Detecting Original Image Using Histogram, DFT and SVM

ACEEE International Journal on Signal and Image Processing Vol 1, No. 1, Jan 2010

Detecting Original Image Using Histogram, DFT
and SVM
T. H. Manjula Devi1, H.S.Manjunatha Reddy,2 K. B. Raja3Venugopal K. R3 and L. M. Patnaik4
1
Department of Telecommunication Engg, Dayananda Sagar College of Engineering, Bangalore
2
Department of Electronics & Communication, Global Academy of Technology
3
Department of Computer Science and Engineering, University Visvesvaraya College of Engineering, Bangalore
4
Vice chancellor, Defence Institute of Advanced Technology, Puna.
thm0303@gmail.com

Abstract specific). In case of universal technique, the embedding
scheme is unknown; so attempt is made to detect the
Information hiding for covert communication is rapidly existence of hidden data. Whereas in embedding specific,
gaining momentum. With sophisticated techniques being the embedding technique is known; therefore payload
developed in steganography, steganalysis needs to be
universal. In this paper we propose Universal Steganalysis
size is estimated. There are Steganalysis Tools available
using Histogram, Discrete Fourier Transform and SVM freely or as commercial software. The Stegdetect Tool,
(SHDFT). The stego image has irregular statistical detects Jsteg, Jphide, invisible secrets, Outguess, F5,
characteristics as compare to cover image. Using Histogram camouflage Steganographic schemes in JPEG images.
and DFT, the statistical features are generated to train One- The Tools developed using the Chi-Square analysis,
Class SVM to discriminate the cover and stego image. performs a statistical attack to detect hidden data in BMP
SHDFT algorithm is found to be efficient and fast since the images.
number of statistical features is less compared to the The success of steganography depends on various
existing algorithm. aspects such as the algorithm used for embedding,
compression algorithm used and alteration of image
properties. LSB is most commonly used technique for
Keywords: Universal Steganalysis, Cover-Medium, Payload,
Redundant Bits, Total Cover Capacity, Histogram, DFT, image steganography, which uses bitwise methods to
SVM. manipulate LSB of the cover image. The minute changes
of LSB are imperceptible to human eye and the method
I. INTRODUCTION of hiding information in LSB can be analogized by
adding noise to the image. Reliable algorithms for
Since long time ago, surreptitious communication compression used in Steganography are Windows Bitmap
and exchange of information is well known. There are (BMP), Graphics Interchange Format (GIF) and Joint
numerous illustrations depicting covertness used in Photographic Experts Group (JPEG), to ensure the
communication. Steganography, water marking, hidden information is not lost after transformation.
cryptography are some of the techniques used for hiding Lossless compression algorithms such as BMP and GIF
information [1]. There are several mediums for hiding formats are chosen for LSB techniques in which
information, such as digital images, text, audio, web modifications are usually made in spatial domain. JPEG
pages, video, etc. Steganography is a skill and discipline is losy compression algorithm where data hiding is
of embedding a secret payload into a cover medium. The usually done in frequency domain.
redundant bits in the cover medium are identified and General Steganalysis method has not been developed
replaced with the payload. The intent of steganography is since every Steganographic method utilizes different
ensuring that the presence of hidden message is methods of embedding the payload. In classical
undetectable. The different techniques used for Steganographic schemes, the security lies in the
embedding of data are Least Significant Bit (LSB), concealment of the encoding technique whereas the
Discrete Cosine Transform (DCT), Discrete Wavelet modern schemes adopt Kerchoff’s Principle of
Transform (DWT), Spread Spectrum (SS) and Palette Cryptography, and hence the security depends on the
technique. There are software tools available over the secret key that is used to encode the payload.
Internet for embedding the payload in digital images and There are several Classifiers being used for pattern
videos viz., S-Tools, J-Steg, Outguess, F5 and Steghide. recognition: Bayesian Multi-Variate, Fischer Linear
Discriminant (FLD), Neural Network (NN), and Support
The steganography is being used by criminals and Vector Machines (SVM). SVM is based on statistical
terrorists to exchange information regarding their illegal learning theory which is highly dependent upon the
activities. The government and other standard kernel functions viz., Gaussian Radial Basis Function,
organizations are using steganalysis to restrain these Polynomial, Exponential Radial Basis Function, Multi-
illegal activities. Steganalysis is to discover and Layer Perceptron, Splines, BSplines, Additive Kernels
recognize the extent of a hidden message and can be and Tensor Product are used for mapping features. SVM
either blind (universal) or non-blind (embedding can be classified as One-Class and Multi-Class. In One-

17
© 2010 ACEEE
DOI: 01.ijsip.01.01.04


Class SVM, only the features of cover images are for images. The statistical regularities exhibited by the
required for classification of cover and stego image natural images are captured by First and Higher Order
whereas in Multi-Class SVM, the features of both the Statistical Model. The FLD is used for classification of
cover and stego-images are required for discriminating the cover and stego image. This model explores its
the cover and stego image. Owing to simple geometry, usefulness in digital forensics including steganography;
SVMs are used more predominantly than Neural distinguish between natural photographs and computer
Networks. Biggest limitation of SVM is choosing of the generated images.
kernel function, speed and size, both in training and Siwei Lyu and Hany Farid [9] have proposed
testing. universal Steganalysis using Higher Order Statistics
Contribution: In this paper, we have proposed (SHS). Wavelet Decomposition based on QMF is used to
SHDFT Universal Steganalysis Method. Histogram and obtain the statistics of an image. The magnitude statistics
Discrete Fourier Transform of color images are computed is obtained from Linear Predictor and phase statistics is
to obtain the PDF moments. In addition, the energy of the obtained from Gaussian Pyramid and Local Angular
image is considered. The non-linear Classifier Support Harmonic Decomposition. One Class and Multi Class
Vector Machine is used for classification of cover and SVM is used for classification. The disadvantage of the
stego image. algorithm is the number of features required to train the
SVM are high. Hence, the time required to detect stego
II. RELATED WORK image is more.

Anderson and Petitcolas [2] have proposed two III. MODEL
mathematical frame works for Steganography:
informatics and theoretical models. The quality of the The steganalysis model is discussed in this section.
retrieved image is poor in this algorithm. Petitcolas et al.,
Steganalysis Model
[3] have discussed a survey on information hiding. The
applications and the terminologies used in information
hiding are elucidated. Certain limitations of information Figure 1 gives the block diagram of SHDFT to
hiding have been portrayed, the JPEG compression, low discriminate cover and stego images using SVM.
pass filtering, cropping, scaling, Additive Gaussian noise, Color image: JPEG and BMP color images are used
rotation, cause distortion in the image. An important for the analysis. It consists of RGB color planes, c ∈ {r,
understanding of these limitations is helpful in g, b}. The color planes are separated and 1D-Histogram
steganalysis. is applied to each of the separated color planes of the
image.
Fridich et al., [4] have presented a steganalytic Histogram: Histograms are the basis for numerous
method to identify stego image generated by F5 spatial domain processing techniques used to provide
steganographic algorithm in JPEG images reliably. The useful image statistics. Image processing operations such
method involves estimation of the cover image histogram as steganography result in changes to the image
from the stego image. The statistical variation of cover histogram. The histogram is applied on each color plane
image is determined using the Least Square Fit, which of the image. The first and second PDF moments i.e., the
compares estimated histogram of DCT coefficients of mean and the variance of the histogram coefficients of
stego-image. This technique can be extended to other each colour plane are calculated which yields 6D
steganographic algorithms that manipulate quantized statistical features.
DCT coefficients. Siwei Lyu and Hany Farid [5] have Discrete Fourier Transform (DFT): The sequence of
described an approach of multi-scale Wavelet N spatial complex coefficients x0, x1,…….., xN-1 is
Decomposition to build Higher Order Statistical Model to transformed into the sequence of N frequency complex
detect hidden data in gray images. Support Vector coefficients X(0), X(1),…….., X(N-1) by the DFT
Machine is used to detect the statistical variations in a according to the formula as given in Equation
test image.
N −1
− i 2πnk
Farid [6] has described a steganalytic method to detect
hidden messages using Wavelet Decomposition by
X(k)= ∑ x(n) exp(
n =0 N
)
building Higher Order Statistical Model of the image. Where n=0, 1… N-1 and k=0, 1… N-1.
The model extracts basic coefficient statistics and error DFT for Histogram of each color channel, c ∈ {r, g, b}
statistics from an optimal magnitude predictor. FLD is is applied. The PDF moments i.e., mean, variance,
used to distinguish the cover and stego image. Siwei Lyu skewness and kurtosis of the DFT coefficients are
and Hany Farid [7] have described a Universal computed which yields 12D statistics. The total energy
Steganalysis Algorithm that exploits strong statistical for DFT coefficients is computed which yields 3D
regularities that exist between the color channels of Statistics.
natural images. The statistical model is generated using SVM: In One-Class SVM, only the features of cover
first and Higher Order Statistics of the image. One Class images are required for classification of cover and stego
SVM is used to detect hidden data. image. The feature vector comprising 24D statistics is
Hany Farid and Siwei Lyu [8] have employed Multi- used to train the SVM.
Scale Wavelet Decomposition to build a statistical model
18
© 2010 ACEEE
DOI: 01.ijsip.01.01.04


generate 100 stego images Using Jphide software for
JPEG images and S-Tools 4.0 software for BMP images.
Color image The payload is of the size 6.0, 4.7, 1.2, 0.3 kilobytes,
corresponding to 100%, 78%, 20% and 5% Of total
cover capacity The capacity will vary according to the
size of the cover image. The steganographic capacity is
1D Histogram Feature Vector then the ratio of the size of the embedded payload to the
total cover capacity. Each stego image is generated with
the same quality factor as that of the original cover image
so as to minimize the statistical variations. Figure 3 and
Table 2 gives the comparison of the detection rate for the
DFT One-Class SVM proposed algorithm SHDFT and existing algorithm SHS
for different embedding rate in BMP images using S-
Tools. It is observed that the percentage of detection rate
Figure 1: Block Diagram of SHDFT increases as embedding rate increases for both the
algorithms. Figure 4 and Table 3 gives the comparison of
Features:We find the mean of the difference between the detection rate for the proposed algorithm SHDFT and
DFT and Histogram for each color plane which yields 3D existing algorithm SHS for different embedding rate in
statistics. SHDFT totally yields 24D characteristics which JPEG images using JPhide and S-Tools. The proposed
are used to discriminate the cover and stego image. algorithm SHDFT requires only 24D statistics for
training SVM as compared to 432D statistics required for
IV. ALGORITHM existing SHS algorithm for the detection of payload[9].
TABLE 1:

ALGORITHM FOR SHDFT

Algorithm for SHDFT
• Input: The test image
• Output: SVM classifies the test image as
cover or stego image.
1. Separation of the color planes, c ∈ {r, g, b} of (a) (b)
the color image.
2. Build Histogram (1D) for each color planes
c ∈ {r, g, b}. Compute the 1st and 2nd moments
i.e., mean and variance of the histogram
coefficients. This yields 6D statistics.
3. Build Discrete Fourier Transform for histogram
of each color planes, c ∈ {r, g, b}. Compute the (c) (d)
1st, 2nd, 3rd and 4th moments i.e., mean variance,
skewness and kurtosis of the DFT coefficients.
This yields 12D statistics. Also compute the
total energy for the DFT coefficients. This
yields 3D more statistics.
4. The 1st order moment Mean of the difference of
histogram and DFT is computed for each color (e) (f)
channel. This yields 3D statistics. Figure 2: The payload images (a)( b), BMP cover
5. The feature vectors obtained from steps 2, 3 and image (c)( d) and JPEG cover image (e)(f).
4 yields 24D features.
6. The feature vectors obtained in steps 5 are used
to train SVM to classify test image.

V. PERFORMANCE ANALYSIS
For the performance analysis we consider 100 color
images of JPEG and BMP formats with 600*400 pixels
in size (86 kilobytes). Figure 2 (a)(b) shows the payload.
Figure2 (c)(d) and (e)( f) gives the cover images of BMP
and JPEG formats respectively. The payload of various
sizes is embedded into the full-resolution cover images to
19
© 2010 ACEEE
DOI: 01.ijsip.01.01.04


VI. CONCLUSION
TABLE 2

EMBEDDING RATE FOR BMP IMAGES USING S-TOOLS In the wake of a number of steganographic techniques
being developed, steganalysis needs to have a universal
approach. We present a steganalysis technique that relies
Embedding Detection Rate (%) on building a Statistical Model using the Histogram and
Rate (%) SHDFT SHS Discrete Fourier Transform. SVM is used for
100 96.42 85.71 classification of stego and cover images. This Model has
78 92.67 78.57 fairly higher detection rates and comparatively
20 89.87 71.42 computation time is less, as the numbers of features
5 86.78 64.28 required to train the SVM is much lesser than other
techniques. In future, the algorithm may be tested for
embedding rate less than 5%.
S-Tools SHDFT
SHS
REFERENCES
100
90
Detection Rate (%)

80
70 [1] D. Kahn, “The History of Steganography,” Lecture Notes
60 in Computer Science, Springer-Verlag, vol. 1174, 1996.
50 [2] R. Anderson and F. Petitcolas, “On the Limits of
40 Steganography,” IEEE Journal on Selected Areas in
30 Communications, vol. 16, no. 4, pp. 474–481, 1998.
20 [3] E. Petitcolas, R. Anderson and M. Kuhn, “Information
10 Hiding – a Survey,” Proceedings of the IEEEs, vol. 87,
0 no. 7, pp. 1062–1078, 1999.
100 78 20 5
[4] J. Fridrich, M. Goljan, and D. Hogea, “Steganalysis of
Embedding Rate (% ) JPEG images: Breaking the F5 Algorithm,” International
Workshop on Information Hiding, Lecture Notes in
Computer Science, Springer-Verlag, vol. 2578, pp. 310-
Figure 3: Comparison of Detection rate for various payload sizes for 323, 2003.
SHDFT and SHS algorithm for BMP images using OC-SVM. [5] S. Lyu and H. Farid, “Detecting Hidden Messages using
Higher Order Statistics and Support Vector Machines,”,
TABLE 3 International Workshop on Information Hiding, Lecture
EMBEDDING RATE FOR JPEG IMAGES USING JPHIDE Notes in Computer Science, Springer-Verlag, vol. 2578,
pp. 340-354,
[6] S. Lyu and H. Farid, “Steganalysis using Color Wavelet
Embedding Detection Rate (%)
Statistics and One-Class Support Vector Machines,” in
Rate (%) SHDFT SHS SPIE Symposium on Electronic Imaging, vol. 5306, pp.
100 98.67 89.91 35-46, 2004.
78 95.73 80.57 [7] E. P. Simoncelli, “Statistical Modelling of Photographic
Images,” in Handbook of Image and Video Processing, A.
20 91.32 76.42 Bovik, Ed. Academic Press, chapter 4.7, Second Edition,
5 86.78 68.28 2005.
[8] Hany Farid and Siwei Lyu, “Higher-order Wavelet
Statistics and their Application to Digital Forensics,”
SHDFT IEEE Workshop on Statistical Analysis in Computer
JPhide SHS Vision, 2003.
[9] Siwei Lyu and Hany Farid, “Steganalysis using Higher-
100
Order Image Statistics,” IEEE Transaction on
90
Information Forensics and Security, vol. 1, pp. 111-119,
80
2006.
Detection Rate (%)

70
60
50
Mrs. Manjula Devi T.H is an
40
Assistant Professor in the
30
Department of
20
10
Telecommunication Engg,
0
Dayananda Sagar College of
100 78 20 5 Engg, Bangalore. She
Embedding Rate (% ) obtained her B.E. degree in
Electronics Engg from
Figure 4: Comparison of Detection rate for various payload sizes for
Bangalore University. Her
SHDFT and SHS algorithm for JPEG images using OC-SVM specialization in Master
degree was Electronics & Communication from

20
© 2010 ACEEE
DOI: 01.ijsip.01.01.04


Bangalore University and currently she is pursuing Economics from Bangalore University and Ph.D. in
Ph.D. in the area of Steganography and Steganalysis Computer Science from Indian Institute of
under the guidance of Dr. K. B. Raja, Asst. Professor, Technology, Madras. He has a distinguished academic
Dept of Electronics and Communication Engg, career and has degrees in Electronics, Economics,
University Visvesvaraya college of Engg, Bangalore. Law, Business Finance, Public Relations,
Her area of interest is in the field of Signal Processing, Communications, Industrial Relations, Computer
Communication and Control Systems. She is the life Science and Journalism. He has authored 27 books on
member of International Microelectronics And Computer Science and Economics, which include
Packaging Society (IMAPS) India Chapter, and Indian Petrodollar and the World Economy, C Aptitude,
society for Technical Education. Mastering C, Microprocessor Programming, Mastering
C++ etc. He has been serving as the Professor and
Mr. H.S.Manjunatha Reddy is a Chairman, Department of Computer Science and
Professor in the department of Engineering, University Visvesvaraya College of
Electronics and Communication Engineering, Bangalore University, Bangalore. During
Engineering, Global Academy his three decades of service at UVCE he has over 200
of Technology, Bangalore. He research papers to his credit. His research interests
obtained his B.E. Degree in include computer networks, parallel and distributed
Electronics from Bangalore systems, digital signal processing and data mining.
University, Bangalore. His
specialization in Master degree
was Digital Electronics from L M Patnaik is a Vice Chancellor, Defence Institute of
Visvesvaraya Technological Advanced Technology (Deemed
University, Belgaum. He is pursuing research in the University), Pune, India. During
area of Steganography and Steganalysis for secured the past 35 years of his service
communication. His area of interest is in the field of at the Indian Institute of
Digital Image Processing, Communication Networks Science, Bangalore. He has over
and Biometrics. He is life member of Indian Society 500 research publications in
for Technical Education, New Delhi. refereed International Journals
and Conference Proceedings.
K B Raja is an Assistant He is a Fellow of all the four
Professor, Dept. of Electronics leading Science and
and Communication Engg, Engineering Academies in
University Visvesvaraya college India; Fellow of the IEEE and the Academy of Science
of Engg, Bangalore University, for the Developing World. He has received twenty
Bangalore. He obtained his national and international awards; notable among them
Bachelor of Engineering and is the IEEE Technical Achievement Award for his
Master of Engineering in significant contributions to high performance
Electronics and Communication computing and soft computing. His areas of research
Engineering from University interest have been parallel and distributed computing,
Visvesvaraya College of mobile computing, CAD for VLSI circuits, soft
Engineering, Bangalore. He was awarded Ph.D. in computing, and computational neuroscience.
Computer Science and Engineering from Bangalore
University. He has over 35 research publications in
refereed International Journals and Conference
Proceedings. His research interests include Image
Processing, Biometrics, VLSI Signal Processing,
computer networks.

K R Venugopal is currently the
Principal and Dean, Faculty of
Engineering, University
Visvesvaraya College of
Engineering, Bangalore
University, Bangalore. He
obtained his Bachelor of
Engineering from University
Visvesvaraya College of
Engineering. He received his
Masters degree in Computer
Science and Automation from Indian Institute of
Science Bangalore. He was awarded Ph.D. in

21
© 2010 ACEEE
DOI: 01.ijsip.01.01.04

Detecting Original Image Using Histogram, DFT and SVM

More Related Content

What's hot (19)

Similar to Detecting Original Image Using Histogram, DFT and SVM (20)

More from IDES Editor (20)

Recently uploaded (20)

Detecting Original Image Using Histogram, DFT and SVM