FedVAE-KD: a privacy-preserving federated learning framework for scalable wireless network intrusion detection

Official URL: https://dx.doi.org/10.1109/BalkanCom71095.2026.11605017

Wireless networks face a critical challenge: effective intrusion detection requires data sharing, yet privacy regulations hinder centralized security solutions. We introduce FedVAEKD, a federated learning framework that combines variational autoencoders with knowledge distillation to enable secure, decentralized threat detection across heterogeneous wireless environments. FedVAE-KD facilitates collaborative model training among multiple operators without exposing raw traffic data. Key contributions include: (i) 17.2× model compression for deployment on resource-constrained edge devices, (ii) robust performance using only 10% of the training data, addressing data scarcity in IoT and mobile systems, and (iii) formal (ϵ,δ)-differential privacy guarantees for secure collaboration. Evaluated on KDD Cup 1999, NSL-KDD, UNSW-NB15, and CICDDOS2019 datasets, our framework achieved up to 99.92% accuracy with a 1.10× improvement in inference speed and demonstrated strong resilience to gradient inversion and model poisoning. FedVAE-KD enables a scalable, privacy-preserving solution for real-world 5G/6G, IoT, and mobile edge network defense.