Page 51 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 5 – Internet of Everything

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 5

RF‑BASED LOW‑SNR CLASSIFICATION OF UAVS USING CONVOLUTIONAL NEURAL NETWORKS

1
1
Ender Ozturk , Fatih Erden , Ismail Guvenc 1
1 Electrical and Computer Engineering, NC State University, Raleigh, NC 27606, United States
NOTE: Ender Ozturk, eozturk2@ncsu.edu

Abstract – Unmanned Aerial Vehicles (UAVs), or drones, which can be considered as a coverage extender for Internet of
Everything (IoE), have drawn high attention recently. The proliferation of drones will raise privacy and security concerns
in public. This paper investigates the problem of classi ication of drones from Radio Frequency (RF) ingerprints at the low
Signal‑to‑Noise Ratio (SNR) regime. We use Convolutional Neural Networks (CNNs) trained with both RF time‑series images
and the spectrograms of 15 different off‑the‑shelf drone controller RF signals. When using time‑series signal images, the CNN
extracts features from the signal transient and envelope. As the SNR decreases, this approach fails dramatically because the
information in the transient is lost in the noise, and the envelope is distorted heavily. In contrast to time‑series representation
of the RF signals, with spectrograms, it is possible to focus only on the desired frequency interval, i.e., 2.4 GHz ISM band, and
ilter out any other signal component outside of this band. These advantages provide a notable performance improvement
over the time‑series signals‑based methods. To further increase the classi ication accuracy of the spectrogram‑based CNN,
we denoise the spectrogram images by truncating them to a limited spectral density interval. Creating a single model using
spectrogram images of noisy signals and tuning the CNN model parameters, we achieve a classi ication accuracy varying
from 92% to 100% for an SNR range from –10 dB to 30 dB, which signi icantly outperforms the existing approaches to
our best knowledge.
Keywords – Convolutional neural networks (CNN), low SNR regime, RF ingerprinting, spectrogram, UAV classi ication

1. INTRODUCTION criminal activities recently with drones involved, and
their small sizes make it dif icult to detect, classify, and
Unmanned aerial vehicles (UAVs) or drones have re‑ interdict them [6, 7]. Latest surveys also demonstrate
cently gained a great deal of interest among researchers that 75% of the subjects exhibit privacy and security
due to unrivaled commercial opportunities in various concerns about all unmanned aerial use cases [8]. In
ields, such as wireless communications, logistics, deliv‑ this regard, Federal Aviation Agency (FAA) of the
ery, search and rescue, smart agriculture, surveillance, United States recently announced a Proposed Rule that
among others [1]. In addition, the recent COVID‑19 out‑
elaborates the future action that would require remote
break revealed the importance of remote operations in
identi ication of unmanned aircraft systems to address
every aspect of life, which may accelerate social accep‑
safety and security concerns [9]. UAVs can be identi ied
tance of drone use cases such as delivery of goods and
through a set of features that uniquely represent them.
medication [2, 3, 4]. With the new advances in airspace
These features can be extracted from various data
regulations and drone‑related technologies, it is expected
sources, such as visual data, acoustic, RF, or radar
that there will be more and more UAVs in the skies for var‑
signals. Each of these source types has its own pros
ious use cases, sharing the airspace with other aerial ve‑
and cons which we will review in the next section. Our
hicles [5]. The increase in daily drone usages can be con‑ contributions with this work are summarized below.
sidered in the context of The Internet of Everything (IoE),
a broader term than Internet of Things (IoT), aiming to • In this study, we develop a Convolutional Neural Net‑
include the entire realm of information sources and des‑ work (CNN)‑based classi ier using both time‑series
tinations in one paradigm. signal images and spectrogram images of 15 differ‑
ent drone controller RF signals to classify drones
Innate advantages of UAVs that make them popular, such of different makes and models. These signals are
as ease of operation and low cost, could also be consid‑ transmitted by proprietary circuit designs and con‑
ered as major disadvantages from security and privacy tain distinct ingerprints of commercially available
perspectives. This motivates detection, classi ication, and drones; they can be exploited by a machine learn‑
tracking of different types of UAVs, and interdicts unau‑ ing model to classify the make and model of the
thorized or malicious UAVs to maintain privacy and secu‑ drone. We use controller signals as the data set [27]
rity. Classi ication of UAVs can also be critical for forensics was already in possession; however, the proposed
use cases, e.g. for identifying a UAV after a malicious ac‑ approach can also be directly applied to the sig‑
tivity (e.g. eavesdropping, espionage) based on the cap‑ nals transmitted from drones to their controllers. A
tured signals of the UAV. There have been many lowchart of the overall procedure is given in Fig. 1.

© International Telecommunication Union, 2021 39