Download presentation
Presentation is loading. Please wait.
Published byMuriel Cunningham Modified over 8 years ago
1
MMC LAB Secure Spread Spectrum Watermarking for Multimedia KAIST MMC LAB Seung jin Ryu 1MMC LAB
2
Contents 1. Introduction 2. Previous Work 3. Watermarking in the Frequency Domain 4. Structure of the Watermark 5. Experimental Results 6. Conclusion 2
3
Introduction Sudden increasing of digitized media Need for copyright enforcement schemes Cryptography provides little protection Digital watermark Intended to complement cryptographic processes Permanently embedded in the data MMC LAB 3
4
Introduction Characteristics Unobtrusive Perceptually invisible Robustness Must be difficult to remove –Common signal processing –Common geometric distortions –Subterfuge Attacks (Collusion and Forgery) Universality The algorithm should apply to the other data format Unambiguousness Unambiguous retrieval of the watermark MMC LAB 4
5
Introduction Building a strong watermark Watermark structure The watermark placed explicitly in the perceptually most significant components Perceptual capacity that allows watermark insertion without perceptual degradation Insertion strategy The watermark drawn from a Gaussian distribution N(0, 1) Offering good protection against collusion MMC LAB 5
6
Previous work Substituting the insignificant bits Inserting an identification string into a digital audio Turner Watermarks which resemble quantization noise Quantization noise is imperceptible to viewers Tanaka et al. DCT of 8 X 8 blocks A triple of frequencies is selected, modified Not based on any perceptual significance Variance between coefficients is small
7
Frequency-based scheme spreads the watermark over the whole spatial extent of the image Processing operations MMC LAB 7 Watermarking in the Frequency Domain Robustness Unobtrusive Unambiguousness
8
Watermarking in the Frequency Domain Lossy compression Eliminates high-frequency components Geometric distortions Rotation, translation, scaling, cropping, etc. Leads to a loss of data in the high-frequency spectral regions Signal distortions D/A-A/D conversion, resampling, requantization, etc. Signal transformations to be undone by using the original image MMC LAB 8
9
Spread spectrum Coding of a Watermark Spread spectrum communications Frequency domain – communication channel Watermark – signal Method Watermark is spread over very many frequency bins The energy in any one bin is very small and undetectable Be inserted imperceptibly in the most significant spectral components of the data –To avoid loss of watermark MMC LAB 9 Watermarking in the Frequency Domain
10
Embedding & Detecting the Watermark MMC LAB 10 Watermarking in the Frequency Domain
11
Structure of the Watermark Description of the Watermarking Procedure MMC LAB 11 DefinitionNotation DocumentD ValuesV = v 1, …, v n WatermarkX = x 1, …, x n (x i is chosen by N(0, 1) Wartermarked valuesV’ = v 1 ’, …, v n ’ Attacked documentD* Attacked valuesV* Corrupted watermarkX* generated by V* and V
12
Inserting and Extracting the Watermark Formulae for computing V’ Determining Multiple Scaling Parameters More or less tolerance to modification is allowed how sensitive the image is? In this paper (2) with a single parameter α = 0.1 MMC LAB 12 Structure of the Watermark v i ’ = v i + αx i (1) v i ’ = v i (1 + αx i ) (2) v i ’ = v i (e αxi ) (3)
13
Choosing the Length, n, of the Watermark The degree to which the watermark is spread out Altered components are increased, the extent to which they must be altered decreases. MMC LAB 13 Structure of the Watermark v i ’ = v i + αx i V i *= v i ’+ r i (r i is white noise with standard deviation σ)
14
Structure of the Watermark Evaluating the Similarity of Watermarks sim(X, X*) is distributed according to N(0, 1) X* X is distributed by N(0, X* X*) Robust Statistics Postprocessing for X* causes the improved performance –x* i = x* i – E i (X*) –x* i = 0 (if | x* i | > tolerance) –x* i = sign(x* i – E i (X*)) MMC LAB 14 sim(X, X * ) =
15
Structure of the Watermark Resilience to Multiple-Document Attacks Average attack Discrete watermarks Easier to completely eliminate Watermarks of 1 or -1, eliminated all but a 2 1-t Uniformly chosen watermark can be removed by only 5 documents MMC LAB 15
16
Structure of the Watermark Resilience to Multiple-Document Attacks Continuous valued watermarks greater resilience to average attacks
17
Structure of the Watermark Embedding process Detecting process MMC LAB 17 2D DCTsort v’=v (1+ w) IDCT & normalize Original image N largest coeff. other coeff. marked image random vector generator wmk seed DCT compute similarity threshold test image decision wmk DCTselect N largest original unmarked image select N largest preprocess –
18
Experimental Result Original and Watermarked images Original imageWatermarked image
19
Experimental Result The response of the watermark detector : 13.4 Image scaling
20
Experimental Result JPEG compression JPEG with 10% quality and 0% smoothing response of the watermark detector: 22.8 JPEG with 5% quality and 0% smoothing response of the watermark detector: 13.9
21
E xperimental Result The response of the watermark detector : 5.2 Dithering x* i = x* i – E i (X*) : 10.5
22
E xperimental Result Cropping The response of the watermark detector : 14.6
23
Experimental Result Cropping with JPEG image JPEG with 10% quality and 0% smoothing response of the watermark detector: 10.6
24
Experimental Result Print, xerox, and scan The response of the watermark detector : 4.0 x* i = x* i – E i (X*) x* i = sign(x* i – E i (X*)) :7.0
25
Experimental Result Rewatermarking
26
Experimental Result Collusion Attack
27
Experimental Result Environment Matlab 256*256 grey Lenna image Experiments Difference between another watermark Quantization error MMC LAB 27 + = sim(X, X*) Normal detecting29.852 After quantization29.8461
28
Conclusion k random numbers N(0,1) as watermark. Perceptually most significant components. maximizes the change of detecting the watermark after attacks. Experiment largest 1000 DCT coefficients Attacks Scaling, JPEG compression, Dithering, Cropping, Printing, xeroxing, and scanning, Rewatermarking, Collusion The correlation with the real watermark has a peak.
29
MMC LAB 29MMC LAB
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.