Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Research ArticleJamming Prediction for Radar Signals Using MachineLearning Methods
Gyeong-Hoon Lee 1 Jeil Jo 2 and Cheong Hee Park 1
1Department of Computer Science and Engineering Chungnam National University Daejeon Republic of Korea2e 2nd Research and Development Institute Agency for Defense Development Daejeon Republic of Korea
Correspondence should be addressed to Cheong Hee Park cheongheecnuackr
Received 7 May 2019 Revised 22 December 2019 Accepted 30 December 2019 Published 24 January 2020
Academic Editor Roberto Di Pietro
Copyright copy 2020 Gyeong-Hoon Lee et al +is is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited
Jamming is a form of electronic warfare where jammers radiate interfering signals toward an enemy radar disrupting the receiver+e conventional method for determining an effective jamming technique corresponding to a threat signal is based on the librarywhich stores the appropriate jamming method for signal types However there is a limit to the use of a library when a threat signalof a new type or a threat signal that has been altered differently from existing types is received In this paper we study twomethodsof predicting the appropriate jamming technique for a received threat signal using deep learning using a deep neural network onfeature values extracted manually from the PDW list and using long short-term memory (LSTM) which takes the PDW list asinput Using training data consisting of pairs of threat signals and corresponding jamming techniques a deep learning model istrained which outputs jamming techniques for threat signal inputs Training data are constructed based on the information in thelibrary but the trained deep learning model is used to predict jamming techniques for received threat signals without using thelibrary +e prediction performance and time complexity of two proposed methods are compared In particular the ability topredict jamming techniques for unknown types of radar signals which are not used in the stage of training the model is analyzed
1 Introduction
Electronic warfare is a military activity that uses the elec-tromagnetic spectrum to attack an enemy or impede enemyassaults and jamming is a form of electronic warfare wherejammers radiate interference signals toward an enemy radarin order to disrupt the receiver Jamming techniques can becategorized largely to noise jamming range deceptionjamming velocity deception jamming and angle deceptionjamming [1] Since the effect of jamming varies dependingon the characteristics of the received radar signal a jammingtechnique that is effective for the threat signal has to beapplied
When a threat signal is received the conventionalmethod to determine a jamming technique is based oninformation in the library which stores the appropriatejamming method for signal types While a jamming methodcan be selected and applied easily when the type and value of
a received threat signal exist in the library there is a limit tothe use of a library when a threat signal of a new type or athreat signal that has been altered differently from existingtypes is received Hence it is necessary to apply a machinelearning model that can predict the appropriate jammingtechnique by learning the received radar signalcharacteristics
An analog radar signal is converted to a digital signal andstored as a pulse description word (PDW) list where variousfeature values including the pulse width (PW) pulse rep-etition interval (PRI) and radio frequency (RF) are recordedaccording to the arrival time of pulses [2] In particular sincePRI and RF are modulated with time and a radar signal hasdifferent modulation types it is necessary to grasp sequentialcharacteristics in PDW In recent years deep learning hasbeen used effectively for sequential data analysis
In this paper we study two methods of predicting ajamming method corresponding to a threat signal using
HindawiSecurity and Communication NetworksVolume 2020 Article ID 2151570 9 pageshttpsdoiorg10115520202151570
deep learning Firstly a deep neural network can be modeledon feature values extracted manually from the PDW list Inthis approach the feature extraction process and thelearning process of a neural network model are performedsequentially Hence an effective feature extraction methodthat will enhance the performance of the model has to beselected appropriately Secondly instead of extracting thefeature values we can learn long short-term memory(LSTM) [3] which uses the PDW list itself as input Despitethe outstanding performance of LSTM in dealing with theshort- and long-term dependence of the sequence data nostudies using LSTM has been conducted to predict how tojam the threat signal as far as we know
+e prediction performance and time complexity of twoapproaches are compared Various feature values includingautocorrelation coefficients [4] and mel frequency cepstralcoefficients (MFCCs) [5] are extracted from the radar signaland a deep neural networks of different structures are testedAlso the process of determining values of model parametersin LSTM is discussed In particular the ability to predictjamming methods for unknown types of radar signals whichare not used to construct the model is analyzed
+e paper is organized as follows in Section 2 the re-search related to radar signal analysis is discussed and inSection 3 the radar signal data and testing setup areexplained In Section 4 the process of determining modelparameters in LSTM is discussed Section 5 explains thejamming prediction model based on feature value extractionand a deep neural network In Section 6 we compare theprediction performance and training time in two ap-proaches Conclusions are made in Section 7
2 Related Work
Generally the methods of extracting feature values fromsequence data can be divided into time-domainmethods andfrequency-domain methods [6] Time-domain methodsextract features such as various statistical values and auto-correlation coefficients Studies in [7ndash9] have used statisticalvalues including the root mean square square root of theamplitude kurtosis value skewness value peak-peak valuecrest factor impact factor margin factor shape factor andkurtosis factor In [4] for fuzzy clustering of time seriesautocorrelation coefficients of different lags were used asfeature values In the frequency domain fast Fouriertransform spectrum analysis or wavelet transform are oftenused for pattern analysis over time MFCC is a common andefficient technique for signal processing which is used invarious domains such as speech recognition speakeridentification and hand gesture recognition [5 10]
Radar signal analysis uses various feature extractionmethods in time and frequency domains in order to obtainthe sequential characteristics with respect to time A study in[11] analyzed deception jamming and target echo using thefeatures of amplitude undulation high-order cumulant andbispectrum Another study compared classification perfor-mance of frequency-modulated (FM) signals from additivewhite Gaussian noise in two cases when using the auto-correlation coefficients in the time domain and when using
the power spectrum in the frequency domain as the input ofa neural network (NN) [12] It showed that at a high signal-to-noise ratio (SNR) the NN classifier performs well foreither case but at very low SNR it performs better when itsinput is the autocorrelation of the signal
Machine learning techniques have been used for radarsignal classification In [13] a 64-dimensional vector for asignal was constructed based on the second difference of thepulse TOAs (time of arrival) and a 3-layer multilayer per-ceptron (MLP) was trained to classify radar signals to fourPRI modulation types +ere was also a study that classifiedradar signals according to the PRI modulation types bydefining a set of five features based on the histogram of pulseintervals and the second difference of TOAs and training aMLP with one hidden layer [14] +e work in [15] inves-tigated the recognition of generic radar data signal trainpulse sources using three classification models neuralnetworks support vector machines (SVM) and randomforests (RF) which were combined with three missing featureimputation techniques such as multiple imputation k-nearest neighbour imputation and bagged tree imputationIt showed overall improved accuracy when the classifier istrained on the bagged tree imputed values In [16] therestricted Boltzmannmachine was used to extract the featureparameters and recognize the radar emitter A bottom-uphierarchical unsupervised learning was used to obtain theinitial parameters and the traditional BP algorithm wasconducted to fine tune the network parameters
Research was also carried out on jamming signal clas-sification using deep learning A study in [17] used a 13-layerartificial neural network composed of a 3-layer convolu-tional neural network (CNN) 2 pooling layers 4 fullyconnected layers and 4 batch normalization layers for theclassification of jamming signals of five types CNN has theadvantage of being able to automatically extract features andlocally use the sequential information of a signal However itis difficult to determine the long-term time dependence ofthe signal because only local information can be obtained In[18] the jammer classification problem in the GlobalNavigation Satellite System (GNSS) bands was treated as ablack-and-white image classification problem based on atime-frequency analysis and image mapping of a jammedsignal GNSS signals interfered with one of the five jammertypes were saved as a black-and-white image SVM and CNNalgorithms were applied for classification of six classes in-cluding one class where the jammer was absent
Deep learning has also been used for various tasks inradar signal processing In [19] recurrent neural network(RNN) model with gated recurrent unit (GRU) was appliedto remove interference and reconstruct transmit signal +eauthors of [20] argued that deep learning offers a unifyingframework to integrate sensing processing and decision-making and applied their approach to the autofocus problemin the synthetic aperture radar (SAR) imaging
3 Radar Signal Data
In this section we explain the representation of radar signalsand experimental setting including data construction
2 Security and Communication Networks
31 Radar Signals A radar signal can be represented as asequence of pulses as shown in Figure 1 For each pulse 5characteristic values pk for PW ik for PRI fk for RF ak forAOA (angle of attack) and mk for AMP (amplitude) areobtained and a radar signal can be stored as a PDW list
X1 p1 i1 f1 a1 m1( 1113857 Xn pn in fn an mn( 11138571113858 1113859
(1)
Generally radar signals have modulations in the RF andPRI in order to evade to be detected RFmodulation refers tochanges in a sequence [f1 fn] of pulse frequencyvalues +ere are 6 types of RF modulation including fixagile hopping sine minus sawtooth and +sawtooth as shown inFigure 2(a) PRI modulation refers to changes in a sequence[i1 in] of pulse repetition interval values and there are 7types of PRI modulation including stable jitter staggerdwell and switch sine minus sawtooth and +sawtooth as shownin Figure 2(b)
32 Experimental Setting With reference to the specifica-tions of the currently operating radar models 2258 signaltypes were configured with different types and magnitudesof RF and PRI Each type had a minimum of 1000 andmaximum of 9000 PDW lists and a total of 6870000PDW lists were composed A total of eight jammingtechniques corresponding to 2258 signal types wereestablished by assigning one jamming technique to eachsignal type based on the information of the library+e goalof a deep learning model is set to predict one jammingtechnique among eight jamming techniques for threatsignal input Hence the output layer in a deep learningmodel consists of eight nodes and each jamming techniqueis encoded as an 8-dimensional binary vector by the one-hot-encoding method where only one component is 1 andthe others are 0 In the training and testing stages of a deeplearning model the values in the output layer which areforwarded from a previous layer are translated as proba-bility values for each jamming technique by the softmaxfunction
A deep learning model is trained with training datawhich is composed of the pairs of radar signals andcorresponding jamming techniques In the testing stageone among eight jamming techniques is predicted as anoutput of the constructed model when a radar signal isgiven as an input To evaluate the prediction performanceof a deep learning model for an unknown radar signal typethat is not used in the training stage as well as for a knownsignal type we randomly divided 2258 signal types intothe set A of 2035 types to be used for the training andtesting and the set B of 223 types to be used for testing ofthe unknown types +e ratio of the two divisions wasapproximately 9 1 For signal types in A dividing PDWlists belonging to each type into 3 groups with a ratio of 7 1 2 three data sets for training validation and testing ofknown types were composed +e training data were usedto train the model and the validation data were used toselect the model parameters and determine the finalmodel Validation data are needed to prevent the trained
model from overfitting with training data Dropout [21] isalso tested which is a method for preventing overfitting byprobably dropping nodes of hidden layers during learn-ing +e test data are not used for training and insteadthey are used to evaluate the performance of the finalmodel Lastly the performance of the final model isassessed using the PDW lists of the unknown radar signaltype in the set B Figure 3 shows the configuration of thedata sets
Based on the configuration of the data sets shown inFigure 3 a flowchart to show the overall organization in theremaining sections is given in Figure 4 for easy readabilityBased on train and validation data of signal types of set A theconstruction of a LSTM model is explained in Section 4 andthe construction of feature extraction and a deep neuralnetwork is described in Section 5 Test data of set A whichrepresents the known signal types and test data of set Bwhich indicates the unknown signal types are used for theperformance comparison of two models in Section 6
4 Construction of a Jamming PredictionModelBased on LSTM
In this section the construction of a LSTM model whichtakes a PDW list of a threat signal as input is explained
41ModelParameters inLSTM Preliminary experiment wascarried out using the small scale of radar signal data to selectthe LSTM model structure and parameter values [22] +edata used in the preliminary test had 1157500 PDW lists fora total of 640 radar signal types and 4 types of jammingmethods +e basic model was composed of an input layerwith 5 nodes for the feature values of AOA AMP RF PRIand PW a LSTM layer consisting of 200 nodes and theoutput layer consisting of 4 nodes A process of determiningan optimization method minibatch size dropout ratio thenumber of the LSTM layers and fully connected layers anddecay ratio of learning rate was performed Table 1 shows theparameters and their values which were tested Since therewere somany combinations of the parameter values we haveadopted a strategy that determines the parameter valuessequentially When testing optimization methods in step 1all other parameters were set to the underlined values in thethird column of Table 1 After the best optimization methodis selected it is fixed in the next steps where the remainingparameter values are determined +is process is performedin sequence up to step 7 in Table 1 +e fourth column ofTable 1 shows the selected parameter values
42 LSTM Model With reference to the model structureconfigured in Table 1 the final LSTM model was composedof an input layer with 3 nodes for RF PRI and PW 2 LSTMlayers with 200 nodes one fully connected layer with 400nodes and an output layer Since one among the eightjamming techniques needs to be predicted for each inputthreat signal the output layer has 8 nodes +rough addi-tional tests it was determined to apply the batch normali-zation in the fully connected layer and gradient clipping for
Security and Communication Networks 3
Pulse
f1 f2 f3
t2t1 t3
p1 p2 p3
i1 i2 i3
Figure 1 A radar signal represented as a sequence of pulses
Fix Agile Hopping
TOA
RF
Sine ndashsawtooth +sawtooth
(a)
Stable Jitter DnsPRI
TOASine ndashsawtooth +sawtooth Stagger
(b)
Figure 2 Various types of RF and PRI (a) 6 types of RF (b) 7 types of PRI
4 Security and Communication Networks
stable learning Figure 5 shows the structure of the finalLSTM model
5 Construction of Jamming Prediction ModelBased on Feature Extraction and a DeepNeural Network
In this section feature extraction from a PDW list of a threatsignal and the construction of a deep neural network areexplained in detail
51 Feature Extraction From sequence data expressed asx1 xn features such as statistical values autocorrelationcoefficients and MFCCs were extracted
Four statistic values of the mean x standard deviation sskewness w and kurtosis k were computed from the se-quence data such as
x 1n
1113944
n
1xt
s
1n
1113944
n
1xt minus x( 1113857
2
11139741113972
w 1n 1113936
n1 xt minus x( 1113857
31113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
32
k 1n1113936
n1 xt minus x( 1113857
41113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
2
(2)
+e autocorrelation coefficient R(τ) represents the linearrelevance between two subsequences with a lag τ which iscomputed by
A 2035 signal types B 223 types
Train4324600 PDW lists
Validation617800
Test1235600
692000PDW lists
Figure 3 +e configuration of the data sets
A 2035 signal typesTrain
4324600 PDW lists
Validation617800 PDW lists
Section 4 constructionof a LSTM model
Section 5 constructionof feature extraction
and a deep neural network
Section 6 comparison of predictionperformance by two models
A 2035 known signal typesTest
1235600 PDW lists
B 223 unknown signal types
Test692000 PDW lists
Figure 4 A flowchart to explain the overall organization
Table 1 Parameters of a LSTM model
Step Parameters Tested values Selectedvalue Note
1 Optimizationmethod
RMSProp AdamAdadelta Adam Initial learning rate was set as 0003 for RMSProp [23] Adam [24] and 10 for
Adadelta [25]2 Minibatch size 50 100 200 200
3 Dropout ratio () 0 10 30 50 0 Dropout ratio means the rate at which the output gate units in a LSTM layerare randomly removed
4 LSTM layers 1 2 2 +e model with 2 LSTM layers showed higher accuracy than the model of 1LSTM layer
5 Fully connectedlayer 0 1 1 +e model with the fully connected layer had a higher accuracy than the
model with no fully connected layer
6 Input features 3 5 3 Higher accuracy was obtained when using three features RF PRI and PWinstead of using 5 features AOA AMP RF PRI and PW
7 Decay ratio Use no use Use When gradually decreasing the learning rate by multiplying 09 to theprevious learning rate per epoch after epoch 10 higher accuracy was obtained
Security and Communication Networks 5
R(τ) 1113936
nminus τ1 xt minus x( 1113857 xt+τ minus x( 1113857
1113936n1 xt minus x( 1113857
2 (3)
Various autocorrelation coefficients can be obtained bychanging the time difference τ [4]
MFCC extracts feature values from the sequence datathrough the process of framing Fast Fourier transform(FFT) mel filter bank log function and inverse FFT [5] Inthe framing step the sequence is divided into segments ofequal size FFT is applied to a subsequence of each segmentto convert to the frequency domain and then the powerspectrum is obtained +e mel spectrum is obtained byapplying mel filters to the power spectrum and the coef-ficients are obtained by inverse FFT after applying the logfunction to the mel spectrum +ese coefficients are calledMFCCs All or some of the coefficients obtained in eachsegment can be used as the feature values [5]
52 Composition of Feature Sets We extracted feature valuesusing RF PRI and PW sequences of a PDW list Bycomputing the mean standard deviation skewness andkurtosis from each sequence a total of 12 feature values wereobtained from a PDW list By increasing the time differenceτ from 1 to 10 by incrementing 1 we extracted 10 auto-correlation coefficients from the RF and PRI sequences andtherefore a total of 20 feature values were extracted from aPDW list Also 10 MFCCs were computed for the RF andPRI sequences so that a total of 20 values were extracted+eautocorrelation coefficients and MFCCs were computedusing Python libraries such as the function acf of statsmodelsand the function mfcc of librosa [26]
Using statistical features as the basic features andcombining the autocorrelation coefficients and MFCCs withbasic features we composed 4 feature sets as shown in
Table 2 and tested prediction performance of a neuralnetwork model on each feature set In the feature set F1 aradar signal is represented with only statistical features +efeature set F2 contains the statistical features and autocor-relation coefficients while F3 is composed of the statisticalfeatures and MFCCs +e feature set F4 contains all of thestatistical features autocorrelation coefficients and MFCCsresulting in a total of 52 feature values
53 Performance of Neural Network Models on VariousFeature Sets To select a neural network structure com-parative testing for 4 different models was performedvarying the activation function and the number of layers asshown in Table 3 Considering test results in Section 41Adam was used as the optimization method and the min-ibatch size was set to 200
+e 4 feature sets in Table 2 and the 4 models in Table 3were used to compare the performance +e total number oftraining epochs was set to 15 and the training data com-posed in Section 32 were used to train the models +eaccuracy of the validation data was evaluated in each epochTable 4 shows the training and validation accuracies for 16combinations of 4 feature sets and 4 NNmodels which wereevaluated in the epoch with the highest accuracy for thevalidation data It revealed that model 4 on feature set F3 wasfound to have the highest performance Hence model 4 onfeature set F3 was chosen as the final model Figure 6 de-scribes the structure of the final deep neural network model
6 Performance Comparison of LSTM Modeland Deep Neural Network Model UsingExtracted Features
In this section we compare the prediction performance andtraining time complexity of two deep learning models fortest data of the known signal types in set A and test data ofthe unknown signal type in set B As explained in Section 322258 signal types were randomly divided into the set A to beused for the training validation and testing and the set B tobe used for testing of the unknown signal types Now werepeat the random division 10 times and measure the av-erage accuracy for performance comparison of two ap-proaches using LSTM and using a deep neural network withextracted features (denoted as DNNEF)
Table 5 shows test accuracy of known signal types in Aand unknown types in B from 10 repeated experiments Intesting of known signal types the average accuracy of 9846was obtained for the NN with extracted features and theaverage accuracy of 9936 was obtained for the LSTMmodel +e average accuracy for the unknown types whenusing the neural network with extracted features was 9245while the average accuracy for the LSTMmodel was 9353In both cases the LSTM model showed a little higheraccuracy
+e paired t-test was carried out in order to determinewhether the higher average accuracy of the LSTM model isstatistically significant +e paired t-test for the accuracy ofknown types gave the p value 00000612 which implies
Output layer 8 nodes
Input layer 3 nodes
RF PRI PWPDW list
Fully connected layer (400 nodes)
LSTM layer (200 nodes)
LSTM layer (200 nodes)
Figure 5 Structures of the final LSTM model
6 Security and Communication Networks
statistical significance in the difference between the accu-racies of the two models in the significance level of 001However for the unknown type accuracy the p value by thepaired t-test was approximately 0215 and so there was nosignificant difference in the significance level of 001 +isresult was thought to be due to the very large deviation in theaccuracy of the LSTM model for the unknown type whichwas a minimum of 8969 and maximum of 9720
In order to compare the learning time of the twomethods the average training time per epoch was measured+e specification of the computer used to measure the timeconsumed was CPU Intel i7-7700 360GHz and RAM320GB and GPU NVIDIA GeForce GTX 1080 Ti Table 6compares the execution time for model training for 15epochs+e process of extracting the features from the PDWlists was conducted once before the training of the deepneural network model As shown in Table 6 the time
consumed for training the LSTM model for 15 epochs wasapproximately 142 times longer than that of the deep neuralnetwork
Table 3 Four deep neural network models
Parameters Model 1 Model 2 Model 3 Model 4Activation function tanh relu tanh reluHidden layers 2 4Hidden units 400Learning rate Initially set as 00005
Learning rate decay Multiplying the previous learning rate by097 for every 1000000 samples
Table 4 +e performance in 16 combinations of feature sets andmodels
Model Feature set Training accuracy()
Validation accuracy()
Model 1
F1 9572 9592F2 9677 9709F3 9735 9778F4 9729 9778
Model 2
F1 9567 9572F2 9681 9730F3 9725 9755F4 9719 9761
Model 3
F1 9689 9735F2 9721 9765F3 9742 9804F4 9730 9778
Model 4
F1 9684 9748F2 9780 9819F3 9782 9847F4 9777 9802
Input layer 32 nodes
Hidden layer (400 nodes)
Output layer 8 nodes
Hidden layer (400 nodes)
Hidden layer (400 nodes)
Hidden layer (400 nodes)
32 feature values extracted from a PDW list
RF PRI PW
Figure 6 Structure of the final deep neural network
Table 5 Test accuracy () of known types and unknown types
DNNEF LSTM modelKnowntypes
Unknowntypes
Knowntypes
Unknowntypes
1 9849 9333 9972 97202 9835 9247 9970 95093 9849 9057 9878 90764 9860 9281 9906 90945 9861 9272 9932 93426 9855 9431 9960 95067 9839 9262 9971 96738 9847 9211 9901 89699 9831 9219 9900 906410 9833 9132 9965 9576Avg 9846 9245 9936 9353
Table 6 Comparison of execution time in seconds
Model Feature extraction Training for 1 epoch Training for 15epochs
DNN 7374233 1932943 36368378LSTM mdash 3454399 51815985
Table 2 Composition of four feature sets
Featuresets
Statisticalfeatures
Autocorrelationcoef MFCC Total
F1 12 mdash mdash 12F2 12 20 mdash 32F3 12 mdash 20 32F4 12 20 20 52
Security and Communication Networks 7
7 Conclusion
In this study we applied deep learning for jamming pre-diction for a radar signal Two methods were comparedusing a deep neural network model based on the featuresextracted from radar signals and using LSTMmodel withoutthe feature extraction process
+e deep neural network requires feature extraction inadvance and this is conducted once before the training ofthe deep neural network model +e training speed of themodel per epoch was faster than that of the LSTM modelHowever the feature extraction method has to be selectedappropriately because the model prediction performancevaries depending on the feature values extracted from thePDW list +e LSTMmodel has the advantages of being ableto perform training and prediction by directly using thePDW list of the radar signal without the extraction of featurevalues +e prediction accuracy of the LSTM model washigher on average than that of the deep neural networkmodel However there is the disadvantage that the trainingtime takes longer than a deep neural network model trainedon the extracted features
Testing results demonstrate that the jamming methodcan be predicted for an unknown type of radar signal with anaverage accuracy of approximately 92 and higher It showsthat deep learning methods can be used effectively forpredicting an appropriate jamming technique for new radarsignals which are not used in model training
Data Availability
+e radar signal data used to support the findings of thisstudy have not been made available because of the policy ofAgency for Defense Development Korea that data distri-bution is limited
Conflicts of Interest
+e authors declare that there are no conflicts of interestregarding the publication of this paper
References
[1] M S R Lothes M B Szymanski and R G Wiley RadarVulnerability to Jamming Artech House Boston MA USA1990
[2] V Iglesias J Grajal O Yeste-Ojeda M Garrido M Sanchezand M Lopez-Vallejo ldquoReal-time radar pulse parameterextractorrdquo in Proceedings of the IEEE Radar ConferenceCincinnati OH USA May 2014
[3] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[4] P DrsquoUrso and E Maharaj ldquoAutocorrelation-based fuzzyclustering of time seriesrdquo Fuzzy Sets and Systems vol 160no 24 pp 3565ndash3589 2009
[5] S Gupta J Jaafar W F wan Ahmad and A Bansal ldquoFeatureextraction using MFCCrdquo Signal amp Image Processing An In-ternational Journal vol 4 no 4 pp 101ndash108 2013
[6] R Shumway and D Stoffer Time Series Analysis and itsApplication with R Examples Springer Berlin Germany2000
[7] Z Xia S Xia L Wan and S Cai ldquoSpectral regression basedfault feature extraction for bearing accelerometer sensorsignalsrdquo Sensors vol 12 no 10 pp 13694ndash13719 2012
[8] T W Rauber F de Assis Boldt and F M Varejao ldquoHet-erogeneous feature models and feature selection applied tobearing fault diagnosisrdquo IEEE Transactions on IndustrialElectronics vol 62 no 1 pp 637ndash646 2015
[9] B R Nayana and P Geethanjali ldquoAnalysis of statistical time-domain features effectiveness in identification of bearingfaults from vibration signalrdquo IEEE Sensors Journal vol 17no 17 pp 5618ndash5625 2017
[10] M B L Muda and I Elamvazuthi ldquoVoice recognition al-gorithm using mel frequency cepstral coefficient (MFCC) anddynamic time warping (DTW) techniquesrdquo Journal ofComputing vol 2 no 3 pp 138ndash143 2010
[11] S Q L Jian-xun and Y Hai ldquoSignal feature analysis andexperimental verification of radar deception jammingrdquo inProceedings of the IEEE CIE International Conference onRadar Chengdu China October 2011
[12] A S A Mendoza and B Flores ldquoClassification of radarjammer FM signals using a neural networkrdquo in Proceedings ofthe SPIE Radar Sensor Technology XXI vol 10188 May 2017
[13] G Noone ldquoA neural approach to automatic pulse repetitioninterval modulation recognitionrdquo in Proceedings of the 1999Information Decision and Control Data and InformationFusion Symposium Signal Processing and CommunicationsSymposium and Decision and Control Symposium Proceedings(Cat No99EX251) Adelaide Australia February 1999
[14] J Kauppi and K Martikainen ldquoAn efficient set of features forpulse repetition interval modulation recognitionrdquo in Pro-ceedings of the IET International Conference on Radar SystemsEdinburgh UK October 2007
[15] I Jordanov N Petrov and A Petrozziello ldquoSupervised radarsignal classificationrdquo in Proceedings of the International JointConference on Neural Networks Vancouver Canada July2016
[16] D Zhou X Wang Y Tian and R Wang ldquoA novel radarsignal recognition method based on a deep restricted Boltz-mann machinerdquo Engineering Review vol 37 pp 165ndash1712017
[17] Z Y Z Wu Y Zhao and H Luo ldquoJamming signals clas-sification using convolutional neural networkrdquo in Proceedingsof the IEEE International Symposium on Signal Processing andInformation Technology (ISSPIT) Bilbao Spain December2017
[18] R Ferre A Fuente and E Lohan ldquoJammer classification inGNSS bands via machine learning algorithmsrdquo Sensorsvol 19 no 22 p 4841 2019
[19] J Mun H Kim and J Lee ldquoA deep learning approach forautomotive radar interference mitigationrdquo 2019 httpsarxivorgabs190306380
[20] E Masen B Yonei and B Yazici ldquoDeep learning for radarrdquoin Proceedings of the 2017 IEEE Radar Conference(RadarConf) Seattle WA USA May 2017
[21] N Srivastava G Hinton A Krizhevsky I Sutskever andR Salakhutdinov ldquoDropout a simple way to prevent neuralnetworks from overfittingrdquo e Journal of Machine LearningResearch vol 15 no 1 pp 1929ndash1958 2014
[22] G Lee ldquoRadar jamming technique prediction using deeplearningrdquo Master thesis Chungnam National UniversityDaejeon Korea 2019
[23] M Zeiler ldquoADADELTA an adaptive learning rate methodrdquo2012 httpsarxivorgabs12125701
8 Security and Communication Networks
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9
deep learning Firstly a deep neural network can be modeledon feature values extracted manually from the PDW list Inthis approach the feature extraction process and thelearning process of a neural network model are performedsequentially Hence an effective feature extraction methodthat will enhance the performance of the model has to beselected appropriately Secondly instead of extracting thefeature values we can learn long short-term memory(LSTM) [3] which uses the PDW list itself as input Despitethe outstanding performance of LSTM in dealing with theshort- and long-term dependence of the sequence data nostudies using LSTM has been conducted to predict how tojam the threat signal as far as we know
+e prediction performance and time complexity of twoapproaches are compared Various feature values includingautocorrelation coefficients [4] and mel frequency cepstralcoefficients (MFCCs) [5] are extracted from the radar signaland a deep neural networks of different structures are testedAlso the process of determining values of model parametersin LSTM is discussed In particular the ability to predictjamming methods for unknown types of radar signals whichare not used to construct the model is analyzed
+e paper is organized as follows in Section 2 the re-search related to radar signal analysis is discussed and inSection 3 the radar signal data and testing setup areexplained In Section 4 the process of determining modelparameters in LSTM is discussed Section 5 explains thejamming prediction model based on feature value extractionand a deep neural network In Section 6 we compare theprediction performance and training time in two ap-proaches Conclusions are made in Section 7
2 Related Work
Generally the methods of extracting feature values fromsequence data can be divided into time-domainmethods andfrequency-domain methods [6] Time-domain methodsextract features such as various statistical values and auto-correlation coefficients Studies in [7ndash9] have used statisticalvalues including the root mean square square root of theamplitude kurtosis value skewness value peak-peak valuecrest factor impact factor margin factor shape factor andkurtosis factor In [4] for fuzzy clustering of time seriesautocorrelation coefficients of different lags were used asfeature values In the frequency domain fast Fouriertransform spectrum analysis or wavelet transform are oftenused for pattern analysis over time MFCC is a common andefficient technique for signal processing which is used invarious domains such as speech recognition speakeridentification and hand gesture recognition [5 10]
Radar signal analysis uses various feature extractionmethods in time and frequency domains in order to obtainthe sequential characteristics with respect to time A study in[11] analyzed deception jamming and target echo using thefeatures of amplitude undulation high-order cumulant andbispectrum Another study compared classification perfor-mance of frequency-modulated (FM) signals from additivewhite Gaussian noise in two cases when using the auto-correlation coefficients in the time domain and when using
the power spectrum in the frequency domain as the input ofa neural network (NN) [12] It showed that at a high signal-to-noise ratio (SNR) the NN classifier performs well foreither case but at very low SNR it performs better when itsinput is the autocorrelation of the signal
Machine learning techniques have been used for radarsignal classification In [13] a 64-dimensional vector for asignal was constructed based on the second difference of thepulse TOAs (time of arrival) and a 3-layer multilayer per-ceptron (MLP) was trained to classify radar signals to fourPRI modulation types +ere was also a study that classifiedradar signals according to the PRI modulation types bydefining a set of five features based on the histogram of pulseintervals and the second difference of TOAs and training aMLP with one hidden layer [14] +e work in [15] inves-tigated the recognition of generic radar data signal trainpulse sources using three classification models neuralnetworks support vector machines (SVM) and randomforests (RF) which were combined with three missing featureimputation techniques such as multiple imputation k-nearest neighbour imputation and bagged tree imputationIt showed overall improved accuracy when the classifier istrained on the bagged tree imputed values In [16] therestricted Boltzmannmachine was used to extract the featureparameters and recognize the radar emitter A bottom-uphierarchical unsupervised learning was used to obtain theinitial parameters and the traditional BP algorithm wasconducted to fine tune the network parameters
Research was also carried out on jamming signal clas-sification using deep learning A study in [17] used a 13-layerartificial neural network composed of a 3-layer convolu-tional neural network (CNN) 2 pooling layers 4 fullyconnected layers and 4 batch normalization layers for theclassification of jamming signals of five types CNN has theadvantage of being able to automatically extract features andlocally use the sequential information of a signal However itis difficult to determine the long-term time dependence ofthe signal because only local information can be obtained In[18] the jammer classification problem in the GlobalNavigation Satellite System (GNSS) bands was treated as ablack-and-white image classification problem based on atime-frequency analysis and image mapping of a jammedsignal GNSS signals interfered with one of the five jammertypes were saved as a black-and-white image SVM and CNNalgorithms were applied for classification of six classes in-cluding one class where the jammer was absent
Deep learning has also been used for various tasks inradar signal processing In [19] recurrent neural network(RNN) model with gated recurrent unit (GRU) was appliedto remove interference and reconstruct transmit signal +eauthors of [20] argued that deep learning offers a unifyingframework to integrate sensing processing and decision-making and applied their approach to the autofocus problemin the synthetic aperture radar (SAR) imaging
3 Radar Signal Data
In this section we explain the representation of radar signalsand experimental setting including data construction
2 Security and Communication Networks
31 Radar Signals A radar signal can be represented as asequence of pulses as shown in Figure 1 For each pulse 5characteristic values pk for PW ik for PRI fk for RF ak forAOA (angle of attack) and mk for AMP (amplitude) areobtained and a radar signal can be stored as a PDW list
X1 p1 i1 f1 a1 m1( 1113857 Xn pn in fn an mn( 11138571113858 1113859
(1)
Generally radar signals have modulations in the RF andPRI in order to evade to be detected RFmodulation refers tochanges in a sequence [f1 fn] of pulse frequencyvalues +ere are 6 types of RF modulation including fixagile hopping sine minus sawtooth and +sawtooth as shown inFigure 2(a) PRI modulation refers to changes in a sequence[i1 in] of pulse repetition interval values and there are 7types of PRI modulation including stable jitter staggerdwell and switch sine minus sawtooth and +sawtooth as shownin Figure 2(b)
32 Experimental Setting With reference to the specifica-tions of the currently operating radar models 2258 signaltypes were configured with different types and magnitudesof RF and PRI Each type had a minimum of 1000 andmaximum of 9000 PDW lists and a total of 6870000PDW lists were composed A total of eight jammingtechniques corresponding to 2258 signal types wereestablished by assigning one jamming technique to eachsignal type based on the information of the library+e goalof a deep learning model is set to predict one jammingtechnique among eight jamming techniques for threatsignal input Hence the output layer in a deep learningmodel consists of eight nodes and each jamming techniqueis encoded as an 8-dimensional binary vector by the one-hot-encoding method where only one component is 1 andthe others are 0 In the training and testing stages of a deeplearning model the values in the output layer which areforwarded from a previous layer are translated as proba-bility values for each jamming technique by the softmaxfunction
A deep learning model is trained with training datawhich is composed of the pairs of radar signals andcorresponding jamming techniques In the testing stageone among eight jamming techniques is predicted as anoutput of the constructed model when a radar signal isgiven as an input To evaluate the prediction performanceof a deep learning model for an unknown radar signal typethat is not used in the training stage as well as for a knownsignal type we randomly divided 2258 signal types intothe set A of 2035 types to be used for the training andtesting and the set B of 223 types to be used for testing ofthe unknown types +e ratio of the two divisions wasapproximately 9 1 For signal types in A dividing PDWlists belonging to each type into 3 groups with a ratio of 7 1 2 three data sets for training validation and testing ofknown types were composed +e training data were usedto train the model and the validation data were used toselect the model parameters and determine the finalmodel Validation data are needed to prevent the trained
model from overfitting with training data Dropout [21] isalso tested which is a method for preventing overfitting byprobably dropping nodes of hidden layers during learn-ing +e test data are not used for training and insteadthey are used to evaluate the performance of the finalmodel Lastly the performance of the final model isassessed using the PDW lists of the unknown radar signaltype in the set B Figure 3 shows the configuration of thedata sets
Based on the configuration of the data sets shown inFigure 3 a flowchart to show the overall organization in theremaining sections is given in Figure 4 for easy readabilityBased on train and validation data of signal types of set A theconstruction of a LSTM model is explained in Section 4 andthe construction of feature extraction and a deep neuralnetwork is described in Section 5 Test data of set A whichrepresents the known signal types and test data of set Bwhich indicates the unknown signal types are used for theperformance comparison of two models in Section 6
4 Construction of a Jamming PredictionModelBased on LSTM
In this section the construction of a LSTM model whichtakes a PDW list of a threat signal as input is explained
41ModelParameters inLSTM Preliminary experiment wascarried out using the small scale of radar signal data to selectthe LSTM model structure and parameter values [22] +edata used in the preliminary test had 1157500 PDW lists fora total of 640 radar signal types and 4 types of jammingmethods +e basic model was composed of an input layerwith 5 nodes for the feature values of AOA AMP RF PRIand PW a LSTM layer consisting of 200 nodes and theoutput layer consisting of 4 nodes A process of determiningan optimization method minibatch size dropout ratio thenumber of the LSTM layers and fully connected layers anddecay ratio of learning rate was performed Table 1 shows theparameters and their values which were tested Since therewere somany combinations of the parameter values we haveadopted a strategy that determines the parameter valuessequentially When testing optimization methods in step 1all other parameters were set to the underlined values in thethird column of Table 1 After the best optimization methodis selected it is fixed in the next steps where the remainingparameter values are determined +is process is performedin sequence up to step 7 in Table 1 +e fourth column ofTable 1 shows the selected parameter values
42 LSTM Model With reference to the model structureconfigured in Table 1 the final LSTM model was composedof an input layer with 3 nodes for RF PRI and PW 2 LSTMlayers with 200 nodes one fully connected layer with 400nodes and an output layer Since one among the eightjamming techniques needs to be predicted for each inputthreat signal the output layer has 8 nodes +rough addi-tional tests it was determined to apply the batch normali-zation in the fully connected layer and gradient clipping for
Security and Communication Networks 3
Pulse
f1 f2 f3
t2t1 t3
p1 p2 p3
i1 i2 i3
Figure 1 A radar signal represented as a sequence of pulses
Fix Agile Hopping
TOA
RF
Sine ndashsawtooth +sawtooth
(a)
Stable Jitter DnsPRI
TOASine ndashsawtooth +sawtooth Stagger
(b)
Figure 2 Various types of RF and PRI (a) 6 types of RF (b) 7 types of PRI
4 Security and Communication Networks
stable learning Figure 5 shows the structure of the finalLSTM model
5 Construction of Jamming Prediction ModelBased on Feature Extraction and a DeepNeural Network
In this section feature extraction from a PDW list of a threatsignal and the construction of a deep neural network areexplained in detail
51 Feature Extraction From sequence data expressed asx1 xn features such as statistical values autocorrelationcoefficients and MFCCs were extracted
Four statistic values of the mean x standard deviation sskewness w and kurtosis k were computed from the se-quence data such as
x 1n
1113944
n
1xt
s
1n
1113944
n
1xt minus x( 1113857
2
11139741113972
w 1n 1113936
n1 xt minus x( 1113857
31113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
32
k 1n1113936
n1 xt minus x( 1113857
41113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
2
(2)
+e autocorrelation coefficient R(τ) represents the linearrelevance between two subsequences with a lag τ which iscomputed by
A 2035 signal types B 223 types
Train4324600 PDW lists
Validation617800
Test1235600
692000PDW lists
Figure 3 +e configuration of the data sets
A 2035 signal typesTrain
4324600 PDW lists
Validation617800 PDW lists
Section 4 constructionof a LSTM model
Section 5 constructionof feature extraction
and a deep neural network
Section 6 comparison of predictionperformance by two models
A 2035 known signal typesTest
1235600 PDW lists
B 223 unknown signal types
Test692000 PDW lists
Figure 4 A flowchart to explain the overall organization
Table 1 Parameters of a LSTM model
Step Parameters Tested values Selectedvalue Note
1 Optimizationmethod
RMSProp AdamAdadelta Adam Initial learning rate was set as 0003 for RMSProp [23] Adam [24] and 10 for
Adadelta [25]2 Minibatch size 50 100 200 200
3 Dropout ratio () 0 10 30 50 0 Dropout ratio means the rate at which the output gate units in a LSTM layerare randomly removed
4 LSTM layers 1 2 2 +e model with 2 LSTM layers showed higher accuracy than the model of 1LSTM layer
5 Fully connectedlayer 0 1 1 +e model with the fully connected layer had a higher accuracy than the
model with no fully connected layer
6 Input features 3 5 3 Higher accuracy was obtained when using three features RF PRI and PWinstead of using 5 features AOA AMP RF PRI and PW
7 Decay ratio Use no use Use When gradually decreasing the learning rate by multiplying 09 to theprevious learning rate per epoch after epoch 10 higher accuracy was obtained
Security and Communication Networks 5
R(τ) 1113936
nminus τ1 xt minus x( 1113857 xt+τ minus x( 1113857
1113936n1 xt minus x( 1113857
2 (3)
Various autocorrelation coefficients can be obtained bychanging the time difference τ [4]
MFCC extracts feature values from the sequence datathrough the process of framing Fast Fourier transform(FFT) mel filter bank log function and inverse FFT [5] Inthe framing step the sequence is divided into segments ofequal size FFT is applied to a subsequence of each segmentto convert to the frequency domain and then the powerspectrum is obtained +e mel spectrum is obtained byapplying mel filters to the power spectrum and the coef-ficients are obtained by inverse FFT after applying the logfunction to the mel spectrum +ese coefficients are calledMFCCs All or some of the coefficients obtained in eachsegment can be used as the feature values [5]
52 Composition of Feature Sets We extracted feature valuesusing RF PRI and PW sequences of a PDW list Bycomputing the mean standard deviation skewness andkurtosis from each sequence a total of 12 feature values wereobtained from a PDW list By increasing the time differenceτ from 1 to 10 by incrementing 1 we extracted 10 auto-correlation coefficients from the RF and PRI sequences andtherefore a total of 20 feature values were extracted from aPDW list Also 10 MFCCs were computed for the RF andPRI sequences so that a total of 20 values were extracted+eautocorrelation coefficients and MFCCs were computedusing Python libraries such as the function acf of statsmodelsand the function mfcc of librosa [26]
Using statistical features as the basic features andcombining the autocorrelation coefficients and MFCCs withbasic features we composed 4 feature sets as shown in
Table 2 and tested prediction performance of a neuralnetwork model on each feature set In the feature set F1 aradar signal is represented with only statistical features +efeature set F2 contains the statistical features and autocor-relation coefficients while F3 is composed of the statisticalfeatures and MFCCs +e feature set F4 contains all of thestatistical features autocorrelation coefficients and MFCCsresulting in a total of 52 feature values
53 Performance of Neural Network Models on VariousFeature Sets To select a neural network structure com-parative testing for 4 different models was performedvarying the activation function and the number of layers asshown in Table 3 Considering test results in Section 41Adam was used as the optimization method and the min-ibatch size was set to 200
+e 4 feature sets in Table 2 and the 4 models in Table 3were used to compare the performance +e total number oftraining epochs was set to 15 and the training data com-posed in Section 32 were used to train the models +eaccuracy of the validation data was evaluated in each epochTable 4 shows the training and validation accuracies for 16combinations of 4 feature sets and 4 NNmodels which wereevaluated in the epoch with the highest accuracy for thevalidation data It revealed that model 4 on feature set F3 wasfound to have the highest performance Hence model 4 onfeature set F3 was chosen as the final model Figure 6 de-scribes the structure of the final deep neural network model
6 Performance Comparison of LSTM Modeland Deep Neural Network Model UsingExtracted Features
In this section we compare the prediction performance andtraining time complexity of two deep learning models fortest data of the known signal types in set A and test data ofthe unknown signal type in set B As explained in Section 322258 signal types were randomly divided into the set A to beused for the training validation and testing and the set B tobe used for testing of the unknown signal types Now werepeat the random division 10 times and measure the av-erage accuracy for performance comparison of two ap-proaches using LSTM and using a deep neural network withextracted features (denoted as DNNEF)
Table 5 shows test accuracy of known signal types in Aand unknown types in B from 10 repeated experiments Intesting of known signal types the average accuracy of 9846was obtained for the NN with extracted features and theaverage accuracy of 9936 was obtained for the LSTMmodel +e average accuracy for the unknown types whenusing the neural network with extracted features was 9245while the average accuracy for the LSTMmodel was 9353In both cases the LSTM model showed a little higheraccuracy
+e paired t-test was carried out in order to determinewhether the higher average accuracy of the LSTM model isstatistically significant +e paired t-test for the accuracy ofknown types gave the p value 00000612 which implies
Output layer 8 nodes
Input layer 3 nodes
RF PRI PWPDW list
Fully connected layer (400 nodes)
LSTM layer (200 nodes)
LSTM layer (200 nodes)
Figure 5 Structures of the final LSTM model
6 Security and Communication Networks
statistical significance in the difference between the accu-racies of the two models in the significance level of 001However for the unknown type accuracy the p value by thepaired t-test was approximately 0215 and so there was nosignificant difference in the significance level of 001 +isresult was thought to be due to the very large deviation in theaccuracy of the LSTM model for the unknown type whichwas a minimum of 8969 and maximum of 9720
In order to compare the learning time of the twomethods the average training time per epoch was measured+e specification of the computer used to measure the timeconsumed was CPU Intel i7-7700 360GHz and RAM320GB and GPU NVIDIA GeForce GTX 1080 Ti Table 6compares the execution time for model training for 15epochs+e process of extracting the features from the PDWlists was conducted once before the training of the deepneural network model As shown in Table 6 the time
consumed for training the LSTM model for 15 epochs wasapproximately 142 times longer than that of the deep neuralnetwork
Table 3 Four deep neural network models
Parameters Model 1 Model 2 Model 3 Model 4Activation function tanh relu tanh reluHidden layers 2 4Hidden units 400Learning rate Initially set as 00005
Learning rate decay Multiplying the previous learning rate by097 for every 1000000 samples
Table 4 +e performance in 16 combinations of feature sets andmodels
Model Feature set Training accuracy()
Validation accuracy()
Model 1
F1 9572 9592F2 9677 9709F3 9735 9778F4 9729 9778
Model 2
F1 9567 9572F2 9681 9730F3 9725 9755F4 9719 9761
Model 3
F1 9689 9735F2 9721 9765F3 9742 9804F4 9730 9778
Model 4
F1 9684 9748F2 9780 9819F3 9782 9847F4 9777 9802
Input layer 32 nodes
Hidden layer (400 nodes)
Output layer 8 nodes
Hidden layer (400 nodes)
Hidden layer (400 nodes)
Hidden layer (400 nodes)
32 feature values extracted from a PDW list
RF PRI PW
Figure 6 Structure of the final deep neural network
Table 5 Test accuracy () of known types and unknown types
DNNEF LSTM modelKnowntypes
Unknowntypes
Knowntypes
Unknowntypes
1 9849 9333 9972 97202 9835 9247 9970 95093 9849 9057 9878 90764 9860 9281 9906 90945 9861 9272 9932 93426 9855 9431 9960 95067 9839 9262 9971 96738 9847 9211 9901 89699 9831 9219 9900 906410 9833 9132 9965 9576Avg 9846 9245 9936 9353
Table 6 Comparison of execution time in seconds
Model Feature extraction Training for 1 epoch Training for 15epochs
DNN 7374233 1932943 36368378LSTM mdash 3454399 51815985
Table 2 Composition of four feature sets
Featuresets
Statisticalfeatures
Autocorrelationcoef MFCC Total
F1 12 mdash mdash 12F2 12 20 mdash 32F3 12 mdash 20 32F4 12 20 20 52
Security and Communication Networks 7
7 Conclusion
In this study we applied deep learning for jamming pre-diction for a radar signal Two methods were comparedusing a deep neural network model based on the featuresextracted from radar signals and using LSTMmodel withoutthe feature extraction process
+e deep neural network requires feature extraction inadvance and this is conducted once before the training ofthe deep neural network model +e training speed of themodel per epoch was faster than that of the LSTM modelHowever the feature extraction method has to be selectedappropriately because the model prediction performancevaries depending on the feature values extracted from thePDW list +e LSTMmodel has the advantages of being ableto perform training and prediction by directly using thePDW list of the radar signal without the extraction of featurevalues +e prediction accuracy of the LSTM model washigher on average than that of the deep neural networkmodel However there is the disadvantage that the trainingtime takes longer than a deep neural network model trainedon the extracted features
Testing results demonstrate that the jamming methodcan be predicted for an unknown type of radar signal with anaverage accuracy of approximately 92 and higher It showsthat deep learning methods can be used effectively forpredicting an appropriate jamming technique for new radarsignals which are not used in model training
Data Availability
+e radar signal data used to support the findings of thisstudy have not been made available because of the policy ofAgency for Defense Development Korea that data distri-bution is limited
Conflicts of Interest
+e authors declare that there are no conflicts of interestregarding the publication of this paper
References
[1] M S R Lothes M B Szymanski and R G Wiley RadarVulnerability to Jamming Artech House Boston MA USA1990
[2] V Iglesias J Grajal O Yeste-Ojeda M Garrido M Sanchezand M Lopez-Vallejo ldquoReal-time radar pulse parameterextractorrdquo in Proceedings of the IEEE Radar ConferenceCincinnati OH USA May 2014
[3] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[4] P DrsquoUrso and E Maharaj ldquoAutocorrelation-based fuzzyclustering of time seriesrdquo Fuzzy Sets and Systems vol 160no 24 pp 3565ndash3589 2009
[5] S Gupta J Jaafar W F wan Ahmad and A Bansal ldquoFeatureextraction using MFCCrdquo Signal amp Image Processing An In-ternational Journal vol 4 no 4 pp 101ndash108 2013
[6] R Shumway and D Stoffer Time Series Analysis and itsApplication with R Examples Springer Berlin Germany2000
[7] Z Xia S Xia L Wan and S Cai ldquoSpectral regression basedfault feature extraction for bearing accelerometer sensorsignalsrdquo Sensors vol 12 no 10 pp 13694ndash13719 2012
[8] T W Rauber F de Assis Boldt and F M Varejao ldquoHet-erogeneous feature models and feature selection applied tobearing fault diagnosisrdquo IEEE Transactions on IndustrialElectronics vol 62 no 1 pp 637ndash646 2015
[9] B R Nayana and P Geethanjali ldquoAnalysis of statistical time-domain features effectiveness in identification of bearingfaults from vibration signalrdquo IEEE Sensors Journal vol 17no 17 pp 5618ndash5625 2017
[10] M B L Muda and I Elamvazuthi ldquoVoice recognition al-gorithm using mel frequency cepstral coefficient (MFCC) anddynamic time warping (DTW) techniquesrdquo Journal ofComputing vol 2 no 3 pp 138ndash143 2010
[11] S Q L Jian-xun and Y Hai ldquoSignal feature analysis andexperimental verification of radar deception jammingrdquo inProceedings of the IEEE CIE International Conference onRadar Chengdu China October 2011
[12] A S A Mendoza and B Flores ldquoClassification of radarjammer FM signals using a neural networkrdquo in Proceedings ofthe SPIE Radar Sensor Technology XXI vol 10188 May 2017
[13] G Noone ldquoA neural approach to automatic pulse repetitioninterval modulation recognitionrdquo in Proceedings of the 1999Information Decision and Control Data and InformationFusion Symposium Signal Processing and CommunicationsSymposium and Decision and Control Symposium Proceedings(Cat No99EX251) Adelaide Australia February 1999
[14] J Kauppi and K Martikainen ldquoAn efficient set of features forpulse repetition interval modulation recognitionrdquo in Pro-ceedings of the IET International Conference on Radar SystemsEdinburgh UK October 2007
[15] I Jordanov N Petrov and A Petrozziello ldquoSupervised radarsignal classificationrdquo in Proceedings of the International JointConference on Neural Networks Vancouver Canada July2016
[16] D Zhou X Wang Y Tian and R Wang ldquoA novel radarsignal recognition method based on a deep restricted Boltz-mann machinerdquo Engineering Review vol 37 pp 165ndash1712017
[17] Z Y Z Wu Y Zhao and H Luo ldquoJamming signals clas-sification using convolutional neural networkrdquo in Proceedingsof the IEEE International Symposium on Signal Processing andInformation Technology (ISSPIT) Bilbao Spain December2017
[18] R Ferre A Fuente and E Lohan ldquoJammer classification inGNSS bands via machine learning algorithmsrdquo Sensorsvol 19 no 22 p 4841 2019
[19] J Mun H Kim and J Lee ldquoA deep learning approach forautomotive radar interference mitigationrdquo 2019 httpsarxivorgabs190306380
[20] E Masen B Yonei and B Yazici ldquoDeep learning for radarrdquoin Proceedings of the 2017 IEEE Radar Conference(RadarConf) Seattle WA USA May 2017
[21] N Srivastava G Hinton A Krizhevsky I Sutskever andR Salakhutdinov ldquoDropout a simple way to prevent neuralnetworks from overfittingrdquo e Journal of Machine LearningResearch vol 15 no 1 pp 1929ndash1958 2014
[22] G Lee ldquoRadar jamming technique prediction using deeplearningrdquo Master thesis Chungnam National UniversityDaejeon Korea 2019
[23] M Zeiler ldquoADADELTA an adaptive learning rate methodrdquo2012 httpsarxivorgabs12125701
8 Security and Communication Networks
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9
31 Radar Signals A radar signal can be represented as asequence of pulses as shown in Figure 1 For each pulse 5characteristic values pk for PW ik for PRI fk for RF ak forAOA (angle of attack) and mk for AMP (amplitude) areobtained and a radar signal can be stored as a PDW list
X1 p1 i1 f1 a1 m1( 1113857 Xn pn in fn an mn( 11138571113858 1113859
(1)
Generally radar signals have modulations in the RF andPRI in order to evade to be detected RFmodulation refers tochanges in a sequence [f1 fn] of pulse frequencyvalues +ere are 6 types of RF modulation including fixagile hopping sine minus sawtooth and +sawtooth as shown inFigure 2(a) PRI modulation refers to changes in a sequence[i1 in] of pulse repetition interval values and there are 7types of PRI modulation including stable jitter staggerdwell and switch sine minus sawtooth and +sawtooth as shownin Figure 2(b)
32 Experimental Setting With reference to the specifica-tions of the currently operating radar models 2258 signaltypes were configured with different types and magnitudesof RF and PRI Each type had a minimum of 1000 andmaximum of 9000 PDW lists and a total of 6870000PDW lists were composed A total of eight jammingtechniques corresponding to 2258 signal types wereestablished by assigning one jamming technique to eachsignal type based on the information of the library+e goalof a deep learning model is set to predict one jammingtechnique among eight jamming techniques for threatsignal input Hence the output layer in a deep learningmodel consists of eight nodes and each jamming techniqueis encoded as an 8-dimensional binary vector by the one-hot-encoding method where only one component is 1 andthe others are 0 In the training and testing stages of a deeplearning model the values in the output layer which areforwarded from a previous layer are translated as proba-bility values for each jamming technique by the softmaxfunction
A deep learning model is trained with training datawhich is composed of the pairs of radar signals andcorresponding jamming techniques In the testing stageone among eight jamming techniques is predicted as anoutput of the constructed model when a radar signal isgiven as an input To evaluate the prediction performanceof a deep learning model for an unknown radar signal typethat is not used in the training stage as well as for a knownsignal type we randomly divided 2258 signal types intothe set A of 2035 types to be used for the training andtesting and the set B of 223 types to be used for testing ofthe unknown types +e ratio of the two divisions wasapproximately 9 1 For signal types in A dividing PDWlists belonging to each type into 3 groups with a ratio of 7 1 2 three data sets for training validation and testing ofknown types were composed +e training data were usedto train the model and the validation data were used toselect the model parameters and determine the finalmodel Validation data are needed to prevent the trained
model from overfitting with training data Dropout [21] isalso tested which is a method for preventing overfitting byprobably dropping nodes of hidden layers during learn-ing +e test data are not used for training and insteadthey are used to evaluate the performance of the finalmodel Lastly the performance of the final model isassessed using the PDW lists of the unknown radar signaltype in the set B Figure 3 shows the configuration of thedata sets
Based on the configuration of the data sets shown inFigure 3 a flowchart to show the overall organization in theremaining sections is given in Figure 4 for easy readabilityBased on train and validation data of signal types of set A theconstruction of a LSTM model is explained in Section 4 andthe construction of feature extraction and a deep neuralnetwork is described in Section 5 Test data of set A whichrepresents the known signal types and test data of set Bwhich indicates the unknown signal types are used for theperformance comparison of two models in Section 6
4 Construction of a Jamming PredictionModelBased on LSTM
In this section the construction of a LSTM model whichtakes a PDW list of a threat signal as input is explained
41ModelParameters inLSTM Preliminary experiment wascarried out using the small scale of radar signal data to selectthe LSTM model structure and parameter values [22] +edata used in the preliminary test had 1157500 PDW lists fora total of 640 radar signal types and 4 types of jammingmethods +e basic model was composed of an input layerwith 5 nodes for the feature values of AOA AMP RF PRIand PW a LSTM layer consisting of 200 nodes and theoutput layer consisting of 4 nodes A process of determiningan optimization method minibatch size dropout ratio thenumber of the LSTM layers and fully connected layers anddecay ratio of learning rate was performed Table 1 shows theparameters and their values which were tested Since therewere somany combinations of the parameter values we haveadopted a strategy that determines the parameter valuessequentially When testing optimization methods in step 1all other parameters were set to the underlined values in thethird column of Table 1 After the best optimization methodis selected it is fixed in the next steps where the remainingparameter values are determined +is process is performedin sequence up to step 7 in Table 1 +e fourth column ofTable 1 shows the selected parameter values
42 LSTM Model With reference to the model structureconfigured in Table 1 the final LSTM model was composedof an input layer with 3 nodes for RF PRI and PW 2 LSTMlayers with 200 nodes one fully connected layer with 400nodes and an output layer Since one among the eightjamming techniques needs to be predicted for each inputthreat signal the output layer has 8 nodes +rough addi-tional tests it was determined to apply the batch normali-zation in the fully connected layer and gradient clipping for
Security and Communication Networks 3
Pulse
f1 f2 f3
t2t1 t3
p1 p2 p3
i1 i2 i3
Figure 1 A radar signal represented as a sequence of pulses
Fix Agile Hopping
TOA
RF
Sine ndashsawtooth +sawtooth
(a)
Stable Jitter DnsPRI
TOASine ndashsawtooth +sawtooth Stagger
(b)
Figure 2 Various types of RF and PRI (a) 6 types of RF (b) 7 types of PRI
4 Security and Communication Networks
stable learning Figure 5 shows the structure of the finalLSTM model
5 Construction of Jamming Prediction ModelBased on Feature Extraction and a DeepNeural Network
In this section feature extraction from a PDW list of a threatsignal and the construction of a deep neural network areexplained in detail
51 Feature Extraction From sequence data expressed asx1 xn features such as statistical values autocorrelationcoefficients and MFCCs were extracted
Four statistic values of the mean x standard deviation sskewness w and kurtosis k were computed from the se-quence data such as
x 1n
1113944
n
1xt
s
1n
1113944
n
1xt minus x( 1113857
2
11139741113972
w 1n 1113936
n1 xt minus x( 1113857
31113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
32
k 1n1113936
n1 xt minus x( 1113857
41113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
2
(2)
+e autocorrelation coefficient R(τ) represents the linearrelevance between two subsequences with a lag τ which iscomputed by
A 2035 signal types B 223 types
Train4324600 PDW lists
Validation617800
Test1235600
692000PDW lists
Figure 3 +e configuration of the data sets
A 2035 signal typesTrain
4324600 PDW lists
Validation617800 PDW lists
Section 4 constructionof a LSTM model
Section 5 constructionof feature extraction
and a deep neural network
Section 6 comparison of predictionperformance by two models
A 2035 known signal typesTest
1235600 PDW lists
B 223 unknown signal types
Test692000 PDW lists
Figure 4 A flowchart to explain the overall organization
Table 1 Parameters of a LSTM model
Step Parameters Tested values Selectedvalue Note
1 Optimizationmethod
RMSProp AdamAdadelta Adam Initial learning rate was set as 0003 for RMSProp [23] Adam [24] and 10 for
Adadelta [25]2 Minibatch size 50 100 200 200
3 Dropout ratio () 0 10 30 50 0 Dropout ratio means the rate at which the output gate units in a LSTM layerare randomly removed
4 LSTM layers 1 2 2 +e model with 2 LSTM layers showed higher accuracy than the model of 1LSTM layer
5 Fully connectedlayer 0 1 1 +e model with the fully connected layer had a higher accuracy than the
model with no fully connected layer
6 Input features 3 5 3 Higher accuracy was obtained when using three features RF PRI and PWinstead of using 5 features AOA AMP RF PRI and PW
7 Decay ratio Use no use Use When gradually decreasing the learning rate by multiplying 09 to theprevious learning rate per epoch after epoch 10 higher accuracy was obtained
Security and Communication Networks 5
R(τ) 1113936
nminus τ1 xt minus x( 1113857 xt+τ minus x( 1113857
1113936n1 xt minus x( 1113857
2 (3)
Various autocorrelation coefficients can be obtained bychanging the time difference τ [4]
MFCC extracts feature values from the sequence datathrough the process of framing Fast Fourier transform(FFT) mel filter bank log function and inverse FFT [5] Inthe framing step the sequence is divided into segments ofequal size FFT is applied to a subsequence of each segmentto convert to the frequency domain and then the powerspectrum is obtained +e mel spectrum is obtained byapplying mel filters to the power spectrum and the coef-ficients are obtained by inverse FFT after applying the logfunction to the mel spectrum +ese coefficients are calledMFCCs All or some of the coefficients obtained in eachsegment can be used as the feature values [5]
52 Composition of Feature Sets We extracted feature valuesusing RF PRI and PW sequences of a PDW list Bycomputing the mean standard deviation skewness andkurtosis from each sequence a total of 12 feature values wereobtained from a PDW list By increasing the time differenceτ from 1 to 10 by incrementing 1 we extracted 10 auto-correlation coefficients from the RF and PRI sequences andtherefore a total of 20 feature values were extracted from aPDW list Also 10 MFCCs were computed for the RF andPRI sequences so that a total of 20 values were extracted+eautocorrelation coefficients and MFCCs were computedusing Python libraries such as the function acf of statsmodelsand the function mfcc of librosa [26]
Using statistical features as the basic features andcombining the autocorrelation coefficients and MFCCs withbasic features we composed 4 feature sets as shown in
Table 2 and tested prediction performance of a neuralnetwork model on each feature set In the feature set F1 aradar signal is represented with only statistical features +efeature set F2 contains the statistical features and autocor-relation coefficients while F3 is composed of the statisticalfeatures and MFCCs +e feature set F4 contains all of thestatistical features autocorrelation coefficients and MFCCsresulting in a total of 52 feature values
53 Performance of Neural Network Models on VariousFeature Sets To select a neural network structure com-parative testing for 4 different models was performedvarying the activation function and the number of layers asshown in Table 3 Considering test results in Section 41Adam was used as the optimization method and the min-ibatch size was set to 200
+e 4 feature sets in Table 2 and the 4 models in Table 3were used to compare the performance +e total number oftraining epochs was set to 15 and the training data com-posed in Section 32 were used to train the models +eaccuracy of the validation data was evaluated in each epochTable 4 shows the training and validation accuracies for 16combinations of 4 feature sets and 4 NNmodels which wereevaluated in the epoch with the highest accuracy for thevalidation data It revealed that model 4 on feature set F3 wasfound to have the highest performance Hence model 4 onfeature set F3 was chosen as the final model Figure 6 de-scribes the structure of the final deep neural network model
6 Performance Comparison of LSTM Modeland Deep Neural Network Model UsingExtracted Features
In this section we compare the prediction performance andtraining time complexity of two deep learning models fortest data of the known signal types in set A and test data ofthe unknown signal type in set B As explained in Section 322258 signal types were randomly divided into the set A to beused for the training validation and testing and the set B tobe used for testing of the unknown signal types Now werepeat the random division 10 times and measure the av-erage accuracy for performance comparison of two ap-proaches using LSTM and using a deep neural network withextracted features (denoted as DNNEF)
Table 5 shows test accuracy of known signal types in Aand unknown types in B from 10 repeated experiments Intesting of known signal types the average accuracy of 9846was obtained for the NN with extracted features and theaverage accuracy of 9936 was obtained for the LSTMmodel +e average accuracy for the unknown types whenusing the neural network with extracted features was 9245while the average accuracy for the LSTMmodel was 9353In both cases the LSTM model showed a little higheraccuracy
+e paired t-test was carried out in order to determinewhether the higher average accuracy of the LSTM model isstatistically significant +e paired t-test for the accuracy ofknown types gave the p value 00000612 which implies
Output layer 8 nodes
Input layer 3 nodes
RF PRI PWPDW list
Fully connected layer (400 nodes)
LSTM layer (200 nodes)
LSTM layer (200 nodes)
Figure 5 Structures of the final LSTM model
6 Security and Communication Networks
statistical significance in the difference between the accu-racies of the two models in the significance level of 001However for the unknown type accuracy the p value by thepaired t-test was approximately 0215 and so there was nosignificant difference in the significance level of 001 +isresult was thought to be due to the very large deviation in theaccuracy of the LSTM model for the unknown type whichwas a minimum of 8969 and maximum of 9720
In order to compare the learning time of the twomethods the average training time per epoch was measured+e specification of the computer used to measure the timeconsumed was CPU Intel i7-7700 360GHz and RAM320GB and GPU NVIDIA GeForce GTX 1080 Ti Table 6compares the execution time for model training for 15epochs+e process of extracting the features from the PDWlists was conducted once before the training of the deepneural network model As shown in Table 6 the time
consumed for training the LSTM model for 15 epochs wasapproximately 142 times longer than that of the deep neuralnetwork
Table 3 Four deep neural network models
Parameters Model 1 Model 2 Model 3 Model 4Activation function tanh relu tanh reluHidden layers 2 4Hidden units 400Learning rate Initially set as 00005
Learning rate decay Multiplying the previous learning rate by097 for every 1000000 samples
Table 4 +e performance in 16 combinations of feature sets andmodels
Model Feature set Training accuracy()
Validation accuracy()
Model 1
F1 9572 9592F2 9677 9709F3 9735 9778F4 9729 9778
Model 2
F1 9567 9572F2 9681 9730F3 9725 9755F4 9719 9761
Model 3
F1 9689 9735F2 9721 9765F3 9742 9804F4 9730 9778
Model 4
F1 9684 9748F2 9780 9819F3 9782 9847F4 9777 9802
Input layer 32 nodes
Hidden layer (400 nodes)
Output layer 8 nodes
Hidden layer (400 nodes)
Hidden layer (400 nodes)
Hidden layer (400 nodes)
32 feature values extracted from a PDW list
RF PRI PW
Figure 6 Structure of the final deep neural network
Table 5 Test accuracy () of known types and unknown types
DNNEF LSTM modelKnowntypes
Unknowntypes
Knowntypes
Unknowntypes
1 9849 9333 9972 97202 9835 9247 9970 95093 9849 9057 9878 90764 9860 9281 9906 90945 9861 9272 9932 93426 9855 9431 9960 95067 9839 9262 9971 96738 9847 9211 9901 89699 9831 9219 9900 906410 9833 9132 9965 9576Avg 9846 9245 9936 9353
Table 6 Comparison of execution time in seconds
Model Feature extraction Training for 1 epoch Training for 15epochs
DNN 7374233 1932943 36368378LSTM mdash 3454399 51815985
Table 2 Composition of four feature sets
Featuresets
Statisticalfeatures
Autocorrelationcoef MFCC Total
F1 12 mdash mdash 12F2 12 20 mdash 32F3 12 mdash 20 32F4 12 20 20 52
Security and Communication Networks 7
7 Conclusion
In this study we applied deep learning for jamming pre-diction for a radar signal Two methods were comparedusing a deep neural network model based on the featuresextracted from radar signals and using LSTMmodel withoutthe feature extraction process
+e deep neural network requires feature extraction inadvance and this is conducted once before the training ofthe deep neural network model +e training speed of themodel per epoch was faster than that of the LSTM modelHowever the feature extraction method has to be selectedappropriately because the model prediction performancevaries depending on the feature values extracted from thePDW list +e LSTMmodel has the advantages of being ableto perform training and prediction by directly using thePDW list of the radar signal without the extraction of featurevalues +e prediction accuracy of the LSTM model washigher on average than that of the deep neural networkmodel However there is the disadvantage that the trainingtime takes longer than a deep neural network model trainedon the extracted features
Testing results demonstrate that the jamming methodcan be predicted for an unknown type of radar signal with anaverage accuracy of approximately 92 and higher It showsthat deep learning methods can be used effectively forpredicting an appropriate jamming technique for new radarsignals which are not used in model training
Data Availability
+e radar signal data used to support the findings of thisstudy have not been made available because of the policy ofAgency for Defense Development Korea that data distri-bution is limited
Conflicts of Interest
+e authors declare that there are no conflicts of interestregarding the publication of this paper
References
[1] M S R Lothes M B Szymanski and R G Wiley RadarVulnerability to Jamming Artech House Boston MA USA1990
[2] V Iglesias J Grajal O Yeste-Ojeda M Garrido M Sanchezand M Lopez-Vallejo ldquoReal-time radar pulse parameterextractorrdquo in Proceedings of the IEEE Radar ConferenceCincinnati OH USA May 2014
[3] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[4] P DrsquoUrso and E Maharaj ldquoAutocorrelation-based fuzzyclustering of time seriesrdquo Fuzzy Sets and Systems vol 160no 24 pp 3565ndash3589 2009
[5] S Gupta J Jaafar W F wan Ahmad and A Bansal ldquoFeatureextraction using MFCCrdquo Signal amp Image Processing An In-ternational Journal vol 4 no 4 pp 101ndash108 2013
[6] R Shumway and D Stoffer Time Series Analysis and itsApplication with R Examples Springer Berlin Germany2000
[7] Z Xia S Xia L Wan and S Cai ldquoSpectral regression basedfault feature extraction for bearing accelerometer sensorsignalsrdquo Sensors vol 12 no 10 pp 13694ndash13719 2012
[8] T W Rauber F de Assis Boldt and F M Varejao ldquoHet-erogeneous feature models and feature selection applied tobearing fault diagnosisrdquo IEEE Transactions on IndustrialElectronics vol 62 no 1 pp 637ndash646 2015
[9] B R Nayana and P Geethanjali ldquoAnalysis of statistical time-domain features effectiveness in identification of bearingfaults from vibration signalrdquo IEEE Sensors Journal vol 17no 17 pp 5618ndash5625 2017
[10] M B L Muda and I Elamvazuthi ldquoVoice recognition al-gorithm using mel frequency cepstral coefficient (MFCC) anddynamic time warping (DTW) techniquesrdquo Journal ofComputing vol 2 no 3 pp 138ndash143 2010
[11] S Q L Jian-xun and Y Hai ldquoSignal feature analysis andexperimental verification of radar deception jammingrdquo inProceedings of the IEEE CIE International Conference onRadar Chengdu China October 2011
[12] A S A Mendoza and B Flores ldquoClassification of radarjammer FM signals using a neural networkrdquo in Proceedings ofthe SPIE Radar Sensor Technology XXI vol 10188 May 2017
[13] G Noone ldquoA neural approach to automatic pulse repetitioninterval modulation recognitionrdquo in Proceedings of the 1999Information Decision and Control Data and InformationFusion Symposium Signal Processing and CommunicationsSymposium and Decision and Control Symposium Proceedings(Cat No99EX251) Adelaide Australia February 1999
[14] J Kauppi and K Martikainen ldquoAn efficient set of features forpulse repetition interval modulation recognitionrdquo in Pro-ceedings of the IET International Conference on Radar SystemsEdinburgh UK October 2007
[15] I Jordanov N Petrov and A Petrozziello ldquoSupervised radarsignal classificationrdquo in Proceedings of the International JointConference on Neural Networks Vancouver Canada July2016
[16] D Zhou X Wang Y Tian and R Wang ldquoA novel radarsignal recognition method based on a deep restricted Boltz-mann machinerdquo Engineering Review vol 37 pp 165ndash1712017
[17] Z Y Z Wu Y Zhao and H Luo ldquoJamming signals clas-sification using convolutional neural networkrdquo in Proceedingsof the IEEE International Symposium on Signal Processing andInformation Technology (ISSPIT) Bilbao Spain December2017
[18] R Ferre A Fuente and E Lohan ldquoJammer classification inGNSS bands via machine learning algorithmsrdquo Sensorsvol 19 no 22 p 4841 2019
[19] J Mun H Kim and J Lee ldquoA deep learning approach forautomotive radar interference mitigationrdquo 2019 httpsarxivorgabs190306380
[20] E Masen B Yonei and B Yazici ldquoDeep learning for radarrdquoin Proceedings of the 2017 IEEE Radar Conference(RadarConf) Seattle WA USA May 2017
[21] N Srivastava G Hinton A Krizhevsky I Sutskever andR Salakhutdinov ldquoDropout a simple way to prevent neuralnetworks from overfittingrdquo e Journal of Machine LearningResearch vol 15 no 1 pp 1929ndash1958 2014
[22] G Lee ldquoRadar jamming technique prediction using deeplearningrdquo Master thesis Chungnam National UniversityDaejeon Korea 2019
[23] M Zeiler ldquoADADELTA an adaptive learning rate methodrdquo2012 httpsarxivorgabs12125701
8 Security and Communication Networks
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9
Pulse
f1 f2 f3
t2t1 t3
p1 p2 p3
i1 i2 i3
Figure 1 A radar signal represented as a sequence of pulses
Fix Agile Hopping
TOA
RF
Sine ndashsawtooth +sawtooth
(a)
Stable Jitter DnsPRI
TOASine ndashsawtooth +sawtooth Stagger
(b)
Figure 2 Various types of RF and PRI (a) 6 types of RF (b) 7 types of PRI
4 Security and Communication Networks
stable learning Figure 5 shows the structure of the finalLSTM model
5 Construction of Jamming Prediction ModelBased on Feature Extraction and a DeepNeural Network
In this section feature extraction from a PDW list of a threatsignal and the construction of a deep neural network areexplained in detail
51 Feature Extraction From sequence data expressed asx1 xn features such as statistical values autocorrelationcoefficients and MFCCs were extracted
Four statistic values of the mean x standard deviation sskewness w and kurtosis k were computed from the se-quence data such as
x 1n
1113944
n
1xt
s
1n
1113944
n
1xt minus x( 1113857
2
11139741113972
w 1n 1113936
n1 xt minus x( 1113857
31113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
32
k 1n1113936
n1 xt minus x( 1113857
41113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
2
(2)
+e autocorrelation coefficient R(τ) represents the linearrelevance between two subsequences with a lag τ which iscomputed by
A 2035 signal types B 223 types
Train4324600 PDW lists
Validation617800
Test1235600
692000PDW lists
Figure 3 +e configuration of the data sets
A 2035 signal typesTrain
4324600 PDW lists
Validation617800 PDW lists
Section 4 constructionof a LSTM model
Section 5 constructionof feature extraction
and a deep neural network
Section 6 comparison of predictionperformance by two models
A 2035 known signal typesTest
1235600 PDW lists
B 223 unknown signal types
Test692000 PDW lists
Figure 4 A flowchart to explain the overall organization
Table 1 Parameters of a LSTM model
Step Parameters Tested values Selectedvalue Note
1 Optimizationmethod
RMSProp AdamAdadelta Adam Initial learning rate was set as 0003 for RMSProp [23] Adam [24] and 10 for
Adadelta [25]2 Minibatch size 50 100 200 200
3 Dropout ratio () 0 10 30 50 0 Dropout ratio means the rate at which the output gate units in a LSTM layerare randomly removed
4 LSTM layers 1 2 2 +e model with 2 LSTM layers showed higher accuracy than the model of 1LSTM layer
5 Fully connectedlayer 0 1 1 +e model with the fully connected layer had a higher accuracy than the
model with no fully connected layer
6 Input features 3 5 3 Higher accuracy was obtained when using three features RF PRI and PWinstead of using 5 features AOA AMP RF PRI and PW
7 Decay ratio Use no use Use When gradually decreasing the learning rate by multiplying 09 to theprevious learning rate per epoch after epoch 10 higher accuracy was obtained
Security and Communication Networks 5
R(τ) 1113936
nminus τ1 xt minus x( 1113857 xt+τ minus x( 1113857
1113936n1 xt minus x( 1113857
2 (3)
Various autocorrelation coefficients can be obtained bychanging the time difference τ [4]
MFCC extracts feature values from the sequence datathrough the process of framing Fast Fourier transform(FFT) mel filter bank log function and inverse FFT [5] Inthe framing step the sequence is divided into segments ofequal size FFT is applied to a subsequence of each segmentto convert to the frequency domain and then the powerspectrum is obtained +e mel spectrum is obtained byapplying mel filters to the power spectrum and the coef-ficients are obtained by inverse FFT after applying the logfunction to the mel spectrum +ese coefficients are calledMFCCs All or some of the coefficients obtained in eachsegment can be used as the feature values [5]
52 Composition of Feature Sets We extracted feature valuesusing RF PRI and PW sequences of a PDW list Bycomputing the mean standard deviation skewness andkurtosis from each sequence a total of 12 feature values wereobtained from a PDW list By increasing the time differenceτ from 1 to 10 by incrementing 1 we extracted 10 auto-correlation coefficients from the RF and PRI sequences andtherefore a total of 20 feature values were extracted from aPDW list Also 10 MFCCs were computed for the RF andPRI sequences so that a total of 20 values were extracted+eautocorrelation coefficients and MFCCs were computedusing Python libraries such as the function acf of statsmodelsand the function mfcc of librosa [26]
Using statistical features as the basic features andcombining the autocorrelation coefficients and MFCCs withbasic features we composed 4 feature sets as shown in
Table 2 and tested prediction performance of a neuralnetwork model on each feature set In the feature set F1 aradar signal is represented with only statistical features +efeature set F2 contains the statistical features and autocor-relation coefficients while F3 is composed of the statisticalfeatures and MFCCs +e feature set F4 contains all of thestatistical features autocorrelation coefficients and MFCCsresulting in a total of 52 feature values
53 Performance of Neural Network Models on VariousFeature Sets To select a neural network structure com-parative testing for 4 different models was performedvarying the activation function and the number of layers asshown in Table 3 Considering test results in Section 41Adam was used as the optimization method and the min-ibatch size was set to 200
+e 4 feature sets in Table 2 and the 4 models in Table 3were used to compare the performance +e total number oftraining epochs was set to 15 and the training data com-posed in Section 32 were used to train the models +eaccuracy of the validation data was evaluated in each epochTable 4 shows the training and validation accuracies for 16combinations of 4 feature sets and 4 NNmodels which wereevaluated in the epoch with the highest accuracy for thevalidation data It revealed that model 4 on feature set F3 wasfound to have the highest performance Hence model 4 onfeature set F3 was chosen as the final model Figure 6 de-scribes the structure of the final deep neural network model
6 Performance Comparison of LSTM Modeland Deep Neural Network Model UsingExtracted Features
In this section we compare the prediction performance andtraining time complexity of two deep learning models fortest data of the known signal types in set A and test data ofthe unknown signal type in set B As explained in Section 322258 signal types were randomly divided into the set A to beused for the training validation and testing and the set B tobe used for testing of the unknown signal types Now werepeat the random division 10 times and measure the av-erage accuracy for performance comparison of two ap-proaches using LSTM and using a deep neural network withextracted features (denoted as DNNEF)
Table 5 shows test accuracy of known signal types in Aand unknown types in B from 10 repeated experiments Intesting of known signal types the average accuracy of 9846was obtained for the NN with extracted features and theaverage accuracy of 9936 was obtained for the LSTMmodel +e average accuracy for the unknown types whenusing the neural network with extracted features was 9245while the average accuracy for the LSTMmodel was 9353In both cases the LSTM model showed a little higheraccuracy
+e paired t-test was carried out in order to determinewhether the higher average accuracy of the LSTM model isstatistically significant +e paired t-test for the accuracy ofknown types gave the p value 00000612 which implies
Output layer 8 nodes
Input layer 3 nodes
RF PRI PWPDW list
Fully connected layer (400 nodes)
LSTM layer (200 nodes)
LSTM layer (200 nodes)
Figure 5 Structures of the final LSTM model
6 Security and Communication Networks
statistical significance in the difference between the accu-racies of the two models in the significance level of 001However for the unknown type accuracy the p value by thepaired t-test was approximately 0215 and so there was nosignificant difference in the significance level of 001 +isresult was thought to be due to the very large deviation in theaccuracy of the LSTM model for the unknown type whichwas a minimum of 8969 and maximum of 9720
In order to compare the learning time of the twomethods the average training time per epoch was measured+e specification of the computer used to measure the timeconsumed was CPU Intel i7-7700 360GHz and RAM320GB and GPU NVIDIA GeForce GTX 1080 Ti Table 6compares the execution time for model training for 15epochs+e process of extracting the features from the PDWlists was conducted once before the training of the deepneural network model As shown in Table 6 the time
consumed for training the LSTM model for 15 epochs wasapproximately 142 times longer than that of the deep neuralnetwork
Table 3 Four deep neural network models
Parameters Model 1 Model 2 Model 3 Model 4Activation function tanh relu tanh reluHidden layers 2 4Hidden units 400Learning rate Initially set as 00005
Learning rate decay Multiplying the previous learning rate by097 for every 1000000 samples
Table 4 +e performance in 16 combinations of feature sets andmodels
Model Feature set Training accuracy()
Validation accuracy()
Model 1
F1 9572 9592F2 9677 9709F3 9735 9778F4 9729 9778
Model 2
F1 9567 9572F2 9681 9730F3 9725 9755F4 9719 9761
Model 3
F1 9689 9735F2 9721 9765F3 9742 9804F4 9730 9778
Model 4
F1 9684 9748F2 9780 9819F3 9782 9847F4 9777 9802
Input layer 32 nodes
Hidden layer (400 nodes)
Output layer 8 nodes
Hidden layer (400 nodes)
Hidden layer (400 nodes)
Hidden layer (400 nodes)
32 feature values extracted from a PDW list
RF PRI PW
Figure 6 Structure of the final deep neural network
Table 5 Test accuracy () of known types and unknown types
DNNEF LSTM modelKnowntypes
Unknowntypes
Knowntypes
Unknowntypes
1 9849 9333 9972 97202 9835 9247 9970 95093 9849 9057 9878 90764 9860 9281 9906 90945 9861 9272 9932 93426 9855 9431 9960 95067 9839 9262 9971 96738 9847 9211 9901 89699 9831 9219 9900 906410 9833 9132 9965 9576Avg 9846 9245 9936 9353
Table 6 Comparison of execution time in seconds
Model Feature extraction Training for 1 epoch Training for 15epochs
DNN 7374233 1932943 36368378LSTM mdash 3454399 51815985
Table 2 Composition of four feature sets
Featuresets
Statisticalfeatures
Autocorrelationcoef MFCC Total
F1 12 mdash mdash 12F2 12 20 mdash 32F3 12 mdash 20 32F4 12 20 20 52
Security and Communication Networks 7
7 Conclusion
In this study we applied deep learning for jamming pre-diction for a radar signal Two methods were comparedusing a deep neural network model based on the featuresextracted from radar signals and using LSTMmodel withoutthe feature extraction process
+e deep neural network requires feature extraction inadvance and this is conducted once before the training ofthe deep neural network model +e training speed of themodel per epoch was faster than that of the LSTM modelHowever the feature extraction method has to be selectedappropriately because the model prediction performancevaries depending on the feature values extracted from thePDW list +e LSTMmodel has the advantages of being ableto perform training and prediction by directly using thePDW list of the radar signal without the extraction of featurevalues +e prediction accuracy of the LSTM model washigher on average than that of the deep neural networkmodel However there is the disadvantage that the trainingtime takes longer than a deep neural network model trainedon the extracted features
Testing results demonstrate that the jamming methodcan be predicted for an unknown type of radar signal with anaverage accuracy of approximately 92 and higher It showsthat deep learning methods can be used effectively forpredicting an appropriate jamming technique for new radarsignals which are not used in model training
Data Availability
+e radar signal data used to support the findings of thisstudy have not been made available because of the policy ofAgency for Defense Development Korea that data distri-bution is limited
Conflicts of Interest
+e authors declare that there are no conflicts of interestregarding the publication of this paper
References
[1] M S R Lothes M B Szymanski and R G Wiley RadarVulnerability to Jamming Artech House Boston MA USA1990
[2] V Iglesias J Grajal O Yeste-Ojeda M Garrido M Sanchezand M Lopez-Vallejo ldquoReal-time radar pulse parameterextractorrdquo in Proceedings of the IEEE Radar ConferenceCincinnati OH USA May 2014
[3] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[4] P DrsquoUrso and E Maharaj ldquoAutocorrelation-based fuzzyclustering of time seriesrdquo Fuzzy Sets and Systems vol 160no 24 pp 3565ndash3589 2009
[5] S Gupta J Jaafar W F wan Ahmad and A Bansal ldquoFeatureextraction using MFCCrdquo Signal amp Image Processing An In-ternational Journal vol 4 no 4 pp 101ndash108 2013
[6] R Shumway and D Stoffer Time Series Analysis and itsApplication with R Examples Springer Berlin Germany2000
[7] Z Xia S Xia L Wan and S Cai ldquoSpectral regression basedfault feature extraction for bearing accelerometer sensorsignalsrdquo Sensors vol 12 no 10 pp 13694ndash13719 2012
[8] T W Rauber F de Assis Boldt and F M Varejao ldquoHet-erogeneous feature models and feature selection applied tobearing fault diagnosisrdquo IEEE Transactions on IndustrialElectronics vol 62 no 1 pp 637ndash646 2015
[9] B R Nayana and P Geethanjali ldquoAnalysis of statistical time-domain features effectiveness in identification of bearingfaults from vibration signalrdquo IEEE Sensors Journal vol 17no 17 pp 5618ndash5625 2017
[10] M B L Muda and I Elamvazuthi ldquoVoice recognition al-gorithm using mel frequency cepstral coefficient (MFCC) anddynamic time warping (DTW) techniquesrdquo Journal ofComputing vol 2 no 3 pp 138ndash143 2010
[11] S Q L Jian-xun and Y Hai ldquoSignal feature analysis andexperimental verification of radar deception jammingrdquo inProceedings of the IEEE CIE International Conference onRadar Chengdu China October 2011
[12] A S A Mendoza and B Flores ldquoClassification of radarjammer FM signals using a neural networkrdquo in Proceedings ofthe SPIE Radar Sensor Technology XXI vol 10188 May 2017
[13] G Noone ldquoA neural approach to automatic pulse repetitioninterval modulation recognitionrdquo in Proceedings of the 1999Information Decision and Control Data and InformationFusion Symposium Signal Processing and CommunicationsSymposium and Decision and Control Symposium Proceedings(Cat No99EX251) Adelaide Australia February 1999
[14] J Kauppi and K Martikainen ldquoAn efficient set of features forpulse repetition interval modulation recognitionrdquo in Pro-ceedings of the IET International Conference on Radar SystemsEdinburgh UK October 2007
[15] I Jordanov N Petrov and A Petrozziello ldquoSupervised radarsignal classificationrdquo in Proceedings of the International JointConference on Neural Networks Vancouver Canada July2016
[16] D Zhou X Wang Y Tian and R Wang ldquoA novel radarsignal recognition method based on a deep restricted Boltz-mann machinerdquo Engineering Review vol 37 pp 165ndash1712017
[17] Z Y Z Wu Y Zhao and H Luo ldquoJamming signals clas-sification using convolutional neural networkrdquo in Proceedingsof the IEEE International Symposium on Signal Processing andInformation Technology (ISSPIT) Bilbao Spain December2017
[18] R Ferre A Fuente and E Lohan ldquoJammer classification inGNSS bands via machine learning algorithmsrdquo Sensorsvol 19 no 22 p 4841 2019
[19] J Mun H Kim and J Lee ldquoA deep learning approach forautomotive radar interference mitigationrdquo 2019 httpsarxivorgabs190306380
[20] E Masen B Yonei and B Yazici ldquoDeep learning for radarrdquoin Proceedings of the 2017 IEEE Radar Conference(RadarConf) Seattle WA USA May 2017
[21] N Srivastava G Hinton A Krizhevsky I Sutskever andR Salakhutdinov ldquoDropout a simple way to prevent neuralnetworks from overfittingrdquo e Journal of Machine LearningResearch vol 15 no 1 pp 1929ndash1958 2014
[22] G Lee ldquoRadar jamming technique prediction using deeplearningrdquo Master thesis Chungnam National UniversityDaejeon Korea 2019
[23] M Zeiler ldquoADADELTA an adaptive learning rate methodrdquo2012 httpsarxivorgabs12125701
8 Security and Communication Networks
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9
stable learning Figure 5 shows the structure of the finalLSTM model
5 Construction of Jamming Prediction ModelBased on Feature Extraction and a DeepNeural Network
In this section feature extraction from a PDW list of a threatsignal and the construction of a deep neural network areexplained in detail
51 Feature Extraction From sequence data expressed asx1 xn features such as statistical values autocorrelationcoefficients and MFCCs were extracted
Four statistic values of the mean x standard deviation sskewness w and kurtosis k were computed from the se-quence data such as
x 1n
1113944
n
1xt
s
1n
1113944
n
1xt minus x( 1113857
2
11139741113972
w 1n 1113936
n1 xt minus x( 1113857
31113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
32
k 1n1113936
n1 xt minus x( 1113857
41113872 1113873
1n1113936n1 xt minus x( 1113857
21113872 1113873
2
(2)
+e autocorrelation coefficient R(τ) represents the linearrelevance between two subsequences with a lag τ which iscomputed by
A 2035 signal types B 223 types
Train4324600 PDW lists
Validation617800
Test1235600
692000PDW lists
Figure 3 +e configuration of the data sets
A 2035 signal typesTrain
4324600 PDW lists
Validation617800 PDW lists
Section 4 constructionof a LSTM model
Section 5 constructionof feature extraction
and a deep neural network
Section 6 comparison of predictionperformance by two models
A 2035 known signal typesTest
1235600 PDW lists
B 223 unknown signal types
Test692000 PDW lists
Figure 4 A flowchart to explain the overall organization
Table 1 Parameters of a LSTM model
Step Parameters Tested values Selectedvalue Note
1 Optimizationmethod
RMSProp AdamAdadelta Adam Initial learning rate was set as 0003 for RMSProp [23] Adam [24] and 10 for
Adadelta [25]2 Minibatch size 50 100 200 200
3 Dropout ratio () 0 10 30 50 0 Dropout ratio means the rate at which the output gate units in a LSTM layerare randomly removed
4 LSTM layers 1 2 2 +e model with 2 LSTM layers showed higher accuracy than the model of 1LSTM layer
5 Fully connectedlayer 0 1 1 +e model with the fully connected layer had a higher accuracy than the
model with no fully connected layer
6 Input features 3 5 3 Higher accuracy was obtained when using three features RF PRI and PWinstead of using 5 features AOA AMP RF PRI and PW
7 Decay ratio Use no use Use When gradually decreasing the learning rate by multiplying 09 to theprevious learning rate per epoch after epoch 10 higher accuracy was obtained
Security and Communication Networks 5
R(τ) 1113936
nminus τ1 xt minus x( 1113857 xt+τ minus x( 1113857
1113936n1 xt minus x( 1113857
2 (3)
Various autocorrelation coefficients can be obtained bychanging the time difference τ [4]
MFCC extracts feature values from the sequence datathrough the process of framing Fast Fourier transform(FFT) mel filter bank log function and inverse FFT [5] Inthe framing step the sequence is divided into segments ofequal size FFT is applied to a subsequence of each segmentto convert to the frequency domain and then the powerspectrum is obtained +e mel spectrum is obtained byapplying mel filters to the power spectrum and the coef-ficients are obtained by inverse FFT after applying the logfunction to the mel spectrum +ese coefficients are calledMFCCs All or some of the coefficients obtained in eachsegment can be used as the feature values [5]
52 Composition of Feature Sets We extracted feature valuesusing RF PRI and PW sequences of a PDW list Bycomputing the mean standard deviation skewness andkurtosis from each sequence a total of 12 feature values wereobtained from a PDW list By increasing the time differenceτ from 1 to 10 by incrementing 1 we extracted 10 auto-correlation coefficients from the RF and PRI sequences andtherefore a total of 20 feature values were extracted from aPDW list Also 10 MFCCs were computed for the RF andPRI sequences so that a total of 20 values were extracted+eautocorrelation coefficients and MFCCs were computedusing Python libraries such as the function acf of statsmodelsand the function mfcc of librosa [26]
Using statistical features as the basic features andcombining the autocorrelation coefficients and MFCCs withbasic features we composed 4 feature sets as shown in
Table 2 and tested prediction performance of a neuralnetwork model on each feature set In the feature set F1 aradar signal is represented with only statistical features +efeature set F2 contains the statistical features and autocor-relation coefficients while F3 is composed of the statisticalfeatures and MFCCs +e feature set F4 contains all of thestatistical features autocorrelation coefficients and MFCCsresulting in a total of 52 feature values
53 Performance of Neural Network Models on VariousFeature Sets To select a neural network structure com-parative testing for 4 different models was performedvarying the activation function and the number of layers asshown in Table 3 Considering test results in Section 41Adam was used as the optimization method and the min-ibatch size was set to 200
+e 4 feature sets in Table 2 and the 4 models in Table 3were used to compare the performance +e total number oftraining epochs was set to 15 and the training data com-posed in Section 32 were used to train the models +eaccuracy of the validation data was evaluated in each epochTable 4 shows the training and validation accuracies for 16combinations of 4 feature sets and 4 NNmodels which wereevaluated in the epoch with the highest accuracy for thevalidation data It revealed that model 4 on feature set F3 wasfound to have the highest performance Hence model 4 onfeature set F3 was chosen as the final model Figure 6 de-scribes the structure of the final deep neural network model
6 Performance Comparison of LSTM Modeland Deep Neural Network Model UsingExtracted Features
In this section we compare the prediction performance andtraining time complexity of two deep learning models fortest data of the known signal types in set A and test data ofthe unknown signal type in set B As explained in Section 322258 signal types were randomly divided into the set A to beused for the training validation and testing and the set B tobe used for testing of the unknown signal types Now werepeat the random division 10 times and measure the av-erage accuracy for performance comparison of two ap-proaches using LSTM and using a deep neural network withextracted features (denoted as DNNEF)
Table 5 shows test accuracy of known signal types in Aand unknown types in B from 10 repeated experiments Intesting of known signal types the average accuracy of 9846was obtained for the NN with extracted features and theaverage accuracy of 9936 was obtained for the LSTMmodel +e average accuracy for the unknown types whenusing the neural network with extracted features was 9245while the average accuracy for the LSTMmodel was 9353In both cases the LSTM model showed a little higheraccuracy
+e paired t-test was carried out in order to determinewhether the higher average accuracy of the LSTM model isstatistically significant +e paired t-test for the accuracy ofknown types gave the p value 00000612 which implies
Output layer 8 nodes
Input layer 3 nodes
RF PRI PWPDW list
Fully connected layer (400 nodes)
LSTM layer (200 nodes)
LSTM layer (200 nodes)
Figure 5 Structures of the final LSTM model
6 Security and Communication Networks
statistical significance in the difference between the accu-racies of the two models in the significance level of 001However for the unknown type accuracy the p value by thepaired t-test was approximately 0215 and so there was nosignificant difference in the significance level of 001 +isresult was thought to be due to the very large deviation in theaccuracy of the LSTM model for the unknown type whichwas a minimum of 8969 and maximum of 9720
In order to compare the learning time of the twomethods the average training time per epoch was measured+e specification of the computer used to measure the timeconsumed was CPU Intel i7-7700 360GHz and RAM320GB and GPU NVIDIA GeForce GTX 1080 Ti Table 6compares the execution time for model training for 15epochs+e process of extracting the features from the PDWlists was conducted once before the training of the deepneural network model As shown in Table 6 the time
consumed for training the LSTM model for 15 epochs wasapproximately 142 times longer than that of the deep neuralnetwork
Table 3 Four deep neural network models
Parameters Model 1 Model 2 Model 3 Model 4Activation function tanh relu tanh reluHidden layers 2 4Hidden units 400Learning rate Initially set as 00005
Learning rate decay Multiplying the previous learning rate by097 for every 1000000 samples
Table 4 +e performance in 16 combinations of feature sets andmodels
Model Feature set Training accuracy()
Validation accuracy()
Model 1
F1 9572 9592F2 9677 9709F3 9735 9778F4 9729 9778
Model 2
F1 9567 9572F2 9681 9730F3 9725 9755F4 9719 9761
Model 3
F1 9689 9735F2 9721 9765F3 9742 9804F4 9730 9778
Model 4
F1 9684 9748F2 9780 9819F3 9782 9847F4 9777 9802
Input layer 32 nodes
Hidden layer (400 nodes)
Output layer 8 nodes
Hidden layer (400 nodes)
Hidden layer (400 nodes)
Hidden layer (400 nodes)
32 feature values extracted from a PDW list
RF PRI PW
Figure 6 Structure of the final deep neural network
Table 5 Test accuracy () of known types and unknown types
DNNEF LSTM modelKnowntypes
Unknowntypes
Knowntypes
Unknowntypes
1 9849 9333 9972 97202 9835 9247 9970 95093 9849 9057 9878 90764 9860 9281 9906 90945 9861 9272 9932 93426 9855 9431 9960 95067 9839 9262 9971 96738 9847 9211 9901 89699 9831 9219 9900 906410 9833 9132 9965 9576Avg 9846 9245 9936 9353
Table 6 Comparison of execution time in seconds
Model Feature extraction Training for 1 epoch Training for 15epochs
DNN 7374233 1932943 36368378LSTM mdash 3454399 51815985
Table 2 Composition of four feature sets
Featuresets
Statisticalfeatures
Autocorrelationcoef MFCC Total
F1 12 mdash mdash 12F2 12 20 mdash 32F3 12 mdash 20 32F4 12 20 20 52
Security and Communication Networks 7
7 Conclusion
In this study we applied deep learning for jamming pre-diction for a radar signal Two methods were comparedusing a deep neural network model based on the featuresextracted from radar signals and using LSTMmodel withoutthe feature extraction process
+e deep neural network requires feature extraction inadvance and this is conducted once before the training ofthe deep neural network model +e training speed of themodel per epoch was faster than that of the LSTM modelHowever the feature extraction method has to be selectedappropriately because the model prediction performancevaries depending on the feature values extracted from thePDW list +e LSTMmodel has the advantages of being ableto perform training and prediction by directly using thePDW list of the radar signal without the extraction of featurevalues +e prediction accuracy of the LSTM model washigher on average than that of the deep neural networkmodel However there is the disadvantage that the trainingtime takes longer than a deep neural network model trainedon the extracted features
Testing results demonstrate that the jamming methodcan be predicted for an unknown type of radar signal with anaverage accuracy of approximately 92 and higher It showsthat deep learning methods can be used effectively forpredicting an appropriate jamming technique for new radarsignals which are not used in model training
Data Availability
+e radar signal data used to support the findings of thisstudy have not been made available because of the policy ofAgency for Defense Development Korea that data distri-bution is limited
Conflicts of Interest
+e authors declare that there are no conflicts of interestregarding the publication of this paper
References
[1] M S R Lothes M B Szymanski and R G Wiley RadarVulnerability to Jamming Artech House Boston MA USA1990
[2] V Iglesias J Grajal O Yeste-Ojeda M Garrido M Sanchezand M Lopez-Vallejo ldquoReal-time radar pulse parameterextractorrdquo in Proceedings of the IEEE Radar ConferenceCincinnati OH USA May 2014
[3] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[4] P DrsquoUrso and E Maharaj ldquoAutocorrelation-based fuzzyclustering of time seriesrdquo Fuzzy Sets and Systems vol 160no 24 pp 3565ndash3589 2009
[5] S Gupta J Jaafar W F wan Ahmad and A Bansal ldquoFeatureextraction using MFCCrdquo Signal amp Image Processing An In-ternational Journal vol 4 no 4 pp 101ndash108 2013
[6] R Shumway and D Stoffer Time Series Analysis and itsApplication with R Examples Springer Berlin Germany2000
[7] Z Xia S Xia L Wan and S Cai ldquoSpectral regression basedfault feature extraction for bearing accelerometer sensorsignalsrdquo Sensors vol 12 no 10 pp 13694ndash13719 2012
[8] T W Rauber F de Assis Boldt and F M Varejao ldquoHet-erogeneous feature models and feature selection applied tobearing fault diagnosisrdquo IEEE Transactions on IndustrialElectronics vol 62 no 1 pp 637ndash646 2015
[9] B R Nayana and P Geethanjali ldquoAnalysis of statistical time-domain features effectiveness in identification of bearingfaults from vibration signalrdquo IEEE Sensors Journal vol 17no 17 pp 5618ndash5625 2017
[10] M B L Muda and I Elamvazuthi ldquoVoice recognition al-gorithm using mel frequency cepstral coefficient (MFCC) anddynamic time warping (DTW) techniquesrdquo Journal ofComputing vol 2 no 3 pp 138ndash143 2010
[11] S Q L Jian-xun and Y Hai ldquoSignal feature analysis andexperimental verification of radar deception jammingrdquo inProceedings of the IEEE CIE International Conference onRadar Chengdu China October 2011
[12] A S A Mendoza and B Flores ldquoClassification of radarjammer FM signals using a neural networkrdquo in Proceedings ofthe SPIE Radar Sensor Technology XXI vol 10188 May 2017
[13] G Noone ldquoA neural approach to automatic pulse repetitioninterval modulation recognitionrdquo in Proceedings of the 1999Information Decision and Control Data and InformationFusion Symposium Signal Processing and CommunicationsSymposium and Decision and Control Symposium Proceedings(Cat No99EX251) Adelaide Australia February 1999
[14] J Kauppi and K Martikainen ldquoAn efficient set of features forpulse repetition interval modulation recognitionrdquo in Pro-ceedings of the IET International Conference on Radar SystemsEdinburgh UK October 2007
[15] I Jordanov N Petrov and A Petrozziello ldquoSupervised radarsignal classificationrdquo in Proceedings of the International JointConference on Neural Networks Vancouver Canada July2016
[16] D Zhou X Wang Y Tian and R Wang ldquoA novel radarsignal recognition method based on a deep restricted Boltz-mann machinerdquo Engineering Review vol 37 pp 165ndash1712017
[17] Z Y Z Wu Y Zhao and H Luo ldquoJamming signals clas-sification using convolutional neural networkrdquo in Proceedingsof the IEEE International Symposium on Signal Processing andInformation Technology (ISSPIT) Bilbao Spain December2017
[18] R Ferre A Fuente and E Lohan ldquoJammer classification inGNSS bands via machine learning algorithmsrdquo Sensorsvol 19 no 22 p 4841 2019
[19] J Mun H Kim and J Lee ldquoA deep learning approach forautomotive radar interference mitigationrdquo 2019 httpsarxivorgabs190306380
[20] E Masen B Yonei and B Yazici ldquoDeep learning for radarrdquoin Proceedings of the 2017 IEEE Radar Conference(RadarConf) Seattle WA USA May 2017
[21] N Srivastava G Hinton A Krizhevsky I Sutskever andR Salakhutdinov ldquoDropout a simple way to prevent neuralnetworks from overfittingrdquo e Journal of Machine LearningResearch vol 15 no 1 pp 1929ndash1958 2014
[22] G Lee ldquoRadar jamming technique prediction using deeplearningrdquo Master thesis Chungnam National UniversityDaejeon Korea 2019
[23] M Zeiler ldquoADADELTA an adaptive learning rate methodrdquo2012 httpsarxivorgabs12125701
8 Security and Communication Networks
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9
R(τ) 1113936
nminus τ1 xt minus x( 1113857 xt+τ minus x( 1113857
1113936n1 xt minus x( 1113857
2 (3)
Various autocorrelation coefficients can be obtained bychanging the time difference τ [4]
MFCC extracts feature values from the sequence datathrough the process of framing Fast Fourier transform(FFT) mel filter bank log function and inverse FFT [5] Inthe framing step the sequence is divided into segments ofequal size FFT is applied to a subsequence of each segmentto convert to the frequency domain and then the powerspectrum is obtained +e mel spectrum is obtained byapplying mel filters to the power spectrum and the coef-ficients are obtained by inverse FFT after applying the logfunction to the mel spectrum +ese coefficients are calledMFCCs All or some of the coefficients obtained in eachsegment can be used as the feature values [5]
52 Composition of Feature Sets We extracted feature valuesusing RF PRI and PW sequences of a PDW list Bycomputing the mean standard deviation skewness andkurtosis from each sequence a total of 12 feature values wereobtained from a PDW list By increasing the time differenceτ from 1 to 10 by incrementing 1 we extracted 10 auto-correlation coefficients from the RF and PRI sequences andtherefore a total of 20 feature values were extracted from aPDW list Also 10 MFCCs were computed for the RF andPRI sequences so that a total of 20 values were extracted+eautocorrelation coefficients and MFCCs were computedusing Python libraries such as the function acf of statsmodelsand the function mfcc of librosa [26]
Using statistical features as the basic features andcombining the autocorrelation coefficients and MFCCs withbasic features we composed 4 feature sets as shown in
Table 2 and tested prediction performance of a neuralnetwork model on each feature set In the feature set F1 aradar signal is represented with only statistical features +efeature set F2 contains the statistical features and autocor-relation coefficients while F3 is composed of the statisticalfeatures and MFCCs +e feature set F4 contains all of thestatistical features autocorrelation coefficients and MFCCsresulting in a total of 52 feature values
53 Performance of Neural Network Models on VariousFeature Sets To select a neural network structure com-parative testing for 4 different models was performedvarying the activation function and the number of layers asshown in Table 3 Considering test results in Section 41Adam was used as the optimization method and the min-ibatch size was set to 200
+e 4 feature sets in Table 2 and the 4 models in Table 3were used to compare the performance +e total number oftraining epochs was set to 15 and the training data com-posed in Section 32 were used to train the models +eaccuracy of the validation data was evaluated in each epochTable 4 shows the training and validation accuracies for 16combinations of 4 feature sets and 4 NNmodels which wereevaluated in the epoch with the highest accuracy for thevalidation data It revealed that model 4 on feature set F3 wasfound to have the highest performance Hence model 4 onfeature set F3 was chosen as the final model Figure 6 de-scribes the structure of the final deep neural network model
6 Performance Comparison of LSTM Modeland Deep Neural Network Model UsingExtracted Features
In this section we compare the prediction performance andtraining time complexity of two deep learning models fortest data of the known signal types in set A and test data ofthe unknown signal type in set B As explained in Section 322258 signal types were randomly divided into the set A to beused for the training validation and testing and the set B tobe used for testing of the unknown signal types Now werepeat the random division 10 times and measure the av-erage accuracy for performance comparison of two ap-proaches using LSTM and using a deep neural network withextracted features (denoted as DNNEF)
Table 5 shows test accuracy of known signal types in Aand unknown types in B from 10 repeated experiments Intesting of known signal types the average accuracy of 9846was obtained for the NN with extracted features and theaverage accuracy of 9936 was obtained for the LSTMmodel +e average accuracy for the unknown types whenusing the neural network with extracted features was 9245while the average accuracy for the LSTMmodel was 9353In both cases the LSTM model showed a little higheraccuracy
+e paired t-test was carried out in order to determinewhether the higher average accuracy of the LSTM model isstatistically significant +e paired t-test for the accuracy ofknown types gave the p value 00000612 which implies
Output layer 8 nodes
Input layer 3 nodes
RF PRI PWPDW list
Fully connected layer (400 nodes)
LSTM layer (200 nodes)
LSTM layer (200 nodes)
Figure 5 Structures of the final LSTM model
6 Security and Communication Networks
statistical significance in the difference between the accu-racies of the two models in the significance level of 001However for the unknown type accuracy the p value by thepaired t-test was approximately 0215 and so there was nosignificant difference in the significance level of 001 +isresult was thought to be due to the very large deviation in theaccuracy of the LSTM model for the unknown type whichwas a minimum of 8969 and maximum of 9720
In order to compare the learning time of the twomethods the average training time per epoch was measured+e specification of the computer used to measure the timeconsumed was CPU Intel i7-7700 360GHz and RAM320GB and GPU NVIDIA GeForce GTX 1080 Ti Table 6compares the execution time for model training for 15epochs+e process of extracting the features from the PDWlists was conducted once before the training of the deepneural network model As shown in Table 6 the time
consumed for training the LSTM model for 15 epochs wasapproximately 142 times longer than that of the deep neuralnetwork
Table 3 Four deep neural network models
Parameters Model 1 Model 2 Model 3 Model 4Activation function tanh relu tanh reluHidden layers 2 4Hidden units 400Learning rate Initially set as 00005
Learning rate decay Multiplying the previous learning rate by097 for every 1000000 samples
Table 4 +e performance in 16 combinations of feature sets andmodels
Model Feature set Training accuracy()
Validation accuracy()
Model 1
F1 9572 9592F2 9677 9709F3 9735 9778F4 9729 9778
Model 2
F1 9567 9572F2 9681 9730F3 9725 9755F4 9719 9761
Model 3
F1 9689 9735F2 9721 9765F3 9742 9804F4 9730 9778
Model 4
F1 9684 9748F2 9780 9819F3 9782 9847F4 9777 9802
Input layer 32 nodes
Hidden layer (400 nodes)
Output layer 8 nodes
Hidden layer (400 nodes)
Hidden layer (400 nodes)
Hidden layer (400 nodes)
32 feature values extracted from a PDW list
RF PRI PW
Figure 6 Structure of the final deep neural network
Table 5 Test accuracy () of known types and unknown types
DNNEF LSTM modelKnowntypes
Unknowntypes
Knowntypes
Unknowntypes
1 9849 9333 9972 97202 9835 9247 9970 95093 9849 9057 9878 90764 9860 9281 9906 90945 9861 9272 9932 93426 9855 9431 9960 95067 9839 9262 9971 96738 9847 9211 9901 89699 9831 9219 9900 906410 9833 9132 9965 9576Avg 9846 9245 9936 9353
Table 6 Comparison of execution time in seconds
Model Feature extraction Training for 1 epoch Training for 15epochs
DNN 7374233 1932943 36368378LSTM mdash 3454399 51815985
Table 2 Composition of four feature sets
Featuresets
Statisticalfeatures
Autocorrelationcoef MFCC Total
F1 12 mdash mdash 12F2 12 20 mdash 32F3 12 mdash 20 32F4 12 20 20 52
Security and Communication Networks 7
7 Conclusion
In this study we applied deep learning for jamming pre-diction for a radar signal Two methods were comparedusing a deep neural network model based on the featuresextracted from radar signals and using LSTMmodel withoutthe feature extraction process
+e deep neural network requires feature extraction inadvance and this is conducted once before the training ofthe deep neural network model +e training speed of themodel per epoch was faster than that of the LSTM modelHowever the feature extraction method has to be selectedappropriately because the model prediction performancevaries depending on the feature values extracted from thePDW list +e LSTMmodel has the advantages of being ableto perform training and prediction by directly using thePDW list of the radar signal without the extraction of featurevalues +e prediction accuracy of the LSTM model washigher on average than that of the deep neural networkmodel However there is the disadvantage that the trainingtime takes longer than a deep neural network model trainedon the extracted features
Testing results demonstrate that the jamming methodcan be predicted for an unknown type of radar signal with anaverage accuracy of approximately 92 and higher It showsthat deep learning methods can be used effectively forpredicting an appropriate jamming technique for new radarsignals which are not used in model training
Data Availability
+e radar signal data used to support the findings of thisstudy have not been made available because of the policy ofAgency for Defense Development Korea that data distri-bution is limited
Conflicts of Interest
+e authors declare that there are no conflicts of interestregarding the publication of this paper
References
[1] M S R Lothes M B Szymanski and R G Wiley RadarVulnerability to Jamming Artech House Boston MA USA1990
[2] V Iglesias J Grajal O Yeste-Ojeda M Garrido M Sanchezand M Lopez-Vallejo ldquoReal-time radar pulse parameterextractorrdquo in Proceedings of the IEEE Radar ConferenceCincinnati OH USA May 2014
[3] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[4] P DrsquoUrso and E Maharaj ldquoAutocorrelation-based fuzzyclustering of time seriesrdquo Fuzzy Sets and Systems vol 160no 24 pp 3565ndash3589 2009
[5] S Gupta J Jaafar W F wan Ahmad and A Bansal ldquoFeatureextraction using MFCCrdquo Signal amp Image Processing An In-ternational Journal vol 4 no 4 pp 101ndash108 2013
[6] R Shumway and D Stoffer Time Series Analysis and itsApplication with R Examples Springer Berlin Germany2000
[7] Z Xia S Xia L Wan and S Cai ldquoSpectral regression basedfault feature extraction for bearing accelerometer sensorsignalsrdquo Sensors vol 12 no 10 pp 13694ndash13719 2012
[8] T W Rauber F de Assis Boldt and F M Varejao ldquoHet-erogeneous feature models and feature selection applied tobearing fault diagnosisrdquo IEEE Transactions on IndustrialElectronics vol 62 no 1 pp 637ndash646 2015
[9] B R Nayana and P Geethanjali ldquoAnalysis of statistical time-domain features effectiveness in identification of bearingfaults from vibration signalrdquo IEEE Sensors Journal vol 17no 17 pp 5618ndash5625 2017
[10] M B L Muda and I Elamvazuthi ldquoVoice recognition al-gorithm using mel frequency cepstral coefficient (MFCC) anddynamic time warping (DTW) techniquesrdquo Journal ofComputing vol 2 no 3 pp 138ndash143 2010
[11] S Q L Jian-xun and Y Hai ldquoSignal feature analysis andexperimental verification of radar deception jammingrdquo inProceedings of the IEEE CIE International Conference onRadar Chengdu China October 2011
[12] A S A Mendoza and B Flores ldquoClassification of radarjammer FM signals using a neural networkrdquo in Proceedings ofthe SPIE Radar Sensor Technology XXI vol 10188 May 2017
[13] G Noone ldquoA neural approach to automatic pulse repetitioninterval modulation recognitionrdquo in Proceedings of the 1999Information Decision and Control Data and InformationFusion Symposium Signal Processing and CommunicationsSymposium and Decision and Control Symposium Proceedings(Cat No99EX251) Adelaide Australia February 1999
[14] J Kauppi and K Martikainen ldquoAn efficient set of features forpulse repetition interval modulation recognitionrdquo in Pro-ceedings of the IET International Conference on Radar SystemsEdinburgh UK October 2007
[15] I Jordanov N Petrov and A Petrozziello ldquoSupervised radarsignal classificationrdquo in Proceedings of the International JointConference on Neural Networks Vancouver Canada July2016
[16] D Zhou X Wang Y Tian and R Wang ldquoA novel radarsignal recognition method based on a deep restricted Boltz-mann machinerdquo Engineering Review vol 37 pp 165ndash1712017
[17] Z Y Z Wu Y Zhao and H Luo ldquoJamming signals clas-sification using convolutional neural networkrdquo in Proceedingsof the IEEE International Symposium on Signal Processing andInformation Technology (ISSPIT) Bilbao Spain December2017
[18] R Ferre A Fuente and E Lohan ldquoJammer classification inGNSS bands via machine learning algorithmsrdquo Sensorsvol 19 no 22 p 4841 2019
[19] J Mun H Kim and J Lee ldquoA deep learning approach forautomotive radar interference mitigationrdquo 2019 httpsarxivorgabs190306380
[20] E Masen B Yonei and B Yazici ldquoDeep learning for radarrdquoin Proceedings of the 2017 IEEE Radar Conference(RadarConf) Seattle WA USA May 2017
[21] N Srivastava G Hinton A Krizhevsky I Sutskever andR Salakhutdinov ldquoDropout a simple way to prevent neuralnetworks from overfittingrdquo e Journal of Machine LearningResearch vol 15 no 1 pp 1929ndash1958 2014
[22] G Lee ldquoRadar jamming technique prediction using deeplearningrdquo Master thesis Chungnam National UniversityDaejeon Korea 2019
[23] M Zeiler ldquoADADELTA an adaptive learning rate methodrdquo2012 httpsarxivorgabs12125701
8 Security and Communication Networks
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9
statistical significance in the difference between the accu-racies of the two models in the significance level of 001However for the unknown type accuracy the p value by thepaired t-test was approximately 0215 and so there was nosignificant difference in the significance level of 001 +isresult was thought to be due to the very large deviation in theaccuracy of the LSTM model for the unknown type whichwas a minimum of 8969 and maximum of 9720
In order to compare the learning time of the twomethods the average training time per epoch was measured+e specification of the computer used to measure the timeconsumed was CPU Intel i7-7700 360GHz and RAM320GB and GPU NVIDIA GeForce GTX 1080 Ti Table 6compares the execution time for model training for 15epochs+e process of extracting the features from the PDWlists was conducted once before the training of the deepneural network model As shown in Table 6 the time
consumed for training the LSTM model for 15 epochs wasapproximately 142 times longer than that of the deep neuralnetwork
Table 3 Four deep neural network models
Parameters Model 1 Model 2 Model 3 Model 4Activation function tanh relu tanh reluHidden layers 2 4Hidden units 400Learning rate Initially set as 00005
Learning rate decay Multiplying the previous learning rate by097 for every 1000000 samples
Table 4 +e performance in 16 combinations of feature sets andmodels
Model Feature set Training accuracy()
Validation accuracy()
Model 1
F1 9572 9592F2 9677 9709F3 9735 9778F4 9729 9778
Model 2
F1 9567 9572F2 9681 9730F3 9725 9755F4 9719 9761
Model 3
F1 9689 9735F2 9721 9765F3 9742 9804F4 9730 9778
Model 4
F1 9684 9748F2 9780 9819F3 9782 9847F4 9777 9802
Input layer 32 nodes
Hidden layer (400 nodes)
Output layer 8 nodes
Hidden layer (400 nodes)
Hidden layer (400 nodes)
Hidden layer (400 nodes)
32 feature values extracted from a PDW list
RF PRI PW
Figure 6 Structure of the final deep neural network
Table 5 Test accuracy () of known types and unknown types
DNNEF LSTM modelKnowntypes
Unknowntypes
Knowntypes
Unknowntypes
1 9849 9333 9972 97202 9835 9247 9970 95093 9849 9057 9878 90764 9860 9281 9906 90945 9861 9272 9932 93426 9855 9431 9960 95067 9839 9262 9971 96738 9847 9211 9901 89699 9831 9219 9900 906410 9833 9132 9965 9576Avg 9846 9245 9936 9353
Table 6 Comparison of execution time in seconds
Model Feature extraction Training for 1 epoch Training for 15epochs
DNN 7374233 1932943 36368378LSTM mdash 3454399 51815985
Table 2 Composition of four feature sets
Featuresets
Statisticalfeatures
Autocorrelationcoef MFCC Total
F1 12 mdash mdash 12F2 12 20 mdash 32F3 12 mdash 20 32F4 12 20 20 52
Security and Communication Networks 7
7 Conclusion
In this study we applied deep learning for jamming pre-diction for a radar signal Two methods were comparedusing a deep neural network model based on the featuresextracted from radar signals and using LSTMmodel withoutthe feature extraction process
+e deep neural network requires feature extraction inadvance and this is conducted once before the training ofthe deep neural network model +e training speed of themodel per epoch was faster than that of the LSTM modelHowever the feature extraction method has to be selectedappropriately because the model prediction performancevaries depending on the feature values extracted from thePDW list +e LSTMmodel has the advantages of being ableto perform training and prediction by directly using thePDW list of the radar signal without the extraction of featurevalues +e prediction accuracy of the LSTM model washigher on average than that of the deep neural networkmodel However there is the disadvantage that the trainingtime takes longer than a deep neural network model trainedon the extracted features
Testing results demonstrate that the jamming methodcan be predicted for an unknown type of radar signal with anaverage accuracy of approximately 92 and higher It showsthat deep learning methods can be used effectively forpredicting an appropriate jamming technique for new radarsignals which are not used in model training
Data Availability
+e radar signal data used to support the findings of thisstudy have not been made available because of the policy ofAgency for Defense Development Korea that data distri-bution is limited
Conflicts of Interest
+e authors declare that there are no conflicts of interestregarding the publication of this paper
References
[1] M S R Lothes M B Szymanski and R G Wiley RadarVulnerability to Jamming Artech House Boston MA USA1990
[2] V Iglesias J Grajal O Yeste-Ojeda M Garrido M Sanchezand M Lopez-Vallejo ldquoReal-time radar pulse parameterextractorrdquo in Proceedings of the IEEE Radar ConferenceCincinnati OH USA May 2014
[3] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[4] P DrsquoUrso and E Maharaj ldquoAutocorrelation-based fuzzyclustering of time seriesrdquo Fuzzy Sets and Systems vol 160no 24 pp 3565ndash3589 2009
[5] S Gupta J Jaafar W F wan Ahmad and A Bansal ldquoFeatureextraction using MFCCrdquo Signal amp Image Processing An In-ternational Journal vol 4 no 4 pp 101ndash108 2013
[6] R Shumway and D Stoffer Time Series Analysis and itsApplication with R Examples Springer Berlin Germany2000
[7] Z Xia S Xia L Wan and S Cai ldquoSpectral regression basedfault feature extraction for bearing accelerometer sensorsignalsrdquo Sensors vol 12 no 10 pp 13694ndash13719 2012
[8] T W Rauber F de Assis Boldt and F M Varejao ldquoHet-erogeneous feature models and feature selection applied tobearing fault diagnosisrdquo IEEE Transactions on IndustrialElectronics vol 62 no 1 pp 637ndash646 2015
[9] B R Nayana and P Geethanjali ldquoAnalysis of statistical time-domain features effectiveness in identification of bearingfaults from vibration signalrdquo IEEE Sensors Journal vol 17no 17 pp 5618ndash5625 2017
[10] M B L Muda and I Elamvazuthi ldquoVoice recognition al-gorithm using mel frequency cepstral coefficient (MFCC) anddynamic time warping (DTW) techniquesrdquo Journal ofComputing vol 2 no 3 pp 138ndash143 2010
[11] S Q L Jian-xun and Y Hai ldquoSignal feature analysis andexperimental verification of radar deception jammingrdquo inProceedings of the IEEE CIE International Conference onRadar Chengdu China October 2011
[12] A S A Mendoza and B Flores ldquoClassification of radarjammer FM signals using a neural networkrdquo in Proceedings ofthe SPIE Radar Sensor Technology XXI vol 10188 May 2017
[13] G Noone ldquoA neural approach to automatic pulse repetitioninterval modulation recognitionrdquo in Proceedings of the 1999Information Decision and Control Data and InformationFusion Symposium Signal Processing and CommunicationsSymposium and Decision and Control Symposium Proceedings(Cat No99EX251) Adelaide Australia February 1999
[14] J Kauppi and K Martikainen ldquoAn efficient set of features forpulse repetition interval modulation recognitionrdquo in Pro-ceedings of the IET International Conference on Radar SystemsEdinburgh UK October 2007
[15] I Jordanov N Petrov and A Petrozziello ldquoSupervised radarsignal classificationrdquo in Proceedings of the International JointConference on Neural Networks Vancouver Canada July2016
[16] D Zhou X Wang Y Tian and R Wang ldquoA novel radarsignal recognition method based on a deep restricted Boltz-mann machinerdquo Engineering Review vol 37 pp 165ndash1712017
[17] Z Y Z Wu Y Zhao and H Luo ldquoJamming signals clas-sification using convolutional neural networkrdquo in Proceedingsof the IEEE International Symposium on Signal Processing andInformation Technology (ISSPIT) Bilbao Spain December2017
[18] R Ferre A Fuente and E Lohan ldquoJammer classification inGNSS bands via machine learning algorithmsrdquo Sensorsvol 19 no 22 p 4841 2019
[19] J Mun H Kim and J Lee ldquoA deep learning approach forautomotive radar interference mitigationrdquo 2019 httpsarxivorgabs190306380
[20] E Masen B Yonei and B Yazici ldquoDeep learning for radarrdquoin Proceedings of the 2017 IEEE Radar Conference(RadarConf) Seattle WA USA May 2017
[21] N Srivastava G Hinton A Krizhevsky I Sutskever andR Salakhutdinov ldquoDropout a simple way to prevent neuralnetworks from overfittingrdquo e Journal of Machine LearningResearch vol 15 no 1 pp 1929ndash1958 2014
[22] G Lee ldquoRadar jamming technique prediction using deeplearningrdquo Master thesis Chungnam National UniversityDaejeon Korea 2019
[23] M Zeiler ldquoADADELTA an adaptive learning rate methodrdquo2012 httpsarxivorgabs12125701
8 Security and Communication Networks
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9
7 Conclusion
In this study we applied deep learning for jamming pre-diction for a radar signal Two methods were comparedusing a deep neural network model based on the featuresextracted from radar signals and using LSTMmodel withoutthe feature extraction process
+e deep neural network requires feature extraction inadvance and this is conducted once before the training ofthe deep neural network model +e training speed of themodel per epoch was faster than that of the LSTM modelHowever the feature extraction method has to be selectedappropriately because the model prediction performancevaries depending on the feature values extracted from thePDW list +e LSTMmodel has the advantages of being ableto perform training and prediction by directly using thePDW list of the radar signal without the extraction of featurevalues +e prediction accuracy of the LSTM model washigher on average than that of the deep neural networkmodel However there is the disadvantage that the trainingtime takes longer than a deep neural network model trainedon the extracted features
Testing results demonstrate that the jamming methodcan be predicted for an unknown type of radar signal with anaverage accuracy of approximately 92 and higher It showsthat deep learning methods can be used effectively forpredicting an appropriate jamming technique for new radarsignals which are not used in model training
Data Availability
+e radar signal data used to support the findings of thisstudy have not been made available because of the policy ofAgency for Defense Development Korea that data distri-bution is limited
Conflicts of Interest
+e authors declare that there are no conflicts of interestregarding the publication of this paper
References
[1] M S R Lothes M B Szymanski and R G Wiley RadarVulnerability to Jamming Artech House Boston MA USA1990
[2] V Iglesias J Grajal O Yeste-Ojeda M Garrido M Sanchezand M Lopez-Vallejo ldquoReal-time radar pulse parameterextractorrdquo in Proceedings of the IEEE Radar ConferenceCincinnati OH USA May 2014
[3] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[4] P DrsquoUrso and E Maharaj ldquoAutocorrelation-based fuzzyclustering of time seriesrdquo Fuzzy Sets and Systems vol 160no 24 pp 3565ndash3589 2009
[5] S Gupta J Jaafar W F wan Ahmad and A Bansal ldquoFeatureextraction using MFCCrdquo Signal amp Image Processing An In-ternational Journal vol 4 no 4 pp 101ndash108 2013
[6] R Shumway and D Stoffer Time Series Analysis and itsApplication with R Examples Springer Berlin Germany2000
[7] Z Xia S Xia L Wan and S Cai ldquoSpectral regression basedfault feature extraction for bearing accelerometer sensorsignalsrdquo Sensors vol 12 no 10 pp 13694ndash13719 2012
[8] T W Rauber F de Assis Boldt and F M Varejao ldquoHet-erogeneous feature models and feature selection applied tobearing fault diagnosisrdquo IEEE Transactions on IndustrialElectronics vol 62 no 1 pp 637ndash646 2015
[9] B R Nayana and P Geethanjali ldquoAnalysis of statistical time-domain features effectiveness in identification of bearingfaults from vibration signalrdquo IEEE Sensors Journal vol 17no 17 pp 5618ndash5625 2017
[10] M B L Muda and I Elamvazuthi ldquoVoice recognition al-gorithm using mel frequency cepstral coefficient (MFCC) anddynamic time warping (DTW) techniquesrdquo Journal ofComputing vol 2 no 3 pp 138ndash143 2010
[11] S Q L Jian-xun and Y Hai ldquoSignal feature analysis andexperimental verification of radar deception jammingrdquo inProceedings of the IEEE CIE International Conference onRadar Chengdu China October 2011
[12] A S A Mendoza and B Flores ldquoClassification of radarjammer FM signals using a neural networkrdquo in Proceedings ofthe SPIE Radar Sensor Technology XXI vol 10188 May 2017
[13] G Noone ldquoA neural approach to automatic pulse repetitioninterval modulation recognitionrdquo in Proceedings of the 1999Information Decision and Control Data and InformationFusion Symposium Signal Processing and CommunicationsSymposium and Decision and Control Symposium Proceedings(Cat No99EX251) Adelaide Australia February 1999
[14] J Kauppi and K Martikainen ldquoAn efficient set of features forpulse repetition interval modulation recognitionrdquo in Pro-ceedings of the IET International Conference on Radar SystemsEdinburgh UK October 2007
[15] I Jordanov N Petrov and A Petrozziello ldquoSupervised radarsignal classificationrdquo in Proceedings of the International JointConference on Neural Networks Vancouver Canada July2016
[16] D Zhou X Wang Y Tian and R Wang ldquoA novel radarsignal recognition method based on a deep restricted Boltz-mann machinerdquo Engineering Review vol 37 pp 165ndash1712017
[17] Z Y Z Wu Y Zhao and H Luo ldquoJamming signals clas-sification using convolutional neural networkrdquo in Proceedingsof the IEEE International Symposium on Signal Processing andInformation Technology (ISSPIT) Bilbao Spain December2017
[18] R Ferre A Fuente and E Lohan ldquoJammer classification inGNSS bands via machine learning algorithmsrdquo Sensorsvol 19 no 22 p 4841 2019
[19] J Mun H Kim and J Lee ldquoA deep learning approach forautomotive radar interference mitigationrdquo 2019 httpsarxivorgabs190306380
[20] E Masen B Yonei and B Yazici ldquoDeep learning for radarrdquoin Proceedings of the 2017 IEEE Radar Conference(RadarConf) Seattle WA USA May 2017
[21] N Srivastava G Hinton A Krizhevsky I Sutskever andR Salakhutdinov ldquoDropout a simple way to prevent neuralnetworks from overfittingrdquo e Journal of Machine LearningResearch vol 15 no 1 pp 1929ndash1958 2014
[22] G Lee ldquoRadar jamming technique prediction using deeplearningrdquo Master thesis Chungnam National UniversityDaejeon Korea 2019
[23] M Zeiler ldquoADADELTA an adaptive learning rate methodrdquo2012 httpsarxivorgabs12125701
8 Security and Communication Networks
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9
[24] T Tieleman and G Hinton ldquoLecture 65-rmsprop divide thegradient by a running average of its recent magnituderdquoCOURSERA Neural Networks for Machine Learning vol 4no 2 2012
[25] D Kingma and J Ba ldquoAdam a method for stochastic opti-mizationrdquo 2014 httpsarxivorgabs14126980
[26] httpswwwpythonorgdownloadsreleasepython-352
Security and Communication Networks 9