ECE643 DIGITAL IMAGE PROCESSING

1
ECE643DIGITAL IMAGE PROCESSING

• Steganalysis versus Splicing detection
• Paper by
• Yun Q. Shi, Chunhua Chen, Guorong Xuan and Wei Su
• By
• 
  Nehal Patel
• Siddharth Samdani

2
Agenda

• Steganography
• Splicing
• Relation between stegnalysis and splicing
  detection
• Current stegnalysis Method
• Apply stegnalysis method to detect spliced images
• Result
• Conclusion

3
Steganography

• From the Greek word steganos meaning covered
  and the Greek word graphie meaning writing
• Steganography is the process of hiding of a
  secret message within an ordinary message and
  extracting it at its destination
• Anyone else viewing the message will fail to
  know it contains hidden/encrypted data

4
Splicing

• Definition
• The spliced image is a composite
  picture generated by combining image fragments
  from the same or different images without further
  post-processing such as smoothing of boundaries
  among different fragments.
• Image splicing is one of the simple
  commonly used image tampering schemes is often
  used as an initial step for image tampering.
• With modern image processing techniques,
  image splicing can be hardly caught by human
  visual system (HVS).

5
Fig.1
A
C
B
B,C Original Images A Spliced Image
6
General comparison(stegnalysis and splicing
detection)

• Different motivation and objectives
• Steganography encodes information bits and then
  embeds bits into cover image where as splicing is
  to replace one or more parts of the host image
  with fragments from the same host image or other
  source image.
• Statistical artifacts are different.
• Both try to reduce difference between cover image
  and modified image.
• Stegnography is more global while splicing is
  more local ( stegnography often embeds data in a
  cover image as widely as possible, while splicing
  just touches the part of host image).
• Splicing generally change the content of a host
  image, therefore the relative change between host
  image and its spliced version is larger.
• Since stego images and spliced images are
  touched, the stegnograhic and splicing operation
  cause disturbance on the smoothes, regularity,
  continuity, consistency, and periodicity of the
  image.
• Above statistical artifacts can be detectable
  using well designed natural image model.

7
Measurement

• The following measurements are used to measure
  the strength of the change brought to the cover
  image or host image.
• Subjective measurement for steganalysis
• HVS (Human visual system)
• Objective measurement for Steganalysis
• BPP(bits per pixel for steganography )
• MSE(mean square error) or PSNR (peak signal to
  noise ratio)
• For Splicing MSE or PSNR can be an objective
  measure candidate.
• Stegnography and splicing both are detectable by
  machine learning schemes.

8
Machine Learing

• A well designed natural image model can separate
  stego or spliced images.
• Image model consists of a feature vector which
  characterize a given image.
• With dataset comprising both natural image and
  non-natural image, universal stegnalysis or
  splicing detection can be carried out under the
  machine learning framework.

9
Image dataset

• The Columbia Image splicing detction Evaluation
  dataset can be used.
• Contains 933 authentic and 912 spliced images
  size of 128 X 128.
• These data sets are created by DVMM(digital video
  and multimedia lab) Columbia university.

10
Image Dataset
11
Classifier, classifications and result analysis

• Classifier (SVM-support vector machine)
• 5/6 of authentic and 5/6 of the spliced images
  are used to train a SVM classifier and remaining
  1/6 of these images are used to test the trained
  classifier.
• ROC (receiver operating characteristic ) curve is
  obtained to demonstrate the performance of
  trained classifiers .
• AUC (Area under the ROC curve ) or TN (true
  negative) and TP (true positive) rate methods
  also can be used to show classifiers performance.

12
Applying natural image models created in
Universal Steganalysis to Splicing Detection

• Some Universal Steganalysis methods
• Hyu and Farids Method
• Shi et al.s method
• Zou et al.s method
• Chen et al.s method

13
An Advanced Natural Image Model to Boost Splicing
Detection Capability

• Novel natural image model

14
Natural Image Model Components

• Multi-size Block Discrete Cosine Transform
  (MBDCT)
• Splicing procedure changes the frequency
  distribution of a host image, these changes are
  reflected by coefficients of BDCT.
• Correlation changes in various patterns and is
  complicated due to a number of factors ( for ex.
  Different host images)
• These changes cannot be captured effectively by
  one single block size BDCT but with various block
  size the MBDCT coefficients can perceive the
  frequency changes in a variety ways.
• The application of is as follows nxn BDCT
• The image is divided into nxn non
  overlapping blocks. Then DCT is applied
  independently on each block, which gives a 2-D
  array consisting of BDCT coefficients of all the
  blocks. Using individual block size corresponding
  BDCT 2-D array is obtained. Each of this BDCT 2-D
  array generates corresponding features.

15
Natural Image Model Components

• Moment Based Features
• The moment based features consist of 1-D and 2-D
  characteristic functions of image 2-D array, its
  prediction error 2-D array and all the wavelets
  sub bands.
• Wavelet analysis, prediction error,
  characteristic functions and 2-D histogram are
  key features of moment based features.
• Wavelet analysis are used due to their superior
  multi-resolution and space-frequency analytical
  capabilities. While wavelet transform is suitable
  to catch local changes in spatial frequency
  domains and hence good for splicing detection.
• The 2-D prediction error array is used to reduce
  the influence caused by diversity and enhance the
  statistical artifacts introduced by splicing.
• The 2-D histogram measures the intensity change
  of pixels with respect to their neighbors and
  thus can reflect statistical effects of splicing
  artifacts more efficiently.

16
Natural Image Model Components

• Markov based features
• Markov based features are able to reflect the
  statistical changes.
• In this image pixels are predicted with the help
  of neighboring pixels and the prediction error
  image is generated by subtracting the prediction
  value from the pixel value.
• The above step gives difference 2-D arrays from
  the given image 2-D array or coefficient 2-D
  array.
• The difference 2-D array is then applied to a
  predefined threshold.
• These 2-D difference array is modeled by Markov
  process and then transition probability matrices
  is calculated for each difference array. The
  values of this matrices are used to build another
  part of natural image model.
• By predicting an image pixel or a BDCT
  coefficient using its immediate neighbor assumes
  that the disturbances caused by splicing can be
  emphasized by prediction error.
• Combining moments based features markov based
  features makes this novel natural image model
  more effective

17
Results

• The implementation results of novel natural image
  model on image dataset
• The averaged ROC curve of 20 experiments obtained
  by applying the proposed natural image model is
  show below, in which ROC curves from experiments
  performed using Universal Steganalysis Methods
  are also included.
• The implementation of novel approach gave TN rate
  91.52(2.19), TP rate 92.86(1.72), accuracy
  92.18(1.30)and averaged AUC 0.9537(0.0112).

18
Result
The ROC curves of applying the natural image model
19
Detecting Real Images

• The results of trained classifier from the 20
  random experiments was used to test the three
  images in Fig. 1.
• The test results are shown in table below, in the
  table it can be seen that among the 60 image
  test, 57 provided correct classification.

20
Conclusion

• The figure shows that with a well designed
  natural image model, stego images spliced
  images can be separated from natural images in a
  feature space.

21
Conclusion

• Different in target and application,
  steganography splicing have some aspects in
  common.
• One such aspect is that both cause the touched
  images to deviate from natural images.
• A novel natural image model, from the state of
  art steganalysis schemes was presented applied
  to splicing detection, which demonstrated
  advancement in splicing detection.
• Lessons learnt from steganalysis can be applied
  to splicing detection, while steganalysis can
  learn something from splicing detection as well.