Architectural Strategies for VFL Model Training Impacts
Explore different options for addressing VFL-related interactions between components like AnLF and MTLF in an NWDAF environment, examining implications and decision-making processes.
Download Presentation
Please find below an Image/Link to download the presentation.
The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author. If you encounter any issues during the download, it is possible that the publisher has removed the file from their server.
You are allowed to download the files provided on this website for personal or commercial use, subject to the condition that they are used lawfully. All files are the property of their respective owners.
The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author.
E N D
Presentation Transcript
Possible Architectural Question (back ground, not vote) How to address the following EN: For an NWDAF impacts of the split into AnLF and MTLF are FFS: Option 1: Assume that AnLF and MTLF are collocated both for NWDAF as VFL server and NWDAF as VFL client and thus do not define any VFL-related interactions between AnLF and MTLF. YES: NO: Option 2: Assume that AnLF and MTLF are collocated both for NWDAF as VFL server and NWDAF as VFL client and thus do not define any VFL-related interactions between VFL client AnLF and VFL client MTLF. Provide error handling when some AnLF tries to request model from MTLF as VFL server (corresponding to 2B and 4B) YES: NO: Option 3: Assume that AnLF and MTLF are collocated for NWDAF as VFL client and thus do not define any VFL-related interactions between VFL server and VFL client. For NWDAF as VFL server, allow that MTLF provisions model (plus vendor specific container with VFL related info about clients) to AnLF (corresponding to case 3) YES: NO: Option 4: Assume that no models can be provisioned from MTLF to AnLF both for NWDAF as VFL server and NWDAF as VFL client. But allow some AnLF that does not obtain the VFL related models to trigger the model training. (corresponding to case 1 and 4) YES: NO: 2
Case 0: VFL training triggered by configuration or internal logic Trigger = internal VFL server Already covered in specification as an existing Note No need to discuss or SoH. Next step: none VFL server AF or NWDAF(MTLF+AnLF) VFL Clients 3. Internal logic/OAM configuration 4.VFL training 3
Case 1: VFL training at AF VFL server by supervising AnLF Supervising AnLF (that also forwards inference requests) SoH : Question1: support of case 1? Yes: No: AF as VFL server VFL clients 1. Determine that model training is required 2. Naf_Model_Training request 3. Train VFL model 4. Model training is ongoing /Model training complete 1- Nnwdaf_analyticsinfo_ subscribe/request 2.Nnwdaf_Analytics_Subscribe/request Impacts Analysis: Superwising AnLF (that also forwards inference requests) is enhanced to supervise VFL training and can determine to start VFL training New service for request VFL training as step 2 3.Starts VFL training 4. Start VFL inference 5.Nnwdaf_Analytics_Notify/response 4
Case 2: VFL training at AF VFL server by MTLF A) MTLF asks AF VFL server to start VFL training B) MTLF rejects with error code no model due to VFL at AF NWDAF (AnLF) NWDAF (MTLF) AF as VFL Server VFL clients ANLF MTLF AF VFL server VFL clients 1. Decides to request to be notified when an ML model Is available 0.Nnwdaf_analyticsinfo_ subscribe/request 2.Nnwdaf_MLmodelProvision_Subscribe Model is not available (VFL Server) 1. Decides to retrieve an ML model from MTLF (if one that supports analytics ID is discovered) 4.Starts VFL training 2.Nnwdaf_MLmodelProvision_Subscribe 3. Error: No model due to VFL at AF 5.VFL training completed 4. Nnwdaf_MLmodelProvisoningNotify ML Model is available 0.Naf_Inference subscribe/request 5. If no model available, start VFL training 1- Nnwdaf_analyticsinfo_ subscribe/request 2. Naf_analyticsinfo subscribe/request 4.Starts VFL training 5. Start VFL inference 6.Naf_analyticsinfo notify/response Impacts Analysis: - If some MTLF supports the Nnwdaf_ML model provision_Subscribe service, it provides a new error code. Impacts Analysis: MTLF needs to determine VFL training and discover AF as VFL server; New service for request VFL training as step 3 5
SoH : Question1: support of 2A? Y N SoH : Question1: support of 2B? Y N 6
Case 3: VFL training at NWDAF VFL server triggered by model retrieval request from AnLF (separated AnLF and MTLF in VFL server ): Then AnLF as VFL inference server asks MTLF as VFL training server to retrieve ML model VFL server for inference (AnLF) VFL server for training (MTLF) SoH : Question3: support of case 3? Yes: No: 1. Determine that model is required, e.g. based on configuration, analytics monitoring or inference request 2. Nnwdaf_MLModelProvision_Subscribe 3. Train VFL model 4. Nnwdaf_MLModelProvision_Notify(VFL server model retrieval info, VFL correlation ID, vendor container VFL context(weights, clients, features etc) Impacts Analysis: - NWDAF as VFL server is split into two parts: AnLF as VFL inference server & MTLF as VFL training server. - Nnwdaf_ML model provision_Notify is enhanced to inform some info about VFL 7
Case 4: VFL training at NWDAF VFL server triggered by model retrieval request from AnLF B) :To retrieve ML model, AnLF discovers a VFL server (NWDAF) and asks it to provide model directly A) :AnLF discovers a VFL server (NWDAF) and asks to be notified when the ML Model is available NWDAF (AnLF) NWDAF (MTLF) (VFL server) VFL clients ANLF VFL server NWDAF VFL clients 1. Decides to request to be notified when an ML model Is available 0.Nnwdaf_analyticsinfo_ subscribe/request 2.Nnwdaf_MLmodelProvision_Subscribe Model is not available 1. Decides to retrieve an ML model from MTLF (if one that supports analytics ID is discovered) 3.Starts VFL training 4. Nnwdaf_MLmodelProvisoningNotify ML Model is available 2.Nnwdaf_MLmodelProvision_Subscribe 3. Error: No model due to VFL 1- Nnwdaf_analyticsinfo_ subscribe/request 2.Nnwdaf_Analytics_Subscribe/request 4.Nnwdaf_analyticsinfo_ subscribe/request 5. If no model available, start VFL training 3.Starts VFL training 4. Start VFL inference 5.Nnwdaf_Analytics_Notify/response Impacts Analysis: - the NWDAF as VFL server support the Nnwdaf_ML model provision_Subscribe service, it provides a new indication that no model - Nnwdaf_ML model provision_Notify is enhanced to inform AnLF that model is available Impacts Analysis: - the NWDAF as VFL server support the Nnwdaf_ML model provision_Subscribe service, it provides a new error code. 8
SoH : Question1: support of 4A? Y N SoH : Question1: support of 4B? Y N 9
Case 5: VFL training triggered by analytics request from NF Trigger = analytics Consumer NF consumer discovers a VFL server (NWDAF) directly Reached consensus as a special case for triggering VFL training. No need to discuss or SoH. 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 The next step: determine how to capture it with the exacting wording, for example to say analytics request from NF consumer may trigger VFL training 3. If no model available, perform VFL training 4.VFL training 5.VFL inference 6. Nnwdaf_Analytics_Notify Impacts analysis: - none 10
Case 6: VFL training triggered by analytics request from AnLF Trigger = analytics Consumer NF consumer discovers AnLF, AnLF discovers VFL server (NWDAF or AF) SoH : Question6: support of case 6 Yes: No: VFL server (AF or NWDAF containing AnLF Consumer NF NWDAF(MTLF+AnLF) ) 0. Nnwdaf_AnalyticsInfo_Request / Nnwdaf_AnalyticsSubscription_Subscribe NEF VFL clients 1: AnLF is enhanced to discover (e.g. or be aware of) VFL server to provide analytics from NRF 2. Nnwdaf/Af/nef_Analytics or Inference_Request (analytics ID) 3. If no model available, perform VFL training 3. Trigger VFL training 4. VFL inference 5a. Analytics/inference result Impacts analysis: - Some specific AnLFs (red one in the picture) are enhanced/preconfigured as an interworking node between 5GC consumer and AF as VFL server, it provides analytics on behalf of VFL server(AF) based on VFL inference result from AF 11
