Year: 2025
Type: article
Abstract: Accurate identification of IoT devices is essential for network management and security enforcement in increasingly heterogeneous environments. Existing fingerprinting approaches rely on static subspace modeling and shallow fusion strategies, which limit their adaptability and discriminative power. In this paper, we propose A-IDFE (Adaptive IoT Device Fingerprint Extraction), a novel deep learning framework that enhances fingerprint robustness, adaptability, and generalization for device identification. A-IDFE introduces (a) a unified encoder with adaptive channel attention to capture heterogeneous traffic subspaces efficiently, (b) a Transformer-based fusion module that models cross-subspace dependencies, (c) a discriminative embedding learning objective using ArcFace to improve inter-class separability, and (d) a robustness-aware training strategy incorporating perturbation and adversarial augmentations. Extensive experiments on public and private IoT datasets demonstrate that A-IDFE outperforms existing methods in both identification accuracy and robustness against noisy, adversarial, and cross domain conditions, making it a promising solution for practical IoT device fingerprinting.
Source: 電腦學刊
Authors: Kang-Qian Huang, Xi-Jun Lin, Jia-Jun Liu, Ying-Kai Zheng, Xing-Wu Wang, Ya-Li Gao
Institution: Beijing University of Posts and Telecommunications
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Funder: China Southern Power Grid
OpenAlex ID: W4413833618