Content for TS 23.288 Word version: 19.3.0

1… 4… 5… 5A… 6… 6.1.3 6.1.4… 6.1.4.4… 6.1.5… 6.1A… 6.1B… 6.1B.2.3… 6.1C 6.2… 6.2.3… 6.2.6… 6.2.6.2 6.2.6.3… 6.2.6.3.3 6.2.6.3.4 6.2.6.3.5 6.2.6.3.6… 6.2.7… 6.2.8… 6.2.9… 6.2.13… 6.2A… 6.2B… 6.2B.3 6.2B.4… 6.2C… 6.2D… 6.2E… 6.2F… 6.2H… 6.2H.2.2… 6.2H.2.3… 6.2H.2.4… 6.3… 6.4… 6.5… 6.6… 6.7… 6.7.3… 6.7.4… 6.7.5… 6.8… 6.9… 6.10… 6.11… 6.12… 6.13… 6.14… 6.16… 6.17… 6.18… 6.19… 6.20… 6.21… 6.22… 6.23… 7… 7.4… 7.7… 7.9… 8… 9… 10… 11… 12… A…

6.2H.2.4 Inference procedure for vertical federated learning 6.2H.2.4.1 Inference procedure for vertical federated learning when NWDAF or Trusted AF is acting as VFL server 6.2H.2.4.2 Inference procedure for vertical federated learning when untrusted AF is acting as VFL server 6.2H.2.4.3 Contents of VFL Inference service 6.2H.3 Contents of ML Model VFL Training services for Vertical Federated Learning 6.2H.4 Contents of ML Model Training services for Vertical Federated Learning
...

6.2H.2.4 Inference procedure for vertical federated learning p. 180

6.2H.2.4.1 Inference procedure for vertical federated learning when NWDAF or Trusted AF is acting as VFL server p. 180

Reproduction of 3GPP TS 23.288, Fig. 6.2H.2.4.1-1: Inference procedure for vertical federated learning when NWDAF or Trusted AF is acting as VFL server

Figure 6.2H.2.4.1-1: Inference procedure for vertical federated learning when NWDAF or Trusted AF is acting as VFL server

The inference procedure when trusted AF is acting as VFL server may be triggered by a request or subscription from a 5GC consumer NF or internal service logic of the AF acting as VFL server. If triggered by internal service logic of the AF acting as VFL server, the steps 0, 1,7 and 8 are skipped.

Step 0.

The analytics consumer NF sends an Analytics request/subscribe (Analytics ID, Target of Analytics Reporting= e.g. UE IDs and optionally both Analytics Reporting Information=Analytics target period and Analytics Filter, Analytics Accuracy Request if the analytics consumer NF requests Analytics Accuracy Monitoring) to NWDAF containing AnLF by invoking a Nnwdaf_AnalyticsInfo_Request or a Nnwdaf_AnalyticsSubscription_Subscribe., providing parameters as defined in clause 6.1.3.

Step 1.

If the NWDAF containing AnLF can be the VFL server to generate the VFL inference results for the requested analytics ID, then step 1 is skipped.

If the NWDAF containing AnLF cannot generate the analytics output, the NWDAF containing AnLF determines the VFL Server AF for the requested analytics.

If the VFL server is trusted AF, the NWDAF containing AnLF sends a request to trusted AF VFL server using Naf_Inference_subscribe/request including Analytics ID, Target of Analytics Reporting = e.g. UE IDs and optionally Analytics Reporting Information=Analytics target period Analytics Filter and Analytics Accuracy Request.

Step 2.

Based on the information received in the step 0 or 1, VFL server decides to initiate the VFL inference procedure with the VFL clients. Before the VFL server initiate the VFL inference procedure, the VFL server may initiate the VFL training procedure if no VFL model is already trained as described in the clause 6.2H.2.3 based on the information received in the step 0 or 1 or VFL server local configuration. VFL Server selects clients(s) using information stored in the VFL training process. The VFL server may select some or no clients, e.g. depending on the accuracy of the VFL model, the contribution to the training result and the current status of the VFL clients.

When no VFL Clients are selected, the VFL server may generates the VFL inference results based only on its local trained ML model associated with the determined VFL correlation ID, skipping the steps 2 - 6 (and 7 if step 1 was also skipped).

When a VFL client is not available for inference, e.g. depending on the contribution to the training result, the accuracy of the VFL model, the contribution to the training result and the current status of the VFL clients, the VFL server may select another available VFL client with a well-trained model not yet selected for inference, from the original training process, or the VFL server may select a new VFL client and start a new training procedure.

VFL server NWDAF or trusted AF determines and sends a VFL Inference request/subscription to each VFL client including the Target of VFL inference = e.g. UE IDs, VFL correlation ID to indicate the VFL client which previously well-trained VFL local model associated with this ID will be used and optionally VFL inference filter, Data time window, Time when intermediate local result is needed, Dataset Statistical Properties, Analytics Metadata Request, as defined in clause 6.2H.2.4.3.

Step 2a.

For each NWDAF VFL client, the VFL Server NWDAF or trusted AF sends an Nnwdaf_VFLInference_Subscribe or Nnwdaf_VFLInference_Request to the VFL client.

Step 2b.

For each trusted AF VFL client, the VFL Server NWDAF sends an Naf_VFLInference_Subscribe or Naf_VFLInference_Request to the VFL client.

Step 2c.

For each untrusted AF VFL client, the VFL Server NWDAF sends an Nnef_VFLInference_Subscribe or Nnef_VFLInference_Request to the NEF serving the AF.

Step 2d.

For each untrusted AF VFL client, the NEF converts any internal identifiers to external identifiers and sends an Naf_VFLInference_Subscribe or Naf_VFLInference_Request to the untrusted AF VFL client.

Step 3.

Each VFL Client collects its local data by using mechanism as specified in clause 6.2 and considering a possibly received data time window to ensure that different VFL clients use the inference data collected in the same time period if the VFL Client does not have local data available already.

Step 4.

Based on the VFL correlation ID, each VFL Client determines the VFL local model to generate the intermediate local inference results.

Step 5.

VFL Client sends the client intermediate local results to the VFL server. If requested in step 2, it includes Analytics Metadata, as defined in clause 6.2H.2.4.3. Alternatively, the VFL client may report their status and other cause value for rejecting the VFL inference process (e.g. overload, target UE moved out of NWDAF serving area) in the VFL inference response service.

The intermediate local results, which are sent from the VFL Client to the VFL Server during the VFL inference process, are the information for the VFL Server to combine and generate the VFL inference results.

If the VFL server used an inference subscription in step 2, step 5 may be repeated.

Step 5a.

Each NWDAF VFL client sends an Nnwdaf_VFLInference_Notify or Nnwdaf_VFLInference_Request response to the VFL Server NWDAF or trusted AF.

Step 5b.

Each trusted AF VFL client sends a Naf_VFLInference_Notify or Naf_VFLInference_Request response to the VFL Server NWDAF.

Step 5c.

Each untrusted AF VFL client sends a Naf_VFLInference_Notify or Naf_VFLInference_Request response to the NEF.

Step 5d.

For each untrusted AF VFL client, the NEF converts any external to internal identifiers and sends an Nnef_VFLInference_Notify or Nnef_VFLInference_Request response to the NWDAF VFL server.

Step 6.

The VFL server may collect its local data and generate the intermediate local inference results. When the VFL Server selected VFL clients to participate in the VFL Inference process, it combines all the intermediate local results to generate the combined inference results based on the VFL correlation ID. The VFL server takes into account the participation of each VFL client during the ML training process and the importance of the intermediate local results when it generates the combined inference results.

The VFL server may compute the VFL accuracy based on all the intermediate local results received from VFL clients and the label, if it receives Analytics accuracy request in step 0. The server also aggregates possibly received analytics metadata.

Step 7.

Depending on request, the NWDAF or trusted AF as VFL server sends Nnwdaf_AnalyticsInfo_Response or Nnwdaf_AnalyticsSubscription_Notify or Naf_Inference_Response/Notify to the consumer (i.e. NWDAF containing AnLF) including the VFL inference results, optionally, VFL accuracy.

Step 8.

The NWDAF containing AnLF may provide VFL accuracy if the consumer NF provides Analytics Accuracy request in step 0.

6.2H.2.4.2 Inference procedure for vertical federated learning when untrusted AF is acting as VFL server p. 183

Reproduction of 3GPP TS 23.288, Fig. 6.2H.2.4.2-1: Inference procedure for vertical federated learning when untrusted AF is acting as VFL server

Figure 6.2H.2.4.2-1: Inference procedure for vertical federated learning when untrusted AF is acting as VFL server

The inference procedure when untrusted AF is acting as VFL server may be triggered by a request or subscription from a 5GC consumer NF or internal service logic of the AF acting as VFL server. If triggerd by internal service logic of the AF acting as VFL server, the steps 0, 1,7 and 8 are skipped.

Step 0.

Same as step 0 in clause 6.2H.2.4.1.

Step 1.

Same as step 1 in clause 6.2H.2.4.1, to NEF using Nnef_Inference_subscribe/request. NEF forwards the subscription request to AF using Naf_Inference_subscribe/request. The NWDAF containing AnLF determine the VFL server AF during the VFL training phase as described in clause 6.2H.2.3.2 or based on the untrusted AF VFL server discovery procedure as described in the clause 6.2H.2.1.2.

Step 2.

Same as step 2 in clause 6.2H.2.4.1, to NEF using Nnef_VFLInference_subscribe/request. An untrusted AF includes the external NWDAF ID and sends the request to the NEF.

NEF converts any received external identifiers to internal identifiers and forwards the subscription request to NWDAF using Nnwdaf_VFLInference_subscribe/request.

Step 2c.

An NWDAF VFL client may send Nnwdaf_VFLInference_Subscribe or Nnwdaf_VFLInference_Request to the other one or more indirect NWDAF VFL client as configured from which it desires to receive the client intermediate inference results.

Step 3.

Same as step 3 in clause 6.2H.2.4.1.

Step 4.

Same as step 4 in clause 6.2H.2.4.1.

Step 5.

Same as step 5 in clause 6.2H.2.4.1, to NEF using Nnwdaf_VFLInference_Notify/Request Response.

NEF converts any received internal identifiers to external identifiers and forwards the subscription notify to the untrusted AF using Nnef_VFLInference_Notify/Request Response.

Step 5c-5d.

The indirect NWDAF VFL clients share the client intermediate inference results with the NWDAF VFL client from which it received request in step 2c. This NWDAF VFL client aggregates the received client intermediate inference results, performs local computation using the aggregated intermediate inference results, then sends one notify message to the NEF by including its client intermediate inference result.

If the VFL server used an inference subscription in step 2, step 5 may be repeated.

Step 6.

Same as step 6 in clause 6.2H.2.4.1.

Step 7.

Same as step 7 in clause 6.2H.2.4.1, to NEF using Naf_Inference_Notify/Request Response.

NEF converts any received internal identifiers to external identifiers and forwards the subscription Notify to NWDAF using Nnef_Inference_Notify/Request Response.

Step 8.

Same as step 8 in clause 6.2H.2.4.1.

6.2H.2.4.3 Contents of VFL Inference service p. 184

The consumers of the VFL inference services may provide the input parameters in Nnwdaf_VFLInference service or Naf_VFLInference service or in Nnef_VFLInference service as listed below:

Analytics ID: identifies the analytics for which the ML Model is requested to be used for inference.
(Only for Inference subscribe service) A Notification Target Address (+ Notification Correlation ID) as defined in clause 4.15.1 of TS 23.502, allowing to correlate notifications received from the VFL client with the subscription.
(for new Inference subscription or inference request) A VFL Correlation ID: indicate the VFL client which previously well-trained VFL local model associated with this ID will be used.
(for new Inference subscription or inference request) Target of VFL inference: indicates the object(s) for which data for VFL inference is requested, i.e. a list of SUPIs, group of UEs identified by a list of Internal-Group-Ids or External-Group-Ids.
(Only for new Inference subscribtion) Subscription Correlation ID.
(only for Nnef_VFLInference service) external NWDAF ID.
[Optional] VFL inference filter: indicates the conditions to be fulfilled for generating intermediate local inference results, e.g. S-NSSAI, Area of Interest. Parameter types in the VFL inference filter are the same as parameter types in the Analytics Filter Information which are defined in procedures. S-NSSAI is only applicable for the Nnwdaf_VFLInference service.
[Optional] Time when intermediate local inference result is needed: indicates to the VFL client the latest time to receive intermediate local inference result.
[Optional] Dataset Statistical Properties: information in order to influence the data selection mechanisms to be used for the generation of an intermediate local inference results. See clause 6.1.3 for details.
[Optional] Data time window: if specified, only events that have been created in the specified time interval are considered for the inference result generation.
[OPTIONAL] Analytics metadata request: indicates a request from VFL server to VFL client to provide the "analytics metadata information" related to the produced intermediate local inference results.

The VFL client provides to the consumer of the VFL inference service operations the output information in as listed below:

(Only for Inference notify) Notification Correlation Information.
Intermediate local inference results: the output of the client VFL local model, which is used for the VFL server to generate the VFL inference results.VFL correlation ID.
Analytics ID.
[Optional] Analytics metadata information: additional information required to aggregate the output analytics for the requested Analytics ID(s):
Number of data samples used for the generation of the output analytics.
Data time window of the data samples.
Dataset Statistical Properties of the analytics output used for the generation of the analytics.
[OPTIONAL] Data source(s) of the data used for the generation of the output analytics.

6.2H.3 Contents of ML Model VFL Training services for Vertical Federated Learning p. 185

The consumers of the ML Model training services may provide the input parameters in Nnwdaf_VFLTraining service or Naf_VFLTraining service or in Nnefe_VFLTraining service as listed below:

Analytics ID: identifies the analytics for which the ML Model is requested to be trained.
At start of Training phase VFL server sends VFL Correlation ID: identifies a VFL process to be executed among the candidate VFL participants.
[OPTIONAL] VFL Interoperability Information that indicates the intermediate model training information that the VFL Server supports and asks the client to verify its support for (e.g. activations, gradients, type of loss), the content of the VFL Interoperability Information is not standardized in this Release.
[OPTIONAL] In training phase, Maximum response time (i.e. the maximum time between VFL clients receive intermediate model training information and send back intermediate training result).
[OPTIONAL] Training Filter Information: enables to select which data for the ML Model training is requested, e.g. S-NSSAI, Area of Interest. Parameter types in the Training Filter Information are the same as or subset of parameter types in the ML Model Filter Information which are defined in clause 6.2A.2.
In training, the Intermediate model training information includes what the Server sends to each client which contains what has been agreed in the preparation phase. The content is not standardized in this Release.
[OPTIONAL] ML model monitoring accuracy check flag: request local ML model accuracy monitoring information which assists VFL server to perform ML model accuracy monitoring.
[OPTIONAL] Feature ID(s) as described in preparation phase and at start of training phase. See definition of Feature ID below.
[OPTIONAL] At start of Training phase, Use case context: indicates the context of use of ML Model.
In training, VFL training iteration number.
In the preparation phase, the initial list of samples selected by the VFL Server.
[OPTIONAL] At start of Training phase, sample ID of the dataset for accuracy monitoring are selected from the sample ID list decided in the preparation phase, identifying the data to be used for generating intermediate training result by VFL client and supporting the VFL Server to determine the VFL ML model accuracy.

NOTE 1:
How the VFL server determine the sample IDs of dataset for accuracy monitoring is up to implementation.
[OPTIONAL] In later steps in training, the VFL server may provide delta from the sample list sent in first step of training procedure to indicate which samples aligned in the preparation phase will not be part of the rest of training procedure.

NOTE 2:
The delta list can only remove samples from the initial sample list.
[OPTIONAL] Checkpoint information: indicates whether model status at the current iteration should be saved as a training checkpoint.

The VFL client provides to the consumer of the ML Model training service operations the output information in as listed below:

In the preparation phase, the list of samples accepted by the VFL client.
[OPTIONAL] In training the VFL client may provide delta from the sample list received in first step of training procedure to indicate which samples will not be part of the rest of training procedure.

NOTE 3:
The delta list can only remove samples from the initial sample list.
[OPTIONAL] As described in preparation phase, the Feature ID(s) that the VFL client supports. A Feature ID indicates what features the VFL client can use for an Analytics ID, the values of the Feature ID are not standardized.
[OPTIONAL] VFL Interoperability Information that indicates the Intermediate model training information that the client supports in response to what the server sent in the subscription request.
In training, the Intermediate training result includes what the Client sends to the server which contains what has been agreed in the preparation phase. The content is not standardized in this Release.
[OPTIONAL] Local ML model accuracy monitoring information.

6.2H.4 Contents of ML Model Training services for Vertical Federated Learning p. 186

The consumers of the ML Model training services may provide the input parameters in Naf_Training service or in Nnef_Training service as listed below. The NWDAF decides which of the optional parameters it forwards when it has received an ML Model provisioning subscription request.

Analytics ID: identifies the analytics for which the ML Model is used.
[OPTIONAL] Use case context: indicates the context of use of the analytics to select the most relevant ML Model.
[OPTIONAL] ML Model Filter Information: indicates the applicable conditions of the trained ML model and enables the consumer to select which ML Model for the analytics is requested, e.g. S-NSSAI, Area of Interest. Parameter types in the ML Model Filter Information are the same as parameter types in the Analytics Filter Information which are defined in procedures.
[OPTIONAL] Target of ML Model Reporting: indicates the object(s) for which ML Model is requested, e.g. specific UEs i.e. a list of SUPIs, a group of UEs i.e. a list of Internal-Group-Ids or any UE (i.e. all UEs).
[OPTIONAL] Requested representative ratio: a minimum percentage of UE(s) in the group whose data is a non-empty set and can be used in the model training when the Target of ML Model Reporting is a group of UEs i.e. a list of Internal-Group-Ids.
[OPTIONAL] ML Model Target Period: indicates time interval [start, end] for which ML Model for the Analytics is requested. The time interval is expressed with actual start time and actual end time (e.g. via UTC time).
[OPTIONAL] Inference Input Data information: contains information about various settings that are expected to be used by the consumer (e.g. NWDAF containing AnLF) during inferences such as:
the "Input Data" that are expected be used, each of them optionally accompanied by metrics that show the granularity with which this data will be used (i.e., a sampling ratio, the maximum number of input values, and/or a maximum time interval between the samples of this input data).

NOTE 1:
This can be a subset of the possible Input Data specified for a certain analytics type.
the data sources that are expected to be used, indicated as a list of NF instance (or NF set) identifiers.
ML Model Reporting Information for the subscription, parameters as per Event Reporting Information Parameter defined in Table 4.15.1-1, TS 23.502.
A Notification Target Address (+ Notification Correlation ID) as defined in clause 4.15.1 of TS 23.502, allowing to correlate notifications with this subscription.
[OPTIONAL] Accuracy level of Interest.
[OPTIONAL] Time when model is needed: indicates the latest time when the consumer expects to receive the ML Model(s).
[OPTIONAL] ML Model Monitoring Information:
- desired ML Model metric.
  
  NOTE 2:
  In this Release, only "ML Model Accuracy" is defined as ML Model metric.
- [OPTIONAL] For subscription ML Model monitoring reporting mode: such as Accuracy reporting interval or pre-determined status. Depending on the reporting mode, the AF reports the ML Model accuracy either periodically or when the ML Model accuracy is crossing an ML Model Accuracy threshold, i.e. the accuracy either becomes higher or lower than the ML Model Accuracy threshold.
- [OPTIONAL] ML Model Accuracy Threshold: indicating the accuracy threshold of the ML Model requested by the consumer (as a kind of pre-determined status). It also can be used as an indication that the AF is triggered to execute the accuracy monitoring operations for the ML Model provisioned to AnLF.
- [OPTIONAL] DataSetTag and ADRF ID if available: indicates the inference data (including input data, prediction and the ground truth data at the time which the prediction refers to) stored in ADRF which can be used by AF to retrain or reprovision of the ML Model.

The AF provides to the consumer the output information as listed below:

For Notifications, the Notification Correlation Information.
For the Analytics ID requested by the service consumer, the following information:
- Indication whether training is ongoing when e.g. accuracy report is provided on request from consumer, or that training is done, and thereby Analytics or Inference can be requested to NF identified by ML Model provider information.
- [OPTIONAL] Spatial validity: indicates Area where the provided ML Model Information applies.
  
  NOTE 3:
  Spatial validity and Validity period are determined by internal logic and it is a subset of AoI if provided in ML Model Filter Information and of ML Model Target Period, respectively.
- [OPTIONAL] ML Model representative ratio: indicating the percentage of UE(s) in the group whose data is used in the ML Model training when the Target of ML Model Reporting is a group of UE(s).
- [OPTIONAL] ML Model Accuracy Information: indicates the accuracy of the ML Model if related ML Model Monitoring Information was provided, which includes:
  - the metric value of the ML Model.