Pattern Recognition 37 (2004) 469 474 www.elsevier.

com/locate/patcog

Hiding data in images by simple LSB substitution

Abstract In this paper, a data hiding scheme by simple LSB substitution is proposed. By applying an optimal pixel adjustment process to the stego-image obtained by the simple LSB substitution method, the image quality of the stego-image can be greatly improved with low extra computational complexity. The worst case mean-square-error between the stego-image and the cover-image is derived. Experimental results show that the stego-image is visually indistinguishable from the original cover-image. The obtained results also show a signicant improvement with respect to a previous work. ? 2003 Pattern Recognition Society. Published by Elsevier Ltd. All rights reserved.
Keywords: Data hiding; LSB substitution

1. Introduction Data hiding is a method of hiding secret messages into a cover-media such that an unintended observer will not be aware of the existence of the hidden messages. In this paper, 8-bit grayscale images are selected as the covermedia. These images are called cover-images. Cover-images with the secret messages embedded in them are called stego-images. For data hiding methods, the image quality refers to the quality of the stego-images. In the literature, many techniques about data hiding have been proposed [15]. One of the common techniques is based on manipulating the least-signicant-bit (LSB) planes by directly replacing the LSBs of the cover-image with the message bits. LSB methods typically achieve high capacity. Wang et al. [6] proposed to embed secret messages in the moderately signicant bit of the cover-image. A genetic algorithm is developed to nd an optimal substitution matrix for the embedding of the secret messages. They also proposed to use a local pixel adjustment process (LPAP) to improve the image quality of the stego-image. Unfortunately,

Corresponding author. E-mail address: [email protected] (C.-K. Chan).

since the local pixel adjustment process only considers the last three least signicant bits and the fourth bit but not on all bits, the local pixel adjustment process is obviously not optimal. The weakness of the local pixel adjustment process is pointed out in Ref. [7]. As the local pixel adjustment process modies the LSBs, the technique cannot be applied to data hiding schemes based on simple LSB substitution. Recently, Wang et al. [8] further proposed a data hiding scheme by optimal LSB substitution and genetic algorithm. Using the proposed algorithm, the worst mean-square-error (WMSE) between the cover-image and the stego-image is shown to be 1 of that obtained by the simple LSB substitu2 tion method. In this paper, a data hiding scheme by simple LSB substitution with an optimal pixel adjustment process (OPAP) is proposed. The basic concept of the OPAP is based on the technique proposed in Ref. [7]. The operations of the OPAP is generalized. The WMSE between the cover-image and the stego-image is derived. It is shown that the WMSE obtained by the OPAP could be less than 1 of that obtained by 2 the simple LSB substitution method. Experimental results demonstrate that enhanced image quality can be obtained with low extra computational complexity. The results obtained also show better performance than the optimal substitution method described in Ref. [8].

0031-3203/$30.00 ? 2003 Pattern Recognition Society. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.patcog.2003.08.007

470

C.-K. Chan, L.M. Cheng / Pattern Recognition 37 (2004) 469 474 Table 1 Worst PSNR for k = 15 by simple LSB substitution k PSNR 1 48.13 2 38.59 3 31.23 4 24.61 5 18.30

The rest of the paper is organized as follows. Section 2 brie y describes the simple LSB substitution. In Section 3, the optimal pixel adjustment process is described and the performance is analyzed. Experimental results are given in Section 4. Finally, Section 5 concludes this paper. 2. Data hiding by simple LSB substitution In this section, the general operations of data hiding by simple LSB substitution method is described. Let C be the original 8-bit grayscale cover-image of Mc Nc pixels represented as C = {xij |0 6 i Mc ; 0 6 j Nc ; xij {0; 1; : : : ; 255}}: M be the n-bit secret message represented as M = {mi |0 6 i n; mi {0; 1}}: (2) (1)

PSNR of the obtained stego-image can be computed by PSNR worst = 10 log10 = 10 log10 2552 WMSE 2552 dB: (2k 1)2 (6)

Table 1 tabulates the worst PSNR for some k = 15. It could be seen that the image quality of the stego-image is degraded drastically when k 4.

Suppose that the n-bit secret message M is to be embedded into the k-rightmost LSBs of the cover-image C. Firstly, the secret message M is rearranged to form a conceptually k-bit virtual image M represented as M = {mi |0 6 i n ; mi {0; 1; : : : ; 2k 1}}; (3)

3. Optimal pixel adjustment process In this section, an optimal pixel adjustment process (OPAP) is proposed to enhance the image quality of the stego-image obtained by the simple LSB substitution method. The basic concept of the OPAP is based on the technique proposed in Ref. [7]. Let pi , pi and pi be the corresponding pixel values of the ith pixel in the cover-image C, the stego-image C obtained by the simple LSB substitution method and the rened stego-image obtained after the OPAP. Let i = pi pi be the embedding error between pi and pi . According to the embedding process of the simple LSB substitution method described in Section 2, pi is obtained by the direct replacement of the k least signicant bits of pi with k message bits, therefore, 2k
i

where n Mc Nc . The mapping between the n-bit secret message M = {mi } and the embedded message M = {mi } can be dened as follows:
k1

mi =
j=0

mik+j 2k1j :

Secondly, a subset of n pixels {xl1 ; xl2 ; : : : ; xln } is chosen from the cover-image C in a predened sequence. The embedding process is completed by replacing the k LSBs of xli by mi . Mathematically, the pixel value xli of the chosen pixel for storing the k-bit message mi is modied to form the stego-pixel xli as follows: xli = xli xli mod 2k + mi : (4)

2k :
i

(7)

The value of such that Interval 1 : Interval 2 : Interval 3 :

can be further segmented into three intervals, 2k1 2k ;

In the extraction process, given the stego-image S, the embedded messages can be readily extracted without referring to the original cover-image. Using the same sequence as in the embedding process, the set of pixels {xl1 ; xl2 ; : : : ; xln } storing the secret message bits are selected from the stego-image. The k LSBs of the selected pixels are extracted and lined up to reconstruct the secret message bits. Mathematically, the embedded message bits mi can be recovered by mi = xli mod 2k : (5)

2k1 6 2k
i

6 2k1 ; (8)

2k1 :

Suppose that all the pixels in the cover-image are used for the embedding of secret message by the simple LSB substitution method. Theoretically, in the worst case, the

Based on the three intervals, the OPAP, which modies pi to form the stego-pixel pi , can be described as follows: Case 1 (2k1 i 2k ): If pi 2k , then pi = pi 2k ; otherwise pi = pi ; Case 2 (2k1 6 i 6 2k1 ): pi = pi ; Case 3 (2k i 2k1 ): If pi 256 2k , then pi = pi + 2k ; otherwise pi = pi .

C.-K. Chan, L.M. Cheng / Pattern Recognition 37 (2004) 469 474

471

Let i = pi pi be the embedding error between pi and pi . i can be computed as follows: Case 1 (2k1 i 2k and pi 2k )
i

Combining Eqs. (6) and (10), we have WMSE = (2k1 )2 WMSE (2k 1)2 when k = 1; WMSE 4 WMSE when k = 2; 9
16 49 64 225

= p i p i = p i 2k p i = 2k1 2k 2k1
i i

2k =

2 k 2k

0:
i

WMSE WMSE

when k = 3; when k = 4:

(11)

Case 2 (2k1
i

2k and pi 2k )
i

= p i p i = pi p i = 2k1
i

2k :
i

Case 3 (2k1 6
i

6 2k1 )
i

= p i p i = pi p i = 2
k1

62
i

k1

Case 4 (2k
i

2k1 and pi 256 2k )

= p i p i = pi + 2 k p i = 2k + 2k 0
i i

+ 2k

2k1 + 2k

Eq. (11) reveals that WMSE 1 WMSE, for k 2; and 2 WMSE 1 WMSE when k = 4. This result also shows 4 that the WMSE obtained by the OPAP is better than that obtained by the optimal substitution method proposed in Ref. [8] in which WMSE = 1 WMSE. 2 Moreover, the optimal pixel adjustment process only requires a checking of the embedding error between the original cover-image and the stego-image obtained by the simple LSB substitution method to form the nal stego-image. The extra computational cost is very small compared with Wangs method [8] which requires huge computation for the genetic algorithm to nd an optimal substitution matrix. 4. Experimental results

2k1 :
k i

Case 5 (2
i

k1

and pi 256 2 )

= p i p i = pi p i = 2k
k1 : i 2

From the above ve cases, it can be seen that the absolute value of i may fall into the range 2k1 | i | 2k only when pi 2k (Case 2) and pi 256 2k (Case 5); while for other possible values of pi , i falls into the range 0 6 | i | 6 2k1 . Because pi is obtained by the direct replacement of the k LSBs of pi with the message bits, pi 2k and pi 256 2k are equivalent to pi 2k and pi 256 2k , respectively. In general, for grayscale natural images, when k 6 4, the number of pixels with pixel values smaller than 2k or greater than 256 2k is insignificant. As a result, it could be estimated that the absolute embedding error between pixels in the cover-image and in the stego-image obtained after the proposed OPAP is limited to 0 6 | i | 6 2k1 : (9)

This section presents experimental results obtained for two cover-image sets. The rst set of cover-images consists of four standard grayscale images, Lena, Baboon, Jet and Scene, each of 512 512 pixels, as depicted in Fig. 1. The second set consists of 1000 randomly

Let WMSE and WMSE be the worst case mean-squareerror between the stego-image and the cover-image obtained by the simple LSB substitution method and the proposed method with OPAP, respectively. According to Eq. (9) WMSE can be derived by WMSE = 1 M c Nc
Mc Nc 1 i=0

(2k1 )2 = (2k1 )2 :

(10)

Fig. 1. The rst set cover-images of size 512 512 pixels.

472

C.-K. Chan, L.M. Cheng / Pattern Recognition 37 (2004) 469 474

Fig. 2. Test image used as the second set of secret message.

generated grayscale images. There are two set of secret messages. The rst set of secret message consists of 1000 randomly generated message of 512 512 k bits, where k refers to the number of LSBs in the cover image pixels that are used to hold the secret data bits. For example, suppose that the last two LSBs of the cover image pixels are used to hold the secret data, then the secret data is of size 512 512 2 = 524 288 bits. The second set consists of the reduced-sized images of the grayscale image Ti as shown in Fig. 2. The reduced-sized images are of size 512 256 pixels (for 4-bit insertion), 384 256 pixels (for 3-bit insertion), 256 256 pixels (for 2-bit insertion) and 256 128 pixels (for 1-bit insertion), respectively. The results of embedding the rst set of secret messages into the rst set of cover-images are listed in Table 2. Referring to Table 2, the column labeled OPAP is our proposed

method with the optimal pixel adjustment process; the column labeled LSB is the simple LSB substitution method; and the column labeled OLSB in the optimal LSB substitution method proposed in Ref. [8]. For the OPAP and LSB methods, the obtained PSNR values are the average values of embedding the 1000 sets random messages into the cover-images. For the OLSB method, for k = 1; 2, the obtained PSNR values are the average values of embedding the 1000 sets random messages into the cover-images, for k = 3, the obtained PSNR values are the average values of embedding the 10 out of 1000 sets random messages into the cover-images while for k = 4, no experiments are conducted due to the large number of searching space for the optimal substitution matrix. The results reveal that our proposed method has much better performance than the LSB and OLSB methods for k = 2 4. The results of embedding the reduced-sized image of Fig. 2 into the rst set of cover-images are listed in Table 3. The results also reveal that our proposed method has much better performance than the LSB and OLSB methods for k = 2 4. Table 4 also shows the percentage of cover image pixels associated with the ve cases: Case 1 Case 2 Case 3 Case 4 Case 5 (2k1 (2k1 (2k (2k
i i

2k and pi 2k ); 2k and pi 2k );
i

(2k1 6
i i

6 2k1 );

2k1 and pi 256 2k ); 2k1 and pi 256 2k ): (12)

Table 2 The results of embedding the random messages into the rst set of cover-images Cover image Lena k 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 OPAP 51.1410 46.3699 40.7271 34.8062 51.1414 46.3691 40.7253 34.8021 51.1405 46.3700 40.7273 34.8065 51.1410 46.3702 40.7270 34.806 LSB 51.1410 44.1519 37.9234 31.7808 51.1414 44.1492 37.9226 31.8588 51.1405 44.1149 37.9557 31.8487 51.1410 44.1497 37.8914 31.8467 OLSB 51.1483 44.1651 37.9467 51.1477 44.1619 37.9480 51.1478 44.1276 37.9978 51.1480 44.1628 37.9849

Baboon

Jet

Scene

C.-K. Chan, L.M. Cheng / Pattern Recognition 37 (2004) 469 474 Table 3 The results of embedding the reduced-sized image of Fig. 2 into the rst set of cover-images Cover image Lena k 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 OPAP 51.1299 46.3707 40.7266 34.8434 51.1415 46.3761 40.7254 34.7853 51.1458 46.3692 40.7241 34.8283 51.1402 46.3674 40.7420 34.7974 LSB 51.1299 44.0216 37.8626 31.2818 51.1415 44.0205 37.8642 31.3307 51.1458 43.9901 37.8898 31.4083 51.1402 44.0176 37.8522 31.3208

473

OLSB 51.1524 44.7638 38.7242 51.1415 44.7440 38.7295 51.1458 44.7354 38.7667 51.1420 44.7656 38.6909

Baboon

Jet

Scene

Table 4 The percentage of cover image pixels associated with the ve cases (Eq. (12)) when the reduced-sized images of Fig. 2 are embedded into the cover-images Cover image Lena k 2 3 4 2 3 4 2 3 4 2 3 4 Case 1 (%) 9.52 14.15 21.30 9.53 14.03 20.78 9.67 13.91 20.31 9.58 14.17 21.01 Case 2 (%) 0 0 0 0.01 0.02 0.05 0 0 0 0 0.01 0.01 Case 3 (%) 86.55 80.86 73.27 86.51 80.90 73.85 86.32 81.20 74.22 86.53 80.78 73.74 Case 4 (%) 3.93 4.99 5.43 3.95 5.05 5.32 4.01 4.89 5.47 3.89 5.04 5.24 Case 5 0 0 0 0 0 0 0 0 0 0 0 0

Baboon

Jet

Scene

when the reduced-sized images of Fig. 2 are embedded into the cover-images. For illustrative purpose, Fig. 3 shows a pair of stego-images obtained by embedding the reduced-sized image Ti of size 512 256 pixels into the cover-image Lena of size 512 512 pixels using the simple LSB method and the proposed OPAP method. From Fig. 3(a) (stego-image obtained by the simple LSB-substitution method), one can see some false contours appearing on the shoulder of Lena. The unwanted artifacts may arise sus-

picion and defeat the purpose of steganography. However, there is no such artifacts appearing on the stego-image (Fig. 3(b)) obtained by the proposed method. The visual quality of stego-images obtained by the proposed method are much better than that of obtained by the simple LSB-substitution method. To further evaluate the performance of the proposed method, the reduced-sized image of Fig. 2 are embedded into 1000 sets randomly generated cover-images and the obtained average PSNR values are listed in Table 5. The

474

C.-K. Chan, L.M. Cheng / Pattern Recognition 37 (2004) 469 474

(a)

(b)

Fig. 3. Stego-images obtained by (a) Simple LSB-substitution method; (b) proposed method, where the secret-image is of size 512 256 pixels (4-bit insertion). Table 5 The results of embedding the reduced-sized image of Fig. 2 into the second set of cover-images Cover image Random k 1 2 3 4 OPAP 51.1410 46.3215 40.6023 34.4868 LSB 51.1410 44.0217 37.8621 31.337

References
[1] A.Z. Tirkel, R.G. Van Schyndel, C.F. Osborne, A digital watermark, Proceedings of ICIP 1994, Austin Convention Center, Austin, Texas, Vol. II, 1994, pp. 86 90. [2] W. Bender, N. Morimoto, A. Lu, Techniques for data hiding, IBM Syst. J. 35 (3/4) (1996) 313336. [3] T.S. Chen, C.C. Chang, M.S. Hwang, A virtual image cryptosystem based upon vector quantization, IEEE Trans. Image Process. 7 (10) (1998) 14851488. [4] L.M. Marvel, C.G. Boncelet, C.T. Retter, Spread spectrum image steganography, IEEE Trans. Image Process. 8 (8) (1999) 10751083. [5] K.L. Chung, C.H. Shen, L.C. Chang, A novel SVD- and VQ-based image hiding scheme, Pattern Recognition Lett. 22 (9) (2001) 10511058. [6] Ran-Zan Wang, Chi-Fang Lin, Ja-Chen Lin, Hiding data in images by optimal moderately signicant-bit replacement, IEE Electron. Lett. 36 (25) (2000) 20692070. [7] Chi-Kwong Chan, L.M. Cheng, Improved hiding data in images by optimal moderately signicant-bit replacement, IEE Electron. Lett. 37 (16) (2001) 10171018. [8] Ran-Zan Wang, Chi-Fang Lin, Ja-Chen Lin, Image hiding by optimal LSB substitution and genetic algorithm, Pattern Recognition 34 (3) (2001) 671683.

results show that similar PSNR values can be obtained for di erent type of cover-images. 5. Conclusion In this paper, a data hiding method by simple LSB substitution with an optimal pixel adjustment process is proposed. The image quality of the stego-image can be greatly improved with low extra computational complexity. Extensive experiments show the e ectiveness of the proposed method. The results obtained also show signicant improvement than the method proposed in Ref. [8] with respect to image quality and computational e ciency.

About the AuthorCHI-KWONG CHAN received the B.Eng. and M.Phil. degrees in Electronic Engineering and the Ph.D. degree in Computer Engineering and Information Technology from City University of Hong Kong in 1996, 1999, and 2003, respectively. His research interests include security, image processing, neural networks, and FPGA implementation.

About the AuthorL.M. CHENG received the B.Sc. degree in Physics and Computer Science and the Ph.D. degree from Kings College London (previously Queen Elizabeth College), University of London in 1979 and 1982, respectively. He had been employed as a research fellow at Kings College, London, Principal Engineer of ERA Technology Ltd. UK, and Project Manager and Senior Consultant at Logica Space and Defence Systems Ltd. UK. He joined City University of Hong Kong in 1989 and is now an Associate Professor at the City University of Hong Kong. His research interests include image processing, security, and neural network.