112 (IJCNS) International Journal of Computer and Network Security,

Vol. 2, No. 10, 2010

A Robustness Steganographic Method Against

Noise For RGB Images Based On PCA
Hamid Dehghani1, Majid Gohari2
1
Malekashtar University of Technology, Department Computer Scienses,
Lavizan Road, Tehran, Iran
[email protected]
2
Malekashtar University of Technology, Department of Secret Communication,
Lavizan Road, Tehran, Iran
[email protected]

are then embedded into the web page through the upper and
Abstract: In this paper an effective color image steganography lower cases of letters in HTML tags. When a watermarked
method based on Principle Component Analysis (PCA) is web page is tampered, the extracted watermarks can detect
proposed. After applying the PCA technique on the RGB the modifications to the web page. A. Abadpour and S.
trichromatic system, eigenimages obtained, the least significant Kasaei [7] use the PCA to compression and watermarking of
bit of the Eigenimages pixels value are replaced with the data color images. However steganography based on PCA
bits. Then inverse PCA applied and stego-image is resulted. The technique is typically a novel and first attempt scheme.
method used the correlation between the colors of RGB image.
The rest of this paper is organized as follows: first, Section 2
Experimental results show that the proposed algorithm robust
against noises, attacks and detection comparable to LSB
PCA technique briefly discussed. The proposed PCA-based
methods that replace the LSBs of pixel value directly. Hope with steganography scheme is specifically described in Section 3.
optimizing the method, makes it an effective steganographic The paper continues with Section 4 which contains the
technique in security communication. experimental results and discussions. Finally, Section 5
concludes the paper.
Keywords: Steganography, Principle Component Analysis,
RGB image, correlation. 2. Principle Component Analysis
Let I be an N×N matrix, denoted by I ∈ , where F
1. Introduction represents either the real number domain or the complex
Steganography is combination of science and art concealing number domain. The first step of PCA is to calculate the
the secret message so that the very existence of it is not covariance matrix V of I, which is defined as
detectable [1]. The major steganography characteristics are
V= ,
providing a larger hidden capacity for data-hiding and
maintaining a good perceived quality [2]. In the Where is the row vector in I, t denotes the transpose
steganography process, first and more important term that operation, and ∈ is the average vector of the row
must regarded is undetectability and then robustness vectors in I, i.e.
(resistance to various image processing methods and
compression) and capacity of the hidden data, that separate .
it from related techniques such as watermarking and
Then eigen decomposition (ED) is applied to V:
cryptography. The steganographic algorithms can be divided
into two groups: spatial/time domain and transform domain V = UL ,
techniques. In the spatial domain in the case of images,
Where denotes the inverse matrix of U. L is a diagonal
secret message directly embedding in pixels value of image.
matrix with eigenvalues of V as its diagonal
The transform domain methods operate in the Discrete
elements , and the columns of U, , , . . . ,
Cosine Transform, Fourier or wavelet transform domains of
, are the eigenvectors of V.
the host signal [2]. The proposed method belongs to
According to theory of linear algebra, the primary
transform domain but color transform domain. Applying
information of V is in the larger eigenvalues and
PCA transform caused images that called eigenimages [3].
corresponding eigenvectors. Without loss of generality, we
In natural RGB images, R-, G- and B-component have high
assume that these diagonal elements in L have been sorted
correlation: , and [4],
in descending order ( . A property of
while eigenimages aren’t correlated and distribution of data
PCA is its optimal signal reconstruction in the sense of
is quite heterogeneous. Hence changing pixels value of RGB
minimum mean square error (MSE) when only a subset of
image caused most sensible variation of changing pixels
eigenvectors, called principal vectors, is used as basis
value of eigenimages [5]. Tow characteristics uncorrelation
vectors of a feature space S:
and data distribution of eigenimages can be used for
effective steganographic. Watermarking based on PCA S = span ( , ,..., ), m N.
transform for tamper-proof of web pages is proposed by
With the feature space S, we can obtain another
Qijun and Hongtao [6]. The proposed scheme generates
representation of the original data I by projecting them into
watermarks based on the PCA technique. The watermarks
the space S. Specifically this is as the following equation:
 (IJCNS) International Journal of Computer and Network Security, 113
Vol. 2, No. 10, 2010

= .[ , ,..., ], i= 1, 2, . . . , N, 4. Implementation Results

Where ∈ can be viewed as the coordination of the One of the error metrics used to compare the host image and
original data in the feature space S. Actually they are what stego-image is Pick Signal to Noise Ratio (PSNR);
we call ‘principal components’, which, to some extent, can
PSNR = 10× dB
distinguish the original data well [7]. In the context of the
proposed stenographic method, we generate principal
Where, M and N are the size of the images. c(i,j) and s(i,j)
components (eigenimages) as m=N=3, which we will show
are sequentially pixel of cover and stego images
in the next part. Further details on PCA can be found in
[11].
Refs. [8, 9]. However, we will not discuss it more in this
paper due to limits of space.
4.1. Image quality
To quantify the invisibility of the embedded secret message,
3. Data embedment and extraction three 512×512 images, namely, Lena, Baboon and pepper,
In order to apply PCA, we have first to construct a matrix, I, shown in Fig. 2, were used as host images. Size of secret
from the given RGB image of size M×N. the elements of R-, message is 35000×3 bits (35000 bits embedded in each of
G- and B-component taken into three rows of the matrix: eigenimages) coded via convolutional coding of rate 4/8.
Table 1 show average PSNR of R-,G- and B-component and
I= . percent of correct retrieval data with different embedding
rate and different LSB locations (bpp: bits per pixel; LSB
PCA is then applied to the matrix I, and we choose all the loc.: LSBs location under embedding)
m=N=3 principal components as eigenimages (PCA1, for the host images. Usually error occurs in lower lsbs. so
PCA2, PCA3). Secret message bits are embedded in the with selecting upper LSB location(s) for imbedding message
pixel value of these images with common LSB stenographic bits, error decrease. But increasing the LSB location caused
methods. In this way rounding of elements of PCA1, PCA2 the PSNR be decreased. We do experiments for 1, 2, 3, 4, 6
and PCA3 images in the transmitter and receiver is needed. and 7 bit(s) per RGB pixel (that been used). Selecting of
The act of rounding caused error. For conquest the error, LSB location is based on the maximum data retrieval and
convolutional coding (Error-Correcting Code) [10] of then for the similar types, maximum PSNR is selected.
message bits applied. After applying invers PCA, stego Since transform domain of proposed method and frequency
image is obtained. Flowchart of data embedding process transform domain methods and also supported format (RGB
shown in Fig. 1(a). format for the color transform domain; Greyscale and JPEG
format for FT, DCT and Wavelet domain) aren’t similar, we
while not compare the proposed method with these methods.

4.2. Embedding capacity

Since, usually, the proposed method is with loss, using
Error-Correcting Codes is needed. Therefore embedding
capacity is depended on image properties and coding rate.
Maximum capacity and PSNR can be optimized by using
adaptive process depended on coding rate and LSB
locations. It is clear that total bits number of secret
message, , after embedding can be calculated by below
equation:

Where is bpp of PCAi and show the used

coding rate of embedding of PCAi in M-by-N RGB image.

The results of Table 1 shown that the proposed algorithm
Figure 1. The embedding and extraction schemas of the has acceptable capacity but (because of coding) lower than
proposed method. (a) Embedding schema, and (b) LSB method.
Extraction schema.
4.3. Robustness against noise
In side of receiver, the extraction process can be carried out In this evaluation, as an active attack, we generate White
by reversing the embedding procedure as the block diagram Gaussian Noise by using wgn(m,n,p) function of Matlab
of Fig. 1(b). From this figure, it is obvious that this PCA is software and added to image components. wgn(m,n,p)
blind. The receiver first compute the eigenimages of stego generates an m-by-n matrix of white Gaussian noise. The p
image and then apply the LSB extraction. After decoding parameter specifies the power of the matrix in bits.
the retrieval data, secret message is obtained. Parameter-p selected to equal the numbers of bits embedded
in a pixel (bpp). Secret message is coded by convolutional
coding with rate 4/8. LSB and proposed methods are
simulated and results gathered for the test images in Table
114 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 10, 2010

2. It can be seen from this Table that the PCA-based

method is highly Robustness against white Gaussian noise
than LSB method.
When we hide secret message bits sequentially in PCAi
pixels lsb caused the change in image domain that dispread
in all of components. So the effect of noise reduced.
Figure 2. Test images

Table 1: The average PSNR and percent of correct retrieval data for different bpp and optimum locations.

bpp – LSB loc. percent of correct retrieval data

PCA1 PCA2 PCA3 PCA1 PCA2 PCA3 PSNR
not used 1–3 not used - 100 - 49.62
1–2 1–2 not used 100 100 - 46.25
Lena
1- 3 1–2 1–2 100 100 99.99 41.60
1–3 2 - 2:3 1–2 100 100 100 40.44
1–2 2 - 2:3 3 - 1:3 100 99.99 99.41 41.81
1–2 3 – 1:3 3 – 1:3 100 99.62 99.35 42.47
not used not used 1–2 - - 100 50.48
1–2 1–2 not used 100 100 - 46.51
Baboon 1–2 2 - 3:4 not used 100 100 - 41.84
1–3 2 - 2:3 1–2 100 100 100 40.59
2 - 3:4 2 - 3:4 2 - 2:3 100 100 100 36.08
2 - 3:4 2 - 3:4 3 - 2:4 99.99 100 99.69 36.31
not used 1–3 not used - 100 - 47.95
not used 2 - 2:3 not used - 100 - 49.56
Pepper not used 2 - 2:3 1–1 - 100 97.71 46.88
1–1 2 - 1:2 1–1 96.76 99.64 98.01 48.26
2 - 2:3 2 - 1:2 2 - 1:2 96.25 97.05 95.78 43.64
2 - 2:3 3 - 1:3 2 - 1:2 96.01 98.63 95.46 42.43
6 6 3 4
Pepper 99.3 98.9 87.1 89.3
Table 2: Percent of correct retrieval data in presence of 2 1 2 3
white Gaussian noise with variance equal to bpp, for LSB
and PCA-based methods (secret message size is 35000×3
bits and coding rate is 4/8) 5. Conclusion
1 2 3 4
bpp bpp bpp bpp In this paper a novel transform domain steganography
Lena 50.2 50.2 49.9 50.0 method based on Principle Component Analysis for secure
7 4 6 3 and robustness steganography was proposed. The main idea
LSB Baboon 50.0 49.6 50.3 49.4 is based on embedding a secret data in the pixels value of
2 0 3 9 obtained images from colour transform domain. The
Pepper 49.7 50.0 50.3 50.1 proposed blind steganography method has the acceptable
8 2 9 1 capacity and reduces the stego image visual distortion by
Lena 98.2 97.5 94.5 80.3 hiding the secret data in eigenimages. Robustness against
PCA- 3 0 0 7 White Gaussian Noise is the major characteristic of the
Based Baboon 99.9 75.7 69.0 78.4 proposed method.
 (IJCNS) International Journal of Computer and Network Security, 115
Vol. 2, No. 10, 2010

Reference

[1] S .Katzenbeisser, F .Petitcolas, Information Hiding

Techniques for Steganography and Digital
Watermarking, Artech HousePublishers, 2000.
[2] A. Cheddad, Digital image steganography: Survey and
analysis of current methods, IEEE Trans. Signal
Processing 90 (2010) 727–752.
[3] Arash Abadpour & Shohreh Kasaei, Color PCA
eigenimages and their application to compression and
watermarking, Image and Vision Computing 26 (2008)
878–890.
[4] Paluse H., Reprensentation of color images in different
color spaces, in the color image processing handbook,
S.J .Sangwine and R.E.N, Eds., 67-90 .Chapman and
hall, London, 1998.
[5] A. Abadpour, Color image processing using principal
component analysis, Master’s thesis, Sharif University
of Technology, Mathematics Science Department,
Tehran, Iran, 2005.
[6] Q. Zhao, H. Lu, PCA-based web pagewatermarking,
Pattern Recognition 40 (2007) 1334 – 1341.
[7] A. Abadpour, S. Kasaei, Color PCA eigenimages and
their application to compression and watermarking,
Image and Vision Computing 26 (2008) 878–890.
[8] J.E. Jackson, A User’s Guide to Principal Components,
Wiley, New York, 1991.
[9] I.T. Jolliffe, Principle Component Analysis, Springer,
New York, 1986.
[10] M. Bossert and F. Hergert, “Hard- and soft-decision
decoding beyond the half minimum distance An
algorithm for linear codes,” IEEE Trans. Inform.
Theory, vol. IT-32, pp. 709–714, Sept. 1986.
[11] Z. Wang, A. Bovik, A universal image quality index,
IEEE Signal Process. Lett. 9 (March 2002) 81–84.

Majid Gohari received the B.S. degrees in

Electrical Engineering from Ferdowsi University
of Mashhad in 2008. He is now Graduate student
of Secret Communication in Malekashtar
University of Technology