanying drawings, the same components are assigned the same reference numerals regardless of the reference numerals, and overlapping descriptions thereof will be omitted.

1 is a diagram for conceptually explaining a federated learning method according to an embodiment.

The federated learning method 10 is an optimized general model ( General Model) may be a learning method for downloading to the client 150 .

The federated learning method 10 may be applied to a communication environment including a plurality of devices. In the federated learning method 10 , the parameter server 100 acquires a machine learning model from the client 150 , and continuously distributes a general model generated based on the acquired machine learning model to the client 150 . ), while maximizing the performance of the machine learning model, it is possible to prevent invasion of privacy of the client 150 .

A plurality of clients 150 train a machine learning model by a local epoch, and after collecting each learned machine learning model in the parameter server 100, a global general learning model is generated and a plurality of clients 150 An operation transmitted to the epochs may be performed as many as epochs (global epoch).

In the federated learning method 10, the local data of the client 150 is not exchanged and/or transmitted, and the model parameter of the machine learning model of the client 150 is transmitted to the parameter server 100, thereby leaking personal information. can prevent

The federated learning method 10 may alleviate the network communication bottleneck by compressing and updating the model parameters of the machine learning model of the client 150 to the model.

The parameter server 100 may generate an optimized general model through element-wise averaging based on the transmitted compressed model parameters, and the client 150 may download the optimized general model. can

2 is a diagram for explaining in detail a method of associative learning.

The client 150 may train the machine learning model using local data (eg, image data). For example, the client 150 may learn the machine learning model by local epochs. The model parameters of the machine learning model for which training is completed may be compressed. Compressed model parameters may be transmitted to the parameter server 100 . Local learning, compression, and transmission operations of the plurality of clients 150 may be update operations.

The parameter server 100 may generate a general model based on the compressed model parameters received from the client 150 . The model parameters of the general model are referred to as general model parameters. The parameter server 100 may generate general model parameters based on a combined model that is a combination of compressed model parameters received from some randomly selected clients 150 among the plurality of clients 150 . The parameter server 100 may generate general model parameters through element-wise averaging. The general model creation operation of the parameter server 100 may be an execute operation.

The client 150 may receive the general model parameters from the parameter server 100 . The client 150 may obtain the model parameters of the machine learning model by restoring the general model parameters. That is, the client 150 may initialize the machine learning model by restoring the optimized general model parameters to be suitable for the machine learning model. The general model parameter reception, machine learning model parameter restoration, and machine learning model initialization operations of the client 150 may be download operations.

3A is a diagram for explaining an update operation of the federated learning method, and FIG. 3B is a flowchart of the update operation.

The client 150 may train a machine learning model based on local data ( 310 ). The client 150 may check whether learning has progressed as much as the local epoch ( 320 ), and may compress the model parameters of the machine learning model on which the learning is completed ( 330 ).

은 클라이언트(150)의 모델 파라미터를 나타내고,

는 압축된 모델 파라미터를 나타낸다.The compression of the model parameters can be expressed as Equation (1). In Equation 1, i represents a specific global epoch, k represents a specific client 150,

represents the model parameters of the client 150,

represents the compressed model parameters.

The model parameter may be a vector of a one-dimensional structure. The client 150 may compress the model parameters by using a neural network capable of reflecting a one-dimensional vector structure as an encoder in order to compress the model parameters while maintaining information. The client 150 may compress model parameters using a convolutional neural network (CNN) as an encoder. For example, the CNN may be Imagenet, CIFAR-10, and CIFAR-100, and two one-dimensional convolutional layers (1D convolutional layer (Conv1D)) and Maxpooling layers pretrained with a public dataset. layer) may be included.

)는 추출된 특징들에 기초하여 맥스풀링 레이어를 통해 생성될 수 있다. 맥스 풀링 레이어는 마스크로 가려진 공간에서 가장 큰 값을 추출함으로써, 복잡도(complexity)를 낮추며 병진 불변성(translational invariance)을 보장할 수 있다.The one-dimensional convolutional layer may extract features of model parameters through a convolutional mask and shared weights. By superimposing one-dimensional convolutional layers, more complex features can be extracted. compressed model parameters (

) may be generated through the maxpooling layer based on the extracted features. The max pooling layer can reduce complexity and guarantee translational invariance by extracting the largest value from a space covered by a mask.

A model parameter may be input to the first 1D convolutional layer, and the 1D convolutional layer may extract low level features using a 1*5 convolutional mask. For example, low-level features may include correlations and associations between two layers of a machine learning model.

The second one-dimensional convolutional layer can extract high level features using a 1*5 convolution mask. For example, high-level features are features that can be inferred based on low-level features, and may include a spatial distribution pattern of model parameters extracted from input data.

The last layer may perform 1*2 max pooling. By leaving one maximum value in the 1*n mask, max bulling reduces the size of data to reduce complexity, and allows linearly slightly shifted data to be recognized as the same data to ensure translational invariance.

The client 150 may transmit the compressed model parameters to the parameter server 100 ( 340 ).

4A is a diagram for explaining an execution operation of the federated learning method, and FIG. 4B is a flowchart of the execution operation.

The parameter server 100 may generate an optimized general model based on the compressed model parameters. The parameter server 100 may randomly select some clients 150 among the clients 150 and generate a general model based on the compressed model parameters received from the selected clients 150 .

는 1*n 벡터인 일반 모델 파라미터이고,

는 일반 모델 파라미터의 m번째 성분이고,

는 k번째 클라이언트(150)로부터 수신한 압축된 모델 파라미터의 m 번째 성분이고,

는 선택된 모든 압축된 모델 파라미터의 m번째 인덱스의 성분 중 0이 아닌 성분의 개수이다. 수학식 2를 풀어서 정리하면 수학식 3과 같을 수 있다.The parameter server 100 may generate a general model through element-wise averaging ( 450 ). The general model parameter of the general model may be expressed as Equation (2). in Equation 2

is the general model parameter which is a 1*n vector,

is the mth component of the general model parameter,

is the m-th component of the compressed model parameter received from the k-th client 150,

is the number of non-zero components among the components of the m-th index of all selected compressed model parameters. By solving Equation 2 and rearranging it, Equation 3 may be the same.

The parameter server 100 divides the compressed model parameters by a non-zero number for each component through element-wise averaging, effectively excluding components without features extracted from the input data, while optimizing for each component. A general model can be derived.

5A is a diagram for explaining a download operation of the federated learning method, and FIG. 5B is a flowchart of the download operation.

는 클라이언트(150)의 기계 학습 모델에 맞게 복원된 모델 파라미터일 수 있다.The client 150 may obtain the model parameters of the machine learning model by restoring the general model parameters received from the parameter server 100 ( 530 ). The restoration of the model parameters can be expressed as in Equation (4). in Equation 4

may be a model parameter restored to fit the machine learning model of the client 150 .

In order to effectively restore the model parameters of the machine learning model based on the optimized general model transmitted from the parameter server 100, the client 150 is a one-dimensional vector of the general model parameter, which is a transpose convolution that can reflect the structure of the parameter. Model parameters may be restored by using a Transposed Convolutional Neural Network (T-CNN) as a decoder. The decoder can restore the model parameters by minimizing noise while maintaining important information of the general model parameters. For example, the transposed convolutional network may be Imagenet, CIFAR-10, and CIFAR-10, and two transposed 1D convolutional layers (Transposed Conv1D) pretrained with a public dataset. ) and an unpooling layer.

The unpooling layer can expand the vector without noise while maintaining the component values of the general model parameters. The vector extended through the unpooling layer may be further extended based on key features of the vector extended through a convolutional mask and shared weights of the transposing one-dimensional convolutional layer.

By overlapping transposed one-dimensional convolutional layers (Transposed Conv1D), a vector can be extended with more sophisticated features. The unpooling layer can guarantee translational invariance without maximizing complexity by adding 0 while maintaining the value as a mask.

The unpooling layer may receive a general model parameter received from the parameter server 100 as an input and perform 1*2 unpooling. Unpooling is the operation of adding one maximum value and n-1 zeros within a 1*n mask. Even if the data size is increased, the complexity does not increase significantly, and linearly slightly shifted data can be recognized as the same data. Thus, translational invariance can be guaranteed.

The first transposed one-dimensional convolutional layer (Transposed Conv1D) may receive the output of the unpooling layer as an input, and may expand the vector to a sparse feature through a 1*5 convolutional mask. For example, sparse features may include interrelationships and associations between two layers of a machine learning model.

The second transposed one-dimensional convolutional layer (Transposed Conv1D) may extract general level features through a 1*5 convolution mask. The general level features are features that can be inferred from the sparse features, and the spatial distribution of model parameters extracted from data may include a pattern.

The client 150 may initialize the machine learning models with the model parameters output from the decoder ( 520 ).

6 shows a federated learning system according to an embodiment.

The federated learning system may perform the federated learning method 10 . The federated learning system may include a parameter server 100 and a client 150 . Although only one client 150 is illustrated in FIG. 6 for convenience of description, the federated learning system may include a plurality of clients 150 . The client 150 may be an electronic device such as a smart phone, a wearable device, or an autonomous vehicle.

The parameter server 100 and the client 150 may perform data communication through a network. The network may be a local area network (LAN) and/or a wide area network (WAN, for example, the Internet) accessible through conventional wired and/or wireless communication technologies.

The parameter server 100 may include a processor 650 and a memory 670 , and the client 150 may include a processor 610 and a memory 630 .

The memories 630 and 670 may store instructions (or programs) executable by the processor. For example, the instructions may include instructions for executing an operation of a processor and/or an operation of each component of the processor.

The memories 630 and 670 may be implemented as a volatile memory device or a nonvolatile memory device.

The volatile memory device may be implemented as dynamic random access memory (DRAM), static random access memory (SRAM), thyristor RAM (T-RAM), zero capacitor RAM (Z-RAM), or twin transistor RAM (TTRAM).

Nonvolatile memory devices include EEPROM (Electrically Erasable Programmable Read-Only Memory), Flash memory, MRAM (Magnetic RAM), Spin-Transfer Torque (STT)-MRAM (Spin-Transfer Torque (STT)-MRAM), Conductive Bridging RAM (CBRAM) , FeRAM(Ferroelectric RAM), PRAM(Phase change RAM), Resistive RAM(RRAM), Nanotube RRAM(Nanotube RRAM), Polymer RAM(Polymer RAM(PoRAM)), Nano Floating Gate Memory (NFGM)), a holographic memory, a molecular electronic memory device, or an Insulator Resistance Change Memory.

The memories 630 and 670 may store a matrix on which an operation included in the neural network is to be performed. The memories 630 and 670 may store operation results generated by processing by the processors 610 and 650 .

The processors 610 and 650 may process data stored in the memories 630 and 670 . Processors 610 and 650 may execute computer readable code (eg, software) stored in memories 630 and 670 and instructions issued by processors 610 and 650 .

The processors 610 and 650 may be hardware-implemented data processing devices having circuits having a physical structure for executing desired operations. For example, desired operations may include code or instructions included in a program.

For example, a data processing unit implemented in hardware may include a central processing unit, a graphics processing unit, a neural network processing unit, a multi-core processor, It may include a multiprocessor, an application-specific integrated circuit (ASIC), and a field programmable gate array (FPGA).

It should be understood that the federated learning method performed by the parameter server 100 and the client 150 is performed through each of the processors 610 and 650, and since the federated learning method has been described in detail with reference to FIGS. 1 to 5B, overlapping The description will be omitted.

The embodiments described above may be implemented by a hardware component, a software component, and/or a combination of the hardware component and the software component. For example, the apparatus, methods, and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate (FPGA) array), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions, may be implemented using a general purpose computer or special purpose computer. The processing device may execute an operating system (OS) and a software application running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For convenience of understanding, although one processing device is sometimes described as being used, one of ordinary skill in the art will recognize that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that can include For example, the processing device may include a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as parallel processors.

Software may comprise a computer program, code, instructions, or a combination of one or more thereof, which configures a processing device to operate as desired or is independently or collectively processed You can command the device. The software and/or data may be any kind of machine, component, physical device, virtual equipment, computer storage medium or apparatus, to be interpreted by or to provide instructions or data to the processing device. , or may be permanently or temporarily embody in a transmitted signal wave. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored in a computer-readable recording medium.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination, and the program instructions recorded on the medium are specially designed and configured for the embodiment, or may be known and available to those skilled in the art of computer software. may be Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

The hardware devices described above may be configured to operate as one or a plurality of software modules to perform the operations of the embodiments, and vice versa.

As described above, although the embodiments have been described with reference to the limited drawings, those of ordinary skill in the art may apply various technical modifications and variations based thereon. For example, the described techniques are performed in an order different from the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (19)

training a machine learning model;
By inputting the first model parameter of the machine learning model into a plurality of one-dimensional convolutional layers included in the machine learning model, a 1*5 convolutional mask of the plurality of one-dimensional convolutional layers and a shared weight ( extracting a feature of the first model parameter using shared weights;
generating a compressed first model parameter using a max pooling layer based on a characteristic of the first model parameter;
transmitting the compressed first model parameter to a parameter server;
receiving a general model parameter generated based on the compressed first model parameter in the parameter server;
obtaining a second model parameter by restoring the general model parameter;
initializing the machine learning model based on the second model parameter
including,
The feature is
low-level features including correlations and associations between a plurality of one-dimensional convolutional layers included in the machine learning model; and
A high-level feature comprising a spatial distribution pattern of the first model parameter
includes,
The general model parameters are
It is generated based on the average for each component of the compressed first model parameter received by the parameter server,
The average for each component is,
The learning method, which is the average of non-zero components among the components of the same index.
delete delete delete According to claim 1,
The obtaining step is
obtaining the second model parameter through a neural network;
A learning method comprising:
According to claim 1,
The first model parameter and the second model parameter are
A one-dimensional vector, a learning method.
receiving compressed model parameters from a plurality of clients;
generating a general model parameter based on an average for each component of the compressed model parameters;
transmitting the general model parameters to the plurality of clients;
including,
The compressed model parameters are
The plurality of clients input the first model parameter of the machine learning model into a plurality of one-dimensional convolutional layers included in the machine learning model, and a 1*5 convolutional mask of the plurality of one-dimensional convolutional layers is obtained. and extracting a feature of the first model parameter using shared weights, and generated using a max pooling layer based on the feature of the first model parameter,
The feature is
low-level features including correlations and associations between a plurality of one-dimensional convolutional layers included in the machine learning model; and
A high-level feature comprising a spatial distribution pattern of the first model parameter
includes,
The average for each component is,
The learning method, which is the average of non-zero components among the components of the same index.
delete 8. The method of claim 7,
The steps to create are
selecting some of the compressed model parameters; and
generating the general model parameter based on the part;
A learning method comprising:
A computer program stored in a computer-readable recording medium in combination with hardware to execute the method of any one of claims 1, 5, 6, 7, and 9.
An electronic device for performing federated learning, comprising:
a memory containing instructions; and
a processor for executing the instructions
including,
When the instructions are executed by the processor, the processor
train a machine learning model,
By inputting the first model parameter of the machine learning model into a plurality of one-dimensional convolutional layers included in the machine learning model, a 1*5 convolutional mask of the plurality of one-dimensional convolutional layers and a shared weight ( using shared weights) to extract a feature of the first model parameter,
generating a compressed first model parameter using a max pooling layer based on the characteristics of the first model parameter;
sending the compressed first model parameter to a parameter server;
receiving a general model parameter generated based on the compressed first model parameter from the parameter server;
recovering the general model parameters to obtain a second model parameter;
initialize the machine learning model based on the second model parameter;
The feature is
low-level features including correlations and associations between a plurality of one-dimensional convolutional layers included in the machine learning model; and
A high-level feature comprising a spatial distribution pattern of the first model parameter
includes,
The general model parameters are
It is generated based on the average for each component of the compressed first model parameter received by the parameter server,
The average for each component is
which is an average of non-zero components among components of the same index.
delete delete delete 12. The method of claim 11,
The processor is
The electronic device obtains the second model parameter through a neural network.
12. The method of claim 11,
The first model parameter and the second model parameter are
A one-dimensional vector, an electronic device.
In the server for performing federated learning,
a memory containing instructions; and
a processor for executing the instructions
including,
When the instructions are executed by the processor, the processor
receiving compressed model parameters from a plurality of clients;
Generating a general model parameter based on the component-by-component average of the compressed model parameters,
sending the general model parameters to the plurality of clients;
The compressed model parameters are
The plurality of clients input the first model parameter of the machine learning model into a plurality of one-dimensional convolutional layers included in the machine learning model, and a 1*5 convolutional mask of the plurality of one-dimensional convolutional layers is obtained. and extracting a feature of the first model parameter using shared weights, and generated using a max pooling layer based on the feature of the first model parameter,
The feature is
low-level features including correlations and associations between a plurality of one-dimensional convolutional layers included in the machine learning model; and
A high-level feature comprising a spatial distribution pattern of the first model parameter
includes,
The average for each component is,
The server, which is the average of the non-zero components of the components of the same index.
delete 18. The method of claim 17,
The processor is
selecting some of the compressed model parameters,
and generate the general model parameter based on the part.

Images