Detecting Codes

Evaluating features for extraction for blind steganalysis classifiers

A presentation of material from this article: http://www.sciencedirect.com/science/article/pii/S0165168408001138

Transcript

3. Steganography 101 011100010011101000101000 Higher-Order Bits Lower-Order Bits
4. Steganography 101 011100010011101000101000 01110001 = 113 01110000 = 112 11110001 = 241       111100010011101000101000 011100000011101000101000
5. Steganography 101 Image Steganography: hiding data in the lower-order bits of digital images “Attack at midnight”
6. Blind Steganalysis Steganalysis ≠ Cryptanalysis With cryptanalysis, the goal is to decode With steganalysis, the goal is to detect A steganographic method is considered “broken” if it can merely be detected that the image has been tampered with Image blind steganalysis needs to simply detect that something in the lower-order bits is fishy
7. 1. Framework for image blind steganalysis 2. Feature extraction for blind steganalysis 3. Analysis of features Agenda
8. Framework Design science artifact!
9. M O T I V A T I O N Which features are the most likely to be useful in blind steganalysis?
10. Image Quality Metrics Examples •  Mean Square Error •  Multiresolution Distance Measure •  Weighted Spectral Distance •  Structural Content •  Cross Correlation •  Weighted Spectral Distance •  Median Block Weighted Spectral Distance •  Normalized Absolute Error •  HVS-based L2 •  Angle Mean •  Image Fidelity
11. PDF Moments of Subband Coefficients Decompose the image by applying high-pass filters •  Horizontal •  Vertical •  Diagonal Also do a low pass filter and recursively decompose  
12. PDF Moments of Subband Coefficients
13. Histogram: A plot of the number of pixels at each color value along a spectrum COM of Histogram Characteristic Functions
14. COM of Histogram Characteristic Functions Different “Centers of Mass”
15. CF Moments of Subband Histograms Perform a 2-level wavelet decomposition But the diagonals now get decomposed as well Do histogram analysis of the subbands
16. Statistical Analysis of Co-occurrence Matrix Co-occurrence Matrix populated by every pair of pixels and represents how close any two pixels are No Steganography Steganography
17. Merging Spatial and DCT features Gradient Energy Pick a direction When  deriving  a  pixel’s   value,  use  it’s  difference   from  the  last  pixel   Ver8cal  Gradient  Energy  =  Sum  of  squares  of  ver8cal   differences  divided  by  #  of  pixels     Horizontal  Gradient  Energy  =  Sum  of  squares  of   horizontal  differences  divided  by  #  of  pixels     Total  Energy  =  Ver8cal  +  Horizontal  
18. Merging Spatial and DCT features DCT coefficients “Every four adjacent blocks of 8x8 pixels were grouped i n t o a m a c r o b l o c k f o r parameter estimation, and then all DCT coefficients except for the DC components in a macroblock can be modeled as Laplacian distribution… Basically break the image into chunks, find the gradient energy of the chunks, reduce via statistics
19. 1: Collect images! 2: Apply steganography algorithms! 3: Extract features! 4: Calculate differences!
20. Results!
21. Criticisms Paper Criticisms Poorly written Second section, evaluating classifiers, that is mostly just a summary, not an experiment Features not really explained, and citations are often secondhand sources
22. Paper Contributions General framework for steganalysis evaluation Good jumping-off point for future steganalysis, knowing which feature is “best” Cited 93 times!
23. Questions ?