Communication related to federated learning

Abstract

One disclosure of the present specification provides a method. The method may comprise the steps of: receiving, from a first network entity, a notification message related to VFL training; and transmitting, to a second network entity, a request message related to the VFL training.

Description

Communication related to federated learning

[0001] This specification relates to mobile communication.

[0002] 3GPP (3rd Generation Partnership Project) LTE (Long-Term Evolution) is a technology designed to enable high-speed packet communication. Many methods have been proposed to achieve LTE goals, such as reducing costs for users and operators, improving service quality, expanding coverage, and increasing system capacity. As high-level requirements, 3GPP LTE demands reduced cost per bit, improved service availability, flexible use of frequency bands, a simple structure, open interfaces, and appropriate power consumption of terminals.

[0003] Work has begun at the ITU (International Telecommunication Union) and 3GPP to develop requirements and specifications for New Radio (NR) systems. 3GPP must identify and develop the technical components necessary to successfully standardize NR in a timely manner, satisfying both urgent market demands and the longer-term requirements presented by the ITU-R (ITU Radio communication sector) IMT (International Mobile Telecommunications)-2020 process. Furthermore, NR must be able to utilize any spectrum band up to at least 100 GHz so that it can be used for wireless communication even in the distant future.

[0004] NR targets a single technical framework that covers all deployment, usage, and requirements, including eMBB (enhanced Mobile Broadband), mMTC (massive Machine Type-Communications), and URLLC (Ultra-Reliable and Low Latency Communications). NR must be forward compatible by nature.

[0005] Communication based on Federated Learning (FL) is being discussed. In the case of Vertical Federated Learning (VFL), unlike Horizontal Federated Learning (HFL), ML models are not shared between the server and the client. For example, if client re-selection is required during the VFL process, the newly selected client may not possess a trained ML model. Consequently, according to conventional technology, there is a problem that VFL is not performed efficiently. For instance, the performance of VFL may be degraded by the new client, and there is a problem that the new client must undergo iterative training. Since training is required whenever the client changes, there is a problem that it takes a long time for the ML model to be used for inference.

[0006] According to one embodiment of the present specification, a method is provided. The method may include the steps of: receiving a notification message related to VFL training from a first network entity; and transmitting a request message related to VFL training to a second network entity.

[0007] According to one embodiment, a device for implementing the above method is provided.

[0008] According to one embodiment of the present specification, a method is provided. The method may include the steps of: receiving a request message related to VFL training from a server related to VFL; and obtaining information related to an ML model or an ML model.

[0009] According to one embodiment, a device for implementing the above method is provided.

[0010] FIG. 1 shows an example of a communication system to which the implementation of the present specification is applied.

[0011] FIG. 2 shows an example of a wireless device to which the implementation of the present specification applies.

[0012] FIG. 3 shows an example of a UE to which the implementation of the present specification applies.

[0013] FIG. 4 shows an example of a 5G system structure to which the implementation of the present specification is applied.

[0014] Figure 5 shows an example of ML model storage within ADRF.

[0015] Figure 6 shows an example of a procedure for ML model retrieval from ADRF.

[0016] FIGS. 7a and FIGS. 7b illustrate a first example of a procedure according to one embodiment of the disclosure of the present specification.

[0017] FIGS. 8a and FIGS. 8b illustrate a second example of a procedure according to one embodiment of the disclosure of the present specification.

[0018] FIG. 9 illustrates an example of operations according to one embodiment of the disclosure of the present specification.

[0019] The following techniques, devices, and systems may be applied to various wireless multiple access systems. Examples of multiple access systems include Code Division Multiple Access (CDMA) systems, Frequency Division Multiple Access (FDMA) systems, Time Division Multiple Access (TDMA) systems, Orthogonal Frequency Division Multiple Access (OFDMA) systems, Single Carrier Frequency Division Multiple Access (SC-FDMA) systems, and Multi-Carrier Frequency Division Multiple Access (MC-FDMA) systems. CDMA may be implemented through wireless technologies such as Universal Terrestrial Radio Access (UTRA) or CDMA2000. TDMA may be implemented through wireless technologies such as Global System for Mobile Communications (GSM), General Packet Radio Service (GPRS), or Enhanced Data Rates for GSM Evolution (EDGE). OFDMA can be implemented through wireless technologies such as IEEE (Institute of Electrical and Electronics Engineers) 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802.20, or E-UTRA (Evolved UTRA). UTRA is part of UMTS (Universal Mobile Telecommunications System). 3GPP (3rd Generation Partnership Project) LTE (Long-Term Evolution) is part of E-UMTS (Evolved UMTS) using E-UTRA.3GPP LTE uses OFDMA in the downlink (DL) and SC-FDMA in the uplink (UL). Evolutions of 3GPP LTE include LTE-A (Advanced), LTE-A Pro, and / or 5G NR (New Radio).

[0020] For convenience of explanation, the implementation of this specification is described primarily in relation to 3GPP-based wireless communication systems. However, the technical characteristics of this specification are not limited thereto. For example, the following detailed description is provided based on a mobile communication system corresponding to a 3GPP-based wireless communication system, but aspects of this specification that are not limited to 3GPP-based wireless communication systems may be applied to other mobile communication systems.

[0021] For terms and technologies used in this specification that are not specifically described, reference may be made to wireless communication standard documents published prior to this specification.

[0022] In this specification, "A or B" may mean "only A," "only B," or "both A and B." Alternatively, in this specification, "A or B" may be interpreted as "A and / or B." For example, in this specification, "A, B or C" may mean "only A," "only B," "only C," or "any combination of A, B and C."

[0023] A slash ( / ) or a comma used in this specification may mean "and / or." For example, "A / B" may mean "A and / or B." Accordingly, "A / B" may mean "only A," "only B," or "both A and B." For example, "A, B, C" may mean "A, B or C."

[0024] In this specification, "at least one of A and B" may mean "only A," "only B," or "both A and B." Additionally, in this specification, the expressions "at least one of A or B" or "at least one of A and / or B" may be interpreted as synonymous with "at least one of A and B."

[0025] Additionally, in this specification, "at least one of A, B and C" may mean "only A," "only B," "only C," or "any combination of A, B and C." Furthermore, "at least one of A, B or C" or "at least one of A, B and / or C" may mean "at least one of A, B and C."

[0026] Additionally, parentheses used in this specification may mean "for example." Specifically, when indicated as "control information (PDCCH)," "PDCCH" may be proposed as an example of "control information." In other words, "control information" in this specification is not limited to "PDCCH," and "PDCCH" may be proposed as an example of "control information." Furthermore, even when indicated as "control information (i.e., PDCCH)," "PDCCH" may be proposed as an example of "control information."

[0027] Technical features described individually within a single drawing in this specification may be implemented individually or simultaneously.

[0028] Although not limited thereto, the various descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this specification may be applied to various fields where wireless communication and / or connectivity between devices (e.g., 5G) is required.

[0029] The present specification will be described in more detail below with reference to the drawings. In the following drawings and / or description, the same reference numerals may refer to the same or corresponding hardware blocks, software blocks, and / or function blocks unless otherwise indicated.

[0030] FIG. 1 shows an example of a communication system to which the implementation of the present specification is applied.

[0031] The 5G usage scenario shown in FIG. 1 is merely an example, and the technical features of this specification may be applied to other 5G usage scenarios not shown in FIG. 1.

[0032] The three main requirement categories for 5G are (1) enhanced Mobile BroadBand (eMBB) category, (2) massive Machine Type Communication (mMTC) category, and (3) Ultra-Reliable and Low Latency Communications (URLLC) category.

[0033] Referring to FIG. 1, the communication system (1) includes wireless devices (100a to 100f), a base station (BS; 200), and a network (300). FIG. 1 illustrates a 5G network as an example of the network of the communication system (1), but the implementation of the present specification is not limited to a 5G system and may be applied to future communication systems beyond a 5G system.

[0034] The base station (200) and the network (300) can be implemented as wireless devices, and a specific wireless device can operate as a base station / network node in relation to another wireless device.

[0035] Wireless devices (100a to 100f) represent devices that perform communication using Radio Access Technology (RAT) (e.g., 5G NR or LTE) and may also be referred to as communication / wireless / 5G devices. Wireless devices (100a to 100f) may include, but are not limited to, robots (100a), vehicles (100b-1 and 100b-2), eXtended Reality (XR) devices (100c), portable devices (100d), home appliances (100e), Internet-Of-Things (IoT) devices (100f), and Artificial Intelligence (AI) devices / servers (400). For example, vehicles may include vehicles with wireless communication capabilities, autonomous vehicles, and vehicles capable of performing communication between vehicles. Vehicles may include unmanned aerial vehicles (UAVs) (e.g., drones). XR devices may include AR (Augmented Reality) / VR (Virtual Reality) / MR (Mixed Reality) devices and may be implemented in the form of HMDs (Head-Mounted Devices) and HUDs (Head-Up Displays) mounted on vehicles, televisions, smartphones, computers, wearable devices, home appliances, digital signs, vehicles, robots, etc. Portable devices may include smartphones, smart pads, wearable devices (e.g., smartwatches or smart glasses), and computers (e.g., laptops). Home appliances may include TVs, refrigerators, and washing machines. IoT devices may include sensors and smart meters.

[0036] In this specification, wireless devices (100a to 100f) may be referred to as User Equipment (UE). The UE may include, for example, a mobile phone, a smartphone, a laptop computer, a digital broadcasting terminal, a PDA (Personal Digital Assistant), a PMP (Portable Multimedia Player), a navigation system, a slate PC, a tablet PC, an ultrabook, a vehicle, a vehicle with autonomous driving capabilities, a connected car, a UAV, an AI module, a robot, an AR device, a VR device, an MR device, a hologram device, a public safety device, an MTC device, an IoT device, a medical device, a fintech device (or financial device), a security device, a weather / environment device, a 5G service-related device, or a device related to the Fourth Industrial Revolution.

[0037] Wireless devices (100a to 100f) can be connected to a network (300) through a base station (200). AI technology may be applied to the wireless devices (100a to 100f), and the wireless devices (100a to 100f) can be connected to an AI server (400) through the network (300). The network (300) can be configured using a 3G network, a 4G (e.g., LTE) network, a 5G (e.g., NR) network, and a network after 5G. The wireless devices (100a to 100f) may communicate with each other through the base station (200) / network (300), but they may also communicate directly (e.g., sidelink communication) without going through the base station (200) / network (300). For example, vehicles (100b-1, 100b-2) can communicate directly (e.g., V2V (Vehicle-to-Vehicle) / V2X (Vehicle-to-everything) communication). Also, IoT devices (e.g., sensors) can communicate directly with other IoT devices (e.g., sensors) or other wireless devices (100a to 100f).

[0038] Wireless communication / connections (150a, 150b, 150c) can be established between wireless devices (100a to 100f) and / or between wireless devices (100a to 100f) and base station (200) and / or between base station (200). Here, the wireless communication / connections can be established through various RATs (e.g., 5G NR), such as uplink / downlink communication (150a), sidelink communication (150b) (or D2D (Device-To-Device) communication), and communication between base stations (150c) (e.g., relay, IAB (Integrated Access and Backhaul)). Through the wireless communication / connections (150a, 150b, 150c), wireless devices (100a to 100f) and base station (200) can transmit / receive wireless signals to / from each other. For example, wireless communication / connection (150a, 150b, 150c) may transmit / receive signals through various physical channels. To this end, based on various proposals in this specification, at least some of the following may be performed: a process for setting various configuration information for transmitting / receiving wireless signals, a process for various signal processing (e.g., channel encoding / decoding, modulation / demodulation, resource mapping / demapping, etc.), and a resource allocation process.

[0039] NR supports multiple numerologies or subcarrier spacings (SCS) to support various 5G services. For example, when the SCS is 15 kHz, it supports a wide area in traditional cellular bands; when the SCS is 30 kHz / 60 kHz, it supports dense-urban areas, lower latency, and wider carrier bandwidth; and when the SCS is 60 kHz or higher, it supports a bandwidth greater than 24.25 GHz to overcome phase noise.

[0040] The NR frequency band can be defined by two types of frequency ranges (FR1, FR2). The numerical values ​​of the frequency ranges may change. For example, the two types of frequency ranges (FR1, FR2) may be as shown in Table 1 below. For convenience of explanation, among the frequency ranges used in the NR system, FR1 may mean "sub 6GHz range" and FR2 may mean "above 6GHz range" and may be referred to as Millimeter Wave (mmW).

[0041] Frequency Range Definition Frequency Range Subcarrier Spacing FR1 450 MHz - 6000 MHz 15, 30, 60 kHz FR2 24 250 MHz - 52600 MHz 60, 120, 240 kHz

[0042] As described above, the numerical values ​​of the frequency range of the NR system may change. For example, FR1 may include a band of 410 MHz to 7125 MHz as shown in Table 2 below. That is, FR1 may include a frequency band of 6 GHz (or 5850, 5900, 5925 MHz, etc.) or higher. For example, the frequency band of 6 GHz (or 5850, 5900, 5925 MHz, etc.) or higher included within FR1 may include an unlicensed band. The unlicensed band may be used for various purposes, for example, for communication for vehicles (e.g., autonomous driving).

[0043] Frequency Range Definition Frequency Range Subcarrier Spacing FR1 4 10 MHz - 7 125 MHz 15, 30, 60 kHz FR2 24 250 MHz - 5 2600 MHz 60, 120, 240 kHz

[0044] Here, the wireless communication technology implemented in the wireless device of this specification may include LTE, NR, and 6G, as well as NarrowBand IoT (NB-IoT) for low-power communication. For example, NB-IoT technology may be an example of Low Power Wide Area Network (LPWAN) technology and may be implemented according to standards such as LTE Cat NB1 and / or LTE Cat NB2, but is not limited to the names mentioned above. Additionally, or generally, the wireless communication technology implemented in the wireless device of this specification may perform communication based on LTE-M technology. For example, LTE-M technology may be an example of LPWAN technology and may be referred to by various names such as eMTC (enhanced MTC). For example, LTE-M technology may be implemented in at least one of various standards such as 1) LTE CAT 0, 2) LTE Cat M1, 3) LTE Cat M2, 4) LTE non-BL (Non-Bandwidth Limited), 5) LTE-MTC, 6) LTE MTC, and / or 7) LTE M, and is not limited to the names mentioned above. Additionally or generally, wireless communication technology implemented in the wireless device of this specification may include at least one of ZigBee, Bluetooth, and / or LPWAN with consideration for low-power communication, and is not limited to the names mentioned above. For example, ZigBee technology may create Personal Area Networks (PANs) related to small / low-power digital communication based on various standards such as IEEE 802.15.4, and may be referred to by various names.

[0045] FIG. 2 shows an example of a wireless device to which the implementation of the present specification applies.

[0046] In FIG. 2, the first wireless device (100) and / or the second wireless device (200) may be implemented in various forms depending on the use example / service. For example, {the first wireless device (100) and the second wireless device (200)} may correspond to at least one of {wireless devices (100a–100f) and base station (200)}, {wireless devices (100a–100f) and wireless devices (100a–100f)} and / or {base station (200) and base station (200)} of FIG. 1. The first wireless device (100) and / or the second wireless device (200) may be composed of various components, devices / parts and / or modules.

[0047] The first wireless device (100) may include at least one transceiver such as a transceiver (106), at least one processing chip such as a processing chip (101), and / or one or more antennas (108).

[0048] The processing chip (101) may include at least one processor, such as a processor (102), and at least one memory, such as a memory (104). Additionally and / or generally, the memory (104) may be placed outside the processing chip (101).

[0049] The processor (102) can control the memory (104) and / or the transceiver (106) and may be configured to implement the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed herein. For example, the processor (102) may process information within the memory (104) to generate a first information / signal and transmit a wireless signal containing the first information / signal through the transceiver (106). The processor (102) may receive a wireless signal containing a second information / signal through the transceiver (106) and process the second information / signal to store the obtained information in the memory (104).

[0050] Memory (104) may be connected to the processor (102) so as to be operable. Memory (104) may store various types of information and / or instructions. Memory (104) may store firmware and / or software code (105) that implements code, instructions, and / or a set of instructions that perform the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this specification when executed by the processor (102). For example, firmware and / or software code (105) may implement instructions that perform the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this specification when executed by the processor (102). For example, firmware and / or software code (105) may control the processor (102) to perform one or more protocols. For example, firmware and / or software code (105) may control the processor (102) to perform one or more wireless interface protocol layers.

[0051] Here, the processor (102) and memory (104) may be part of a communication modem / circuit / chip designed to implement RAT (e.g., LTE or NR). A transceiver (106) may be connected to the processor (102) and transmit and / or receive a wireless signal through one or more antennas (108). Each transceiver (106) may include a transmitter and / or receiver. The transceiver (106) may be interchangeably used with an RF (Radio Frequency) unit. In this specification, the first wireless device (100) may represent a communication modem / circuit / chip.

[0052] The second wireless device (200) may include at least one transceiver such as a transceiver (206), at least one processing chip such as a processing chip (201), and / or one or more antennas (208).

[0053] The processing chip (201) may include at least one processor, such as a processor (202), and at least one memory, such as a memory (204). Additionally and / or alternatively, the memory (204) may be placed outside the processing chip (201).

[0054] The processor (202) can control the memory (204) and / or the transceiver (206) and may be configured to implement the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed herein. For example, the processor (202) may process information within the memory (204) to generate a third information / signal and transmit a wireless signal containing the third information / signal through the transceiver (206). The processor (202) may receive a wireless signal containing a fourth information / signal through the transceiver (206) and process the fourth information / signal to store the obtained information in the memory (204).

[0055] Memory (204) may be connected to the processor (202) so as to be operable. Memory (204) may store various types of information and / or instructions. Memory (204) may store firmware and / or software code (205) that implements code, instructions, and / or sets of instructions that perform descriptions, functions, procedures, proposals, methods, and / or flowcharts disclosed in this specification when executed by the processor (202). For example, firmware and / or software code (205) may implement instructions that perform descriptions, functions, procedures, proposals, methods, and / or flowcharts disclosed in this specification when executed by the processor (202). For example, firmware and / or software code (205) may control the processor (202) to perform one or more protocols. For example, firmware and / or software code (205) may control the processor (202) to perform one or more wireless interface protocol layers.

[0056] Here, the processor (202) and memory (204) may be part of a communication modem / circuit / chip designed to implement a RAT (e.g., LTE or NR). A transceiver (206) may be connected to the processor (202) and transmit and / or receive a wireless signal through one or more antennas (208). Each transceiver (206) may include a transmitter and / or receiver. The transceiver (206) may be interchangeably used with an RF unit. In this specification, the second wireless device (200) may represent a communication modem / circuit / chip.

[0057] Hereinafter, hardware elements of the wireless device (100, 200) will be described in more detail. Although not limited thereto, one or more protocol layers may be implemented by one or more processors (102, 202). For example, one or more processors (102, 202) may implement one or more layers (e.g., functional layers such as a PHY (physical) layer, a MAC (Media Access Control) layer, an RLC (Radio Link Control) layer, a PDCP (Packet Data Convergence Protocol) layer, an RRC (Radio Resource Control) layer, and an SDAP (Service Data Adaptation Protocol) layer). One or more processors (102, 202) may generate one or more PDUs (Protocol Data Units), one or more SDUs (Service Data Units), messages, control information, data, or information according to the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this specification. One or more processors (102, 202) may generate a signal (e.g., baseband signal) including a PDU, SDU, message, control information, data, or information according to the description, function, procedure, proposal, method, and / or operation flowchart disclosed in this specification and provide it to one or more transceivers (106, 206). One or more processors (102, 202) may receive a signal (e.g., baseband signal) from one or more transceivers (106, 206) and may obtain a PDU, SDU, message, control information, data, or information according to the description, function, procedure, proposal, method, and / or operation flowchart disclosed in this specification.

[0058] One or more processors (102, 202) may be referred to as a controller, a microcontroller, a microprocessor, and / or a microcomputer. One or more processors (102, 202) may be implemented by hardware, firmware, software, and / or a combination thereof. For example, one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), one or more Digital Signal Processing Devices (DSPDs), one or more Programmable Logic Devices (PLDs), and / or one or more Field Programmable Gate Arrays (FPGAs) may be included in one or more processors (102, 202). For example, one or more processors (102, 202) may be composed of a set of communication control processors, application processors (APs), electronic control units (ECUs), central processing units (CPUs), graphic processing units (GPUs), and memory control processors. One or more memories (104, 204) may be connected to one or more processors (102, 202) and may store various forms of data, signals, messages, information, programs, codes, instructions, and / or commands. One or more memories (104, 204) may be composed of Random Access Memory (RAM), Dynamic RAM (DRAM), Read-Only Memory (ROM), Erasable Programmable ROM (EPROM), flash memory, volatile memory, non-volatile memory, hard drive, register, cache memory, computer read storage media, and / or combinations thereof.One or more memories (104, 204) may be located inside and / or outside of one or more processors (102, 202). Additionally, one or more memories (104, 204) may be connected to one or more processors (102, 202) through various technologies such as wired or wireless connections.

[0059] One or more transceivers (106, 206) may transmit user data, control information, wireless signals / channels, etc., as described in the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this specification to one or more other devices. One or more transceivers (106, 206) may receive user data, control information, wireless signals / channels, etc., as described in the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this specification from one or more other devices. For example, one or more transceivers (106, 206) may be connected to one or more processors (102, 202) and may transmit and receive wireless signals. For example, one or more processors (102, 202) may control one or more transceivers (106, 206) to transmit user data, control information, wireless signals, etc., to one or more other devices. Additionally, one or more processors (102, 202) can control one or more transceivers (106, 206) to receive user data, control information, wireless signals, etc. from one or more other devices.

[0060] One or more transceivers (106, 206) may be connected to one or more antennas (108, 208). Additionally and / or generally, one or more transceivers (106, 206) may include one or more antennas (108, 208). One or more transceivers (106, 206) may be configured to transmit and receive user data, control information, wireless signals / channels, etc., mentioned in the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed herein through one or more antennas (108, 208). In this specification, one or more antennas (108, 208) may be a plurality of physical antennas or a plurality of logical antennas (e.g., antenna ports).

[0061] One or more transceivers (106, 206) can convert received user data, control information, wireless signals / channels, etc. from RF band signals to baseband signals in order to process received user data, control information, wireless signals / channels, etc. using one or more processors (102, 202). One or more transceivers (106, 206) can convert processed user data, control information, wireless signals / channels, etc. from baseband signals to RF band signals using one or more processors (102, 202). To this end, one or more transceivers (106, 206) may include (analog) oscillators and / or filters. For example, one or more transceivers (106, 206) can up-convert an OFDM baseband signal into an OFDM signal through an (analog) oscillator and / or filter under the control of one or more processors (102, 202) and transmit the up-converted OFDM signal at a carrier frequency. One or more transceivers (106, 206) can receive an OFDM signal at a carrier frequency and down-convert the OFDM signal into an OFDM baseband signal through an (analog) oscillator and / or filter under the control of one or more processors (102, 202).

[0062] Although not illustrated in FIG. 2, the wireless device (100, 200) may include additional components. The additional components (140) may be configured in various ways depending on the type of the wireless device (100, 200). For example, the additional components (140) may include at least one of a power unit / battery, an input / output (I / O) device (e.g., audio I / O port, video I / O port), a driving unit, and a computing unit. The additional components (140) may be connected to one or more processors (102, 202) through various technologies, such as wired or wireless connections.

[0063] In an implementation of the present specification, the UE may operate as a transmitting device in the uplink and as a receiving device in the downlink. In an implementation of the present specification, the base station may operate as a receiving device in the UL and as a transmitting device in the DL. For technical convenience, it is generally assumed that the first wireless device (100) operates as a UE and the second wireless device (200) operates as a base station. For example, a processor (102) connected to, mounted on, or released to the first wireless device (100) may be configured to perform UE operations according to an implementation of the present specification or to control a transceiver (106) to perform UE operations according to an implementation of the present specification. A processor (202) connected to, mounted on, or released to the second wireless device (200) may be configured to perform base station operations according to an implementation of the present specification or to control a transceiver (206) to perform base station operations according to an implementation of the present specification.

[0064] In this specification, the base station may be referred to as Node B, eNode B, or gNB.

[0065] FIG. 3 shows an example of a UE to which the implementation of the present specification applies.

[0066] Referring to FIG. 3, the UE (100) can correspond to the first wireless device (100) of FIG. 2.

[0067] The UE (100) includes a processor (102), memory (104), transceiver (106), one or more antennas (108), a power management module (141), a battery (142), a display (143), a keypad (144), a SIM (Subscriber Identification Module) card (145), a speaker (146), and a microphone (147).

[0068] The processor (102) may be configured to implement the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed herein. The processor (102) may be configured to control one or more other components of the UE (100) to implement the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed herein. Layers of a wireless interface protocol may be implemented in the processor (102). The processor (102) may include an ASIC, other chipsets, logic circuits, and / or data processing devices. The processor (102) may be an application processor. The processor (102) may include at least one of a DSP, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), and a modem (modulator and demodulator). An example of the processor (102) is the SNAPDRAGON manufactured by Qualcomm®. TM Series processor, EXYNOS made by Samsung® TM Series processors, A Series processors made by Apple®, HELIO made by MediaTek® TM Series processors, ATOM made by Intel® TM It can be found in series processors or corresponding next-generation processors.

[0069] Memory (104) is coupled to the processor (102) so as to be operable and stores various information for operating the processor (102). Memory (104) may include ROM, RAM, flash memory, memory card, storage medium and / or other storage device. When the implementation is implemented in software, the technology described herein may be implemented using modules (e.g., procedures, functions, etc.) that perform the descriptions, functions, procedures, proposals, methods and / or operation flowcharts disclosed herein. Modules may be stored in memory (104) and executed by the processor (102). Memory (104) may be implemented within the processor (102) or outside the processor (102), in which case it may be communicatively coupled to the processor (102) through various methods known in the technology.

[0070] A transceiver (106) is coupled to operate with a processor (102) and transmits and / or receives a wireless signal. The transceiver (106) includes a transmitter and a receiver. The transceiver (106) may include a baseband circuit for processing a wireless frequency signal. The transceiver (106) controls one or more antennas (108) to transmit and / or receive a wireless signal.

[0071] The power management module (141) manages the power of the processor (102) and / or the transceiver (106). The battery (142) supplies power to the power management module (141).

[0072] The display (143) outputs the result processed by the processor (102). The keypad (144) receives input to be used by the processor (102). The keypad (144) can be displayed on the display (143).

[0073] A SIM card (145) is an integrated circuit for securely storing an International Mobile Subscriber Identity (IMSI) and associated keys, and is used to identify and authenticate a subscriber in a mobile device such as a mobile phone or computer. Additionally, contact information can be stored on many SIM cards.

[0074] The speaker (146) outputs sound-related results processed by the processor (102). The microphone (147) receives sound-related input to be used by the processor (102).

[0075] FIG. 4 shows an example of a 5G system structure to which the implementation of the present specification is applied.

[0076] The 5G system (5GS) structure consists of the following network functions (NF).

[0077] - AUSF (Authentication Server Function)

[0078] -AMF (Access and Mobility Management Function)

[0079] - DN (Data Network), for example, operator services, internet access, or third-party services

[0080] - USDF (Unstructured Data Storage Function)

[0081] - NEF (Network Exposure Function)

[0082] - I-NEF (Intermediate NEF)

[0083] - NRF (Network Repository Function)

[0084] - NSSF (Network Slice Selection Function)

[0085] - PCF (Policy Control Function)

[0086] - SMF (Session Management Function)

[0087] - UDM (Unified Data Management)

[0088] - UDR (Unified Data Repository)

[0089] - UPF (User Plane Function)

[0090] - UCMF (UE radio Capability Management Function)

[0091] - AF (Application Function)

[0092] - UE (User Equipment)

[0093] - (R)AN ((Radio) Access Network)

[0094] - 5G-EIR (5G-Equipment Identity Register)

[0095] - NWDAF (Network Data Analytics Function)

[0096] - CHF (CHarging Function)

[0097] 또한, 다음과 같은 네트워크 기능이 고려될 수 있다.

[0098] - N3IWF (Non-3GPP InterWorking Function)

[0099] - TNGF (Trusted Non-3GPP Gateway Function)

[0100] - W-AGF (Wireline Access Gateway Function)

[0101] Figure 4 shows the 5G system structure in a non-roaming case using a reference point representation showing how various network functions interact with each other.

[0102] In Figure 4, UDSF, NEF, and NRF are not described for clarity of the point-to-point diagram. However, all network functions shown can interact with UDSF, UDR, NEF, and NRF as needed.

[0103] For clarity, the connection between UDR and other NFs (e.g., PCF) is not shown in FIG. 4. For clarity, the connection between NWDAF and other NFs (e.g., PCF) is not shown in FIG. 4.

[0104] The 5G system structure includes the following reference points.

[0105] - N1: Reference point between UE and AMF.

[0106] - N2: Reference point between (R)AN and AMF.

[0107] - N3: Reference point between (R)AN and UPF.

[0108] - N4: Reference point between SMF and UPF.

[0109] - N6: Reference point between the UPF and the data network.

[0110] - N9: Reference point between two UPFs.

[0111] The following reference points show the interactions that exist between the NF services of NF.

[0112] - N5: Reference point between PCF and AF.

[0113] - N7: Reference point between SMF and PCF.

[0114] - N8: Reference point between UDM and AMF.

[0115] - N10: Reference point between UDM and SMF.

[0116] - N11: Reference point between AMF and SMF.

[0117] - N12: Reference point between AMF and AUSF.

[0118] - N13: Reference point between UDM and AUSF.

[0119] - N14: Reference point between two AMFs.

[0120] - N15: Reference point between PCF and AMF for non-roaming scenarios, reference point between PCF and AMF of the visited network for roaming scenarios.

[0121] - N16: Reference point between two SMFs (in the case of roaming, between the SMF of the visited network and the SMF of the home network)

[0122] - N22: Reference point between AMF and NSSF.

[0123] In some cases, two NFs may need to be connected to each other to service the UE.

[0124] In 3GPP Release 18, discussions on HFL (Horizontal Federated Learning) took place among NWDAFs.

[0125] 3GPP Release 19 included a study on Vertical Federated Learning (VFL) between NWDAF and AF in the FS_AIML_CN study item (3GPP TR 23.700-84 v2.0.0: Study on Core Network Enhanced Support for Artificial Intelligence (AI) / Machine Learning (ML)). Based on the conclusions of the study, specification work for the AIML_CN Work Item is currently being discussed. (e.g., 3GPP SP-240991: Core Network Enhanced Support for Artificial Intelligence (AI) / Machine Learning (ML))

[0126] The conclusion of Key Issue #2: 5GC Support for Vertical Federated Learning in FS_AIML_CN of TR 23.700-84 is as follows.

[0127] Conclusion on KI#2: 5GC Support for Vertically Federated Learning

[0128] P#2.1: New features related to VFL may include the following:

[0129] P#2.1.1 VFL Server: The VFL Server may be an NWDAF or AF that integrates local training results for local ML model updates during the VFL training process. Additionally, the VFL Server can discover and select VFL clients to tune the VFL training process. In the VFL inference process, the VFL Server can aggregate local inference results from VFL clients to generate the final VFL inference result. The VFL Server can transmit the final VFL inference result to the consumer. There is only one VFL Server for each VFL process.

[0130] P#2.1.2 VFL Client: A VFL client may be an NWDAF or AF that holds a local dataset and performs local training and inference in response to a request from a VFL server. Multiple VFL clients may exist during the VFL training and inference process.

[0131] P#2.1.3 When the VFL server is NWDAF, the VFL process is associated with an analysis ID (e.g., VFL model training for analysis ID, VFL inference for analysis ID).

[0132] P#2.1.4 When the VFL server is an AF, the VFL process is associated with internal AF processes (e.g., VFL model training for internal AF processes, VFL inference for internal AF processes).

[0133] P#2.1.5: 5GC can support vertical federated learning (VFL) in the following scenarios, for example, federated learning techniques that do not exchange or share local datasets or ML models:

[0134] - VFL between NWDAFs within a single PLMN.

[0135] - VFL between AF and NWDAF(s) within a single PLMN.

[0136] NWDAF, the VFL server, can determine whether to use VLF to provide a specific analysis ID based on internal logic and operator policies.

[0137] P#2.1.6: NWDAF and AF may be the only NFs capable of interacting with any of the above VFL functions (e.g., VFL server and VFL client).

[0138] P#2.2: Regarding the registration and discovery of VFL entities, the following examples are described:

[0139] P#2.2.1 An NWDAF that is a VFL client may register with an NRF with an NF profile that includes VFL function information (e.g., VFL function type (e.g., VFL client)). For an AF that is not trusted as a VFL client, the NEF registers based on the NRF's settings within the NF profile information for the AF as specified in S6.2.2.3 of TS 23.288 V19.0.0, and includes VFL function information (VFL function type (e.g., VFL client)) as part of the NF profile information for the AF.

[0140] P#2.2.2 An NWDAF acting as a VFL server can select candidate VFL client(s) and / or AF(s) from the NRF for the VFL training process (via NEF profiles in the case of untrusted AFs).

[0141] P#2.2.3 AF (via NEF in the case of untrusted AF) can select candidate NWDAF(s) as VFL client(s) for the VFL training process in NRF as a VFL server.

[0142] P#2.3: Regarding sample alignment for VFL, the following example is explained:

[0143] P#2.3.1: When an NWDAF acts as a VFL server, the NWDAF triggers sample alignment, can query sample availability from NWDAFs or AFs acting as VFL clients, and creates an interclause between samples.

[0144] P#2.3.2 In the case of a VFL where the NWDAF acts as a VFL server and the AF acts as a VFL client, the NEF may participate in sample alignment based on information mapping (e.g., internal versus external information). Subsequently, the NWDAF determines the final list of participants supporting the samples participating in VFL training as a VFL server.

[0145] P#2.3.3: For AF acting as a VFL server and NWDAF acting as a VFL client, AF performs sample alignment, selects samples to use in the training process, and generates an interclause of samples.

[0146] P#2.3.4: Feature description information can be registered with the NRF or configured locally in the NWDAF or AF. Feature description information can be negotiated between the VFL server and the VFL client when performing feature alignment. Feature alignment is optional in the VFL process.

[0147] P#2.4: Regarding the VFL training process, the following example is explained:

[0148] P#2.4.1: Either NWDAF or AF can act as a VFL server and initiate the VFL training process with VFL clients.

[0149] P#2.4.2: If an untrusted AF participates in the VFL training process, the interaction between that AF and the NWDAF takes place via the NEF. If the NWDAF acts as a VFL server, the NWDAF can receive labels from the AF.

[0150] P#2.4.3: The VFL server assigns an identifier to associate participants during the VFL training and subsequent VFL inference processes, which is associated with distributed ML models during the VFL co-model training process.

[0151] P#2.4.4: The VFL client computes the intermediate results of the local ML models participating in VFL training, and the VFL client reports the intermediate results to AF or NWDAF (acting as the VFL server).

[0152] P#2.4.5: A VFL client can provide intermediate results (e.g., gradient information, loss information) to other VFL clients at the direction of a VFL server.

[0153] P#2.4.6: An AF or NWDAF acting as a VFL server aggregates intermediate results from VFL clients, trains a local model, calculates intermediate results based on a local ML model, and transmits intermediate results to VFL clients participating in the joint VFL training process.

[0154] P#2.4.7: After processing the received intermediate results (which may include convergence reports), the VFL server calculates various intermediate training information (e.g., gradient information, loss information) to update its own local model and the VFL client's model during the VFL training process, sends the updates to the VFL client, and the VFL server / client updates the local ML model based on the received information. The VFL server determines when the VFL training process ends and then notifies the VFL client that the training has ended.

[0155] P#2.5: Regarding the VFL inference process, the following example is explained:

[0156] P#2.5.1: In a scenario where an NWDAF acts as a VFL server, an NF consumer of the analysis can obtain the necessary output based on the VFL inference process generated through the VFL inference process between the VFL server and the corresponding AF or the NWDAF acting as a VFL client:

[0157] - If NWDAF operates as a VFL server, NWDAF can trigger the VFL inference step after receiving an analysis subscription or request and deliver the analysis results to the NF consumer.

[0158] P#2.5.2: For AFs acting as VFL servers, the AF acting as a VFL server can initiate an inference process with the corresponding NWDAF (VFL client). The AF can have the inference start triggered by a 5GC consumer (e.g., an NWDAF containing an AnLF). If the AF is not trusted, the interaction takes place through the NEF.

[0159] P#2.5.3: Before performing VFL inference, the NWDAF acting as the VFL server may determine the corresponding AF and / or NWDAF as VFL clients for the inference process based on the same identifier used in the VFL training process. Relevant details will be specified in the prescriptive phase.

[0160] P#2.5.4: The VFL inference process may be controlled by a set of requirements. For example, the determination of whether all VFL participants associated with the same identifier are required for the VFL inference process may be based on accuracy requirements, VFL signal and load costs, contribution weights of each client, and the temporal availability of VFL participants' outputs.

[0161] P#2.5.5: When performing VFL model performance monitoring, inference data can be used for model retraining, which is consistent with R18.

[0162] As in P#2.4.1, VFL scenarios assume two scenarios: one where NWDAF acts as a server, and another where AF acts as a server. When NWDAF acts as a server, both AF and NWDAF can be included in the Client. When AF acts as a server, only NWDAF can be included in the Client.

[0163] As in P#2.4.2, for untrusted AFs, the interaction between NWDAFs and AFs can be achieved through the Network Exposure Function (NEF).

[0164] According to P#2.4.3, the VFL Server may assign a VFL Correlation ID to a VFL Process. The VFL Correlation ID can be used to associate participants (VFL Server, VFL Client(s)) of VFL training and a subsequent VFL Inference process. This is linked to distributed ML models in VFL's joint model training process.

[0165] According to prior art such as Clauses 6.2B.5, 6.2B.6, and 6.2B.7 of TS 23.288 V19.0.0, an NWDAF Containing MTLF (Model Training logical function) can store a Machine Learning (ML) Model in an Analytics Data Repository Function (ADRF) and / or delete a stored ML Model. Another NWDAF Containing MTLF (Model Training logical function) or NWDAF Containing AnLF (Analytics logical function) can retrieve an ML Model stored in the ADRF.

[0166] Explains an example of storing ML models within ADRF.

[0167] The following drawings are made to illustrate a specific example of the present specification. The names of specific devices or specific signals / messages / fields described in the drawings are presented as examples, and therefore the technical features of the present specification are not limited to the specific names used in the following drawings.

[0168] Figure 5 shows an example of ML model storage within ADRF.

[0169] The procedure illustrated in Fig. 5 can be used to store an ML model in an ADRF using an NWDAF containing an MTLF.

[0170] NWDAF containing MTLF may be, for example, NWDAF containing MTLF.

[0171] 0. NWDAF containing MTLF can decide to save the ML model to ADRF.

[0172] For example, the MTLF included in NWDAF can decide to save the ML model to ADRF based on the MTLF policy.

[0173] 1. An NWDAF containing an MTLF may call the Nadrf_MLModelManagement_StorageRequest service to request that an ADRF store ML model(s) (a set of) models. An NWDAF containing an MTLF may optionally include a list of allowed NF instances for ML model identifiers in the Nadrf_MLModelManagement_StorageRequest as described in TS 33.501 V18.6.0.

[0174] 2. ADRF can download ML models.

[0175] ADRF can locally maintain the association between the ML model identifier, the NF instance ID of the NWDAF including the MTLF, and the list of allowed NF instances (if present in step 1).

[0176] [Optional] If the request includes an ML model address instead of an ML model, ADRF can download the ML model based on the ML model address and save it locally.

[0177] 3. ADRF can send a Nadrf_MLModelManagement_StorageRequest response message containing an ML Model storage result indication to a consumer.

[0178] Referring to Figure 6, an example of ML model retrieval in ADRF is described.

[0179] The following drawings are made to illustrate a specific example of the present specification. The names of specific devices or specific signals / messages / fields described in the drawings are presented as examples, and therefore the technical features of the present specification are not limited to the specific names used in the following drawings.

[0180] Figure 6 shows an example of a procedure for ML model retrieval from ADRF.

[0181] The procedure shown in Fig. 6 can be used by consumers to retrieve ML models from ADRF (NWDAF including MTLF and NWDAF including AnLF).

[0182] 1. An NF consumer may request an ML model from an NWDAF containing an MTLF, and / or send a subscription related to an ML model to an NWDAF containing an MTLF.

[0183] For example, an NWDAF service consumer (e.g., NWDAF containing AnLF or NWDAF containing MTLF) can call the Nnwdaf_MLModelProvision_Subscribe service or the Nnwdaf_MLModelInfo_Request service to subscribe to or request trained ML model(s) associated with analytics ID(s).

[0184] 2. An NWDAF containing an MTLF can determine whether to retrieve a set of ML model(s) associated with an analysis ID(s) from the ADRF.

[0185] If the NWDAF containing the MTLF authorizes the NF consumer to directly retrieve the ML model(s) stored in the ADRF, steps 3 and 4 may be omitted.

[0186] If the NWDAF containing the MTLF needs to retrieve a set of ML models corresponding to the analysis ID requested in step 1 from the ADRF, and the NF consumer is agnostic about the location where the ML models are stored, steps 3 and 4 are performed.

[0187] For reference, regarding NWDAF and ADRF including MTLF certifying NF consumers, TS 33.501 V18.6.0 may be referenced.

[0188] 3. An ADRF service consumer (including MTLF and NWDAF) may request an ML model stored in ADRF by calling the Nadrf_MLModelManagement_RetrievalRequest request service operation (containing a stored transaction identifier or one or more unique ML model identifiers).

[0189] 4. ADRF can verify service consumers (including MTLFs) as described in Annex X.10 of TS 33.501 V18.6.0. If verification is successful, ADRF can send a Nadrf_MLModelManagement_RetrievalRequest response (including the address of the ML model file of the model file stored in ADRF) service operation.

[0190] 5. An NWDAF containing an MTLF may send notifications and / or responses to an NWDAF service consumer, including an Analytics ID tuple and one or more unique ML model identifiers and ML model information tuples. The ML model information may include an ML model file address or an ADRF(Set) ID. The ADRF(Set) ID may be included only if the NWDAF containing the MTLF has authorized the NF consumer to retrieve ML models stored in the ADRF in Step 2. If the ADRF(Set) ID is provided and the NWDAF containing the MTLF has authorized the NF service consumer to retrieve all ML models corresponding to a specific Storage Transaction ID, the Storage Transaction ID may be provided. In other cases, the NWDAF containing the MTLF may provide only the ML model identifier.

[0191] 6. If, in Step 5, the NWDAF service consumer (NWDAF containing AnLF or NWDAF containing MTLF) receives an ADRF (Set) ID (location where the ML model requested in Step 1 is stored) and / or an ML Model provide indicator, the NWDAF service consumer may call the Nadrf_MLModelManagement_RetrievalRequest (stored transaction identifier or one or more unique ML model identifiers) service operation to get the ML model stored in ADRF.

[0192] 7. ADRF can validate the service consumer as described in Annex X.10 of TS 33.501 V18.6.0. If validation is successful, ADRF can send the Nadrf_MLModelManagement_RetrievalRequest response (ML model identifier(s) and model file address(s) stored in ADRF) to the NWDAF service consumer.

[0193] In Rel-18, to determine whether ML Model sharing exists between NWDAFs, the ML Model Interoperability indicator and ML Model Interoperability Information parameter were defined in TS 23.288 V19.0.0.

[0194] - The ML Model Interoperability indicator may include a list of NWDAF providers (vendors) from which ML models can be imported from NWDAFs containing MTLFs. Additionally, the ML Model Interoperability indicator may indicate that NWDAFs containing MTLFs support interoperable ML models requested from vendors included in the list.

[0195] For example, the ML Model Interoperability indicator can represent a list of vendors with whom NWDAF can share ML models. For instance, an NWDAF (e.g., NWDAF1) created by vendor A, which provides NWDAF, may be able to share ML models with vendors A, B, C, and D. In this case, NWDAF1 can display "vendor list: A, B, C, D" as the ML Model Interoperability indicator. For example, in this case, NWDAF1 may also receive shared ML models from vendor B's NWDAF.

[0196] [Optional] ML Model Interoperability Information. This is vendor-specific information that conveys, for example, the requested model file format, model execution environment, etc. Since the encoding, format, and value of ML Model Interoperability Information are vendor-specific, they are not specified. If necessary for sharing purposes, the encoding, format, and value of ML Model Interoperability Information may be agreed upon among vendors.

[0197] For example, ML Model Interoperability information may include information related to an ML Model that can be transmitted or received between compatible vendors. For example, ML Model Interoperability information may be vendor-specific information, such as one or more of the model file type, execution environment, encoding, and / or format. ML Model Interoperability information may be non-standardized information.

[0198] An FL server (e.g., VFL server) can receive an ML Model Interoperability indicator from an existing client NWDAF. Based on the ML Model Interoperability indicator, the FL server (e.g., VFL server) can select an NWDAF to receive an ML Model from the existing client NWDAF. Based on the ML Model Interoperability information, the selected NWDAF can select an ML Model provided by the existing NWDAF.

[0199] At least one of a server, AnLF, MTFL, an NWDAF containing AnLF, and / or an NWDAF containing MTLF may use an ML Model Interoperability indicator when searching for an NWDAF Containing MTLF to share an ML Model via NRF. ML Model Interoperability Information is used when an NWDAF containing AnLF or an NWDAF containing MTLF requests an ML Model from an NWDAF containing MTLF through ML Model Provisioning.

[0200] According to conventional operation, the Server NWDAF may send a request message to the NRF to discover Client NWDAFs. In this case, the NRF may include the Clients' ML Model Interoperability indicators in the information related to the candidate Client NWDAFs. The NRF may send the information related to the candidate Client NWDAFs to the Server NWDAF.

[0201] At the SA2#164 meeting, a discussion was held regarding a general training procedure in which the NWDAF operates as a VFL Server and AFs and NWDAFs operate as Clients; however, due to a lack of time, the relevant CR (S2-2409416) was not approved and the discussion was postponed to the next meeting.

[0202] The following Editor's Note was scheduled to be added to the document.

[0203] Editor's Note: Whether and how to maintain the Vertical Federation Learning process, which includes the dynamic reselection, addition, or removal of VFL client NWDAF(s), is for further study (FFS).

[0204] This Editor's Note implies that there is a need to discuss measures to maintain the VFL process when VFL Clients change during VFL training. Specification work for similar scenarios was also conducted in Release 18's Horizontal Federated Learning (HFL), and the procedure for maintaining Federated Learning is described in TS 23.288 V19.0.0, section 6.2C.2.3. For example, the FL Server NWDAF includes actions that trigger the re-selection, addition, and deletion of FL Client NWDAFs, as well as the discovery of new FL Client NWDAFs and the processes for FL Clients to dynamically join or leave an ongoing FL.

[0205] In the HFL procedure of Rel-18 TS 23.288 V19.0.0, FL Clients may wish to no longer participate in FL. For example, FL Clients may wish to no longer participate in FL due to reasons such as high NF load, time availability changes, or capability changes (e.g., no longer supporting FL). In this case, the Client may notify the FL Server that it no longer wishes to participate. The FL Server may exclude the Client from the training Clients and select another Client. The FL Server may obtain the NF load of the relevant FL Client NWDAF based on NF load analytics or receive the NF load directly from the FL Client NWDAF. Additionally, if the FL Client NWDAF does not send a response to the training within the maximum response time requested by the FL Server, the FL Server may decide to exclude the Client from the FL process based on that information (e.g., that the FL Client NWDAF did not respond to the training within the maximum response time).

[0206] Cases requiring client reselection like this can also occur in VFL. Meanwhile, in the case of conventional HFL, the FL Server NWDAF provides an ML model for training to the FL Client NWDAF. Additionally, in the case of conventional HFL, when the FL Server NWDAF selects an FL Client NWDAF, it can select an FL Client NWDAF that can share the ML model from the FL Server NWDAF. The new FL Client NWDAF can receive the model from the server, perform training immediately, and deliver the results to the server.

[0207] Communication based on Federated Learning (FL) is being discussed. In the case of Vertical Federated Learning (VFL), unlike Horizontal Federated Learning (HFL), ML models are not shared between the server and the client.

[0208] For example, unlike HFL, VFL does not share ML models between the server and the client, and each client has a different local ML model. For instance, in the case of VFL, each VFL client has a different local ML model depending on the feature it is responsible for.

[0209] According to the prior art, there is a problem that VFL is not performed efficiently. For example, if client re-selection is required during the VFL process, the newly selected client may not have a trained ML model. As a result, the performance of VFL may be degraded by the new client, and there is a problem that the new client must undergo iterative training. Since training is required whenever the client changes, there is a problem that it takes a long time for the ML model to be used for inference.

[0210] Specifically, in the case of VFL, even when models cannot be shared between the server and the client, between servers, or between clients, the client or server trains using their respective local ML models and shares only the training results. Therefore, in the case of VFL, there is no need to share ML models between the VFL Server and the VFL Client. There may be instances during the VFL process where client reselection is required (e.g., when a client can no longer participate in FL due to changes in NF load or time availability). Consequently, in such cases, a client newly selected by the VFL Server may not possess a previously trained model. As a result, for the newly selected client (or its ML model) to participate in inference, it is necessary for the newly selected client to train with existing VFL participants using an Initial ML Model. The VFL client may receive the Initial ML Model from the server, or the VFL client may possess the Initial ML Model locally.

[0211] According to conventional technology, a new client can perform VFL with clients that already possess trained models using an Initial ML Model. In this case, since the Initial ML Model is not sufficiently trained, it can affect the overall VFL performance (e.g., accuracy). Furthermore, to use a new client (or the new client's ML model) for inference, iterative training is required until the performance of the new client's ML model converges. For reference, when iterative training is performed on an ML model, there may be a point where the parameter values ​​of the ML model no longer change significantly. In such a case, the performance of the ML model can be said to have converged. The criteria for determining whether the performance of an ML model has converged may vary depending on the algorithm or implementation.

[0212] Accordingly, in the case of VFL, training is required whenever some clients are changed, so there is a problem that it takes a long time for the changed client (or the ML model of the changed client) to be used for inference.

[0213] Therefore, in federated learning (e.g., VFL), there is a need for a method to reduce training time when a client changes and / or to utilize an existing trained ML model to quickly use the changed client (or the ML model of the changed client) for inference.

[0214] In this specification, examples of methods to improve the maintenance of the VFL Process in 3GPP 5GC (or 6G or later next-generation mobile communication technologies) are described. For example, when client reselection occurs, a method for client reselection that can efficiently maintain the VFL Process is proposed.

[0215] The client resection method for VFL proposed in the disclosure of this specification may be composed of a combination of one or more of the following operations / configurations / steps.

[0216] For the procedures and / or messages described in the following examples, conventional procedures / messages may be used, conventional procedures / messages may be extended and used, or new procedures / messages may be defined and used.

[0217] In this specification, Server, Server, VFL server, FL server, and FL server NWDAF may be used as terms with the same meaning.

[0218] This specification focuses on the proposed content. Where the same operation as the prior art is performed, the description of the prior art may be omitted.

[0219] The VFL Correlation ID in this specification may be used as a term with the same meaning as VFL Model Correlation ID, VFL Session ID, etc.

[0220] In this specification, when AF operates as a VFL Server, VFL Clients may include only network entities related to analysis (e.g., NWDAF).

[0221] If a network entity related to analysis (e.g., NWDAF) acts as a VFL Server, VFL Clients may include AF and / or the network entity related to analysis (e.g., NWDAF).

[0222] For reference, in the disclosure of this specification, NWDAF may be an example of a network entity related to analysis. In other words, the scope of the disclosure of this specification is not limited by the designation NWDAF. For example, descriptions and / or actions related to NWDAF may apply to network entities related to analysis.

[0223] When NWDAF operates as a VFL Server, the provisions proposed in this specification may be applied between clients where the VFL Client is NWDAF.

[0224] In this specification, if the AF is an untrusted AF, the AF can communicate with the Core Network through the NEF.

[0225] In this specification, terms such as VFL process, VFL operation, VFL task, and VFL operation may be used interchangeably.

[0226] In this specification, a VFL process may include one or more of VFL training and VFL inference operations. However, the scope of a VFL process in the disclosure of this specification is not limited thereto. For example, in the disclosure of this specification, a VFL process may include various VFL-related operations (e.g., aggregation of results from other VFL clients by a VFL Client).

[0227] The VFL Server can assign a VFL Correlation ID to VFL processes with VFL Clients.

[0228] If the VFL Server is an NWDAF, the VFL Correlation ID assigned by the NWDAF can be unique within the PLMN.

[0229] If the VFL Server is an AF, the VFL Correlation ID may include information related to the Application ID. In this case, the VFL Process can be uniquely identified on the network using only the VFL Correlation ID. Otherwise, duplicate VFL Correlation IDs may be assigned per AF. In this case, the VFL Server and / or client can uniquely identify the corresponding VFL Process by considering both the VFL Correlation ID and the Application ID. Therefore, if the VFL Server is an AF, the Application ID may be transmitted along with the VFL Correlation ID between the VFL Server and the Client. For example, the VFL Server may send the VFL Correlation ID (+Application ID) to the Client. The Client may also send the VFL Correlation ID (+Application ID) to the VFL Server.

[0230] The VFL Correlation ID (+ Application ID) can be stored separately from the Analytics ID or together with the Analytics ID.

[0231] It is assumed that the NWDAF participating in the VFL includes AnLF and / or MTLF.

[0232] The VFL server may provide an Initial ML Model for the VFL to the Client, and each Client NWDAF may have a pre-configured Initial ML Model.

[0233] The VFL Server can trigger the reselection, addition, or deletion of VFL Clients based on information regarding the status of VFL Client NWDAFs during the VFL Process and / or its internal judgment. Additionally, the VFL Server can perform discovery of new VFL Client NWDAFs based on information regarding the status of VFL Client NWDAFs and / or its internal judgment. When client reselection is required in VFL in this manner, some participating clients may be excluded and new clients may join. In this case, training time is required for the new clients to train on the VFL using an Initial ML Model until the performance of the Initial Model converges. To reduce this training and / or training time, a method to utilize an existing trained ML Model for inference may be considered.

[0234] ML models may be interoperable between clients. In some implementations, training time for clients newly joining the VFL may be saved by allowing new clients to use the ML models already trained by removed clients. As such, in some implementations, when the server reselects a client, the server may provide the new client with information that allows it to retrieve the trained ML models. Additionally, in some implementations, the server may prioritize selecting clients that can use existing trained ML models.

[0235] Information that can retrieve a previously trained ML model may include the first example of the disclosure of this specification (e.g., [1] an example using the ID of a Client NWDAF with a trained model), and / or the second example of the disclosure of this specification (e.g., [2] an example using the ID of an ADRF (Set) where a trained model is stored).

[0236] Hereinafter, the first, second, and third examples of the disclosure of this specification are described. In the following examples and in the drawings of FIGS. 7a through 8b, [1] may be a description related to the first example of the disclosure of this specification. [2] may be a description related to the second example of the disclosure of this specification. [3] may be a description related to the third example of the disclosure of this specification.

[0237] 1. First example of the disclosure of this specification

[0238] The first example of the disclosure in this specification is [1] an example using a Client NWDAF ID with a trained model.

[0239] For example, a new client can retrieve a previously trained ML model based on a Client NWDAF ID that has a trained model.

[0240] The VFL server may reselect another client due to the exclusion of a specific client from the VFL process (or for other reasons). In this case, the VFL server may preferentially select a new (interoperable) client from among the candidate clients that can continue to use the model trained by the excluded client.

[0241] To this end, when the Client NWDAF to be excluded requests termination from the server, it may send the ID of the trained ML Model currently in use and / or the ML Model Interoperability indicator to the server. The server may also store the ID of the Client NWDAF to be excluded. For reference, the ML Model Interoperability indicator may be transmitted from the NRF to the VFL Server during the process of discovering the Client NWDAF through the NRF.

[0242] The VFL Server may discover new client NWDAFs responsible for features of existing clients that are excluded from the VFL Process. In this case, when the VFL Server discovers candidates for clients that can receive shared existing trained ML models, it may consider the ML Model Interoperability indicator.

[0243] The VFL Server may request VFL candidate Clients to prepare for VFL training. In this case, the VFL Server may provide the IDs of existing Client NWDAFs to be excluded, and additionally provide the candidate Clients with the ML Model IDs and / or VFL Correlation IDs. Based on this, the VFL Server may inform the candidate Clients of information regarding which ML Models associated with which VFL Correlation IDs can participate in the VFL Process. Each candidate Client may attempt ML Model Provisioning based on the information provided by the Server.

[0244] 2. Second example of the disclosure of this specification

[0245] A second example of the disclosure of this specification is [2] an example of using an ADRF (Set) ID in which a trained model is stored.

[0246] For example, a new client can retrieve a previously trained ML model based on the ADRF (Set) ID where the trained model is stored.

[0247] Each Client NWDAF may also store the trained ML Model in the ADRF according to the procedure of ML Model Storage in ADRF in Clause 6.2B.5 of TS 23.288 V19.0.0 (or the example in Fig. 5).

[0248] In this case, the Client NWDAF to be excluded may request termination as in [1]. Meanwhile, in the second example of the disclosure of this specification, the Client NWDAF to be excluded may save the ML Model trained so far to the ADRF before requesting termination. The Client NWDAF to be excluded may inform the VFL server of the ADRF (Set) ID and Storage Transaction Identifier. Additionally, the Client NWDAF to be excluded may also send the ML Model ID and / or ML Model Interoperability indicator to the VFL server.

[0249] As in the first example of the disclosure of this specification (e.g., [1]), the VFL Server may discover new VFL Client NWDAFs, and the VFL Server may also consider the ML Model Interoperability indicator.

[0250] The VFL Server may request each Client to prepare for training. At this time, the VFL Server provides the Client with the ADRF (Set) ID and Storage Transaction ID for ML Model retrieval, and the VFL Server may additionally provide the Client with the VFL Correlation ID and / or ML Model ID.

[0251] Each Client can check whether it can retriev a trained ML Model through Nadrf_MLModelManagement_RetrievalRequest based on the ADRF (Set) ID and Storage Transaction ID and / or ML Model ID received from the server.

[0252] Hereinafter, examples applicable to the first example of the disclosure of the present specification and / or the second example of the disclosure of the present specification are described.

[0253] According to the first example of the disclosure of this specification (e.g., [1]) and / or the second example of the disclosure of this specification (e.g., [2]), each candidate client may perform the following actions. For example, each candidate client may attempt to retrieve an ML model trained by a client NWDAF to be excluded based on information provided by the server (or information received from the server). And if the server provides an Initial ML model to the candidate client, each candidate client may check whether the model provided by the server and / or an Initial ML model pre-configured internally are also available.

[0254] Each candidate Client NWDAF can provide available ML Model ID information along with its eligibility to participate in VFL in response to a training preparation request.

[0255] Based on the responses of each candidate Client NWDAF, the server can check if there is a Client that can use an existing trained ML Model ID. The server can select Clients that can use an existing trained ML Model ID to participate in the VFL process preferentially.

[0256] 3. Third example of the disclosure of this specification

[0257] The third example of the disclosure in this specification describes an example in which the [3] Server selects the first Client NWDAF by considering the ML Model Interoperability indicator.

[0258] Even if a new VFL Client is reselected according to the first example of the disclosure of this specification (e.g., [1]) and / or the second example of the disclosure of this specification (e.g., [2]), training through an Initial ML Model may be required. For example, the VFL Server may not be able to find a new VFL Client that satisfies both of two conditions (e.g., i) can receive ML Model sharing from the Client to be excluded from the VFL or from ADRF, and ii) can take charge of the VFL features of the Client to be excluded. In this case, training through an Initial ML Model becomes required.

[0259] According to conventional operation, when the server performs NF Discovery through NRF, the server can receive the ML Model Interoperability indicator of each candidate client. Therefore, when the VFL Server selects a client for a specific feature of the VFL, it can select a Client NWDAF in preparation for the case where the client is excluded from the VFL process. For example, based on the ML Model Interoperability indicator, the VFL Server can preferentially select a candidate client as a Client NWDAF if there are alternative clients capable of ML Model sharing.

[0260] In this case, the Server may store alternative Client NWDAFs for each Client NWDAF and ML Model Interoperability indicator information for each (e.g., ML Model Interoperability indicator information for each alternative Client NWDAF).

[0261] In the third example of the disclosure of this specification, the operation after the server selects a client may be the same as in the first example of the disclosure of this specification (e.g., [1]) and / or the second example of the disclosure of this specification (e.g., [2]). Additionally, according to the third example of the disclosure of this specification, the discovery process for new clients in the first example of the disclosure of this specification (e.g., [1]) and / or the second example of the disclosure of this specification (e.g., [2]) may be omitted.

[0262] Hereinafter, with reference to FIGS. 7a and 7b and FIGS. 8a and 8b, examples of procedures in which the first example (e.g., [1]) and / or the second example (e.g., [2]) of the disclosure of the present specification are applied are described. Additionally, the third example (e.g., [3]) of the disclosure of the present specification may also be applied. For example, each of the examples in FIGS. 7a and 7b and FIGS. 8a and 8b may be an example of a procedure in which at least one of the first to third examples of the disclosure of the present specification is combined.

[0263] The following drawings are made to illustrate a specific example of the present specification. The names of specific devices or specific signals / messages / fields described in the drawings are presented as examples, and therefore the technical features of the present specification are not limited to the specific names used in the following drawings.

[0264] FIGS. 7a and FIGS. 7b illustrate a first example of a procedure according to one embodiment of the disclosure of the present specification.

[0265] Figures 7a and 7b are examples related to VFL client reselection and trained ML model retrieval.

[0266] 0. The VFL Server can perform discovery via NRF to select a client for the initial VFL. At this time, the VFL Server may receive ML Model Interoperability indicators for each candidate client.

[0267] When selecting a client for a specific feature of VFL, the VFL Server may prepare for the case where the client is excluded from the VFL process. For example, based on the ML Model Interoperability indicator, the VFL Server may prioritize selecting a candidate client as the Client NWDAF if there are alternative clients capable of ML Model sharing. In this case, the Server may store the ML Model Interoperability indicator information for each of the alternative Client NWDAFs associated with each Client NWDAF.

[0268] 1-2. The VFL server may send request messages or subscription messages related to VFL training to one or more clients (e.g., one or more VFL clients NWDAF). For example, the request message or subscription message related to VFL training may be a VFL Training Request or a VFL Training Subscribe. The request message or subscription message related to VFL training may include one or more of an Analytics ID, a VFL correlation ID, and / or VFL related Info.

[0269] One or more clients (e.g., one or more VFL clients NWDAF) may send response messages or notification messages related to VFL training to the VFL server. Response messages or notification messages related to VFL training may be VFL Training Response or VFL training Notify. For example, response messages or notification messages related to VFL training may include one or more of an analysis ID, VFL correlation IDs, ML model interoperability indicator, information related to termination (e.g., a flag for a termination request), ADRF (set) ID, and / or a Storage Transaction ID.

[0270] During VFL training, the VFL Server may receive requests from some VFL Clients stating that they no longer wish to participate in VFL. For example, a VFL client that wishes to no longer participate in VFL may send a response message or notification message related to VFL training to the VFL Server containing information related to termination (e.g., a flag for a termination request). Examples in FIGS. 7a and 7b show that the leftmost VFL client among the VFL clients performing Step 2 sends information related to termination (e.g., a flag for a termination request).

[0271] In some implementations, at Step 1, the VFL Server may send a request to some Clients to exclude them from the VFL. In this case, the request message or subscription message related to VFL training at Step 1 may include information related to termination.

[0272] If the VFL Server is an AF, message transmission between the AF and the Core Network is done through NEF.

[0273] In some implementations according to the first example of the disclosure of this specification, [1] a VFL Client to be excluded from VFL may inform the VFL server of a trained ML Model ID along with a flag for a Termination Request. For example, a response message or notification message related to VFL training in step 2 may include at least one of information related to termination (e.g., a flag for a Termination Request) and / or a trained ML Model ID.

[0274] In some implementations according to the second example of the disclosure of this specification, [2] a VFL Client to be excluded from the VFL may inform the VFL server of the trained ML Model ID, the ADRF (set) ID where the ML Model is stored, and the associated Storage Transaction Identifier, along with a flag for a Termination Request. For example, a response message or notification message related to the VFL training of step 2 may include at least one of information related to termination (e.g., a flag for a Termination Request), the trained ML Model ID, the ADRF (set) ID where the ML Model is stored, and / or the associated Storage Transaction Identifier.

[0275] A VFL Client to be excluded from VFL may additionally send an ML Model Interoperability indicator for the ML Model to the VFL server. For example, where [1] and / or [2] apply, a VFL Client to be excluded from VFL may include an ML Model Interoperability indicator in response messages or notification messages related to VFL training.

[0276] For reference, since the server already knows the Analytics ID and VFL Correlation ID, response messages or notification messages related to VFL training may not include the Analytics ID and VFL Correlation ID.

[0277] 3. The VFL server can store information received from the terminating client NWDAF.

[0278] For example, the VFL Server may store information received along with a Termination Request from a Client being excluded from the VFL (e.g., terminating client) according to [1] and / or [2].

[0279] In some implementations according to the first example of the disclosure of this specification, [1] the VFL Server may store the Client NWDAF ID and ML Model ID of the client to be excluded.

[0280] In some implementations according to the second example of the disclosure of this specification, [2] the VFL Server stores the ADRF (Set) ID, Storage Transaction Identifier, and ML Model ID received from the Client NWDAF ID to be excluded.

[0281] VFL Clients to be excluded from VFL can also additionally store information related to ML Model Interoperability indicators for ML Models.

[0282] In steps 4 and 5, client discovery and / or client selection may be performed.

[0283] 4-5. The Server can perform NF Discovery with NRF to select a new candidate Client NWDAF.

[0284] For example, the server may send a discovery-related request message (e.g., Nnrf_NFDiscovery Request) to the NRF. The discovery-related request message may include one or more of the following information: service area, NF type (NWDAF), ML Model Interoperability indicator (e.g., may be the same as the ML Model Interoperability indicator received from the client to be excluded), VFL Client capability, and VFL-related information.

[0285] For example, the NRF may send a response message related to discovery (e.g., Nnrf_NFDiscovery Response) to the server. The response message related to discovery may include candidate NWDAF ID(s) and ML Model Interoperability indicators (e.g., ML Model Interoperability indicators corresponding to each of the candidate NWDAF ID(s).

[0286] At this time, the server may select a VFL client capable of ML Model sharing with the VFL client to be excluded as a candidate by using the ML Model Interoperability indicator received from the VFL client to be excluded. For reference, in the disclosure of this specification, the VFL client capable of ML Model sharing may be an NWDAF associated with the vendor included in the ML Model Interoperability indicator transmitted by the VFL client to be excluded. For example, the VFL client to be excluded may be NWDAF 1 of vendor A, and the ML Model Interoperability indicator of the NWDAF may include vendors A and B. In this case, the ML Model Interoperability indicator of NWDAF 2 of vendor B may include vendors A, B, and C. In such a case, the VFL client capable of ML Model sharing with the VFL client to be excluded (e.g., NWDAF 1) may include NWDAF 2.

[0287] Alternatively, the process according to steps 4 and 5 may be omitted, and the server may select a new client candidate from among the candidate clients NWDAF that have already been previously discovered.

[0288] In some implementations according to the third example of the disclosure of this specification, [3] the VFL Server may know the Client(s) that can share ML Model with the Client to be excluded from the VFL. In this case, the NF Discovery process through NRF may be omitted.

[0289] 6. The server may send a request message or subscribe message related to VFL training to one or more candidate clients (one or more clients included in the new VFL clients in the example of FIG. 7a and 7b). For example, the request message or subscribe message related to VFL training may be a VFL training Request or a VFL training Subscribe. The request message or subscribe message related to VFL training may include one or more of an analysis ID, a VFL correlation ID, a (initial) ML model ID, a trained ML model ID, and / or VFL-related information. The request message or subscribe message related to VFL training may also include a previous client NWDAF ID, or an ADRF (set) ID and a storage transaction ID.

[0290] Before actual training, the Server may send a preparation message for training to each Client to select a Client from among the candidate Clients to actually participate in the training. For example, the Server may include information related to preparation (e.g., a preparation flag) in a request message (or subscribe message) related to VFL training in accordance with conventional operations. The request message (or subscribe message) related to VFL training may also include a VFL Correlation ID representing the VFL Process.

[0291] The Server may include the Initial ML Model and Interoperability Information to be used for training by each Client in the request message or subscription message related to VFL training. For example, the Interoperability Information may include VFL Interoperability Information and / or VFL Interoperability Indicator. If a pre-configured ML Model is available, the Client may choose, based on its internal judgment, whether to use the pre-configured Initial ML Model or the ML Model received from the Server.

[0292] In some implementations according to the first example of the disclosure of this specification, [1] the Server may inform one or more candidate clients of the Client NWDAF ID of a previous client (e.g., a client that requested termination) in VFL. For example, the Server may include the Client NWDAF ID of the client that requested termination in a request message (or subscribe message) related to VFL training. The Client NWDAF ID may be information to allow each candidate Client to attempt to retrieve an ML Model from the Client NWDAF that requested termination. The Server may also include the trained ML Model ID and / or VFL Correlation ID used by the previous Client NWDAF along with the previous Client NWDAF ID in the request message (or subscribe message) related to VFL training.

[0293] In some implementations according to the second example of the disclosure of this specification, [2] the Server may inform one or more candidate clients of the ADRF (Set) ID and Storage Transaction Identifier in which the ML Model used by the previous client (e.g., Client NWDAF who requested termination) in VFL is stored. For example, the Server may include the ADRF (Set) ID and Storage Transaction Identifier in a request message (or subscribe message) related to VFL training. Additionally, the Server may also send the trained ML Model ID and / or VFL Correlation ID together to one or more candidate clients.

[0294] In the following, in some implementations according to the first example of the disclosure of this specification, steps 7 and 8 may be performed. If only the second example of the disclosure of this specification or the third example of the disclosure of this specification applies, steps 7 and 8 may be omitted.

[0295] 7. In some implementations according to the first example of the disclosure of this specification, [1] the server may have included a Client NWDAF ID to be excluded in step 6 (e.g., the client ID of the client to be excluded, the client NWDAF ID of the previous client, or the client NWDAF ID of the terminated client). In this case, one or more candidate clients may perform a provisioning request for a trained ML Model based on the Client NWDAF ID and determine whether the ML Model can be retrieved. For example, one or more candidate clients may send a model provisioning request message to the client associated with the Client NWDAF ID based on the Client NWDAF ID. The request message may include the trained ML Model ID and / or VFL Correlation ID received in step 6.

[0296] 8. In some implementations according to the first example of the disclosure of this specification, [1] a prior Client NWDAF (or a excluded client NWDAF, or a terminated client NWDAF) may provide an ML Model to a candidate Client NWDAF. In this case, the prior Client NWDAF may transmit the ID of the trained ML Model and ML Model Information to the candidate Client NWDAF. For example, the prior Client NWDAF may transmit an ML Model Provisioning Response containing the ID of the ML Model and ML Model Information to one or more candidate clients.

[0297] If the request message in Step 7 contains only the VFL Correlation ID without the ML Model ID, the previous VFL Client NWDAF may transmit the ML Model ID used for the VFL Correlation ID to one or more candidate clients. For example, the previous VFL Client NWDAF may have transmitted the trained ML Model ID to the VFL server in Step 2, but the VFL server may not have transmitted the trained ML Model ID to one or more candidate clients in Step 6. In this case, even if the VFL server does not transmit the trained ML Model ID in Step 6, the VFL server may transmit the VFL Correlation ID. In this case, one or more candidate clients may request the ML Model used in the corresponding VFL process from the previous VFL Client NWDAF based on the VFL Correlation ID. Then, the previous VFL Client NWDAF may transmit the ML Model used in the VFL process associated with the VFL Correlation ID to one or more candidate clients.

[0298] ML Model Information may include an ML Model file address or an ADRF (Set) ID.

[0299] For example, if ML Model Information includes an ML Model file address, one or more candidate clients may download an ML model from that address (e.g., ML Model file address) based on the ML Model file address. For example, one or more candidate clients may download an ML model from a previous client via the ML Model file address.

[0300] If the ML Model Information includes an ADRF (Set) ID, a Storage Transaction ID may also be included. For example, in this case, the ML Model Information may include an ADRF (Set) ID and a Storage Transaction ID. Here, since the ML Model ID is transmitted in Step 8, the Storage Transaction ID may be included optionally. In this case, one or more candidate clients may additionally perform the action of receiving the ML model from the ADRF. For example, if the ML Model Information includes an ADRF (Set) ID, one or more candidate clients may perform Step 9.

[0301] 9. In some implementations according to the second example of the disclosure of this specification, [2] the server may have transmitted an ADRF (Set) ID to one or more candidate clients in step 6. In this case, one or more candidate clients may determine whether they can retriev a trained ML Model from ADRF using at least one of the Storage Transaction Identifier, ML Model ID, and / or VFL Correlation ID parameters based on the ADRF (Set) ID. In accordance with the procedure for ML Model retrieval from ADRF in Clause 6.2B.7 of TS 23.288 V19.0.0 (e.g., the procedure according to the example in FIG. 6), if the verification of the Client requesting the ML Model is successful, ADRF may transmit the ML Model (ML Model file address) to the Client.

[0302] 10. Candidate client NWDAFs can determine whether they can participate in VFL based on the server's VFL Preparation request and send a response. For example, one or more candidate client NWDAFs can send response messages or notification messages (e.g., VFL Training Response or VFL Training Notify) related to VFL training to the server.

[0303] For example, a response message or notification message related to VFL training may include one or more pieces of information such as an analysis ID, a VFL correlation ID, information regarding whether to join the VFL, and / or (supported) ML model ID(s).

[0304] A response message or notification message related to VFL training may include a list of supported ML Model IDs (e.g., (supported) ML Model ID(s)). A response message or notification message related to VFL training may also include ML Model ID information available in the Client NWDAF that the server saved in step 3.

[0305] 11. The VFL server can select a client based on a supported ML model ID provided by new clients (e.g., one or more candidate clients).

[0306] For example, the VFL server can make a final selection of a Client NWDAF from among the Clients that can join the VFL. When the VFL server selects a Client NWDAF, it may prioritize selecting a Client that has sent the ML Model ID used by a previously joined VFL Client NWDAF to an ML Model that is available (or supported).

[0307] 12. The VFL server may send request messages or subscription messages related to VFL training to existing client NWDAFs that will perform training, and to new client NWDAFs that will participate in training with client NWDAFs. For example, the request message or subscription message related to VFL training may be a VFL Training Request or a VFL Training Response. The request message or subscription message related to VFL training may include one or more pieces of information such as an analysis ID, a VFL correlation ID, and VFL-related information.

[0308] For example, the VFL server can send training requests to new Client NWDAFs to participate in training with existing Client NWDAFs. For example, in the examples of FIGS. 7a and 7b, an example is illustrated in which the VFL server sends a training request to one client NWDAF.

[0309] 13. The VFL server may send an unsubscribe message related to VFL training or a request message related to VFL training containing a termination flag to Client NWDAF, who decided in step 2 not to participate in VFL training any further. By sending an unsubscribe message or a request message related to VFL training, the VFL server may inform the Client that it has been excluded from training. For example, an unsubscribe message related to VFL training may include one or more of an analysis ID and a VFL correlation ID. For example, a request message related to VFL training may include one or more of an analysis ID, a VFL correlation ID, and information related to termination (e.g., a termination flag).

[0310] The following drawings are made to illustrate a specific example of the present specification. The names of specific devices or specific signals / messages / fields described in the drawings are presented as examples, and therefore the technical features of the present specification are not limited to the specific names used in the following drawings.

[0311] FIGS. 8a and FIGS. 8b illustrate a second example of a procedure according to one embodiment of the disclosure of the present specification.

[0312] Figures 8a and 8b are examples related to VFL client reselection and trained ML model retrieval.

[0313] 0. It can be performed in the same manner as step 0 of FIGS. 7a and FIGS. 7b.

[0314] 1-2. The VFL server may send a subscription message related to VFL training to one or more clients (e.g., one or more VFL clients NWDAF). For example, the subscription message related to VFL training may be Nnwdaf_VFLTraining_Subscribe. The subscription message related to VFL training may include one or more of an Analytics ID, a VFL correlation ID, a Notification Target Address (or a Notification Correlation ID and a Notification Target Address), and / or an iteration number.

[0315] For reference, in the disclosure of this specification, the Notification Correlation ID may be an ID assigned by the Subscribing entity (e.g., the VFL server in Step 1). Subsequently, when another entity (e.g., the entity to which the subscription is targeted) (e.g., one or more clients in Step 1) notifies, it may send a notification message containing the correlation ID to the Subscribing entity. Based on this, the Subscribing entity can determine which subscriber the notification is for. The Notification Target Address may be an address to receive the notification.

[0316] One or more clients (e.g., one or more VFL clients NWDAF) may send a notification message related to VFL training to the VFL server. A notification message related to VFL training may be Nnwdaf_VFLTraining_Notify. For example, a notification message related to VFL training may include one or more of the following: Notification Target Address (or Notification Correlation ID and Notification Target Address), [1][2] Trained Model ID (e.g., Trained ML model ID), [1][2] ML model Interoperability indicator, information related to a termination request (e.g., flag for a termination request), [2] (ADRF (set) ID, and / or Storage Transaction ID).

[0317] During VFL training, the VFL Server may receive requests from some VFL Clients stating that they no longer wish to participate in VFL. For example, a VFL client that wishes to no longer participate in VFL may send a response message or a notification message related to VFL training to the VFL Server containing information related to termination (e.g., a flag for a termination request). Examples in FIGS. 8a and 8b show the VFL client located furthest to the left among the VFL clients performing Step 2 sending information related to termination (e.g., a flag for a termination request).

[0318] In some implementations, at Step 1, the VFL Server may send a request to some Clients to exclude them from VFL. In this case, the subscription message related to VFL training at Step 1 may include information related to termination.

[0319] If the VFL Server is an AF, message transmission between the AF and the Core Network is done through NEF.

[0320] In some implementations according to the first example of the disclosure of this specification, [1] a VFL Client to be excluded from VFL may inform the VFL server of a trained ML Model ID along with a flag for a Termination Request. For example, a notification message related to VFL training in step 2 may include at least one of information related to a termination request (e.g., a flag for a Termination Request) and / or a trained ML Model ID.

[0321] In some implementations according to the second example of the disclosure of this specification, [2] a VFL Client to be excluded from the VFL may inform the VFL server of at least one of the trained ML Model ID, the ADRF (set) ID where the ML Model is stored, and / or the associated Storage Transaction Identifier, along with a flag for a Termination Request. For example, a notification message related to the VFL training of step 2 may include information related to the termination request (e.g., a flag for a Termination Request), the trained ML Model ID, the ADRF (set) ID where the ML Model is stored, and / or the associated Storage Transaction Identifier.

[0322] In some implementations according to the first example of the disclosure of this specification and / or the second example of the disclosure of this specification, [1][2] a VFL Client to be excluded from VFL may additionally send an ML Model Interoperability indicator for an ML Model to the VFL Server. For example, where [1] and / or [2] apply, a VFL Client to be excluded from VFL may include an ML Model Interoperability indicator in a notification message related to VFL training.

[0323] 3. The VFL server can store information received from the terminating client NWDAF.

[0324] For example, the VFL Server may store information received along with a Termination Request from a Client being excluded from the VFL (e.g., terminating client) according to [1] and / or [2].

[0325] In some implementations according to the first example of the disclosure of this specification, [1] the VFL Server may store the Client NWDAF ID and ML Model ID of the client to be excluded.

[0326] In some implementations according to the second example of the disclosure of this specification, [2] the VFL Server stores the ADRF (Set) ID, Storage Transaction Identifier, and ML Model ID received from the Client NWDAF ID to be excluded.

[0327] In some implementations according to the first example of the disclosure of this specification and / or the second example of the disclosure of this specification, [1][2] a VFL Client to be excluded from VFL may additionally store information related to an ML Model Interoperability indicator for an ML Model.

[0328] In steps 4 and 5, client discovery and / or client selection may be performed.

[0329] 4-5. The Server can perform NF Discovery with NRF to select a new candidate Client NWDAF.

[0330] For example, the server may send a search-related request message (e.g., Nnrf_NFDiscovery Request) to the NRF. The search-related request message may include one or more of the following information: service area, NF type (NWDAF), [1] [2] ML Model Interoperability indicator (e.g., may be the same as the ML Model Interoperability indicator received from the client to be excluded), VFL Client capability, and / or VFL interoperability indicator.

[0331] For reference, NFs with the same VFL interoperability indicator can understand each other's feature IDs and VFL configuration information. For example, the previously explained ML Model Interoperability indicator can indicate whether clients can share models with each other. The VFL interoperability indicator can indicate NFs that can perform VFL with each other. For instance, VFL-related operations may only be possible between NFs with the same VFL interoperability indicator.

[0332] For example, the NRF may send a response message related to discovery (e.g., Nnrf_NFDiscovery Response) to the server. The response message related to discovery may include candidate NWDAF ID(s) and ML Model Interoperability indicators (e.g., ML Model Interoperability indicators corresponding to each of the candidate NWDAF ID(s).

[0333] At this time, the server may select a VFL client capable of ML Model sharing with the VFL client to be excluded as a candidate by using the ML Model Interoperability indicator received from the VFL client to be excluded. For reference, in the disclosure of this specification, the VFL client capable of ML Model sharing may be an NWDAF associated with the vendor included in the ML Model Interoperability indicator transmitted by the VFL client to be excluded. For example, the VFL client to be excluded may be NWDAF 1 of vendor A, and the ML Model Interoperability indicator of the NWDAF may include vendors A and B. In this case, the ML Model Interoperability indicator of NWDAF 2 of vendor B may include vendors A, B, and C. In such a case, the VFL client capable of ML Model sharing with the VFL client to be excluded (e.g., NWDAF 1) may include NWDAF 2.

[0334] Alternatively, the process according to steps 4 and 5 may be omitted, and the server may select a new client candidate from among the candidate clients NWDAF that have already been previously discovered.

[0335] For reference, the ML Model Interoperability indicator may be included in a request message related to discovery (e.g., Nnrf_NFDiscovery Request) in some implementations according to the first example of the disclosure of this specification and / or the second example of the disclosure of this specification.

[0336] 6. The server may send a request message related to VFL training (e.g., Nnef_VFLTraining_Request message) to one or more candidate clients (one or more clients included in the new VFL clients in the examples of FIG. 7a and 7b). The request message related to VFL training includes one or more of the analysis ID, VFL correlation ID, (initial) ML model ID, [1][2] trained ML model ID, VFL interoperability information, Feature ID(s), and / or VFL Interoperability indicator, and may include [1] previous client NWDAF ID, or [2] ADRF (Set) ID, Storage Transaction ID.

[0337] Before actual training, the Server may send a request message related to VFL training (e.g., Nnef_VFLTraining_Request message) to each client to prepare for training and to select the client that will actually participate in the training from among the candidate clients. The request message related to VFL training may also include a VFL Correlation ID representing the VFL Process.

[0338] The server may include the initial ML model to be used for training and interoperability information in the request message related to VFL training for each client. If there is a pre-configured ML model, the client may choose, based on its internal judgment, whether to use the pre-configured initial ML model or the ML model received from the server.

[0339] In some implementations according to the first example of the disclosure of this specification, [1] the Server may inform one or more candidate clients of the Client NWDAF ID of the client that requested termination from VFL. For example, the Server may include the Client NWDAF ID of the client that requested termination in a request message related to VFL training. The Client NWDAF ID may be information to allow each candidate Client to attempt to retrieve an ML Model from the Client NWDAF that requested termination. The Server may also include the trained ML Model ID and / or VFL Correlation ID used by the previous Client NWDAF, along with the previous Client NWDAF ID, in the request message related to VFL training.

[0340] In some implementations according to the second example of the disclosure of this specification, [2] the Server may inform one or more candidate clients of the ADRF (Set) ID and Storage Transaction Identifier in which the ML Model used by Client NWDAF, which requested termination from VFL, is stored. For example, the Server may include the ADRF (Set) ID and Storage Transaction Identifier in a request message related to VFL training. Additionally, the Server may also send the trained ML Model ID and / or VFL Correlation ID together to one or more candidate clients.

[0341] In the following, in some implementations according to the first example of the disclosure of this specification, steps 7 and 8 may be performed. If only the second example of the disclosure of this specification or the third example of the disclosure of this specification applies, steps 7 and 8 may be omitted.

[0342] 7. In some implementations according to the first example of the disclosure of this specification, [1] the server may have included a Client NWDAF ID to be excluded in step 6 (e.g., the client ID of the client to be excluded, the client NWDAF ID of the previous client, or the client NWDAF ID of the terminated client). In this case, one or more candidate clients may perform a provisioning request for a trained ML Model based on the Client NWDAF ID and determine whether the ML Model can be retrieved. For example, one or more candidate clients may send a model provisioning request message to the client associated with the Client NWDAF ID based on the Client NWDAF ID. The request message may include the trained ML Model ID and / or VFL Correlation ID received in step 6.

[0343] 8. In some implementations according to the first example of the disclosure of this specification, [1] a prior Client NWDAF (or a excluded client NWDAF, or a terminated client NWDAF) may provide an ML Model to a candidate Client NWDAF. In this case, the prior Client NWDAF may transmit the ID of the trained ML Model and ML Model Information to the candidate Client NWDAF. For example, the prior Client NWDAF may transmit an ML Model Provisioning Response containing the ID of the ML Model and ML Model Information to one or more candidate clients.

[0344] If the request message in Step 7 contains only the VFL Correlation ID without the ML Model ID, the previous VFL Client NWDAF may transmit the ML Model ID used for the VFL Correlation ID to one or more candidate clients. For example, the previous VFL Client NWDAF may have transmitted the trained ML Model ID to the VFL server in Step 2, but the VFL server may not have transmitted the trained ML Model ID to one or more candidate clients in Step 6. In this case, even if the VFL server does not transmit the trained ML Model ID in Step 6, the VFL server may transmit the VFL Correlation ID. In this case, one or more candidate clients may request the ML Model used in the corresponding VFL process from the previous VFL Client NWDAF based on the VFL Correlation ID. Then, the previous VFL Client NWDAF may transmit the ML Model used in the VFL process associated with the VFL Correlation ID to one or more candidate clients.

[0345] ML Model Information may include an ML Model file address or an ADRF (Set) ID.

[0346] For example, if ML Model Information includes an ML Model file address, one or more candidate clients may download an ML model from that address (e.g., ML Model file address) based on the ML Model file address. For example, one or more candidate clients may download an ML model from a previous client via the ML Model file address.

[0347] If the ML Model Information includes an ADRF (Set) ID, a Storage Transaction ID may also be included. For example, the ML Model Information may include an ADRF (Set) ID and a Storage Transaction ID. Here, since the ML Model ID is transmitted in Step 8, the Storage Transaction ID may be included optionally. In this case, one or more candidate clients may additionally perform the action of receiving the ML model from the ADRF. For example, if the ML Model Information includes an ADRF (Set) ID, one or more candidate clients may perform Step 9.

[0348] 9. It can be performed in the same manner as step 9 of FIGS. 7a and FIGS. 7b.

[0349] 10. Candidate client NWDAFs can determine whether they can participate in the VFL based on the server's VFL Preparation request and send a response. For example, one or more candidate client NWDAFs can send a response message related to VFL training (e.g., Nnwdaf_VFLTraining_Request response) to the server.

[0350] For example, a response message related to VFL training may include one or more of the following information: information regarding whether to join VFL, supported VFL interoperability information, a list of samples accepted by the client, and / or (supported) ML model ID(s).

[0351] The response message related to VFL training may include a list of supported ML Model IDs (e.g., (supported) ML Model ID(s)). The response message related to VFL training may also include ML Model ID information available in the Client NWDAF that the server saved in step 3.

[0352] 11. It can be performed in the same manner as step 11 of FIGS. 7a and FIGS. 7b.

[0353] 12. The VFL server may send a subscription message related to VFL training (e.g., Nnwdaf_VFLTraining_Subscribe) to existing client NWDAFs that will perform training and to new client NWDAFs that will participate in training with client NWDAFs. The subscription message related to VFL training may include one or more of the following information: analysis ID, VFL correlation ID, Notification Target Address (or Notification Correlation ID and Notification Target Address), and / or iteration number.

[0354] For example, the VFL server can send training requests via Nnwdaf_VFLTraining_Subscribe to Client NWDAFs that will newly participate in training with existing Client NWDAFs. For example, in the examples of FIGS. 8a and 8b, an example is illustrated in which the VFL server sends a training request to one client NWDAF.

[0355] 13. The VFL server may send an unsubscribe message related to VFL training (e.g., Nnwdaf_VFLTraning_Unsubscribe) to Client NWDAF, who decided in step 2 not to participate in VFL training any further. By sending an unsubscribe message related to VFL training, the VFL server can inform the Client that it has been excluded from training. For example, the unsubscribe message related to VFL training may include a subscription correlation ID.

[0356] For example, the subscription correlation ID may be an ID indicating which subscription is being unsubscribed from when a VFL server or another NF unsubscribes. For instance, in response to a subscriber's subscription, another NF may assign a Subscription Correlation ID. When the subscriber updates or unsubscribes, they may send a message containing the Subscription Correlation ID to another NF.

[0357] In the disclosure of this specification, a specific Client may be excluded from the VFL process during the VFL process, and the VFL server may select a new candidate Client. In this case, according to one embodiment of the disclosure of this specification, the VFL server may efficiently select a Client capable of participating in the VFL process using a previously trained ML Model.

[0358] The following drawings are made to illustrate a specific example of the present specification. The names of specific devices or specific signals / messages / fields described in the drawings are presented as examples, and therefore the technical features of the present specification are not limited to the specific names used in the following drawings.

[0359] FIG. 9 illustrates an example of operations according to one embodiment of the disclosure of the present specification.

[0360] For reference, the procedure illustrated in FIG. 9 is merely an example, and the scope of disclosure of this specification is not limited by the example of FIG. 9.

[0361] For example, regarding the example of FIG. 9, the operations described in the examples of FIG. 1 through 8b may also be applied. For example, even if the operations, contents, etc. are not directly described in the example of FIG. 9, the operations, contents, etc. described in various examples of the disclosure of this specification may be applied.

[0362] The operations illustrated in FIG. 9 are merely examples, and the scope of disclosure of this specification is not limited to the operations illustrated in FIG. 9.

[0363] For reference, in the example of FIG. 9, the server may be a server that supports operations related to FL. For example, the server may be a VFL server that supports operations related to VFL.

[0364] For reference, in the example of FIG. 9, the first network entity may be the first VFL client. For example, the first network entity may be the first VFL client NWDAF. The first network entity may be a client participating in the VFL (or VFL process).

[0365] For reference, in the example of FIG. 9, the second network entity may be the second VFL client. For example, the second network entity may be the second VFL client NWDAF. The second network entity may be a candidate client that intends to newly join the VFL (or VFL process).

[0366] For reference, in the example of FIG. 9, one first network entity and one second network entity are illustrated, but this is merely an example, and the scope of the disclosure of this specification is not limited to one first network entity and one second network entity. For example, the description in FIG. 9 may also apply to one or more first network entities (e.g., one or more VFL clients) and one second network entity (e.g., one or more VFL clients).

[0367] In step (S901), the first network entity can send a notification message to the server.

[0368] For example, the first network entity may send a notification message related to VFL training to the server. The notification message may include information related to a termination request. The notification message may further include i) information about the trained ML model (e.g., ID), or ii) information about the trained ML model (e.g., ID), ADRF information (e.g., ID), and storage transaction information (e.g., ID).

[0369] In some implementations, the notification message may further include an ML model interoperability indicator.

[0370] In some implementations, the server may select at least one candidate client to which an ML model trained by a first network entity can be shared, based on an ML model interoperability indicator. A second network entity may be included in the at least one candidate client.

[0371] In some implementations, the server may send a discovery request message to the NRF that includes an ML model interoperability indicator. The server may receive a discovery response message from the NRF that includes information (e.g., ID) related to one or more candidate clients and an ML model interoperability indicator related to each of the information (e.g., ID) related to one or more candidate clients. Based on the discovery response message, the server may select at least one candidate client that includes a second network entity.

[0372] In step (S902), the server can send a request message to a second network entity.

[0373] For example, the server can send a request message related to VFL training to a second network entity.

[0374] For example, the request message may include i) client information (e.g., ID) of the first network entity and further include at least one of VFL correlation information (e.g., ID) or the trained ML model information (e.g., ID). Or the request message may include ii) ADRF information (e.g., ID) and further include at least one of the trained ML model information (e.g., ID), the VFL correlation information (e.g., ID) or the stored transaction information (e.g., ID).

[0375] For example, i) based on at least one of the above VFL correlation information (e.g., ID) or the above trained ML model information (e.g., ID) and the client information (e.g., ID) of the first network entity, or ii) based on at least one of the above trained ML model information (e.g., ID), the above VFL correlation information (e.g., ID) or the above stored transaction information (e.g., ID) and the above ADRF information (e.g., ID), the second network entity may obtain information related to the ML model or the ML model.

[0376] For example, information related to an ML model may be information related to an ML model trained by a first network entity. For example, information related to an ML model may include an ML Model file address and / or ADRF (Set) information (e.g., ID). If information related to an ML model includes ADRF (Set) information (e.g., ID), information related to an ML model may include ADRF (Set) information (e.g., ID) and Storage Transaction information (e.g., ID).

[0377] For example, the ML model may be an ML model trained by the first network entity.

[0378] In some implementations, based on the fact that the notification message includes trained ML model information (e.g., ID), the request message may include client information (e.g., ID) of the first network entity and may further include at least one of VFL correlation information (e.g., ID) or the trained ML model information (e.g., ID).

[0379] In some implementations, based on the fact that a notification message includes trained ML model information (e.g., ID), ADRF information (e.g., ID), and stored transaction information (e.g., ID), a request message may include the ADRF information (e.g., ID) and further include at least one of the trained ML model information (e.g., ID), the VFL correlation information (e.g., ID), or the stored transaction information (e.g., ID).

[0380] In some implementations, at least one of the above VFL correlation information (e.g., ID) or the above trained ML model information (e.g., ID), and the client information of the first network entity (e.g., ID) may be used for the second network entity to obtain information related to the ML model from the first network entity.

[0381] For example, the second network entity may transmit an ML model provision request message to the first network entity that includes at least one of the VFL correlation information (e.g., ID) or the trained ML model information (e.g., ID). The second network entity may receive a response message from the first network entity that includes the trained ML model information (e.g., ID) and information related to the ML model.

[0382] In some implementations, at least one of the trained ML model information (e.g., ID), the VFL correlation information (e.g., ID), or the stored transaction information (e.g., ID), and the ADRF information (e.g., ID), and a second network entity may be used to retrieve an ML model from the ADRF.

[0383] For example, the second network entity can retrieve an ML model from the ADRF based on the ADRF information (e.g., ID) and the stored transaction information (e.g., ID).

[0384] In some implementations, the server may receive a response message related to the VFL training from a second network entity. For example, the response message may include ML model information (e.g., ID) supported by the second network entity.

[0385] For example, the server may determine whether to select the second network entity as a client of the VFL process based on supported ML model information (e.g., ID).

[0386] For example, based on the fact that the second network entity is selected as a client of the VFL process, the server can send a subscription message related to VFL training to the second network entity.

[0387] For example, the server can send an unsubscribe message related to VFL training to the first network entity.

[0388] The first network entity and the second network entity can support operations related to the client of the VFL processor.

[0389] According to the first example of the disclosure of this specification, a Client NWDAF ID with a trained ML Model may be used. For example, when a Client NWDAF to be excluded from the VFL Process requests termination, it may transmit the ID of the trained ML Model currently in use and / or an ML Model Interoperability indicator to the server. The VFL Server may perform discovery for new VFL Client NWDAFs to be added to the VFL Process. In this case, the VFL Server may consider the ML Model Interoperability indicator when searching for candidates for Clients who can receive sharing of the existing trained ML Model. For example, the VFL Server may request preparation for VFL training from new candidate Client NWDAFs responsible for the features of the existing Client being excluded from the VFL Process. In this case, the VFL Server may transmit the ID of the Client NWDAF of the existing Client being excluded to the candidate Client NWDAFs. The VFL Server may additionally transmit the ML Model ID and / or VFL Correlation ID to the candidate Client NWDAFs.

[0390] According to the second example of the disclosure of this specification, an ADRF (Set) ID in which a trained ML model is stored may be used. For example, a Client NWDAF to be excluded from the VFL Process may store the ML models trained so far in the ADRF before requesting termination. The Client NWDAF may transmit the ADRF (Set) ID and Storage Transaction Identifier to the VFL Server. Additionally, the Client NWDAF may also transmit the ML Model ID and / or ML Model Interoperability indicator to the VFL Server. As in the second example of the disclosure of this specification, the VFL Server may perform discovery for new VFL Client NWDAFs. The VFL Server may request each Client to prepare for training. In this case, the VFL Server may transmit the VFL Correlation ID and Storage Transaction ID along with the ADRF (Set) ID for ML Model retrieval instead of the ID of the Client NWDAF to be excluded. The VFL Server may also transmit the VFL Correlation ID, the ID of the excluded Client NWDAF (e.g., previous client NWDAF ID), and the ML Model ID. Each Client can check whether it can retriev a trained ML Model via a Nadrf_MLModelManagement_RetrievalRequest using the ADRF (Set) ID and Storage Transaction ID and / or ML Model ID received from the server.

[0391] According to the first and / or second examples of the disclosure of this specification, each candidate client may attempt to retrieve an ML model trained by a client NWDAF to be excluded based on information transmitted by the server. And if the server has provided an Initial ML Model to the candidate client, each candidate client may check whether the said Model and a previously internally configured Initial ML Model are available. Each candidate client NWDAF may provide the VFL server with information on available (or supported) ML Model IDs along with whether it is eligible to participate in the VFL in response to a training preparation request. Based on the responses of each candidate client NWDAF, the server may select to prioritize the participation of a client in the VFL process if there is a client that can use an existing trained ML Model ID.

[0392] According to the third example of the disclosure of this specification, the Server may consider an ML Model Interoperability indicator when selecting the initial Client NWDAF. For example, when the VFL Server selects a Client for a specific feature of the VFL, it may select a Client NWDAF in anticipation of the case where the Client is excluded from the VFL Process. For example, based on the ML Model Interoperability indicator, the VFL Server may preferentially select a candidate Client as the Client NWDAF if there are alternative Clients capable of ML Model sharing. In this case, the Server may store alternative Client NWDAFs for each Client NWDAF and information regarding their respective ML Model Interoperability indicators. In the third example of the disclosure of this specification, the operation after the Server selects the client may be the same as in the first example of the disclosure of this specification (e.g., [1]) and / or the second example of the disclosure of this specification (e.g., [2]). Additionally, according to the third example of the disclosure of this specification, the discovery process for new clients in the first example of the disclosure of this specification (e.g., [1]) and / or the second example of the disclosure of this specification (e.g., [2]) may be omitted.

[0393] This specification may have various effects.

[0394] For example, VFL can be effectively supported.

[0395] For example, if client reselection is required in VFL, a previously trained ML model can be used. Consequently, resources and time for iterative training can be reduced, and the model can be rapidly utilized for VFL inference. For instance, resources and time wasted performing iterative training until the performance of the VFL process converges for a new client using an Initial ML Model can be reduced. By using a previously trained ML model for a new client, the ML model can be rapidly utilized for VFL inference.

[0396] For reference, the operation of the terminal (e.g., UE, etc.) described in this specification may be implemented by the device of FIGS. 1 to 3 described above. For example, the terminal may be the first device (100) or the second device (200) of FIG. 2. For example, the operation of the terminal described in this specification may be processed by one or more processors (102 or 202). The operation of the terminal described in this specification may be stored in one or more memories (104 or 204) in the form of an instruction / program (e.g., instruction, executable code) executable by one or more processors (102 or 202). One or more processors (102 or 202) may control one or more memories (104 or 204) and one or more transceivers (105 or 206) and execute the instruction / program stored in one or more memories (104 or 204) to perform the operation of the terminal (e.g., UE) described in the disclosure of this specification.

[0397] Additionally, instructions for performing the operation of the terminal described in the disclosure of this specification may be stored in a non-volatile computer-readable storage medium. The storage medium may be included in one or more memories (104 or 204). And, the instructions recorded in the storage medium may perform the operation of the terminal described in the disclosure of this specification by being executed by one or more processors (102 or 202).

[0398] For reference, the operation of a network node (e.g., AMF, SMF, UPF, AF, VFL server, FL server, NF consumer, consumer, NWDAF including MTLF, ADRF, NEF, NRF, VFL Client NWDAF, etc.) or a base station (e.g., NG-RAN, gNB, RAN, (R)AN, etc.) described in this specification may be implemented by the device of FIGS. 1 to 3, which will be described below. For example, the network node or base station may be the first device (100) or the second device (200) of FIG. 2. For example, the operation of the network node or base station described in this specification may be processed by one or more processors (102 or 202). The operation of the terminal described in this specification may be stored in one or more memories (104 or 204) in the form of an instruction / program (e.g., instruction, executable code) executable by one or more processors (102 or 202). One or more processors (102 or 202) can control one or more memories (104 or 204) and one or more transceivers (106 or 206) and execute instructions / programs stored in one or more memories (104 or 204) to perform the operation of a network node or base station as described in the disclosure of this specification.

[0399] Additionally, instructions for performing the operation of a network node or base station described in the disclosure of this specification may be stored in a non-volatile (or non-transient) computer-readable storage medium. The storage medium may be contained in one or more memories (104 or 204). And, the instructions recorded in the storage medium may perform the operation of a network node or base station described in the disclosure of this specification by being executed by one or more processors (102 or 202).

[0400] Although preferred embodiments have been described by way of example above, the disclosure of this specification is not limited to such specific embodiments, and may be modified, changed, or improved in various forms within the scope of the spirit and claims of this specification.

[0401] In the exemplary system described above, methods are described based on a flowchart as a series of steps or blocks, but are not limited to the order of the described steps, and some steps may occur in a different order or simultaneously with other steps as described above. Furthermore, a person skilled in the art will understand that the steps shown in the flowchart are not exclusive, and that other steps may be included, or that one or more steps of the flowchart may be omitted without affecting the scope of rights.

[0402] The claims described in this specification may be combined in various ways. For example, the technical features of the method claims in this specification may be combined to be implemented as a device, and the technical features of the device claims in this specification may be combined to be implemented as a method. Furthermore, the technical features of the method claims and the technical features of the device claims in this specification may be combined to be implemented as a device, and the technical features of the method claims and the technical features of the device claims in this specification may be combined to be implemented as a method. Other implementations are within the scope of the following claims.

Claims

1. A step of receiving a notification message related to Vertical Federated Learning (VFL) training from a first network entity, The above notification message includes information related to a termination request; and It includes the step of transmitting a request message related to the above VFL training to a second network entity, and A method wherein the above request message comprises: i) client information of the first network entity and further comprises at least one of VFL correlation information or trained Machine Learning (ML) model information; or ii) Analytics Data Repository Function (ADRF) information and further comprises at least one of the trained ML model information, VFL correlation information, or storage transaction information.

2. In Paragraph 1, A method based on the fact that the above notification message includes the above-mentioned trained ML model information, wherein the above request message includes client information of the first network entity and further includes at least one of VFL correlation information or the above-mentioned trained ML model information.

3. In Paragraph 1, A method based on the fact that the above notification message includes the above trained ML model information, the above ADRF information, and the above stored transaction information, wherein the above request message includes the above ADRF information and further includes at least one of the above trained ML model information, the above VFL correlation information, or the above stored transaction information.

4. In Paragraph 1 or 2, At least one of the above VFL correlation information or the above trained ML model information, and the client information of the above first network entity are used by the above second network entity to obtain information related to the ML model from the above first network entity.

5. In Paragraph 1 or 3, A method in which at least one of the above-mentioned trained ML model information, the above-mentioned VFL correlation information, or the above-mentioned stored transaction information, and the above-mentioned ADRF information are used by the second network entity to retrieve an ML model from the ADRF.

6. In any one of paragraphs 1 through 5, The above notification message further includes an ML model interoperability indicator, and Based on the ML model interoperability indicator, the method further includes the step of selecting at least one candidate client to which the ML model trained by the first network entity can be shared, and A method in which the second network entity is included in the at least one candidate client.

7. In paragraph 6, the above-mentioned selecting step is, A step of sending a discovery request message containing an ML model interoperability indicator to a Network Repository Function (NRF); A step of receiving a search response message from the NRF comprising information related to one or more candidate clients and an ML model interoperability indicator related to each of the information related to one or more candidate clients; and A method further comprising the step of selecting at least one candidate client based on the above search response message.

8. In any one of paragraphs 1 through 7, The method further includes the step of receiving a response message related to the VFL training from the second network entity. A method in which the above response message includes ML model information supported by the second network entity.

9. In Paragraph 8, A method further comprising the step of determining whether to select the second network entity as a client of the VFL process based on the supported ML model information.

10. In Paragraph 9, A method further comprising the step of transmitting a subscription message related to VFL training to the second network entity based on the fact that the second network entity is selected as a client of the VFL process.

11. In any one of paragraphs 1 through 10, A method further comprising the step of sending an unsubscribe message related to the VFL training to the first network entity.

12. In any one of paragraphs 1 through 10, A method in which the first network entity and the second network entity support operations related to a client of a VFL processor.

13. In the device, At least one transmitter / receiver; At least one processor; and It includes one or more memories that store instructions and can be connected to operate with one or more processors, and The above at least one processor is: a device adapted to perform a method according to any one of claims 1 to 12.

14. At least one processor; and It includes at least one memory that stores instructions and is operablely electrically connected to at least one processor. The above-mentioned at least one processor is: an apparatus adapted to perform the method according to any one of claims 1 to 12.

15. As a non-transitory computer-readable medium (CRM) recording instructions, The above instructions, when executed by one or more processors, cause the one or more processors to perform a method according to any one of claims 1 through 12, a CRM.

16. A step of receiving a request message related to VFL training from a server related to Vertical Federated Learning (VFL), The above request message comprises: i) client information of a first network entity and further comprises at least one of VFL correlation information or trained Machine Learning (ML) model information; or ii) Analytics Data Repository Function (ADRF) information and further comprises at least one of the trained ML model information, the VFL correlation information, or storage transaction information; and A method comprising the step of obtaining information related to an ML model or an ML model.

17. In Paragraph 16, The above-mentioned acquisition step is, A step of transmitting to the first network entity a request message for providing an ML model, comprising at least one of the VFL correlation information or the trained ML model information; and A method further comprising the step of receiving a response message from the first network entity, the response message including the trained ML model information and information related to the ML model.

18. In Paragraph 16, The above-mentioned acquisition step is, A method further comprising the step of retrieving the ML model from the ADRF based on at least one of the above-mentioned trained ML model information, the above-mentioned VFL correlation information, or the above-mentioned stored transaction information, and the above-mentioned ADRF information.

19. In any one of paragraphs 16 through 18, The method further includes the step of transmitting a response message related to the above VFL training to a server related to the above VFL, and A method in which the above response message includes ML model information supported by the second network entity.

20. In Paragraph 19, A method further comprising the step of receiving a subscription message related to the above VFL training from a server related to the above VFL.

21. In any one of paragraphs 16 through 20, A method in which the first network entity and the second network entity support operations related to a client of a VFL processor.

22. In a device, the device is: One or more transmitters / receivers; One or more processors; and It includes one or more memories that store instructions and can be connected to operate with one or more processors, and The above-mentioned at least one processor is: a device adapted to perform the method according to any one of claims 16 to 21.

Images