Detection, Tracking, and Interdiction for Amateur Drones

Detecția, monitorizarea și interzicerea dronelor utilizate de amatori

Enabled by recent technological advances, miniaturization, and open source hardware/software initiatives, use of
unmanned aerial vehicles (UAVs) for a wide range of applications is quickly transforming from a futuristic idea to
reality. Pervasive use of drones for future smart cities, however, also raises several technical and societal
concerns and challenges that must be addressed, as UAVs can be used to conduct physical attacks (e.g., via
explosives) and cyber attacks (e.g., hacking a critical infrastructure). Vital information can be obtained from
drones by using network exploits and malware- based attacks. Drones with high dependency on GPS can be
hijacked using GPS spoofing, to be used against people and infrastructure [1].
As the cost of commercial small UAVs is decreasing, the number of hobbyists using them is increasing rapidly.
Thus, it is becoming difficult to identify unauthorized drones and bring them down when necessary. Small drones
are a danger to civilian aircraft, and there have been reports of near accidents involving UAVs and aircraft around
airports. It is critical that small UAVs are restricted from entering protected areas, and an unauthorized UAV that
may be a potential threat should be detected, interdicted, and brought down in a safe area.
In this article, we provide a unified review of the techniques for detecting, tracking, and interdicting small
unauthorized UAVs around restricted areas. As shown in Fig. 1, we consider that UAV detectors are deployed
around the area of interest; the detectors can be ground units as well as other monitoring UAVs (see, e.g., the
Aerial Dragnet concept from the Defense Advanced Research Projects Agency, DARPA, in the United States). If
the trajectory of the UAV remains outside the protected area, it is considered to be safe. If the trajectory crosses
into the protected area, the drone needs to be detected and brought down outside the protected zone.

Threats from Unauthorized Drones

In this section we review some common cyber and physical threats introduced by unauthorized drones, giving
examples from popularly used small commercial UAVs.

Physical Threats from Unauthorized Drones

A major threat from small UAVs is their use in direct physical attacks on people, property, and infrastructure.
Small UAVs may, for example, carry explosives, radioactive materials, or guns that may be used by terrorist
groups, countries, and even individuals with malicious intent, to carry out targeted attacks on people and
infrastructure. For instance, in 2015 a quadcopter crashed on the lawn of the White House, which (despite being
unintentional) shows the potential dangers that can be introduced by unauthorized small UAVs. Organized crime
groups can use small UAVs to carry illegal information or material, such as smuggling drugs across national
borders. Targeted physical attacks to a country’s critical infrastructure, such as its power grid, can enable follow-
up cyber attacks.
Another major concern related to the use of drones is that of privacy. It is easy to mount a camera or recording
device to capture information, and this can violate personal privacy. To overcome such concerns, the Center of
Democracy and Technology (CDT) has asked the FAA to issue rules on privacy, and recommends using a data
collection statement to know whether the information collected will be retained, used, or disclosed. The most
suitable means to maintain privacy was termed Privacy by Design (PbD), which helps in maintaining standards
and provides remedies for security breaches. By adopting PbD principles, privacy intrusion becomes limited, and
privacy can be ensured at an early stage [2].

Cyber Threats to/from Unauthorized Drones

Due to their agile nature, drones can be used to approach the vicinity of a cyber-physical system to perform cyber
attacks, such as hijacking (e.g., other drones), spoofing, “sniffing,” and hacking. For example, many commercial,
lower-end drones use WiFi technology for communication between the drone and ground station controller, which
may be vulnerable to security breaches and attacks from other drones. An unencrypted Wi-Fi link used with
lower-end commercial drones such as the Parrot AR.Drone allows any individual to connect and potentially hack
the drone (see e.g., Skyjack software) [2]. Moreover, a de-authentication attack can be carried out by one drone to
other drones, e.g., using the aircrack-ng software which sends disassociate packets to connected clients (such as
those remotely controlling the drone) for disconnecting them from the drone’s WiFi controller connection. The
popular DJI Phantom 3 is also susceptible to such WiFi attacks, while (to the best knowledge of the authors) such
weaknesses have not yet been reported in the newer DJI Phantom 4 drone.
Another way drones are commonly used is to sniff (eavesdrop) wireless signals that may include private
information. For example, a small UAV such as a DJI Phantom, with a WiFi Pineapple mounted on it, can be used
to easily sniff wireless signals from peoples’ houses. Built-in tools such as airdoump-ng, sslstrip can be used to
dissect the wireless data. The basic idea is to land the drone on a building/balcony to collect data, and afterward
bring it back to the originating location. The connection with the WiFi Pineapple can be maintained over a 3G
connection by creating an SSH relay server.
The data captured by the WiFi Pineapple is time stamped and includes MAC address, received signal strength
indicator (RSSI), and manufacturer’s name of the WiFi enabled devices. Such data can be used to identify and
track individuals in public space (e.g., based on daily/weekly behavior patterns), which can introduce privacy and
security concerns. Based on the WiFi probe requests captured, Fig. 2 provides an example for the number of
different types of devices that we have detected around a UAV’s flight trajectory around Florida International
University. In just a few minutes, over one hundred unique mobile WiFi devices were identified. Such data
collected at city scale and longer time durations by unauthorized drones may have serious privacy implications.
Detection of Unauthorized Drones
In this section, since a good majority of drone detection techniques rely on the use of RF signals, we first discuss
some key RF propagation characteristics that may impact drone detection/tracking accuracy. We then review key
techniques for drone detection: ambient RF signals, radar, acoustic signals, computer vision, and data fusion.

Radio Propagation Channel Considerations

As is well known, radio propagation characteristics are frequency-dependent. Thus, knowledge of the specific
frequency band(s) used by drones for signaling is a prerequisite for effective detection. The primary channel
characteristic of interest for detection and jamming is attenuation, because for a given value of transmit power
and receiver sensitivity, path loss determines the detection or effective jamming range. Since path loss also
depends on the environment, path loss models, and consequently detection range, are also environment-specific.
For example, path loss differs dramatically between open rural settings and dense (or “cluttered”) urban
environments [3]. Often, path loss models specify the presence or absence of a line of sight (LOS) between
transmitter and receiver, so one should account for the LOS or non-LOS (NLOS) condition in evaluating detection
or jamming. For drones, the presence/absence of LOS is a direct function of elevation angle and communication
link type (e.g., air-ground or air-air). Traditional air-ground channels typically have relied on LOS, and most
modern applications of drones also aim to place the aircraft well above ground, hence increasing the probability of
LOS. In such cases, diffraction can be more prevalent than shadowing, although for very low altitudes, terrestrial
obstacle shadowing models may be used. Additional work is required for specific drone shadowing models in
various settings.
Multiple path loss models exist for the unlicensed bands in different environments. Most of these are for cellular
and indoor applications, but to first order, for low-altitude (and slow-moving) drones, the cellular models can be
used as initial approximations. But new air-ground path loss models for low-altitude drones, in multiple
environments, should be developed. The obvious reason for this is that drones will not only hover in high locations
away from local obstacles (mimicking cellular base stations), but will also move in three dimensions throughout an
environment. Hence, effects such as diffraction and waveguiding may be more important for some drone
operational conditions than in typical cellular channels. For LOS channels with no clutter very near the ground
receiver, quasi-deterministic “tworay” models have been shown to be fairly accurate [3], with measured standard
deviations from the log-distance linear fit less than 3 dB. These models have an LOS component and Earth
surface reflection, which can be estimated from geometry, along with a number of multipath components that are
modeled statistically.
Also of interest is channel dispersion attributable to multipath propagation. For example, in some cluttered
settings, a strong multipath reflection may be the only detectable signal from a low-power drone at low altitude.
This raises the topic of antenna directionality. The use of directional antennas can increase link range and
suppress, or enhance, multipath components. This potential advantage comes with the potential disadvantage of
a narrower field of view, requiring more time if the drone signal must be found by searching spatially. As higher
frequency bands are used, more sophisticated arrays for fast multiple- beam scanning will become feasible.
The mobility of drones also has effects on the propagation channel. Mobility causes spatial variation of signal
intensity as well as Doppler effects. Since Doppler shifts are directly proportional to both velocity and carrier
frequency, as these increase, Doppler shifts can be significant, particularly for narrowband signals. As an
example, for a carrier frequency at the low end of the millimeter-wave (mmWave) band (e.g., 30 GHz), a small
drone moving at its top speed of, say, 30 m/s, can cause a Doppler shift of 3 kHz. This may be an appreciable
fraction of a narrowband signal's bandwidth (and is approximately 25 percent of the current LTE subcarrier
bandwidth).

Radio Frequency Signals from UAVs

In [4], possible solutions and challenges of inexpensive RF-based drone detection were discussed, and
experimental low-cost feasibility demonstrations were provided. Detection methods were classified as:
1. Analyzing the reflected signal from the propellers
2. Eavesdropping on the communication between the drone and its controller
3. Analyzing the reflection patterns from vibration of the drone’s body.
It was shown via experiments that the first approach is feasible for distances below 3 m, and the second approach
works for distances below 50 m. These limits, of course, depend on established transmit power limits and the
propagation environment. In [5], it was shown that the hopping sequence in popular frequency hopping spread
spectrum physical layer signals can be extracted using software defined radios (SDRs). Since most commercial
radio controlled UAVs use this type of physical layer, getting access to the hopping sequence opens the way to
not only detect and track these UAVs more reliably, but also to jam and spoof them. Unique RF transmission
patterns from popular commercially available UAVs can be used to train a classifier (e.g., a neural network),
which can then be used to classify a specific UAV in real time.
In [6], three prevalent commercial RF physical layers for remote UAV control are characterized. An analog FM
transmitter’s radio footprint and communication protocol were compared with two more advanced methods. The
spread spectrum implementation was decoded and shown to have an RF footprint distinguishable from that of
FM. These two were further compared with a DJI Phantom 3 link, and it was suggested that bandwidth and
central frequency can be used to detect and classify the type of UAV. Although most commercially available small
drones employ unlicensed frequency bands and often WiFi signaling, with the prevalence and continued growth of
SDRs, knowledgeable drone users will eventually configure their own radio links, with customized physical layers,
and if needed, customized medium access control (MAC) layers. Depending on signaling complexity, this could
be possible in nearly any frequency band from very high frequency (VHF) bands up to (at present) approximately
6 GHz. Such novel radio signaling will present much more challenging cases for detection.

Radar
When there is no active RF radiation from a drone (e.g., fully/partially autonomous operation), or if such RF
signals cannot be detected reliably, use of RF based radar techniques can be an appealing alternative for
detecting/tracking drones. As opposed to, say, optical sensors, radars are active sensors, and they can operate
day or night. The mmWave radars, in particular, have very large bandwidths, and hence have a high resolution
that can facilitate accurate detection, classification, and tracking of unauthorized UAVs. A major challenge with
radar-based detection of microdrones is that they have very small radar cross section (RCS), and they fly at lower
altitudes and at lower speeds compared to larger drones [7]. Use of mmWave radars enables one to precisely
analyze the Doppler spectrum, which can provide information for classifying the type of the drone, understanding
whether they carry any payload (e.g., explosives), and distinguishing them from other flying objects/animals such
as birds.
Use of radar techniques for detection, classification, and tracking of micro-drones has been recently explored in
the literature. For example, in [7], the authors considered the use of multi-static radars operating at 2.4 GHz to
analyze micro-Doppler signatures of micro-drones (DJI Phantom Vision 2) with different payloads, and classify
drones into no payload, 200 g payload, and 500 g payload categories. A transmit power around 23 dBm was
used, along with horizontally polarized antennas with 24 dBi gain and 10°  10° beamwidth, at radar to target
distances less than 100 m. The proposed feature detection techniques consistently achieve more than 90 percent
classification accuracy.
In [8], frequency-modulated continuous-wave (FMCW) radar operating at 35 GHz was used to classify DJI
Phantom 2 (quadcopter) and DJI S1000 (octocopter) drones, at radar ranges of 30–90 m. Authors observed wide
Doppler sidelobes in range-Doppler profiles due to the high carrier frequency. On the other hand, information
about the drone’s range, radial velocity, size, type, and even shape can be extracted (also altitude if two antennas
are used). The authors in [9] presented an mmWave radar operating at 94 GHz at 20 dBm transmit power, a
bandwidth of 1 GHz, and a range resolution of 15 cm. The radar can cover distances from 10 m up to several
hundred meters. In [10], the authors considered the use of a 16-receive, 16-transmit multiple-input multiple-output
(MIMO) radar to generate 256 virtual elements to detect and track drones. The radar operates at 36 GHz central
frequency, and uses pulses of duration 500 ns. A hypothesis testing framework was developed to decide whether
a drone exists within a certain range/angular cell. Experimental results by the authors show that the target drone
can be tracked accurately at distances exceeding 100 m and at radial velocity of about 5 m/s.
For proof of concept purposes, we conducted preliminary experiments with Time Domain P410 UWB radar
equipment to evaluate range estimation accuracy to a small quadcopter. The UWB transceivers have an
instantaneous bandwidth of 2.2 GHz (between 3.1 GHz and 5.3 GHz), transmitted power from the radio was
limited to –14.5 dBm to comply with FCC regulations, and at the receiver a sampling resolution of 61 ps is utilized.
A large number of channel impulse response (CIR) realizations based on the reflected radar signals from the UAV
were captured at the UWB radar receiver considering a static channel environment, and consecutive CIRs were
subtracted from each other to filter out the multipath components from static objects. From these data, statistics
(e.g., variance and mean) of the Doppler shifts can be calculated, which are then used to find likelihood functions
and subsequently range estimates at each time instant using a maximum likelihood estimator; these estimates
are observed to be discontinuous across different timescales. Finally, a Kalman filter was used to track the range
of the target UAV. Results show that with time domain Kalman filtering, the noisy range estimates can be
smoothed to improve accuracy. Range estimates at multiple radar locations can be used to triangulate a target
UAV’s location. Directional antennas and beamforming techniques can be used to increase the radar detection
range.

Acoustic Sensors
Acoustic processing methods are a seemingly obvious method of detecting and tracking small UAVs. After all, we
are all too familiar with the distinctive buzzing noise made by a small UAV that is flying nearby. However, there is
more than meets the ear to these acoustic methods. Several research groups [11–13] have considered acoustic
sensors and have demonstrated the ability to detect small UAVs (quad-copters) in ranges from 20 m [11] up to
600 m [12] with a 99.5 percent probability of detection and a 3 percent false positive rate.
The two main differentiators in prior work are the signal processing method and the type of acoustic sensor
employed. In terms of sensors, microphone arrays [11–13] have some advantages over a single microphone: they
offer not only the chance to detect the presence of a drone, but through beamforming they can also find its
azimuth and elevation, thus allowing for angle of arrival localization by using two or more microphone arrays.
Additionally, the coherent detection enabled by beamforming also increases the signalto- noise ratio, which in turn
increases the range and rate of correct detection and tracking. The authors of [12] used a tetrahedral
configuration for the microphones with each microphone at the tips of a tetrahedral pyramid with microphones
approximately 27 cm apart from each other, showing a 99.5 percent probability of detection for ranges smaller
than 600 m. In contrast, in [13], a commercial product (Distran Omni360) used 120 elements in a spherical
arrangement to reliably detect drones between 150 m and 290 m depending on the drone type.
Signal processing methods range from searching for a known signature [11–13] using the well-known signal
processing technique of spectrograms, to deep neural networks trained specifically to distinguish small UAVs
from background noise. In terms of advantages, acoustic detection and tracking can be completely passive, and
relatively inexpensive (although system performance does benefit up to a point from an increase in the number
and quality of microphones as well as the overall computational power). Range of reliable detection depends not
only on the type of acoustic sensors, but also on environmental factors, for example, the presence and intensity of
wind as well as background noise. Additionally, accurate detection and tracking relies on a comprehensive
database of acoustic signatures, which in turn is vulnerable to new methods of propulsion with signatures that
may differ substantially from the ones in the database.

Computer Vision
Another approach for detecting and tracking unauthorized drones is computer vision. In the last few years, a wide
variety of long-standing problems in machine learning, artificial intelligence, and computer vision have seen
significant improvements. Average detection precision at the ImageNet Large Scale Visual Recognition
Competition (ILSVRC) had reached 0.80 in 2016 from 0.22 in 2013. Detection and tracking of drones based on
computer vision has important advantages, including:
1. Being able to detect drones that do not have RF transmission
2. Use of passive and cheap optical sensors
3. Excellent inherent directional accuracy
For faster and more accurate detection, visual and thermal cameras with different fields of views (FOVs) can be
used simultaneously: wider FOV can be used for coverage, while narrower FOV can be used for better
detection/classification accuracy. Infrared and laser-based techniques are alternative optical approaches for
detection of drones. For example, for detecting drones at night, use of shortwave infrared cameras was recently
demonstrated in [14].
A major disadvantage of optical sensors is that they operate poorly when visibility is limited (e.g., in the presence
of clouds, fog, and dust). They also cannot operate in NLOS conditions such as in urban environments. To
improve the detection accuracy in such scenarios, camera-based techniques are typically used in combination
with other earlier discussed detection methods, such as acoustic sensors [15] and radar [9].

Sensor Fusion
As should be apparent from earlier discussions, different technologies available for detection and tracking of
UAVs have various trade-offs related to cost, accuracy, precision, range, energy efficiency (critical if sensors
operate on batteries), portability (e.g., sensors deployed at other UAVs), and complexity (Table 1). For example,
while some sensor technologies can only operate very well in LOS conditions (e.g., computer vision), some others
can also operate in NLOS environments (e.g., RF-based approaches). For accurate and quick detection/tracking
of UAVs, data fusion techniques that can simultaneously use information from multiple types of sensors carry
critical importance (see, e.g., [9, 15] for joint use of 94 GHz mmWave radar and acoustic sensors, respectively,
with high resolution optical cameras), and this constitutes an open research area.

Interdiction of Unauthorized Drones

After detection of unauthorized UAVs (and potentially after localizing/tracking them), there can be multiple ways to
interdict unauthorized drones, depending on the vulnerability of the UAVs to different techniques. In this section,
we briefly review some of the common approaches, which are also summarized in Fig. 5. One way to interdict a
drone is to jam its remote control signal (commonly in the 2.4 GHz industrial, scientific, and medical, ISM, band)
and/or GPS signal using a jammer gun, as illustrated in Fig. 5a. In another approach, shown in Fig. 5b, an
operator fires a cannister with large nets, and as soon as the drone’s rotors become tangled, a parachute brings
the drone safely down to the ground. As discussed previously, certain types of drones may be vulnerable to de-
authentication attacks as in Fig. 5c, which can be used for breaking the communication link between the drone
and its operator.

As shown in Fig. 5d, Scotland's metropolitan police are training eagles to take down suspicious drones. This low-
cost idea does not require use of any other devices and has no danger to civilians. Eagles were made to consider
drones as prey so that they catch them and place them in a safe area. One other way in which a drone could be
brought down is by using another drone. For example, as shown in Fig. 5e, Tokyo police created a drone squad
for privacy breaches. The police department was provided net-carrying drones that will trap suspicious drones
flying in the vicinity. The GPS spoofing in Fig. 5f is another possible cyber attack that has been shown to work
effectively on drones [17], which can be implemented even with low-cost SDRs. The communication links in
drones include incoming signals from GPS satellites, signals indicating the drone’s presence, and a two-way link
between the ground station and the drone. The basic idea in GPS spoofing is to transmit fake GPS coordinates to
the control system of the drone. This will hijack the drone, and subsequently it will be completely under the control
of the attacker. A successful attack is conducted when the attacker is very close to the drone, or equivalently by
using a directional antenna with narrow beamwidth aiming at the drone. Due to lacking an authentication
mechanism, civilian drones can be attacked easily by delaying signals, whereas attacks on military drones are
complicated due to the use of authentication mechanisms.