fully-distributed spectrum sensing: application to
TRANSCRIPT
Fully-distributed spectrum sensing:application to cognitive radio
Philippe Ciblat
Dpt Comelec, Télécom ParisTech
Joint work with F. Iutzeler (PhD student funded by DGA grant)
Cognitive radio principle
Spectrum is, at a first glance, entirely used
However, at given time, an assigned subband can be free
⇒ white space
Two kinds of users
− Primary : have paid for using an pre-assigned subband− Secondary : are allowed to use a white space
Insert secondary users into white spaces
How detecting the presence of a primary user ?
Philippe Ciblat Fully-distributed spectrum sensing 2 / 15
Hidden terminal issue
Primary Receiver
Primary Transmitter
Secondary user
Secondary user
Philippe Ciblat Fully-distributed spectrum sensing 3 / 15
Hidden terminal issue
Primary Receiver
Primary Transmitter
Secondary user
Secondary user
Philippe Ciblat Fully-distributed spectrum sensing 3 / 15
Hidden terminal issue
Reception disturbance
Primary Receiver
Primary Transmitter
Secondary user
Secondary user
Problem : a secondary user is disturbing the primary receiver
Solution : secondary users have to cooperate to detect the primary user
Philippe Ciblat Fully-distributed spectrum sensing 3 / 15
Two ways for cooperating
Centralized detection (Fully)-Distributed detection
Fusioncenter
Primary Transmitter
Primary Receiver
Secondary user
Primary Transmitter
Primary Receiver
Secondary user
Detection with more than one sensors
If fusion center is available, centralized (also called distributed) detection
If fusion center is not available, (fully)-distributed detection
− robust against nodes attack− simple network management
Philippe Ciblat Fully-distributed spectrum sensing 4 / 15
System model
{
H0 (absence of primary user) : yk (n) = bk(n) k = 1, · · · ,KH1 (presence of primary user) : yk (n) = xk (n) + bk (n) n = 1, · · · ,Ns
with
secondary user index k and time index n
bk(n) Gaussian with variance σ2k known at secondary user k
{xk (n)}n coming from primary user known at secondary user k
Performance metric
Detection probability : PD = P(H1|H1)
False alarm probability : PFA = P(H1|H0)
Goal : minimizing PFA such that PD ≥ Ptarget
D
Remarks :
If {xk (n)}n unknown but Gaussian ⇒ Energy detector
Hard detection (local decision and then voting) not considered
Philippe Ciblat Fully-distributed spectrum sensing 5 / 15
Reminder on (soft) centralized detection
Optimal test : Log-Likelihood Ratio (LLR)
Λ(y) = log
(
p(y|H1)
p(y|H0)
)
H1
≷H0
µ, with µ chosen for ensuring Ptarget
D
Application to our practical case :
T (y) =1
K
K∑
k=1
tk (yk)H1
≷H0
η, with tk (yk) =yT
k xk
σ2k
and yk = [yk (1), · · · , yk (Ns)]T, xk = [xk (1), · · · , xk (Ns)]
T, (.)T = transpose.
Threshold computation
η =√ςT Q
(−1)(
Ptarget
D
)
+ mT
where Q(−1) is the inverse of the Gaussian tail function, and
mT = Ns
(
1
K
K∑
k=1
SNRk
)
and ςT =Ns
K
(
1
K
K∑
k=1
SNRk
)
.
Philippe Ciblat Fully-distributed spectrum sensing 6 / 15
Fully-distributed detection (I)
Gossiping step of duration NgSensing step of duration Ns
Tk ≈ averageℓ (tℓ)tℓ = yT
ℓ xℓ/σ2
ℓ for each node ℓ
Question : How computing the average of tℓ in a distributed way
⇒ Gossiping (also called consensus) algorithms
Gossiping algorithm description : an example (Pairwise Gossip)
x(0) = [x1(0), · · · , xK (0)]T : initial values
At time t , a node i wakes up and calls one
of its neighbor j . Then
xi(t + 1) = (xi(t) + xj(t))/2
xj(t + 1) = (xi(t) + xj(t))/2
⇒ x(t + 1) = W(t)x(t)t→∞→ xaverage1
T1(y)...
TK (y)
= P
t1(y1)...
tK (yK )
with P = (pkℓ)k,ℓ=1,··· ,K the
gossiping algorithm matrix
after Ng iterations.
Philippe Ciblat Fully-distributed spectrum sensing 7 / 15
Fully-distributed detection (II)
The final test function at node k is
Tk(y) =K∑
ℓ=1
pkℓyTℓxℓ
σ2ℓ
H1
≷H0
ηk ,
where the threshold (for pre-defined Ptarget
D ) is given by
ηk =√ςk Q
(−1) (P
target
D
)
+ mk
with
mk = Ns
K∑
ℓ=1
pkℓSNRℓ and ςk = Ns
K∑
ℓ=1
p2kℓSNRℓ.
Problem
Threshold not computable in a distributed way due to the terms p2kℓ.
Philippe Ciblat Fully-distributed spectrum sensing 8 / 15
Two approaches for threshold computation (I)
Approach 1 : distributed with knowledge of K
η(1)k =
√
ς(1)k Q
(−1) (P
target
D
)
+ mk with ς(1)k =
Ns
K
K∑
ℓ=1
pkℓSNRℓ.
Approach 2 : fully-distributed
Using Sum-Weight-like gossip in order to perform the average and the sum.
z := Qt, w(1) := Q1, w(e) := Qe
where
Q the gossip algorithm matrix after Ng iterations,
e the K -sized vector whose first component is 1 and the others 0.
Each node k calculates the k -th component of
zp = z ⊘ w(1) = Pt → taverage1 and zs = z ⊘ w
(e) = St → tsum1
where
⊘ the elementwise division.
P = diag (1 ⊘ Q1)Q and S = diag (1 ⊘ Qe)Q.
Philippe Ciblat Fully-distributed spectrum sensing 9 / 15
Two approaches for threshold computation (II)
The threshold is then as follows
η(2)k =
√
ς(2)k Q
(−1) (P
target
D
)
+ mk with ς(2)k = Ns
(
∑Kℓ=1 pkℓSNRℓ
)2
∑Kℓ=1 skℓSNRℓ
.
Remarks
Algorithm fully distributed since the number of nodes not required.
New threshold does not ensure the target probability of detection
PD(k) = Q
√
ς(2)k
ςk
Q(−1) (
Ptarget
D
)
.
Philippe Ciblat Fully-distributed spectrum sensing 10 / 15
Numerical illustrations
Setup :
T = Ns + K = 128 with Ns = Ng = 64.
Ptarget
D = 0.99
K = 10
Considered algorithms : energy-based algorithm
centralized detection
pairwise gossip (PG) with centralized threshold
pairwise gossip with Approach 1 based distributed threshold
broadcast sum-weight gossip (BWG) with Approach 2 based distributed
threshold
Philippe Ciblat Fully-distributed spectrum sensing 11 / 15
Performance analysis
PFA and PD versus mean SNR PFA and PD versus Ns
Fully-distributed algorithm performs well
Sensing time equivalent to gossiping time
Philippe Ciblat Fully-distributed spectrum sensing 12 / 15
Hidden terminal context
Hidden terminal configuration PFA and PD versus T
Fast convergence for fully-distributed algorithm
Hidden terminal issue is fixed
Philippe Ciblat Fully-distributed spectrum sensing 13 / 15
Comparison with existing approach
Training based algorithm
Comparison with “A. Sayed, Distributed detection over adaptive networks
using diffusion adaptation, IEEE Trans. Signal Processing, May 2011”.
In this paper, sensing and gossiping steps are interleaved
Outperforming existing approach
More relevant choice of threshold
Philippe Ciblat Fully-distributed spectrum sensing 14 / 15
Conclusion
New fully-distributed detection algorithm
Outperforms existing approaches
Other works :
Max-consensus : asynchronous algorithm
Average-consensus : fast algorithm (BWG) and theoretical analysis
... distributed optimization (to be done)
References :F. Iutzeler and P. Ciblat, “Fully-distributed spectrum sensing : application to cognitive radio”, submitted for publicationto Eusipco, 2013.F. Iutzeler, P. Ciblat, and W. Hachem, “Analysis of Sum-Weight-like algorithms for averaging in Wireless SensorNetworks”, accepted for publication to IEEE Trans. Signal Processing.F. Iutzeler, P. Ciblat, and J. Jakubowicz, “Analysis of max-consensus algorithms in wireless channels”, IEEE Trans.Signal Processing, vol. 60, no. 11, pp. 6103-6107, November 2012.
Philippe Ciblat Fully-distributed spectrum sensing 15 / 15