Federated learning framework for prediction based load distribution in 5G network slicing

Abstract

The 5G technology brings transformative changes across sectors like healthcare, automotive, and entertainment by integrating massive IoT networks and supporting dense device connectivity. Network slicing in 5G further ignites the capability by allowing tailored virtual networks for specific applications, enhancing operational efficiency and user experience across diverse scenarios. In this paper we propose a framework to use Federated Learning (FL) in 5G network slicing to support service assignment. The aim is to optimize the network traffic allocation among various slices. It first predicts the load on each network slice and then the incoming traffic is allocated to a slice which is most suitable and not heavily loaded. The DeepSlice dataset on 5G slicing is horizontally splited into multiple segments to train a federated CNN model which are deployed across multiple clients. The model is analyzed with varying number of clients and parameters such as accuracy and loss are observed. The performance of federated approach is compared with centralized approach of prediction keeping essential hyper parameters unchanged. Outcomes in terms of training and testing is presented for better interpretation of the proposed framework. Observation shows that the federated learning outperform the centralized technique in accuracy as well as loss.