DETECTION OF DSSS SIGNAL IN NON-

COOPERATIVE COMMUNICATIONS

Liang Chang, Fuping Wang, Zanji Wang

Laboratory on circuits and system, Department of Electrical Engineering, Tsinghua University
Beijing, China, 100084
Email: chang-103gmails.tsinghua.edu.cn
Abstract Direct sequence spread spectrum (DSSS) signal is SNR environment. But it has a very high computational
widely used both in military and commercial communications. complexity. The self-correlation based methods use the
Because of the low power spectrum density, it is very difficult cyclic-stationary characteristic of the DSSS signal, which
to detect in non-cooperative communications. In this paper, a has a lower computational complexity. The method
novel detection algorithm was proposed based on Haar proposed by [4] is the representative method, which has the
wavelet transform (HWT). The characteristics of DSSS signal best performance among all the autocorrelation based
after HWT was analyzed firstly, then the detection algorithm methods.
was implemented by calculating the second order moments of
The wavelet transform can characterize the cyclic-
autocorrelations in wavelet domain. Simulation results show
that the proposed method performs well in low SNR, and it is
stationary characteristic of DSSS signal; meanwhile it has
strong capacity of de-noising. So it's suitable for dealing
better than the representative method.
with DSSS signal by wavelet transform. This paper
I. INTRODUCTION sufficiently utilizes the above character of wavelet
transform. The representation of DSSS signal after wavelet
Direct sequence spread spectrum (DSSS) signal is transform is analyzed firstly. Then both in the conditions of
widely used both in military and commercial fields. It has noise-only present and signal with noise present, the
very low power spectral density, which results in the characterization of second order moment of autocorrelation
transmission below the noise level. So in the field of is given in wavelet domain. Based on the above analysis,
non-cooperative communications, especially in the domain we propose a robust detection statistic and show the
of electronic warfare, the interception of DSSS signal is still detection process. As a byproduct, the symbol period can
a challenging and important problem to be solved. also be estimated.
In non-cooperative communications, the receiver has The paper is organized as follows. In section 2, we
minimal prior knowledge of the DSSS signal: such as chip give the model of DSSS signal, and then the representation
rate, baud rate and spread sequence. So the interception is of DSSS signal after wavelet transform is analyzed. In
usually implemented by the following steps: The first step section 3 the detection statistic and the process of the
is to detect whether there is DSSS signal below the noise detector is proposed. In section 4 we briefly analyzed the
level. When the DSSS signal is present, the modulation statistical characterization of the detector. Section 5
parameters (chip rate and baud rate) are extracted from the evaluates the detector's performance by computer
intercepted signal, and then the spread spectrum sequence simulation. Finally in section 6 the conclusion is drawn.
can be estimated. Finally the de-spreading and
demodulation process can be completed after the above II. SIGNAL MODEL AND ITS WAVELET TRANSFORM
steps. Above all, the detection of DSSS signal is the most
important, which is the basis of the other steps. A. Signal Model
So far there are three kinds of detection methods: We assume the carrier frequency is known to the
Energy radiometry[2], autocorrelation based methods[3,5] receiver. The small frequency offset will not affect the
and spectrum correlation based methods[4]. Energy detector. So without loss of generality, we just deal with the
radiometry is the simplest way, but it does not utilize the baseband model of DSSS signal. Each transmitted symbol
characteristic of DSSS signal, which results in the poor is modulated by a period of pseudo-random sequence. So
performance when the signal to noise ratio is unknown. the DSSS signal can be modeled as follows:
Spectrum correlation method performs well even in low r(t) = Sm(t - lX)p(t - lx) + n(t) (La)
 In the above equation S denotes the signal power; O<t<1,
T7 is a random time delay. n(t)2 is the complex additive F- t (4)
Tr(t) =+t -l<t< 1,
Gaussian noise with power cn . p(t) is the spread 00 t >lor t <-1.
sequence, which is the periodic extension of one period
spread sequence p0 (t) . They can be define as follows: If we fix scale a as a constant, equation (3) is just a
one-dimension function of variable 'r When
p (t) =Epo (t - ITO) ( b -=kT and r.kTo(kEN), (eiV.-ei ) in
equation (3) can be expressed as A=ej0 (ej'n -e7 n+1),
Po(t =Pnq(tlTc -n) .c (1< n < N- 1) . When r = kTo , the expression is
n=1
Bk =eiOej -e +1eJ( . So equation (3) can be expressed
In the above equation, To = NT is the symbol as a periodic extension form:
period; N is the spread
{}N factor; TC is the chip width of
spread sequence; {P 1 is one period spread sequence HwT, (a,)zYQ(r -kkT)ejk ±\SaBjTr -( -kTo)J (5.a)
with value + 1, q(t5 'is a unit rectangular function.
m(t) is the transmitted data which has the following
representation:
While = J AnTr L-( -nTc)
Q(S) (5.b)

m(t) = E m,q(t / To - k) (1 .d)

k=,, III. STRUCTURE OF THE DETECTOR
{ mk } is the modulated sequence defined by the We assume a long baseband DSSS signal r(t) is
corresponding constellation. The normally used modulation intercepted by the receiver. The signal after WT is denoted
types are BPSK and QPSK. as y(t).Then we divide y(t) into M segment. So the
autocorrelation of the m-th segment can be expressed as:
B. Wavelet Transform of DSSS Signal
We consider the continuous wavelet transform (CWT) Ry(-) (r) =
I 'T.
y(t)y* (t -,r)dt (6)
T,
with definitions as follows:
Then consider the second order moment of auto-
t (2) correlations:
CWT(a,)= jJr(t)V= a jdt
p (r) 1 m
M 2 (7)
In the above equation a is the wavelet scale, T is the
time delay. r(t) is the signal to be analyzed. Vf (t) is Then another statistic can be defined as:
called mother wavelet. The typical mother wavelet include
(p(r)) Ir>p
Morlet, Haar and Daubechies wavelet. The paper chooses A (r) = mean ( (8)
the haar wavelet, which has the simplest form, and also has
the strong ability of de-noising and transient In section 4 we will prove that when T is multiple of
characterization. the symbol rate, a peak value in p(r) and p2 (r) will
appear. So the detection statistic can be defined as follows:
In [6] the Haar wavelet transform (HWT) of MPSK
signal in median carrier frequency is given. The DSSS Pd = max (p2 (r)) (9)
signal is essentially a MPSK signal. So with a little
modification, we can get the form of baseband DSSS signal By picking the series of peak value and calculating
in wavelet domain as follows: their space, the symbol period can also be obtained. Figure
1 gives the structure of the algorithm.
HW ='aZ(ej]- -eY )Tr L( -nT)J
v(a,c) -
(3) Input r t Divide into
signal M semnt0
WT ytS S
.
o
mnso
auto corre lati on
Cumulate
he results
In equation (3), V. = (p, + (k ° yp, is the modulation
phase of the nth chip with value belongs to {+27}, while Symbol period
estimation i
(k is the corresponding modulation phase of the kth symbol
with value belongs to {271 M}. T and a has the same
definition as above. Tr (t) is a unit triangle function Threshold
defined as follows: calculate

Figure 1 structure of the detector

 IV. STATISTIC ANALYSIS B. Signal Plus Noise Present
When the signal is hidden in the noise, based on the
In this section we analysis both in the circumstance of character of cyclic stationary in DSSS signal, there will be a
noise-only present and the signal plus noise present, the series of peak values when is multiple of the symbol
statistic characterization of the detector respectively. How period. The autocorrelation at r = kTo (k E N) can be
to calculate the threshold is drawn in section C. expressed as follows:
A. Noise-Only Present
R (kTO) - f s(t)s (t -r)dt (19)
We assume the noise as a complex form of w
n(t) = nI (t) + jnQ (t) . So the integration in equation (6) can
be express as:
= 1
w
jfw s LAn2Tr22 (t-nT) ei( i-Oik)dt
i=-o n=O a

y(t)y (t - r) = n, (t)n1 (t - r) - nQ (t)nQ (t -r) +

j [n, (t - f)nQ (t) - nI (t)nQ (t -i)I If we denotesRs()Ls: 2 nT)Jsoequation(I9)
(10)
can be expressed as:
We discuss the real and imaginary part respectively,
(20)
Rs (kTO)
s
Ye j(o,-O"i) I R(t-kTO)dt
T
w
the mean and variance can be obtained as following: =

TW i=-.0

{Re[y(t)y*(t r)]} E {Im[y(t)y*(t =)]} 0

E ( 11) After a few mathematic manipulations, the mean of its
square modulus reduces to:
Var {Re [y(t)y (t r)]} = Var {Im [y(t)y* (t r)}
- -
(21)
E[n(t)n(t r)] + E[n (t)n (t r)] c4 / 2
_
=
(12) mpHOiz uE{|Rs(Dj<,T2)12 ( a1) S2 N

From equation (21), the larger scale a we choose,

From equation (6) and the central limit theorem, the and the bigger the peak value will be. So the de-noising
real and imaginary part of Ry (r) approximately obeys the ability of WT in a large scale will improve the performance
Gaussian distribution, with mean and variance as follows: of detection. Meanwhile we can estimate the symbol period
by extracting the series of peak value.
E{Re[R, (r)]} = EIm[R (r)]} = (13)
C. Detection Threshold
(14) Based on the analysis in section A, we can obtain the
Var {Re[RY(r)]} = Var {Im[RY(,r)],} = probability of false alarm (PF) in statistic p2 (r) as
follows:
So R (r) 12 = Re2[Ry (r)] + Im2 [Ry (r)] obeys the
exponent distribution. The probability density function (pdf) A-m (22)
HO
P,- A-1
can be expressed as following: PF,12 =Q 07 HO
=Q
P2

f {1 2J2 e2u y> 0 (15)

m )

O y <O The total PF of the detector can be expressed as:

Utilizing the independence of R () 2 for P'F =1(1- F) (23)

different m, we can obtain:
In equation (23) Nw denotes the number of samples in
m =E[p(r)]=
E ELR(r)1 2=2 =T (16) every segment. In the detection process, we can firstly set
the total PF of the detector, and then calculate the PF of
&72 = I
Var [I Ry (r) 2 ] = MT2 (17) statistic p2(r) from equation (23). So the detection
threshold can be obtained by equation (22). The threshold
can also be used to estimate the symbol period.
So the ,mean and variance of p2 (r) is:
(1 8.a) V. SIMULATION RESULTS
M
P2 =E{P2(r)}=
{ E{jp(r)}
} Ep(r)}
Ej
1
In the simulation experiment it is assumed that the
2
P2 VaIP2(rC)}
=Var ()}Var -(r)}
{pO
E 2{p(r)} M
(1 8.b) modulation type of transmitted data is BPSK, the spread
sequence is Gold sequence with spread factor 63, which is
generated by two m-sequence with characteristic algorithm in [5](algorithm 2) with their correct probability
polynomial A(x)= i +X2+ X3+ X4+ x5 and (x)= I +x +x2 +x5+x6 in different symbol length. Similar to the results of
respectively. The sample rate is 8 times of the chip rate, and experiment 2, algorithm 1 is much better than algorithm 2.
the wavelet scale is 8. The total data is 400 symbols with It can also be seen that with the decrease of the symbol
each segment of 2000 samples. length, the performance distinction between algorithm 1
and 2 will increase.
Experiment 1: The input SNR is -15dB. The PF is set
to 0.01 and 0.05 respectively. The threshold is calculated by
the Newman-Person criteria. We use Monte-Carlo
simulations both in the condition of noise-only present
(Conditionl) and signal plus noise present (Condition2).
The detection result is DSSS signal exist (reusltl) or
non-DSSS signal exist(result2) of DSSS signal. In every PF
we run the simulation 1000 times independently. The
performance table can be obtained below.
TABLE I: Performance Analysis of the detector

SNR= PF-0.01 PF-0.05 Number of symbols

-13dB Resultl Result2 Resultl Result2 Figure 3 Correct probability of symbol period estimation
Conditionl 1 0 1 0 VI. CONCLUSION
Condition2 0.004 0.996 0.054 0.946
Due to the DSSS signal's pseudo-random charac-
SNR= PF-0.01 PF0-05 teristic, it is very difficult to be detected in low SNR
-15dB Resultl Result2 Resultl Result2 environment. The paper proposes a blind detection method
Conditionl 0.418 0.581 0.745 0.254
based on the wavelet transform. The algorithm sufficiently
utilizes the capacity of cyclic-stationary characterization
Condition2 0.007 0.998 0.055 0.944 and de-noising of the wavelet transform. So the proposed
algorithms can robustly detection the DSSS signal in low
It can be seen clearly that from table 1 the simulated
SNR, meanwhile estimate the symbol period. Simulation
PF well matches the preset PF, so the theoretical analysis of
results verify the algorithm's effectiveness and superiority
section 4 is correct. It can also be seen that in SNR=-13dB
against other typical algorithm.
and PT-0.01, the total wrong probability is lessn thanl .
Experiment 2: Denote the proposed algorithm as REFERENCE
algorithm 1, the algorithm in [5] as algorithm 2. We [1]. John G. Proakis, "Digital Communications", Fourth Edition, Mac
investigate the ROC of the detector both in input SNR Graw Hill, 2003
-16dB and -18dB. It could be seen clearly that algorithm 1 [2]. Davidovici.S, and Kanterakis.E.G. "Radiometric Detection of Direct-
has evident better performance than algorithm 2. Sequence Spread-Spectrum Using Interference Excision", IEEE
Journal on Selected Areas in Communications, Vol.7, No,4. pp
576-588, May, 1989.
[3]. Polydoros.A, and Weber.C.L, "Detection Performance Consider-
ations for Direct-Sequence and Time-Hopping LPI Waveforms".
0.8 IEEE Journal on Selected Areas in Communications, Vol.3, No.5, pp
727-744, Sep, 1985
[4]. Gardner W.A, "Signal interception: a unifying theoretical framework
0.3 114
-*--
SNR=-16dB,Algonthm
SNR=-16dB,Algonthm 2
1 _ for feature detection", IEEE Trans on Commun. Vol 36, 897-906,
0A2 | + SNR=-18dB,Algonthm 1 1988
l
0.1 +~~~~~1 SNR=-16dB,Algonthm 2_ [5]. G. Burel, C. Bouder and 0. Berder Detection of Di- rect Sequence
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Spread Spectrum Trans- missions without Prior Knowledge Global
PF
COM Vol 1 pp 236-239, 2001.
Figure 2 ROC of the detector
[6]. K.C.Ho, W.Prokopiw and Y.T.Chan Modulation identification of
Experiment 3: Investigate the performance of symbol digital signals by the wavelet transform [J]. IEE Proc-Radar, Sonar
period estimation. The input SNR was set to -15dB. Navig, Vol.147, No.4, August 2000.
Compare the proposed algorithm (algorithm 1) and the