Robust Drone Detection and Classiﬁcation from Radio Frequency

Signals Using Convolutional Neural Networks

Stefan Gl

uge

1 a

, Matthias Nyfeler

1 b

, Nicola Ramagnano

, Claus Horn

1 c

and Christof Sch

upbach

3 d

Institute of Computational Life Sciences, Zurich University of Applied Sciences, 8820 W

adenswil, Switzerland

Institute for Communication Systems, Eastern Switzerland University of Applied Sciences,

8640 Rapperswil-Jona, Switzerland

armasuisse Science + Technology, 3602 Thun, Switzerland

Keywords:

Deep Learning, Robustness, Signal Detection, Unmanned Aerial Vehicles.

Abstract:

As the number of unmanned aerial vehicles (UAVs) in the sky increases, safety issues have become more

pressing. In this paper, we compare the performance of convolutional neural networks (CNNs) using ﬁrst, 1D

in-phase and quadrature (IQ) data and second, 2D spectrogram data for detection and classiﬁcation of UAVs

based on their radio frequency (RF) signals. We focus on the robustness of the models to low signal-to-noise

ratios (SNRs), as this is the most relevant aspect for a real-world application. Within an input type, either IQ

or spectrogram, we found no signiﬁcant difference in performance between models, even as model complexity

increased. In addition, we found an advantage in favor of the 2D spectrogram representation of the data. While

there is basically no performance difference at SNRs ≥ 0 dB, we observed a 100% improvement in balanced

accuracy at −12 dB, i.e. 0.842 on the spectrogram data compared to 0.413 on the IQ data for the VGG11

model. Together with an easy-to-use benchmark dataset, our ﬁndings can be used to develop better models for

robust UAV detection systems.

1 INTRODUCTION

Drones, or civil UAVs, have evolved from hobbyist

toys to commercial systems with many applications.

As more drones ﬂy in the sky, safety issues are be-

coming more pressing. Regulations and technical so-

lutions (such as transponder systems) are needed to

safely integrate UAVs into the airspace. However,

even with a standard airspace integration, drones can

still pose serious threats. Safety regulations can be

circumvented by technical and human error or delib-

erate misuse. To protect critical infrastructure such

as airports, drone detection and classiﬁcation systems

are needed that do not depend on the cooperation of

the UAV. Various technologies such as audio, video,

radar, or RF scanners have been proposed for this task

(Kunze and Saha, 2022).

In this paper, we investigate different approaches

https://orcid.org/0000-0002-7484-536X

https://orcid.org/0000-0001-7929-7625

https://orcid.org/0000-0003-1557-7913

https://orcid.org/0000-0001-5822-3360

for the detection and classiﬁcation of drones based

on their RF signals. We compare the performance of

CNNs using two different representations of the input

data: ﬁrst, raw IQ data, without requiring much pre-

processing (except for windowing and normalization)

and second, spectrogram data computed with consec-

utive Fourier transforms for the real and imaginary

parts of the signal. In terms of performance, we focus

on the robustness of the models to low SNRs, as this

is the most relevant aspect for a real-world application

of the system. To facilitate future model development,

we provide an easy-to-use benchmark dataset.

In the next section, we brieﬂy review related work

in this research area, followed by a description of the

data collection and data preprocessing procedure in

Section 2. Section 3 describes the model architectures

and their training/validation method. The resulting

performance metrics are presented in Section 4 and

further discussed in Section 5.

496

Glüge, S., Nyfeler, M., Ramagnano, N., Horn, C. and Schüpbach, C.

Robust Drone Detection and Classiﬁcation from Radio Frequency Signals Using Convolutional Neural Networks.

DOI: 10.5220/0012176800003595

In Proceedings of the 15th International Joint Conference on Computational Intelligence (IJCCI 2023), pages 496-504

ISBN: 978-989-758-674-3; ISSN: 2184-3236

1.1 Related Work

A literature review of machine learning (ML) ap-

proaches for drone detection and classiﬁcation that

refelcts the state of the art in 2019, is provided

by Taha and Shoufan (Taha and Shoufan, 2019).

The authors provide an overview of approaches for

radar, visual data, acoustic data, and RF-based sys-

tems. Therefore, we brieﬂy describe how the ﬁeld has

evolved since 2019 and reassess our use case, namely

RF data in a noisy environment.

The openly available DroneRF dataset (Allahham

et al., 2019) has been used in several works (Al-

Sa’d et al., 2019; Swinney and Woods, 2020; Zhang,

2021). It contains RF recordings from three drones

in four ﬂight modes (i.e., on, hovering, ﬂying, video

recording). It was recorded using universal soft-

ware radio peripheral (USRP) software-deﬁned radio

(SDR) transceivers. Signals that could be considered

noise in the 2.4 GHz band (Bluetooth, Wi-Fi) were

not recorded. Furthermore, the dataset contains only

time series data, and not the complex IQ signals.

Together with the dataset, the authors proposed

three deep neural networks to detect the presence of

a drone, the presence of a drone and its type, and the

presence of a drone, its type, and its ﬂight mode. The

average accuracy is reported to be 99.7% for the 2-

class problem, 84.5% for the 4-class problem, and

46.8% for the 10-class problem, respectively (Al-Sa’d

et al., 2019).

Medaiyese et al. (Medaiyese et al., 2021) propose

a semi-supervised framework for UAV detection us-

ing wavelet analysis. Accuracy between 86% and

97% was achieved at SNRs of 30 dB and 18 dB, while

it dropped to chance level for SNRs below 10 dB to

6 dB. They recorded a dataset consisting of six differ-

ent types of drones, as well as two Bluetooth devices

and two Wi-Fi devices. Unfortunately, this dataset is

not openly available.

The openly available DroneDetect V2 dataset was

created by Swinney and Woods (Swinney and Woods,

2021). It contains raw IQ data recorded with a

BladeRF SDR. Seven drone models were recorded

in three different ﬂight modes (on, hovering, ﬂy-

ing). Measurements were also repeated with different

types of noise, such as interference from a Bluetooth

speaker, a Wi-Fi hotspot, and simultaneous Bluetooth

and Wi-Fi interference. The dataset does not include

measurements without drones, which would be nec-

essary to evaluate a drone detection system. The re-

sults in (Swinney and Woods, 2021) show that Blue-

tooth signals are more likely to interfere with detec-

tion and classiﬁcation accuracy than Wi-Fi signals.

Overall, frequency domain features extracted from a

Figure 1: Recording of drone signals in the anechoic cham-

ber. A DJI Phantom 4 Pro drone with the DJI Phantom

GL300F remote control.

CNN where shown to be more robust than time do-

main features in the presence of interference.

Unlike most deep learning approaches, Ge et al.

(Ge et al., 2021) focus on pre-processing and combin-

ing signals from two frequency bands before feeding

them into a neural network classiﬁer to improve clas-

siﬁcation accuracy, which is reported to improve from

46.8% to 91.9%.

Zhang et al. (Zhang et al., 2023) used downlink

video data and focused on data augmentation with

various environmental signals. They applied data

segmentation techniques of spectral features with a

ResNet architecture for the RF data with a bandwidth

of 100 MHz with SNRs ranging from 20 dB down to

0 dB. The accuracy at the lowest SNR of 0 dB was still

around 70%.

2 MATERIALS

2.1 Data Acquisition

To record pure drone signals without any interference,

the drone remote control and, if present, the drone

itself were placed inside the anechoic chamber, see

Figure 1. The signals were received by a LogPer an-

tenna and sampled and stored by an Ettus Research

USRP B210. In the static measurement, the respec-

tive signals of the remote control (TX) alone or with

the drone (RX) were measured. In the dynamic mea-

surement, one person at a time was inside the ane-

choic chamber and operated the remote control (TX)

to generate a signal that is as close to reality as pos-

sible. All signals were recorded at a sampling fre-

quency of 56 MHz (highest possible real-time band-

width). All drone models with recording parameters

are listed in Table 1, including both uplink and down-

link signals.

To access the robustness of a drone detection

Robust Drone Detection and Classiﬁcation from Radio Frequency Signals Using Convolutional Neural Networks

497

Table 1: Transmitters and receivers recorded in the dataset and their respective labels. Additionally, we show the class

label used, the center frequency (GHz), the channel spacing (MHz), the burst duration (ms), and the repetition period of the

respective signals (ms).

Transmitter Receiver Label Center Freq. Spacing Duration Repetition

DJI Phantom GL300F DJI Phantom 4 Pro DJI 2.44175 1.7 2.18 630

Futaba T7C - FutabaT7 2.44175 2 1.7 288

Futaba T14SG Futaba R7008SB FutabeT14 2.44175 3.1 1.4 330

Graupner mx-16 Graupner GR-16 Graupner 2.44175 1 1.9/3.7 750

Bluetooth/Wi-Fi Noise - Noise 2.44175

Taranis ACCST X8R Receiver Taranis 2.440 1.5 3.1/4.4 420

Turnigy 9X - Turnigy 2.445 2 1.3 61, 120-2900

model, we further considered three types of noise

and interference. First, Bluetooth/Wi-Fi noise was

recorded using the hardware setup described above.

The measurements were performed in a public and

busy university building. In this open recording setup,

we had no control over the exact number or types of

active Bluetooth/Wi-Fi devices and the actual trafﬁc

in progress.

Second, artiﬁcial white Gaussian noise was used,

and third, receiver noise from the USRP was used

at various gain settings without the antenna attached.

This should prevent the ﬁnal model from misclassi-

fying quantization noise in the absence of a signal,

especially at low gain settings.

2.2 Data Preparation

To reduce memory consumption and computational

effort, we reduced the bandwidth of the signals by

downsampling from 56 MHz to 14 MHz using the

SciPy (Virtanen et al., 2020) signal.decimate function

with an 8th order Chebyshev type I ﬁlter.

The drone signals come in short bursts with some

low power gain or background noise in between. For

training purposes, we divided the signals into vectors

of 16384 samples (≈ 1.2 ms). Only vectors contain-

ing a burst, or at least a partial burst, were considered

for training. This was achieved by applying an energy

threshold as shown in Figure 2.

The selected drone signal vectors x with i ∈

{1,... k} were normalized to a carrier power of 1 per

sample, i.e. only the part of the signal vector contain-

ing drone bursts was considered for the power cal-

culation (m samples out of k). This was achieved

by identifying the bursts as the samples where a

smoothed energy was above a threshold, as shown in

Figure 3. The signal vectors x are thus normalized by

ˆx(i) = x (i) /

∑

|x(i)|

. (1)

Noise vectors (Bluetooth, Wi-Fi, Ampliﬁer,

Gauss) n with samples i ∈ {1,...k } were normalized

Figure 2: 10ms signal drone signal from DJI PhantomPro4

showing a signal burst. Only the vectors with average en-

ergy (orange) above the threshold (red) were used for train-

ing. The start/end of the considered vectors of 16384 sam-

ples are shown as green dashed lines. The large y-values are

due to the not yet normalized raw int16 data.

Figure 3: Drone signal vector with carrier normalization.

Only those samples whose smoothed energy (orange) is

above the threshold (red) are used for the normalization.

to a mean power of 1 using

ˆn(i) = n(i)/

∑

|n(i)|

. (2)

To train robust models, we mixed the drone signal

vectors with noise at different SNRs. Since the signal

NCTA 2023 - 15th International Conference on Neural Computation Theory and Applications

498

carrier power and the noise power were both normal-

ized to 1, we added separate normalized noise vec-

tors ˆn, not considered in the noise class in the training

dataset, to each normalized signal vector ˆx as

ˆy(i) =



√

k · ˆx(i) + ˆn(i)



√

k +1

, with k = 10

SNR/10

, (3)

to generate the normalized vectors y at different SNRs

used for training and validation.

Given the mixed normalized IQ vectors, we

computed the spectrograms using the ScyPi sig-

nal.sepctrogram function with a Tukey window. Of-

ten only the absolute values of the spectrum are used,

but in this study the full complex spectrum was con-

sidered to preserve phase information.

2.3 Benchmark Dataset

To facilitate future model development, we provide

an easy-to-use benchmark dataset

together with a

code example to get started

. The dataset consists of

the non-overlapping signal vectors of length 16384,

which corresponds to ≈ 1.2 ms at 14 MHz. We also

added Labnoise (Bluetooth, Wi-Fi, Ampliﬁer) and

Gaussian noise to the dataset.

After normalization, the drone signals were mixed

with either Labnoise (50%) or Gaussian noise (50%).

The noise class was created by mixing Labnoise and

Gaussian noise in all possible combinations (i.e., Lab-

noise + Labnoise, Labnoise + Gaussian noise, Gaus-

sian noise + Labnoise, and Gaussian noise + Gaus-

sian noise). For the drone signal classes, as for the

noise class, the number of samples for each level of

SNR is evenly distributed over the interval of SNRs

∈ [−20,30] dB in steps of 2 dB, i.e., 3792-3800 sam-

ples per SNR level.The resulting number of samples

per class is shown in Table 2.

After data normalization and mixing, we com-

puted the power spectrum of each sample with con-

secutive Fourier transforms with non-overlapping

segments of length 128 for the real and imaginary

parts of the signal. That is, the two IQ signal vec-

tors ([2 × 16384]) are represented as two matrices

([2 × 128 ×128]). Figure 4 shows four samples of

the data set for different types of drones at different

SNRs.

https://www.kaggle.com/datasets/sgluege/noisy-

drone-rf-signal-classiﬁcation

https://github.com/sgluege/Noisy-Drone-RF-Signal-

Classiﬁcation

3 METHODS

3.1 Model Architecture and Training

We tested different conﬁgurations of the Visual Ge-

ometry Group (VGG) CNN architecture (Simonyan

and Zisserman, 2015). The main idea of this archi-

tecture is to use multiple layers of small (3 ×3) con-

volutional ﬁlters instead of larger ones. This is in-

tended to increase the depth and expressiveness of

the network, while reducing the number of parame-

ters. The VGG architecture consists of several vari-

ants, such as VGG11 to VGG19, which differ in

the number of convolutional layers (11 and 19, re-

spectively). The VGG architecture achieves state-of-

the-art results on the ImageNet (Deng et al., 2009)

dataset, which contains 14 million images belonging

to 1000 classes, and outperforms many previous mod-

els, such as AlexNet (Krizhevsky et al., 2012) and ZF-

Net (Zeiler and Fergus, 2014). In addition to image

classiﬁcation, the architecture is also widely used as a

feature extractor for other computer vision tasks such

as object detection, face recognition (Parkhi et al.,

2015), and semantic segmentation (Long et al., 2015).

To be able to process 1D IQ data, we adapted

the VGG architecture from using Conv2D layers

to Conv1d, MaxPool2d to MaxPool1d, and Adap-

tiveAvgPool2d to AdaptiveAvgPool1d, respectively.

For the spectrogram input data the VGG architecture

can be used as is. For the dense classiﬁcation layer,

we used 256 linear units followed by 7 linear units at

the output (one unit per class).

For network training, we used a stratiﬁed 5-fold

train-validation-test split. In each fold, we trained a

network using 80% and 20% of the available samples

for each class for training and testing, respectively.

Repeating the stratiﬁed split ﬁve times ensures that

each sample was in the test set once in each exper-

iment. Within the training set, 20% of the samples

were used as the validation set during training.

Model training was performed for 25 epochs with

a batch size of 64. PyTorch’s (Paszke et al., 2019) im-

plementation of stochastic gradient descent optimiza-

tion (Bottou, 1999) was used with a ﬁxed momentum

of 0.9. In addition, we applied a learning rate decay

by a factor of 0.1 if the validation loss did not im-

prove within the last 3 epochs of training. The initial

learning rate was set to 0.005.

3.2 Model Evaluation

After each training epoch, the model was evaluated

on the validation set. The model with the highest bal-

anced accuracy was then saved and later evaluated on

Robust Drone Detection and Classiﬁcation from Radio Frequency Signals Using Convolutional Neural Networks

499

Table 2: Number of samples in the different classes in the benchmark dataset.

Class DJI FutabaT14 FutabaT7 Graupner Taranis Turnigy Noise

#samples 2194 6938 3661 6481 16546 10333 52552

(a) Turnigy at SNR 22 dB. (b) FutubaT14 at SNR 0 dB.

Figure 4: Spectrogram and IQ data samples from the benchmark dataset for different drones at different SNRs (a-c) and noise

(d).

the withheld test data. The performance of the models

on the test data was evaluated in terms of classiﬁca-

tion accuracy and balanced accuracy.

Accuracy is the simplest metric, measuring the

proportion of correct predictions out of the total num-

ber of observations. It is simply calculated as the

number of correct classiﬁcations divided by the total

number of samples. However, accuracy can be mis-

leading if the data is unbalanced. In our case, the

noise class is overrepresented in the dataset (see Ta-

ble 2).

Balanced accuracy is deﬁned as the average of the

recall obtained for each class, i.e. it gives equal weight

to each class regardless of how frequent or rare it is.

4 RESULTS

Table 3 shows the mean ±standard deviation of accu-

racy and balanced accuracy on the test data, obtained

in the 5-fold cross-validation of the different models.

The models using the IQ data at the input consistently

perform 10% worse than those using the spectrogram

data.

NCTA 2023 - 15th International Conference on Neural Computation Theory and Applications

500

Within an input type, either IQ or spectrogram,

there is no signiﬁcant difference in performance be-

tween the models, even as the model complexity in-

creases from VGG11 to VGG19.

The number of epochs for training (#epochs)

shows when the highest balanced accuracy was

reached in the validation. It can be seen that

the less complex models (e.g. VGG11) need more

time/epochs compared to the more complex models.

However, the resulting classiﬁcation performance is

the same.

To further assess the robustness of the models to

noise, we computed the accuracy and balance accu-

racy separately for each noise level in the dataset. Fig-

ure 5 shows the resulting 5-fold mean (balanced) ac-

curacy over SNRs ∈ [−20, 30]dB in 2 dB steps. Note

that we do not show the standard deviation to keep

the plots readable. In general, we observe a degrada-

tion in performance from 0 dB down to near chance

level at −20 dB. At the lowest SNR level, we observe

a large difference between the accuracy and the bal-

anced accuracy. For example, about 0.65 for accu-

racy and 0.3 for balanced accuracy, for models trained

on the spectrogram data. The reason for this is the

overrepresentation of the noise class in the dataset, to-

gether with the fact that the vast majority of misclas-

siﬁcations occur between noise and drones and not

between different types of drones. Figure 6 shows the

confusion matrix for the VGG11 model with spectro-

gram data (VGG11 SPEC) at SNR −20 dB for a sin-

gle validation on the test data. It illustrates the fact

that the model mainly misclassiﬁes drone signals as

noise, which is to be expected at such a low SNR.

To access the spread of the results at low SNR lev-

els, we show the 5-fold mean ± standard deviation of

accuracy and balanced accuracy in Tables 4 and 5 for

SNR 0 dB, −6 dB, −12 db, and −18 dB.

5 DISCUSSION

In our experiment, we saw a signiﬁcant advantage in

favor of the 2D spectrogram representation of the data

over the 1D IQ representation. While there is no per-

formance gap at SNRs ≥ 0 dB, we observed a huge

difference at lower SNRs (see Figure 5). For example

at −12 dB, there is still a decent balanced accuracy of

0.842 on the spectrogram data compared to 0.413 on

the IQ data for the VGG11 model (see Table 5).

The obvious question is why the spectrogram rep-

resentation seems to be easier to separate in noisy

conditions. Part of the explanation may lie in the basic

architecture of the VGG model. It was originally de-

veloped for image classiﬁcation tasks, so it should not

(a)

(b)

Figure 5: Mean accuracy a) and balanced accuracy b) ob-

tained in the 5-fold cross-validation of the different models

on the drone classiﬁcation task over the SNR. The different

types of input data are shown as solid lines for spectrogram

(SPEC) and dashed lines for IQ.

be surprising that it is better suited to the problem of

2D spectrogram data compared to 1D IQ data. How-

ever, given the complexity of the models, one might

assume that they can learn the necessary features from

1D representations. Apparently, in our case, this as-

sumption only holds true for SNRs≥ 0.

Since drones have a much narrower bandwidth

than noise, one might expect them to be easier to de-

tect in frequency space. However, Fourier transform-

ing the IQ vectors (16384 samples) did not give bet-

ter performance at SNRs< 0 dB. This conﬁrms the as-

sumption that a deep CNN can learn the necessary ﬁl-

ters for a Fourier transform. However, the combined

time- and frequency-domain information in the spec-

trogram seems to help the network focus on both fre-

quency information and the temporal structure of the

signal bursts.

Given our benchmark dataset, it is possible to op-

timize the model side of the problem and perhaps ﬁnd

a model architecture with comparable or better per-

formance using the IQ data, for example with neural

Robust Drone Detection and Classiﬁcation from Radio Frequency Signals Using Convolutional Neural Networks

501

Table 3: Mean ± standard deviation of the accuracy (Acc.) and the balanced accuracy (balanced Acc.) obtained in 5-fold

cross-validation of the different models in the drone classiﬁcation task for different types of input data, i.e. spectrogram

(SPEC) and IQ. The best result is highlighted. An indication of the model training time is given with the mean ± standard

deviation of the number of training epochs (#epochs). The number of trainable parameters (#params) indicates the complexity

of the model.

Input Model Acc. balanced Acc. #epochs #params

VGG11 0.911 ±0.026 0.790 ±0.005 26.6 ±0.800 3.47 ·10

VGG13 0.906 ±0.031 0.784 ±0.004 22.8 ±4.354 3.53 ·10

VGG16 0.893 ±0.023 0.780 ±0.009 23.4 ±2.871 5.30 ·10

VGG19 0.911 ±0.037 0.783 ±0.005 23.4 ±5.426 7.07 ·10

SPEC

VGG11 0.981 ±0.012 0.900 ±0.005 20.4 ±3.072 9.35 ·10

VGG13 0.981 ±0.008 0.897 ±0.004 16.2 ±2.135 9.54 ·10

VGG16 0.988 ±0.002 0.899 ±0.005 17.8 ±1.720 14.85 ·10

VGG19 0.989 ±0.005 0.898 ±0.003 17.4 ±0.800 20.16 ·10

Table 4: Mean ± standard deviation of the accuracy obtained in 5-fold cross-validation of the different models in the drone

classiﬁcation task for different types of input data, i.e. spectrogram (SPEC) and IQ, at different noise levels.

SNR 0 dB −6 dB −12 dB −18 dB

Input Model

VGG11 0.948 ±0.006 0.851 ±0.017 0.683 ±0.017 0.526 ±0.010

VGG13 0.950 ±0.006 0.856 ±0.011 0.644 ±0.027 0.517 ±0.009

VGG16 0.954 ±0.010 0.843 ±0.015 0.634 ±0.025 0.521 ±0.024

VGG19 0.952 ±0.005 0.845 ±0.009 0.642 ±0.018 0.525 ±0.012

SPEC

VGG11 0.987 ±0.003 0.960 ±0.007 0.911 ±0.010 0.732 ±0.006

VGG13 0.984 ±0.005 0.964 ±0.010 0.902 ±0.016 0.733 ±0.017

VGG16 0.985 ±0.006 0.967 ±0.010 0.910 ±0.006 0.724 ±0.011

VGG19 0.985 ±0.002 0.968 ±0.006 0.902 ±0.016 0.723 ±0.028

Table 5: Mean ± standard deviation of the balanced accuracy obtained in the 5-fold cross-validation of the different models

in the drone classiﬁcation task for different types of input data, i.e. spectrogram (SPEC) and IQ, at different noise levels.

SNR 0 dB −6 dB −12 dB −18 dB

Input Model

VGG11 0.956 ±0.006 0.731 ±0.028 0.413 ±0.012 0.225 ±0.015

VGG13 0.961 ±0.012 0.737 ±0.041 0.350 ±0.025 0.211 ±0.009

VGG16 0.960 ±0.006 0.706 ±0.044 0.344 ±0.043 0.201 ±0.019

VGG19 0.954 ±0.009 0.713 ±0.024 0.350 ±0.023 0.217 ±0.016

SPEC

VGG11 0.984 ±0.009 0.937 ±0.006 0.842 ±0.035 0.444 ±0.027

VGG13 0.985 ±0.005 0.951 ±0.012 0.819 ±0.036 0.454 ±0.001

VGG16 0.984 ±0.007 0.950 ±0.014 0.829 ±0.020 0.441 ±0.032

VGG19 0.983 ±0.006 0.951 ±0.006 0.812 ±0.021 0.431 ±0.023

architecture search approaches (Chen et al., 2018).

Furthermore, we have seen that the confusion at

low SNRs mainly occurs between the noise class

and the drones, and not between the different drones

themselves (see Figure 6). This is particularly rele-

vant for the application of drone detection systems in

security sensitive areas. The ﬁrst priority is to detect

any kind of UAV, regardless of its type. Since it seems

to be comparatively easy to learn the class separation

at high SNRs, one can shift the focus during learning

by redistributing the samples towards lower SNRs in-

stead of the equal distribution we used.

Optimizing the data collection and preprocessing

itself is beyond the scope of this work. The hardware

setup was chosen for the development of a rather sim-

ple and low-budget drone detection system (consumer

grade notebook with GPU + SDR). Several parame-

ters, such as sampling frequency, length of input vec-

tors, etc., were set to allow real-time detection with

a limited amount of memory and computing power.

That is, data acquisition, preprocessing, and model

inference should not take signiﬁcantly longer than the

signal being processed (≈ 1.2 ms per sample in our

case).

NCTA 2023 - 15th International Conference on Neural Computation Theory and Applications

502

Figure 6: Confusion matrix of a run of the VGG11 model

with spectrogram data at the input (VGG11 SPEC) at SNR

−20 dB. The average accuracy is 0.63 and the average bal-

anced accuracy is 0.28.

REFERENCES

Al-Sa’d, M. F., Al-Ali, A., Mohamed, A., Khattab, T., and

Erbad, A. (2019). Rf-based drone detection and iden-

tiﬁcation using deep learning approaches: An initia-

tive towards a large open source drone database. Fu-

ture Generation Computer Systems, 100:86–97.

Allahham, M. S., Al-Sa’d, M. F., Al-Ali, A., Mohamed, A.,

Khattab, T., and Erbad, A. (2019). Dronerf dataset: A

dataset of drones for rf-based detection, classiﬁcation

and identiﬁcation. Data in Brief, 26:104313.

Bottou, L. (1999). On-Line Learning and Stochastic Ap-

proximations, pages 9–42. Cambridge University

Press, USA.

Chen, L.-C., Collins, M. D., Zhu, Y., Papandreou, G.,

Zoph, B., Schroff, F., Adam, H., and Shlens, J.

(2018). Searching for efﬁcient multi-scale architec-

tures for dense image prediction. In Proceedings of the

32nd International Conference on Neural Information

Processing Systems, NIPS’18, page 8713–8724, Red

Hook, NY, USA. Curran Associates Inc.

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-

Fei, L. (2009). Imagenet: A large-scale hierarchical

image database. In 2009 IEEE conference on com-

puter vision and pattern recognition, pages 248–255.

Ieee.

Ge, C., Yang, S., Sun, W., Luo, Y., and Luo, C. (2021).

For rf signal-based uav states recognition, is pre-

processing still important at the era of deep learning?

In 2021 7th International Conference on Computer

and Communications (ICCC), pages 2292–2296.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012).

Imagenet classiﬁcation with deep convolutional neu-

ral networks. In Pereira, F., Burges, C., Bottou, L.,

and Weinberger, K., editors, Advances in Neural In-

formation Processing Systems, volume 25. Curran As-

sociates, Inc.

Kunze, S. and Saha, B. (2022). Drone classiﬁcation with a

convolutional neural network applied to raw iq data.

Institute of Electrical and Electronics Engineers Inc.

Long, J., Shelhamer, E., and Darrell, T. (2015). Fully con-

volutional networks for semantic segmentation. In

2015 IEEE Conference on Computer Vision and Pat-

tern Recognition (CVPR), pages 3431–3440.

Medaiyese, O. O., Ezuma, M., Lauf, A. P., and Guvenc,

I. (2021). Semi-supervised learning framework for

uav detection. volume 2021-September, pages 1185–

1190. Institute of Electrical and Electronics Engineers

Inc.

Parkhi, O. M., Vedaldi, A., and Zisserman, A. (2015). Deep

face recognition. In Xianghua Xie, Mark W. Jones, G.

K. L. T., editor, Proceedings of the British Machine Vi-

sion Conference (BMVC), pages 41.1–41.12. BMVA

Press.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., Desmaison, A., Kopf, A., Yang, E., De-

Vito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,

Steiner, B., Fang, L., Bai, J., and Chintala, S. (2019).

Pytorch: An imperative style, high-performance deep

learning library. In Wallach, H., Larochelle, H.,

Beygelzimer, A., d'Alch

e-Buc, F., Fox, E., and Gar-

nett, R., editors, Advances in Neural Information Pro-

cessing Systems 32, pages 8024–8035. Curran Asso-

ciates, Inc.

Simonyan, K. and Zisserman, A. (2015). Very deep con-

volutional networks for large-scale image recognition.

In Bengio, Y. and LeCun, Y., editors, 3rd Interna-

tional Conference on Learning Representations, ICLR

2015, San Diego, CA, USA, May 7-9, 2015, Confer-

ence Track Proceedings.

Swinney, C. J. and Woods, J. C. (2020). Unmanned

aerial vehicle ﬂight mode classiﬁcation using convo-

lutional neural network and transfer learning. In 2020

16th International Computer Engineering Conference

(ICENCO), pages 83–87.

Swinney, C. J. and Woods, J. C. (2021). Rf detection and

classiﬁcation of unmanned aerial vehicles in environ-

ments with wireless interference. pages 1494–1498.

Institute of Electrical and Electronics Engineers Inc.

Taha, B. and Shoufan, A. (2019). Machine learning-based

drone detection and classiﬁcation: State-of-the-art in

research. IEEE Access, 7:138669–138682.

Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M.,

Reddy, T., Cournapeau, D., Burovski, E., Peterson, P.,

Weckesser, W., Bright, J., van der Walt, S. J., Brett,

M., Wilson, J., Millman, K. J., Mayorov, N., Nel-

son, A. R. J., Jones, E., Kern, R., Larson, E., Carey,

C. J., Polat,

I., Feng, Y., Moore, E. W., VanderPlas,

J., Laxalde, D., Perktold, J., Cimrman, R., Henriksen,

I., Quintero, E. A., Harris, C. R., Archibald, A. M.,

Ribeiro, A. H., Pedregosa, F., van Mulbregt, P., and

SciPy 1.0 Contributors (2020). SciPy 1.0: Fundamen-

tal Algorithms for Scientiﬁc Computing in Python.

Nature Methods, 17:261–272.

Zeiler, M. D. and Fergus, R. (2014). Visualizing and under-

standing convolutional networks. In Fleet, D., Pajdla,

T., Schiele, B., and Tuytelaars, T., editors, Computer

Vision – ECCV 2014, pages 818–833, Cham. Springer

International Publishing.

Robust Drone Detection and Classiﬁcation from Radio Frequency Signals Using Convolutional Neural Networks

503

Zhang, H., Li, T., Li, Y., Li, J., Dobre, O. A., and Wen,

Z. (2023). Rf-based drone classiﬁcation under com-

plex electromagnetic environments using deep learn-

ing. IEEE Sensors Journal, 23(6):6099–6108.

Zhang, Y. (2021). Rf-based drone detection using machine

learning. In 2021 2nd International Conference on

Computing and Data Science (CDS), pages 425–428.

NCTA 2023 - 15th International Conference on Neural Computation Theory and Applications

504