Droppix: Towards More Realistic Video Fingerprinting
Przemysław Bła
´
skiewicz
1,2
, Marek Klonowski
1,2
and Piotr Syga
1,2
1
Wrocław University of Science and Technology, Poland
2
Vestigit sp.z o.o., Poland
Keywords:
Watermarking, Video, Copyright Management, Intellectual Property Protection.
Abstract:
We present preliminary results for a way of video fingerprinting that is different from typical methods based
on paradigms from watermarking of still images. Our approach is based on modifying some fragments of the
clip in a carefully chosen manner. We show the robustness of our approach against a number of typical of
attacks. The marks introduced by our modifications are almost imperceptible to the viewer and their impact
can be adjusted. Finally, our protocol is computationally light and can be combined with others schemes as an
extra security layer.
1 INTRODUCTION
With a rapid growth of multimedia market, the need
for preventing piracy is stronger than ever before. An
illegal redistribution of copyrighted content causes
significant losses to its producers and providers. This
is the main motivation for constructing more and
more advanced audio and video fingerprinting tech-
niques, that introduce individual marks into each copy
of a clip. Such mark (fingerprint) can be then linked
with a legitimate receiver. If a video is illegally circu-
lated, e.g. retransmitted, the copy can attributed to its
associated legitimate user.
One of the fingerprinting techniques for multi-
media is watermarking which leverages the fact that
bandwidth of digital video/image signals is much
higher that the amount of information available to
the human eye. The idea is to embed in each copy
of a movie or picture marks that are invisible to the
viewer, however can be easily detected by the copy-
right owner.
Designing methods of fingerprinting one needs to
take into account that the watermarked content can
be subject to various attacks aimed at removing them
without significant degradation of quality. That is, an
attacker aims at destroying the embedded message,
yet preserving the commercial value of the original
data. There are many transformations that could be
This Work Is Partially Supported by the National
Centre for Research and Development (Ncbir) Project
POIR.01.01.01-00-1032/18 and Polish National Science
Centre – Project UMO-2018/29/B/ST6/02969.
viewed as an attack on the watermark, including rota-
tions, flipping along the vertical axis or using various
filters (see e.g. (Asikuzzaman and Pickering, 2018)
for a list of possible attacks). In principle, an ade-
quate fingerprinting method needs to provide at least
two features The first is transparency quality of
the image with embedded fingerprinting information
cannot be significantly worse than that of the origi-
nal. Ideally, the watermarked content is indistinguish-
able from the original one. Second fundamental issue
is robustness it is expected that the fingerprinting
method should be immune to all attacks that preserve
reasonable quality of the output. That is, the adver-
sary could be capable of removing the characteristic
mark only if the video or image is changed to such
an extent that it becomes useless for commercial pur-
poses.
While fingerprinting methods designed for static
pictures seem to be very mature, there is still many
problems with securing video content. Many papers
treat video material as a sequence of static pictures
and a typical attempt to fingerprint a video is to re-
place one such picture with any efficient “static” wa-
termark. Even though such approach is theoretically
efficient, it usually does not work in practice. The
majority of today’s video content is represented as
a GoP structure (ISO/IEC 23009-1, 2012; ISO/IEC
14496-12, 2011), with carefully chosen key frames
(IFrames) and vast majority of finally displayed im-
ages is represented as difference between the frames
(PFrames or BFrames). While it allows dropping re-
dundant data, an efficient embedding a picture into a
468
Bła
´
skiewicz, P., Klonowski, M. and Syga, P.
Droppix: Towards More Realistic Video Fingerprinting.
DOI: 10.5220/0009876104680476
In Proceedings of the 17th International Joint Conference on e-Business and Telecommunications (ICETE 2020) - SECRYPT, pages 468-476
ISBN: 978-989-758-446-6
Copyright
c
2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
video can be impossible, since fragile changes can be
immediately removed by the encoder before finger-
printing marks are detected.
Our Contribution. In this paper we test an ap-
proach to video fingerprinting that is more suitable
for real-life video content. The basic idea is to care-
fully modify some fragments of the video in a way
that is characteristic for each copy of a given film.
Deletion is however not realised by removing frames
but rather fading out some short fragments. This ap-
proach preserves the quality of the original video in
a much better way and is almost imperceptible to the
viewers if implemented properly. In our method the
copy of a film can be detected even by cam-coding
a film displayed on a wide classes of devices. Our
protocol does not require any heavy computations for
embedding or detection, unlike some new and very
promising watermarking protocols based on machine
learning (rather for still images than video material).
To the best of our knowledge, the only similar ap-
proach to video fingerprinting is presented in (Lee
et al., 2006) wherein the authors present a nice idea
of removing single frames from a movie. Note that
in contrast to the paper, we discuss the impracticality
of removing frames (cf. Sect. 3) and instead we mod-
ify their luminance. Moreover, we investigate a wider
spectrum of attacks than Lee et al. (cf. Sect 4).
2 PREVIOUS AND RELATED
WORK
Methods for Static Pictures. Methods of robust
fingerprinting of digital content have been studied for
a long time. Majority of commonly used techniques
can be found in survey (Potdar et al., 2005). The
earliest methods of utilising transparent watermarks
relied on negligible modifications made to the spa-
tial domain of the image (e.g. modifying least sig-
nificant bits (LSB) of the colour intensity (Fridrich,
1998) followed by (Tjokorda Agung B.W et al., 2012)
that enables RLE compression or statistical testing of
the mean intensity of the marked pixels (Nikolaidis
and Pitas, 1998)). Due to the robustness requirement,
coping with the effects of compression was of partic-
ular interest (Sebé et al., 2000). Often, when the in-
tensity modification was made in the spatial-domain,
the authors resort to converting colour space (e.g., to
YCbCr (Patvardhan et al., 2018)), that allows more
significant (i.e., more robust) modifications without
significant visible changes. Note that transparency is
not always required; the authors in (Ping Wah Wong
and Memon, 2001) allowed a parameter that is re-
sponsible for visibility or the transparency of the wa-
termark, depending on the embedder’s intentions. In
order to allow more robust and transparent modifica-
tions, many researchers turned to frequency-domain
alterations. Due to the nature of popular compression
algorithms, Discrete Cosine Transform was a primar-
ily investigated method (Hwang et al., 2003; Chang
and Tsan, 2005). In (Kingsbury, 2000), the author
introduced a concept of utilising Dual-tree complex
wavelet transform. This approach was followed ex-
tensively by (Mabtoul et al., 2007; Alkhathami et al.,
2013; Bhatnagar and Wu, 2013; Zebbiche et al., 2018)
among others. Another widely used technique is Sin-
gular Value Decomposition (SVD) (Gupta and Raval,
2012; Makbol and Khoo, 2013; Makbol and Khoo,
2014). Hybrid approaches, mixing the basic space-
domain methods like LSB with frequency-domain ap-
proaches were presented e.g., in (Sheth and Nath,
2016), watermarking in frequency-domain with utili-
sation of SVD (Gaur and Srivastava, 2017) or utilising
Hadamard transform with Schur decomposition (Li
et al., 2018). A new approach to watermarking was
introduced with the advent of neural networks and
deep learning. The authors of (Tsai and Liu, 2011)
combine the idea of wavelet transforms with utilis-
ing neural networks. In (Zhu et al., 2018) a novel
framework for end-to-end approach utilising an en-
coder and decoder with nosier layers was introduced.
The authors of (Wen and Aydore, 2019) presented the
notion of adversarial training, whereas (Zhong and
Shih, 2019) proposed a fully automated system for
images captured directly from the camera. An addi-
tional neural network was used in (Luo et al., 2020)
so that distortions improving robustness were added.
Methods for Motion Pictures. Majority of algo-
rithms for marking video material are based on a di-
rect application of methods constructed for still pic-
tures. The main flaw of direct mapping of the static
image watermarking algorithms to motion pictures is
that the time-domain compression results, quite of-
ten, in deletion or distortion of the embedded water-
mark. Nevertheless, the authors of many watermarks
designed for videos followed the way paved by static
images, e.g. DCT (Sun et al., ), DT-CWT (Coria et al.,
2008), 3D Discrete Fourier Transform (Deguillaume
et al., 1999). In (Noorkami and Mersereau, 2005)
main focus was watermarking in H.264 compressed
domain and the authors of (Xu et al., 2011) use DCT
in order to gain robustness against H.264/AVC com-
pression. Somewhat different, and quite similar to
the one described in this paper (cf. 3) approach was
presented by (Lee et al., 2006), where certain frames
Droppix: Towards More Realistic Video Fingerprinting
469
from a movie are removed so that the pattern of
missing frames allows unique user identification. An
object-based watermarking was introduced in (Swan-
son et al., 1997) and extended to scene in (Swanson
et al., 1998), whereas in (Lee and Seo, ) an adaptive
modulation is used. A robust, visual watermarking
has been proposed in (Zhang et al., 2007). More in-
formation on the techniques used in video watermark-
ing may be found in (Doerr and Dugelay, 2003; Bhat-
tacharya et al., 2006; Chang et al., 2011; Asikuzza-
man and Pickering, 2018).
3 DESCRIPTION OF THE
ALGORITHM
The general intent while working on the algorithm
was to embed a unique sequence into the movie, so
that the numbers retrieved from the sequence uniquely
identify the user leaking the movie. Such approach is
fundamentally different to the approach of adapting
static image watermarking to embed entire informa-
tion in a single frame. A natural idea to embed such
watermark is to add some mark on selected frames
and the number of the frames create a unique se-
quence. In order to cope with compression (both of
a single frame and time-domain) and provide some
level of robustness against attacks, one can go to the
extreme and remove certain frames (Lee et al., 2006).
Such approach has two major downsides: it is diffi-
cult to implement in a compressed movie (e.g. AVC
or HEVC) and often frame removal is simulated by
replacing it with a black one, which allows easy lo-
calisation by the adversary and “filling up the hole” if
they have access to another copy of the movie. Alter-
natively, the attacker can black-out few more frames
in order to mislead the decoder. To fix it we can en-
sure that each copy not only has a sequence of frames
that are “missing” (either blacked-out or removed) but
also there is a frame that is not removed in only this
single copy. However, the second issue arises with the
fix such idea does not scale well with the number
of users. An average feature movie lasts around 100
minutes, which results in roughly 144000 frames. If
there needs to be a single frame that is unique to a
given user, the watermarking for only 100000 possi-
ble users requires removing the majority of the movie
for all the users. Due to the problems described above
and specific challenges presented by AVC or HEVC
compression when trying to remove a frame (espe-
cially problematic are the cases when two consecu-
tive frames in a single GOP are removed), instead of
removing the frames, replacing them with a black one
or leaving some visual macroartifact, we decided to
test if modification of the luminance of a frame would
provide adequate marking capabilities.
Droppix. We start our algorithm with generating
S , a sequence of the frames to be watermarked. It
uniquely identifies a user and is generated as a code-
word of their unique number. In the final version
of the algorithm Reed-Solomon codes or Levenshtein
distance may be used to extend S to provide sufficient
correcting power for cases when original frames from
S have not all been detected by the decoder or ad-
ditional ones have been falsely identified. Next, we
decode the movie in order to obtain a sequence of
frames. For each frame s S we perform the same
operations (Alg. 8). We start by converting the frame
into CIELAB colour space. Next, we multiply the L
component (lightness) of each pixel by a predefined
constant α. Our test showed that in order for a ro-
bust, yet transparent watermark, the values between
0.92 and 0.97 can be used. We proceed with convert-
ing the frame back to BGR colour space and replacing
the resulting frame instead of s. When all frames in S
were modified, we code the movie back to the original
container using compression algorithm.
Data: M – movie to be watermarked, k – user
key (e.g. device ID), α – watermark
strength parameter
Result: M – watermarked movie
1 S SequenceGen(k);
2
h
f
1
, f
2
, . . . , f
n
i
Decode(M);
3 for i
{
1, 2, . . . , n
}
do
4 if i S then
5
ˆ
f BGR2LAB( f
i
);
6
ˆ
f
L
ˆ
f
L
· α;
7 f
i
LAB2BGR(
ˆ
f );
8 M Encode(
h
f
1
, f
2
, . . . , f
n
i
);
Algorithm 1: Droppix: watermark embedding.
In order to decode the watermark, we need also
the original movie and the set of parameters used by
the embedding algorithm. We start by decoding both
movies to obtain respective sequences of frames. We
proceed by comparing pairs of corresponding frames
using structural similarity (Zhou Wang et al., 2004)
of their grayscale versions. Next, we calculate the
mean square error of the similarity distances for each
frame-pair in an array mse. In order to retrieve the
information on the frames that were watermarked we
establish a threshold of considered values and local-
ising the peaks of the normalised errors. The indexes
of the peaks are then subject to the reverse S generat-
ing algorithm in order to correct possible mistakes in
SECRYPT 2020 - 17th International Conference on Security and Cryptography
470
retrieving the information and to recover the user ID
of the leaker. The idea of the decoding procedure is
shown in Algorithm 2.
Data: M – original movie,
ˆ
M – copy of the
movie with potential watermark, K
set of user keys, t – acceptance
threshold
Result: k – user ID
1 SEtab[];
2
h
f
1
, f
2
, . . . , f
n
i
Decode(M);
3
ˆ
f
1
,
ˆ
f
2
, . . . ,
ˆ
f
n
Decode(
ˆ
M);
4 for i
{
1, 2, . . . , n
}
do
5 g RGB2GRAY( f
i
);
6 ˆg RGB2GRAY(
ˆ
f
i
);
7 if StructuralSimilarity(g, ˆg)<t
then
8 mse MSE(g, ˆg);
9 SEtab.append(mse);
10 medianmedian(SEtab);
11 SEtabmedian;
12 PfindPeakIndexes();
13 kSequenceGen
1
(P);
Algorithm 2: Droppix: watermark decoding.
4 EXPERIMENTAL RESULTS
Corpus. In order to test our solution we gathered
a corpus of clips from various commercial television
broadcasts to be our testing body. They all were
H.264 encoded HD quality 120 seconds long frag-
ments. In each of those samples a watermark was em-
bedded using our method, and for each such marking
the numbers of frames that we used for watermarking
were stored.
Watermarking. Next, we checked how our water-
marking procedure changes values of mean square er-
ror (MSE), structural similarity (SS) and peak signal-
to-noise ratio (PSNR) calculated with respect to non-
watermarked clips. Fig. 1, shows the influence of
droppix parameter α on those indicators for an exem-
plary clip. One can tell that as α approaches 1, there
is smaller influence of the watermark on the parame-
ters, and hence detection success rate can be expected
to drop. More importantly, the watermarking leaves
a clear peak in all three values, a property we want
to maintain when the clip undergoes a modification
attack.
Visual inspection of the influence of α resulted
in the following observations. For value 0.9 the
0
20
40
60
80
100
120
140
160
180
0 5 10 15 20 25 30 35 40 45
MSE
Framenumber
0.90
0.92
0.95
(a) MSE
24
26
28
30
32
34
36
38
40
42
0 5 10 15 20 25 30 35 40 45
PSNR
Framenumber
0.90
0.92
0.95
(b) PSNR
0.95
0.955
0.96
0.965
0.97
0.975
0.98
0.985
0 5 10 15 20 25 30 35 40 45
Structuralsimilarity
Framenumber
0.90
0.92
0.95
(c) SS
Figure 1: Influence of watermarking on MSE, PSNR and
SS values for different α.
modified frame can be slightly noted as flicker, with
0.95 a (subjective) perception is marginal and de-
pends heavily on the particular scene the frame be-
longs to. On the other hand, as discussed below, for
values above 0.95 the success rate of detection dete-
riorates rapidly. Example original and watermarked
frames (for α = 0.92) and their difference are shown
in Fig. 2. Note that difference between “clean” and
watermarked frame is a uniquely almost-black rectan-
Droppix: Towards More Realistic Video Fingerprinting
471
(a) Original frame
(b) Watermarked frame
(c) Difference – uniformly almost black
Figure 2: Example frame from testing clip.
gle, since the luminance of all pixels has been reduced
by the same (small) amount.
Attacks. The aim of the attack in our scenario is to
change PSNR, MSE and SS characteristics of the clip,
so that finding the frames used for watermarking be-
comes impossible.
From many possible attacks possible we chose six
to give a general idea about the performance of our
scheme. These attacks are easy to perform and as
such are more available to less experienced attackers,
and there might be more leaked content modified by
such simple means. The attacks we chose are in the
list below, and for each we provide a short rationale
for choosing it. All attacks kept other characteristics
(resolution, framerate, etc.) of the movie unchanged.
1. Transcoding from H.264 to MPEG-2 format and
downsampling to SD. Both these formats are pop-
ular and can provide good quality video. Due
to different methods of compression, transcod-
ing can effectively erase pixel-level modifications
used in other watermarking techniques. Since our
approach modifies lightness (also a visual prop-
erty), we checked how different compression al-
gorithm influences detection.
2. Vertical flip. This method is frequently applied to
counter watermarking techniques relying on spa-
tial information within a frame. Importantly, ver-
tical flip can be visually acceptable for the viewer
as little content is lost and there are many exam-
ples of such content on the internet.
3. Globally increasing brightness. Since lightness of
a frame is our embedding technique, such attack
might limit our detection capability.
4. Slight rotation (1 degree). Similar to the vertical
flip above, this attack yields visually acceptable
results to the viewer, yet is able to mis-lead wa-
termarking schemes utilising spacial information
of a frame. Because we utilise structural similar-
ity between frames as one of the decision factors,
we wanted to verify how such slight change might
affect detection.
5. Blurring. This is one of the common attacks
against watermarks embedded in pixel structure
of the frame. To test our approach with this at-
tack, we selected (subjectively) maximal parame-
ters for blurring both chrominance and luminance
channels so that the result was still acceptable to
the viewer as HD content.
6. Frame averaging. This is another popular tech-
nique of attack against watermarking. Essentially,
it is a low-pass filter applied to the movie so that
fast/sharp movement, which the human eye is not
capable of identifying is smoothed out. In our
tests we transformed each frame to be the average
of two consecutive frames, yielding still accept-
able results to the viewer.
All attacks were performed using ffmpeg in ver-
sion 4.2.2 and the implementation of the watermark-
ing and detection algorithms was done in Python with
OpenCV ver. 4.2.0 bindings and SciKit ver. 0.16.2
Testing Framework. Each clip from the corpus
was first watermarked. Next, the content was modi-
fied according to the attack scenario. Finally, both the
original clip and the watermarked and modified con-
tent were passed to watermark detection. In that pro-
cess, the PSNR, structural similarity and mean-square
error (MSE) for each pair of frames was calculated.
Next, frame numbers suspected as ones encoding the
watermark were selected. Finally, the chosen frame
numbers were matched against those stored for par-
ticular clip at the moment of watermark insertion. If
SECRYPT 2020 - 17th International Conference on Security and Cryptography
472
all frames have been found and no other the trial was
deemed success. Any additional frames except those
determined by watermarking yielded a false-positive.
And finally, when not all watermarked frames were
found, the trial was considered failed.
Results. For the sake of presentation, the results are
shown for clips where watermark was embedded us-
ing frames number 10, 15, 20 and 22, which allows vi-
sual inspection of obtained data. There were 28 such
clips. Data presented in Fig. 1, 3, 4 and 5 correspond
to the same clip, so that relative change between non-
attacked and attacked/modified video can be traced.
Overall results are shown in Tab. 1. Columns
“success” and “FP” show the number of clips where
the watermark was identified, and where additional
frames were also selected, respectively. The column
“extra” presents the mean number of extra frames de-
tected and the column σ shows its standard devia-
tion.
It can be seen that the watermark embedding is
little influenced by brightness modification, and much
less still by transcoding and down-sampling. Rotation
by a small degree makes the scheme more prone to
false-positives as well as limits the success rate. Ver-
tical flip attack seems to be most effective against our
method, however, on closer inspection it was revealed
that MSE values calculated in this case were smaller
then average for frames used for watermarking than
for regular ones. This is exactly the opposite from
what has been observed for other attacks and conse-
quently our decoding algorithm was not able to detect
those negative peaks. This is illustrated in Fig. 3.
-600
-500
-400
-300
-200
-100
0
100
200
300
400
0 5 10 15 20 25 30 35 40 45
Deviation
Framenumber
vflip
brightness
mpeg2
rotate
blur
averaging
Figure 3: MSE deviation from average per frame for dif-
ferent attacks on the same movie. Watermark is inserted in
frames 10, 15, 20 and 22.
In general, the flipping attack has the same impact
on PSNR and structural similarity indicators in that it
produces opposite orientation of peaks as compared
to other studied attacks (Figs. 4 and 5). Therefore, for
10
15
20
25
30
35
40
0 5 10 15 20 25 30 35 40 45
PSNR
Framenumber
vflip
brightness
mpeg2
rotate
blur
averaging
Figure 4: PSNR of frames for different attacks on the same
movie. Watermark is inserted in frames 10, 15, 20 and 22.
0.4
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30 35 40 45
Structuralsimilarity
Framenumber
vflip
brightness
mpeg2
rotate
blur
averaging
Figure 5: Structural similarity of attacked frames to origi-
nals for different attacks on the same movie. Watermark is
inserted in frames 10, 15, 20 and 22.
this particular attacks our detection algorithm must be
adequately adjusted.
Interpretation of Results. Our preliminary results
on the proposed method for watermarking suggest
that it is immune against some popular and common
attacks on video content. While some aspects of the
detection method require a more evolved approach,
the tests support our claim about potential of this
method, particularly as a one-of-many methods used
concurrently on the same video content.
Building a fully functional detector requires more
samples and wider range of attack cases, or employ-
ing ANN as a decision maker based on a sequence of
tuples of the form (PSNR, MSE, SS) for each frame.
Different clips have different thresholds by which to
decide whether a given frame pertains to the water-
mark or not. As a consequence, the detector should
determine acceptance/rejection parameters on the fly
and for particular clip, with only some initial startup.
Because our research was about testing the idea,
our testing scenario did not use redundancy codes
Droppix: Towards More Realistic Video Fingerprinting
473
Table 1: General detection results for 28 clips.
α 0.9 0.92 0.95
Attack success FP extra σ success FP extra σ success FP extra σ
brightness 28 0 28 0 28 9 5.88 1.36
vflip 0 27 3.22 2.79 0 25 3.25 2.81 0 27 3.28 2.71
mpeg2 + SD 28 3 1 0 28 3 1 0 28 2 1 0
rotate 23 28 4.14 2.42 20 27 3.96 2.8 16 27 4.07 2.22
blur 27 8 8 5.07 26 12 8.58 4.31 24 18 8.78 4.37
averaging 26 5 2.4 1.36 25 7 4.14 3 23 11 6.36 3.08
nor did we use any correlation between numbers
of frames used for embedding. The use of Reed-
Solomon codes and the like would allow rising α =
0.94 even higher (since 0.98 produced more failed at-
tempts due to not all frames being detected). And
because in our tests not detecting all frames used
for embedding was considered failure, some standard
method of recovering missing information would fur-
ther rise the success rate.
As we mentioned above, a broader framework can
be devised where there is another scheme, that co-
exists in the movie and is resistant to other attacks,
for which our scheme is susceptible.
5 CONCLUSION
In our paper we presented an approach to finger-
printing of video materials that diverges from a usual
paradigm. The proposed protocol provides immu-
nity against some common attacks and is adequate
for contemporary multimedia format. Moreover, the
transparency of our scheme is acceptable enough to
be commercially acceptable and can be traded off
for better detection rates. The presented protocol
is somehow “orthogonal” to and independent from
many other methods of typical fingerprinting. There-
fore we believe that our approach can also be used
as an extra security layer together with other, classi-
cal protocols. The synergy of such combination can
lead to a very strong immunity against various attacks.
One path of experimentation is that of immunizing
the scheme against attack using a certain number of
copies of the movie.
Additionally, as a future work we leave out
enhancing the method so that it allows blind-
identification. The only reason the original footage is
required is for estimation of MSE between the frames,
however as one of every few IFrames is watermarked
by increasing its luminance channel, it will be an out-
lier among frames describing the same scene. Locali-
sation of the outliers in a sliding window would allow
identification of the marked frames without a refer-
ence clip, and due to its significant robustness against
averaging it is feasible to adjust the parameters so that
blind detection is possible without significant loss in
accuracy.
REFERENCES
Alkhathami, M., Han, F., and Van Schyndel, R. (2013). Fin-
gerprint image watermarking approach using dtcwt
without corrupting minutiae. In 2013 6th Inter-
national Congress on Image and Signal Processing
(CISP), volume 03, pages 1717–1723.
Asikuzzaman, M. and Pickering, M. R. (2018). An
overview of digital video watermarking. IEEE Trans-
actions on Circuits and Systems for Video Technology.
Bhatnagar, G. and Wu, Q. J. (2013). Biometrics inspired
watermarking based on a fractional dual tree com-
plex wavelet transform. Future Generation Computer
Systems, 29(1):182 195. Including Special section:
AIRCC-NetCoM 2009 and Special section: Clouds
and Service-Oriented Architectures.
Bhattacharya, S., Chattopadhyay, T., and Pal, A. (2006).
A survey on different video watermarking techniques
and comparative analysis with reference to h.264/avc.
In 2006 IEEE Int. Symposium on Consumer Electron-
ics.
Chang, H. T. and Tsan, C. L. (2005). Image watermarking
by use of digital holography embedded in the discrete-
cosine-transform domain. Appl. Opt., 44.
Chang, X., Wang, W., Zhao, J., and Zhang, L. (2011). A
survey of digital video watermarking. In 2011 Sev-
enth International Conference on Natural Computa-
tion, volume 1, pages 61–65.
Coria, L. E., Pickering, M. R., Nasiopoulos, P., and Ward,
R. K. (2008). A video watermarking scheme based on
the dual-tree complex wavelet transform. IEEE Trans-
actions on Information Forensics and Security, 3.
Deguillaume, F., Csurka, G., O’Ruanaidh, J. J., and Pun,
T. (1999). Robust 3D DFT video watermarking. In
Security and Watermarking of Multimedia Contents,
volume 3657. Int. Society for Optics and Photonics.
Doerr, G. and Dugelay, J.-L. (2003). A guide tour of video
watermarking. Signal Processing: Image Communi-
cation, 18(4):263 282. Special Issue on Technolo-
gies for Image Security.
SECRYPT 2020 - 17th International Conference on Security and Cryptography
474
Fridrich, J. (1998). Image watermarking for tamper detec-
tion. In Proceedings 1998 International Conference
on Image Processing. ICIP98 (Cat. No.98CB36269),
volume 2, pages 404–408 vol.2.
Gaur, S. and Srivastava, V. K. (2017). A hybrid rdwt-dct and
svd based digital image watermarking scheme using
arnold transform. In 2017 4th Int. Conf. on Signal
Processing and Integrated Networks (SPIN).
Gupta, A. K. and Raval, M. S. (2012). A robust and se-
cure watermarking scheme based on singular values
replacement. Sadhana, 37(4):425–440.
Hwang, D.-C., Bae, K.-H., Lee, M.-H., and Kim, E.-S.
(2003). Real-time stereo image watermarking using
discrete cosine transform and adaptive disparity maps.
In Tescher, A. G., Vasudev, B., Jr., V. M. B., and Di-
vakaran, A., editors, Multimedia Systems and Applica-
tions VI, volume 5241, pages 233 – 242. International
Society for Optics and Photonics, SPIE.
ISO/IEC 14496-12 (2011). Information technology – Cod-
ing of audio-visual objects. Standard, International
Organization for Standardization, Geneva, CH.
ISO/IEC 23009-1 (2012). Information technology Dy-
namic adaptive streaming over HTTP (DASH). Stan-
dard, International Organization for Standardization,
Geneva, CH.
Kingsbury, N. (2000). A dual-tree complex wavelet
transform with improved orthogonality and symmetry
properties. In Proceedings 2000 Int. Conf. on Image
Processing, volume 2.
Lee, S.-W. and Seo, D.-I. Novel robust video watermarking
algorithm based on adaptive modulation.
Lee, Y., Kim, C., and Lee, S. (2006). Video fingerprinting
based on frame skipping. In Proceedings of the Inter-
national Conference on Image Processing, ICIP 2006,
October 8-11, Atlanta, Georgia, USA. IEEE.
Li, J., Yu, C., Gupta, B. B., and Ren, X. (2018). Color image
watermarking scheme based on quaternion hadamard
transform and schur decomposition. Multimedia Tools
and Applications, 77(4):4545–4561.
Luo, X., Zhan, R., Chang, H., Yang, F., and Milanfar,
P. (2020). Distortion Agnostic Deep Watermarking.
arXiv e-prints, page arXiv:2001.04580.
Mabtoul, S., Hassan, E., Elhaj, I., and Aboutajdine, D.
(2007). Robust color image watermarking based on
singular value decomposition and dual tree complex
wavelet transform. In 2007 14th IEEE International
Conference on Electronics, Circuits and Systems.
Makbol, N. M. and Khoo, B. E. (2013). Robust blind im-
age watermarking scheme based on redundant discrete
wavelet transform and singular value decomposition.
AEU-International Journal of Electronics and Com-
munications, 67(2):102–112.
Makbol, N. M. and Khoo, B. E. (2014). A new robust and
secure digital image watermarking scheme based on
the integer wavelet transform and singular value de-
composition. Digital Signal Processing, 33:134–147.
Nikolaidis, N. and Pitas, I. (1998). Robust image water-
marking in the spatial domain. Signal Processing, 66.
Noorkami, M. and Mersereau, R. M. (2005). Compressed-
domain video watermarking for h.264. In IEEE Int.
Conf. on Image Processing 2005.
Patvardhan, C., Kumar, P., and Vasantha Lakshmi, C.
(2018). Effective color image watermarking scheme
using ycbcr color space and qr code. Multimedia Tools
and Applications, 77(10):12655–12677.
Ping Wah Wong and Memon, N. (2001). Secret and public
key image watermarking schemes for image authenti-
cation and ownership verification. IEEE Transactions
on Image Processing, 10(10):1593–1601.
Potdar, V. M., Han, S., and Chang, E. (2005). A survey of
digital image watermarking techniques. In INDIN ’05.
2005 3rd IEEE International Conference on Industrial
Informatics, 2005., pages 709–716.
Sebé, F., Domingo-Ferrer, J., and Herrera, J. (2000).
Spatial-domain image watermarking robust against
compression, filtering, cropping, and scaling. In Goos,
G., Hartmanis, J., van Leeuwen, J., Pieprzyk, J., Se-
berry, J., and Okamoto, E., editors, Information Secu-
rity, pages 44–53. Springer Berlin Heidelberg.
Sheth, R. K. and Nath, V. V. (2016). Secured digital image
watermarking with discrete cosine transform and dis-
crete wavelet transform method. In 2016 International
Conference on Advances in Computing, Communica-
tion, Automation (ICACCA) (Spring), pages 1–5.
Sun, J., Yang, N., Liu, J., Yang, X., Li, X., and Zhang, L.
Video watermarking scheme based on spatial relation-
ship of dct coefficients.
Swanson, M. D., Bin Zhu, and Tewfik, A. H. (1998).
Multiresolution scene-based video watermarking us-
ing perceptual models. IEEE Journal on Selected Ar-
eas in Communications, 16(4):540–550.
Swanson, M. D., Zhu, B., Chau, B., and Tewfik, A. H.
(1997). Object-based transparent video watermark-
ing. In Proceedings of First Signal Processing Society
Workshop on Multimedia Signal Processing.
Tjokorda Agung B.W, Adiwijaya, and Permana, F. P.
(2012). Medical image watermarking with tamper
detection and recovery using reversible watermarking
with lsb modification and run length encoding (rle)
compression. In 2012 IEEE Int. Conf. on Commu-
nication, Networks and Satellite (ComNetSat).
Tsai, H.-H. and Liu, C.-C. (2011). Wavelet-based image
watermarking with visibility range estimation based
on hvs and neural networks. Pattern Recognition, 44.
Wen, B. and Aydore, S. (2019). ROMark: A Robust Wa-
termarking System Using Adversarial Training. arXiv
e-prints, page arXiv:1910.01221.
Xu, D., Wang, R., and Wang, J. (2011). A novel watermark-
ing scheme for h.264/avc video authentication. Signal
Processing: Image Communication, 26(6):267 – 279.
Zebbiche, K., Khelifi, F., and Loukhaoukha, K. (2018). Ro-
bust additive watermarking in the dtcwt domain based
on perceptual masking. Multimedia Tools and Appli-
cations, 77(16):21281–21304.
Zhang, J., Ho, A. T. S., Qiu, G., and Marziliano, P.
(2007). Robust video watermarking of h.264/avc.
IEEE Transactions on Circuits and Systems II: Ex-
press Briefs, 54(2):205–209.
Droppix: Towards More Realistic Video Fingerprinting
475
Zhong, X. and Shih, F. Y. (2019). A Robust Image Wa-
termarking System Based on Deep Neural Networks.
arXiv e-prints, page arXiv:1908.11331.
Zhou Wang, Bovik, A. C., Sheikh, H. R., and Simoncelli,
E. P. (2004). Image quality assessment: from error
visibility to structural similarity. IEEE Transactions
on Image Processing, 13(4):600–612.
Zhu, J., Kaplan, R., Johnson, J., and Fei-Fei, L. (2018).
HiDDeN: Hiding Data with Deep Networks. In The
European Conference on Computer Vision (ECCV).
SECRYPT 2020 - 17th International Conference on Security and Cryptography
476