Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

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

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Deep Learning based Drone Classification


using Radar Cross Section Signatures at
mmWave Frequencies
Rui Fu1, Mohammed Abdulhakim Alabsi2(Senior Member IEEE), Ki-Hwan Kim3, Young-sil
Lee3, Ahmed Abdulhakim Alabsi4, Hoon Jae Lee5
1
Blockchain Laboratory of Agricultural, Vegetables,Weifang University of Science and Technology,Weifang,Shandong,262700, China
2
Department of Computer Engineering, Graduate School, Dongseo University, 47 Juryero, Sasanggu, Busan, 47011, Korea
3
International College, Dongseo University, Busan 47011, Korea
4
Department of Smart Computing, Kyungdong University, 46 4gil, Gangwondo, 24764, Bongpo, Gosung, Korea
3
Division of Information and Communication Engineering, Dongseo University, 47 Juryero, Sasanggu, Busan, 47011, Korea

Corresponding author: A.A.A, H.J.L (e-mail: absiahmed@kduniv.ac.kr, hjlee@dongseo.ac.kr).


This research was a part of the project titled 'Marine digital AtoN information management and service system development(1/5) (20210650)', funded by the
Ministry of Oceans and Fisheries, Korea..”

ABSTRACT This paper presents drone classification at millimeter-wave (mmWave) radars using the deep
learning (DL) technique. The adoption mmWave technology in radar systems enables better resolution and
aid in detecting smaller drones. Using radar cross-section (RCS) signature enables us to detect malicious
drones and suitable action can be taken by respective authorities. Existing drone classification converts the
RCS signature into images and then performs drone classification using a convolution neural network (CNN).
Converting every signature into an image induces additional computation overhead; further CNN model is
trained considering fixed learning rate. Thus, when using CNN-based drone classification under a highly
dynamic environment exhibit poor classification accuracy. This paper present an im- proved long short-term
memory (LSTM) by introducing a weight optimization model that can reduce computation overhead by not
allowing the gradient to not flow through hidden states of the LSTM model. Further, present adaptive learning
rate optimizing (ALRO) model for training the LSTM model. Experiment outcome shows LSTM-ALRO
achieves much better drone detection accuracies when compared with the existing CNN-based drone
classification model.

INDEX TERMS Convolutional neural network, drone detection, micro doppler signature (MDS),
millimeter-wave, radar cross-section, unmanned aerial vehicle.

I. INTRODUCTION for the prevention of natural disasters [1], search and rescue
Every new technology invented brings about positive and operations by gaining information [2] and providing essential
negative impacts to modern civilization. Recently, unmanned materials, border surveillance, and providing security to users
aerial vehicles (UAVs) such as drones have seen a constant in the city [3], [4], [20].
rise in adoption in a different domain; this is because the drone In the last few years, there have been numerous episodes when
has the capability of providing real-time video streaming and drone operators violated regulations and have used drones in
image capturing features. The drone has reached a capability restricted areas [5], [20]. Such incidents can lead to
of taking autonomous decisions based on different states; thus, catastrophic economic consequences and, more importantly,
it has become the center of attraction employing artificial can present a real danger to people's lives. Nowadays, the
intelligence techniques in provisioning different applications drone regulation topic is under development and, possibly,
to military and civilians. Further, with technological growth, common rules will be applied for all countries in the EU in the
drones can be operated using smartphones rather than a near future. However, there might be custom-built
traditional remote controller. The drones have been adopted unregistered drones flying in restricted areas. Undoubtedly, it

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

is more difficult to detect drones due to their small dimensions Following a similar framework to the measurements above,
relative to other moving objects, for example, cars. Typical [13] fill the gap and present comprehensive 3D measurements
maximum radar cross-section (RCS) values for consumer cars of RCS characteristics of nine UAVs as well as the RCS of the
vary around average values of 18 dBsm at 23-27 GHz and 25 Li-Po batteries, since the batteries are the immense reflective
dBsm at 76-81 GHz [6]. On the other hand, typical RCS values object in the drone configuration. In [13], studied quasi-3D
for drones are from -15 to -20 dBsm in the X-band [7] and RCS signatures of a diverse range of drone models over the
smaller than -20 dBsm at 30-37 GHz [8]. frequency band of 26-40 GHz, which may provide essential
In this context, a new challenge is to enrich such small flying material for the drone database [18], [19]. Using this
robots with radar capabilities for exploring unknown measurement this work aim at developing an automatic
environments and for tracking targets with higher accuracy. detection drone using deep learning technique [25], [26]. Here
An additional requirement is to accomplish all this by reducing we model the LSTM model to perform drone classification as
the time needed to complete the mission [9]. Nowadays, such it is very efficient in addressing gradient descent problems
operations are mainly performed by fusing all the information with respect. However, different paths induce the magnitude
coming from onboard sensors as, for example, vision-based of gradient imbalance because the computational graph is
(camera or video) or inertial sensors [10]. Nevertheless, such polynomial in nature with respect to time. As a result, the
technologies fail in scarce visibility conditions or in harsh gradients from the linear path are affected for higher weight
propagation environments, like indoors. A possible solution to matrices. Here we present an improved weight optimization
overcome such shortcomings is to embed UAVs with model that can reduce computation overhead by not allowing
millimeter-wave (mm-wave) radars, which due to the reduced the gradient to not flow through hidden states of the LSTM
wavelength, can be significantly miniaturized and integrated model. Existing deep learning-based drone classification
onboard, and can improve the detection and tracking models [19], [25] are learned using Adam optimization model
capabilities of targets [11]. with fixed learning rate; as a result, impacts drone
The use of mmWave frequencies [21] in the radar systems has classification accuracies when introduced in highly dynamic
some advantages, e.g., better resolution, and makes it possible and uncertain environments [26]. Thus, for improving drone
to detect smaller objects. Based on the measured RCS classification accuracies in this paper we present the adaptive
signatures the flying object can be identified, and necessary learning rate optimization (ALRO) model.
actions can be undertaken by the authorities. Cellular The significance of drone classification using the LSTM-
operators have already begun deploying the new infrastructure ALRO model is described below. The LSTM-ALRO based
required for 5G NR [22], and soon 5G mobile edge computing drone classification model reduces computation overhead for
(MEC) [23] servers operating at mmWave frequencies will be drone classification by introducing a weight optimization
densely deployed in cities [24]. Adopting this infrastructure model. The modeling of the adaptive learning rate
for drone detection would result in great cost savings since the optimization model further aid in achieving better drone
work of provisioning and deploying the equipment would detection accuracies considering both smaller and larger
already be handled by cellular operators. Recent work has drones. The LSTM-ALRO model achieves better accuracies,
shown that these base stations could operate as a radar with precision, recall, and F1-measure outcomes for performing
small modifications [12]. drone classification for both small and large drones.
In [13] provided extensive measurement results of diverse The paper is organized as follows. In section II, Deep learning-
drone models at one possible 5G frequency band along with a based Drone classification at mmWave frequency. In section
study of the particular drone's frame materials. In addition, we III, simulation result and analysis is discussed. Finally, the
estimate the effect of different drone parts on the total RCS significance of research work is discussed and also presents
values through simulation results. A large number of the the future direction of enhancing the drone classification
research papers (e.g. [8], [14], [15]), in which authors study model.
the RCS signatures of drones at different frequencies, show
the relevance and impact of this topic. For example, in [14], II. DEEP LEARNING-BASED DRONE CLASSIFICATION
the authors present measurement results of the DJI Phantom AT MMWAVE FREQUENCY
drone using a 94 GHz radar system. In [16], the RCS of two This section presents deep learning-based drone classification
consumer drone models is presented in the 5.8 to 8.2 GHz at mmWave frequency. First, the system model used for
frequency range. Another research work [8], provides performing drone classification is presented. Second, present
wideband RCS measurements of non-metallic objects at 30- the deep learning models for classifying drones. Finally, an
37 GHz. RCS results obtained using a Ku band radar are adaptive learning model for enhancing drone classification
presented in [17]-[19]. However, in the previous research accuracies is performed.
activities, only a handful of drone models (with a limited range
of materials) were studied, and many frequencies were not A. SYSTEM MODEL
covered. In addition, there is a lack of RCS covering full or Here different kind of drone moves through a monitoring area
quasi-3D space. and drone classification is done using RCS measurement

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

collected using mmWave. The RCS is measured considering The matrix 𝑃𝑃𝑢𝑢 defines a function of long-short term model
the different frequencies and these measurement values are gates, hidden and cell states; this is because 𝑃𝑃𝑢𝑢 is composed of
trained using LSTM-ALRO (Long short-term memory- full gradient ascend due to temporal path mentioned in Eq. (1).
Adaptive learning rate optimization) model in order to carry A noticeable feature is the entire gates and hidden state 𝑖𝑖𝑢𝑢 are
out drone classification tasks. constraint with respect to outcome obtained through tan hot
sigmoid activation functions. However, on the other side, the
B. LSTM MODEL FOR DRONE CLASSIFICATION cell state 𝑑𝑑𝑢𝑢 progress elementwise with Eq. (1) and are
Here we present a new LSTM model that can efficiently constraint by session instance 𝑢𝑢; thus, with determinate 𝑑𝑑0 , the
address the gradient descent problems. The LSTM for values 𝑃𝑃𝑢𝑢 are constrained. Similarly, the matrix 𝑄𝑄𝑢𝑢 is
enhancing the structure of gradients with respect to time is composed of full gradients which ascend due to residual paths.
described using the following equation The element of the matrix 𝑄𝑄𝑢𝑢 are set of weight in a linear
function 𝑋𝑋ℎ𝑖𝑖 , 𝑋𝑋𝑔𝑔𝑔𝑔 , 𝑋𝑋𝑗𝑗𝑗𝑗 , 𝑋𝑋𝑝𝑝𝑝𝑝 . If the weights are set very large
𝑑𝑑𝑢𝑢 = 𝑔𝑔𝑢𝑢 ⨀𝑑𝑑𝑢𝑢−1 + 𝑗𝑗𝑢𝑢 ⨀ℎ𝑢𝑢 (1) element in 𝑄𝑄𝑢𝑢 can become extremely large and it becomes
even worse when we multiply 𝑄𝑄𝑢𝑢 in Eq. (12), because product
𝑖𝑖𝑢𝑢 = 𝑝𝑝𝑢𝑢 ⨀𝑡𝑡𝑡𝑡𝑡𝑡ℎ(𝑑𝑑𝑢𝑢 ) (2) operation will result in polynomial, and it is difficult to bound
them with respect to time. Thus, 𝑃𝑃𝑢𝑢 get inhibited in
where 𝑑𝑑𝑢𝑢 defines the cell state and 𝑖𝑖𝑢𝑢 represent the hidden comparison with 𝑄𝑄𝑢𝑢 because of the existence of gradient
state, ⨀ signify point-wise product and 𝑔𝑔𝑢𝑢 , 𝑗𝑗𝑢𝑢 , 𝑝𝑝𝑢𝑢 and ℎ𝑢𝑢 component magnitudes imbalance; thus, it is problematic of
represent gates. suppressing the gradient components of 𝑃𝑃𝑢𝑢 .
The gate ℎ𝑢𝑢 is obtained through the following equations In order to address the above problems here, we optimize the
ℎ𝑢𝑢 = tanh�𝑋𝑋ℎ𝑖𝑖 𝑖𝑖𝑢𝑢−1 + 𝑋𝑋ℎ𝑦𝑦 𝑦𝑦𝑢𝑢 + 𝑐𝑐ℎ � (3) gradient value in such a way that 𝑃𝑃𝑢𝑢 is not intimidated when
Then, the gate 𝑔𝑔𝑢𝑢 is defined using the following equation 𝑄𝑄𝑢𝑢 is extremely high. Thus, aid in multiplying 𝑄𝑄𝑢𝑢 between
𝑔𝑔𝑢𝑢 = 𝜎𝜎�𝑋𝑋𝑔𝑔𝑔𝑔 𝑖𝑖𝑢𝑢−1 + 𝑋𝑋𝑔𝑔𝑔𝑔 𝑦𝑦𝑢𝑢 + 𝑐𝑐𝑔𝑔 � (4) above zero to one to reduce the magnitude. The proposed
Similarly, the gate 𝑗𝑗𝑢𝑢 is obtained using the following equation weight optimization model is shown in algorithm 1. The
𝑗𝑗𝑢𝑢 = 𝜎𝜎�𝑋𝑋𝑗𝑗𝑗𝑗 𝑖𝑖𝑢𝑢−1 + 𝑋𝑋𝑗𝑗𝑗𝑗 𝑦𝑦𝑢𝑢 + 𝑐𝑐𝑗𝑗 � (5) proposed weight optimization model doesn’t allow the
Then, the gate 𝑝𝑝𝑢𝑢 is defined using the following equation gradient to flow through different 𝑖𝑖𝑢𝑢 states in an independent
𝑝𝑝𝑢𝑢 = 𝜎𝜎�𝑋𝑋𝑝𝑝𝑝𝑝 𝑖𝑖𝑢𝑢−1 + 𝑋𝑋𝑝𝑝𝑝𝑝 𝑦𝑦𝑢𝑢 + 𝑐𝑐𝑝𝑝 � (6) manner considering probability 1 − 𝒫𝒫 , where 𝒫𝒫 ∈ [0,1] as
Generally, an optimization 𝛼𝛼(𝑖𝑖𝑈𝑈 ) are utilized as the outcome hyper-parameter used for the optimization process.
at instance 𝑢𝑢 with respect to that the loss ℓ𝑢𝑢 is computed and
is mathematical defines as Algorithm 1. Weight optimization model
ℓ𝑢𝑢 ≔ ℓ�𝛼𝛼(𝑖𝑖𝑈𝑈 )� (7) Step 1. Start
A notable characteristic of the LSTM model is the use of Step 2. Input. {𝑦𝑦𝑢𝑢 }𝑈𝑈𝑢𝑢−1 , 𝑖𝑖0 , 𝑑𝑑0 , 𝒫𝒫
Step 3. ℓ = 0
recursive association among cell states 𝑑𝑑𝑢𝑢 that is linear in Step 4. ∀ 1 ≤ 𝑢𝑢 ≤ 𝑈𝑈 do
nature. This linear association aids the gradients to flow for the Step 5. If Bernoulli(𝒫𝒫) == 1 then
loner period. However, the weight matrices 𝑋𝑋ℎ𝑖𝑖 , 𝑋𝑋𝑔𝑔𝑔𝑔 , 𝑋𝑋𝑗𝑗𝑗𝑗 , 𝑋𝑋𝑝𝑝𝑝𝑝 Step 6. 𝚤𝚤̃𝑢𝑢−1 ←stop-gradient(𝑖𝑖𝑢𝑢−1 )
Step 7. Else
in LSTM model computational graph are polynomial in nature Step 8. 𝚤𝚤̃𝑢𝑢−1 ← 𝑖𝑖𝑢𝑢−1
which grows with respect to time; thus, from different paths Step 9. 𝑖𝑖𝑢𝑢 , 𝑑𝑑𝑢𝑢 ←LSTM(𝑦𝑦𝑢𝑢 , 𝚤𝚤̃𝑢𝑢−1 , 𝑑𝑑𝑢𝑢−1 )
induces magnitude of gradient imbalance, thus for higher Step 10. ℓ ← 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙�𝛼𝛼(𝑖𝑖𝑢𝑢 )�
weight matrices it affects the gradients from linear paths. As a Step 11. ℓ ← ℓ + ℓ𝑢𝑢
Step 12. Obtain ℓ
result, affecting drone classification performance. Step 13. Stop
In addressing the aforementioned problems here we introduce
the backpropagation mathematical operation of the LSTM
Now we show that gradient component obtained through 𝑄𝑄𝑢𝑢
model. First, let fix 𝑥𝑥 be an element of the weight matrix
get reduced using gradient loss function outcome from
𝑋𝑋ℎ𝑖𝑖 , 𝑋𝑋𝑔𝑔𝑔𝑔 , 𝑋𝑋𝑗𝑗𝑗𝑗 , 𝑋𝑋𝑝𝑝𝑝𝑝 , 𝑋𝑋ℎ𝑦𝑦 , 𝑋𝑋𝑔𝑔𝑔𝑔 , 𝑋𝑋𝑗𝑗𝑗𝑗 , 𝑋𝑋𝑝𝑝𝑝𝑝 . Define,
algorithm 1. 𝑈𝑈 𝑈𝑈
𝑃𝑃𝑢𝑢 = �𝐺𝐺�𝑢𝑢 0𝑜𝑜 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑�𝑙𝑙̃𝑢𝑢 �� (8) dd 𝑑𝑑𝑖𝑖 𝑈𝑈
Let 𝑎𝑎𝑢𝑢 = � 𝑢𝑢 ; 𝑢𝑢 ; 1𝑈𝑈 𝑜𝑜 � and 𝑎𝑎�𝑢𝑢 are equivalent with
d𝑥𝑥 d𝑥𝑥
respect to 𝑎𝑎𝑢𝑢 when weight optimization model with
𝑄𝑄𝑢𝑢 = �0𝑜𝑜 𝛽𝛽𝑜𝑜 0𝑜𝑜 � (9)
probability 1 − 𝒫𝒫 is applied during backpropagation; thus, the
⎛ ⎞ 𝑎𝑎�𝑢𝑢 becomes
𝑎𝑎𝑢𝑢 = ⎜ 𝑑𝑑𝑖𝑖𝑢𝑢 ⎟ (10)
𝑎𝑎�𝑢𝑢 = (𝑃𝑃𝑢𝑢 + 𝛽𝛽𝑢𝑢 𝑄𝑄𝑢𝑢 ) (13)
Then, 𝑎𝑎𝑢𝑢 can be defined as follows
+ (𝑃𝑃𝑢𝑢−1
𝑎𝑎𝑢𝑢 = (𝑃𝑃𝑢𝑢 ⎝
+ 𝑄𝑄𝑢𝑢 )𝑎𝑎
⎠ 𝑢𝑢−1 (11)
+ 𝛽𝛽𝑢𝑢−1 𝑄𝑄𝑢𝑢−1 ) … (𝑃𝑃2
By iterating the above equation, we can rewrite it as follows
+ 𝛽𝛽2 𝑄𝑄2 )𝑎𝑎�𝑢𝑢
𝑎𝑎𝑢𝑢 = (𝑃𝑃𝑢𝑢 + 𝑄𝑄𝑢𝑢 )(𝑃𝑃𝑢𝑢−1 + 𝑄𝑄𝑢𝑢−1 ) … (𝑃𝑃2 (12)
where 𝛽𝛽𝑢𝑢 , 𝛽𝛽𝑢𝑢−1 , … , 𝛽𝛽2 are identically and independently
+ 𝑄𝑄2 )𝑎𝑎1
distributed and other variables are similar to Eq. (12).

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

Eq. (13) doesn’t allow the gradient to pass-through 𝑖𝑖𝑢𝑢 states Step 5. While
Step 6. 𝛾𝛾𝑗𝑗+1 ← 𝛾𝛾0 𝒟𝒟�𝓌𝓌𝑗𝑗 �
of LSTM model with probabilities of 1 − 𝒫𝒫 in a stochastic
Step 7. 𝑗𝑗 ← 𝑗𝑗 + 1
manner as described below Step 8. End while
ℱ𝛽𝛽2 ,…,𝛽𝛽𝑢𝑢 [𝑎𝑎�𝑢𝑢 ] = (𝑃𝑃𝑢𝑢 + 𝒫𝒫𝑄𝑄𝑢𝑢 ) (14) Step 9. Obtain 𝛾𝛾𝑗𝑗
+ (𝑃𝑃𝑢𝑢−1 + 𝒫𝒫𝑄𝑄𝑢𝑢−1 ) … (𝑃𝑃2 Step 10. Stop
+ 𝒫𝒫𝑄𝑄2 )𝑎𝑎�1 The adaptive learning rate optimization model achieves
In this work the 𝑄𝑄𝑢𝑢 is dropped in a stochastic manner in the better convergence when employed with the Adam
gradient component. Thus, aid in reducing the 𝑄𝑄𝑢𝑢 in optimization model of Eq. (16) as described below
comparison with the standard LSTM model and also aid in 𝓂𝓂�𝑗𝑗 (18)
reducing computation overhead by not allowing the gradient 𝓌𝓌𝑗𝑗+1 = 𝓌𝓌𝑗𝑗 − 𝛾𝛾

to not flow through hidden states 𝑖𝑖𝑢𝑢 of LSMT model. Further, �𝓋𝓋
�𝑗𝑗 + 𝛿𝛿
the existing model has shown using Adam optimizer for batch-
𝑗𝑗
based learning achieve good learning efficiency; However, the 𝓂𝓂𝑗𝑗 ⁄�1 − 𝜇𝜇1 � (19)
existing model is trained using constant learning rate; thus, = 𝓌𝓌𝑗𝑗 − 𝛾𝛾 ,
when adopted to classifying drone achieves very poor results. �𝓋𝓋𝑗𝑗 ⁄�1 − 𝜇𝜇2𝑗𝑗 � + 𝛿𝛿 ′
For addresses in the next sub-section, we introduce an adaptive
learning rate optimization model for achieving better drone
classification accuracies. (20)
�1 − 𝜇𝜇2𝑗𝑗 𝓂𝓂𝑗𝑗
= 𝓌𝓌𝑗𝑗 − 𝛾𝛾 𝑗𝑗
.
C. ADAPTIVE LEARNING RATE OPTIMIZATION MODEL �1 − 𝜇𝜇1 � �𝓋𝓋𝑗𝑗 + 𝛿𝛿 ′
FOR DRONE CLASSIFICATION
This section introduces an adaptive learning rate optimization
model for the drone classification model. The Adam optimizer where 𝜇𝜇1 and 𝜇𝜇2 are some constant; if 𝑗𝑗 is significantly
model is a gradient-based method by combining both 𝑗𝑗 𝑗𝑗
large, then �1 − 𝜇𝜇2 ��1 − 𝜇𝜇1 � is considered to equal to
RMSprop and Adagrad which is widely used to train various
classification and regression models. The update rule of Adam one. Therefore, the adaptive learning rate optimization
methodology is obtained using the following equation model can be obtained as follows
𝓌𝓌𝑗𝑗 (21)
𝓌𝓌𝑗𝑗+1 = 𝓌𝓌𝑗𝑗 − 𝛾𝛾𝑗𝑗 ,
𝓂𝓂
� 𝑗𝑗 (15) �𝓋𝓋𝑗𝑗 + 𝛿𝛿 ′
𝓌𝓌𝑗𝑗+1 = 𝓌𝓌𝑗𝑗 − 𝛾𝛾
�𝓋𝓋
�𝑗𝑗 + 𝛿𝛿 ′ Where
𝛾𝛾𝑗𝑗 = 𝛾𝛾0 𝒟𝒟�𝓌𝓌𝑗𝑗 � (22)
Where where (𝛾𝛾0 is a constant),
𝜕𝜕𝒟𝒟�𝓌𝓌𝑗𝑗 � (16) 𝑗𝑗 (23)
𝓂𝓂𝑗𝑗 = 𝜇𝜇1 𝓂𝓂𝑗𝑗−1 + (1 − 𝜇𝜇1 ) , 𝜕𝜕𝒟𝒟�𝓌𝓌𝑗𝑗−𝑘𝑘+1 �
𝜕𝜕𝓌𝓌 𝓂𝓂𝑗𝑗 = (1 − 𝜇𝜇1 ) � 𝜇𝜇1𝑘𝑘−1 ,
𝜕𝜕𝓌𝓌
𝑘𝑘=1
2 (17)
𝜕𝜕𝒟𝒟�𝓌𝓌𝑗𝑗 � and
𝓋𝓋𝑗𝑗 = 𝜇𝜇2 𝓋𝓋𝑗𝑗−1 + (1 − 𝜇𝜇2 ) � � , 𝑗𝑗 2 (24)
𝜕𝜕𝓌𝓌 𝜕𝜕𝒟𝒟�𝓌𝓌𝑗𝑗−𝑘𝑘+1 �
𝓋𝓋𝑗𝑗 ⁄(1 − 𝜇𝜇2 ) � 𝜇𝜇2𝑘𝑘−1 � �
𝜕𝜕𝓌𝓌
𝓂𝓂𝑗𝑗 𝓋𝓋𝑗𝑗 𝑘𝑘=1
Where 𝓂𝓂
� 𝑗𝑗 = , and 𝓋𝓋
�𝑗𝑗 = . The Adam optimizer
𝑗𝑗
𝜇𝜇1 𝜇𝜇1
𝑗𝑗
In this work, the learning rate of constant 𝛾𝛾 is changed
model computes exponential moving average for every to function 𝛾𝛾𝑗𝑗 − 𝛾𝛾0 𝒟𝒟�𝓌𝓌𝑗𝑗 � to make sure the learning rate is
gradient. Similarly, computes the squares of the gradient; optimized in an adaptive manner. Using Eq. (21), the
then optimize the learning rate by taking ratios of both sequence �𝓌𝓌𝑗𝑗 � can be obtained and 𝓂𝓂𝑗𝑗 and 𝓋𝓋𝑗𝑗 are
gradients. The standard Adam optimizer model generally obtained using Eq. (23) and (24), respectively. The �𝓌𝓌𝑗𝑗 �
decreases the learning rate or keeps it unchanged and can’t
converges and satisfies the minimal value of optimization
be neglected when it reaches local minima. Thus, an
function; this, aid in achieving better drone detection
adaptive learning rate mechanism using optimization
classification accuracies.
functions is introduced which is described in Algorithm 2.
III. SIMULATION RESULT AND ANALYSIS
Algorithm 2. Adaptive learning rate optimization model
Here experiment is conducted for performing automatic
Step 1. Start
Step 2. 𝛾𝛾0 Configure learning rate drone classifying using the proposed LSTM-ALRO model.
Step 3. 𝒟𝒟(𝓌𝓌). Optimization function with parameter 𝓌𝓌 Here receiver operating curve (ROC) performance metric
Step 4. 𝑗𝑗 ← 0(Initialize instance) such as accuracy, precision, recall, and F1-Score is used for

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

validating LSTM-ALRO based drone classification model. with frequency ranging from 26 GHz-40 GHz of Helicopter
Here the resulting outcome obtained is compared with Kyosho drone is shown in Fig. 3. Here the RCS measurement
existing GoogLeNet and CNN-based drone classification of different drones for the frequency range of 28 GHz and 38
models [1] in terms of accuracies. GHz is shown in Fig. 4. Based on the measurement the drone
is differentiated into two groups, smaller drones are composed
A. MEASUREMENT MODEL of 6 drones which are putted into group I, and the remaining 4
This work uses RCS measurement data collected from [13]. drones are putted into group II.
Here they constructed an anechoic chamber for measuring
antenna patterns that guarantee better RCS measurement.
The graphical representation of the measurement model is
shown in Fig. 1. The transmitter 𝑇𝑇𝑇𝑇 and receiver 𝑅𝑅𝑅𝑅 of
dual-polarized Vivaldi horn antenna with quasi-monostatic
radar are placed in one end of the shielded anechoic
chamber. A VNA is utilized for the generation and
recording of the received signal. The frequency is set in the
range of 26 − 40 GHz with a bandwidth capacity of 1
kHz, transmit power is set to 20 𝑑𝑑𝑑𝑑𝑑𝑑, and the distance 𝑅𝑅
among the drone to be tested (DTBT) and test antennas are
5.8𝑚𝑚. Herewith step size of 1 the frequency is varied, and
measurement is collected. The drone to be tested (DTBT)
is placed on top of a rotating pillar opposite to the chamber
than antennas. Tow axes of rotation are done to DTBT
using stepped motors. Initially, the drone is rotated in the
azimuth plane (i.e., x-axis). Then, it is moved toward the
center axis (i.e., y-axis). Angular positions are varied from
𝜃𝜃 ∈ [−90° , 90° ] and 𝜃𝜃 ∈ [0° , 180° ] with steps size of 1° .
For each drone, the bottom hemisphere was measured
aiding and reducing measurement time and provide a
realistic drone detection mechanism where radar could be
placed in the ground.

FIGURE 2. Photographs of the empirically studied objects (i.e.,


multirotor platforms, helicopter, battery) [13].

FIGURE 1. Schematic representation of measurement model [13].

B. DRONE USED FOR PERFORMING AUTOMATIC


DRONE CLASSIFICATION
Here we used measurements obtained from one helicopter,
one radio-controlled (RC), and eight multi-rotor drones. The
images of drones used are shown in Fig. 2 and the
corresponding drone size and material used for designing
drones are described in Table I. The sample RCS measurement

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

FIGURE 3. Sample RCS measurement of Helicopter Kyosho drone.

TABLE 1: TYPES OF DRONES

FIGURE 5. Monostatic RCS measurement of drones from Group II,


considering azimuth and elevation angles. The Left and center columns
indicate RCS measurement at 28 GHz and 38 GHz, respectively of the
corresponding drone (right column).

Case 1. Here experiment is conducted to perform drone


classification using a small drone (i.e., group 1 drones). Here
experiment is conducted to classify each drone with respect to
other drones; thus, it is a binary classification problem.
Receiver operating curve and ROC performance metrics such
as accuracies, precision, recall, and F1-score are used for
analyzing the drone classification model using LSTM-ALRO.
The ROC curve outcome for detecting different drones is
shown in Fig. 6. The ROC performance metric at 26 GHz is
shown in Table II and the ROC performance metric at 40 GHz
is shown in Table III. Comparative analysis considering 26
GHz and 40 GHz is shown in Fig. 7. From the result obtained
we can see the LSTM-ALRO based drone classification
achieve very good detection performance for both smaller and
larger frequency level. However, high detection accuracies are
achieved larger frequency size.

FIGURE 4. Monostatic RCS measurement of drones from Group I,


considering azimuth and elevation angles. The Left and center columns
indicate RCS measurement at 28 GHz and 38 GHz, respectively of the
corresponding drone (right column).

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

(a) F450 (d) P4P

(e) Parrot
(b) Heli FIGURE: 6 ROC outcomes for performing Group I drone classification
using LSTM-ALRO model corresponds to measurements at 28 GHz.

TABLE.2 GROUP I DRONE CLASSIFICATION ROC ANALYSIS CORRESPONDS


TO MEASUREMENT AT 28 GHZ
Type of Accuracy Precision Recall F1-Score
drones
F450 0.972375691 1 0.834254144 0.909638554
Heli 0.968692449 0.845070423 0.994475138 0.913705584
Mavic 0.998158379 0.989071038 1 0.994505495
P4P 0.990791897 0.980874317 0.991712707 0.986263736
Parrot 0.992633517 1 0.955801105 0.97740113
Average 0.984530387 0.963003156 0.955248619 0.9563029

TABLE.3 GROUP, I DRONE CLASSIFICATION ROC ANALYSIS CORRESPONDS


TO MEASUREMENT AT 40 GHZ
Type of drones Accuracy Precision Recall F1-Score
F450 1 1 1 1
Heli 1 1 1 1
(c) Mavic Mavic 1 1 1 1
P4P 1 1 1 1
Parrot 1 1 1 1
Average 1 1 1 1

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

Group II Drone Classification ROC Group II Drone Clasification ROC


Analysis Analysis
1.01 1

1 0.999

0.99
0.998
0.98
0.997

(%)
0.97
(%)

0.996
0.96
0.995
0.95
0.994
0.94

0.93 0.993
Accuracy Accuracy Precision Precision Recall (26 Recall (40 F1-Score F1-Score Accuracy Accuracy Precision Precision Recall (28 Recall (40 F1-Score F1-Score
(26 GHz) (40 GHz) (26 GHz) (40 GHz) GHz) GHz) (26 GHz) (40 GHz) (28 GHz) (40 GHz) (28 GHz) (40 GHz) GHz) GHz) (28) (40 GHz)
% 0.9845304 1 0.963 1 0.9552 1 0.9563 1 % 0.9980663 0.9991713 0.9959053 0.998274 0.9951657 0.9979282 0.9955094 0.9981005

FIGURE 8. ROC analysis considering 26 GHz and 40 GHz for Group II


drone classification.
FIGURE 7. ROC analysis considering 26 GHz and 40 GHz for Group I
drone classification.
Case 3. Here experiment is conducted to perform drone
Case 2. Here experiment is conducted to perform drone classification using both smaller and larger drones (i.e., both
classification using a larger drone (i.e., group 2 drones). Here group 1 and group 2 drones). Here experiment is conducted to
experiment is conducted to classify each drone with respect to classify each drone with respect to other drones; thus, it is a
other drones; thus, it is a binary classification problem. binary classification problem. Receiver operating curve
Receiver operating curve performance metrics such as performance metrics such as accuracies, precision, recall, and
accuracies, precision, recall, and F1-score are used for F1-score are used for analyzing the drone classification model
analyzing the drone classification model using LSTM- ALRO. using LSTM- ALRO. The ROC performance metric at 26
The ROC performance metric at 26 GHz is shown in Table IV GHz is shown in Table VI and the ROC performance metric
and the ROC performance metric at 40 GHz is shown in Table at 40 GHz is shown in Table VII. Comparative analysis
V. Comparative analysis considering 26 GHz and 40 GHz is considering 26 GHz and 40 GHz is shown in Fig. 8. From the
shown in Fig. 8. From the result obtained we can see the result obtained we can see the LSTM-ALRO based drone
LSTM-ALRO based drone classification achieve very good classification achieve very good detection performance for
detection performance for both smaller and larger frequency both smaller and larger frequency level.
level. However, high detection accuracies are achieved larger
frequency size. TABLE:6 GROUP II DRONE CLASSIFICATION ROC ANALYSIS CORRESPONDS
TO MEASUREMENT AT 26 GHZ
Type of drones Accuracy Precision Recall F1-Score
TABLE.4 GROUP II DRONE CLASSIFICATION ROC ANALYSIS CORRESPONDS
TO MEASUREMENT AT 28 GHZ F450 1 1 1 1
Type of Accuracy Precision Recall F1-Score
drones Heli 1 1 1 1
Hexa 0.998895028 1 0.994475138 0.997229917 Mavic 1 1 1 1
M100 0.998895028 0.997245179 1 0.99862069
Walkera 0.997237569 0.986376022 1 0.993141289 P4P 1 1 1 1
Y600 0.997237569 1 0.986187845 0.993045897
Average 0.998066298 0.9959053 0.995165746 0.995509448 Parrot 1 1 1 1
Hexa 1 1 1 1
TABLE.5 GROUP II DRONE CLASSIFICATION ROC ANALYSIS CORRESPONDS
M100 1 1 1 1
TO MEASUREMENT AT 40 GHZ
Type of Accuracy Precision Recall F1-Score Walkera 1 1 1 1
drones
Hexa 0.999447514 1 0.997237569 0.998616874 Y600 1 1 1 1
M100 0.999447514 0.99862069 1 0.999309869 Average 1 1 1 1
Walkera 0.998895028 0.997237569 0.997237569 0.997237569
Y600 0.998895028 0.997237569 0.997237569 0.997237569
TABLE: 7 GROUP II DRONE CLASSIFICATION ROC ANALYSIS CORRESPONDS
Average 0.999171271 0.998273957 0.997928177 0.99810047
TO MEASUREMENT AT 28 GHZ
Type of drones Accuracy Precision Recall F1-Score
F450 0.999654696 1 0.994475138 0.997229917
Heli 0.999654696 0.994505495 1 0.997245179
Mavic 1 1 1 1
P4P 1 1 1 1

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

Parrot 1 1 1 1 Further, enhance the LSTM-ALRO model to improve drone


Hexa 0.999654696 1 0.997237569 0.998616874 classification accuracies.
M100 0.999654696 0.99862069 1 0.999309869
Walkera 0.999654696 0.997245179 1 0.99862069 ACKNOWLEDGMENT
R.F., M.A.A.-A. and A.A.A.-A. contributed to the main idea
Y600 0.999654696 1 0.997237569 0.998616874
and the methodology of the research. R.F., M.A.A.-A.
Average 0.999769797 0.998930151 0.998772253 0.998848823
designed the experiment, performed the simulations, and
wrote the original manuscript. R.F., A.A.A.-A., K.H.K., and
Y.S.L. contributed significantly to improving the technical
Drone Classification ROC Analysis and grammatical contents of the manuscript. R.F., A.A.A.-A.,
1.0002
and H.J.L. reviewed the manuscript and provided valuable
1 suggestions to further refine it. All authors have read and
0.9998
agreed to the published version of the manuscript. Supervision
0.9996
0.9994 H.J.L.
0.9992
(%)

0.999
REFERENCES
0.9988
[1] L. Marconi, The SHERPA project: Smart collaboration between
0.9986
0.9984
humans and ground-aerial robots for improving rescuing activities
0.9982
in alpine environments, Proc. IEEE Int. Symp. Safety, Security,
0.998 Rescue Robot (2012) 1–4.
Accurac Accurac Precision Precision Recall Recall F1-Score F1-Score
y (26 y (40 (26 (40 (26 (40 (26 (40 [2] A. Merwaday, Improved throughput coverage in natural disasters:
GHz) GHz) GHz) GHz) GHz) GHz) GHz) GHz) Unmanned aerial base stations for public-safety communications,
Series1 1 0.9997698 1 0.9989302 1 0.9987723 1 0.9988488
IEEE Veh. Technol. Mag 11 (4) (2016) 53–60.
FIGURE 9. ROC analysis considering 26 GHz and 40 GHz for drone [3] C. Wang, Autonomous navigation of UAVs in large-scale complex
classification. environments: A deep reinforcement learning approach, IEEE
Trans. Veh. Technol 68 (3) (2019) 2124–2136.
IV. CONCLUSION [4] A. Guerra, F. Guidi, D. Dardari, P. M. Djurić, Reinforcement
Learning for UAV Autonomous Naviga- tion, Mapping and Target
In [13] collected RCS measurement considering different
Detection, 2020 IEEE/ION Position, Location and Navigation
measurement angles through mmWave frequency. Similar to Symposium (PLANS) (2020) 1004–1013.
[13] in [19] collected signature through polarimetry Ku-band [5] L. Schroth (2019). [link]. URL https://www.droneii.com/the-
frequency modulated continuous wave (FMCW) radar drone-market-2019-2024-5-things-you-need-to-know
[6] K. Geary, J. S. Colburn, A. Bekaryan, S. Zeng, B. Litkouhi, M.
system. They converted the entire signal information into the
Murad, Automotive radar target characterization from 22 to 29
image and then these images were trained using a convolution GHz and 76 to 81 GHz, Proc. IEEE Radar Conf. (RadarCon)
neural network and then the classification of the drone is (2013) 1–6.
performed. However, these models induce huge computation [7] A. D. D. Quevedo, F. I. Urzaiz, J. G. Menoyo, A. A. Lopez, Drone
detection and RCS measurements with ubiquitous radar, Proc.
overhead as it collects a large amount of RCS data and that
Int. Conf. Radar (RADAR) (2018) 1–6.
convert into an image and then performs classification using [8] M. Röding, G. Sommerkorn, S. Häfner, R. Müller, R. S. Thomä,
GoogLeNet and CNN deep learning models. Using J. Goerlich, K. Garhammer, Fully polarimetric wideband RCS
GoogLeNet the drone classification accuracies is 89.967% and measurements for small drones, Proc. 11th Eur. Conf. Antennas
using CNN they achieved accuracies of 96.62%. The drone Propag. (EUCAP) (2017) 3926–3930.
[9] P. H. ugler, Radar taking off: New capabilities for UAVs, IEEE
exhibits dynamic behavior and operates in an uncertain Microw. Mag 19 (7) (2018) 43–53.
environment; thus, requires an adaptive learning model. [10] I. Guvenc, Detection, tracking, and interdiction for amateur drones,
However, existing models are trained considering fixed IEEE Commun. Mag 56 (4) (2018) 75–81.
learning rates, thus the result achieved is not satisfactory. In [11] P. H. ugler, M. Geiger, C. Waldschmidt, 77 GHz radar-based
altimeter for unmanned aerial vehicles, Proc. IEEE Radio
this paper, we presented an improved Deep learning model Wireless Symp (2018) 129–132.
namely LSTM-ALRO which can adaptively work well under [12] D. Solomitckii, M. Gapeyenko, V. Semkin, S. Andreev, Y.
a highly uncertain and dynamic environment for performing Koucheryavy, Technologies for efficient amateur drone detection
varied classes of drone classification (i.e., including both small in 5G millimeter-wave cellular infrastructure, IEEE Commun.
Mag 56 (1) (2018) 43–50.
and large drones). The LSTM-ALRO based drone [13] V. Semkin, Analyzing Radar Cross Section Signatures of Diverse
classification model achieves an accuracy of 99.88% which is Drone Models at mmWave Frequencies, IEEE Access 8 (2020)
significantly higher than that of GoogLeNet and CNN –based 48958–48969.
drone classification models. The result proves the adaptive [14] M. Caris, W. Johannes, S. Sieger, V. Port, S. Stanko, Detection of
small UAS with W-band radar, Proc. 18th Int. Radar Symp
nature of the LSTM-ALRO based drone classification model (2017) 1–6.
with respect to dynamic characteristic of drone. [15] M. Pieraccini, L. Miccinesi, N. Rojhani, RCS measurements and
Future work would consider testing the LSTM-ALRO ISAR images of small UAVs, IEEE Aerosp. Electron. Syst. Mag
considering more diverse type of drone and also considering 32 (9) (2017) 28–32.
[16] A. Herschfelt, C. R. Birtcher, R. M. Gutierrez, Y. Rong, H. Yu, C.
multiple measurement considering more diverse angle. A. Balanis, D. W. Bliss, Consumer- grade drone radar

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3115805, IEEE Access

crosssection and micro-Doppler phenomenology, Proc. IEEE Ki Hwan Kim is an associate Professor in internati
Radar Conf. (Radar- Conf) (2017) 981–985.
onal College, Dongseo University, South Korea. He
[17] J. Ochodnicky, Z. Matousek, M. Babjak, J. Kurty (2017).
[18] B. & Kim, H.-S. & Kang, Park, Seong-Ook, Drone received his B.S., M.S., and PhD. degree in Compu
Classification Using Convolutional Neural Networks With Merged
ter Networking from Dongseo University, Republic
Doppler Images, IEEE Geoscience and Remote Sensing Letters. PP
(2016) 1–5. of Korea in 2020. He research interests are cryptog
[19] B. K. Kim, H. S. Kang, S. Lee, S. O. Park, Improved Drone
Classification Using Polarimetric Merged- Doppler Images, raphy, Information security and Side-Channel Attack(SCA).
IEEE Geoscience and Remote Sensing Letters (2021).
[20] Hildmann, E. Kovacs (2019).
[21] M. & Polese, Bertizzolo, & Lorenzo, L. & Bonati, Young Sil Lee is an associate Professor in internati
Gosain, T. Melodia (2020).
onal College, Dongseo University, South Korea. Sh
[22] A. Guerra, D. Dardari, P. M. Djurić, Dynamic Radar Network of
UAVs: A Joint Navigation and Tracking Approach, IEEE Access e received her B.S. degree and M.S. degree in Elect
8 (2020) 116454–116469.
[23] J. Zheng, T. Yang, H. Liu, T. Su, L. Wan, Accurate Detection and rical engineering, at Dongseo University, Busan, So
Localization of Unmanned Aerial Vehicle Swarms-Enabled uth Korea in 2006 and 2010 respectively. She recei
Mobile Edge Computing System, IEEE Transactions on Industrial
Informatics 17 (2021) 5059–5067. ved her Ph.D. degree in Ubiquitous IT from Dongse
[24] F. Guidi, D. Dardari, Radio Positioning with EM Processing of o University, Busan, South Korea in 2015. Her research area covers secu
the Spherical Wavefront, IEEE Trans- actions on Wireless
Communications. rity including healthcare system, RFID or WSN technologies. She works
[25] O. & Medaiyese, Ezuma, & Martins, A. & Lauf, for Department of Computer Engineering of Dongseo University as.
I. Guvenc (2021).
[26] A. Guerra, F. Guidi, D. Dardari, P. M. Djurić, Reinforcement
Learning for UAV Autonomous Naviga- tion, Mapping and Target
Ahmed A. Al-Absi is an associate Professor in S
Detection, 2020 IEEE/ION Position, Location and Navigation
Symposium (PLANS), 2020 1004–1013. mart Computing department, Kyungdong Univers
ity Global Campus, South Korea. Dr. Al-Absi rec
eived his Ph.D. degree in Computer Science (spec
Rui FU is a Professor in the institute of intelligent
ializing in Big Data Processing) from Dongseo U
manufacturing at Weifang University of Science an
niversity, Korea. He received his Master of Comp
d Technology, China. She received her (MS) degre
uter Science degree from University Utara Malaysia- Malaysia in 2011,
e in System Theory from Qingdao University -Chin
and a bachelor’s degree in computer applications from Bangalore Univer
a in 2012-2015. She earned her Ph.D. degree in the
sity- India in 2007. He has more than ten years of experience in teaching
Department of Information and Communication En
and university lecturing in the areas of database design and computer al
gineering at Dongseo University, Korea. Her research interests include
gorithms. In the field of research and publication, Dr. Al-Absi’s has publ
Artificial Intelligence, VANET, UAVs/Drone, Deep learning, Logistics
ished numbers of research papers in peer-reviewed international journals
Transportation and Mathematics.
and conferences. His research interests include Database Systems, Big
Data, Hadoop, Cloud computing, Distributed systems, Parallel computin
Mohammed Abdulhakim Alabsi received his BS g, High-performance computing, VANET, and bioinformatics.
degree in Computer Application from Bangalore U
niversity in India. He earned his (MS) degree at Do Hoon-Jae Lee received the B.S., M.S., and Ph.D.
ngseo University- South Korea in 2018. Currently, degree in Electrical Engineering from Kyungpook
he is a Ph.D. candidate in the Department of Inf-zo national university in 1985, 1987 and 1998, respe
rmation and Communication Engineering at Dongs ctively. He had been engaged in the research on c
eo University, Korea. His research interests include IoT, VANET, UAV, ryptography and network security at Agency for D
AI, Cryptology, Network Security, Computer Networks and Digital Co efense Development from 1987 to 1998. Since 2002 he has been workin
mmunications. g for Department of Computer Engineering of Dongseo University as an
associate professor, and now he is a full professor. His current research i
nterests are in security communication system, side-channel attack, USN
& RFID security. He is a member of the Korea institute of Information s
ecurity and cryptology, IEEE Computer Society, IEEE Information Theo
ry Society etc.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/

You might also like