WO2023120776A1 - Device-to-device knowledge transmission method using proxy dataset in federated learning, and system therefor - Google Patents

Abstract

A device-to-device knowledge transmission method using a proxy dataset in federated learning, according to the present invention, comprises the steps of: downloading proxy data; generating, by a management server, management server output knowledge by machine learning input data by using a first artificial neural network function; generating, by first to nth devices, a plurality of pieces of device output knowledge by machine learning input data by using a second artificial neural network function; and if a preset number of times is reached, stopping the machine learning, transmitting the management server output knowledge to the first to nth devices, and transmitting the plurality of pieces of device output knowledge to the management server.

Description

Method and system for transferring knowledge between devices using proxy data set in federated learning

The present invention relates to a method and system for transferring knowledge between devices using a proxy data set in federated learning.

The present invention is supported by group research support (R&D) of the Ministry of Science and ICT (Task number: 1711135177, task number: 2020R1A4A1018607, research task name: Meta Federated Learning-based mobile edge computing system core structure development, task management institution: National Research Foundation of Korea, Project executing agency: Kyunghee University (International Campus) Industry-Academic Cooperation Foundation, research period: 2020.07.01.~2023.02.28.) and Information and Communication Technology Innovation Talent Fostering (R&D) (Task identification number: 1711139517, task number: 2021-0- 02068-001, Research project title: Research and development of artificial intelligence innovation hub, Project management agency: Information and Communication Planning and Evaluation Institute, Project executing agency: Kyunghee University (International Campus) Industry-University Cooperation Foundation, Research period: 2020.07.01.~2025.12.31.) It was derived from a study conducted as part of On the other hand, there is no property interest of the Korean government in any aspect of the present invention.

Recent mobile applications are equipped with various AI functions such as AI-based cameras, extended reality (XR), and intelligent assistants based on user data. However, AI functions that use centralized machine learning (ML) as above inevitably have concerns about leakage of personal information data.

Accordingly, the development of AI functions using machine learning (ML) based on federated learning (FL) that can prevent leakage of personal information data is continuously progressing.

However, the AI function using machine learning based on federated learning (FL) as described above may deteriorate the function of machine learning learning because there is heterogeneity of learning data in the central computing device and various user devices, and consequently, the machine learning learning model difference occurs.

Accordingly, there is a need for a technique for preventing the leakage of personal information data and at the same time improving the performance of machine learning learning in a central computing device and a user device.

The technical problem to be solved by the present invention is a method and system for transferring knowledge between devices using a proxy dataset in federated learning to prevent leakage of personal information data in an AI learning process using machine learning based on federated learning (FL). It is about.

In addition, the technical problem to be solved by the present invention is to use a proxy dataset in federated learning to improve the learning function of machine learning while minimizing the communication data load in the AI learning process using machine learning based on federated learning (FL). It relates to a method and system for transferring knowledge between devices.

A knowledge transfer method between devices using a proxy dataset in federated learning according to an embodiment of the present invention includes the steps of downloading proxy data, machine learning of input data using a first artificial neural network function in a management server, and outputting the information to the management server. Generating knowledge, generating a plurality of device output knowledge by machine learning input data using a second artificial neural network function in the first to nth devices, and stopping and managing machine learning when a preset number of times is reached and transmitting server output knowledge to first to nth devices and transmitting a plurality of device output knowledge to a management server.

In addition, the step of machine learning the input data using the first artificial neural network function in the management server according to an embodiment of the present invention to generate management server output knowledge may include a plurality of n-th devices provided from the first to n-th devices. It includes receiving output knowledge and generating a first loss function based on the nth management server output knowledge and proxy data generated by the first artificial neural network function based on the input data.

In addition, the step of generating management server output knowledge by machine learning the input data using the first artificial neural network function in the management server according to an embodiment of the present invention includes a plurality of n-th device output knowledge and n-th management server output Generating a first normalization function based on the knowledge and proxy data, generating a first artificial neural network function based on the first loss function and the first regularization function, and machine learning the input data in the first artificial neural network function. and generating an n+1th management server output knowledge.

In addition, the step of generating a plurality of device output knowledge by machine learning the input data using the second artificial neural network function in the first to nth devices according to an embodiment of the present invention is n+1 provided by the management server. Receiving management server output knowledge, uploading a plurality of n-th device personal data generated from each of the first to n-th devices, and each of a plurality of n-th device output knowledge generated based on the input data and a plurality of n-th device output knowledge and generating a plurality of second loss functions based on each of n device personal data.

In addition, the step of generating a plurality of device output knowledge by machine learning the input data using the second artificial neural network function in the first to nth devices according to an embodiment of the present invention, each of the plurality of nth device output knowledge , Generating a plurality of second normalization functions based on the n+1th management server output knowledge and proxy data, a plurality of second artificial neural network functions based on the plurality of second loss functions and the plurality of second normalization functions Generating and machine learning the input data based on a plurality of second artificial neural network functions and generating a plurality of n+1th device output knowledge.

In addition, when a preset number of times is reached according to an embodiment of the present invention, machine learning is stopped, the management server output knowledge is transmitted to the first to n th devices, and a plurality of device output knowledge is transmitted to the management server. The step of doing is, when the preset number of times is reached, stopping machine learning of the input data in the first artificial neural network function and transmitting the n+1 management server output knowledge to the first to nth devices and reaching the preset number of times In this case, stopping the machine learning of the input data in the second artificial neural network function and transmitting the output knowledge of the plurality of n+1 management devices to the management server. In addition, a knowledge transfer system between devices using a proxy dataset in federated learning according to an embodiment of the present invention uses proxy data and a first artificial neural network function to perform machine learning on input data and generate management server output knowledge. Includes server, proxy data and a second artificial neural network function to machine learn input data and generate a plurality of device output knowledge, and a control server that stops machine learning when a preset number of times is reached do.

In addition, the management server according to an embodiment of the present invention includes a first input unit that receives output knowledge from a plurality of n-th devices from first to n-th devices, and a first input unit generated by a first artificial neural network function based on input data. Generating a first loss function based on n management server output knowledge and proxy data, generating a first normalization function based on a plurality of n th device output knowledge, n th management server output knowledge, and proxy data, and generating a first loss function and a first modeling unit for generating a first artificial neural network function based on a first normalization function, and output knowledge of the n+1 management server, which is a result of machine learning of input data from the first artificial neural network function, to the first to nth devices. It includes a first output unit that outputs to.

In addition, each of the 1st to nth devices according to an embodiment of the present invention includes a second input unit for receiving the n+1th management server output knowledge, and a device that has learned from the pre-stored data set of the nth individual and the proxy data set. A device data set storage unit for uploading the knowledge of, a second loss function is generated based on the n-th device output knowledge generated based on the input data and the n-th device personal data stored in advance, and the n-th device output knowledge, n+1-th device output knowledge A second modeling unit for generating a second normalization function based on the management server output knowledge and proxy data, and generating a second artificial neural network function based on the second loss function and the second regularization function, and the second artificial neural network function and a second output unit for performing machine learning on the input data and generating n+1 th device output knowledge.

In addition, the control server according to an embodiment of the present invention stops machine learning of the input data in the first artificial neural network function when the preset number of times is reached, and transfers the n+1 management server output knowledge to the first to nth devices. and when the preset number of times is reached, the machine learning of the input data is stopped in the second artificial neural network function, and a plurality of n+1 management device output knowledge is controlled to be transmitted to the management server.

In addition, it includes a non-transitory recording medium that can be printed as a computer on which a program for executing a knowledge transfer method between devices using a proxy dataset in federated learning according to an embodiment of the present invention is recorded.

The method and system for transferring knowledge between devices using proxy datasets in federated learning according to the present invention can prevent leakage of personal information data in the AI learning process using machine learning based on federated learning (FL).

In addition, the knowledge transfer method and system between devices using proxy datasets in federated learning according to the present invention minimizes the load of communication data and increases the learning speed in the AI learning process using machine learning based on federated learning (FL). It can improve the learning function of machine learning.

1 is a diagram of a knowledge transfer system between devices using a proxy dataset in federated learning according to an embodiment of the present invention.

2 is a diagram illustrating a process of performing machine learning using an artificial neural network function in a management server and a plurality of devices according to an embodiment of the present invention.

3 is a flowchart illustrating a process of performing machine learning in a management server according to an embodiment of the present invention.

4 is a flowchart illustrating a process of performing machine learning using an artificial neural network function in a plurality of devices according to an embodiment of the present invention.

Hereinafter, various embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may be embodied in many different forms and is not limited to the embodiments set forth herein.

In order to clearly describe the present invention, parts irrelevant to the description are omitted, and the same reference numerals are assigned to the same or similar components throughout the specification. Therefore, the reference numerals described above can be used in other drawings as well.

In addition, since the size and thickness of each component shown in the drawings are arbitrarily shown for convenience of explanation, the present invention is not necessarily limited to the shown bar. In the drawing, the thickness may be exaggerated to clearly express various layers and regions.

In addition, the expression "the same" in the description may mean "substantially the same". That is, it may be the same to the extent that a person with ordinary knowledge can understand that it is the same. Other expressions may also be expressions in which "substantially" is omitted.

1 is a diagram of a knowledge transfer system between devices using a proxy dataset in federated learning according to an embodiment of the present invention.

In federated learning according to an embodiment of the present invention, the knowledge transfer system 1 between devices using a proxy dataset may include a management server 20 and a plurality of devices 40.

The management server 20 may include a first input unit 22 , a first modeling unit 23 , and a first output unit 24 .

The first input unit 22 may download proxy data. The first input unit 22 may receive input of a plurality of device output knowledge (g1, ..., gn, see FIG. 2 ), which is a result of machine learning by the plurality of devices 40 .

The first modeling unit 23 includes a plurality of device output knowledge (g1, ..., gn) provided from the first input unit 22, proxy data, and the first modeling unit 23 (or global (global) A first artificial neural network for machine learning the input data (x1) based on the management server output knowledge (gs, see FIG. 2), which is the result of machine learning from the first artificial neural network function (ANN(1)) stored in advance in the model). The function ANN(1) can be recreated.

Specifically, the first modeling unit 23 includes the management server output knowledge (gs), which is a result of machine learning from the first artificial neural network function (ANN(1)) stored in advance in the first modeling unit 23, and uploaded proxy data. A first loss function may be generated based on.

The first loss function generated by the first modeling unit 23 is the difference between the downloaded proxy data and the management server output knowledge (gs), which is the result of machine learning from the first artificial neural network function (ANN(1)) stored in advance. As a function, it can be generated using knowledge distillation (KD).

In addition, the first modeling unit 23 includes management server output knowledge (gs), which is a result of machine learning in the first artificial neural network function (ANN(1)), downloaded proxy data, and a plurality of device output knowledge (g1, . .., a first normalization function may be generated based on gn).

The first normalization function generated by the first modeling unit 23 is the management server output knowledge (gs), which is a result of machine learning in the first artificial neural network function (ANN(1)) based on the downloaded proxy data, and a plurality of devices. It represents the normalized value between output knowledge (g1, ..., gn) as a function, and can be generated using Kullback-Leiber divergence (KLD) and Jensen-Shannon divergence divergence.

The Kullback-Leiber divergence (KLD) is a function used to calculate the difference between two probability distributions. The information entropy that would occur if data were sampled using another distribution that approximated the ideal distribution. means the difference between The Jensen-Shannon divergence divergence uses the Kullback-Leiber divergence (KLD) as a distance concept, and calculates the difference between two information entropies based on the Kullback-Leiber divergence (KLD). and how to average it.

The first modeling unit 23 may regenerate the first artificial neural network function ANN(1) for machine learning of the input data x1 based on the first loss function and the first normalization function.

The first output unit 24 provides the management server output knowledge gs, which is a result of machine learning of the input data x1 based on the first artificial neural network function ANN(1), to a plurality of devices 40. can The management server output knowledge gs output from the first output unit 24 may be provided to a plurality of devices 40 through the network 30 .

The plurality of devices 40 may include a device data storage unit 41 , a second input unit 42 , a second modeling unit 43 , and a second output unit 44 .

The device data storage unit 41 may store device personal data of the device 40 . At this time, the device personal data relates to data of a user using the device 40 and may correspond to various types such as images, texts, and voice files. Device personal data stored in the device data storage unit 41 of the plurality of devices 40 may be different according to user data.

The second input unit 42 may download proxy data. The second input unit 42 may receive management server output knowledge (gs), which is a result of machine learning in the management server 20 .

The second modeling unit 43 includes management server output knowledge (gs) provided from the second input unit 42, device personal data provided from the device data storage unit 41, and previously stored in the second modeling unit 43. Based on the device output knowledge (g1, ..., gn), which is the result of machine learning in the second artificial neural network function (ANN (2 (1), ..., ANN (2 (n)), input data (x2) A second artificial neural network function (ANN(2(1), ..., ANN(2(n))) for machine learning can be recreated.

Specifically, the second modeling unit 43 may generate a second loss function based on the personal device data provided from the device data storage unit 41 and the downloaded proxy data.

The second loss function generated by the second modeling unit 43 represents the difference between the downloaded proxy data and the device personal data provided from the device data storage unit 41 as a function, and has the same knowledge as the first artificial neural network function. It can be produced using knowledge distillation (KD).

In addition, the second modeling unit 43 outputs device output knowledge (g1, ..., which is a result of machine learning from the second artificial neural network function (ANN(2(1), ..., ANN(2(n))). gn), the uploaded proxy data, and the management server output knowledge (gs) provided from the management server 20 may generate a second normalization function.

The second normalization function generated by the second modeling unit 43 is machine learning from the second artificial neural network function (ANN(2(1), ..., ANN(2(n))) based on the downloaded proxy data. As a function of the normalization value between the resultant device output knowledge (g1, ..., gn) and management server output knowledge (gs), the second artificial neural network function (ANN(2(1), ..., ANN(2 (n)) can be generated using the same Kullback-Leiber divergence (KLD) and Jensen-Shannon divergence divergence.

The second modeling unit 43 generates a second artificial neural network function (ANN(2(1), ..., ANN(2 (n)) can be reproduced.

The second output unit 44 outputs device output knowledge (which is a result of machine learning of the input data (x2) based on the second artificial neural network function (ANN(2(1), ..., ANN(2(n))) g1, ..., gn can be provided to the management server 20. Device output knowledge (g1, ..., gn) output from the second output unit 44 is managed through the network 30 It may be provided to the server 20.

The control server 50 may stop performing machine learning of the input data x1 in the first artificial neural network function ANN(1) when the preset number of times is exceeded. The control server 50 may stop machine learning the input data (x2) in the second artificial neural network function (ANN(2(1), ..., ANN(2(n))) when the preset number of times is exceeded. there is.

2 is a diagram illustrating a process of performing machine learning using an artificial neural network function in a management server and a plurality of devices according to an embodiment of the present invention.

The first artificial neural network function (ANN(1)) generated by the first modeling unit 23 included in the management server 20 includes the first input layer IL(1) and the first hidden layer HL(1) , and the first output layer OL(1).

Input data x1 may be applied to the first input layer IL( 1 ). Management server output value data zs may be generated while the input data x1 applied to the first input layer IL(1) passes through the first hidden layer HL(1). The management server output value data zs generated through the first hidden layer HL(1) may be applied to the first output layer OL(1).

The first hidden layer HL(1) may be configured by various activation functions, and the input data x1 applied to the first hidden layer HL(1) is dependent on various activation functions. The weight value may be multiplied by and converted into management server output value data (zs).

Hereinafter, the weight value multiplied by the input data x1 by an activation function is referred to as characteristic value data (or es).

At this time, the management server output value data (zs) includes information about the characteristic value data (es) multiplied by the input data (x1) by an activation function.

The management server output value data zs generated in the first output layer OL( 1 ) of the first modeling unit 23 may be applied to the first output unit 24 .

At this time, the first output unit 24 transmits the management server output knowledge gs to a plurality of devices ( 40) can be provided.

The second artificial neural network function (ANN(2(1)), ..., ANN(2(n))) generated by the second modeling unit 43 included in the plurality of devices 40 is the second input layer (IL(2)), a second hidden layer (HL(2)), and a second output layer (OL(2)).

Input data x2 may be applied to the second input layer IL( 2 ). Device output value data z1, ..., zn may be generated while the input data x2 applied to the second input layer IL(2) passes through the second hidden layer HL(2). The device output value data z1, ..., zn generated through the second hidden layer HL(2) may be applied to the second output layer OL(2).

The second hidden layer HL(2) may be configured by various activation functions, and the input data x2 applied to the second hidden layer HL(2) is dependent on various activation functions. The weight value may be multiplied by and converted into device output value data (z1, ..., zn).

Hereinafter, each of the weight values multiplied by the input data x2 by an activation function in the plurality of devices 40 is referred to as characteristic value data (or e1, ..., en).

At this time, the device output value data (or z1, ..., zn) is the characteristic value data (or e1, ..., en ) contains information about

A plurality of device output value data (z1, ..., zn) generated in the second output layer (OL(2)) of the second modeling unit 43 may be applied to the second output unit 44.

At this time, the second output unit 44 is based on a plurality of device output value data (z1, ..., zn) including information on the characteristic value data (e1, ..., en) through the network 30 Thus, a plurality of device output knowledge (g1, ..., gn) can be provided to the management server 20.

Referring to FIGS. 1 and 2 together, the management server 20 converts the weight value multiplied to the input data (x1) by an activation function into characteristic value data (or, es) or characteristic value data (es). Since the management server output knowledge (gs) is provided to a plurality of devices 40 based on the device output value data (zs) including information about, the communication speed can be increased by minimizing the load of communication data.

The plurality of devices 40 use the weight value multiplied by the input data (x2) by an activation function as characteristic value data (e1, ..., en) or characteristic value data (e1, ..., en ) Since a plurality of device output knowledge (g1, ..., gn) is provided to the management server 20 based on a plurality of device output value data (z1, ..., zn) including information about communication data The communication speed can be increased by minimizing the load.

The second artificial intelligence neural network (ANN(2(1)), ..., ANN(2(n))) generated by the second modeling unit 43 of the plurality of devices 40 is a plurality of devices 40 However, as described above, the plurality of devices 40 provide the management server 20 with characteristic value data (e1, ..., en) or device output value data (z1, ..., zn), device personal data is not transmitted. Accordingly, since the user's personal information data is not transmitted to the management server 20 in the plurality of devices 40, leakage of personal information data can be prevented.

The second artificial intelligence neural network (ANN(2(1)), ..., ANN(2(n))) generated by the second modeling unit 43 of the plurality of devices 40 is A first artificial intelligence neural network (ANN(1)) generated based on the management server output knowledge (gs) based on the provided management server output value data (zs) and generated by the first modeling unit 23 of the management server 20 is generated based on a plurality of device output knowledge (z1, ..., zn) based on device output value data (z1, ..., zn) provided from a plurality of devices (40), so learning from the management server (20) The heterogeneity of the learning data to be learned and the learning data to be learned from the plurality of devices 40 can be reduced, so that the function of machine learning learning can be increased.

3 is a flowchart illustrating a process of performing machine learning in a management server according to an embodiment of the present invention.

Proxy data can be downloaded in step S10.

Specifically, the first input unit 22 of the management server 20 may download proxy data.

In step S11, a plurality of n-th device output knowledge provided by a plurality of devices may be input.

Specifically, the first input unit 22 of the management server 20 is based on the second artificial neural network function (ANN(2(1), ..., ANN(2(n)) in the plurality of devices 40). A plurality of generated n-th device output knowledge (g1, ..., gn) may be input.

In step S12, a first loss function may be generated using the nth management server output knowledge (gs) calculated through the first artificial neural network function (ANN(1)) and proxy data.

Specifically, the first modeling unit 23 of the management server 20 is based on the n-th management server output knowledge (gs), which is a result of machine learning in the first artificial neural network function (ANN(1)), and uploaded proxy data. Thus, a first loss function can be generated.

n-th management server output knowledge (gs) calculated through the first artificial neural network function (ANN(1)) in step S13, a plurality of n-th device output knowledge (g1, ..., gn), and proxy data A first normalization function can be generated using

Specifically, the first modeling unit 23 of the management server 20 inputs n-th management server output knowledge (gs), which is a result of machine learning in the first artificial neural network function (ANN(1)), in step S11. A normalization function may be generated using the received plurality of n-th device output knowledge (g1, ..., gn) and proxy data.

In step S14, a first artificial neural network function may be regenerated based on the first loss function and the first normalization function.

Specifically, the first modeling unit 23 of the management server 20 generates a first artificial neural network function (ANN) based on the first loss function generated in step S12 and the first normalization function generated in step S13. (1)) can be reproduced.

In step S15, the regenerated first artificial neural network function (ANN(1)) may perform machine learning on the input data and generate the n+1 management server output knowledge (gs).

Specifically, the input data (x1) may be applied to the first input layer (IL(1)) of the first modeling unit 23 of the management server 20, and the first artificial neural network function (ANN(1)) Through this, the n+1th management server output knowledge (gs) can be generated.

In step S16, the n+1th management server output knowledge may be provided to each of a plurality of devices.

Specifically, the control server 50 stops machine learning in the first modeling unit 23 when the preset number of times of learning is reached, and the first output unit 24 outputs the n+1 management server output knowledge (gs). may be provided to each of the plurality of devices 40 .

4 is a flowchart illustrating a process of performing machine learning using an artificial neural network function in a plurality of devices according to an embodiment of the present invention.

Proxy data can be downloaded in step S20.

Specifically, the second input units 42 of the plurality of devices 40 may download proxy data.

In step S21, the n+1th management server output knowledge provided by the management server may be input.

Specifically, the second input unit 42 of the plurality of devices 40 may receive n+1th management server output knowledge gs provided from the management server 20 .

In step S22, a plurality of n-th device personal data generated by a plurality of devices may be uploaded.

Specifically, the plurality of devices 40 may upload the plurality of n-th device personal data stored in the device data storage unit 41 .

In step S23, a plurality of second loss functions may be generated based on the plurality of n-th device output knowledge and the plurality of n-th device personal data calculated through the second artificial neural network function.

Specifically, the second modeling unit 43 of the plurality of devices 40 uses the nth device personal data uploaded from the device data storage unit 41 and the second artificial neural network function (ANN(2(1), ... , A plurality of second loss functions may be generated based on a plurality of n-th device output knowledge (g1, ..., gn) calculated through ANN (2 (n)).

For example, the second modeling unit 43 of any one device 40 is the nth device personal data uploaded from the device data storage unit 41 of any one device 40 and any one device 40 ) A second loss function may be generated based on the n-th device output knowledge calculated through the second artificial neural network function.

Alternatively, the second modeling unit 43 of the other device 40 stores the nth device personal data uploaded from the device data storage unit 41 of the other device 40 and the other device 40. A second loss function may be generated based on the n-th device output knowledge calculated through the second artificial neural network function.

In step S25, a plurality of second normalization functions are generated by using the n-th device output knowledge calculated through the second artificial neural network function, the n+1-th management server output knowledge provided by the management server, and proxy data. can

Specifically, the second modeling unit 43 of the plurality of devices 40 is the second artificial neural network function (n-th, which is the result of machine learning from ANN(2(1), ..., ANN(2(n))) A normalization function may be generated using the device output knowledge (g1, ..., gn), the n+1th management server output knowledge (gs) provided by the management server 20, and proxy data.

In step S26, a second artificial neural network function may be regenerated based on the second loss function and the second normalization function.

Specifically, the second modeling unit 43 of the plurality of devices 40 calculates the second artificial neural network function (based on the second loss function generated in step 24 and the second normalization function generated in step S27). ANN(2(1), ..., ANN(2(n)) can be recreated.

In the second artificial neural network function regenerated in step S27, input data may be machine-learned and a plurality of n+1 device output knowledge may be generated.

Specifically, the input data x2 may be applied to the second input layer IL2(1), ..., IL2(n) of the second modeling unit 43 of the plurality of devices 40, and A plurality of n+1th device output knowledge (g1, ..., gn) can be generated through the 2 artificial neural network functions (ANN(2(1), ..., ANN(2(n)).

In step S28, a plurality of n+1th device output knowledge and n+1th management server output knowledge may be compared.

Specifically, the control server 50 stops machine learning in the second modeling unit 43 when the preset learning number is reached, and the second output unit 44 outputs n+1 management server output knowledge to the management server ( 20) can be provided.

The drawings and detailed description of the present invention referred to so far are only examples of the present invention, which are only used for the purpose of explaining the present invention, and are used to limit the scope of the present invention described in the meaning or claims. It is not. Therefore, those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the appended claims.

Claims (11)

downloading proxy data; generating management server output knowledge by machine learning input data using a first artificial neural network function in a management server; generating a plurality of device output knowledge by machine learning the input data using a second artificial neural network function in the first to nth devices; and Stopping the machine learning when a preset number of times is reached, transmitting the management server output knowledge to the first to nth devices, and transmitting the plurality of device output knowledge to the management server, A knowledge transfer method between devices using a proxy dataset in federated learning. According to claim 1, The step of generating management server output knowledge by machine learning input data using a first artificial neural network function in the management server, receiving a plurality of n-th device output knowledge provided by the first to n-th devices; and Generating a first loss function based on the nth management server output knowledge and the proxy data generated by the first artificial neural network function based on the input data, A knowledge transfer method between devices using a proxy dataset in federated learning. According to claim 2, Generating the management server output knowledge by machine learning the input data using the first artificial neural network function in the management server, generating a first normalization function based on the plurality of n-th device output knowledge, the n-th management server output knowledge, and the proxy data; generating the first artificial neural network function based on the first loss function and the first normalization function; and Further comprising the step of machine learning the input data in the first artificial neural network function and generating n + 1 management server output knowledge. A knowledge transfer method between devices using a proxy dataset in federated learning. According to claim 3, Generating a plurality of device output knowledge by machine learning the input data using a second artificial neural network function in the first to nth devices, receiving n+1 th management server output knowledge provided by the management server; Uploading a plurality of n-th device personal data generated in each of the first to n-th devices; and Generating a plurality of second loss functions based on each of a plurality of n-th device output knowledge generated based on the input data and each of the plurality of n-th device personal data, A knowledge transfer method between devices using a proxy dataset in federated learning. According to claim 4, Generating a plurality of device output knowledge by machine learning the input data using a second artificial neural network function in the first to nth devices, generating a plurality of second normalization functions based on the plurality of nth device output knowledge, the n+1th management server output knowledge, and the proxy data; generating a plurality of second artificial neural network functions based on the plurality of second loss functions and the plurality of second normalization functions; and Further comprising the step of machine learning the input data based on the plurality of second artificial neural network functions and generating a plurality of n + 1 device output knowledge. A knowledge transfer method between devices using a proxy dataset in federated learning. According to claim 5, When a preset number of times is reached, stopping the machine learning, transmitting the management server output knowledge to the first to nth devices, and transmitting the plurality of device output knowledge to the management server, stopping machine learning of the input data in the first artificial neural network function and transmitting the n+1th management server output knowledge to the first to nth devices when the preset number of times is reached; and Stopping machine learning of the input data in the second artificial neural network function when the preset number of times is reached and transmitting the plurality of n+1 management device output knowledge to the management server, A method for transferring knowledge between devices using a proxy dataset in federated learning A management server for machine learning input data and generating management server output knowledge using proxy data and a first artificial neural network function; first to nth devices machine-learning the input data and generating a plurality of device output knowledge using the proxy data and a second artificial neural network function; and Including a control server that stops the machine learning when a preset number of times is reached, A knowledge transfer system between devices using proxy datasets in federated learning. According to claim 7, The management server, a first input unit that receives output knowledge from a plurality of n-th devices from the first to n-th devices; Based on the input data, a first loss function is generated based on the nth management server output knowledge generated by the first artificial neural network function and the proxy data; A first normalization function is generated based on the plurality of nth device output knowledge, the nth management server output knowledge, and the proxy data, and the first artificial neural network is generated based on the first loss function and the first normalization function. a first modeling unit that generates a function; and Including a first output unit for outputting n + 1 th management server output knowledge, which is a result of machine learning of the input data in the first artificial neural function, to the first to n th devices, A knowledge transfer system between devices using proxy datasets in federated learning. According to claim 8, Each of the first to nth devices, a second input unit for receiving the n+1 th management server output knowledge; a device data set storage unit for uploading the plurality of n-th device output knowledge obtained by machine learning based on the previously stored n-th device personal data and the proxy data; A second loss function is generated based on the n-th device output knowledge generated based on the input data and the previously stored n-th device personal data, and the n-th device output knowledge, the n+1 management server output knowledge, and a second modeling unit generating a second normalization function based on the proxy data and generating the second artificial neural network function based on the second loss function and the second normalization function; and And a second output unit for machine learning the input data based on the second artificial neural network function and generating n + 1 th device output knowledge. A knowledge transfer system between devices using proxy datasets in federated learning. According to claim 9, The control server, When the preset number of times is reached, the first artificial neural network function stops machine learning of the input data, transmits the n+1 management server output knowledge to the first to nth devices, and reaches the preset number of times. In this case, controlling the second artificial neural network function to stop machine learning of the input data and transmit the output knowledge of the plurality of n+1 management devices to the management server. A data delivery system using proxy data in federated learning. Possible non-transitory recording media printed on a computer on which a program for executing the knowledge transfer method between devices using a proxy dataset in the federated learning of claim 1 is recorded.

Images