10 April: fyi -- research, Rennes

Index of April 2008 | Index of year: 2008 | Full index



Robustness against desynchronization of watermarking signals using
universal decoding and their connections with watermarking security

Digital watermarking aims at embedding an imperceptible watermark in
a multimedia signal, in order to identify, e.g., its owner, or its
legal user. During the past ten years, a huge effort has been done to
design robust schemes especially against geometrical attacks. De-
synchronization has been one of the most challenging issues in
watermarking since its inception, and still remains as a hard problem
lacking of "universal" solutions. For the moment, only ad-hoc
countermeasures have been proposed:
- Embedding of pilot signals in order to synchronize the decoder,
- watermarking of invariant coefficients,
- pre-registration before embedding and decoding.
These countermeasures perform correctly from a robustness point of
view, even if theoretical works show their non-optimality. Yet, most
importantly, they introduce security flaws in watermarking schemes.
Some pilot signals are constant reference signals, easily estimated
by an average process. Others have a periodic structure, which is
easily removable. Invariant features have the inconvenient to be very
few which prevents the spreading of the watermark over long
sequences. This low dimensionality reduces de facto the size of the
secret key. Registration is a public process, and the attacker can
also apply it or delude it.

Recently, the attention has moved towards the paradigm of universal
decoding. For instance, recent works by Moulin consider the attacking
channel as a cascade of a fixed memoryless channel (e.g. a Gaussian
channel) and an invertible mapping representing the geometrical
attack, which is parameterized by P. Roughly speaking, a decoder is
said to be universal if it performs as well as a decoder that knows
the parameter P. In some cases, universal decoders with affordable
complexity can be shown to exist.
Limitations of the present analysis are:
- The analysis is made assuming that P takes values in a discrete
set, that grows sub-exponentially on the dimensionality of the
watermarked signals (hence, the case of continuous-valued P is not
considered).
- The decoder does not exploit the fact that the attacking channel
may remain the same for several observations (in Digital
Communications, this is usually termed as the "coherence time" of the
channel).

We initiated several years ago a study of the security of
watermarking schemes [1], which leaded to the more recent paper [2],
and we propose to overcome these limitations through an
identification of the security purposes and the de-synchronization
problem. More precisely, the formulation of the universal decoding
problem will first be stated for a very simple setup dealing with
high dimension vectors: the memoryless channel is simply regarded as
the identity function, and the geometrical attack consists in a
random cycling shift (by the vector P) of the watermarked signal.
From previous theoretical analysis, we know that for this particular
channel, even for continuous-valued P, correct decoding of the
embedded message is possible whenever the number of observations is
sufficiently large. In other words, this geometrical attack has, in
theory, no impact on the capacity of the hidden channel.

The goal of the postdoctoral is to invent more practical designs of
universal decoding. Then, the explicit analysis of their loss in
decoding performance (under this simple geometrical attack) and of
the security levels will be carried out. This will allow us to extend
the notion of watermarking security to a more general setup. This
would give not a definitive answer but at least some arguments about
this very basic but yet unsolved question: Is there an inherent trade-
off between robustness and security?

References:
[1] Watermarking Security: Theory and Practice, F. Cayre, C.
Fontaine, T. Furon. IEEE Transactions on Signal Processing, Volume
53, Number 10, pp 3976-3987 (Oct 2005), special issue "Supplement on
Secure Media III".
[2] Exploiting security holes in lattice data hiding, L. Perez-
Freire, F. Perez-Gonzalez. In Information Hiding, IH'07, Lecture
Notes in Computer Science, Saint-Malo, France, 11-13 June 2007.
Springer-Verlag.

Contact:
Caroline Fontaine http://www.irisa.fr/temics/staff/fontaine/
Teddy Furon http://www.irisa.fr/temics/staff/furon/

------------------------------------------------------------------------
------
Institute:
INRIA Rennes http://www.irisa.fr/

Team-project:
ASPI http://www.irisa.fr/aspi/
TEMICS http://www.irisa.fr/temics/

Duration: 12 months

Subject:
Rare event analysis for digital watermarking

This posdoc offer is within the Nebbiano project ( http://anr-
nebbiano.irisa.fr ) funded by the ANR (National French Reserch
Agency). The proposed subject is about probabilistic approaches to
watermarking. Watermarking is a technology which embeds in a digital
file a key, or message, almost undetectable by others. The
robustness of such embeddings is much discussed in the related
literature, but the reliability issue is much less addressed. By
reliability here we mean, in the context of Digital Rights
Managements (DRM), the probability of accusing an innocent user of
being the source of an unauthorized copy of the protected content.

As far as we know, almost nothing concerning estimation of low
probability of errors has been done by the watermarking community
although there exist better methods than the simple Monte-Carlo
simulations. They have been successfully applied, for instance, to
estimate frequencies of packet losses in digital telecommunication,
for a long time now. The proposed work is to study both theoretically
and by simulation the reliability issues about these watermarking
schemes. Depending on the applicant, the work can be more oriented on
simulation or theoretical results.

The applicant should hold a PhD at the date of beginning. To apply
send your application (as soon as possible) including a CV, a list of
publications and a statement of research to one of the contacts below.

Contact:
Frédéric Cérou http://www.irisa.fr/aspi/fcerou/
Teddy Furon http://www.irisa.fr/temics/staff/furon/


Index of April 2008 | Index of year: 2008 | Full index