Digital Modulation Recognition Using Support

Vector Machine Classifier

Hussam Mustafa and Miloš Doroslovaþki
Department of Electrical and Computer Engineering
The George Washington University
Washington, DC 20052

Abstract-We propose four features to classify amplitude shift the likelihood function to do recognition. They are
keying with two levels and four levels, binary phase shift optimal in the sense of the minimum probability of
keying, quadrature phase keying, frequency shift keying misclassification.
with two carriers and four carriers. After that we present a In this paper we present the signal model we assume
new method of classification based on support vector
(Section 2), as well as the new proposed features (Section
machine (SVM) that uses the four proposed features. We
study the performance of SVM classifier and compare it to 3) for modulation recognition. Further, we describe
the previous work done in the literature on the digital briefly the support vector machine (SVM) algorithm
modulation classification problem. (Section 4) and discuss the construction of SVM classifier
(Section 5). Finally, we present and comment on
I. INTRODUCTION simulation results (Section 6).

Recognition of modulation in received signals is II. SIGNAL MODEL

important for many applications, such as signal
interception, interference identification, electronic First we consider the following complex baseband
warfare, enforcement of civilian spectrum compliance, signal
radar and intelligent modems. The modulation recognition
methods can be divided into two categories. The first is r (k ) x ( k )  n( k ) (1)
modulation recognition with prior information available.
The information provides knowledge of signal parameters where x(k ) is the transmitted signal
such as amplitude, carrier frequency, symbol rate, pulse
shape, initial phase, channel characteristic and noise N 1
power. The second, and more challenging, is modulation x(k ) ¦a e n
j (T n T c )
p (k  nT ) (2)
recognition without any prior information about signal n 0

parameters.
In the past years there have been different approaches and (an , T n ) are the magnitude and phase of a modulation
to solve the modulation recognition problem. These constellation point. c is the initial phase. p ( k  nT ) is
approaches can be classified in three groups. The first the pulse shape function and T is the symbol rate. n(k) is
group includes approaches that use memoryless assumed to be complex white Gaussian noise with power
nonlinearities and detect the spectrum lines occurring for V2 .
specific modulation types [1]. The second group includes
the feature based approaches, where the recognition is III. CLASSIFICATION FEATURES
divided into two stages. The first stage maps the signal
into a smaller feature domain; usually the feature domain The features used in this paper are based on two main
is independent of the signal’s parameters. The second processing steps. The first step is the multiplication of two
stage does the classification of the signal by comparing consecutive signal values. The second step is the
the measured values of features to a priori collocation of statistical characterization of the quantity obtained in the
the feature values for each modulation type [2], [3] and first step. Based on these steps we choose the following
[4]. And the third are the decision theoretic approaches, features to distinguish between modulations:
where in [5,6] all the signal parameters are assumed
known to the receiver. However in [7] the classifier does
1 M 1
not need to know the initial phase. These approaches use 1) ¦ Im( s (k ) s (k  1)) o ASK and PSK/FSK
M k 0

0-7803-8622-1/04/$20.00 ©2004 IEEE 2238

 where M is the number of samples in the realization and * SVC can be applied to separable and non-separable
represents the conjugate operator. data points. In the non-separable case the algorithm adds
one more design parameter. The parameter is the weight
2) Kurtosis(Re( s ( k ) s ( k  1))) o ASK2/ASK4/PSK2 of the error caused by the points present in the wrong
class region. In our application we face this issue in the
3) ª Kurtosis(Re( s (k ) s (k  1))) º o PSK2/PSK4 low SNR cases. On the other hand, in the high SNR cases,
« Kurtosis(Im( s ( k ) s ( k  1))) »
¬ ¼ the algorithm takes its simplest separable case version.
Second maximum ( FFT (s(k)) Another degree of freedom in the SVC is the kernel
4) o FSK2/FSK4 function used. In our application, since we are dealing
Third maximum ( FFT (s(k))
with one and two dimensional features, we used linear and
Based on these features we constructed the polynomial-of-power-2 kernels. Finally the number of
classification tree shown in Fig.1. data points used in the training procedure is also another
parameter that needs to be determined before constructing
IV. SUPPORT VECTOR MACHINE (SVM) the SVM classifier.

SVM is an empirical modeling algorithm that can be V. CLASSIFICATION USING SVM

applied in classification problems. The first objective of
the Support Vector Classification (SVC) is the Fig.2 presents the probability of correct classification
maximization of the margin between the two nearest data of 2000 binary phase shift keying (PSK2) and quadrature
points belonging to two separate classes. The second phase shift keying (PSK4) signals using different numbers
objective is to constrain that all data points belong to the of training points. In this figure we present two cases. The
right class. It is a two-class solution which can use multi- first case is the separable data case (SNR=5dB) where all
dimensions features. The two objectives of the SVC the data points are separated completely and there are no
problem are then incorporated into an optimization misclassified data points. In this case we see that as we
problem. This is done by constructing the dual and primal increase the number of training points, the probability of
problem of the classical Lagrangian problem with correct classification converge towards 1. It should be
transferring the constraint of the second objective to noted that since we are dealing with a two-dimensional
become constraints on the Lagrange variables. The feature, the minimum number of training points needed to
complete derivation of SVC is given in [8-9]. determine the SVM classifier is 3 [8-9]. The second case
is the nonseparable case (SNR=0dB) where some of the
data points are not separated from the data points
Received Signal corresponding to the other class. In this case again, as we
increase the number of training points, we achieve better
probability of correct classification. However, as we
Feature 1 continue increasing the number of training points, we do
not converge towards a specific value; instead we oscillate
< .01

Feature 4

Feature 2
FSK2 FSK4
>Threshold 1 <Threshold 2 <Threshold
1&
> Threshold 2

ASK2 ASK4

Feature 3

>Threshold 3 <Threshold 3

PSK4 PSK4 Fig. 2. The probability of correct classification of 2000 PSK2 and PSK4
signals using different numbers of training points. p=.05 and
Fig. 1. Proposed recognition tree Tc  [0, 2S ] .

2239
around it. This is due to the fact that as we increase the the confusion matrix of the SVM classifier of the same
number of training points, we also increase the number of simulation example presented in Fig.3-5. The results in
misclassified data points which affect the determination of the table are limited to SNR 0 dB and 5dB. Fig. 6 and 7
the discriminating curve. From the simulation results we present the probability of misclassification of 3000 two-
choose for the SVM classifier 25 training points as good level amplitude shift keying (ASK2), four-level amplitude
candidate. At 25 training points we achieve the shift keying (ASK4), PSK2, PSK4, two-carrier frequency
convergence in the separable case and an acceptable shift keying (FSK2) and four-carrier frequency shift
performance in the nonseparable case. keying (FSK4) signals for p={.05,.1}. Each signal has a
Due to the simplicity of the data structure we have, the sampling frequency of 10,000 and time duration of 4
value of C did not affect the classifier structure. From the seconds. In the case of FSK2 the two carrier frequencies
simulation results we did not find much of a difference are {2000,3000} samples/seconds. The center frequency
when we changed the value of C from [1,100]. In our for FSK4 is 2500 with frequency separation of 500
simulation we choose for C the value of 1. samples/seconds.
To determine the kernel used to construct the SVM
classifier in the simulation, we tested two kernels, the
linear and second-order polynomial. In the linear kernel
case the SVM classifier is a straight line separating the
two classes. In the second-order polynomial case the SVM
classifier is a parabolic curve separating the two classes.

VI. SIMULATION AND DISSCUSION

In this section we compare the proposed SVM classifier

to three previously discussed classifiers: the maximum
likelihood classifier proposed by [6-7]; the qLLR
classifier proposed by [5] and finally the cumulant-based
classifier proposed by [2]. Also we compare SVM
classifier to another two proposed classifiers based on the
classification tree in Fig.1: fixed threshold classifier and
dynamic threshold classifier. The dynamic threshold is
determined by the value of SNR. In order to compare Fig. 3. Probability of classification error (Pe) for 2000 PSK2 and PSK4
fairly these classifiers with the SVM classifier, we need to signals for different SNRs. ‘ML’ represents the maximum likelihood
determine the amount of information needed from the classifier, ‘Dynamic tree’ is the proposed dynamic threshold classifier,
‘Fixed tree’ is the proposed fixed threshold classifier, ‘Poly’ is the qLLR
receiver point of view in order for the classifier to operate. classifier, ‘Swami’ is the cumulant-based classifier and ‘SVM’ is the
In the case of the maximum likelihood classifier, the proposed SVM classifier. p=.05.
receiver needs to know all the signal parameters and the
noise power. In the case of signal parameters, this
includes the value of the constellation points and the Actual
random initial phase (c). In the case of qLLR classifier, Classification Modulation PSK2 PSK4
cumulant-based classifier, dynamic threshold and SVM, Output
the noise power and all signal parameters (except the PSK2 573 197
value of the constellation points and the random initial PSK4 427 803
phase) must be known to the receiver. Finally in the case
of fixed threshold classifier, the same scenario as for Table 1: Confusion Matrix of SVM algorithm for SNR=0dB and p=.05.
SVM classifier applies here, except that the receiver does
not need to know the noise power. We now present the
simulation examples in which we compare all six
classifiers. Actual
Classification Modulation PSK2 PSK4
Fig.3-5 present the probability of misclassification as a Output
function of SNR of 2000 PSK2 and PSK4 signals at
p={.05,.1,.2} where p is the ratio of symbol rate to PSK2 1000 0
sampling rate. For each signal we choose sampling PSK4 0 1000
frequency of 500 samples/second; time duration of 4
Table 2: Confusion Matrix of SVM algorithm for SNR=5dB and p=.5.
seconds; random T c  [0, 2S ] ; and each constellation point
has equal probability of occurrences. Tables 1-6 present

2240
Fig. 4. Probability of classification error (Pe) for 2000 PSK2 and PSK4 Fig. 6. Probability of classification error (Pe) for 3000 ASK2, ASK4,
signals for different SNRs. Acronyms are the same as for Figure 2. p= .1. PSK2, PSK4, FSK2 and FSK4 signals for different SNRs. Acronyms are
the same as for Figure 2. p= .05.

Actual
Classification Modulation PSK2 PSK4
Output

PSK2 616 201

PSK4 384 799
Table 3: Confusion Matrix of SVM algorithm for SNR=0dB and p=.1.

Fig. 7. Probability of classification error (Pe) for 3000 ASK2, ASK4,

PSK2, PSK4, FSK2 and FSK4 signals for different SNRs. Acronyms are
the same as for Figure 2. p= .1.

Fig. 5. Probability of classification error (Pe) for 2000 PSK2 and PSK4 Actual
signals for different SNRs. Acronyms are the same as for Figure 2. p= .2. Modulation
ASK2 ASK4 PSK2 PSK4 FSK2 FSK4
Classification
Output
ASK2 0 0 0 0 0 0
Actual
ASK4 10 490 14 4 0 0
Classification Modulation PSK2 PSK4 PSK2 490 2 476 5 0 0
Output PSK4 0 0 10 491 0 0
FSK2 0 0 0 0 500 0
PSK2 589 250 FSK4 0 0 0 0 0 500
PSK4 411 750 Table 5: Confusion matrix of SVM algorithm for SNR 0dB.p=.05.

Table 4: Confusion Matrix of SVM algorithm for SNR=0dB and p=.2.

2241
 Actual [7] W. Wei “Classification of Digital Modulation Using Constellation
Modulation Analyzes.” Ph.D. dissertation. Univ. of Southern California. 1998.
ASK2 ASK4 PSK2 PSK4 FSK2 FSK4 [8] C. Burges “A Tutorial on Support Vector Machines for Pattern
Classification Recognition,” Data Mining and Knowledge Discovery, Vol. 2,
Output 1998, pp 121-167.
ASK2 500 0 0 0 0 0 [9] N. Cristianini and J. Shawe-Taylor An Introduction to Support
ASK4 0 500 0 0 0 0 Vector Machines and other Kernel-Based Learning Methods,
PSK2 0 0 500 0 0 0 Cambridge University Press, 2000.
PSK4 0 0 0 500 0 0
FSK2 0 0 0 0 500 0
FSK4 0 0 0 0 0 500
Table 6: Confusion matrix of SVM algorithm for SNR 5dB.p=.05.

From Fig. 3-7 it is clear that the maximum likelihood

classifier has the best performance among the compared
classifiers. Following the maximum likelihood classifier,
the qLLR comes second for small values of p. However as
p increases, the performance of qLLR classifier
deteriorates. The reason for that is as we increase the
value of p, we decrease the number of samples in the
averaging process used in the qLLR classifier. This
affects the approximation used in the algorithm.
The simulation results also show that dynamic threshold
classifier outperforms the fixed threshold classifier. This
is due to the curvature of the kurtosis curves at
SNR<10dB. In the case of cumulant-based classifier, from
the figures it is clear that the performance of the algorithm
is independent of p.
Finally the SVM classifier shows robust performance
over all simulations (whether distinguishing PSK2 from
PSK4 or applied to the classification tree proposed in
Fig.1) for different values of p. In the SNR=0dB area the
SVM classifier outperforms the dynamic tree classifier
and cumulant-based classifier. This is due to the fact that
the SVM classifier is modified such that it can be used on
nonseparable data.

REFERENCES

[1] J. Reichert. “Automatic classification of communication signals

using higher order statistics”, ICASSP, vol.5, New York, NY,
USA. 1992 pp.221-4.
[2] A. Swami, B.M. Sadler. “Hierarchical digital modulation
classification using cumulants”, IEEE Transactions on
Communications, vol.48, no.3, pp.416-29, March 2000.
[3] K.C. Ho, W. Prokopiw, YT. Chan. “Modulation identification of
digital signals by the wavelet transform”, IEE Proceedings:
Radar, Sonar & Navigation, vol.147, no.4, pp.169-76, Aug. 2000.
[4] H. Deng, M. Doroslovacki, H. Mustafa, X. Jinghao, K. Sunggy.
“Automatic digital modulation classification using instantaneous
features”, 2002 IEEE International Conference on Acoustics,
Speech, and Signal Processing. Proceedings IEEE, vol.4, 2002,
pp.IV4168.
[5] A. Polydoros, K. Kim. “On the detection and classification of
quadrature digital modulations in broad-band noise”, IEEE
Transactions on Communications, vol.38, no.8, pp.1199-211 Aug.
1990.
[6] W. Wei, J. M. Mendel. “Maximum-likelihood classification for
digital amplitude-phase modulations”, IEEE Transactions on
Communications, vol.48, no.2, pp.189-93, Feb. 2000.

2242