Method and device for diffusing communicatively efficient artificial intelligence models for improving performance of federated learning system

The FedDif method optimizes federated learning by using local diffusion to manage Non-IID data in federated learning systems, improving performance and reducing communication costs through efficient model spreading and updating, addressing the inefficiencies caused by Non-IID data characteristics.

WO2024237418A1PCT designated stage expired Publication Date: 2024-11-21IND UNIV COOP FOUND HANYANG UNIV ERICA CAMPUS
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
PCT/KR2023/095031
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-18
Filing Date
2023-07-13
Publication Date
2024-11-21

AI Technical Summary

Technical Problem

Federated learning systems face performance degradation due to non-independent and identically distributed (Non-IID) data characteristics among users, leading to inefficiencies in model learning and communication costs.

Method used

The implementation of a communication-efficient artificial intelligence model diffusion method, known as FedDif, which uses a local diffusion process to optimize the degree of learning (DoL) and IID distance, allowing for efficient model spreading and updating between user equipment (UEs) through device-to-device communication, thereby minimizing communication costs and improving learning performance.

Benefits of technology

FedDif enhances the learning performance of federated learning systems by up to 10.37% compared to baseline FL with non-IID settings, while reducing communication costs and the number of model transmissions, thus addressing the weight divergence and personal information leakage issues.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure KR2023095031_21112024_PF_FP_ABST
    Figure KR2023095031_21112024_PF_FP_ABST
Patent Text Reader

Abstract

The present specification relates to a method for operating a user equipment (UE) for diffusing artificial intelligence models in a federated learning (FL) system. The method may comprise the steps of: receiving a global model copy from a central server; constructing a local model on the basis of the global model copy and acquiring a first local model by training the local model by using the acquired data; updating the first local model to a second local model through the local training based on local diffusion; and transmitting the second local model to the central server.
Need to check novelty before this filing date? Find Prior Art

Description

Communication-Efficient Artificial Intelligence Model Diffusion Method and Device for Improving the Performance of Federated Learning Systems

[0001] The present invention relates to a communication-efficient artificial intelligence model diffusion method and device for improving the performance of a federated learning system, and more specifically, to a communication-efficient artificial intelligence model diffusion method and device for improving the performance of a federated learning system, which includes a novel diffusion strategy for exchanging artificial intelligence models between users prior to collation of artificial intelligence models.

[0002]

[0003] In the next-generation mobile communication system, as many communication devices require large communication capacity, various technologies are being proposed to satisfy the requirements of enhanced mobile broadband (eMBB) service, massive Machine Type Communications (mMTC) service, and Ultra-Reliable and Low Latency Communications (URLLC) service as improved services compared to the existing radio access technology (RAT).

[0004] Furthermore, artificial intelligence or machine learning may be considered in next-generation mobile communication systems, and federated learning (FL) may be applied based on these technologies. Federated learning is a new learning paradigm that addresses the privacy leak problem of centralized learning. It can protect personal privacy, reduce base station load through distributed processing, and reduce traffic between base stations and terminals. Federated learning is a distributed learning technology that trains models using distributed data held by multiple users. In other words, federated learning can utilize distributed data from multiple users. Here, if a specific user with the characteristic of non-independent and identically distributed (Non-IID) exists among the multiple users, federated learning performance may deteriorate. The following describes a method to improve this.

[0005]

[0006] [Prior Art Literature]

[0007] [Patent Document]

[0008] Korean Patent Publication No. 10-2022-0168128 A

[0009]

[0010] This specification relates to a communication-efficient artificial intelligence model diffusion method and device for improving the performance of a federated learning system.

[0011] This specification relates to a method and device for performing optimization by utilizing the degree of learning (DoL) of an artificial intelligence model and the independent and identically distributed (IID) distance for proximity to the identity distribution of the DoL to improve the performance of a federated learning system.

[0012] This specification relates to a method and apparatus for improving performance by adding a local diffusion process to a federated learning system.

[0013] This specification relates to a specific operating method and device for a local diffusion process.

[0014]

[0015] According to one embodiment of the present specification, a method for operating a user equipment (UE) for artificial intelligence model diffusion in a federated learning (FL) system may include: receiving a global model copy from a central server; constructing a local model based on the global model copy, and performing learning on the local model using the acquired data to obtain a first local model; updating the first local model to a second local model through local learning based on local diffusion; and transmitting the second local model to the central server.

[0016] In addition, according to one embodiment of the present specification, a user equipment (UE) for performing artificial intelligence model diffusion in a federated learning (FL) system includes a memory, a transceiver, and a control unit for controlling the memory and the transceiver, wherein the control unit can receive a global model copy from a central server, construct a local model based on the global model copy, perform learning on the local model through the acquired data to obtain a first local model, update the first local model to a second local model through local learning based on local diffusion, and transmit the second local model to the central server.

[0017] In addition, according to one embodiment of the present specification, a method for operating a central server for spreading an artificial intelligence model in a federated learning (FL) system may include a step of transmitting a copy of a global model to a plurality of UEs participating in federated learning, wherein each of the plurality of UEs builds a local model based on the copy of the global model, performs learning on the local model through data acquired by each of the plurality of UEs, and determines a next learning UE by receiving a candidate IID distance from each of the plurality of UEs based on the local learning, a step of transmitting scheduling information based on the determined next learning UE to the plurality of UEs participating in federated learning, and a step of obtaining a local model learned based on local learning from the plurality of UEs and updating the global model.

[0018] In addition, according to one embodiment of the present specification, a central server for performing artificial intelligence model diffusion in a federated learning (FL) system includes a memory, a transceiver, and a control unit for controlling the memory and the transceiver, wherein the control unit transmits a copy of a global model to a plurality of UEs participating in the federated learning, each of the plurality of UEs constructs a local model based on the copy of the global model, performs learning on the local model through data acquired by each of the plurality of UEs, receives candidate IID distances from each of the plurality of UEs based on the local learning to determine a next learning UE, transmits scheduling information based on the determined next learning UE to the plurality of UEs participating in the federated learning, and obtains a local model learned based on local learning from the plurality of UEs to update the global model.

[0019] Additionally, the following may be commonly applied:

[0020] According to one embodiment of the present specification, a UE may be a UE participating in federated learning based on a central server.

[0021] Additionally, according to one embodiment of the present specification, when local learning is performed in the first diffusion round, the UE transmits a Degree of Learning (DoL) to a plurality of UEs participating in federated learning based on a central server, wherein the DoL may represent a cumulative data distribution learned through data acquired based on a local model.

[0022] Additionally, according to one embodiment of the present specification, the DoL may be transmitted to multiple UEs participating in federated learning based on a broadcasting scheme.

[0023] Additionally, according to one embodiment of the present disclosure, the DoL may be transmitted to a central server via an uplink control channel, and the central server may transmit the DoL to a plurality of UEs participating in federated learning via a downlink control channel.

[0024] In addition, according to one embodiment of the present specification, a UE may receive a DoL from each of a plurality of UEs participating in federated learning, reflect data state information (DSI) of a first local model to each DoL to generate a preliminary DoL, and derive a first candidate Independent and Identically Distributed (IID) distance based on the preliminary DoL and transmit the result to a central server.

[0025] Additionally, according to one embodiment of the present specification, the central server derives bid price information by deriving a difference value between a second candidate IID derived in a second diffusion round and a first candidate IID, wherein the second diffusion round is a diffusion round prior to the first diffusion round, and the central server derives bid price information for each of a plurality of UEs participating in federated learning to determine the next learning UE through UE evaluation.

[0026] Additionally, according to one embodiment of the present disclosure, the central server may further consider channel state information (CSI) for each of the plurality of UEs to determine the next learning UE through UE evaluation.

[0027] Additionally, according to one embodiment of the present specification, when a UE is determined to be the next learning UE, the UE may transmit the first local model to a plurality of UEs participating in federated learning based on a device to device (D2D) manner.

[0028] Additionally, according to one embodiment of the present specification, when the UE receives a local model of the next learning UE from the next learning UE based on the D2D method, the UE can perform learning based on the local model of the next learning UE to obtain a second local model.

[0029]

[0030] This specification has the effect of providing a communication-efficient artificial intelligence model diffusion method for improving the performance of a federated learning system.

[0031] This specification has the effect of improving performance by performing optimization using the Degree of Learning (DoL) of an artificial intelligence model and the IID (independent and identically distributed) distance for proximity to the identity distribution of the DoL to improve performance of a federated learning system.

[0032] This specification has the effect of improving performance by adding a local diffusion process to a federated learning system.

[0033] This specification has the effect of providing a specific method of operation of a local diffusion process.

[0034] The effects that can be obtained from this specification are not limited to the effects mentioned above, and other effects that are not mentioned can be clearly understood by a person having ordinary skill in the art to which the present invention pertains from the description below.

[0035]

[0036] Figure 1 is a diagram illustrating a network environment applied to the present disclosure.

[0037] Figure 2 is a drawing showing a device configuration applied to the present disclosure.

[0038] Figure 3 is a diagram illustrating an existing artificial intelligence model learning paradigm applicable to the present disclosure.

[0039] Figure 4 is a diagram illustrating a federated learning paradigm applicable to the present disclosure.

[0040] FIG. 5 is a diagram illustrating an operation in which personalized data applicable to the present disclosure degrades learning performance.

[0041] FIG. 6 is a diagram illustrating a method for minimizing diffusion efficiency by minimizing the IID distance for UEs having a non-IID data set applicable to the present disclosure.

[0042] Figure 7 is a graph comparing learning performance and communication cost applicable to the present disclosure.

[0043] Figure 8 is a graph comparing learning performance and communication cost applicable to the present disclosure.

[0044] Figure 9 is a graph comparing learning performance and communication cost applicable to the present disclosure.

[0045] Figure 10 is a diagram showing a FedDif operation process applicable to the present disclosure.

[0046] Figure 11 is a flowchart illustrating a communication-efficient artificial intelligence model assurance method for improving the performance of a federated learning system applicable to the present disclosure.

[0047] Figure 12 is a flowchart illustrating a communication-efficient artificial intelligence model diffusion method for improving the performance of a federated learning system applicable to the present disclosure.

[0048]

[0049] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The detailed description set forth below, together with the accompanying drawings, is intended to illustrate exemplary embodiments of the present invention and is not intended to represent the only embodiments in which the present invention may be practiced. The following detailed description includes specific details to provide a thorough understanding of the present invention. However, one of ordinary skill in the art will appreciate that the present invention may be practiced without these specific details.

[0050] The following embodiments combine components and features of the present invention in a predetermined form. Each component or feature may be considered optional unless explicitly stated otherwise. Each component or feature may be implemented without being combined with other components or features. Furthermore, some components and / or features may be combined to form an embodiment of the present invention. The order of operations described in the embodiments of the present invention may be changed. Some components or features of one embodiment may be included in another embodiment or may be replaced with corresponding components or features of another embodiment.

[0051] Specific terms used in the following description are provided to aid in understanding the present invention, and the use of such specific terms may be changed to other forms without departing from the technical spirit of the present invention.

[0052] In some cases, to avoid obscuring the concepts of the present invention, well-known structures and devices are omitted or depicted in block diagram form focusing on the core functions of each structure and device. Furthermore, the same components are described using the same reference numerals throughout this specification.

[0053] Additionally, while the terms "first" and / or "second" may be used herein to describe various components, these components should not be limited by these terms. These terms are only intended to distinguish one component from another; for example, without departing from the scope of the concepts of this specification, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component.

[0054] Additionally, throughout the specification, when a part is said to "include" a component, this does not exclude other components, but rather implies the inclusion of other components, unless otherwise specifically stated. Furthermore, terms such as "unit" and "part" used in the specification mean a unit that processes at least one function or operation, which may be implemented by a combination of hardware and / or software.

[0055] Figure 1 is a diagram showing a network environment applied to the present disclosure.

[0056] Referring to FIG. 1, devices (110, 120, 130, 140) may be connected via a network (150). For example, devices (110, 120, 130, 140) may be connected to at least one device or server via a wireless network (150) and may transmit and receive data therethrough. In addition, devices (110, 120, 130, 140, or user devices) may be mobile devices or fixed devices. Specifically, devices (110, 120, 130, 140) may be mobile devices such as smart phones, tablets, or wearable devices. In addition, devices (110, 120, 130, 140) may be fixed devices such as computers, laptops, and PCs. As another example, the devices (110, 120, 130, 140) may be IoT (internet of things) devices, VR (virtual reality) / AR (augmented reality) devices, and other devices, and are not limited to a specific embodiment.

[0057] Additionally, the server may be a device capable of providing content or services by being connected to one or more devices via a network (150). As a specific example, the server may provide services or responses to requests from at least one device connected via the network (150). Here, the server may be linked based on software or applications installed on at least one device, thereby providing services.

[0058] FIG. 2 is a diagram illustrating a device configuration applicable to the present disclosure. Referring to FIG. 2, the device (210) may include a control unit (211), a transceiver unit (212), and a memory (213). Furthermore, the device (210) may further include other configurations in addition to the aforementioned configurations. The device (210) of FIG. 2 may transmit and receive data through communication with another device (220) and may be a user equipment (UE). For example, the device (210) may refer to a smartphone, a smart pad, a laptop, a PC, and other communication-capable devices, but may not be limited to a specific device.

[0059] For example, the control unit (211) of the device (210) may be configured to execute and process commands related to the operation or driving of the device (210). The control unit (211) may be a logical entity that controls the transceiver unit (212) and other components, and is not limited to a specific embodiment. The control unit (211) of the device (210) may be configured to process commands of a computer program by performing arithmetic, logic, and input / output operations. For example, commands may be provided to the control unit (211) based on signals stored in the memory (213) or acquired through the transceiver unit (212), and the control unit (211) may perform operations based thereon.

[0060] The transceiver (212) may provide a function for communicating with at least one of another device (220) and a server via a network. For example, the other device (220) may also include a control unit (221), a transceiver (222), and a memory (223). In addition, the other device (220) may also be the aforementioned smart device, a PC, or other communication-capable device, and is not limited to a specific embodiment.

[0061] In addition, as an example, the device (210) may include an input unit (not shown) and an output unit (not shown). Here, the input unit may be a configuration that provides a keyboard, mouse, touchpad, camera, and other input signals, and the output unit may be a configuration that provides a display, speaker, and other output signals. However, the present invention may not be limited thereto. The external output device may include a device such as a display, speaker, or haptic feedback device.

[0062] The memory (213) is a non-transitory computer-readable recording medium, and may include a permanent mass storage device such as a random access memory (RAM), a read only memory (ROM), a disk drive, a solid state drive (SSD), or a flash memory, and is not limited to a specific form. In addition, the memory (213) may store commands or program codes related to the driving or operation of the device (210). In addition, the operating system or other software of the device (210) may be stored in the memory (213). In addition, as an example, the device (210) may use software loaded from a separate computer-readable recording medium. Here, the computer-readable recording medium may include a floppy drive, a disk, a tape, a DVD / CD-ROM drive, a memory card, and other recording media, and is not limited to a specific embodiment.

[0063] Hereinafter, the user equipment (UE) as the device (210, 220) of FIG. 2 is described based on the user equipment. In the following, each UE can communicate with a base station or a central server, and an artificial intelligence model can be built in the base station or the central server and shared with each UE. For example, the base station or the central server may be a base station of a 5G communication system, a 6G base station for a next-generation communication system, or any other type of node or server. In other words, the base station or the central server is merely an entity that is connected to the UE and learns an artificial intelligence model based on federated learning, and may not be limited to a specific form. In the following, for the convenience of explanation, it is referred to as a central server. However, this is only a name for the convenience of explanation and is not limited thereto, and may take other forms.

[0064] A federated learning (FL) system can be applied to jointly train a single AI model based on multiple UEs in a wireless communication network or wireless local area network (WLAN). Here, the UE can include not only smartphones but also any terminal equipped with computing capabilities capable of performing AI learning, producing personalized data, and connecting to a network. Each of the devices in FIG. 2 described above can correspond to a UE.

[0065] Here, as an example, the AI ​​model may be an AI model that uses personalized data. As a specific example, the AI ​​model may be trained based on a classification problem that classifies data into C classes (C-class classification problem). For example, the C-class classification problem may refer to learning to predict whether input data belongs to one of C classes, where C may be the number of classes. Here, the AI ​​model that operates based on the C-class classification problem may learn a function that analyzes the input data and predicts the probability of belonging to each class, and the training may be performed using the FL described above. However, the classification problem of classifying into C classes is only one example and is not limited thereto, and may be applied to other AI models.

[0066] FL is an artificial intelligence model of a central server that copies a global model and transmits it to multiple UEs, and each UE can use its own data to learn the global model it has received.

[0067] More specifically, multiple UEs can train their own local models based on the global model using their own data. Each of the multiple UEs can then transmit information about the trained local models to a central server. The central server can then update the global model based on the local models received from each of the multiple UEs. By repeating the above-described series of processes, the FL can train the global model without collecting data from the UEs (or users).

[0068] For example, FIG. 3 is a diagram illustrating an existing artificial intelligence model learning paradigm applicable to the present disclosure, and FIG. 4 is a diagram illustrating a federated learning paradigm applicable to the present disclosure. Referring to FIG. 3, an artificial intelligence model (310) may be built in a central server (or base station, 330). The central server (330) may collect user data (340) from each of a plurality of UEs (320) and learn the artificial intelligence model (310) based on the user data. Here, as an example, the existing artificial intelligence model learning paradigm may form an independent and identically distributed (IID) data set in the process of collecting and verifying user data (340) in the central server (330) and learn the artificial intelligence model based on the IID data set. However, the above-described method has a problem that the user's data privacy may be violated by the central server management entity and that enormous communication costs may be incurred in the process of collecting user data.

[0069] On the other hand, referring to FIG. 4, in the federated learning paradigm, the central server (330) shares the constructed artificial intelligence model (310) with a plurality of UEs (320), and each of the plurality of UEs (320) can construct a local model based on the artificial intelligence model (310) received from the central server (330) and perform learning on its own. Thereafter, the plurality of UEs (320) transmit the results of the learning to the central server (330), and the central server (330) can update the artificial intelligence model (310) through the learning result information received from each UE (320). That is, since the central server (330) does not directly obtain user data (340) for each of the plurality of UEs (320), the user's data privacy is protected, and the traffic between the central server (330) and the plurality of UEs (320) can be reduced.

[0070] However, each of the plurality of UEs (320) may acquire user data (340) based on its own purpose and tendency. That is, when comparing the user data (340) for the plurality of UEs (320), there may exist a UE that has a characteristic of non-independent identity distribution (non-IID) that is biased toward a specific class or produces very small data. Here, the learning result information transmitted to the central server (330) by the UE (320) that performs local learning based on the non-IID data set may cause the learning performance of the global model to be significantly reduced by the artificial intelligence model (310) of the central server.

[0071] More specifically, the training of the AI ​​model (310) is performed through optimization that minimizes the error between the training data set and the AI ​​model's predictions. Here, if the difference in the learning results of each UE is significant, the central server (330) may not be able to optimize for error minimization when training the AI ​​model (310) using the learning result information, resulting in performance degradation.

[0072] As an example, FIG. 5 is a diagram illustrating an operation in which personalized data applicable to the present disclosure degrades learning performance. Referring to FIG. 5(a), a central server can obtain local model learning information (520) from each of a plurality of UEs. The central server can update an artificial intelligence model through an optimization operation that minimizes errors based on the local model learning information (520) of the plurality of UEs. In other words, an artificial intelligence model with minimized errors can be derived based on the local model learning information (520). Here, if the local model learning information (520) has an independent identity distribution (IID), the above-described optimization may be performed with little performance degradation. On the other hand, referring to FIG. 5(b), if the local model learning information (520) has a non-IID, the above-described optimization may be performed with performance degradation.

[0073] In the following, considering the above, a method for solving the problem of learning performance degradation due to the non-IID characteristics of user data of UEs in the FL system is described. Specifically, the problem of learning performance degradation due to the non-IID characteristics can be solved by correcting the learning path of the model so that it goes to the optimal point during the learning process using user data of each UE. That is, before the central server collects local models from each UE as AI models, the local models can learn the user data of other UEs and correct the learning path. In other words, the local models can be diffused (Federated Diffusion, hereinafter referred to as FedDif) before AI model collection so that they can learn the user data (or data sets) of other UEs, and a specific method for this is described below.

[0074] For example, in existing AI learning paradigms, optimization can be performed based on stochastic gradient descent (SGD). SGD is an optimization algorithm that adjusts model weights using the gradient of a loss function. SGD calculates the gradient of the loss function for only one randomly selected data sample and updates the model weights. This allows for rapid model learning because calculations are performed on only a subset of samples rather than the entire data set. SGD can utilize data samples organized based on mini-batch size in large data sets, and the learning speed and accuracy can be controlled by adjusting the mini-batch size. For example, in existing AI learning paradigms, a single IID data set is divided into mini-batches of data samples, and optimization is performed only once for each mini-batch. SGD-based optimization may incur some loss in optimization performance, but the learning speed and computational complexity can be measured.

[0075] Below, we describe a method for updating a federated learning model, considering the SGD method described above. For example, in federated learning, each UE (or user) can be treated as a single mini-batch. Here, a UE needs to ensure that the cumulative data distribution it has learned for its AI model becomes the identity distribution. For example, a UE can ensure that the learned cumulative data distribution becomes the identity distribution by applying a local model of another specific UE to its learned cumulative data distribution. To achieve this, a FedDif operation may be required, which efficiently propagates the AI ​​model to a specific UE among multiple UEs that have performed training, and this is described in detail below.

[0076] For example, as local models spread throughout the system based on FedDif, the number of transmissions for local models can increase, leading to a linear increase in the communication costs incurred during training due to this operation. Therefore, it is necessary to minimize the communication costs and the number of model transmissions used in a single communication.

[0077] For example, to minimize the communication cost used in a single communication, local model transmission may be performed based on device-to-device communications (D2D communications) in a wireless local area network (WLAN). Here, the wireless local area network may be a communication performed based on a local area network, and the communication may be performed based on WiFi or other technologies. For example, local model information may be exchanged between terminals based on the wireless local area network. As another example, D2D communications may be performed based on a wireless local area network. D2D communications may be referred to as sidelink, V2X, and other names, but are not limited to a specific form. D2D communications may support communication between terminals based on 5G communications in a wireless local area network. Furthermore, for example, data exchange between terminals may be possible based on D2D communications or sidelink communications in a next-generation communication system, and is not limited to a specific form.

[0078] That is, local model information can be exchanged between terminals to receive local model information from other UEs so that the cumulative data distribution learned by the UE becomes an identity distribution. Here, in order to minimize the number of transmissions of the AI ​​model, the degree of learning (DoL), which indicates the cumulative data distribution learned by the UE based on the AI ​​model, and the IID distance (IID distance), which indicates how close the DoL is to the identity distribution, can be utilized.

[0079] For example, diffusion efficiency can refer to the rate of reduction in the IID distance relative to the communication cost incurred by all UEs participating in the FL. Based on the above, an auction can be designed to find the next trainer UE that can maximize diffusion efficiency within the system through auction theory. Auction theory can be a method in which multiple entities participate in the bidding and a specific entity is selected through an auction that allows for the best result among the multiple entities. For example, in the auction, the evaluation of each UE for an AI model can be determined based on the rate of reduction in the IID distance based on the learning rate. Here, each UE can quantify how much its own data set can reduce the IID distance for each AI model's DoL and submit a bid to the central server.

[0080] The central server can use inter-UE wireless channel state information (CSI) and bidding information to calculate diffusion efficiency and select the next learning UE that maximizes diffusion efficiency for each AI model. Then, by utilizing the information from the next learning UE, diffusion configurations can be applied to other UEs, maximizing diffusion efficiency.

[0081]

[0082] Based on the above, FedDif can prevent deviations from the optimal learning path that can occur during the training and collation of AI models in FL. Specifically, deviations from the optimal learning path can occur when the weights of the AI ​​model diverge compared to the weights of the optimized model, resulting in a divergence problem.

[0083] Here, FedDif can be a series of processes that utilize the IID distance to select the next learning UE that minimizes the IID distance from the current DoL of the AI ​​models, considering that IID data sets can be generated by combining non-IID data sets. Based on the above, FedDif can resolve the weight divergence problem that can arise due to non-IID data sets.

[0084] As an example, FIG. 6 is a diagram illustrating a method for minimizing diffusion efficiency by minimizing the IID distance for UEs having a non-IID data set applicable to the present disclosure. Referring to FIG. 6(a) and FIG. 6(b), the graphs show whether the IID distance and diffusion efficiency converge to 0 as diffusion rounds progress. Here, in the simulation environment of FIG. 6(a) and FIG. 6(b), the number of UEs is 10, and they can be evenly distributed in a circular system with a radius of 250 m. However, this is merely an example for convenience of explanation and may not be limited thereto.

[0085] For example, the UE's data set can be allocated by dividing the CIFAR-10 data set into 10 non-IID data sets based on the Dirichlet distribution. Furthermore, wireless channel and frame modeling can be based on, but not limited to, the channel model and numerology specified in the 5G standard. Furthermore, the transmission power used in communication with the central server and D2D communication can be set to 40 dBm and 23 dBm, respectively. Since model transmission and reception between UEs is performed via D2D communication, signal strength attenuation due to distance may be more severe. However, since the distance between UEs is relatively shorter than the distance between the central server and the UE, high-quality communication can be achieved even with low transmission power.

[0086] For example, FedDif can include two parameters that can control the degree of diffusion by considering hyperparameter tuning and communication parameters for learning artificial intelligence models. The first parameter is the minimum tolerable IID distance (MID). ) In the case where the data of UEs within the system are distributed in various ways for each class, the IID distance can theoretically converge to 0, but in reality, errors may exist. Therefore, The IID distance can be a minimum threshold that stops diffusion when it decreases, and can be set by the administrator. As shown in Figures 6(a) and 6(b), as diffusion rounds progress, the IID distance and diffusion efficiency can converge to 0. Based on this, a specific method for efficient artificial intelligence model diffusion to improve the performance of a federated learning system is described later in Figure 10.

[0087] Also, as an example, Fig. 7 is a graph comparing learning performance and communication costs applicable to the present disclosure. Referring to Fig. 7, The smaller the number, the higher the learning performance of the AI ​​model, but more diffusion can be performed repeatedly.

[0088] Additionally, the second parameter is the Minimum tolerable quality of service (MQS). ) may be. As described above, if the next learning UE is located on the opposite side of the system after the auction for diffusion is conducted, the distance between the current UE and the next learning UE may be very long. Therefore, communication costs may increase exponentially. Considering the above, a minimum threshold value may be required to diffusion the model while ensuring a certain level of communication quality, and this can be set by the administrator.

[0089] As an example, Fig. 8 is a graph comparing learning performance and communication cost applicable to the present disclosure. Referring to Fig. 8(a) and Fig. 8(b), the difference in learning performance is small, but As this increases, the number of diffusion iterations required for optimization may increase. For example, in the simulation environment of Figure 8, the system size is relatively small, so the change is not significant. However, as the system size increases, the difference in performance for the corresponding parameter may become larger.

[0090] Additionally, as an example, the degree of non-IID of a data set is determined by the concentration parameter (α) of the Dirichlet distribution, and the smaller the size, the further it may deviate from the IID distribution. Figure 9 is a graph comparing learning performance and communication cost applicable to the present disclosure. Referring to Figures 9(a) and 9(b), as α decreases, the learning performance itself significantly decreases, and the communication cost for achieving the target learning performance may increase significantly. However, FedDif can always exhibit higher performance than the basic FL.

[0091] For example, the exchange of AI models between existing UEs can be randomly performed by the central server. Here, the weight divergence problem described above may not be resolved. For example, if only users with increasing IID distances remain in the system, the learning path of the central server's AI model may deviate from the optimal path compared to the default FL. On the other hand, since FedDif is a process that solves the problem of maximizing diffusion efficiency, it may not perform model diffusion that deviates from the optimal path.

[0092] As another example, previously, if a local model was transmitted to a central server without considering communication costs, the central server could randomly shuffle the global model and transmit it to the UEs without collating it. However, if UEs participating in the aforementioned learning process are scheduled together with UEs not participating, the communication quality of the UEs not participating in the learning process may deteriorate. On the other hand, FedDif not only theoretically guarantees a solution to the weight divergence problem when performing diffusion, but also optimizes diffusion for communication efficiency, enabling it to solve the problem more efficiently than conventional techniques.

[0093]

[0094] FIG. 10 is a diagram illustrating a FedDif operation process applicable to the present disclosure. Referring to FIG. 10, a central server may broadcast a copy of a global model to multiple UEs based on an artificial intelligence model. For example, the process of the central server broadcasting a copy of the global model to multiple UEs may be a global initialization (1010) process. Thereafter, the multiple UEs may transmit local model information learned based on their respective data sets to the central server. In other words, the central server may collect local model information learned from each of the multiple UEs, and this process may be a global aggregation (1030) process. Here, as an example, a local diffusion (1020) process may be further performed between the global initialization (1010) process and the global aggregation (1030) process in FIG. 10. Specifically, the process of training an AI model through a central server's global initialization (1010), local diffusion (1020), and global aggregation (1030) processes may be performed based on a regular cycle. For example, T communication rounds may be performed for federated learning. Figure 10 may be the tth communication round among T communication rounds. However, this is merely an example for convenience of explanation and may not be limited thereto.

[0095] Here, for example, a local diffusion process (1020) may be performed after a global initialization (1010) process in one communication round. The goal of local diffusion (1020) may be to preserve user privacy, a fundamental principle of FL, while minimizing communication costs and maximizing learning performance through minimal diffusion. Therefore, the detailed processes constituting local diffusion (1020) need to ensure that the UE's data sample itself or its distribution cannot be inferred. Considering the above, the detailed processes may utilize the DoL described above as the cumulative distribution of the model's learning data, rather than the UE's data distribution. The DoL may be a learning diagram for an AI model to represent the cumulative data distribution it has learned, as described above.

[0096] Transmission and reception steps that take local diffusion into account can be performed via a D2D overlay mode that also utilizes channels for cellular mobile communications. That is, the UE can perform D2D communications using the same channel used for communication with a central server (or base station). Alternatively, D2D communications can operate in a mode scheduled and controlled by a central server (or base station) or in a mode directly scheduled and controlled by the UE, although this may not be limited to a specific mode.

[0097] Here, FedDif needs to be performed considering the data set status, the model's learning level, the channel status for communication, and other factors. For example, referring to FIG. 10, there may be a central server and multiple UEs participating in the FL within the system. As a specific example, there are N terminals in the system, It may be considered that a representative terminal participates in FL. However, this is merely an example for convenience of explanation and may not be limited to this.

[0098] The data set of the ith terminal participating in FL is a random variable Probability density function of is generated by can be expressed as . Here, The size of the data set is Since the data sets between UEs have non-IID characteristics, the distribution of terminals i and j is satisfies. The data state information (DSI) of terminal i is the ratio of data samples for each class. It can be expressed as . Federated learning performs a total of T communication rounds, and in the t-th communication round, the local diffusion is A number of diffusion rounds can be performed. Here, the set of UEs that the local model m passes through in the kth diffusion round can be expressed as a diffusion chain, as shown in Mathematical Formula 1 below.

[0099] [Mathematical Formula 1]

[0100]

[0101]

[0102] In mathematical expression 1 can be a UE that trains a local model m in the kth diffusion round. Also, the diffusion chain The size of all data sets trained can be as shown in Equation 2 below. In addition, the DoL of the local model m in the kth diffusion round can be as shown in Equation 3 below.

[0103] [Equation 2]

[0104]

[0105]

[0106] [Equation 3]

[0107]

[0108]

[0109] Here, DoL, similar to DSI, can be the proportion of all data samples learned by the local model m in k diffusion rounds. Here, DoL The IID distance can be the difference from the identity probability distribution and can be defined as the Wasserstein-1 distance (Earth-mover distance) as in Equation 4 below. For example, the Wasserstein-1 distance can be a method for calculating the distance between two probability distributions. However, it is not limited thereto, and various functions such as Kullback-Leibler divergence (KLD) or Jenssen-Shannon distance (JSD) can also be utilized. For example, based on the above, only the similarity between the probability distribution and the identity distribution needs to be calculated. Therefore, a function that can express the similarity can be applied and may not be limited to a specific form. For example, in Equation 4, μ and can be the identity probability distribution and its DSI.

[0110] [Equation 4]

[0111]

[0112]

[0113] Next, referring to FIG. 10, the local diffusion process may include DoL broadcasting (1021), candidates of IID distance reporting (1022), diffusion configuration (1023), and model transmission (1024), and a local training process based thereon. For example, the DoL broadcasting (1021) process may be a process in which all UEs transmit the DoL of the local model they have trained to neighboring UEs based on a broadcasting method. For example, each UE may find a UE that can reduce the IID distance when training the local model based on the DoL of the local model received based on the broadcasting method. For example, the DoL broadcasting (1021) process may be performed by broadcasting the DoL to neighboring UEs based on a beacon message or other short message.

[0114] As another example, each UE may transmit the DoL to the central server (or base station) via the physical uplink control channel (PUCCH), and after the central server collects the DoL, the DoL information for the entire model may be rebroadcast via the physical downlink control channel (PDCCH). As another example, since the central server is aware of the UEs participating in the FL, it may communicate with the designated UEs to transmit the above-described information, and may not be limited to a specific embodiment. That is, for each UE, the DoL of the local model may be transmitted to at least one UE participating in the FL. Thereafter, in the candidate IID distance reporting process (1022), each UE may generate a preliminary DoL by reflecting the data state information (DSI) for its own data set to the transmitted DoL. Thereafter, each UE may calculate a candidate IID distance (CID) for the generated preliminary DoL and transmit it to the central server (or base station). The central server can derive the difference between the IID distance for the DoL calculated in the previous diffusion round and the received candidate IID distance, and perform an evaluation of the UEs. The central server can then organize an auction based on the evaluation of the UEs, as described above.

[0115] In the diffusion configuration (1023) process, the next trainer UE may be determined based on an auction, and scheduling for local model exchange may be performed. Specifically, the central server may determine a bidding price based on the evaluation of UEs and perform a winner selection algorithm with channel state information (CSI) collected from the cellular mobile communication system. For example, Table 1 below may be an example of a winner selection algorithm. However, the winner selection algorithm may not be limited to Table 1 below.

[0116] Specifically, the central server can calculate diffusion efficiency based on the following mathematical expression 5 by utilizing the bid price and CSI.

[0117]

[0118] [Equation 5]

[0119]

[0120] Here, the bid price can be as shown in the following mathematical formula 6, and in mathematical formula 5 can be the same as mathematical expression 7. For example, may be the bandwidth required for model transmission that can be produced via CSI.

[0121] [Equation 6]

[0122]

[0123]

[0124] [Equation 7]

[0125]

[0126]

[0127] In mathematical formula 7 is a UE that trains a local model m in (k-1) diffusion rounds. UE trained in k diffusion rounds The signal-to-noise ratio (SNR) between the two can be the same as the CSI information.

[0128] For example, the central server can utilize the calculated diffusion efficiency to calculate the optimal matching between the current UE and the next learning UE through the Hungarian algorithm (Kuhn-Munkres algorithm). The central server can then perform diffusion scheduling using the calculated matching and a basic radio resource allocation algorithm. The model transmission (1024) process and the local training process may be processes in which local models are transmitted and received between UEs based on the central server's scheduling, and learning is performed based on the local models received by each UE.

[0129] For example, DoL broadcasting (1021), candidate IID distance reporting (1022), and diffusion configuration (1023) are very small control message transmissions, so they do not use a lot of transmission resources and communication costs can be very low. However, model transmission (1024) may use a lot of transmission resources because local model transmission and reception are performed. Considering the above, a method to minimize the number of diffusion rounds may be required. As an example, Table 2 below may be an algorithm for the overall operation process of FedDif that minimizes the number of diffusion rounds, but may not be limited thereto.

[0130]

[0131] FIG. 11 is a flowchart illustrating a communication-efficient AI model assurance method for improving the performance of a federated learning system applicable to the present disclosure. Referring to FIG. 11, a user equipment (UE) for AI model diffusion in a federated learning (FL) system can receive a global model copy from a central server (S1110). Thereafter, the UE can build a local model based on the global model copy and perform learning on the local model using data it has acquired to obtain a first local model (S1120). In other words, the first local model may refer to a local model on which the UE has performed learning using data it has acquired. Thereafter, the UE can update the first local model to a second local model through local learning based on local diffusion (S1130). For example, if the UE is selected as the next learning UE based on the local diffusion process, the first local model and the second local model may be identical. On the other hand, if the UE is not selected as the next learning UE and acquires a local model from another UE, the first local model may be updated to a second local model by reflecting the local model of the other UE. That is, the second local model may be a local model learned by reflecting the local model of the next learning UE. Thereafter, the UE may transmit the second local model to the central server (S1140). Here, as an example, the UE may include a memory, a transceiver, and a control unit that controls the memory and the transceiver, and the control unit may perform the above-described operation. As an example, the UE may be a UE participating in federated learning based on the central server. In addition, if local learning is performed in the first diffusion round, the UE may transmit a Degree of Learning (DoL) to a plurality of UEs participating in federated learning based on the central server.Here, DoL can represent the cumulative data distribution learned through data acquired based on the local model, as described above.

[0132] As a specific example, the DoL may be transmitted to multiple UEs participating in federated learning based on a broadcasting method, or may be transmitted to a central server via an uplink control channel and then transmitted by the central server to multiple UEs participating in federated learning, as described above.

[0133] Thereafter, the UE can receive DoLs from each of the multiple UEs participating in the federated learning, and generate a preliminary DoL by reflecting the data set state information (DSI) of the first local model in each DoL. Furthermore, the UE can derive a first candidate IID distance based on the preliminary DoL and transmit it to the central server.

[0134] The central server can derive bid price information by calculating the difference between the second candidate IID derived in the second diffusion round and the first candidate IID. The second diffusion round may be a diffusion round preceding the first diffusion round.

[0135] The central server can derive bid price information for each of the multiple UEs participating in federated learning and determine the next learning UE through UE evaluation. For example, the central server can further consider channel state information (CSI) for each of the multiple UEs and determine the next learning UE through UE evaluation, as described above.

[0136] Here, if the UE is determined to be the next learning UE, the UE can transmit the first local model to multiple UEs participating in federated learning based on a device-to-device (D2D) scheme. Conversely, if the UE is not determined to be the next learning UE, the UE can receive the local model of the next learning UE from the next learning UE based on a D2D scheme. Thereafter, the UE can perform learning based on the local model of the next learning UE to acquire the second local model.

[0137] Figure 12 is a flowchart illustrating a communication-efficient artificial intelligence model diffusion method for improving the performance of a federated learning system applicable to the present disclosure.

[0138] Referring to FIG. 12, in a federated learning (FL) system, a central server for artificial intelligence model diffusion can transmit a copy of a global model to multiple UEs participating in federated learning (S1210). Here, each of the multiple UEs can build a local model based on the copy of the global model, and perform learning on the local model through data acquired by each of the multiple UEs.

[0139] Thereafter, when local learning is performed based on local diffusion, the central server can receive candidate IID distances from each of a plurality of UEs based on local learning to determine a next learning UE. (S1220) Thereafter, the central server can transmit scheduling information based on the determined next learning UE to a plurality of UEs participating in federated learning. (S1230) Thereafter, the central server can obtain a local model learned based on local learning from a plurality of UEs to update a global model. (S1240) Here, the central server can include a memory, a transceiver, and a control unit that controls the memory and the transceiver, and the control unit can perform the above-described operation.

[0140] For example, the central server can obtain candidate IID distances from each of multiple UEs participating in federated learning to derive bid price information, and perform UE evaluation by considering the bid price information and CSI to determine the next learning UE, as described above.

[0141] Here, the next learning UE can transmit its local model to multiple UEs participating in federated learning based on the D2D method. The central server can then obtain updated local models from each of the multiple UEs participating in federated learning during the global aggregation process and update the global model.

[0142] As described above, according to one embodiment of the present invention, a new diffusion strategy (Federated Diffusion, FedDif) for a machine learning (ML) model is proposed to address the weight divergence and privacy leakage issues caused by non-IID data in a federated learning system. In FedDif, a user terminal propagates a local model to neighboring user terminals via device-to-device (D2D) communication. This FedDif allows the local model to experience different distributions before parameter aggregation. Furthermore, it is theoretically demonstrated that FedDif can avoid the weight dispersion problem. Furthermore, a communication-efficient diffusion strategy for ML models is proposed, which can determine the trade-off between learning performance and communication costs based on auction theory. Performance evaluation results show that FedDif improves the test accuracy of the global model by 10.37% compared to a baseline FL with non-IID settings. Furthermore, FedDif reduces the number of consumed subframes by 1.28 to 2.85 times compared to state-of-the-art methods excluding model compression. FedDif also improved the number of transmitted models by a factor of 1.43 to 2.67 compared to state-of-the-art methods.

[0143] The embodiments of the present invention described above may be implemented through various means. For example, the embodiments of the present invention may be implemented through hardware, firmware, software, or a combination thereof.

[0144] In the case of hardware implementation, the method according to embodiments of the present invention may be implemented by one or more ASICs (Application Specific Integrated Circuits), DSPs (Digital Signal Processors), DSPDs (Digital Signal Processing Devices), PLDs (Programmable Logic Devices), FPGAs (Field Programmable Gate Arrays), processors, controllers, microcontrollers, microprocessors, etc.

[0145] When implemented using firmware or software, the methods according to embodiments of the present invention may be implemented in the form of modules, procedures, or functions that perform the functions or operations described above. The software code may be stored in a memory unit and executed by a processor. The memory unit may be located within or outside the processor and may exchange data with the processor using various known means.

[0146] The detailed description of the preferred embodiments of the present invention disclosed as described above has been provided to enable those skilled in the art to implement and practice the present invention. While the present invention has been described above with reference to preferred embodiments, it will be understood by those skilled in the art that various modifications and variations can be made therein without departing from the spirit and scope of the present invention as set forth in the claims below. Accordingly, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein. Furthermore, while the preferred embodiments of the present specification have been illustrated and described above, the present specification is not limited to the specific embodiments described above, and various modifications may be made by those skilled in the art to which the present invention pertains without departing from the gist of the present specification as claimed in the claims. Furthermore, such modifications should not be individually understood from the technical spirit or perspective of the present specification.

[0147] In addition, both the product invention and the method invention are described in the specification, and the description of both inventions may be applied supplementarily as needed.

[0148]

[0149] [Explanation of symbols]

[0150] 210: Device

[0151] 211: Control unit of the device

[0152] 212: Transmitter and receiver of the device

[0153] 213: Device memory

[0154] 220: Device

[0155] 221: Control unit of the device

[0156] 222: Transmitter and receiver of the device

[0157] 223: Device memory

[0158] 310: Artificial Intelligence Model

[0159] 320: User Equipment (UE)

[0160] 330: Central Server

[0161] 340: User data

Claims

1. In a method of operating a user equipment (UE) for spreading an artificial intelligence model in a federated learning (FL) system, Step of receiving a copy of the global model from the central server; A step of constructing a local model based on the above global model copy and performing learning on the local model using the acquired data to obtain a first local model; A step of updating the first local model to a second local model through local learning based on local diffusion; and A terminal operation method comprising a step of transmitting the second local model to the central server.

2. In paragraph 1, A terminal operation method, wherein the UE is a UE that participates in the federated learning based on the central server.

3. In paragraph 2, When the local learning is performed in the first diffusion round, the UE transmits the DoL (Degree of Learning) to multiple UEs participating in the federated learning based on the central server. A terminal operation method in which the above DoL represents a cumulative data distribution learned through the acquired data based on the above local model.

4. In paragraph 3, A terminal operation method, wherein the above DoL is transmitted to the plurality of UEs participating in the federated learning based on a broadcasting method.

5. In paragraph 3, The above DoL is transmitted to the central server via the uplink control channel, A terminal operation method, wherein the central server transmits the DoL to the plurality of UEs participating in the federated learning through a downlink control channel.

6. In paragraph 3, The UE receives a DoL from each of the plurality of UEs participating in the federated learning, Generate a preliminary DoL by reflecting the data set state information (DSI) of the first local model in each of the above DoLs, and A terminal operation method for deriving a first candidate IID (Independent and Identically Distributed) distance based on the above preliminary DoL and transmitting the same to the central server.

7. In paragraph 6, The central server derives bid price information by deriving a difference value between the second candidate IID derived in the second diffusion round and the first candidate IID, wherein the second diffusion round is a diffusion round prior to the first diffusion round, A terminal operation method in which the central server derives bid price information for each of the plurality of UEs participating in the federated learning and determines the next learning UE through UE evaluation.

8. In paragraph 7, A terminal operation method, wherein the central server further considers channel state information (CSI) for each of the plurality of UEs and determines the next learning UE through UE evaluation.

9. In paragraph 8, A terminal operation method in which, when the UE is determined to be the next learning UE, the UE transmits the first local model to the plurality of UEs participating in the federated learning based on a D2D (device to device) method.

10. In paragraph 8, A terminal operation method in which, when the UE receives a local model of the next learning UE from the next learning UE based on a D2D method, the UE performs learning based on the local model of the next learning UE to obtain the second local model.

11. In a method of operating a central server for spreading an artificial intelligence model in a federated learning (FL) system, A step of transmitting a copy of a global model to a plurality of UEs participating in the federated learning, wherein each of the plurality of UEs builds a local model based on the copy of the global model, and performs learning on the local model through data acquired by each of the plurality of UEs; A step of receiving candidate IID distances from each of a plurality of UEs based on local learning and determining the next learning UE; A step of transmitting scheduling information based on the determined next learning UE to the plurality of UEs participating in the federated learning; and A central server operation method comprising a step of obtaining a local model learned based on the local learning from the plurality of UEs and updating the global model.

12. In a user equipment (UE) that performs artificial intelligence model diffusion in a federated learning (FL) system, memory; Transmitter and receiver; and Including a control unit that controls the above memory and the above transmitter / receiver, The above control unit, Receive a copy of the global model from the central server, A local model is built based on the above global model copy, and a first local model is obtained by performing training on the local model using the acquired data. Update the first local model to a second local model through local learning based on local diffusion, and A terminal that transmits the second local model to the central server.

13. In paragraph 12, The above UE is a UE that participates in the federated learning based on the central server, When the local learning is performed in the first diffusion round, the UE transmits the DoL (Degree of Learning) to multiple UEs participating in the federated learning based on the central server. The above DoL is a terminal that represents a cumulative data distribution learned through the acquired data based on the local model.

14. In paragraph 13, The above DoL is transmitted to the plurality of UEs participating in the federated learning based on a broadcasting method.

15. In paragraph 13, The above DoL is transmitted to the central server via the uplink control channel, A terminal in which the central server transmits the DoL to the plurality of UEs participating in the federated learning through a downlink control channel.

16. In paragraph 13, The UE receives a DoL from each of the plurality of UEs participating in the federated learning, Generate a preliminary DoL by reflecting the data set state information (DSI) of the first local model in each of the above DoLs, and A terminal that derives a first candidate IID (Independent and Identically Distributed) distance based on the above preliminary DoL and transmits it to the central server.

17. In paragraph 16, The central server derives bid price information by deriving a difference value between the second candidate IID derived in the second diffusion round and the first candidate IID, wherein the second diffusion round is a diffusion round prior to the first diffusion round, The central server derives bid price information for each of the plurality of UEs participating in the federated learning, and performs UE evaluation using the bid price information and channel state information (CSI) for each of the plurality of UEs to determine the next learning UE.

18. In paragraph 17, A terminal that transmits the first local model to the plurality of UEs participating in the federated learning based on a D2D (device to device) method when the UE is determined to be the next learning UE.

19. In paragraph 17, A terminal in which, when the UE receives a local model of the next learning UE from the next learning UE based on a D2D method, the UE performs learning based on the local model of the next learning UE to obtain the second local model.

20. In a central server that performs artificial intelligence model diffusion in a federated learning (FL) system, memory; Transmitter and receiver; and Including a control unit that controls the above memory and the above transmitter / receiver, The above control unit Transmitting a copy of the global model to a plurality of UEs participating in the federated learning, wherein each of the plurality of UEs builds a local model based on the copy of the global model, and performs learning on the local model through data acquired by each of the plurality of UEs. Based on local learning, the next learning UE is determined by receiving candidate IID distances from each of multiple UEs, Transmit scheduling information based on the determined next learning UE to the plurality of UEs participating in the federated learning, and A central server that obtains a local model learned based on the local learning from the plurality of UEs and updates the global model.

Citation Information

Patent Citations

  • Display panel and electronic apparatus comprising the same

    KR1020240083957A

  • Technique to perform neural network architecture search with federated learning

    US20210374502A1

  • Device and method for signal transmission in wireless communication system

    WO2022124729A1

  • KR20230066020A