FEDAKD: Federated Edge-Assisted Anomaly-Aware Knowledge Distillation for 5G Intrusion Detection

ICLR 2026 Conference Submission16957 Authors

Keywords: federated learning, knowledge distillation, edge-assisted learning, anomaly-aware sampling

Abstract: The rise of 5G networks has exponentially increased the complexity and volume of network traffic, thereby strengthening the challenges in ensuring robust intrusion detection. Federated Learning (FL) emerges as a promising paradigm for collaborative anomaly detection, enabling multiple distributed clients to train a shared model without exchanging raw data, thus preserving privacy. However, FL in 5G environments wrestles with class imbalance, heterogeneous anomaly distributions, and constrained computational resources at edge devices. To address these issues, we propose a novel Federated Edge-Assisted Anomaly-Aware Knowledge Distillation (FEDAKD) framework designed for 5G network intrusion detection. FEDAKD integrates anomaly-aware sampling, teacher-student transformer architectures, and advanced aggregation techniques such as FedProx to enhance model performance while minimizing computational overhead. We conduct extensive evaluations on a 5G-specific intrusion dataset, demonstrating that FEDAKD outperforms baseline methods, including centralized training, Federated Averaging, and non-transformer classifiers, achieving higher weighted F1 scores and more accurate detection of various attack types. The results of the experiment underscore FEDAKD's efficacy in delivering scalable, privacy-preserving, and high-performance intrusion detection in modern 5G networks.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 16957