Design of Tiny Contrastive Learning Network With Noise Tolerance for Unauthorized Device Identification in Internet of UAVs

Document Type

Article

Publication Title

IEEE Internet of Things Journal

Abstract

Artificial intelligence enhanced Internet of unmanned aerial vehicles (UAVs) is a promising network to achieve the complicated vehicular tasks and construct intelligent communication networks. One of the critical tasks is to guarantee a secure network access while achieving tradeoff between accuracy and latency through lightweight deployment on resource-limited and hardware-constrained UAVs. To address this issue, a novel noise-tolerant radio frequency fingerprinting (NT-RFF) based on tiny machine learning (TinyML) scheme is proposed, which amalgamates contrastive learning and data augmentation, aiming to improve the generalization ability of unauthorized device identification (UDI). Particularly, we first exploit the augmentation technique to enhance the legitimate training data sets under the circumstance of varying signal-to-noise ratios, facilitating an enhanced and diversified data sets. Second, a synthesis of contrastive learning and supervised learning is employed to attain comprehensive global learning. We design a new contrastive loss criteria to capture relevant information from the samples collected over the air. Besides, we design a categorical cross-entropy loss criteria by which supervisory information can be leveraged from associated labels. Finally, quantification is utilized to enhance model efficiency and achieve an optimal balance between accuracy and latency within computing and energy resource-limited UAVs. Experimental results demonstrate that the proposed tiny NT-RFF which only contains about 25-30% quantitative parameters can maintain excellent performance and improve the UDI accuracy greatly compared with the traditional machine learning-based RFF schemes. Moreover, the remarkable results showcase that our proposed framework attains a substantial increase in identification accuracy compared to the data augmentation and contrastive learning-based RFF and DASL-RFF methods, exhibiting improvements of 14.16% and 5.17%, respectively.

First Page

20912

Last Page

20929

DOI

10.1109/JIOT.2024.3376529

Publication Date

6-15-2024

Keywords

Contrastive learning, data augmentation, tiny machine learning (TinyML), unauthorized device identification (UDI), unmanned aerial vehicles (UAVs)

This document is currently not available here.

Share

COinS