HUBFIRE - A MULTI-CLASS SVM BASED JPEG

STEGANALYSIS USING HBCL STATISTICS AND FR INDEX

Veena H. Bhat

, Krishna S., P. Deepa Shenoy, Venugopal K. R.

Department of Computer Science and Engineering, University Visvesvaraya College of Engineering, Bangalore, India

L. M. Patnaik

Vice Chancellor, Defence Institute of Advanced Technology, Pune, India

Keywords: Steganography, Steganalysis, Support Vector Machines, Huffman coding.

Abstract: Blind Steganalysis attempts to detect steganographic data without prior knowledge of either the embedding

algorithm or the ‘cover’ image. This paper proposes new features for JPEG blind steganalysis using a

combination of Huffman Bit Code Length (HBCL) Statistics and File size to Resolution ratio (FR Index);

the Huffman Bit File Index Resolution (HUBFIRE) algorithm proposed uses these functionals to build the

classifier using a multi-class Support Vector Machine (SVM). JPEG images spanning a wide range of

resolutions are used to create a ‘stego-image’ database employing three embedding schemes – the advanced

Least Significant Bit encoding technique, that embeds in the spatial domain, a transform-domain embedding

scheme: JPEG Hide-and-Seek and Model Based Steganography which employs an adaptive embedding

technique. This work employs a multi-class SVM over the proposed ‘HUBFIRE’ algorithm for statistical

steganalysis, which is not yet explored by steganalysts. Experiments conducted prove the model’s accuracy

over a wide range of payloads and embedding schemes.

1 INTRODUCTION

Steganography is the art or practice of concealing a

message, image or file within another image or file

in such a way that the sender would be able to

communicate to the intended recipient covertly.

Steganalysis is the science of detecting messages

(payload) hidden using steganography; it often deals

with scrutinizing the carrier media for anomalies or

non-ideal artifacts that are introduced in the process

of steganography. While cryptography conceals data

by encrypting the message, steganography achieves

privacy by hiding the very existence of the message

in an innocent looking cover object (Fridrich et al.,

2007). According to the terminology as agreed in the

First International Workshop on Information Hiding

(Pfitzmann, 1996), the embedding of a text or

‘payload’ in a ‘cover image’ gives a ‘stego-image’.

Transductive Support Vector Machines (SVM),

developed by Vladimir Vapnik, a kernel-based

learning technique, have their roots in a geometrical

interpretation of the classification problem to find a

separating hyperplane (classifier) mapped on to a

higher dimensional input space by optimization of a

convex cost function (Vapnik, 1998 and Kristin,

2000). SVMs based on Structural Risk Minimization

principle (Vapnik, 1995), work on maximizing the

error margin of the training set, and have been

effectively used for classification and regression of

real-world data. SVMs have been effectively used in

linear and non-linear pattern recognition, regression

and classification problems.

2 RELATED WORK

A stego process is defined as a ε-secure process if

the Kullback-Leibler divergence ∂ between the

probability density functions of the cover document

cover

and those of this very same content embedding

a message p

stego

is less than ε:

∂( p

cover

stego

) 1≤ ε.

(1)

Veena H Bhat works as Professor of Information

Technolo

at IBS-Ban

alore, India.

447

H. Bhat V., S. K., Shenoy P., K. R. V. and M. Patnaik L. (2010).

HUBFIRE - A MULTI-CLASS SVM BASED JPEG STEGANALYSIS USING HBCL STATISTICS AND FR INDEX.

In Proceedings of the International Conference on Security and Cryptography, pages 447-452

DOI: 10.5220/0002989004470452

 SciTePress

The process is called ‘secure’ if ε = 0, and in this

case the steganography is perfect, creating no

statistical differences by embedding of the message

(Miche et al.,

2009), steganalysis would then be

impossible.

Support Vector Machines (SVM) are one of the

most effective methods for pattern recognition and

multivariate classification (Joachims, 1998; Liang,

2004). A binary SVM identifies whether an instance

belongs to a class ensuring a minimum

generalisation error with a hyperplane to distinguish

the two classes. Multi-class SVM algorithms such as

OVA (One Versus All) use winner-takes-all strategy

(Boser et al., 1992) and OAO (One Against One)

that use max-wins voting method (Liang 2004) are

in vogue. Binary SVM are used for Steganalysis as

proposed by Lyu and Farid using higher order

statistics (Lyu and Farid, 2002). Some of the blind

steganalysis techniques developed using multi-class

SVM for classification include analysis based on

run-length histograms (Dong and Tan, 2008).

Quantized DCT features are used as input features to

a SVM classifier (Pevny and Fridrich, 2007).

Steganalysis using SVM classifier for a huge

number of input features, discussed by Xuezeng et

al.,(2008) employs features extracted from

histograms.

3 TESTED EMBEDDING

SCHEMES

Steganographic techniques, based on the nature of

embedding scheme adopted, are classified into

spatial domain, frequency domain and adaptive

steganography. The stego-image database is

populated, for all these three embedding schemes.

3.1 LSB – Least Significant Bit

Among the several techniques employed for

steganography, LSB Steganography is a popular

spatial domain technique because of its robustness,

fine concealment, high steganographic-embedding

capacity and easy realization. In this work, we have

employed a scheme that uses both sequential

embedding and scattered embedding (Advanced

LSB technique) (Chen et al., 2006). To build the

stego-image database we have used the open source

Matlab code, for this technique by Luigi Rosa.

3.2 JPHS – JPEG Hide and Seek

JPEG Hide-&-Seek (JPHS) is a transform domain

tool designed for JPEG files and lossy compression.

JPHS uses least significant bit overwriting of the

Discrete Cosine Transform (DCT) coefficients used

by the JPEG algorithm. Though JPHS fails to

preserve the DCT histogram statistics, its statistical

detectability is among the lowest based on the

results reported in (Tomas, 2007).

The stego-image database is built using the open

source Matlab code for this technique, namely

‘steganoh’, designed by Francisco Echegorri. This

program hides a text file in a grayscale image by

disordering it into lines and columns using the

Ranpermut-Encryption algorithm.

3.3 MBS – Model Based

Steganography

Unlike LSB and JPHS embedding techniques, the

MBS Technique is an adaptive technique, proposed

by Sallee, which tries to model statistical properties

of an image and preserve them during embedding

process (Sallee, 2005). The embedding operation

employs a nonadaptive arithmetic decoder and the

coefficients in each histogram bin are modified with

respect to embedding rule, while the global

histogram and symbol probabilities are preserved.

Attacks such as Blockiness (Ullerich and Westfield,

2008) can detect this embedding scheme. Our stego-

image database for MBS is built using the open

source Matlab code (Sallee, 2005).

Table 1: First Order Histogram statistics for different

embedding schemes.

First

Order

Statistics

Cover LSB JPHS MBS

Mean

4.2948 4.2947 4.3011 4.2951

Variance

0.9936 0.9937 0.9918 0.993

Skewness

-0.8237 -0.8235 -0.8192 -0.8221

Kurtosis

4.2406 4.2397 4.249 4.2364

Energy

0.3438 0.3438 0.3447 0.3438

Entropy

1.2852 1.2853 1.2849 1.2852

File size

(kb)

163 164 167 150

The stego-images created as mentioned in

sections 3.1, 3.2 and 3.3 and the cover images were

checked for first order histogram statistics, table 1

shows some of the observations for images of

resolution 2048X1536 for a maximum payload.

From table 1, it can be deduced that in the absence

of reference to cover images, it would be difficult to

differentiate between stego and cover images using

SECRYPT 2010 - International Conference on Security and Cryptography

448

first order histogram statistics for blind steganalysis.

Throughout this paper, for convenience we

address these techniques as LSB, JPHS and MBS.

4 HBCL STATISTICS AND FR

INDEX

4.1 HBCL - Huffman Bit Code Length

Statistics

JPEG, a lossy compression format, employs

sequential Huffman Encoding for data compression

wherein symbols (DCT coefficients in this case) are

taken and encoded with variable length codes that

are assigned based on statistical probabilities. A

grayscale image employs 2 Huffman tables, 1 each

for AC and DC portions. When a JPEG image is

embedded with a particular payload, certain non-

ideal JPEG artefacts are introduced in the given

image, though the enormity of this deviation varies.

On artefacts the nature of the DC HBCL statistics of

the images in the populated image-database, we

have considered HBCL statistics from two to five

bits only, as the variation in HBCL from the 6

bit

onwards is negligible.

One of the scoring features of Huffman coding

algorithm is its ‘unique prefix property’ that is no

code is a prefix to any other code, making the codes

assigned to the symbols unique. This fact further

supports our choice of the HBCL statistics for our

evaluation features to be efficient as the JPEG

artefacts introduced by steganography on an image

becomes unique and hence can be predicted using a

suitable classifier or a prediction model. In our work

we extracted these statistics using Huffman

decoding for a JPEG image in Matlab.

4.2 FR Index – File size to Resolution

Ratio

When a raw image is compressed by JPEG

compression, based on the resolution of the image,

its quality and compression ratio the resulting JPEG

file takes up a particular file size. This indicates that

the file size and the resolution of an image and

further its quality are interrelated. Thus the ratio of

file size of the image in bytes to that of its resolution

is found to be unique and in a certain range for a

given resolution, this functional is termed ‘FR

Index’ and used as one of the inputs to build the

prediction model. Table 4 shows the range of FR

Index for the resolutions used in our database.

5 IMPLEMENTATION

5.1 Image Database

For effective performance evaluation of a

steganalysis technique, the data set employed in the

experiment, which validates the prediction model for

its sensitivity and specificity, is very important. This

work adopts JPEG grayscale images, as it is harder

Figure 1: Processing details to create the image database.

to detect hidden data in them compared to color

images where steganalysis can utilize dependencies

between color channels.

This proposed model uses a dataset of images

that would include those images often found on the

internet and private domain, at the same taking care

to create the entire image dataset within the quality

range of 57 to 72, but since one has a freedom in

selecting the quantization table when compressing

an image using the JPEG algorithm, there is no

standard definition of a quality factor. Therefore the

quality factor of the images in the dataset is

approximated by deploying the publicly available

‘JPQ-JPG Quality Estimator’ essentially, JPQ

estimates the quality factor of the image by

comparing its quantization table to the suggested

quantization table in the JPEG standard that employs

optimized Huffman tables. This quality estimation

program assumes the images to be subsampled at

2:1:1 and an approximate recompression error

ranging from -0.49 to +0.44, (which can be

considered to have a negligible effect on the JPEG

features) (Bhat, H, V. et al., 2010).

Among the 10 image sets (Basic Dataset) of 100

images each evaluated, the least resolution is 75X75

and the highest being 2048X1536 which was taken

from Jegou’s database (Herve, 2008). Table 2

illustrates details of the image database used in our

performance evaluation. The Basic Set lists the

different resolutions used with 100 images in each

set; the Resized Set lists the set of images

HUBFIRE - A MULTI-CLASS SVM BASED JPEG STEGANALYSIS USING HBCL STATISTICS AND FR INDEX

449

correspondingly resized from column 1 to validate

against the JPEG Resizing Error. The range of the

FR Index for each set is also mentioned.

5.1.1 Cover Images and Stego Images

For each tested method, the cover grayscale JPEG

images and several stego grayscale JPEG images

embedded with different payloads are prepared. By a

cover image, it is understood as an image into which

no message is embedded. LSB, JPHS and MBS were

the embedding schemes employed with three

different payloads in each scheme and for each

resolution. Thus for the entire set of 2,000 cover

images, 18,000 stego images with different

parameters were generated and evaluated.

Table 2: Image dataset details.

Basic Set FR Index Resized

set

FR Index

75X75 0.12 – 0.59 - -

130X130 0.09 – 0.45 100X100 0.06 – 0.42

214X214 0.09 – 0.42 200X200 0.08 – 0.34

384X256 0.08 – 0.37 300X300 0.06 – 0.32

481X321 0.06 – 0.31 400X400 0.07 – 0.47

512X512 0.03 – 0.41 500X500 0.04 – 0.52

720x480 0.07 – 0.43 600X600 0.07 – 0.37

800X600 0.03 – 0.31 700X700 0.05 – 0.46

1024X768 0.01 – 0.03 800X800 0.02 – 0.05

2048X

1536

0.004 – 0.03

900X900 0.01 – 0.05

1KX1K 0.005 – 0.03

In order to check for consistency of the features

extracted for the model and to validate this over

JPEG Resizing Error, the basic dataset is resized.

Table 2 describes the nature of resizing performed

and related observations. Thus, a set of 2000 JPEG

cover images are obtained on which the consistency

of the prediction model is evaluated.

5.1.2 Payloads

Three different payloads for LSB, JPHS and MBS

each are tested for the 2,000 images. In case of LSB

and MBS embedding schemes, a random data in

corresponding to the payload size is embedded,

where as in JPHS a text file of the respective size is

embedded after encryption. In case of MBS, we used

a code that uses optimized Huffman tables and the

embedding rate on an average was found to be

1.0431 bits per change (bpc).

For images with resolution 100X100 and less,

340 bytes, 500 bytes and 720 bytes were the three

payload sizes used; however our prediction model

produced consistent results for payload sizes less

than 300 bytes too. For the remaining resolutions up

to that of 2048X1536, we use payloads of sizes 1024

bytes (1 KB), 2048 bytes (2KB) and 5120 bytes

(5KB). Thus the prediction model is evaluated over

20,000 images (inclusive of cover and stego

images).

5.2 Input Functionals used in the

Model

In the first phase, to predict steganography, the

resolution of the image, file size, quality and HBCL

statistics are considered. Though quality does not

play a pivotal role in building the prediction model,

it helps to build the image database. Resolution and

file size of an image are considered because the file

size is a direct consequence of the resolution of an

image and the JPEG compression performed on the

image. Hence, these factors are in a way affected by

the JPEG compression (the quantization and the

Huffman compression employed).

Table 3: Correlation values for the input attributes selected

for steganalysis.

HBCL statistics and the reason for choosing

them are explained in section 4.1. Correlation

between these factors is analyzed to check their

interdependency. Table 3 illustrates the correlation

between the functionals over the entire image

database. ‘2

HBCL’ indicates the number of

Huffman code of 2 bit length, ‘3 HBCL’ for number

of 3 bit length codes and so on.

5.3 Classifier

Multi-class Support Vector Machine (Burges, C.,

1998; Hsu, Chih-Wei., Lin, Chih-Jen., 2002) is used

as a classifier with number of classes set to 4

wherein class 1 indicates cover images, class 2; LSB

encoding, class 3; JPHS, class 4; MBS scheme. The

kernel function opted is Gaussian with suitable λ

and cost function with the size of the quadratic

programming set to 100. The gamma and cost

parameters, (c,γ) for the SVM classification are

estimated using grid-search over a multiplicative

SECRYPT 2010 - International Conference on Security and Cryptography

450

grid of 2

–2

to 2

over a sampling method of 10 fold

cross-validation. This is implemented using an open

source Matlab toolbox, SVM-KM.

Let K

be the entire set of ‘cover’ images in the

database. K

⊂K

is a set of ten different resolutions

of images; this set K

is embedded with a payload

employing three different steganographic methods as

described in section 3 thus generating a ‘stego-

image’ dataset denoted as K

. The set K

TRAIN

is used

to train the multi-class SVM where K

TRAIN

= K

∪ K

A dataset K

TEST

= K

∩ K

is used to test the trained

model in sets of 600 instances with 150 per class

such that K

TRAIN

∩ K

TEST

= Null. One such test is

illustrated in the figure 2.

Table 4: HUBFIRE Algorithm.

Input: The test image

Output: The image classified as genuine or stego

image.

1. Data Generation

a. Identify the resolution and file size of an

image for the sets, K

TEST

and K

TRAIN

b. Populate the image database consisting of

cover and stego images – LSB, JPHS,

MBS with payloads of 1KB, 2KB and

5KB.

2. Data Preprocessing

a. JPEG Huffman decoder is used to extract

four HBCL statistical data for each image

of the training set, K

TRAIN

b. Basic properties of the image – filesize,

quality, resolution are recorded.

c. Calculate the FR Index for each image.

3. The functionals created through the data

preprocessing step 2 is used to train the

Multi-class SVM with number of classes set

to 4.

4. The image to be tested is processed against

the model trained in step 3.

6 RESULTS

Figure 2 illustrates the confusion matrix of the One-

Versus-All SVM algorithm used to classify. The

model is highly sensitive to stego-images as against

the cover images which increases the false positive

response of the system.

The embedding schemes chosen to train the

model span over different domains as described in

section 3 making the model an efficient blind

steganalysis technique. It is observed that the trained

model is highly sensitive to MBS in comparison

with other embedding schemes.

The model has to be fine tuned to reduce the

false positive rate so as to give a reliable detection

Figure 2: Confusion matrix of HUBFIRE, the multi-class

classifier.

even with respect to cover image.

7 PERFORMANCE ANALYSIS

The results of the proposed model are compared

with the work of Pevny and Fridrich (2006).

A comparison is illustrated in the table 5. The

proposed model is found to have a high false

positive rate however it is highly sensitive for

classification among the stego-images, thus the

model serves as a reliable ‘stego-classifier’.

8 FUTURE WORK AND

CONCLUSIONS

Our future work will include optimising the multi-

class SVM model, by calculating the cost function to

reduce the false positive rate. The payloads

considered in this work are 1KB, 2KB and 5KB for

each of the embedding schemes, however, the model

does not analyse the response of the trained model

for each of the embedded payload individually. This

would reflect upon the sensitivity of the model, with

respect to varying payloads. Multi-class SVMs with

different kernels of radial basis function and sigmoid

function need to be explored.

As explained in section 6, the model is reliable

in stego-image detection that is embedded with

steganographic schemes that span over several

domains. The model gives an efficiency of 100%

with respect to detection of MBS.

HUBFIRE - A MULTI-CLASS SVM BASED JPEG STEGANALYSIS USING HBCL STATISTICS AND FR INDEX

451

Table 5: Confusion matrix of the multi-class classifier for the testing set. The leftmost column contains the embedding

algorithm. The remaining columns show the results of the classification of HUBFIRE and the model proposed by Pevný

and Fridrich (2006).

HUBFIRE (Pevný and Fridrich, 2006)

Cover LSB JPHS MBS Cover LSB JPHS MBS

Cover 70.67% 26.67% 2.67% 0.00% 96.45% 0.12% 0.20% 1.44%

LSB 0.00% 99.33% 0.67% 0.00% 0.08% 99.08% 0.53% 0.08%

JPHS 0.00% 0.67% 99.33% 0.00% 0.20% 0.12%

98.32

0.56%

MBS 0.00% 0.00% 0.00% 100.00% 1.44% 1.56% 0.72% 94.44%

REFERENCES

Bhat, H, V., Krishna, S., Shenoy, P, D., Venugopal, K, R.,

Patnaik, L, M., 2010. JPEG Steganalysis using HBCL

Statistics and FR Index. To be published in the

Proceedings of Pacific Asia Workshop on Intelligence

and Security Informatics.

Burges, C., 1998. A Tutorial on Support Vector Machines

for Pattern Recognition. Data Mining and Knowledge

Discovery, 2, 121–167.

Boser, B., Guyon, I., Vapnik, V., 1992. A Training

Algorithm for Optimal Margin Classifiers.

Proceedings of the Fifth Annual Workshop on

Computational Learning Theory, 5, 144–152.

Chen, Ming.,Zhang, Ru., Niu, Xinxin., Yang, Yixian.,

2006. Analysis of Current Steganography Tools:

Classification & Features. Proceedings of the 2006

International Conference on Intelligent Information

Hiding and Multimedia Signal Processing, 384-387.

Cristianini, N., Shawe-Taylor, J., 2000. An Introduction

to Support Vector Machines, Cambridge University

Press.

Herve, Jegou., Matthijs, Douze., Cordelia, Schmid., 2008.

Hamming Embedding and Weak Geometry

Consistency for Large Scale Image Search.

Proceedings of the Tenth European Conference on

Computer Vision, 304-317.

Hsu, Chih-Wei., Lin, Chih-Jen., 2002. A Comparison of

Methods for Multiclass Support Vector Machines.

IEEE Transactions on Neural Networks, 415-425.

Joachims, T., 1998. Text categorization with support

vector machines: Learning with many relevant

features. Proceeding of the Tenth European

Conference on Machine Learning (ECML), 137–142.

Kristin, P, Bennet., Campbell, Colin., 2000. Support

Vector Machines: Hype or Hallelujah ? SIGKDD

Explorations, 2, 1-13.

Lyu, Siwei., Farid, Hany., 2002. Detecting Hidden

Messages Using Higher-Order Statistics and Support

Vector Machines. Proceedings of the Fifth

International Workshop on Information Hiding, 2578,

340-354.

Miche, Yoan., Bas, Patrick., Lendasse, Amaury., Jutten,

Christian., Simula, Olli., 2009. Reliable Steganalysis

using a Minimum Set of Samples and Features.

EURASIP Journal of Information Security.

Pevný, Tomas., Fridrich, J., 2006. Multi-class Blind

Steganalysis for JPEG Images, Computer Science and

Software Engineering , 939-942.

Pevný, Tomas., Fridrich, J., 2007. Merging Markov and

DCT features for Multi-class JPEG Steganalysis,

Proceedings SPIE - Electronic Imaging, Security,

Steganography, and Watermarking of Multimedia

Contents, 03–04.

Pfitzmann, B., 1996. Information Hiding Terminology.

Proceedings of the First International Workshop on

Information Hiding, 347-350.

Sallee, P., Model-based Methods for Steganography and

Steganalysis, 2005. International Journal of Image

Graphics, 167–190.

Ullerich, Christian., Westfeld, Andreas., 2008. Weakness

of MB2. Digital Watermarking, 127-142.

Vapnik, N, Vladimir., 1995. Nature of Statistical Learning

Theory, Springer.

Vapnik, N, Vladimir., 1998. Statistical Learning Theory,

York: Wiley.

Xuezeng, Pan., Li, Zhuo., Jian, Chen., Jiang, Xiaoning.,

Zeng, Xianting., 2008. A New Blind Steganalysis

Method for JPEG Images. International Conference

on Computer Science and Software Engineering. 939-

942.

SECRYPT 2010 - International Conference on Security and Cryptography

452