Content for TR 22.876 Word version: 19.1.0
7.2 Asynchronous FL via direct device connection
7.2.1 Description
7.2.2 Pre-conditions
7.2.3 Service Flows
7.2.4 Post-conditions
7.2.5 Existing features partly or fully covering the use case functionality
7.2.6 Potential New Requirements needed to support the use case
...
...
7.2 Asynchronous FL via direct device connection p. 28
7.2.1 Description p. 28
Federated Learning (FL) is an important machine learning service. Due to the Synchronous FL [8], Sync-FL, requires a strict communication quality for each UE in order to get all the intermediate results to FL server in time, the Synchronous FL is sometimes vulnerable to the unpredicted wireless condition and the divergence of UEs' capabilities. Therefore, the Asynchronous FL [9], Async-FL, has been widely used in many circumstances. The main idea of Async-FL is to let UE report its result whenever it is ready and the FL server will refresh the model without waiting for all intermediates results are collected. The Sync-FL and Async-FL have pros and cons as the Table 7.2-1 shows.
| Sync-FL | Async-FL | |
|---|---|---|
| Total computation workload | Lower. | Higher. The UE will get a new model for training when it uploads the result without waiting for other UE's result. So the computation work load in each UE can be increased. |
| Communication requirement | Higher. All UEs shall report its result before next FL round starts | Lower. |
7.2.2 Pre-conditions p. 28
The direct device connection can be used to realize the Async-FL. As the Figure 7.2-1 shows, for some UEs who are in a bad coverage it can use the indirect network connection to communicate with Parameter Server (PS). The communication requirement via indirect network connection can be relaxed i.e. no need to transmitted all UEs training result with a restricted timing.
For each member UE, it can send its training result to the PS via either direct network connection or indirect network connection, and the PS will send a new model to the member UE without waiting for other Aggregate UEs' result (i.e. Async-FL).
7.2.3 Service Flows p. 29
- The Parameter Server (PS) distributes the global model to FL member UEs via direct network connection or indirect network connection; For the UEs in bad coverage, it can use the indirect network connection to perform a Async-FL with PS.
- When receiving the training result from member UE, the relay UE sends it to the Parameter server immediately to get a new model for the member UE. Due to the relay UE has a limited QoS for its own network connection (PDU session), the relay UE needs to determine the QoS for indirect network connection for each of member UEs based on an aggregated QoS (QoS upper limit) for the group of members served by the relay UE.
- The Async-FL will be performed until the model accuracy reached a certain threshold.
7.2.4 Post-conditions p. 29
Thanks to the indirect network connection, the FL server can still use the valuable data stored in UEs who are out of coverage with the method of Async-FL. The model training is finally finished with expected model performance.
The charging for an Remote UE using an Indirect 3GPP Communication will be done.
7.2.5 Existing features partly or fully covering the use case functionality p. 29
In clause 3.1 of TS 22.261 (v19.2.0)
aggregated QoS: QoS requirement(s) that apply to the traffic of a group of UEs.
In clause 6.9 of TS 22.261 (v19.2.0)
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 (v19.2.0),
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).
In clause 4.8 of TS 22.115 (V18.0.0) on "Charging Requirements for Indirect 3GPP Communication"
This section describes the requirements to enable operator collection of charging data for an Evolved ProSe Remote UE and Relay UE using an Indirect 3GPP Communication. The requirements also apply in the roaming case.
The 3GPP core network shall be able to collect charging data for an Evolved ProSe Remote UE which accesses the 3GPP core network through an Indirect 3GPP Communication.
7.2.6 Potential New Requirements needed to support the use case p. 30
[P.R. 7.2-001]
5GS shall be able to support an aggregated QoS for a group UEs served by a relay UE.
[P.R. 7.2-002]
5GS shall be able to provision an aggregated QoS to a relay UE for a group-based service.
[P.E. 7.2-003]
Based on 3rd party request and user consent, the 5G system shall be able to expose information (e.g. candidate UEs) to an authorized 3rd party to assist the 3rd party for UE member(s) selection in a group of UEs (e.g. UEs of a FL group), for UEs using direct or indirect network connection. E.g. the 3rd party request may include the expected QoS as a criteria for UE member selection.
