Advanced VFL Training Trigger Condition Insights

Triggering VFL training Way Forwards 1

For trigger condition for VFL training or VFL inference, we had a lots of discussion in Jan emeeting and S2-2501134 from Ericsson, etc had been technically endorsed; Several companies brought discussion papers about trigger condition for VFL training or VFL inference S2-2501477(Nokia), 7 cases and corresponding solutions are discussed. S2-2501483(Ericsson), discussion to agree the endorsed CR involving 3 cases; S2-2501547(Lenovo), 3 cases similar as the ones in Ericsson s CR for VFL training are proposed and suggested to capture in TS; S2-2502035(HW), 3 cases similar as the ones in Ericsson s CR are mentioned, but proposed to document them in the clause of inference. S2-2501984(ZTE), 5 cases are mentioned, 1-3 cases are basic 4-5 are also OK to capture. Proposal 1: Step1: get confirmation on the basis of the endorsed CR: 1. the trigger for (VFL) training can be a) internal AnLF and b) Internal MTLF c) internal AF and even d) Analytics request from Consumer. All these cases shall be covered. 2. There could be multiple models for an Analytics ID, and each Model may be trained with a different method. 1. Models for an AnalyticsID may be trained using HFL or VFL or internally in MTLF. 2. Special case where Certain Analytics ID is trained using VFL only. Step2: Where and how to document them can be decided based on the cases agreed. 2

Case 0: VFL training triggered by OAM configuration or internal logic Trigger is internal VFL server VFL server NWDAF (MTLF+AnLF) VFL Clients 3. Internal logic/OAM configuration 4.VFL training Way Forward: Since no one disagreed this case, we may keep it in TS, more explicit clarification on wording may be needed. Suggested wording: XX (or discussed during the meeting) 3

Case 1: VFL training triggered by analytics request from NF or AnLF Trigger = Consumer B) NF consumer discovers AnLF, AnLF discovers VFL server (NWDAF or AF) A) NF consumer discovers a VFL server (NWDAF) directly VFL server (AF or NWDAF ) Proxy NWDAF containing AnLF Analytics Consumer (e.g. PCF) VFL server NWDAF (MTLF+AnLF) Consumer NF VFL Clients NEF 0. Nnwdaf_AnalyticsInfo_Request / Nnwdaf_AnalyticsSubscription_Subscribe 1. discovers NWDAF(AnLF) at NRF to provide Analytics VFL clients 2. Nnwdaf_Analytics_Subscribe 1: AnLF is enhanced to discover (e.g. or be aware of) VFL server to provide analytics from NRF 3. If no model available, perform VFL training 2. Nnwdaf/Af/nef_AnalyticsInfo_Request / Nnwdaf_AnalyticsSubscription_Subscribe 4.VFL training 5.VFL inference 6. Nnwdaf_Analytics_Notify 3. Trigger VFL training 4. VFL inference Back ground info: 1-A is corresponding to case 0d in endorsed paper, and implicitly supported in inference part (6.2H.2.4); 1-B is broadening solution in training into inference part (6.2H.2.4); Technically, no one has disagreed case 1A can be trigger for VFL training (though an EN exists on the need); Case A) can better be part of the inference requests, then refer to training procedure in step 4). Case B) it is part of the inference procedure. It is new as a trigger for training and need to be agreed. Some company concerned where to document this case, i.e. in VFL training clause or in VFL inference clause, or both? Because this case indeed involve both VFL training and VFL inference processes and combined them together. Ericsson requests offline to clarify that: a specific analytics uses only VFL and the AnLF knows that. 5a. Nnwdaf_AnalyticsInfo_Response Way forwards: keep this case in TS, but discuss where and how to document with some essential clarification if needed 4

Case 2: VFL training triggered by model provisioning/training request from AnLF or MTLF Covers all triggers a), b) and d) and deal with multiple models for same Analytics ID A) To retrieve ML model, AnLF discovers a VFL server (NWDAF) directly B) To retrieve ML model, AnLF discovers MTLF, then MTLF discovers a VFL server (AF) Traditional NWDAF (AnLF) Proxy NWDAF (MTLF) AF Traditional NWDAF (AnLF) NWDAF (MTLF) (is an VFL server) VFL clients VFL clients (VFL server) 0.Nnwdaf_analyticsinfo_ subscribe/request 0.Nnwdaf_analyticsinfo_ subscribe/request 1. Decides to retrieve an ML model from MTLF, 1. Decides to retrieve an ML model from MTLF, 2.Nnwdaf_Mlmode lProvision_Subscribe 2.Nnwdaf_MLmodelProvision_Subscribe No model 3.VFL training 4.Starts VFL training 3.Starts VFL training 5a.VFL training complete 5b.VFL training complete VFL training comple Background info: 2-A is corresponding to case 0c in endorsed paper. Wherein, step 1 may be decided based on AnLF s internal requirement of ML model or based on analytics request from analytics consumer (i.e. dash line in step 0). 2-B is corresponding to case 0a+0b in endorsed paper, wherein either Consumer, AnLF or MTLF triggers model training but MTLF takes decision on how to train model as in earlier releases. People have different view on this case 2, some companies can not accept this case 2 (neither A nor B). Some one comment even this case is accepted, they should be documented into clause for VFL inference. way forwards: Step 1: discuss whether we need this case or not. This is part of the endorsed CR, as such we wonder why we question this case only. Step 2: if yes, then decide where to document and how. (please see next slide about potential documentation proposal ) 5

Ericssons proposal on documentation and rewording: For 2-A (corresponding to case 0c in the endorsed paper): in training 0c was added for the case where the NWDAF is the VFL Server, (not the AF), then AnLF contacts MTLF that happens to be the VFL Server and can start training with the VFL Clients. For 2-B (corresponding to case 0a, 0b in the endorsed paper : in training & in inference 1. In training procedure, we add the possibility in step 0a and b to add an optionalflag indicating inference shall be started when training is done and result shall be sent to consumer . This is to not force AnLF to request Inference if this is part of Inference procedure. 2. In Inference procedure a reference to step 0a and b is done, stating that the flag shall be set. Please Ericsson can clarify further to help people to about this proposal on this page. Ericsson s response: Case 2-A is described according to our understanding Case 2-B addition is if we agree that an inference/analytics request is used in Inference procedure as trigger for training and referenced to training procedure using ML Model provisioning A call flow showing compromise by Ericsson is shown in the next page. 6

Case 3: Proposal on compromise: NWDAF (AnLF) NWDAF (MTLF) NWDAF (VFL client) Trusted AF (VFL client) Untrusted AF (VFL client) NF NEF VFL server ... ... 0a. Nnwdaf_MLModelProvision_Subscribe Response to subscription [MTLF decides either 0b) respond no model will be available or oc) ask AnLF to request Training directly to FL server] NWDAF (VFL client) Untrusted AF (VFL client) To cover the basic pre- Rel-19 cases handling of Model provisioning should be covered. All triggers should be part of solution. 0b. Naf_Training_Subscribe 0c. Naf_Training_Subscribe 0d. Nnwdaf_MLModelProvision_Subscribe [ask for VFL no model will be shared] 0e.Nnwdaf_AnalyticsSubscription_Subscribe Response to subscription 1. VFL Client Selection and VFL preparation phase 0a. Covers the case when MTLF decides what model and training method to use based on the parameters sent from AnLF. This is similar as we have in pre-Rel-19. Repeat until training completes 2. Start VFL training / Start another VFL training round 2a. Nnwdaf_VflTraining_Subscribe 2b. Naf_VflTraining_Subscribe 2c. Nnef_VflTraining_AFClient_Subscribe 2d. Naf_Vfl_Training_Subscribe 3. Collect input data 4. Local model Update 5. Local model information reporting 5a. Nnwdaf_MLModel_Training_Notify or Nnwdaf_MLModelTrainingInfo_Request response. 0b is the case when MTLF triggers model training using VFL (can be triggered internally in MTLF or may be triggered by 0a). This is similar as for HFL. 5b. Naf_Vfll_Training_Notify 5c. Naf_Vfl_Training_Notify 5d. Nnef_Vfl_Training_AFClient_Notify 6. Combine intermediate results and calculate training results 7a. Evaluate results 0c Is covering Nokia s suggested case that a new service (Naf_Training) can be triggered by AnLF. This may be triggered directly by AnLF or if agreed after MTLF has suggested it to do so. 7b. Naf_VFLTraining_Notification 7c. Nnwdaf_MLModelProvision_Subscribe 7d. Nnwdaf_MLModelProvision_Subscribe/unsubscribe 7e. Naf_VFLTraining_Subscribe/unsubscribe 7f. Nnwdaf_MLModelProvision_Notification 7g. Nnwdaf_MLModelProvision_Subscribe/unsubscribe 7h. Update local model and continue or terminate the VFL training process 8. VFL training Termination 8a. Nnwdaf_VFLTraining_Unsubscribe 0d Same as 0c but for an MTLF as VFL server. 8b. Naf_VflTraining_Unsubscribe 8c. Nnef_VflTraining_AFClient_Unsubscribe 8d. Naf_Vfl_Training_Unubscribe 0e is the special case in endorsed CR 9. store model 9. store model 7