Robust Motion Watermarking
based on
Multiresolution Analysis
Tae-hoon Kim
Jehee Lee
Sung Yong Shin
Korea Advanced Institute of
Science and Technology
Introduction

Watermarking


Embedding signature into the media data
Applications of watermarking




Ownership protection (robust watermarking )
Data authentication
Fingerprinting
Secret data hiding
………
Objectives

Robust watermarking for motion data



Imperceptible
Non-invertible
Robust to attacks

smoothing, cropping,
scaling, type conversion,
quantization, adding noise,
adding another watermark,
…
Ownership Protection with Watermark
insertion
+
original motion
extraction
-
registered
suspect motion
watermark
watermarked
motion
analysis of
similarity
extracted
watermark
registration
suspect motion
Previous Work

[Schyndel et al. 1994]


[Tanaka et al. 1990]


Embedding noise-like watermarks
[Cox et al. 1997]


Modifying the least significant bits
Introducing spread-spectrum for images
[Praun et al. 1999]

Employing spread-spectrum for 3D meshes
Spread Spectrum Watermarking

Embedding a watermark with redundancy
original signal
watermark signal
insertion
+
watermarked
signal
Properties of spread spectrum:
 JR (jam resistance)
 LPI (low probability of intercept)
Spread Spectrum Approaches
image

Images


frequency
domain
watermarked
image
[Cox et al. 1997]
Discrete cosine transform
Modifying the most important coefficients
Spread Spectrum Approaches
3D mesh
basis functions
watermarked
mesh
basis
function

3D meshes

Multiresolution analysis
[Praun et al. 1999]
Our Approach

Spread spectrum watermarking for
motion
Motion data = bundle of motion signals of
position or orientation
…
…
motion data
motion signal
Our Approach
Problem:
Difficult to obtain frequency information
from the motion data due to complication
caused by orientations
Solution:
Extracting frequency information from
multiresolution representation
Multiresolution Representation

Representing at multiple resolutions


Hierarchy of successive smoother and
coarser signals
Hierarchy of displacement maps
m(3)
m(2)
m(1)
m(0)
Decomposition
m( n 1)
Reduction
m(n )
Expansion
d( n 1)
Reduction : smoothing, followed by down-sampling
Expansion : up-sampling, followed by smoothing

Both of them can be realized by spatial masking
[Lee2000]
Representation and Reconstruction

Representation
d
m

(n )
( n 1)
d( n  2 )
…
( n 1)
m( n  2)
…
m
d
( 0)
( 0)
m
Reconstruction
d
m
(n )
( n 1)
m
( n 1)
…
…
d
(1)
m
(1)
d ( 0)
m( 0)
Motion Watermarking
Based on multiresolution analysis

Watermark insertion

Watermark extraction

Analysis of similarity between
inserted and extracted watermarks
Watermark Insertion

Decomposing motion signal
m(0)
coarse base signal
d (0)
d (1)
…
original signal
d (n-1)
detail coefficients
Multiresolution
Representation
Watermark Insertion

Perturbing the largest coefficients
watermark
coefficient
m(0)
coarse base signal
(u, v)  (1   wi )(u, v)
(0)
dd(0)
scaling
parameter
(1)
dd(1)
…
original signal
(n-1)
d(n-1)
detail coefficients
the altered
i-th largest
coefficient
((uu, v))
Watermark Insertion

Reconstructing the motion signal
m(0)
coarse base signal
d(0)
d(1)
…
original signal
d(n-1)
detail coefficients
watermarked
signal
Watermark Insertion
Perturbation of coefficient
 Embedding watermark into wide range

+
original motion
watermark
signal
watermarked
motion
Watermark Extraction

Registering original and suspect motion

Using dynamic time warping
[Bruderlin1995]
original
signal
original
signal
suspect
signal
registered
suspect
signal
dynamic
time warping
resampling
Watermark Extraction

Decomposing motion signals
m(0)
m*(0)
coarse base signal
coarse base signal
d (0)
d*(0)
original signal
d (1)
d*(1)
…
…
d (n-1)
detail coefficients
suspect signal
d*(n-1)
detail coefficients
Watermark Extraction

Comparing watermarked coefficients
m(0)
m*(0)
coarse base signal
coarse base signal
d (0)
(u, v)
d*(0)
comparing
d*(1)
d (1)
…
…
(u* , v* )
d (n-1)
d*(n-1)
detail coefficients
detail coefficients
Watermark Extraction

Extracting suspect watermark

Obtaining
from
w  ( w , w , ..., w )
*
*
1
*
2
w  ( w1 , w2 , ..., wm )
(u* , v* )  (1  wi* )(u, v)
scaling
parameter
*
m
Analysis of Similarity

Computing false-positive probability

False-positive probability (Pfp ):
Probability of incorrectly asserting that the data
is watermarked when it is not

Using Student’s t-test

From correlation
 w, w 
*
Experimental Results
Data A
Data B
Data C
Data D
Experimental Results

Original Motion and Watermarked Motion

Fly Spin Kick
Experimental Results

Original Motion and Watermarked Motion

Blown Back
Experimental Results

Results for various attacks

Adding noise attack on Fly Spin Kick
Experimental Results

Results for various attacks

Adding the second watermark on Fly Spin Kick
Experimental Results

Results for various attacks

Smoothing attack on Blown Back
Experimental Results

Results for various attacks

Time warping attack on Blown Back
Experimental Results
Conclusion and Future Works

Watermarking schemes for motion data




Spread spectrum approach
Using multiresolution motion analysis
Robust to attacks
Future works



Consideration for other attacks
Blind detection
Watermark extraction from rendered images
Q/A : False-negative Probability

False-negative Probability
Probability of failing to detect watermarked
data


lesser important than false-positive probability
More difficult to analyze since it depends on
the type and magnitude of attacks
Q/A : Non-invertible Watermark

Generating non-invertible watermark
w  {w1 , w2 ,..., wm }
randomly selected from N (0,1)

seeded by cryptographic hash function with
(original data + owner’s key)
original data
owner’s key
hashed
value
random
numbers
w1 , w2 ,..., wm