Transcript Detecting CFA Interpolation

Exposing Digital Forgeries in
Color Filter Array Interpolated
Images
By Alin C. Popescu and Hany Farid
Presenting - Anat Kaspi
The Goal



Low cost high resolution digital camera, sophisticated
photo editing  Digital media can be manipulated very
easily
Fake images…
Photos no longer hold the
unique stature as a definitive
recording of events
Automatically detecting digital forgeries in any portion of
an image

In contrast to other approaches: watermark, signature
Drawback: must be inserted at time of recording
The Technique
Digital forgeries may leave no visual clues but they may
alter the underlying statistics of an image

Color image consists of three channels containing
samples from different bands of the color spectrum

Most digital cameras are equipped with only a single
color sensor and use Color Filter Array (CFA)

The other two missing colors
must be estimated from the
neighboring to obtain three
channel color images – CFA Interpolation
The Technique (Cont.)


A subset of samples, within a color channel, are
correlated to neighboring samples
The correlations are periodic since the color filters
arranged in a periodic pattern
Presence or lack of correlation produced by
CFA interpolation can be used to detect forgery

There are many CFA Interpolation algorithms
 Bilinear and Bicubic, Median Filter, Gradient Based,
Adaptive Color Plane and more…
Example Bilinear interpolation
The Estimated samples are perfectly correlated to their neighbors
The Method - EM Algorithm





Two step iterative algorithm
We have two models: M1, M2
Outputs:
Probability Map – detect if a color image is a result of CFA
interpolation
Linear coefficients – used to distinguish between different
CFA interpolation
Results




CFA interpolation of their creation
Each color channel was independently blurred with 3x3
binomial filter
Down sample by factor of two in each direction
Re sampled onto Bayer array and CFA interpolated
Collected 100 images: 50 of resolution 512x512, 50 of
resolution 1024x1024
Gradient
3x3 median
No CFA
interpolation
Results



Detecting Localized Tampering
Composite images – splicing the non CFA image and the
same CFA interpolated image
Plausible forgery created using Adobe Photoshop
Sensitivity and Robustness




Tested the sensitivity of the model to
typical distortions that may conceal
trace of tampering
JPEG compression, additive white
Gaussian noise, Gamma correction
Robustness
Measure of similarity between
probability maps of each color channel
vs. synthetically generated probability
maps
Results: bilinear, bicubic, smooth hue,
variable number of gradient - 100%,
Median 99%, ACP 97%
Discussion
Advantages
 The technique works in the absence of any digital
watermark or signature
 Simple linear model to capture the correlation produced
by CFA interpolation
 Shown efficacy
Drawbacks
 Can be attacked by resampleing onto CFA and then
reinterpolating - requires knowledge of camera CFA
pattern