SLArch: Arquitetura de Split Learning orientada a métricas de rede para desempenho de Redes Móveis B5G/6G

Abstract

Split Learning (SL) is a collaborative learning technique in which a neural network model is partitioned between client and server, enabling training without the need to share the original data. This study investigates the integration of SL with B5G/6G mobile networks using the ns-3 simulator and a convolutional neural network (CNN) trained on the MNIST dataset. We evaluated model-performance metrics, such as accuracy, as well as network indicators including packet loss, latency, throughput, and energy consumption. The results demonstrate that increasing transmit power reduces latency and improves model accuracy. The SL model exhibited performance variability with distance but maintained satisfactory accuracy (>80%) at distances up to 160 m in the highest-power scenario. These findings demonstrate the viability of SL as an enabling technology for next-generation mobile networks, optimising distributed training in communicationconstrained settings.