Chapter: Capita Selecta
Fingerprinting and Watermarks
Electronic watermarking is a new research area, combining aspects of
digital signal processing, cryptography, statistical communication theory
and human perception.
It aims at embedding additional data into clear content (images, audio etc)
in a way that is difficult to remove.
Principal applications of electronic watermarks are in
copyright enforcement, automatic metering of asset usage in multi-media applications,
piracy tracing, and in
providing additional information, such as image captions.
New multi-media networks and services facilitate the distribution of content,
but at the same time make copying and copyright piracy simple.
Here we see a clear need to embed copyright data, such as the ownership or
the identity of the authorized user in an indelable way.
In case of a legal dispute
over copy rights, this provides a proof of the origin of the material,
and it can provide a tool to formally establish which
user received and illegally re-disseminated the material.
This need particularly exists
for audio, video (frames) and still images in broadcast multimedia networks.
Similar requirements also exist for executable computer programs
or information distribution through packaged (storage) media.
One application can be to to protect and enforce film
copyright infringements, for instance with Digital Video Disc.
Some abiguity exists over terms used.
In particular, we distinguish between watermarking,
fingerprinting, posing a signature or label
and transmitting over a subliminal channel
By watermarking we mean the process of electronically attaching the
identity (secret or clear, analogue or digital) of the owner of the copyright
in a way that is difficult to erase. Watermarking is comparable to placing an
electronic stamp on the document. In some cases, the word watermark is used
exclusively for a hidden identity code.
Watermarking differs from authentication or digital signature
that proves to a receiver that the message could only have come from one particular
transmitter. Mostly, authentication messages in the form of conventional hash
functions can easily be deleted by a pirate who wishes to use copyrighted material
for illegal purposes. The goal is to give the copyright owner of a digital image
(or other piece of information) the possibility to attest technically the origin
of the image. Watermarking does not address authentication explicitly, although
certain forms of watermarks can be used for authentication as well.
By fingerprinting we mean the process of attaching the identity of
(crypto-) decoder to a signal in a way that is difficult to erase. This allows
the copyright owner to trace pirates (or at least their decoder) if the clear
signal is re-disseminated illegally.
The main difference between watermarks and fingerprints is whether the identity
of the transmitter or that of the recipient is embedded in the signal. In
many technical solutions, this also determines where (thus, in which device)
the identifier is being added to the clear signal.
The requirements for fingerprinting are contradictory. On one hand the broadcaster
or copyright holder may want to easily recognize the fingerprint, preferably
visually. This allows easy tracing of a decoder that is used for illegal purposes.
This approach is very similar to the commonly used watermarking by means of
the logo of a TV station that is continuously visible in one of the corners
of the screen. On the other hand, the fingerprint should be hidden, in order
not to disturb paying viewers with program-unrelated messages on their screen,
or to avoid that any pirate can detect and erase the fingerprint electronically.
In the latter case, it may involve specific equipment to detect and decode
PC Windows tool to add a pseudo-random noise sequence to an image and to detect
this watermark by correlation.
Watermarking as a mechanism for copy control
Watermarking is regarded as a useful tool for copy control, e.g. for (DVD or
DVB) video or audio A
watermark preferably is undeletable by a hacker, perceptually and statistically
invisible and resistant to addition of random noise. With statistically invisible
is meant that the watermark cannot be deleted by statistical analysis (like
Kalman filtering) and lossy compression algorithms (like JPEG and MPEG). Another
name for watermarking is "Tattooing". Watermarks can represent registration
numbers of the copyright holders, terminal numbers of the machine which was
used to duplicate the image, registration number of legal users.
Embedding and extracting watermarks requires additional hardware, for example,
a watermark based on masking will require image edges extraction, a watermark
hidden in the frequency domain will require many fast Fourier transforms.
Another problem is the prove of the legal validity of the watermark in case
of court pursuits. The energy required to prove the legitimacy of the watermark
is directly related to the watermark quality.
For example, a watermark can be embedded in a picture by using the following
method. Bits, which represent registration numbers etc., can be hidden in
the picture by modificating the luminance values inside blocks of nxn pixels.
Each block can contain for example one bit of the additional information.
Other methods modify the DCT coefficients generated by JPEG or MPEG coders.
A watermark can also be hidden in audio signals.
link to MP2 files
|Audio Engineering Society AES workshop
on copy management, May 1998|
Audio clips by
- Paul Jessop, Technical Director of International Federation of
- Marcus Erne, ETH Zurich, on watermarking and
- Don Jerrard on legal issues.
hiding in audio signals provides a special challenge because the human auditory
system is extremely sensitive. The pertibations in a sound file can be detected
as low as one part in ten million (-80dB). Although the limit of perceptible
noise increases as the noise contents of the host audio signal increases,
the typical allowable noise level is very low. The human auditory system has
very low sensitivity to the phase of the sound. Unfortunately, this hole has
been exploited by numerous audio compression algorithms. According to experiments,
8 bits of hidden data per second can be added to host sounds with quiet backgrounds.
Twice as much data can be added to host sounds with noisy background by using
Data can also be hidden in digital document images, for example by using
word and line shifting. Each document recipient (i.e. a subscriber) receives
a document containing a unique set of fingerprint marks. Each mark corresponds
to an imperceptible horizontal displacement of a textual object. Since the
information is not observable upon casual inspection of a document, a recipient
may not be aware of its presence. A recovered, unauthorized document copy
can be traced to the original, authorized recipient.