Kernel Fisher Discriminant for Steganalysis of JPEG Hiding Methods

1
Kernel Fisher Discriminant forSteganalysis of
JPEGHiding Methods

• Jeremiah Harmsen
• William Pearlman
• Rensselaer Polytechnic Institute
• Center for Image Processing Research

harmsj_at_rpi.edu
2
Overview

• Motivation
• JPEG Coding Process
• Kernel Fisher Discriminant/Least Squares
• Detection schemes
• Coefficient indices
• Spatial Method
• Conclusion

3
JPEG Coding
4
Processing Sequence
5
Kernel Fisher Discriminant

• Kernel Method
• Mapping
• Kernel
• Kernel Fisher
• Nonlinear Extension of Fisher
• Maximize interclass mean
• Minimize intraclass scatter
• Kernel Least Squares
• KFLD equivalent to KLS projection onto /- 1

6
Features
Frequency Division
Q. Coeff. Index Histograms
0
1
5
6
14
15
0
2
4
7
13
16
26
3
8
. . .
12
17
25
9
11
18
24
10
19
23
20
22
26
25
24
21

27
63
7
Experiments

• F5 Hiding Algorithm (rev. 12)
• Modifies Quant. DCT Coeff. Indicies
• 512x512 Color
• 10, 20, , 100 Embeddings
• F5 calculated capacity
• Kernel Fisher Discriminant
• Concatenated Q. DCT histograms
• Exponential Kernel
• LS-SVMLab
• ?2, ? found using cross-validation

8
Fast Results
9
Spatial Method

• Kernel Fisher Discriminant
• Features
• Fridrich
• Estimate embedding rate
• Longer processing path
• Feature Vector
• ?1, ?2, ?3
• ?k probability kth DCT element used in
  embedding

10
Spatial Results
11
Conclusions

• Large embedding detectable in indices
• Low embedding rate detectable using spatial method

12
Questions?

• harmsj_at_rpi.edu
• http//www.rpi.edu/harmsj

The support of the Air Force Research Labs,
Center for Integrated Transmission and
Exploitation is gratefully acknowledged.
13
(No Transcript)
14
Index Histogram
15
(No Transcript)
16
Fisher Linear Discriminant

• Maximize interclass mean
• Minimize intraclass variance