Steganalysis in Technicolor - Boosting WS Detection of Stego Images ...
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor Boosting WS Detection of Stego Images from CFA-Interpolated Covers
Matthias Kirchner and Rainer Böhme University of Münster
IEEE ICASSP | Florence, Italy | May 8, 2014
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale images abound
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale images abound I a more colorful world raises questions about steganographic security
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale images abound I a more colorful world raises questions about steganographic security 1 how plausible are grayscale images?
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale 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?
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale 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
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
WS Steganalysis
[Fridrich & Goljan, 2004; Ker & Böhme, 2008]
ˆ of uniform LSB replacement embedding in estimates the embedding rate p grayscale images (0) n ˆ= p
n 2X
n
i=1
(p)
(−1) xi
(p)
xi
(0)
− ˆxi
x x (p) xˆ (0) p
cover object ∈ Z stego object ∈ Zn cover estimate embedding rate
living knowledge WWU Münster
I
3 /16
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
WS Steganalysis
ˆ of uniform LSB replacement embedding in estimates the embedding rate p grayscale images (0) n ˆ= p
I
[Fridrich & Goljan, 2004; Ker & Böhme, 2008]
n 2X
n
(p)
(−1) xi
(p)
xi
(0)
− ˆxi
i=1
x x (p) xˆ (0) p
cover object ∈ Z stego object ∈ Zn cover estimate embedding rate
cover estimate: linear prediction from spatial neighboorhood − 41
FKB8 :
1 2 − 41
Matthias Kirchner
1 2
− 41
0 21 1 1 2 −4
FLS8 :
b a b
a 0 a
b a b
living knowledge WWU Münster
I
3 /16
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
WS Steganalysis
n 2X
n
(p)
wi (−1) xi
(p)
xi
(0)
− ˆxi
i=1
x x (p) xˆ (0) p w
cover object ∈ Z stego object ∈ Zn cover estimate embedding rate vector of weights
cover estimate: linear prediction from spatial neighboorhood − 41
FKB8 :
I
[Fridrich & Goljan, 2004; Ker & Böhme, 2008]
ˆ of uniform LSB replacement embedding in estimates the embedding rate p grayscale images (0) n ˆ= p
I
3 /16
1 2 − 41
1 2
− 41
0 21 1 1 2 −4
FLS8 :
b a b
a 0 a
b a b
enhanced variants for various assumptions about cover source and embedding strategies Matthias Kirchner
living knowledge WWU Münster
I
Steganalysis in Technicolor
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
4 /16
CFA Interpolation typical digital images are captured with a color filter array Bayer pattern I
red channel
Matthias Kirchner
at least 2/3 of all pixels are interpolated
green channel
blue channel
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
4 /16
CFA Interpolation typical digital images are captured with a color filter array Bayer pattern
red channel
Matthias Kirchner
I
at least 2/3 of all pixels are interpolated
I
intra-channel and inter-channel dependencies
green channel
blue channel
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
4 /16
CFA Interpolation typical digital images are captured with a color filter array Bayer pattern I
at least 2/3 of all pixels are interpolated
I
intra-channel and inter-channel dependencies 0
bilinear interpolation:
Fgreen :
Fred :
red channel
Matthias Kirchner
green channel
blue channel
1 4
1 4
1
0 1 4
0
1 4
0
1 4 1 2 1 4
1 2
1 4 1 2 1 4
1 1 2
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
5 /16
Spatial Neighborhood Categories CFA configuration defines characteristic pixel neighborhood categories
Bayer pattern
green channel
“raw”
Matthias Kirchner
red channel
interpolated
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
5 /16
Spatial Neighborhood Categories CFA configuration defines characteristic pixel neighborhood categories x {C} = (xi | i ∈ C) C ∈ {R, 4N, 4D, 2H, 2V} Bayer pattern
green channel
red channel
4D 2H 2V
4N
Matthias Kirchner
“raw”
living knowledge WWU Münster
I I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
6 /16
Spatial Neighborhood Predictability I
OLS linear prediction from all 3 × 3 neighborhoods ( FLS8 ) per category
I
prediction error 12
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
RMS
8
6 4
bilinear interpolation, green channel, 7408 images (size: 512×512)
2
0
4N
R
Matthias Kirchner
global
living knowledge WWU Münster
10
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
6 /16
Spatial Neighborhood Predictability I
OLS linear prediction from all 3 × 3 neighborhoods ( FLS8 ) per category
I
prediction error direct neighbors
12
10
3 4
4
2
0
4N
R
Matthias Kirchner
global
4N
R
4N
R
4N
R
4N
R
4N
R
1 2
4N
R
4N
R
4N
1 4
R
4N
R
4N
R
4N
R
4N
R
4N
6 diagonal neighbors
RMS
8
estimated coefficients
0 − 41
bilinear interpolation, green channel, 7408 images (size: 512×512)
− 21
4N
R
global
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
6 /16
Spatial Neighborhood Predictability I
OLS linear prediction from all 3 × 3 neighborhoods ( FLS8 ) per category
I
prediction error
I
10
“perfect” predictability of interpolated pixels
1 2 1 4
6 diagonal neighbors
RMS
8
3 4
4
2
0
4N
R
Matthias Kirchner
global
0 − 41
bilinear interpolation, green channel, 7408 images (size: 512×512)
− 21
4N
R
global
living knowledge WWU Münster
direct neighbors
12
estimated coefficients
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
7 /16
CFA-Inspired Linear Pixel Predictors different neighborhood categories, C, call for tailored predictors (p) option 1: estimated from image per category, FLS8 x {C}
green channel
red channel 4D 2H 2V
4N Matthias Kirchner
R
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
7 /16
CFA-Inspired Linear Pixel Predictors different neighborhood categories, C, call for tailored predictors (p) option 1: estimated from image per category, FLS8 x {C} (p) option 2: fixed pre-set filter kernels, FC x {C} F4N
bilinear interpolation: 0 green channel
1 4
red channel
0
1 4
0 1 4
F2H 0
0
1 4
1 2
0
0
0 0 0
0 1 2
0
4D 2H 2V
1 4
0 4N Matthias Kirchner
R
F2V
F4D
1 4
0 0 0
1 4
0 1 4
0 0 0
1 2
0 1 2
0 0 0
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
8 /16
CFA-WS Steganalysis utilization of CFA neighborhood relations in individual color channels (p) 2 X (p) ˆC = (−1) xi xi − FC x (p) i p |{C}|
x (p) FC
{i∈C}
p
green channel
red channel 4D 2H 2V
4N Matthias Kirchner
R
stego object ∈ Zn predictor for category C ∈ {R, 4N, 4D, 2H, 2V} embedding rate
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
8 /16
CFA-WS Steganalysis utilization of CFA neighborhood relations in individual color channels x (p) FC
(p) 2 X (p) ˆC = wi (−1) xi xi − FC x (p) i p |{C}|
{i∈C}
p w green channel
red channel 4D 2H 2V
4N Matthias Kirchner
R
stego object ∈ Zn predictor for category C ∈ {R, 4N, 4D, 2H, 2V} embedding rate vector of weights
I
aggregation to a combined ˆ is equivalent to estimate p assigning weights
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
9 /16
Experimental Setup
10, 000 BOSSBase grayscale images (size: 512 × 512), sampled onto a Bayer grid to apply plain bilinear CFA interpolation
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
9 /16
Experimental Setup
10, 000 BOSSBase grayscale images (size: 512 × 512), sampled onto a Bayer grid to apply plain bilinear CFA interpolation I 3, 400 raw color image patches (size: 512 × 512) from the Dresden Image Database (Nikon D70), each processed with dcraw bilinear interpolation and proprietary content-adaptive Adobe Lightroom interpolation
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
9 /16
Experimental Setup
10, 000 BOSSBase grayscale images (size: 512 × 512), sampled onto a Bayer grid to apply plain bilinear CFA interpolation I 3, 400 raw color image patches (size: 512 × 512) from the Dresden Image Database (Nikon D70), each processed with dcraw bilinear interpolation and proprietary content-adaptive Adobe Lightroom interpolation I
uniform LSB embedding
I
exclude covers with >5 % flat blocks (size: 3 × 3 )
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
10 /16
Steganalysis Results (I) – Estimation 0.1
plain bilinear interpolation, green channel
MAE
0.05
0.005 4N
R 4N LS8
4N LS8 global KB8
0.001
0
0.1
0.2
0.3
embedding rate p
Matthias Kirchner
0.4
0.5
living knowledge WWU Münster
0.01
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
11 /16
Steganalysis Results (II) – Estimation 0.1
plain bilinear interpolation, red channel
MAE
0.05
0.005 4D
2V 4D LS8
R 2V LS8
LS8
global KB8 0.001
0
0.1
0.2
0.3
embedding rate p
Matthias Kirchner
0.4
0.5
living knowledge WWU Münster
0.01
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
12 /16
Steganalysis Results (III) – Estimation 0.1
dcraw bilinear interpolation, green channel
MAE
0.05
0.005 4N
R 4N LS8
4N LS8 global KB8
0.001
0
0.1
0.2
0.3
embedding rate p
Matthias Kirchner
0.4
0.5
living knowledge WWU Münster
0.01
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
13 /16
Steganalysis Results (IV) – Detection increasing mismatch between CFA modelling assumptions and reality lets FKB8 gain advantage (p = 0.01)
F
bilinear
adaptive
plain N=7,408
dcraw N=3,316
Lightroom N=3,166
FP50 EER
FP50 EER
FP50
0.23 0.33 0.26 0.34
0.08 0.19 0.09 0.20
0.15 0.20 0.10 0.16
0.19 0.30 0.11 0.23
EER
Standard WS (KB8) FKB8 FLS8
0.35 0.41 0.38 0.43
Proposed CFA-WS (4N) F4N FLS8
Matthias Kirchner
0.01 0.05 0.01 0.04
green channel
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook color image steganography is hard
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook color image steganography is hard—and largely unexplored
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook color image steganography is hard—and largely unexplored
I
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook I
color image steganography is hard—and largely unexplored
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function I bad news: image forensics research
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook I
color image steganography is hard—and largely unexplored
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function I bad news: image forensics research I good news: counter-forensics research
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook I
color image steganography is hard—and largely unexplored
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function I bad news: image forensics research I good news: counter-forensics research I join forces?
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook I
color image steganography is hard—and largely unexplored
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function I bad news: image forensics research I good news: counter-forensics research I join forces? I
come up with plausible communication channels for grayscale images?
Matthias Kirchner
living knowledge WWU Münster
I
Plausible Grayscale Images?
Plausible Grayscale Images?
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor* Boosting WS Detection of Stego Images from CFA-Interpolated Covers
Thanks for your attention!
*Similarities with existing companies and trademarks are accidental.
Steganalysis in Technicolor Boosting WS Detection of Stego Images from CFA-Interpolated Covers
Matthias Kirchner and Rainer Böhme University of Münster
IEEE ICASSP | Florence, Italy | May 8, 2014
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale images abound
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale images abound I a more colorful world raises questions about steganographic security
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale images abound I a more colorful world raises questions about steganographic security 1 how plausible are grayscale images?
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale 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?
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
2 /16
Are Steganographers Colorblind?
research community seems to live in a monochromatic world where grayscale 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
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
WS Steganalysis
[Fridrich & Goljan, 2004; Ker & Böhme, 2008]
ˆ of uniform LSB replacement embedding in estimates the embedding rate p grayscale images (0) n ˆ= p
n 2X
n
i=1
(p)
(−1) xi
(p)
xi
(0)
− ˆxi
x x (p) xˆ (0) p
cover object ∈ Z stego object ∈ Zn cover estimate embedding rate
living knowledge WWU Münster
I
3 /16
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
WS Steganalysis
ˆ of uniform LSB replacement embedding in estimates the embedding rate p grayscale images (0) n ˆ= p
I
[Fridrich & Goljan, 2004; Ker & Böhme, 2008]
n 2X
n
(p)
(−1) xi
(p)
xi
(0)
− ˆxi
i=1
x x (p) xˆ (0) p
cover object ∈ Z stego object ∈ Zn cover estimate embedding rate
cover estimate: linear prediction from spatial neighboorhood − 41
FKB8 :
1 2 − 41
Matthias Kirchner
1 2
− 41
0 21 1 1 2 −4
FLS8 :
b a b
a 0 a
b a b
living knowledge WWU Münster
I
3 /16
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
WS Steganalysis
n 2X
n
(p)
wi (−1) xi
(p)
xi
(0)
− ˆxi
i=1
x x (p) xˆ (0) p w
cover object ∈ Z stego object ∈ Zn cover estimate embedding rate vector of weights
cover estimate: linear prediction from spatial neighboorhood − 41
FKB8 :
I
[Fridrich & Goljan, 2004; Ker & Böhme, 2008]
ˆ of uniform LSB replacement embedding in estimates the embedding rate p grayscale images (0) n ˆ= p
I
3 /16
1 2 − 41
1 2
− 41
0 21 1 1 2 −4
FLS8 :
b a b
a 0 a
b a b
enhanced variants for various assumptions about cover source and embedding strategies Matthias Kirchner
living knowledge WWU Münster
I
Steganalysis in Technicolor
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
4 /16
CFA Interpolation typical digital images are captured with a color filter array Bayer pattern I
red channel
Matthias Kirchner
at least 2/3 of all pixels are interpolated
green channel
blue channel
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
4 /16
CFA Interpolation typical digital images are captured with a color filter array Bayer pattern
red channel
Matthias Kirchner
I
at least 2/3 of all pixels are interpolated
I
intra-channel and inter-channel dependencies
green channel
blue channel
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
4 /16
CFA Interpolation typical digital images are captured with a color filter array Bayer pattern I
at least 2/3 of all pixels are interpolated
I
intra-channel and inter-channel dependencies 0
bilinear interpolation:
Fgreen :
Fred :
red channel
Matthias Kirchner
green channel
blue channel
1 4
1 4
1
0 1 4
0
1 4
0
1 4 1 2 1 4
1 2
1 4 1 2 1 4
1 1 2
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
5 /16
Spatial Neighborhood Categories CFA configuration defines characteristic pixel neighborhood categories
Bayer pattern
green channel
“raw”
Matthias Kirchner
red channel
interpolated
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
5 /16
Spatial Neighborhood Categories CFA configuration defines characteristic pixel neighborhood categories x {C} = (xi | i ∈ C) C ∈ {R, 4N, 4D, 2H, 2V} Bayer pattern
green channel
red channel
4D 2H 2V
4N
Matthias Kirchner
“raw”
living knowledge WWU Münster
I I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
6 /16
Spatial Neighborhood Predictability I
OLS linear prediction from all 3 × 3 neighborhoods ( FLS8 ) per category
I
prediction error 12
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
RMS
8
6 4
bilinear interpolation, green channel, 7408 images (size: 512×512)
2
0
4N
R
Matthias Kirchner
global
living knowledge WWU Münster
10
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
6 /16
Spatial Neighborhood Predictability I
OLS linear prediction from all 3 × 3 neighborhoods ( FLS8 ) per category
I
prediction error direct neighbors
12
10
3 4
4
2
0
4N
R
Matthias Kirchner
global
4N
R
4N
R
4N
R
4N
R
4N
R
1 2
4N
R
4N
R
4N
1 4
R
4N
R
4N
R
4N
R
4N
R
4N
6 diagonal neighbors
RMS
8
estimated coefficients
0 − 41
bilinear interpolation, green channel, 7408 images (size: 512×512)
− 21
4N
R
global
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
6 /16
Spatial Neighborhood Predictability I
OLS linear prediction from all 3 × 3 neighborhoods ( FLS8 ) per category
I
prediction error
I
10
“perfect” predictability of interpolated pixels
1 2 1 4
6 diagonal neighbors
RMS
8
3 4
4
2
0
4N
R
Matthias Kirchner
global
0 − 41
bilinear interpolation, green channel, 7408 images (size: 512×512)
− 21
4N
R
global
living knowledge WWU Münster
direct neighbors
12
estimated coefficients
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
7 /16
CFA-Inspired Linear Pixel Predictors different neighborhood categories, C, call for tailored predictors (p) option 1: estimated from image per category, FLS8 x {C}
green channel
red channel 4D 2H 2V
4N Matthias Kirchner
R
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
7 /16
CFA-Inspired Linear Pixel Predictors different neighborhood categories, C, call for tailored predictors (p) option 1: estimated from image per category, FLS8 x {C} (p) option 2: fixed pre-set filter kernels, FC x {C} F4N
bilinear interpolation: 0 green channel
1 4
red channel
0
1 4
0 1 4
F2H 0
0
1 4
1 2
0
0
0 0 0
0 1 2
0
4D 2H 2V
1 4
0 4N Matthias Kirchner
R
F2V
F4D
1 4
0 0 0
1 4
0 1 4
0 0 0
1 2
0 1 2
0 0 0
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
8 /16
CFA-WS Steganalysis utilization of CFA neighborhood relations in individual color channels (p) 2 X (p) ˆC = (−1) xi xi − FC x (p) i p |{C}|
x (p) FC
{i∈C}
p
green channel
red channel 4D 2H 2V
4N Matthias Kirchner
R
stego object ∈ Zn predictor for category C ∈ {R, 4N, 4D, 2H, 2V} embedding rate
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
8 /16
CFA-WS Steganalysis utilization of CFA neighborhood relations in individual color channels x (p) FC
(p) 2 X (p) ˆC = wi (−1) xi xi − FC x (p) i p |{C}|
{i∈C}
p w green channel
red channel 4D 2H 2V
4N Matthias Kirchner
R
stego object ∈ Zn predictor for category C ∈ {R, 4N, 4D, 2H, 2V} embedding rate vector of weights
I
aggregation to a combined ˆ is equivalent to estimate p assigning weights
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
9 /16
Experimental Setup
10, 000 BOSSBase grayscale images (size: 512 × 512), sampled onto a Bayer grid to apply plain bilinear CFA interpolation
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
9 /16
Experimental Setup
10, 000 BOSSBase grayscale images (size: 512 × 512), sampled onto a Bayer grid to apply plain bilinear CFA interpolation I 3, 400 raw color image patches (size: 512 × 512) from the Dresden Image Database (Nikon D70), each processed with dcraw bilinear interpolation and proprietary content-adaptive Adobe Lightroom interpolation
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
9 /16
Experimental Setup
10, 000 BOSSBase grayscale images (size: 512 × 512), sampled onto a Bayer grid to apply plain bilinear CFA interpolation I 3, 400 raw color image patches (size: 512 × 512) from the Dresden Image Database (Nikon D70), each processed with dcraw bilinear interpolation and proprietary content-adaptive Adobe Lightroom interpolation I
uniform LSB embedding
I
exclude covers with >5 % flat blocks (size: 3 × 3 )
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
10 /16
Steganalysis Results (I) – Estimation 0.1
plain bilinear interpolation, green channel
MAE
0.05
0.005 4N
R 4N LS8
4N LS8 global KB8
0.001
0
0.1
0.2
0.3
embedding rate p
Matthias Kirchner
0.4
0.5
living knowledge WWU Münster
0.01
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
11 /16
Steganalysis Results (II) – Estimation 0.1
plain bilinear interpolation, red channel
MAE
0.05
0.005 4D
2V 4D LS8
R 2V LS8
LS8
global KB8 0.001
0
0.1
0.2
0.3
embedding rate p
Matthias Kirchner
0.4
0.5
living knowledge WWU Münster
0.01
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
12 /16
Steganalysis Results (III) – Estimation 0.1
dcraw bilinear interpolation, green channel
MAE
0.05
0.005 4N
R 4N LS8
4N LS8 global KB8
0.001
0
0.1
0.2
0.3
embedding rate p
Matthias Kirchner
0.4
0.5
living knowledge WWU Münster
0.01
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
13 /16
Steganalysis Results (IV) – Detection increasing mismatch between CFA modelling assumptions and reality lets FKB8 gain advantage (p = 0.01)
F
bilinear
adaptive
plain N=7,408
dcraw N=3,316
Lightroom N=3,166
FP50 EER
FP50 EER
FP50
0.23 0.33 0.26 0.34
0.08 0.19 0.09 0.20
0.15 0.20 0.10 0.16
0.19 0.30 0.11 0.23
EER
Standard WS (KB8) FKB8 FLS8
0.35 0.41 0.38 0.43
Proposed CFA-WS (4N) F4N FLS8
Matthias Kirchner
0.01 0.05 0.01 0.04
green channel
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook color image steganography is hard
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook color image steganography is hard—and largely unexplored
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook color image steganography is hard—and largely unexplored
I
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function
living knowledge WWU Münster
I
Matthias Kirchner
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook I
color image steganography is hard—and largely unexplored
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function I bad news: image forensics research
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook I
color image steganography is hard—and largely unexplored
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function I bad news: image forensics research I good news: counter-forensics research
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook I
color image steganography is hard—and largely unexplored
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function I bad news: image forensics research I good news: counter-forensics research I join forces?
Matthias Kirchner
living knowledge WWU Münster
I
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor
14 /16
Summary and Outlook I
color image steganography is hard—and largely unexplored
good news: substantial steganalysis performance boosts hinge on sufficient information about CFA interpolation function I bad news: image forensics research I good news: counter-forensics research I join forces? I
come up with plausible communication channels for grayscale images?
Matthias Kirchner
living knowledge WWU Münster
I
Plausible Grayscale Images?
Plausible Grayscale Images?
W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER
Steganalysis in Technicolor* Boosting WS Detection of Stego Images from CFA-Interpolated Covers
Thanks for your attention!
*Similarities with existing companies and trademarks are accidental.
