IEEE - Aerospace and Electronic Systems - May 2022 - 20

A Review of Security Incidents and Defence Techniques Relating to the Malicious Use of Small UASs
features from spectrogram images. Mohajerin et al. [85]
used a mixture of simulated UAS, bird, and plane radar
tracks and a range of statistical features for classification.
The work, however, does not fully take into account atmospheric
and environmental effects. Fioranelli et al. [86] used
multistatic RADAR and micro-Doppler signature analysis
to classify between UASs carrying different payloads. They
achieve accuracy over 90% using centroid features from the
micro-Doppler signature.
Ritchie et al. [87] observed that micro-UAS electromagnetic
scattering caused by the rotor blades produces
strong variations with azimuth and frequency. Jahangir
and Baker [88] used the holographic radar that uses a 2D
antenna array to create a multibeam 3D surveillance sensor.
Decision tree, a machine learning classifier, is used to
distinguish between false tracks successfully and has overall
detection of 88%. Lunden and Koivunen [89] used a
deep CNN to extract features from high range resolution
profiles (HRRPs) for multistatic radars. They show that
large target aircraft are distinguishable using this method
but the work does not extend to smaller targets, such as
UASs. Torvik et al. [90] used polarimetric parameters to
address the problem with radars distinguishing between
birds and UASs, which are comparable in size. They focus
on reducing critical detection time by using polarimetric
features with a nearest neighbor classifier. Oh et al. [91]
performed the automatic multicategory classification of
mini-UASs using empirical-mode decomposition. Statistical
and geometrical features are fed to machine learning
classifier support vector machine for prediction.
Kim et al. [92] were one of the first to consider a CNN
with merged Doppler images to perform classification of
UASs. They found frequency domain features to be more
robust than micro-Doppler signatures, which prior work had
concentrated on. Mendis et al. [93] used a deep belief network
and spectral correlation functions (SCFs) from a
Doppler radar. SCFs are chosen due to their resilience to
noise. They show that environments where SNR is less than
0 high levels ofaccuracy can be maintained. However, their
work does not consider UASs in motion but only in static
positions. Ren and Jiang [94] highlighted that existing
micro-Doppler spectrogram, cepstrogram, and cadence representations
do not include any phase information but only
magnitude. They address this using a 2D complex Fourier
transform and in doing so improve error rates for cadence
velocity diagrams. Zhang et al. [95] enhanced the microDoppler
signature robustness by using STFT spectrograms
in theK-band andX-band with principal component analysis
for feature extraction. A support vector machine is again
used as the machine learning classifier and they show that
the fusion ofmultiple radar sensors produces a higher accuracy
than a single radar feed. Regev et al. [96] highlighted
the issue that UASs present radars due to their low radar
cross section and when detection has occurred, distinguishing
the UAS from birds, or by UAS type. They use a
20
multilayer perceptron (MLP) neural network to classify
UAS type using the baseband signal from the radar return.
Fuhrmann et al. [97] classified the UAV types-quadcopters,
octocopters, helicopters, and fixed wing platforms of
various sizes using features extracted from time frequency
transforms including STFT, cadence velocity, and cepstrograms.
Using support vector machine classification accuracy
is produced at over 96%. Oh et al. [91] used the empiricalmode
decomposition for UAS classification. They use eight
statistical and geometrical features and SVM as the machine
learning classifier. They utilize the unique patterns in microDoppler
produced by the motion ofthe UAS blades. Ma et al.
[98] investigated entropies Shannon, spectral, log energy,
approximate, fuzzy, and permutation to enhance mini-UAS
classification accuracy. Support vector machine again is used
as themachine learning classifier. They show higher accuracy
than compared work but with increased computational power
requirements. Sun et al. [99] classified and localized UASs
using micro-Doppler signatures and dimensionality reduction.
Sun et al. show robust feature selection, which works at
lower frequencies. Habermann et al. [100] introduced a new
type of feature and use point cloud features generated from
the radar return to classify helicopters and UASs. The classification
uses anMLP artificial neural networks and can classify
between different types ofhelicopters and UAS, even in very
low-SNR environments.
Wang et al. [101] compared CNN detection methods
with traditional CFAR methods. The CNN had the highest
performance especially in low-SNR environments. They
also found that coincidence detection could improve CNN
results. However, their work is limited to simulated data and
is yet to be tested on real radar emissions. Samaras et al.
[102] highlighted issues with existing research requiring
long illumination times, and therefore, a tracking radar
architecture rather than a surveillance radar. This is due to
existing methods being based on the Fourier spectra. Samaras
et al. propose a deep learning solution for surveillance
radar data to distinguish between UASs, birds, and noise,
producing accuracy of95%. Choi and Oh [103] used a deep
CNN to classify micro-Doppler signatures from different
UASs. Wan et al. [104] proved that HRRP feature extraction
for automatic target recognition using a CNN and spectrograms.
However, again this work does not extend to smaller
targets, such as UASs. Guo et al. [105] used a 1D CNN to
overcome prior issues with HRRP sensitivity. Again the
work does not consider smaller UAS targets. Chen et al.
[106] developed motion models for UASs and flying birds
and then calculate the variance in the time domain of the
model occurrence probability to estimate the target before
identifying and classifying it. The model results are validated
using ground truth radar data from airport and coastal
environments.
Messina et al. [107] automatically classifiedUASs using
machine learning and surveillance radar signals. They show
classification between birds, planes, and cars at higher than
IEEE A&E SYSTEMS MAGAZINE
MAY 2022

IEEE - Aerospace and Electronic Systems - May 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - May 2022