Side-Informed Steganography with Additive Distortiondde.binghamton.edu/tomasD/pdf/SI-slides.pdf · 2015-11-12 · Side-Informed Steganography with Additive Distortion ... EBS [Wang,

Side-Informed Steganography with Additive Distortion

Tomáš Denemark and Jessica Fridrich

Tomáš Denemark and Jessica Fridrich Side-Informed Steganography with Additive Distortion 1 / 21

Side-Informed Steganography

P

RAWhigh resolutionhigh bit depthuncompressed

. . .

precover

U

T

resizedgrayscale

DCT domain. . .

unquantizedcover

X

Q

e

quantized cover

quantization error

ρ

embeddingcosts

Y

stego



P


. . .

precover

U

T

resizedgrayscale

DCT domain. . .

unquantizedcover

X

Q

e

quantized cover

quantization error

ρ

embeddingcosts

Y

stego



P


. . .

precover

U

T

resizedgrayscale

DCT domain. . .

unquantizedcover

X

Q

e

quantized cover

quantization error

ρ

embeddingcosts

Y

stego



P


. . .

precover

U

T

resizedgrayscale

DCT domain. . .

unquantizedcover

X

Q

e

quantized cover

quantization error

ρ

embeddingcosts

Y

stego



P


. . .

precover

U

T

resizedgrayscale

DCT domain. . .

unquantizedcover

X

Q

e

quantized cover

quantization error

ρ

embeddingcosts

Y

stego



P


. . .

precover

U

T

resizedgrayscale

DCT domain. . .

unquantizedcover

X

Q

e

quantized cover

quantization error

ρ

embeddingcosts

Y

stego


Previous Art

GIFEmbedding-while-dithering [Fridrich, IHW 1999]

JPEGPerturbed Quantization [Fridrich, ACM MMSec 2004]

MMEx [Kim, IHW 2006]

BCHopt [Sachnev, ACM MMSec 2009]

EBS [Wang, ICASSP 2012]

NPQ [Huang, ACM IH&MMSec 2013]

SI-UNIWARD [Holub, ACM IH&MMSec 2013]

UED [Guo, TIFS 2014]

UERD [Guo, TIFS 2015]


Previous Art (cont’d)

Embedding limited to binary operationEither rounding as is or "to the other side"

Changing an element "to the other side"has positive cost

MMEx:ρij = 1− 2|eij |

BCHopt: (simplified)ρij = (q(1− 2|eij |)/2)2,q is the quantization step

SI-UNIWARD:ρij = (1− 2|eij |)ρ

(J−UNIWARD)ij [Uij ]− 1

[Uij ] = Xij

[Uij ] + 1

Uij

eij


Proposed Method – Cost Modulation

Applicable to any additive stego scheme Athat uses costs ρ(A)ij

Ternary embedding instead of binary

General formula for modulating the costs

ρij = (|Uij − Yij | − |Uij − Xij |) ρ(A)ij

=

(1− 2 |eij |) ρ

(A)ij Yij = Xij + sign (eij)

0 Yij = Xij

ρ(A)ij Yij = Xij − sign (eij)

[Uij ]− 1

[Uij ] = Xij

[Uij ] + 1

Uij

eij


Proposed Method – Cost Generation

P

Costs extracted from the unquantized cover ratherthan the quantized cover

U

T

X

Q

e

ρ Y


Experimental Setup

Precover source:BOSSBase v1.01 consisting of 10000 full resolution RAW image files

Unquantized covers:Images converted using ufraw to RGB TIFF. All further processing was donein Matlab rather than ImageMagick.

Feature set:Spatial Rich Model (dim 34671) [Fridrich, TIFS 2012]J+SRM [Kodovský, SPIE 2012]

Classifier:Ensemble of FLDs [Kodovský, TIFS 2012]

Performance measure:Average out-of-bag error EOOB (estimate of PE = 1

2(PFA + PMD))


SI and Processing Considered

Spatial domainHILL [Li, IEEE ICIP 2014]

S-UNIWARD [Holub, EURASIP 2014]

Resizing, Color Conversion, Quantization

JPEG domainJ-UNIWARD [Holub, EURASIP 2014]

JPEG Compression


Experiment 1 - Resizing

RAWfull-size

color

8 bitfull-sizegrayscale

convert,gray

P

double512× 512

grayscale

crop,resize

(Lanczos 3)

U

ufraw output: 24bit TIFF


Experiment 1 – Resizing (HILL)

0 0.1 0.2 0.3 0.4 0.50

0.2

0.4

Payload (bpp)

EO

OB

originalbinaryternaryρ from Xρ from U


Experiment 1 – Resizing (S-UNIWARD)

0 0.1 0.2 0.3 0.4 0.50

0.2

0.4

Payload (bpp)

EO

OB

originalbinaryternaryρ from Xρ from U


Resizing – Selection Channel

HILL 0.4 bpp SI-HILL

Last operation before quantizing: Resizing with Lanczos 3 in Matlab


Experiment 2 – Color Conversion

RAWfull-size

color

24 bit512× 512

color

convert,crop

P

double512× 512

grayscale

0.2989R+

0.5870G+

0.1140B

U



Experiment 2 – Color Conversion (HILL)

0 0.1 0.2 0.3 0.4 0.50

0.2

0.4

Payload (bpp)

EO

OB

originalbinaryternaryρ from U


Experiment 2 – Color Conversion (S-UNIWARD)

0 0.1 0.2 0.3 0.4 0.50

0.2

0.4

Payload (bpp)

EO

OB



Experiment 3 – Quantization

RAWfull-size

color

16 bit512× 512

grayscale

convert,gray,crop

P

double512× 512

grayscale

quantizeto 8 bit

U



Experiment 3 – Quantization (HILL)

0 0.1 0.2 0.3 0.4 0.50

0.2

0.4

Payload (bpp)

EO

OB



Experiment 3 – Quantization (S-UNIWARD)

0 0.1 0.2 0.3 0.4 0.50

0.2

0.4

Payload (bpp)

EO

OB



Experiment 4 – JPEG Compression

RAWfull-size

color

8 bit512× 512

grayscale

convert,gray,crop,resize

P

double512× 512

grayscale

DCT

U



Experiment 4 – JPEG Compression (J-UNIWARD, QF 75)

0 0.1 0.2 0.3 0.4 0.5 0.60

0.2

0.4

Payload (bpnzac)

EO

OB

originalbinaryternaryρ from Xρ from USI-UNI


Conclusion

We present general steganographic method of using side-information for

any domain

any cost-based steganography

any transformation with quantization

Improves on previous state-of-the-art by

allowing ternary embedding (effective when quantization is fine)

extracting the costs from the unquantized cover (effective whenquantization is coarse)

Source codes available at dde.binghamton.edu/download


dde.binghamton.edu/download

Side-Informed Steganography with Additive Distortiondde.binghamton.edu/tomasD/pdf/SI-slides.pdf · 2015-11-12 · Side-Informed Steganography with Additive Distortion ... EBS [Wang,

Documents