hash
stringlengths
32
32
doc_id
stringlengths
5
12
section
stringlengths
5
1.47k
content
stringlengths
0
6.67M
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.7.2.3 Selecting UEs for reporting locations based on PEI
Figure 6.7.2.3-1: Selecting UEs for reporting locations based on PEI 1. The UE registers via NG-RAN node at an AMF. 2. The AMF obtains subscription data from the UDM that contain user consent for data collection and an indication whether the UE is suitable to serve as a PRU and whether the PRU is stationary. 3. The AMF...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.7.2.4 Inference of location for NG-RAN assisted positioning
Figure 6.7.2.4-1: Inference of location for NG-RAN assisted positioning 0. A trigger to determine location as specified in of TS 23.273 [7] occurs at an LMF, for instance an incoming Nlmf_Location_DetermineLocation service operation. The LMF determines to use NG-RAN assisted direct AIML positioning with LMF side model....
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.7.2.5 Inference of location for UE assisted positioning
Figure 6.7.2.5-1: Inference of location for UE assisted positioning 0. A trigger to determine location as specified in of TS 23.273 [7] occurs at an LMF, for instance an incoming Nlmf_Location_DetermineLocation service operation. The LMF determines to use UE assisted direct AIML positioning with LMF side model. 1. The ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.7.2.6 Monitoring accuracy of model for NG-RAN assisted location measurements
Figure 6.7.2.6-1: Monitoring accuracy of model for NG-RAN assisted location measurements 1.-10. Same steps as in Figure 6.7.2.1-1. The model training entity is replaced by the model accuracy monitoring entity, which may be the same as the model training entity or a different entity, for instance an LMF or NWDAF/MTFLF. ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.7.3 Impacts on services, entities and interfaces
Model training entity (LMF or NWDAF/MTLF): NOTE: The entity performing the model training is expected to be selected in the conclusion phase. - Performs a bulk subscription for location model training measurements towards the AMF. - Receive measurements and ground truth location. - Train model. - Store input data and t...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.8 Solution #8: MTLF-based model performance monitoring for AI/ML positioning
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.8.1 Description
This solution addresses Key Issue #1 "Enhancements to LCS to support Direct AI/ML based Positioning" regarding how to monitor model performance for ML models used for direct AI/ML based positioning. And assumes LMF with an AI/ML model for inference capability. According to the definition of TS 23.288 [5], an NWDAF cont...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.8.2 Procedures
Figure 6.8.2-1: Procedure of MTLF-based model performance monitoring for AI/ML positioning 1. The AF/Consumer NF subscribes to the LMF to request AI/ML Positioning. 2. The LMF performs UE location data collection from UE, RAN, AMF and GMLC as described in TS 23.273 [7]. Editor's note: The details on enhancements to tho...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.8.3 Impacts on services, entities and interfaces
MTLF: - Supports to monitor AI/ML positioning model performance. - Supports to generate new or re-trained AI/ML positioning models. LMF: - Supports to provide inference of AI/ML positioning to MTLF. RAN: - Supports to provide UE location data to LMF. UE: - Supports to provide UE location data to LMF.
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.9 Solution #9: new solution for KI#1 support monitoring the performance of AI model
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.9.1 Description
This solution resolves the how to monitor and evaluate the performance of an AI/ML models in Key Issue #1 Enhancements to LCS to support Direct AI/ML based Positioning. PRU (Positioning Reference Unit) defined for LCS service is reused to support LMF to monitor and evaluate the performance of an AI/ML model. In this so...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.9.2 Procedures
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.9.2.1 Performance of AI/ML model monitoring based on PRU
Figure 6.9.2.1-1 performance of AI/ML model monitoring based on PRU 1. AMF receives LCS service request from UE (i.e. MO-LR), or from LCS Client (i.e. MT-LR), or AMF itself. 2-3. AMF selects a LMF and sends Location request to LMF that determines to select AI/ML model based positioning method to compute the final UE lo...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.9.2.2 Procedure of PRU information collection for AI/ML model training
Figure 6.9.2.2-1: Procedure of PRU information collection for AI/ML model training 1. NWDAF invoke an Nnrf_NFDiscovery Request service operation to an NRF. The service operation includes a PRU existence indication and an AoI. 2. The NRF selects one or more PRU serving LMFs based on the PRU existence indication and the ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.9.3 Impacts on services, entities and interfaces
Editor's note: It is FFS to capture impacts on existing 3GPP nodes and functional elements. NWDAF: - Support to discover the PRU serving LMFs, and request the PRU known location and associated PRU location measurement(s) directly from the PRU serving LMFs. LMF: - Support for AI/ML Model inference to compute the final U...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.10 Solution #10: Direct AI/ML based positioning with NWDAF assistance
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.10.1 Description
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.10.1.1 General
This solution addresses Key Issue #1 "Enhancements to LCS to support Direct AI/ML based Positioning". In this solution, the existing NWDAF framework as specified in TS 23.288 [5] is reused for Direct AI/ML positioning in 5GC as follows: - The NWDAF containing MTLF trains the ML model for Direct AI/ML positioning. - The...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.10.1.2 Input data
To train the ML model, the NWDAF containing MTLF collects UE positioning related input data as list in Table 6.10.1.2-1. Table 6.10.1.2-1: Data collected by NWDAF for training the ML model for UE positioning Information Source Description UE ID LMF SUPI of the target UE. Timestamp LMF A time stamp when the location was...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.10.2 Procedures
Figure 6.10.2-1: Procedure for NWDAF providing ML model for direct UE positioning 1a. [In the case of 5GC-MT-LR procedure] The AF or LCS client sends a request to the GMLC for positioning of a target UE, and the GMLC invokes the Namf_Location_ProvidePositioningInfo service operation towards the AMF, as specified in TS ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.10.3 Impacts on services, entities and interfaces
NWDAF: - The NWDAF containing MTLF trains the ML model for Direct AI/ML positioning. - The NWDAF containing MTLF collects UE positioning related data from e.g. the LMF, AF and AMF, to (re)train the ML model for Direct AI/ML positioning. LMF: - Requests or subscribes to ML model for Direct AI/ML positioning from the NWD...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.11 Solution #11: Data collection procedure for LMF-side model training
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.11.1 Functional Description
This solution proposes to reuse the UE positioning procedure defined in TS 23.273 [7] to identify target LMF(s) for measurement data collection. After receiving the data collection request, the LMF(s) can send the measurement data and PRU information to NWDAF for LMF-side model training directly(i.e. without passing GM...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.11.2 Procedures
Figure 6.11.2-1: Procedure of data collection for LMF-side model training 1. NWDAF sends measurement data/PRU information collection request to GMLC with target UE ID(s) and its NF instance ID. Editor's note: Whether to perform data collection per UE, per area or both is FFS. 2. The GMLC invokes a Nudm_UECM_Get service...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.11.3 Impacts on existing services, entities and interfaces
To implement the solution above, the NWDAF shall provide its NF instance ID and target UE ID(s) in the training data collection request. The GMLC shall forward the request to AMF and then AMF shall forward the request to LMF. The LMF needs to prepare training data according to the received target UE ID(s) and send the ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.12 Solution #12: new solution for KI#1 support the data collection for AI model training based on authorization
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.12.1 Description
This solution resolves how to trigger or authorize the positioning data collection for AI/ML model training for Key Issue #1 Enhancements to LCS to support Direct AI/ML based Positioning. In this solution, it proposes to define a subscription data of authorization for positioning data collection of a UE for AI/ML mode ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.12.2 Procedures
Figure 6.12.2-1: Data collection for AI/ML mode training with authorization 1. NWDAF subscribes the data collection for AI model training service to LMF, to collect the measurements data/location data of one UE when the LCS Service is triggered in future. If NWDAF has the target UE ID(s), e.g. SUPI, or GPSI, then NWDAF...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.12.3 Impacts on services, entities and interfaces
Editor's note: It is FFS to capture impacts on existing 3GPP nodes and functional elements. LMF: - Support for AI/ML Model inference to compute the final UE location. - Check with UDM whether the positioning data of the UE is allowed for AI/ML model training. UDM: - Enhanced to authorize the positioning data collection...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.13 Solution #13: Sample/Feature alignment and general training procedure for the VFL
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.13.1 Description
This solution is proposed to address Key Issue #2: 5GC Support for Vertical Federated Learning and based on the Use Case #5. In Use Case #5, two scenarios are identified: - Scenario 1: NWDAF initiates VFL training process. - Scenario 2: AF initiates VFL training process. This solution focuses on scenario 2 where AF is ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.13.2 Procedures
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.13.2.1 Sample and feature alignment procedure
In VFL, it is required VFL Active Participant(VFL server) and VFL Passive Participant(VFL client) have same samples and different features before the model training. This solution proposed a preparation process used to have a negotiation between VFL Active Participant(VFL server) (i.e. AF) and VFL Passive Participant(V...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.13.2.2 General training procedure for the VFL between the VFL Active Participant(VFL server) (i.e. AF) and VFL Passive Participant(VFL client) (i.e. NWDAF)
In the process of VFL model training, the AF and NWDAF(s) exchange intermediate data (e.g. intermediate training result, loss information). It is assumed that the VFL Active Participant(VFL server) has ground truth data or labels or is able to collect ground truth data or labels. It is also assumed that the VFL Passive...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.13.2.3 Inferencing procedure for the VFL between the VFL Active Participant (i.e. AF) and VFL Passive Participant (i.e. NWDAF(s))
Figure 6.13.2.3-1 1. VFL Active Participant(VFL server) (i.e. AF) may determine to perform VFL inference using the VFL well-trained ML model which is represented by VFL model correlation ID. When there is not a VFL Passive Participant available for VFL inference which has been involved for VFL training of the VFL model...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.13.3 Impacts on services, entities and interfaces
NWDAF: - Enhancement to support the VFL model sample and feature alignment with the AF to make sure the sample is the same, but the feature for the sample is different between AF and NWDAF. - Based on the model and loss function provided by the AF, performing local training and update for the VFL model. NEF: - Enhancem...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.14 Solution #14: General procedure for NWDAF initiated Vertical Federated Learning between NWDAF(s) and AF(s)
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.14.1 Description
This clause specifies how NWDAF(s) and AF(s) can leverage VFL technique to train joint ML model to address the Key Issue #2: 5GC Support for Vertical Federated Learning. The training data is always a key element for ML model. In general, more data used in model training, better performance the model would achieve. Due ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.14.2 Procedures
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.14.2.1 Training Procedure
The figure 6.14.2.1-1 below shows general training procedure for Vertical Federated Learning between NWDAF(s) and AF(s). Figure 6.14.2.1-1: General training procedure for VFL between NWDAF(s) and AF(s) 0. The consumer sends a subscription request to VFL server NWDAF to train an ML model, including Analytics ID, ML mode...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.14.2.2 Inference Procedure
Figure 6.14.2.2-2: General inference procedure for VFL between NWDAF(s) and AF(s) 1. The first step is the same as the existing one captured in clause 6.4.4 of TS 23.288 [5] since the NF consumer doesn't need to be aware whether the NWDAF performs VFL or normal analytics for the requested analytics ID. Consumer NF send...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.14.3 Impacts on services, entities and interfaces
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.14.3.1 Training Impact
NWDAF: - Supports selecting the VFL clients and executing the VFL procedure. - Supports performing sample alignment. NRF: - Supports the registration and discovery of entities which participate in VFL.
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.14.3.2 Inference Impact
NWDAF: - Enhanced to initiate the VFL inference procedure to support the request analytics. - Collects the intermediate results from AF and generates the final inference result based on the AF and NWDAF intermediate inference result. NEF: - Enhanced to support the VFL inference related information exchange between NWDA...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.15 Solution 15: Support for vertical federated learning: Model Training and Inference
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.15.1 Description
Editor's note: This clause will describe the solution principles and architecture assumptions for corresponding key issue(s). Sub-clause(s) may be added to capture details. Vertical federated learning or feature-based federated learning is applicable to the cases that two data sets share the same sample space but diffe...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.15.2 Procedures
Editor's note: This clause describes high-level procedures and information flows for the solution.
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.15.2.1 Sample/Feature alignment
Feature/Sample alignment is supported by a VFL server receiving a model training request for VFL operation. The VFL Server on reception of a ML training request may perform the following tasks: - Identify other VFL function(s) that perform VFL operation based on the model requested. - Identify the data availability (fe...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.15.2.2 VFL training procedure
After the alignment is complete the VFL active participant coordinates the VFL training process. The task of the VFL active participant during the training process are proposed to be: - Train the model containing the label data based on intermediate values - Provide Gradient/Losses to Passive Participants - Determine c...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.15.2.3 Distributed Inference
Once a model consumer (i.e. an AnLF) is informed that the a model is trained using VFL, the consumer sends a inference request to the VFL Active Participant. The VFL active participant sends inference requests to each VFL participant and aggregates the received result to derive an aggregate inference output. The VFL ac...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.15.3 Impacts on services, entities and interfaces
Editor's note: This clause captures impacts on existing services, entities and interfaces. MTLF supports: - New functionality to support alignment of samples for the VFL process and selecting participants for the VFL training process. - Coordinating the VFL training by acting as an Active Participant. - Joining a VFL t...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.16 Solution #16: Support for VFL with NWDAF and AF as Participants
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.16.1 Description
This solution addresses use case #4 and key issue #2. Unlike traditional centralized learning approaches, where data is pooled together in a single location, or Horizontal Federated Learning (HFL), where different entities contribute similar types of data about different samples, VFL allows for the collaborative traini...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.16.2 Procedures
NOTE 1: In this solution, the VFL server coordinates the VFL operation and acts as active participant with access to labels. NOTE 2: VFL clients in this solution can be either VFL active participants with access to labels or VFL passive participants without access to labels. NOTE 3: In this solution, NWDAF instances co...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.16.3 Impacts on services, entities and interfaces
NWDAF: - New service to support data alignment. - Enhancements to Nnwdaf_MLModelTraining service (or new service(s)) to support VFL preparation as well as sharing of intermediate training and inference results. NEF/AF: - New service to support data alignment. - New service(s) to support VFL preparation as well as shari...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.17 Solution #17: NF discovery and selection for VFL
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.17.1 Description
The VFL process may include NWDAF and external AF. This solution proposes to do NF discovery process through NRF. The NWDAF can report/update its VFL capability to NRF and external AF can report its VFL capability to NRF via NEF. After reporting the VFL capability, the VFL server can select VFL client afterwards. Below...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.17.2 Procedures
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.17.2.1 External AF VFL capability report
The 2-side VFL capability reporting process of external AF can be conducted as follows: Figure 6.17.2-1: VFL capability report for External AF 1. External AF reports or updates its VFL capability(i.e. VFL server, VFL client, VFL server and client) and its AF ID to NEF. 2. NEF store the VFL capability and AF ID of exter...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.17.2.2 NWDAF VFL capability report
The 2-side VFL capability reporting process of NWDAF can be conducted as follow: Figure 6.17.2-2: VFL capability report for NWDAF 1. NWDAF reports/updates its VFL capability (i.e. VFL server, VFL client, VFL server and client) to NRF in its NF profile. 2. NRF stores the NF profile of NWDAF. 3. The NRF response with ope...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.17.2.3 Client selection process of VFL
The client selection process of VFL can be conducted as follows: Figure 6.17.2-3: Client selection process of VFL 0. The procedure above assumes the VFL server selection is done, or no need for VFL server selection. 1. The VFL Server NWDAF discovers and selects other NWDAF(s) containing MTLF as VFL Client NWDAF(s) from...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.17.3 Impacts on existing services, entities and interfaces
To implement the solution above, The AF and NWDAF shall report its VFL capability to NRF. The NEF shall be involved in helping external AF for the VFL capability reporting. On the other side, the VFL server shall have the ability to select VFL client(s) base on the criteria mentioned in clause 6.17.2.3. The NRF shall h...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18 Solution #18: Vertical Federated Learning between NWDAF and AF
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.1 Description
This solution is for Key Issue#2: 5GC Support for Vertical Federated Learning. If an ML model needs to be trained on local data set(s) from data source(s) (e.g. NF and AF), which have different feature spaces for the same samples (e.g. UE IDs), and some data cannot be obtained by the model training logic function direc...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2 Procedures
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.1 Discovery and selection of VFL clients
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.1.1 Discovery and selection of AF(s) if NWDAF as the VFL server
Figure 6.18.2.1.1-1: Registration and discovery of AF(s) for VFL 1-3. The AF registers its NF profile as defined in clause 5.2.7.2.2 of TS 23.502 [3], with the difference it includes the Supported VFL capability information (VFL capability type (e.g. VFL Server (VFL active participants) and/or VFL Clients (VFL passive ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.1.2 Discovery and selection of NWDAF(s) if AF as the VFL server
Figure 6.18.2.1.2-1: Registration and discovery of NWDAF for VFL 1-3. The NWDAF registers its NF profile as defined in clause 5.2.7.2.2 of TS 23.502 [3], with the difference it includes the Supported VFL capability information (VFL capability type (e.g. VFL Server (VFL active participants) and/or VFL Clients (VFL passi...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.2 VFL training procedure between NWDAF and AF
The following principles are defined for the VFL Joint ML Model training procedures: - If AF is involved in the VFL Joint ML Model training process, it is assumed that such AF is capable to perform model training. - If NEF is required for the interaction with 3rd party AFs, NEF services are extended in order to properl...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.2.1 VFL training procedure if NWDAF acts as the VFL server
Figure 6.18.2.2.1-1: VFL training procedure when NWDAF acts as the VFL server 1. If an ML model needs to be trained on local data set(s) from data source(s) (e.g. NF and AF), which have different feature spaces for the same samples (e.g. UE IDs), the NWDAF as the VFL server determines to use Vertical Federated Learning...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.2.2 VFL training procedure if AF acts as the VFL server
Figure 6.18.2.2.2-1: VFL training procedure when AF acts as the VFL server 1. If the ML model needs to be trained on local data set(s) from data source(s) (e.g. NF and AF), which have different feature spaces for the same samples (e.g. UE IDs), AF as the VFL server determines to use Vertical Federated Learning mechanis...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.3 VFL training procedure among NWDAFs
The following principles are defined for the VFL Joint ML Model training procedures: - An NWDAF (VFL Server) with VFL Capability for performing VFL ML Training can initiate the process VFL joint Training Process. NOTE 1: The procedures illustrated in 6.18.2.3-1 show the NWDAF (VFL Server), i.e. Active VFL Participant, ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.4 VFL inference procedure between NWDAF and AF
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.4.1 VFL inference procedure if NWDAF acts as the VFL server
Figure 6.18.2.4.1-1: VFL inference procedure when NWDAF acts as the VFL server 1. To obtain analytics, the NF consumer sends analytics info request to NWDAF with analytics ID, Target of Analytics Reporting and analytics filter. Editor's note: Whether the NWDAF in this step is the NWDAF acted as VFL sever in VFL model t...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.3.2 VFL inference procedure if AF acts as the coordinator
Figure 6.18.2.4.2-1: VFL inference procedure when AF acts as the VFL server 1. To obtain analytics, the NF consumer sends analytics info request to AF with analytics ID, Target of Analytics Reporting and analytics filter. 2. For the analytics info request, if the AF as the VFL server decides to do inference based on th...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.2.5 VFL inference procedure among NWDAFs
Figure 6.18.2.5-1 illustrated the procedure for analytics subscription at an NWDAF with an Active participant role (i.e. VFL Server) and the generation of the analytics output based on VFL inference process. NOTE: The procedure in Figure 6.18.2.5-1 supports the Use Case #4. The proposed procedure for VFL inference is b...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.18.3 Impacts on services, entities and interfaces
NWDAF: - Determine to use Vertical Federated Learning mechanism to train ML model if some features cannot be obtained directly from data producer AF. - Discover and select AF to participate the Vertical Federated Learning procedure. - Interact with AF(s) to perform VFL training and VFL inference. AF: - Determine to use...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.19 Solution #19: VFL inference procedure between NWDAF and AF
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.19.1 Description
This solution is proposed for KI#2 and Scenario 2 (i.e. AF initiated Scenario) of use case#5 to support the VFL inference procedure between NWDAF and AF. In the VFL model inference, the ML models are aligned, "VFL Passive Participant" NF(s), i.e. NWDAF, send intermediate results to "VFL Active Participant" NF, i.e. AF....
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.19.2 Procedures
Figure 6.19.2-1: VFL model inference on AF initiated scenario This clause describes the AF initiated VFL inference procedure to use ML model that trained in the VFL model training procedure. This procedure supports the two cases: case 1) the AF was involved in the VFL model training procedure, and case 2) AF was not in...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.19.3 Impacts on Existing Nodes and Functionality
For AF: - support to request the intermediate results of local ML models from NWDAF (s). - support to align the ML model between NWDAF. For NWDAF: - support to compute the intermediate results, report the intermediate results to the AF.
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.20 Solution #20: Inference procedure for the Vertical Federated Learning between NWDAF(s) and AF(s)
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.20.1 Description
This solution resolves KI#2: 5GC Support for Vertical Federated Learning, focusing on the procedures for VFL Inference. While each party in HFL uses the trained global model to make inferences, the parties in VFL have to collaborate to make inferences and each party may have only a sub-model. The ML Model Training for ...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.20.2 Procedures
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.20.2.1 Inference Procedure Initiated by the NWDAF
Figure 6.20.2.1-1: Inference procedure for the VFL initiated by NWDAF 0. Each NWDAF containing AnLF and AF which support VFL Inference is registered to the NRF with its NF Profile, which includes Analytics ID(s), Address information of NWDAF, Service Area, VFL Inference capability and Time interval supporting VFL Infer...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.20.3 Impacts to Services, Entities and Interfaces
NWDAF: - Support new Service operation between NWDAF(s) containing AnLF and/or AF(s) for VFL Inference (e.g. Nnwdaf_MLInferenceInfo_Request/Response, Nnwdaf_MLInference_Subscribe/Notify). - New capability indication for AnLF (VFL Inference Server, VFL Inference Client). - MTLF can provide the ML Model for VFL Inference...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21 Solution #21: Vertical Federated Learning for support of Application Layer QoE
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21.1 Key Issue mapping
This solution addresses KI#2 "5GC Support for Vertical Federated Learning".
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21.2 Description
Application Layer Quality of Experience (QoE) is affected by factors in different domains, such as network information (e.g. slice performance, applied policies), UE information (UE power consumption status, trajectory, user context) and Application Layer information (applied codec, QoE metric). The challenge is to use...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21.2.1 VFL for Application QoE provisioning
In this model of VFL, sub-models running in the network and the AF generates customized analytics based on an on-going session to influence the AF and UE application to optimize QoE experienced by the user. The process can be carried by either AF as active participant or NWDAF as active participant. The ML models are s...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21.2.2 VFL for Network parameters provisioning
In this model of Vertical Federated Learning, sub-models running in the network and AF generates customized analytics to pre-provision the network to support future QoE requirements of the user. The active participant (AF or NWDAF) performs network provisioning based on custom analytics generated by a ML algorithm oper...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21.3 Procedures
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21.3.1 Application provisioning for customized QoE
Figure 6.21.3.1-1: Call flow for application QoE provisioning 1. AF initiates the preparation phase for VFL for analytics ID Collaborative Analytics. This phase includes alignment on UE ID(s) and negotiation of active participant role. This step is always initiated by the AF irrespective of whether AF is the active par...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21.3.2 Network parameters provisioning
Figure 6.21.3.2‑1: Call flow for network parameter provisioning 1. AF initiates the preparation phase for VFL for analytics ID Collaborative Analytics. This phase includes alignment on UE ID(s) and negotiation of active participant role. This step is always initiated by the AF irrespective of whether AF is the active p...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.21.4 Impacts on services, entities and interfaces
AF: Initiates VFL UE ID alignment: - As active participant: - Provisions bottom layer processing information; - provisions gradients during training phase; - receives intermediate results. - As passive participant: - Receives bottom layer processing information; - Receives gradients during training phase; - Exposure of...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.22 Solution #22: Vertical Federated learning considering internal NWDAF architecture
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.22.1 Description
This solution addresses key issue 2 and use cases 4 and 5. It assumes that vertical federated learning is controlled by an NWDAF acting as server (or active participant). Clients (or passive participants) can either be other NWDAFs or AFs. No separate Controller is assumed. NOTE 1: This solutions assumes that the NWDAF...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.22.2 Procedures
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.22.2.1 VFL model training and inference involving MTLF, AnLF, ADRF and AF
Figure 6.22.2.1-1: VFL model training and inference involving MTLF, AnLF, ADRF and AF 1. The AnLF acting as VFL inference server registers its capability to act as VFL inference server in the NRF together with Analytics IDs it supports and a vendor ID. 2. The AnLF acting as VFL inference client registers its capability...
c8f50ee24a14933c33290076ea3625ad
23.700-84
6.22.2.2 VFL sample alignment for UE samples
Figure 6.22.2.2-1: VFL sample alignment for UE samples 1. Before starting a vertical federated learning, the MTLF acting as VFL training server determines an initial list of target UEs to be used as sample for vertical federated learning. For each candidate UE, steps 2 to 11 are performed. 2. The MTLF acting as VFL tra...