Effective VFL Training Strategies Backed by Multiple Companies

Triggering VFL training Way Forwards 1

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: discuss which cases can be agreed first from technical point of view. Step2: Where and how to document them can be decided based on the cases agreed. 2

Scenario Description: - Analytics consumer discovers NWDAF(AnLF) to provide analytics (i.e. specific analytics ID); - If the discovered NWDAF(AnLF) is an NWDAF that can act as VFL server Case I: Back ground info: - This case has been mentioned by all DPs from companies, no one disagreed it Case 5 of Nokia s DP; Scenario 3 of Lenovo DP; Case 0d in Ericsson s endorsed CR; The 1st case in table 2 of HW s DP Case 3 of ZTE s DP - This case has been supported in the existing TS 23.288( clause 6.2.2H.4), some further wording refinements may be need ,e.g. to clarify VFL training can be stared if has not, before the VFL inference. Analytics Consumer (e.g. PCF) VFL server NWDAF (MTLF+AnLF) VFL Clients 1. discovers NWDAF(AnLF) at NRF to provide Analytics 2. Nnwdaf_Analytics_Subscribe 3. If no model available, perform VFL training 4.VFL training 5.VFL inference 6. Nnwdaf_Analytics_Notify Way Forward: Since no one disagreed this case, we keep it still be there in TS, and do some further wording refinements as above. 3

Scenario Description: - Analytics consumer discovers NWDAF(AnLF) to provide analytics - If the discovered NWDAF(AnLF) can not be the VFL server for that analytics ID - Then if the NWDAF(AnLF) discovers an VFL server (a. VFL server NWDAF or, b.VFL server AF) from NRF to provides analytics Case II: A): NWDAF as VFL server B): AF as VFL server NWDAF containing AnLF NWDAF VFL server Consumer NF NWDAF containing AnLF Consumer NF AF VFL server 0. Nnwdaf_AnalyticsInfo_Request / Nnwdaf_AnalyticsSubscription_Subscribe 0. Nnwdaf_AnalyticsInfo_Request / Nnwdaf_AnalyticsSubscription_Subscribe NEF VFL clients VFL clients 1: AnLF is enhanced to discover (e.g. or be aware of) VFL server to provide analytics from NRF 1: AnLF is enhanced to discover (e.g. or be aware of) VFL server to provide analytics from NRF 2. Nnwdaf_AnalyticsInfo_Request / Nnwdaf_AnalyticsSubscription_Subscribe 2. Nnef_Inference/analytics_Request 3. Trigger VFL training 3. Trigger VFL training 4. VFL inference 4. VFL inference 5a. Nnwdaf_AnalyticsInfo_Response 5a. Nnef_Inference/analytics_Response Back ground info: - This case has been mentioned by: Case 6-1(AF as VFL server) and 7-1(NWDAF as VFL server) of Nokia s DP (Nokia is positive) Similar as case 4-5 in ZTE s DP but there with more restrictions (ZTE conditionally supports this Case II) - This case has been covered in the existing TS 23.288( clause 6.2.2H.4), some further wording refinements may be need ,e.g. to clarify VFL training can be stared if has not, before the VFL inference. Way Forward: Since no one disagreed this case, we may keep it still be there in TS, and do some further wording refinements as above?? Otherwise, we can SoH 4

Scenario Description: - Analytics consumer discovers NWDAF(AnLF) to provide analytics - If the discovered NWDAF(AnLF) can not be the VFL server for that analytics ID - Then if the NWDAF(AnLF) discovers an NWDAF(MTLF) from NRF to provides the ML model, and if the discovered NWDAF(MTLF) can be the VFL server for that analytics ID. Case III: Analytics Consumer (e.g. PCF) NWDAF (AnLF) NWDAF (MTLF) (is an VFL server) Back ground info: - This case has been mentioned by DPs from companies Case 7-2 of Nokia s DP (not acceptable for Nokia) Scenario 1 of Lenovo DP (Lenovo supports) Case 0c in Ericsson s endorsed CR;(E supports) The 2nd-acase in table 2 of HW s DP (HW is negative?) ZTE can not accept this case III - This case is a new one to document in TS 23.288, proposed but not agreed in SA2#166AH VFL clients 1. Detect NWDAF at NRF via Analytics ID 2. Nnwdaf_Analytics_Subscribe 3. Decides to retrieve an ML model from MTLF, 4.Nnwdaf_MLmodelProvision_Subscribe No model 5.Starts VFL training VFL training comple 6.Nnwdaf_Analytics_Subscribe 7. VFL inference - SoH if support or against this case - support - against - Way forward according to the SoH result: - xx - xx 5

Scenario Description: - Analytics consumer discovers NWDAF(AnLF) to provide analytics - If the discovered NWDAF(AnLF) can not be the VFL server for that analytics ID - Then if the NWDAF(AnLF) discovers an NWDAF(MTLF) from NRF to provides the ML model, and if the discovered NWDAF(MTLF) is still not an VFL server for that analytics ID. - The NWDAF(MTLF) further asks an VFL server (AF) to do VFL training. Case IV: Analytics Consumer (e.g. PCF) NWDAF (AnLF) Proxy NWDAF (MTLF) AF VFL clients (VFL server) Back ground info: - This case has been mentioned by DPs from companies Case 4a of Nokia s DP (Nokia is negative) Scenario 2 of Lenovo DP (Lenovo supports) Case 0a+0b in Ericsson s endorsed CR;(E supports) The 2nd-bcase in table 2 of HW s DP (HW is positive) ZTE can not accept this case IV - This case is a new one to document in TS 23.288, proposed but not agreed in SA2#166AH 1. Detect NWDAF at NRF via Analytics ID 2. Nnwdaf_Analytics_Subscribe 3. Decides to retrieve an ML model from MTLF, 4.Nnwdaf_Mlmode lProvision_Subscribe 5.VFL training 6.Starts VFL training 7a.VFL training complete 7b.VFL training complete 8. Naf_Inference/Analytics_Subscribe 9.VFL inference - SoH if support or against this case - support - against - Way forward according to the SoH result: - xx - xx 6