Content for TR 22.876 Word version: 19.1.0
6.2 5GS assisted transfer learning for trajectory prediction
6.2.1 Description
6.2.2 Pre-conditions
6.2.3 Service Flows
6.2.4 Post-conditions
6.2.5 Existing features partly or fully covering the use case functionality
6.2.6 Potential New Requirements needed to support the use case
6.2.6.1 Potential Functionality Requirements
6.2.6.2 Potential KPI Requirements
...
...
6.2 5GS assisted transfer learning for trajectory prediction p. 21
6.2.1 Description p. 21
AIML model transfer learning is beneficial for lowing cost and raising effective when training a model using a target UE based on a pre-training model. The principle of transfer learning is to use the knowledge from the source domain to train a model in the target domain to achieve more expedient and higher accuracy efficiency [25].
![Copy of original 3GPP image for 3GPP TS 22.876, Fig. 6.2-1: AI/ML model transfer learning from source UE to target UE [26]](../img/tinv-22-876-6.2-1.gif)
Figure 6.2-1: AI/ML model transfer learning from source UE to target UE [26]
(⇒ copy of original 3GPP image)
(⇒ copy of original 3GPP image)
Since the AI model is a kind of knowledge, when the centralized application server acquires enough number of AIML model used by UEs, it may perform a backward inference/inversion attacks [27] to derive the feature of UE's local data set, which means a privacy risk exists. In order to resolve the privacy concern for transfer learning, the model transfer via direct device connection is a better to be used so that the network node (e.g. application server) cannot acquire the AIML model used by UE and no way to do backward inference.
6.2.2 Pre-conditions p. 21
Alice is a customer of intelligent-driving service provided by company-A. She lives in Chaoyang district in Beijing and driving to her office building in CBD every working day. By using the intelligent driving service, Alice's car can predict the trajectory of neighbouring vehicles (as Figure 6.2-2 shows), so as to pre-alert Alice of some potential collision and Alice can decide whether to steer, accelerate, or any other driving operation.
![Copy of original 3GPP image for 3GPP TS 22.876, Fig. 6.2-2: Qualitative results using model of trajectory prediction: the orange trajectory represents the observed 2s. Red represents ground truth for the next 3 seconds and green represents the multiple forecasted trajectories for those 3s [24]](../img/tinv-22-876-6.2-2.gif)
Figure 6.2-2: Qualitative results using model of trajectory prediction: the orange trajectory represents the observed 2s. Red represents ground truth for the next 3 seconds and green represents the multiple forecasted trajectories for those 3s [24]
(⇒ copy of original 3GPP image)
(⇒ copy of original 3GPP image)
An AIML model can be for the object recognition and prediction, the model is offered by company-A and customers of company-A have signed "smart driving project" (an agreement for AIML model sharing and improvement).
6.2.3 Service Flows p. 22
- Bob bought a car equipped with intelligent driving functionality and he would like to use auto-driving for his daily driving, so he applies to company-A to offer the intelligent-driving service.
- Company-A needs to install certain AIML model to Bob's car while use Bob's local data to train the model. The company-A identified Alice's model to be shared to Bob's car. In order to minimize privacy issue, the "smart driving project" signed by customer only allows the model to be transferred among users directly instead of letting application server to acquire and forward it.
- Company-A requests 5G system to transmit the AIML model for intelligent driving from Alice's car to Bob's car via direct device connection at a proper time (e.g. when the direct device connection can be established)
- When acquiring the AI model from Alice's car, Bob's car performs "fine-tuning" operation of transfer learning based on the local data to tune the model to be better used for its own intelligent-driving service.
6.2.4 Post-conditions p. 22
Thanks to 5GS assisted AIML model transfer via direct device connection, Bob's car efficiently gets an ideal AIML model for intelligent-driving by means of transfer learning.
6.2.5 Existing features partly or fully covering the use case functionality p. 22
In clause 6.9 of TS 22.261 v18.6.1
The 5G system shall support different traffic flows of a remote UE to be relayed via different indirect network connection paths.
The connection between a remote UE and a relay UE shall be able to use 3GPP RAT or non-3GPP RAT and use licensed or unlicensed band.
The connection between a remote UE and a relay UE shall be able to use fixed broadband technology.
The 5G system shall be able to provide indication to a remote UE (alternatively, an authorized user) on the quality of currently available indirect network connection paths.
The 5G system shall be able to maintain service continuity of indirect network connection for a remote UE when the communication path to the network changes (i.e. change of one or more of the relay UEs, change of the gNB).
The 5G system shall be able to support a UE using simultaneous indirect and direct network connection mode.
The 5G system shall enable the network operator to authorize a UE to use indirect network connection. The authorization shall be able to be restricted to using only relay UEs belonging to the same network operator. The authorization shall be able to be restricted to only relay UEs belonging to the same application layer group.
In clause 6.40 of TS 22.261 v18.6.1
Subject to user consent, operator policy and regulatory requirements, the 5G system shall be able to expose information (e.g. candidate UEs) to an authorized 3rd party to assist the 3rd party to determine member(s) of a group of UEs (e.g. UEs of a FL group).
6.2.6 Potential New Requirements needed to support the use case p. 23
6.2.6.1 Potential Functionality Requirements p. 23
[P.R.6.2-001]
Based on user consent and 3rd party request, operator policy, the 5G system shall support a means to authorize specific UEs to transmit data (e.g. AI-ML model tansfer for a specific application) via direct device connection in a certain location and time.
[P.R.6.2-002]
Subject to user consent and operator policy, the 5G system shall be able to expose information to an authorized 3rd party to assist the 3rd party to determine candidate UEs for data transmission via direct device connection (e.g. for AIML model transfer).
6.2.6.2 Potential KPI Requirements p. 23
[P.R.6.2-003]
The 5G system shall be able to support transmitting an AI/ML model via direct device connection fulfilling the KPIs for transmission of typical AIML model for trajectory prediction and object recognition [24] [28] in Table 6.2-1.
| Payload size | Latency for model transmission (NOTE 1) | Transmission Data rate | |
|---|---|---|---|
| LaneGCN | 15 MByte | 3 seconds | 5 MByte/s |
| ResNet-50 | 25 MByte | 3 seconds | 8.33 MByte/s |
| ResNet-152 | 60 MByte | 3 seconds | 20 Mbyte/s |
| PointPillar | 18 MByte | 3 seconds | 6 MByte/s |
| SECOND | 20 MByte | 3 seconds | 6.67 MByte/s |
| Part-A2-Free | 226 MByte | 3 seconds | 75.33 MByte/s |
| Part-A2-Anchor | 244 MByte | 3 seconds | 81.33 MByte/s |
| PV-RCNN | 50 MByte | 3 seconds | 16.67 MByte/s |
| Voxel R-CNN (Car) | 28 MByte | 3 seconds | 9.33 MByte/s |
| CaDDN (Mono) | 774 MByte | 3 seconds | 248 MByte/s |
|
NOTE 1:
The transfer learning does not have a very high requirement for transmission latency since it is not a real-time inference service, hence it assumes the model transmission via direct device connection should be finished in 3 seconds.
|
|||
