5GS Assistance for Federated Learning Member Selection in Trajectory Prediction Scenarios

In federated learning (FL) over the 5G system (5GS), a group of user equipments (UEs) that are FL participants with diverse local datasets can accelerate the model training and enhance the model generalization. It is impractical to blindly select a large number of FL participants in order to derive a well-trained model. Instead, it would be more beneficial to select a small number of FL participants (i.e. UEs) with specific characteristics given the considerations of the mobile communication resource constraints. However, how to effectively select suitable UEs as FL participants is a challenging problem for application server on mobile targets, since it requires prior information about user characteristics, such as user location information and mobility-related information. This paper proposes a FL member selection method based on the assistance of a 3rd Generation Partnership Project (3GPP) 5GS to address the above problem. In particular, the trajectory prediction is considered as a typical scenario where 5GS has the ability to identify the list of best candidate UEs to participate in the FL according to the criteria as specified by the FL server. This paper illustrates the effective 5GS support for the application AI services from the Network for AI perspective.