Professional Documents
Culture Documents
Digital Watermarking: An Introduction: Edward J. Delp
Digital Watermarking: An Introduction: Edward J. Delp
An Introduction
Edward J. Delp
Purdue University
School of Electrical and Computer Engineering
Purdue Multimedia Testbed
Video and Image Processing Laboratory (VIPER)
West Lafayette, Indiana
email: ace@ecn.purdue.edu
http://www.ece.purdue.edu/~ace
EPICS Spring 2003 Slide 1
Outline
• Provide an introduction to watermarking and data
hiding and overview its use
• Describe how security techniques may/will impact
multimedia systems
• Auditing (fingerprinting)
– Who did what and when?
Attacker Is this a
pirate
device?
Authentication
• Authentication
• Hashing
• Time-stamping
• Watermarking
• Scenario
– an owner places digital images on a network
server and wants to “protect” the images
• Goals
– verify the owner of a digital image
– detect forgeries of an original image
– identify illegal copies of the image
– prevent unauthorized distribution
Watermark
YD ( i ) = X D ( i )(1 + aW )
• T = user-defined threshold
• If S > T, image is authentic
X u ,v + J u ,vWu ,v , X u ,v > J u ,v
Yu ,v =
X u ,v , otherwise
Y1 = X + W1
– Create counterfeit original, XF
X F = Y1 − W2
– Y1 now appears to be a marked version of XF
⇒ Y1 = X F + W2
• In personal computing:
– RGB (Red, Green, Blue) color space is common
– Each color component assigned an integer value
between 0 to 255, or 8 bits
– (0,0,0) corresponds to darkest black; (255,255,255)
corresponds to brightest white
Red 1 1 1 1 0 0 0 0
Green 1 1 0 1 1 0 0 0
RGB values of
Blue 1 1 0 0 1 1 1 0 pixel
Index R G B
94 96 2 4 0 0 0 0
23 240 32 24 1 32 84 2
85 2 33 25
96 97 35 27 239 160 2 233
240 240 216 206
241 239 200 102 RGB values of
Pixel value == index pixel
254 255 255 255
in color table
255 128 127 124
Cover Image
Steganographic
Encoder
Secret (S-Tools)
Message
Stego Image
Stego Key
EPICS Spring 2003 Slide 56
Digital Image Steganography
• Cover Image = Original image
• Message = Data (files) to be hidden
• Stego Image = Altered image containing hidden data
• Stego Key = Secret needed to embed or recover message
Extracted
Message
Steganographic
Stego Image Decoder
(S-Tools)
Stego Image
Stego Key
EPICS Spring 2003 Slide 63
Message Processing
• Goal: Encrypt (scramble) the message
K
Passphrase
MD5 (128-bit)
K2
64-bit buffer
Initialization
Red 1 1 1 1 0 0 0 x
Green 1 1 0 1 1 0 0 x
RGB values of
Blue 1 1 0 0 1 1 1 x pixel
Color Index c c c c c x x x
RGB values of
pixel
Selects 1 of the 32 Set to match 3 bits
colors used in the of the encrypted
reduced color image message
• Decoding is straightforward
• StirMark
– http://www.cl.cam.ac.uk/~fapp2/watermarking/image
_watermarking/stirmark/
← Original “Girls”
Altered “Girls” →
?
coefficients are zero 4
5
6
7
• Non-zero coefficients are coded:
– Location (zig-zag)
– Quantization index/value
• Zero coefficients not coded
1 2 3 4 5 6 7 8 9 10 11 12 13
I B B P B B P B B P B B I
1 4 2 3 7 5 6 10 8 9 13 11 12
EPICS Spring 2003 Slide 101
Compressed Video Data
• I-Pictures: “JPEG”, no motion vectors
Header Encoded DCT Coefficients of Picture
• Disadvantages:
– Computationally expensive
– Compression can damage watermark
– Must insert watermark with excess strength
– Watermark embedder does not know compression
parameters
– Re-compression can degrade video further
EPICS Spring 2003 Slide 103
Watermarking Compressed Video
• Compressed-domain approach
– Partially decode the compressed video
– Insert watermark by altering syntactic elements of
video (such as DCT coeffs)
– Re-assemble the compressed video stream
• Disadvantages:
– Must parse compressed video data during watermark
embedding
– Watermark insertion is constrained by allowable
syntax / semantics of compressed video stream
EPICS Spring 2003 Slide 105
Watermarking Issues
• General watermarking issues:
– Capacity, robustness, perceptibility, security
– Synchronization
– Attacks
– Computational complexity
Original 3 +3 -2 -2 0 0 2 0
Coded Signal
Watermark +1 +1
Drift
Compensation
Signal -1 -1
Watermarked
Signal 3 +3 -1 -3 0 1 1 0
Reconstructed
Watermarked
Signal 3 6 5 2 2 3 4 4
• Drift compensation
Watermarked Attacked
EPICS Spring 2003 Slide 118
Synchronization Attack Example
• Temporal synchronization
– Initial synchronization
– Re-synchronization after bit/decoding errors
– Attacks: Frame deletion, insertion, transposition,
averaging
• Techniques
– Sliding correlator
– Templates
∑ ∑ ∑ W ( x , y , t )Y ( x − x , y − y , t − t )
x y t
0 0 0
– Disadvantages:
• Templates must be easily detectable, and thus vulnerable to
attack and removal
• Template embedding adds distortion in the watermarked
video, affecting perceptual quality
K0 K0 K0 K0
• Time-invariant watermark
– Very high temporal redundancy
– Temporal synchronization is trivial (not needed)
– Low security
K0 K1 K2 K3
K0 K1 K2 K3
K T-1
Watermark Temporal
Generator Watermark Key for Redundancy
Next Frame Control
K(t+1)
Finite State
Machine
time
a=8, ß=2 0 1 2 3 4 5 6 7 8 9 10 11 12 13
Numbers indicate frame index
EPICS Spring 2003 Slide 131
Frame Analyzer
• Analyzes the watermarked frames, output is a vector of
feature values
– Used with the state machine to generate the next
watermark when needed
– Allows the watermark to be video dependent
State 4 State 3
Ki+1 = Ki+c3Z Ki+1 = Ki+c2Y
Test Picture
Z(t)
Watermark Watermark
Detector Generator
Watermark
Memory
(Queue)
State,
Key
EPICS Spring 2003 Slide 134
Watermark Detection Protocol
• Detector does not know a, ß
• The detector tries the following keys:
– The initial key K0
– Every key value stored in the queue
100
90
80
70
60
50
40
30
20
10
0
30 / 5
30 / 3
30 / 1
150 / 5
150 / 3
150 / 1
300 / 5
300 / 3
300 / 1
150 / 10
300 / 10
30 / 10
100
90
80
70
60
50
40
30
20
10
0
150 / 10
300 / 10
30 / 10
150 / 5
150 / 3
150 / 1
300 / 5
300 / 3
300 / 1
30 / 5
30 / 3
30 / 1
100
90
80
70
60
50
40
30
20
10
0
30 / 10
150 / 10
300 / 10
150 / 5
150 / 3
150 / 1
300 / 5
300 / 3
300 / 1
30 / 5
30 / 3
30 / 1
100
90
80
70
60
50
40
30
20
10
0
150 / 10
300 / 10
30 / 10
30 / 5
30 / 3
30 / 1
150 / 5
150 / 3
150 / 1
300 / 5
300 / 3
300 / 1
Redundancy ( Period [pictures] / Repeat [pictures] )
100
90
80
70
60
50
40
30
20
10
0
150 / 10
300 / 10
30 / 10
150 / 5
150 / 3
150 / 1
300 / 5
300 / 3
300 / 1
30 / 5
30 / 3
30 / 1
1.5
0.5
0
0 50 100 150 200 250
P icture Number
EPICS Spring 2003 Slide 143
Combined Attack
5% 10% 25%
Percent Watermarked Pictures Detected
100
90
80
70
60
50
40
30
20
10
0
150 / 10
30 / 10
30 / 5
150 / 5
Redundancy ( Period [pictures] / Repeat [pictures] )