5.46.1 General

This clause describes the list of 5GC enablers to support the following AI/ML operations in the Application layer over the 5G System:

• AI/ML operation splitting between AI/ML endpoints;
• AI/ML model/data distribution and sharing;
• Distributed/Federated Learning.

NOTE 1: Requirements on 5G System assistance to AI/ML operations in the Application layer are specified in clause 6.40 of TS 22.261 [2].

The AF may subscribe to NEF monitoring events as described in clause 4.15.3.2.3 of TS 23.502 [3] in order to assist its application AI/ML operations. For example, the AF may subscribe to session inactivity time monitoring event in order to assist the AI/ML application server in scheduling available UE(s) to participate in the AI/ML operation (e.g. Federated Learning). In addition, the AF may subscribe to NEF to be notified on the traffic volume exchanged between the UE and the AI/ML application server in order to assist the AF with the transfer of AI/ML data.

The AF that aims to provide an AI/ML operation may request assistance from the 5GC as described in clause 5.46.2. The AF initially provides a list of target member UE(s) and at least one filtering criterion, when subscribing to the NEF to be notified about a subset list of UE(s) (i.e. list of candidate UE(s)) that fulfil certain filtering criteria. Details of the procedures are described in clause 4.15.13 of TS 23.502 [3]. This subset list of UE(s) may become the member UE(s) used in the AI/ML operation depending on the AF's final decision, considering its internal logic. Alternatively, the AF may select the list of UEs for the AI/ML operation without NEF involvement as described in Annex I of TS 23.502 [3]: in this case, the AF determines a list of UEs without any assistance from the NEF and may use, for example, NWDAF analytics to assist with the AI/ML operation over 5G System.

The AF may request the network to provide a recommended time window for the AI/ML operation using the Planned Data Transfer with QoS (PDTQ) requirements and procedures as described in clause 6.1.2.7 of TS 23.503 [45] and in clause 4.16.15 of TS 23.502 [3].

At the time or during the AI/ML operation, the AF may request the serving NEF to provide QoS for a list of UEs. Each UE is identified by its UE IP address. The AF may subscribe to QoS Monitoring which may include also Consolidated Data Rate monitoring as described in clause 5.45 and in clause 4.15.6.13 of TS 23.502 [3] for those AF requests for QoS that result in a successful resource allocation. The AF provides QoS parameters that are derived from the performance requirements listed in clause 7.10 of TS 22.261 [2].

As a result of the AF subscription to NEF to provide the subset list of UE(s) that fulfil certain filtering criteria, the AF may be notified about changes in the subset list of UE(s) and in such a case the AF may determine a updated list of UEs used in the AI/ML operation from the new subset list of UE(s) provided by NEF and the AF may request a new recommended time window for the AI/ML operation using the Planned Data Transfer with QoS (PDTQ) requirements as described in clause 6.1.2.7 of TS 23.503 [45] and in clause 4.16.15 of TS 23.502 [3]. The AF may request the NEF to provide QoS for the updated list of UEs, each identified by UE IP address, that results in QoS resources previously allocated to some UEs to be released, while QoS resources for other UEs to be allocated and QoS monitoring to be initiated.

The AF may also subscribe to, or request Network Data Analytics as defined in TS 23.288 [86], such as End-to-end data volume transfer time analytics, DN performance analytics, Network performance analytics, UE mobility analytics, WLAN performance analytics etc. in order to assist its AI/ML operations.

The AF hosting an AI/ML based application may provision the Expected UE Behaviour parameters captured in Table 4.15.6.3-1 of TS 23.502 [3] and/or the Application-Specific Expected UE Behaviour parameter(s) captured in Table 4.15.6.3f-1 of TS 23.502 [3] to the 5GC. The parameters may be provisioned with corresponding confidence and/or accuracy levels and a threshold may also be provided to the UDM by the NF (e.g. AMF or SMF) subscribing to such externally provisioned parameters as described in clause 4.15.6.2 of TS 23.502 [3].

In addition, the following principles apply when 5GS assists the AI/ML operation at the application layer:

• AF requesting 5GS assistance to AI/ML operations in the application layer shall be authorized by the 5GC using the existing mechanisms.

NOTE 2: In this Release, assistance to AI/ML operations in the application layer is not supported for roaming UEs.

NOTE 3: Policy and charging control as defined in TS 23.503 [45] can be used for traffic related to AI/ML operations as described in clause 6.40.1 of TS 22.261 [2]. Capabilities based on Flow Based Charging can be used together with an appropriate PCF configuration for charging differentiation between AI/ML traffic and other type of traffic from the same application. As long as the AF can provide different filter information as described in TS 23.503 [45] for the AI/ML traffic and the other type of traffic from the same application in the procedures utilized for the resource request, the PCF can assign different rating groups (i.e. charging key) or rating group + service-ids based on an appropriate PCF configuration. The UPF will then handle such traffic accordingly. This enables charging of AI/ML traffic according to operator's policies.

• Application AI/ML decisions and their internal operation logic reside at the AF and UE application client and is out of scope of 3GPP.
• Based on application logic, it is the application decision whether to request assistance from 5GC, e.g. for the purpose of selection of Member UEs that participate in certain AI/ML operation.
• In this Release, 5GS assistance to AI/ML operations in the application layer is conducted within a single slice.