Data hiding for multimedia transmission

Among the possible applications of data hiding, the exploitation of a hidden
communication channel for improved transmission, particularly video
transmission, is gaining more and more consensus. Data hiding can be
helpful for video transmission in several ways. From the source coding
point of view, it can help to design more powerful compression schemes
where part of the information is transmitted by hiding it in the coded
bit stream. For instance, chrominance data could be hidden within the
bit stream conveying luminance information. Alternatively, the audio data
could be transmitted by hiding it within the video frame sequence (audio in
video). From a channel coding perspective, data hiding can be exploited to
improve the resilience of the coded bit stream with respect to channel errors
(self-correcting or self-healing images/video). As a matter of fact, redundant
information about the transmitted video could be hidden within the
coded bit-stream and used for video reconstruction in case channel errors
impaired the bit-stream.
2.3.1 Data compression
Traditionally, data hiding and data compression are considered contradictory
operations, in that each of them seems to obstruct the goal of the
other. As a consequence, a great deal of research has been devoted to finding
an appropriate compromise between the two goals, e.g. by designing
a watermarking algorithm which survives data compression. Despite this
apparent contradiction, some authors started investigating the possibility
of improving the effectiveness of existing data compression schemes by encoding
only part of the information and hiding the remaining information
within the coded bit-stream itself. For instance, the possibility of hiding
the chrominance part of an image within luminance information has been
investigated with rather good results, in that for a given compression rate,
a better fidelity to the original image is obtained. Another possibility consists
in hiding the audio signal of a video within the visual part of the video
stream.
From a theoretical point of view, one may wonder whether, in the presence
of ideal, perceptually lossless compression, data hiding is still possible
or not. The answer to this question is not easy at all. First of all the notion
of ideal perceptually lossless compression must be clarified. For example,
we may say that a perceptually lossless compression algorithm is ideal if it
removes all the information contained in the digital asset which can not be
perceived by a human observer. It is readily seen, that the above definition
precludes the possibility that data hiding and ideal compression coexist to
gether, given that the ultimate goal of any data hiding scheme is to modify
the host asset in such a way that it contains a piece of information which,
though present, can not be perceived by the human senses. Unfortunately
(or fortunately, depending on the point of view), the above definition is by
far too heuristic to lead to a mathematical formulation of the coding problem.
When a more practical definition of perceptually lossless ideal coding
is given, the possibility of designing a data hiding scheme which survives
ideal compression remains an open issue, mainly due to the particular nature
of perceptual equality between assets. As a matter of fact, perceptual
equality is not an equality in a strict mathematical sense since the transitive
property is not satisfied, nor the perceptual distance between assets
is a true distance in a mathematical sense, since the triangular inequality
does not hold, thus making the theoretical analysis of data hiding in the
presence of ideal, perceptually lossless compression extremely difficult.
At a more practical level, data hiding can represent a new way to overcome
the imperfections of current compression algorithms, which, far from
being ideal as they are, do not remove all the perceptual redundancy6 contained
within the to-be-compressed asset.
From an application perspective, the most stringent requirement is watermark
capacity, since the larger the capacity the higher the effectiveness
of the source coding algorithm. On the contrary, robustness is not an issue
at all, provided that data hiding is performed in the compressed domain,
or simultaneously to data compression. This is not the case, if data hiding
precedes compression, since in this case the hidden data must survive compression.
Protocol level requirements are rather obvious: blind watermark
detection is required, as well as the adoption of a readable watermarking
scheme.
2.3.2 Error recovery
A problem with the transmission of data in compressed form is the vulnerability
of the coded bit stream to transmission errors. This is the case
with most of the compression standards, including JPEG for still images,
MPEG and H.263 for digital video or MP3 for audio. For example, in
MPEG-2 video, a single bit error can cause a loss of synchronization that
will be visible over an entire group of pictures (GOP). To cope with the
fragility of compressed data, channel coding is usually adopted to enable
error detection or correction. This always corresponds to the introduction
of a controlled amount of redundancy. Redundancy can either be introduced
at the transmission level, by relying on error correcting codes, or at
the application level, i.e. by modifying the syntax of the coded bit stream,
in the attempt to make it more resilient against errors. Though the above
solutions considerably improve the quality of the reconstructed data in the
presence of errors, all of them share two basic drawbacks: i) usually they
are not standard-compliant (even if new standards make provision for error
resilient compression, backward compatibility with previous standard
is often lost); ii) the net available bit-rate decreases to make room for the
redundancy.
A possible alternative consists in performing error detection and concealment
at the decoder side. For instance, in video transmission, temporal
concealment may be applied in the attempt to reconstruct the missed information
from past frames, or the data in the present frame may be used to
reconstruct lost information (spatial concealment). Nevertheless, it is obviously
impossible to exactly recover the original content of a video frame,
e.g. in the presence of occlusions, once the corresponding part of the bit
stream has been lost.
Data hiding represents an alternative approach to the problem: the
redundant information is hidden within the compressed stream and, possibly,
used by the decoder to recover from errors. For instance, a low
quality version of the compressed asset may be transmitted through the
hidden channel to enable the reconstruction of the information that was
lost because of channel errors. In some cases, it is only important to detect
errors, e.g. to ask the retransmission of data, then the hidden data can be
used as in authentication applications, with tampering being replaced by
transmission errors. Note that with data hiding, backward standard compliance
is automatically achieved, since the hidden data is simply ignored
by a decoder which is not designed to exploit it. As to the preservation of
the net bit-rate available for payload transmission, it has to be noted that,
though unperceivable, the watermark always introduces a certain amount
of distortion which decreases the PSNR of the encoded data. Such a loss
in PSNR should be compared to the PSNR loss caused by the reduction
of the net bit-rate consequent to the use of conventional forward error correction
techniques, or to transmission of the redundant information at the
application level.
As for joint source coding and data hiding, even in this case, the actual
possibility of replacing error correcting codes with data hiding (or improving
the capability of error correcting codes via data hiding methodologies)
is not easy to asses from a theoretical point of view. As a matter of fact,
results from rate distortion theory and Shannon's theorem on channel coding
seem to indicate that no improvement has to be expected by using
data hiding for error correction. Nevertheless, real data transmission conditions
are far from the ideal conditions assumed in information theory: thechannel is not AWGN, source and channel coding are not ideal, asymptotic
analysis does not always hold, PSNR is not a correct measure of perceptual
degradation. At a more practical level, then, data hiding is likely to bring
some advantages with respect to conventional error handling techniques.
Even in this case, applications requirements are less demanding than in
copyright protection applications. The most stringent requirement regards
capacity, in that the higher the capacity the larger amount of redundancy
can be transmitted, thus increasing robustness against errors. For example,
the transmission of a low resolution version of a 512 x 512 gray level image
may require the transmission of 4096 pixel values, for a total required
capacity of about 10 Kbit (we assumed that each pixel requires at least
2.5 bits to be coded), which is a rather high value. Conversely, robustness
is not a major concern, even if it is obviously required that the hidden
information survives transmission errors. It is also obvious that the use
of a blind watermarking algorithm is required. The adoption of readable
watermarking is also mandatory, unless the hidden information is only used
to detect errors, without attempting to correct them.
2.4 Annotation watermarks
Despite digital watermarking is usually looked at as a mean to increase
data security (be it related to copyright protection, authentication or reliable
data transmission), the ultimate nature of any data hiding scheme
can be simply regarded as the creation of a side transmission channel, associated
to a piece of work. Interestingly, the capability of the watermark
to survive digital to analog and analog to digital conversion leads to the
possibility of associating the side channel to the work itself, rather than to
a particular digital instantiation of the work. This interpretation of digital
watermarking paves the way for many potential applications, in which
the watermark is simply seen as annotation data, inserted within the host
work to enhance its value. The range of possible applications of annotation
watermarks is a very large one, we will just describe a couple of examples
to give the reader a rough idea of the potentiality of digital watermarking
when this wider perspective is adopted. Note that the requirements
annotation watermarks must satisfy, can not be given without carefully
considering application details. In many cases, watermark capacity is the
most important requirement, however system performance such as speed
or complexity may play a predominant role. As to robustness, the requirements
for annotation watermarks are usually much less stringent that those
raised by security or copyright protection applications.
2.4.1 Labelling for data retrieval
Content-based access to digital archives is receiving more and more attention,
due to the difficulties in accessing the information stored in very large,
possibly distributed, archives. By letting the user specify the work he is
looking for, by roughly describing its content at a semantic level, many
of the difficulties usually encountered during the retrieval process can be
overcome. Unfortunately, it is very difficult for a fully automated retrieval
engine to analyze the data at a semantic level, thus virtually all contentbased
retrieval systems developed so far fail to provide a true access to the
content of the database. A possibility to get around this problem consists
in attaching to each work a description of its semantic content. Of course,
producing a label describing the semantic content of each piece of work is
a very time consuming operation, thus it is essential that such a label is
indissolubly tied to the object it refers to, regardless of the object format,
and its analog or digital nature. In this context, digital watermarking may
provide a way whereby the labelling information is indissolubly tied to the
host work, regardless of the format used to record it. When the work moves
from an archive to a new one, possibly passing from the analog domain,
the information describing the content of the work travels with the work
itself, thus avoiding information loss due to format modification. To exemplify
the advantages of data hiding with respect to conventional data
labelling, let us consider the archival of video sequences in MPEG-4 format.
An annotation watermark could be hidden within each video object
forming the MPEG-4 stream. For instance, the name of an actor could
be hidden within the corresponding video object. If the marked object is
copy-edited to create a different video sequence, the hidden label is automatically
copied with the object thus avoiding the necessity of labelling it
again. Similarly, if the object is pasted to a new video after going in the
analog and back to the digital domain, the annotation watermark is not
lost, thus making the semantic labelling of the new video easier.
2.4.2 Bridging the gap between analog and digital objects
A clever way to exploit the side communication channel made available by
digital watermarking, consists in linking any analog piece of work to the
digital world. The smart image concept, derived by the MediaBridge system
developed by Digimarc Corporation, is an example of such a vision of
digital watermarking. According to the smart image paradigm, the value
of any image is augmented by embedding within it a piece of information
that can be used to link the image to additional information stored on the
Internet. For example, such an information can be used to link a picture on
a newspaper to a web page further exploring the subject of the article the
image appears in. The actual link to the Internet is activated by showing
the printed picture to a video camera connected to a PC; upon watermark
extraction the URL of the web site with the pertinent information is retrieved
and the connection established. More generally, the information
hidden within the piece of work is dormant until a suitable software reads
it, then it may be used to control the software which retrieved the watermark,
to link the object to additional information, to indicate the user how
to get additional services, or to provide the user with a secret information
to be used only upon the payment of a fee. Watermark retrieval itself,
may be conditioned to the payment of a fee, thus providing a conditional
access mechanism that can be exploited in commercial applications, e.g.
bonus programme applications, where the gathering of a certain number of
watermarks is the access key to a discount programme.
2.5 Covert communications
Covert communication is the most ancient application of data hiding, since
it traces backs at least to the ancient Greeks, when the art of keeping a message
secret was used for military applications. Indeed, it is often invoked
that the first example of covert communication is narrated by Herodotus,
who tells the story of a message tattooed on .the shaved head of a slave:
the slave was sent through the enemy's lines after his hair was grown again,
thus fooling the enemy. Even if it is likely that the history of covert communication
started well before Herodotus' time, the art of keeping a message
secret is called steganography, from the Greek words are^ai/o^ (covered)
and ^paipeiv (writing). As opposed to cryptography, the ultimate goal of
a covert communication scheme is to hide the very existence of the hidden
message. In this case, the most important requirement is the imperceptibility
requirement, where imperceptibility assumes a wider sense, in that
it is essential that the presence of the message can not be revealed by any
means, e.g. through statistical analysis. In steganography, the most important
requirement after security (undetectability) is capacity, even if it
is obvious that the less information is embedded into the carrier signal, the
lower the probability of introducing detectable artifacts during the embedding
process.
A covert communication scheme is often modelled by considering the
case of a prisoner who wants to communicate with a party outside the
prison. To avoid any illegal communication, the warden inspects all the
messages sent by the prisoner and punishes him every time he discovers
that a secret message was hidden within the cover message (even if he is
not able to understand the meaning of the hidden message). Once casted in
a statistical framework, the prisoner problem can be analyzed by using tools
derived from information theory, and the possibility of always establishing
a secure covert channel demonstrated. The capacity of the covert channel
can also be calculated. It is important to point out that according to
the prisoner and the warden model, the prisoner is free to design the host
message so to facilitate the transmission of the hidden message, a condition
which does not hold in many practical applications where the sender is not
allowed to choose the host message.
Despite its ancient origin, and although a great deal of research has
been carried out aiming at designing robust watermarking techniques, very
little attention has been paid to analyzing or evaluating the effectiveness of
such techniques for steganographic applications. Instead, most of the work
developed so far has focused on analyzing watermarking algorithms with
respect to their robustness against various kinds of attacks attempting to
remove or destroy the watermark. However, if digital watermarks are to be
used in steganography applications, the detectability of watermark presence
must be investigated carefully, since detection by an unauthorized agent
would defeat the ultimate purpose of the covert communication channel.
2.6 Further reading
The necessity of considering buyer's rights in addition to those of the seller
in fingerprinting-based copy-protection systems was first pointed out by
[186]. Such a problem was lately analyzed by N. Memon and P. W. Wong
in [149], where they first introduced the IBS copy protection protocol.
The DVD copy protection protocol we briefly discussed in section 2.1.3,
is part of a complex system devised by an international pool of consumer
electronics companies, to protect the digital distribution of copyrighted
video. More details about such a system may be found in [32, 141].
An early formalization of data authentication relying on (semi-)fragile
watermarking may be found in [126], whereas for a detailed list of requirements
fragile authentication-oriented watermarking must satisfy the reader
is referred to [79].
Authentication of video surveillance data through digital watermarking
is thoroughly discussed in [23]. In the same paper, the general mathematical
framework for data authentication discussed in section 2.2.2 was first
introduced.
The notion of compressive data hiding, i.e. the possibility of exploiting
data hiding technology to improve coding efficiency, was formalized by
P. Campisi, D. Kundur, D. Hatzinakos and A. Neri in [35], where the
advantages obtained by hiding the chrominance components of an image
within the luminance bit stream are shown. Such a concept, though, was
already present in earlier works in which the possibility of hiding the audio
component of a video within its visual component was advanced [209].
The possibility of exploiting data hiding to improve the reliability of
multimedia transmission in the presence of errors has been explored by
several researchers. For some practical examples illustrating the potentiality
of such a strategy, readers may refer to [17, 80, 188, 203].
The potentialities of annotation watermarking are still largely unexplored,
partly because research was mainly focused on security-oriented
applications, partly because for this kind of application watermarking just
represents an additional way of solving problems which could be addressed
through different technologies as well. Readers interested in this particular
kind of application may refer to [68] where a survey of possible applications
of annotation watermarks is given, and [65, 198], where the smart image
concept is illustrated.
An insightful mathematical formalization of the covert communication
problem may be found in the seminal work by C. E. Shannon [197]. For
a good survey of covert communication through digital watermarking, the
reader.