Research Article
RSEN-RFF: Deep Learning-Based RF Fingerprint Recognition in Noisy Environment
@INPROCEEDINGS{10.1007/978-3-030-90196-7_9, author={Zhaonan Du and Di Liu and Jiawen Zhang and Di Lin and Yuan Gao and Jiang Cao}, title={RSEN-RFF: Deep Learning-Based RF Fingerprint Recognition in Noisy Environment}, proceedings={Artificial Intelligence for Communications and Networks. Third EAI International Conference, AICON 2021, Xining, China, October 23--24, 2021, Proceedings, Part I}, proceedings_a={AICON}, year={2021}, month={11}, keywords={RF fingerprint recognition IoT security Deep learning}, doi={10.1007/978-3-030-90196-7_9} }- Zhaonan Du
Di Liu
Jiawen Zhang
Di Lin
Yuan Gao
Jiang Cao
Year: 2021
RSEN-RFF: Deep Learning-Based RF Fingerprint Recognition in Noisy Environment
AICON
Springer
DOI: 10.1007/978-3-030-90196-7_9
Abstract
As an emerging technology of Internet of Things security, radio frequency (RF) fingerprint identification technology can be used to identify wireless devices and meet the needs of the Internet of Things regarding user access control. Machine learning and deep learning have been applied to recognize a mobile device by extracting and analyzing its RF fingerprinting characteristics due to their powerful feature learning and representational abilities. However, the performance and accuracy of the learning algorithm will degrade dramatically in the circumstances of high-intensity noise (low signal-to-noise ratio, low SNR). To address this problem, this paper proposes an attention-based residual network algorithm, named RSEN-RFF, to train and recognize the RF fingerprint characteristics of lightweight mobile devices, and compared the proposed algorithm with classic convolutional neural networks (LeNet5, etc.), which are the most widely used algorithms for IoT device identification. Unlike other machine learning methods based on feature engineering, deep models use neural networks to solve the characterization of RF fingerprint features without the need for a process of feature extraction based on professional knowledge. The results show that the fitting speed and recognition accuracy of proposed algorithm are better than those of the previous algorithm under the condition of low SNR.
