Radar Jamming Recognition ; Models, Methods, And Prospects

This article has been accepted for publication in IEEE Journal of Selected Topics in Applied Earth Observations and
Remote Sensing. This is the author's version which has not been fu
content may change prior to final publication. Citation information: DOI 10.1109/JSTARS.2024.3522951
1
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
Radar Jamming Recognition:

Models, Methods, and Prospects
Zan Wang, Zhengwei Guo, Gaofeng Shu, Member, IEEE and Ning Li, Member, IEEE
Abstract—In the modern warfare with complex and changeable jamming in radar electronic counter measures (ECM)
electromagnetic environment, radar jamming is getting more [9][10][11], jamming effects are becoming more realistic,
complex and realistic, which poses a serious threat to radar, diverse, and complex, seriously affecting the reconnaissance
jamming recognition has become a hot topic in the field of capabilities of the radar. In complex electromagnetic
electronic countermeasures. To make effective anti-jamming environments, the recognition of radar jamming is crucial in
measures, numerous jamming recognition methods have been achieving timely battlefield situational awareness and
proposed. This paper presents a systematic review of jamming implementing effective countermeasures. Therefore, radar
recognition for this topic. Specifically, firstly building a system jamming recognition is one of the core steps in electronic
framework for jamming models, including deception jamming,
battlefield situational awareness, significantly impacting
suppression jamming, and smart jamming, thoroughly
subsequent decision-making.
explaining the operational mechanisms. Then, recognition
methods based on traditional machine learning are summarized, However, due to the widespread deployment of various
and delves into the advantages and disadvantages of feature radar systems, the electromagnetic environment has become
extraction methods and classifiers. Furthermore, the focus shifts exceedingly intricate [12]. The recognition of radar jamming
to neural network-based methods, such as shallow neural signals has encountered unprecedented challenges, primarily
network methods and deep neural network methods. In manifested in three aspects: 1) A diverse array of jamming
particular, restricted sample strategies are also discussed as signals is gradually emerging. The DRFM jamming system
potential future directions. Finally, concluding on the current can rapidly collect radar transmission signals [13], restores
status of jamming recognition methods and the prospects for them with remarkable fidelity, and modulates them for
future work are made. This paper provides a reference for the retransmission, thereby creates highly realistic false targets
research of radar jamming recognition. and expands the coverage of jamming. Such jamming methods
achieve deception or suppression effects simultaneously,
Index Terms—Radar jamming recognition, feature extraction, making it difficult for radar systems to effectively detect and
traditional machine learning, neural network, limited sample recognize real targets. This poses a severe threat to the normal
strategies. operation of radar systems. 2) An explosive growth in
jamming signal density has been witnessed recently. With the
I. INTRODUCTION expansion of radar applications and functionalities, the volume
W ITH the development of electronic warfare technology,

the competition between electronic jamming and
counter-jamming has become increasingly intense [1]. Radar
of jamming signals has increased dramatically [14].
Simultaneously, the emergence of intelligent devices has
contributed to a steep rise in pulse density within the
can achieve target detection, location, tracking and guidance electromagnetic environment. In the face of increasingly
of weapon systems, which is widely equipped on various scarce spectrum resources, the dense distribution of jamming
weapon systems and can substantially improve the combat signals is challenging to alleviate. 3) An increasing number of
effectiveness of traditional equipment [2][3][4]. The rapid pulse sequences are becoming highly interleaved. As
development of sensor technology and digital radio frequency electronic warfare progresses towards integrated and
memory (DRFM) technology has resulted in a geometric collaborative operations across multiple domains [15], the
growth in the variety of jamming methods [5][6][7]. Jamming likelihood of overlapping jamming from various sources
methods, as the tactics of electronic warfare, disrupts, deceives, escalates. Traditional recognition methods are no longer
or impairs the performance and functionality of enemy radar adequate for the rapidly evolving battlefield conditions. There
systems by transmitting specific electromagnetic signals, is an imperative to investigate intelligent recognition methods
thereby countering and weakening the effectiveness of radar to furnish comprehensive intelligence support for subsequent
technology [8]. Due to the increasingly prominent role of military operations.
The increasing complexi ty of the electromagnetic
This work was supported by Natural Science Foundation of Henan under environment is accelerating the advancement of methods for
Grant 242300421170. radar jamming recognition. Since the birth of electronic
The authors are with the School of Computer and Information Engineering, warfare, the recognition of radar jamming has consistently
Henan University, Kaifeng 475004, China, and with Henan Key Laboratory of
Big Data Analysis and Processing, Kaifeng 475004, China, and also with
been a focal point for researchers, resulting in a lot of novel
Henan Province Engineering Research Center of Spatial Information research findings. Fig. 1 primarily presents statistical data on
Processing, Kaifeng 475004, China. (E-mail: wangzan@henu.edu.cn; research papers related to radar jamming recognition over the
gzw@henu.edu.cn; gaofeng.shu@henu.edu.cn; hedalining@henu.edu.cn). past three decades, papers from sources such as Institute of
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 IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. This is the author's version which has not been fu
2
60 57 jamming, suppression jamming, and smart jamming. In
paperNum
Section III, traditional machine learning-based jamming
50 recognition methods are discussed, covering feature extraction
and analysis as well as classifier design. Section IV presents
Number of papers
40 recognition methods based on neural networks, including

35 shallow neural networks and deep neural networks. Besides,
summarizing jamming recognition methods under sample-
30
constrained conditions. Section V presents the experiments
and results. Section Ⅵ presents the discussion about radar
20
jamming recognition. Finally, in Section Ⅶ, conclusions and
future prospects are provided.
10 7 7 Others
3 IETIOP 7.2%
1 1 2.7%
3.6%
0 Elsevier
-1995 996-2000 001-2005 006-2010 011-2015 016-2020 021-2024 3.6%
1990 1 2 2 2 2 2
Year Springer
Fig. 1. Statistical graph of papers about radar jamming recognition over the 3.6%
past 30 years. The counted papers are from Institute of Electrical and Wiley
5.4%
Electronics Engineers (IEEE), Molecular Diversity Preservation International
(MDPI), John Wiley & Sons, Inc. (Wiley), and so on.
MDPI
11.7%
Electrical and Electronics Engineers (IEEE), Molecular IEEE
62.2%
Diversity Preservation International (MDPI), John Wiley &
Sons, Inc. (Wiley), and so on. It's evident that a great number IEEE MDPI Wiley Springer
of methods for jamming recognition have sprung up. Fig. 2 Elsevier IET IOP Others
depicts the distribution of published papers by publishers.
From this chart, it is evident that the IEEE publisher contains Fig. 2. Distribution of published papers divided by publishers.
the largest portion of papers, with 62.2% of the papers from
IEEE, 11.7% from MDPI, 5.4% from Wiley, 3.6% from Journal
Conference
Springer, 3.6% from Elsevier, 3.6% from IET, and 2.7% from IEEE
IOP, suggesting that IEEE can be considered the main MDPI

publisher in this field. Fig. 3 shows the number of journals and
Wiley
conferences in publishers. Fig. 4 is the keyword map of the
Publisher
selected papers. It is clear that the common keywords are Springer

“radar jamming recognition”, “active jamming”, Elsevier
“convolutional neural network”, and so on.
It can be concluded that this topic will undergo further IET
exploration, and in the coming years, we can expect to see IOP

more research publications in this area. However, these works
Others
are often scattered and tend to be based on specific scenarios
or aimed at specific radar jamming recognition, and they 0 10
Number of papers
20 30 40
usually lack more comprehensive and systematic analysis and Fig. 3. Number of journals and conferences in publishers.
summary.
In this paper, we systematically present the framework for
jamming model in radar systems and comprehensively review
the relevant literature on radar jamming recognition.
Additionally, we provide a detailed summary and analysis of
recent advancements in radar jamming recognition,
emphasizing both traditional machine learning-based methods
and neural network-based methods developed in recent years.
Finally, the paper offers the discussion, conclusion and
forecast of the challenges and future directions in radar
jamming recognition.
In response to the complex and diverse radar jamming
methods, approaches based on traditional machine learning
and neural network have gradually emerged, achieving
excellent recognition results. To provide a comprehensive
exploration of these advancements, the remaining sections of
this paper are organized as shown in Fig. 5. The structure of
the rest of the paper is as follows: Section II introduces the Fig. 4. Keyword map of selected papers.
framework of radar jamming model, including deception
3
Deception Jamming Suppression jamming can be categorized according to the
Framework of methods of influence, including aiming jamming (AJ),
Suppression Jamming
Jamming Model
Decision Tree
blocking jamming (BJ), and sweeping jamming (SJ). Besides,
Smart Jamming suppression jamming can be also classified based on different
Support Vector
Machine modulation methods such as radio frequency jamming (RFJ),
Feature Extraction
and Analysis Random Forest
amplitude modulation jamming (AMJ), frequency modulation
Traditional Machine
Radar Jamming
Learning-Based jamming (FMJ), phase modulation jamming (PMJ), pulse
Recognition Recognition Methods Classifier Design Bayesian Classifier jamming (PJ), and comb spectrum jamming (CSJ).
Data Augmentation
Smart jamming, utilizing DRFM technology, effectively
Shallow Neural
Network combines both deception and suppression jamming effects.
Neural Network-
Transfer Learning This dual functionality renders smart jamming particularly
Deep Neural
Based Recognition
Network Metric Learning
resistant to conventional suppression methods, marking it as a
Methods
highly advanced and formidable tactic in electronic warfare.
Limited Sample
Strategies
Meta Learning Common examples include noise product jamming (NPJ) and
Zero-Shot Learning noise convolution jamming (CNJ).
Fig. 5. Structure of the paper.
In radar systems, linear frequency modulation (LFM)
signals are employed and can be represented as:
II. FRAMEWORK OF JAMMING MODEL = rect exp 2 + , (1)
With the advancement of DRFM technology, the jamming where represents the range time, rect(⋅) is the rectangular
signals relayed are often highly coherent with the actual target function, is the pulse width, is the carrier frequency, is
echo signals. These signals can pass through matching filters, the chirp rate.
achieving the same gain effect as the true target echoes. This
results severely disrupting the normal detection operations of A. Deception Jamming
radar systems, and the anti-jamming capability of radar is 1) False Target Class
facing huge challenges [16]. The prompt and precise The false target class jamming signals usually refer to the
recognition of radar jamming is crucial for radars to gain forwarding after adding certain delay, frequency shift,
implement effective countermeasures. Radar jamming can be and phase shift parameters to the radar signal intercepted by
be divided into active jamming and passive jamming based on the jammer, so that the radar may falsely detect and track the
the source of energy. Additionally, the radar jamming can be false target as real target. The typical false target jamming
categorized into deception jamming, suppression jamming, signals include RDJ, VDJ, and ADJ. These models can be
and smart jamming according to the operational mechanisms represented as:
[17] [18] [19] [20]. Furthermore, the further classification of = exp 2 +Δ ! × − Δ$ ! + %
+ Δ% +j − Δ$ ! '
jamming signals relies on modulation methods and the ways (2)
is the power amplification factor, % is the power

of influence [21] [22] [23] [24]. The framework of radar
where
amplification factor, Δ$ =( , Δ =( , Δ% =( , they are
jamming model is illustrated in Fig. 6. Deception jamming [17]
regards false targets as genuine ones and is typically achieved
the jamming modulation functions for range deception,
velocity deception, and angle deception, ( is the constant.
through the process of interception, modulation, and
retransmission, simulating echo characteristics to create well-
focused false targets. Suppression jamming [18] involves 2) Gate Pull Off Class
transmitting high-power jamming noise signals to suppress the Gate pull off class jamming signals periodically and
normal reception and display of radar, effectively drowning gradually disengage the jamming signal from under the real
out genuine signals. Smart jamming [23] [24] not only could target parameters, including RGPOJ, VGPOJ, and RVGPOJ.
avoid the process of coherent radar for suppression jamming The difference between gate pull off jamming and false target
to improve the utility of jamming power, but also own better jamming lies in the modulation of additional parameters such
jamming effect than pure deception jamming. as time delay and frequency shift according to certain rules.
Deception jamming can be primarily categorized into three The process of generating jamming can be divided into three
classes based on modulation methods: false target class, gate stages: capture, pull-off, and cessation. In the capture phase,
pull off class, and dense false target class. Specifically, the the jammer intercepts radar pulses and re-emits the signals
typical false target class jamming signals include range with the same doppler frequency but increased power. During
deception jamming (RDJ), velocity deception jamming (VDJ), the pull-off phase, the doppler frequency of the jamming
and angle deception jamming (ADJ). The gate pull off class pulses is modulated, either upward or downward, to ensure
jamming signals include range gate pull off jamming (RGPOJ), differentiation between actual and deception targets. In the
velocity gate pull off jamming (VGPOJ), range-velocity gate cessation phase, the jammer ceases the transmission of
pull off jamming (RVGPOJ). In addition, the dense false deception signals once the radar tracks the spurious target,
target class jamming signals include smeared spectrum losing sight of the actual target. The signal models can be
jamming (SMSPJ), chopping and interleaving jamming represented as:
(C&IJ), dense false target jamming (DFTJ), and interrupted = ) ⋅ exp * +% + − − Δ − Δ , - + %-'
⋅ exp 2 Δ , .
sampling repeater jamming (ISRJ). (3)
4
Radar Jamming
Deception Jamming Suppression Jamming Smart Jamming
Different Methods of Modulation Different Methods of Modulation Different Methods of Influence Different Methods of Modulation
False Target Gate Pull Off

Class Class
Noise Noise
Modulation Class Modulation Class
Aiming
Range Velocity Angle Range Gate Velocity Range-Velocity Jamming
Deception Deception Deception Pull Off Gate Pull Off Gate Pull Off Radio Noise
Jamming Jamming Jamming Jamming Jamming Jamming Frequency Frequency Blocking
Jamming Modulation Jamming
Jamming Noise Noise
Dense False Target Class Other... Convolution Product
Noise Amplitude Noise Phase Sweeping Jamming Jamming
Modulation Modulation Jamming
Jamming Jamming
Spectrum Chopping and Dense False Comb
Interrupted Sampling Pulse
Smearing Interleaving Target Spectrum Other... Other... Other...
Repeater Jamming Jamming
Jamming Jamming Jamming Jamming
where ) represents the amplitude of the radar jamming signal.

Fig.6. Framework of radar jamming model.
1-1
% + − − Δ − Δ , - represents the phase term that t = 2 j6 −7 ,

/
(4)
6=1
9
j6
depicts the phase variation of the jamming signal, is the
time when the radar pulse is received, Δ stands for the = exp 8 :, 0 ⩽ ⩽ ,/
/
(5)
intrinsic time delay of the jammer, and Δ , can encompass where / denotes the number of sub-pulse, 9 denotes the
range drag or other time delay components, % is the initial bandwidth, j6 is the sub-pulse expression.
phase of the jamming signal. Δ , is the frequency shift term ● Chopping and Interleaving Jamming (C&IJ): It is to
that describes the doppler frequency shift of the jamming acquire and store part of the radar signal of a complete
signal. The method can incorporate velocity drag, range drag, pulse width radar chirp signal, and then perform signal
or combination of both, as shown in Table I.
extension to obtain the complete signal. Slice
3) Dense False Target Class
reconstruction jamming involves slicing the radar signal
Dense false target jamming signals are achieved by the
at equal intervals and reconstructing it by repeatedly
jammer intercepting radar signals, delaying, modulating,
and retransmitting them to create multiple false targets, filling the slices into adjacent intervals to create the
thereby effectively obscuring the real target among a jamming signal. The signal model can be represented as:
=-1
=2 ? ,
multitude of false targets. Common types of dense false
−
(@
target jamming include: SMSPJ, C&IJ, DFTJ, ISRJ. (6)
>=0
The expression for the first slicing process ?
● Smeared Spectrum Jamming (SMSPJ): It is composed of
multiple sub-pulses, and each sub-pulse is generated by is given as:
sampling radar signals at intervals. In the time domain, B-1
the pulse signal is reduced by the factor of /, resulting in
? = A2 u + −7 −u + − −7 D,
0
2 @ 2 ( @
(7)
1 6=0
the compressed pulse width that is of the radar
where @ represents the number of sampled sub-pulses, (

transmitted signal. According to the principles of linear
frequency modulation signals, the modulation slope of the
sub-pulse is / times of that of the radar transmitted signal.
denotes the number of time slots for each segment. The
original pulse is divided into (@ segments.
The signal model can be represented as:
TABLE I
CHARACTERISTICS OF GATE PULL OFF CLASS JAMMING MODELS
Jamming ) EFG H Characteristics
%R follows a uniform
)R Δ
Change the phase,
RGPOJ dR
distribution over [0, 2π] affect the distance measurement of target.
%V follows a uniform
)V Δ
Change doppler frequency,
VGPOJ K
distribution over [0, 2π] affect the velocity measurement of target
%RV follows a uniform
)R-V
Change both the phase and doppler frequency shift
RVGPOJ ∆ dRV
distribution over [0, 2π] simultaneously
5
● Dense False Target Jamming (DFTJ): The radar jammer stringent, making it relatively easier to implement in
generates DFTJ by delaying, stacking, and forwarding the engineering. The center frequency of SJ is not constant but
intercepted radar signal many times, capable of forming continuously varies in a periodic manner over time, resulting
multiple false targets to consume radar resources. The in a periodic impact on the radar receiver. According to the
signal model can be expressed as: different ways of modulation, it can be divided into the
L following types:
= 2 OM ⋅ − − $M , (8) 1) Noise Modulation Class
MN0
The noise modulation class jamming signals involve
where P is the forwarding times of the jammer, OM is the modulating Gaussian white noise onto the linear frequency-
magnitude of jamming modulation for the Q th forwarding, modulated signal of the radar [10], utilizing a substantial
and and $M represent the delays of the real target and the Q amount of modulated noise to create suppression effect.
th false target, respectively. Common modulation methods include RFJ, NAMJ, NFMJ,
● Interrupted Sampling Repeater Jamming (ISRJ): After and NPMJ. The signal models can be expressed as:
intercepting a large time-width signal, the jammer first = X) + O)1 Y
b
∙ [\ ] ^S_ + 92 a c d
e dT
samples and stores a small segment of the signal,
immediately forwards it after modulation and `L (10)
amplification, then receives and samples the next segment +% + f gL c Y,
of the signal, forwards it again, and repeats the process where ) is a constant amplitude, )1 is the amplitude of
until the entire pulse ends. The signal model can be modulating signal, _ represents the center frequency, `L is
expressed as: the frequency modulation factor, gL is the phase modulation
L R
− /P + / + Q 6 factor, % is the initial phase, c ⋅ is the modulated noise, O,
= 2 2 rect S T −m 6 , 9 9, and f are constants. The specific signal models for the four
6
MN0 1N
where W is the number of slices and 6 and P represent the
types of jamming signals are configured as shown in Table III.
2) Pulse jamming (PJ)
width of the slices and the number of forwarding, respectively. The signal typically refers to pulse signals modulated onto
the high-frequency carrier wave, using rectangular pulse
B. Suppression Jamming jamming as the input signal. The signal model can be
According to the different ways of influence, it is divided represented as:
+∞
= 2 rect −/ Or [\ ] 2 + % !,
into three types of jamming, the characteristics are shown in
Table II. The AJ is processed by frequency traction, the center r (11)
Where Or and
frequency of the radar receiving system is detected, and then n=-∞
the parameters of the jammer are adjusted, and the jamming espectively represent the amplitude, and
bandwidth covers the radar bandwidth. BJ bandwidth can be repetition time of the rectangular pulse jamming, r is the
greater than five times the radar receiver bandwidth, and the carrier frequency, % is the initial phase.
requirements on the jamming center frequency are less
TABLE II
CHARACTERISTICS OF SUPPRESSION JAMMING MODELS
Jamming st Ft Characteristics
AJ 9r ≈ 2∼5 9x r ≈ +y Slightly larger than the radar bandwidth
BJ 9r >59x r ≈ +y Much larger than the radar bandwidth

Δz Δ
9r = 2∼5 9x r =y + + × − ∈X0, z Y
z
2
SJ The center frequency changes periodically
Note: : radar signal center frequency; r : jamming center frequency; 9x : radar signal bandwidth; 9r : jamming bandwidth; y : frequency deviation; ∆ ] :swept
frequency range; z : sweep cycle.
TABLE III
CHARACTERISTICS OF NOISE-MODULATION CLASS SUPPRESSION JAMMING MODELS
Jamming A B C Parameter Characteristics
RFJ 1 0 0 )1 : Rayleigh distribution Limited bandwidth noise signal
NAMJ 1 0 0 )1 ∼N(0, σn2) Amplitude modulated by noise
NFMJ 0 1 0 `L : Frequency modulation slope Instantaneous frequency modulated by noise
NPMJ 0 0 1 PM : Phase modulation coefficient Instantaneous phase modulated by noise
6
(a) LFM (b) RDJ (c) VDJ (d) ADJ
(e) RGPOJ (pull off period) (f) VGPOJ (pull off period) (g) RVGPOJ (pull off period) (h) SMSPJ
(i) C&IJ (j) ISRJ (k) DFTJ (l) AJ
(m) BJ (n) SJ (o) RFJ (p) NAMJ
(q) NFMJ (r) NPMJ (s) CSJ (t) PJ
(u) NPJ (v) CNJ (w) ISRJ+CSJ (x) C&IJ+SMSPJ

Fig. 7. Time-domain waveforms of the above LFM and jamming signals.
7
(a) LFM (b) RDJ (c) VDJ (d) ADJ
(e) RGPOJ (pull off period) (f) VGPOJ (pull off period) (g) RVGPOJ (pull off period) (h) SMSPJ
(i) C&IJ (j) ISRJ (k) DFTJ (l) AJ
(m) BJ (n) SJ (o) RFJ (p) NAMJ
(q) NFMJ (r) NPMJ (s) CSJ (t) PJ
(u) NPJ (v) CNJ (w) ISRJ+CSJ (x) C&IJ+SMSPJ

Fig. 8. Time-Frequency spectrograms of the above LFM and jamming signals.
8
3) Comb spectrum jamming (CSJ): quantization bits in DRFM quantization effects. The methods
Each sub-pulse overlaps in the time domain, presenting the achieved good recognition results when the quantization bits
comb-shaped distribution in the frequency domain. By were less than four. However, with the advancement of
controlling key parameters such as the number of frequency electronic technology, the quantization bits of DRFM have
points and frequency spacing, it is possible to flexibly adjust significantly increased, rendering the harmonic effects
the number and spacing of false targets. The signal model can produced less observable. Consequently, numerous scholars
be expressed as follows: have shifted the focus to the generation and mechanisms of
M radar jamming signals, aiming to recognize the jamming based
= )|}MB rect 2 (> ] cos 2 > , (12)
on the inherent differences within the signals. Specifically, the
|}MB
>=1
initial step involves extracting the characteristic parameters of
where )|}MB and |}MB respectively represent the amplitude
the jamming signals, followed by the recognition of the
and pulse width of the jamming, Q denotes the number of
signals through the application of appropriate classifiers.
frequency points, > and (> respectively represent the A. Feature Extraction and Analysis
frequency and amplitude of the − ℎ point. In the context of jamming signal recognition, it is crucial to
C. Smart Jamming extract more typical and highly discriminative features for
radar jamming signals, and the features are usually analyzed in
By modulating the radar signal intercepted by the jammer the time, frequency, time-frequency, or transform domains.
with noise, it simultaneously achieves the effects of deception The existing typical jamming features can be categorized into
jamming and suppression jamming. Common smart jamming various mapping domains, as summarized in Table IV.
signals are NPJ and CNJ. Typically, the time-domain sequences of received signals are
The NPJ model can be expressed as:
= −$ ⋅/ ,
transformed into the desired mapping domain or kept in the
(13)
where / represents the Gaussian white noise with the mean
original domain. Subsequently, the specific characteristics of
of 0 and the variance of σ2 , $ denotes the time delay, and “⋅”
jamming signals are captured and analyzed.
1) Time Domain
indicates the multiplication process. Some scholars have introduced complexity theory and other
The CNJ model can be represented as: concepts into the field of jamming recognition, leading to the
= −$ ∗/ , (14) proposal of new time-domain feature parameters. These
where / represents the Gaussian white noise with the mean parameters include time-domain box dimension [31], L-Z
of 0 and the variance of σ2 , $ denotes the time delay, and “*” complexity [31], amplitude fluctuations [32], and higher-order
indicates the convolution process. cumulants [32], among others. Xiong et al. [31] embarked on
Finally, The LFM waveform is used as the radar an analysis of time-domain sequences involving radio
transmission signal, where the signal width T = 20 µs, frequency noise jamming and distance deception jamming,
bandwidth B = 10 MHz, sampling rate Fs = 20 MHz, and the from which they extracted features such as time-domain box
length of each time-domain sample is 2000 points. The time- dimension and L-Z complexity. Li et al. [32] compared the
domain waveforms and time–frequency spectrograms of the distinct characteristics of deception jamming and target echoes,
above LFM and jamming signals are shown in Fig. 7 and 8, focusing on amplitude fluctuations and higher-order
respectively. cumulants. The comparison provided insights to differentiate
between genuine and false targets based on feature disparities.
III. TRADITIONAL MACHINE LEARNING-BASED RADAR Tian et al. [33] proposed an algorithm based on the product
JAMMING RECOGNITION METHODS spectrum matrix. The method calculates the product spectrum
under different Pulse Repetition Intervals (PRI) and arranges it
DRFM, as the effective means of jamming, reproduces
into a two-dimensional matrix in the slow-time domain. Du et
jamming signals based on intercepted real signals [25] [26]. In
al. [34] proposed that during a coherent processing interval,
the early research about radar jamming recognition, lots of
pulse echo data is coherently demodulated and organized into
researchers attempted to utilize the signal distortion caused by
a two-dimensional matrix. By examining the sample matrix in
the quantization characteristics of DRFM during the
both the fast-time and slow-time domains, and comparing its
modulation process for radar jamming recognition. Hill et al.
bilateral symmetry, the type of the jamming can be recognized.
[27] focused on the quantization characteristics of DRFM
Xu et al. [35] extracted features including singular spectral
deception jamming, used higher-order statistical parameters to
entropy, sample entropy as the basis for classification. In
extract characteristic features of deception jamming,
addition, skewness [36] and kurtosis [36] in the time domain
pioneering the recognition of deception jamming from the
are also commonly used as expert prior knowledge to
perspective of DRFM phase quantization. Berger et al. [28]
recognize different types of jamming. Wei et al. [37]
discovered that DRFM causes signal center frequency offset
established a feature set with time domain skewness, time
and the generation of harmonics around the center frequency
domain kurtosis to recognize jamming. Chen et al. [38]
when performing velocity gate pull-off jamming. Additionally,
extracted features such as time domain skewness, time domain
the spectral characteristics of signals when DRFM performs
kurtosis, time domain instantaneous features, and entropy
range gate pull-off jamming are derived and detailed
features. They divided the training set according to the degree
mathematical models are provided. Subsequently, Greco et al.
of noise impact of the sample to train the classifier, then
[29-30] conducted in-depth research on the impact of
achieved effective real-time recognition and noise robustness
9
in identifying amplitude modulation, frequency modulation, Analyzing non-stationary signals with time-varying
and radio frequency jamming signals. frequencies requires extending the concept of the global
Although extracting feature parameters from the time spectrum to techniques that describe local frequency variations.
domain offers benefits such as reduced computational demand, Joint time-frequency analysis using time-frequency
simpler complexity, and superior real-time performance, there transformations [47] improves the analysis of non-stationary
are limitations. In the time domain, signals can easily overlap signals by revealing time-varying information embedded
and are susceptible to noise interference. At times, they may within the signal. When analyzing signals in the time-
not provide detailed insights into the frequency components of frequency joint domain, it provides not only information about
the signals. the signal in the time and frequency domains but also insights
2) Frequency Domain into the time-frequency characteristics of the signals. This
To better understand the frequency characteristics of signals, approach allows for an intuitive observation of the relationship
frequency domain analysis is a commonly adopted method. between time and frequency. Extracting feature parameters
This method involves breaking down the signal on the from the time-frequency domain offers strong noise resistance
frequency axis to reveal its spectral information, such as and stability. However, it comes at the cost of significantly
frequency components, distribution, and power spectrum. increased computational complexity compared to one-
Researchers have leveraged Fast Fourier Transformation (FFT) dimensional feature extraction methods in the time domain or
to extract features from the frequency domain, aiming to frequency domain. Therefore, reducing computational
capture the spectrum or power spectrum of the signal. Xu et al. complexity in engineering applications is the challenge that
[36] extracted frequency stability, normalized maximum needs to be addressed. Time-frequency analysis methods are
instantaneous amplitude variance, normalized instantaneous mainly categorized into linear time-frequency analysis and
frequency variance, intra-pulse correlation coefficient to quadratic time-frequency analysis.
recognize different types of jamming. Wei et al. [37] Linear time-frequency transformation methods primarily
established a feature set with frequency domain skewness, and include the short time Fourier transform (STFT) [48] and the
frequency domain envelope fluctuation of the jamming signals. wavelet transform (WT) [49]. While the traditional Fourier
Jiang et al. [39] extracted the frequency domain correlation transform lacks the ability to capture local time-frequency
dimensions of intermittently sampled jamming as feature features, STFT [50] addresses this limitation by employing a
parameters. Subsequently, they used spectral box window function to characterize local information. The
dimensionality [40] as the feature parameter for classification window size determines the scale for capturing local features,
and obtained results for distinguishing between jamming and with a narrow window providing better time resolution and a
real echo. Miao et al. [41] extracted mean frequency, centroid wider window enhancing frequency resolution. However, once
frequency, root mean square frequency, standard deviation of the window function is selected, the shape and length of the
frequency, frequency domain skewness, frequency domain time-frequency window remain fixed, making it challenging to
kurtosis, and frequency domain envelope fluctuation as the simultaneously achieve the best time and frequency resolution.
basis for distinguishing jamming. Shi et al. [42] used the Subsequently, many researchers applied STFT to the field of
carrier factor coefficient (C parameter), the ratio of the jamming signal recognition. They utilized the STFT to
variance to the square of its mean (R parameter), the flatness transform the original jamming signal into time-frequency
spectral coefficient (Fse), and the third-order bias coefficient images and performed recognition based on the features of the
(b3) to differentiate between types of jamming. time-frequency images [48]. WT employs wavelet bases as
However, frequency domain analysis has some drawbacks, time-domain window functions. The width of the wavelet
such as the loss of time-domain information, inability to bases varies with frequency, automatically narrowing the time
handle non-stationary signals, and high computational window as the frequency increases, thereby improving
complexity. In practical applications, a key limitation of resolution and overcoming the limitations of the fixed window
frequency domain analysis is its inability to handle non- function in STFT. This results in better time-frequency local
stationary signals, which are prevalent in many real scenarios. analysis capabilities. Peng et al. [49] extracted features for
3) Time-Frequency Domain pseudo-targets based on distance, velocity, and angle, then
Traditionally, radar signals have been analyzed in either the conducted wavelet analysis and extracted wavelet entropy
time domain or the frequency domain. However, one of the features to reflect the complexity and randomness of the signal,
key features commonly observed in radar data is the variation ultimately utilizing the features as effective features for
of frequency over time [43]. To better capture these time- jamming recognition. The Choi-Williams distribution (CWD)
varying characteristics, researchers utilize a variety of time- [51] is a bilinear time-frequency transform known for its
frequency domain features, including wavelet energy spectral ability to provide high-resolution representations of non-
entropy [35], Rényi entropy [36], scale center of gravity x [41], stationary signals in the joint time-frequency domain.
scale center of gravity y [41], center moment u22 [41], center Differing from linear time-frequency transformations,
moment u44 [41], variance [41], Zernike moments [44], Gabor quadratic time-frequency analysis provides enhanced time-
atomic time-frequency [45], singular value distributions [46], frequency resolution, even though it involves cross-terms.
two-dimensional instantaneous frequency parameters [46]. Common quadratic time-frequency analysis methods include
These features provide joint time and frequency information the Wigner-Ville Distribution (WVD) [52], Smooth Pseudo
about the signal, making the analysis of radar signals in non- Wigner-Ville Distribution (SPWVD) [53], Gabor transform
stationary environments more effective. (GT) [54], and so on. Among these, the WVD has attracted
significant attention from researchers due to its excellent time-
10
TABLE IV
TYPICAL JAMMING FEATURES AND EXTRACTION METHODS
Mapping Extraction
Typical features Advantages Disadvantage
domains method
box dimension [31], L-Z complexity [31],
amplitude fluctuations [32], higher-order
cumulants [32], sample entropy [35], singular
spectral entropy [35], envelope fluctuation High computing speed, Not suitable for low frequencies and
Time domain Sampling
parameter [36], kurtosis [36][38], skewness good real-time performance abrupt parameter changes
[36][38], entropy features [38], normalized
instantaneous amplitude variance [38], normalized
instantaneous frequency variance [38]
frequency stability [36], normalized maximum
instantaneous amplitude-frequency [36], intra-
pulse correlation coefficient [36], correlation
dimension [37], spectral box dimensionality [39],
Complex calculation,
Frequency kurtosis [39], mean frequency [40], centroid High accuracy,
FFT greatly affected by noise,
domain frequency [40], root mean square frequency [40], detailed frequency information
and only steady-state data is applied
standard deviation of frequency [40], skewness
[40][41], frequency domain envelope fluctuation
[39][40], power spectral entropy [41], C
parameter[42], R parameter [42], Fse [42], b3 [42]
Window function width cannot
No cross-terms,
STFT adaptively change according to
low computational complexity
frequency
High time-frequency Severely affected by
WVD
wavelet energy spectral entropy [35], Rényi concentration cross-term jamming
entropy [36], scale center of gravity x [41], scale
center of gravity y [41], center moment u22 [41], Good noise resistance,
Time- SPWVD Poor time-frequency concentration
center moment u44 [41], variance [41], Zernike fewer cross-terms
Frequency
moments [44], Gabor atomic time-frequency [45], Adaptive window,
domain Depends on wavelet base function,
singular value distributions [46], two-dimensional WT time-frequency localization,
instantaneous frequency parameters [46] poor adaptability
variable resolution
Depends on wavelet base Significantly affected by noise and
GT
function, poor adaptability parameter selection
High-resolution representation, Computational complexity,
CWD
reduced cross-Term parameter sensitivity
bispectrum feature [32], variance [57], High resolution,
information entropy [57], mean [58], BT Computationally complex
good noise resistance
variance[58], root mean square [58], box
dimension [58], waveform factor [58], peak factor
Transform
[58], kurtosis[58], margin [58], comentropy [58],
domain Multi-scale analysis, Affected by parameter selection,
the fractional-domain peak-to-average power ratio FRFT
[61], fractional-domain envelope fluctuation [61], time-frequency localization computationally complex
moment skewness coefficient [61], peak similarity
coefficient [61].
frequency concentration properties. Unlike methods relying on effectively extract detailed time-frequency features from
sliding window functions, the WVD is immune to variations signals. Wavelet energy spectral entropy [35] and Rényi
in time and frequency resolution and boasts the best energy entropy [36] in the time frequency domain are commonly used
concentration. However, when analyzing signals with multiple as expert prior knowledge to recognize different types of
components, it introduces severe interference in the form of jamming. Miao et al. [41] use image processing techniques to
cross-terms. Optimizing the kernel function is an effective extract feature parameters of the grayscale images of jamming
improvement method to address the issue of cross-term signals, including scale center of gravity x [41], scale center of
jamming. For example, SPWVD reduces interference between gravity y [41], center moment u22 [41], center moment u44 [41],
signal components by applying a smoothing kernel (such as a and variance [41], thereby distinguishing six types of jamming.
Gaussian kernel) [55][56]. The Gabor Transform (GT) In order to effectively deal with the major three types of
decomposes the signal into time-frequency atoms using a set deception jamming. Yang et al. [44] proposed a recognition
of windows that are shifted in both the time and frequency algorithm based on Zernike moment feature of time-frequency
domains. By doing so, GT provides a more localized time- distribution, with a focus on the analysis of a pulse signal
frequency representation, which helps to reduce the cross-term during the pull-off period. They selected SPWVD to mitigate
jamming. Nevertheless, this comes at the cost of sacrificing the cross-term interference inherent. Through the computation
time-frequency energy concentration and increasing of Zernike moments, they achieved recognition of RGPOJ,
computational complexity. Nevertheless, this comes at the cost VGPOJ, and RVGPOJ. In addition, in order to recognize
of sacrificing time-frequency energy concentration and dense false target jamming, Hao et al. [45] proposed a
increasing computational complexity. recognition method based on time-frequency atomic
By applying time-frequency analysis methods, we can decomposition theory, they extracted the autocorrelation
11
function of the jamming and then adaptively constructed a The FRFT often referred to as generalized Fourier
time-frequency atom sub-dictionary. The approach enables the transform [59], is capable of handling non-linear and non-
rapid extraction of Gabor atomic time-frequency parameters of stationary signals that exhibit overlap in the time-frequency
the SMSPJ and C&IJ. However, the method seemed plane. By appropriately rotating the signal along the time axis,
somewhat limited in handling various jamming types. Sun et FRFT not only captures time-domain information but also
al. [46] extracted singular value distributions and two- provides insights into the frequency domain. The method
dimensional instantaneous frequency parameters from the makes it highly effective in mitigating cross-terms [60],
WVD time-frequency representation, creating two- endowing the FRFT with comprehensive time and frequency
dimensional features for the recognition of six types of domain processing abilities. Lin et al. [61] fully exploited the
jamming. The experimental results show that even in the case energy concentration characteristics of FRFT to process five
of low JNR and small number of training samples, the time- typical jamming signals by selecting the optimal rotation angle,
frequency two-dimensional eigenvalues extracted show successfully extracting characteristic parameters such as
obvious differences. envelope fluctuation, moment skewness coefficient, and peak
Due to significant differences in feature performance, the similarity coefficient. Simulation results show that the
time-frequency features are used as classification criteria, recognition algorithm based on FRFT domain feature
offering a novel approach to extracting jamming features. It is extraction can effectively recognize typical jamming.
worth noting that quadratic time-frequency analysis methods As we can see from the above, the application of
can enhance both time and frequency resolution. However, the appropriate time-frequency analysis methods within the time
multiplication process introduces cross time-frequency domain, frequency domain, time-frequency domain, and
distributions, which are akin to interactions between the transform domain enables the extraction of extensive feature
squared sums of two components. The cross-terms can lead to information from jamming signals. However, large-scale data
significant jamming in the analysis of the signal's time- often contains redundant information, which can significantly
frequency components. The jamming from cross-terms can be increase computation time. This not only hinders the
mitigated through the optimization of kernel functions. improvement of recognition accuracy but also risks giving rise
4) Transform Domain to the “curse of dimensionality”. Therefore, it is crucial to
In addition to time-frequency features, transform domain consider effectively reducing the dimensionality of the data
features have become the important direction for extracting before feeding it into the classifiers for recognition.
jamming feature parameters. Commonly extracted features in Dimensionality reduction can be approached from two
the transform domain include bispectrum feature [32], perspectives: On the one hand, feature selection, which
variance [57], information entropy [57], mean [58], involves using evaluation criteria to identify the optimal
variance[58], root mean square [58], box dimension [58], subset of variables, helps eliminate redundancy, simplifies the
waveform factor [58], peak factor [58], kurtosis[58], margin training process, and enhances the efficiency of the model. On
[58], comentropy [58], the fractional-domain peak-to-average the other hand, mapping high-dimensional data to a lower-
power ratio [61], fractional-domain envelope fluctuation [61], dimensional space preserves essential information, thereby
moment skewness coefficient [61], and peak similarity reducing the data dimension and computational cost. Tian et al.
coefficient [61]. These features provide valuable insights into [33] adopted non-negative matrix factorization (NMF) for
the jamming signals, thus enhancing overall recognition feature extraction, selected optimal features for recognition,
performance. and further analyzed the separability of feature parameters,
Common approaches include using various transformation which helped enhance recognition performance. In the same
methods to obtain feature parameters in different domains for year, Zhou et al. [62] filtered out the noise modes with less
jamming signal recognition. Methods such as bispectral original information through mode selection. Subsequently,
transformation (BT) and fractional Fourier transformation features with better separation were selected from the features
(FRFT) are among the means used, providing new avenues for extracted from the remaining modes by the interclass
extracting jamming feature parameters and increasing the divisibility method. Therefore, when reducing data
ways to gather jamming information. dimensionality, it is necessary to choose the appropriate
As a popular method in high-order spectral estimation, method based on actual conditions, considering whether
bispectrum analysis has been widely used in various fields due enough information can be retained after reduction and other
to the excellent performance in suppressing Gaussian noise factors like computational efficiency.
and retaining the amplitude and phase information of signal.
B. Classifier Design
By applying bispectrum transform to the signal and extracting
bispectrum feature, variance, and information entropy as the Using correlation methods to extract features, followed by
feature factors, the jamming signals can be effectively selection, allows them to be placed into a classifier for
recognized [32][57]. Zhou et al. [58] introduced an algorithm recognition. Selecting the appropriate one from various types
that utilizes BT to extract diagonal slice features such as mean, of classifiers can contribute to improve jamming recognition
variance, root mean square, box dimension, waveform factor, accuracy. Commonly used classifiers include the decision tree
peak factor, kurtosis, margin, and comentropy for deception (DT) [63], support vector machine (SVM) [64], random forest
jamming recognition. Experimental results demonstrate that (RF) [65], bayesian classifier (BC) [66], and more. Table V
BT can effectively extract relevant features, thereby shows the literature on traditional machine learning-based
facilitating the recognition of jamming signals. recognition methods.
12
1) Decision Tree classification, they achieved an overall recognition rate of over
DT classifier is a machine learning algorithm used for 95% in the JNR range of −10 dB to 0dB, demonstrating high
classification and regression tasks. It is a supervised learning noise resistance of the method. Xin et al. [71] combined with
method that constructs a model in the form of a tree structure, the features in the time domain and frequency domain, as well
where each internal node represents a decision or a test on an as the texture features in the grayscale symbiosis matrix of the
attribute, each branch represents the outcome of the test, and time-frequency graph, then fed the features into the SVM, the
each leaf node represents a class label or a regression value method can reach 97.2% at 0 dB and converge to 100% at 3
[67]. Xu et al. [36] extracted a set of highly separable features dB. Compared to DT, SVM classifiers only require the
from typical jamming and deception jamming in the time extraction of jamming features without the need for manually
domain, frequency domain, and time-frequency domain. They setting classification thresholds for feature parameters, saving
used the features for classification with a decision tree and a significant amount of human resources.
compared it to SVM. The simulation results showed that the 3) Random Forest
decision tree had higher recognition rate than SVM. However, RF classifier combines the principles of bagging ensemble
there were significant differences in recognition rates for learning theory [72] with the random subspace method [73]. It
different jamming signals. To address the issues with is an ensemble learning model based on decision trees [74].
traditional decision tree-based jamming recognition classifiers, RF comprises multiple decision trees, primarily trained
which require prior knowledge and manual intervention, Wei through bagging ensemble learning, the final classification
et al. [37] established a feature set with multidimensional result is determined through a combination of individual
parameters extracted from the time domain and frequency decision tree outputs and voting. RF addresses the bottleneck
domain of the jamming signals. They used a fuzzy C-means issues associated with decision trees, making it well-suited for
(FCM) clustering algorithm to automatically design a decision parallel and scalable high-dimensional data classification. It
tree, avoiding the need for any prior knowledge when also exhibits robustness to noise and outliers, functioning as
selecting decision thresholds. The jamming recognition discriminative and non-linear model. RF supports
accuracy reached 90% when the jamming-to-noise ratio (JNR) classification, regression, and multi-classification problems
exceeded 10 dB. However, the approach did not consider the while offering strong interpretability. Du et al. [75] proposed a
scenarios with low JNR. Later, Chen et al. [38] aimed to method for wideband unconventional jamming recognition,
balance real-time performance and noise robustness, they used extracting features from visibility graphs and employing the
rough set analysis to determine whether jamming signals were RF classifier for unconventional jamming recognition.
affected by noise and then fed the results into the relevant DT Experimental results indicate that recognition probabilities
for secondary classification. Even when JNR was as low as 0 exceed 90% when the JNR is greater than 0 dB. Xu et al. [35]
dB, the recognition rate for AMJ, FMJ, and RFJ jamming extracted features including power spectral entropy, wavelet
exceeded 90%. It can be observed that the decision tree-based energy spectral entropy, singular spectral entropy, sample
radar jamming recognition classification methods are entropy as the basis for classification. They combined the
straightforward and easy to understand. However, these features with the RF classifier to recognize AMJ, FMJ, and
methods typically rely on the single-feature parameters at PMJ. Using the RF classifier generally outperforms individual
classification nodes, which can significantly impact DT in terms of recognition capability. However, it requires
classification results based on selection of feature parameters more training time and larger memory space when handling
and classification thresholds. large-scale datasets.
2) Support Vector Machine 4) Bayesian Classifier
SVM [68] has been extensively applied in fields such as BC is known for its fast training on massive datasets,
signal processing and image recognition. It works by using support for incremental training, and significant advantages in
kernel functions to map input vectors into a higher- data processing and radar signal recognition [76] [77]. It
dimensional feature space, thus enabling classification of non- specifies conditional independence assumptions through
linear problems in this space. Zhou et al. [62] introduced a conditional probabilities, allowing complex problems to be
feature extraction algorithm for deception jamming decomposed into simpler models for efficient processing. The
recognition and compound deception jamming recognition models learned are also easily interpretable, but the
based on variational mode decomposition (VMD). By feeding computational complexity is high, making it challenging to
the extracted features into SVM, they achieved an average implement. Li et al. [78] extracted three feature parameters,
recognition rate of 97.17% at SNR of 0 dB. To address the namely information entropy, exponential entropy, and norm
challenge where there is a significant overlap of sample points entropy, with the aim of distinguishing between deception and
in the jamming signal space, Wang et al. [69] proposed the suppression jamming, they employed a naive bayes classifier
signal feature-space support vector machine (SF-SVM) (NBC) that utilized the differences in these features for
classification algorithm. When SNR is − 3 dB, the method classification purposes. Zhou et al. [58] extracted
could recognize more than 80% of six types of jamming multidimensional and distinct features, combined with the BC,
signals, marking a substantial improvement in recognition achieving excellent recognition results with the recognition
accuracy. Wang et al. [70] extracted features from frequency rate of 90% at JNR = 0 dB. Increasing the dimensionality of
domain and time-frequency domain, and they constructed a features within a certain range can effectively improve
three-dimensional feature vector using spectral box dimension, jamming recognition. Shi et al. [42] proposed one method that
singular value entropy, and wavelet packet energy entropy. By approximated conditional feature probabilities via data
inputting the feature vectors into a multi-class SVM for
13
augment and kernel density estimation (KDE) from small success, the effectiveness of certain methods is not always
labeled dataset to construct NBC. The recognition rate exceeds satisfactory. Table VI shows the advantages and disadvantages
75% at a JNR of −10 dB. Dai et al. [79] extracted frequency of traditional machine learning-based recognition methods.
entropy and peak-to-average ratio, constructing joint time- The main reasons for this are as follows: feature extraction is
frequency features and establishing the minimum risk NBC typically tailored to specific tasks or datasets, and the diverse
for target and jamming recognition. nature of jamming signals makes it increasingly challenging to
Although these studies have achieved some degree of
TABLE V
PAPERS BASED ON TRADITIONAL MACHINE LEARNING METHODS
NO. Author Date Jamming Extract feature Classifier Performance
time domain: kurtosis, skewness, normalized JNR: 0dB,

Chen et
38 2022 AMJ, FMJ, RFJ instantaneous amplitude variance, normalized DT Recognition Rate: more
al.
instantaneous frequency variance, entropy features than 90%
Hao et JNR>10dB, Recognition
45 2019 SMSPJ, C&IJ Gabor time-frequency dictionary SVM
al. Rate: more than 90%
bispectrum transform domain: mean value, JNR: −3dB, Recognition
Zhou et RGPOJ, VGPOJ, variance, root mean square, box dimension, Rate: more than80%;
58 2021 NBC
al. RVGPOJ waveform factor, peak factor, kurtosis, Margin JNR: 0dB, Recognition
factor, comentropy Rate:90%
time domain: skewness, kurtosis, envelope
fluctuation parameter;
Training Samples:300,
Xu et AMJ, FM BJ, FM SJ, DDJ, frequency domain: frequency stability, normalized
36 2019 DT Average Recognition Rate:
al. MFTJ, ISRJ, SNJ maximum instantaneous amplitude, intra-pulse
more than 98%
correlation coefficient;
time-frequency domain: Rényi entropy
AMJ, FMJ, CSJ, Coherent
Fuzzy JNR: 10dB,
Wei et Suppressing, Front Copy, time domain: kurtosis, skewness;
37 2019 Clustering Recognition Rate: more
al. Middle Copy, Back Copy, frequency domain: skewness, envelope fluctuation
DT than 90%
DFTJ, ISRJ
RGPOJ, VGPOJ,
RVGPOJ,
SNR: 0dB,
Zhou et RGPOJ+VGPOJ, energy entropy, information entropy, box
62 2013 SVM Average Recognition
al. RGPOJ+RVGPOJ, dimension, standard deviation, kurtosis
Rate:97.17%
VGPOJ+RVGPOJ,
RGPOJ+VGPOJ+RVGPOJ
space spectral bandwidth, space time peak-to-
STJ, MTJ, average ratio, space power spectrum flatness, space
Wang SNR: 6dB,
69 2017 NJ, IJ, Linear chirp amplitude spectrum peak-to-average ratio, space SF-SVM
et al. Recognition Rate:95%
jamming, CSJ fractional degree of aggregation, space
compression gain
JNR: −10dB,
Wang STJ, MTJ, LFMJ, box dimension, singular value entropy,
70 2022 SVM Recognition Rate：more
et al. PBJ, CSJ wavelet packet energy entropy
than 90%
time domain: skewness, kurtosis, envelope factor, JNR:0 dB, Recognition
peak-to-average ratio;
Xin et STJ, MTJ, AMJ, Rate :97.2%；
71 2023 frequency domain: skewness, kurtosis, zero-center SVM
al. FMJ, LFMJ JNR:3dB, Recognition
normalized power spectrum maximum, envelope
factor Rate converge to 100%
average degree, average clustering coefficient, JNR:>0dB,
Du et ISRJ, ISCJ,
75 2019 newman assortativity coefficient, normalized RF Recognition Rate: more
al. PDTJ, SNCJ, NPJ
network-structure entropy than 90%
Xu et power spectral entropy, wavelet energy spectral SNR:0dB,
35 2020 FMJ, AMJ, PMJ RF
al. entropy, singular spectral entropy, sample entropy Recognition Rate:80%
JNR: −10dB,
STJ, MTJ, NJ, Broadband
frequency domain: f , R, Fse, b3
Shi et jamming, Comb frequency
42 2021 NBC Recognition Rate: more
al. jamming, Sweeping
than 75%
frequency jamming
NAMJ, NFMJ, Sine AMJ,
Dai et JSR: 10dB,
79 2022 Square AMJ, Triangle frequency entropy, peak-to-peak ratio ONBCMR
al. Recognition Rate: 80.6%
AMJ, Sine FMJ
Note：Deception jamming: front copy jamming, Middle Copy, Back Copy, interrupted-sampling circularly jamming (ISCJ), partial-pulse dense transmitted
jamming (PDTJ). Suppression jamming: single tone jamming (STJ), multi tone jamming (MTJ), partial band jamming (PBJ), narrow-band jamming (NJ),
LFMJ, Coherent Suppressing, impulse jamming (IJ), sparse noise convolutional jamming (SNCJ), Broadband jamming, Comb frequency jamming, Sweeping
frequency jamming, NAMJ, NFMJ, Sine AMJ, Square AMJ, Triangle AMJ, Sine FMJ. f : carrier factor coefficient; R: the ratio of the variance to the square of
the mean of the envelope; Fse: flatness spectral coefficient; b3: 3-order bias coefficient.
14
TABLE VI
THE ADVANTAGES AND DISADVANTAGES OF TRADITIONAL MACHINE LEARNING METHODS
Classifier Advantages Disadvantages

Easy to understand and interpret, Prone to overfitting, not suitable for high-
DT [36][37][38][63][67]
simple computation dimensional datasets
Effective handling of high-dimensional spaces, High computational complexity,
SVM [62][64][68][69][70][71]
strong generalization ability requires parameter tuning
RF Good parallelism and scalability for high-dimensional data, Less interpretable,
[35][65][72][73][74][75] evaluates feature importance longer training time
BC Suitable for small sample data, Requires prior knowledge,
[42][58][66][76][77][78][79] provides probability estimates assumes feature independence
extract key features. Moreover, an excessive number of decline in the classification performance. The BP neural
extracted features can lead to increased redundancy. The lack network has a strong nonlinear mapping capability and allows
of generality across different tasks or datasets may result in a for the setting of the number of hidden layers and the number
decrease in model performance when applied to a new dataset, of neurons in each layer based on the specific problem. Ruo-
necessitating the re-extraction of certain features. Secondly, Ran et al. [84] extracted features from the time domain and
many classification models require a large amount of labeled frequency domain and established a classifier based on the BP
samples for pre-training to achieve better generalization. neural network. The results indicated that as the JNR
However, obtaining a significant number of labeled signal
increased, the recognition rate of various composite jamming
samples in advance is highly impractical. Additionally,
signals improved. When the JNR is above 10 dB, the
labeling a large number of samples can be an exceedingly
recognition rate could exceed 90%, though the JNR
time-consuming and labor-intensive task, which can
substantially impact the performance of jamming recognition. significantly affected the performance. Sun et al. [46] fed
time-frequency domain features into the BP network,
IV. NEURAL NETWORK-BASED RADAR J AMMING achieving a recognition rate of nearly 90% for six types of
RECOGNITION METHODS jamming when the JNR was at 0 dB. But the BP neural
network has a slow learning speed and is prone to getting
Due to the advantages of neural network methods, such as
trapped in local minima. The RBF neural network has a
automatic feature learning and the ability to handle various data certain robustness, can effectively handle noisy data, and
distributions without constraints, an increasing number of
reduces computational load. Peng et al. [49] subjected
researchers are applying neural network approaches to jamming
jamming signals to wavelet analysis, extracting wavelet
signal recognition. This paper focuses on introducing jamming
entropy, and forwarded it to the RBF neural network for
recognition methods and the performance in jamming recognition
automatic classification, resulting in faster computation speeds
using shallow and deep neural networks.
and better real-time capabilities. However, the RBF network is
A. Shallow Neural Network Methods overly dependent on parameter selection. It largely relies on
Common shallow networks include the back propagation the choice of basis functions and parameter settings, which
the Kohonen self-organizing feature map neural network [80], requires extensive experimentation and experience to select
(BP) neural network [81], and the radial basis function (RBF) appropriate parameters. Table VII shows the papers on
neural network [82]. The Kohonen neural network possesses shallow neural network methods.
the ability for distributed information storage and can perform B. Deep Neural Network Methods
self-organizing feature analysis of data. It supports clustering
In recent years, deep learning [85] has emerged as a leading
analysis for large datasets but requires the selection of trend in big data analysis, achieving significant breakthroughs
appropriate parameters such as network structure and learning in computer vision tasks such as image recognition [86],
rate. Ma et al. [83] employed the Kohonen neural network as a speech recognition [87], and object detection [88]. Building on
classifier, effectively distinguishing angular false target these advancements, deep learning has been introduced into
jamming. However, if the parameters of the Kohonen neural jamming recognition and has shown promising results
network are not chosen properly, which might lead to a [89][90][91][92]. In contrast to traditional machine learning-
TABLE VII
PAPERS BASED ON SHALLOW NEURAL NETWORKS
NO. Author Date Jamming Preproccess Model Performance
SIR: −10dB~0dB,
83 Ma et al. 2005 Repeating ADJ WVD Kohonen
Recognition Rate :98%
Ruo-Ran Multi-domain JNR:10dB, Recognition Rate: more
84 2016 RFJ, NAMJ, NFMJ, NPMJ, RGPOJ BP
et al. features than 90%
Narrow-band noise jamming, Coherent noise JNR: 0dB,
46 Sun et al. 2021 WVD BP
jamming, Sample jamming, ISRJ, RGPOJ, DFTJ Recognition Rate: close to 90%
Wavelet entropy JNR:6dB, Recognition Rate :90%;
49 Peng et al. 2020 VDJ, DDJ, ADJ RBF
features JNR:9dB, Recognition Rate: 98%
Note：Deception jamming: Sample jamming. Suppression jamming: Narrow-band noise jamming, Coherent noise jamming.
15
based recognition methods that rely on manually extracted the recognition accuracy remained as high as 97.9%.
features, the neural network methods can autonomously learn However, the computational efficiency was relatively low.
advanced features from complex jamming. By leveraging Later, Zhao et al. [93] imported the measured one-dimensional
jamming features, deep learning methods can adeptly tackle echo data into CNN directly, demonstrating the effectiveness
the complex challenges of jamming recognition. Consequently, of CNN for jamming recognition with recognition accuracy
numerous deep neural networks have been developed to exceeding 99.7%. In the same year, Lin et al. [98] used CNN
extract jamming features, they can autonomously extract to identify the typical jamming and spoofing jamming,
jamming features using convolutional kernels, and offer an
achieving recognition rates of over 99%.
end-to-end solution that eliminates the necessity for manual
feature design and extraction. Popular deep neural network Meng et al. [99] separately fed the real and imaginary parts
architectures [93][94][95][96] include the convolutional neural of jamming data into CNNs with the same model structure and
network (CNN) and the recurrent neural network (RNN). Fig. parameters for feature extraction, and the features from both
9 gives relevant deep learning models. branches were then merged. The approach demonstrated
However, the type of features extracted from deep networks impressive performance, achieving an average recognition
may vary. For instance, different deep networks can extract accuracy of 98.52% when the training sample size was 100.
features in the time domain, frequency domain, time- To reduce parameter redundancy and accelerate the processing
frequency domain，and transform domain. In this section, we time, Zhang et al. [100] proposed a fast recognition method
systematically review the jamming recognition methods based for fast complex-valued CNN (F-CV-CNN) radar jamming
on deep learning. Depending on the feature domain, deep signal based on pruning, and the recognition accuracy was
networks for jamming recognition can be further divided into 94.42% when the number of training samples was 150. To
time domain feature networks, frequency domain feature automatically recognize the jamming signal components of
networks, time-frequency domain feature networks, transform compound jamming signals, Meng et al. [101] proposed a
domain feature networks and multi-domain feature fusion multi-label class representation complex-valued convolutional
networks. Table VIII shows the papers on deep learning with neural network (ML-CR-CV-CNN) with an end-to-end
sufficient samples. manner. At JNR > 10 dB, the accuracy rate of the model based
1) Time domain feature network on multi-label classification reaches stable values of more than
Wu et al. [97] effectively extracted jamming features using 95%. Later, long short-term memory (LSTM) networks
CNN and mapped the extracted features to a two-dimensional attracted extensive attention in the field of classification [102],
space using multidimensional scaling (MDS), subsequently it can automatically extract features from signals and learn the
identifying jamming signals. Even under low SNR of −5 dB, long-range sequential dependencies. It is a type of RNN model
Raw data
(a)
224×224×64
112×112×128
x
56×56×256 Forget Gate Input Gate Output Gate
weight layer
Ct-1 Ct
28×28×512 14×14×512 7×7×512 x tanh
1×1×4096
F(x) relu
identity
ft it ot
weight layer σ σ tanh σ
convolution+ ReLU
Wf Wi Wc Wo
max pooling F(x)+x
ht-1 ht
fully connected +ReLU relu
xt
softmax
(b) (c) (d)
Fig.9. A selection of relevant deep learning models. (a) The LeNet Network. (Image from [93]). (b) The Visual Geometry Group Network. (Image from [95). (c)
The residual neural network block. (Image from [111]). (d) The LSTM unit. (Image from [104]).
16
used for time-series data processing, the cells can store values dimensional convolutional residual network to classify and
from any past instances and are not severely affected by the discriminate time-frequency image time series. This method
problem of gradient diminishing [103]. Qin et al. [104] used achieved classification and discrimination for five types of
the signal sequences in gate pull off jamming class as inputs to jamming signals, with recognition rates exceeding 95% when
an LSTM model, learning the dependencies between signal the JNR was greater than −10 dB. Lan et al. [111] fed two
frames. The model accurately recognized gate pull-off types of signals into Resnet50 for training: one being multi-
jamming and exhibited excellent generalization performance. pulse jamming combined with echo, and the other being multi-
2) Frequency domain feature network pulse jamming combined with the original signal. When the
In response to the challenge of non-stationary signals being jamming-to-signal ratio (JSR) was 3 dB and JNR was −4 dB,
unsuitable for feature extraction using Fourier transforms. Qu the recognition accuracy rate both reached 99%. Wang et al.
et al. [105] extracted power spectral features for suppression- [112] used STFT to generate time-frequency images for six
type composite jamming. They utilized jamming recognition different types of jamming signals with JNR below 0 dB. They
network known as JRNET, which incorporated a fusion of employed the ResNet-50 network to classify and identify
residual blocks and asymmetric convolution blocks. The various jamming signals, achieving the recognition rate
approach enhanced the capability to recognize subtle features. exceeding 90% when JNR was above 0 dB. However, due to
While it successfully identified jamming types, it was unable the abundance of information contained in the time-frequency
to capture relative positional information and was confined to images, the global representation risked being overlooked.
specific jamming signals, making it challenging to address CNNs required the continual stacking of convolutional layers
composite jamming scenarios. Due to the presence of a large and pooling operations to expand the receptive fields for
amount of jamming information in the frequency domain, Z. global information extraction, resulting in bulky recognition
Dang [106] used frequency domain features and deep neural networks with increased computational complexity and
networks to enhance radar jamming recognition. They input parameter counts. In recent years, vision transformers have
the frequency domain data of jamming signals into a complex- been advancing rapidly. Compared to convolutional networks,
valued ResNet, for JNR greater than −4 dB, the recognition transformers can model long-range dependencies via the self-
rate for each type of jamming signal approached 100%. The attention mechanism and capture information from different
recognition network performed well under low JNR conditions. representational subspaces through multi-head self-attention
3) Time-frequency domain feature network [113]. Lang et al. [114] integrated the global representation of
Focusing solely on the global features of the signals in jamming time-frequency data, leveraging the strengths of both
either the time domain or the frequency domain would result transformers and CNN. They utilized the multi-head self-
in the neglect of its local information, which is insufficient to attention mechanism from transformers to merge global and
meet the requirements of signal processing and analysis. To local information within the time-frequency images,
capture the local information of the signal, it is necessary to enhancing the feature extraction capabilities of the recognition
compute two-dimensional joint function of time and frequency. network. Remarkably, they achieved the recognition accuracy
Many studies have accomplished this by converting the of 99.5% at the JNR of − 6 dB. Swin Transformer (SwinT)
original signal into time-frequency spectrogram [107] and reduces the computational and memory costs substantially
utilizing networks to extract time-frequency domain features. through the shifted window attention design and hierarchical
Liu et al. [108] utilized time-frequency feature images and representation [115]. It is one of the important advancements
inputted them into a pre-trained CNN for classification. in vision transformer architectures. Sha et al. [51] proposed a
Simulation results indicated that the model achieved the diffusion model as the preprocessing module to denoise the
accuracy of 98.67% for nine types of jamming signals within input signals, then utilized SwinT to extract global features
the JNR range of 0 to 8 dB. Nonetheless, further research is from the CWD graphs. Overall, the recognition rate for all
needed to explore the adaptability and robustness to various jamming types exceeds 95% at JNR is −4 dB.
jamming signals. Cai et al. [109] constructed a spectrum To reduce model complexity and the number of parameters,
waterfall containing white gaussian noise signals and proposed Qiu et al. [116] proposed the hybrid dilated convolution (HDC)
a CNN-based recognition algorithm for three types of method. HDC is a variation of regular convolution operations,
jamming signals. preserving the advantages of local connections and weight
Using the spectral waterfall diagram to convert the occupied sharing found in typical CNN models. The innovation
channels by jamming devices into images, the algorithm facilitated jamming recognition. Zhang et al. [117] adopted
achieved a recognition accuracy of around 90% at the JNR of the concept of object detection for jamming recognition,
30 dB, but the jamming signals were relatively simple. Wang overcoming the traditional need for separate designs targeting
et al. [110] utilized STFT to extract time-frequency domain different jamming types and parameters. They performed
features and constructed a CNN model for jamming signal time-frequency analysis and normalization on jamming signals
recognition. When the JNR was between 4 dB and 16 dB, the and used the normalized time-frequency images through STFT
recognition rate exceeded 94%. Li et al. [95] introduced a as inputs to the YOLO network. The model simultaneously
method based on STFT-CNN to extract jamming features. identified each jamming component and provided precise
They employed an improved VGGNet-16 network model for time-frequency localization information. When JNR exceeded
feature learning and training and achieved composite jamming 0 dB, both the accuracy and recall rates of jamming detection
recognition for two superimposed jamming signals with the exceeded 85%. While some research lacked emphasis on
recognition rate of 97.7%. Wang et al. [96] utilized a three- lightweight network designs, Chen et al. [118] obtained time-
17
TABLE VIII
PAPERS BASED ON DEEP NEURAL NETWORKS WITH SUFFICIENT SAMPLES
Wu et JNR: −5dB,
97 2017 Audio jamming, NJ, PJ, SJ, SSJ Sampling CNN
al. Recognition Rate: 97.9%
Zhao et Clutter, SFJ, PMJ 1MHz, FMJ 1MHz, One-dimensional Recognition Rate:
93 2019 LeNet
al. FMJ 10MHz, SJ 5M, SJ 400M echo data more than 99.7%
RFJ, NAMJ, NFMJ, MFTJ, SMSPJ, C&IJ, Two-dimensional JNR: −5~30 dB,
98 Lin et al. 2017 2D⁃CNN
ISRJ, SFNCJ recurrence plot Recognition Rate:99%
Train Sample:100,
Meng et
99 2021 ISRJ, AJ, BJ, SJ, DFTJ, Smart jamming Real, Imaginary 1D-CNN Average Recognition Rate:
al.
98.52%
Zhang et Noise, ISRJ, AJ, BJ, SJ, DDJ, DFTJ, SNJ, CJ, Train Sample: 150, Recognition
100 2021 Real, Imaginary F-CV-CNN
al. CJ+ISRJ, DFTJ+SNJ, DDJ+SJ Rate :94.42%
One-dimensional
Meng et Compound jamming (Component: ISRJ, JNR > 10 dB, Recognition Rate:
101 2023 time–domain ML-CV-CNN
al. NAMJ, NFMJ, DFTJ) more than 95%
signals
JSR: 3 dB~7 dB,
Qin et
104 2019 RGPOJ, VGPOJ, RVGPOJ Sampling LSTM Average Recognition Rate: more
al.
than 95%
AMN, FMN, PMN, SFN, AFN, APN, ASN, Power spectrum JNR: −20 dB,
105 Qu et al. 2020 JRNET
FPN, FSN, PSN features Recognition Rate: 85.92%
JNR: −4 dB,
Complex-valued
106 Z. Dang 2020 STJ, MTJ, PBJ, NFMJ, LFMJ, PJ FFT Recognition Rate:
ResNet
Approximately 100%
STFT， JNR:0～8 dB,
NAMJ, NFMJ, CNJ, CPJ, SMSPJ,
108 Liu et al. 2019 CNN Average Recognition
C&IJ, MFTJ, ISRJ, CSJ OS⁃CFAR
Rate :98.67%
JNR:30 dB,
109 Cai et al. 2019 SJ, STJ, MTJ Spectrum waterfall CNN
Recognition Rate:90%
Wang et Narrow pulse jamming, MFTJ, JNR: 4dB~16dB, Average
110 2019 STFT CNN
al. Blanket jamming Recognition Rate :94%
SNR: −2db~16dB, Recognition
95 Li et al. 2020 PBJ, STJ, MTJ, LFMJ, NAMJ, NFMJ STFT VGGNet⁃16
Rate: more than 93%
Wang et JNR> −10dB, Recognition
96 2021 CNJ, NPJ, C&IJ, SMSPJ, ISJ STFT 3D-ResNet34
al. Rate: more than 95%
Lan et RFTJ, VFTJ, ISDJ, PDTJ, ISRJ, WDTJ, JNR: −4dB, Recognition
111 2021 STFT Resnet50
al. SMSPJ, RGPOJ, VGPOJ, RVGPOJ Rate :99%.
Wang et JNR>0dB, Recognition Rate:
112 2021 CNJ, CSJ, SMSPJ, C&IJ, ISRJ, RGPOJ STFT Resnet50
al. more than 90%
JNR: −6 dB,
AJ, BJ, SJ, DDJ, VDJ, ISRJ, SMSPJ, DDJ +
Lang et Lightweight
114 2022 ISRJ, DDJ + SMSPJ, VDJ + ISRJ, VDJ + STFT
al. JR-TFViT Recognition Rate: 99.5%
SMSPJ, ISRJ + SMSPJ
Sha et NAMJ, NFMJ, CNJ, NPJ, CSJ, SMSPJ, C&IJ, JNR: −4dB, Recognition Rate:
51 2023 CWD Diff-SwinT
al, ISRJ more than 95%
Qiu et JNR:0db~15dB, Average
116 2022 RFJ, NAMJ, NFMJ, ISRJ, SMSPJ, CNJ STFT Gesture9-Net
al. Recognition Rate：90%
Zhang et NFMJ, CSJ, Narrow pulse jamming, ISRJ, JNR:0db~15dB, Average
117 2022 STFT YOLO v5
al. DFTJ, SNJ Recognition Rate: 90%
Chen et Lightweight SIR: −20dB, Average

118 2023 ISDRJ, ISRRJ, DDJ, DFTJ, SMSPJ, CNJ STFT
al. CNN Recognition Rate: 100%
JNR: 5dB,
Zhu et
121 2020 NFMJ, NAMJ, RFT, VFT, CSN, PSN, SL SPWVD MLAMC Average Recognition Rate: more
al.
than 90%
FPFJ, PPDFJ, ISRJ, SNJ, SJ, de-chirp, STFT, JNR: 10~30dB,
122 Lv et al. 2024 MLCJRN
Noise suppression jamming GSCT Recognition Rate: 92.25%
Zhou et JNR:0 dB, Recognition Rate：
123 2023 NPJ, CSJ, SMSPJ, ISRJ, MFTJ STFT CNN
al. more than 90%
Wang et Bispectral diagonal JSR> −7.5 dB, Recognition

124 2021 ISDJ, ISRJ, ISCJ, PDTJ, SMSPJ, WDTJ Resnet
al. slice Rate:80%
SNR: −5dB, Recognition Rate:

Ren et
125 2022 MFTJ, NFMJ, SNJ, DDJ+VDJ FRFT Resnet
al.
more than 97%.
18
STFT, Modulus, DFCN

Shao et Noise, ISRJ, AJ, BJ, SJ, RDJ, DFTJ, SNJ, CJ, JNR:30~60dB,
127 2020 Phase, Real, (1D-CNN,
al. CJ+ISRJ, RDJ+ Sweep, DFTJ+ SNJ Average Recognition Rate: 96%
Imaginary 2D-CNN, DNN)
Kong et STFT, RD domain MDFRCNN-AM JSR:3~5dB, Average

128 2022 DDJ, DFTJ, ISRJ, SNJ, VDJ, RVJDJ, SMSPJ
al. feature map (2D-CNN） Recognition Rate: 99%
Real,
Imaginary, ISR:10～50dB,
Frequency SNR: −10～10dB,
Cao et spectrum, Power Information fusion
129 2023 NAMJ, NFMJ, NPMJ, RDJ, ISDJ, ISRJ WideResNet28_2
al. spectrum, Time- hyperparameter d:13
Frequency Average Recognition Rate:
distribution 88.06%
Matrix
NAMJ, NPJ, CNJ, CSJ, SMSP, C&IJ, ISRJ, JNR= −9 ~ 6 dB,
Zhou et Mff-Net,
130 2023 MFTJ, FDA, CSJ+SMSP, CSJ+C&IJ, FRFT, STFT Average Recognition Rate: more
al. If-Net
CSJ+ISRJ, SMSP+ C&IJ than 90%
Note ： Deception jamming : interrupted-sampling circularly jamming (ISCJ), interrupted-sampling directly jamming (ISDJ), partial-pulse dense transmit
jamming (PDTJ), range false target jamming (RFTJ), velocity false target jamming (VFTJ), whole-pulse dense transmit jamming (WDTJ), interrupted-sampling
directly jamming (ISDJ), interrupted-sampling repeater jamming (ISRRJ), range false target modulation (RFT), velocity false target modulation (VFT), slicing
modulation (SL), full-pulse forwarding jamming (FPFJ), full-pulse dense forwarding jamming (FPDFJ), partial-pulse dense forwarding jamming (PPDFJ), range-
velocity joint deception jamming (RVJDJ), frequency diverse array (FDA). Suppression jamming: Audio jamming, narrow-band jamming (NJ), spread
spectrum jamming (SSJ), Single-frequency jamming (SFJ), PMJ 1MHz, FMJ 1MHz, FMJ 10MHz, SJ 5M, SJ 400M, amplitude modulation noise (AMN),
frequency modulation noise (FMN), phase modulation noise (PMN), sweeping frequency noise (SFN), AFN (AMN and FMN compound jamming), APN (AMN
and PMN compound jamming), ASN (AMN and SFN compound jamming), FPN (FMN and PMN compound), FSN (FMN and SFN compound jamming), PSN
(PMN and SFN compound jamming), Narrow pulse jamming, Blanket jamming, partial band jamming (PBJ), noise suppression jamming, Sweep jamming (SJ).
Smart jamming: smart noise jamming (SFNCJ or SNJ), noise product jamming (CPJ), convolutional smart noise modulation (CSN), product smart noise
modulation (PSN). Other: Chaff jamming (CJ), gray-scale compression transformation (GSCT), Clutter.
frequency images of jamming regions using STFT and lightweight and efficient multi-label compound jamming
established lightweight CNN models for recognizing six types recognition network (MLCJRN). The model combines the
of typical jamming signals. To optimize the balance between lightweight ShuffleNet v2, the well-designed efficient hybrid
prediction accuracy and computational cost, Wang et al. [119] attention (EHA) module based on efficient channel attention
developed a Reinforcement Learning-based Adaptive Forward (ECA), and the multi-label threshold optimization algorithm.
Propagation (RLAFP) method to dynamically adjust the depth Through the combination, the model achieves a remarkable
of computation. The advent of reinforcement learning [120] overall recognition accuracy of 92.25% for six types of single-
addresses policy optimization issues, and the trained model measured jamming data, specifically within the JNR range of
acquires autonomous learning capabilities. The method 10 to 30 dB. Zhou et al. [123] introduced a strategy that
comprises two components: a main network designed for involves separation before recognition, with the goal of
multiple outputs, which allows for the early stopping of test addressing the labeling limitations inherent in supervised
samples, and a policy network responsible for determining the learning methods. They separated composite jamming signals,
optimal depth of forward propagation based on input processed them through CNN for recognition, and determined
characteristics. Additionally, a time-frequency component the specific jamming types within composite jamming signals.
perception module was constructed, enabling the network to When the SNR was 0 dB, recognition accuracy of the model
compute only the locations with time-frequency components exceeded 90%. However, the approach only considered
within the input time-frequency images, while disregarding overlapping jamming at the same time-frequency point and
unimportant areas, thus significantly reducing redundant was susceptible to noise jamming.
computation. To address the issues such as increased model 4) Transform domain feature network
size and reduced performance in traditional multi-class In addition to the time domain, frequency domain, and
methods, Zhu et al. [121] introduced a deep multi-label based time-frequency domain, some discernible data features can
multiuser automatic modulation classification framework also be captured in the transform domain, and the recognition
(MLAMC) for compound signals recognition. They designed task can be completed through the classical neural network.
a threshold optimization algorithm to improve the Wang et al. [124] introduced a method based on residual
performance, which can identify 127 combined scenarios of neural networks. They extracted bispectral features from
seven types of radar jamming signals. However, the approach jamming signals and employed bispectral diagonal slices as
relied on training datasets, and recognition accuracy inputs to the network, thus achieving the recognition of
significantly decreased when faced with composite signals not spoofing jamming. When JSR exceeded − 7.5 dB, the
previously designated as labels. Further from the engineering recognition rate exceeded 80%. Ren et al. [125] introduced a
point of view, because of the complex network structure, the jamming recognition method based on fractional Fourier
multi-label network may not be suitable for embedded devices transform and residual convolutional neural networks, the
with limited computing power and memory. To apply results show that the recognition effect of multiple jamming
effectively to the resource-constrained airborne pulse signal signals under different jamming to signal ratio is higher than
real-time analysis instruments, Lv et al. [122] proposed a 90%, when SNR was −5 dB, the accuracy was more than 97%.
19
5) Multi-domain feature fusion network To address the issue of limited samples, numerous jamming
The above methods all involve feeding single-domain recognition methods have been proposed. Although there isn't
features into a deep learning model for recognition. However, a unified classification standard in the literature to define these
single-domain features contain limited information. Later, recognition approaches, we can categorize them based on the
many researchers utilized neural networks [126] to fuse multi- most commonly used methods in this paper. Specifically, we
domain features such as time domain, frequency domain, and primarily classify jamming recognition methods into two types:
time-frequency domain to further improve recognition Few-shot Learning (FSL) [131] and Zero-shot Learning (ZSL)
accuracy. [132][133]. Both of these methods focus on handling limited
Shao et al. [127] proposed a fusion model (DFCN) labeled data in the field of radar active jamming recognition,
composed of three subnetworks (namely, 1D-CNN, 2D-CNN, but their objectives and application methods differ. Few-shot
and fusion network). The 1D-CNN is used to extract deep learning (FSL) does not rely on large amounts of samples;
features from the raw radar jamming signals. To extract time- instead, it learns prior knowledge to quickly identify new tasks
frequency features, the jamming signals are subjected to STFT with only a few samples. It constitutes an N-way K-shot
to obtain time-frequency spectrograms. Deep time-frequency learning paradigm, where each task consists of N categories,
features are extracted using 2D-CNN, and the fusion network and each category has K labeled samples. By utilizing a small
deeply integrates the features extracted by CNN. However, it number of labeled samples, the model is trained to rapidly
does not consider the low JNR scenario. Considering that the adapt to new, previously unseen tasks. Since the concept of
doppler domain also contains rich feature information, Kong few-shot learning was introduced, many scholars have
et al. [128] proposed a hybrid neural network model that proposed various methods over the years, including Data
combines radar echo range-doppler domain and time- Augmentation (DA) [134], Transfer Learning (TL) [135],
frequency domain to recognize single-deception jamming and Metric Learning (ML)[136], and Meta-Learning (M-L)[137].
compound deception jamming under extended targets. This Although few-shot learning significantly mitigates the
hybrid model can coordinate multiple learning models from dependency of deep learning algorithms on large amounts of
different domains, incorporating attention mechanisms and annotated data, in many scenarios, there may still be cases
multi-domain feature fusion, thus enhancing the capability for where some sample data does not belong to any known
correct detection and recognition. Cao et al. [129] introduced a category or certain categories have no sample data at all. In
radar jamming intelligent recognition model based on multi- contrast, zero-shot learning (ZSL) [138] focuses on predicting
domain information fusion. They extracted key features such new categories that have no labeled samples at all. By
as real parts, imaginary parts, spectra, and power spectra, leveraging auxiliary information, zero-shot learning can
subsequently fused them into a new multi-domain information directly achieve classification of unseen categories. Table IX
fusion matrix. The matrix is then fed into the shows the papers on deep neural networks with restricted
WideResNet28_2 network, achieving recognition accuracy of sample.
88.06%. To realize the recognition of a variety of classical 1) Data Augmentation
jamming signals and compound signals, Zhou et al. [130] DA is considered an intuitive way to address the above-
designed multi-feature fusion network model (Mff-Net) and mentioned issues and has been proven to be beneficial for
information fusion network model (If-Net) based on FRFT training machine learning models in deep architectures [134],
images and time-frequency images, average recognition especially for small datasets. Recently, more advanced data
accuracy was more than 90% when JNR= −9 ~ 6 dB. augmentation methods for few-shot learning have been
researched, such as the methods based on generative
C. Restricted Sample Strategies
adversarial networks (GANs). GANs are a prominent solution
In recent years, research on radar jamming recognition has for modifying and reconstructing semantic information in
mostly been based on supervised deep learning methods. images, including tasks like image extension, image
These methods require a large amount of jamming samples, restoration, image style transfer, and high-definition image
and effective jamming recognition models with low generation. The models are the type of unsupervised learning
generalization errors can only be obtained when trained on network introduced by Goodfellow [139]. They generate
extensive jamming datasets. When the number of jamming credible samples through the adversarial relationship between
samples in the training dataset is limited, deep learning-based a generator and a discriminator, providing researchers with
radar jamming recognition models are prone to overfitting, different approaches to jamming recognition, as shown in Fig.
leading to larger generalization errors in test environments. 10. Yuan et al. [140] proposed the concept of intelligent
Consequently, the recognition accuracy of jamming jamming using deep GANs to generate high-fidelity signals,
recognition models in testing environments tends to be lower. laying the foundation for further research in cognitive
In real radar battlefield environments, it is challenging to electronic warfare. Since the radar fails to accurately
acquire a significant number of jamming samples, and recognize new types of jamming that were not included in the
manually labeling a large number of jamming samples is both training set upon detection, Li et al. [141] introduced one
time-consuming and labor-intensive. Therefore, deep learning- model called generative adversarial network pseudo-labeled
based radar jamming recognition models struggle to obtain convolutional neural network (GAN-PL-CNN). This model
sufficient sample sizes for practical applications, and the establishes datasets for known and unknown types of jamming
models often suffer from issues such as overfitting, limited and achieves semi-supervised radar jamming recognition
generalization ability, lack of robustness, and a lack of using an adaptive pseudo-labeling method. It intelligently
feedback mechanisms. increases labeled samples, and as the number of labeled
20
samples increases, the recognition performance improves, and 2) Transfer learning
the ability to exclude unknown types becomes stronger. Ding In recent years, TL has seen successful applications in
et al. [142] proposed an innovative fusion algorithm based on various fields, especially in domains with limited training
conditional generative adversarial network (CGAN) and samples, such as remote sensing. Transfer learning [135] is
convolutional neural network (CNN), when training samples considered a set of techniques aimed at importing useful
were 30, JNR within the range of −20 dB to 20 dB, achieving information learned from source data into target data,
recognition accuracy of 85%. significantly reducing the demand for training samples.
Achieving high accuracy with deep learning algorithms in
Real Data scenarios with small number of samples can be challenging.
Fig. 11 illustrates the learning process of transfer learning.
Random Noise Currently, transfer learning can effectively enhance model
generalization performance, leading researchers to explore
novel transfer learning-based approaches to achieve precise
Real
Fake Data recognition within a limited training dataset. Model-based
Fake transfer learning is one of the widely used transfer learning
Discriminator
methods. The core idea involves transferring pre-trained
network models (such as AlexNet [143], VGGNet [144],
Update parameter ResNet [145], etc.) from a source domain to a target domain.
Generator The method enables the recognition model in the target
domain to fully leverage the pre-trained model parameters.
Update parameter Typically, the pre-trained model serves as the initial values for
Fig.10. Generative adversarial network model. the target domain model or as a component of it. The transfer
TABLE IX
PAPERS ON DEEP NEURAL NETWORKS WITH RESTRICTED SAMPLE
Ding et CWJ_A, CWJ_W, AMJ, NBNJ, Train samples:30, JNR: −20~20dB,
142 2023 PSD CGAN
al. MTJ, LFMJ, PPNJ, NAMJ Recognition Rate: 85%
JNR: −8 dB, Recognition Rate: more than 97.5%;
Bilinear
Xiao et AMJ, FMJ, N-CNJ, R-CNJ,
JNR:−2 dB, Recognition Rate: close to100%
146 2021 SPWVD EfficientNet
al. SMSPJ, C&IJ, ISRJ
Network
ISRJ, DFTJ, SNJ, BJ, Spot
Hou et AlexNet, JNR:30dB~60dB, Train Samples: 1%, Recognition
147 2022 jamming, LSJ, DFTJ+SNJ, SPWVD
al. SqueezeNet Rate: 98.92%
ISRJ+LSJ
Model based on
Zhou et NAMJ, NFMJ, RFJ, RGPOJ,
148 2022 STFT OpenMax and JNR>8 dB, Recognition Rate:90%
al. VGPOJ, SMSPJ, C&IJ, ISRJ
confidence score
Luo et ISRJ, DDJ, DFTJ, SNJ, DDJ + JNR:30~60 dB, 3% of the samples were labeled,
149 2023 STFT JR-TFSAD
al. ISRJ, DDJ + SNJ, SNJ + DFTJ Recognition Rate:90%
ISRJ, C&I, DDJ, VDJ, SNJ,
Luo et DDJ+ISRJ, DDJ+SNJ, DDJ+VDJ, JNR:0~30 dB, training size: 5%,
150 2024 STFT JR-ACAR
al. VDJ+ISRJ, BJ (measured), DFTJ Recognition Rate: about 98.25%
(measured)
Zou et RGPOJ, CNJ, NPJ, SMSPJ, C&IJ, SNR: 0 ~ 10 dB, JSR: −5 ~ 15 dB, Accuracy for 50
151 2023 STFT MobileViT_CA
al. ISRJ, CSJ samples per class:96.58%
JNR:30~60 dB, 5% train samples,
ISRJ, DDJ, DFTJ, SNJ, DDJ + OA: about 95%, AA: about 93.74%, F1 score:
155 Lv et al. 2022 STFT WECNN-TL
SNJ, DFTJ + SNJ, DDJ+ISRJ about 93.54%, Recall: about 93.55%, Kappa: about
94.43%
JNR: −8 dB,
DDJ, VDJ, DFTJ, ISDJ, ISRJ,
Wu et
160 2023 ISLJ, DDJ+VDJ, VDJ+ISDJ, DWT SSWAN
al. Recognition Rate: more than 96.3%
DDJ+ISRJ, DDJ+ISDJ
Tian et Prototypical SNR: 6 dB,

161
al.
2021 RGPO, VGPO, RVGPO WT network Recognition Rate: more than 90%
RMT, VMT, RVMT, C&I, ISRJ, five-shot case,

Xiao et
162
al.
2024 SMSP, SNJ, RGPO, VGPO, STFT PSPNet Recognition Rate (simulation data): 97.02%,
RVGPO Recognition Rate (measured data): 98.49%
Note：Deception jamming: intermittent sampling loop jamming (ISLJ), range multifalse target (RMT), velocity multifalse target (VMT), range-VMT (RVMT).
Suppression jamming: continuous wave jamming by amplitude (CWJ_A), continuous wave jamming by frequency (CWJ_W), narrow band noise jamming
(NBNJ), periodic pulse noise jamming (PPNJ), linear sweeping jamming (LSJ), Spot jamming. Smart jamming: Gaussian white noise modulation convolution
jamming (N-CNJ), rectangle wave modulation convolution jamming (R-CNJ). Other: power spectral density (PSD).
21
learning approach circumvents the lengthy phase of the model Convolutional layers Fully connected layers
starting from scratch, thereby effectively reducing the learning flatten
costs for the target domain model, improving training speed,  
and enhancing training performance. Fine-tuning is a ImageNet
technique that utilizes pre-trained model parameters as initial
values and then trains a new model on top of them, it can
enhance the generalization capabilities of models, but it Transfer Learning
flatten
requires additional computational resources and time. Xiao et
al. [146] proposed a jamming recognition model based on  
New
Bilinear EfficientNet and attention mechanism, using images
EfficientNet-B3 and transfer learning. The method first
obtained the time-frequency image of jamming signals  convolution
through time-frequency analysis. Then, it automatically Fig.11. Learning process of transfer learning.
extracted deep features from the time-frequency images using
a bidirectional network with attention mechanisms. Ultimately, scenarios with small sample sizes [153] [154]. Lv et al. [155]
the approach achieved radar jamming recognition with an proposed a weighted ensemble CNN and transfer learning
overall recognition rate exceeding 97.5% at a JNR of −8 dB. (WECNN-TL), the model was developed to fully leverage the
Hou et al. [147] introduced an efficient method based on real part, imaginary part, magnitude, and phase in time-
transfer learning. They employed time-frequency analysis frequency spectrum data. However, the model required multi-
methods such as STFT, WVD, and SPWVD, and modified dimensional data to recognize jamming types, leading to a
pretrained networks like AlexNet and SqueezeNet. This complex network structure with many parameters.
approach not only had the ability to train models with small Additionally, training multiple models in the ensemble
sample sizes but also achieved higher accuracy while network could require significant time and computational
significantly reducing training time and memory usage. As resources, making it less suitable for practical device
new jamming signals continually emerge, and it's impossible deployment.
to anticipate them in advance, Zhou et al. [148] presented two 3) Metric Learning
models for radar jamming open set recognition (OSR) that can When training samples are limited, neural networks often
accurately classify known jamming and distinguish unknown suffer from overfitting, which can degrade the classification
jamming in the case of small samples. When the JNR was accuracy on test samples. To address this issue and effectively
greater than 8 dB, both models achieved high average handle classification tasks within the feature space of samples,
accuracy, with recognition rates for jamming signals reaching it is common in machine learning systems to measure the
up to 90%. However, this method could only detect or reject similarity between different samples using suitable metric
unknown signals, but could not effectively identify them. To methods. Consequently, metric learning [136] has become an
fully leverage time-frequency information from jamming important research direction in the field of machine learning.
signals, Luo et al. [149] proposed jamming recognition ML aims to make samples of the same class close to each
network called joint representation time-frequency self- other and samples of different classes far apart by modeling
attention distillation (JR-TFSAD). The network incorporated the distribution of distances between samples. Prototype
global information from time-frequency spectrum, achieving learning belongs to metric learning, it is a clustering algorithm
recognition rate of up to 90% for most jamming signals when derived from the k-nearest neighbor (KNN) classifier. Chen et
labeled samples accounted for only 3%. To mine the complete al. [156] introduced hollow convolutional prototype learning
information of the jamming signals for enhancing both with a new hybrid loss function, and they proposed a hybrid
accuracy and robustness, Luo et al. [150] proposed a few-shot attention and feature reuse net (HAFRNet), the method has
adaptive confidence aggregation and cross-modal refinement considerable classification accuracy of the known classes and
jamming recognition (JR-ACAR) method. When training size rejection performance of the unknown classes.
was 5%, the overall accuracy was about 98.25%. To reduce Currently, some of the most effective metric learning
the network parameter volume while maintaining high methods involve learning an embedding space based on neural
recognition accuracy, accelerating network inference speed, networks, where some metrics are used to compute the
and reducing hardware requirements, Zou et al. [151] similarity between samples. The methods are known as deep
proposed a recognition algorithm called MobileViT_CA that metric learning methods, and the Siamese networks are
combined the lightweight network (MobileViT) and popular examples of metric learning techniques commonly
coordinate attention, achieving 96.58% accuracy rate for 50 used in recognition tasks, as illustrated in Fig. 12. Koch et al.
samples per class. [157] proposed a small-sample recognition method using
In addition to the above approaches, to enhance model Siamese neural networks. It employed a Siamese structure to
predictive performance and robustness and achieve better calculate the similarity between input samples and trained a
generalization, some researchers have combined transfer convolutional network on the training set to measure the
learning with ensemble learning [152]. By integrating multiple similarity between two samples. Finally, it used a similarity
models in various ways, such as voting, averaging, and estimation network to calculate the similarity of test samples
boosting, to compensate for individual model limitations and for specific identification. Wang et al. [158] introduced a
obtain improved generalization performance. This approach model consisting of two CNNs with shared parameters. The
has been successfully applied to radar target recognition in model incorporated three loss functions: cross-entropy,
22
Kullback-Leibler (KL) divergence for feature extraction, and M-L involves training meta-learner models across various
Euclidean distance. These losses allowed the network to learn learning tasks, allowing them to solve new learning tasks with
relationships between samples to enhance recognition minimal training samples [137]. And it is primarily
accuracy. Shao et al. [159] presented CNN-based Siamese characterized by iterative learning across tasks, where meta-
network for radar jamming signal classification. It separated learners rapidly adapt in each task, and the feedback from the
the real and imaginary data of paired input signals, conducted tasks is used to adjust the learning strategy of the meta-learner
feature extraction using one-dimensional CNN with shared [163]. Specifically, in meta-learning, a learning algorithm
parameters, and used Manhattan distance to measure the known as the meta-learner is employed to learn task-specific
similarity between two samples in feature space. It determined learners. This learning algorithm is trained on a set of similar
the category of unknown jamming signals through similarity tasks, aiming to maximize the overall generalization ability of
probabilities. Wu et al. [160] proposed the Siamese Squeeze all task learners. The meta-learning process is illustrated in Fig.
Wavelet Attention Network (SSWAN), which utilized the 13. Liu et al. [164] introduced intelligent representations,
lightweight SqueezeNet as the network backbone and including SPWVD, FRFT, and constellation maps, as input for
combined a Siamese structure with deep squeezing and a model-agnostic meta-learning networks, and utilized federated
wavelet attention mechanism. The network achieved a learning to train distributed networks for jamming recognition.
recognition accuracy of over 96.3% when the JNR was −8 dB. When JNR was 0 dB, the recognition performance exceeded
Tian et al. [161] proposed a prototypical Siamese network 90%, resulting in enhanced generalization performance and
based on radar feature parameters for multiresolution jamming improved reliability.
recognition, and the experiment results demonstrated that the
proposed network can achieve a good recognition performance Meta learning
with a small number of samples. To make full use of the
information within jamming signals and further enhance the
performance of jamming recognition in few-shot scenarios, Learning Task1 New Task1
Xiao et al. [162] proposed a pretraining and self-supervised
fine-tuning-based prototypical network (PSPNet). PSPNet is Learning Task2 New Task2
an improved prototypical network that incorporates
pretraining and self-supervised fine-tuning, achieving high
accuracies of 97.02% and 98.49% respectively in five-shot Learning Task3 New Task3
settings using both simulated and measured data.
Learning Taskn New Taskn

Input 1
Extract Fig.13. Process of meta learning.

. .
feature1 . .
. . 5) Zero-Shot Learning
. .
The ability to infer and recognize new occurrences based on
Share Weight Similarity previously acquired knowledge was formally introduced as
ZSL [138]. ZSL does not create separate training datasets but
Input 2
focuses on learning a mapping from the data space to the
Extract Distance- semantic space [165]. It involves learning the mapping on
feature2 . .
. .
metric labeled source data with the hope that it can be transferred to
.
.
.
.
target data. During testing, when target data becomes available,
the model maps the data to a shared semantic space. Then, it
Fig.12. Learning process of Siamese networks. uses the predefined relationship between the semantic space
and the category space to map the semantic space data to the
Siamese networks provide a clear learning approach by category space for target recognition, as illustrated in Fig. 14.
organizing data in pairs and training a binary classifier to Due to its inferential capacity and the reduced need for
determine whether input samples belong to the same category. extensive manual labeling, ZSL is crucial for recognizing
However, Siamese networks still face some challenges. Firstly, unknown jamming signals. Zhang et al. [166] introduced the
the metric learning approach in Siamese networks is CNN-Zero framework, which learns representations in the
susceptible to noise. If there are outliers in the training data, semantic attribute space and distinguishes relationships
Siamese networks may still consider the impact of these noisy between region features. They constructed a hybrid loss
points on the predicted category of the sample to be predicted. function comprising attribute distance loss, cross-entropy loss,
Secondly, this training method is relatively inefficient, due to and reconstruction loss. With JNR ranging from 10 dB to 15
pairwise combinations, the size of the training data effectively dB, the average recognition rate for eight jamming types
increases by several orders of magnitude, potentially reached 88%, outperforming supervised neural networks.
introducing redundant information that can impact the final Later, Zhang et al. [167] applied the concept of ZSL to
performance. propose an open-world recognition method called the RCAE-
4) Meta-Learning OWR algorithm. During supervised training, they built a
residual convolutional autoencoder network structure to
23
extract semantic information solely from the known training transform points is 100. Moreover, the input images of the
set. In the unsupervised classification phase, the test set network are proportionally resized to 128×128.
containing both known and unknown signals are fed into the TABLE Ⅹ
trained encoder. They employed distance-based recognition DATA INFORMATION
methods to classify jamming signals. When JNR exceeded 10 Jamming Parameters Value range
dB, the recognition rates for seven known jamming signals RDJ False target delay 10μs
and two unknown jamming signals surpassed 92%. However, VDJ Radial velocity 100m/s
there are still limitations and drawbacks that need to be ADJ Angular deviation 2 degrees
addressed. Currently, the emphasis is primarily placed on pull-off period 2000m/s
common jamming signals, while other jamming signals exist. capture 2ms
Considering open-world recognition of radar jamming under RGPOJ pull-off 10ms
and cessation 2ms
class-imbalanced conditions will be a challenging yet valuable amplitude 1.5
task for future research. pull-off velolcity 0.01MHz/s
capture 2ms
VGPOJ pull-off 10ms
Source task Source task cessation 2ms
data space class space amplitude 1.5
capture 2ms
Semantic pull-off 10ms
RVGPOJ
cessation 2ms
Space amplitude 1.5
Target task Target task SMSPJ Number of sub-pulses 5
data space class space C&IJ
Number of pulses 2~4
Number of time slots 2~5
Number of forwards 1~4
Fig.14. Process of Zero-Shot learning. ISRJ Slice period 10μs
Jamming duty cycle 0.2,0.25,0.3,0.5
Number of false target 3~6
DFTJ
V. EXPERIMENTS AND RESULTS False target delay 10μs
AJ Jamming bandwidth 20MHZ
In this section, we provide a detailed introduction to the
BJ Jamming bandwidth 60MHZ
jamming dataset. Secondly, we describe the experimental
SJ Sweep cycle 40μs
settings of the classic methods and then define several
performance evaluation metrics. Finally, the experimental RFJ Gaussian white noise Mean:0, Variance:1
results are presented and analyzed. NAMJ Gaussian white noise Mean:0, Variance:1
Gaussian white noise Mean:0, Variance:1
NFMJ
A. Dataset Modulation coefficient 10
Gaussian white noise Mean:0，Variance:1
In this paper, we set the bandwidth of LFM signal B = 10 NPMJ
Modulation coefficient 1
MHz, time-width T = 20 µs, and sampling rate Fs =20 MHz. CSJ Number of comb 3~7
This paper simulates the deception jamming ， suppression Pulse amplitude 5
PJ
jamming, and smart jamming introduced in Section II and Initial phase 0
combines two kinds of jamming for experiments, the time- NPJ Gaussian white noise Mean:0, Variance:1
domain waveforms and time-frequency spectrograms of the CNJ Gaussian white noise Mean:0, Variance:1
above LFM and jamming signals are shown in Fig. 7 and 8, ISRJ+CSJ
Determined by specific jamming
C&IJ+SMSPJ
respectively. The parameters of the radar jamming generated
in this article are shown in Table Ⅹ, which means that each B. Experimental Settings
jamming signal is randomly generated within a specific During the training phase, Pytorch is employed as the
parameter range to ensure the diversity of the samples. Among platform, where an RTX 1080 Ti GPU is used to speed up the
these types of jamming, a variety of jamming types are network training. In the experiment, this article chooses
selected to construct a diverse dataset, including SMSPJ, C&IJ, overall accuracy (OA), F1 score (F1), Kappa, Flops, and
ISRJ, DFTJ, NAMJ, NFMJ, CSJ, NPJ, CNJ, combinations Params to quantitatively measure the performance of various
such as ISRJ+CSJ and C&IJ+SMSPJ. Each sample contains a methods in the radar jamming recognition. The main
true target echo with SNR = 0 dB and random location hyperparameters of the networks are set as follows: the
variation, and the JNR of the radar jamming signal is varied optimizer is SGD, the maximum number of training epochs is
from 0 to 10 dB. The dataset also includes nonjamming set to 100, the learning rate is 0.0001, and the batch size is 32.
samples, and the number of samples for each class is 500, so Random selected 80% labeled samples of each class constitute
the total number of samples is N = 6000. The STFT is training sets for model training, and the remaining samples
performed on each sample to obtain the time-frequency constitute the test sets.
domain dataset, in which the input time-domain sample is
divided into 32 segments with eight-point overlap between the C. Results and Analysis
adjacent two segments. The Hamming window is used as Table XI presents the OA, F1, Kappa, Flops, Params of the
window function φ(t) of STFT, and the number of Fourier SVM [64], RF [65], VGGNet-16 [95], ResNet50 [112], DFCN
24
TABLE XI
RECOGNITION RESULTS FOR JAMMING TYPE OF DIFFERENT METHODS
\ SVM[64] RF[65] VGGNet-16[95] ResNet50[112] DFCN[127] MobileViT_CA[151]

Nonjamming 98.00 100.00 97.00 100.00 99.00 100.00
SMSPJ 36.00 80.00 100.00 99.00 100.00 100.00
C&IJ 65.00 72.00 98.00 97.00 100.00 92.00
ISRJ 40.00 79.00 86.00 92.00 91.00 90.00
DFTJ 67.00 94.00 98.00 98.00 99.00 100.00
NAMJ 100.00 100.00 100.00 100.00 100.00 100.00
NFMJ 68.00 99.00 100.00 100.00 100.00 100.00
CSJ 56.00 95.00 72.00 59.00 82.00 93.00
NPJ 100.00 100.00 100.00 98.00 100.00 99.00
CNJ 100.00 100.00 100.00 100.00 100.00 100.00
ISRJ+CSJ 77.00 95.00 71.00 72.00 84.00 92.00
C&IJ+SMSPJ 40.00 70.00 99.00 96.00 98.00 100.00
OA(%) 70.58 90.33 93.42 92.58 96.08 97.17
F1(%) 70.57 90.29 93.41 92.57 96.07 97.16
Kappa(%) 67.91 89.45 92.82 91.91 95.73 96.91
Flops (GB) \ \ 5.06 1.35 0.43 0.45
Params(MB) \ \ 65.10 23.53 0.99 5.07
(a) (b) (c)
(d) (e) (f)

Fig.15.Confusion matrix obtained for each method. (a) SVM[64]. (b) RF[65]. (c) VGGNet-16[95]. (d) ResNet50[112].
(e) DFCN[127]. (f) MobileViT_CA[151].
[127], and MobileViT_CA [151]. The SVM algorithm uses a than others, and the feature extraction process is severely
radial basis function (RBF) as kernel function, and the RF time-consuming in RF and SVM algorithms based on manual
algorithm contains 100 decision trees. The extracted features feature extraction. The OA of the MobileViT_CA is 97.17%,
include time domain mean, time domain variance, box which is higher than SVM, RF, VGGNet-16, ResNet50, and
dimension, approximate entropy, fractional low-order moment, DFCN. The Flops and parameters of the DFCN are only 0.43
frequency domain skewness, and frequency domain kurtosis. GB and 0.99 MB. The Flops and Params of the DFCN are
In addition, other methods recognize jamming through time- lower than other methods. In addition, the OAs of the NAMJ
frequency spectrograms. The experimental results in Table XI and CNJ are 100% in several methods as well, which reveals
show that the performance of the recognition methods. that the time domain, frequency domain, and time-frequency
As shown in Table XI, the neural network-based radar domain features are effective to recognize jamming.
jamming recognition methods have significant advantages To provide a more intuitive demonstration of the model's
over the traditional machine learning-based methods in OA performance for different jamming types, the column
with the JNR is from 0 dB to 10 dB and SNR is 0dB. normalized confusion matrixis presented in Fig.15 to analyze
Specifically, the OA of the SVM is 70.58%, which is lower the proportion of samples predicted to be positive that are
25
positive. The darker the color of the diagonal line in the smart jamming. Radar jamming recognition methods are
confusion matrix, the higher the percentage of correct categorized into two groups: the methods based on traditional
recognition. In the confusion matrix, SVM is represented by machine learning and the methods based on neural network.
lighter colors, indicating lower recognition accuracy, while Traditional machine learning methods essentially rely on
MobileViT_CA achieves the highest recognition accuracy. manually extracting features that are subsequently fed into the
corresponding classifiers, but the approach has drawbacks
Ⅵ. DISCUSSION such as heavy reliance on prior expert knowledge, difficulty in
feature extraction, and low recognition accuracy.
In this paper, we provided a comprehensive review of the
Subsequently, neural network-based methods are introduced
current development of radar jamming recognition. We started
for jamming recognition. The methods treat the neural
by building the framework of radar jamming model, including
network model structure as the backbone, adjusting training
deception jamming, suppression jamming, and smart jamming.
parameters to automatically extract features. To enhance
This framework provides a novel perspective to classify the
recognition accuracy, especially in situations where training
radar jamming in various jamming types. Our exploration
samples are scarce, the neural network-based methods
revealed a notable trend: the increasing complexity of
incorporate various techniques such as data augmentation,
jamming signals. Subsequently, latest developments in the
transfer learning, and metric learning. However, the method
recognition methods based on traditional machine learning are
based on neural network also have some limitations. Firstly,
presented, including feature extraction and classifier design.
the performance of neural network models heavily depends on
After that, the emergence of neural network-based approaches,
the quality and quantity of training data. Obtaining labeled
particularly shallow and deep neural networks, has
samples from the same distribution as jamming signals in real
revolutionized the field, offering unprecedented accuracy and
scenarios is often challenging, making it costly to train high-
efficiency. Furthermore, the strategies developed to address
performance models. Secondly, the internal mechanisms of
the challenge of restricted samples are crucial in real-world
models are difficult to interpret. The neural network models
applications where labeled data is limited. The integration of
typically consist of millions of parameters with complex
the advanced techniques holds significant potential for
interrelationships, making it challenging to explain how the
enhancing radar jamming recognition systems, paving the way
model makes decisions. Therefore, while machine learning
for more robust and intelligent solutions in the future.
methods have made significant progress in jamming
Our findings provide new insights in radar jamming
recognition, further research is needed to overcome the
recognition. Specifically, when confronted with complex
limitations and challenges.
jamming signals, manually extracting features becomes
The increasingly complex electromagnetic environment
challenging. Traditional machine learning methods tend to
poses higher demands on jamming recognition methods.
perform well only when information is abundant. However, in
However, based on the relevant literature over the past 30
scenarios where sample information is limited, neural
years, there are significant limitations in existing research.
network-based recognition methods have proven to be almost
Future directions for research in radar jamming recognition
equally effective and significantly superior to traditional
can consider the following:
machine learning-based methods. Nevertheless, it's important
• Extract representative features for new types of jamming
to note that neural networks typically comprise numerous
parameters and computational units, resulting in considerable signals to enhance recognition capabilities. With the
computational resource consumption during training and emergence of new, highly realistic jamming signals, feature
inference. Additionally, the interpretability is limited, and extraction is often done using empirical or heuristic methods.
generalization capabilities require enhancement. Therefore, it The most important features are sometimes challenging to
is essential to comprehensively consider the factors and select capture, and the extracted features may be subjective and
an appropriate recognition method. speculative, often focusing on specific types of jamming.
One of the limitations of this work is the insufficient Further research is needed on how to select representative and
comprehensive analysis of the dataset. Although the radar discriminative features and choose the most effective feature
jamming signals are represented, more data around the world sets. As new jamming types continue to appear in future
would be better for diversity. Several studies have claimed to electronic warfare scenarios, studying the features of the new
provide measured jamming data with a high enough resolution types of jamming will be crucial to enhance radar's resistance
to facilitate recognition. As we gain access to more open- to jamming.
source and precise datasets, we will prioritize comparing the • Expanding the modeling of signals under realistic
performance of modern recognition methods to ensure conditions is essential to improve anti-jamming technology
accurate and reliable radar jamming recognition. This research and development. Many simulations simplify the
continuous effort will ensure a more accurate and reliable actual radar systems, considering only jamming signals and
analysis of radar jamming recognition. noise. They often do not account for the complex overlap of
multiple jamming signals in time, space, and frequency. In
Ⅶ. CONCLUSION AND PROSPECTS practice, jamming signals, noise, and target signals are often
mixed and challenging to distinguish. Future research should
This paper provides a systematic review of research consider more realistic radar receiving systems or analyze
achievements in the field of radar jamming recognition. The real-world field data to establish a comprehensive research
study analyzes the generation mechanisms and mathematical framework that closely resembles actual electronic warfare
models of deception jamming, suppression jamming, and
26
environments. The method will foster the advancement of Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-20,
2024.
anti-jamming technology.
[12] C. Liu, R. Wu, Z. He, X. Zhao, H. Li, and P. Wang, “Modeling and
• Improving the ability to recognize unknown jamming analyzing electromagnetic interference signal in complex battlefield
signals is essential. Most neural network-based recognition environments,” in Proceedings 2015 International Conference on
methods make a closed-world assumption, assuming that the Communications, Signal Processing, and Systems, 2016, pp. 351-361.
[13] M. A. Ali, H. Ahmed, R. Saadia, W. Ahmed, and F. Tahir, “Electronic
types of jamming signals in the testing phase are present in the deception jamming: False target generation in radars,” in Proceedings
training phase. When new types of jamming signals occur, International Conference on Recent Advances in Electrical Engineering
existing methods may incorrectly recognize them as one of the & Computer Sciences (RAEE & CS), 2022, pp. 1-6.
known types. To tackle the problem, research could focus on [14] H. Bang, W. Wang, S. Zhang, and Y. Liao, “FDA-based space-time-
frequency deceptive jamming against SAR imaging,” IEEE
data augmentation, transfer learning, zero-shot learning, and Transactions on Aerospace and Electronic Systems, vol. 58, no. 3, pp.
various other approaches. However, challenges such as 2127-2140, 2021.
insufficient intelligence, low levels of autonomy, and a lack of [15] Q. Wen, “An overview of the study of the complexity of the complex
versatility might surface. It is urgent to find methods to electromagnetic environments,” in Proceedings 8th International
Symposium on Computational Intelligence and Design (ISCID), vol. 1,
efficiently train network models under conditions of limited 2015, pp. 245-250.
labeled samples. In the future, the effective integration of [16] W. Liu, J. Meng, and L. Zhou, “Impact analysis of DRFM-based active
traditional machine learning and neural network-based jamming to radar detection efficiency,” The Journal of Engineering,
approaches into radar jamming recognition holds potential to vol. 2019, no. 20, pp. 6856-6858, 2019.
[17] D. Cheng, Z. Liu, Z. Guo, G. Shu, and N. Li, “A Repeater-Type SAR
augment the intelligence and autonomy of machine learning Deceptive Jamming Method Based on Joint Encoding of Amplitude and
systems, subsequently propelling radar countermeasure Phase in the Intra-Pulse and Inter-Pulse,” Remote Sensing, vol. 14, no.
technology towards enhanced intelligence and autonomy. 18, pp. 4597, 2022.
[18] N. Li, D. Cheng, P Lu, G. Shu, and Z. Guo, “Smart jamming against
SAR based on non-linear frequency-modulated signal,” IEEE
ACKNOWLEDGEMENT Transactions on Aerospace and Electronic Systems, vol. 59, no. 4, pp.
The authors would like to thank both the anonymous 3588-3605, 2022.
[19] Z. Liu, D. Cheng, N. Li, L. Min, and Z. Guo, “Two-Dimensional
reviewers and the editors for the valuable comments and Precise Controllable Smart Jamming Against SAR via Phase Errors
suggestions that helped in improving the clarity of this article. Modulation of Transmitted Signal,” IEEE Geoscience and Remote
Sensing Letters, vol. 21, pp. 1-5, 2024.
[20] D. C. Schleher, “Electronic warfare in the information age,” Norwood,
REFERENCES MA: Artech House, 1999.
[1] N. Li and Y. Zhang, “A survey of radar ECM and ECCM,” IEEE [21] H. Liu, B. Zhao, L. Huang, and L. Zhang, “Single Channel SAR
Transactions on Aerospace and Electronic Systems, vol. 31, no. 3, pp. Deception Jamming Recognition Based on Feature Map Learning,” in
1110-1120, 1995. Proceedings 5th International Conference on Information
[2] J. Yan, H. Liu, B. Jiu, Z. Liu, and Z. Bao, “Joint detection and tracking Communication and Signal Processing (ICICSP), 2022, pp. 662-668.
processing algorithm for target tracking in multiple radar system,” [22] L. Zhang, J. Ren, and T. Li, “Time-varying jamming modeling and
IEEE Sensors Journal, vol. 15, no. 11, pp. 6534–6541, Nov. 2015. classification,” IEEE transactions on signal processing, vol. 60, no. 7,
[3] L. Wang, W. Zhu, Y. Zhang, Q. Liao, and J. Tang, “Multi-target pp. 3902-3907, Jul. 2012.
detection and adaptive waveform design for cognitive MIMO radar,” [23] X. Xu, Y. Bao, and H. Zhou, “Technology of smart noise jamming
IEEE Sensors Journal, vol. 18, no. 24, pp. 9962–9970, Dec. 2018. based on convolution modulation,” Modern Radar, 2007.
[4] Y. Chen, Y. Wang, F. Qu, and W. Li, “A graph-based track-before- [24] Y. Zhang, “Technology of smart noise jamming based on multiplication
detect algorithm for automotive radar target detection,” IEEE Sensors modulation.” in Proceedings International Conference on Electric
Journal, vol. 21, no. 5, pp. 6587–6599, Mar. 2021. Information and Control Engineering, 2011, pp. 4557-4559.
[5] Q. Li, L. Zhang, Y. Zhou, S. Zhao, N. Liu, and J. Zhang, [25] W. D. Blair and M. Brandt-Pearce, “Discrimination of target and
“Discrimination of active false targets based on Hermitian distance for RGPO echoes using frequency diversity,” in Proceedings the Twenty-
distributed multiple-radar architectures,” IEEE Access, vol. 7, pp. Ninth Southeastern Symposium on System Theory, 1997.
71872–71883,2019. [26] F. Chen, R. Li, L. Ding, L. Liu, L. Dai, and G. Deng, “A method
[6] H. Yu, N. Liu, L. Zhang, Q. Li, J. Zhang, S. Tang, and S. Zhao, “An against DRFM dense false target jamming based on jamming
interference suppression method for multistatic radar based on noise recognization,” in Proceedings IET International Radar Conference,
subspace projection,” IEEE Sensors Journal, vol. 20, no. 15, pp. 8797– 2015, pp. 1-4.
8805, Aug. 2020. [27] P. C. J. Hill and V. Truffert, “Statistical processing techniques for
[7] M. Tan, C. Wang, B. Xue, and J. Xu, “A novel deceptive jamming detecting DRFM repeat-jam radar signals,” in Proceedings IEE
approach against frequency diverse array radar,” IEEE Sensors Journal, Colloquium on Signal Processing Techniques for Electronic Warfare,
vol. 21, no. 6, pp. 8323–8332, Mar. 2021. 1992, pp. 1-6.
[8] Q. Sun, T. Shu, M. Tang, K. Yu, and W. Yu, “Effective moving target [28] S. D. Berger, “Digital radio frequency memory linear range gate stealer
deception jamming against multichannel SAR-GMTI based on multiple spectrum,” IEEE Transactions on Aerospace and Electronic Systems,
jammers,” IEEE Geoscience and Remote Sensing Letters, vol. 17, no. 3, vol. 39, no. 2, pp. 725-735, 2003.
pp. 441–445, Mar. 2020. [29] M. Greco, F. Gini, A. Farina, and V. Ravenni, “Effect of phase and
[9] D. Ben and Z. Yan, “Study on active jamming method for ka-band range gate pull-off delay quantisation on jammer signal,” IEE
radar seeker,” in Proceedings IEEE 5th Advanced Information Proceedings Radar, Sonar Navigation, vol. 153, no. 5, pp. 454-459,
Technology, Electronic and Automation Control Conference (IAEAC), 2006.
2021, pp. 1403-1407. [30] M. Greco, F. Gini, and A. Farina, “Radar detection and classification of
[10] Y. Liu, T. Li, and Z. Gu, “Research on SAR active deception jamming jamming signals belonging to a cone class,” IEEE transactions on
scenario generation technique,” in Proceedings Fifth International signal processing, vol. 56, no. 5, pp. 1984-1993, 2008.
Conference on Instrumentation and Measurement, Computer, [31] Y. Xiong, M. Liu, B. Tang, and M. Sun, “Feature-extracting method for
Communication and Control (IMCCC), 2015, pp. 152-156. radar compound jamming signal based on box dimension and LZ
[11] N. Li and X. Hu, “Ultrawideband Mutual RFI Mitigation Between SAR complexity,” Journal of Data Acquisition & Processing, vol. 23, no. 6,
Satellites: From the Perspective of European Sentinel-1A,” IEEE pp. 663-667, 2008.
27
[32] J. Li, Q. Shen, and H. Yan, “Signal feature analysis and experimental convolutional neural networks,” in 25th European Signal Processing
verification of radar deception jamming,” in Proceedings IEEE CIE Conference (EUSIPCO), 2017, pp. 326-330.
International Conference on Radar, 2011, pp. 230-233. [53] W. Si, J. Li, P. Yuan and Y Li, “Study on time-frequency characteristic
[33] X. Tian, B. Tang, and G. Gui, “Product spectrum matrix feature of PD Pulse using Wigner-Ville Distributions,” Conference Record of
extraction and recognition of radar deception jamming,” International the 2008 IEEE International Symposium on Electrical Insulation,
Journal of Electronics, vol. 100, no. 12, pp. 1621-1629, 2013. Vancouver, BC, Canada, 2008, pp. 355-357.
[34] C. Du, Y. Zhao, L. Wang, B. Tang, and Y. Xiong, “Deceptive multiple [54] J. Yao, “Complete Gabor transformation for signal representation,”
false targets jamming recognition for linear frequency modulation IEEE Transactions on Image Processing, vol. 2, no. 2, pp. 152-159,
radars,” The Journal of Engineering, vol. 2019, no. 21, pp. 7690-7694, April 1993.
2019. [55] J. Jeong, W. J. Williams, “Kernel design for reduced interference
[35] J. Xu, S. Ying, and H. Li, “Gps interference signal recognition based on distributions,” IEEE Transactions on Signal Processing, vol. 40 no. 2,
machine learning,” Mobile Networks and Applications, vol. 25, no. 6, pp. 402–412. 1992.
pp. 2336-2350, 2020. [56] H. Lui, and NV. Shuley, “Joint time-frequency analysis of ultra
[36] C. Xu, L. Yu, Y. Wei, and P. Tong, “Research on active jamming wideband radar signals,” IEEE International Conf. on Signal Processing
recognition in complex electromagnetic environment,” in Proceedings and Communication Systems. 2007, pp 1-8.
IEEE International Conference on Signal, Information and Data [57] J. Li, B. Tang, and Q. Lv, ‘‘Bispectrum feature extraction used in
Processing (ICSIDP), 2019, pp. 1-5. deceptive jamming modes recognition,’’ Electronic Science and
[37] Y. Wei, Y. Li, and J. Zhang, “Radar jamming recognition method based Technology University, vol. 38, no. 3, pp. 329–332, Mar. 2009.
on fuzzy clustering decision tree,” in Proceedings IEEE International [58] H. Zhou, C. Dong, R. Wu, X. Xu and Z. Guo, “Feature fusion based on
Conference on Signal, Information and Data Processing (ICSIDP), Bayesian decision theory for radar deception jamming recognition,”
2019, pp. 1-5. IEEE Access, vol. 9, pp. 16296-16304, 2021.
[38] J. Chen, Y. Liu, and H. Song, “Active blanket jamming identification [59] D. H. Bailey and P. N. Swarztrauber, “The fractional Fourier transform
method based on rough set and decision tree,” in Proceedings 21st and applications,” SIAM Review, vol. 33, no. 3, pp. 389-404, 1991.
International Symposium on Distributed Computing and Applications [60] H. M. Ozaktas, O. Arikan, M. A. Kutay and G. Bozdagt, “Digital
for Business Engineering and Science (DCABES), 2022, pp. 141-145. computation of the fractional Fourier transform,” IEEE Transactions on
[39] Y. Jiang, M. He, C. Yu, and B. Wang, “A novel method of interrupted- Signal Processing, vol. 44, no. 9, pp. 2141-2150, 1996.
sampling repeater jamming recognition based on correlation [61] J. Lin and M. Gao, “Research on Active Jamming Recognition Method
dimension,” Fire and Command and Control, vol. 41, no. 7, pp. 152- Based on Fractional Fourier Transform,” in Proceedings IEEE 6th
156, 2016. Advanced Information Technology, Electronic and Automation Control
[40] Y. Jiang, M. He, C. Yu, and B. Wang, “Method of Interrupted-sampling Conference (IAEAC), 2022, pp. 137-143.
repeater jamming recognition based on box dimension,” Modern [62] H. Zhou, Z. Wang, R. Wu, X. Xu, and Z. Guo, “Jamming Recognition
Defense Technology, vol. 44, no. 2, pp. 157-164, 2016. Algorithm Based on Variational Mode Decomposition,” IEEE Sensors
[41] X. Miao, “Research on Radar Active Jamming Signal Identification Journal, 2023.
Technology,” Master’s thesis, Xidian University, 2023. [63] Y. Li, M. Dong, and R. Kothari, “Classifiability-based omnivariate
[42] Y. Shi, X. Lu, Y. Niu, and Y. Li, “Efficient jamming identification in decision trees,” IEEE Transactions on Neural Networks, vol. 16, no. 6,
wireless communication: Using small sample data driven naive bayes pp. 1547-1560, 2005.
classifier,” IEEE Wireless Communications Letters, vol. 10, no. 7, pp. [64] Q. Zhao and J. C. Principe, “Support vector machines for SAR
1375-1379, 2021. automatic target recognition,” IEEE Transactions on Aerospace and
[43] C. Wu and B. Chen, “A recognition algorithm of VGPO jamming based Electronic Systems, vol. 37, no. 2, pp. 643-654, 2001.
on discrete chirp-Fourier transform,” EURASIP Journal on Advances in [65] A. Paul, D. P. Mukherjee, P. Das, A. Gangopadhyay, A. R. Chintha,
Signal Processing, vol. 2020, no. 37, pp. 1-18, 2020. and S. Kundu, “Improved random forest for classification,” IEEE
[44] X. Yang, H. Ruan, and H. Feng, “A recognition algorithm of deception Transactions on Image Processing, vol. 27, no. 8, pp. 4012-4024, 2018.
jamming based on image of time-frequency distribution,” in [66] P. Shiguihara, A. D. A. Lopes, and D. Mauricio, “Dynamic Bayesian
Proceedings 7th IEEE International Conference on Electronics network modeling, learning, and inference: a survey,” IEEE Access,
Information and Emergency Communication (ICEIEC), 2017, pp. 275- vol. 9, pp. 117639-117648, 2021.
278. [67] H. Zhang, X. Ma, and J. Liu, “Radar Signal Sorting System Based on
[45] Z. Hao, W. Yu, and W. Chen, “Recognition method of dense false Multi-domain Feature Extraction and Decision Tree Algorithm,” in
targets jamming based on time ‐ frequency atomic decomposition, ” Proceedings IEEE International Conference on Artificial Intelligence
The Journal of Engineering, vol. 2019, no. 20, pp. 6354-6358, 2019. and Computer Applications (ICAICA), 2021, pp. 1214-1220.
[46] P. Sun, J. Yu, and W. Hao, “Research on Radar Active Jamming [68] Z. Li, X. Hao, H. Chen, and P. Li, “Target signal recognition for CW
Recognition Based on 2-D Time-Frequency Features,” in Proceedings Doppler proximity radio detector based on SVM,” in Proceedings
3rd International Academic Exchange Conference on Science and International Conference on Mechatronic Sciences, Electric
Technology Innovation (IAECST), 2021, pp. 777-781. Engineering and Computer (MEC), 2013, pp. 1160-1163.
[47] T. Thayaparan, L. Stankovic, M. Amin, V. Chen, L. Cohen, and B. [69] G. Wang, Q. Ren, and Y. Su, “The interference classification and
Boashash, “Editorial Time-frequency approach to radar detection, recognition based on SF-SVM algorithm,” in Proceedings IEEE 9th
imaging, and classification,” IET Signal Processing, vol. 4, no. 3, pp. International Conference on Communication Software and Networks
197-200, 2010. (ICCSN), 2017, pp. 835-841.
[48] J. Yang, Z. Bai, J. Hu, Y. Yang, Z. Xian, X. Hao, and K. Kwak, “Time- [70] G. Wang, Y. Kang, H. Wang, J. Sun, J. Wang, and Z. Zhang,
frequency Analysis and Convolutional Neural Network Based Fuze “Communication jamming signal recognition method based on fractal
Jamming Signal Recognition,” in Proceedings 25th International dimension and information entropy,” in Proceedings 4th International
Conference on Advanced Communication Technology (ICACT), 2023, Conference on Machine Learning, Big Data and Business Intelligence
pp. 277-282. (MLBDBI), 2022, pp. 116-123.
[49] R. Peng, P. Dong, and C. Meng, “Deceptive jamming recognition based [71] M. Xin and Z. Cai, “A feature fusion-based communication jamming
on wavelet entropy and RBF neural network,” Journal of Physics: recognition method,” Wireless Networks, vol. 29, pp. 2993-3004, 2023.
Conference Series, IOP Publishing, vol. 1486, no. 2, pp. 022003, 2020. [72] B. Krawczyk, M. Woźniak, and G. Schaefer, “Cost-sensitive decision
tree ensembles for effective imbalanced classification,” Applied Soft
[50] Z. Geng, H. Yan, J. Zhang, and D. Zhu, “Deep-learning for radar: A
Computing, vol. 14, pp. 554-562, 2014.
survey,” IEEE Access, vol. 9, pp. 141800-141818, 2021.
[73] M. P. Hosseini, A. Hajisami, and D. Pompili, “Real-time epileptic
[51] M. Sha, D. Wang, F. Meng, W. Wang, Y. Han, “Diff-SwinT: An
seizure detection from EEG signals via random subspace ensemble
Integrated Framework of Diffusion Model and Swin Transformer for
learning,” in Proceedings IEEE International Conference on Autonomic
Radar Jamming Recognition”, Future Internet ,vol. 15, no.12, pp. 374,
Computing (ICAC), 2016, pp. 209-218.
2023.
[74] L. Breiman, “Random Forests”, Machine learning, vol. 45, no. 1, pp. 5-
[52] J. Brynolfsson and M. Sandsten, “Classification of one-dimensional
32, 2001.
non-stationary signals using the Wigner-Ville distribution in
28
[75] C. Du and B. Tang, “Novel unconventional-active-jamming recognition International Conference on Electronic Information Technology and
method for wideband radars based on visibility graphs,” Sensors, vol. Computer Engineering, 2021, pp. 449-454.
19, no. 10, pp. 2344, 2019. [97] Z. Wu, Y. Zhao, Z. Yin, and H. Luo, “Jamming signals classification
[76] G. E. Smith, K. Woodbridge, and C. J. Baker, “Naïve Bayesian radar using convolutional neural network,” in Proceedings IEEE International
micro-Doppler recognition,” in Proceedings International Conference Symposium on Signal Processing and Information Technology
on Radar, 2008, pp. 111-116. (ISSPIT), 2017, pp. 062-067.
[77] H. Yamaguchi, T. Osafune, M. Tanaka, and H. Okuda, “Non-coherent [98] J. Lin and X. Fan, “Radar active jamming recognition based on
radar signal detection based on Bayesian theory,” in Proceedings recurrence plot and convolutional neural network,” in Proceedings
International Conference on Radar, 2008, pp. 586-591. IEEE 4th Advanced Information Management, Communicates,
[78] Z. Li and W. Song, “Entropy Theory Method for Active Jamming Electronic and Automation Control Conference (IMCEC), 2021, pp.
Perception,” Signal Process, vol. 12, pp. 1652-1656, 2017. 1511-1515.
[79] J. Dai, X. Hao, Z. Li, P. Li and X. Yan, “Adaptive target and jamming [99] Y. Meng, L. Yu, and Y. Wei, “Multiple information cognition of
recognition for the pulse Doppler radar fuze based on a time-frequency interrupted sampling repeater jamming in complex scenes,” in
joint feature and an online-updated naive Bayesian classifier with Proceedings CIE International Conference on Radar (Radar), 2021, pp.
minimal risk,” Defence Technology, vol. 18, no. 3, pp. 457-466, 2022. 2176-2179.
[80] M. Torma, “Kohonen self-organizing feature map and its use in [100] H. Zhang, L. Yu, Y. Chen, and Y. Wei, “Fast complex-valued CNN for
clustering,” in Proceedings ISPRS Commission III Symposium: Spatial radar jamming signal recognition,” Remote Sensing, vol. 13, no. 15, pp.
Information from Digital Photogrammetry and Computer Vision, 1994, 2867, 2021.
vol. 2357, pp. 830-835. [101] Y. Meng, L. Yu, and Y. Wei, “Multi-Label Radar Compound Jamming
[81] S. Zhang, Y. Tao, Y. Zhao, and Y. Chen, “Research on radar jamming Signal Recognition Using Complex-Valued CNN with Jamming Class
evaluation method based on BP neural network,” in Proceedings 2nd Representation Fusion,” Remote Sensing, vol. 15, no. 21, pp. 5180,
International Conference on Electrical and Electronic Engineering 2023.
(EEE 2019), 2019, pp. 300-305. [102] J. Li, A. Mohamed, G. Zweig, and Y. Gong, “LSTM time and
[82] X. Qiang, Z. Wei-gang, and B. Yuan, “Jamming style selection for frequency recurrence for automatic speech recognition,” in Proceedings
small sample radar jamming rule base,” in Proceedings IEEE IEEE Workshop on Automatic Speech Recognition and Understanding
International Conference on Signal Processing, Communications and (ASRU), 2015, pp. 187-191.
Computing (ICSPCC), 2018, pp. 1-5. [103] A. Pourranjbar, G. Kaddoum, and W. Saad, “Jamming pattern
[83] X. Ma, J. Qin, and J. Li, “Pattern recognition-based method for radar recognition over multi-channel networks: A deep learning approach,” in
anti-deceptive jamming,” Journal of Systems Engineering and Proceedings 55th Asilomar Conference on Signals, Systems, and
Electronics, vol. 16, no. 4, pp. 802-805, 2005. Computers, 2021, pp. 305-308.
[84] F. Ruo-Ran, “Compound jamming signal recognition based on neural [104] Y. Qin, J. Yang, M. Zhu, Y. Li, “Fast recognition of pull-off jamming
networks,” in Proceedings 6th International Conference on using LSTM,” in Proceedings IEEE International Conference on Signal,
Instrumentation & Measurement, Computer, Communication and Information and Data Processing (ICSIDP), 2019, pp. 1-5.
Control (IMCCC), 2016, pp. 737-740. [105] Q. Qu, S. Wei, S. Liu, J. Liang, and J. Shi, “JRNet: Jamming
[85] J. Schmidhuber, “Deep learning in neural networks: An overview,” recognition networks for radar compound suppression jamming
Neural networks, vol. 61, pp. 85-117, 2015. signals,” IEEE Transactions on Vehicular Technology, vol. 69, no. 12,
[86] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification pp. 15035-15045, 2020.
with deep convolutional neural networks,” in Proceedings [106] Z. Dang, “Research on the Technology of Wireless Communication
Communications of the ACM, 2012, pp.84–90. Interference Signal Identification and Processing Based on Deep
[87] W. Feng, N. Guan, Y. Li, X. Zhang, and Z. Luo, “Audio visual speech Learning,” Master’s thesis, Electronic Science and Technology
recognition with multimodal recurrent neural networks,” in Proceedings University, 2023.
International Joint Conference on Neural Networks (IJCNN), 2017, pp. [107] T. Kuang, H. Chen, L. Han, R. He, W. Wang, and G. Ding, “Abnormal
681-688. signal recognition with time-frequency spectrogram: a deep learning
[88] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Region-based approach,” 2022, arXiv:2205.15001.
convolutional networks for accurate object detection and semantic [108] Q. Liu and W. Zhang, “Deep learning and recognition of radar jamming
segmentation,” IEEE Transactions on Pattern Analysis and Machine based on CNN,” in Proceedings 12th international symposium on
Intelligence, vol. 38, no. 1, pp. 142-158, 2016. computational intelligence and design (ISCID), 2019, pp. 208-212.
[89] N. Zhang, Y. Li, Y. Shi, and J. Shen. “A CNN-Based Adaptive [109] Y. Cai, K. Shi, F. Song, Y. Xu, X. Wang, and H. Luan, “Jamming
Federated Learning Approach for Communication Jamming pattern recognition using spectrum waterfall: A deep learning method,”
Recognition,” Electronics, vol. 12, no. 16, pp. 3425, 2023. in Proceedings IEEE 5th International Conference on Computer and
[90] X. Tang, X. Zhang, J. Shi, S. Wei, and L. Yu, “SAR deception jamming Communications (ICCC), 2019, pp. 2113-2117.
target recognition based on the shadow feature,” in Proceedings 25th [110] Y. Wang, B. Sun, and N. Wang, “ Recognition of radar active ‐
European Signal Processing Conference (EUSIPCO), 2017, pp. 2491- jamming through convolutional neural networks, ” The Journal of
2495. Engineering, vol. 2019, no. 21, pp. 7695-7697, 2019.
[91] C. Shi, Q. Zhang, T. Lin, Z. Liu, and S. Li, “Recognition of Micro- [111] X. Lan, T. Wan, K. Jiang, Y. Xiong, and B. Tang, “Intelligent
Motion Jamming Based on Complex-Valued Convolutional Neural recognition of chirp radar deceptive jamming based on multi-pulse
Network,” Sensors, vol. 23, no. 3, pp. 1118, 2023. information fusion,” Sensors, vol. 21, no. 8, pp. 2693, 2021.
[92] R. Peng, W. Wei, D. Sun, and G. Wang, “A positive-unlabeled radar [112] J. Wang, W. Dong, Z. Song, and Q. Fu, “Identification of radar active
false target recognition method based on frequency response features,” interference types based on three-dimensional residual network,” in
IEEE Signal Processing Letters, vol. 30, pp. 1067-1071, 2023. Proceedings 3rd International Academic Exchange Conference on
[93] Q. Zhao, Y. Liu, L. Cai, and Y. Lu, “Research on electronic jamming Science and Technology Innovation (IAECST), 2021, pp. 167-172.
identification based on CNN,” in Proceedings IEEE International [113] K. Han, Y. Wang, H. Chen, G. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y.
Conference on Signal, Information and Data Processing (ICSIDP), Xu, “A Survey on Vision Transformer”, IEEE transactions on pattern
2019, pp. 1-5. analysis and machine intelligence, vol. 45, no. 1, pp. 87–110, 2022.
[94] J. Wang, W. Dong, and Q. Fu, Z. Song, “Radar jamming classification [114] B. Lang and J. Gong, “JR-TFViT: A Lightweight Efficient Radar
and recognition technology based on deep learning,” International Jamming Recognition Network Based on Global Representation of the
Conference on Signal Image Processing and Communication (ICSIPC Time-Frequency Domain,” Electronics, vol. 11, no. 17, pp. 2794, 2022.
2021), vol. 11848, pp. 201-207, 2021. [115] A. Lin, B. Chen, J. Xu, Z. Zhang, G. Lu, D. Zhang, “DS-TransUNet:
[95] M. Li, Q. Ren, and J. Wu, “Interference classification and identification Dual Swin Transformer U-Net for Medical Image Segmentation.” IEEE
of TDCS based on improved convolutional neural network,” in Journal Transactions on Instrumentation and Measurement, vol. 71, pp. 1–15,
of Physics: Conference Series, 2020, vol. 1651, no. 1, pp. 012155. 2022.
[96] J. Wang, W. Dong, and Z. Song, “Radar active jamming recognition [116] L. Qiu and Y. Fan, “A radar jamming recognition method based on
based on time-frequency image classification,” in Proceedings 5th hybrid dilated convolution,” in Proceedings 3rd International
Conference on Computer Vision, Image and Deep Learning &
29
International Conference on Computer Engineering and Applications [137] T. Hospedales, A. Antoniou, P. Micaelli, and A. Storkey, “Meta-
(CVIDL & ICCEA), 2022, pp. 692-695. learning in neural networks: A survey,” IEEE Transactions on Pattern
[117] J. Zhang, Z. Liang, C. Zhou, Q. Liu, and T. Long, “Radar compound Analysis and Machine Intelligence, vol. 44, no. 9, pp. 5149-5169, 2022.
jamming cognition based on a deep object detection network,” IEEE [138] Z. Fu, T. Xiang, E. Kodirov, and S. Gong, “Zero-shot learning on
Transactions on Aerospace and Electronic Systems, 2022. semantic class prototype graph,” IEEE transactions on pattern analysis
[118] X. Chen, Y. Liu, and C. Wang, “DRFM-Based Jamming Signal and machine intelligence, vol. 40, no. 8, pp. 2009-2022, 2017.
Recognition Method Guided by Target Detection,” Procedia Computer [139] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S.
Science, vol. 221, pp. 1013-1020, 2023. Ozair, A. Courville, and Y. Bengio, “Generative Adversarial
[119] P. Wang, “Research on Wireless Communication Interference Networks,” Communications of the ACM, vol. 63, no. 11, pp. 139-144,
Identification Technology Based on Deep Learning,” Doctoral 2020.
dissertation, Electronic Science and Technology University, 2023. [140] J. Yuan, Y. Bu, S. Yang, and Q. Chi, “Refined recognition and
[120] Z. Chen and M. He, “Deep reinforcement learning principles and intelligent smart interference of radar signal,” in Proceedings IEEE Int.
practice,” Beijing: People's Posts and Telecommunications Publishing Conf. Unmanned Syst. (ICUS), 2019, pp. 616-622.
House, 2019. [141] H. Li, X. Fang, L. Zhang, H. Kang, and W. Zhang, “Semi-supervised
[121] M. Zhu, Y. Li, Z. Pan, and J. Yang, “Automatic modulation recognition Open-set Recognition of Radar Active Jamming,” in Proceedings CIE
of compound signals using a deep multi-label classifier: A case study International Conference on Radar (Radar), 2021, pp. 2168-2171.
with radar jamming signals,” Signal Processing, vol. 169, pp. 107393, [142] X. Ding, Y. Zhang, G. Li, N. Ye, Y. Guo, T. Mabuchi, H. Anzai, and
2020. K. Yang, “Few-Shot Recognition and Classification of Jamming Signal
[122] Q. Lv, H. Fan, J. Liu, Y. Zhao, M. Xing, and Y. Quan, “Multilabel via CGAN-Based Fusion CNN Algorithm,” 2023, arXiv:2311.05273.
Deep Learning-Based Lightweight Radar Compound Jamming [143] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification
Recognition Method,” IEEE Transactions on Instrumentation and with deep convolutional neural networks,” Communications of the
Measurement, vol. 73, pp. 1-15, 2024. ACM, vol. 60, no. 6, pp. 84-90, 2017.
[123] H. Zhou, L. Wang, M. Ma, and Z. Guo, “Compound Radar Jamming [144] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
Recognition Based on Signal Source Separation,” Signal Processing, large-scale image recognition,” 2014, arXiv:1409.1556.
pp. 109246, 2024. [145] K. Chen, S. Zhang, L. Zhu, S. Chen, and H. Zhao, “Modulation
[124] K. Wang, Z. Dong, T. Wan, K. Jiang, W. Xiong, and X. Fang, recognition of radar signals based on adaptive singular value
“Research on radar active deception jamming identification method reconstruction and deep residual learning,” Sensors, vol. 21, no. 2, pp.
based on RESNET and bispectrum features,” in Proceedings 449, 2021.
International Conference on Computer Engineering and Application [146] Y. Xiao, J. Zhou, Y. Yu, and L. Guo, “Active jamming recognition
(ICCEA), 2021, pp. 491-495. based on bilinear EfficientNet and attention mechanism,” IET Radar
[125] M. Ren, B. Cheng, and P. Gao, “Active Jamming Signal Recognition Sonar & Navigation, vol. 15, no. 9, pp. 957-968, 2021.
based on Residual Neural Network*,” in Proceedings IEEE [147] Y. Hou, H. Ren, Q. Lv, L. Wu, X. Yang, and Y. Quan, “Radar-
International Conference on Signal Processing, Communications and Jamming Classification in the Event of Insufficient Samples Using
Computing (ICSPCC), 2022, pp. 1-4. Transfer Learning,” Symmetry, vol. 14, no. 11, pp. 2318, 2022.
[126] X. Zhang and X. Liu, “Interference Signal Recognition Based on Multi- [148] Y. Zhou, S. Shang, X. Song, S. Zhang, T. You, and L. Zhang,
Modal Deep Learning,” in Proceedings 7th International Conference on “Intelligent Radar Jamming Recognition in Open Set Environment
Dependable Systems and Their Applications (DSA), 2020, pp. 311-312. Based on Deep Learning Networks,” Remote Sensing, vol. 14, no. 24,
[127] G. Shao, Y. Chen, and Y. Wei, “Deep fusion for radar jamming signal pp. 6220, 2022.
classification based on CNN,” IEEE Access, vol. 8, pp. 117236- [149] Z. Luo, Y. Cao, T. S. Yeo, Y. Wang, and F. Wang, “Few-shot radar
117244, 2020. jamming recognition network via time-frequency self-attention and
[128] Y. Kong, X. Wang, C. Wu, X. Yu, and G. Cui, “Active deception global knowledge distillation,” IEEE Transactions on Geoscience and
jamming recognition in the presence of extended target,” IEEE Remote Sensing, vol. 61, pp. 1-12, 2023.
Geoscience and Remote Sensing Letters, vol. 19, pp. 1-5, 2022. [150] Z. Luo, Y. Cao, T. S. Yeo, and F. Wang, “Few-Shot Radar Jamming
[129] F. Cao, Z. Gao, C. He, X. Feng, J. Xu, C. Xue, and J. Qin, “Intelligent Recognition Network via Complete Information Mining,” IEEE
identification of radar active jamming type based on multi-domain Transactions on Aerospace and Electronic Systems, vol. 60, no. 3, pp.
information fusion,” in Journal of Physics: Conference Series, 2023, 3625-3638, 2024.
vol. 2480, no. 1, pp. 012015. [151] W. Zou, K. Xie, and J. Lin, “Light-weight deep learning method for
[130] H. Zhou, L. Wang, and Z. Guo, “Recognition of radar compound active jamming recognition based on improved MobileViT,” IET
jamming based on convolutional neural network,” IEEE Transactions Radar, Sonar & Navigation, vol. 17, pp. 1299-1311, 2023.
on Aerospace and Electronic Systems, vol. 59, no. 6, pp. 7380-7394, [152] G. Cheng, B. Yan, P. Shi, K. Li, X. Yao, L. Guo, and J. Han,
2023. “Prototype-CNN for few-shot object detection in remote sensing
[131] Y.Wang, Q.Yao Q, J. Kwok, L. Ni, “Generalizing from a few images,” IEEE Transactions on Geoscience and Remote Sensing, vol.
examples: A survey on few-shot learning,” ACM computing surveys 60, pp. 1-10, 2021.
(csur), vol. 53, no. 3, 2020. [153] F. Zhang, Y. Wang, J. Ni, Y. Zhou, and W. Hu, “SAR target small
[132] C. H. Lampert, H. Nickisch, and S. Harmeling, “Learning to detect sample recognition based on CNN cascaded features and AdaBoost
unseen object classes by between-class attribute transfer,” in 2009 rotation forest,” IEEE Geoscience and Remote Sensing Letters, vol. 17,
IEEE conference on computer vision and pattern recognition. IEEE, no. 6, pp. 1008-1012, 2020.
2009, pp. 951–958. [154] X. Nie, R. Gao, R. Wang, and D. Xiang, “Online multiview deep forest
[133] R. Socher, M. Ganjoo, C. D. Manning, and A. Ng, “Zero-shot learning for remote sensing image classification via data fusion,” IEEE
through cross-modal transfer,” Advances in neural information Geoscience and Remote Sensing Letters, vol. 18, no. 8, pp. 1456-1460,
processing systems, vol. 26, 2013. 2021.
[134] S. C. Wong, A. Gatt, V. Stamatescu, and M. D. McDonnell, [155] Q. Lv, Y. Quan, W. Feng, M. Sha, S. Dong, and M. Xing, “Radar
“Understanding Data Augmentation for Classification: When to deception jamming recognition based on weighted ensemble CNN with
Warp?” in Proceedings International Conference on Digital Image transfer learning,” IEEE Transactions on Geoscience and Remote
Computing: Techniques and Applications (DICTA), 2016, pp. 1-6. Sensing, vol. 60, pp. 1-11, 2022.
[135] Y. Sun, Y. Wang, H. Liu, N. Wang, and J. Wang, “SAR target [156] X. Chen, Z. Zhao, X. Ye, S. Zheng, C. Lou, and X. Yang, “Efficient
recognition with limited training data based on angular rotation Open-Set Recognition for Interference Signals Based on Convolutional
generative network,” IEEE Geoscience and Remote Sensing Letters, Prototype Learning,” Applied Sciences, vol. 12, no. 9, pp. 4380, 2022.
vol. 17, no. 11, pp. 1928-1932, 2020. [157] G. Koch, R. Zemel, and R. Salakhutdinov, “Siamese neural networks
[136] J. V. Davis, B. Kulis, P. Jain, S. Sra, and I. S. Dhillon, “Information- for one-shot image recognition,” ICML deep learning workshop, vol. 2,
theoretic metric learning,” in Proceedings 24th international conference no. 1, 2015.
on Machine learning, 2007, pp. 209-216. [158] P. Wang, Y. Cheng, B. Dong, and H. Xu, “Convolutional Neural
Network-Based Interference Recognition,” in Proc. 2020 IEEE 20th
International Conference on Communication Technology (ICCT),
2020, pp. 1296-1300.
30
[159] G. Shao, Y. Chen, and Y. Wei, “Convolutional neural network-based Gaofeng Shu (M’23) received the B.S.
radar jamming signal classification with sufficient and limited
degree from Wuhan University, Wuhan,
samples,” IEEE Access, vol. 8, pp. 80588-80598, 2020.
[160] Z. Wu, T. Wang, Y. Cao, M. Zhang, and L. Yang, “Radar active Hubei, China, in 2015, and the Ph.D.
deception jamming recognition based on Siamese squeeze wavelet degree from the Department of Space
attention network,” IET Radar, Sonar & Navigation, vol. 17, no. 12, Microwave Remote Sensing Systems,
2023.
Aerospace Information Research Institute,
[161] X. Tian, B. Chen, and Z. Zhang, “Multiresolution Jamming
Recognition with Few-shot Learning,” in Proceedings CIE International Chinese Academy of Sciences, Beijing,
Conference on Radar, 2021, pp. 2267-2271. China, in 2021. From 2019 to 2020, he
[162] S. Xiao, S. Zhang, M. Jiang, and W. Q. Wang, “PSPNet: Pretraining studied as a visiting Ph.D. student in the
and Self-Supervised Fine-Tuning-Based Prototypical Network for
Università degli Studi di Roma “La Sapienza”, Rome, Italy. In
Radar Active Deception Jamming Recognition with Few Shots,” IEEE
Geoscience and Remote Sensing Letters, vol. 21, pp. 1-5, 2024. 2021, he joined the School of Computer and Information
[163] H. Xu, J. Wang, H. Li, D. Ouyang, and J. Shao, “Unsupervised meta- Engineering, Henan University, Kaifeng, China. His research
learning for few-shot learning,” Pattern Recognition, vol. 116, pp. interests include synthetic aperture radar (SAR) imaging,
107951, 2021.
orbital angular momentum (OAM), and electromagnetic
[164] M. Liu, Z. Liu, W. Lu, Y. Chen, X. Gao, and N. Zhao, “Distributed
few-shot learning for intelligent recognition of communication vortex technology in synthetic aperture radar (SAR) systems.
jamming,” IEEE Journal of Selected Topics in Signal Processing, vol.
16, no. 3, pp. 395-405, 2021.
[165] Z. Ye, F. Hu, F. Lyu, L. Li, and K. Huang, “Disentangling semantic-to-
Ning Li (S’15-M’16) received the B.S.
visual confusion for zero-shot learning,” IEEE Transactions on
Multimedia, vol. 24, pp. 2828-2840, 2022. degree in electronics information
[166] N. Zhang, J. Shen, Y. Shi, and Y. Li, “CNN-Zero: A Zero-Shot engineering from Northeast Forestry
Learning Framework for Jamming Identification,” in Proceedings 22nd University, Harbin, China, in 2009, the
International Conference on Communication Technology (ICCT),
M.S. degree in communication and
2022, pp. 1126-1131.
[167] Y. Zhang, Z. Zhao, and Y. Bu, “Radar Active Jamming Recognition information system from the Nanjing
under Open World Setting,” Remote Sensing, vol. 15, pp. 4107, 2023. University of Aeronautics and Astronautics,
Nanjing, China, in 2012, and the Ph.D.
degree at the Institute of Electronics,
Chinese Academy of Sciences (IECAS), Beijing, China, in
Zan Wang was born in Henan, China, in 2015. In July 2015, he joined the Department of Space
1998. She is currently pursuing the M.S. Microwave Remote Sensing System, IECAS, Beijing, China,
degree with the College of Computer and where he was an assistant professor. Since December 2017, he
Information Engineering, Henan has been a full professor with the School of Computer and
University, Kaifeng, China. Her research Information Engineering, Henan University, Kaifeng, China.
interests include synthetic aperture radar His research interests include synthetic aperture radar (SAR)
(SAR) signal processing and SAR and inverse SAR imaging algorithms and autofocusing
electronic countermeasure. techniques, SAR polarimetric theory, and SAR image
processing. Dr. Li was a recipient of the Special Prize of
President Scholarship for Postgraduate Students from the
Zhengwei Guo received her B.S. degree University of Chinese Academy of Sciences in 2015.
in radio technology engineering from
Huazhong University of Science and
Technology, Wuhan, China, in 1984 and
then joined Henan University, Kaifeng,
China, as an Assistant Professor. Since
2008, Zhengwei Guo served as a
professor in School of Computer and
Information Engineering of Henan University. Her research
focus on management system software, data processing
system, and information security technology.Her projects on
SAR image applications and information processing system
have been supported by National Natural Science Foundation
of China and Department of Science and Technology of
Henan Province. Guo has published several papers including
peer reviewed journals,conference proceedings, patents,
software copyrights, and received the First-Prize Scientific
and Technological.

Radar Jamming Recognition ; Models, Methods, And Prospects

Uploaded by

Copyright:

Available Formats

Radar Jamming Recognition ; Models, Methods, And Prospects

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Radar Jamming Recognition ; Models, Methods, And Prospects

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in IEEE Journal of Selected Topics in Applied Earth Observations and

Radar Jamming Recognition:

W ITH the development of electronic warfare technology,

40 recognition methods based on neural networks, including

IOP, suggesting that IEEE can be considered the main MDPI

selected papers. It is clear that the common keywords are Springer

exploration, and in the coming years, we can expect to see IOP

is the power amplification factor, % is the power

Deception Jamming Suppression Jamming Smart Jamming

False Target Gate Pull Off

where ) represents the amplitude of the radar jamming signal.

% + − − Δ − Δ , - represents the phase term that t = 2 j6 −7 ,

where @ represents the number of sampled sub-pulses, (

AJ 9r ≈ 2∼5 9x r ≈ +y Slightly larger than the radar bandwidth

BJ 9r >59x r ≈ +y Much larger than the radar bandwidth

Jamming A B C Parameter Characteristics

RFJ 1 0 0 )1 : Rayleigh distribution Limited bandwidth noise signal

NAMJ 1 0 0 )1 ∼N(0, σn2) Amplitude modulated by noise

NFMJ 0 1 0 `L : Frequency modulation slope Instantaneous frequency modulated by noise

NPMJ 0 0 1 PM : Phase modulation coefficient Instantaneous phase modulated by noise

(a) LFM (b) RDJ (c) VDJ (d) ADJ

(i) C&IJ (j) ISRJ (k) DFTJ (l) AJ

(m) BJ (n) SJ (o) RFJ (p) NAMJ

(q) NFMJ (r) NPMJ (s) CSJ (t) PJ

(u) NPJ (v) CNJ (w) ISRJ+CSJ (x) C&IJ+SMSPJ

(a) LFM (b) RDJ (c) VDJ (d) ADJ

(i) C&IJ (j) ISRJ (k) DFTJ (l) AJ

(m) BJ (n) SJ (o) RFJ (p) NAMJ

(q) NFMJ (r) NPMJ (s) CSJ (t) PJ

(u) NPJ (v) CNJ (w) ISRJ+CSJ (x) C&IJ+SMSPJ

time domain: kurtosis, skewness, normalized JNR: 0dB,

Classifier Advantages Disadvantages

Chen et Lightweight SIR: −20dB, Average

Wang et Bispectral diagonal JSR> −7.5 dB, Recognition

SNR: −5dB, Recognition Rate:

STFT, Modulus, DFCN

Kong et STFT, RD domain MDFRCNN-AM JSR:3~5dB, Average

Tian et Prototypical SNR: 6 dB,

RMT, VMT, RVMT, C&I, ISRJ, five-shot case,

Learning Taskn New Taskn

Extract Fig.13. Process of meta learning.

\ SVM[64] RF[65] VGGNet-16[95] ResNet50[112] DFCN[127] MobileViT_CA[151]

(a) (b) (c)

(d) (e) (f)

You might also like