principles of digital watermarking ingemar j. cox, matt l. miller, and jeffrey a bloom
TRANSCRIPT
Principles of Digital Watermarking
Ingemar J. Cox, Matt L. Miller, and Jeffrey A Bloom
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Course outline
Part 1: Introduction and Applications
Part 2: Basic Algorithms and Concepts
Part 3: Advanced Watermarking
Course outline
Part 1: Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Definitions and Applications: Outline
Definitions of watermarking
Properties of watermarking systems
Watermarking applications
Conclusions
Definitions and Applications
Definitions of watermarking
Definitions and Applications
Without common definitions, various approaches and technologies cannot be compared.
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Definitions of watermarking: Our definition
Definitions and Applications
Watermarking is the practice of unobtrusively modifying a work of art (image, song, software
program, geometric model, etc.) to embed a message about that work.
Multimedia watermarking is the practice of imperceptibly altering a work (image, song, etc.)
to embed a message about that work.
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Definitions of watermarking: Basic design of a system
Originalwork
Message(regarding
work)
Watermarkembedder
Watermarkedwork
(looks likeoriginal)
Watermarkdetector
Detectedmessage
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Definitions of watermarking: Other definitions
Imperceptibility is not always considered essential (allows for visible watermarking).
Sometimes more broadly defined as any data hiding (i.e. hidden data need not relate to work).
Sometimes more narrowly defined as owner identification (watermarks must indicate identity of owner).
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Definitions of watermarking: Related terms
Data hiding: any technology for preventing adversaries from perceiving or finding data.
Steganography: keeping the existence of messages secret by hiding them within objects, media, or other messages.
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Definitions of watermarking: Related terms
Watermarking is the practice of unobtrusively modifying a work of art (image, song, software program, geometric model, etc.) to embed a
message about that work.
Steganography is the practice of undetectably modifying a work to embed a message.
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Original
Undetectable
Unobtrusive
Properties of systems
Understanding, comparing, and selecting watermarking approaches or technologies takes place in the context of system properties.
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
List of properties to be discussed Embedding effectiveness
Fidelity
Data payload
Blind vs. informed detection
False positive rate
Robustness
SecurityDefinitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
A note before we begin …
When we say “random work”, we mean a work drawn from an application-dependent distribution of works. Examples: x-rays, animation, natural image, classical music, speech, etc.
When we say “random watermark”, we mean a watermark message drawn from the set of possible messages.
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Embedding effectiveness
A system’s embedding effectiveness is theprobability it will succeed in embeddinga random watermark in a random work.
Randomwork
Randommessage
Watermarkembedder
Watermarkdetector
MessageMessagedetecteddetectedcorrectly?correctly?
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Embedding effectiveness
Why might embedding effectiveness be less than 100 percent? In some cases, it is not possible to embed required
amount of information imperceptibly. Actual implementations usually involve some
round-off and truncation before watermarked work is stored, which sometimes make watermark undetectable.
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Properties of systems: Fidelity
A system’s fidelity is the perceptual similaritybetween marked and unmarked works.
Randomwork
Randommessage
Watermarkembedder
Watermarkedwork
Humanobserver
worksworksappearappear
sufficientlysufficientlysimilar?similar?
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Properties of systems: Data payload
A system’s data payload is the amount ofinformation that it can embed in asingle work.
Randomwork
Watermarkembedder
Watermarkedwork
Random message
Definitions and Applications
01101001…01101001…
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Blind vs. informed detection
An informed detector requires someinformation about the original, unwatermarkedwork. A blind detector does not.
Originalwork
Message Watermarkembedder
Watermarkdetector
Required byRequired byinformed detectorinformed detector
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Properties of systems: False positive rate
A system’s false positive rate is thefrequency with which it is expected to detectwatermarks in unwatermarked works.
Random,unwatermarked
work
Watermarkdetector
WatermarkWatermarkdetected?detected?
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Properties of systems: False negative rate
A system’s false negative rate is thefrequency with which it is expected to NOT detectwatermarks in watermarked works.
Random,watermarked
work
Watermarkdetector
WatermarkWatermarkdetected?detected?
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Properties of systems: Robustness
A watermark’s robustness is its ability tosurvive normal processing (e.g. lossycompression, noise reduction, etc.).
Randomwork
Randommessage
Watermarkembedder
WatermarkdetectorNormal
processing
MessageMessagedetecteddetectedcorrectly?correctly?
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Properties of systems: Security
A watermark’s security is its ability toresist hostile attacks, specifically designedto defeat the purpose of the watermark.
Types of attacks Unauthorized embedding (forgery) Unauthorized detection Unauthorized removal
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Security – unauthorized embedding
Randomwork
Forgedmessage
Watermarkdetector
Forged Forged MessageMessagedetecteddetected
WatermarkedworkUnauthorized
embedding by an adversary
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Security – unauthorized removal
Randomwork
Randommessage
Watermarkembedder
WatermarkdetectorHostile
processing by an adversary
MessageMessagedetecteddetectedcorrectly?correctly?
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Security – unauthorized detection
Definitions and Applications
OriginalWork
Message Watermarkembedder
AdversaryAdversarycan detectcan detectmessage?message?
Attempt atdetection by
adversary
Watermarking applications
Watermarking may be appropriate for applications in which data about a work must be imperceptibly embedded. Different applications place different requirements on system properties.
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
List of examples discussed
Broadcast Monitoring
Owner Identification
Proof of Ownership
Transaction Tracking
Content Authentication
Copy Control
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Broadcast monitoring
Watermarkembedder
Watermarkdetector
Broadcasting system
Content wasContent wasbroadcast!broadcast!
Originalcontent
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Broadcast monitoring
Verify advertising broadcasts (1997 scandal in Japan)
Verify royalty payments ($1000 of unpaid royalties to actors per hour of broadcast)
Catch instances of piracy
Monitor when and whether content is transmitted over broadcast channels, such as television or radio
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Owner identification
Originalwork
Distributedcopy
Watermarkdetector
Alice isAlice isowner!owner!
Watermarkembedder
Alice
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus Help honest people identify rightful owner
Notify people of copyright In US, until 1988, such notice was required to
retain copyright Since 1988, presence of notice increases possible
reward in lawsuits
Owner identification
Watermark identifies owner of copyright, similar to a copyright notice
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Owner identification
The problem with text: this well-known image …
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Owner identification
… is a pirated part of a larger image.
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Proof of ownership
Originalwork
Distributedcopy
Watermarkdetector
Alice isAlice isowner!owner!
Watermarkembedder
Alice
BobDefinitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Proof of ownership
Differs from owner identification in two ways Intended to carry burden of proof Watermark need not be detectable by anyone
other than owner (allows informed detection)
Watermark is used to prove ownership in a court of law
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Transaction tracking
Originalwork
HonestBob
Watermarkdetector
B:Evil BobB:Evil Bobdid it!did it!
Watermark A
EvilBob
Unauthorizedusage
Watermark B
Alice
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Transaction tracking
Identifying pirates (DiVX corporation)
Identifying information leaks (M. Thatcher, movie dailies)
Watermarks record transaction histories of content, typically identifying first authorized recipient
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Transaction tracking
The MPAA estimates that piracy costs the US film industry $3B per year
One source of material is the annual distribution of Oscar screeners to the 5,803 voting members of the Academy
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Transaction tracking
Thomson system enabled the MPAA to distribute individually-watermarked VHS and DVD screeners to its 5,803 eligible voting members
Screeners appeared on the internet The Last Samurai Something's Gotta Give Mystic River
Actor Carmine Caridi expelled from MPAA
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Content authentication
Watermark embedder
Watermark detector
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Content authentication
Exact authentication: work is inauthentic if even one bit has changed
Selective authentication: work is inauthentic only if significantly changed
Tell-tale watermarks/localization: identify what changes have been made
Watermark is used to detect modifications applied to cover work
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Copy control
Record control: recording devices contain detectors and refuse to record copyrighted material
Playback control: players contain detectors and refuse to play pirated material
Watermarks indicate whether content may be copied
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Copy controlCompliantrecorder
CompliantplayerLegal copy
Illegal copy
Playback control
Record control
Definitions and Applications
Non-compliantrecorder
Conclusions
Definitions and Applications
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Conclusions: Stuff not covered
Erasability (whether watermark can be perfectly removed)
Cipher and watermark keys
Modification and multiple watermarks
Cost
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Conclusions: Take Away
Watermarking may be appropriate for applications in which data about a work must be imperceptibly embedded.
Different applications place different requirements on system properties.
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Conclusions: Take Away
Key properties include Embedding effectiveness Fidelity Data payload Blind vs. informed detection False positive rate Robustness Security
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Digression: The politics of DRM
Why does Hollywood care about piracy? Loss in revenue
But some level of piracy actually stimulates sales
Evidence that peer-to-peer file sharing affects sales is mixed But has been used to control the evolution of the digital
market
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Digression: The politics of DRM
Why do computer and consumer electronics companies care about DRM?
Need content owners to provide content in new digital formats
Conflict of interests Customers don’t want DRM Legal and business contracts impose DRM
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Course outline
Part 1: Definitions and Applications
Part 2: Basic Algorithms and Concepts
Part 3: Informed Watermarking
Course outline
Part 2: Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Basic Algorithms and Concepts: Outline
Algorithmic building blocks
Robustness issues
Security issues
Conclusions
Algorithmic building blocksOver the past 5 to 10 years of research, several ideas have emerged as basic building blocks of watermarking systems.
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
A simple watermark embedder
Given … Watermark pattern, w Cover image co
Embedding strength
Compute watermarked image, cw, as
wcc ow α
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
A watermarked version of this …
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
…looks like this
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed detection
Given Possibly watermarked image c Original cover image co
Subtract original to obtain watermark pattern (if present)
present is watermarkif αwccw on
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Linear correlation test
Use linear correlation to determine whether
Linear correlation defined as
If c = co + n, then zlc(wn,w) 0
If c = co + w + n, then zlc(wn,w) zlc(w,w)
wwn α
yxyxNN
zyx
, ,11
,,
lc wwwwww nnn
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Blind detection
If w is chosen so that zlc(co,w) is likely to be close to 0, then zlc (c,w) zlc(wn,w).
No need to subtract out co before computing linear correlation.
White noise pattern tends to have low-magnitude correlation with any image.
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Interpreting system geometrically
Media space – a high-dimensional space in which each point corresponds to a work. 256 256 grayscale image 65,536 dimensions
(one for each pixel). 5 second mono audio clip, sampled at 44,100Hz
220,500 dimensions (one for each sample)
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
2d pictures of media space
Several possible interpretations Abstraction of high-dimensional space (just pretend
media space is really 2d) Projection of media space Slice of media space
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Picture of media space
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Algorithmic building blocks: Watermark in media space
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Geometric interpretation of zlc()
zlc(c,w) is just dot product of c and w divided by N
Dot product of c and w is cosine of angle between them, times their magnitudes
If |w| = 1, then dot product is projection of c onto direction of w
Comparing zlc(c,w) against a threshold leads to detection region with planar boundary
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Geometric interpretation of zlc()
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Now that we have a basic system …
… let’s consider a problem: What happens when we change the contrast of the image?
cwn = cw , where is some scalar value
zlc(cwn,w) zlc(cw,w)
If < 1, detection value might drop below threshold
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Solution: normalized correlation
Normalize correlation by magnitudes of vectors
Scaling has no effect on znc(c,w)
wc
wcwc
,ncz
wcwc
wcwc ,
, ncnc zz
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Geometric interpretation of znc()
znc(c,w) correlation is just cosine of angle between c and w
Comparing znc(c,w) against a threshold is equivalent to comparing angle against a threshold
Result: detection region with conical boundary
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Geometric interpretation of znc()
acos( znc(c,w) )
w
c
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Another problem
What happens if the image is spatially shifted a little?
Detection value will depend on autocorrelation function of watermark pattern.
White noise pattern has close to zero autocorrelation.
Watermark is unlikely to be detected.
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Possible solution
Watermark Fourier-magnitude instead of pixel values
Fourier-magnitudes are invariant to translation
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Possible solution
To embed Take FFT of image and compute magnitudes Add w to magnitudes Scale FFT coefficients of image to new magnitudes
and take inverse FFT
To detect Take FFT of image and compute magnitudes Compute normalized correlation between
magnitudes and w
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Watermark extraction
We can view the preceding system as comprising two basic parts A watermark extraction process that maps points in
media space to points in some marking space (Fourier-magnitude space, in this case)
A simple watermarking system that operates in marking space
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Reasons for watermark extraction
Increase robustness Project into distortion invariant space Invert distortions Reduce noise
Reduce computational cost
Increase security Key-based extraction
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
“Transform domain” watermarking
Many authors categorize watermarking systems by transforms included in their extraction processes, e.g. … “Spatial-domain watermarking” (no transform) “DCT-domain watermarking” “Wavelet-domain watermarking” Etc.
But …
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
“Transform domain” watermarking
… the transform alone says little about how the system works If T is a linear, energy-preserving transform, then
zlc( T(c),w ) = zlc( c,T-1(w) ) Thus a linear-correlation-based system in domain
T is the same as a spatial-domain system with a different watermark pattern
It is the nonlinearities in the extraction process that distinguish a system’s behavior
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Perceptual shaping
Basic idea: amplify watermark in areas where the cover work can mask noise
w
co
cw
Perceptualmodel
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Image before embedding
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Without perceptual shaping
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
With perceptual shaping
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Early approach: invert shaping in detector (shown here for informed detector)
Detection after perceptual shaping
c
co
wn
Perceptualmodel
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Detection after perceptual shaping
Not necessary to invert perceptual shaping Distortion of watermark pattern degrades detection
value for given watermark scaling value, , but … … possible to use larger value of because
pattern is better hidden
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Geometric view of perceptual shaping
Region ofacceptable
fidelity Shapedwatermark
vector
Original(unshaped)watermark
vector
Basic Algorithms and Concepts
Robustness IssuesThe robustness of a watermark is its ability to survive normal processing.
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Watermarked image is corrupted by additive noise
Linear Correlation
Linear Correlation (matched filtering) is optimal when noise is AWGN.
Additive Noise
ncc wwn
nwcwcwz wwnlc ,
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Watermarked image is subjected to a change in contrast
Linear Correlation
For < 1, this scaling decreases the detection value.
How can we select a threshold?
Valumetric Scaling
wwn cc
wwnlc cwcwz ,
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Valumetric Scaling
wn
wn
c
c
Unit Sphere
w
w
Normalized Correlation
Independent of vector magnitude
Describes the cosine of the angle between the vectors
-1 znc +1
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Quantization
Transform
Quantization
EntropyCoding
uncompressedwork
compressedwork
Quantization noise cannot be modeled as additive white noise
There are current efforts to model quantization noise
Eggers and Girod
Appendix B.5
Canonical TransformCoder Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Synchronization
Geometric distortion in imagery translation, rotation, zoom, aspect
ratio, skew, perspective distortion, warp
Temporal distortion in audio time delay, time scaling
Video can suffer from both geometric and temporal misalignments
Noise due to synchronization errors is not well modeled as additive white noise.
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Synchronization Approaches
Exhaustive Search Detection applied at all possible
temporal/geometric distortions Negative impact on false positive probability Usually requires too much computation
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Synchronization Approaches
Synchronization Synchronization pattern is embedded along
with the payload-carrying pattern. Registration to synchronization pattern prior to
detection. Negative impact on fidelity and security
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Synchronization Approaches
Implicit Synchronization Watermark location in time or space is relative
to extracted features Example: audio reference pattern added
between salient points
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Synchronization Approaches
Invariance Design patterns that are invariant to
desynchronization Example: Use of Fourier magnitude in
watermark extraction process for shift invariance
Basic Algorithms and Concepts
Security IssuesThe security of a watermark is its ability to resist hostile attacks specifically designed to defeat the purpose of the watermark.
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Robustness
Security against unauthorized removal requires robustness to any process that maintains fidelity Desynchronization Attacks Noise Removal Attacks
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Mosaic Attack
Image is broken into many small rectangular patches
Each patch is too small for reliable detection
Patches are displayed in a table such that patch edges are adjacent
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Collusion Attacks
Many different works, same watermark
Many different watermarks, same work
Simple example: averaging Average of many different works gives an estimate
of the watermark Average of many copies of the same work reduces
the strength of each watermark
212121
2
1
2
1
2
1
2wwcwcwc
ccooo
ww
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Copy Attack
Watermark is “copied” from one work to another
Unauthorized Embedding
Example: apply a watermark removal attack to obtain an estimate of the watermark, add to fake.
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Ambiguity Attack
Create the appearance that a watermark has been added to someone else’s work
Example: define fake watermark pattern and subtract from the distributed image. This is the fake “original.” Difference between distributed and Bob’s original
contains Bob’s watermark Difference between distributed and Alice’s original
contains Alice’s watermark
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pusAlice’s
original(real)
Bob’s“original”
(fake)
Alice’sdetector
Bob’sdetector
Alice isAlice isowner!owner!
Bob isBob isowner!owner!
Distributedcopy
Alice
Bob
Ambiguity Attack
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Solution to ambiguity attack: Alice uses system that cannot be hacked May be possible to implement by making
watermark dependent on cryptographic hash of original work
Strictly-speaking, provides proof of ancestry, rather than proof of ownership
Ambiguity Attack
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Sensitivity Analysis
Technique for removing watermark when adversary has black box detector
Estimate the normal to the detection region surface boundary at some point
Assume that this normal indicates a short path out of the detection region
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Sensitivity Analysis
w
Detection region
watermarkedwork
work A
attackedwork
1 Find a work that lies on the detection boundary
2 Approximate the normal to the detection boundary
3 Scale and add the normal to the watermarked work
Basic Algorithms and Concepts
Conclusions
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Conclusions: Major stuff not covered
Message-coding for multi-bit watermarks
Non correlation-based watermarking “Constraint-based” watermarking (can usually
be recast as correlation-based) Quantization-based watermarking (will be
covered in part 3)
Authentication methods
ROC Curves
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Conclusions: Take Away
Linear correlation (matched filtering) is optimal for detecting a signal in AWGN
Most processing is not well modeled as AWGN
Normalized correlation provides robustness to amplitude changes
Helpful to think of a work as a point in a high dimensional space
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Conclusions: Take Away
Watermark extraction: project a work to another space for embedding and/or detection
Perceptual modeling can improve fidelity and allow for stronger embedding
Robustness to desynchronization is an difficult problem
Collusion attacks and sensitivity analysis are significant security challenges
Basic Algorithms and Concepts
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Outline
Part 1: Definitions and Applications
Part 2: Basic Algorithms and Concepts
Part 3: Informed Watermarking
Part 3: Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed watermarking: Outline
Idea of informed watermarking
Informed shaping
Informed coding
Conclusions
Informed watermarking
Idea of informed watermarking
Informed watermarking is the practice of using information about the cover work during watermark coding and shaping.
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Blind coding & shaping
Blindcoding
Blindshaping(scaling)
WatermarkedWork
OriginalWork
Message
Watermark embedder
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed shaping
Blindcoding
Informedshaping
WatermarkedWork
OriginalWork
Message
Watermark embedder
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed coding & shaping
Informedcoding
Informedshaping
WatermarkedWork
OriginalWork
Message
Watermark embedder
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Central insight
Watermarking with informed embedder and blind detector = communication with side information at the transmitter Shannon’s model: transmitter has knowledge of
channel’s noise characteristics In watermarking, cover Work = (part of) noise Theoretical results for this type of channel should
apply to watermarking
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Consequences
Informed shaping alone Allows more precise control of fidelity/robustness
tradeoff
Informed coding + informed shaping Greatly increases payload for a given
fidelity/robustness performance Alternatively, improves fidelity/robustness
performance for a given payload
Informed watermarking
Informed shaping
The cover work can be used to inform perceptual shaping. It can also be used to adjust watermark pattern for maximal robustness.
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Basic approach
Design detector (we’ll use the linear-correlation detector from Part 2)
Treat detection algorithm and parameters as given
Design best embedder we can
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Embedding problem
The embedder is capable of producing any image
Objective: produce an image within the intersection of a region of acceptable fidelity and the detection region
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Embedding problem
w
Region ofacceptable
fidelity
Any point inthis area is asuccessfulembedding
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Embedding problem
Several possible approaches Maximize robustness for a given fidelity Maximize fidelity for a given robustness
Either approach requires Estimate of fidelity Estimate of robustness
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Simple embedding method
Assume MSE indicates fidelity (better estimates lead to perceptual shaping)
Assume robustness is monotonic function of linear correlation
Under these assumptions, blind embedding achieves maximum “robustness” for given “fidelity”
Alternatively …
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Simple embedding method
… we can minimize fidelity impact while embedding for a constant “robustness”
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Estimating robustness
Simple assumption: robustness is monotonic function of detection value True for linear correlation Not true for other detection measures
For normalized correlation, we have obtained good results by estimating amount of white noise that may be added before watermark is likely to be lost
Informed watermarking
Informed coding
Significantly larger data payloads can be embedded if the mapping between messages and watermark patterns is dependent on the cover work.
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed coding: Outline
Writing on dirty paper (problem studied by M. Costa)
Dirty-paper codes
Application of dirty-paper codes to watermarking
Experimental results
Conclusions
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed coding: Writing on dirty paper
M. Costa studied a “dirty-paper channel” Obtain a piece of paper with normally-distributed
dirt Write a message using limited ink Send message, acquiring more dirt along the way Recipient cannot distinguish dirt from ink How much information can we send?
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed coding: The dirty-paper channel
Transmitter Receiver
Firstnoise
m
Secondnoise
m’
s
n
x y
x limited by power constraint:
i
pi 2x
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed coding: Costa’s result
First noise has no effect on channel capacity
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed coding: Dirty-paper codes
Basic idea Dirty-paper code = code in which each
message is represented by several alternative code vectors
From the set of vectors that represent the desired message, choose the one, u, that is closest to the first noise, s
Transmit a function of u and s, for example x = u - s
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Coding for a simplified channel
Consider a simplified version of the dirty-paper channel First noise has only two possible values, s1 and s2
(i.e. there are only two possible patterns of dirt on the paper)
Remainder of channel is the same
If s1 is sufficiently different from s2, then Costa’s result is easy to obtain
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Coding for a simplified channel
DC E
A B
F G DC E
A B
F GThis group ofcode vectors
centered on s1
This group ofcode vectors
centered on s2
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed coding: Dirty-paper codes
For full dirty-paper channel:
Try to design a dirty-paper code in which, within the power-constraint around every possible s,
there is at least one code vector for each message.
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Coding for full dirty-paper channel
Code must ensure that, within the power-constraint around every possible s, there is at least one code vector for each message.
Capacity cannot be achieved transmitting x = u – s.
Costa transmits x = u – s, where is a carefully-chosen constant
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Application to watermarking
Costa’s proof does not translate directly to watermarking In watermarking, noise is not Gaussian Non-Gaussian noise necessitates non-spherical
detection regions (e.g. cones)
Lessons from Costa Use dirty-paper codes Use non-trivial informed embedding
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Practical dirty-paper codes
Costa’s code is generated randomly Requires exhaustive search during encoding and
decoding Practical for only very small data payloads
Lattice code is most-studied practical code Chen & Wornell (“Dither Index Modulation”,
“Quantization Index Modulation”) Eggers, Su, & Girod (“Scalar Costa Scheme”)
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Informed coding: Lattice codes
Each dimension in marking space encodes one symbol, usually one bit
Bit encoded by choosing between two quantization points
10 0 0 01 1
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Properties of lattice codes
Typically much higher capacity than correlation-based systems (> 1000 bits)
Not usually as robust as correlation-based systems Correlation-based systems have better
payload/robustness tradeoff when noise is high Lattice codes susceptible to changes in image
brightness or audio volume
Informed watermarking
Conclusions
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Conclusions: Stuff not covered
Informed-embedding for multi-bit watermarks
Syndrome coding
Application of informed-coding to correlation-based systems
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Conclusions: Take Away
Informed shaping Embedder may choose any point in the detection
region for the desired message Best to base choice on an estimate of robustness
Informed coding Define several patterns for each message, and
embed the one that’s closest to the cover work In theory, capacity of watermarking might be
unaffected by distribution of cover works
Informed watermarking
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Future directions
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Future directions
Research Informed coding
Quantization index modulation Syndrome coding Trellis coding
Robustness Non-random processes
• Esp. geometric distortions
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Future directions
Research Security
Collusion attacks Others …
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Future directions
Commercial applications Transaction tracking
Movie screeners Digital cinema
Broadcast monitoring Metadata
Lyrics in MP3 files
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Future directions
Commercial applications Authentication
Cameras Surveillance video Medical imagery
Enhancements to legacy systems 3D HDTV – Benoit Macq
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Future directions
Commercial applications? Copy control Proof of ownership
UC
L A
dast
ral P
ark
Post
gra
duate
Cam
pus
Future directions
Commercial applications Must be based on a service or product
Not technology Similar to commercial applications of cryptography