Robust Undetectable Interference Watermarks

1
Robust Undetectable Interference Watermarks

• Ryszard Grzaslewicz (WUT)
• Miroslaw Kutylowski (WUT)
• Jaroslaw Kutylowski (HNI)
• Wojciech Pietkiewicz (WUT)

2
Introduction

• Motivation
• undetectable watermarks cannot be detected in an
  image
• only the owner can prove his rights to an image
  using a secret private key
• not suited for web crawlers
• Key features of our scheme
• watermarks encoded in spatial domain
• resistant against attack preserving distance
  between points (filtering, rotation, JPEG
  compression)
• resistance against some attacks changing distance
  between points
• watermark can be reconstructed from a small part
  of image
• original image not needed for reconstruction
• large computational effort for reconstruction
  needed

3
Physical motivation

• Interference Youngs experiment

4
Scheme overview
5
Construction of simple interference image

• Watermark image
• all black with some white points
• white points are light source
• set of white points

6
Construction of simple interference image

• Interference image
• interference image is placed at distance on
  top of watermark image
• each light source from influences each
  point of interference image
• distance and determine strengthof
  influence
• this is approximation of real physicalinterferenc
  e image
• drawback visible pattern

7
Construction of encrypted interference image

• Interference image
• consider the influence of one point from
• previously it influenced a point at distance
  with
• a ring at distance consists of cells of size
• intensity of each cell defined by value of hash
  function
• the angle is taken modulo repetition of
  the same sequence
• key is needed for computation of values
• interference image does not containvisible
  patterns

8
Scheme overview
9
Reconstruction of watermark

• General idea
• determine the intensity of each point
  of the watermark image
• take the watermarked image (cover image
  interference image)
• generate interference image for point
• compare them
• if there is strong similarity assign
  a high value

Low similarity small pixel value
High similarity large pixel value
Example with simple interference image not with
encrypted usually the encrypted would be used
watermarked image
interference image for
10
Reconstruction of watermark

• Idea (continued)
• look for the highest valued points
• check whether these points form a valid watermark
• use equilateral triangles
• certain number of triangles of specific edge
  length must be found to form a watermark
• Properties
• for the points actually in
• there will be a large similarity between the
  watermarked image and the interference image
• there are points to check
• for each point operations are needed to
  compare the images
• key is needed for reconstruction

11
Reconstruction of watermark

• Additional operations
• the reconstruction must be repeated for all
  rotations with degree
• this yields resistance against rotations of the
  image
• the scale factor of the image must be determined
• take a small part of the image
• there should be at least one white point from
  in this part
• check different scale factors and perform
  reconstruction of this image part for it
• determine scale factor with largest peak value
  this peak corresponds to a white point from the
  watermark
• perform full reconstruction with this scale factor

12
Experimental results

• Evaluation against StirMark 3
• 77 of 89 tests passed
• Cropping all passed
• Removing rows/columns only lighter one passed
• Flip all passed
• Scaling all passed
• Change aspect ratio all passed
• Rotation with cropping for rotation smaller
  than 30 passed
• Rotation with cropping and scalling for
  rotation smaller than 30 passed
• Shearing only simple ones passed
• Linear transformations not passed
• StirMark not passed
• Gaussian filtering all passed
• Sharpening all passed
• Median filtering all passed
• LRAttack all passed
• JPEG compression all passed

13
Conclusions

• Results
• robustness against attacks basing on filtering
  and local editionsof image shown
• robustness against some linear transformations
  (rotating, scaling) shown
• Further work
• methods for detection of general linear
  transformations needed
• methods for detection of nonlinear
  transformations needed

14
Thank you for your attention!