SCIENTIFIC COMPUTING/HPC SOURCE presents an educational webcast
“Virtualized Research Ecosystems” allow researchers from across
the world to collaborate on important research initiatives. Our expert
from the Research Industry, Michael Elliot, offers up specific benefits
and application uses from the field of virtualized research.
This webinar packs a lot of information into 30 minutes.
You don’t want to miss it!
Examining the landscape of solutions for
Virtualized Research
Ecosystems
Now On Demand at:
www.scientificcomputing.com/ecosystems
Microsoft Paint leads to obvious forgeries,
Hollywood and the gaming industry have
spent much treasure and effort to create
photorealistic images of people, places and
situations that have never existed in reality.
Developing automated tools capable of
detecting manipulated images is the goal of
digital image forensics. Like many automated processes, the approach is initially
modeled after the way humans perform a
task and then optimizes and advances the
process by leveraging the unique advantages of algorithms in speed and capacity.
Our human eyes are easily tricked by
optical illusions, smoke and mirrors, but
our human sense of hearing is much more
acute at detecting fakes, errors and differ-
ences. When you find a Web site offering
free downloads of your favorite musical
artists’ biggest hits and instead get record-
ings of a cruise-ship cover band, your
brain can determine things are different
very quickly. Not overly concerned with
the volume and intensity of the music,
our brains are fantastic frequency analyz-
ers that make exact predictions of what
combination of frequencies to expect as
the performance unfolds. Digital image
forensics algorithms “listen” to an image
through the conversion of image intensities
into image frequencies through mathemat-
ical transforms. While “listening” to an
image, the algorithms create a profile of
what frequencies to expect. In most cases,
the algorithm is not seeking to compare
a new image to an existing image, but
analyzing each sub-region of an image to
determine if it fits with the frequency pro-
file of the entire file. When a random kitten
is pasted into an image and made to fit
“visually” using tools such as Photoshop,
the image forensics algorithms can quickly
discover the alteration, much like a human
audiophile can notice an attempt to alter a
record by using an X-Acto knife to splice
in a new section of grooved vinyl.
After altering an original image, it can be
copied into a new digital file in an attempt
to smooth out the rough edges surrounding
the alterations. When saved into a lossy
image file format that performs compres-
sion on image features, such as the format
created by the Joint Photographic Experts
Group (JPEG), some frequency information
is lost, while the picture looks identical to
the eye. Each time the image is saved, addi-
tional information is discarded by the com-
pression algorithm and begins to degrade
the visible quality of the image, much like
the photocopied office memes popular in the
1970s. A process known as error level anal-
ysis (ELA) examines a digital image to deter-
mine how degraded the image has become
due to multiple saves (copies) into the lossy
format. ELA is used to discover regions of
an image that have been compressed to a
different degree. This is similar to looking
for a new cloth patch that has been sewn
onto a pair of well-worn jeans.
Current research in the area of digital
image forensics is developing better ways
to convert image files into frequencies, such
as using wavelet transforms in addition
to more traditional cosine transforms and
more sensitive methods for determining
if each area of an image belongs to the
whole. Similar to specific “moments” in
time, “image moments” determine how
well intensity values belong to specific
regions of an image. Project “Ghiro,”
available at www.getghiro.org, is an open
source project that performs automated
digital forensic analysis on user images,
including the extraction of image metadata,
GPS localization information, image type
and MIME (Multipurpose Internet Mail
Extensions) information and Error Level
Analysis (ELA), outfitting users with very
powerful tools. About the only downside
to the wide availability of these advanced
analysis instruments is discovering how
ubiquitous Photoshop and Autotune have
become in their respective industries.
William Weaver is an associate professor
in the Department of Integrated Science,
Business, and Technology at La Salle
University. He may be contacted at
editor@ScientificComputing.com.
