SECURE AND ROBUST COPYRIGHT PROTECTION FOR

H.264/AVC BASED ON SELECTED BLOCKS DCT

K. Ait Saadi

, A. Bouridane

and H. Meraoubi

Centre de Développement des Technologies Avancées, Division Architecture des Systèmes

Cité 20 Août 1956, Baba Hassen, Alger, Algeria

School of Electronics, Electrical Engineering and Computer Science, Institute of Electronics

Communications and Information Technology, Queen's UniversityBelfast, Belfast BT7 7NN, U.K.

Keywords: Copyright protection, H.264/AVC, digital watermarking.

Abstract: This paper proposes a new block based DCT selection and a robust video watermarking algorithm to hide

watermark is first quantized and securely inserted. To achieve invisibility and robustness, the high entropy

DCT 4x4 blocks within the macroblocks are selected to minimise the distortion caused by the embedded

watermark and then scrambled using Linear Congruential Generator (LCG) technique. This approach leads

to a good robustness by maintaining good visual quality of the watermarked sequences. The experimental

results demonstrate the effectiveness of the algorithm against some attacks such as re-compression by the

H.264 codec, transcoding and scaling.

1 INTRODUCTION

Nowadays, most digital applications such as Internet

multimedia, wireless video, personal video

recorders, video-on-demand, videophone and

videoconference face two main problems: (i) to

improve the computation speed of the compression

to meet bandwidth criteria and best video quality as

possible; (ii) how to protect the copyright of the

digital products. To address the first point,

H.264/AVC video compression standard, which is

the latest and most advanced video system to date

and developed jointly by ITU and MPEG, provides a

far more efficient solution for compressing video

than any other compression method available. It

typically outperforms all existing standards by a

factor of three to four especially in comparison to

MPEG-2 (Wiegand, 2003). For the second point,

there has been significant interest in watermarking

which is used for owner identification, royalty

payments, and authentication by determining

whether the data has been tampered with in any

manner from its original form (Sang-Kwang, 2000).

As a result, a large number of watermarking

schemes have been proposed to hide copyright

marks and other information for different video

codecs, however a few works related to H.264/AVC

can be found in the literature. Recent research into

H.264/AVC included the fast implementation of

integer discrete cosine transform (DCT) and variable

block size motion compensation (Fan, 2006).

However, higher compression ratio leads to the

difficulty in balancing among trade-off requirements

for watermarking H.264/AVC video data (Zhang,

2003) (Qiu, 2004).

The idea of compressed-domain watermarking of

videos is not new (Zhang, 2005) (Zhang, 2007) and

this paper proposes a new scheme of blocks DCT

selection and a robust video watermarking

algorithms to hide copyright information in the

compressed domain of the emerging video coding

standard H.264/AVC. The greyscale watermark can

be treated as watermarks for copyright protection of

company trademarks or logos.

2 THE PROPOSED VIDEO

WATERMARKING SYSTEM

Our proposed H.264 watermarking algorithm is

based on Qiu et al. technique (Qiu, 2004) where they

proposed a hybrid watermarking scheme in the DCT

351

Ait Saadi K., Bouridane A. and Meraoubi H. (2008).

SECURE AND ROBUST COPYRIGHT PROTECTION FOR H.264/AVC BASED ON SELECTED BLOCKS DCT.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 351-355

DOI: 10.5220/0001932203510355

 SciTePress

Figure 1: Watermark embedding scheme.

domain and a fragile watermarking in motion

vectors. Our approach operates in the DCT domain

at the macroblock level of I-frames. I-frames are

chosen for watermark embedding because their

existence is crucial for the video signal. Also, P- and

B-frames are highly compressed by motion

compensation and there is less capacity for

embedding a watermark into them. The luminance

component of macroblocks in an I-frame is intra-

coded in 16×16 or 4×4 intra-prediction modes. Each

4 × 4 block of residual data is transformed by an

integer transform after the intra-prediction stage. If

the macroblock is coded in the 16 × 16 intra-

prediction mode, the dc coefficients of all 4 × 4

blocks are transformed by a 4 × 4 Hadamard

transform after the 4 × 4 integer transform to further

decorrelate these coefficients [1]. We only embed

the watermark in the quantized ac residuals of the

luminance component of 4 × 4 intra-predicted

macroblocks. We do not embed the watermark in the

16 × 16 intra-predicted macroblocks for two reasons.

First, the 16×16 intra-prediction mode is used for

smooth regions of the frame, and watermark

embedding causes visible artefacts there. Second,

the extra Hadamard transform for this macroblock

decorrelates the dc coefficients much further so that

many of these coefficients are set to zero.

2.1 Embedding Process

The overall scheme of the proposed embedding

process is shown in Figure. 1. The proposed

algorithm embeds the watermark in one quantized ac

coefficient Xq(u,v) in the high frequency along the

diagonal positions (i.e., u=v) of a high entropy and

scrambled blocks within a macroblock. In order to

survive the recompression, the watermark signal

W(u,v) must be strong enough to survive the

quantization (Qiu, 2004), so that:

1]QP),v,u(W[quant)v,u(W

≥=

(1)

where the quant[.] denotes the quantization

parameter, QP denotes the quantization parameter

and (u, v) denotes a position in a 4x4 block Bk.

The watermark W(u,v) is quantized with

different Qp where each is associated to different

Qstep. The embedding is done with the Qstep giving

the minimum distance d

min

between the coefficient

quantified Xq(u,v) and the Wq(u,v). Our experiments

show that the ac coefficients in diagonal positions

are more stable than others. Effectively, we tested

the insertion with all the quantized ac coefficients

within the block and the best ac coefficient that

verifies the invisibility of the watermark is found at

position (1,1). The insertion is performed by

replacing the Xq(1,1) by the watermarked coefficient

as follow :

{

}

⎩

⎨

⎧

0 Wnif 0

1Wn if )v,u(W),v,u(Xmax

(2)

where wn is the bit to be embedded. We notice the

ac coefficient Xq(u, v) is cleared if ‘0’ is embedded.

It can be justified by the fact that the Xq(u, v) is zero

in most cases. It will not introduce significant

artefacts

To obtain robust and invisible watermarks, the

blocks with high entropy are selected for

embedding.

Define a finite set of data D = {x

, ..., x

}

generated according to a probability distribution P =

Bitstrea

4x4

Block

Key

4x4

Integer

DCT

Intra/Inter

Prediction &

Mode Decision

Entropy

coding

Quantization

Watermark

Quantization

Embedding

process

⊕

Calculationofentropy,

Selection and Scrambling

of the high entropy

blocks

Inverse Entropy,

Quantization & DCT

Frame

Buffer

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

352

, ..., p

}. Shannon’s entropy (Thiemer, 2006) of P

is given by:

plog.p)P(H

∑

−=

(3)

From the source-channel point of view H(P)

stands for the average size of the code necessary to

transmit data from D when these are generated by P.

In our experiments D is the set of gray level values

taken from a current block (in practice N=256).

Given a block B(u, v), where (u, v) is its location in

a macroblock of 16×16 pixels. Let H(B(u, v)), be the

entropy of the gray level distribution in B(u, v). We

derive a slight variation of (3) as following:

255

plog.p))v,u(B(H

∑

−=

(4)

{pj} is the probability distribution of the gray level

values in B(u, v).

Once the entropies blocks of each macroblock

are calculated, they are ranked in descending order

to choose those blocks with high value of entropy.

To improve the robustness of this approach, the

position G(i) of the selected high entropy blocks

within a macroblock are randomly scrambled using

the simplest and fastest random generator called

“Linear Congruential Generator” (LCG) (Raymond,

2006) such as:

M mod )cr )1i(G*m()i(G +−=

(5)

where m, cr and M represent the multiplier, the

increment and the modulus, respectively. They are

chosen to maximise the period which cannot exceed

M and which is equal to 16 in our case. G(0) is a key

for the insertion and extraction of the watermark.

2.2 Detection Process

Watermark detection is performed after entropy

decoding: the bitstream is partially decoded to obtain

the transformed ac coefficients. This is followed by

applying the LCG with the same key used in the

insertion process to select the watermarked blocks.

For each selected block containing DCT coefficients

in the I-frames, the watermark bit is determined as

follows:

(6)

3 EXPERIMENTS AND RESULTS

The proposed watermarking technique has been

integrated into the H.264 J-M-7.6 reference software

(H.264/AVC Joint Model 7.6 (JM-7.6) Reference

Software). The standard test video clips include

Foreman, Claire, Stefan. All frames are coded at 30

frames/s at 372 kbits/s.

All video clips are coded in QCIF format

(176x144). P and B frames are not used in our

experiments because there are very few nonzero

DCT coefficients due to the efficient compression

performance. A small 16x16 grayscale watermark is

used for the experiment (Figure. 2). In each I-

frames, the watermark to be inserted must be

adjusted with the blocks into the macroblock. By

embedding the watermark in the higher entropy

blocks within a macroblock, a total of 256 bits are

embedded as copyright owner’s signature in a frame

QCIF (176x144) resolution according to the position

of the blocks generated by the LCG. By comparing

to [5] where the total of 99 bits are embedded in a

frame, the embedding process developed increases

the capacity of insertion about 157 bits by frame

while still maintaining a high visual quality. The

experiment shows us that only the QP = [28,32,36],

corresponding to typical QPs for low bit-rate

applications are suitable to obtain the

imperceptibility and the integrity of the watermark.

Figure. 3 shows the insertion tests performed on

different values of QP: for QP less than 28, the

watermark is not completely removed, beyond QP =

36 the watermark is visible.

Figure 2: Logo image as watermark.

Figure 3: Extracted watermark with different values of

QP.

Figure. 4 shows the insertion tests performed on

different values of QP for the videos "Claire" and

"Container" and their correspondent PSNR (dB),

respectively. The quality degradation is clearly

shown on the picture for the insertion performed

with QP greater than 36. Figure. 5 illustrates the

average PSNR comparison results for a set of test

sequences.

QP = 16 QP = 24 QP = 26 QP = 28

⎩

⎨

⎧

≥

∧

otherwise 0

0)v,u(X if 1

SECURE AND ROBUST COPYRIGHT PROTECTION FOR H.264/AVC BASED ON SELECTED BLOCKS DCT

353

QP = 24

PSNR= 40.38 dB

QP = 28

PSNR= 40.00 dB

QP = 40

PSNR= 39.02 dB

QP =48

PSNR= 36.28 dB

On average, the watermarking leads to a

decrease of approximately 0.06 dB to 0,13 dB and in

the experiments, no visible noise can be observed in

the test video sequences. Figure. 6 shows Foreman

and Container as examples of video watermarked

sequences with PSNR (dB).

Some of general video manipulations are

performed as attacks to evaluate the effectiveness of

the robust watermark: two codecs video H.264/AVC

and H.263 are applied to re-compress the

watermarks videos, the transcoding from the YUV

(4:2:0) to RGB (4:4:4) format and in the third

experiment, we investigated robustness to scaling.

This attack rescales the video sequences from the

QCIF (276x144) resolution to the CIF (352×288)

resolution. Table 1 shows the reconstructed

watermark from the marked and attacked Claire and

Container sequences with the PSNR (dB). The

reconstructed watermark after watermarking is

highly correlated to the original pattern. This

demonstrates and supports the feasibility of the

proposed approach.

Figure 4: Watermarked Claire sequence with different

values of QP.

4 CONCLUSIONS

In this paper, we have proposed a watermarking

algorithm for H.264 that is robust to some general

video manipulations. We have achieved this goal by

calculating the entropy of the gray level distribution

in the blocks to determine which of these are well

textured and thus suitable for watermark embedding.

To improve the robustness, we used a key-dependent

algorithm (LCG) to randomly select the high entropy

blocks within the macroblock. The simulation results

shows that we increased the payload about of 157

bits by frame comparing to (5) where the total of bits

are 99 without a noticeable change in perceptual

quality. Thus, future work leads to verify the

robustness of the proposed approach against others

common image processing.

Figure 5: PSNR (dB) of marked and unmarked CIF-size of

Foreman, Claire, Stefan, Coastguard, Flower, Bridge,

Container (denoted by For, Cla, Ste, Coa, Flo, Bri, Con in

the horizontal axis) at 372 kbit/s.

Figure 6: The 3

frames of unmarked and marked

Container (a) and Foreman (b) sequences (at 372 kbits/s,

QCIF size).

Unmarked (37,94dB) (a) Marked (37, 85 dB)

Unmarked (37,45 dB) (b) Marked (37, 39 dB)

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

354

Table 1: Reconstructed watermark after manipulations and

the corresponding correlations.

Unmarked

Test

Sequences

Claire

PSNR=41.29

Container

PSNR=37.9

4 dB

Original

watermark

Extracted

from

watermarked

video

ρ=1

PSNR=41,00 dB

ρ=1

PSNR=37.88

Extracted

after re-

compressed

with H.264

coder

ρ=1

PSNR=39.47dB

ρ=0.97

PSNR=37.65

Extracted

after re-

compressed

with H.263

coder

ρ=0.35

PSNR =35.69dB

ρ=0.44

PSNR=32.02

Extracted

after

transcoding in

the RVB

(

4:4:4

)

format

ρ=0.93

PSNR= 38.08dB

ρ=0.97

PSNR=37.58

Extracted

after scaling

ρ=0.97

PSNR=38.62 dB

ρ=0.94

PSNR=37.36dB

REFERENCES

Wiegand, T., Sullivan, G. J., Bjøntegaard, G., and Luthra,

G. J. 2003. Overview of the H.264/AVC video coding

standard. IEEE Trans. Circuits Syst. Video

Technol.13, 7 (Jul. 2003), 560–576.

Sang-Kwang, L., and Yo-Sung, H., 2000. Digital Audio

Watermarking In the Cepstnun Domain. IEEE

Transactions on Consumer Electronics. 46, 3 (Aug.

2000).

Fan, C. P. 2006. Fast 2-dimensional 4x4 forward integer

transform implementation for H.264/AVC. IEEE

Trans. Circuits Syst. II. Exp. Briefs. 53, 3 (Mar.2006),

174–177.

Zhang, J. and Ho, A. T. S. 2003. An efficient digital

image-in-image watermarking algorithm using the

integer discrete cosine transform (IntDCT). In Proc.

IEEE Joint Conf. 4th Int. Conf. Info. Commun. Signal

Process. and 4th Pacific-Rim Conf. Multimedia. 2

(Dec. 2003), 1163–1167.

Qiu, G., Marziliano, P., Ho, A. T. S. D., He, J., and Sun,

Q. B. 2004. A hybrid watermarking scheme for

H.264/AVC video. In Proc. 17th Int. Conf. Pattern

Recogn., Cambridge, U.K., (Aug. 2004).

Zhang, J., and Ho, A. T. S. 2005. Robust Digital Image-in-

video Watermarking for the, Emerging H.264/AVC

Standard. IEEE Workshop on Signal Processing

Systems Design and Implementation. (Nov. 2005),

657-662.

Zhang, J., and Ho, A. T. S. 2007. Robust video

watermarking of H.264/AVC. IEEE Trans. Circuit and

Systems. 54, 2 (Feb. 2007), 205–209.

Thiemer, H., Sahbi, S., and Steinebach, M. 2006. Using

Entropy for Image and Video Authentication

Watermarks. In IS and T/SPIE Conference on

Security, Steganography and Watermarking of

Multimedia Contents (SPIE), San Jose, California,

6072-607218 (Jan. 2006).

Raymond, S., Lee, T., and Lam, H. W. S. A. 2006.

Chaotic Real-time Cryptosystem using a Switching

Algorithmic-based Linear Congruential Generator

(SLCG). IJCSNS International Journal of Computer

Science and Network Security, 6, 8B (Aug. 2006).

H.264/AVC Joint Model 7.6 (JM-7.6) Reference Software.

[Online].Available:

http://iphome.hhi.de/suehring/tml/download/ol.

Bowman,

SECURE AND ROBUST COPYRIGHT PROTECTION FOR H.264/AVC BASED ON SELECTED BLOCKS DCT

355