Detection and Classification of UAVs Using RF Fingerprintsin The Presence of Wi-Fi and Bluetooth Interference

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2019.2955889, IEEE Open
Journal of the Communications Society
1
Detection and Classification of UAVs Using RF Fingerprints

in the Presence of Wi-Fi and Bluetooth Interference
Martins Ezuma, Fatih Erden, Chethan Kumar Anjinappa, Ozgur Ozdemir, Member, IEEE, and
Ismail Guvenc, Senior Member, IEEE
This paper investigates the problem of detection and classification of unmanned aerial vehicles (UAVs) in the presence of wireless
interference signals using a passive radio frequency (RF) surveillance system. The system uses a multistage detector to distinguish
signals transmitted by a UAV controller from the background noise and interference signals. First, RF signals from any source are
detected using a Markov models-based naïve Bayes decision mechanism. When the receiver operates at a signal-to-noise ratio (SNR)
of 10 dB, and the threshold, which defines the states of the models, is set at a level 3.5 times the standard deviation of the preprocessed
noise data, a detection accuracy of 99.8% with a false alarm rate of 2.8% is achieved. Second, signals from Wi-Fi and Bluetooth
emitters, if present, are detected based on the bandwidth and modulation features of the detected RF signal. Once the input signal
is identified as a UAV controller signal, it is classified using machine learning (ML) techniques. Fifteen statistical features extracted
from the energy transients of the UAV controller signals are fed to neighborhood component analysis (NCA), and the three most
significant features are selected. The performance of the NCA and five different ML classifiers are studied for 15 different types
of UAV controllers. A classification accuracy of 98.13% is achieved by k-nearest neighbor classifier at 25 dB SNR. Classification
performance is also investigated at different SNR levels and for a set of 17 UAV controllers which includes two pairs from the same
UAV controller models.
Index Terms—Interference, machine learning, Markov models, RF fingerprinting, unmanned aerial vehicles (UAVs), UAV detection
and classification.
I. I NTRODUCTION frequency (RF) signals from UAV controllers. In [7], drone

pilots are identified by analyzing RF signals captured from
NMANNED aerial vehicles (UAVs), or drones, are be-
U coming ubiquitous in modern society. The recent popu-
larity of UAVs is mainly due to the advancement in micro-
the drone controllers. The pilots’ behavioral biometrics can be
identified from the captured signals using machine learning
(ML) techniques. The ML algorithms are trained using the
electro-mechanical systems-based precision sensors, such as
RF signals when the controller is handled by the legitimate
inertial motion units and gyroscopes, which are used for guid-
owner of the device. That way, it is possible to identify
ance, navigation, and control of UAVs. Consequently, UAVs
different drones and their pilots. However, while the behavioral
have become relatively cheap and affordable. They are finding
biometrics of drone pilot is an important information for
new applications in areas such as surveillance, smart policing,
drone detection, an adversary could be anyone whose behavior
search and rescue missions, infrastructure inspections, package
metrics we have no prior knowledge of. Therefore, in order
delivery, and precision agriculture [2]. Judging by the current
to accurately detect and identify an adversary drone, one
trend in UAV applications, it is expected that UAVs will
should focus on identifying the intrinsic signature of the drone
become an integral part of modern society. However, there
controller itself. These intrinsic signatures can be extracted
are security and privacy issues associated with the ubiquity of
from the RF signals transmitted by the UAV controllers and
UAVs.
referred to as the RF fingerprints of the controllers.
In recent times, UAVs have been used in ways that intro- Since the communication signals of most commercial and
duce a threat to public safety [3]. There have been several hobby grade UAVs are transmitted in the same frequency band
instances where hobby drones have been used to transport as Wi-Fi and Bluetooth transmissions, it becomes challenging
illegal drugs across prison walls. In addition, drones have to detect and identify RF signals from the UAV controllers
carried out espionage attacks which pose serious risk to public in the presence of these interferers. Moreover, surveillance
safety. Recently, drones operated by dissidents have flown and electronic warfare systems should be able to differentiate
into sensitive national infrastructures like nuclear reactors UAVs from different manufacturers. For instance, the identity
and airports [4]. Moreover, drones are becoming tools for of a UAV can provide useful information about the payload,
cyberattack and terrorism. For instance, Wi-Fi sniffing UAVs operational range, control signal characteristics (e.g., for jam-
can eavesdrop on smartphone users and steal sensitive data ming such signals), and the threat capability of the associated
without being detected [5], [6]. UAV. Accurate identification of UAVs is also important in
Considering the security and privacy issues associated with digital forensic analysis of aerial threats.
UAVs, accurate detection and classification of these vehicles In this work, we propose a multistage UAV detection and
are vital to public safety and national security. One promising an ML-based classification system for identifying 17 different
technique for UAV detection is based on the analysis of radio UAV controllers in the presence of wireless interference, i.e.,
Wi-Fi and Bluetooth devices. The multistage UAV detection
This work has been supported in part by NASA under the Federal Award
ID number NNX17AJ94A. This paper was presented in part at the IEEE system consists of two detectors. The first detector employs
Aerospace Conference, Big Sky, Montana, Mar. 2019 [1]. First three authors a two-state Markov model based naïve Bayes algorithm in
of this paper have equal contributions. deciding if the captured data contains RF signals or not.
All the authors are with the Department of Electrical and Computer
Engineering, North Carolina State University, Raleigh, NC 27606 (e-mail: Once an RF signal is detected, the second stage detector
{mcezuma, ferden, canjina, oozdemi, iguvenc}@ncsu.edu). decides if the signal comes from a UAV controller or an
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2019.2955889, IEEE Open
2
interference source. Given that the detected RF signal is from achieve an average accuracy of 94.43% and 96.13%,
an interference source, the source class is identified as Wi-Fi respectively. However, in [1], DA and NN achieve an
or Bluetooth. On the other hand, if the detected signal is from average accuracy of 88.15% and 58.49%, respectively,
a UAV controller, the signal is transferred to the ML-based when used to classify 14 UAV controllers.
classification system to determine the make and model of the 3) We study the confusion that results when attempting to
UAV controller. In an earlier work [1], the authors proposed classify UAV controllers of the same make and model.
a system for detecting and classifying 14 different UAV con- This is important in digital forensic analysis and de-
trollers. The system design assumes the absence of interference tecting decoys in surveillance systems. To investigate
signals. However, this assumption is not always correct. The this confusion, we included two pairs of identical UAV
contributions of the current work are summarized as follows: controllers to a pool of 13 different UAV controllers. That
is, we capture control signals from 17 UAV controllers
1) The paper investigates the problem of detecting and and evaluate the ability of the proposed classification
classifying signals from UAV controllers in the presence system at different SNRs. For an SNR of 25 dB, kNN
of co-channel wireless interference. We consider inter- and RandF achieve accuracy of 95.53% and 95.18%,
ference from Wi-Fi and Bluetooth sources and describe respectively, when the three most significant RF features
a methodology to detect the UAVs. The interference are used. To the best of our knowledge, past studies on
detection ensures the proposed UAV detection system is UAV classification using RF techniques considers only
robust against false alarms and missed target detection. In a limited number of different make and model UAV
addition, in [1], we used two fixed thresholds, positioned controllers, often less than 10 [8], [9].
at ± 3σ, to transform the captured signal into three- The remainder of the paper is organized as follows:
state Markov models, where σ is the standard deviation Section II provides a brief overview of the related work.
of the noise signal in the environment. However, in the Section III describes the multistage detection system, and
current work, we use a single threshold to transform Section IV introduces the methodology to detect Wi-Fi and
the captured signal into two-state Markov models, which Bluetooth interference signals. Feature extraction and the RF
reduces overall complexity. We also define a procedure fingerprinting-based UAV classification system are explained
to determine the optimum threshold value based on the in Section V. The experimental setup and data capture tech-
available training data. It turns out that better detection nique is described in Section VI while the detection and
accuracy can be achieved when a single but properly classification results are presented in Section VII. The paper
selected threshold is used to generate the Markov models. is concluded in Section VIII.
At an SNR of 10 dB, the current work achieves a
detection accuracy of 99.8% using a threshold that is II. R ELATED W ORK
3.5 times the standard deviation. However, in [1], the
UAV detection and classification through RF signals can be
detection accuracy is 84% under the same SNR condition.
grouped into two major headings: RF physical layer features-
Besides, in the current study, we evaluate the detection
based and RF medium access control (MAC) layer features-
performance for different thresholds based on the false
based techniques. In general, these techniques use an RF sens-
alarm rate (FAR).
ing device to capture the RF communication signal between a
2) We introduce the concept of energy transient for the
UAV and its controller.
extraction of RF-based features and show how effective it
is for the classification of the UAV controller signals. The
energy transient is computed using the representation of A. RF Physical Layer Features-Based Techniques
the RF signals in energy-time-frequency domain. From Most of the techniques classified within this category rely on
the energy transient, 15 statistical features are extracted the physical layer characteristics of the RF transmission from
for the UAV classification. The performance of five a UAV to its controller (or vice versa), such as the amplitude
different ML algorithms are compared using the proposed envelope or the spectrum of the RF signal. These techniques
RF fingerprinting technique. In addition, we investigate are sometimes referred to as RF fingerprinting techniques
the neighborhood component analysis (NCA) as a prac- because they utilize the unique characteristics of the RF signals
tical algorithm for feature selection in the classification for the detection and classification of the UAVs. Experimental
problem. The classification results using the three most investigations show that most of the commercial UAVs have
significant features, selected by the NCA, are compared unique RF signatures which is due to the circuitry design and
with those when all the 15 RF features are used. We also modulation techniques employed. Therefore, RF fingerprints
evaluate the classification performance at different signal- extracted from the UAV or its remote controller signals can
to-noise ratios (SNRs). For an SNR of 25 dB, the results be used as a basis for the detection and classification of the
show that the k-nearest neighbor (kNN) and random UAVs.
forest (RandF) machine learning algorithms are the best In [10], RF fingerprints of the UAV’s wireless control
performing classifiers, achieving accuracy of 98.13% and signals are extracted by computing the amplitude envelope
97.73%, respectively, when the three most significant RF- of the signal. The dimensionality of the processed signal is
based features are used for the classification of 15 UAV reduced by performing principal component analysis (PCA),
controllers. In comparison, the kNN classifier achieves and the lower-dimensional data is fed into an auxiliary
an accuracy of 96.3% when used to classify 14 UAV classifier Wasserstein generative adversarial networks (AC-
controllers [1]. Furthermore, in the current work, for WGANs). The AC-WGANs achieves an overall classification
the case of 15 UAV controllers, DA and NN classifiers rate of 95% when four different types of UAVs are considered.
3
UAV High-resolution oscilloscope

Wi-Fi interference sources
Captured
waveform
UAV remote
controller
Bluetooth interference source
Fig. 1. The scenario of the RF-based UAV detection system. The passive RF surveillance system listens for the signal transmitted between the controller and
the UAV. The environment contains signals from Wi-Fi and Bluetooth interference devices which operate in the same frequency band with the UAV and its
remote controller.
In [11], drones are detected by analyzing the RF background In general, a major concern with the Wi-Fi fingerprinting
activities along with the RF signals emitted when the drones techniques is the privacy. This is because the same Wi-
are operated in different modes. Afterward, RF spectrum of the Fi detection system can spoof Wi-Fi traffic data from a
drone signal is computed using the discrete Fourier transform smartphone user or a private Wi-Fi network. In addition, only
(DFT). The drone classification system is designed by training a limited number of commercial drones employ Wi-Fi links
a deep neural network with the RF spectrum data of different for video streaming and control. Most commercial drones use
drones. The system shows an accuracy of 99.7% when two proprietary communication links.
drones are classified, 84.5% with four drones, and 46.8% with Besides RF and Wi-Fi fingerprinting techniques, several
ten drones. other techniques have been investigated for UAV detection, in-
In [12], an industry integrated counter-drone solution is cluding radar-based techniques, acoustic techniques, and com-
described. The solution is based on a network of distributed puter vision techniques [14]. However, as discussed in [14],
RF sensors. In this system, RF signals from different UAV traditional radar systems are not so effective in detecting UAVs
controllers are detected using an energy detector. Afterward, with small radar cross sections, and acoustic and computer
the signals of interest are classified using RF spectral shape vision-based techniques are greatly impaired by ambient en-
correlation features. Besides, distributed RF sensors make it vironmental conditions. In contrast, RF techniques are not
possible to localize the UAV controller using time difference limited by these problems. We start by describing the design
of arrival (TDoA) or multilateration techniques. However, this of the multistage detector of our proposed RF-based system.
industrial solution is quite expensive.
III. M ULTISTAGE UAV S IGNAL D ETECTION
We consider the scenario shown in Fig. 1, where a pas-
B. RF MAC Layer Features-Based Techniques sive RF surveillance system listens for the control signals
There are many UAVs that use Wi-Fi protocol for video transmitted between a UAV and its remote controller. The
streaming and control. The techniques categorized under this main hardware components of the surveillance system are
heading use MAC layer features, such as packet statistics, 2.4 GHz RF antenna and a high-frequency oscilloscope, which
for detection and classification of the Wi-Fi controlled UAVs. is capable of sampling the captured data at 20 GSa/s. Instead
These techniques are sometimes referred to as Wi-Fi finger- of an oscilloscope, a standard software-defined radio like the
printing techniques. Thus, the RF detection system consists universal software radio peripheral (USRP) can also be used
primarily of a Wi-Fi packet-sniffing device, which can inter- for data capture. In order to avoid aliasing, the data capture
cept the Wi-Fi data traffic between a UAV and its remote device should be able to sample the captured data above the
controller. In [9], unauthorized Wi-Fi controlled UAVs are Nyquist rate. In this study, since we are interested in capturing
detected by a patrolling drone using a set of Wi-Fi statistical RF data in the 2.4 GHz band, the data capture device should
features. The extracted features include MAC addresses, root- be able to sample at a rate of at least 5 MSa/s. Besides, if the
mean-square (RMS) of the Wi-Fi packet length, packet dura- RF surveillance system is passive as described in Fig. 1, then
tion, average packet inter-arrival time, among others. These it increases the stealth attribute of the detection system. This
features are used to train different ML algorithms which implies, the system shown in Fig. 1 can detect an adversary
perform the UAV classification task. In [9], the random tree UAV while itself remaining undetected by the UAV. This
and random forest classifiers achieve the best performance as stealth attribute is vital in electronic warfare environments
measured by the true positive and false positive rates. where low probability of intercept (LPI) emitters are very
In [13], drone presence is detected by eavesdropping on valuable. Furthermore, the passive RF detection system has
Wi-Fi channels between the drone and its controller. The an advantage over a radar system in terms of the maximum
system detects drones by analyzing the impact of their unique detection range. This is because why a radar would have to
vibration and body shifting motions on the Wi-Fi signals transmit pulses and listen for the backscattering (echo) from
transmitted by the drone. The system achieves accuracy above the target, the passive RF detector only needs to listen for the
90% at 50 meters. signals from the target.
4
0.2 0.2 0.2 0.2
Amplitude (Volts)
Amplitude (Volts)
Amplitude (Volts)
Amplitude (Volts)
0.1 0.1 0.1 0.1
0 0 0 0
-0.1 -0.1 -0.1 -0.1
-0.2 -0.2 -0.2 -0.2

0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time (s) 10-4 Time (s) 10-4 Time (s) 10-4 Time (s) 10-4
(a) (b) (c) (d)
0.2 0.2 0.2 0.2
Amplitude (Volts)
Amplitude (Volts)
Amplitude (Volts)
Amplitude (Volts)
0.1 0.1 0.1 0.1
0 0 0 0
-0.1 -0.1 -0.1 -0.1
-0.2 -0.2 -0.2 -0.2

0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
(e) (f) (g) (h)
0.2 0.2 0.2 0.2

Amplitude (Volts)
Amplitude (Volts)
Amplitude (Volts)
Amplitude (Volts)
0.1 0.1 0.1 0.1
0 0 0 0
-0.1 -0.1 -0.1 -0.1
-0.2 -0.2 -0.2 -0.2
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
(i) (j) (k) (l)
Fig. 2. RF signals captured from eight different UAV controllers and four different UAVs while on flight: (a) Graupner MC-32, (b) Spektrum DX6e, (c)
Futaba T8FG, (d) DJI Phantom 4 Pro, (e) DJI Inspire 1 Pro, (f) JR X9303, (g) Jeti Duplex DC-16, (h) FlySky FS-T6, (i) DJI Matrice 600 UAV, (j) DJI
Phantom 4 Pro UAV, (k) DJI Inspire 1 Pro UAV, (l) DJI Mavic Pro.
Since most commercial UAVs operate in the 2.4 GHz band, can be exploited for identifying the source UAV controller. The
the passive RF surveillance system is designed to operate flowchart in Fig. 3 provides a high-level graphical description
in this frequency band. However, this also corresponds to of the entire system. The first step in detecting and identifying
the operational band of Wi-Fi and mobile Bluetooth devices. a UAV controller is data capture. Usually, the captured raw sig-
Therefore, in real wireless environment, signals from these nal has a large size and is often very noisy. Therefore, before
wireless sources will act as interference to the detection of detection and classification, the signals are first pre-processed
the UAV control signals. Also, in such real environment, the using multiresolution analysis. Next, the processed signals are
presence of noise may further reduce the chance of correctly transferred to the multistage detection system, which consists
detecting the UAV signals when present. of two stages. In the first stage, the detector employs naïve
Given the scenario in Fig. 1, the passive RF surveillance Bayesian hypothesis test in deciding if the captured signal is an
system has to decide if the captured data comes from a RF signal or noise. If the decision is positive, the second stage
UAV controller, an interference source, or background noise. detector is activated to decide if the captured RF signal comes
In the case where the captured data comes from a UAV from an interference source or a UAV controller. This detector
controller, the detection system should be able to correctly uses bandwidth analysis and modulation-based features for
classify the UAV controller. However, if the detected signal interference detection. If the detected RF signal is not from a
is from an interference source, the detection system should Wi-Fi or Bluetooth interference source, it is presumed to be
be able to correctly identify the source, i.e., a Wi-Fi or a signal transmitted by a UAV controller. Consequently, the
a Bluetooth device. Therefore, the detection problem is detected signal is transferred to an ML-based classification
a multi-hypothesis problem. For such problems, it is well system for accurate identification of the UAV controller.
known that computational complexity increases as the number
of hypothesis increases. Consequently, the multi-hypothesis
detection problem can be simplified by using a multistage A. Pre-processing Step: Multiresolution Analysis
sequential detector. In this system, each detection stage is a Captured RF data are pre-processed by means of wavelet-
simple binary hypothesis test which is much easier to solve. based multiresolution analysis. It has been established that
Fig. 2 illustrates sample RF signals captured from eight dif- multiresolution decomposition using discrete wavelet trans-
ferent UAV controllers and four different UAVs (on flight). The form (DWT) like the Haar wavelet transform is effective for
figure shows each signal has different characteristics, which analyzing the information content of signals and images [15].
5
4
104 4
10
4
Capture
Data
2 2
y [n]
0 0
-2 -2
No
-4 -4
Multiresolution yT [n] Yes
Naïve Bayesian Signal 0 1 2 3 4 5 0 0.2 0.4 0.6 0.8 1 1.2
Analysis Detector Detected ? 106 10
6
(a) (b)
Yes
Fig. 5. (a) The sampled raw data y[n] captured from the remote controller
Compute
Bandwith No
Modulation
Bluetooth of a DJI Phantom 4 Pro UAV using an oscilloscope with a sampling rate
≥ 20 MHz Signal ? of 20 GSa/s, and (b) the transformed data yT [n] obtained at the output of
Features
Yes the two-level Haar wavelet filter. Due to successive downsampling, yT [n] has
Yes No about 3.8 × 106 fewer data samples than y[n].
the bias in the signals, leading to a higher detection accuracy,

Wi-Fi Bluetooth
Source Compute
Compute Source which is required in applications like UAV threat detection.
Energy
Spectrogram
Trajectory Fig. 5 shows the effect of the Haar wavelet decomposition
on a sample signal captured from the controller of the DJI
Phantom 4 Pro UAV. It is clear from the figure that the wavelet
Machine
UAV Training Learning
Extract RF
Fingerprints
transform removes the bias in the signal alignment and reduces
Database Classifier
the data size. It will be shown in Section V, the transformation
also preserves the characteristics of the original waveform.
After the pre-processing step, the data is transferred to the
UAV first stage of the detection system, where we decide if the
Controller
ID
captured data is an RF signal or noise.
Fig. 3. The system flowchart providing a graphical description of information B. Naïve Bayes Decision Mechanism for RF Signal Detec-
processing and flow of data through the system. tion
In this stage, we first model the pre-processed RF data,
Input
g [n] 2
d1 [n] yT [n], using two-state Markov models for “RF signal" and
y [n] d2 [n] Output
g [n] 2 “noise" classes. This allows us to compute the likelihood
a1 [n]
yT [n]
h [n] 2 that the captured data come from either the signal or noise
a2 [n] class. According to the Bayesian decision theory, the optimum
h [n] 2
detector is the one that maximizes the posterior probability.
Fig. 4. The two-level discrete Haar wavelet transform for pre-processing of Mathematically, let C ∈ {0, 1} be an index denoting the class
the captured raw data. of the pre-processed RF data yT [n], where C = 1 when the
captured raw signal y[n] is an RF signal, and C = 0 otherwise.
Let SyT = [SyT (1), SyT (2), . . . , SyT (N )]> be the state vector
In this work, multiresolution decomposition of the captured representation of the given test data yT [n] containing N
RF data are carried out using the two-level Haar transform as samples, with SyT (i) ∈ {S1 , S2 }, i = 1, 2, .., N , and S1 and S2
shown in Fig. 4. Using this transform, the raw input signal being the two states in the Markov models. Then, the posterior
is decomposed into subbands, and important time-frequency probability of the RF signal class given SyT is
information can be extracted at different resolution levels [16]. P (SyT |C = 1)P (C = 1)
In the first level, the input RF data are split into low- and high- P (C = 1|SyT ) = , (1)
P (SyT )
frequency components by means of the half band low-pass
(h[n]) and high-pass (g[n]) filters, respectively. This process is where P (SyT |C = 1) is the likelihood function conditioned
followed by a dyadic decimation, or downsampling, of the fil- on C = 1, P (C = 1) is the prior probability of the RF
ter outputs to produce the approximate coefficients, a1 [n], and signal class, and P (SyT ) is the evidence. A similar expression
detail coefficients, d1 [n]. In the second level, a1 [n] coefficients holds for the posterior probability P (C = 0|SyT ). In practice,
are further decomposed in a similar manner, and the generated since the evidence is not a function of C, it can be ignored.
d2 [n] coefficients are taken as the final output (yT [n]). Then, Therefore, we are only interested in maximizing the numerator
yT [n] is input to the multistage detection system. Moving from in (1). That is,
left to right in Fig. 4, we get coarser representation of the b = arg max P (Sy |C)P (C).
captured RF data. The output RF data will have fewer samples C T (2)
C
due to the successive downsampling of the input RF data. This We decide that the captured signal belongs to an RF signal
reduces the computational complexity of the overall process. (i.e., C = 1), if
Multiresolution analysis is also useful in detecting weak
signals in the presence of background noise and removing P (SyT |C = 1)P (C = 1) ≥ P (SyT |C = 0)P (C = 0). (3)
P11=0.0338 P12=0.0245 P22=0.9171 6
S1 S2
For the detection experiment, we collected an equal number P11=0.0338 P12=0.0245 P22=0.9171
of RF and noise signals. Therefore, it is rational to assume P21=0.0245
the prior probabilities of the RF signal and noise classes are S1 S2
equal, then the decision rule in (3) reduces to Signal Class
P21=0.0245
P (SyT |C = 1) ≥ P (SyT |C = 0). (4)
(a)
Therefore, for a given test data, we need to compute and Signal Class
P11=0.544 P12=0.1870 P22=0.0816
compare the likelihood probabilities P (SyT |C = {0, 1}). First,
in order to compute the likelihood probability for the RF
S1 S2
signal and noise classes, we use large amount of training data P12=0.1870
captured from multiple UAV controllers, Wi-Fi routers, mo- P11=0.544 P22=0.0816
bile Bluetooth emitters, and background noise. This training P21=0.1870
S1 S2
data set is stored in a database as shown in Fig. 3. Since (b)
Noise Class
the captured RF data (after sampling) is a discrete time- Fig. 6. Two-state Markov model and associated state transition probabilities
P21=0.1870
varying waveform, we can model it as a stochastic sequence using δ = 3.5σ for (a) the RF signal class, and (b) the noise class.
of states/events. The likelihood probability of such a state Noise Class
sequence can be computed based on the transitions between
the states of the generated Markov models. is significantly higher than the other transition probabilities.
A two-state Markov model for a given signal yT [n] can Based on these observations, the differences between the
be generated by mapping each sample in the signal to one state transition probabilities of each class can be utilized to
of the two states (S1 and S2 ). The samples whose absolute determine the class of a captured test signal.
amplitudes are less than or equal to a predetermined threshold Consequently, the likelihood of the test signal being an RF
δ are considered as in state S1 , while the samples with signal can be calculated as follows:
absolute amplitude greater than δ are considered as in state N −1
S2 . Mathematically, the state transformation is performed as Y
P (SyT |C = 1) = p(SyT (n) → SyT (n + 1)|C = 1)
follows: ( n=1
S1 , |yT [n]| ≤ δ Y T (i,j)
SyT (n) = . (5) = N
T PC=1 (i, j) (7)
S2 , |yT [n]| > δ
i,j={1,2}
Based on the above rule, it is straightforward to transform =
Y N
ij
pij;C=1 .
yT [n] into the state vector, SyT . Once SyT is obtained, the
i,j={1,2}
probability of a transition between any two states is calculated.
Note that the state vector is generated based on the amplitude The product of the conditional transition probabilities in the
of the signal samples in the wavelet domain. The choice of δ above equation gives the likelihood of obtaining the state
in (5) depends on the operating SNR of the system and will vector SyT given the hypothesis C = 1 is true. The log-
be discussed in Section VII-A. The transition count matrix, likelihood of the above expression is:
T N , and the transition probability matrix, T P , are defined as X
follows: log (P (SyT |C = 1)) = Nij log(pij;C=1 ). (8)
i,j={1,2}
N11 N12 p11 p12 TN
TN = ,TP = =P , (6) Similarly, the log-likelihood of the signal coming from a noise
N21 N22 p21 p22 i,j Nij
class is calculated by
respectively, where Nij is the number of transitions from state X
Si to Sj among all samples of yT [n], and pij = P (Si → Sj ) log (P (SyT |C = 0)) = Nij log(pij;C=0 ). (9)
is the probability of a transition from state Si to Sj . The i,j={1,2}
matrix T P is obtained by normalizing the T N matrix with
The decision will be favored to C = 1, if
the total number of samples in the signal. It is expected that
log(P (SyT |C = 1)) ≥ log(P (SyT |C = 0)); otherwise,
the transition probabilities generated for the signal class (UAV,
C = 0. We discuss the detection results in Section VII-A. If
Wi-Fi, and Bluetooth) and the noise class will be significantly
the captured test signal belongs to the RF signal class, then the
different at modest SNR levels. Also, the choice of δ in (5)
second stage detector is invoked to identify UAV controller-
dictates the transition probabilities for both the signal and
type signals. Otherwise, the system continues sensing the
noise class. In Section VII-A, the threshold δ is expressed
environment for the presence of signals as shown in Fig. 3.
in terms of the estimated standard deviation (σ) of the prepro-
cessed noise data captured from the environment. Moreover,
during the experiments, data is captured within a short time IV. D ETECTION OF W I -F I A ND B LUETOOTH
window (0.25 ms), thus we assume the environmental noise I NTERFERENCE
is stationary during this interval. In recent times, there has been interest in detecting Wi-
Fig. 6 shows the two-state Markov models for the RF Fi and Bluetooth signals [17]. In [18], a new technique is
signal and noise classes obtained from the training data using proposed for classifying Wi-Fi and Bluetooth interference
δ = 3.5σ. From Fig. 6(a), we see that for the signal class, p22 signals in the 2.4 GHz band. The technique uses the Hidden
is significantly higher than p11 , p12 , and p21 . On the other Markov Model (HMM) to model sequences or periodicity
hand, from Fig. 6(b), we see that for the noise class, p11 in the captured signal. The expectation-maximization (EM)
7
106 106 106

12 2.5 2.5
10 2 2
8
1.5 1.5
FFT
FFT
FFT
6
1 1
4
0.5 0.5
2
0 0 0
-100 -80 -60 -40 -20 0 20 40 60 80 100 -3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Frequency (MHz) Frequency (MHz) Frequency (MHz)
(a) (b) (c)
Fig. 7. Bandwidth analysis of (a) Wi-Fi signal, (b) Bluetooth signal from Motorola e5 cruise, and (c) Spektrum DX5e UAV controller signal.
algorithm is used to learn the parameters of the HMM models. TABLE I

The proposed system achieved accuracy above 88%. The S PECIFICATIONS OF THE W I -F I AND B LUETOOTH S TANDARDS .
major drawback is the fact that the EM algorithm often Bluetooth (IEEE Wi-Fi (IEEE
Standard
converges to a local maximum. Therefore, the EM algorithm 802.15.1 WPAN) 802.11 WLAN)
might fail in computing a consistent estimate of the parameters Center frequency (GHz) 2.4 2.4/ 5
Bandwidth (MHz) 1 20/ 40/ 80/ 160
of the model. Besides, in the context of RF-based UAV de- PHY modulation [20], [21] GFSK/FSK/DPSK DSSS/ OFDM
tection in urban environments, where the Wi-Fi and Bluetooth Range (m) variable >50
signals are considered as interference, there has been minimal Data rate (Mbps) 2 variable
research efforts. Fortunately, these interference signals are well
standardized and can be identified by using the knowledge
of their specifications. Table I provides a brief summary of We consider a zero-crossing GFSK/FSK demodulator. It is
the specifications for Wi-Fi and Bluetooth transmissions. It is known that the Bluetooth GFSK/FSK signal is transmitted in
obvious that the signal bandwidth and the modulation type are burst consisting of M data bits dm ∈ {−1, +1}, each bit
two important features for identifying the Wi-Fi and Bluetooth having a period Tb and average energy per bit Eb [19]. A
signals. The second stage detector exploits these features for general model for such a signal is given as:
r
detecting these interference sources. 2Eb
s(t) = . cos(2πfo t + ϕ(t, α) + ϕo ) + n(t), (10)
The first step in deciding if the detected signal is a wireless Tb
interference or not is to perform bandwidth analysis. This where ϕ(t, α) is a phase modulating function, ϕo is an
is because Wi-Fi signals can be easily identified by their arbitrary phase constant, fo is the operational frequency,
bandwidth. According to Table I, Bluetooth 2.0 signals have and n(t) is the channel noise component. The zero-crossing
a bandwidth of 1 or 2 MHz, Wi-Fi signals have a bandwidth demodulator considered herein for Bluetooth interference
of 20 MHz (or more) while all the UAV controller signals in detection is able to detect the time instants at which the
our database have bandwidth less than 10 MHz. Therefore, if signal s(t) is equal to zero and has a positive slope, i.e.,
the detected RF signal has a bandwidth equal or greater than the zero-crossings. When a Bluetooth device transmits at
20 MHz, it is classified as a Wi-Fi signal. Bandwidth analysis the basic rate using the standard GFSK/FSK modulation,
is performed by taking the Fourier transform of the resampled one symbol represents one bit. Therefore, the time interval
signal. Fig. 7 shows the result of the bandwidth analysis of a between consecutive zero-crossings is a measure of the
typical Wi-Fi, a Bluetooth (from Motorola e5 cruise), and a symbol duration of the Bluetooth signal.
UAV (Spektrum DX5e) controller signal. Fig. 8 shows the results of the zero-crossing demodulation
If the detected signal has a bandwidth less than 20 MHz, it is of a Bluetooth signal from Motorola e5 cruise. The captured
assumed to be transmitted either from a Bluetooth interference Bluetooth signal and its fast Fourier transform (FFT) are
source or a valid UAV controller. Since most mobile Bluetooth shown in Fig. 8(a) and Fig. 8(b), respectively. From Fig. 8(b),
devices employ Gaussian frequency shift keying GFSK/FSK we see that the transmit frequency of the Bluetooth device
modulation, it is reasonable to detect and discriminate these is 2.4 GHz. Afterward, the signal is shifted and resampled by
devices by means of modulation features. In this study, two 1/2000. The FFT of the resampled signal is shown in Fig. 8(c).
GFSK/FSK modulation features, namely, frequency deviation This figure shows that the bandwidth of the Bluetooth signal
and symbol duration, will be used to discriminate Bluetooth is around 2 MHz, which is far less than 20 MHz. Next, the
signals. Frequency deviation is a measure of the maximum resampled signal is demodulated by taking the derivative of
difference between the peak frequency in the GFSK/FSK its phase angle, and the start point of the demodulated signal
signal and the center frequency. On the other hand, sym- is estimated using the Higuchi algorithm [22]. The Higuchi
bol duration is the minimum time interval in the observed algorithm detects the start point of the signal by measuring
Bluetooth waveform or pulse. Therefore, using a GFSK/FSK the fractal dimensions of the signal. Once the start point is
demodulator, these features can be extracted and used as the detected, the frequency deviation is estimated as one half the
basis for Bluetooth signal detection. peak-to-peak frequency of the demodulated signal. Fig. 8(d)
8
104 109 106

4 2.5
4
2
2
Digitized amplitude
3
1.5
FFT
FFT
0 2.4 GHz
2
1
-2 1 0.5
-4 0 0
0 0.02 0.04 0.06 0.08 0.1 -5 -4 -3 -2 -1 0 1 2 3 4 5 -3 -2 -1 0 1 2 3
Time (ms) Frequency (GHz) Frequency (MHz)
(a) (b) (c)
106 30
1
2 25
Frequency of occurrence
Start-point 0.8
Frequency (Hz)
1 20 Symbol
Sample value
551.12 kHz 0.6 duration=0.5 s

0 15
0.4
-1 10
0.2 5
-2
-3 0 0
0.02 0.04 0.06 0.08 0.1 0 0.01 0.02 0.03 0.04 0.5 1.0 1.5 2.0 2.5 3.0
Time (ms) Time (ms) Zero-crossing interval ( sec)
(d) (e) (f)
Fig. 8. Extraction of the modulation features of a Bluetooth interference signal from Motorola e5 cruise mobile device using zero-crossing demodulation
technique: (a) Raw signal, (b) FFT of the raw signal, (c) FFT of the shifted and resampled signal (by 1/2000), (d) the demodulated signal showing a
peak-to-peak frequency of 551.12 kHz, (e) binary signal, and (f) histogram of the time-interval between consecutive zero-crossings in the modulated signal.
shows a plot of the demodulated signal and the estimated start

point which is obtained using the Higuchi algorithm. From
1200
the figure, the peak to peak frequency of the demodulated
signal is estimated as 551.12 kHz, and therefore, the frequency
1000
deviation is 275.56 kHz.
Frequency Deviation (kHz)
In order to estimate the symbol duration, the demodulated 800

signal is converted to a binary signal by using the mean
as a threshold. Fig. 8(e) shows the binary signal, where 600
binary one represents a positive frequency deviation, and a Bluetooth
binary zero represents a negative frequency deviation. Then, cluster
400
we compute the derivative of the binary signals to locate the
zero-crossings. To ensure we accurately compute the symbol 200
duration, we compute the histogram of the time intervals Bluetooth
UAV controllers
between consecutive zero-crossings. This is necessary because
0
channel distortions will cause some deviation in these in- 0 0.5 1 1.5 2 2.5
tervals. Fig. 8(f) shows the histogram of the time intervals Symbol Duration ( s)
between consecutive zero-crossings for the Bluetooth signal.
The estimated symbol duration is 0.5 µs. To validate the joint Fig. 9. Feature space showing the symbol duration and frequency deviation
discriminating ability of these modulation features, Bluetooth of the signals from several mobile Bluetooth devices and UAV controllers.
signals from six mobile phones and signals from nine UAV Each UAV controller is represented by a circular marker of a unique color.
controllers are collected. The mobile phones are Iphone 7,
Iphone XR, LG X charge, Motorola G Play, Motorola e5 of the Bluetooth signals from different mobile phones. All
cruise, and Samsung Galaxy Note 9. The UAV controllers the Bluetooth signals have a symbol duration of 0.5µs and
considered are Jeti Duplex DC-16, Spektrum DX5e, Spektrum a frequency deviation of less than 350 kHz. Therefore, the
DX6e, Spektrum DX6i, Spektrum JR X9303, FlySky FS-T6, frequency deviation and symbol duration can be used as fea-
Graupner MC-32, HK-T6A, and Turnigy 9X. The UAV signals tures in a simple maximum likelihood classifier for identifying
are frequency modulated as well. Therefore, all the collected Bluetooth interference signals. If the detected signal is not
signals are demodulated using the zero-crossing technique. from a Bluetooth interference source, it is presumed to be an
Fig. 9 shows the feature space of the demodulated Bluetooth emission from a UAV controller and transferred to the UAV
and UAV controller signals. The figure shows a clear clustering classification system.
9
TABLE II
140 0
S TATISTICAL F EATURES .
120 Features Formula Measures
Power/frequency (dB/Hz)
Mean (µ) 1 PN Central tendency
N Pi=1 i
x
100
-50 Absolute mean (x̄) 1 N Central tendency
N i=1 |xi |
80
h i1
Standard 1 PN 2 2 Dispersion
i=1 (xi − x̄)
2.4 GHz
deviation(σT ) N −1
60 PN 3 Asymmetry/shape
Skewness (γ) i=1 (xi −x̄)
-100
(N −1)σT 3 descriptor
40 PN
Entropy (H) − i=1 xi log2 xi Uncertainty
h P i1 Magnitude/Average
20 Root mean square 1 N 2 2
(xrms ) N i=1 xi power
-150 h P 1 2
i
Root (xr ) 1 N Magnitude
i=1 |xi |
0 1 2 3 4 5 2
N
Frequency (GHz) PN 4 Tail/shape
Kurtosis (k) i=1 (xi −x̄)
(N −1)σT 4 descriptor
(a) PN
1
Variance N i=1 (xi − µ)2 Dispersion
Peak value (xpv ) max(xi ) Amplitude
Waveform ampli-
4 Peak to peak (xppv ) max(xi ) − min(xi )
tude
xrms
Normalized energy trajectory
Shape factor(xsf ) Shape descriptor

x x̄
max
Crest factor xrms
Peak extremity
3 xmax
Impulse factor Impulse
Energy x x̄
max
transient Clearance factor xr
Spikiness
2
the maximum energy values along the time-axis. From this

1
distribution, we estimate the energy transient by searching
for the most abrupt change in the mean or variance of the
0 normalized energy trajectory. The term transient is defined as
0 10 20 30 40 50
a sudden change in the waveform of the signal which could
Time ( s)
be due to modulation in amplitude, frequency or phase. A
(b) transient contains unique information of the signal and can
Fig. 10. (a) The spectrogram and, (b) the energy trajectory of the UAV be exploited in classification tasks. Accurate detection of the
controller signal shown in Fig. 5. start point of time-domain transients is critical and highly
dependent on the environmental noise level. If the SNR is
considerably low, transient start point may not be detected
V. UAV C LASSIFICATION U SING RF F INGERPRINTS properly and this may result in extracting features that do not
The input to the ML classifiers are the RF-based features represent the signal. Due to this problem, we propose to use
extracted from the energy-time-frequency domain representa- energy transient by using an analogy between the transients in
tion of the UAV controller signals. For this representation, we the time-domain and energy-time-frequency domain. For the
use the spectrogram method. The spectrogram of any signal RF signal in Fig. 5, the normalized energy trajectory and the
is computed using the squared magnitude of the discrete time corresponding energy transient are shown in Fig. 10(b).
short-time Fourier transform (STFT) Once the energy transient is detected, RF fingerprints (a

∞
X
2
set of 15 statistical features) are extracted. Each feature is
Spectrogram(m, ω) = yT [k]w[k − m]e −jωk a physical descriptor of the energy transients and can provide
, (11)

valuable information for ML-based classification of the signals
k=−∞
captured from different UAV controllers. Table II gives the
where yT [n] is the pre-processed signal captured by the list of the extracted features used in this study. The features
surveillance system, m is discrete time, ω is the frequency, extracted from 17 UAV controllers are used to train five
and w[n] is a sliding window function that acts as a filter. different ML algorithms: kNN, RandF, discriminant analysis
The spectrogram analysis of the captured RF signals can (DA), support vector machine (SVM), and neural networks
reveal the transmit frequency of the signal as well as the (NN). Since some of the features may be correlated, therefore
frequency hopping patterns. Fig. 10(a) shows the spectrogram redundant, we also perform feature selection to reduce the
of the signal captured from the remote controller of the DJI computational cost of the classification algorithm.
Phantom 4 Pro UAV (the signal in Fig. 5). In computing the
spectrogram, the signal is divided into segments of length 128
with an overlap of 120 samples between adjoining segments. A. Feature Selection Using NCA
Then, a Hamming window is used, followed by a 256-point The NCA algorithm is a nearest neighbor-based feature
DFT. The spectrogram shows that the transmit frequency of weighting algorithm, which learns a feature weighting vector
the signal is 2.4 GHz. by maximizing a leave-one-out classification accuracy using
The spectrogram displays the energy distribution of the a gradient based optimizer. It is a non-parametric, embedded,
signal along the time-frequency axis. Therefore, the energy and supervised learning method for feature selection. NCA
trajectory can be computed from the spectrogram by taking learns the weighting vector/matrix by which the primary data
10
0.14 TABLE III

UAV C ATALOGUE .
0.12
Normalized feature weight
0.1
Make Model Make Model
Inspire 1 Pro DX5e
0.08 Matrice 100 DX6e
DJI Matrice 6001 Spektrum DX6i
0.06 Phantom 4 Pro1 JR X9303
Phantom 3
0.04
Futaba T8FG Graupner MC-32
0.02
HobbyKing HK-T6A FlySky FS-T6
Turnigy 9X Jeti Duplex DC-16
0
or
an R r
da ak2 ot
vi k
e sis
Sh Ku ce
ss
n
ew or
S
py
F r
Fa k
Va tion
n
st cto
de ea
ls ea
a
ea
ce M
an Pe Ro
ct
ct
Sk act
ne
n
tro
Ab Me
ap rto
rd P
ria
pu P
Fa
C Fa
M
En
variance and standard deviation. On the other hand, entropy,

s.
re
which measures the uncertainty in the data set, is the least

ar
Im
le
C
St
significant feature. Based on these results, the ML algorithms

Feature index
can safely discard the less important features and still achieve
Fig. 11. NCA ranking of all the 15 RF fingerprints extracted from 17 UAV good (even better) classification performance. This is because
controllers. discarding the less significant features reduces the chance of
overfitting. In addition, for large-scale classification problems,
there can be huge computational saving in training and testing
are transformed into a lower-dimensional space [23]. In this the classifiers with fewer number of features.
lower-dimensional space, the features are ranked according to
a weight metric, with the more important features receiving VI. E XPERIMENTAL S ETUP AND DATA C APTURE
higher weight values.
Given a set of training samples representing the different During the experiments, RF signals are captured from 17
UAV controllers, U = {(x1 , Y1 ), . . . , (xi , Yi ), . . . , (xn , Yn )}, UAV controllers, six mobile Bluetooth devices (smart phones),
where xi is a p-dimensional feature vector extracted from the and a Wi-Fi router. Table III gives the catalogue of the UAV
energy transient, Yi ∈ {1, 2, . . . , C} are the corresponding controllers from eight different manufacturers. All the UAV
class labels, and C is the number of classes. Then the controllers transmit control signals in the 2.4 GHz frequency
NCA learns the feature weighting vector w by maximizing a band. In particular, a pair of UAV controllers from DJI Matrice
regularized objective function f (w) with respect to the weight 600 and DJI Phantom 4 Pro models are used while only one of
of each features. The regularized objective function is defined the other controller type is used. This is important for forensic
as: and security analysis to investigate the confusion that would
  arise when a target recognition system attempt to distinguish
n n p between UAV controllers of the same make and model. For
1 X X X
f (w) = pij Yij − λ wr2  , where the remaining part of the study we will refer to the pair of DJI
n i=1
j=1,i6=j r=1 Matrice 600 as DJI M600 Mpact and DJI M600 Ngat. Simi-
(
Pn k(dw (xi ,xj )) if i 6= j larly, the pair of Phantom 4 Pro controllers will be referred to
pij = j=1,i6=j k(dw (xi ,xj )) , (12) as DJI Phantom 4 Pro Mpact and DJI Phantom 4 Pro Ngat.
0, if i = j Fig. 12 shows the indoor and outdoor experimental scenar-
ios. In each case, the RF passive surveillance system detects
(
1, if Yi = Yj
Yij = , signals transmitted by the UAV controllers and the interference
0, otherwise sources. Due to space limitations, only the results of the indoor
n is the number of samples in the feature set, λ is the experiments will be reported. The experimental RF passive
regularization term, wr is a weight associated with the rth surveillance system consists of a 6 GHz bandwidth Keysight
feature, and pij is the probability with which each point MSOS604A oscilloscope with a maximum sampling frequency
xi selects another point xj as its reference neighbor and of 20 GSa/s, 2 dBi omnidirectional antenna (for short distance
inherits the class label of the latter [24].PThe parameter Yij detection), and 24 dBi Wi-Fi grid antenna (for longer distance
p 2 detection). The antennas operate in the 2.4 GHz frequency
is an indicator function, dw (xi , xj ) = r=1 wr |xir − xjr |
is a weighted distance function between xi and xj , and band. Furthermore, to ensure only signals in the 2.4 GHz band
k(a) = exp( σa ) is some kernel function. Thus, NCA is a are captured, the output of the receiving antenna is passed
kernel-based feature selection algorithm that selects the most through a 2.4 GHz bandpass filter. The addition of a 2.4 GHz
descriptive and informative features by optimizing (12) using bandpass filter also removes out of band interference signals.
gradient update techniques. Therefore, only Wi-Fi and Bluetooth interference signals are
considered in this study. Moreover, the detection range of
Fig. 11 shows the results of the NCA ranking of 15 features
the RF surveillance system can be further improved by using
extracted from the 17 UAV controllers. The experimental
a combination of high-gain receive antennas and low-noise
setup and structure of the captured data are described in
power amplifiers (LNAs).
Section VI. In Fig. 11, we see that NCA ranks the RF
The receiver antenna continuously senses the environment
fingerprints according to their weight values. It turns out that
for the presence of RF signals. During both the training and
the shape factor is the most discriminative feature in the
feature set. The next significant feature is the kurtosis which 1 A pair of these controllers is used in this study. For all other controllers,
describes the tailedness of the energy trajectory curve. Next are only one of each type is considered.
11
presented in Fig. 13. The selected thresholds are functions

of the standard deviation (σ) of the preprocessed noise data
and the FAR specification. The value of σ is estimated after
performing multiresolution analysis (wavelet preprocessing)
of a concatenation of several noise data captured from the
environment. On the other hand, FAR, also known as the
probability of false detection, is the percentage of false alarms
per the number of non-events.
Fig. 13 shows that at very low SNR, such as −10 dB, the
detection accuracy is generally very low irrespective of the
threshold. As a result, in case of low-level signals (where
signals completely buried in the noise), the probability of
(a) missed detection increases. Besides, for a given SNR, it is
observed that the set threshold also affects the performance of
the detection system. For instance, when the system operates at
an SNR of 2 dB, a threshold of δ = 0.1σ will achieve a detec-
tion accuracy of above 99%. However, the threshold δ = 0.1σ
yields to a FAR of 100%. Therefore, a very low threshold
value will result in a high percentage of misclassification of the
noise data as signals. Furthermore, for the given SNR of 2 dB,
an increase in the threshold value to δ = 1.1σ will reduce the
detection accuracy and FAR to 96.6% and 14.8%, respectively.
Further increase in the threshold to δ = 2.5σ will greatly
reduce the detection accuracy and FAR to 40.4% and 3.2%,
respectively. Therefore, the optimum threshold depends on the
(b) operating condition and the requirements on the FAR. Besides,
the input impedance of the oscilloscope places a fundamental
Fig. 12. (a) Indoor and (b) outdoor experimental scenarios for UAV signal limit on the sensitivity of the passive detection system used in
detection.
this study.
test phases, data capture is performed in real-time followed by In addition, Fig. 13 shows that better detection performance
breaking into windows of a specific duration. Afterward, the (with low FAR) can be achieved if the detector operates at
captured data are automatically saved in MATLAB extension higher SNRs (above 8 dB) and threshold δ ∈ [2.5σ, 4.1σ]. For
format (.mat) in a cloud database for post-processing. We note instance, when the receiver operates at an SNR of 10 dB with
that, for accurate detection of the RF signals, the window a threshold δ = 3.5σ, the detection accuracy becomes 99.8%
length should be small enough so that the transitions that and, FAR drops to 2.8%. Although a continuous increase in
characterize the RF signals are not dominated by those of the threshold will further reduce the FAR, it will not always
the noise signals. On the other hand, for the classification guarantee a better detection accuracy, especially when the
process, the window length should be kept large enough such receiver operates at SNRs of less than 18 dB. This is because
that the energy transient can be extracted properly. Based on the dissimilarity between the transition matrices of the RF
these considerations, the window length is set to 0.25 ms. signal and noise classes reduces as δ increases beyond some
For each controller, 100 RF signals, each of which contains optimum value. Therefore, there is a high chance of detection
5000k samples (spanning a period of 0.25 ms), are collected. error as δ increases indefinitely.
During the experiment, the data was partitioned with the ratio Once a signal has been detected, the bandwidth and the
p=0.2. We used 80% for training (training (60%) + cross- modulation-based features are estimated as described in Sec-
validation (20%)) and 20% for testing. We set aside the tion IV. This information is used to decide if the signal comes
test data and performed cross-validation to train the machine from a UAV controller or any of the known interference
learning models. The cross validation avoids over-fitting and sources (Wi-Fi and Bluetooth sources). Given that the detected
helps to remove the bias in the training phase. To be specific, signal comes from a UAV controller, it is sent to the ML-based
we used k-fold cross-validation (with k = 5), where data is classification system for accurate identification. Classification
divided into k subsets, where each time, one of the k subsets results are discussed next.
is used as the validation/test set and the remaining k-1 subsets
are used for training the model. The error is averaged over all B. UAV Classification Results
k trials to get the total effectiveness of the model. The final
test error was recorded by averaging the test error for different For the classification problem, 15 statistical features given
Monte-Carlo simulations. in Table II are extracted. Feature selection is performed using
the NCA algorithm as described in Section V-A.
VII. R ESULTS To validate the result efficiency of the NCA and the ML
classifiers, 10 Monte Carlo simulations are run on the test
A. Detection Results dataset. On one hand, all the 15 features are used for the UAV
Detection performance of the proposed system is assessed controller classification problem. On the other hand, only three
for different SNRs and threshold choices, and the results are most significant features are used according to the NCA weight
12
100
90
[0.1 ,0.5 ]
80
[3.1 ,4.1 ]
70
Detection Accuracy (%)
60 =0.1 (FAR = 100% )

[1.1 ,2.5 ]
=0.5 (FAR = 75.8% )
50 =1.1 (FAR = 14.8% )
=1.5 (FAR = 10.2% )
40
=2.1 (FAR = 5.6% )
=2.5 (FAR = 3.2% )
30
=3.1 (FAR =3.0% )
20 =3.5 (FAR = 2.8% )
Increasing
=4.1 (FAR = 2.6% )
10 =15 (FAR =1.01% )
[15 ,20 ]
=20 (FAR = 0.61% )
0
-10 -8 -6 -4 -2 0 2 4 6 8 10 12 14 16 18 20 22 24
SNR (dB)
Fig. 13. The signal detection accuracy of the Markov model-based naïve Bayesian detector versus SNR for different values of δ.
TABLE IV
P ERFORMANCE OF THE ML C LASSIFICATION A LGORITHMS AT 25 dB SNR. 100 S AMPLE
S IGNALS FROM E ACH UAV C ONTROLLER I S C APTURED WITH 80% U SED FOR T RAINING
AND 20% FOR T ESTING (PARTITION R ATIO = 0.2). T HE S ELECTED RF F INGERPRINTS
A RE : S HAPE FACTOR , K URTOSIS , AND VARIANCE .
# of controllers Classifier Accuracy (%)2 Computational Time (s)2
All Feat. Selected Feat. All Feat. Selected Feat.
kNN 97.30 98.13 24.85 24.57
DA 96.30 94.43 19.42 18.58
15 SVM 96.47 91.67 119.22 111.02
NN 96.73 96.13 38.73 38.14
RandF 98.53 97.73 21.37 20.89
kNN 95.62 95.53 26.16 25.13
DA 92.77 88.12 19.36 18.90
17 SVM 93.82 87.88 139.94 141.68
NN 92.88 93.03 46.04 43.33
RandF 96.32 95.18 24.71 24.84
ranking shown in Fig. 11. These are the shape factor, kurtosis to train this model. More details can be found in [25],
and variance. The classification experiments are run separately [26]. Some of the critical hyper-parameters for the machine
for the case of 15 and 17 UAV controllers. Here, the number learning models (for the 17 controller case) after the Bayesian
of controllers represents the number of classes considered. In hyperparameter optimization are listed below:
the case of 15 controllers, all the controllers are of a different • kNN: Number of neighbors = 10, Distance metric =
model. However, in the case of 17 controllers, a pair of DJI mahalanobis
Matrice 600 (labeled as DJI Matrice 600 Mpact and DJI • DA: Type = Linear, Delta = 0.15146, Gamma =
Matrice 600 Ngat) and a pair DJI Phantom 4 Pro controllers 0.00016419
(labeled as DJI Phantom 4 Pro Mpact and DJI Phantom 4 Pro • SVM: Coding: onevsone, Lambda = 3.9941e-08, Learner
Ngat) are considered in addition to 13 different models. = Logistic
We used the Bayesian optimization method to obtain • NN: Double layer: Number of hidden nodes (Layer 1) =
the best hyper-parameters for our machine learning models. 45, Number of hidden nodes (Layer 2) = 15, Learning
Bayesian optimization has become a successful tool for hyper- rate = 0.30103, Activation functions = radbas
parameter optimization of machine learning algorithms, such • RandF: Bagged Ensemble with 60 bagged decision trees
as support vector machines or deep neural networks. The
algorithm internally maintains a Gaussian process model of 2 Both the accuracy and total computation time are the average of the 10
the objective function and uses objective function evaluations Monte Carlo simulations.
13
Table IV provides the classification accuracy of all five ML 100

algorithms. With the exception of the kNN and NN classifiers,
Classification Accuracy (%)

the table shows that the classification accuracy is only slightly
95
higher when all the features are used as compared to when
only the three selected features are used. Table IV shows
that it takes each ML classifier between 18-142 s to classify 90
the set of test signals extracted from all 17 UAV controllers.
The time taken is lesser when we use only the selected
features obtained from the NCA algorithm. The savings in 85
time, though little, will scale greatly as the number of UAVs
(to be classified) increases. This time saving may be critical
in aerial surveillance systems, where the response time to 80
kNN DA SVM NN RandF
effectively neutralizing a threat is very small. The outdoor Classifier
experimental test shows that it takes less than 0.5 s to detect (a)
and identify the RF signals from a single UAV controller which
is 200 m away. This timing is significantly small because 100
most commercial and hobby grade UAVs travel at a very

slow speed. For instance, the maximum speed of DJI Phantom
3 UAV is about 16 m/s. This means the proposed passive 95
detection system can detect this UAV hundred of miles away

before it gets into harm’s way. Moreover, the complexity of the
90
proposed system lies only in the training phase. It took several
hours to train the system using thousands of signals captured
from all 17 controllers at different SNR. Once trained, the 85
system has a good performance to complexity ratio. Moreover,
training is performed only once. In the test phase, only the
relevant features are used for classification. Hence, the results 80
kNN DA SVM NN RandF
in Table IV validate the decision to perform feature selection Classifier
using the NCA algorithm. (b)
Table IV shows the RandF classifier yields the highest
classification accuracy when all the features are used. For the Fig. 14. Box plot analysis of the classification accuracy of the ML classifiers
case of 15 and 17 controllers, RandF achieves an accuracy using the three selected features (shape factor, kurtosis, and variance) with
(a) 15 controllers, and (b) 17 controllers.
of 98.53% and 96.32%, respectively. Therefore, when all the
features are used, RandF is the best performing classifier. It is
followed by the kNN classifier, which achieves an accuracy of
97.30% and 95.62% with 15 and 17 controllers, respectively. Carlo simulations. Comparing the box plots in Fig. 14(a) and
The DA classifier is the least optimal when all the features Fig. 14(b), we see that the box plot metrics for each classifier
are utilized. On the other hand, when only the three selected are lower in the case of 17 controllers as compared to the
features are used, the kNN classifier performs the best with case of 15 controllers. This will be further investigated with
an accuracy of 98.13% and 95.53% for 15 and 17 controllers, the help of the confusion matrix. In addition, the box plots
respectively. It is followed by the RandF classifier which an reveal the presence of outliers in the performance of the SVM
accuracy of 97.73% and 95.18% with 15 and 17 controllers, and NN classifiers. These outliers suggest that for a given
respectively. When only the three most significant features test signal, SVM and NN classifiers could produce accuracy
are used, the least optimal classifier is SVM. We also note values well below the average values reported in Table IV.
that the DA classifier has the shortest computational time This observation raises the concern about the reliability of
whereas the SVM classifier has the longest computational these classifiers for the UAV controller classification problem.
time. The DA classifier computes and decomposes the class The SNR of the detected signal is an important factor that
covariance matrices. Fortunately, MATLAB has optimized influences the accuracy of the classifiers. Fig. 15 shows the
subroutines/functions for matrix computation and decomposi- accuracy versus SNR for the kNN, RandF and DA classifiers.
tion. As a result, DA is very fast and efficiently implemented For signals with SNR in the interval between 15 and 25 dB,
in MATLAB. However, SVM is slow because it has several the kNN is slightly better than the RandF for the case of
key parameters that need to be optimized to achieve the best 15 controllers. In the same SNR region, the RandF performs
classification performance. SVM hyperparameter optimization best for the case of 17 controllers. In this SNR range, the
searches for different kernel functions (Sigmoid, linear, RBF, DA classifier has the worst performance. On the other hand,
etc). This process is time-consuming. for SNR between 4 and 15 dB, the performance of the DA
Table IV provides only the average classification accuracy classifier improves significantly, outperforming the kNN and
results. A more detailed summary can be obtained from a RandF classifiers when 15 controllers are considered. This
box plot analysis shown in Fig. 14. Each box plot gives a is an interesting observation since DA is known to have
summary of the performance of a classifier in terms of the the shortest computational time. However, for SNR between
minimum, first quartile, median (red horizontal line), third 0 to 4 dB, the RandF classifier has the best performance.
quartile, and the maximum accuracy values over 10 Monte In general, the accuracy of all the classifiers increases with
14
is detected, it is identified using RF fingerprints along with

100
the ML-based classification techniques. Reducing the number
of required features with the help of NCA, the study shows
90
that it is possible to achieve an accuracy of 98.13% in
classifying 15 different controllers using only three features

80
in a kNN classifier. It is also shown that the proposed system
70
can even classify the same make and model UAV controllers
without much compromising the overall accuracy. In addition,
60 the detection and classification performance of the proposed
system is tested for a range of SNR levels. In each task, the
kNN-15
50 kNN-17
system is shown to be safe for SNR levels of above 10 dB.
DA-15 Future studies will present the detection of UAVs directly
DA-17 from the UAV signals in outdoor scenarios and consider the
40
RandF-15
RandF-17 potential of sensor fusion for improved UAV detection.
30
25 20 15 10 5 0
SNR (dB) R EFERENCES
Fig. 15. Classification accuracy versus SNR for kNN, RandF and DA [1] M. Ezuma, F. Erden, C. K. Anjinappa, O. Ozdemir, and I. Guvenc,
classifiers using the three selected RF fingerprints (shape factor, kurtosis, and “Micro-UAV detection and classification from RF fingerprints using
variance) as features for training and testing the ML classifiers. machine learning techniques,” in Proc. IEEE Aerosp. Conf., Big Sky,
Montana, Mar. 2019.
[2] H. Shakhatreh, A. H. Sawalmeh, A. Al-Fuqaha, Z. Dou, E. Almaita,
I. Khalil, N. S. Othman, A. Khreishah, and M. Guizani, “Unmanned
SNR. Therefore, to ensure accurate identification of the UAV aerial vehicles (UAVs): A survey on civil applications and key research
controller, it is best to operate the receiver at SNR above challenges,” IEEE Access, vol. 7, pp. 48 572–48 634, Apr. 2019.
[3] I. Guvenc, O. Ozdemir, Y. Yapici, H. Mehrpouyan, and D. Matolak, “De-
15 dB, in which case, kNN and RandF are the optimal tection, localization, and tracking of unauthorized UAS and jammers,”
classifiers for the datasets. Fig. 15 also shows that for all SNR, in Proc. IEEE/AIAA Digit. Avionics Syst. Conf. (DASC), St. Petersburg,
the accuracy plot is slightly lower when 17 controllers are FL, Sept. 2017, pp. 1–10.
[4] A. Solodov, A. Williams, S. Al Hanaei, and B. Goddard, “Analyzing the
considered as compared to the case of 15 controllers. threat of unmanned aerial vehicles (UAV) to nuclear facilities,” Security
The confusion matrix gives an idea of what a classifier is J., vol. 31, no. 1, pp. 305–324, Feb. 2018.
getting right and the type of errors it makes. Fig. 16 shows the [5] A. A. Alajmi, A. Vulpe, and O. Fratu, “UAVs for Wi-Fi receiver mapping
and packet sniffing with antenna radiation pattern diversity,” Wirel. Pers.
confusion matrices of the classifiers: kNN, RandF and DA for Commun., vol. 92, no. 1, pp. 297–313, Jan. 2017.
the case of 15 and 17 remote controllers. On the vertical axis [6] B. Nassi, A. Shabtai, R. Masuoka, and Y. Elovici, “Sok-security and
of each confusion matrix is the output class or the prediction privacy in the age of drones: Threats, challenges, solution mechanisms,
and scientific gaps,” arXiv preprint arXiv:1903.05155, Mar. 2019.
of the classifier while the horizontal is the target class or true [7] A. Shoufan, H. M. Al-Angari, M. F. A. Sheikh, and E. Damiani, “Drone
label. From the confusion matrices in Fig. 16, we observe pilot identification by classifying radio-control signals,” IEEE Trans. Inf.
that in the case of 17 controllers, the degree of confusion Forensics Secur., vol. 13, no. 10, pp. 2439–2447, Oct. 2018.
[8] Z. Shi, M. Huang, C. Zhao, L. Huang, X. Du, and Y. Zhao, “Detection
around the DJI controllers is relatively higher as compared to of LSSUAV using hash fingerprint based SVDD,” in Proc. IEEE Int.
the case of 15 controllers. This is because in the former, we Conf. Commun. (ICC), Paris, France, May 2017, pp. 1–5.
intentionally included two pairs of identical DJI controllers [9] I. Bisio, C. Garibotto, F. Lavagetto, A. Sciarrone, and S. Zappatore,
(DJI Matrice 600 MPact, DJI Matrice 600 Ngat, DJI Phantom “Unauthorized amateur UAV detection based on WiFi statistical finger-
print analysis,” IEEE Commun. Mag., vol. 56, no. 4, pp. 106–111, Apr.
4 Pro Mpact, and DJI Phantom 4 Pro Ngat). Consequently, 2018.
there are some confusions among these four controllers leading [10] C. Zhao, C. Chen, Z. Cai, M. Shi, X. Du, and M. Guizani, “Classification
to a slight reduction in the classification accuracy in the case of small UAVs based on auxiliary classifier wasserstein GANs,” in Proc.
IEEE Global Telecomm. Conf. (GLOBECOM), Abu Dhabi, UAE, Dec.
of 17 controllers. However, the kNN and RandF classifiers 2018, pp. 206–212.
still achieves an average accuracy of 95.53% and 95.18%, re- [11] M. F. Al-Sa’d, A. Al-Ali, A. Mohamed, T. Khattab, and A. Erbad, “RF-
spectively. Therefore, these classifiers are robust in identifying based drone detection and identification using deep learning approaches:
An initiative towards a large open source drone database,” Future Gener.
UAV controllers of the same make and model. On the other Comput. Syst., vol. 100, pp. 86–97, May 2019.
hand, the DA classifier is characterized by several more con- [12] T. Boon-Poh, “RF techniques for detection, classification and
fusions among different controllers which reduces its average location of commercial drone controllers,” https://tekmarkgroup.
com/eshop/image/catalog/Application/AEROSPACE/Paper-5_
accuracy to 88.12% in the case of 17 remote controllers. Thus, Techniques-for-Detection-Location-of-Commercial-Drone-Controllers_
while the kNN and RandF seem to be the best classifiers, the 2017-Malaysia-AD-Symposium.pdf, Tech. Rep., July 2017.
DA classifier still performs well for the given dataset. [13] P. Nguyen, H. Truong, M. Ravindranathan, A. Nguyen, R. Han, and
T. Vu, “Matthan: Drone presence detection by identifying physical
signatures in the drone’s RF communication,” in Proc. ACM Int. Conf.
VIII. C ONCLUSION Mobile Sys., Appl., Services (ACM MobiSys), Niagara Falls, NY, June
2017, pp. 211–224.
In this paper, the problem of detecting and classifying RF [14] I. Guvenc, F. Koohifar, S. Singh, M. L. Sichitiu, and D. Matolak, “De-
signals from different UAV controllers is investigated. The tection, tracking, and interdiction for amateur drones,” IEEE Commun.
Mag., vol. 56, no. 4, pp. 75–81, Apr. 2018.
detection system is designed to operate in the presence of [15] S. G. Mallat, “A theory for multiresolution signal decomposition: the
wireless interference from Wi-Fi and Bluetooth sources. These wavelet representation,” IEEE Trans. Pattern Anal. Mach. Intell., no. 7,
interference signals are detected using a multistage detector, pp. 674–693, July 1989.
[16] A. Bultan and R. A. Haddad, “System identification with denoising,” in
which estimates the bandwidth and modulation features of the Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), vol. 1,
detected RF signals. Once the signal from a UAV controller Istanbul, Turkey, June 2000, pp. 576–579.
Output class Output class Output class
D D D
JI D JI JI D
Sp H Ph J Ph DJ Ph J
an I M Sp H Sp H
Je e o G t a D Je an I M
D Je an I M
D
ti kt bb r S S S ti ek ob G
S S ti ek ob G
S S
D F ru y au F p pe pe DJ om tric DJI JI
k D tru by ra
F
Sp p p D tomatric DJ JI D tru by ra
F
Sp p p D tomatric DJ JI
D u p ly m i n e F l m k up u e ek ek JI 4 e I M In F l m k up u e ek ek JI 4 e I M In
JI T s pn ut k kt kt I P 4 P e 6 M Ins
le u k JR g e ab tru ru ru h r 0 at pi D up D up
D I x rn y a D
JI
I
le Tu ysk JRing ne tab ktru tru tru Ph Pr 60 at spi
x rn y a D
JI
I
le Tu ysk JRing ne tab ktru tru tru Ph Pr 60 at spi
x rn y a
D JI D nsp D ig F X9HK r M a T m m Dm D ntoo M0 M rice re 1 D JI D nsp D ig F X9HK r M a T m m Dm D ntoo M0 M rice re 1 D JI D nsp D ig F X9HK r M a T m m Dm D ntoo M0 M rice re 1
JI M JI ir C y S- 3 -T C 8 D X X m p p 1 P C y S- 3 -T C 8 D X X m p p 1 P C y S- 3 -T C 8 D X X m p p 1 P
-1 9 T 0 6 -3 FG X6 6 5 a a 0 JI M JI ir a a JI M JI ir a a
Ph at M e 6 X 6 3 A 2 i e e 3 ct ct 0 ro Ph at M e -1 9 T 0 6 -3 FG X6 6 5
6 X 6 3 A 2 e e 3 ct ct 00 ro Ph at M e -1 9 T 0 6 -3 FG X6 6 5
6 X 6 3 A 2 e e 3 ct ct 00 ro
an ric atr 1 P an ric atr 1 P i an ric atr 1 P i
to e 6 ice ro to e 6 ice ro to e 6 ice ro
m 0 1 m 0 1 m 0 1
4 0 0 4 0 0 4 0 0
D Pr Mp 0 D Pr Mp 0 D Pr Mp 0
J o a J o a J o a
Sp I Ph M ct Sp I Ph M ct Sp I Ph M ct
e a pa e a pa e a pa
Sp ktr nto ct Sp ktr nto ct Sp ktr nto ct
e um m e um m e um m
Sp ktru DX 3 Sp ktru DX 3 Sp ktru DX 3
increases with increasing value of ρ.

ek m 5e ek m 5e ek m 5e
tr DX tr DX tr DX
G Fut um 6e
r
H au ab DX
G Fut um 6e
H rau ab DX
G Fut um 6e
H rau ab DX
ob p a 6
Sp by ne T8 i o a o a
Sp bby pne T8 6i Sp bby pne T8 6i
ek kin r M FG ek kin r M FG ek kin r M FG
tru g C
Target class
tru g C tru g C
Target class
Target class
m HK -32 m HK -32 m HK -32
Fl JR -T6 Fl JR -T6 Fl JR -T6
ys X9 A ys X9 A ys X9 A
(e) DA with 15 controllers

Je ky 3 ky 3 ky 3
(a) kNN with 15 controllers

Je Je
(c) RandF with 15 controllers

ti Tu FS 03 ti Tu FS 03 ti Tu FS 03
D r - D r - D r -
up ni T up ni T up ni T
le gy 6 le gy 6 le gy 6
x 9 x 9 x 9
D X D X D X
C
DA, acc. = 94.43%, timeElapsed:18.58s

C C
kNN, acc. = 98.13%, timeElapsed:24.57s
-1 -1 -1
6
RandF, acc. = 97.73%, timeElapsed:20.89s

6 6
1-
1-
1-
0
1
0
1
0
1
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
Output class Output class Output class

D DJI D DJI D DJI
JI P D D JI P D D JI P D D
Sp H Ph ha J JI Sp H Ph ha J JI Sp H Ph ha J JI
Je ek ob G a n I M Ma D Je Je
ti ti ek ob G
S
an nt I M Ma D ti ek ob G
S
an nt I M Ma D
D tru by ra SpSp Sp D nto tom at tricDJ JI
D t b r S D t b r S
D up Fly m kin up Fu ek ek ek JI m 4 rice e I M In D F l
ru y au F p peSpe DJ tomom atr tricDJI JI
k D F l
ru y au F p peSpe DJ tomom atr tricDJI JI
k
J I l e Tu s J g n ta tr tr tr Ph 4 P 6 as
up
T y m in p u ek k k I P 4 ice e M Ins
n t t
up
T y m in p u ek k k I P 4 ice e M Ins
n t t
D D JI le u sk JR g e ab tru ru tru h 4 P Pr 6 60 at pi D JI le u sk JR g e ab tru ru tru h 4 P Pr 6 60 at pi

JI D Ins x rn ky R H er ba umum um an Pr ro 60 00 tricpire
D ig F X9 K M T o 0 JI D Ins x rn y a JI D Ins x rn y a
C y S- 3 -T C 8 D DX DX tom N Mp N Mp e 1 1 P D ig F X9HK r M a T m m Dm D ntoro o M00 0 Mricere 1 D ig F X9HK r M a T m m Dm D ntoro o M00 0 Mricere 1
D D Ma JI M pire - - F X g
16 9X T6 03 6A 32 G 6 6e 5e 3 ga ac a ac 00 ro D D Ma JI M pire C y S- 3 -T C 8 D X X m N p N p 1 P
-1 9 T 0 6 -3 FG X6 6 5 g a g a D D Ma JI M pire C y S- 3 -T C 8 D X X m N p N p 1 P
-1 9 T 0 6 -3 FG X6 6 5 g a g a
JI JI tri a 1
P c i t t t t JI JI tri a 1
c 6 X 6 3 A 2 i e e 3 at ct at ct 00 ro JI JI tri a 1
c 6 X 6 3 A 2 i e e 3 at ct at ct 00 ro
D ha Ma e 6 tric Pr P P
JI n tri 0 e o D ha Ma e 6 tric Pr D ha Ma e 6 tric Pr
Ph tom ce 0 M 10 JI n tri 0 e o JI n tri 0 e o
Ph tom ce 0 M 10 Ph tom ce 0 M 10
an 4 60 p 0 an 4 60 p 0 an 4 60 p 0
to P 0 ac to P 0 ac to P 0 ac
m ro N t m ro N t m ro N t
D 4 M gat
J P D 4 M gat
J P
D 4 M gat
J P
Sp I P ro pac
h N t Sp I P ro pac
h N t
Sp I P ro pac
h N t
Spektr ant ga Spektr ant ga Spektr ant ga
e u o t ek um om t ek um om t
Sp ktrum D m 3 Sp tru D 3 Sp tru D 3
ek m X5 ek m X5 ek m X5
t t t
G F ru DX e G Fu rumDX6e G Fu rumDX6e
H ra uta m D 6e H ra ta D e H ra ta D e
ob up ba X
Sp b n T 6 o b o b
Sp bb upn a T X6 Sp bb upn a T X6
ek yk er 8F i ek yk er 8F i ek yk er 8F i
Target class
Target class
Target class
tru ing M G tru ing M G tru ing M G

m H C-3 m H C-3 m H C-3
Fl JR K-T 2 Fl JR K-T 2 Fl JR K-T 2
(f) DA with 17 controllers

ys X 6A
Je k 9 ys X 6A ys X 6A
(b) kNN with 17 controllers
Je k 9 Je k 9
ti T y F 30
D u S 3
up rn -T
(d) RandF with 17 controllers ti T y F 30
D u S 3
ti T y F 30
D u S 3
le igy 6 up rn -T up rn -T
le igy 6 le igy 6
x 9 x 9 x 9
D X
DA, acc. = 88.12%, timeElapsed:18.90s
C D X
C
D X
C
kNN, acc. = 95.53%, timeElapsed:25.13s
-1 -1 -1
6
RandF, acc. = 95.18%, timeElapsed:24.84s
6 6
1-
1-
1-
0
1
0
1
0
1
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
matrices, the colorbar is used to specify the degree of confusion in terms of the confusion probability ρ. Moving down the colorbar, the degree of confusion
Fig. 16. Confusion matrices of kNN, RandF and DA classifiers using the three selected RF fingerprints (shape factor, kurtosis, and variance). In the confusion
15
16
[17] S. Rayanchu, A. Patro, and S. Banerjee, “Airshark: detecting non-wifi Tech. Rep., Oct. 2017.
RF devices using commodity WiFi hardware,” in Proc. ACM Internet [22] R. Esteller, G. Vachtsevanos, J. Echauz, and B. Litt, “A comparison of
Meas. Conf. (ACM IMC), Berlin, Germany, Nov. 2011, pp. 137–154. waveform fractal dimension algorithms,” IEEE Trans. Circuits Syst. I.
[18] Z. Weng, P. Orlik, and K. J. Kim, “Classification of wireless interference Fundam. Theory Appl., vol. 48, no. 2, pp. 177–183, Feb. 2001.
on 2.4 ghz spectrum,” in Proc. IEEE Wireless Communications and [23] J. Goldberger, G. E. Hinton, S. T. Roweis, and R. R. Salakhutdinov,
Networking Conference (WCNC), Istanbul, Turkey. IEEE, 2014, pp. “Neighbourhood components analysis,” in Proc. NeurIPS, Vancouver,
786–791. Canada, Dec. 2004, pp. 513–520.
[19] T. Scholand and P. Jung, “Bluetooth receiver with zero-crossing zero- [24] W. Yang, K. Wang, and W. Zuo, “Neighborhood component feature
forcing demodulation,” IET Electron. Lett., vol. 39, no. 17, pp. 1275– selection for high-dimensional data.” JCP, vol. 7, no. 1, pp. 161–168,
1277, Aug. 2003. Jan. 2012.
[20] N. Instrument, “Introduction to Bluetooth device testing: From the- [25] J. Snoek, H. Larochelle, and R. P. Adams, “Practical bayesian optimiza-
ory to transmitter and receiver measurements,” http://download.ni.com/ tion of machine learning algorithms,” in Advances in neural information
evaluation/rf/intro_to_bluetooth_test.pdf, Tech. Rep., Sept. 2016. processing systems, 2012, pp. 2951–2959.
[21] Tektronix, “Wi-Fi: Overview of the 802.11 physical layer and [26] A. Klein, S. Falkner, S. Bartels, P. Hennig, and F. Hutter, “Fast bayesian
transmitter measurements,” https://www.tek.com/document/primer/ optimization of machine learning hyperparameters on large datasets,”
wi-fi-overview-80211-physical-layer-and-transmitter-measurements, arXiv preprint arXiv:1605.07079, 2016.

Detection and Classification of UAVs Using RF Fingerprintsin The Presence of Wi-Fi and Bluetooth Interference

Uploaded by

Copyright:

Available Formats

Detection and Classification of UAVs Using RF Fingerprintsin The Presence of Wi-Fi and Bluetooth Interference

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Detection and Classification of UAVs Using RF Fingerprintsin The Presence of Wi-Fi and Bluetooth Interference

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Detection and Classification of UAVs Using RF Fingerprints

I. I NTRODUCTION frequency (RF) signals from UAV controllers. In [7], drone

UAV High-resolution oscilloscope

Bluetooth interference source

0.2 0.2 0.2 0.2

-0.1 -0.1 -0.1 -0.1

-0.2 -0.2 -0.2 -0.2

(a) (b) (c) (d)

0.2 0.2 0.2 0.2

-0.1 -0.1 -0.1 -0.1

-0.2 -0.2 -0.2 -0.2

(e) (f) (g) (h)

0.2 0.2 0.2 0.2

-0.1 -0.1 -0.1 -0.1

-0.2 -0.2 -0.2 -0.2

(i) (j) (k) (l)

the bias in the signals, leading to a higher detection accuracy,

106 106 106

(a) (b) (c)

algorithm is used to learn the parameters of the HMM models. TABLE I

104 109 106

(a) (b) (c)

551.12 kHz 0.6 duration=0.5 s

(d) (e) (f)

shows a plot of the demodulated signal and the estimated start

In order to estimate the symbol duration, the demodulated 800

Shape factor(xsf ) Shape descriptor

the maximum energy values along the time-axis. From this

0.14 TABLE III

variance and standard deviation. On the other hand, entropy,

which measures the uncertainty in the data set, is the least

significant feature. Based on these results, the ML algorithms

presented in Fig. 13. The selected thresholds are functions

60 =0.1 (FAR = 100% )

Table IV provides the classification accuracy of all five ML 100

Classification Accuracy (%)

Classification Accuracy (%)

detection system can detect this UAV hundred of miles away

is detected, it is identified using RF fingerprints along with

classifying 15 different controllers using only three features

increases with increasing value of ρ.

(e) DA with 15 controllers

(a) kNN with 15 controllers

(c) RandF with 15 controllers

DA, acc. = 94.43%, timeElapsed:18.58s

RandF, acc. = 97.73%, timeElapsed:20.89s

Output class Output class Output class

D D JI le u sk JR g e ab tru ru tru h 4 P Pr 6 60 at pi D JI le u sk JR g e ab tru ru tru h 4 P Pr 6 60 at pi

tru ing M G tru ing M G tru ing M G

(f) DA with 17 controllers

You might also like