A Hierarchical Networking and Privacy-Preserving Federated Learning Framework for 5G Networks

Abstract

Artificial intelligence (AI) has been widely envisioned as a key enabler for 5G and beyond networks. To integrate AI into mobile networks, the third generation partnership (3GPP) introduces the network data analytics function (NWDAF) starting from Release 15 to support “in-network” learning and inference, and further supports federated learning (FL) from Release 16 to protect data privacy. However, practical deployment of federated learning in 5G networks still faces challenges of high communication overhead and potential risks of model and data leakage. Motivated by these challenges, we propose a hierarchical networking and privacy-preserving federated learning (HiNP-FL) framework for 5G networks. Specifically, in the HiNP-FL framework, 1) we propose the hierarchical NWDAF based FL mechanism to reduce FL communication overhead in 5G networks; 2) based on multi-party polynomial evaluation (OMPE), we design a FL model and data privacy protection mechanism for the hierarchical FL mechanism; 3) we validate the privacy protection capability of the HiNP-FL framework through privacy analysis, and testify its effectiveness in terms of model accuracy and communication efficiency through extensive experiments.