Federated transfer learning-based intrusion detection system in 5G networks
The development of Intrusion Detection Systems (IDS) for the Internet of Things (IoT) and 5G networks is rapidly advancing. This study investigates the application of federated architectures to train detection models while preserving data privacy by eliminating the need for data sharing among device...
| Autores: | , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/451974 |
| Acceso en línea: | https://hdl.handle.net/2117/451974 https://dx.doi.org/10.1016/j.eswa.2025.130868 |
| Access Level: | acceso abierto |
| Palabra clave: | Intrusion detection system Federated learning Federated transfer learning Convolutional neural network Internet of things UNSW-NB15 dataset Bot-IoT dataset Weight sharing Robust aggregation functions Geometric median |
| Sumario: | The development of Intrusion Detection Systems (IDS) for the Internet of Things (IoT) and 5G networks is rapidly advancing. This study investigates the application of federated architectures to train detection models while preserving data privacy by eliminating the need for data sharing among devices. We propose a Federated Transfer Learning (FTL) model tailored for scenarios with unbalanced nodes, enhancing the detection capabilities for unknown attacks compared to conventional Federated Learning (FL) approaches. Utilizing the Bot-IoT dataset as the source domain and the UNSW-NB15 dataset as the target domain, our experiments reveal significant improvements in detection performance. Specifically, nodes characterized by lower proportions of malicious traffic demonstrate up to a 62.614% enhancement in detecting unknown attacks, increasing detection rates from 19.090% to 81.704%. Moreover, our findings indicate that FTL not only improves the identification of unknown threats but also maintains robust performance in detecting both attacks and benign traffic. Notably, the minimum accuracy achieved by the most imbalanced node reaches 0.912, in contrast to 0.741 with standard FL models. These results highlight the potential of FTL to train robust models across distributed nodes while ensuring privacy, thereby contributing to improved security measures in IoT and 5G networks. |
|---|