You are on page 1of 16

J. Vis. Commun. Image R.

38 (2016) 2944

Contents lists available at ScienceDirect

J. Vis. Commun. Image R.


journal homepage: www.elsevier.com/locate/jvci

A robust watermarking framework for High Efficiency Video Coding


(HEVC) Encoded video with blind extraction process q
Tanima Dutta , Hari Prabhat Gupta
Dept. of Computer Science and Engineering, Indian Institute of Technology (BHU) Varanasi, India

a r t i c l e i n f o a b s t r a c t

Article history: Digital watermarking is an efficient and promising means for copyright protection of multimedia objects.
Received 7 March 2015 Digital videos are stored and transmitted in a compressed format, which has drawn a great deal of atten-
Accepted 9 December 2015 tion in compressed domain watermarking for video. The embedding and extraction of watermark bits in
Available online 2 March 2016
compressed domain, therefore does not require complete decoding and re-encoding of the compressed
video. There have been several compression standards for video. High Efficiency Video Coding (HEVC),
Keywords: a successor to H.264 Advanced Video Coding (AVC), is the latest standard for video compression with high
Blind extraction
compression efficiency. In this paper, we propose a robust watermarking framework with a blind extrac-
Compressed domain
Copyright protection
tion process for HEVC encoded video. A readable watermark is embedded invisibly in 4  4 intra pre-
Digital video dicted blocks of the HEVC encoded video. Our watermarking framework enforces security by exploring
High Efficiency Video Coding (HEVC) the spatio-temporal characteristics of the compressed video and a random key for selection of embedding
Key extraction regions. We also analyze the strengths of different compressed domain features of HEVC encoded video
Robust watermarking for implementing the embedding algorithm. Experimental results demonstrate that the proposed work
Security restrict the increase in video bit rate and degradation of perceptual quality. The proposed framework
can also survive filtering, compressions, and noise additions maintaining good quality and robustness.
2016 Elsevier Inc. All rights reserved.

1. Introduction many applications, complete decoding of video sequences is not


always suitable. Consequently, compressed video watermarking
In the digital world of today, humans can arbitrarily and easily [3,4] has gained more attention. In addition, the embedding and
access or distribute digital media from networks due to the rapid extraction of watermark in compressed domain has less computa-
advancement in network technology. The protection of intellectual tional overhead [3] since complete decoding and re-encoding of
property becomes important for the society and needs more atten- the video stream is not required for both embedding and extrac-
tion. This has been a serious threat to digital media producers and tion of watermark bits. Several video compression standards have
copyright owners to protect media from a potential intruder to been emerging in the last few years. The aim of each standard is to
avoid loss in business [1]. Digital watermarking is a favorable provide more compressed data with better perceptual quality (or
method for copyright protection, where a digital code (generically simply quality). High Efficiency Video Coding (HEVC), also known
called watermark) is embedded in the digital media and typically as H.265, is the most efficient and latest compression standard
contains the information about the copyright owner, the creator used in a wide range of applications [5]. HEVC has been designed
of the media, or the authorized customer. to address essentially all existing applications of H.264/MPEG-4
Introducing watermarks in digital videos can be useful to safe- Advanced Video Codec (AVC). HEVC mainly focus on increasing
guard copyright. The watermark can be inserted either in uncom- video resolution and the use of parallel processing architectures.
pressed (raw) video [2] or compressed video [3,4]. Video signals HEVC provides improved compression efficiency and support for
are often stored and transmitted in a compressed format. Applica- differentiated and premium 4 K content and multi-view encoding
tion of uncompressed watermarking techniques for compressed for broadcast and online vendors to bundle more channels/ content
video sequences, however, needs complete decoding and re- on existing delivery mediums.
encoding of the video for watermark embedding or detection. In To minimize computational overheads and visual impacts, a
solution for encryption and watermarking compliant with H.264/
AVC and HEVC video coding standards is designed in [6]. The com-
q
This paper has been recommended for acceptance by M.T. Sun. pressed domain watermarking algorithms for H.264/AVC encoded
Corresponding author. video are not very suitable for watermarking in HEVC encoded

http://dx.doi.org/10.1016/j.jvcir.2015.12.007
1047-3203/ 2016 Elsevier Inc. All rights reserved.
30 T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944

video since unlike H.264/AVC, (1) HEVC has coding tree blocks Ho [9] proposed an embedding scheme for video watermarking,
(CTBs) of size 16  16, 32  32, and 64  64; (2) HEVC has 35 intra where the authentication information is represented by a prepro-
modes in total with 1/32th sample accuracy (bilinear) and adaptive cessed binary sequence and embedded into mid frequency coeffi-
smoothing filter for most reference samples; (3) Symmetric predic- cients in I frame. To improve the security in watermark, the sign
tion block (PB) partitioning (like intra) and asymmetric PB parti- of the coefficients is changed based on the watermark. The water-
tioning; (4) In merge mode, one derived motion vector candidate marking scheme presented in [9] improved in [10] by focusing on
is selected based on neighbors of temporal and spatial motion vec- gray scale patterns and characters. Qiu et al. [11] proposed embed-
tors to save bits (allows direct-like modes); (5) HEVC has more ding the robust watermark onto quantized discrete cosine trans-
transform sizes, such as, 4  4, 8  8, 16  16, and 32  32; and form (DCT) coefficients of intra frames and the fragile watermark
(6) out of 35 intra prediction modes, 1, 2, and 434 are angular into motion vectors of inter frames. Kuo and Lo [12] improved
modes. the video watermarking technique proposed in [11] by choosing
In this paper, we propose a novel video watermarking frame- more appropriate locations for embedding both robust and fragile
work for the HEVC encoded video that has the following salient watermarks in H.264 video during video encoding.
features: In [13], the watermark is embedded by replacing some bits in
the bit-stream directly, but the pre-embedding analysis is compu-
1. Quality: The degradation of the quality of the watermarked tationally complex. The same authors proposed a robust and non-
video due to embedding distortion is minimized by appropri- blind watermarking technique by using the Watson Visual Model
ately selecting intra predicted blocks for embedding of water- [14] for embedding in I frame [15]. The proposed non-blind tech-
mark bits. nique [15] extended for the P frame in [16], where watermark bits
2. Robustness: We propose a robust algorithm for selecting blocks are embedded in all nonzero coefficients of P frame.
for watermark embedding based on spatio-temporal features of A blind method to embed in I frames of the compressed video is
the compressed video. The algorithm minimizes the synchro- proposed in [3]. Block selection is performed based on the value of
nization error. The watermark embedding and extraction algo- prediction modes and the number of nonzero quantized coeffi-
rithms are also robust in nature. We also propose a framework cients (NNZ) of a 4  4 block. The nonzero coefficients are changed
consisting of a public key and a private key to prevent from the to zero value for hiding watermark bits. A data hiding framework is
self collusion attack. The public key is extracted from spatio- designed in [17] to select the embedding region using the concepts
temporal characteristics of the compressed video. of forbidden-zone-data-hiding. The parity of sign of the coefficients
3. Security: We secure the watermarking framework by using a and the value of mid frequency coefficients are changed for water-
random key to select candidate blocks from a set of blocks marking in I frame [18]. In [19], the watermark embedded in non-
selected previously based on spatio-temporal characteristics zero coefficients of P frame in compressed domain to have a better
of the compressed video. perceptual quality of the watermarked video and a limited increase
4. Blindness: A blind watermarking framework does not require in video bit rate. In [20], a non-blind H.264/CAVLC structure-
the original video at the decoder to extract watermark bits. preserving substitution watermarking algorithm is proposed that
We propose a blind watermarking framework where a com- embeds watermark by bit substitutions in the motion vector differ-
pressed and encrypted private key is sent to the authorized cli- ences of non-reference frames. Su et al. [21] proposed a non-blind
ent through a secured channel to extract the watermark. This method to embed watermark in I frames and P frames. A water-
minimizes the transmission overhead of sending a location mark is embedded using spread spectrum and coefficients are
map of huge size. selected using the Watson Visual Model [14].
5. Bit rate: Our watermarking framework allows a limited increase
in video bit rate by embedding watermark bits only in nonzero  Motivation: The paper is motivated by the following limitations
quantized coefficients. observed in the literature on video watermarking for HEVC
6. Complexity: We use compressed domain features to extract the encoded video. In [7], the scheme is highly sensitive to the quan-
public key, select blocks for embedding, and embed the water- tization parameter (QP) value since the watermark is embedded
mark. This decreases the computational overhead. in the LSB of DCT coefficients. In [8], a watermark is embedded
during the encoding phase of the HEVC/H.265 video. HEVC is
The rest of the paper is organized as follows: In the next section, built on the platform of H.264/AVC, therefore, the paper is also
we briefly discuss the literature on video watermarking for HEVC motivated by the following limitations observed in the existing
and H.264/AVC encoded video. Section 3 proposes the framework literature for watermarking in H.264/AVC encoded video. The
for compressed domain watermarking in the HEVC encoded video. methods proposed in [9,10,18] are fragile to re-encoding and
We present the experimental results to evaluate the performance of recompression attacks and the preprocessing phase is computa-
the proposed work in Section 4 and conclude the paper in Section 5. tionally expensive. Moreover, the work in [9] simply applied the
MPEG-style algorithm to H.264/AVC and did not tailor to the
characteristics of H.264/AVC. The schemes described in
2. Related work and motivation [15,16,22] are non-blind in nature, so original video is required
to extract the watermark at the decoder. In addition, complete
A very limited literature exists on watermarking in HEVC decompression and recompression of the video is required in
encoded video and thus, watermarking in HEVC encoded video is [15,16] for watermarking that restricted the scope of applica-
still in its rudimentary phase. Swati et al. has proposed a fragile tions. The watermarking technique presented in [15] requires
watermarking scheme, where the watermark is embedded in the a computationally expensive prediction process for watermark
least significant bit (LSB) of the nonzero quantized coefficients in embedding and has higher increase in video bit rate. The blind
the HEVC encoded video [7]. Ogawa et al. propose a watermarking algorithm in [3] has changed the nonzero coefficients to zero
method for HEVC/H.265 video streams that embeds information value for hiding the watermark that significantly degrades the
during the encoding phase of the video [8]. Many literatures exist quality. In [17], the error correction capability features of
on compressed domain watermarking for H.264/AVC encoded the Reed-Solomon codes are exploited to provide robustness.
video. In this paper, we also focus on review of literature on com- The method proposed in [21] is non-blind that embed water-
pressed domain watermarking in H.264 encoded video. Zhang and mark in I frames and P frames of H.264/AVC encoded video.
T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944 31

+
Partially

Rate of change in Modes


Selected Encoded Random 0.04
Watermark HEVC Key
Video
0.03 +
Selection of Public Private Compessed
different Key Key Watermarked 0.02
compressed
Public domain features HEVC +
Key Video 0.01 +
Extraction Palette
Generation + +
Block Selection +
0
+ +
Blind Extraction 0 1 2 3 4 5 6 7 8
Candidate Block
Private Number of NonZero Quantized Residuals (NNZ)
Key Selection
Generation HEVC (a) Change in Prediction Modes with NNZ
Embedding Extracted Uncompessed
Feature Selection Watermark Watermarked
HEVC Video

Rate of change in Modes


Compessed 0.4
Watermarked Blind
Embedding +
Video 0.3 +
+
(a) Embedding Framework (b) Extraction Framework 0.2 +
+
Fig. 1. Block diagram of the proposed watermarking framework for HEVC encoded +
video.
0.1 +
+ +
+ +
The compressed domain watermarking methods proposed in 0
0 2 4 6 8 10 12 14 16 18 20
[19,23] have the following limitations:
Motion Vector Field
(1) The watermark is embedded only in inter-predicted 4  4
blocks of P frames in H.264/AVC encoded video, which limits
(b) Change in Prediction Modes with Motion Vector
the embedding capacity; (2) the robustness threshold is selected Fig. 2. Change in Prediction Modes with NNZ and Motion Vector. The change in
randomly and not optimized; (3) the method is fragile to syn- luminance intra prediction modes decreases with the increase in the value of NNZ
chronization error; and (4) the size of the location map is huge, in a block. Similarly, the change in luminance intra prediction modes is higher for
no motion (static) and high motion blocks.
which adds transmission overhead. In [20], the quality of the
watermarked video is poor due to change in motion vectors of
the blocks. therefore, if any attacker wants to manipulate I frames, the percep-
In summary, there exist only few contributions in the literature tual quality of the video degrades highly [3,17,15]. Next, we select
on compressed domain watermarking for HEVC encoded video. appropriate luminance component of 4  4 intra predicted blocks
Considering the limitations of HEVC and H.264/AVC watermarking in I frames for invisibly embedding a bipolar watermark sequence
in the literature, we propose a robust watermarking framework 1; 1. If a smooth region of a frame is selected as embedding
with a blind extraction process for HEVC encoded video. The region for watermark embedding, the framework may cause visible
increase in video bit rate is also minimized. The compressed artifacts [3]. An efficient algorithm for selecting a suitable embed-
domain features are only explored for selecting embedding region, ding region of a frame uses the video content to get appropriate
embedding watermark, and extracting the public key. The frame- features, which would provide a bound on the degradation of per-
work is made secured by use of a random key for selecting the ceptual quality and increase in video bit rate. However, these fea-
blocks for watermark embedding. tures need to be robust enough to avoid the possibility of changes
after watermark insertion or re-encoding [3]. In other words, the
appropriate features would be robust to prevent desynchroniza-
3. Proposed framework tion during watermark extraction in the HEVC encoded video.

In this section, we propose a robust watermarking framework Synchronization error in intra coded blocks in HEVC encoded video:
for HEVC encoded video. The proposed framework first selects can- The increase in desynchronization in the watermark detection
didate blocks in I frames, known as embedding region, for embed- process due to embedding distortion in compressed domain
ding a watermark in HEVC encoded video. Next, we have watermarking is known as synchronization (re-encoding) error.
presented the watermark embedding and extraction algorithms. The intra prediction modes have block sizes of 32  32 down to
We also illustrate the extraction of the public key and the estima- 4  4 in HEVC encoded video. There are multiple prediction
tion of embedding thresholds. The block diagram of the proposed modes in I frames so it is very hard to track the change in pre-
framework is shown in Fig. 1. diction modes.

3.1. Embedding region We consider the following criteria for selecting the 4  4 intra
predicted blocks in I frames of HEVC encoded video.
The efficiency of a watermarking framework depends on the
selection of an embedding region. For selecting the embedding  Smooth regions of a frame: In I frame of the HEVC encoded
region, we first select appropriate I frames where watermark can video, a block is said to be a SMOOTH mode block, if it has
be embedded. We work on I frames since such frames convey a 32  32, 32  16, 16  32, or 16  16 prediction mode. The
wide range of information compared to P and B frames and blocks which have such prediction modes are not used for
32 T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944

watermark embedding. These blocks are used for smooth such frames are intra coded. The extraction of pseudo motion
regions of a frame and watermark embedding in a smooth vectors of I frames based on motion vectors at the same location
region may cause visible artifacts. In addition, due to embed- in the nearest P frame of the previous group of pictures (GOP) is
ding distortion, the blocks in the smooth region have a tendency described in [24] as follows:
to change into different mode blocks. This increases desynchro-
nization in the watermark extraction process at the decoder.
Therefore, our framework avoids the SMOOTH mode blocks
for watermark embedding. 1. Divide P frames into non-overlapping blocks.
 Number of nonzero quantized coefficients: The number of 2. Calculate motion vectors for all blocks in P frames of previ-
nonzero quantized coefficients (NNZ) plays an important role ous GOP.
in selecting blocks for embedding that can minimize synchro- 3. Smooth all motion vectors by using a 3  3 median filter.
nization error [3]. Part (a) of Fig. 2 shows the relationship 4. Motion vectors of intra-coded blocks have zero values in P
between the rate of change in prediction modes and NNZ in I frames, so estimate motion vector for intra-coded blocks
frames. The result presented in the figure is with 80% confi- from neighboring blocks in the respective largest coding
dence level though the error bars are not visible in the plots. unit (LCU).
Fig. 2(a) is based on the nature of CIF and QCIF videos. The figure 5. Assign pseudo motion vector to each block in I frames by
illustrates that the synchronization error decreases as the value interpolating motion vectors at the same location in the
of the NNZ is increased from 0 to 8. It can be observed from nearest P frame.
the figure that the blocks with higher values of NNZ may be
selected for watermark embedding. With the increase in the
value of NNZ, the number of blocks available for watermark Procedure 1. Block Selection
embedding decreases. However, the exact number of blocks

for embedding depends on the spatial characteristics of the Finally, the pseudo motion vector for every block in the I frame
video [3]. An appropriate threshold, denoted by N t , would be is obtained. If the value of pseudo motion vector of a block is
therefore selected based on the spatial characteristics of each greater than zero, then the block can be selected for watermarking.
sequence. For example, Fig. 2(a) shows that a block having Part (b) of Fig. 2 shows the relationship between the rate of change
NNZ greater than 5 has less synchronization error, i.e., N t 5. in prediction modes and pseudo motion vectors. It illustrates that
 Motion information: Embedding uncorrelated watermarks in the embedding in low motion regions provides an acceptable
successive frames usually results in annoying temporal flicker robustness against synchronization error due to embedding. An
and visual artifacts. Motion information is necessary for avoid- appropriate threshold, denoted by Mt , would therefore be selected
ing such temporal flicker. The degradation of quality of water- for embedding the watermark, where M t > 0. All blocks are
marked video due to any perturbation in a low motion region selected for watermark embedding, where the value of nonzero
is less than that of a higher motion region or static region (no pseudo motion vector is less than Mt . The value of threshold M t
motion). The value of motion vectors in I frames are zero since depends on the relationship between the rate of change in
T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944 33

prediction modes and pseudo motion vector. In this paper, we con- 3.2. Key extraction
sider the value of M t as 20 based on the Fig. 2(b). The result pre-
sented in the figure is with 80% confidence level though the error In each Iframe, candidate blocks are selected for embedding
bars are not visible in the plots. Fig. 2(b) is based on the nature using a random key from the blocks that already selected in Proce-
of CIF and QCIF videos. dure 1. The candidate blocks are those blocks where the watermark
is actually embedded. The selection of candidate blocks using a
Procedure 2. Function MagRound pseudo random key enhances the security of the proposed frame-
work. Different frames in a video will have different keys, which
are generated randomly using a pseudo random number generator
to avoid self-collusion attack [25]. The security of the algorithm is
imposed by randomly selecting candidate blocks using Procedure 3
from the set of blocks selected in Procedure 1.
If the watermark is embedded in the same location in correlated
I frames then the embedding algorithm becomes vulnerable to the
self-collusion attack [25]. Thus, a very long sequence is required.
Transmitting a long sequence, however, would make the algorithm
impractical. Such problem can be solved by generating the
sequence from a combination of a public key, which is extracted
from some features of the LCU, i.e., a block of size 64  64 and a
private key that is possessed by the copyright owner. The public
key is extracted from each LCU and pass to a cryptographic system
with the private key.
The compressed domain features available from HEVC codec
information are used for generating the public key without any fur-
ther decoding. The public key would be extracted from some fea-
tures of the LCU that cannot be changed by the attacker without
degrading the perceptual quality of the video. To make the
 Quality: The proposed procedure for embedding region selec- extracted public key robust, some features of the LCU, which are
tion incorporates a visual threshold, denoted by V t , for main- sensitive to human eyes, can be used. One such feature is DC coef-
taining acceptable perceptual quality of the watermarked ficients of the LCU. If DC coefficients themselves are used for public
video. The coefficients obtained from DCT transform are divided key extraction, then the attacker could change DC coefficients by
into DC coefficient and AC coefficients in a 4  4 block. We work the same amount, which will make watermark detection impossi-
on AC coefficients since such coefficients are less sensitive to ble. In I frames, changing a nonzero DC value to zero or zero DC
embedding error compared to DC coefficients. Let AC 1 and AC 2 value to nonzero will degrade the quality of the watermarked
are first two nonzero AC coefficients in a DCT transformed video and even create blocking artifacts [3]. Another feature is
block. If the absolute difference between AC 1 and AC 2 of a block the chrominance prediction modes of the LCU.
is greater than the threshold V t , then that block may be used for A 256 bits long public key extracted from each LCU is based on
watermark embedding. In Section 3.5, we estimate the value of DC components and chrominance prediction (chroma) modes in
visual threshold V t for HEVC encoded videos. the LCU as shown in Fig. 3. The first 64 bits give the DC coefficient
of every 4  4 block in the LCU. The last 192 bits show chroma
The complete process of selecting 4  4 intra predicted blocks modes of 4  4 blocks in each LCU. Each chroma mode is repre-
that are suitable for watermark embedding is given by Procedure sented by 3 bits.
1, where the function MagRound() is described in Procedure 2. The private key is estimated from the candidate block locations
and the public key. The security of the watermarking systems are
Procedure 3. Candidate Block Selection generally not made as complex as cryptographic systems. In this
34 T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944

work, public key and private key are encrypted using modulo 2 value of watermark bit is 1 and the magnitude value of AC 1 is less
arithmetic. The encrypted keys are compressed losslessly using than or equal to the magnitude value of AC 2 , then the value of mod-
run length coding. The public key is available to all clients, whereas ified coefficients are denoted by AC 01 AC 1 D signD  Rt and
the private is sent to the authorized client through secure channel. AC 02 AC 2  D  signD  Rt , where the value of D is
The complete process of key extraction is elaborated in Procedure 4. MagRoundAC 1 ; AC 2 . Similarly, if the value of watermark bit is 1
The number of candidate blocks selected in Procedure 3 and the and the magnitude value of AC 1 is greater than the magnitude value
number of blocks selected in Procedure 1 are denoted by m and n, of AC 2 , then the value of modified coefficients are denoted by AC 01
respectively. A random key is used to select m blocks out of n
AC 1  D  signD  Rt and AC 02 AC 2 D signD  Rt . If the value
blocks, i.e., candidate blocks. In the proposed framework, the cryp-
of watermark bit is 1 and the magnitude value of AC 1 is greater
tographic space for security is n Cm . As the value of m increases, the
than AC 2 or value of watermark bit is 1 and the magnitude value
proposed watermarking framework will be more robust.
of AC 1 is less than or equal to AC 2 , then no changes will be made.
If the value of Rt is large, the absolute difference between AC 1
Procedure 4. Key Extraction
and AC 2 will be large, which increases the robustness, but the

3.3. Watermark embedding perceptual quality of the watermarked video will degrade signifi-
cantly. In this section, the watermark embedding algorithm pro-
A bipolar watermark sequence 1; 1 is embedded invisibly in posed in [19,23] is modified and the formal algorithm of
candidate blocks in I frames of HEVC encoded video using com- proposed watermark embedding in HEVC encoded video in com-
pressed domain features. In the proposed watermarking frame- pressed domain is facilitated by Procedure 5. In Section 3.5, we
work, we embed a watermark bit in a candidate block by estimate the robustness threshold Rt for different video sequences.
changing the value of first two nonzero AC coefficients, i.e. AC 1
and AC 2 . Procedure 5. Watermark Embedding

To increase robustness of our watermarking framework, we use a If watermark is not embedded in appropriate coefficients, then
robustness threshold, denoted by Rt , where Rt > 0. The proposed embedding algorithm may not robust and the watermarked video
watermark embedding procedure is described as follows: If the may have visual artifacts. Therefore, the selection of appropriate
T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944 35

Table 1
Robustness threshold and visual threshold.

Video Rt D Vt H.264/AVC HEVC


PSNR eD PSNR eD
Coastguard 1 15 5 29.73 0.5 35.03 0.6
610 10 29.68 0.7 34.24 0.8
1115 15 28.92 1 34.02 1
3 15 5 29.71 0.8 34.69 0.7
610 10 29.04 1 33.88 0.9
1115 15 28.85 0.7 33.21 1
5 15 5 30.01 1 34.83 0.8
610 10 29.59 0.8 34.02 1
1115 15 29.49 0.7 32.57 0.7
Garden 1 15 5 28.53 0.7 35.44 0.7
610 10 27.45 1 34.87 0.9
1115 15 26.67 0.8 34.11 1
3 15 5 28.49 0.7 34.24 0.5
610 10 28.3 0.9 33.61 0.7
1115 15 28.28 1 33.14 1
5 15 5 28.7 1 33.85 0.8
610 10 28.3 0.8 32.98 1
1115 15 28.21 0.7 31.94 0.7
Husky 1 15 5 29.71 0.6 36.02 0.6
610 10 29.67 0.8 35.74 0.8
1115 15 28.22 1 35.04 1
3 15 5 29.69 0.8 35.64 0.7
610 10 29.04 1 34.61 1
1115 15 28.08 0.6 33.86 0.8
5 15 5 29.63 0.7 35.08 0.8
610 10 28.19 1 34.24 1
1115 15 27.44 0.9 33.37 0.9
Soccer 1 15 5 29.8 0.5 35.29 0.8
610 10 29.1 0.7 33.98 1
1115 15 28.92 1 32.89 0.7
3 15 5 29.09 0.7 34.91 0.8
610 10 28.78 0.9 33.79 1
1115 15 28.11 1 32.94 0.6
5 15 5 29.03 0.8 34.43 0.7
610 10 28.12 1 33.86 1
1115 15 27.94 0.7 32.42 0.6

coefficients or features for watermark embedding is very crucial. In


a candidate block, one or more compressed domain features, such Zero or Nonzero DC Value
as NNZ, sign and value of the quantized coefficients, and sign and
value of motion vectors, may be used for watermark embedding. 0 1 ... 63 64 65 ... 255
We use the following criteria for selection of appropriate features:
Chroma Prediction Mode
 Magnitude of low frequency nonzero quantized AC coeffi-
cients: Low frequency coefficients are more stable to synchro- Fig. 3. Extraction of public key from DC coefficients and chroma prediction modes
nization error [15]. Such coefficients can be changed to embed of each LCU.

watermark bits in a block [15,19]. It can maintain the degrada-


tion of quality and increase in video bit rate and synchroniza-
tion error.
 Sign of AC coefficients: In [9,10,18], the sign of AC coefficients
 NNZ value: In [3], the number of nonzero quantized AC coeffi-
is changed to embed watermark bits, which degraded the qual-
cients (NNZ) is decreased by changing the nonzero coefficients
ity of the watermarked video highly. Therefore, the sign of the
to zero value to embed a watermark bit, which increases syn-
coefficients is not changed to embed watermark.
chronization error. A block, which is actually predicted from
 High frequency coefficients: High frequency AC coefficients
such a watermarked block, may have a better match with a
are mostly reset to zero due to synchronization error or by
neighbor of that watermarked block at the decoder. Moreover,
the use of filters [26].
changing nonzero coefficients to zero degrades quality signifi-
cantly. Similarly, changing zero coefficients to nonzero will also
In summary, only low frequency nonzero quantized AC coeffi-
increase desynchronization at the decoder. Furthermore, chang-
cients in 4  4 blocks of I frames are suitable for watermark
ing zero coefficients to nonzero will result in significant
embedding. To reduce complexity, the first two nonzero quantized
increase in video bit rate. Therefore, the value of NNZ of a block
AC coefficients, namely AC 1 and AC 2 , in zigzag sequence order in a
is not changed to embed watermark.
4  4 block are changed for embedding a watermark bit. Water-
 DC coefficients: DC coefficients are not used for watermark
mark embedding in such coefficients will not create significant
embedding because the quality of the watermarked video will
visual artifacts. The values of such coefficients do not change much
degrade significantly and may even cause blocking artifacts
on desynchronization due watermarking noise.
[16].
36 T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944

3.4. Watermark extraction portional to the visual threshold V t . If the value of V t is small, then
the embedding capacity is reduced.
The watermark extraction is performed at the decoder after Based on the desired quality, capacity, and robustness of the
entropy decoding. The public key is scrambled using the private watermarking framework, a percentage of 4  4 blocks, namely k,
key to generate the resultant sequence (palette) which specifies containing a high value of D, is selected for embedding. In this case,
the candidate block locations used for embedding. A palette is a by choosing a higher value for k, the embedding capacity and
binary file sequence that contains the location formation of candi- robustness will increase but the quality will decrease. For a lower
date block [3]. It is used in the decoder for watermark extraction. In value of k, the embedding capacity and robustness will decrease
the proposed framework, a palette is generated from the combina- while the quality will increase. To obtain a target value of s, the
tion of a private key and a public key at the decoder. In most liter- endurance function of D distribution, denoted by eD , is utilized as
ature [18,15,3,17], the palette is sent to the decoder through follows:
secured channels. In the proposed work, only a compressed private X
key is sent securely to the authorized client for extraction of water- eD s PD < s pD < k; 1
mark bits. In absence of palette at the decoder, the candidate D<s

blocks will be selected using Procedure 1 and the random key,


which is used in Procedure 2. However, in absence of palette, the where the endurance function depends on the probability distribu-
robustness of the watermarking framework decreases. Due to syn- tion function. Based on Eq. (1), for the same k, different values of s
chronization error the location of the blocks selected at the deco- will be achieved depending on the spatial activity of the sequences.
der will not be exactly same as that of the locations selected at The achieved value for s is utilized in order to establish V t and Rt .
the encoder. The estimated values of V t and Rt are used to select suitable 4  4
In an I frame, the first two nonzero coefficients of a candidate blocks for embedding. Table 1 shows the values of V t and Rt for dif-
block of the watermarked video are denoted by AC 01 and AC 02 . If ferent video sequences. We have changed the value of D; Rt , and V t ,
the magnitude value of AC 01 is greater than AC 02 , the extracted to get the value of eD . In each video sequence, for each value of Rt ,
the value of D and V t are chosen, where the value of eD is maximum.
watermark bit will be 1, otherwise the value of watermark bit
is 1. The formal procedure for watermark extraction is facilitated The value of average PSNR over payload 2 f100; 200; 300g is esti-
mated for both HEVC and H.264/AVC videos. The value of D is taken
by Procedure 6.
in the range of 15. Similarly, the value of Rt 2 f1; 3; 5g and
V t 2 f5; 10; 15g. The results in the table are shown for four different
Procedure 6. Extract and Authenticate
CIF and QCIF video sequences, such as Coastguard, Garden, Husky,

3.5. Visual threshold and robustness threshold and Soccer. As the resolution of the compressed video change, the
spatial characteristics as well as temporal characteristics per 4  4
The visual threshold (denoted by V t in Section 3.1) is used to block changes. As a result, the value of number of nonzero quan-
control the degradation of perceptual quality of the watermarked tized coefficients and motion vector of a block may change. There-
video. If the value of V t is small, then degradation of quality will fore, the value of V t and Rt will also change for High Definition (HD)
be minimized, but the number of blocks selected for embedding videos. Similarly, if the configuration parameters of HECV or H.264/
will be reduced. Furthermore, the quality of the watermarked AVC codec changes, the spatial characteristics as well as temporal
video also partially depends on robustness threshold (denoted by characteristics per block will also change and as a result the value
Rt in Section 3.3). If the value of Rt is small, then robustness will of V t and Rt will change. In Table 1, the corresponding values of
decrease, but the degradation of quality will be reduced. With PSNR are highlighted in bold whenever the value of eD is 1.
the increase in the value of Rt , the absolute value of the modified
coefficient (i.e., AC 01 and AC 02 ) will be high. Therefore, higher value 4. Experimental results
of Rt increases the robustness of the proposed framework, but
may significantly degrade the quality. In other words, if the value The proposed watermark framework is implemented using
of D is small, then higher value of Rt can be used. If the value of H.265/HEVC reference software HM9.0 [27]. The advanced tools
D in a block is large, then smaller value of Rt may be used. In such in anchor HM9.0 have already been thoroughly studied in JCT-
a scenario, the robustness of the proposed framework will be high VC. Simulations were performed under common conditions set
and degradation of perceptual quality will also reduced. However, by the JCT-VC without any optimization. The All-intra Main config-
the maximum value of D can be V t . Higher value of Rt , therefore, uration is used in the experiments, where the RDQ and RDQTS are
implies smaller value of V t . Intuitively, embedding capacity is pro- disabled. Common conditions are intended to reflect the typical
T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944 37

Fig. 4. Embedding capacity using Procedure 1 for (a) Husky (b) Soccer (c) Flower and (d) Mobile videos (HEVC and H.264/AVC).

Table 2 bit-stream and is the configuration used under which coding effi-
PSNR of unwatermarked encoded video sequences for HEVC and H.264/AVC of QCIF, ciency of proposals are evaluated. We select the configuration of
CIF, and HD format. random-access-main with closed-GOP. The value of intra refresh
Sequence H.264/AVC HEVC type is set to 2.
PSNR (dB) PSNR (dB) The video sequences in common interchange format (CIF) and
Flower 28.55 34.93
quarter CIF (QCIF) [28,29] at payload = {100, 150, 200, 250, 300}
Ice 27.63 34.67 on an average of 100 frames are coded using a dyadic high-delay
Coastguard 28.49 34.58 hierarchical prediction structure with a GOP having intra frame
Garden 27.02 33.54 period value as 8 and 4 active reference pictures, where payload
Paris 26.35 35.71
is the amount of watermark bits actually embedded in a video
Trevor 28.21 33.89
Husky 28.38 33.27 sequence. The input videos in our experiments have different prop-
Football 28.91 35.75 erties, such as highly textured or less textured, low motion or high
Soccer 28.31 34.56 motion, with or without change in the shot, and uniform or noisy
Mobile 28.41 35.01
background. The quantization parameter (QP) is set to 28 [22].
Old town cross 54.09 64.91
Park joy 53.11 61.11
The values of V t and Rt are determined by using Eq. (1). The values
Crowd run 56.92 68.28 of M t and N t are determined by using Fig. 2. The proposed water-
Ducks take off 56.32 65.82 marking framework is also implemented using H.264/AVC refer-
In to tree 57.01 68.61 ence software JM17.2 and compared with the existing literature
38 T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944

Table 3
Results for PSNR, VQM, and BIR using the proposed watermarking framework.

Sequence Payload H.264/AVC HEVC


2 3
PSNR (dB) VQM  10 BIR  10 PSNR (dB) VQM  102 BIR  103
Flower 100 28.09 2.99 1.24 34.45 1.84 0.76
150 28.08 3.39 1.92 33.96 2.15 1.44
200 27.99 3.63 1.49 33.45 2.38 1.01
250 27.92 3.78 1.87 33.32 2.93 1.39
300 27.88 5.06 1.42 32.69 3.31 0.94
Ice 100 27.13 3.58 1.58 34.07 2.03 1.1
150 27.12 3.62 1.66 34 2.47 1.18
200 27.08 4.79 1.27 33.36 3.28 0.79
250 27.06 5.53 2.69 33.15 3.44 2.21
300 27.01 4.03 1 31.67 3.68 0.52
Coastguard 100 28.06 3.27 1.97 34.01 2.28 1.49
150 28 3.36 1.44 33.83 2.37 0.96
200 27.95 3.42 1.06 32.86 2.51 0.58
250 27.83 3.69 1.38 32.02 2.86 0.9
300 27.81 4.07 1.6 31.85 3.04 1.12
Garden 100 26.86 5.18 1.23 33.27 2.52 0.75
150 26.83 5.27 2.19 33.02 2.73 1.71
200 26.82 5.33 1.99 32.63 3.51 1.51
250 26.74 5.38 2.09 31.55 3.99 1.61
300 26.34 5.31 1.27 30.94 4.23 0.79
Paris 100 26.06 4.32 1.51 35.37 2.48 1.03
150 25.95 4.3 1.07 34.08 2.63 0.59
200 25.74 4.28 1.88 34.76 2.72 1.4
250 25.62 4.08 1.89 33.27 3.52 1.41
300 25.34 3.87 1.89 32.72 3.9 1.41
Trevor 100 28.1 4.6 1.65 33.41 2.25 1.17
150 27.38 4.26 0.92 32.92 2.84 0.44
200 27.31 3.68 2.19 32.22 3.22 1.71
250 27.17 3.77 1.81 31.23 3.43 1.33
300 27 3.47 1.72 30.99 3.5 1.24
Husky 100 28.16 5.38 1.93 35.27 2.77 1.45
150 27.83 5.28 2.09 35.03 3.13 1.61
200 27.22 5.23 2.19 34.62 3.51 1.71
250 26.94 5.09 2.11 33.76 3.92 1.63
300 26.68 4.81 1.2 32.95 4.26 0.72
Football 100 28.56 4.23 1.49 35.37 2.48 1.01
150 28.24 4.13 1.37 34.08 2.03 0.89
200 28.05 3.98 2.03 34.76 2.46 1.55
250 27.92 3.91 1.96 33.52 2.8 1.48
300 27.81 3.87 1.89 32.65 3.4 1.41
Soccer 100 28 4.46 1.75 34.21 2.15 1.27
150 27.78 4.22 1.53 34.02 2.34 1.05
200 27.52 3.98 2.22 33.32 2.75 1.74
250 27.27 3.75 2.01 32.76 3.23 1.53
300 27.08 3.66 1.82 32.35 3.32 1.34
Mobile 100 28.01 3.16 1.68 34.91 1.41 1.2
150 27.68 3.32 1.55 34.72 2.55 1.07
200 27.32 3.91 2.23 33.75 2.99 1.75
250 27.07 4.05 2.11 32.96 3.13 1.63
300 26.78 4.16 1.97 32.55 3.36 1.49

[3,18,15,17], which are based on H.264/AVC and an appropriate In the HEVC encoded video, NNZ is comparatively less than
similar configuration of HEVC is used for coding video sequence H.264/AVC video in most of the frames. Therefore, most of the
using H.264. We kept the payload same for every sequence H.264/AVC encoded video sequences have higher embedding
encoded in HEVC and H.264. capacity than HEVC encoded video. Embedding capacity is inver-
sely proportional to the motion threshold M t . Furthermore, embed-
4.1. Embedding capacity ding capacity is proportional to the visual threshold. Therefore,
embedding capacity is a function of k, denoted by Fk. The embed-
The number of blocks suitable for watermark embedding is the ding capacity, denoted by C E , is given by
embedding capacity of the video. In this work, we select the blocks, C E Fk;
which are suitable for watermark embedding, based on the spatio-
temporal characteristics of a video. Thus, the embedding capacity where the value of k / Nt M
1
t
, k > eD s / V t , and eD is given in Eq.
of the proposed work depends on block selection criteria of Proce- (1). The embedding capacity of a video depends on the spatial and
dure 1. Total number of blocks suitable for embedding selected in a temporal characteristics of that video per 4  4 block. Therefore, if
GOP in different HEVC and H.264/AVC video sequences are most blocks of a video have higher texture and are in motion, then
depicted in Fig. 4. that video will have higher embedding capacity.
T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944 39

Fig. 5. Perceptual quality is measured using average PSNR and VQM as shown in part (a) and part(b) of figure, respectively. Part (c) of figure shows BIR to estimate the video
bit rate.

4.2. Visual quality, peak signal-to-noise ratio, and bit increase rate of 102 , which indicate that the watermarked videos have accept-
able quality. Video quality is a characteristic of a video passed
In this section, we measure the perceptual quality of the water- through a video transmission/processing system, a formal or infor-
marked video using Peak Signal-to-Noise Ratio (PSNR) and Visual mal measure of perceived video degradation (typically, compared
Quality Metric (VQM) at payload = {100, 150, 200, 250, 300}. We to the original video). The application VQM creates an objective
also measure the increase in video bit rate using the metric Bit comparison of video codecs and measures the perceptual effects
Increase Rate (BIR). Bit Increase Rate (BIR) is defined as the percent- of video impairments, including blurring, jerky motion, global
age of bit rate increase per embedded bit [19], i.e., noise, block distortion and color distortion, and combines them
 into a single metric. It shows a high correlation with subjective
number of bits in watermarked video
BIR video quality assessment. Higher value implies higher degradation
payload  number of bits in original video
 of quality. When the payload is very less, the perceptual quality of
number of bits in original video the watermarked video is almost same as the unwatermarked one.
  100; 2
payload  number of bits in original video As the payload increases, the amount of degradation of quality also
where payload is the amount of watermark bits actually embedded increases. However, the degradation is quality also depends on the
in a video sequence. embedding region, where watermark is embedded. PSNR also con-
PSNR, VQM, and BIR of the proposed framework for HEVC and firms that the proposed embedding scheme does not sacrifice the
H.264/AVC encoded video are shown in Table 3. The values of PSNR quality very much. The degradation of the quality in terms of PSNR
of unwatermarked encoded video sequences for HEVC and H.264/ due to watermark embedding is about or less than 1 dB for all
AVC are shown in Table 2. VQM results in Table 3 are in the order videos. This confirms that the degradation will not be perceived
40 T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944

efficiency. The perceptual quality of the HEVC encoded water-


marked video is better than that of the H.264/AVC encoded video
for same payload. Similarly, increase in video bit rate in HEVC
encoded video is less than that of the H.264/AVC encoded water-
marked video.
The perceptual quality of the watermarked video is compared
with recent existing schemes based on the Peak Signal-to-Noise
Ratio (PSNR) and Visual Quality Metric (VQM). Increase in video
bit rate is also calculated compared with existing schemes. Fig. 5
(a) and (b) shows the average PSNR and average VQM in the case
of the proposed framework and the existing work [3,18,15,17].
The results are the average over five different payloads, i.e., {100,
150, 200, 250, 300}. The results for HEVC encoded video confirm
that the degradation of quality of the watermarked video will not
be perceived by human eyes. The result also shows that the pro-
posed framework for H.264/AVC encoded video outperforms other
schemes in perceptual quality for most of the videos. However, the
number of nonzero coefficients in Coastguard and Paris videos is
very less as the frames are highly correlated. Therefore, the syn-
chronization error is high in such blocks as shown in Fig. 2(a). This
Fig. 6. Change in quality and robustness of the watermarked video, which is
quantized with QP = 28, due to re-encoding by changing QP in a range from 20 to causes higher perceptual error. In Trevor video, the number of
36. change in shots is high. Therefore, different thresholds will be more
appropriate for different shots. Thus, watermark embedding in
by human eyes. In case of BIR, the increase in video rate is in the such frames by using a single threshold for visual quality increases
order of 103 which is nominal. HEVC shows better quality and less the degradation of the quality. The perceptual quality of the water-
increase in video bit rate for most of the videos due to its coding marked video for HEVC is much better than H.264/AVC encoded

Fig. 7. Change in quality and robustness by changing parameters for four different image processing attacks.
T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944 41

Fig. 8. Average robustness against re-encoding attacks. Part (a) shows the robustness against recompression error. Part (b) shows the robustness against decreases of QP from
28 to 26. Part (c) shows the robustness against increases QP from 28 to 30.

video. The proposed watermark embedding algorithm and embed- ent innocent or malicious attacks on the video. The robustness of a
ding region selection algorithm are designed in such a way that the watermarking method [19] is given by
degradation of quality of the watermarked video will be minimum.
Robustness 1  BER  100; 3
In case of BIR, Fig. 5(c) shows that the increase in the video bit rate
is negligible. It also compares the proposed framework with the where Bit Error Rate (BER) is defined as the frequency of bit errors
existing literature [3,18,15,17]. If the zero coefficients are changed when detecting a multi-bit watermark message [30], i.e.,
to nonzero or the motion vectors are changed to embed water-
mark, then the increase in video bit rate is significant. In the pro- number of error bits
BER : 4
posed watermarking framework, only two nonzero coefficients in total number of bits sent
a block are perturbed to embed a watermark bit. Neither the value BER is used for evaluating the robustness of the proposed work
of NNZ, motion vectors or zero coefficients of a block is changed. In against different attacks and compared with the existing water-
HEVC encoded videos, increase in the video bit rate is reduced marking schemes [18,15,3,17].
since HEVC provides appropriately 4050% reduction in the bit rate The information about the embedding locations is saved as a
at the same visual quality compared to H.264. palette during the embedding process [3,18,15,17]. This palette is
provided to the decoder for the extraction process. In the absence
of palette, if a watermarked location is not detected or a non-
4.3. Robustness against attacks watermarked position is selected incorrectly, the synchronization
in the watermark sequence will be lost. This will decrease the
In this section, we measure robustness of the proposed water- robustness of the a watermarking scheme. Palettes are used to
marking framework against different attacks. Robustness can be know about the embedding locations during the watermark
defined as how efficiently a watermarking system can resist differ- extraction process for both the proposed framework and the
42 T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944

Table 4
Robustness against four different kinds of attacks for different video sequences.

Attacks Parameter Sequence Schemes


H.264/AVC HEVC [18] [15] [3] [17]
Salt and pepper Salt and pepper noise density = 0.001 Flower 83 88 65 67 81 75
Ice 82 86 67 66 77 72
Coastguard 76 80 61 63 75 72
Garden 82 86 65 68 77 78
Paris 83 87 69 69 82 72
Trevor 80 86 66 68 87 74
Husky 81 86 67 69 85 75
Football 82 86 67 70 76 76
Soccer 80 85 68 68 81 74
Mobile 79 84 69 67 80 73
Gaussian noise Gaussian noise density = 0.001 Flower 76 80 66 63 77 77
Ice 78 82 65 64 74 76
Coastguard 77 82 62 65 78 74
Garden 82 86 63 67 74 76
Paris 75 81 66 66 76 74
Trevor 77 81 61 64 81 73
Husky 76 83 65 65 81 74
Football 79 85 67 69 80 77
Soccer 82 84 61 67 78 75
Mobile 80 85 62 68 79 76
Circular averaging filter Circular averaging filter r = 0.06 Flower 73 81 61 61 79 74
Ice 78 83 63 63 75 70
Coastguard 78 84 61 64 79 72
Garden 81 86 66 66 78 73
Paris 80 85 63 67 79 75
Trevor 79 83 65 65 83 74
Husky 78 86 66 68 83 71
Football 80 87 68 67 82 70
Soccer 81 86 69 68 79 73
Mobile 82 88 68 67 80 74
Gaussian filter Gaussian filter r = 0.05 Flower 74 80 62 61 78 73
Ice 76 82 64 64 77 72
Coastguard 77 83 60 63 79 72
Garden 82 87 65 66 78 75
Paris 81 85 64 65 80 70
Trevor 80 84 66 66 82 74
Husky 75 87 67 69 81 71
Football 82 86 66 64 83 72
Soccer 80 84 68 65 80 74
Mobile 81 85 69 68 79 73

existing literature [3,18,15,17]. Thus, presence of palette increases and radius of gaussian filter and circular averaging filter are shown
robustness in watermark detection. in Fig. 7. HEVC encoded video mostly provides better quality and
In [22], it is shown that by increasing the quantization param- robustness against different noise additions with respect to
eter (QP), the compression efficiency increases but the quality of H.264 video. Furthermore, HEVC video has better quality and
the watermarked video decreases. To get acceptable quality with robustness when blurred lightly using averaging filter and poor
highly compressed video, the QP of 28 is chosen [22]. The robust- quality when blurred significantly in comparison to H.264/AVC
ness of the proposed framework are evaluated against re-encoding encoded video. Similarly, HEVC video has better quality and
attack [3], such as recompression error and changing the value of robustness when blurred lightly using a gaussian filter in compar-
quantization parameter. The proposed watermarking framework ison to H.264/AVC encoded video.
is also tested against different noise additions and filtering attacks, Next, Fig. 8 compares average robustness results for re-
such as salt and pepper noise, gaussian noise, circular averaging fil- encoding attacks, i.e., recompression error and changing QP from
ter, and gaussian filter. 28 to 26 and 28 to 30, respectively, with the existing literature
We first depicted the robustness results of the proposed frame- [3,18,15,17] providing palette to the decoder for extraction of
work by changing the value of QP in Fig. 6. Initially, the watermark watermark bits. In case of recompression error, neither the value
is embedded when the value of QP is 28. The watermarked video is of QP nor any other parameter is changed. The watermarked video
re-encoded by changing the value of QP from 28 to a range of 20 is re-encoded with similar parameters and changes in synchroniza-
36. We also check the change in quality using PSNR by changing tion in watermark bits are noted. HEVC encoded videos provides
the value of QP. It is clear from the result that the HEVC encode higher robustness than H.264/AVC encoded videos in case of re-
video sequences have better quality compare to H.264/AVC encode encoding attacks. The proposed framework for H.264/AVC outper-
video sequences when QP is changed. The results in the figure are forms in most of the videos. The shot change in Trevor video is fre-
the mean over all sequences which are mentioned in the Table 3. quent. A fixed threshold value for different shots does not give a
HEVC encoded videos are more robust to re-encoding attack. good result. However, our results for H.264 videos are comparable
Fig. 6 shows the results where the quality and robustness are plot- with the existing literature. We can achieve better results by
ted for watermarked video with respect to unwatermarked video changing the values of thresholds M t , N t , V t , and Rt when a change
for both H.264/AVC and HEVC. Similarly, changing the variance in the shot is encountered. It is clear from the figure that the pro-
for salt and pepper noise, standard deviation of gaussian noise, posed framework is highly robust against recompression errors.
T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944 43

Table 5
Average PSNR, average BIR, and average robustness against recompression error (re-encoding attack) for five different HEVC encoded HD videos with two different resolutions.

HD video 1080p Resolution 2160p Resolution


3
PSNR (dB) BIR  10 Robustness PSNR (dB) BIR  103 Robustness

Old town cross 53.36 0.35 91 64.26 0.23 93


Park joy 52.22 0.34 95 60.69 0.22 95
Crowd run 56.23 0.28 94 67.73 0.21 92
Ducks take off 55.86 0.32 93 65.21 0.19 91
In to tree 56.81 0.29 92 68.04 0.18 90

The proposed watermarking framework for both HEVC and H.264/ for PSNR, BIR and robustness against recompression error for HD
AVC encoded video outperforms the existing literature. videos are not compared with any existing literature.
Finally, Table 4 illustrates the average robustness of the pro-
posed watermarking framework against noise additions, such as,
salt and pepper noise and gaussian noise and blurring using circu- 5. Conclusion
lar averaging filter and gaussian filter, at the decoder for extraction
of the watermark and compared with the existing literature In this paper, we proposed a novel watermarking framework
[3,18,15,17]. A palette is used for the proposed framework as well with a blind extraction process that embeds a readable watermark
as existing literature. HEVC encoded video outperforms against dif- invisibly into the I frames of HEVC encoded video. A comprehen-
ferent noise additions and filtering attacks in comparison to H.264/ sive description of the proposed video watermarking framework,
AVC encoded video. It is observed from the result that the proposed including block selection, public key extraction, watermark
work for H.264/AVC encoded video shows better results than exist- embedding, and watermark extraction, is given. The experiments
ing literature for most of the videos. An appropriate selection of are conducted to demonstrate that the proposed watermarking
embedding region and changing low frequency nonzero coefficient framework is robust against image processing attacks and resistant
to embed watermark make the proposed framework robust. In against re-encoding attacks. The watermarking security is imposed
short, HEVC encoded video provides better robustness against by randomly selecting the candidate blocks. We conducted a series
image processing attacks using the proposed watermarking of experiments to prove its effectiveness with respect to perceptual
framework. quality and increase in video bit rate. The use of compressed
domain features in selecting blocks for watermark embedding,
4.4. Performance on High Definition (HD) videos the embedding and extraction algorithms, and the extraction of
public key reduces the computational complexity. The proposed
HEVC has adopted many advanced encoding tools. The size of framework is implemented on different video resolutions. In
high definition (HD) video is a massive. There is no way for its con- future, the proposed framework can be extended to provide a com-
venient transmission. Instead, the video is compressed to reduce mon platform to compressed domain watermarking for a video
the amount of data into a more manageable form. A high lossy encoded using different codecs, i.e., codec independent platform
compression technique will minimize the size of the video, but and may address more classes of attacks.
degrade the quality significantly. HEVC is an efficient compression
technique for compressing HD videos. It can significantly improve
the compression performance of HD videos and save half of the bit References
rate compared to the H.264/AVC for the same perceptual video
[1] Y. Tew, K. Wong, An overview of information hiding in H.264/AVC compressed
quality. Thus, the benefit of HEVC in the streaming space is that video, IEEE Trans. Circ. Syst. Video Technol. 24 (2) (2014) 305319.
it can relate the encoding of existing standard-definition (SD) [2] F. Hartung, B. Girod, Watermarking of uncompressed and compressed video,
and HD content better than that of H.264 by enabling cost savings Sig. Process. 66 (3) (1998) 283301.
[3] A. Mansouri, A. Aznaveh, F. Torkamani, F. Kurugollu, A low complexity video
and the ability to stream higher quality video to lower bit-rate
watermarking in H.264 compressed domain, IEEE Trans. Inf. Forens. Sec. 5 (4)
connections. (2010) 649657.
In this section, we implement the proposed watermarking [4] C. Lu, C. Hsu, Near-optimal watermark estimation and its countermeasure:
framework on HD videos. We evaluate the results for quality, antidisclosure watermark for multiple watermark embedding, IEEE Trans. Circ.
Syst. Video Technol. 17 (4) (2007) 454467.
increase in video bit rate using average PSNR and average BIR, [5] G. Sullivan, J. Ohm, W.-J. Han, T. Wiegand, Overview of the high efficiency
respectively using HEVC encoding for different HD videos [28] at video coding (HEVC) standard, IEEE Trans. Circ. Syst. Video Technol. 22 (12)
payload = {100, 150, 200} as shown in Table 5. We also evaluate (2012) 16491668.
[6] A. Boho, G. Van Wallendael, A. Dooms, J. De Cock, G. Braeckman, P. Schelkens,
the average robustness against recompression error (re-encoding B. Preneel, R. Van De Walle, End-to-end security for video distribution: the
attack) for HD videos using HEVC encoding and are shown in combination of encryption, watermarking, and video adaptation, IEEE Sig.
Table 5. The resolution of HD videos are 1080p and 2160p. The Process. Magaz. 30 (2) (2013) 97107.
[7] S. Swati, K. Hayat, Z. Shahid, A watermarking scheme for high efficiency video
results show that the degradation of the quality of the water- coding (HEVC), PLoS One 9 (08) (2014) e105613.
marked video and the increase in video bit rate are reduced. It is [8] K. Ogawa, G. Ohtake, Watermarking for HEVC/H.265 stream, in: Proceedings of
clear from the results that the proposed watermarking framework ICCE, 2015, pp. 102103.
[9] J. Zhang, A.T. Ho, Efficient video authentication for H.264/AVC, in: Proceedings
for HD videos is robust against recompression error. As the resolu- of ICICIC 3, 2006, pp. 4649.
tion of a compressed video changes, the number of nonzero quan- [10] J. Zhang, A. Ho, G. Qiu, P. Marziliano, Robust video watermarking of H.264/AVC,
tized coefficients and the value of motion vector of a block will also IEEE Trans. Circ. Syst. II: Exp. Briefs 54 (2) (2007) 205209.
[11] G. Qiu, P. Marziliano, A. Ho, D. He, Q. Sun, A hybrid watermarking scheme for
change. In Section 3, the values of M t and N t in Fig. 2 and the values
H.264/AVC video, in: Proceedings of ICPR 04, 2004, pp. 865869.
of V t and Rt in the proposed algorithm are estimated based on the [12] T. Kuo, Y. Lo, A hybrid scheme of robust and fragile watermarking for H.264/
nature of CIF and QCIF videos. The existing literature [16,31,22,19] AVC video, in: Proceedings of BMSB, 2010, pp. 16.
used for comparison also have used CIF and QCIF videos. The [13] M. Noorkami, R. Mersereau, Compressed-domain video watermarking for
H.264, in: Proceedings of ICIP 2, 2005, pp. 890893.
threshold values in the existing literature are based on CIF and [14] A.B. Watson, DCT quantization matrices visually optimized for individual
QCIF videos. Therefore, for the sake of a fair comparison, the results images 1913, 1993, pp. 202216.
44 T. Dutta, H.P. Gupta / J. Vis. Commun. Image R. 38 (2016) 2944

[15] M. Noorkami, R.M. Mersereau, A framework for robust watermarking of [22] J. Li, H. Liu, J. Huang, Y.Q. Shi, Reference Index-based H.264 video
H.264-encoded video with controllable detection performance, IEEE Trans Inf watermarking scheme, ACM Trans. Multim. Comput. Commun. Appl. 8 (2S)
Forens Sec 2 (1) (2007) 1423. (2012) 33:133:22.
[16] M. Noorkami, R. Mersereau, Digital video watermarking in P-frames with [23] T. Dutta, Robust Compressed Domain Video Watermarking for H.264/AVC, Ph.
controlled video bit-rate increase, IEEE Trans. Inf. Forens. Sec. 3 (3) (2008) D. Thesis, Indian Institute of Technology Guwahati, 2014.
441455. [24] T. Dutta, A. Sur, S.K. Nandi, MCRD: motion coherent region detection in H.264
[17] E. Esen, A. Alatan, Robust video data hiding using forbidden zone data hiding compressed video, in: Proceedings of ICME, 2013, pp. 16.
and selective embedding, IEEE Trans. Circ. Syst. Video Technol. 21 (8) (2011) [25] B. Furht, O. Marques, Handbook of Video Databases: Design and Applications,
11301138. first ed., CRC Press, 2003.
[18] D. Xu, R. Wang, J. Wang, A novel watermarking scheme for H.264/AVC [26] M. Noorkami, Secure and Robust Compressed-Domain Video Watermarking
video authentication, Sig. Process.: Image Commun. 26 (6) (2011) 267 for H.264, Ph.D. Thesis, Georgia Institute of Technology, 2007.
279. [27] (2012) HEVC Reference Software <http://hevc.kw.bbc.co.uk/trac/browser/
[19] T. Dutta, A. Sur, S. Nandi, A robust compressed domain video watermarking in jctvc-hm/tags>.
P-frames with controlled bit rate increase, in Proceedings of NCC 2013, 2013, [28] (2015) Xiph.org Video Test Media [derfs collection] <https://media.xiph.org/
pp. 15. video/derf/>.
[20] T. Stutz, F. Autrusseau, A. Uhl, Non-blind structure-preserving substitution [29] (2015) Database: Images and Video Clip <http://see.xidian.edu.cn/
watermarking of H.264/CAVLC inter-frames, IEEE Trans. Multim. 16 (5) (2014) vipsl/databaseVideo.html>.
13371349. [30] D. Salomon, G. Motta, D. Bryant, An Engineers guide to Automated Testing of
[21] P. Su, C. Wu, I. Chen, C. Wu, Y. Wu, A practical design of digital video High-Speed Interfaces, Artech House, 2010.
watermarking in H.264/AVC for content authentication, Sig. Process.: Image [31] G. Feng, G. Wu, Motion vector and mode selection based fragile video
Commun. 26 (89) (2011) 413426. watermarking algorithm, in: Proceedings of ASID, 2011, pp. 7376.

You might also like