steganalysis in technicolor - binghamton...
TRANSCRIPT
Steganalysis in TechnicolorBoosting WS Detection of Stego Images from CFA-Interpolated Covers
Matthias Kirchner and Rainer BöhmeUniversity of Münster
IEEE ICASSP | Florence, Italy |May 8, 2014
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 2 /16
Are Steganographers Colorblind?
I research community seems to live in a monochromatic world wheregrayscale images abound
I a more colorful world raises questions about steganographic security
1 how plausible are grayscale images?2 how much can steganalysts gain from color information?
I this work: WS steganalysis and bilinear CFA interpolation
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 2 /16
Are Steganographers Colorblind?
I research community seems to live in a monochromatic world wheregrayscale images abound
I a more colorful world raises questions about steganographic security
1 how plausible are grayscale images?2 how much can steganalysts gain from color information?
I this work: WS steganalysis and bilinear CFA interpolation
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 2 /16
Are Steganographers Colorblind?
I research community seems to live in a monochromatic world wheregrayscale images abound
I a more colorful world raises questions about steganographic security
1 how plausible are grayscale images?
2 how much can steganalysts gain from color information?
I this work: WS steganalysis and bilinear CFA interpolation
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 2 /16
Are Steganographers Colorblind?
I research community seems to live in a monochromatic world wheregrayscale images abound
I a more colorful world raises questions about steganographic security
1 how plausible are grayscale images?2 how much can steganalysts gain from color information?
I this work: WS steganalysis and bilinear CFA interpolation
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 2 /16
Are Steganographers Colorblind?
I research community seems to live in a monochromatic world wheregrayscale images abound
I a more colorful world raises questions about steganographic security
1 how plausible are grayscale images?2 how much can steganalysts gain from color information?
I this work: WS steganalysis and bilinear CFA interpolation
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 3 /16
WS Steganalysis [Fridrich & Goljan, 2004; Ker & Böhme, 2008]
I estimates the embedding rate p̂ of uniform LSB replacement embedding ingrayscale images
p̂ =2n
n∑i=1
(−1) x(p)i
(x(p)i − x̂(0)
i
) x (0) cover object ∈ Zn
x (p) stego object ∈ Zn
x̂ (0) cover estimatep embedding rate
w vector of weights
I cover estimate: linear prediction from spatial neighboorhood
FKB8 :
− 14
12 −
14
12 0 1
2− 1
412 −
14
FLS8 :
b a ba 0 ab a b
I enhanced variants for various assumptions about cover source andembedding strategies
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 3 /16
WS Steganalysis [Fridrich & Goljan, 2004; Ker & Böhme, 2008]
I estimates the embedding rate p̂ of uniform LSB replacement embedding ingrayscale images
p̂ =2n
n∑i=1
(−1) x(p)i
(x(p)i − x̂(0)
i
) x (0) cover object ∈ Zn
x (p) stego object ∈ Zn
x̂ (0) cover estimatep embedding rate
w vector of weights
I cover estimate: linear prediction from spatial neighboorhood
FKB8 :
− 14
12 −
14
12 0 1
2− 1
412 −
14
FLS8 :
b a ba 0 ab a b
I enhanced variants for various assumptions about cover source andembedding strategies
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 3 /16
WS Steganalysis [Fridrich & Goljan, 2004; Ker & Böhme, 2008]
I estimates the embedding rate p̂ of uniform LSB replacement embedding ingrayscale images
p̂ =2n
n∑i=1
wi (−1) x(p)i
(x(p)i − x̂(0)
i
) x (0) cover object ∈ Zn
x (p) stego object ∈ Zn
x̂ (0) cover estimatep embedding ratew vector of weights
I cover estimate: linear prediction from spatial neighboorhood
FKB8 :
− 14
12 −
14
12 0 1
2− 1
412 −
14
FLS8 :
b a ba 0 ab a b
I enhanced variants for various assumptions about cover source andembedding strategies
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 4 /16
CFA InterpolationI typical digital images are captured with a color filter array
Bayer pattern
red channel green channel blue channel
I at least 2/3 of all pixels are interpolated
I intra-channel and inter-channel dependencies
Fgreen :
0 14 0
14 1 1
40 1
4 0
Fred :
14
12
14
12 1 1
214
12
14
bilinear interpolation:
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 4 /16
CFA InterpolationI typical digital images are captured with a color filter array
Bayer pattern
red channel green channel blue channel
I at least 2/3 of all pixels are interpolated
I intra-channel and inter-channel dependencies
Fgreen :
0 14 0
14 1 1
40 1
4 0
Fred :
14
12
14
12 1 1
214
12
14
bilinear interpolation:
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 4 /16
CFA InterpolationI typical digital images are captured with a color filter array
Bayer pattern
red channel green channel blue channel
I at least 2/3 of all pixels are interpolated
I intra-channel and inter-channel dependencies
Fgreen :
0 14 0
14 1 1
40 1
4 0
Fred :
14
12
14
12 1 1
214
12
14
bilinear interpolation:
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 5 /16
Spatial Neighborhood Categories
I CFA configuration defines characteristic pixel neighborhood categories
I x{C} = (xi | i ∈ C) C ∈ {R, 4N, 4D, 2H, 2V}
Bayer pattern green channel red channel
“raw” interpolated
4N
4D
2H
2V
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 5 /16
Spatial Neighborhood Categories
I CFA configuration defines characteristic pixel neighborhood categoriesI x{C} = (xi | i ∈ C) C ∈ {R, 4N, 4D, 2H, 2V}
Bayer pattern green channel red channel
“raw”
interpolated
4N
4D
2H
2V
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 6 /16
Spatial Neighborhood PredictabilityI OLS linear prediction from all 3× 3 neighborhoods (FLS8 ) per category
I prediction error
4N R global
0
2
4
6
8
10
12
RMS
I estimated coefficients
14
12
34
dire
ctne
ighb
ors
4N R global
− 12
− 14
0
diag
onal
neig
hbor
s
bilinear interpolation, green channel,7408 images (size: 512×512)
4N R 4N R 4N
R 4N R 4N R
4N R 4N R 4N
R 4N R 4N R
4N R 4N R 4N
“perfect” predictabilityof interpolated pixels
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 6 /16
Spatial Neighborhood PredictabilityI OLS linear prediction from all 3× 3 neighborhoods (FLS8 ) per category
I prediction error
4N R global
0
2
4
6
8
10
12
RMS
I estimated coefficients
14
12
34
dire
ctne
ighb
ors
4N R global
− 12
− 14
0
diag
onal
neig
hbor
s
bilinear interpolation, green channel,7408 images (size: 512×512)
4N R 4N R 4N
R 4N R 4N R
4N R 4N R 4N
R 4N R 4N R
4N R 4N R 4N
“perfect” predictabilityof interpolated pixels
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 6 /16
Spatial Neighborhood PredictabilityI OLS linear prediction from all 3× 3 neighborhoods (FLS8 ) per category
I prediction error
4N R global
0
2
4
6
8
10
12
RMS
I estimated coefficients
14
12
34
dire
ctne
ighb
ors
4N R global
− 12
− 14
0
diag
onal
neig
hbor
s
bilinear interpolation, green channel,7408 images (size: 512×512)
4N R 4N R 4N
R 4N R 4N R
4N R 4N R 4N
R 4N R 4N R
4N R 4N R 4N
“perfect” predictabilityof interpolated pixels
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 7 /16
CFA-Inspired Linear Pixel Predictors
I different neighborhood categories, C, call for tailored predictors
option 1: estimated from image per category, FLS8
(x (p){C}
)
option 2: fixed pre-set filter kernels, FC(x (p){C}
)
green channel red channel
R4N
4D
2H
2V
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 7 /16
CFA-Inspired Linear Pixel Predictors
I different neighborhood categories, C, call for tailored predictors
option 1: estimated from image per category, FLS8
(x (p){C}
)option 2: fixed pre-set filter kernels, FC
(x (p){C}
)
green channel red channel
R4N
4D
2H
2V
F4N
0 14 0
14 0 1
40 1
4 0
F4D
14 0 1
40 0 014 0 1
4
F2H
0 0 012 0 1
20 0 0
F2V
0 12 0
0 0 00 1
2 0
bilinear interpolation:
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 8 /16
CFA-WS Steganalysis
I utilization of CFA neighborhood relations in individual color channels
p̂C =2|{C}|
∑{i∈C}
(−1) x(p)i
(x(p)i −FC
(x (p))
i
) x (p) stego object ∈ Zn
FC predictor for categoryC ∈ {R, 4N, 4D, 2H, 2V}
p embedding rate
w vector of weights
green channel red channel
R4N
4D
2H
2V
I aggregation to a combinedestimate p̂ is equivalent toassigning weights
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 8 /16
CFA-WS Steganalysis
I utilization of CFA neighborhood relations in individual color channels
p̂C =2|{C}|
∑{i∈C}
wi (−1) x(p)i
(x(p)i −FC
(x (p))
i
) x (p) stego object ∈ Zn
FC predictor for categoryC ∈ {R, 4N, 4D, 2H, 2V}
p embedding ratew vector of weights
green channel red channel
R4N
4D
2H
2V
I aggregation to a combinedestimate p̂ is equivalent toassigning weights
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 9 /16
Experimental Setup
I 10, 000 BOSSBase grayscale images (size: 512× 512), sampled onto aBayer grid to apply plain bilinear CFA interpolation
I 3, 400 raw color image patches (size: 512× 512) from the Dresden ImageDatabase (Nikon D70), each processed with dcraw bilinear interpolationand proprietary content-adaptive Adobe Lightroom interpolation
I uniform LSB embedding
I exclude covers with >5 % flat blocks (size: 3× 3 )
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 9 /16
Experimental Setup
I 10, 000 BOSSBase grayscale images (size: 512× 512), sampled onto aBayer grid to apply plain bilinear CFA interpolation
I 3, 400 raw color image patches (size: 512× 512) from the Dresden ImageDatabase (Nikon D70), each processed with dcraw bilinear interpolationand proprietary content-adaptive Adobe Lightroom interpolation
I uniform LSB embedding
I exclude covers with >5 % flat blocks (size: 3× 3 )
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 9 /16
Experimental Setup
I 10, 000 BOSSBase grayscale images (size: 512× 512), sampled onto aBayer grid to apply plain bilinear CFA interpolation
I 3, 400 raw color image patches (size: 512× 512) from the Dresden ImageDatabase (Nikon D70), each processed with dcraw bilinear interpolationand proprietary content-adaptive Adobe Lightroom interpolation
I uniform LSB embedding
I exclude covers with >5 % flat blocks (size: 3× 3 )
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 10 /16
Steganalysis Results (I) – Estimation
0 0.1 0.2 0.3 0.4 0.50.001
0.005
0.01
0.05
0.1
4N R4N 4NLS8 LS8
global KB8
plain bilinear interpolation, green channel
embedding rate p
MAE
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 11 /16
Steganalysis Results (II) – Estimation
0 0.1 0.2 0.3 0.4 0.50.001
0.005
0.01
0.05
0.1
4D 2V R4D 2VLS8 LS8 LS8
global KB8
plain bilinear interpolation, red channel
embedding rate p
MAE
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 12 /16
Steganalysis Results (III) – Estimation
0 0.1 0.2 0.3 0.4 0.50.001
0.005
0.01
0.05
0.1
4N R4N 4NLS8 LS8
global KB8
dcraw bilinear interpolation, green channel
embedding rate p
MAE
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 13 /16
Steganalysis Results (IV) – Detection
I increasing mismatch between CFA modelling assumptions and reality letsFKB8 gain advantage
(p = 0.01) bilinear adaptive
plain dcraw LightroomN=7,408 N=3,316 N=3,166
F FP50 EER FP50 EER FP50 EER
Standard WS (KB8)
FKB8 0.35 0.41 0.23 0.33 0.08 0.19FLS8 0.38 0.43 0.26 0.34 0.09 0.20
Proposed CFA-WS (4N)
F4N 0.01 0.05 0.15 0.20 0.19 0.30FLS8 0.01 0.04 0.10 0.16 0.11 0.23 green channel
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 14 /16
Summary and Outlook
I color image steganography is hard
—and largely unexplored
I good news: substantial steganalysis performance boosts hinge onsufficient information about CFA interpolation function
I bad news: image forensics researchI good news: counter-forensics researchI join forces?
I come up with plausible communication channels for grayscale images?
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 14 /16
Summary and Outlook
I color image steganography is hard—and largely unexplored
I good news: substantial steganalysis performance boosts hinge onsufficient information about CFA interpolation function
I bad news: image forensics researchI good news: counter-forensics researchI join forces?
I come up with plausible communication channels for grayscale images?
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 14 /16
Summary and Outlook
I color image steganography is hard—and largely unexplored
I good news: substantial steganalysis performance boosts hinge onsufficient information about CFA interpolation function
I bad news: image forensics researchI good news: counter-forensics researchI join forces?
I come up with plausible communication channels for grayscale images?
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 14 /16
Summary and Outlook
I color image steganography is hard—and largely unexplored
I good news: substantial steganalysis performance boosts hinge onsufficient information about CFA interpolation function
I bad news: image forensics research
I good news: counter-forensics researchI join forces?
I come up with plausible communication channels for grayscale images?
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 14 /16
Summary and Outlook
I color image steganography is hard—and largely unexplored
I good news: substantial steganalysis performance boosts hinge onsufficient information about CFA interpolation function
I bad news: image forensics researchI good news: counter-forensics research
I join forces?
I come up with plausible communication channels for grayscale images?
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 14 /16
Summary and Outlook
I color image steganography is hard—and largely unexplored
I good news: substantial steganalysis performance boosts hinge onsufficient information about CFA interpolation function
I bad news: image forensics researchI good news: counter-forensics researchI join forces?
I come up with plausible communication channels for grayscale images?
Matthias Kirchner
livin
gkn
owle
dge
WW
UM
ünst
er
WESTFÄLISCHEWILHELMS-UNIVERSITÄTMÜNSTER Steganalysis in Technicolor 14 /16
Summary and Outlook
I color image steganography is hard—and largely unexplored
I good news: substantial steganalysis performance boosts hinge onsufficient information about CFA interpolation function
I bad news: image forensics researchI good news: counter-forensics researchI join forces?
I come up with plausible communication channels for grayscale images?
Matthias Kirchner