Multimedia Watermarking
Techniques
Frank Hartung and Martin Kutter
ECE 738
In Young Chung
Outline
1.
2.
3.
4.
Terminology
Requirement
Basic Watermarking Principles
Watermarking techniques for
1. Text Document
2. Image
3. Video
4. Audio
5.other media
5.
Conclusion
Terminology

Watermarking

Data hiding and data embedding

Fingerprinting and Labeling

Embedded signatures

Visible watermarks
techniques that allow secret communication, usually by embedding or hiding
the secret information
applications where the existence of the embedded data are publicly known, but
there is no need to protect it
special application of watermarking related to copyright protection
stand for watermark in early publication
visible pattern, like logos, inserted into images or video
Watermarking Requirements
1)
2)
3)
As much information as possible
 Capacity
Only be accessible by authorized parties
by means of cryptographic key
 security
Resist against hostile attacks
 Robusteness
Invisibility
Imperceptibility
Robustness
Capacity
4)
Imperceptibility
Supplemental W. Requirements
1)
2)
Watermark Recovery may or may not
allowed to use the original data.
Real time watermarking requirements
e.g., video fingerprinting (We will see this
later)
Design Issues





Watermark Security and Keys
Robustness
Imperceptability
Capacity
Watermark Detection
Robustness
Capacity
Imperceptibility
Basic Watermarking Principles

Three issues in the design of a watermarking system
1) Design of the Watermark signal
W  f0 (I , K )
or W  f 0 ( I , K , X )
2) Design of the embedding method
Y  f1 ( X ,W )
3) Design of the corresponding extraction method
^
I  g( X ,Y , K )
•
•
^
or
I  g (Y , K )
Where, W: watermark signal, I: watermark information.
K: key, X: host data, Y: watermarked signal
Basic Watermarking Principles

The public or secret key is used to enforce security

Many watermarking schemes use spread-spectrum methods
they add a PN signal with low amplitude to the host data.

Correlator is used for watermark detection
Watermarking Techniques
1. Text Document Watermarking
Text Document Watermarking
Two methods to hide information
1)
in the semantics; in the meaning and ordering
or the words
2)
in the format
* In the layout and the appearance
- Example of word shifting coding -
Text Document Watermarking;
Three coding methods
1)
2)
Line shift coding
*Assumption; lines are uniformly spaced
 doesn’t need original for watermark extraction
word-shift coding
*Assumption; space between words are usually variable
 needs original for Watermark extraction
3)
feature coding; slightly modifies the features

Goal: making watermark removal more expensive than obtaining
the right to copy from the copy right owner
Watermarking Techniques
2. Image Watermarking
Why is Image watermarking so important?
1) There is a large demand and productions
2) Most of watermarking research
 much more than video or audio watermark
Common ideas for Image
watermarking

A lot of watermarking methods are very similar and differ
only in parts

Three topics 1) watermark signal design
2) embedding
3) recovery (Detection)
I.
Watermark signal design
The watermark signal; typically a pseudorandom signal with
low amplitude
e.g., Gaussian, uniform, or bipolar pdf
* The watermark signal: often designed in spatial domain,
sometimes in DCT or block-wise DCT
Common ideas of Image
watermarking
ii.
Signal embedding (where?)
* embed watermark signal mostly to the
luminance channel alone
* sometimes in color channels
* in the spatial domain
* or in the DCT,DFT and DWT
(full-image DCT or block wise DCT domain)
advantage in terms of visibility and security
Common ideas of Image
watermarking

Argue?
about embedding domain
 Low, medium or high freq.?

For maximum robustness
 embed watermark signal adaptively
where the host data populate (typically the
low frequency)
Common ideas of Image
watermarking
iii.
Recovery (Detection)
* Usually done by correlation method; a
correlation receiver or a matched filter
Acknowledgements

The Following slides organized by Author
name

The author of this paper put stress on
embedding domain/methods So, we will
mainly deal with embedding domain/method of
numerous watermarking methods
Tirkel

Publication: “Electrical Water Mark” in 1993

Proposal
m-sequence PN code embedding in LSB plane

* To gain full access to the LSB plane without much distortion
 compress original image to 7 bits through histogram manipulation

the decoding process use the unique and
optimal autocorrelation of m-sequences
Matsui and Tanaka

Publication: “Video Steganography: How to secretly Embed
a Signature in a picture”

Proposal
Predictive code schemes using key table
* exploit correlation between adjacent pixels by coding the
prediction error instead of coding the individual gray scale
value

 i  X i  X i 1

And embed a watermark in forms of a binary string
Smith

New approach

Digital watermarking and digital
modulation (especially, direct sequence
spread spectrum modulation) share
similar concepts

More in depth analysis of 2-D amplitude
modulation was given by Hernandez
Bender


I.
Proposal
2 Methods
Patchwork
1. randomly selected pairs of pixels (ai , bi ) Are
used to hide 1 bit by increasing the ai’s by one
and decreasing the bi’s by one
2. The expected value of the sum of N pixel
pairs
ContII.
Texture Block Coding
1. Watermark is embedded by copying one
image texture block to another area in the
image with similar texture
2. Recovery- autocorrelation
* Remarkable point- high robustness to any
kind of distortion, since both image area
distorted in a similar way autocorrelation still
works
Pitas and Kaskalis

Proposal

Signature casting on digital images
- based on same basic idea as the patchwork
1. The watermark S  {sm,n } is the same size as the original
image (here, # of ones = # of zero)
2. The original image is divided into two sets A and B of
equal size
I : Original Image , xmn : Luminance value
Contn3. The watermark is superimposed by changing
the elements of the subset A by positive factor k
A'  {xmn  k , xmn  A}
*k is positive integer
4. The watermarked image is given by the union of
A’ and B
Langelaar


Proposal
Block base spatial watermarking
- Improved version of previous method
1.
2.
3.
4.
5.
The image is tiled into square blocks (8x8)
Each block is selected pseudorandomly
To embed “1”, k x P added to the block
To embed “0”, k x P subtracted to the block
Each selected block has a PN pattern P
* k: scaling factor, P : PN pattern
Bruyndonckx


Proposal
Watermarking with the use of pixel classification
Purpose- increase the performance of the block base spatial
watermarking methods
1. Select blocks (PN) & classify the block based on three types of
contrast between zones; hard, progressive and noise contrast
2. Each zone subdivided into two categories A and B based on
gird defined by the coder
3. Each pixel is assigned to one of four zone/category
combinations e.g., 1/A,…2/B
Cont4. A bit b is embedded by modifying the zone/category means
to satisfy the following constraints
m1*A ...m2* B: the modified zone/category mean values
S: the watermark embedding strength
5. the modification of the mean values is done by applying equal
luminance variations for all pixels belonging to the same zone
7. To increase robustness the authors suggest to perform
redundant bit embedding and use error correcting codes
6. Good robustness to JPEG compression is reported
Kutter

Proposal
Improved spread-spectrum watermarking in the spatial
domain
* Exclusively works with the blue image component, in the
RGB color space maximize the watermarking strength and
minimize visual artifacts
* preprocess the image prior to watermark decoding
increased robustness, applicable to any spread-spectrum
spatial domain watermarking
Cont1. A single bit b is embedded at a
pseudorandomly selected location (I,j) by either
adding or subtracting (Amplitude modulation)
Where Bi, j describes the blue value at location (I,j)
Li , j :the Luminance at the same location
 : the embedding strength
Cont2. To recover an embedded bit, an estimate of original value is
computed
c : the size of the cross-shaped neighborhood
3. The bit value is determined by looking at the sign of the difference
^
 i , j  Bi , j  Bi , j
*To increase robustness, (Author suggest) each signature bit is
embedded several times and to extract, the sum of all differences  i, j is
used
Macq

Proposal
Watermarking adapted to the HVS using masking and
modulation
1. The watermark in spatial domain is low-pass filtered,
frequency modulated, masked and then added to the host
image
2. A secret key is used to determine the modulation freq.
3. Masking uses an extension of the masking phenomena
for monochromatic signals, called gratings

Recovery- demodulation followed by a correlation function
Voyatzis and Pitas

Proposal
Watermarking by inserting logo like patterns using torus
automorphism
* A 2-D torus automorphism: a kind of spatial
transformation
* It is defined by
* Iterated action of A on a point
system expressed by
 0 form a dynamical
Cont* This system mixes the point in chaotic way.
* Under certain circumstances, the
automorphisms may have periodicity
1-iteration

2-iteration
10-iteration
T-iteration
…
Cont
How/where to embed?
1. Watermark is mixed using the automophism
AN
2. then, overlaid on a selected block in the original image
e.g., LSB

Recovery- extracting the mixed watermark
then, reconstruct the watermark using the automorphism AT  N
* Where, T is the automorphism period.
Raymond and Wolfgang


Proposal
Watermarking technique to verify image
authenticity based on an approach similar to the
m-sequence approach
1. A random sequence generated is mapped
from {0,1} to {-1,1}, arranged into a suitable
block and added to image

Recovery – overlays the watermarked block with
the watermark block and compute inner product
and compares the result with ideal value
Cont-


The test statistics

If the watermark is unchanged ,
=0
When  is greater than a defined tolerance, the
block fails the watermark test

is defined as

Chen and Wornell


Proposal
Quantized Index Watermark (QIM)
*Please recall even-odd embedding and predictive coding
scheme using table
*Not based on spread-spectrum modulation but,quantization
modulation
*Based on a set of N-dimensional quantizers.
*Quantizers – satisfy a distortion constraint
- each reconstruction values from one
quantizer are ‘’far away’’ from the others
1. The message to be transmitted is used as in index for quantizer
section
2. Selected quantizer is used to embed information
3. Spatial or DCT domain used
Koch


Proposal
Efficient watermarking in DCT domain for JPEG (first
introduction)
1. The image is divided into square blocks of size 8x8 for
which the DCT is computed
2. From a pseudoramdomly selected blocks, a pair of
midfrequency coefficients are selected from 12 predetermined
pairs
3. To embed a bit- the coefficient are modified such that the
difference between them (a pair of coefficient) is positive or
negative
* Good robustness – JPEG (Q=50%)
Swanson


Proposal
DCT domain watermarking technique, Based on frequency
masking of DCT blocks
1. Input image is split up in to square matrix and DCT is
computed
2. Frequency mask is computed based on the knowledge
that a masking grating raises the visual threshold for signal
gratings around the mask freq.
3. The resulting perceptual mask is scaled and multiplied by
PN(DCT)-sequence
4. This watermark is added to the corresponding DCT block
Podilchuk


Proposal
Perceptual watermarking using the just noticeable difference
(JND) to determine an image-dependent watermark modulation
mask
I u ,v ; the transform coefficients of the original image
wu ,v ; the computed JND based on visual models
JNDu ,v
; the watermark values

Recovery- based on the correlation
* Robust to JPEG compression, cropping, scaling and additive
noise
Boland and Cox

Proposal
Frequency-domain watermarking (First)

perceptual adaptive methods which is based on modulation

1. Generate the watermark with statistical distribution ; e.g.,
N(0,1)
2. The watermark is inserted into the image
Cont* ; determine the strength
* The watermark embedded 1000 strongest DCT
coefficients

Detection- given by the normalized correlation
coefficient
* Boland propose a similar techniques – DCT,
DWT, Walsh-Hadamard, FFT
Summary







Several different image watermarking methods
Most watermarking methods are based on the same basic
principle
- small, pseudorandom changes are applied to selected
region; spatial or transform domain
Recovery – correlation-like similarity measures.
Usually, the number of modified coefficient is much larger
than the number of bits to be encoded
Embedding domain have a influence on the robustness
 spatial – less robust to noise like attack / E.g.- JPEG
- more robust to cropping, translating.
 freq. – less robust to cropping which destroy the
embedding water mark (DCT,DFT,DWT)
Watermarking Techniques
3. Video Watermarking
Common idea for video watermarking

Video sequences consists of a series of consecutive and
equally time-spaced still images
 in general, very similar with image watermarking so,
image watermark method is applicable to video directly

Important differences
* available signal space;
for image; very limited
for video; much larger signal space (# of pixels)
*video watermark imposes real or near real-time
watermarking system
 complexity issue is much more important
Cont
The structure of video as a sequence of images give rise to
particular attacks
 frame averaging, frame dropping and frame swapping
(Only in video)

Two competing requirements
*A good watermarking scheme
1. may recover the full watermark from a short part of the
sequence
2. distribute watermark information over several consecutive
frame to have robustness against frame dropping
 depends on application

Compressed/ uncompressed video
Hartung and Girod


Proposal
Watermarking of compressed video for fingerprint
application
* Used spread spectrum approach and added an additive
watermark into video
* The watermark is generated using a PN signal with the
same dimension as the video signal
*Each information bit is redundantly embedded into many
pixels
1. For each compressed video fame, 8x8 DCT transformed
watermark signal is added to DCT coefficient of the video
Cont*This method done for I,P, and B frames
* Rate control- by comparing each encoded watermarked
DCT coefficient versus the corresponding encoded
unwatermarked coefficient
(because video steam use variable length coding, the
watermarked signal may or may not need more bits
encoding than the unwatermarked one)
 If more bits required, the coefficient is not used for
embedding

Recovery – correlation using the same PN sequence used
for generation of watermark
Jordan


Proposal
Watermarking of compressed video that embed information in
motion vector of motion -compensated prediction schemes.
1. Motion vectors slightly modified in pseudorandom way

Watermark embedding/detection available as long as the video
is in compressed format

After decompression?
 the watermark still be recovered by recompressing and
detecting

Artifacts? Because the blocks pointed to by the original and the
modified vector are very similar no visible artifacts
Hsu and Wu


Proposal
Watermarking for compressed video using middle-freq DCT
* Extension of their image watermarking method
1. modifies middle freq DCT coefficients in relation to spatially
(I-frames) or temporally ( for P- and B- blocks) neighboring
block
* Prior to embedding, the watermark signal is spatially
scrambled to make it robust to cropping

Drawback- for watermark extraction, original video and
watermark should be known
Langelaar


Proposal
2 methods
1) data hiding method
1. adds the label directly in the MPEG bit stream by
replacing variable length codes (VLC) of DCT coefficient
* In MPEG-2 code tables there are pairs of code which
represent the same run and levels that deviate only by one
from each other
We can choose one as bit “1” the other as “0”

Drawback- the label can be easily removed by
decompression and recompression
Cont2) watermarking
1. for each bit to be embedded, a set of n 8x8-block is
pseudorandomly taken from the video frame
(here, n typically 16~64)
2. Seudorandomly divide into two subsets of equal size.
3. For each subset, the energy of the high-freq. DCT
coefficient is measured
4. In order to embed the bit, the energy of the high-freq
coefficient in one or the other subset is reduced by
removing high freq. coefficients
Cont-
Block diagram of watermark embedding into DCT coefficients of compressed video

Recovery - 1. select the same set of blocks
2. divide it into the same subsets
3. Compare the energy of the high-freq.
coefficient in each of the subsets
* Here, We can use the secret key for the selection of blocks
Cont
Drawbacks
1. robustness is limited
2. Re-encoding increases the error rate of the embedded
signal much
Swanson


Proposal
Multiscale watermarking method working on uncompressed
video
*Same scheme as Image watermarking
1. video sequence is segmented into scenes
2. temporal wavelet transform is applied to each scenes,
yielding temporal low-pass and high-pass frames
3. the watermark is embedded into each of the temporal
components (here, low-freq.)
4.Inverse transform the watermarked components to get the
watermarked video
Cont
Interesting properties
The watermark has some components that change over time,
since they are embedded in low frequency coefficient
 this allow robustness against frame averaging, frame
dropping

Drawback- very high complexity
Linnartz


Proposal
Embed information in the GOP structure of the
MPEG-2 compressed video
*There is a maximum distance between two
successive I-frames
*The frame type signaled in the frame header
and can be switched randomly from frame to
frame.
Cont* A typical GOP in display order (N=12)
* If I-frame is fixed we have 2048 variations
*However, most available video codecs never use most of the admissible
GOP structure
 purposely use irregular GOP structure, that are very unlikely, to embed
information

Drawback- use this method only during compression
- decompress and recompression will remove this information
completely

Merit - Low complexity
Darmstaedter


Proposal
Embed a spatial-domain low-pass spread spectrum
watermark into 8x8 pixel blocks of video sequences
1. The blocks are classified according to their activity
*Low activity are not watermarked
2. Low-pass pseudorandom pattern is added to each block
* Each block conveys one bit watermark information
Cont* The block repeated over several blocks and
several frames for robustness

Recovery-done in spatial domain after
decompression using correlation concept with
threshold
Deguillaume


Proposal
Spread spectrum watermark into 3-D
block of video using 3-D DFT
* Embed a spread spectrum watermark
into 3-D blocks of video by employing a
3-D DFT and add to the transform
coefficients
Busch


Proposal
Apply a still-image watermarking method
working on DCT blocks to video sequences
1. The watermarks are embedded into the
luminance component of umcompressed video
and retrieved after decompression
*In order to improve the invisibility of the
watermarks, especially at edge, blocks are
selected depend on activity (high-activity block
selected)
Cont
Recovery- 0~50% error rate are reported
depend on the sequence
the author propose to embed the watermark
into several consecutive frames
(over 50frame a few percent error rate
reported)
Kalker


Proposal
Video Watermark for video broadcasting
monitoring application
*Called JAWS (just another watermarking system)
* For low complexity, both watermarking and
detection are performed in the spatial domain
watermarking before compression
detection after decompression
Cont* The watermark size is 128x128 and repeated (tiled) to fill
whole video frame
* To avoid visible artifacts, the watermark is, on a pixel-bypixel basis, scaled with scale factor
* Scale factor is derived from an activity measure
* Activity measure is computed using a Laplacian high-pass
filter

Detection- correlation detector is used
- in case of presence of spatial shift, a search
over all possible shift is performed
(128x128 position)
Summary

The proposed methods span a wide complexity
range from low complexity to considerable
complexity
 In general, the more complex methods
provide higher robustness

Most methods operate on uncompressed video;
a few methods embed watermark directly into
compressed video
e.g., DCT embedding or the motion vector
embedding, GOP structure embedding
Watermarking Techniques
3. Audio Watermarking
Common ideas for audio
watermarking

Compared to image and video, audio
signal has much less samples per time
interval
 amount of information which can be
embedded is much lower

HAS (Human Audible System) is much
more sensitive than the HVS
Boney


Propose
A spread spectrum approach
* PN sequence is filtered in several stage to
exploit masking effects of the HAS
* The watermark is low-pass filtered by using
full audio compression/decompression scheme
to guarantee that it survive audio compression
Tilki and Beex


Propose
interactive television application where they embed information into
the audio components of a television signal
1. The information to be embedded is partitioned in blocks of 35bits
2. Each information bit is modulated using a sinusoidal carrier of a
specific frequency with low amplitude and added to audio signal
 if the sinusoidal carrier for a specific bit is present=> “1”
 otherwise => “0”
* The frequencies of the sinusoidal carrier are above 2.4 kHz
 To reduce interference from the audio signal , the audio signal is
attenuated at frequencies above 2.4 kHz
Bender


Propose
phase coding
; use the phase information as a data space
1. For encoding, a Fourier Transform is applied and the
phase value of each freq. component are lined up as matrix
2. Binary information can be added into this matrix by
modifying the phase component.
* HAS is not very sensitive to the phase distortion of the
sound
Watermarking Techniques
4. other Multimedia DATA
Ohbuchi


Propose
embed visible and invisible watermarks into 3-D polygonal
models
* This model comprise primitives like points, lines, ploygons,
and ployhedrons
* Modify geometry or topology for watermarking

2 methods
1) pseudorandomly selects sets of four adjacent triangles
* embed information by displacing the vertices of the four
triangles
 up to 1% of the shortest edge of the rectangular
bounding box of the entire 3-D model
Cont
2) pseudorandomly selects tetrahedron from the
mesh and embeds information in the volume
ratio of consecutive tetrahedron by modification
of the vertices
Hartung

Proposal

A spread-spectrum method for watermarking of MPEG-4 FAP’s
*MPEG-4 features model-based animation 3-D head modules using
FAP (Facial Animation Parameter)
*There are FAP like “rotate head”, “open mouth” or “raise right cornerlip”
1. In order to embed information, the parameters first have to be
estimated from the sequence
2. The watermarks are embedded into the animation parameter
*Adaptive amplitude attenuation prevent visible distortion of animated
head level.
Cont* Interesting point is that the watermark is not
embedded in pixels but in the semantics( the
way the head and face move)
Conclusion





We reviewed the most important aspects
 design requirements, system issues, and
techniques for digital watermarking
We have elaborated on numerous watermarking
techniques for still images, video, audio, text,
and other multimedia data.
Majority of techniques are similar and based on
modulation with a PN signal.
Hypothesis test using correlation is used in the
watermark recovery
Although working systems are available,
research has to continue
Thanks for your attention
- Questions? -