550 likes | 719 Views
Detect Digital Image Forgeries. Ting-Wei Hsu. History of photo manipulation. 1860 the portrait of Lincoln is a composite of Lincoln ’ s head and John Calhoun ’ s body. History of photo manipulation. 1917: “ Cottingley fairies. History of photo manipulation.
E N D
Detect Digital Image Forgeries Ting-Wei Hsu
History of photo manipulation • 1860 the portrait of Lincoln is a composite of Lincoln’s head and John Calhoun’s body
History of photo manipulation • 1917: “Cottingley fairies
History of photo manipulation • 1930s: Stalin had disgraced comrades airbrushed out of his pictures
History of photo manipulation • 1936: same story with Mao
History of photo manipulation • 1936: same story with Mao
History of photo manipulation • Oprah Winfrey head on Ann-Margret
History of photo manipulation • 1994: O.J. Simpson’s mug shot modified to appear moremenacing
History of photo manipulation • April 2003: This digital composite of a British soldier in Basra, gesturing to Iraqi civilians urging them to seek cover,
History of photo manipulation • February 2004: Senator John Kerry and Jane Fonda sharing a stage at an anti-war rally emerged during the 2004 Presidential primaries as Senator Kerry was campaigning for the Democratic nomination.
History of photo manipulation • March 2004
History of photo manipulation • February 2008:
History of photo manipulation • August 2007
History of photo manipulation • November 2007
Cue in Forgeries Detection • Light Transport Difference • Acquisition Difference • Model Detect
Detect inconsistencies in Lighting • If the photo was composited, it’s often difficult to match the lighting conditions from individual photographs.
Color Model • Assumption: • the surface of interest is Lambertian • the surface has a constant reflectance value • the surface is illuminated by a point light source infinitely far away
Image Intensity Model • R : constant reflectance value • N(x,y) : 3 vector representing the surface normal at (x ,y) • A : constant ambient light • L : surface normal
Detect Duplicated Image Region • A common manipulation in tampering with an image is to copy and paste portions of the image to conceal a person or object in the scene.
Forgeries Using Duplicated Image • Applying PCA on small fixed size image block. • Reduce dimension representation • This representation is robust to minor variations in the image due to additive noise or lossy compression • Do lexicographic sorting
Results • Take 10 seconds in 512*512 image using 3 GHz processor
Detect by Tracking Re-sample • Processing in making forgeries often necessary to resize or rotate. • Assume resizing by linear or cubic interpolation method.
Resample • Resample by factor of 4/3
Resample • Use EM algorithm to estimate
Rotated and Resized • Upsampled by 15% and rotated by 5% • Rotated by 5% and upsampled by 15%
PATTERN NOISE & DETECTION OF ITS PRESENCE • Detection of digitally manipulated images based on the sensor pattern noise . • Detection whether image take from same camera or from another region.
PATTERN NOISE & DETECTION OF ITS PRESENCE • Most digital camera with CCD or CMOS use color filter array (CFA)
PRNU • Photo-response non-uniformity noise • Dominate part of the pattern noise in nature images. • PNU – pixel non-uniformity : different sensitivity of pixel to light • Caused by stochastic inhomogenities present in silicon wafer
Noise Model • xij : signal from light • ηij: random shot noise • cij: dark current • εij: read-out noise
Learn PNU • F : denoising filtering • Training by more than 50 picture
Detect • Random select n region with m masks • Estimate