Collusion-Resistant Fingerprinting for Compressed Multimedia Avinash L. Varna1, Shan He1, Ashwin Swaminathan1, Min Wu1, Haiming Lu2 and Zengxiang Lu2 1
ECE department, Univ. of Maryland, College Park, USA
2 Research
Institute of Info. Tech., Tsinghua University, Beijing, China
International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2007
Digital Fingerprinting
Unauthorized leak of information
How to trace traitors? Î Digital Fingerprinting Simple and effective attack - collusion by a group of users
Collusion resistant fingerprinting for Multimedia
Uncoded Fingerprints [Wang et al. ‘05] Joint Coding and Embedding [He and Wu ‘06]
Mostly for uncompressed host signals
First work on collusion-resistant fingerprinting for compressed multimedia Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
2
Motivation for Compressed Domain Fingerprinting Cable TV distribution System Set Top Box Decrypt Decompress Embed Fingerprint
Compressed video
Pirated Copy Collude
Online Music/Video Store Compressed File User Database of compressed video
Online Store Server Embed Fingerprint
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
3
Issues w/ Fingerprinting Compressed Signals Quantization step size = 6 A
66
-36
114
B
66
-36
114
C
72
-42
108
Average
68
-40
112
Compress
66
-42
114
Extracted Fingerprint
0
0
0
Compressed Host
66
-42
114
Fingerprinted Copies
Fingerprinted Signal
66.66
-41.06 111.51
After Recompression
66
-42
114
Difficult to catch colluders Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
4
System Model Compressed File User Database of compressed video Compressed host signal S Sj = mΔ
Fingerprinted Copies {X( k ) }k∈Sc
Online Store Server Embed Fingerprint
Fingerprint Embedding
Δe
Gaussian noise n
User index i
Collusion Attack
Fingerprinted Signal X(i)
Compress
Compress
Colluded Signal
Δc
V
Attacked Signal Z
Focus on one DCT frequency band
Compression Æ Quantization
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
5
System Model
Compression by Fingerprint Embedder
Compression by colluders
Stronger Compression (Δe < Δ) Æ more distortion Weaker Compression (Δe > Δ) Æ increased bandwidth same level of compression (Δe = Δ)
Stronger Compression (Δc < Δ) Æ more distortion Weaker Compression (Δc > Δ) Æ increased bandwidth, higher probability of detection same level of compression (Δc = Δ)
Collusion attacks – averaging and other order statistics based attacks
Focus on averaging, median, and minimum
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
6
Orthogonal FP for Compressed Multimedia
Embedding
Extend Spread Spectrum based Orthogonal fingerprinting
Fingerprint components Wj(i) i.i.d. Gaussian ⎛ S + W (i ) ⎞ X = round ⎜ ⎟× Δ Δ ⎝ ⎠ (i )
Visual distortion constraint
2 (i ) ⎡ E X − S ⎤ < D(Δ ) ⎢⎣ ⎥⎦
Detection
Correlation based non-blind detection to catch one colluder with high confidence 1 q = arg max h(Z − S), W (i ) i =1,2,..., N M h – preprocessing
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
7
Results for Orthogonal Fingerprints
Results independent of host distribution
System Parameters
104
fingerprint length =
D(Δ) = 15, WNR=0 dB
Minimum Uncompressed Δ=1 Median Averaging Δ=4Minimum
0.8
Δ ↑ ⇒ Pd ↓ Averaging is the most effective attack Low collusion resistance
Δ=6Median Averaging
0.6
Pd
1
1024 users
Around 7 for Δ = 6 (QF = 75)
0.4
0.2
0
5
10
15 20 20 No. of colluders
25 25
30 30
Δ=1 Averaging Δ=6 Collusion
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
8
A Closer Look at Colluded Fingerprints 1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
-12
-6
0
6
12
0 -12 -9
Embedded Fingerprint Δ=6
0
3
6
9
12
Averaging Δ=6
0 -24
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
-10
0
10
Embedded Fingerprint Δ=1
-3
1
0
-6
0 -4
-2
0
2
Averaging Δ=1
4
0
-18
-12
-6
0
Minimum Δ=6
-12
-10
-8
-6
6
12
25 colluders
-4
Minimum Δ=1
Averaging more effective than minimum Δ = 6 - Embedded fingerprint is discrete Î easily removed after collusion Δ = 1 - Embedded fingerprint more continuous Î more difficult to remove Need to make the fingerprint more continuous Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
9
Anti-Collusion Dither (ACD) Add a dither (random) sequence before embedding
Make compressed host more continuous
⎛ S j + d j + W j(i ) ⎞ ⎡ −Δ +Δ ⎤ X ' = round ⎜ d × Δ ∈ ; , ⎟⎟ j ⎢ ⎜ Δ ⎣ 2 2 ⎥⎦ ⎝ ⎠ (i ) j
Embedded Fingerprint also more continuous
0.7
0.7
0.5
0.5
0.3
0.3
0.1
0.1
0
-12
-6
0
6
Without dither
12
0
Distribution of host signal at various stages of the embedding process
-12
-6
0
6
12
With dither
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
10
Results with ACD for One DCT Channel Detection: subtract dither and use correlation detection q = arg max
i =1,2,..., N
1 h(Z − S − d), W (i ) M
Collusion resistance increases to ~30 colluders (from ~7) Performance similar to uncompressed host signal case
Pd
With Dither
Δ= 6
1 0.9 0.8
Without Dither
0.7 0.6
Averaging without ACD Median without ACD Minimum without ACD Minimum with ACD Averaging with ACD Median with ACD
0.5 0.4 0.3 0.2 0.1 0
5
10
15
20
25
30
No. of colluders Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
11
Fingerprinting System for Compressed Hosts
Fingerprint Embedding Dither Compressed Host
+
De-quantize
+
Embed Fingerprint
Compress
PSNR X dB
Fingerprint Detection Dither
Suspicious copy (Compressed)
De-quantize
+
Preprocessing -
Correlation Detection
Host
HVS model used to determine embeddable coefficients
Energy in frequency band chosen based on HVS model
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
12
Results for Fingerprinting Compressed Images Comparison of visual quality (embedding PSNR = 42 dB)
Lena (QF=75)
Without ACD
With ACD
Comparison of file-size (bytes) Image
Original Size
Lena
Fingerprinted Image Size Without ACD
With ACD
10,513
10,502
10,531
Baboon
17,197
17,199
17,192
Barbara
12,586
12,588
12,560
Î Fingerprinted file-size almost the same as original Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
13
Results for Fingerprinting Compressed Images
Combine fingerprint information from all frequency bands
System Parameters
1024 users
Embedding PSNR = 42 dB
WNR = 0 dB
Pd improves significantly for all attacks
Better collusion resistance
Without ACD ~7 colluders
With ACD > 20 colluders
With Dither
Without Dither
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
14
Summary
Examined the problem of fingerprinting compressed multimedia
Extended Gaussian spread spectrum fingerprinting to compressed case
Poor collusion resistance at moderate levels of compression Averaging is the most effective attack
Proposed Anti-Collusion Dithering technique to improve collusion resistance
Collusion resistance approximately triples Perceptual quality and bandwidth not affected
Collusion-Resistant Fingerprinting for Compressed Multimedia: Varna, He, Swaminathan, Wu, Lu, Lu
15
Collusion-Resistant Fingerprinting for Compressed Multimedia Avinash L. Varna1, Shan He1, Ashwin Swaminathan1, Min Wu1, Haiming Lu2 and Zengxiang Lu2 http://www.ece.umd.edu/~varna 1
ECE department, Univ. of Maryland, College Park, USA
2 Research
Institute of Info. Tech., Tsinghua University, Beijing, China
International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2007
