The 5G+ networks are going through extensive overhauling at both the radio access network (RAN) and core network (CN) to truly support the envisioned 5G+ use cases with ever so increasingly stringent quality of service requirements. Although many academic and industrial advancements have been realized, there are still needs for innovative changes vital for 5G+ connectivity, applications, and services. Consequently, the 3GPP is ushering in the next paradigm shift for CNs. CNs, which consist of several network functions (NFs) working cooperatively to service a RAN, are expected to exploit artificial intelligence (AI) in the form of an all-seeing-eye AI engine known as the Network Data Analytics Function (NWDAF). Its purpose is to optimize and automate network operation and management. However, accurate analytics require extensive machine learning (ML) training and as such, in this article, we present a distributed collaborative ML approach to facilitate said training. Accordingly, we present two variations of a distributed NWDAF architecture to enhance NF data localization, improve security, reduce control overhead during model training, shorten the training time, and finally, enhance the accuracy of the trained models by virtue of local testing on real-time network data. The first architecture distributes the global model training by exploiting the unique learning domain of each local model for each network function instance in a federated learning-based scheme while the second partitions the actual training network between multiple entities for each network function in a split learning-based scheme.
Read full abstract