Robust Motion Watermarkingbased onMultiresolution Analysis

Tae-hoon KimJehee 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

watermark

registration

extractedwatermark

extraction

+

-

analysis ofsimilarity

original motion watermarked motion

registeredsuspect motion suspect motion

Previous Work

[Schyndel et al. 1994] Modifying the least significant bits

[Tanaka et al. 1990] Embedding noise-like watermarks

[Cox et al. 1997] 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

insertion+

watermarked signal

watermark signal

original signal

insertion+

watermarked signal

watermark signal

Properties of spread spectrum: JR (jam resistance) LPI (low probability of intercept)

Spread Spectrum Approaches

Images [Cox et al. 1997] Discrete cosine transform Modifying the most important coefficients

image watermarkedimage

frequencydomain

Spread Spectrum Approaches

3D meshes [Praun et al. 1999] Multiresolution analysis

3D mesh basis functions watermarkedmesh

basis function

Our Approach

Spread spectrum watermarking for motion

motion signal

…

motion data

…

Motion data = bundle of motion signals of position or orientation

Our Approach

Problem:Difficult to obtain frequency information from the motion data due to complicati

on caused by orientations

Solution:Extracting frequency information frommultiresolution representation

Multiresolution Representation

Representing at multiple resolutions Hierarchy of successive smoother and

coarser signals Hierarchy of displacement maps

(3)m (2)m (1)m (0)m

Decomposition

Reduction : smoothing, followed by down-samplingExpansion : up-sampling, followed by smoothing

Both of them can be realized by spatial masking [Lee2000]

)(nm

)1( nm

Reduction Expansion

)1( nd)(nm

)1( nm

Reduction Expansion

)1( nd)(nm

)1( nm

Reduction Expansion

)1( nd

Representation and Reconstruction

Representation

)(nm )1( nm

)1( nd

)2( nm

)2( nd

)(nm )1( nm (1)m

)0(m

)0(d…

…

)1( nd )1(d

)0(m

)0(d

Reconstruction…

…

Motion Watermarking

Based on multiresolution analysis

Watermark insertion

Watermark extraction Analysis of similarity between

inserted and extracted watermarks

Watermark Insertion

Decomposing motion signal

original signal

MultiresolutionRepresentation

(0)d(1)d

(n-1)d

…

coarse base signal

detail coefficients

(0)m

Watermark Insertion

Perturbing the largest coefficients

original signal

( , )u v

(0)d(1)d

(n-1)d

…

coarse base signal

detail coefficients

(0)m

the i-th largest coefficient

(0)d(1)d

(n-1)d

coarse base signal

detail coefficients

…

(0)m

altered coefficient

( , ) u v

( , ) (1 )( , )iw u v u v

scaling paramete

r

watermark

coefficient

Watermark Insertion

Reconstructing the motion signal

original signal

(0)d(1)d

(n-1)d

coarse base signal

detail coefficients

…

(0)m

watermarked signal

Watermark Insertion

Perturbation of coefficient Embedding watermark into wide

range

original motion

+

watermarksignal

watermarked motion

Watermark Extraction

Registering original and suspect motion Using dynamic time warping

[Bruderlin1995]

dynamic time warping

resampling

originalsignal

suspectsignal

originalsignal

registeredsuspect signal

Watermark Extraction

Decomposing motion signals

original signal

suspect signal

(0)d(1)d

(n-1)d…

coarse base signal

detail coefficients

(0)m

*(0)d*(1)d

*(n-1)d

coarse base signal

detail coefficients

…

*(0)m

Watermark Extraction

Comparing watermarked coefficients

(0)d(1)d

(n-1)d…

coarse base signal

detail coefficients

(0)m

*(0)d*(1)d

*(n-1)d

coarse base signal

detail coefficients

…

*(0)m

( , )u v

),( ** vu

comparing

Watermark Extraction

Extracting suspect watermark

Obtaining from

scaling paramete

r

),)(1(),( *** vuvu iw

) ..., , ,( **2

*1

*mwwww

) ..., , ,( 21 mwwww

Analysis of Similarity

Computing false-positive probability False-positive probability (Pfp ):

Probability of incorrectly asserting that the datais watermarked when it is not

Using Student’s t-test From correlation * ,ww

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 ProbabilityProbability 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

randomly selected from seeded by cryptographic hash function with

(original data + owner’s key)

(0,1)N

1 2{ , ,..., }mw w ww

original dataowner’s key

hashed value

random numbers

1 2, ,..., mw w w