Forensic Data Hiding Optimized for JPEG 2000 Dieter Bardyn, Johann - - PowerPoint PPT Presentation

Forensic Data Hiding Optimized for JPEG 2000

Dieter Bardyn, Johann A. Briffa, Ann Dooms and Peter Schelkens May 18, 2011

Overview

◮ Image adaptive data hiding

Overview

◮ Image adaptive data hiding

Overview

◮ Image adaptive data hiding ⇒ tuned to image statistics

Overview

◮ Image adaptive data hiding ⇒ tuned to image statistics

◮ better fidelity

Overview

◮ Image adaptive data hiding ⇒ tuned to image statistics

◮ better fidelity ◮ better robustness

Overview

◮ Image adaptive data hiding ⇒ tuned to image statistics

◮ better fidelity ◮ better robustness

◮ New technique suited for JPEG2000 compressed media

Overview

◮ Image adaptive data hiding ⇒ tuned to image statistics

◮ better fidelity ◮ better robustness

◮ New technique suited for JPEG2000 compressed media ◮ IDS codes for synchronization

The Basics

Quantization Index Modulation

◮ Embed information M into a coverwork c by modifying its

content imperceptibly

The Basics

Quantization Index Modulation

◮ Embed information M into a coverwork c by modifying its

content imperceptibly

◮ Embed m = 0 or 1 in a sample x using Scalar QIM1

xw = Q

• x − m∆

2

• + m∆

2

1Chen and Wornell

The Basics

Quantization Index Modulation

◮ Embed information M into a coverwork c by modifying its

content imperceptibly

◮ Embed m = 0 or 1 in a sample x using Scalar QIM1

xw = Q

• x − m∆

2

• + m∆

2 ◮ Choose samples (coefficients) to embed log2(M) bits

1Chen and Wornell

The Basics

Perceptual Shaping

◮ Psychovisual studies on perceptually similar signals

2Distortion should remain imperceptible

The Basics

Perceptual Shaping

◮ Psychovisual studies on perceptually similar signals ◮ Complex models of human visual system

2Distortion should remain imperceptible

The Basics

Perceptual Shaping

◮ Psychovisual studies on perceptually similar signals ◮ Complex models of human visual system ◮ First applied to quantization in compression schemes

2Distortion should remain imperceptible

The Basics

Perceptual Shaping

◮ Psychovisual studies on perceptually similar signals ◮ Complex models of human visual system ◮ First applied to quantization in compression schemes ◮ How to apply to Scalar QIM?

2Distortion should remain imperceptible

The Basics

Perceptual Shaping

◮ Psychovisual studies on perceptually similar signals ◮ Complex models of human visual system ◮ First applied to quantization in compression schemes ◮ How to apply to Scalar QIM?

◮ Operate in transform domain 2Distortion should remain imperceptible

The Basics

Perceptual Shaping

◮ Psychovisual studies on perceptually similar signals ◮ Complex models of human visual system ◮ First applied to quantization in compression schemes ◮ How to apply to Scalar QIM?

◮ Operate in transform domain ◮ Determine maximum allowable2 distortion ǫ 2Distortion should remain imperceptible

The Basics

Perceptual Shaping

◮ Psychovisual studies on perceptually similar signals ◮ Complex models of human visual system ◮ First applied to quantization in compression schemes ◮ How to apply to Scalar QIM?

◮ Operate in transform domain ◮ Determine maximum allowable2 distortion ǫ ◮ Determine quantizer stepsize ∆ to be ǫ

2

2Distortion should remain imperceptible

Perceptual Shaping and Data Hiding

Blind data hiding

◮ Watermark extractor does not need original data (key-based)

Perceptual Shaping and Data Hiding

Blind data hiding

◮ Watermark extractor does not need original data (key-based) ◮ No performance loss3

3Data Hiding Codes, Moulin and Koeter

Perceptual Shaping and Data Hiding

Blind data hiding

◮ Watermark extractor does not need original data (key-based) ◮ No performance loss3 ◮ Perceptual Shaping ⇒ image dependent coefficient selection

3Data Hiding Codes, Moulin and Koeter

Perceptual Shaping and Data Hiding

Blind data hiding

◮ Watermark extractor does not need original data (key-based) ◮ No performance loss3 ◮ Perceptual Shaping ⇒ image dependent coefficient selection ◮ Use mask values to select coefficients

3Data Hiding Codes, Moulin and Koeter

Perceptual Shaping and Data Hiding

Blind data hiding

◮ Watermark extractor does not need original data (key-based) ◮ No performance loss3 ◮ Perceptual Shaping ⇒ image dependent coefficient selection ◮ Use mask values to select coefficients ◮ Compare to threshold determined by payload size

3Data Hiding Codes, Moulin and Koeter

Perceptual Shaping and Data Hiding

Perceptual Shaping: Lewis-Barni

◮ Lewis-Barni mask on DWT coefficients

qθ

l (i, j)

= Θ(l, θ)∆(l, i, j)Ξ(l, i, j)0.2

Perceptual Shaping and Data Hiding

Perceptual Shaping: Lewis-Barni

◮ Lewis-Barni mask on DWT coefficients

qθ

l (i, j)

= Θ(l, θ)∆(l, i, j)Ξ(l, i, j)0.2

◮ Θ depends on resolution level and orientation

Perceptual Shaping and Data Hiding

Perceptual Shaping: Lewis-Barni

◮ Lewis-Barni mask on DWT coefficients

qθ

l (i, j)

= Θ(l, θ)∆(l, i, j)Ξ(l, i, j)0.2

◮ Θ depends on resolution level and orientation ◮ ∆ measures local brightness

Perceptual Shaping and Data Hiding

Perceptual Shaping: Lewis-Barni

◮ Lewis-Barni mask on DWT coefficients

qθ

l (i, j)

= Θ(l, θ)∆(l, i, j)Ξ(l, i, j)0.2

◮ Θ depends on resolution level and orientation ◮ ∆ measures local brightness ◮ Ξ factors in texture activity

Perceptual Shaping and Data Hiding

Perceptual Shaping: Lewis-Barni

◮ Lewis-Barni mask on DWT coefficients

qθ

l (i, j)

= Θ(l, θ)∆(l, i, j)Ξ(l, i, j)0.2

◮ Θ depends on resolution level and orientation ◮ ∆ measures local brightness ◮ Ξ factors in texture activity

◮ + Accurate representation of HVS.

Perceptual Shaping and Data Hiding

Perceptual Shaping: Lewis-Barni

◮ Lewis-Barni mask on DWT coefficients

qθ

l (i, j)

= Θ(l, θ)∆(l, i, j)Ξ(l, i, j)0.2

◮ Θ depends on resolution level and orientation ◮ ∆ measures local brightness ◮ Ξ factors in texture activity

◮ + Accurate representation of HVS. ◮ + DWT Based

Perceptual Shaping and Data Hiding

Perceptual Shaping: Lewis-Barni

◮ Lewis-Barni mask on DWT coefficients

qθ

l (i, j)

= Θ(l, θ)∆(l, i, j)Ξ(l, i, j)0.2

◮ Θ depends on resolution level and orientation ◮ ∆ measures local brightness ◮ Ξ factors in texture activity

◮ + Accurate representation of HVS. ◮ + DWT Based ◮ - High Complexity.

Perceptual Shaping and Data Hiding

Perceptual Shaping: Solanki

◮ Solanki mask on DCT coefficients of 8 by 8 blocks

Eblock =

7

• i,j=0

||C(i, j)||2 − ||C(0, 0)||2

Perceptual Shaping and Data Hiding

Perceptual Shaping: Solanki

◮ Solanki mask on DCT coefficients of 8 by 8 blocks

Eblock =

7

• i,j=0

||C(i, j)||2 − ||C(0, 0)||2

◮ + Low Complexity

Perceptual Shaping and Data Hiding

Perceptual Shaping: Solanki

◮ Solanki mask on DCT coefficients of 8 by 8 blocks

Eblock =

7

• i,j=0

||C(i, j)||2 − ||C(0, 0)||2

◮ + Low Complexity ◮ - DCT based

Perceptual Shaping and Data Hiding

Perceptual Shaping: Solanki

◮ Solanki mask on DCT coefficients of 8 by 8 blocks

Eblock =

7

• i,j=0

||C(i, j)||2 − ||C(0, 0)||2

◮ + Low Complexity ◮ - DCT based ◮ - Block based

Perceptual Shaping and Data Hiding

Perceptual Shaping: Tree Based

◮ Tree based mask on DWT coefficients

Etree(l, θ, i, j) =

l−1

• k=1+a

2l−k−1

• x,y=0

||I θ

k (i + x, j + y)||2 ,

(1)

Perceptual Shaping and Data Hiding

Perceptual Shaping: Tree Based

◮ Tree based mask on DWT coefficients

Etree(l, θ, i, j) =

l−1

• k=1+a

2l−k−1

• x,y=0

||I θ

k (i + x, j + y)||2 ,

(1)

◮ + Low Complexity

Perceptual Shaping and Data Hiding

Perceptual Shaping: Tree Based

◮ Tree based mask on DWT coefficients

Etree(l, θ, i, j) =

l−1

• k=1+a

2l−k−1

• x,y=0

||I θ

k (i + x, j + y)||2 ,

(1)

◮ + Low Complexity ◮ + DWT based

Perceptual Shaping and Data Hiding

Perceptual Shaping: Tree Based

◮ Tree based mask on DWT coefficients

Etree(l, θ, i, j) =

l−1

• k=1+a

2l−k−1

• x,y=0

||I θ

k (i + x, j + y)||2 ,

(1)

◮ + Low Complexity ◮ + DWT based ◮ + Good visual performance

Perceptual Shaping and Data Hiding

Perceptual Shaping: the masks (a) Lewis-Barni (b) Solanki (c) Tree Based

Perceptual Shaping and Data Hiding

Insertion, Deletion and Substitution Codes (IDS)

◮ Synchronization issues modeled by IDS channel

Perceptual Shaping and Data Hiding

Insertion, Deletion and Substitution Codes (IDS)

◮ Synchronization issues modeled by IDS channel ◮ Conventional ECC expect a substitution-only channel

Perceptual Shaping and Data Hiding

Insertion, Deletion and Substitution Codes (IDS)

◮ Synchronization issues modeled by IDS channel ◮ Conventional ECC expect a substitution-only channel ◮ We use an improved Davey-MacKay construction:

Perceptual Shaping and Data Hiding

Insertion, Deletion and Substitution Codes (IDS)

◮ Synchronization issues modeled by IDS channel ◮ Conventional ECC expect a substitution-only channel ◮ We use an improved Davey-MacKay construction:

◮ outer non-binary error-correcting code

Perceptual Shaping and Data Hiding

Insertion, Deletion and Substitution Codes (IDS)

◮ Synchronization issues modeled by IDS channel ◮ Conventional ECC expect a substitution-only channel ◮ We use an improved Davey-MacKay construction:

◮ outer non-binary error-correcting code ◮ sparse code

Perceptual Shaping and Data Hiding

Insertion, Deletion and Substitution Codes (IDS)

◮ Synchronization issues modeled by IDS channel ◮ Conventional ECC expect a substitution-only channel ◮ We use an improved Davey-MacKay construction:

◮ outer non-binary error-correcting code ◮ sparse code ◮ pseudo-random binary marker sequence

Overview

• f the complete system

◮ Payload: 300 bits ◮ Modified coefficients: 3000 (rate 1/10 IDS code)

Results

IDS performance

◮ High error rate, especially for Tree-based ◮ Still within capabilities of ECC

Results

Decoding performance

◮ Robustness as good as Lewis-Barni, at reduced complexity ◮ Poor robustness of Solanki – domain mismatch

Results

Other attacks

◮ We have seen effect of JPEG 2000 compression

Results

Other attacks

◮ We have seen effect of JPEG 2000 compression ◮ Effect of other attacks is similar:

Results

Other attacks

◮ We have seen effect of JPEG 2000 compression ◮ Effect of other attacks is similar:

◮ JPEG compression

Results

Other attacks

◮ We have seen effect of JPEG 2000 compression ◮ Effect of other attacks is similar:

◮ JPEG compression ◮ AWGN noise addition

Results

Visual Performance

PSNR SSIM Lewis-Barni 59 dB 1.0000 Solanki 55 dB 0.9996 Tree Based 59 dB 1.0000

(d) Lewis Barni (e) Solanki (f) Tree Based

Conclusion

◮ Data hiding and IDS codes to solve synchronization issues

Conclusion

◮ Data hiding and IDS codes to solve synchronization issues ◮ Novel perceptual mask

Conclusion

◮ Data hiding and IDS codes to solve synchronization issues ◮ Novel perceptual mask

◮ Low complexity

Conclusion

◮ Data hiding and IDS codes to solve synchronization issues ◮ Novel perceptual mask

◮ Low complexity ◮ Good visual performance

Conclusion

◮ Data hiding and IDS codes to solve synchronization issues ◮ Novel perceptual mask

◮ Low complexity ◮ Good visual performance

◮ Readily applicable for forensic applications

Appendix: Bibliography (1/2)

• 1. Chen, B. and Wornell, G., “Quantization Index Modulation: A

class of provably good methods for digital watermarking and information embedding”, IEEE Trans. Inf. Theory 47(4), 1423–1443 (2001).

• 2. Moulin, P. and Koetter, R., “Data-Hiding Codes”, Proc. of

IEEE 93, 2083–2126 (2005).

• 3. Solanki et al., “Robust image-adaptive data hiding using

erasure and error correction”, IEEE Trans. Im. Proc. 13(12), 1627–1639 (2004).

• 4. Barni et al., “Improved Wavelet-based watermarking through

pixel-wise masking”, IEEE Trans. Im. Proc. 10(5), 783–791 (2001).

Appendix: Bibliography (2/2)

• 5. Lewis, A. and Knowles, G., “Image compression using the 2-d

wavelet transform”, IEEE Trans. Im. Proc. 1(2), 224–250 (1992).

• 6. Davey, M.C. and MacKay, D.J.C., “Reliable communication
• ver channels with insertions, deletions and substitutions”,

IEEE Trans. Inf. Th. 47(2), 687–698 (2001).