Optimizing network slice classification using federated learning and hybrid fusion with knowledge distillation approach

Abstract

With the advancement of 5G and the anticipation of 6G networks, dynamic and efficient network slicing has become crucial for meeting diverse Quality of Service (QoS) requirements across IoT applications. This study proposes a multi-stage framework that combines traditional machine learning, federated learning (FL), and knowledge distillation to classify network slices—eMBB, URLLC, and mMTC—using real-world traffic features. The approach begins with a federated learning model trained across decentralized, non-IID client datasets using average, max, and min aggregation strategies to ensure privacy and robustness. To enhance performance and reduce model complexity, an XGBoost teacher model generates soft labels and structural features which are transferred to a student neural network via knowledge distillation and feature fusion. This fusion model achieves a classification accuracy of 98%, outperforming conventional classifiers and the FL models in isolation. The methodology offers a scalable, privacy-preserving solution for real-time slice classification in next-generation mobile networks, making it particularly suitable for latency-sensitive and resource-constrained edge environments.