Detection and Removal of Hidden Data in Images Embedded with Quantization Index Modulation - Computer Network Security - page 367

camera. Embedding positions, number of DCT coefficients involved, and watermark-
ing intensity were selectable depending on the robustness requirement and the amount
of watermark data. The embedding capacity was one bit per 8×8 block.

5.1 Detectability

For the one-layer transform-domain embedding, discreteness of the histogram is eas-
ily identifiable and cannot be missed. A total of 70 images sized 256×256 and marked
with the DCT/QIM technique were tested, and all showed clear discreteness in the
histogram of relevant DCT components. The estimated quantization steps were in
agreement with the actually used

∆ as shown in Table 1.

Table 1. Quantization steps and the estimated results

Estimated ∆

Selected DCT
index N

∆ actually used in
embedding

Mean Standard deviation

20 20 20.06 0.15

35 30 30.09 0.21

51 35 35.84 0.97

For the dual-transform/QIM embedding, detection was done as follows.
Step 1: Caculate energy of every AC component for each block, V 2 , V 3 , … , V 64 .
Step 2: Search for the local maxima M i .
Step 3: Caculate a predicted value M i ' from V i - 2 , V i - 1 , V i +1 and V i +2 :

M ' = 2 V − V + 2 V − V .

(9)

i i − 1 i − 2 i + 1 i − 2

Step 4: Caculate the parameter i , which is proportional to the prediction error:

( )

1 4 M − M '

( )

i i

α = M − M ' = .

(10)

i i i

V + V + V + V

V

i − 2 i − 1 i + 1 i + 2

i

Step 5: Compare i with a preset threshold T to make a decision. The threshold
may vary from database to database. Here we have chosen T = 2.5.
A group of 91 images in was used in the experiment. The results are given in Table
2, showing that the detection performance is dependent upon both the strength and the
frequency-domain location of the embedding. The proposed method performs better
when the hidden data are embedded in coefficients with a relatively high index.

Table 2. Performance of dual-transform/QIM watermark detection.

Actually used
∆

Number of false
alarms*
29 28 80 11 1
30 35 86 5 1
40 40 89 2 1
45 35 91 0 1
* False alarm means that a coefficient is erroneously judged as marked but were actually not.

Coefficient
index N

Number of
detected marks

Number of
missings