Steganography and Block-based Quantitative Steganalysismcl.usc.edu/wp-content/uploads/2014/07/seminar_slide.pdf · Introduction Deﬁnition What are steganography and steganalysis?

Steganography and Block-based Quantitative Steganalysis

Haiqiang Wang

Advisor: C.-C. Jay Kuo

Viterbi School of Engineering, USC

07/26/2014

Haiqiang Wang Steganography and Steganalysis 1 / 38

Outline

1 IntroductionDefinitionInformation Hiding

2 Research on steganographyTransform domainSpatial domain

3 Research on SteganalysisMachine Learning ApproachCurse of Dimensionality

4 Block-based Quantitative SteganalysisWhy block-basedBlock-based steganalysis approach


Introduction Definition

What are steganography and steganalysis?

DefinitionSteganography is the art of communicating in a way which hides theexistence of the communication.

Steganalysis is the science to detect, or estimate the hidden data fromobserved data with little knowledge about steganography algorithm.

Figure 1: Communication with invisible ink


Introduction Definition

Framework of Steganography and Steganalysis

Figure 2: Components of the steganographic channel

Cover X: input image; Stego Y: output image

Payload p: relative message length (bpp: bit per pixel)

HOLUB, V. CONTENT ADAPTIVE STEGANOGRAPHY?DESIGN AND DETECTION. Diss. State University of New York, 2014.


Introduction Information Hiding

Common information hiding techniques

Figure 3: Information hiding classification

V. Nagaraj., et al. "Overview of Digital Steganography Methods and Its Applications." IJAST 60 (2013): 45-58.



Examples of Watermarking

Figure 4: Watermarking used in documentation



Examples of Watermarking

Figure 5: Perceptible and Imperceptible Watermarking used in video



Example of Steganography

Figure 6: Steganography using UNIWARD



Comparison with Watermarking

Table 1: Comparison of Watermarking and Steganography

Watermarking SteganographyGoal copyright protection covert communication

Information host or owner any kind of informationPerceptible either visible or imperceptible statistically undetectable

Receiver point-to-multiple points point-to-pointCapacity not important important

Robustness important not necessary



Why we study Steganography and Steganalysis?

Used in military purpose, intelligence services, or by terrorists via publicaccess channelIn October 2001, An article from the New York Times claims theterrorists use covert communication to prepare and execute the 11September 2001 terrorist attack

Figure 7: Publications number


Research on steganography Transform domain

How to embed information while minimizing distortion?

Transform domain (wavelet, DCT) and spatial domainJP Hide&Seek, Jsteg, MBS1, MMX, nsF5, OutGuess and PerturbedQuantization (PQ)LSB, HUGO, WOW and UNIWARD

Brute-force embedding and empirical embeddingContent adaptive spatial domain embedding performs better

Typical algorithms (PQ, LSB) and state-of-the-art (UNIWARD).

Pevny, T., et al."From blind to quantitative steganalysis." SPIE Electronic Imaging. 2009.


Research on steganography Transform domain

Perturbed Quantization in DCT domain

Think of EE 669 homework 2 (JPEG cpmpression)

8 ∗ 8 image blocks, shifted by −128, two-dimensional DCT, divided byquantization matrix, round to nearest decimal and entropy coding

Fractional part is around 0.5, either round up or down

Disadvantage: change the pixel value more than 1

Figure 8: Key idea of Perturbed Quantization

Fridrich, Jessica., et al."Perturbed quantization steganography." Multimedia Systems 11.2 (2005): 98-107.


Research on steganography Spatial domain

Least Significant Bit Replacement

Figure 9: LSB replacement embedding

Ker, Andrew D. "Steganalysis of embedding in two least-significant bits." ITIFS on 2.1 (2007): 46-54.



Weakness of LSB

Figure 10: Embedding changes statistics



UNIWARD–UNIversal WAvelet Relative Distortion

Minimize a well defined embedding distortion functionEmbedding in noisy regions, complex texture and avoid smooth regions

Figure 11: Steganography using UNIWARD

Denemark, Tomas, et al. "Further Study on the Security of S-UNIWARD." SPIE Electronic Imaging. International Society for Opticsand Photonics, 2014.



Filter banks

β = {K(1),K(2),K(3)} to evaluate smoothness in multiple directions

K(1) = h · gT ,K(2) = g · hT ,K(3) = g · gT

Figure 12: low-pass and high-pass filter



Wavelet decomposition

Directional residuals is the convolution between the filter and image:

Wk = Kk ? X (1)

The distortion of changing one pixel is defined as:

D(X,Y) =

3∑k=1

∑u,v

|Wkuv(X)−Wk

uv(Y)|ε+ |Wk

uv(X)|(2)

kth decomposition, Wkuv is the uvth wavelet coefficient

For each filter, compute the relative wavelet coefficient change w.r.t. thecover image



Embedding cost function

The additive approximation (with subscript "A") of distortion function is:

DA(X,Y) =

n1∑i=1

n2∑j=1

ρij(X,Yij)[Xij 6= Yij] (3)

ρij(X,Yij) is the distortion of changing ijth pixel:

ρij(X,Yij) = D(X,Yij) (4)



Embedding cost

Figure 13: Cover image and embedding distortion



Why does it work?

D(X,Y) =

3∑k=1

∑u,v

|Wkuv(X)−Wk

uv(Y)|ε+ |Wk

uv(X)|(5)

Pixel in noisy region has large wavelet coefficients, embedding distortionis small

Even one smooth direction |Wkuv(X)| will lead to large embedding

distortion

Selected pixels are hard to model in all directions



Embedding distribution of UNIWARD

Figure 14: Content adaptivity of UNIWARD


Research on Steganalysis Machine Learning Approach

Two approaches

Statistical signal detectionDerives the detector from statistical modelAttacks specific embedding algorithm and needs sufficient samples

Standard machine learning approachEnsemble classifierSupport Vector Machine (SVM): binary estimationSupport Vector Regression (SVR): quantitative estimation

Ker, Andrew D., et al. "Moving steganography and steganalysis from the laboratory into the real world." Proceedings of the first ACMworkshop on Information hiding and multimedia security. ACM, 2013.



Ensemble classifier

Lower complexity but can handle large feature set

Also known as bootstrap aggregation or bagging

Figure 15: Diagram of ensemble classifier

HOLUB, V. CONTENT ADAPTIVE STEGANOGRAPHY: DESIGN AND DETECTION. Diss. State University of New York, 2014.



Support vector regression

Principle: Use regression tools to learn the relationship between featurevector and payload

Assumption: Feature changes predictably with payload

ψ̂ = arg minψ∈F

1L

L∑i=1

e(ψ(xi), yi) (6)

xi = f(ci) ∈ Rd is feature vector computed from image ci embedded withrate yi ∈ [0, 1].

Find a mapping function ψ̂ : Rd 7→ [0, 1] that minimizes the estimationerror

Pevny, T., et al."From blind to quantitative steganalysis." SPIE Electronic Imaging. 2009.



Support vector regression

Figure 16: SVR used in quantitative steganalysis


Research on Steganalysis Curse of Dimensionality

Feature dimension problem

Table 2: Feature for Steganalysis (Spatial domain)

Name Dimension Published Year AuthorSPAM 548 2010 T.Pevny and J. Fridrich

CDF 1234 2010 J. Kodovsky and J. FridrichSRM 34671 2012 J. Fridrich and J. Kodovsky

PSRM 12870 2013 V. Holub and J. FridrichCSR 1183 2014 T. Denemark and J. Fridrich


Block-based Quantitative Steganalysis Why block-based

Why block-based approach

Ideas from video coding, 16*16, 16*8, 8*4, . . .Extremly high dimension feature to handle heterogeneous images

Group homogeneous blocks togetherKnowledge about embedding distribution (ROI)

Noisy region, edge and complex texture and avoid smooth region

Ability to achieve better payload estimation


Block-based Quantitative Steganalysis Why block-based

Region of Interest

(a) ROI at payload = 0.25 (b) Embedding dist. at payload = 0.25

Figure 17: ROI and embedding distribution at big payload


Block-based Quantitative Steganalysis Block-based steganalysis approach

Block-based steganalysis approach

1 Block classification2 Different feature of different block group3 Next: support vector regression4 Future work: adaptive block partition



Block classification

3 block types, smooth, edge and texture

Figure 18: Cover image and embedding distortion



Block classification

Figure 19: Distortion map and classification grid



Feature extraction

1183 CSR feature: Content-selective residual

Figure 20: CSR feature components

Denemark, Tomas, et al. "Further Study on the Security of S-UNIWARD." SPIE Electronic Imaging. International Society for Opticsand Photonics, 2014.



Feature extraction

Residual order and truncation (Th and Tc):

Embedding probability

pij =exp(−λρij)

1 + exp(−λρij)(7)



Feature extraction

Pixel classes: compare pij with two parametersType L: when pij(X, α) > tLType s: when pij(X, α) < ts

Example: 1st 1D histogram2d+1 = 4 pixel classes, [s s], [s L], [L s] and [L L].22 = 4 to 3 in 1st, 23 = 8 to 6 in 2nd, and 24 = 16 to 10 in 3rd



Feature reduction

Too large feature for 32*32 blocks

Figure 21: CSR feature of smooth, edge and texture blocks



Feature reduction

PCA to reduce each feature subset

ANOVA to select feature set for each blocks

Figure 22: Reduced feature dimension with PCA


Future work

Future work

Quantitative Steganalysis using SVR

Adaptive block partition


Future work

Thanks


Steganography and Block-based Quantitative Steganalysismcl.usc.edu/wp-content/uploads/2014/07/seminar_slide.pdf · Introduction Deﬁnition What are steganography and steganalysis?

Documents