A New Efficient Robustness Evaluation Approach for Video
Watermarking based on Crowdsourcing
Asma Kerbiche
1
, Saoussen Ben Jabra
1
, Ezzeddine Zagrouba
1
and Vincent Charvillat
2
1
Lab. RIADI - Team of Research SIIVA, Higher Institute of Computer Science, University Tunis El Manar, Ariana, Tunisia
2
Lab. IRIT - Team of Research VORTEX, ENSEEIHT - INP TOULOUSE, University of Toulouse, Toulouse, France
Keywords:
Watermarking, Crowdsourcing, Evaluation, Robustness, Attacks, Camcording, Attack’s Game.
Abstract:
Signature robustness is the most important criteria that must verify a watermarking approach. However, exist-
ing watermarking evaluation protocols always tested simple attacks like rotation, cropping, and compression
but did not consider many dangerous attacks such as camcording which is more and more used for videos. In
this paper, a new robustness evaluation approach for video watermarking is proposed. It is based on on-line at-
tack’s game using crowdsourcing technique. In fact, the proposed game is provided to different users who will
try to destruct an embedded signature by applying one or many combined attacks on a given marked video.
Switch the choice of the users, the most important attacks can be selected. In more, users must not destroy
the visual quality of the marked video to evaluate the tested watermarking approach. Experimental results
show that the proposed approach permits to evaluate efficiently the robustness of any video watermarking. In
addition, obtained results verify that camcording attack is very important in video watermarking evaluation
process.
1 INTRODUCTION
The fast networks development and the evolution of
new compression standards have facilitated copying,
transmission and distribution of digital data such as
text, image and video. Therefore, piracy and illegal
use of these data has become easier and can cause a
significant economic impact. Different techniques are
appeared like cryptography, steganography but they
cant protect efficiently digital contents. Watermark-
ing is then proposed to resolve these problems. It con-
sists to embed a signature into data and to try to detect
it after any manipulation done on marked data. Many
watermarking approaches are developed for different
types of media and especially for video where the pro-
tection researches attracted more and more scientific
community. In fact, an efficient video watermark-
ing must verify two main constraints: invisibility, it
means that original and marked video must be iden-
tical and robustness against different attacks. This
last one is evaluated by applying different attacks on
marked video and then trying to detect the embedded
signature. However, despite the evolution of water-
marking algorithms and hacking techniques, their ro-
bustness is always evaluated only against classic and
simple attacks. In fact, existing evaluation protocols
use the usual attacks such us geometric transforma-
tions, noise addition and filtering but, nowadays, all
watermarking approaches must resist to this type of
attacks. With new technologies development, other
types of attacks are appeared and observed in the real
world. These attacks must be considered both at de-
velopment of watermarking approach and at evalua-
tion step. The most important and malicious attack
is camcording which consists of recording movies in
theaters with a Smartphone or a camera that can be
followed by colorimetric transformations, a compres-
sion or deformation manipulation.
In this paper, a new robustness evaluation ap-
proach is proposed. It is based on crowdsourcing
technique which allows using the collective intelli-
gence of many users. In fact, this approach consists
of an attacks game where different users will try to
destroy a mark inserted into a given video by apply-
ing various combinations of available attacks, while
maintaining a good visibility of the video. This in-
teraction will allows evaluating the watermarking ap-
proach applied on the test video and selecting the
most interesting attacks.
The remainder of this paper is organized as fol-
lows: in the next section, a state of the art of the
existing evaluation protocols of video watermarking
Kerbiche, A., Jabra, S., Zagrouba, E. and Charvillat, V.
A New Efficient Robustness Evaluation Approach for Video Watermarking based on Crowdsourcing.
DOI: 10.5220/0005726201670173
In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 3: VISAPP, pages 169-175
ISBN: 978-989-758-175-5
Copyright
c
2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
169
algorithms and crowdsourcing technique will be pre-
sented. The proposed approach will be detailed in the
second section. The selected attacks will be descriped
in the third section. Experimental results and evalu-
ation will be provided in the fourth section. Finally,
conclusion and perspectives are drawn in Section 5.
2 STATE OF ART
The goal of the proposed work is to select the most
important attacks which can be applied on a marked
video to evaluate any watermarking approach. This
selection is based on crowdsourcing technique.
2.1 Existing Evaluation Protocols
Several protocols have been proposed to evaluate
video watermarking algorithms. However, these pro-
tocols generally use the same traditional attacks for
robustness evaluation. The European project Certi-
mark (Rollin, ) is proposed to realize generic tests to
evaluate image and video watermarking methods. In
the case of a video watermarking, this protocol tested
different attacks which are: compression, digital to
analog and analog to digital conversion (D / A and
A / D), lossy storage formats conversion, logos or
captions adding, geometric transformations like rota-
tion, translation, Cropping and scaling, multiple wa-
termarks, noise and collusion. Several other protocols
have been proposed for image watermarking algo-
rithms and can be applied to video watermarking such
as StirMark benchmark project (Petitcolas, 2000) and
”BOWS” project (Break Our Watermarking System)
(Bas and Furon, ) which is based on multiple con-
straints like capacity, invisibility, speed and robust-
ness. In spite of the efficiency of the existing pro-
tocols, they test only simple attacks where a classic
watermarking approach can resist and they didn’t in-
tegrate the most dangerous and real attacks which can
easily destruct the embedded signature.
2.2 Crowdsourcing Technique
Crowdsourcing technique has become a widely ap-
plied practice in the context of innovation and prob-
lem solving (Stanoevska-Slabeva, 2011), that’s why
researches related to this technique have become
a dynamic and vibrant research area. Howe (J.,
2006), classifies crowdsourcing in three main cate-
gories based on type of task outsourced to the crowd.
The first class presents the idea game which is es-
sentially just a massive call for ideas such as IBM
Jam. In fact, in 2006, IBM initiated a global idea
jam related to the question how to best use and ef-
ficiently commercialize existing technological devel-
opments in the company, the task of the crowd was to
brainstorm about potential new ways how technology
developed at IBM might be applied to enhance exist-
ing or develop new products. The second category
is crowdsourced problem solving where the problem
is broadcasted to a large undefined network of poten-
tial solvers. The goal is to create a complete solution
by integrating complementary contributions from the
crowd by defining of clear interfaces. Xie et al. (Xie
et al., 2005) propose new method to detect user inter-
est maps and extract user attention objects from the
image browsing log using crowdsourcing. A smart
image viewer was developed based on user interest
analysis and a second experiment was carried out to
study how users behave with such a viewer. This
approach is more efficient than image-analysis based
methods and can better represent users’ actual inter-
est. Based on the fact that the viewing experience
on the mobile devices can be improved by determin-
ing important and interesting regions within the video
(regions of interest, or ROIs) and displaying only the
ROIs to the viewer, Carlier et al (Carlier et al., 2010)
propose an alternative paradigm to infer ROIs from
a video by crowdsourcing from a large number of
users through their implicit viewing behavior using a
zoom and pan interface, and infer the ROIs from their
collective wisdom. A retargeted video, consisting of
relevant shots determined from historical users’ be-
havior, can be automatically generated and replayed
to subsequent users who would prefer a less interac-
tive viewing experience. A user study shows that this
automatically retargeted video is of comparable qual-
ity to one handcrafted by an expert user. Finally the
last category is prediction markets in which investors
from the crowd buy and sell futures related to some
expected outcome.
3 PROPOSED EVALUATION
APPROACH
The main originality of the proposed approach is to
use crowdsourcing technique for evaluation. In fact,
it consists of an attacks’ game which is based on
users’ actions when they try to destroy signature from
marked videos. This game makes users free to choice
one or a combination of attacks and to apply them on
videos. This allows detecting the most important ma-
nipulations which can be dangerous for a watermark-
ing algorithm. In more, thanks to using real users,
this permits to simulate the behavior of a pirate when
he wants to destroy a signature contrary to the use
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
170
of a robot where we cannot know the type of ma-
nipulations which can be applied in reality. The pro-
posed approach allows also applying different attacks
without damaging visual quality of the marked and
attacked video. In fact, a visual threshold is defined
to oblige the users to reserve the visual contents of
the video to be attacked. General architecture of the
proposed approach is presented in the figure 1.
Figure 1: General architecture of the proposed approach.
The second originality of the proposed approach
is to use camcording attack to evaluate the robust-
ness of a given watermarking approach. In fact, cam-
cording is an important and much applied attack for
video applications but it is not considered in other
evaluation protocols. This attack consists of film-
ing the video projected or displayed using a Smart-
phone or a camera and then broadcast it after hav-
ing applied some transformations to destroy the wa-
termark that has been inserted. Due to camcord-
ing attack, illegal copies of movies have become an
important concern of the film industry and technol-
ogy development in recent years. Some works were
then proposed, these last years, to guarantee robust-
ness against this attack such as the algorithm devel-
oped by the research and development unit of Ko-
dak (CHANDRAMOULI et al., 2001) that showed its
worth against the camcording by embedding a signa-
ture containing the movie theater where the screening
took place and the time and release date. In more,
for helping researchers to develop watermarking ap-
proaches robust to camcording attack, Philipp Sch-
aber et al. (Schaber et al., 2014) have developed a
tool that simulates a reacquisition of content with a
camcorder.
The proposed attack’s game consists of an inter-
face containing a fun challenge: ”Find the best attacks
that removes the watermark without radically degrad-
ing video. This interface is available to users and
it allows them to apply to the watermarked videos a
combination of attacks that we have identified as most
important and dangerous. The proposed interface is
shown in Figure 2 and is decomposed in three main
parts: the first one contains a preview of the water-
marked video, the second contains the list of selected
attacks and the last part shows a view of the attacked
watermarked video. Each user has three attempts to
destroy the signature. Firstly, he must choice the
type of test video which can be camcordered or not.
Then, he can apply a compression on chosen video
by choosing a compression rate. Finally, he can apply
other types of attacks. After every attack application,
the visual quality of attacked video will be calculated
and if it is lower than a defined threshold, the user
must choice other attacks.
4 CHOSEN ATTACKS
Based on investigations made with film experts, five
attacks are chosen to be tested in the proposed inter-
face. The first attack is camcording which is realized
with four different camera positions: in front of the
screen, then right, then left and finally down of the
screen as shown in Figure 3.
The second chosen attack is MPEG 4 compression
which is the most popular compression standard and
presents the better method for high definition broad-
casting. Due to the importance of this compression
system, any efficient watermarking algorithm must be
able to resist to this attack at least in the low com-
pression ratio. For this reason the second step in our
game of attacks is to allow users to choose to work
with the compressed or not compressed watermarked
video and to choose one of three compression rates
1000 kbit/s, 500 kbit/s and 200 kbit/s (Figure4).
The third attack is the deformation that users can
apply, in the case of camcording videos. It consists to
change videos reframing (Figure 5) by selecting area
coordinates they want to rectify.
The fourth attack is cropping that consists of ex-
tracting a region from the video’s frame by cutting
horizontally or vertically images from the video. This
attack can completely destroy the watermark. In our
proposed game, users can select the video area they
wish to preserve on condition to guarantee the good
visual quality of the cropped video (Figure 6).
A New Efficient Robustness Evaluation Approach for Video Watermarking based on Crowdsourcing
171
Figure 2: Attacks game interface.
Finally, the latest attack available in the proposed
game is changing color of the video by adding red,
green or blue color to the original video frames (Fig-
ure 7).
5 EXPERIMENTAL RESULTS
To test the proposed robustness protocol and to study
and analyze the choice of the different users, we
choose to evaluate and compare two existing video
watermarking schemes: the first one is based on
multi-frequency insertion in feature regions proposed
in our previous work (Kerbiche et al., 2012) and the
second is based on wavelet transform (Chan and Lyu,
2003). These two algorithms presented good robust-
ness against the usual video watermarking attacks
such as rotation, noise addition, frame deleting and
compression... These algorithms were applied to the
two colors video Stefan and Big Buck Bunny. We
Figure 3: Camcording attack: (a) camera in front of the
screen, (b) camera in the left of the screen, (c) camera down
of the screen, (d) camera in the left of the screen.
have exposed the proposed game to 50 users who have
3 attempts to destroy the mark by combining attacks
while keeping a good visibility of the video. In fact,
after each attempt a measure of the quality was calcu-
lated and compared to a threshold before the valida-
tion of the user’s attempt.
We choose to use the structural similarity (SSIM)
as a quality’s measure. It is a method for measur-
ing the similarity between two images. The SSIM in-
dex is a full reference metric and it is designed to im-
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
172
Figure 4: Compression MPEG 4 : (a) 200 kbit/s, (b) 500
kbit/s, (c) 1000 kbit/s.
Figure 5: Deformation of camcordered watermarked video:
(a) Selection of the coordinates of the region, (b) The se-
lected region, (c) Rectified region.
Figure 6: Cropping attack.
Figure 7: Colors changing of the watermarked video: (a)
Red adding, (b) Green adding, (c) Blue adding.
prove on traditional methods like peak signal-to-noise
ratio (PSNR) and mean squared error (MSE), which
have proven to be inconsistent with human eye per-
ception. The difference with other techniques such as
MSE or PSNR is that these approaches estimate per-
ceived errors, but SSIM considers image degradation
as perceived change in structural information. Struc-
tural information is the idea that the pixels have strong
inter-dependencies especially when they are spatially
close. These dependencies carry important informa-
tion about the structure of the objects in the visual
scene. The SSIM metric is calculated on various win-
dows of an image. The measure between two win-
dows x and y of common size N × N is:
SSIM(x,y) =
(2µ
x
µ
y
+ c
1
)(2σ
xy
+ c
2
)
(µ
2
x
+ µ
2
y
+ c
1
)(σ
2
x
+ σ
2
y
+ c
2
)
(1)
Where µ
x
presents the average of x; µ
y
the average of
y; σ
2
x
the variance of x; σ
2
y
the variance of y and σ
xy
the
covariance of x and y. The two variables c
1
=(k
1
L)
2
and c
2
=(k
2
L)
2
permit to stabilize the division with
weak denominator. Finally, L is the dynamic range
of the pixel-values (typically this is 2
]bits per pixel
-1)
and the values of k
1
and k
2
are respectively 0.01 and
0.03 by default.
This similarity study will be applied before each
validation of user. Indeed, if SSIM value is less than
0.4, user attempt will not be validated and the combi-
nation he applied will not be considered.
A New Efficient Robustness Evaluation Approach for Video Watermarking based on Crowdsourcing
173
5.1 Users Interaction
To identify the most used and important attacks we
have registered the different choices of each user. We
observed that users have attempted to test all the at-
tacks in order to see their impacts on the video, af-
ter that, most of them have chosen the camcordered
video (44 have chosen the Stefan camcordered video
and 48 have chosen the Big Buck Bunny camcordered
video). The users which have chosen the camcordered
video with camera in the left, in right and down of the
screen they have all applied the deformation to adjust
the crop of the video. For compression, most users
have chosen this attack with a 500 kbit/s compression
rate in order to avoid degradation of the video visual
quality. Figure 8 shows for each rate, the number of
users which have chosen it.
Figure 8: Number of choice for each rate of compression.
Finally, for the last two attacks, users have been
careful to not degrade the visibility of the video and
especially the visibility of the moving objects. Fig-
ure 9 shows for each attack the number of choice.
According to this curve, we can notice that the most
common attacks are camcording, compression, defor-
mation and cropping.
Figure 9: Number of choice for each attack.
5.2 Video Watermarking Algorithm
Robustness
The video watermarking algorithm based on multifre-
quency insertion in the region of interests (Kerbiche
et al., 2012), showed a high performance against all
the combinations of attacks. In fact, this algorithm
could always detect the presence of the mark in the
video after each test. Figure 10 shows the result of
some combinations of attacks applied on the water-
marked video.
Figure 10: Attacked video watermarked (Kerbiche et al.,
2012): (a) camcordered video with camera in front of the
screen + MPEG-4 Compression 200 kb/s + cropping, (b)
camcordered video with camera in the left of the screen
+ MPEG-4 Compression 200 kb/s + cropping + red color
adding +15.
For video watermarking algorithm based on
wavelet transform (Chan and Lyu, 2003), despite its
robustness against common attacks, 17 users have de-
stroy the mark inserted in the video while maintaining
good visibility. Table 1 shows some combinations ap-
plied by those users who have succeeded to destroy
the signature and SSIM values for each combination
where (1) is camcordered video with camera in front
of the screen + compression 500 kbit/s + Cropping,
(2) camcordered video with camera in the right of
the screen + deformation + compression 500 kbit/s +
Cropping, (3) camcordered video with camera in the
left of the screen + deformation + compression 500
kbit/s, (4) camcordered video with camera in the left
of the screen + deformation + compression 500 kbit/s
+ Cropping + changing colors, (5) camcordered video
with camera in the left of the screen + deformation
+ compression 500 kbit/s + Cropping and finally (6)
camcordered video with camera down of the screen
+ deformation + compression 500 kbit/s + Cropping.
Figure 10 shows the result of two combinations of at-
tacks applied on the watermarked video.
Table 1: SSIM values of the watermarked video which users
have destroy the mark.
(1) (2) (3) (4) (5) (6)
SSIM 0.57 0.51 0.53 0.437 0.431 0.47
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
174
Figure 11: Attacked video watermarked (Chan and Lyu,
2003): (a) camcordered video with camera in the left of the
screen + deformation + compression 500 kbit/s, (b) cam-
cordered video with camera down of the screen + deforma-
tion + compression 500 kbit/s + Cropping.
6 CONCLUSION
In this paper we proposed a new robustness evalua-
tion approach based on crowdsourcing in order to im-
prove the evaluation protocols of video watermarking
algorithms. This approach consists of a game which
allows users to apply different combinations of im-
portant and most used attacks as camcording, defor-
mation, adding color and MPEG-4 compression on
watermarked videos in order to destroy the embedded
mark. This game permits to evaluate existing video
watermarking algorithms and also to identify the most
important attacks based on the choices of different
users. In fact, experimental tests have shown that the
proposed evaluation approach is more efficient than
other existing methods. In fact, it allowed us to com-
pare two video watermarking algorithms which were
ranked as two robust methods but only one of them
have resisted to the combinations of attacks applied
in the proposed game. In addition, it has proved the
importance of the camcording attack which is very
dangerous for video watermarking algorithms and is
often not considered in existing evaluation protocols.
As a perspective for this work, we will complete it in
order to develop a new evaluation protocol of video
watermarking techniques. In fact, other attacks that
may present a risk can be added to the proposed game
and also visibility evaluation must be included. The
game will be beneficial to both sides. Indeed, it will
evaluate the watermarking algorithms; in addition, the
study of different users’ choices as well as the reac-
tion of the watermarking algorithm against their de-
struction attempts could be used to improve the wa-
termarking algorithms to resist against these attacks.
REFERENCES
Bas, P. and Furon, T. Project bows2, www.http://bows2.ec-
lille.fr/.
Carlier, A., Charvillata, V., Ooi, W., and Morin., R. G. G.
(2010). Crowdsourced automatic zoom and scroll for
video retargeting, in acm multimedia.
Chan, P. P.-W. and Lyu, M. R. (2003). A dwt-based digi-
tal video watermarking scheme with error correcting
code, in icics’03. volume 2836.
CHANDRAMOULI, R., MEMON, N., and RABBANI, M.
(2001). Invisible watermarking for digital cinema, in
digital watermarking.
J., H. (2006). The rise of crowdsourcing, wired magazine.
volume 14.
Kerbiche, A., Jabra, S. B., and Zagrouba, E. (2012). A
robust video watermarking based on image mosac-
ing and multi-frequential embedding, in ieee interna-
tional conference on intelligent computer communi-
cation and processing.
Petitcolas, F. A. P. (2000). Watermarking schemes evalua-
tion, in i.e.e.e. signal processing. volume 17.
Rollin, C. Certimark. www.certimark.org.
Schaber, P., Kopf, S., Wesch, C., and Effelsberg, W. (2014).
A camcorder copy simulation as watermarking bench-
mark for digital video, in acm multimedia systems
conference.
Stanoevska-Slabeva, K. (2011). Enabled innovation: Instru-
ments and methods of internet-based collaborative in-
novation, in the 1st berlin symposium on internet and
society.
Xie, X., Liu, H., Goumaz, S., and ying Ma., W. (2005).
Learning user interest for image browsing on small-
formfactor devices, in sigchi conference on human
factors in computing systems.
A New Efficient Robustness Evaluation Approach for Video Watermarking based on Crowdsourcing
175