Content for  TS 23.288  Word version:  19.2.0

[OPTIONAL] VFL training is triggered in the VFL server for the following cases:

• Based on local configuration and agreement among vendors and/or application providers participating in the same group for specific VFL task(s).
• Triggered by either Analytics consumer, AnLF via ML model provisioning or MTLF internal (VFL Server is a trusted AF):

The NWDAF acting as VFL server determines the VFL clients that participate in VFL procedure in the VFL clients discovery and preparation phase as described in the clause 6.2H.2.1 and clause 6.2H.2.2.

To start the VFL training, the VFL server sends a request to start the training to all selected VFL clients. The request includes VFL correlation ID, at least the parameters negotiated during the preparation phase and theVFL training iteration number set to 0, and a Notification Correlation ID. Optionally, the VFL Server includes parameters according to clause 6.2H.3. If the VFL procedure continues in subsequent iterations, the VFL server sends a request for a new VFL training iteration containing an incremented a VFL training iteration number and intermediate model training information to each of the VFL clients for next round of VFL training. The VFL Server, based on internal logic, may provide checkpoint information according to clause 6.2H.3. Step 2a. Step 2b. Step 2c. Step 2d.

[Optional] Each VFL client collects its local data by using the current mechanism if the VFL client has no local data already available. The data used by each VFL Client is collected as per alignment information.

During VFL training procedure, each VFL client further trains the local ML model associated with the same VFL Correlation ID based on their own collected or available data and when applicable (e.g. after the first round of training) and possible intermediate model training information distributed by the VFL server in the previous training iteration. Each VFL Client computes and reports the client intermediate training result of the local ML model to the VFL server. VFL client(s) may also report a delta from the initial list of samples according to clause 6.2H.3. Based on internal logic, the VFL server may indicate if the training has to resume from a previous checkpoint by using its training iteration round ID.

Each VFL client reports the computed client intermediate training result of the local ML model to the VFL server. The Notification Correlation ID and a VFL training iteration number. A client may indicate in message for a request to leave the VFL. Step 5a. Step 5b. Step 5c. Step 5d.

The VFL server may collect the local data and generate its own local intermediate training result. The VFL Server computes the intermediate model training information (e.g. gradient information or loss information that may contain loss function or loss value) based on the VFL Client(s) intermediate training result(s) received in step 4, its own local intermediate results and the label. The intermediate model training information is used for updating the models of VFL clients and/or the model of the VFL Server. Different intermediate model training information may be computed for different VFL clients and for the VFL Server itself. The VFL server may locally compute contribution weights for each VFL client participating in the VFL pocess and store them for subsequent use, e.g. during inference. How the VFL computes the VFL contribution weights is up to implementation. The VFL server may also decide to compute the global ML model metric (e.g. ML model accuracy) based on all the intermediate training result received from VFL clients and the label at any time of VFL training. When determining to generate the ML model metric, the VFL server includes an indication to indicate to the VFL clients to use the sample ID of the dataset for accuracy monitoring to calculate intermediate training results in Nnwdaf_VFLTraining_Subscribe service operation. Corresponding to the sample IDs for accuracy monitoring, each VFL client collects local data as input data for local ML model trained in step 4, then the VFL client generates the intermediate training result based on the trained local ML model and the input data, and it provides the intermediate training result to VFL server. The VFL server computes the ML model metric (e.g. VFL accuracy) based on the received intermediate training result and the label, which are both corresponding to the sample IDs for accuracy monitoring. The VFL server may optionally take account the local ML model accuracy monitoring information received from VFL client(s) when computing global ML Model metric.

[Optional] The VFL server evaluates (e.g. based on the convergence of a loss function or loss value and/or if the pre-set iteration number is reached and/or global ML model metric is stable) whether VFL Training process converged. If the VFL Server evaluates the VFL Training process did not converge, the VFL Server determines another round of VFL training is required and repeats step 2 - 6. If the VFL Server evaluates the VFL training process converged, it determines the VFL Training is completed. In this case, the VFL Server terminates the current VFL training process via step 7. The VFL training termination decision may be also made as follows: Step 7b, 7c, 7f and 7h. Based on the consumer request, the VFL server notifies VFL status report to the consumer, to update the metric value to the consumer periodically (e.g. a certain number of training rounds or at fixed periods) or dynamically when some pre-determined status is achieved (e.g. the ML Model Accuracy threshold is achieved or training time expires).The status report may include global model metric (e.g. ML model accuracy). If NWDAF including MTLF has forwarded a subscription request to VFL server AF in step 0b, and it did not add Notification target set to the first NWDAF containing AnLF, it adds the AF ID, in the Notification to the first NWDAF containing AnLF. Steps 7d, 7e, 7g and 7i. The consumer decides whether the current model can fulfil the requirement, e.g. ML model metric is satisfactory for the consumer and determines to either unsubscribe or continue the training process. Based on the subscription request sent from the consumer, the VFL server updates or terminates the current VFL training process.

The VFL server sends VFL training termination message to VFL Client if it decides to terminate the VFL training process, the termination message contains VFL Correlation ID and may contain the intermediate model training information to each of the VFL clients. Step 8a. Step 8b. Step 8c. Step 8d.

The VFL Server, stores VFL correlation ID, the local trained ML Model, the mapping information of the VFL correlation ID to the following parameters: Analytics ID related to the VFL training process, locally trained Model. Additionally, the VFL server stores the VFL client information (including the NF ID of the VFL Client), which may be used to determine associated VFL client in the VFL inference. Each VFL client updates local ML model based on the intermediate model training information, if received and stores VFL correlation ID, the locally trained ML Model, the mapping information of the VFL correlation ID to locally trained Model

[CONDITIONAL] Same as in step 0 in clause 6.2H.2.3.1, when the AF is the VFL Server, but AF is replaced with untrusted AF, and Nnef_Training_Subscribe offered by NEF is used. NEF forwards the subscription request to AF using Naf_Training_Subscribe. In addition:

• Either based on the information received or internal configuration, VFL server decides to initiate VFL training procedure, or
• case E) same as step 0d, but the Analytics consumer contacts the NWDAF containing AnLF that wants to perform inference and does not have a model, it discovers VLF Server from NRF that is a VFL Server, then request the VFL Server via NEF to perform inference. The VFL server may trigger VFL training if no corresponding VFL training has been performed.

Same as step 1 in Figure 6.2H.2.3.1-1.

To start VFL training, the VFL server do same as in step 2 in Figure 6.2H.2.3.1-1, using Nnef_VFLTraining_Subcribe. Step 2c.

[Optional] Same as step 3 in Figure 6.2H.2.3.1-1.

Same as step 4 in Figure 6.2H.2.3.1-1.

Same as step 5 in Figure 6.2H.2.3.1-1. Step 5a. Step 5b. Step 5c-5d.

[Optional] Same as step 6 in Figure 6.2H.2.3.1-1.

Same as step 7 in Figure 6.2H.2.3.1-1. Same as in step 0 in clause 6.2H.2.3.1, when the AF is the VFL Server, but AF is replaced with untrusted AF, and Nnef_Training_Notify offered by NEF is used. NEF forwards the subscription request to AF using Naf_Training_Notify.

Same as step 8 of Figure 6.2H.2.3.1-1. However, sub steps in that Figure are not applicable. Step 8a. Step 8b. Step 8c.

Same as step 9 of Figure 6.2H.2.3.1-1.