The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
129
strength ( CT 4- /T ) and inversely proportional to the
watermark length(n) and embedding strength( A ).
When A * n is fixed, the noise strength (a 2 + p 1 ) can’t be
too big, otherwise, the rate of false alarms will be so big that it
is impracticable to use the discrimination rule to judge whether
there is watermark or not.
When n = 1000, /1 = 1.0,under the random noise attack which
has the statistic characteristic of mean ¡i =0, variance <j =
2.CK 4.CK 6.CK 8.0, the detection result is as figure 2, here, the
500th to-be-detected information contain watermark. From
figure 2 we can get that: the rate of false alarms is small when
variance <j is 2.04.0 and will increase as the variance <j
increase.
1
0.5
0
-0.5
500 1000
variance=2.0
500 1000
variance=4.0
500 1000
variance=6.0
500 1000
variance=8.0
Figure 2. The influence of variance G to detection
In order to get good attack-resistance capability, the “3cr”rule
is erected. It means that there will be no probability of false
alarms when > 3\fnG -This rule request that:
2
A*n_ >
2
3*jn(jU 2 +G 2 )
The 7(y + (j 2 ) can be treated as the random noise strength,
when the watermark detection need to resist the random noise
below yj(jU 0 2 + cr Q 2 ) , the right value of A, Tl must be set to
satisfy the follow inequality
^- > 3 * yjn(jL¿ 2 +G 2 ) (4)
(5)
From the formula (4) and (5), it is known that in order to
promote the attack-resistance capability of detection, we can
increase the value 2 and n ,but considering the central
limit theorem, the value of n can’t be too small, experiment
testify that n should be no less than 800.
4. THE INFLUENCE OF MEAN ^ OF RANDOM
NOISE TO THE WATERMARK DETECTION
Data’s translation has a bad influence on the watermark
extraction and detection (YANG Cheng-song, 2007), Data’s
translation can be treat as the mean (jl) of the random noise.
In order to display the influence of the mean (jLl) of random
noise to the watermark detection,
let g = 10, n — 1000, A — 1.0, change the value of ¡U , the
variation of false alarms is showed in figure 3.
Figure 3. The influence of noise mean to the false alarm
probability
From figure 3 we can get that the mean {JU) of random noise
has a big influence on the watermark extraction and detection.
In order to eliminate the bad influence of mean {JU) to the
watermark detection, restriction is set in the process of
watermark generation. According to the probability
p{w(k) — 1} — —,P{w(k) — —1} — — , a series of
watermark is created, the watermark which meets the inequality
k=n-1
^ w(k) ~ 0 is chosen.
*=o
Random noise can be expressed as g(k) — JU + g\k) , here
g'(k)~N(0,a 2 )
At the same time, we can control the false alarms under certain
threshold p by control A ^ n as the formula: