Slides
Slides
Slides
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
I know what you have done<br />
in your last<br />
HOTO HOP Session<br />
An Introduction to Digital Image Forensics<br />
Thomas Gloe<br />
Matthias Kirchner *<br />
Faculty of<br />
Computer Science<br />
TU Dresden<br />
ICSI Berkeley<br />
21|01|2010
This is actually not from a horror movie . . .<br />
◮ We have some clues that an alien species has begun to infiltrate (wo)mankind<br />
watch out for:<br />
⊲ two left feet<br />
Gloe & Kirchner Digital Image Forensics slide 1
This is actually not from a horror movie . . .<br />
◮ We have some clues that an alien species has begun to infiltrate (wo)mankind<br />
watch out for:<br />
⊲ two left feet<br />
⊲ missing belly button<br />
Gloe & Kirchner Digital Image Forensics slide 1
This is actually not from a horror movie . . .<br />
◮ We have some clues that an alien species has begun to infiltrate (wo)mankind<br />
watch out for:<br />
⊲ two left feet<br />
⊲ missing belly button<br />
⊲ overly long legs<br />
Gloe & Kirchner Digital Image Forensics slide 1
This is actually not from a horror movie . . .<br />
◮ We have some clues that an alien species has begun to infiltrate (wo)mankind<br />
watch out for:<br />
⊲ two left feet<br />
⊲ missing belly button<br />
⊲ overly long legs<br />
⊲ “ghost” shadows<br />
more examples can be found at<br />
http://photoshopdisasters.blogspot.com<br />
Gloe & Kirchner Digital Image Forensics slide 1
. . . yet it is as scary, at least.<br />
◮ more serious harm of image<br />
manipulations in areas like:<br />
⊲ political debates<br />
⊲ science<br />
⊲ evidence in court<br />
⊲ . . .<br />
former German Foreign Minister Steinmeier<br />
submission to the 2009 DOCMA award (Matthias Kleemann)
. . . yet it is as scary, at least.<br />
◮ more serious harm of image<br />
manipulations in areas like:<br />
⊲ political debates<br />
⊲ science<br />
⊲ evidence in court<br />
⊲ . . .<br />
Sciene 27 July 2007, Vol. 317, no. 5837, p. 450
. . . yet it is as scary, at least.<br />
◮ more serious harm of image<br />
manipulations in areas like:<br />
⊲ political debates<br />
⊲ science<br />
⊲ evidence in court<br />
⊲ . . .
. . . yet it is as scary, at least.<br />
◮ more serious harm of image<br />
manipulations in areas like:<br />
⊲ political debates<br />
⊲ science<br />
⊲ evidence in court<br />
⊲ . . .<br />
lots of interesting examples can<br />
be found at Hany Farid’s website<br />
www.cs.dartmouth.edu/farid/<br />
research/digitaltampering<br />
submission to the 2009 DOCMA award (Christine Gerhardt)
“Edit and share from anywhere”
Authenticity of images in the digital age?<br />
◮ Where is the picture coming from?<br />
Approaches<br />
◮ (How) Has the image been processed?<br />
◮ cryptography or digital watermarks<br />
⊲ marks needs to be added / embedded “actively” (high-end digital cameras)<br />
Gloe & Kirchner Digital Image Forensics slide 4
Authenticity of images in the digital age?<br />
◮ Where is the picture coming from?<br />
Approaches<br />
◮ (How) Has the image been processed?<br />
◮ cryptography or digital watermarks<br />
⊲ marks needs to be added / embedded “actively” (high-end digital cameras)<br />
◮ digital image forensics<br />
⊲ does not assume any knowledge about the original image<br />
⊲ is based on a statistical analysis of the image content<br />
N (µ, σ)<br />
Σ → min<br />
statistical analysis<br />
source?<br />
original?<br />
Gloe & Kirchner Digital Image Forensics slide 4
Digital image forensics: a first glimpse<br />
◮ digital image forensics<br />
assess the authenticity<br />
of digital images by exploiting<br />
their inherent statistical<br />
characteristics<br />
source identification<br />
manipulation detection<br />
Gloe & Kirchner Digital Image Forensics slide 5
Digital image forensics: a first glimpse<br />
◮ digital image forensics<br />
assess the authenticity<br />
of digital images by exploiting<br />
their inherent statistical<br />
characteristics<br />
class<br />
‘legitimate’<br />
model<br />
What means source?<br />
source identification<br />
manipulation detection<br />
device<br />
What means manipulation?<br />
‘malicious’<br />
Gloe & Kirchner Digital Image Forensics slide 5
Digital image forensics: a first glimpse<br />
images represent reality<br />
⊲ empirical science<br />
◮ digital image forensics<br />
assess the authenticity<br />
of digital images by exploiting<br />
their inherent statistical<br />
characteristics<br />
class<br />
‘legitimate’<br />
model<br />
What means source?<br />
source identification<br />
manipulation detection<br />
device<br />
What means manipulation?<br />
‘malicious’<br />
Gloe & Kirchner Digital Image Forensics slide 5
Digital image forensics: a first glimpse<br />
images represent reality<br />
⊲ empirical science<br />
◮ digital image forensics<br />
assess the authenticity<br />
of digital images by exploiting<br />
their inherent statistical<br />
characteristics<br />
requires a large body of<br />
reference data samples<br />
class<br />
‘legitimate’<br />
model<br />
What means source?<br />
source identification<br />
manipulation detection<br />
device<br />
What means manipulation?<br />
‘malicious’<br />
Gloe & Kirchner Digital Image Forensics slide 5
Digital image forensics: a first glimpse<br />
images represent reality<br />
1 ⊲ empirical science 2<br />
3<br />
◮ digital image forensics<br />
assess the authenticity<br />
of digital images by exploiting<br />
their inherent statistical<br />
characteristics<br />
requires a large body of<br />
reference data samples<br />
class<br />
‘legitimate’<br />
model<br />
What means source?<br />
source identification<br />
manipulation detection<br />
device<br />
What means manipulation?<br />
‘malicious’<br />
Gloe & Kirchner Digital Image Forensics slide 5
Interesting. Tell me more!<br />
Image forensics may exploit<br />
◮ artifacts of processing operations<br />
⊲ resampling · copy & paste · inconsistent lightning · double compression<br />
◮ characteristics of the source device<br />
scene<br />
⊲ e. g. digital camera:<br />
lens<br />
lens<br />
distortion<br />
filter<br />
R<br />
G<br />
G<br />
B<br />
CFA layout<br />
sensor<br />
hot pixels,<br />
sensor noise<br />
◮ metadata (has to be handled with care though)<br />
⊲ EXIF header · thumbnails<br />
black box model, camera response function<br />
color<br />
interpolation<br />
interpolation<br />
scheme<br />
post<br />
processing<br />
quantization<br />
table<br />
digital image<br />
Gloe & Kirchner Digital Image Forensics slide 6
Two introductory examples<br />
◮ digital camera identification<br />
based on sensor noise<br />
Gloe & Kirchner Digital Image Forensics slide 7
Two introductory examples<br />
◮ digital camera identification<br />
based on sensor noise<br />
?<br />
Gloe & Kirchner Digital Image Forensics slide 7
Two introductory examples<br />
◮ digital camera identification<br />
based on sensor noise<br />
?<br />
Gloe & Kirchner Digital Image Forensics slide 7
Two introductory examples<br />
◮ digital camera identification<br />
based on sensor noise<br />
≈<br />
Gloe & Kirchner Digital Image Forensics slide 7
Two introductory examples<br />
◮ digital camera identification<br />
based on sensor noise<br />
≈<br />
◮ copy & paste detection<br />
Gloe & Kirchner Digital Image Forensics slide 7
Two introductory examples<br />
◮ digital camera identification<br />
based on sensor noise<br />
≈<br />
◮ copy & paste detection<br />
presumed original<br />
Gloe & Kirchner Digital Image Forensics slide 7
1<br />
After all, it’s all about bits and<br />
bytes, like in computer forensics.<br />
Is digital image forensics<br />
really a science on its own?
1<br />
After all, it’s all about bits and<br />
bytes, like in computer forensics.<br />
Is digital image forensics<br />
really a science on its own?<br />
digital image forensics = computer forensics
By the way, what is computer forensics?
By the way, what is computer forensics?<br />
1<br />
1<br />
1100 0 0<br />
0<br />
1<br />
0<br />
1
By the way, what is computer forensics?<br />
52 51 51 51 49<br />
49 40 36 34 33<br />
55 48 40 33 23<br />
62 58 45 33 22<br />
66 62 53 34 22<br />
0<br />
0<br />
0<br />
1000 1 1<br />
0<br />
0<br />
0
By the way, what is computer forensics?<br />
52 51 51 51 49<br />
49 40 36 34 33<br />
55 48 40 33 23<br />
62 58 45 33 22<br />
66 62 53 34 22<br />
1<br />
1<br />
1<br />
0010 1 1<br />
0<br />
0<br />
0
Digital forensics: proposed ontology<br />
forensics<br />
1 1 1 1 0 1 0 0 0 0 0<br />
0<br />
1<br />
0<br />
0<br />
0<br />
0<br />
1 0 0 0 0<br />
digital<br />
0 0<br />
forensics<br />
1 0 0<br />
1<br />
1<br />
0<br />
0<br />
1<br />
0<br />
analog forensics<br />
1 1 0 1 0 0 0 0 0 1 1<br />
1<br />
1<br />
1<br />
1<br />
0 0 1<br />
computer 1 1 1<br />
forensics 1 1 0<br />
0 1 1<br />
1<br />
0<br />
0<br />
1<br />
1<br />
0<br />
0<br />
1<br />
1 multimedia 0 1 1 1<br />
0 1(image) 0 1 0<br />
0 forensics 0 1 1 1<br />
1 1 1 1 1<br />
1 1 1 digital 0 0 evidence 1 1 1 0 0 0 physical evidence<br />
1 1 0 1 0 1 1 1 1 1 1<br />
Gloe & Kirchner Digital Image Forensics slide 10
Digital forensics: proposed ontology<br />
forensics<br />
0 0 0 0 1 0 1 0 0 0 0<br />
0<br />
0<br />
0<br />
1<br />
1<br />
0<br />
0 1 1 1 0<br />
digital<br />
0 1<br />
forensics<br />
0 0 1<br />
0<br />
1<br />
0<br />
0<br />
1<br />
1<br />
analog forensics<br />
0 0 1 1 1 1 0 1 0 1 0<br />
0<br />
0<br />
1<br />
1<br />
1 1 0<br />
computer 0 1 1<br />
forensics 0 1 1<br />
0 0 1<br />
1<br />
0<br />
0<br />
1<br />
0<br />
1<br />
0<br />
1<br />
1 multimedia 1 0 0 0<br />
1 1(image) 0 1 1<br />
0 forensics 0 1 1 1<br />
0 1 0 1 1<br />
1 0 0 digital 0 1 evidence 0 0 1 0 0 1 physical evidence<br />
1 1 0 0 1 1 1 0 0 1 1<br />
finite sequence of discrete and<br />
perfectly observable symbols<br />
Gloe & Kirchner Digital Image Forensics slide 10
WARNING!<br />
The following slides<br />
intentionally draw a very<br />
black-and-white<br />
picture
Computer forensics = Image forensics<br />
computer forensics image forensics<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
Gloe & Kirchner Digital Image Forensics slide 12
Computer forensics = Image forensics<br />
computer forensics image forensics<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
WWW<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
Gloe & Kirchner Digital Image Forensics slide 12
Computer forensics = Image forensics<br />
computer forensics image forensics<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
◮ digital evidence is not linked<br />
to the outside world<br />
WWW<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
Gloe & Kirchner Digital Image Forensics slide 12
Computer forensics = Image forensics<br />
computer forensics image forensics<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
◮ digital evidence is not linked<br />
to the outside world<br />
WWW<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
Gloe & Kirchner Digital Image Forensics slide 12
Computer forensics = Image forensics<br />
computer forensics image forensics<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
◮ digital evidence is not linked<br />
to the outside world<br />
WWW<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
◮ digital evidence is linked<br />
to the outside world<br />
Gloe & Kirchner Digital Image Forensics slide 12
Computer forensics: a closer look<br />
processing<br />
digital<br />
data<br />
Gloe & Kirchner Digital Image Forensics slide 13
suspicious<br />
traces?<br />
Computer forensics: a closer look<br />
processing<br />
digital<br />
data<br />
Gloe & Kirchner Digital Image Forensics slide 13
suspicious<br />
traces?<br />
Computer forensics: a closer look<br />
processing<br />
digital<br />
data<br />
reality<br />
◮ digital evidence is stored in the<br />
finite automaton each computer<br />
represents<br />
◮ number of states in a closed<br />
system is finite<br />
Gloe & Kirchner Digital Image Forensics slide 13
suspicious<br />
traces?<br />
Computer forensics: a closer look<br />
processing<br />
digital<br />
data<br />
reality<br />
◮ digital evidence is stored in the<br />
finite automaton each computer<br />
represents<br />
◮ number of states in a closed<br />
system is finite<br />
Gloe & Kirchner Digital Image Forensics slide 13
suspicious<br />
traces?<br />
Computer forensics: a closer look<br />
processing<br />
digital<br />
data<br />
reality<br />
◮ digital evidence is stored in the<br />
finite automaton each computer<br />
represents<br />
◮ number of states in a closed<br />
system is finite<br />
◮ non-negligible chance that a<br />
computer is left in a state which<br />
perfectly erases all traces<br />
Gloe & Kirchner Digital Image Forensics slide 13
Image forensics: a closer look<br />
processing<br />
digital<br />
image<br />
Gloe & Kirchner Digital Image Forensics slide 14
Image forensics: a closer look<br />
original?<br />
source<br />
(device) ?<br />
processing<br />
digital<br />
image<br />
Gloe & Kirchner Digital Image Forensics slide 14
Image forensics: a closer look<br />
original?<br />
source<br />
(device) ?<br />
processing<br />
digital<br />
image<br />
sensor<br />
◮ sensors capture parts of the reality and<br />
transform them into digital representations<br />
Gloe & Kirchner Digital Image Forensics slide 14
Image forensics: a closer look<br />
original?<br />
source<br />
(device) ?<br />
processing<br />
digital<br />
image<br />
sensor<br />
◮ sensors capture parts of the reality and<br />
transform them into digital representations<br />
◮ reality is incognizable: ultimate knowledge<br />
whether a digital image reflects reality or<br />
not cannot exist<br />
Gloe & Kirchner Digital Image Forensics slide 14
Image forensics: a closer look<br />
original?<br />
source<br />
(device) ?<br />
processing<br />
digital<br />
image<br />
sensor<br />
◮ sensors capture parts of the reality and<br />
transform them into digital representations<br />
◮ reality is incognizable: ultimate knowledge<br />
whether a digital image reflects reality or<br />
not cannot exist<br />
◮ image forensics = empirical science<br />
Gloe & Kirchner Digital Image Forensics slide 14
Sensors: a source of uncertainty<br />
◮ projection of reality to discrete symbols means a dimensionality reduction<br />
Gloe & Kirchner Digital Image Forensics slide 15
Sensors: a source of uncertainty<br />
◮ projection of reality to discrete symbols means a dimensionality reduction<br />
◮ image forensics has to cope with an additional source of uncertainty<br />
degrees of freedom<br />
Gloe & Kirchner Digital Image Forensics slide 15
Sensors: a source of uncertainty<br />
◮ projection of reality to discrete symbols means a dimensionality reduction<br />
◮ image forensics has to cope with an additional source of uncertainty<br />
◮ what kind of common<br />
post-processing is<br />
legitimate / tolerable?<br />
Gloe & Kirchner Digital Image Forensics slide 15<br />
?
Models: yet another dimensionality reduction<br />
◮ models make projection of reality to<br />
discrete symbols tractable with formal<br />
methods<br />
◮ typical models in image forensics:<br />
⊲ sensor noise follows a Gaussian distribution<br />
⊲ connected regions of identical pixel values are<br />
unlikely to occur in original images<br />
Gloe & Kirchner Digital Image Forensics slide 16
Models: yet another dimensionality reduction<br />
◮ models make projection of reality to<br />
discrete symbols tractable with formal<br />
methods<br />
◮ typical models in image forensics:<br />
⊲ sensor noise follows a Gaussian distribution<br />
⊲ connected regions of identical pixel values are<br />
unlikely to occur in original images<br />
Gloe & Kirchner Digital Image Forensics slide 16<br />
p
Models: yet another dimensionality reduction<br />
◮ models make projection of reality to<br />
discrete symbols tractable with formal<br />
methods<br />
◮ typical models in image forensics:<br />
⊲ sensor noise follows a Gaussian distribution<br />
⊲ connected regions of identical pixel values are<br />
unlikely to occur in original images<br />
projection to a<br />
1-dimensional<br />
variable<br />
◮ models of reality are yet another dimensionality reduction<br />
◮ quality of forensic methods depends on the quality of the employed model!<br />
Gloe & Kirchner Digital Image Forensics slide 16<br />
p
Models: yet another dimensionality reduction<br />
◮ models make projection of reality to<br />
discrete symbols tractable with formal<br />
methods<br />
◮ typical models in image forensics:<br />
⊲ sensor noise follows a Gaussian distribution<br />
⊲ connected regions of identical pixel values are<br />
unlikely to occur in original images<br />
projection to a<br />
1-dimensional<br />
variable<br />
◮ models of reality are yet another dimensionality reduction<br />
◮ quality of forensic methods depends on the quality of the employed model!<br />
Gloe & Kirchner Digital Image Forensics slide 16<br />
p
2<br />
Now, after all the theory, what<br />
are you really doing all day long?<br />
A practical view on digital image forensics
Fundamentals of digital image forensics<br />
real scene<br />
analog 2D<br />
document<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
analog 2D<br />
document<br />
digitalization<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
analog 2D<br />
document<br />
digitalization<br />
digital image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
analog 2D<br />
document<br />
digitalization<br />
digital image<br />
manipulation<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
analog 2D<br />
document<br />
digitalization<br />
digital image<br />
manipulation<br />
output<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
analog 2D<br />
document<br />
digitalization<br />
digital image<br />
manipulation<br />
output<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
analog 2D<br />
document<br />
digitalization<br />
digital image<br />
manipulation<br />
output<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
digitalization<br />
digital image<br />
manipulation<br />
output<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
digitalization<br />
digital image<br />
manipulation<br />
output<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
digitalization<br />
device<br />
characteristics<br />
digital image<br />
manipulation<br />
output<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
digitalization<br />
device<br />
characteristics<br />
digital image<br />
file format<br />
manipulation<br />
output<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
digitalization<br />
device<br />
characteristics<br />
digital image<br />
file format<br />
manipulation<br />
processing<br />
artifacts<br />
output<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
digitalization<br />
device<br />
characteristics<br />
digital image<br />
file format<br />
manipulation<br />
processing<br />
artifacts<br />
output<br />
device<br />
characteristics<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
synthetic 3D scene<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
texture<br />
digitalization<br />
device<br />
characteristics<br />
digital image<br />
file format<br />
manipulation<br />
processing<br />
artifacts<br />
output<br />
device<br />
characteristics<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
synthetic 3D scene<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
texture<br />
digitalization<br />
device<br />
characteristics<br />
rendering<br />
digital image<br />
file format<br />
manipulation<br />
processing<br />
artifacts<br />
output<br />
device<br />
characteristics<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
synthetic 3D scene<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
texture<br />
digitalization<br />
device<br />
characteristics<br />
rendering<br />
missing device<br />
characteristics<br />
digital image<br />
file format<br />
manipulation<br />
processing<br />
artifacts<br />
output<br />
device<br />
characteristics<br />
re-digitalization<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Fundamentals of digital image forensics<br />
synthetic 3D scene<br />
real scene<br />
lightning, shadows<br />
analog 2D<br />
document<br />
surface characteristics<br />
texture<br />
digitalization<br />
device<br />
characteristics<br />
rendering<br />
missing device<br />
characteristics<br />
digital image<br />
file format<br />
manipulation<br />
processing<br />
artifacts<br />
output<br />
device<br />
characteristics<br />
re-digitalization<br />
⊲ too complex to be<br />
studied as a whole<br />
⊲ focus on small,<br />
tractable problems<br />
analog image<br />
Gloe & Kirchner Digital Image Forensics slide 18
Image source identification<br />
goal: determine either the correct class, the correct model / manufacturer,<br />
or even the actual device that took a particular image<br />
Gloe & Kirchner Digital Image Forensics slide 19<br />
?
Image source identification<br />
goal: determine either the correct class, the correct model / manufacturer,<br />
or even the actual device that took a particular image<br />
Gloe & Kirchner Digital Image Forensics slide 19<br />
?
Image source identification<br />
goal: determine either the correct class, the correct model / manufacturer,<br />
or even the actual device that took a particular image<br />
Gloe & Kirchner Digital Image Forensics slide 19<br />
?
Intra- and inter-model similarities<br />
◮ device identification – examined characteristic is unique for each device<br />
⊲ CCD / CMOS sensor noise [Lukáˇs et al., 2005]<br />
inter-model<br />
intra-model similarity intra-model similarity<br />
similarity<br />
low for device identification<br />
high for model identificatio<br />
low similarity<br />
low for device identification<br />
high for model identfication<br />
Gloe & Kirchner Digital Image Forensics slide 20
Intra- and inter-model similarities<br />
◮ device identification – examined characteristic is unique for each device<br />
⊲ CCD / CMOS sensor noise [Lukáˇs et al., 2005]<br />
◮ model identification – examined characteristic is similar for all devices of one<br />
particular model and differs between different models<br />
inter-model<br />
intra-model similarity intra-model similarity<br />
similarity<br />
low for device identification<br />
high for model identification<br />
low similarity<br />
low for device identification<br />
high for model identification<br />
Gloe & Kirchner Digital Image Forensics slide 20
Camera model identification<br />
◮ most probably, no single feature can distinguish between different camera models<br />
⊲ find a set of features with desirable properties<br />
Black box model [Kharrazi et al., 2004; Gloe et al., 2009]<br />
◮ overall 34 features to characterize images from the same camera model<br />
⊲ 12 color features to capture particularities of for instance white balancing<br />
⊲ 9 wavelet domain noise features<br />
⊲ 13 additional ‘image quality metrics’ for a general purpose description<br />
scene<br />
lens<br />
image quality<br />
metrics<br />
filter<br />
R<br />
G<br />
G<br />
B<br />
sensor<br />
image quality metrics,<br />
wavelet statistics<br />
color<br />
interpolation<br />
post<br />
processing<br />
color features color features<br />
digital image<br />
Gloe & Kirchner Digital Image Forensics slide 21
Camera model identification<br />
◮ most probably, no single feature can distinguish between different camera models<br />
⊲ find a set of features with desirable properties<br />
Black box model [Kharrazi et al., 2004; Gloe et al., 2009]<br />
◮ overall 34 features to characterize images from the same camera model<br />
⊲ 12 color features to capture particularities of for instance white balancing<br />
⊲ 9 wavelet domain noise features<br />
⊲ 13 additional ‘image quality metrics’ for a general purpose description<br />
scene<br />
lens<br />
image quality<br />
metrics<br />
filter<br />
R<br />
G<br />
G<br />
B<br />
sensor<br />
image quality metrics,<br />
wavelet statistics<br />
color<br />
interpolation<br />
◮ machine learning algorithms (SVM) for classification<br />
post<br />
processing<br />
color features color features<br />
digital image<br />
Gloe & Kirchner Digital Image Forensics slide 21
Some results on camera model identification<br />
◮ large-scale test with ≈ 15 000 images, stemming from 25 different models and<br />
overall 73 devices [Gloe et al., 2009]
Camera model identification: overall results<br />
◮ test set contains at least two devices per model<br />
◮ 60 % of the images for training the classifier, 40 % for testing<br />
◮ classification accuracy over all models: 96.42 %<br />
identified as<br />
camera model S DT DS DSs DC FZ H T W E F G I L M N O R µ<br />
Coolpix S710 S 100,00 - - - - - - - - - - - - - - - - - -<br />
D200 DT - 93,73 - - - 0,52 0,10 - 0,05 2,77 - - 2,61 - - - 0,05 - 0,16<br />
D70 DS - - 84,57 15,43 - - - - - - - - - - - - - - -<br />
D70s DSs - 0,11 67,51 30,90 - - - - - 0,85 - - 0,63 - - - - - -<br />
DCZ 5.9 DC - - 0,72 0,64 98,42 - - - - 0,12 - - - - - - - - 0,10<br />
DMC-FZ50 FZ - 1,14 - - - 97,31 0,50 0,58 0,32 - 0,06 - - - - 0,09 - - -<br />
DSC-H50 H - - - - - 0,20 98,23 0,20 1,28 - - 0,10 - - - - - - -<br />
DSC-T77 T - 0,74 - - - 3,43 2,04 91,90 1,81 - - 0,09 - - - - - - -<br />
DSC-W170 W - 0,65 - - - - 0,52 0,26 98,06 - - 0,52 - - - - - - -<br />
EX-Z150 E - - 0,19 0,08 - - - - - 99,71 - - - 0,03 - - - - -<br />
FinePixJ50 F - - - - - 0,06 - - - - 99,94 - - - - - - - -<br />
GX100 G - - - - - - - - - - - 100,00 - - - - - - -<br />
Ixus70 I - - 0,79 0,47 - - - - - - - - 98,26 0,42 - - - - 0,05<br />
L74wide L - - 0,04 0,04 - - - - - - - - 1,25 98,66 - - - - -<br />
M1063 M 0,07 0,67 - 0,01 - 0,01 - - - 0,04 - - 0,01 - 99,19 - - - -<br />
NV15 N 0,47 0,47 - - - 4,07 - - 0,05 - - - 0,05 - - 92,85 0,09 - 1,96<br />
OptioA40 O 3,97 0,13 0,09 - - 0,17 - 0,13 - 0,52 - 0,22 0,17 - 0,09 2,28 90,86 - 1,38<br />
RCP-7325XS R - - - - 0,21 - - - - - - - - - - - - 99,79 -<br />
µ1050SW µ - 0,29 - - - 0,12 - - - - - - 0,14 0,10 - 0,38 0,02 - 98,95<br />
Gloe & Kirchner Digital Image Forensics slide 23
Camera model identification: overall results<br />
◮ test set contains at least two devices per model<br />
◮ 60 % of the images for training the classifier, 40 % for testing<br />
◮ classification accuracy over all models: 96.42 %<br />
◮ mis-classifications between almost equal models of the same manufacturer<br />
identified as<br />
camera model S DT DS DSs DC FZ H T W E F G I L M N O R µ<br />
Coolpix S710 S 100,00 - - - - - - - - - - - - - - - - - -<br />
D200 DT - 93,73 - - - 0,52 0,10 - 0,05 2,77 - - 2,61 - - - 0,05 - 0,16<br />
D70 DS - - 84,57 15,43 - - - - - - - - - - - - - - -<br />
D70s DSs - 0,11 67,51 30,90 - - - - - 0,85 - - 0,63 - - - - - -<br />
DCZ 5.9 DC - - 0,72 0,64 98,42 - - - - 0,12 - - - - - - - - 0,10<br />
DMC-FZ50 FZ - 1,14 - - - 97,31 0,50 0,58 0,32 - 0,06 - - - - 0,09 - - -<br />
DSC-H50 H - - - - - 0,20 98,23 0,20 1,28 - - 0,10 - - - - - - -<br />
DSC-T77 T - 0,74 - - - 3,43 2,04 91,90 1,81 - - 0,09 - - - - - - -<br />
DSC-W170 W - 0,65 - - - - 0,52 0,26 98,06 - - 0,52 - - - - - - -<br />
EX-Z150 E - - 0,19 0,08 - - - - - 99,71 - - - 0,03 - - - - -<br />
FinePixJ50 F - - - - - 0,06 - - - - 99,94 - - - - - - - -<br />
GX100 G - - - - - - - - - - - 100,00 - - - - - - -<br />
Ixus70 I - - 0,79 0,47 - - - - - - - - 98,26 0,42 - - - - 0,05<br />
L74wide L - - 0,04 0,04 - - - - - - - - 1,25 98,66 - - - - -<br />
M1063 M 0,07 0,67 - 0,01 - 0,01 - - - 0,04 - - 0,01 - 99,19 - - - -<br />
NV15 N 0,47 0,47 - - - 4,07 - - 0,05 - - - 0,05 - - 92,85 0,09 - 1,96<br />
OptioA40 O 3,97 0,13 0,09 - - 0,17 - 0,13 - 0,52 - 0,22 0,17 - 0,09 2,28 90,86 - 1,38<br />
RCP-7325XS R - - - - 0,21 - - - - - - - - - - - - 99,79 -<br />
µ1050SW µ - 0,29 - - - 0,12 - - - - - - 0,14 0,10 - 0,38 0,02 - 98,95<br />
Gloe & Kirchner Digital Image Forensics slide 23
Camera model identification: overall results<br />
◮ test set contains at least two devices per model<br />
◮ 60 % of the images for training the classifier, 40 % for testing<br />
◮ classification accuracy over all models: 96.42 %<br />
◮ mis-classifications between almost equal models of the same manufacturer<br />
identified as<br />
camera model S DT DS DSs DC FZ H T W E F G I L M N O R µ<br />
Coolpix S710 S 100,00 - - - - - - - - - - - - - - - - - -<br />
D200 DT - 93,73 - - - 0,52 0,10 - 0,05 2,77 - - 2,61 - - - 0,05 - 0,16<br />
D70 DS - - 84,57 15,43 - - - - - - - - - - - - - - -<br />
D70s DSs - 0,11 67,51 30,90 - - - - - 0,85 - - 0,63 - - - - - -<br />
DCZ 5.9 DC - - 0,72 0,64 98,42 - - - - 0,12 - - - - - - - - 0,10<br />
DMC-FZ50 FZ - 1,14 - - - 97,31 0,50 0,58 0,32 - 0,06 - - - - 0,09 - - -<br />
DSC-H50 H - - - - - 0,20 98,23 0,20 1,28 - - 0,10 - - - - - - -<br />
DSC-T77 T - 0,74 - - - 3,43 2,04 91,90 1,81 - - 0,09 - - - - - - -<br />
DSC-W170 W - 0,65 - - - - 0,52 0,26 98,06 - - 0,52 - - - - - - -<br />
EX-Z150 E - - 0,19 0,08 - - - - - 99,71 - - - 0,03 - - - - -<br />
FinePixJ50 F - - - - - 0,06 - - - - 99,94 - - - - - - - -<br />
GX100 G - - - - - - - - - - - 100,00 - - - - - - -<br />
Ixus70 I - - 0,79 0,47 - - - - - - - - 98,26 0,42 - - - - 0,05<br />
L74wide L - - 0,04 0,04 - - - - - - - - 1,25 98,66 - - - - -<br />
M1063 M 0,07 0,67 - 0,01 - 0,01 - - - 0,04 - - 0,01 - 99,19 - - - -<br />
NV15 N 0,47 0,47 - - - 4,07 - - 0,05 - - - 0,05 - - 92,85 0,09 - 1,96<br />
OptioA40 O 3,97 0,13 0,09 - - 0,17 - 0,13 - 0,52 - 0,22 0,17 - 0,09 2,28 90,86 - 1,38<br />
RCP-7325XS R - - - - 0,21 - - - - - - - - - - - - 99,79 -<br />
µ1050SW µ - 0,29 - - - 0,12 - - - - - - 0,14 0,10 - 0,38 0,02 - 98,95<br />
⊲ What happens if we have to consider substantially more camera models or<br />
post-processing of the images?<br />
Gloe & Kirchner Digital Image Forensics slide 23
Camera device identification<br />
◮ most probably, one single feature can distinguish between specific camera devices<br />
Camera ID from sensor noise [Lukáˇs et al., 2005, 2006]<br />
◮ CCD / CMOS sensor elements convert light into a digital signal, a process subject to<br />
various noise sources<br />
⊲ temporal noise: varies throughout all images from a camera<br />
⊲ spatial noise: similar for all images of the very same camera,<br />
but different between distinct devices<br />
Gloe & Kirchner Digital Image Forensics slide 24
Camera device identification<br />
◮ most probably, one single feature can distinguish between specific camera devices<br />
Camera ID from sensor noise [Lukáˇs et al., 2005, 2006]<br />
◮ CCD / CMOS sensor elements convert light into a digital signal, a process subject to<br />
various noise sources<br />
⊲ temporal noise: varies throughout all images from a camera<br />
⊲ spatial noise: similar for all images of the very same camera,<br />
but different between distinct devices<br />
fixed pattern<br />
noise (FPN)<br />
⊲ additive noise, mainly due<br />
to dark noise<br />
⊲ flattened inside camera<br />
photo-response<br />
non-uniformity (PRNU)<br />
⊲ multiplicative noise, due to varying<br />
light-sensitivity of sensor elements<br />
⊲ hard to correct in cameras<br />
Gloe & Kirchner Digital Image Forensics slide 24
Camera device identification<br />
◮ most probably, one single feature can distinguish between specific camera devices<br />
Camera ID from sensor noise [Lukáˇs et al., 2005, 2006]<br />
◮ CCD / CMOS sensor elements convert light into a digital signal, a process subject to<br />
various noise sources<br />
⊲ temporal noise: varies throughout all images from a camera<br />
⊲ spatial noise: similar for all images of the very same camera,<br />
but different between distinct devices<br />
fixed pattern<br />
noise (FPN)<br />
⊲ additive noise, mainly due<br />
to dark noise<br />
⊲ flattened inside camera<br />
photo-response<br />
non-uniformity (PRNU)<br />
⊲ multiplicative noise, due to varying<br />
light-sensitivity of sensor elements<br />
⊲ hard to correct in cameras<br />
camera<br />
fingerprint<br />
Gloe & Kirchner Digital Image Forensics slide 24
How to obtain a digital camera fingerprint<br />
◮ camera-specific PRNU fingerprint is approximated by a reference noise pattern<br />
⊲ take at least 50 images from the same camera (ideally: perfectly lit, bright scenes)<br />
⊲ apply a de-noising filter to estimate the noise term in in each of these images<br />
⊲ take the pixel-wise average of the noise residuals<br />
digital camera stack of images noise residuals averaged noise residuals<br />
(reference noise pattern)<br />
◮ for practical applications, it is recommended to use a more sophisticated maximum<br />
likelihood estimation procedure [Chen et al., 2007, 2008]<br />
Gloe & Kirchner Digital Image Forensics slide 25
A typical noise residual Wavelet de-noising filter [Miçhak et al., 1999]<br />
◮ de-noising filters are not perfect ⊲ residual retains some image content<br />
Faculty of Computer Science, TU Dresden noise residual<br />
Gloe & Kirchner Digital Image Forensics slide 26
Reference noise pattern of a Canon S70<br />
noise pattern averaged over 300 images<br />
◮ reference noise pattern is free of<br />
image content<br />
◮ also contains information about the<br />
camera model (CFA artifacts, . . . )<br />
[Filler et al., 2008]<br />
enlarged detail with pixel defects<br />
Gloe & Kirchner Digital Image Forensics slide 27
correlation coefficient<br />
Was it your camera that took this image?<br />
◮ cameras are identified by measuring the similarity of the noise residual of a<br />
questionable image and all known reference noise pattern<br />
⊲ correlation coefficient or (more preferably) peak to correlation energy<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
reference pattern: S70<br />
Canon S70<br />
Canon S45<br />
Epson 1240U<br />
Ixus IIs<br />
1 image index<br />
350<br />
Closed camera set problem:<br />
◮ choose the camera with the highest<br />
similarity<br />
Open camera set problem:<br />
◮ no decision if maximum similarity is<br />
below a certain threshold<br />
Gloe & Kirchner Digital Image Forensics slide 28
correlation coefficient<br />
Was it your camera that took this image?<br />
◮ cameras are identified by measuring the similarity of the noise residual of a<br />
questionable image and all known reference noise pattern<br />
⊲ correlation coefficient or (more preferably) peak to correlation energy<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
reference pattern: S70<br />
Canon S70<br />
Canon S45<br />
Epson 1240U<br />
Ixus IIs<br />
1 image index<br />
350<br />
Closed camera set problem:<br />
◮ choose the camera with the highest<br />
similarity<br />
Open camera set problem:<br />
◮ no decision if maximum similarity is<br />
below a certain threshold<br />
◮ large-scale test with more than 10 6 images from ≈ 7000 individual cameras:<br />
97.6 % correct identification at a false acceptance rate of < 3 × 10 −5 [Goljan et al., 2009]<br />
Gloe & Kirchner Digital Image Forensics slide 28
Image source identfication: Summary<br />
◮ discussed methods are only the tip of the iceberg<br />
◮ further approaches include:<br />
⊲ sensor dust [Dirik et al., 2007]<br />
⊲ JPEG quantization tables [Farid, 2008]<br />
⊲ chromatic abberation [Van et al., 2007]<br />
⊲ demosaicing artifacts [Swaminathan et al, 2007]<br />
⊲ and many more . . .<br />
◮ current literature concludes that sensor noise is the method of choice whenever we<br />
have access to the questionable device (or a number of reference images)<br />
⊲ extensions to flatbed scanners [Gloe et al., 2007]<br />
⊲ robust against JPEG compression; may even survive printing and re-scanning<br />
⊲ but: computational issues with de-synchronization (cropping, rotation)<br />
◮ device model identification beneficial whenever we don’t have direct access to the<br />
questionable device<br />
⊲ but: does it scale with the number of examined models?<br />
Gloe & Kirchner Digital Image Forensics slide 29
Need a break?<br />
submission to the 2009 DOCMA award (Bernd Busche)
Need a break? Ooops. . .<br />
submission to the 2009 DOCMA award (Bernd Busche)
Image manipulation detection<br />
Digital image forensics checks for<br />
◮ missing or inconsistent device characteristics<br />
◮ presence of processing artifacts<br />
But what actually constitutes a manipulation?
Image manipulation detection<br />
Digital image forensics checks for<br />
◮ missing or inconsistent device characteristics<br />
◮ presence of processing artifacts<br />
But what actually constitutes a manipulation?<br />
‘malicious’ post-processing ‘legitimate’ post-processing
Image manipulation detection<br />
Digital image forensics checks for<br />
◮ missing or inconsistent device characteristics<br />
◮ presence of processing artifacts<br />
But what actually constitutes a manipulation?<br />
‘malicious’ post-processing ‘legitimate’ post-processing<br />
⊲ (mostly) local changes<br />
⊲ splicing<br />
⊲ copy & paste<br />
⊲ . . .<br />
⊲ content-preserving global changes<br />
⊲ denoising<br />
⊲ compression<br />
⊲ contrast enhancement<br />
definitions depend on established habits and conventions
Image manipulation detection<br />
Digital image forensics checks for<br />
◮ missing or inconsistent device characteristics<br />
◮ presence of processing artifacts<br />
But what actually constitutes a manipulation?<br />
‘malicious’ post-processing ‘legitimate’ post-processing
Image manipulation detection<br />
Digital image forensics checks for<br />
◮ missing or inconsistent device characteristics<br />
◮ presence of processing artifacts<br />
But what actually constitutes a manipulation?<br />
‘malicious’ post-processing ‘legitimate’ post-processing<br />
pictures: Andrea Sommer, Doc Baumann
Image manipulation detection<br />
Digital image forensics checks for<br />
◮ missing or inconsistent device characteristics<br />
◮ presence of processing artifacts<br />
But what actually constitutes a manipulation?<br />
research focuses on the<br />
detection of basic image<br />
processing primitives<br />
‘malicious’ post-processing ‘legitimate’ post-processing<br />
pictures: Andrea Sommer, Doc Baumann
Resampling detection<br />
◮ image manipulations often rely on geometric transformations<br />
(scaling, rotation, . . . ) of images or parts thereof<br />
Gloe & Kirchner Digital Image Forensics slide 32<br />
Andrea Sommer, Doc Baumann
Resampling detection<br />
◮ image manipulations often rely on geometric transformations<br />
(scaling, rotation, . . . ) of images or parts thereof<br />
◮ resampling to a new image grid; involves an interpolation step<br />
◮ interpolation introduces periodic linear correlations between neighboring pixels<br />
s1,1 s1,2<br />
s2,1 s2,2<br />
bilinear upsampling<br />
(×2)<br />
0.5<br />
0.5<br />
0.5 0.5<br />
s1,1 s1,2 s1,3<br />
s2,1 s2,2 s2,3<br />
s3,1 s3,2 s3,3<br />
0.5 0.5<br />
◮ resampling artifacts can be detected by the analysis of linear predictor residue<br />
[Popescu & Farid, 2005], [Kirchner, 2008]<br />
Gloe & Kirchner Digital Image Forensics slide 32<br />
0.5<br />
0.5
Resampling detection scheme<br />
t − 2 t − 1 t t + 1 t + 2<br />
Gloe & Kirchner Digital Image Forensics slide 33
Resampling detection scheme<br />
example: linear upsampling (×2)<br />
s(ωt ′ ) = P<br />
t h(ωt′ − t)s(t)<br />
t − 2 t − 1 t t + 1 t + 2<br />
Gloe & Kirchner Digital Image Forensics slide 33
Resampling detection scheme<br />
ωt ′ − ω ωt ′<br />
example: linear upsampling (×2)<br />
ωt ′ + ω<br />
t − 2 t − 1 t t + 1 t + 2<br />
◮ predictor residue: e(ωt ′ ) = s(ωt ′ ) −<br />
KX<br />
αks(ωt ′ + ωk) (α0 := 0)<br />
k=−K<br />
◮ large absolute prediction errors indicate minor degree of linear dependence<br />
Gloe & Kirchner Digital Image Forensics slide 33
Resampling detection scheme<br />
ωt ′ − ω ωt ′<br />
example: linear upsampling (×2)<br />
ωt ′ + ω<br />
t − 2 t − 1 t t + 1 t + 2<br />
◮ predictor residue: e(ωt ′ ) = s(ωt ′ ) −<br />
KX<br />
αks(ωt ′ + ωk) (α0 := 0)<br />
k=−K<br />
◮ large absolute prediction errors indicate minor degree of linear dependence<br />
◮ optimal weights in the example: α = (0.5, 0, 0.5)<br />
⊲ every second sample defaults to 0<br />
Gloe & Kirchner Digital Image Forensics slide 33
Resampling detection scheme<br />
ωt ′ − ω ωt ′<br />
example: linear upsampling (×2)<br />
ωt ′ + ω<br />
t − 2 t − 1 t t + 1 t + 2<br />
◮ predictor residue: e(ωt ′ ) = s(ωt ′ ) −<br />
KX<br />
αks(ωt ′ + ωk) (α0 := 0)<br />
k=−K<br />
◮ large absolute prediction errors indicate minor degree of linear dependence<br />
◮ optimal weights α can be determined with an EM algorithm [Popescu & Farid, 2005]<br />
or set to a fixed linear filter mask [Kirchner, 2008]<br />
Gloe & Kirchner Digital Image Forensics slide 33
Resampling detection scheme<br />
ωt ′ − ω ωt ′<br />
example: linear upsampling (×2)<br />
ωt ′ + ω<br />
t − 2 t − 1 t t + 1 t + 2<br />
◮ predictor residue: e(ωt ′ ) = s(ωt ′ ) −<br />
KX<br />
αks(ωt ′ + ωk) (α0 := 0)<br />
k=−K<br />
◮ large absolute prediction errors indicate minor degree of linear dependence<br />
◮ optimal weights α can be determined with an EM algorithm [Popescu & Farid, 2005]<br />
or set to a fixed linear filter mask [Kirchner, 2008]<br />
◮ p-map: p(ωt ′ ) ∝ exp (−σ|e(ωt ′ )| τ )<br />
◮ measure for the strength of linear dependence<br />
Gloe & Kirchner Digital Image Forensics slide 33
Original<br />
105 %<br />
120 %<br />
Typical detection results<br />
p-map DFT (p-map)<br />
* each spectrum graph has been individually normalized and processed with a maximum filter<br />
◮ resampling causes periodic<br />
pattern in the p-map and<br />
distinct peaks in the p-map’s<br />
DFT<br />
◮ peak position is characteristic for<br />
resampling parameters<br />
◮ typical resampling detectors<br />
employ a frequency domain<br />
peak detector<br />
Gloe & Kirchner Digital Image Forensics slide 34
It’s not only the biggest potato that counts<br />
www.bountyfishing.com<br />
◮ website awards a nice price to the largest fish caught by registered users<br />
◮ problem: ph fishing attacks with Photoshop<br />
Gloe & Kirchner Digital Image Forensics slide 35
It’s not only the biggest potato that counts<br />
www.bountyfishing.com<br />
◮ website awards a nice price to the largest fish caught by registered users<br />
◮ problem: ph fishing attacks with Photoshop<br />
◮ resampling detector unveils cheating<br />
Gloe & Kirchner Digital Image Forensics slide 35
Color filter array interpolation<br />
◮ typical digital cameras use only one CCD / CMOS sensor<br />
and a color filter array (CFA) to capture full-color images<br />
◮ missing color information is estimated from surrounding<br />
genuine elements in the raw image<br />
GR<br />
G<br />
GR<br />
G<br />
GR<br />
RAW image full-color image<br />
G<br />
GB<br />
G<br />
GB<br />
G<br />
GR<br />
G<br />
GR<br />
G<br />
GR<br />
G<br />
GB<br />
G<br />
GB<br />
G<br />
GR<br />
G<br />
GR<br />
G<br />
GR<br />
CFA interpolation / demosaicing<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
Gloe & Kirchner Digital Image Forensics slide 36<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R
Color filter array interpolation<br />
◮ typical digital cameras use only one CCD / CMOS sensor<br />
and a color filter array (CFA) to capture full-color images<br />
◮ missing color information is estimated from surrounding<br />
genuine elements in the raw image<br />
GR<br />
G<br />
GR<br />
G<br />
GR<br />
RAW image full-color image<br />
G<br />
GB<br />
G<br />
GB<br />
G<br />
GR<br />
G<br />
GR<br />
G<br />
GR<br />
G<br />
GB<br />
G<br />
GB<br />
G<br />
GR<br />
G<br />
GR<br />
G<br />
GR<br />
CFA interpolation / demosaicing<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
R<br />
demosaiced images<br />
exhibit specific inter-pixel<br />
correlation artifacts<br />
Gloe & Kirchner Digital Image Forensics slide 36
Example application of CFA artifacts [Popescu & Farid, 2005]<br />
Dresden Palace
Example application of CFA artifacts [Popescu & Farid, 2005]<br />
Dresden Palace
Example application of CFA artifacts [Popescu & Farid, 2005]<br />
◮ periodic artifacts in the linear predictor residue (p-map)<br />
Dresden Palace<br />
CFA peak<br />
DFT(p-map)
Double JPEG compression<br />
Main idea / background<br />
◮ digital images are often already JPEGs before being further processed<br />
◮ digital images are often re-saved as JPEGs after processing<br />
◮ usually, we will see different quantization tables<br />
Gloe & Kirchner Digital Image Forensics slide 38
Double JPEG compression<br />
Main idea / background<br />
◮ digital images are often already JPEGs before being further processed<br />
◮ digital images are often re-saved as JPEGs after processing<br />
◮ usually, we will see different quantization tables<br />
quantization of<br />
DCT coefficients<br />
QFpre<br />
pre-compression post-compression<br />
quantization<br />
factor q0<br />
» –<br />
c<br />
cq0 =<br />
q0<br />
=<br />
QFpost<br />
quantization<br />
factor q1<br />
" » – #<br />
c q0<br />
c (q0,q =<br />
1)<br />
q0 q1<br />
Gloe & Kirchner Digital Image Forensics slide 38
Double JPEG compression artifacts<br />
Quantization of DCT coefficients: a toy example<br />
q 0<br />
c ∈ R<br />
q 0cq 0 ∈ Z<br />
cq 0 =<br />
" #<br />
c<br />
Gloe & Kirchner Digital Image Forensics slide 39<br />
q 0
Double JPEG compression artifacts<br />
Quantization of DCT coefficients: a toy example<br />
q 0<br />
q 1<br />
c ∈ R<br />
q 0cq 0 ∈ Z<br />
q1c (q0 ,q1 ) ∈ Z<br />
cq 0 =<br />
" #<br />
c<br />
q 0<br />
c (q0 ,q 1 ) =<br />
" #<br />
cq0<br />
Gloe & Kirchner Digital Image Forensics slide 39<br />
q 1
Double JPEG compression artifacts<br />
Quantization of DCT coefficients: a toy example<br />
q 0<br />
q 1<br />
c ∈ R<br />
q 0cq 0 ∈ Z<br />
q1c (q0 ,q1 ) ∈ Z<br />
◮ double quantization leads to empty DCT coefficient histogram bins<br />
◮ periodicity depends on q0 / q1 ab [Popescu & Farid, 2004; Lukáˇs et al., 2003]<br />
no artifacts for q1 = n q0<br />
cq 0 =<br />
" #<br />
c<br />
q 0<br />
c (q0 ,q 1 ) =<br />
" #<br />
cq0<br />
Gloe & Kirchner Digital Image Forensics slide 39<br />
q 1
Church of our Ladies, Dresden<br />
A typical example . . .<br />
(1, 1) DCT coefficient histograms<br />
Q = 90<br />
-200 0<br />
200<br />
Q = (80, 90)<br />
-200 0<br />
200<br />
Q = (95, 90)<br />
-200 0<br />
200<br />
Gloe & Kirchner Digital Image Forensics slide 40
. . . and a practical application<br />
submission to the 2009 DOCMA award (Julia Enkelmann)
. . . and a practical application<br />
JPEG ghosts [Farid, 2009]<br />
submission to the 2009 DOCMA award (Julia Enkelmann)
Image manipulation detection: summary<br />
◮ discussed methods are only the tip of the iceberg<br />
◮ further approaches include:<br />
⊲ copy & paste<br />
[Lukáˇs et al., 2003; Popescu & Farid, 2004; . . . ]<br />
⊲ lightning analysis [Jonson & Farid, 2007]<br />
⊲ blocking artifacts [Li et al., 2009]<br />
◮ there is no single catch-all detector<br />
⊲ sensor noise [Lukáˇs et al., 2007]<br />
⊲ chromatic aberrations [Johnson & Farid, 2006]<br />
⊲ and many more . . .<br />
◮ image forensics provides a whole toolbox of different methods that aim at particular<br />
aspects of image manipulation<br />
but: semantic analysis of the results remains to be a task for the human investigator<br />
Gloe & Kirchner Digital Image Forensics slide 42
3Digital image forensics<br />
requires images.<br />
A lot of them.<br />
Test datasets in digital image forensics
How does your algorithm know that this<br />
image is authentic?<br />
Gloe & Kirchner Digital Image Forensics slide 44<br />
accompanying authentic submission to the 2009 DOCMA award (Xaver Klaußner)
How does your algorithm know that this<br />
image is authentic?<br />
Training, training, training, . . .<br />
◮ requires a vast amount of test data<br />
◮ scientifically sound results should rely on publicly available data<br />
Approaches to obtain test data<br />
◮ capture the images on your own<br />
⊲ time-consuming and redundant<br />
◮ use specific databases (mostly designed for other applications)<br />
⊲ images may have undesired properties<br />
◮ download<br />
⊲ images may have even more undesired properties<br />
⊲ no knowledge about the processing history available<br />
⊲ persistence and privacy issues<br />
Gloe & Kirchner Digital Image Forensics slide 44
Dresden Image Database [Gloe & Böhme, 2010]<br />
◮ image database, specifically designed for applications in digital image forensics<br />
◮ altogether more than 14 000 digital camera<br />
images, stemming from<br />
⊲ 73 devices<br />
⊲ 23 camera models<br />
⊲ 14 major camera manufacturers<br />
◮ several devices for most model<br />
◮ full range of consumer, semi-professional and<br />
professional digital cameras<br />
◮ main foucs: image source identification<br />
⊲ body of images allows to study characteristics specific to either manufacturer,<br />
model, or device<br />
⊲ images of each scene captured with each device<br />
◮ applications in manipulation detection [Kirchner & Gloe, 2009]<br />
Gloe & Kirchner Digital Image Forensics slide 45
How to capture 14 000 images<br />
. . .<br />
2<br />
2<br />
ICSI 2010<br />
2A<br />
3 image forensics<br />
3<br />
3A<br />
4<br />
4<br />
ICSI 2010<br />
4A<br />
set of cameras<br />
5 image forensics<br />
5<br />
5A<br />
6<br />
6<br />
ICSI 2010<br />
6A<br />
7 image forensics<br />
◮ each 2 tripods per indoor / outdoor location to capture 2 scenes<br />
◮ 3 different focal lengths per scene and device<br />
. . . and a lot of time<br />
7<br />
7A<br />
8<br />
8<br />
ICSI 2010<br />
8A<br />
9 image forensics<br />
9<br />
9A .<br />
. .
How to capture 14 000 images<br />
. . .<br />
2<br />
2<br />
ICSI 2010<br />
2A<br />
3 image forensics<br />
3<br />
3A<br />
4<br />
4<br />
ICSI 2010<br />
4A<br />
set of cameras<br />
5 image forensics<br />
5<br />
5A<br />
6<br />
6<br />
ICSI 2010<br />
6A<br />
7 image forensics<br />
◮ each 2 tripods per indoor / outdoor location to capture 2 scenes<br />
◮ 3 different focal lengths per scene and device<br />
. . . and a lot of time and hot beverages<br />
7<br />
7A<br />
8<br />
8<br />
ICSI 2010<br />
It looks cold?<br />
8A<br />
9 image forensics<br />
9<br />
9A .<br />
. .
min / max temperature [ ◦ C]<br />
8<br />
4<br />
0<br />
-4<br />
-8<br />
-12<br />
-16<br />
-20<br />
It was cold!<br />
1.1. 11.1. 21.1. 31.1. 10.2. 20.2.<br />
day<br />
◮ photo-shooting took place during the<br />
coldest period of the year 2009<br />
Gloe & Kirchner Digital Image Forensics slide 47
DOCMA Award<br />
Background<br />
◮ creating convincing forgeries is a time-consuming task<br />
◮ researchers are usually no Photoshop experts<br />
◮ typical test images do not reflect “real-life conditions”<br />
How can we learn to detect real forgeries from real<br />
Photoshop users if we don’t have them in our test<br />
database?
DOCMA Award<br />
Background<br />
◮ creating convincing forgeries is a time-consuming task<br />
◮ researchers are usually no Photoshop experts<br />
◮ typical test images do not reflect “real-life conditions”<br />
How can we learn to detect real forgeries from real<br />
Photoshop users if we don’t have them in our test<br />
database?<br />
DOCMA Award 2009<br />
◮ competition amongst (semi-)professional Photoshop<br />
users to create an arbitrary newspaper story,<br />
supported by a convincing image manipulation<br />
◮ initiator: Doc Baumann and his DOCMA magazine for<br />
professional Photoshop users<br />
◮ images (+ originals) are made available to researchers<br />
◮ around 100 interesting submissions<br />
Doc Baumann
DOCMA Award: some examples<br />
Fake<br />
Peter Hebler
DOCMA Award: some examples<br />
Original<br />
Peter Hebler
DOCMA Award: some examples<br />
Fake<br />
Bernd Busche
DOCMA Award: some examples<br />
Original<br />
Bernd Busche
DOCMA Award: some examples<br />
Fake<br />
Peter Wienerroither
DOCMA Award: some examples<br />
Original<br />
Peter Wienerroither
4Now that we have<br />
the theory behind,<br />
the practical algorithms,<br />
and the test data,<br />
everything should be fine?
4Now that we have<br />
the theory behind,<br />
the practical algorithms,<br />
and the test data,<br />
everything should be fine?<br />
Attacks against digital image forensics
Introducing attacks<br />
✔ existing forensic schemes seem to work well under laboratory conditions<br />
✔ usage in practical applications is documented<br />
✘ What if there was a smart counterfeiter, who knew about our techniques?<br />
◮ attacks: methods to systematically mislead digital image forensics<br />
Goals of an attacker<br />
◮ hide traces of image<br />
manipulations<br />
Benefits of studying attacks<br />
◮ eventually better detectors<br />
◮ erase evidence about the source of an image<br />
◮ pretend another source device of an image<br />
◮ protect the source of an image (e .g. retain anonymity)<br />
Gloe & Kirchner Digital Image Forensics slide 52
Digital forensics: proposed ontology<br />
forensics<br />
0 1 1 1 1 1 0 0 0 0 1<br />
1<br />
0<br />
1<br />
1<br />
0<br />
0<br />
1 1 0 0 0<br />
digital<br />
0 0<br />
forensics<br />
1 1 0<br />
0<br />
0<br />
1<br />
0<br />
1<br />
1<br />
analog forensics<br />
1 0 1 0 0 1 1 0 0 1 1<br />
0<br />
1<br />
1<br />
0<br />
0 0 1<br />
computer 1 0 1<br />
forensics 0 0 1<br />
0 1 1<br />
1<br />
1<br />
1<br />
0<br />
1<br />
0<br />
0<br />
0<br />
1 multimedia 0 0 1 1<br />
1 1(image) 1 0 1<br />
0 forensics 0 0 1 0<br />
1 0 1 1 1<br />
1 0 0 digital 1 1 evidence 1 1 0 1 1 1 physical evidence<br />
0 0 1 0 1 1 0 0 0 1 1<br />
forgeability<br />
b=<br />
counter-forensics<br />
Gloe & Kirchner Digital Image Forensics slide 53
Digital forensics: proposed ontology<br />
forensics<br />
0 1 0 1 0 1 0 1 0 0 1<br />
0<br />
1<br />
0<br />
1<br />
1<br />
1<br />
0 0 0 0 1<br />
digital<br />
0 1<br />
forensics<br />
0 1 1<br />
1<br />
1<br />
0<br />
1<br />
0<br />
0<br />
analog forensics<br />
0 1 1 1 0 0 1 1 0 0 0<br />
”physical evidence cannot be wrong,<br />
0<br />
0<br />
1<br />
0<br />
1 1 1<br />
computer 0 0 1<br />
forensics 0 0 1<br />
0 0 0<br />
1<br />
0<br />
1<br />
0<br />
0<br />
1<br />
1<br />
0<br />
0 multimedia 0 1 0 1<br />
0 1(image) 1 1 0<br />
1 forensics 0 1 1 0<br />
1 1 0 0 1<br />
it cannot perjure itself,<br />
it cannot be wholly absent”<br />
[Kirk, 1953]<br />
1 0 0 digital 1 1 evidence 0 0 0 1 1 1 physical evidence<br />
1 0 0 1 1 1 1 1 1 1 0<br />
forgeability<br />
b=<br />
counter-forensics<br />
Gloe & Kirchner Digital Image Forensics slide 53
Counter-forensics: computer forensics<br />
leave<br />
traces<br />
valid state invalid state<br />
Gloe & Kirchner Digital Image Forensics slide 54
Counter-forensics: computer forensics<br />
leave<br />
traces<br />
eliminate<br />
traces<br />
valid state invalid state valid state<br />
Gloe & Kirchner Digital Image Forensics slide 54
Counter-forensics: computer forensics<br />
leave<br />
traces<br />
eliminate<br />
traces<br />
valid state invalid state valid state<br />
valid states are perfectly known<br />
or can be recorded before<br />
Gloe & Kirchner Digital Image Forensics slide 54
Counter-forensics: computer forensics<br />
leave<br />
traces<br />
preemptively<br />
avoid traces<br />
eliminate<br />
traces<br />
valid state invalid state valid state<br />
valid states are perfectly known<br />
or can be recorded before<br />
Gloe & Kirchner Digital Image Forensics slide 54
Counter-forensics: computer forensics<br />
leave<br />
traces<br />
virtualization<br />
preemptively<br />
avoid traces<br />
eliminate<br />
traces<br />
valid state invalid state valid state<br />
valid states are perfectly known<br />
or can be recorded before<br />
Gloe & Kirchner Digital Image Forensics slide 54
Counter-forensics: image forensics<br />
leave<br />
traces<br />
preemptively<br />
avoid traces<br />
eliminate<br />
traces<br />
valid state invalid state valid state<br />
invalidity depends on<br />
the model of reality<br />
virtualization<br />
valid states are perfectly known<br />
or can be recorded before<br />
Gloe & Kirchner Digital Image Forensics slide 55
Counter-forensics: image forensics<br />
leave<br />
traces<br />
preemptively<br />
avoid traces<br />
eliminate<br />
traces<br />
valid state invalid state valid state<br />
invalidity depends on<br />
the model of reality<br />
virtualization is not possible<br />
valid states are not perfectly known<br />
or can be recorded before<br />
and cannot be recorded before<br />
Gloe & Kirchner Digital Image Forensics slide 55
Digital forensics: proposed ontology<br />
forensics<br />
1 0 0 0 1 0 1 1 0 1 1<br />
1<br />
0<br />
0<br />
0<br />
0<br />
0<br />
1 1 0 0 0<br />
digital<br />
1 0<br />
forensics<br />
1 1 0<br />
0<br />
0<br />
1<br />
1<br />
1<br />
0<br />
1 1 1 1 0 1 0 1 1 1 1<br />
1<br />
1<br />
1<br />
0<br />
0 1 1<br />
computer 1 0 0<br />
forensics 0 0 1<br />
0 0 0<br />
1<br />
1<br />
1<br />
0<br />
1<br />
1<br />
1<br />
1<br />
1 multimedia 1 0 0 1<br />
1 1(image) 0 0 0<br />
0 forensics 1 1 0 1<br />
0 0 1 1 0<br />
0perfect 1 1 crime 0 1<br />
1 possible 0 1 0 1<br />
1<br />
1<br />
1 1 1 1 1<br />
compete for<br />
the<br />
1 0<br />
best<br />
0<br />
model<br />
0 0<br />
forgeability<br />
b=<br />
counter-forensics<br />
analog forensics<br />
perfect crime<br />
impossible<br />
Gloe & Kirchner Digital Image Forensics slide 56
Attacks: practical considerations<br />
Who is that guy in the pixel next to you?<br />
◮ successful attacks against particular forensic algorithms are available<br />
[Gloe et al., 2007; Kirchner & Böhme, 2008, 2009]<br />
◮ defeating a whole set of forensic approaches is a much more challenging task<br />
⊲ attacks may introduce new detectable artifacts<br />
⊲ attacks may interfere with each other<br />
⊲ competition for the best model<br />
Gloe & Kirchner Digital Image Forensics slide 57
Attacks: practical considerations<br />
Who is that guy in the pixel next to you?<br />
◮ successful attacks against particular forensic algorithms are available<br />
[Gloe et al., 2007; Kirchner & Böhme, 2008, 2009]<br />
◮ defeating a whole set of forensic approaches is a much more challenging task<br />
⊲ attacks may introduce new detectable artifacts<br />
⊲ attacks may interfere with each other<br />
⊲ competition for the best model<br />
Attacks in terms of plausible<br />
post-processing<br />
◮ in practice, plausible image statistics<br />
might be enough<br />
⊲ downscaling and / or lossy compression<br />
are likely to smooth out and destroy most<br />
of the subtle traces we are looking for<br />
Gloe & Kirchner Digital Image Forensics slide 57
fin<br />
Concluding remarks
Summary<br />
◮ image forensics is a science to assess the authenticity of digital images<br />
image forensics is about statistical analysis of images<br />
⊲ main ingredients: device characteristics and manipulation artifacts<br />
⊲ there exists a variety of different approaches to determine the source of an image<br />
or to detect manipulations<br />
image forensics is an empirical science<br />
⊲ the better our model of reality, the more we can trust our tools<br />
image forensics requires rigor testing on large data sets<br />
⊲ since we cannot model reality entirely, we need as much observations as possible<br />
Gloe & Kirchner Digital Image Forensics slide 59
Multimedia forensics: a growing field<br />
◮ publications on multimedia forensics per year<br />
source: Hany Farid<br />
Digital Forensic Database<br />
8<br />
before<br />
4<br />
2003<br />
12<br />
2004<br />
21<br />
2005<br />
36<br />
2006<br />
Gloe & Kirchner Digital Image Forensics slide 60<br />
65<br />
2007<br />
71<br />
2008<br />
87<br />
2009
Multimedia forensics: a growing field<br />
◮ publications on multimedia forensics per year<br />
source: Hany Farid<br />
Digital Forensic Database<br />
first publication<br />
from Dresden<br />
8<br />
before<br />
4<br />
2003<br />
12<br />
2004<br />
21<br />
2005<br />
36<br />
2006<br />
Gloe & Kirchner Digital Image Forensics slide 60<br />
65<br />
2007<br />
71<br />
2008<br />
87<br />
2009
A broader view<br />
2<br />
2<br />
7<br />
7<br />
ICSI 2010<br />
2A<br />
ICSI 2010<br />
7A<br />
3<br />
3<br />
8<br />
8<br />
ICSI 2010<br />
3A<br />
ICSI 2010<br />
8A<br />
4<br />
4<br />
9<br />
9<br />
ICSI 2010<br />
4A<br />
ICSI 2010<br />
9A<br />
5<br />
5<br />
10<br />
10<br />
ICSI 2010<br />
5A<br />
ICSI 2010<br />
10A<br />
6<br />
6<br />
11<br />
11<br />
ICSI 2010<br />
6A<br />
ICSI 2010<br />
Gloe & Kirchner Digital Image Forensics slide 61<br />
11A
I know what you have done<br />
in your last<br />
HOTO HOP Session<br />
An Introduction to Digital Image Forensics<br />
Thomas Gloe<br />
Matthias Kirchner *<br />
Faculty of<br />
Computer Science<br />
TU Dresden<br />
ICSI Berkeley<br />
21|01|2010
Computer forensics in a broader sense<br />
◮ computers interact with their environment<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
WWW<br />
Gloe & Kirchner Digital Image Forensics slide 63
Computer forensics in a broader sense<br />
◮ computers interact with their environment<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
WWW<br />
WWW<br />
WWW<br />
◮ computers can be part of a network<br />
Gloe & Kirchner Digital Image Forensics slide 63<br />
WWW<br />
WWW<br />
WWW<br />
WWW
Computer forensics in a broader sense<br />
◮ computers interact with their environment<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
WWW<br />
WWW<br />
WWW<br />
◮ computers can be part of a network<br />
◮ computers can be sensors itself<br />
Gloe & Kirchner Digital Image Forensics slide 63<br />
WWW<br />
WWW<br />
WWW<br />
WWW
Computer forensics in a broader sense<br />
◮ computers interact with their environment<br />
physical evidence<br />
digital evidence<br />
1 0 0 1<br />
1 1 0 1<br />
WWW<br />
WWW<br />
WWW<br />
WWW<br />
◮ computers can be part of a network<br />
◮ computers can be sensors itself<br />
◮ computers leave physical evidence<br />
Gloe & Kirchner Digital Image Forensics slide 63<br />
WWW<br />
WWW<br />
WWW<br />
WWW
Image sources<br />
⊲ Photoshop logo (title,64) http://commons.wikimedia.org/wiki/File:Photoshop_CS4.svg<br />
⊲ Porta advertisement (1) DOCMA 32, January 2010, http://www.docma.info<br />
⊲ models (1) http://photoshopdisasters.blogspot.com<br />
⊲ Iranian missile test (7) http://www.spiegel.de<br />
⊲ hard drive (9,61) http://commons.wikimedia.org/wiki/File:Open_hard-drive.jpg<br />
⊲ floppy disk (13,54) http://commons.wikimedia.org/wiki/GNOME_Desktop_icons<br />
⊲ core memory (13) http://commons.wikimedia.org/wiki/File:KL_CoreMemory.jpg<br />
⊲ digital image (14,55) http://commons.wikimedia.org/wiki/File:Crystal_Project_lphoto.png<br />
⊲ newspaper (18) http://commons.wikimedia.org/wiki/Nuvola/apps<br />
⊲ camera1, scanner, monitor, printer, 3D (18,19,20,25) http://commons.wikimedia.org/wiki/Crystal_Clear<br />
⊲ video camera, color palette (18,19) http://commons.wikimedia.org/wiki/Category:Crystal_Project<br />
⊲ beamer (18) http://www.oxygen-icons.org<br />
⊲ Blender logo (18) http://commons.wikimedia.org/wiki/Blender_(software)<br />
⊲ camera2 (19,20) http://commons.wikimedia.org/wiki/Camera<br />
⊲ George Bush (57) http://www.snopes.com/photos/politics/bushbook.asp<br />
⊲ fingerprints (61) http://www.lanl.gov/news/albums/chemistry/fingerprint.jpg<br />
⊲ handcuffs (61) http://commons.wikimedia.org/wiki/File:Handcuffs01_2003-06-02.jpg