Image Authentication by Detecting Traces of Demosaicing

1
Image Authentication by Detecting Traces of
Demosaicing

• June 23, 2008
• Andrew C. Gallagher1,2
• Tsuhan Chen1
• Carnegie Mellon University1 Eastman
  Kodak Company2

2
The Problem Authentication

• Good News Computer Graphics and Image
  Manipulation tools are rapidly advancing.

• Bad News How can we confirm that an image is
  authentically captured by a digital camera?

Image Credit Columbia photographic images and
photorealistic computer graphics dataset.
3
Computer Graphic vs. Photographic
Photo-Realistic Computer Graphics (PRCG)
Photographic Images (PIM)
Image Credit Columbia photographic images and
photorealistic computer graphics dataset.
4
Local Forgeries
Authentic image
Locally Modify Content or insert newContent
(Photographic or PRCG)
Locally Forged Image
5
Goals and Approach

• Our Goals
• Distinguish between Photographic (PIM) and
  Computer Graphic (PRCG)
• Find and Localize Forgeries

• Our Approach
• We focus on the image processing differences
  between digital cameras and computer graphics.
• We detect local traces of CFA interpolation.

6
Contributions

• PIM versus PRCG
• Hardware specific features vs. image physics or
  texture features (Ng et al. 2005, Lyu and Farid
  2005)
• Finding the demosaicing parameters is not
  necessary. (vs. learning with EM as in Popescu
  and Farid 2005).
• Excellent (best) performance on a standard test
  set using interpolation detection.
• We test with actual JPEG images from digital
  cameras.

7
Contributions

• Detecting Local Forgeries
• We show CFA detection is useful for accurately
  localizing suspicious regions.
• We show results on forgeries created from real
  digital camera images.
• The images are available for research.

8
Image Formation

• Digital Cameras

Sharpen Noise Cleaning
Hardware Correction
Balance Tone
Render
JPEG
A/D
Lens
Sensor
9
CFA Interpolation
CFA Interpolation

• Digital Cameras Use Color Filter Arrays
• Interpolation is required
• In general, missing pixels are a linear
  combination of neighbors
• Interpolation can be detected (Gallagher 2000,
  Popescu and Farid 2005).

10
Detecting Traces of CFA Interpolation
Canon EOS JPEG
EstimateVariance
Detect PeakStrength

• CFA Traces survive camera processing(even
  compression)
• Peak Strength

11
PRCG versus PIM
PIM. Distinct Peak at w p
PRCG. No Distinct Peak at w p
12
Results PRCG vs. PIM

• Columbia Image Set
• 800 PIM Digital Camera Images (JPEGs)
• 800 PRCG Photorealistic Computer Graphic
• Previous Approaches
• Texture statistics (wavelets) Lyu and Farid
  (2005)
• Geometric and Physical Features Ng et al. (2005)
• Our Feature Peak Strength

13
Results PRCG vs. PIM

• Performance as a function of region size

14
Results PRCG vs. PIM

• JPEG Quality Factor

Quality Factor 99
15
Results PRCG vs. PIM

• JPEG Quality Factor

Quality Factor 20
16
Results PRCG vs. PIM

• Classification Errors

PIM misclassified as PRCG
PRCG misclassified as PIM
17
Detecting Local Forgeries
Canon EOS JPEG
EstimateVariance
Detect PeakStrength

• Peak is computed locally (64x256)
• Forged regions usually wont have CFA traces.
• Suspicious regions have low .

18
Localizing Forgeries
SuspiciousRegions
Authentic
Forged
Analysis
Good results on all three images.
Images are Available at http//amp.ece.cmu.edu/pe
ople/Andy/authentication.html
19
Discussion

• CFA traces are destroyed by resizing
• CFA interpolation could be forged by a
  sophisticated forger.
• Many tests will likely be necessary to detect
  forgeries.

20
Conclusions

• We propose an elegant CFA interpolation detection
  for
• Distinguishing PIM from PRCG
• Localizing forged image regions
• Recovering the CFA parameters is not necessary.
• Our results are the best yet on a standard image
  set.