A Multimedia Tracing Traitors Scheme Using Multi-level Hierarchical
Structure for Tardos Fingerprint Based Audio Watermarking
Faten Chaabane, Maha Charfeddine and Chokri Ben Amar
REGIM: REsearch Groups on Intelligent Machines, University of Sfax, National Engineering School of Sfax (ENIS),
BP 1173, Sfax, 3038, Tunisia
Keywords:
Tracing Traitors, Hierarchical, Tardos, Watermarking, Fingerprint, Computational.
Abstract:
This paper presents a novel approach in tracing traitors field. It proposes a multi-level hierarchical structure to
the used probabilistic fingerprinting code; the well known Tardos code. This proposed structure is performed
to address the problem of computational costs and time of Tardos code during its accusation step. The gener-
ated fingerprint is embedded in the extracted audio stream of the media by an audio watermarking technique
operating in the frequency domain. The watermarking technique represents an original choice compared to
the existing works in the literature. We assume that the strategy of collusion is known, we compare then the
performance of our tracing traitors framework against different types of attacks. We show in this paper how
the proposed hierarchy and the watermarking layer have a satisfying impact on the performance of our tracing
system.
1 INTRODUCTION
Several manipulations like copying, editing and dif-
fusing multimedia tracks through the internet and
Peer to Peer networks do not represent any challenge
even to simple users but constitute a dangerous phe-
nomenon for the software industry. The purpose of
researchers was to find mechanisms performing copy-
right protection. First works were proposed in water-
marking field which consists in embedding a specific
message in the digital content to protect it from fraud.
In the recent years, with the evolution of cloud and
networks, watermarking techniques can be easily cir-
cumvented by a group of experienced users who try
to cooperate together by applying more complex col-
lusion attacks in order to create and share illegally a
new copy with unknown fingerprint. This type of ma-
nipulations has led to the development of collusion re-
silient secure fingerprints. These codes, associated to
the watermarking technique, are able to trace traitors
who collude together (Charpentier et al., 2011).In
tracing traitors field, there are two major schemes:
static tracing scheme and dynamic one (Laarhoven,
2013).The choice of the tracing strategy depends on
the number of needed fingerprints in the scheme to
capture dishonest persons. In the VOD context, the
video supplier generates a single individual finger-
print specific to each authorized user and inserts it
in every sold release. When an illegal copy is dis-
tributed, he is able to discover the forgery, to trace at
least one colluder and so to delete this suspicious copy
to prevent any other redistribution. In the dynamic
scenario, such as the online streaming systems, the
media distributor has the ability to generate a novel
set of fingerprints when collusion is discovered and
thus disconnect accused persons from the system.
In our work, we are interested in the VOD context,
and in figure 1 below; we present the generic static
tracing scheme. We distinguish, in the first side, the
generation step where the fingerprint, based on the
Tardos code is constructed and diffused to n users
u
i{1..n}
. In the distributor side, a group of collud-
ers colluder
j{1..c}
mix their copies and participate to-
gether in the construction of a suspicious copy with
unknown fingerprint. Once the suspicious copy is de-
tected, the video supplier proceeds by extracting the
colluded fingerprint and analyzing it to retrieve dis-
honest users .
The paper is arranged as follows: the second sec-
tion consists in an overview of related works in trac-
ing field. In section 3, we detail our tracing algo-
rithm based on generating a multi-level hierarchical
fingerprint and its embedding in the stream sound of
a video. In section 4, the performance and the secu-
rity of our proposed tracing system is shown. Finally,
we give a conclusion and some future perspectives.
289
Chaabane F., Charfeddine M. and Ben Amar C..
A Multimedia Tracing Traitors Scheme Using Multi-level Hierarchical Structure for Tardos Fingerprint Based Audio Watermarking.
DOI: 10.5220/0005066602890296
In Proceedings of the 11th International Conference on Signal Processing and Multimedia Applications (SIGMAP-2014), pages 289-296
ISBN: 978-989-758-046-8
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Figure 1: The general static tracing traitors scheme.
2 RELATED WORKS
Unauthorized manipulations on digital media present
an important challenge. The tracing traitors tech-
niques, called also transactional watermarking tech-
niques were proposed by researchers as a suitable so-
lution to trace back illegal media redistribution. The
principle of these systems is to hide by a watermark-
ing technique a specific message in every sold release.
The choice of the watermark is not arbitrary; it is a
mathematical code, known as secure collusion finger-
printing code. Its structure has to resist to almost all
types of attacks and needs huge payload (Charpentier
et al., 2011) (Charpentier et al., 2009). Recent works
proposed in tracing field try to improve the tracing
performance and respond to several requirements tied
to the length of inserted code whenever the collusion
size increases . In our previous paper (Chaabane et al.,
2013), we have dressed a survey of existing finger-
printing techniques and we have tried to classify these
techniques according to the improvement they pro-
pose. In some works, like in (Hayashi et al., 2007),
the hierarchical structure was proposed in fingerprint
generation process, assigning users to a group is made
randomly without considering relationships inter or
intra groups which doesnt present an optimal choice
to ameliorate the Tardos performances. In (Ye et al.,
2013),authors analyze users relationships in a social
network to construct a multi-level hierarchical finger-
print for digital content diffusion through the internet
and Peer to Peer networks. The resulting fingerprint
of each user is the combination of BS code as an outer
the Tardos code as an inner code. However, the lim-
itation of this work is that it tests its proposal sys-
tem only against the majority vote attack. Some other
contributions (Hamida et al., 2011) have proven that
reducing the complexity of the decoding step have a
positive impact on the accusation process thus they
propose to use a hierarchical embedded fingerprint,
this hierarchy in (Hamida et al., 2011) is inspired
from (Wang et al., 2004) and is based on regroup-
ing users according to their social and geographic be-
longings under the assumption that users having the
same characteristics have more probability to cooper-
ate together than with others. In (Shahid et al., 2013),
the proposed tracing scheme embeds the well known
fingerprinting code, the Tardos code (Tardos, 2003)
(Laarhoven and de Weger, 2011), in video signals
in compressed domains; it used the H.264/AVC as a
compression standard and the spread spectrum as a
watermarking technique. The weakness of this work
is that the number of users used in experimentation
didnt exceed 100 whereas in VOD applications it can
reach many thousand.
Some works try to optimize the accusation functions
(Furon et al., 2009) of the Tardos code to ameliorate
its robustness to the worst collusion attacks. In (Des-
oubeaux et al., 2011), a tracing algorithm is proposed
where the accusation function of Tardos code is im-
proved to suit to a specific watermarking technique:
the zero bit watermarking technique, this combina-
tion provides short fingerprint for great number of
users but needs some adjustment in the tracing pro-
cess. According to related works, we have focused
on rising to the challenge of ameliorating robustness
results and accusation rates of improved Tardos code,
thus we try to decrease the complexityof Tardos com-
puting by proposing a multi-level hierarchical finger-
print and we embed it by an original robust water-
marking technique. Our proposed framework will be
explained below.
3 THE PROPOSED ALGORITHM
In our tracing traitors framework, as shown in figure
2, , we distinguish three main steps: the multi-level
hierarchical Tardos fingerprint generation, its embed-
ding in the extracted audio stream and the tracing step
which occurs in the distributor side and in which a
tracking operation is applied to accuse at least one
of the colluders participating to the collusion scheme.
We will explain later each step separately.
3.1 Multi-level Fingerprint Generation
Step
Reducing the search space of dishonest users by as-
signing a user to a specific group represents a suitable
solution to face the Tardos accusation costs. The user
assignment to a group can be used to counter differ-
ent types of coalitions: temporal, geographic, social,
etc. In the hierarchy we embrace, each chosen con-
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
290
Figure 2: The proposed 5-level hierarchy for the fingerprint.
straint corresponds to a level. We detail our construc-
tion more precisely in the following.
3.1.1 Temporal Constraint
This constraint is tied to the time a video stays ac-
cessed by viewers. The frequency of the access to a
video in a VOD platform depends on its popularity. In
the beginning, when a video belonging to the Top 10
is added to a VOD platform, users are very curious
and the frequency accessing is very important, this
behavior changers later to decrease significantly over
time. In (Choi et al., 2012), the viewer interest de-
creases from 100% to less than 10% during a 4-month
period. Thats why, in the time level,we choose to rep-
resent the four first months by four groups where each
group is one month spent by a video in VOD applica-
tion.
3.1.2 Geographic Constraint
According to this constraint, we assume that two users
belonging to the same geographic place are more able
to collude together than with other users from other
regions. We have studied later in (Chaabane et al.,
2013), according to BSA report, that the software
piracy is more important in some countries than in
others.
3.1.3 Social Constraints
We have enhanced our study with statistics shared
by the NPD, National Purchase Diary Group, known
for its consumer market research. The NPD has
studied the media traffic in one of the most popular
Figure 3: The proposed tracing traitors framework.
VOD service, Netflix, and has shown that audience
behavior changes depending on the age and the
gender, the highest rate of users in a week is noted
with persons under 15 years old age, this study
demonstrates also that men are less interested with
VOD services than women.Thus, according to these
statistics, we remind the 5-level hierarchy we adopted
in (Chaabane et al., 2013), where each criterion is
represented by a level in the hierarchy. The first
level is the time level where we assume that the most
important period for a video life in a VOD platform
is about 4 months (Choi et al., 2012), (Liu et al.,
AMultimediaTracingTraitorsSchemeUsingMulti-levelHierarchicalStructureforTardosFingerprintBasedAudio
Watermarking
291
2014). Thats why, we considerate four groups in this
way:in the first group, users curiosity is moderately
important, it increases in the second month to reach
the maximal audience interest, it decreases later in
the third month to the minimal bound in the fourth
one. The second level in the hierarchy represents
the geographic criterion where we propose to divide
VOD users to two essential regions:Zone
A where
the piracy phenomenon is very important, especially
in Asia and Africa continents, and Zone
B where
the piracy phenomenon is less important especially
in Europe, America and Austria. We divide each
continent to two groups of principal countries in
the third level. In the fourth and the fifth levels,
we are interested in social criteria, respectively the
age and the gender ones. Once the hierarchy is
fixed, we form communities so that users having the
same characteristics belong to the same community
and have more probability to collude together in a
forgery attempt than with users belonging to other
communities. The resulting fingerprint for each user
is the concatenation of his community identifier
concatenated to his personal identifier which is
encoded with Tardos code.
Final identifer = id level1 + id level2+ ..
+id level5+ personal id
(1)
This hierarchy strategy is taken especially to reduce
computational costs and time during the Tardos accu-
sation process. Instead of parsing the totality of em-
bedded codewords, we search firstly the committee
having the highest similarity to the group identifier of
the colluded codeword, and then we compute users
scores belonging to it. The VOD context does not re-
quire that the tracing step should be performed in real
time. The whole operations of decoding and accusa-
tion are made offline. In the next section, we describe
the used watermarking technique.
3.2 Embedding Strategy: Audio
Watermarking Technique using
DCT Transform
Embedding collusion secure fingerprint codes with a
robust watermarking technique has necessarily im-
pacts in a tracing scheme, mainly against different
types of attacks. In our tracing approach, we propose
to use an audio watermarking technique described in
details in (Charfeddine et al., 2010) and then applying
DCT to them. The watermark is inserted in the middle
frequencies of each block. The watermark is hidden
especially in the sample closest to the averagevalue of
a localized middle frequencies band. The neural Net-
work is used here to be trained to retrieve from eight
neighbors samples the nearest sample having the best
embedding position. The resulting watermarked sig-
nal is obtained after applying the IDCT. The detection
step is the inverse process of the insertion one. Once
the multi-level fingerprint is embedded, the video is
diffused. The tracing process is launched then in the
distributor side.
3.3 Tracing Process
As shown in figure 2, the whole framework can be
divided into multiple phases: when the supplier de-
tects a copy with unknown fingerprint Y, he decides
to trace back colluders by analyzing the watermark
to extract its ID
group
and retrieving its belonging to a
community. The tracing is performed by the Tardos
code, which represents our forensic code. This proba-
bilistic code consists in generating firstly a set of den-
sity probabilities {p
1
··· p
m
} ∈ {0··· 1} and then con-
structing a matrix of n codewords X
ji
of length m with
Prob[X
ji
=1] = p
i
.In our approach, we use the sym-
metric function to decode whatever is the collusion
strategy. The correlation between the suspicious fin-
gerprint and codewords and the score S
j, j∈{1...n}
per
user are computed with ε
1
and ε
2
are respectively the
false positive and the false negative error probabili-
ties:
S
j
=
m
∑
i=1
g(Y
i
, X
ji
, p
i
) (2)
g(1, 1, p) = g(0, 0, 1− p) =
s
(1− p)
p
g(1, 0, p) = g(0, 1, 1− p) = −
r
p
(1− p)
(3)
It is applied only to users in the selected commu-
nity. We assume that users in the same commu-
nity have more probability to collude together (Des-
oubeaux et al., 2012) than with users belonging to an-
other community. We compute the score Sj per user in
the corresponding community. The principle of Tar-
dos accusation is to compute the similarity between
the codeword X
j
and the suspicious word Y for each
position j.The user whose score is higher than the
threshold Z is thus accused. Steps of the Tardos accu-
sation process are more detailed in (Chaabane et al.,
2013). At the last, as a result of our tracing algorithm,
we compute the detection rate of our system, if we
obtain a detection rate det rate close to 100%, the ac-
cusation is successful, which means that the number
of retrieved users Retrieved collu is equal to the real
number of Our tracing algorithm is briefly detailed
below; we assume that inputs are respectively: the
code length m, the collusion size c, the threshold Z
and the suspicious word Y. colluders c.
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
292
Figure 4: Embedding step of the used audio watermarking technique.
Begin
Procedure calcul_rate(m,Z: Real; c: Int,Y: Vector;
var det_rate, Sj: Real);
ID_community := ID_group(Y);
nb := size(ID_community);
While (j < c) and (j < nb) do
sj :=score (xj, Y, m) ;
if (sj > Z) then
return xj;
Retrived_collu := Retrieved_collu+1;
end
j := j+1;
end
det_rate := retrieved_coll / c;
End.
4 EXPERIMENTAL RESULTS
The real challenge of a tracing traitors scheme is to
cover the gap between theoretical and practical re-
sults. Optimizing the fingerprinting code parameters
and preserving the robustness even if the collusion
size increases are the most important requirements for
a tracing traitors scheme. According to the studies
made by (Choi et al., 2012) the most required videos
in a VOD platform are films, TV reality programs,
sport competitions and music clips. For this reason,
we have chosen four avi video samples belonging to
the four types. The extracted audio files are of dura-
tion of 5min16s. In our paper, we show experimental
results obtained by ice
film.avi and we prove firstly
the robustness of our scheme against all types of at-
tacks and then we vary collusion size and we present
the detection results.
Figure 5: Snapshots samples of experimented videos
respectively tv
prog.avi, ice film.avi, music clip.avi and
sport.avi.
4.1 Tracing Traitor Attacks
Attacks in tracing traitor field are classified according
to the layer they attack. We distinguish as a first class
attacks on the watermarking scheme, which are ap-
plied to destroy the watermark, this type of attacks re-
flect the robustness of the watermarking technique. In
experimentation section, we test the robustness of our
scheme against compression (128, 96 and 64 Kbps)
and some Stirmark attacks. The second class consists
in a scenario made by a group of users to construct a
false video copy with unknown fingerprint. Colluders
participating to such strategies try to make their iden-
tification too hard to the supplier. In our work, we
suppose that the type of colluders strategy is known
and we are interested to the most common collusion
attacks (Hamida et al., 2011): majority/minority vote,
all1 and all0 attacks.
In the table below, we present for each collusion
attack an example to describe it.
AMultimediaTracingTraitorsSchemeUsingMulti-levelHierarchicalStructureforTardosFingerprintBasedAudio
Watermarking
293
Table 1: The most common collusion attacks.
Attack strategy Example
Majority vote attack
0 1 0
1 0 0
1 1 1
1 1 0
Minority vote attack
0 1 0
1 0 0
1 1 1
0 0 1
Allone attack
0 1 0
1 0 0
1 1 1
1 1 1
Allzero attack
0 1 0
1 0 0
1 1 1
0 0 0
4.2 Robustness and Inaudibility Results
We use two major criteria in this experimental part:
NC, the Normalized Cross Correlation, which value
reflects the similarity between the original watermark
and the detected one, and the SNR, the Signal to
Noise Ratio, which is an objective measure to show
the quality of the audio after the insertion step.
SNR = 10× log
∑
i
Y
2
i
∑
i
Y
i
−Y
′
2
!
(4)
Y and Y are respectively the original audio and the
watermarked one.
NC =
∑
N
x=1
W (x) ·W
′
(x)
q
∑
N
x=1
W (x)
2
·
∑
N
x=1
W
′
(x)
2
(5)
W and W are respectively the original watermark and
the detected one.
We remind that Tardos parameters m and c are tied by
the equation below:
m = 2Π
2
c
2
ln
1
ε
1
(6)
In the table above, we show the different values of
collusion size c, the false alarm probability and the
code length variations. We vary also the total number
of users from 10, 100 and 1000.
Table 2: Collusion size, false alarm probability and code
length variations.
Collusion size False alarm value Code length
c=5 ε
1
= 10
−3
m=3455
c=5 ε
1
= 10
−4
m=4935
c=5 ε
1
= 10
−7
m=8390
Figure 6: Robustness against Stirmark and compression
attacks with Tardos parameters: n=100, ε
1
= 10
−3
, c =
5, SNR = 53.0262.
Figure 7: Detection rates for the two structures with Tardos
parameters: n=100,ε
1
= 10
−3
, c = 5, 100trials.
Figure 8: Tracing time for the two types of structures in the
case for clip.avi, Tardos parameters: n=100,ε
1
= 10
−3
, c =
5, 100trials.
In order to test our system performance against
different types of attacks, we compare it to a nonhier-
archical fingerprint under the same parameters the fin-
gerprint length m = 3455, the number of users n=100
and the collusion size c=5. The detection rate and
the CPU time are important tracing criteria. We can
remark from figures 6, 7, 8, 9, 10, 11 and 12 that
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
294
Figure 9: Detection rates for the two structures with Tardos
parameters: n=1000,ε
1
= 10
−4
, c = 5, 100trials.
Figure 10: Tracing time for the two types of structures with
Tardos parameters: n=1000,ε
1
= 10
−4
, c = 5, 100trials.
Figure 11: Detection rates for the two structures
with Tardos parameters: n=1000,,ε
1
= 10
−7
, c =
5forclip.avi, 100trials.
the detection rates of our hierarchical fingerprint are
preserved while the nonhierarchical structure perfor-
mance decreases notably if the user number increases
and the false probability decreased.
5 CONCLUSIONS
In this paper, we have presented our contribution in
tracing traitors field. We have proposed a multilevel
hierarchical fingerprint to reduce tracing and compu-
Figure 12: Tracing time for the two types of structures with
Tardos parameters: n=1000,,ε
1
= 10
−7
, c = 5100trials.
tational costs of Tardos code; we have proposed also
to hide this fingerprint with a robust audio watermark-
ing technique which is a choice different from pro-
posed works in literature. Our proposed system has
provided good robustness against different types of
attacks and good detection rates. In future work, we
propose to assume that the collusion strategy is un-
known and we will try to estimate it.
REFERENCES
Chaabane, F., Charfeddine, M., and Amar, C. B. (2013). A
survey on digital tracing traitors schemes. In Informa-
tion Assurance and Security, pages 85–90.
Charfeddine, M., El’arbi, M., Koub`aa, M., and Amar, C. B.
(2010). Dct based blind audio watermarking scheme.
In SIGMAP, pages 139–144.
Charpentier, A., Fontaine, C., Furon, T., and Cox, I. J.
(2011). An asymmetric fingerprinting scheme based
on tardos codes. In Information Hiding, pages 43–58.
Charpentier, A., Xie, F., Fontaine, C., and Furon, T. (2009).
Expectation maximization decoding of tardos proba-
bilistic fingerprinting code. In Media Forensics and
Security, page 72540.
Choi, J., Reaz, A. S., and Mukherjee, B. (2012). A survey
of user behavior in vod service and bandwidth-saving
multicast streaming schemes. IEEE Communications
Surveys and Tutorials, 14(1):156–169.
Desoubeaux, M., Guelvouita, G. L., and Puech, W. (2011).
Probabilistic fingerprinting codes used to detect traitor
zero-bit watermark. In SPIE Proceedings Vol. 7880:
Media Watermarking, Security, and Forensics III.
Desoubeaux, M., Guelvouita, G. L., and Puech, W. (2012).
Fast detection of tardos codes with boneh-shaw types.
In Proc. SPIE 8303, Media Watermarking, Security,
and Forensics.
Furon, T., P´erez-Freire, L., Guyader, A., and C´erou, F.
(2009). Estimating the minimal length of tardos code.
In Information Hiding, pages 176–190.
Hamida, A. B., Koub`aa, M., and Nicolas, H. (2011). Hi-
erarchical traceability of multimedia documents. In
AMultimediaTracingTraitorsSchemeUsingMulti-levelHierarchicalStructureforTardosFingerprintBasedAudio
Watermarking
295
Computational Intelligence in Cyber Security, pages
108–113.
Hayashi, N., Kuribayashi, M., and Morii, M. (2007).
Collusion-resistant fingerprinting scheme based on
the cdma-technique. In IWSEC, pages 28–43.
Laarhoven, T. (2013). Dynamic traitor tracing schemes, re-
visited. CoRR, abs/1307.0214.
Laarhoven, T. and de Weger, B. (2011). Optimal symmetric
tardos traitor tracing schemes. CoRR, abs/1107.3441.
Shahid, Z., Chaumont, M., and Puech, W. (2013).
H.264/avc video watermarking for active fingerprint-
ing based on tardos code. Signal, Image and Video
Processing, 7(4):679–694.
Tardos, G. (2003). Optimal probabilistic fingerprint codes.
In STOC, pages 116–125.
Wang, Z. J., Wu, M., Trappe, W., and Liu, K. J. R. (2004).
Group-oriented fingerprinting for multimedia foren-
sics. EURASIP J. Adv. Sig. Proc., 2004(14):2153–
2173.
Ye, C., Ling, H., Zou, F., and Lu, Z. (2013). A new finger-
printing scheme using social network analysis for ma-
jority attack. Telecommunication Systems, 54(3):315–
331.
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
296
