digital fingerprinting for g g p g f g g p g f multimedia forensics
TRANSCRIPT
Digital Fingerprinting for Digital Fingerprinting for g g p g fg g p g fMultimedia ForensicsMultimedia Forensics
K.J. Ray Liu, Wade Trappe*, Z. Jane Wang**, and Min Wu
ECE Dept., University of Maryland, College Park* WINLAB/ECE, Rutgers University
Include research sponsored in part by Air Force Research Lab and NSF; and
** ECE, University of British Columbia, Canada
March 2005
p p y ;contributions from Hong Zhao, Shan He, Hongmei Gou, and Zang Li.
Digital Fingerprinting and Tracing TraitorsDigital Fingerprinting and Tracing Traitors
Leak of information as well as alteration and repackaging poses serious threats to government operations and commercial markets – e.g., pirated content or
classified documentstudio
Th L d f
Alicew1
w2
SellSell
Promising countermeasure:
The Lord ofthe Ring
Bob
Carl
w2
w3
grobustly embed digital fingerprints– Insert ID or “fingerprint” (often through conventional watermarking)
to identify each userto identify each user– Purpose: deter information leakage; digital rights management(DRM)– Challenge: imperceptibility, robustness, tracing capability
Digital Fingerprinting for Multimedia Forensics 2
Case Study: Tracing Movie Screening CopiesCase Study: Tracing Movie Screening Copies
Potential civilian use for digital rights management (DRM) Copyright industry – $500+ Billion business ~ 5% U.S. GDPpy g y $5 5
Alleged Movie Pirate Arrested (23 January 2004)– A real case of a successful deployment of 'traitor-tracing'
h i i h di i l lmechanism in the digital realm– Use invisible fingerprints to protect screener copies of pre-release
movies
Carmine Caridi Russell friends … Internetw1Last Samurai
Hollywood studio traced pirated version
Digital Fingerprinting for Multimedia Forensics 4
http://www.msnbc.msn.com/id/4037016/
Embedded Fingerprinting for MultimediaEmbedded Fingerprinting for Multimedia
Multimedia Document
Customer’s ID: Alice
Fingerprinted CopyFingerprinted CopyMultimedia Document
Customer’s ID: Alice
Fingerprinted CopyFingerprinted CopyEmbedded Finger-
embedembedDigital Fingerprint101101 …101101 …
Distribute to Alice
g p pyg p py
embedembedDigital Fingerprint101101 …101101 …
Distribute to Alice
g p pyg p pygprinting
AliceAlice
BobBobUnauthorized Unauthorized rere--distributiondistribution
AliceAlice
BobBobUnauthorized Unauthorized rere--distributiondistribution
Multi-user Attacks
Collusion Attack Collusion Attack (to remove fingerprints)(to remove fingerprints)
Colluded CopyColluded CopyFingerprinted docfor different users
Collusion Attack Collusion Attack (to remove fingerprints)(to remove fingerprints)
Colluded CopyColluded CopyFingerprinted docfor different users
ExtractExtract AliceIdentifyIdentifyExtractExtract AliceIdentifyIdentifyExtract Extract FingerprintsFingerprints
Suspicious Suspicious CopyCopy
101110 …101110 …
Codebook
Alice, Bob, …
Identify Identify TraitorsTraitors
Extract Extract FingerprintsFingerprints
Suspicious Suspicious CopyCopy
101110 …101110 …
Codebook
Alice, Bob, …
Identify Identify TraitorsTraitorsTraitor
Tracing
Digital Fingerprinting for Multimedia Forensics 5
AntiAnti--Collusion FingerprintingCollusion Fingerprinting
Build anti-collusion fingerprinting to trace traitors and colluders Potential impact: gather digital evidence and trace culpritsPotential impact: gather digital evidence and trace culprits
– Deter unauthorized dissemination in the first place let the bad guys know their high risk of being caught
– Support secure information sharing in government, military, and intelligence operations
– Commercial use for digital rights management (DRM)
Traditional colluder tracing is designed for generic data– E g Boneh-Shaw (1995)E.g Boneh-Shaw (1995)
Multimedia poses unique challenges and opportunities– Fingerprint should be embedded seamlessly and robustly
Digital Fingerprinting for Multimedia Forensics 6
– Multimedia-specific characteristics may be exploited
Outline of the TutorialOutline of the Tutorial
Preliminaries – Review of Robust Embedding– Overview of Collusion strategies
Orthogonal Fingerprinting – Collusion Resistance Analysis Group-oriented Fingerprintingp g p g Coded Fingerprinting
Data Embedding in Curves for Fingerprinting Maps Secure and Efficient Distribution of Fingerprinted Video
Digital Fingerprinting for Multimedia Forensics 7
Preliminary: Robust Embedding and CollusionPreliminary: Robust Embedding and CollusionPreliminary: Robust Embedding and CollusionPreliminary: Robust Embedding and Collusion
Digital Fingerprinting for Multimedia Forensics 8
Background: Invisible WatermarkBackground: Invisible Watermark
10011010 …10011010 …
© Copyright …
Digital Fingerprinting for Multimedia Forensics 9
py g
Spread Spectrum Embedding Procedure
( )iii wJNDxy
Human Visual
Watermark wi ~ N (0,1)
JND
iiiy
DCT
Visual Model
IDCT
x y
Watermarked ImageOriginal Image
Key Points: Human perceptual model for imperceptibility Use spreading and noise-like watermark for robustness and
Digital Fingerprinting for Multimedia Forensics 10
Use spreading and noise-like watermark for robustness and imperceptibility
Detection of Spread Spectrum WatermarkDetection of Spread Spectrum Watermark
Detection can be formulated as a hypothesis testing problem. Optimal detector can be calculated from assumptions onOptimal detector can be calculated from assumptions on
distortion (host media and noise from attacks). If distortion is i.i.d. N(0, d
2) then optimal detector is a correlator:
presenceabsence
H0: <(x + noise), w >H1: < y + noise, w > = < w + (x + noise), w >
orH0 ( i ) i
b/bli d 22TT
TN
H0: < (x + noise) - x, w > = < noise, w > H1: < y + noise - x, w > = < w + noise, w >
b=1b=-1
./(:blind-non
;/:blind22
22
www)y
wwy
TdN
dT
N
T
T
Digital Fingerprinting for Multimedia Forensics 11
TN Antipodal Code Modulation
Fingerprinting upon SpreadFingerprinting upon Spread--Spectrum EmbeddingSpectrum Embedding
Spread-spectrum embedding– Provide very good tradeoff on imperceptibility and robustness, y g p p y ,
esp. under non-blind detection– Typical watermark-to-noise (WNR) ratio: -20dB in blind
detection 0dB in non-blind detectiondetection, 0dB in non-blind detection
Choice of modulation schemes# of fingerprints
Orthogonal modulation jj uw
B
# of fingerprints= # of ortho. bases
(Binary) coded modulation
for or
i
iijj b1
uw
1,0ijb 1ijb
Digital Fingerprinting for Multimedia Forensics 12
# of fingerprints >> # of orthorgonal bases
Collusion Attacks by Multiple UsersCollusion Attacks by Multiple Users
Collusion: A cost-effective attack against multimedia fingerprints– Users with same content but different fingerprints come together to g p g
produce a new copy with diminished or attenuated fingerprints– Fairness: Each colluder contributes equal share through averaging,
interleaving and nonlinear combininginterleaving, and nonlinear combining
Originally fi i d
Alice ChrisBob
Originally
Alice Chris
fingerprintedcopies
fingerprintedcopies
Collusion byaveraging 1/3
Colluded copy
Cut-and-paste attack
Digital Fingerprinting for Multimedia Forensics 14
Colluded copyColluded copy
Order Statistic Based Nonlinear CollusionOrder Statistic Based Nonlinear CollusionAlice Bob Chris
Originally fingerprintedcopies
172 173 176174 175Multi-usercollusion Minimum MaximumMedian (Min+Max)/2
C ll d d
Digital Fingerprinting for Multimedia Forensics 15
Colluded copy
Linear vs. Nonlinear CollusionLinear vs. Nonlinear Collusion
Linear collusion by averaging is simple and effective Colluders can output any value between the minimum and maximumColluders can output any value between the minimum and maximum
values, and have high confidence that such spurious value is within the range of JND.
I id li ll i ll Important to consider nonlinear collusion as well. Order statistics based nonlinear collusions
kk wgJNDxygV )()( )()(
– Analyze detection stat.’s distribution
m in m a x
m in m a x m in m a x
; ; ; a v e m e d ia nj j j jV V V V
V V V
CC SkjjjSkjj wgJNDxygV )()( )()(
– Many nonlinear collusions can be modeled as attacks by averagingplus additive noise
m in m a x m in m a x
m o d m in m a x
m in
,
w .p .
j j j
n e g m e dj j j j
jr a n d n e g
V a v e r a g e V V
V V V V
V pV
Digital Fingerprinting for Multimedia Forensics 16
p=> Focus on averaging collusion
m a x
pw .p . 1
jr a n d n e gj
j
pV
V p
Averaging and Nonlinear Collusions (cont’d)Averaging and Nonlinear Collusions (cont’d)
Thresholding detector is robust to different types of attacks:averaging collusion; order-statistic based (min, max, …)
Rationale from detector’s view pointpDetection statistics of averaging and many nonlinear collusions are (approx.) Gaussian distributions with ( pp )same mean
=> Yield similar performance if the overall distortion is the same.
1( , )
1j cg j S
s d
s d
nonlinear attacks
average attacks
Digital Fingerprinting for Multimedia Forensics 18
2c
jj SK
s d average attacks
Orthogonal FingerprintingOrthogonal FingerprintingOrthogonal FingerprintingOrthogonal Fingerprinting
Digital Fingerprinting for Multimedia Forensics 19
Orthogonal FingerprintingOrthogonal Fingerprinting
Straightforward concept and easy to implement– Prior works extended from watermarking by Cox et al., Stone, g y , ,
Killian et al.– Advantage in distinguishing individual fingerprints
Di d i fi i i d i ll i– Disadvantage in fingerprint attenuation during collusion
Tradeoff:Orthogonality helps distinguish individual users– Orthogonality helps distinguish individual users
– Correlated fingerprints help resist collusion=> Orthogonal fingerprinting can be good for a small user groupg g p g g g p
How many colluders out of how many users can we trace accurately using orthogonal fingerprinting?
Digital Fingerprinting for Multimedia Forensics 20
Z. Wang, M. Wu, H. Zhao, W. Trappe, and K.J.R. Liu: "Collusion Resistance of Multimedia Fingerprinting Using Orthogonal Modulation", accepted by IEEE Trans. on Image Proc., June 2004 (to appear in 2005).
Orthogonal Fingerprinting and CollusionOrthogonal Fingerprinting and Collusion
2( ) ~ (0, ), for 1,..., .j j
dd i N i N
y s x
2 2
,
: || || / || ||j k j k
WNR
s s
s d
1 1the correlator ( ) / || ||, for 1,...,T
N jT j j n y s s1 1
c c
j jj S j SK K
y y z s d 2
2
( ) || ||, , ,
(|| || / , ), if ( ( ) | , )
(0, ), o.w.
N j
d cN K c
d
j j
N K j Sp T j H S
N
y
s
Digital Fingerprinting for Multimedia Forensics 21
How Many Colluders Can We Trace?How Many Colluders Can We Trace? Two issues limit the anti-collusion capability:
– Orthogonal fingerprints get attenuated with more colludersl d t d d d t ti t ti ti di t ll d leads to reduced detection statistics corresponding to colluders
– Probability of false alarm increases as the total # of users increases the detector correlates with more orthogonal fingerprint signals the detector correlates with more orthogonal fingerprint signals,
each corresponding to a user to identify which user is involved assume the same threshold is used on detection statistics false alarm depends on the maximum detection statistics with non- false alarm depends on the maximum detection statistics with non
involved users, which becomes larger as # users increases
Tracing Capability: How many colluders out of how many users are sufficient to break down a fingerprinting system?sufficient to break down a fingerprinting system?
To meet desired probability of detection (Pd) & false alarm (Pfp)– We can analyze the maximum allowable colluders
> Thi id d i id li t fi i ti t f
Digital Fingerprinting for Multimedia Forensics 22
=> This provides design guidelines to fingerprinting systems for applications with different protection requirements
Formulations for Max. Number of ColludersFormulations for Max. Number of Colluders
Threshold detector: (indices of colluders) 1,..,
ˆ arg max{ ( ) }Nj nT j h
j
(indices of colluders)
Performance criteria: ˆ{ } 1 (1 ( / ))n Kfp r c dP P S Q h j
S stem req irement:
|| || /ˆ{ } 1 (1 ( ))Kd r c
d
h s KP P S Q
j
System requirement:fp
d
PP
maxK The maximum number of colluders
allowed by a fingerprinting system
– The desired Pfp determines an expression of threshold for detector– The desired Pd determines the maximum # of colluders allowed by
d y g p g y
Digital Fingerprinting for Multimedia Forensics 23
The desired Pd determines the maximum # of colluders allowed by the fingerprinting system
Bounds for Max. Number of ColludersBounds for Max. Number of Colluders
Lower bound and Upper bound for Kmax
Obtained by analytic approximations on Q functions– Obtained by analytic approximations on Q-functions
max 2 2 2min{ , }, where L LN NK n K K
~ N
max 2 2 2
max 1
{ , },log( /(2 log(2 )))
min{ , }, where (1 1 )
L LH
H H K
h n n
NK n K K
h Q
log( )n
(1 1 )KLh Q
– two auxiliary variables are defined as2
1
log(2 )
(1 1 )
L
nL
h n
NK
h Q
Digital Fingerprinting for Multimedia Forensics 24
(1 1 )Lh Q
Results on Colluder BoundsResults on Colluder Bounds
28K– Stringent requirement: correctly identify at least one colluders without falsely
accusing any– The colluder tracing capabilities for a thousand-user system is limited to several
max 28K
Digital Fingerprinting for Multimedia Forensics 25
The colluder tracing capabilities for a thousand user system is limited to several dozens colluders
Collusion Resistance: Kmax vs. WNR; Kmax vs. Pfp
based on numerical calculations
Digital Fingerprinting for Multimedia Forensics 26
based on numerical calculations
Performance Criteria for Different NeedsPerformance Criteria for Different Needs
Why? different goals and interests; a decision-making system consists of many components.a decision making system consists of many components.
Capture one:– To identify at least one colluder with high confidence without accusing
innocent usersinnocent users.
Capture more:– To catch more colluders, possibly at a cost of falsely accusing more– Tradeoff between the expected percentage of colluders that are
successfully captured and the expected percentage of innocent users that are falsely placed under suspicion.
Capture all:– Tradeoff: measured by an efficiency rate, which describes the amount of
expected innocents accused normalized by the number of colluder and
Digital Fingerprinting for Multimedia Forensics 27
expected innocents accused normalized by the number of colluder and the probability of capturing all colluders.
Performance Criteria for Different NeedsPerformance Criteria for Different Needs
Catch one– With high accuracyWith high accuracy– Measured by Pd and Pfp discussed earlier
C h Catch morethe expected fraction of innocents falsely suspected: ( / )
|| || /th t d f ti f ll d f ll t d ( )
i dr Q hh KQ
s
Catch all
|| ||the expected fraction of colluders successfully captured: ( )cd
r Q
the expected number of innocents captured the expected number of colluders captured
ˆ
R
Digital Fingerprinting for Multimedia Forensics 28
( )d r cP P S j
Collusion ResistanceCollusion Resistance
Catch morei ir
1( )i dh Q
NK
System requirement Collusion Resistance
c cr max 1 1( ) ( )i c
KQ Q
Catch allDifferent sets of performance criteria were studied. It seems that an orthogonal fingerprinting system can resist to the collusion attacks based oncollusion attacks based on a few dozen independent copies.
Digital Fingerprinting for Multimedia Forensics 29
GroupGroup--Oriented FingerprintingOriented FingerprintingGroupGroup--Oriented FingerprintingOriented Fingerprinting
Z. Jane Wang, M. Wu, W. Trappe, and K. J. Ray Liu, “Group-Oriented Fingerprinting for Mul-
Digital Fingerprinting for Multimedia Forensics 30
g, , pp , y , p g p gtimedia Forensics", EURASIP Journal on Applied Signal Processing (special issue on MultimediaSecurity and Rights Management), 2004:14, pp. 2153-2173, 2004.
GroupGroup--Oriented ForensicsOriented Forensics
Overcome the limitations of orthogonal fingerprinting– Recall: orthogonal FP treats everybody equallyRecall: orthogonal FP treats everybody equally– Orthogonal strategy has to suspect more to accurately find a
colluder
Colluders often come together in some foreseeable groups– Due to their geographic, social, or other connections
Our approach: design users’ FP in a correlated way– Cluster users into groups based on prior knowledgeg p p g
Intra-group collusion is more likely than inter-group
– Revise orthogonal FP and add correlation to the same group to
Digital Fingerprinting for Multimedia Forensics 31
g g phelp narrow down the suspicion group
Proposed Group FingerprintingProposed Group FingerprintingDesign of collusion-resistant fingerprinting systems:
Design of anti-collusion fingerprints to trace traitors and colluders
2
xsy ijij
Design of detection schemes
||||energy equal ;,,...,1for ),,0(~)( 2
sss liNiNid
lmij
d
Solution: construct intra-group FP in two parts and use threshold detector
Digital Fingerprinting for Multimedia Forensics 32
Solution: construct intra-group FP in two parts, and use threshold detector (at desired intra-group false alarm) to avoid estimating ki
Group Fingerprint DesignGroup Fingerprint Design
Orthogonal modulation between groups– Design L orthogonal sub-systems to represent independent groupsg g y p p g p– M users per group => Total: n = M x L users
Assumption: users in the same group are equally likely to p g p q y ycollude with each other.
Real-valued code modulation within a groupd l l i i hi 1 – Introduce equal correlation within a group
1, 2[ ,..., ]the correlation matrix of { } is
i i iM
i j s
S s s ss R
11
1
s
R
– Each fingerprint consists of common one and individual one:
,{ }i j s 1
)0(~}{where1 21 Niid Iaeeaes
Digital Fingerprinting for Multimedia Forensics 33
),0( },,...,{ where,1 1 NuiiMiiijij Niid Iaeeaes Can be viewed as a real-valued fingerprint code
ExampleExample Basic idea: first identify groups containing colluders,
then identify colluders within each possible guilty group @
ROC Curves Pd vs. Pfp under different collusion settingsConstraint: equal energy 22
02 ||||}||{||}||{|| syy EE c
Digital Fingerprinting for Multimedia Forensics 36
Collusion Resistance of Group FP: Collusion Resistance of Group FP: Kmax vs. n
310fpP
0.8dP Group: upper bound
Orth.
Kmax of the proposed scheme is larger than that of the orthogonal scheme (the solid line), when n is large.
Difference between the lower bound and upper bound is due to the fact that ki=K/| i| in our simulations (symmetric collusion pattern)
Digital Fingerprinting for Multimedia Forensics 37
ki K/| i| in our simulations (symmetric collusion pattern). The smaller the number of guilty groups, the better chance performance.
Extension: TreeExtension: Tree--based Fingerprint Designbased Fingerprint Design
Use tree structure to construct fingerprints combining shared and distinct components
Unified view of fingerprint construction using code modulation– With hierarchically organized basis vectors
Allow for real valued codes– Allow for real-valued codes
a aa1 a2
a11a12 a21 a22
a111 a112 a113a121 a122 a123 a211 a212 a213 a221 a222 a223
s111 s112 s113s121 s122 s123 s211 s212 s213 s221 s222 s223
Digital Fingerprinting for Multimedia Forensics 38
i1i2 i3 i4
C d d Fi i tiC d d Fi i tiCoded FingerprintingCoded Fingerprinting
Digital Fingerprinting for Multimedia Forensics 39
Recall: Modulation Schemes for FingerprintingRecall: Modulation Schemes for Fingerprinting
Spread-spectrum embedding– Provide very good tradeoff on imperceptibility and robustness, y g p p y ,
esp. under non-blind detection– Typical watermark-to-noise (WNR) ratio: -20dB in blind
detection 0dB in non-blind detectiondetection, 0dB in non-blind detection
Choice of modulation schemes# of fingerprints
Orthogonal modulation jj uw
B
# of fingerprints= # of ortho. bases
(Binary) coded modulation
for or
i
iijj b1
uw
1,0ijb 1ijb
Digital Fingerprinting for Multimedia Forensics 40
# of fingerprints >> # of orthorgonal bases
Assumptions & Goals for BonehAssumptions & Goals for Boneh--Shaw Codes (’95)Shaw Codes (’95)
Detectable vs. Undetectable 111 111 111000 000 111
Collusion Scheme: – For each bit position, colluders can manipulate theFor each bit position, colluders can manipulate the
detectable bit value to any possible state, even make it unreadable
M ki A i Marking Assumption– Users cannot change the state of an undetected bit without
rendering the object useless.g j
Goal:– Given the fingerprint string of a suspicious media, output
Digital Fingerprinting for Multimedia Forensics 41
g p g p , pone colluder.
Coded Fingerprints by BonehCoded Fingerprints by Boneh--Shaw ConstructionShaw Construction
[I]Primitive
[II]Enlarge via
[IV]Permute within
[III]Compose withPrimitive
Code
1 1 1 A
n-1 Step-I
Enlarge viaRepetition
Permute withinEach Block
Compose withRandom Code
A B B iD CStep-III
1 1 1 A0 1 1 B0 0 1 C0 0 0 D
nA B B iD A C iiB A D iii: : : : :
DDC:
CAB:
C D B A C vii
111 111 111 A000 111 111 B000 000 111 C000 000 000 D
Step-II
111111111 000111111 000111111 000000 111000000 000[ ]A B B D C
111 111 111 A100 010 001 C
000 000 000 D
111 111 111 110 111 001 110 111 001 000 000 001010000 100[ ]Step-IV
Permute
Fingerprint code for User-i to insert into host datafingerprint
Digital Fingerprinting for Multimedia Forensics 42
g p
Coded Fingerprinting: Prior Work and New IssuesCoded Fingerprinting: Prior Work and New Issues
Collusion-secure codes by Boneh and Shaw ’98– Targeted at generic data with “Marking assumptions”g g g p
~ an abstraction of collusion model
– Codes are too long to be reliably embedded & extracted (Su et al.) ~ millions bits for 1000 users millions bits for 1000 users
– Focus on tracing one of the colluders
New issues with multimedia– “Marking assumptions” may no longer hold …– Some code bits may become erroneously decoded due to strong
noise and/or inappropriate embeddingpp p g– Can choose appropriate embedding to prevent colluders from
arbitrarily changing the embedded fingerprint bits
W ll d ibl
Digital Fingerprinting for Multimedia Forensics 43
Want to trace as many colluders as possible
Spreading + Combinatorial Coded FingerprintingSpreading + Combinatorial Coded Fingerprinting
Overall idea of embedded combinatorial fingerprinting– Explore unique issues associated with multimedia in fingerprint
encoding, embedding & detection– Use appropriate embedding to prevent arbitrary change on code
1st bit1st bit 2nd bit2nd bitB 1st bit1st bit 2nd bit2nd bit
......
B
iiijj b
1uw 1ijb
Build correlated fingerprints in two steps– Binary Anti-collusion fingerprint codes resist up to K colluders
any subset of up to K users share a unique set of code bits
– Use antipodal coded modulation to embed fingerprint codesi th l d t
Digital Fingerprinting for Multimedia Forensics 44
via orthogonal spread spectrum sequences shared bits get sustained and used to identify colluders
1616--bit ACC for Detecting bit ACC for Detecting 3 Colluders Out of 203 Colluders Out of 20
User-1 ( -1,-1, -1, -1, 1, 1, 1, 1, …, 1 ) ( -1, 1, 1, 1, 1, 1, …, -1, 1, 1, 1 ) User-4
Collude by AveragingUniquely Identify User 1 & 4
Embed fingerprint via HVS-based spread spectrum embedding in block-DCT domain
Digital Fingerprinting for Multimedia Forensics 45
Extracted fingerprint code ( -1, 0, 0, 0, 1, …, 0, 0, 0, 1, 1, 1 )
ACC Codes Under Averaging CollusionACC Codes Under Averaging Collusion
User-1 User-4 User-8
Averaging of multimedia domain leads to averaging in code-domain, and corresponds to AND operation after thresholding
Can distinguish colluded bits from sustained bits statistically with appropriate modulation and embedding and the sustained bits are unique with respect to
Digital Fingerprinting for Multimedia Forensics 46
modulation and embedding, and the sustained bits are unique with respect to colluder set
AntiAnti--Collusion Fingerprint CodesCollusion Fingerprint Codes
Simplified assumption (can be relaxed by using soft detection)– Fingerprint codes follow logic-AND operation after collusiong p g p
K-resilient AND ACC code– A binary code C={c1, c2, …, cn} – The logical AND operation of any combination of up to K codevectors is
distinct from the AND of any other combinations of up to K codevectorsExample: {(1110), (1101), (1011), (0111)}
1111000
ACC code via combinatorial design– Balanced Incomplete Block Design (BIBD)
101010111001101111000
Si l E l
010110101100110011110CSimple Example
ACC code via (7,3,1) BIBD for handling up to 2 colluders among 7 users
Digital Fingerprinting for Multimedia Forensics 47
1001011
Balanced Incomplete Block Design (BIBD)Balanced Incomplete Block Design (BIBD)
Construction example of (7,3,1) BIBD code– X={1,2,3,4,5,6,7}
1 2 3 4 5 6 7x x x{ , , , , , , }
– A={123, 145, 246, 167, 347, 257, 356}
( k ) BIBD i (k 1) ili t AND ACC
x x xx x x
x x x (v,k,)-BIBD is an (k-1)-resilient AND ACC
– Defined as a pair (X,A) X is a set of v points
x x xx x x
x x xf p
A is a collection of blocks of X, each with k points every pair of distinct points is in exactly blocks
vv2
x x x
– # codewords = # blocks
Code length for n=1000 users: O( n0.5 ) ~ dozens-to-hundreds bits6
kkvvn 2
Digital Fingerprinting for Multimedia Forensics 48
– Shorter than prior art by Boneh-Shaw O( (log n)6) ~ millions bits
Colluder Detector Design: Two ApproachesColluder Detector Design: Two Approaches
Hard Detection:– Detect the bit values and then estimate colluders from these valuesDetect the bit values and then estimate colluders from these values– Uses the fact that the combination of codevectors uniquely identifies
colluders– Everyone is suspected as guilty and each ‘1’ bit narrows down set
Soft Detection – possible candidates for soft detection:– Sorting: Use the largest detection statistics to optimize likelihood
function to first determine bit values, then estimate colluder set.
– Likelihood-based sequential method: Iteratively update the likelihood function and directly identify the colluder set.
Digital Fingerprinting for Multimedia Forensics 49
ACC Experiment with Gaussian SignalsACC Experiment with Gaussian Signals
– Higher threshold captures more colluders, but suspects more innocents– Soft decoding gives more accurate colluder identification than hard decoding– Joint decoding and colluder identification gives better performance than
separating the two steps
Digital Fingerprinting for Multimedia Forensics 51
– Sequential colluder identification (Simp. AM) gives a good tradeoff between performance and computational complexity
Digital Fingerprinting Example for Tracing TraitorDigital Fingerprinting Example for Tracing Traitor
Our software toolbox demonstrates
llcollusion-resistant fingerprinting of an aerial image captured by UAV Predator overby UAV Predator over Pentagon
Digital Fingerprinting for Multimedia Forensics 53
Recent Work: Exploiting ECC for FingerprintingRecent Work: Exploiting ECC for Fingerprinting
Examines Error Correcting Codes (ECC) based fingerprinting based ona joint coding and embedding framework– Symbol-based fingerprint codes are attractive for frame-based multimedia
(such as video and audio)– Identify the performance advantage and disadvantage of ECC based
fingerprinting compared with Orthogonal fingerprinting
Proposed two new joint coding and embedding techniquesP t d b t b ddi– Permuted subsegment embedding
– Group based joint coding and embedding (GRACE) fingerprinting
Main Results (see ICASSP’05 Monday afternoon session)Main Results (see ICASSP 05 Monday afternoon session)– Substantially improve collusion resistance of ECC based fingerprinting
and provide better trade-off b/w collusion resistance & efficient detectionDemonstrates advantages by joint coding and embedding in fingerprint
Digital Fingerprinting for Multimedia Forensics 54
– Demonstrates advantages by joint coding and embedding in fingerprint design
Similarity between Collusion and MU Comm.Similarity between Collusion and MU Comm.
The Fingerprint Collusion Problem is similar to Multiuser Communication– The colluded signal is simply the host signal plus a mixture of watermarksg p y g p
Collusion Fingerprint Problem Synchronous CDMA Channel
colluderaisuserjthif1}1,0{1
wy
j
n
jjjc d
K
1,1
)()()(1
b
tntsbAty
k
n
kkkk
F d i ti f CDMA h ld h
tsfingerprinj wj
sequencessignature)( tsk
k
For good communication performance: CDMA sequences should have minimum interference between each other. Low Cross-Correlation is Good!
The similarity between Collusion and MU Comm. suggests that good CDMA ld b d fi i t !
Digital Fingerprinting for Multimedia Forensics 55
sequences would be good fingerprints!
ACC Built from Interference AvoidanceACC Built from Interference Avoidance
How to assign M fingerprints in N dimensions to facilitate colluder detection?
M ≤ N: assign orthogonal fingerprints because they are uncorrelated
M > N: the fingerprints are correlated. How do we find the least correlated set S of size N×M?– Minimize Total Squared Correlation (TSC):
n
i
n
jj
TiTSC
1 1
2)( ss– Welch Bound: TSC is lower bounded by M2/N
– WBE sequence set:N
MTSCNMT
2
NISS
i j1 1
– WBE sequence set is known to be optimal in terms of user capacity in synchronous code-division multiple access (CDMA)
NN
Digital Fingerprinting for Multimedia Forensics 56
y p ( )
– One approach to get WBE sequence set: Eigen-algorithm
Detection of WBE FingerprintsDetection of WBE Fingerprints
Vector collusion model– Project the signals, noise, and watermarks into the watermark spacej g , , p
nSK
T T: sufficient statistics for detection
S: code matrix: the coll sion indicator ector
– Goal: estimate from the sufficient statistics T: the collusion indicator vector
Maximum likelihood detection
2||||minargˆ ST
– Minimize the distance between the observed vector T and
||||minarg
SK
T
SK
Digital Fingerprinting for Multimedia Forensics 57
– Integer least square problem… Sphere decoding may be used!K
WBEWBE--ACC Performance Compared with BIBD ACCACC Performance Compared with BIBD ACC
Spreading factor N=10000 (i.e. length of basis). Gaussian i.i.d noise 20 users, 16 dimensions, 3 colluders20 users, 16 dimensions, 3 colluders WBE fingerprints scheme shows improvement in both Pd and Pfa
P b bilit f t d t ti P b bili f f l i1
0 2
0.25ACC-SortingACC-AM WBE S h
Probability of correct detection Pd Probability of false accusation Pfa
0.8
0.9
p d
0 1
0.15
0.2
p fa
WBE-Sphere
0.7 ACC-SortingACC-AM WBE-Sphere
0.05
0.1
Digital Fingerprinting for Multimedia Forensics 58
-25 -23 -21 -19 -17 -150.6
WNR-25 -23 -21 -19 -17 -150
WNR
A Unified FrameworkA Unified Framework
Unified view of fingerprint construction using code modulation
O th l fi i ti B1. Orthogonal fingerprinting:
– Equivalent code matrix B = identity matrix
2. Coded fingerprinting: a compact way to represent users
i
iijj ubw1
2. Coded fingerprinting: a compact way to represent users– Boneh-Shaw’s and error correction code based construction – Combinatorial based ACC code: take advantage of shared bits
3. General correlated fingerprints– Allow real-valued codes such as in group-oriented fingerprinting
Key: how to strategically introduce correlation and separationbetween codes to capture individual and multiple colluders
Ref: M Wu W Trappe Z Wang and K J R Liu: "Collusion Resistant Fingerprinting for
Digital Fingerprinting for Multimedia Forensics 59
Ref: M. Wu, W. Trappe, Z. Wang, and K.J.R. Liu: Collusion Resistant Fingerprinting for Multimedia", IEEE Signal Processing Magazine, Special Issue on Digital Rights Management, pp.15-27, March 2004.
Conclusions from AntiConclusions from Anti--Collusion FP StudiesCollusion FP Studies
Important to design anti-collusion fingerprint for multimediaC ll i i t ff ti tt k i t fi i ti– Collusion is a cost-effective attack against fingerprinting
– Anti-collusion fingerprints can allow us trace traitor and deter unauthorized information leakage
Good news– We can tolerate about a few dozens colluders– We can accommodate more users through use of ACC
ChallengesO fi d h ll d t i t th t– One may find enough colluders to circumvent the system
– More research will lead to better designs and increase the collusion resistance
Digital Fingerprinting for Multimedia Forensics 60
Data Hiding in Curves forData Hiding in Curves forg fg f
Fingerprinting Map DocumentFingerprinting Map Document
Digital Fingerprinting for Multimedia Forensics 61
Interested in Curves?Interested in Curves?
Curve in our everyday life– Appear in a number of documents, e.g. maps and art/engr designpp , g p g g– Increasing popular with writings/drawings from pen-based input
devices (such as Tablet PC)
Storage formats of curves– Raster format
Represent as an image with (binary) intensity Represent as an image with (binary) intensity information of a 2D array of pixels
– Vector format A piecewise linear approximation with such
geometric primitives as vertex points and lines
– Conversion between different formats
Digital Fingerprinting for Multimedia Forensics 62
– Conversion between different formats
Protecting MapsProtecting Maps
Map represents important geospatial information – Ubiquitous in military, intelligence, and commercial operations.q y g p
Traditional way to protecting maps: add intentional errors– Wrongly spell the names; bend roads in a wrong way; add fake lakes/hills
Disadvantages of traditional protection– Substantially alter the geospatial
meanings conveyed by a mapmeanings conveyed by a map– For fingerprinting use, a small amount
of intentional errors can be quite easily identified and removed from severalidentified and removed from several fingerprinted copies of a map
Need a modern protection
Digital Fingerprinting for Multimedia Forensics 63
– Less intrusive; resist collusion & format conversion
Digital Fingerprinting of MapsDigital Fingerprinting of Maps
Fingerprinting for tracing traitors– Different fingerprints are embedded in different copies beforeDifferent fingerprints are embedded in different copies before
distributing them to a number of recipients– For discouraging illegal distribution and tracing traitors in case of
information leakinginformation leaking
Challenge of fingerprinting mapsThe fingerprint must be embedded in a visually non intrusive way– The fingerprint must be embedded in a visually non-intrusive way
– Embedded fingerprints should not be obliterated by attacks multi-user collusion attacks
i di i i li d l i geometric distortions, e.g., rotation, scaling, and translation format & D/A-A/D conversion, such as printing-scanning
– Good news: the original unmarked version is usually available to
Digital Fingerprinting for Multimedia Forensics 64
Good news: the original unmarked version is usually available to detector
Existing Work on Watermarking MapsExisting Work on Watermarking Maps
Watermarking map images: text-based geometric normalization
Watermarking vector maps Watermarking vector maps– Fourier descriptors of polygonal lines: have visible distortions
– Spectral analysis of mesh models Complex and cannot be easily applied to maps beyond urban areas
– 3-D watermarking works using mesh and surface modeling Provide enlightening analogy for 2-D work Very limited work on robust approach and validation results
for 2-D, esp. on collusion-resistant fingerprinting & format conversion
Watermarking general binary images Watermarking general binary images– Often rely on minor manipulation of pixels– Fragile => difficult to deal with RST issues and survive D/A-A/D
Of d f h i i d i
Digital Fingerprinting for Multimedia Forensics 65
– Often used for authentication and annotation
New Approach to Fingerprinting MapsNew Approach to Fingerprinting Maps
Select fingerprinting domain and embedding method to address the imperceptibility and robustness requirement
Feature domain: coordinates of control points in the B-spline representation of curvesrepresentation of curves– These form salient features representing the shape of the curve
Embedding: spread spectrum embedding & correlation detection Embedding: spread spectrum embedding & correlation detection– Good tradeoff between imperceptibility/robustness (esp. nonblind detection)– The general collusion resistance have been studied for independent and
d d fi i ti i d t b ddicoded fingerprinting using spread spectrum embedding orthogonal fingerprinting is good for a moderate size of user group
Resist printing scanning & D/A A/D: b hi h i i i t ti
Digital Fingerprinting for Multimedia Forensics 66
Resist printing-scanning & D/A-A/D: by high precision registration
BB--splines Representation of Curvessplines Representation of Curves
[ ],
0( ) ( )
nB
i i ki
t B t
p c0i
BB--spline spline Basis FunctionBasis FunctionControl Control
PointsPoints
Curve Curve PointsPoints
B-splines: Piecewise polynomial functions giving local approximation for a curve w/ small# parameters B-spline basis functions are defined recursively
)()()()( 111 tBtttBtt kikikii
otherwisettt
tB iii ,0
,1)( 1
1,
1,)()()()(
)(1
1,1
1
1,,
ktt
tBtttt
tBtttB
iki
kiki
iki
kiiki
Control points
Digital Fingerprinting for Multimedia Forensics 67
Provide weights on the basis function to locally control the shape of the curve rendered
Control Points of BControl Points of B--spline Representationspline Representation
Least-square approach to estimate control points – Given a set of properly chosen samples on the curve, control p p y p ,
points can be estimated using the least squares technique
Properties of control points Properties of control points– Serve as a compact set of salient
features for curves => not affectedby noisy scanning boundaries
– Provide local control on shape– Invariant to affine transformations– Invariant to affine transformations
Digital Fingerprinting for Multimedia Forensics 68
Data EmbeddingData Embedding
Spread spectrum additive embedding– Use noise-like sequences as digital fingerprints– Represent different users using (approximately) orthogonal sequences
May generate independent seq. via pseudorandom number generator
– A scaled version of the fingerprint sequence is added to the coordinates ofA scaled version of the fingerprint sequence is added to the coordinates of the set of control points
A fingerprinted curve can be obtained by synthesizing from the fi i t d t l i tfingerprinted control points
Fingerprint SequenceEmbedded
' c c w
Original Curve
MarkedControl Points
MarkedCurve
OriginalControl Points
Digital Fingerprinting for Multimedia Forensics 69
'x x x
y y y
c c wc c w
Data DetectionData Detection
Register with the original unmarked curve Extract control points of the test curve and compute the changes to Extract control points of the test curve and compute the changes to
arrive at an estimated fingerprint sequence
OriginalControl Points
Test CurveRegistered
TestControl Points
Control Points
SpecificUser
Correlation-based Threshold Testing
Fingerprint SequenceEstimated
( ) /
Correlation-based statistics and threshold testing
( ) /( ) /
x x x
y y y
w c cw c c
23)1(2
11log
n
Z : correlation coefficient
n: number of control points
Digital Fingerprinting for Multimedia Forensics 70
Follow approx. unit-variance Gaussian distribution (threshold: around 3~6)
Fidelity and RobustnessFidelity and Robustness
Where does the robustness come from?– Control points represent salient features of curves: analogous to p p g
perceptually significant spectral components in images– Spread spectrum embedding
S li f t Scaling factor α– Larger α larger distortion, more robustness
determined by application’s requirement; we used 0.5 in our tests
The number of control points– Determine the length of the fingerprint sequence
Use an appropriate number of control points to fit the original curve well– Use an appropriate number of control points to fit the original curve well Too few: details of the curve lost; Too many: over-fitting The number of control points is about 5~8% of the total number of
curve pixels in our tests
Digital Fingerprinting for Multimedia Forensics 71
curve pixels in our tests– Place knots and control points according to the curvature
Fingerprinting Simple CurvesFingerprinting Simple Curves
Original Curve
(captured by TabletPC)
Detection Statistics
Typical threshold is 3~6 for f l l f 10 3 10 9
Fingerprinted Curve
(100 control points)
Digital Fingerprinting for Multimedia Forensics 72
( p y ) false alarm of 10-3 ~ 10-9( p )
Fingerprinting Topographic MapFingerprinting Topographic Map
1100x1100 Original Map Fingerprinted Map
9 curves that are long enough are marked
Digital Fingerprinting for Multimedia Forensics 73
9 curves that are long enough are markedTotally, there are 1331 control points used to carry the fingerprint
Fidelity DemonstrationFidelity Demonstration
original and fingerprinted curves are overlaidoriginal: blue; marked: red
Digital Fingerprinting for Multimedia Forensics 74
original: blue; marked: red
Resistant to Collusion AttacksResistant to Collusion Attacks
2-User Interleaving Attack5-User Averaging Attack
. . .
Digital Fingerprinting for Multimedia Forensics 75
Resistant to PrintingResistant to Printing--andand--scanningscanning
Original Curve
Detection Statistics
( 100 control points;l i t ti )manual registration )
Digital Fingerprinting for Multimedia Forensics 76
Fingerprinted curve after Printing-and-Scanning
Automated HighAutomated High--Precision RegistrationPrecision Registration
Inherent non-uniqueness of B-spline control points– A curve can be well approximated by different sets of control points
Iterative Alignment-Minimization (IAM) algorithm– Formulate jointly undoing geometric transform
d i i h i i l f l iand retrieving the original set of control points as an optimization problem
Test Control Points
Step 1Step 1 Step 2Step 2
Initial Estimationof Sample Points
Test RasterCurve
EstimatedSample Points
Super Curve Alignment
Step 1Step 1 Step 2Step 2
Original Ali d
Projection
Digital Fingerprinting for Multimedia Forensics 77
OriginalSample Points Nearest Neighbor
RuleAligned
Raster Curve
Step 3Step 3
An Example Using IAM: Aligning Sample PointsAn Example Using IAM: Aligning Sample Points
Original and test curve Test sample points Test sample points
Digital Fingerprinting for Multimedia Forensics 82
Original and test curve Test sample points after 1 iteration
Test sample points after 15 iterations
An Example Using IAM: (cont’d)An Example Using IAM: (cont’d)
iterations iterations
Transform Parameters Detection Statistics
Digital Fingerprinting for Multimedia Forensics 83
Resilience to Collusion plus PrintingResilience to Collusion plus Printing--andand--ScanningScanning
Detection statistic for one-user print-scan is 24.93 on the 9-curve map Four user averaging collusion + printing-scanningg g p g g
Print using a HP laser printer Scan back by a Canon scanner with 360dpi resolution
@
Digital Fingerprinting for Multimedia Forensics 84
@
Summary and Conclusion on Curve FingerprintingSummary and Conclusion on Curve Fingerprinting
Important to design anti-collusion fingerprint for multimedia– Collusion is a cost-effective attack against fingerprintingCollusion is a cost-effective attack against fingerprinting– Anti-collusion fingerprint can allow us trace traitor and deter
unauthorized information leakage
New data hiding technique for curves – Parameterize curves using the B-spline model and embed data
through spread-spectrum based perturbation on control pointsthrough spread spectrum based perturbation on control points– High fidelity in embedding (w.r.t. the original curve)– Resilient to collusion and printing-and-scanning attack
Potential applications for protecting maps and drawings– Fingerprinting for traitor tracing
Digital Fingerprinting for Multimedia Forensics 85
Secure and Efficient Multicast Distribution ofSecure and Efficient Multicast Distribution ofSecure and Efficient Multicast Distribution of Secure and Efficient Multicast Distribution of Fingerprinted VideoFingerprinted Video
Digital Fingerprinting for Multimedia Forensics 86
MotivationMotivation
Streaming applications with traitor tracing requirement Distribution of fingerprinted multimedia over networks Distribution of fingerprinted multimedia over networks
– Unicast: not scalable– Traditional multicast technology: the same data to multiple users
Goal of the streaming service provider: – Reduce the communication cost– Accommodate as many users as possible– Protect the secrecy of the multimedia content as well as the embedded
fingerprintsfingerprints
Design secure fingerprint multicast schemes!Design secure fingerprint multicast schemes!
Digital Fingerprinting for Multimedia Forensics 87
Previous WorkPrevious Work
Watercast (Brown et al. 99)– Trusted intermediaries embed their unique IDs before forwarding the packets
through the network
WHIM (Judge et al. 00)Trusted routers forward differently fingerprinted packets to different users– Trusted routers forward differently fingerprinted packets to different users
Selective Encryption (Wu et al. 97)– Fingerprints were embedded in DC coeff. of I frames– Selectively encrypt and unicast part of the MPEG stream to each user
Multicast Watermark Protocol (Chu et al. 02)A li d h Sh fi i d d i f b f b i– Applied Boneh-Shaw fingerprint code design on a frame by frame basis
Joint fingerprinting and decryption (Kundur et. al. 04)– Fingerprint design similar to Boneh-Shaw, on a pixel by pixel basis
Digital Fingerprinting for Multimedia Forensics 88
Fingerprint design similar to Boneh Shaw, on a pixel by pixel basis
New Issues on Fingerprint MulticastNew Issues on Fingerprint Multicast
Prior work: – Only applicable to trace a few colludersOnly applicable to trace a few colluders– Adjust fingerprint design to fit the communication framework
New issues: – Video streaming applications: a large number of users, and
t ti ll l b f ll dpotentially a large number of colluders– Consider applications requiring strong traitor tracing capability – Utilize the multicast technology to fit the fingerprint designUtilize the multicast technology to fit the fingerprint design
Digital Fingerprinting for Multimedia Forensics 89
A General Fingerprint Multicast SchemeA General Fingerprint Multicast Scheme
Spread spectrum embedding– Resistant to many attacks y– Not all coefficients are embeddable due to perceptual constraints
An example: percentage of embeddable coefficientsMi A i 18% C h 28% Fl 34%Miss America: 18%, Carphone: 28%, Flower: 34%
– A non-embeddable coefficient has the same value in all copies.
General fingerprint multicast scheme:General fingerprint multicast scheme:– Reduce the comm. cost in distributing the non-embeddable coefficients
Shared non-embeddable coefficientsmulticasti l fi i d ffi i i Uniquely fingerprinted coefficients unicast
– Can be used with most spread spectrum embedding based fingerprinting systems
Digital Fingerprinting for Multimedia Forensics 90
A General Fingerprint Multicast Scheme (cont’d)A General Fingerprint Multicast Scheme (cont’d)
AliceAlice
embedembed
embedembedBob
Bob
The Lord of the Rings
Chris
embedembed
Chris
Digital Fingerprinting for Multimedia Forensics 91
MPEGMPEG--2 Based General Fingerprint Multicast Scheme2 Based General Fingerprint Multicast Scheme
M ti P iti f b dd bl ffi i t Motion vectors, quan. factors, other side info. N
VLC
Positions of embeddable coefficients Multicast to all users
mK
Encrypt
VLC-1
Coded DCT
Q-1 VLCQEmbeddable?
NY
Unicast to user u(i)
( )iK
mKEncrypt
User u(i)’s fingerprint
coeff.K
for user u(i)
The Fingerprint Embedding and Distribution Process for Video on Demand Applications at
Digital Fingerprinting for Multimedia Forensics 92
Process for Video on Demand Applications at the Server’s Side
MPEGMPEG--2 Based General Fingerprint Multicast Scheme2 Based General Fingerprint Multicast Scheme
Multicasted mKVLC-1 Q-1
VLC-1 Q-1
Multicasted DCT coeff.
Decrypt
Decrypt
mKReconstruct
the 8*8bl k
IDCT
0C Q
Unicasted DCT coeff.
yp( )iK
DCT block
Positions of b dd bl
Motion compensation
embeddable coefficients Motion vectors
The Decoding Process at the Receiver’s Side
Digital Fingerprinting for Multimedia Forensics 93
Bandwidth EfficiencyBandwidth Efficiency
Measurement of bandwidth efficiency:– General Internet Routing: hop-based link usageGeneral Internet Routing: hop-based link usage
– Compare the communication cost of two schemes: Pure unicast scheme: General fingerprint multicast scheme:
puCfmC
– Communication cost ratio: The smaller the communication cost ratio, the more efficient
pufmfm CC /
the general fingerprint multicast scheme.
Digital Fingerprinting for Multimedia Forensics 94
Tested Video SequencesTested Video Sequences
Miss America Carphone Flower
R=1.3 bpp
Digital Fingerprinting for Multimedia Forensics 95
Communication Cost RatioCommunication Cost Ratio
Digital Fingerprinting for Multimedia Forensics 96
M users; Rate =1.3 bpp.
Joint Fingerprint Design and DistributionJoint Fingerprint Design and Distribution
General fingerprint multicast scheme:– Utilizes the perceptual constraints on spread spectrum embeddingUtilizes the perceptual constraints on spread spectrum embedding– Multicast the non-embeddable coefficients to all users– Can be used with most spread spectrum embedding based
fingerprinting systems
A joint fingerprint design and distribution scheme– Explores the special structure of the fingerprint designExplores the special structure of the fingerprint design– Also multicast some fingerprinted coefficients that are shared by a
subgroup of users to them
Digital Fingerprinting for Multimedia Forensics 97
– Further improve the bandwidth efficiency
GroupGroup--Oriented Fingerprint DesignOriented Fingerprint Design
Motivation:– Some users are more likely to collude with each other than theSome users are more likely to collude with each other than the
others due to geographical or social reasons– Tree structure to explore the hierarchical relationship between
users
Group oriented fingerprint design1 2a– Group users who are more
likely to collude witheach other together
1a a
1,1a 1,2a 2,1a 2,2aeach other together
– More robust against collusion attacks
1,1,1a 1,1,2a 1,1,3a 1,2,1a 1,2,2a 1,2,3a 2,1,1a 2,1,2a 2,1,3a 2,2,1a 2,2,2a 2,2,3a(1)u (2)u (3)u (4)u (5)u (6)u (7)u (8)u (9)u (10)u (11)u (12)u
Digital Fingerprinting for Multimedia Forensics 98
1,1U 1,2U 2,2U2,1UBy Wang et. al. ’04
Joint Fingerprint Design and DistributionJoint Fingerprint Design and DistributionUser u(1) X1
(1) X2 (1) X3
(1)
S S S
=
1a 2a
S1 S2 S3
a11 a2
1 a31
+
1,1a 1,2a 2,1a 2,2a
1,1,1a 1,1,2a 1,1,3a 1,2,1a 1,2,2a 1,2,3a 2,1,1a 2,1,2a 2,1,3a 2,2,1a 2,2,2a 2,2,3a
a21,1 a3
1,1
a1,1,1
a a a a a a a a a a a a(1)u (2)u (3)u (4)u (5)u (6)u (7)u (8)u (9)u (10)u (11)u (12)u
User u(2) X1 (2) X2
(2) X3 (2)
S1 S2 S3
=
Host Signal S
S1 S2 S3
a11 a2
1
a21,1
a31
a31,1
+
Digital Fingerprinting for Multimedia Forensics 99
1 2 3
a1,1,2
Joint Fingerprint Design and Distribution (cont’d)Joint Fingerprint Design and Distribution (cont’d)
User u(1)
S1+a1 S2+a1+a1,1 S3+a1+a1,1+a1,1,1
Multicast to all users
Multicast to u(1)~u(6)
Unicast to u(1)
Unicast to u(2)Multicast to u(1)~u(3)
1a 2a
1,1a 1,2a 2,1a 2,2a
S1+a1 S2+a1+a1,1 S3+a1+a1,1+a1,1,2
1,1,1a 1,1,2a 1,1,3a 1,2,1a 1,2,2a 1,2,3a 2,1,1a 2,1,2a 2,1,3a 2,2,1a 2,2,2a 2,2,3a(1)u (2)u (3)u (4)u (5)u (6)u (7)u (8)u (9)u (10)u (11)u (12)u
User u(2)
1 2 3u u u u u u u u u u u u
Digital Fingerprinting for Multimedia Forensics 100
Secure Fingerprint Multicast: PerformanceSecure Fingerprint Multicast: PerformancePure Unicast General Fingerprint
MulticastJoint Fingerprint Design
and Distribution
Fi i d i A M d T b d fi iFingerprint design Any Most spread spectrum embedding based
Tree-based fingerprint design
Comm. cost ratio 1 16% ~ 52% 13% ~ 39%
Computation complexity
MG: 0RL: 1
MG: 1RL: 2
MG: 13 ~ 125RL : 4 ~ 5
Collusion resistance Same under fairness constraintsCollusion resistance Same under fairness constraintsSuggested
ApplicationsSmall
number of users
Large number of users, sequences with fewer
embeddable coeff.
Large number of users, sequences with much
more embeddable coeff. (e.g., miss america) (e.g., flower)
–Total number of users: 1000 ~ 10000.
Digital Fingerprinting for Multimedia Forensics 101
–MG: total number of multicast groups in the system–RL: number of channels that each receiver has to listen to
Summary and Future Work on Fingerprint MulticastSummary and Future Work on Fingerprint Multicast
Summary– Reduce the bandwidth requirement by 48% ~ 87%Reduce the bandwidth requirement by 48% ~ 87% – Protect the secrecy of the video content as well as that of the
fingerprints– Preserve the collusion resistance of the embedded fingerprints
k Future work– Other important aspects of the general fingerprint multicast:
Quality of service management Quality of service management Fingerprint multicast over heterogeneous networks to users
with different processing capability
Digital Fingerprinting for Multimedia Forensics 102
Tutorial Summary and Recent PublicationsTutorial Summary and Recent Publications
Understanding the linear and nonlinear collusions:H. Zhao, M. Wu, Z. Wang, and K.J.R. Liu: "Nonlinear Collusion Attacks on Independent Multimedia Fingerprints," IEEE Trans. on Image Proc., to appear in May 2005.Multimedia Fingerprints, IEEE Trans. on Image Proc., to appear in May 2005.
Analyzing the limit on collusion resistance by orthogonal fingerprinting:Z. Wang, M. Wu, H. Zhao, W. Trappe, and K.J.R. Liu: "Collusion Resistance of Multimedia Fingerprinting Using Orthogonal Modulation", IEEE Trans. on Image Proc., to appear in June 2005.
Leveraging prior info. on collusion pattern via group fingerprinting:Z. Wang, M. Wu, W. Trappe, and K.J.R. Liu: "Group-Oriented Fingerprinting for Multimedia Forensics" EURASIP Journal on Applied Signal Processing Special Issue on Multimedia Forensics , EURASIP Journal on Applied Signal Processing, Special Issue on Multimedia Security and Rights Management, Nov. 2004.
Proposing a new framework promoting the joint coding, embedding, and detection, with a demonstration through new combinatorial based ACC, g
W. Trappe, M. Wu, Z. Wang, K.J.R. Liu, “Anti-Collusion Fingerprinting for Multimedia,” IEEE Trans. on Sig. Proc., Special issue on Data Hiding in Digital Media & Secure Content Delivery, 51 (4), pp.1069-1087, April 2003.
Explorations on ECC fingerprinting map/curve FP & FP multicast
Digital Fingerprinting for Multimedia Forensics 103
Explorations on ECC fingerprinting, map/curve FP, & FP multicastSee papers in ICASSP 2005, & upcoming papers in Trans. Sig/Image Proc.
Coming soon 200Fall 2005
Chapter 1 IntroductionChapter 2 Preliminaries on Data EmbeddingChapter 3 Collusion AttacksChapter 4 Orthogonal Fingerprinting and Collusion ResistanceChapter 4 Orthogonal Fingerprinting and Collusion ResistanceChapter 5 Group-Oriented FingerprintingChapter 6 Anti-Collusion Coded (ACC) FingerprintingChapter 7 Secure Fingerprint Multicast for Video StreamingChapter 8 Fingerprinting Curves
To be published by Hindawi Publishing in Fall 2005
Digital Fingerprinting for Multimedia Forensics 104
To be published by Hindawi Publishing in Fall 2005EURASIP Book Series on Signal Processing and Communications
Contact Information of Tutorial Instructors
K.J. Ray Liu [email protected]://www.isr.umd.edu/Labs/CSPL/kjrliu/kjrliu.html
Wade Trappe [email protected] http://www.winlab.rutgers.edu/~trappe/
j @ bZ. Jane Wang [email protected]://www.ece.ubc.ca/~zjanew/
Min Wu minwu@eng umd eduMin Wu [email protected]://www.ece.umd.edu/~minwu/
Digital Fingerprinting for Multimedia Forensics 105