Cross-Silo Federated Domain Adaptation for Trustworthy and Privacy-Preserving IoT Security in 5G-to-6G Wireless Systems

Abstract

The convergence of large-scale Internet of Things (IoT) deployments with emerging 5G-to-6G wireless systems introduces unprecedented security challenges due to extreme network heterogeneity, evolving traffic dynamics, and stringent privacy requirements. Deep learning (DL)-based intrusion detection systems (IDS) often suffer from severe performance degradation when deployed across heterogeneous domains, while centralized training is increasingly impractical in 6G multistakeholder ecosystems due to data privacy and trust constraints. This paper proposes a unified, privacy-preserving, and trustworthy security framework that integrates cross-silo federated learning (FL), progressive domain adaptation (DA), and knowledge distillation (KD) to enable generalizable intrusion detection for next-generation wireless systems. The framework first performs federated pre-training across distributed IoT silos to construct a global backbone model without sharing raw data, supporting privacy-by-design principles essential for 6G architectures. A progressive two-stage DA strategy based on correlation alignment is then employed to mitigate distributional discrepancies between IoT and 5G-to-6G traffic domains. Finally, KD compresses the adapted teacher model into a lightweight student network suitable for real-time deployment on resource-constrained edge devices. Extensive evaluations on the IDSIoT2024 and 5G-NIDD datasets demonstrate that the proposed approach achieves over 90% accuracy on the source domain while maintaining 78-80% accuracy and F1-scores above 0.82 on an unseen target domain, with a compact model size of $0.19 M B$ and sub-millisecond inference latency. These results highlight the framework’s effectiveness in delivering trustworthy, privacy-aware, and adaptive security intelligence, positioning it as a practical enabler for AI-enabled 6G wireless systems.