CC BY
0
Review of RF-based drone classification: techniques, datasets, and challenges
Boban Sazdic-Jotica, Ivan Pokrajacb, Jovan Bajceticc, Nenad Stefanovicd
a Military Technical Institute, Belgrade, Republic of Serbia, e-mail: boban.sazdic.jotic@vs.rs, corresponding author, ORCID iD: https://orcid.org/0000-0001-7239-174X
b Military Technical Institute, Belgrade, Republic of Serbia, e-mail: ivan.pokrajac@vs.rs, ORCID iD: https://orcid.org/0000-0002-2573-5543 c Khaoticen, Belgrade, Republic of Serbia, e-mail: bajce05@gmail.com, ORCID iD: https://orcid.org/0000-0002-3070-2594 d Serbian Armed Forces, Center for Applied Mathematics and Electronics, Belgrade, Republic of Serbia, e-mail: nenad.stefanovic@vs.rs, ORCID iD: https://orcid.org/0000-0002-0222-8311
doi https://d0i.0rg/l 0.5937/vojtehg72-49286
FIELD: computer sciences, telecommunications ARTICLE TYPE: review paper
Abstract:
Introduction/purpose: This article analyzes the publicly available literature on drone classification in the radio frequency domain, focusing on detection and identification. Drones are increasingly used for illegal purposes, making classification techniques crucial. This review paper covers passive radio frequency sensors, classification techniques, and datasets that highlight the challenges.
Methods: Researchers are developing antidrone solutions because drones have become valuable tools and targets for illegal activities. Due to the scope of the subject matter, the review included only the classification of drones via passive radio frequency sensors with a description of the classification techniques (set of algorithms, methods, and procedures) and
NOTE: This work, with a similar title (Drone classification based on radio frequency: techniques, datasets, and challenges), has been presented at the 10th International Scientific Conference on Defensive Technologies OTEH 2022, Belgrade, Serbia, pp.314320, October 13-14.
C76!)
the datasets used for performance testing. This study introduces a new categorization and offers deeper insights into publicly available drone classification techniques.
Results: Based on the results of this study, it is apparent that deep learning algorithms are presently the most effective approach to addressing the challenge of drone classification within the radio frequency domain. One of the primary obstacles is the absence of a comprehensive standard for classifying drones in the radio frequency domain, which should be based on end-user requirements. Additionally, the results of two ablative experiments highlight the preprocessing of raw I/Q radio signals as an essential step in drone classification.
Conclusion: In summary, the proposed categorization provides a valuable tool for literature review. Deep learning is the most effective technique for drone classification, but publicly available datasets with drone radio signals are limited. The key strength of this study is that it represents the first review of publicly available datasets with drone radio signals.
Keywords: deep learning, drone, detection, classification, identification, radio frequency.
Introduction
Unmanned aerial systems (UAS), primarily commercial off-the-shelf (COTS) ones, have become increasingly popular due to their numerous applications, ranging from commercial to military purposes. However, as technology continues to improve and UAS become more affordable, they have also become more susceptible to criminal and terrorist activities. This has made it vital to have effective antidrone (ADRO) systems in place to protect sensitive areas and critical infrastructure. ADRO systems use a combination of sensors, including optoelectronic, acoustic, radar, and radio frequency devices, to monitor and detect UAS. These systems have three core subsystems: monitoring, mitigation, and command and control (C2). To ensure maximum protection against UAS threats, ADRO systems must incorporate different procedures such as detection, spoofing, jamming, and mitigation (Hassanalian & Abdelkefi, 2017; Ding et al, 2018). In summary, ADRO systems are essential for safeguarding against the potential harm that UAS can cause. By incorporating various sensors and procedures, these systems can effectively detect and neutralize UAS threats, making them a critical tool for security forces. The ongoing military operations in Ukraine and the Gaza Strip serve as evidence that the engagement of UAS worth less than tens of thousands of dollars can effectively neutralize crucial weapons and military assets of the enemy.
CO h-
</> <1J
d JD
"TO -C
0
T3 c ro
ui <1J </) ro ro
T3
01 <1J
o <1J
o
_03 o <1J c o
T3 T3 <1J </) ro
.Q LL
a:
M—
o
<1J
a:
ro
-l-J
<1J
m
O —>
T3 N <0
CM <u
_00 cm"
O >
■ST CM o cm
of
UJ
a:
ZD O o
_J
<
o
X
o
LU
I—
>-
a: <
i—
< -j
CD >o
X LU I—
O
o >
0
The ADRO system relies on various sensors combined with detection or warning procedures to detect UAS, also known as drones. These procedures serve two purposes - early warning (detection) and identification of the detected drones, with optional tracking, which provides inputs for the next stage of the ADRO system. Spoofing is one of the approaches used in the next phase but is not compulsory. The ADRO system can deceive drones by sending false radio signals, typically an emergency landing signal. If the spoofing fails, the ADRO system can engage in jamming procedures, where the drone control and navigation signals are disturbed by posing substantial interference. Finally, the ADRO system can use the mitigation procedure to destroy or capture malicious drones.
Although ADRO systems use various strategies and sensors, radar and radio frequency (RF) sensors are the most practical applications for primary drone detection. RF sensors have several advantages over radar technology. One of the most significant benefits is that they do not emit electromagnetic radiation, making them a safer and covert option. Additionally, RF sensors have a more extensive detection range and can be used to counteract UAS that use jamming techniques. RF sensors are passive devices that only receive RF signals from UAS, which are present in almost every scenario. In contrast, radar is an active device that emits electromagnetic energy, making it unsuitable for specific situations. The detection range of RF sensors depends on the environment and the transmitter power of the UAS but is usually comparable to radar. Another advantage of RF sensors is that they can combine with jammers, which can be used for spoofing and jamming if required. RF sensors are also more versatile than other sensors, as they can be utilized for various purposes such as communication protocol detection, drone MAC address detection, feature extraction, or direct use with multiple classification algorithms.
The characteristics mentioned above, especially those offered by RF sensors, have been verified through a quantitative comparative analysis based on data from (Drone Industry Insights UG, 2023) and (Butterworth-Hayes, 2023). The results of such analysis are shown in Figure 1 and include 295 different antidrone (ADRO) systems. Notably, most ADRO systems engage optoelectronic sensors (35.8%), RF sensors (31.24%), and radars (28.13%), while acoustic sensors (4.83%) are the least represented. The availability of cameras and the variety of computer vision (CV) algorithms are the main reasons that optoelectronic sensors (OES) take the first place. However, a more comprehensive and quality data analysis determines that RF sensors and radars are the primary choice for
modern ADRO systems due to their better characteristics. Indeed, this means that cutting-edge ADRO systems have an RF sensor and radar to detect the drone, while the OES is used mainly for closer identification of the drone type.
4,83%
28,13%
31,24%
□ RF
□ OES
□ Radar
□ Acoustic
35,80%
Figure 1 - Quantitative comparison of sensor types in ADRO systems
■4
co
!± CP
<u
d JD
.c
o
T3
<u 0) ro ro
T3
oí <u
o <u
RF-based ADRO systems can extract valuable information by intercepting signals between drones and ground controllers. This information can be used to detect and identify drones based on their communication protocols and MAC addresses. However, a significant drawback of such systems is that they require prior knowledge of the protocols and addresses which may not be available for custom-made drones. The ADRO system can also extract features from intercepted RF signals for detection and uses frequency or joint time-frequency signal representation (TFSR) of I/Q data to prepare inputs for the classification algorithm.
Unfortunately, ambient RF noise, multipath, and customized drones that operate autonomously without a communication link between the drone and the ground controller can make RF-based detection difficult. Real-time monitoring of RF signals can also be challenging due to the complexity of the RF domain. Moreover, it is worth pointing out that drone RF communication can be categorized into three main groups: command and control (uplink), telemetry and video (downlink), and guidance communications. The first two groups use various frequencies, while global navigation signals are used for guidance and navigation. In such a complex environment, RF sensors must be agile, have high-speed scanning performance, be susceptible, and have a high dynamic range within the whole frequency range.
Section 2 of the paper categorizes and describes the literature to provide a comprehensive overview of relevant studies. Section 3
ro
0
01 0) _03 o <u
T3
<u
ro
.Q
OH
<D
<u Q!
<u m
T3 N <0
CM <u
CM hO >
■ST CM o CM
of
UJ
a:
ZD O o
_J
<
o
X
o
LU
I—
>-
a: <
(D <
-J
CD >Q
X LU I—
o
o >
&
compares different classification techniques, and section 4 presents the results of the comparative analysis of the most relevant papers. Finally, section 5 provides the conclusion.
Review of literature and RF techniques categorization
Available studies introduced different classification techniques (approaches) based on the RF sensors. We created a new categorization of these RF techniques for classification according to the following:
- The method of input data processing:
- classic engineering techniques require prior feature extraction in combination with a decision threshold mechanism;
- advanced engineering techniques engage procedures for classification purposes without prior feature extraction (feature extraction is implemented in Al-based algorithms together with the learning process); and
- hybrid engineering techniques present a combination of the previous ones.
- The type of input data:
- techniques with classification algorithms that use the MAC address information as input data;
- techniques with classification algorithms that use the protocol information as input data;
- techniques with classification algorithms that use the features of RF signals as input data; and
- techniques with classification algorithms that use the entire received I/Q RF signal as input data.
It is important to note that both rules can categorize one technique, specifically, a classic engineering approach that uses protocol information as input. The categorization presented in this research paper is based on the most significant research papers available in the literature over the past five years.
A comprehensive literature dataset on counter-unmanned aircraft systems (C-UAS) is available for review in (Sazdic-Jotic, 2024a). This dataset offers information and insights into C-UAS technology, applications, and trends. The framework for this literature review was adopted from (Alzubaidi et al, 2021), starting from the identification, screening, and selection stages, which included almost two hundred (199) research papers in the last five years from publishers such as IEEE,
Nature, Springer, ACM, Elsevier, and MDPI. All reviewed papers have tackled the drone classification problem (detection and identification) using RF sensors. Each paper was categorized as a novel approach, survey/review, dataset description, or regulation/standard, with mandatory remarks explaining specific ADRO procedures (e.g., jamming, spoofing, direction finding).
RF techniques according to the method of input data processing
In classical engineering techniques, extracting features from intercepted RF signals is mandatory for data preparation and processing. This step is crucial as only features extracted with a decision threshold mechanism can be used for classification. However, the feature extraction process can be complex and time-consuming, requiring profound engineering skills and adaptation to the nature of the input data. Authors (Lv & Wang, 2019) used standard deviation analysis, maximum slope analysis, and accumulation in azimuth direction as statistical features for drone detection and direction finding. They also employed principal component analysis (PCA) and empirical mode decomposition (EMD) based wavelet transform (WT) methods to cope with additive Gaussian white noise. A similar approach was presented in (Ezuma et al, 2020) involving fifteen (15) different statistical features of RF signals. On the other hand, (Fu et al, 2018) presented the cyclostationarity signature of the drone RF signal and the pseudo-Doppler principle for the classification issue with a single-channel universal software radio peripheral (USRP) receiver. Authors (Nguyen et al, 2018) described a passive drone detection system (Matthan) based on two critical physical signatures of drones: body shifting and vibration. However, the Matthan approach faces a range of constraints, making it impractical for implementation. Another unique way is the received signal strength indicator (RSSI) ratio that engages various factors, such as propagation channels and fading effects, and can help accurately determine the location of a drone.
Artificial intelligence (AI) algorithms, particularly deep learning (DL), provide an alternative approach to the problem of drone classification. Unlike classical engineering techniques, advanced engineering techniques such as DL algorithms use the entire received I/Q RF signal, perform preprocessing steps, and send all data to the learning process without prior feature extraction. This approach is more robust and scalable, but a significant amount of input data is required for the training process, which can sometimes be a disadvantage. Fully connected deep neural networks (FC-DNN) and convolutional neural networks (CNN) are two
CO hoi" <1J
c JD
"TO -C
0
T3 c ro
01 <1J 0) ro ro
T3
oi <1J
o <1J
o
_03 o <1J c o
T3 T3 <1J 0) ro
.Q LL
a:
M—
O
<1J
a:
ro
<1J
m
o —>
T3 N
ro OT
prominent DL algorithms for drone classification in the publicly available literature. In addition, transformer models - a distinctive DL algorithm with an attention mechanism - are increasingly utilized to attain outstanding outcomes. Authors (Al-Sa'd et al, 2019; Sazdic-Jotic et al, 2022) employed FC-DNN for this purpose, while (Al-Emadi & Al-Senaid, 2020; Allahham et al, 2020; Ozturk et al, 2020; Basak et al, 2021; Mokhtari et al, 2021, 2022; Nguyen et al, 2021) used CNN. Authors (Basak et al, 2023) proposed a unique deep residual network-based autoencoder framework for known drone signal classification, novelty detection, and clustering (DE-FEND).
The quantitative comparison of the techniques that exploit RF sensors according to the method of processing input data is presented in Figure 2.
2%
□ classic techniques
□ advanced techniques
□ hybrid techniques
60%
38%
Figure 2 - Quantitative comparison of RF techniques according to the method of input
data processing
Notably, more than half (60%) of all research studies rely on advanced techniques, while 38% use classic engineering techniques. The reason behind the absence of classical engineering techniques in the analysis is that only studies from the last five years were considered. This time frame shows a significant increase in advanced techniques, now preferred over classical methods. Hence, it is not unsurprising that there is a lack of classical engineering techniques in the analysis. In the overall research, advanced engineering techniques such as DL and machine learning (ML) algorithms account for 47% and 13% of all research. Compared to the results presented in (Sazdic-Jotic et al, 2023), this is a growth of 9.3% for advanced techniques in favor of classic techniques.
It is also worth mentioning that the authors (Abeywickrama et al, 2018; Shi et al, 2018; Shorten et al, 2018; Basak & Scheers, 2019; Bisio et al, 2021; Bhattacherjee et al, 2022; Lofu et al, 2023) present classical and advanced engineering techniques in combination with direction-finding (DF) methods. The specific hardware and software implementation of the RF-based DF of UAS is presented in (Abeywickrama et al, 2018) because
the authors used a single-channel RF sensor and a four-element antenna array in combination with a sparse denoising autoencoder deep neural network (SDAE-DNN). However, it is essential to note that some authors (Zhang et al, 2018; Ezuma et al, 2020; Swinney & Woods, 2021; Medaiyese et al, 2022) use a hybrid engineering technique or a combination of classical and advanced techniques. In (Zhang et al, 2018), the authors used extracted features (the slope, kurtosis, and skewness) of the drone RF signal as an input for an FC-DNN. Moreover, in (Medaiyese et al, 2022), the authors performed feature extraction and used ML algorithms (Logistic Regression). An interesting approach was presented by (Ezuma et al, 2020), where the authors extracted fifteen statistical features from the UAS RF signal and engaged them with five different ML classifiers at various SNR levels.
While advanced techniques may offer a broader range of benefits than classical techniques, they also have certain drawbacks that must be considered. It is essential to thoroughly assess the advantages and disadvantages of each approach before deciding which one to use in a particular situation. Advanced engineering techniques often need help with several challenges that can hinder their successful implementation. One of the most significant obstacles is the requirement for unique training scenarios, which means that engineers must develop customized training models for each specific use case. Additionally, the availability of datasets can be limited, making it challenging to train models effectively. Finally, the transfer learning process can be complex, especially when applying previously learned knowledge to new scenarios. These factors can make it tricky to implement advanced engineering techniques successfully. Moreover, a fusion of different sensor data is also a particular solution to the drone classification problem. This approach involves fusing information from multiple sensors, such as radar, cameras, RF, and acoustic sensors, to create a more comprehensive view of the drone characteristics. This method can provide a more reliable drone classification by analyzing the drone shape, size, movement, and RF signals. It is a promising solution gaining traction in the drone detection and defense industry.
RF techniques according to the type of input data
RF sensors receive an RF signal from a UAS for various purposes. Four techniques exist for detecting and identifying drones based on the input data type. The first technique uses classification algorithms to detect and identify the MAC address of the transceiver device in a drone. In contrast, the second technique employs classification algorithms to detect and identify the communication protocol between drones and ground
CO hui <1J
c JD
"TO -C
o
T3 c ro
ui <1J </) ro ro
T3
ui <1J
o <1J
o
_03 o <1J c o
T3 T3 <1J </) ro
.Q LL
a:
M—
o
<1J
a:
ro
<1J
m
O —>
T3 N CO OT
<u
2
O >
■ST
2
O
2
Cti UJ Cd ZD O o
_J
<
o
X
o
LU
I—
>-
a: <
(D <
-J
CD >o
X LU I—
O
o >
&
control devices. These techniques are the least represented in the literature because they have significant limitations and shortcomings. Both approaches use received and demodulated RF signals to obtain information about the MAC address of the RF transceiver installed in the drone and the type of communication protocol unique to certain types of drones. The information obtained is used for the detection and identification of drones. (Schiller et al, 2023) analyzed the security and privacy of the DJI's tracking protocol (DronelD), presenting sixteen vulnerabilities that can be adopted. In addition, there are more hardware implementations of ADRO systems based on this technique. Authors (Haluza & Cechak, 2016) performed device and protocol identification through data format analysis. In (Sciancalepore et al, 2020), features such as packet inter-arrival time and size were analyzed, while (Stoica et al, 2020) studied eight protocols to classify UAS. The technique with classification algorithms based on protocol recognition is more efficient than the previous one, proven by practical implementations of such ADRO systems. (Oh et al, 2020) also proposed long-range (LoRa) modulation exploiting wireless communication protocol for drone identification.
Furthermore, techniques with classification algorithms that use features of RF signals as input data are more present in the literature. Essential studies that exploited features have been mentioned because this is a mandatory step for classical engineering techniques. Nevertheless, an increasing number of research papers in the literature deal with the entire intercepted I/Q RF signal. Faster hardware and improved computing power have allowed it to exploit the full power of DL algorithms created for a considerable amount of data. As a result, the techniques with algorithms that classify the entire received I/Q RF signal as input data are becoming widely present solutions providing excellent results. RF sensors are used to record the raw I/Q RF signal, and different preprocessing steps are taken to prepare input data for the classifier. Some authors (Al-Sa'd et al, 2019; Sazdic-Jotic et al, 2022) calculated the magnitude or phase spectrum to obtain 1-D (vector) data with corresponding labels, while others (Ozturk et al, 2020; Basak et al, 2021; Nguyen et al, 2021; Mandal & Satija, 2023) used more complex TFSRs, such as spectrograms or scalograms, to obtain 2-D (image) representations of intercepted I/Q RF signals with corresponding labels for classification purposes. Figure 3 illustrates one 2-D TfSr obtained from RF activities in the 2.4 GHz range.
Figure 3 - Spectrogram of two RF drone signals
This TFSR is a spectrogram of the RF signal when two drones operate simultaneously at 2.4 GHz. Two emissions are visually distinctive in Figure 3: the command and control (fixed frequency, FF) and telemetry and video (frequency hopping, FH) emissions. Subsequent image processing techniques based on morphological operations of the obtained spectrograms can extract all emissions. Moreover, it is possible to determine the technical parameters of the detected radio signals (the number of the detected radio signals, the total width of the frequency range, and the width of the frequency range of one channel). The technical parameters estimation of the detected radio signals must be divided into two branches, i.e., analysis of FH or FF emissions. This way, the channel raster, the FH hop durations, and the time between two successive FH hops can be estimated. In the case of the DJI Phantom IV Pro drone, one FF and FH emission was detected with a total band of 75 MHz in the spectral domain. This drone has the highest hop duration (6 ms) with a simple FH emission comparable to a sweep frequency signal. In contrast, the DJI Mavic 2 Zoom and Enterprise drones have three types of FH emissions, which depend on operation modes (a drone is connecting to the flight controller, a drone is hovering, a drone is flying, and a drone is flying with recording) (Sazdic-Jotic, 2024b).
CT>
oo CO cp
<u
c JD
"TO -C
o
T3
<u </) ro ro
T3
oi <u
o <u
ro
0
01 0) _03 o <u
T3
<u
ro
.Q
OH
<D >
<u OH
"ro <u 03
Different DL models are used depending on the preparation method of input data. In (Mokhtari et al, 2021; Sazdic-Jotic et al, 2022), the authors used an FC-DNN and CNN for single drone classification (detection and type identification) and multiple drone detection. Moreover, (Ozturk et al, 2020) examined CNN accuracy with SNR dependency, showing that classification is feasible. The quantitative comparison of the techniques that exploit RF sensors according to the type of input data is presented in Figure 4.
T3 N CO OT
50% 40% 30% 20% 10% 0%
42,66%
40,56%
14,69%
2,10% i-1
PHY
COMM
Features
I/Q
Figure 4 - Quantitative comparison of the RF techniques according to the type of input
data
It is important to note that the techniques that use features and raw I/Q RF signals as input data are the most exploited, with 42.66% and 40.56%, respectively. The techniques that use MAC addresses (PHY) and communication protocols (COMM) for drone classification are the least concurrent in the reviewed literature. Compared to the results presented in (Sazdic-Jotic et al, 2023), this is an increase of 10.8% for the I/Q RF signal as input data. This fact is unsurprising due to the expansion of the DL and ML research papers. More interesting is that more than 95% of all papers that use the whole received I/Q RF signal as input data use advanced engineering techniques.
Datasets and comparison of RF techniques
The main goal of this research was to review and categorize all available RF-based drone classification research papers and datasets. Extensive research was conducted to support and highlight the best RF-based drone detection technique by performing a comprehensive comparative analysis. The most relevant research papers were utilized, and the comparison was made based on the dataset used and engineering techniques. The publicly available studies verified on the DroneRF dataset, and the VTI_DroneSET were presented and classified according to our categorization and results from three experiments. This method was intentionally employed to compare different approaches to detecting or identifying the same number of classes.
It is worth noting that only a few publicly available datasets contain drone RF signals. Only two datasets have the records of RF signals from industrial, scientific, and medical (ISM) radio bands, but only one has multiple drones that operate simultaneously. Additionally, some authors
(Ezuma et al, 2021) and (Vuorenmaa et al, 2020) used ground controllers for classification, which can be valuable in various research. Furthermore, RF receivers generate vast amounts of data during the recording process, leading to massive datasets. It is essential to consider that such an amount of information can be a disadvantage in certain situations due to the requirement for excellent computers, storage, and GPUs - Table 1 lists publicly available datasets containing RF signals from UAS.
Table 1 - The RF drone publicly available datasets
Reference Number of UAS Multiple drones 2.4 GHz 5.8 GHz Size [GB]
(Allahham et al, 2019) 3 - + - 3.75
(Soltani et al, 2020) unknown - + - 4.5
(Vuorenmaa et al, 2020) 10 - + + 3,800
(Sazdic-Jotic et al, 2021) 3 + + + 3.35
(Ezuma et al, 2021) 17 - + - 124.13
(Medaiyese et al, 2023) unknown - + - 66.26
(Swinney & Woods, 2022) 7 - + - 65.76
■4
co
!± Ci
<u
c JD
-C
0
T3 c ro
01 <u </) ro ro
T3
oi <u
o <u
ro
0
01 tn
J5 o <u c O T3 T3
<u
O)
ro
.Q
OH
<u >
<U
OH
<u m
O
T3 N <0 œ
It is worth mentioning that there are additionally publicly available RF fingerprinting datasets (Al-Shawabka et al, 2020) and (Soltani et al, 2020) that can be used for automatic modulation classification in ADRO systems. Such datasets can be used to train the ADRO system to filter out non-interesting RF signals (Wi-Fi, Bluetooth) and concentrate on drone RF signal classification procedures.
Authors (Allahham et al, 2019) presented a DroneRF dataset incorporating three drones, recorded in four operating modes in only one ISM band (2.4 GHz). This dataset was used in over 60% of the reviewed literature. The studies whose results were verified on the DroneRF dataset are presented in Table 2.
The inscription "H" stands for hybrid and "A" for advanced engineering techniques. The inscription "F" stands for features and "R" for raw I/Q RF signal. It is important to note that no classic engineering techniques are employed on the DroneRF dataset. Moreover, for drone detection, there
are several approaches with excellent results that are lower for drone identification and flight mode identification.
Table 2 - Comparative analysis of publicly available studies verified on the DroneRF
dataset
Reference Technique Features Drone detection Type identification Flight mode identification
(Al-Sa'd et al, 2019) A R 99.7 84.5 46.8
(Allahham et al, 2020) A R 100.0 94.6 87.4
(Al-Emadi & Al-Senaid, 2020) A R 99.8 85.8 59.2
(Akter et al, 2020) A R - 92.5 -
(Swinney & Woods, 2020) A F 100.0 90.4 87.5
(Zhang, 2021) A R 100.0 98.7 79.2
(Kihg et al, 2022) A F 100.0 98.6 95.1
(Huynh-The et al, 2022) A R 99.9 98.6 95.3
(Donatus et al, 2023) A R 99.7 94.2 81.5
(Mohammed et al, 2023) A R 99.6 96.9 96.0
Similarly, (Sazdic-Jotic et al, 2021) introduced a dataset with three drones recorded in four operating modes in two ISM bands (2.4 and 5.8 GHz). Moreover, this dataset contains records of multiple (two and three) drones operating at the same time simultaneously.
Such data makes the VTI_DroneSET unique because there is no such dataset in the available literature.
Three studies whose results were verified on the VTI_DroneSET are presented in Table 3.
Table 3 - Comparative analysis of publicly available studies verified on the
VTI DroneSET
Reference Technique Features Drone detection Type identification Multiple drone detection
2.4GHz 5.8GHz 2.4GHz 5.8GHz 2.4GHz 5.8GHz
(Sazdic-Jotic et al, 2022) A R 98.6 99.8 96.1 95.7 96.2 97.3
(Mokhtari et al, 2021) A R - - 100.0 - - -
(Mokhtari et al, 2022) A R - - 99.9 - - -
(Sazdic-Jotic et al, 2023) A R 98.6 100.0 98.8 95.7 99.0 99.2
■4
CO h-!± Ci
<1J
c JD
-C
o
T3
<u </) ro ro
T3
oi <u
It is worth mentioning that the VTI_DroneSET provides multiple drone detection on real RF signals rather than simulated RF signals. It is essential to note that authors (Sazdic-Jotic et al, 2022) engaged FC-DNN while (Mokhtari et al, 2021), (Mokhtari et al, 2022) and (Sazdic-Jotic, 2024b) investigated CNN and CNN with recurrent layers (CRNN), respectively.
It is crucial to consider the input data and the corresponding preprocessing steps for adequate results of comparative analysis. Authors (Sazdic-Jotic et al, 2022) used power spectrum data obtained from raw I/Q radio signals for TFSR. Other authors engaged spectrograms as images for input data and achieved better results because they kept more valuable features for CNN/CRNN learning and testing. Moreover, through ablative analysis, it is possible to establish which TFSR is the best for processing raw I/Q radio signals when CNN is engaged as a neural network. The accuracy results of such an experiment are presented in (Sazdic-Jotic, 2024b) and shown in Table 4. The ablative experiment was performed with the AlexNet model introduced by (Krizhevsky et al, 2017) for the 2.4 GHz frequency band.
The results shown in Table 4 show that TFSR choice is essential for CNN accuracy. Moreover, the Short-Time Fourier Transform (STFT) method outperformed the Continuous Wavelet Transform (CWT) and the Wigner-Ville Decomposition (WVD) methods for all scenarios in the presented ablative experiment.
o <u
ro
0
01 </) J5 o <u c o
T3 T3
<u </) ro
.Q
a:
<D
<u a:
<u m
o
T3 N
ro OT
Table 4 - Comparative analysis of accuracy for different TFSRs in 2.4 GHz in ablative
experiments
the AlexNet / Drone Type Multiple
TFSR detection identification drone detection
STFT 97.3 96.6 99.1
CWT 95.3 96.2 97.9
WVD 86.6 86.0 96.6
An additional ablative experiment compares CNN accuracy with the radio signal segment length on which TFSR is performed. The accuracy results of such an ablative experiment for drone detection are presented in (Sazdic-Jotic, 2024b) and shown in Table 5.
Table 5 - Comparative analysis of the accuracy for different lengths of input data in
ablative experiments
Segment length [samples] Segment duration [ms] Drone detection (the AlexNet model / 2.4 GHz frequency band)
100,000 0.67 97.3
200,000 1.34 100.0
700,000 4.69 100.0
This ablative experiment was performed within the AlexNet model, the STFT method, with radio signals in the 2.4 GHz frequency band and with three-segment lengths (100,000, 200,000, and 700,000 samples). It can be observed that with the increase in the segment length of the drone radio signal, the accuracy improves. Moreover, the detection accuracy is 100% when a segment length is bigger than 200,000 samples.
Challenges in RF-based drone classification
Drone classification procedures encounter a complex challenge regardless of the sensor type and input data. As discussed in the paper (Aledhari et al, 2021), one solution, such as optoelectronic and radio sensors, uses sensor fusion. However, sensor fusion is not widely utilized in this application due to its complex implementation. Moreover, radio sensors are often paired with diverse DL algorithms to navigate the
complex and frequently unpredictable RF environment. Such a robust combination enables the ADRO system to classify and identify drones accurately, even in challenging conditions.
At (Ozturk et al, 2020; Medaiyese et al, 2022; Noh et al, 2022; Mohammed et al, 2023), great emphasis is placed on signal preprocessing as a crucial step in enhancing the accuracy and reliability of detecting and identifying drones. By carefully processing and analyzing incoming signals, it is possible to uncover hidden patterns and extract valuable information that would have otherwise been missed. This meticulous approach enables researchers to deliver superior results and stay ahead of the curve in the fast-evolving field of drone technology with improved accuracy of DL algorithms.
Finally, the multistage classification presented in (Medaiyese et al, 2022) is often employed to enhance the accuracy of the classification process and minimize the inclusion of superfluous input data. Typically, the initial stage entails detecting the drones and classifying the specific drone types. The final stage of identifying drone behavior is generally not required in practical applications.
Conclusion
This study provides a comprehensive analysis of the existing research on drone classification in the radio frequency domain and explores the potential of deep learning algorithms in addressing this issue. According to the findings, the proposed algorithms exhibit promising results in effectively resolving the problem of drone classification. However, further research is necessary to evaluate their practical implementation and testing in real-world antidrone systems. According to the findings in our review, the most effective method for categorizing drones in the radio frequency domain is through deep learning techniques. Nonetheless, it is crucial to remember that most of the research in this area is still in the experimental phase and needs to be implemented practically. One of the significant challenges is the need for a general specification of a drone classification system based on real-world requirements and experience from combat engagement.
Future research should focus on merging multiple datasets or evaluating classification techniques on different datasets to address this issue. Additionally, it is crucial to investigate the new multimodal deep learning algorithm that combines various features and raw I/Q radio signals for more accurate drone classification. Overall, this study contributes to drone classification in the radio frequency domain and highlights the need for further research.
CO
hp
</>
e n el
-C
c d n
ro
01 t e
</) ro
t
ro
d
oi
e u
c e
o
_03 o
e n ro rd d e
</) ro
.Q LL
a:
M—
o
e
a:
ro
<1J
m
O —>
T3 N CO OT
CM <1>
</) _00
cm" h-
Ô >
■sf CM o CM
OH UJ Cd
:d O o
_J
<
o
X
o
LU
I—
>-
Q! <
I—
< _J
CD >o
X LU I—
o
o >
0
References
Abeywickrama, S., Jayasinghe, L., Fu, H., Nissanka, S. & Yuen, C. 2018. RF-based Direction Finding of UAVs Using DNN. In: 2018 IEEE International Conference on Communication Systems (ICCS), Chengdu, China, pp.157-161, December 19-21. Available at: https://doi.org/10.1109/ICCS.2018.8689177.
Akter, R., Doan, V-S., Tunze, G.B., Lee, J.-M. & Kim, D.-S. 2020. RF-Based UAV Surveillance System: A Sequential Convolution Neural Networks Approach. In: 2020 International Conference on Information and Communication Technology Convergence (ICTC), Jeju, South Korea, pp.555-558, October 21-23. Available at: https://doi.org/10.1109/ICTC49870.2020.9289281.
Aledhari, M., Razzak, R., Parizi, R.M. & Srivastava, G. 2021. Sensor Fusion for Drone Detection. In: 2021 IEEE 93rd Vehicular Technology Conference (VTC2021-Spring), Helsinki, Finland, pp.1-7, April 25-28. Available at: https://doi.org/10.1109/VTC2021 -Spring51267.2021.9448699.
Allahham, M.S., Al-Sa'd, M.F., Al-Ali, A., Mohamed, A., Khattab, T. & Erbad, A. 2019. DroneRF dataset: A dataset of drones for RF-based detection, classification and identification. Data in Brief, 26, art.number:104313. Available at: https://doi.org/10.1016/j.dib.2019.104313.
Allahham, M.S., Khattab, T. & Al-Sa'd Mohamed, A.M. 2020. Deep Learning for RF-Based Drone Detection and Identification: A Multi-Channel 1-D Convolutional Neural Networks Approach. In: 2020 IEEE International Conference on Informatics, IoT, and Enabling Technologies (ICIoT), Doha, Qatar, pp.112-117, February 02-05. Available at: https://doi.org/10.1109/ICIoT48696.2020.9089657.
Al-Emadi, S. & Al-Senaid, F. 2020. Drone Detection Approach Based on Radio-Frequency Using Convolutional Neural Network. In: 2020 IEEE International Conference on Informatics, IoT, and Enabling Technologies (ICIoT), Doha, Qatar, pp.29-34, February 02-05. Available at: https://doi.org/10.1109/ICIoT48696.2020.9089489.
Al-Sa'd, M.F., Al-Ali, A., Mohamed, A., Khattab, T. & Erbad, A. 2019. RF-based drone detection and identification using deep learning approaches: An initiative towards a large open source drone database. Future Generation Computer Systems, 100, pp.86-97. Available at: https://doi.org/10.1016/j.future.2019.05.007.
Al-Shawabka, A., Restuccia, F., D'Oro, S., Jian, T., Rendon, B.C., Soltani, N., Dy, J., Ioannidis, S., Chowdhurz, K. & Melodia, T. 2020. Exposing the Fingerprint: Dissecting the Impact of the Wireless Channel on Radio Fingerprinting. In: IEEE INFOCOM 2020 - IEEE Conference on Computer Communications, Toronto, ON, Canada, pp.646-655, July 06-09. Available at: https://doi.org/10.1109/INF0C0M41043.2020.9155259.
Basak, S., Rajendran, S., Pollin, S. & Scheers, B. 2021. Drone classification from RF fingerprints using deep residual nets. In: 2021 International Conference on COMmunication Systems & NETworkS (COMSNETS), Bangalore, India, pp.548-555, January 05-09. Available at:
https://doi.org/10.1109/C0MSNETS51098.2021.9352891.
Basak, S., Rajendran, S., Pollin, S. & Scheers, B. 2023. Autoencoder based framework for drone RF signal classification and novelty detection. In: 2023 25th International Conference on Advanced Communication Technology (ICACT), Pyeongchang, South Korea, pp.218-225, February 19-22. Available at: https://doi.org/10.23919/ICACT56868.2023.10079363.
Basak, S. & Scheers, B. 2019. Passive radio system for real-time drone detection and DoA estimation. In: 2018 International Conference on Military Communications and Information Systems (ICMCIS), Warsaw, Poland, pp.1-6, May 22-23. Available at: https://doi.org/10.1109/ICMCIS.2018.8398721.
Bhattacherjee, U., Ozturk, E., Ozdemir, O., Guvenc, I., Sichitiu, M.L. & Dai, H. 2022. Experimental Study of Outdoor UAV Localization and Tracking using Passive RF Sensing. In: WiNTECH '21: Proceedings of the 15th ACM Workshop on Wireless Network Testbeds, Experimental evaluation & CHaracterization. New Orleans, USA, pp.31-38, January 31 - February 04. Available at: https://doi.org/10.1145/3477086.3480832.
Bisio, I., Garibotto, C., Haleem, H., Lavagetto, F. & Sciarrone, A. 2021. On the Localization of Wireless Targets: A Drone Surveillance Perspective. IEEE Network, 35(5), pp.249-255. Available at:
https://doi.org/10.1109/MNET.011.2000648.
Butterworth-Hayes, P. 2023. The Counter UAS Directory and Buyer's Guide. Unmanned Airspace [online]. Available at:
https://www.unmannedairspace.info/wp-content/uploads/2023/04/Counter-UAS-directory.-April-2023.v2.pdf [Accessed: 05.12.2023].
Ding, G., Wu, Q., Zhang, L., Lin, Y., Tsiftsis, T.A. & Yao, Y.D. 2018. An Amateur Drone Surveillance System Based on the Cognitive Internet of Things. IEEE Communications Magazine, 56(1), pp.29-35. Available at: https://doi.org/10.1109/MCOM.2017.1700452.
Donatus, R.E., Dere, M.D., Donatus, I.H., Chiedu, U.O., Abdulrazaq, M.B. & Ohemu, M.F. 2023. Development of an Optimised Neural Network Model for RF Based UAV Detection and Identification. Dutse Journal of Pure and Applied Sciences, 8(4b), pp.82-91. Available at: https://doi.org/10.4314/dujopas.v8i4b.10.
Drone Industry Insights UG. 2023. Commercial Drone Reports. Droneii.com [online]. Available at: https://droneii.com/drone-reports [Accessed: 28.12.2023].
Ezuma, M., Erden, F., Anjinappa, C.K., Ozdemir, O. & Guvenc, I. 2020. Detection and Classification of UAVs Using RF Fingerprints in the Presence of Wi-Fi and Bluetooth Interference. IEEE Open Journal of the Communications Society, 1, pp.60-76. Available at:
https://doi.org/10.1109/OJCOMS.2019.2955889.
Ezuma, M., Erden, F., Anjinappa, C.K., Ozdemir, O. & Guvenc, I. 2021. Drone remote controller RF signal dataset. IEEEDataPort, 11 February. Available at: https://doi.org/10.21227/ss99-8d56.
Fu, H., Abeywickrama, S., Zhang, L. & Yuen, C. 2018. Low-Complexity Portable Passive Drone Surveillance via SDR-Based Signal Processing. IEEE Communications Magazine, 56(4), pp.112-118. Available at: https://doi.org/10.1109/MCOM.2018.1700424.
CO
hp
</>
e n el
-C
c d n
ro
01 t e
</) ro
t
ro
d
oi
e u
c e
o
_03 o
e n ro rd d e
</) ro
.Q LL
cc
M—
o
e
a:
ro
<1J
m
O —>
■O TH N
ro OT
2 e u s
_00
,2 7
lo V
,4 2 0 2
of
UJ Cd ZD
O
C
_j
A C
H C
LU
T
>-
R
A T
<
CD >Q
X LU I—
o
o >
0
Haluza, M. & Cechak, J. 2016. Analysis and decoding of radio signals for remote control of drones. In: 2016 New Trends in Signal Processing (NTSP), Demanovska dolina, Slovakia, pp.1-5, October 12-14. Available at: https://doi.org/10.1109/NTSP.2016.7747781.
Hassanalian, M. & Abdelkefi, A. 2017. Classifications, applications, and design challenges of drones: a review. Progress in Aerospace Sciences, 91, pp.99-131. Available at: https://doi.org/10.1016Zj.paerosci.2017.04.003.
Huynh-The, T., Pham, Q.-V., Nguyen, T.-V., Da Costa, D.B. & Kim, D.-S. 2022. RF-UAVNet: High-Performance Convolutional Network for RF-Based Drone Surveillance Systems. IEEE Access, 10, pp.49696-49707. Available at: https://doi.org/10.1109/ACCESS.2022.3172787.
Krizhevsky, A., Sutskever, I. & Hinton, G.E. 2017. ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6), pp.84-90. Available at: https://doi.org/10.1145/3065386.
Kilie, R., Kumbasar, N., Oral, E.A. & Ozbek, I.Y. 2022. Drone classification using RF signal based spectral features. Engineering Science and Technology, an International Journal, 28, art.number:101028. Available at: https://doi.org/10.1016/jjestch.2021.06.008.
Lofu, D., Di Gennaro, P., Tedeschi, P., Di Noia, T. & Di Sciascio, E. 2023. URANUS: Radio Frequency Tracking, Classification and Identification of Unmanned Aircraft Vehicles. IEEE Open Journal of Vehicular Technology, 4, pp.921-935. Available at: https://doi.org/10.1109/0JVT.2023.3333676.
Lv, X. & Wang, Z. 2019. An Improved RF Detection Algorithm Using EMD-based WT. KSII Transactions on Internet and Information Systems, 13(8), pp.3862-3879. Available at: https://doi.org/10.3837/tiis.2019.08.003.
Mandal, S. & Satija, U. 2023. Time-Frequency Multiscale Convolutional Neural Network for RF-Based Drone Detection and Identification. IEEE Sensors Letters, 7(7), art.number:7003304. Available at:
https://doi.org/10.1109/LSENS.2023.3289145.
Medaiyese, O., Ezuma M., Lauf, A. & Adeniran, A. 2023. Cardinal RF (CardRF): An Outdoor Uav/Uas/Drone RF Signals With Bluetooth And Wifi Signals Dataset. IEEEDataPort, 08 May. Available at: https://doi.org/10.21227/1xp7-ge95.
Medaiyese, O.O., Ezuma M., Lauf, A.P. & Guvenc, I. 2022. Wavelet transform analytics for RF-based UAV detection and identification system using machine learning, Pervasive and Mobile Computing, 82, art.number:101569. Available at: https://doi.org/10.1016/j.pmcj.2022.101569.
Mohammed, K.K., El-Latif, E.I.A., El-Sayad, N.E., Darwish, A., & Hassanien, A.E. 2023. Radio frequency fingerprint-based drone identification and classification using Mel spectrograms and pre-trained YAMNet neural. Internet of Things, 23, art.number:100879. Available at:
https://doi.org/10.1016/j.iot.2023.100879.
Mokhtari, M., Bajcetic, J., Sazdic-Jotic, B. & Pavlovic, B. 2021. RF-based drone detection and classification system using convolutional neural network. In: 2021 29th Telecommunications Forum (TELFOR), Belgrade, Serbia, pp.1-4,
November 23-24. Available at: 89
https://doi.org/10.1109/TELFOR52709.2021.9653332.
Mokhtari, M. Bajcetic, J., Sazdic-Jotic, B. & Pavlovic, B. 2022. Developing a convolutional recurrent neural network for drone classification. In: 9th Small
p
Systems Simulation Symposium, Nis, Serbia, pp.85-89, February 28 - March 02 [online]. Available at: http://ssss.elfak.rs/proceedings-2022/ [Accessed: & 25.12.2023]. J
Nguyen, H.N., Vomvas, M., Vo-Huu, T. & Noubir, G. 2021. Spectro-Temporal RF Identification using Deep Learning. arXiv:2107.05114. Available at: https://doi.org/10.48550/arXiv.2107.05114.
Nguyen, P., Truong, H., Ravindranathan, M., Nguyen, A., Han, R. & Vu, T. 2018. Cost-Effective and Passive RF-Based Drone Presence Detection and $ Characterization. GetMobile: Mobile Computing and Communications, 21(4), ro
d
e
pp.30-34. Available at: https://doi.org/10.1145/3191789.3191800.
Noh, D.-I., Jeong, S.-G., Hoang, H.-T., Pham, Q.-V., Huynh-The, T., Hasegawa, M., Sekiya, H., Kwon, S.-Y., Chung, S.-H. & Hwang, W.-J. 2022. Signal Preprocessing Technique With Noise-Tolerant for RF-Based UAV Signal hc Classification. IEEE Access, 10, pp.134785-134798. Available at: " https://doi.org/10.1109/ACCESS.2022.3232036.
Oh, J., Lim, D.-W. & Kang, K.-M. 2020. Unmanned Aerial Vehicle ac Identification Success Probability with LoRa Communication Approach. In: 2020 ^ IEEE 31st Annual International Symposium on Personal, Indoor and Mobile Radio j§ Communications, London, UK, pp.1-6, August 31 - September 03. Available at: £ https://doi.org/10.1109/PIMRC48278.2020.9217172.
Ozturk, E., Erden, F. & Guvenc, I. 2020. RF-Based Low-SNR Classification of UAVs Using Convolutional Neural Networks. arXiv:2009.05519. Available at: https://doi.org/10.48550/arXiv.2009.05519.
Sazdic-Jotic, B. 2024a. CUAS_Literature_DataSet. Mendeley Data, 20 February. Available at: https://doi.org/10.17632/6jc6w2v5nf.2.
Sazdic-Jotic, B. 2024b. Klasifikacija dronova u radio-frekvencijskom domenu primenom algoritama dubokog ucenja. Ph.D. thesis. Belgrade, Serbia: University > of Defence in Belgrade, Military Academy (in Serbian). Re
Sazdic-Jotic, B., Pokrajac, I., Bajcetic, J., Bondzulic, B., Joksimovic, V., -5 Sevic, T. & Obradovic, D. 2021. VTI_DroneSET_FFT. Mendeley Data, 04 "S January. Available at: https://doi.org/10.17632/s6tgnnp5n2.3. cci
Sazdic-Jotic, B., Bondzulic, B., Pokrajac, I., Bajcetic, J. & Mohammed, M. 2023. Drone classification based on radio frequency: techniques, datasets, and challenges. In: 10th International Scientific Conference on Defensive Technologies, OTEH 2023, Belgrade, Serbia, pp.314-320, October 13-14 [online]. & Available at: http://www.vti.mod.gov.rs/oteh/elementi/rad/031.pdf [Accessed: 05.02.2024].
Sazdic-Jotic, B., Pokrajac, I., Bajcetic, J., Bondzulic, B. & Obradovic, D. 2022. Single and multiple drones detection and identification using RF based deep learning algorithm. Expert Systems with Applications, 187, art.number:115928. Available at: https://doi.org/10.1016/j.eswa.2021.115928.
d e
ro
.Q
OH
e
T3
2 e u s
_00
,2 7
lo V
,4 2 0 2
of
UJ Cd ZD
O
C
_j
A C
H C
LU
T
>-
R
A T
<
CD >Q
X LU I—
o
o >
0
Schiller, N., Chlosta, M., Schloegel, M., Bars, N., Eisenhofer, T., Scharnowski, T., Domke, F., Schonherr, L. & Holz, T. 2023. Drone Security and the Mysterious Case of DJI's DronelD. In: Network and Distributed System Security (NDSS) Symposium, San Diego, CA, USA, pp.1-17, February 27 - March 03. Available at: https://doi.org/10.14722/ndss.2023.24217.
Sciancalepore, S., Ibrahim, O.A., Oligeri, G. & Di Pietro, R. 2020. PiNcH: an effective, efficient, and robust solution to drone detection via network traffic analysis. Computer Networks, 168, art.number:107044. Available at: https://doi.org/10.1016/j.comnet.2019.107044.
Shi, X., Yang, C., Xie, W., Liang, C., Shi, Z. & Chen, J. 2018. Anti-Drone System with Multiple Surveillance Technologies: Architecture, Implementation, and Challenges. IEEE Communications Magazine, 56(4), pp.68-74. Available at: https://doi.org/10.1109/MCOM.2018.1700430.
Shorten, D., Williamson, A., Srivastava, S. & Murray, J.C. 2018. Localisation of Drone Controllers from RF Signals using a Deep Learning Approach. In: PRAI 2018: Proceedings of the International Conference on Pattern Recognition and Artificial Intelligence, Union Township, NJ, USA, pp.89-97, August 15-17. Available at: https://doi.org/10.1145/3243250.3243272.
Soltani, N., Reus-Muns, G., Salehi, B., Dy, J., Ioannidis, S. & Chowdhury, K. 2020. RF Fingerprinting Unmanned Aerial Vehicles with Non-Standard Transmitter Waveforms. IEEE Transactions on Vehicular Technology, 69(12), pp.15518-15531. Available at: https://doi.org/10.1109/TVT.2020.3042128.
Stoica, P., Basak, S., Molder, C. & Scheers, B. 2020. Review of Counter-UAV Solutions Based on the Detection of Remote Control Communication. In: 2020 13th International Conference on Communications (COMM), Bucharest, Romania, pp.233-238, June 18-20. Available at: https://doi.org/10.1109/COMM48946.2020.9142017.
Swinney, C.J. & Woods, J.C. 2020. Unmanned Aerial Vehicle Flight Mode Classification using Convolutional Neural Network and Transfer Learning. In: 2020 16th International Computer Engineering Conference (ICENCO), Cairo, Egypt, pp.83-87, December 29-30. Available at: https://doi.org/10.1109/ICENCO49778.2020.9357368.
Swinney, C.J. & Woods, J.C. 2021. RF Detection and Classification of Unmanned Aerial Vehicles in Environments with Wireless Interference. In: 2021 International Conference on Unmanned Aircraft Systems (ICUAS), Athens, Greece, pp.1494-1498, June 15-18. Available at: https://doi.org/10.1109/ICUAS51884.2021.9476867.
Swinney, C.J. & Woods, J.C. 2022. DroneDetect Dataset: A Radio Frequency Dataset Of Unmanned Aerial System (UAS) Signals For Machine Learning Detection & Classification. IEEEDataPort 17 May. Available at: https://doi.org/10.21227/5jjj-1m32.
Vuorenmaa, M., Marin, J., Heino, M., Turunen, M., Riihonen, T. & Parlin, K. 2020. Radio-Frequency Control and Video Signal Recordings of Drones. Zenodo, 05 November. Available at: https://doi.org/10.5281/zenodo.4264467.
Zhang, H., Cao, C., Xu, L. & Gulliver, T.A. 2018. A UAV Detection Algorithm g
Based on an Artificial Neural Network. IEEE Access, 6, pp.24720-24728. ^
Available at: https://doi.org/10.1109/ACCESS.2018.2831911. £
Zhang, Y. 2021. RF-based Drone Detection using Machine Learning. In: 2021 2nd International Conference on Computing and Data Science (CDS),
Stanford, CA, USA, pp.425-428, January 28-29. Available at: &
https://doi.org/10.1109/CDS52072.2021.00079. J
ro .c
Revisión de la clasificación de drones basada en RF: técnicas, conjuntos de datos y desafíos
Boban Sazdic-Jotica, autor de correspondencia, Ivan Pokrajac3, "S
Jovan Bajceticb, Nenad Stefanovicc ro
ta
a Instituto Técnico Militar, Belgrado, República de Serbia
Khaoticen, Belgrado, República de Serbia <ü
c Fuerzas Armadas de Serbia, Centro de Matemáticas Aplicadas y Electrónica, Belgrado, República de Serbia
CAMPO: ciencias de computación, telecomunicación TIPO DE ARTÍCULO: artículo de revisión
a c
o <u
Resumen:
Introducción/objetivo: Este artículo analiza la bibliografía disponible públicamente sobre la clasificación de drones en el dominio de la radiofrecuencia, centrándose en la detección e identificación. Los drones se utilizan cada vez más con fines ilegales, lo que hace que las técnicas de clasificación sean cruciales. Este artículo de revisión cubre sensores pasivos de radiofrecuencia, técnicas de clasificación y conjuntos de datos que destacan los desafíos. RF
Métodos: Los investigadores están desarrollando soluciones anti drones porque los drones se han convertido en herramientas valiosas y objetivos para actividades ilegales. Debido al alcance del tema, la revisión incluyó solo la clasificación de drones mediante sensores pasivos de radiofrecuencia con una descripción de las técnicas de clasificación (conjunto de algoritmos, métodos y procedimientos) y los conjuntos de datos utilizados para las pruebas de rendimiento. Este estudio introduce una nueva categoría y ofrece información más profunda sobre las técnicas de clasificación de drones disponibles públicamente.
Resultados: Con base en los resultados de este estudio, es evidente que los algoritmos de aprendizaje profundo son actualmente el enfoque más eficaz para abordar el desafío de la clasificación de drones dentro del dominio de la radiofrecuencia. Uno de los principales obstáculos es la ausencia de un estándar integral para clasificar los drones en el dominio de la radiofrecuencia, que debería basarse en los requisitos del usuario final. Además, los resultados de dos experimentos ablativos destacan el
preprocesamiento de señales de radio I/Q sin procesar como un paso esencial en la clasificación de drones.
Conclusión: En resumen, la clasificación propuesta proporciona una herramienta valiosa para la revisión de la bibliografías. El aprendizaje profundo es la técnica más eficaz para la clasificación de drones, pero los conjuntos de datos disponibles públicamente con señales de radio de drones son limitados. La fortaleza clave de este estudio es que representa la primera revisión de conjuntos de datos disponibles públicamente con señales de radio de drones.
Palabras claves: aprendizaje profundo, drone, detección, clasificación, identificación, radiofrecuencia.
Обзор классификации дронов на основании радиочастотного диапазона: методы, наборы данных и вызовы
Бобан Саздич-Йотича, корреспондент, Иван Покраяца, Йован Байчетич6, Ненад Стефановичв
а Военно-технический институт, г. Белград, Республика Сербия
б ООО «Khaoticen», г. Белград, Республика Сербия
в Вооруженные силы Республики Сербия, центр прикладной математики и электроники, г. Белград, Республика Сербия
РУБРИКА ГРНТИ: 28.23.37 Нейронные сети (Искусственный интеллект) ВИД СТАТЬИ: обзорная статья
Резюме:
Введение/цель: За последнее десятилетие многократно возросло использование беспилотных летательных аппаратов или дронов как в коммерческих (гражданских или любительских), так и в функциональных (военных или промышленных) целях. В связи с этим в данной статье представлен всесторонний обзор общедоступной и актуальной литературы по классификации (обнаружению и идентификации) дронов и их радиочастотного спектра. Особое внимание уделено алгоритмам глубокого обучения и результатам, полученным из общедоступного набора данных VTI_DroneSET.
Методы: Благодаря значительному прогрессу дроны стали полезным инструментом в осуществлении различных целей. Дополнительным преимуществом является то, что они стали дешевле и доступнее, вследствие чего увеличилась опасность от использования дронов в противозаконной деятельности. Такое развитие событий вызвало повышенное вовлечение исследователей в разработку решений по борьбе с дронами. С учетом большого объема общедоступных исследований в данной статье рассматривается исключительно вид дронов,
собирающих данный с помощью пассивных радиочастотных § датчиков с описанием используемых методов (набора алгоритмов, методов и процедур) и наборов данных, <2 используемых в испытаниях эффективности. Для понимания проблемы классификации дронов был проведен количественный и качественный анализы методами технического анализа и & обработки радиосигналов. Количественные показатели с .§ графическим изображением использовались в систематизации собранных статей, в то время как для определения возможности классификации дронов по радиочастотному диапазону использовались алгоритмы глубокого обучения. Помимо этого, в данной статье представлены вызовы и ограничения $ классификации дронов на основании радиосигналов. го
Результаты: Результаты данного исследования доказывают, <л что алгоритмы глубокого обучения в настоящее время являются наиболее эффективным подходом к решению проблемы классификации дронов по радиочастотному диапазону. Однако следует отметить, что большинство современных исследований имеют экспериментальный характер, следовательно, они ограничены в практическом применении. о Главной проблемой является отсутствие общей спецификации в классификации дронов по радиочастотному диапазону, основанной на требованиях, исходящих из ежедневной практики.
Выводы: Вклад данного исследования заключается в систематизации всех доступных работ, посвященных классификации дронов по радиочастотному диапазону, и I представлении некоторых возможностей алгоритмов глубокого обучения. Можно сделать вывод, что предложенные алгоритмы £ могут быть использованы в решении этой проблемы, а в ближайшем будущем можно будет испытать на практике в реальных сценариях антидрон систему защиты от беспилотных летательных аппаратов.
Ключевые слова: глубокое обучение, дроны, обнаружение, классификация, идентификация, радиочастоты. т
тз
Преглед класификаци]е дронова у радио-фреквенци]ском домену: технике, скупови података и изазови ^
Бобан Сазди^отиЧ, аутор за преписку, Иван Покра]аца, м
Jован Ба]четиЧ, Ненад Стефанович а Во]нотехнички институт, Београд, Република Срби]а б КИаойсеп, Београд, Република Срби]а
в Во]ска Срби]е, Центар за приме^ену математику и електронику, Београд, Република Срби]а
см ф
(Л _00
с\Г г*-
о >
■чг" Г о Г
он ш он и О
о <
о
X
о ш
н
>-
он с
I-
сл с
о >о
X ш н
о
о >
ф
ОБЛАСТ: рачунарске науке, телекомуникаци]е КАТЕГОРИJА (ТИП) ЧЛАНКА: прегледни рад
Сажетак:
Увод/цил: КоришЯече беспилотних ваздухоплова - дронова, током последче децени}е многоструко се увеПало, како за комерци]алне (цивилне или аматерске), тако и за функционалне (во}не или индустри}ске) потребе. У овом истраживачком раду представлен }е свеобухватан преглед }авно доступне и актуелне литературе о класификаци]и (детекцща и идентификаци}а) дронова у радио-фреквенци}ском домену. Посебан аспект представла}у алгоритми за дубоко учече и резултати ко\и су доби}ени са }авно доступним скупом (базом) података VTI_DroneSET
Методе: Захвалу}уЬи значащим унапре^ечима дронови су постали корисна средства за различите намене. Додатна погодност }есте што су ]ефтини]и и приступачни]и за коришЯече, што представка опасност од чихове злоупотребе. Стога }е по}ачано ангажоваче истраживача на разво}у антидрон решена. Због обима }авно доступних истраживача, ова] рад }е обухватио само класификаци}у дронова путем пасивних радио-фреквенци]ских сензора са описом коришЯених техника (скуп алгоритама, метода и процедура) и скупова података ко\и се користе за тестираче перформанси. Ради разумевача проблема класификаци}е дронова извршена }е квантитативна и квалитативна анализа са методама техничке анализе и обраде радио-сигнала. Квантитативни показатели са графичким илустраци}ама коришЯени су за систематизаци}у прикуплених радова, док су за утвр^иваче могуЬности класификаци}е дронова у радио-фреквенци}ском домену коришЯени алгоритми дубоког учеча. Штавише, представлени су изазови и ограничеча кяасификаци}е дронова на основу радио-сигнала. Резултати: Показано }е да су алгоритми дубоког учеча тренутно на}боле решече за решаваче питача класификаци}е дронова у радио-фреквенци}ском домену. Ме^утим, веЯина савремених истраживача }е експериментална и има ограничену практичну имплементацщу. Посебан проблем представла недостатак опште спецификаци]е за класификаци]у дронова у радио-фреквенци}ском домену на основу захтева из свакодневног искуства.
Заклучак: Допринос овог истраживача }е у систематизации свих доступних радова ко]и се баве кяасификаци}ом дронова у радио-фреквенци}ском домену, као и у приказу неких могуЬности алгоритама дубоког учеча. Може се заклучити да се предложени алгоритми могу искористити за наведену примену, те да }е у
наредном периоду могуче тестирати практичне g имплементаци]е, као и вршити тестира^е у реалним сценари]има употребе антидрон система.
Къучне речи: дубоко уче^е, дрон, детекци}а, класификаци}а, идентификаци}а, радио-фреквенци}а.
m
C JD
Paper received on: 15.02.2024. °
Manuscript corrections submitted on: 07.06.2024. Paper accepted for publishing on: 08.06.2024.
© 2024 The Authors. Published by Vojnotehnicki glasnik / Military Technical Courier $ (www.vtg.mod.gov.rs, BTr.MO.ynp.cp6). This article is an open access article distributed under the -2 terms and conditions of the Creative Commons Attribution license "o (http://creativecommons.org/licenses/by/3.0/rs/). &
о <u
ro о
<л
(Л
_оз о <u
T3
<u
ro
.Q
OH
<u >
<U
OH "rô Ф
m
тз N ГО
œ
