A novel image hiding approach based on correlation analysis for secure multimodal biometrics

ARTICLE IN PRESS Journal of Network and Computer Applications 33 (2010) 247–257

Contents lists available at ScienceDirect

Journal of Network and Computer Applications journal homepage: www.elsevier.com/locate/jnca

A novel image hiding approach based on correlation analysis for secure multimodal biometrics Miao Qi a,b, Yinghua Lu b,, Ning Du a, Yinan Zhang a, Chengxi Wang a, Jun Kong a,c, a

Computer School, Northeast Normal University, Changchun, China Faculty of Chemistry, Northeast Normal University, China c Key Laboratory for Applied Statistics of MOE, Northeast Normal University, China b

a r t i c l e in fo

abstract

Article history: Received 20 May 2009 Received in revised form 17 November 2009 Accepted 7 December 2009

This paper proposes a novel multimodal biometric images hiding approach based on correlation analysis, which is used to protect the security and integrity of transmitted multimodal biometric images for network-based identification. Compared with existing methods, the correlation between the biometric images and the cover image is first analyzed by partial least squares (PLS) and particle swarm optimization (PSO), aiming to make use of the abundant information of cover image to represent the biometric images. Representing the biometric images using the corresponding content of cover image results in the generation of the residual images with much less energy. Then, considering the human visual system (HVS) model, the residual images as the secret images are embedded into the cover image using middle-significant-bit (MSB) method. Extensive experimental results demonstrate that the proposed approach not only provides good imperceptibility but also resists some common attacks and assures the effectiveness of network-based multimodal biometrics identification. & 2009 Elsevier Ltd. All rights reserved.

Keywords: Correlation analysis Partial least squares Particle swarm optimization Middle-significant-bit Multimodal biometrics

1. Introduction Biometrics is an emerging technology that utilizes distinct physiological or behavioral characteristics to determine or verify the identity of an individual. Due to the outstanding features of uniqueness, reliability and stability, biometrics has been widely applied to secure identification/verification systems and has replaced traditional recognition methods. Although biometricsbased identification methods have many advantages over traditional methods, biometric data themselves cannot provide secrecy and security. Ratha et al. (2001) outlined the potential eight basic sources of attacks to the biometric verification system. For instance, the biometric data being transmitted over the network is vulnerable to potential attacks, which can alter the content of biometric data and degrade the performance of biometric systems. Thus, protecting the security and integrity of the biometric data is a critical issue for ensuring valid biometric identification. Recently, making use of information hiding techniques, such as watermarking and steganography, to protect the security and integrity of the transmitted biometric data has been becoming an  Corresponding author. Tel.: + 86 13756158633; fax: +86 431 85696533.  Corresponding author at: Computer School, Northeast Normal University,

Changchun, China. Tel.: + 86 13756158633; fax: + 86 431 85696533. E-mail addresses: [email protected] (M. Qi), [email protected] (Y. Lu), [email protected] (J. Kong). 1084-8045/$ - see front matter & 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.jnca.2009.12.004

active topic. Ratha et al. (2000) described a blind data hiding method to protect fingerprint images with wavelet-packet scalar quantization standard. Jain et al. (2002) introduced an amplitude modulation-based watermarking method, in which a bit stream of eigenface coefficients was embedded into the selected fingerprint image. Hiding biometric information in a digital image using watermarking technique was described in the literature (Mohamed, 2007). The proposed technique gave a quite simple solution for inserting a secure authentication watermarking in dispersed-dot halftone images, and the hidden biometric data was extracted accurately from the carrier image. Vatsa et al. (2004) presented a case where a face image was used as the cover image and an iris code as the watermark for multibiometric verification using four kinds of watermarking algorithms. Meanwhile, several types of attacks were studied to evaluate the robustness of various algorithms in their work. Recently, they presented a 3-level RDWT biometric watermarking algorithm (Vatsa et al., 2009) to embed the voice biometric MFC coefficients in a color face image. The watermarking algorithm used adaptive user-specific watermarking parameters to improve the performance of hiding method. Also, the experimental results indicated that the proposed algorithm was robust against some frequency and geometric attacks. Noore et al. (2007) presented a watermarking technique using face and demographic text data as multiple watermarks for verifying the chain of custody and protecting the integrity of a fingerprint image. The watermarks were embedded into the coefficients of DWT. Combining DWT

ARTICLE IN PRESS 248

M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

and LSB, literature (Vatsa et al., 2006) proposed a multimodal biometric image watermarking scheme that embedded the feature vectors of a face image into a fingerprint image in order to verify the integrity of biometric data. The results demonstrated that the proposed method was more robust against a set of attacks. Khan et al. (2007) introduced a chaotic secure contentbased hidden transmission scheme using DWT. They encrypted biometric data using chaotic encryption to guarantee the system was more secure and protected from the copy attack. Analyzing existing biometric data hiding methods, almost all of them adopt the idea of digital watermarking and hide one or more biometric image/s or its/their features into another biometric image directly based on transform domain for verification. These methods are robust against some types of attacks but the hiding capacity is not high. For example, literature (Vatsa et al., 2004) embedded 10  100 bits to the grayscale face image of size 1024  768 and the average capacity is about 0.0013 bit/pixel. It embedded 128  128 face template into the grayscale fingerprint of size 512  512 in Vatsa et al. (2006), where the average capacity is about 0.0625 bit/pixel. However, due to their real or supposed secret characteristics of biometric images, when transmitting biometric images as carriers there is the risk that attackers will try to intercept or destroy them resulting in degrading the security of transmission. For security purpose, the steganography technique, which hides the very existence of secret communications (Chang et al., 2009) and allows hiding large amounts of information within image, is employed to embed multimodal biometric images into the public transmitted image in this paper. Regarding information hiding methods, researchers recently are paying more attention to the similarity between secret image and cover image for enhancing the imperceptibility and hiding capacity using block–block scheme (Kermani, 2005; Wang and Tsai, 2007; Shen and Hsu, 2007; Hedieh Sajedi and Mansour Jamzad, 2008). In these methods, both the secret image and the cover image were divided into fixed non-overlapping image blocks in advance, and then computed their similarity. In addition, to insert the secret information into the cover image with transparency and robustness, the characteristics of the human visual system (HVS) are often considered in the process of hiding (Kutter and Winkler, 2002; Eyadat and Vasikarla, 2005; Qi et al., 2008; Lee and Tsaia, 2009). The perceptual masking of the HVS is used to determine perceptually significant positions for embedding robust and transparent information. Thus, the embedding strength is adaptive to the features of the cover image and could guarantee maximum-possible imperceptivity. In this paper, a novel biometric images hiding approach with high capacity is proposed, which aims to protect the multimodal biometric images (palmprint and iris) for secret and secure transmission. The cover images, as transmission carriers, have no specific and visual relations to the biometric images, which can assure the security compared with existing biometric information hiding methods. On the sender side, the sample images are first divided into non-overlapping regions for each modality. Then, the related regions are located in the cover image to reconstruct the non-overlapping regions best through correlation analysis using PLS and PSO. After correlation analysis, the residual images, which cannot be represented by the regions, are generated with much less energy. Second, considering the human visual system (HVS) model, the residual images as the secret images are embedded into the middle-significant-bit plane (MSB) of cover image for resisting some common attacks, which also guarantees the quality of stego-image simultaneously. Finally, the biometric images are extracted for identification on the receiver side. The extensive experiments prove the proposed multimodal biometric images hiding method exhibits the following advan-

tages: (1) high hiding capacity, the average capacity is about 0.1094 bit/pixel; (2) perfect stego-image quality, the PSNR is higher than 49; (3) good robustness, the stego-image is resilient to some common attacks such as frequency and geometric attacks. The remainder of this paper is organized as follows. Section 2 describes the relative methods for correlation analysis briefly. Section 3 proposes the process of proposed biometric images hiding method. The multimodal biometric identification is depicted in Section 4. Section 5 presents the extensive experimental results, followed by the conclusions and future work in Section 6.

2. Relative methods 2.1. Partial least squares Partial least squares (PLS) regression (Wold, 1985) is a multivariate data analysis method developed from the practical application, which is mainly used for regression model between multi-dependent variables and multi-independent variables. Compared with ordinary multiple regressions, PLS possesses many advantages, such as avoiding the harmful effects of multicollinearity, and being capable of building the models when the number of observation is less than the number of variables, etc. The goal of PLS regression is to predict Y from X and to describe their common structure. Let X be the mean-centered n by m data matrix of n observations on m predictor variables, Y be the mean-centered n by r data matrix of n observations on r response variables. In PLS, X and Y are decomposed using a given number of latent variables as follows: X ¼ TP T þE; Y ¼ UQ T þ F;

ð1Þ

where T and P are the score and loading for X, U and Q are the score and loading for Y, E and F are the residual for X and Y. The first score vector t1 = Xw1 with the constraint wT1w1 = 1, is the linear combination of predictor data X that has maximum covariance with the y-scores u1 = Yc1 with the constraint cT1c1 =1. w1 and c1 are the eigenvectors corresponding to the largest eigenvalues of matrices XTYYTX and YTXXTY, respectively. The regression equation of X and Y respect to t1 and u1 can be described as: X ¼ t1 pT1 þE; Y ¼ u1 qT1 þ F; T

2

ð2Þ T

2

where p1 ¼ X t1 =Jt1 J , q1 ¼ Y u1 =Ju1 J . 2.2. Particle swarm optimization Particle swarm optimization (PSO) is introduced by Kennedy and Eberhart (1995). It is an evolutionary metaheuristic inspired by the flocking behavior of birds, which has successfully been used to solve optimization problems in some fields (Sun, 2009; Zhou et al., 2009; Oliveira and Schirru, 2009). In the PSO algorithm, each potential solution, called a particle, owns a random generated velocity that directs the particle through the problem space by the fitness value. For the optimization process, each particle is initialized with a random position (Xid), velocity (Vid), and a fitness value evaluated with a predefined fitness function. Then, the fitness value of each particle is updated based on the local best value (Pid) and global best value (Pgd). The local best value is the best solution that the particle has achieved in the current stage. The global best value is the overall best solution tracked by the particle swarm optimizer. After

ARTICLE IN PRESS M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

finding the local and global best values, the particle updates its position and velocity using the following equation: new old ¼ Vid þ C 1  r1  ðPid Xid Þ þ C2  r2  ðPgd Xid Þ; Vid new old new Xid ¼ Xid þ Vid ;

ð3Þ

where C1 and C2 are two positive constants, i represents the number of particles and d is the dimension in the solution space, r1 and r2 are two random values. Initially, PSO generates a swarm of random particles and then searches for the optimal solution by updating particles. The PSO process will be terminated when the termination criteria or maximum number of iterations has been attained.

249

Second, the biometric data are hidden into the cover image with a secret key to form the stego-image for transmission. In the receiver, the biometric data are first extracted from the received stego-image using the secret key. Then, the extracted biometric data are inputted to the classifier for personal identification. Our study mainly focuses on the data hiding part of Fig. 1. The palmprint and iris biometric images are taken as secret transmitted biometric data. The objective of hiding is to protect their integrity and security for effective network-based identification. The flowchart of the proposed biometric images hiding approach based on correlation analysis is shown in Fig. 2.

3.1. Correlation analysis 3. The proposed image hiding approach In this paper, a novel biometric data hiding approach is proposed for network-based personal identification. The proposed approach is applicable to universal scenarios as Ratha (2003). An applicable scenario of the proposed approach is shown in Fig. 1. In the sender, the biometric data (such as face, fingerprint, and palmprint) to be identified are obtained first. Simultaneously, a cover image is selected randomly from the cover image database. The cover images, as the transmission carriers, are unrelated to the secret information in content and are taken from general nature scenes. Nevertheless, the images with abundant texture image are used in our study aiming to make use of the cover image information to best represent the biometric images.

In this paper, the objective of correlation analysis is to make use of the abundant information of cover image to represent the secret images. The detailed process of correlation analysis between palmprint and cover image is described as follows, which is the same for analyzing the correlation between iris and the cover image. Step 1: The palmprint P (m  n) is divided into k nonoverlapping sub-images Pk(ml  nl) and randomly select a cover image from the cover image database as a transmission carrier. Step 2: For each sub-image Pk, PLS is used to pick out a region Yk(m1  n1) in the cover image where it can best represent Pk in content. That is, the region can reconstruct the sub-image adequately. The position of Yk is obtained by the optimization

Cover images Carrier selection Biometric data acquisition

Data hiding

Secret key

Stego-image

The Sender

Transmission Cover images

Data extraction

Secret key

The Receiver

Identification Result Fig. 1. Applicable scenario of the proposed approach.

Residual Palmprint

Palmprint

Correlation Analysis

Iris

Correlation Analysis

HVS

Cover image

Hiding

Residual Iris Fig. 2. The flowchart of proposed biometric images hiding approach.

Stego-image

ARTICLE IN PRESS 250

M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

method PSO and the fitness function is defined as: Min Funk ¼

ml X n1 X

jPk ði; jÞPk0 ði; jÞj;

i ¼ 1j ¼ 1 T

Pk0 ¼ t~ k q~ k ;

q~ k ¼

ð4Þ

YkT t~ k ; t~ k ¼ Pk ck ; 2 :t~ k :

where P0 k presents reconstructed palmprint sub-image corresponding to Pk, ckis the eigenvector corresponding to the largest eigenvalue of matrix YkTPkPkTYk. Step 3: Once the region is located, the residual sub-image Ek can be obtained by: Ek ¼ Pk Pk0 :

ð5Þ

Step 4: Repeat the above steps for other sub-images. Then the residual images are combined into one residual image E as a secret image. t~ k and the positions of related regions are recorded as secret key for reconstruction. Due to the randomness of PSO and random selection of cover image, the secret key provides strong randomness and it is

’

P1

P2

P1

P3

P4

P3

’

P2

’

E1

E2

’

E3

E4

P4

Fig. 3. (a) Palmprint, (b) cover image, (c) reconstructed palmprint, and (d) secret image.

difficult to be deciphered by attackers. Fig. 3 shows an example of correlation analysis, where white squares in Fig. 3(b) are the related regions. Because E is the difference result, some pixels may be negative. To visualize E with much less energy, the modality of absolute value of E is shown as Fig. 3(d). The energy histograms of original, reconstructed and residual palmprints are shown in Fig. 4(a–c) respectively. We can observe that the energy of Fig. 4(c) is much less than Fig. 4(a) obviously.

3.2. Human visual system model To enhance the impermeability of the proposed hiding approach, the human visual system (HVS) model proposed in Lee and Tsaia (2009) is also considered in this study. That is, the positions of embedding are located using the HVS model but different embedding style compared with Lee and Tsaia (2009). Two HVS characteristics, texture and luminance, are considered to reduce image distortion in stego-image. The number of usable embedding bits of each pixel into the cover image is determined by its neighborhood and grayscale value. The maximum number of data-embeddable bits at each pixel is recorded in a map table. The generation of the map table can be described as given below: Given a grayscale image g, let MAX and MIN represent the maximum and minimum intensity in a 3  3 block with g(i,j) as the center. The map T of maximum number of data-embeddable bits of g(i,j) considering the texture characteristics is computed by Tði; jÞ ¼ ½log2 ðD=2Þ ¼ ½log2 D1 ¼ ½log2 ðMAXMINÞ1:

ð6Þ

The map L of maximum number of data-embeddable bits of g(i, j) considering the luminance characteristics is computed as

Fig. 4. The energy histograms of Fig. 3(a), (c) and (d).

ARTICLE IN PRESS M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

follows: 8 4; if gði; jÞ Z 160; > > > > > > < 3; if gði; jÞ Z 72; 2; if gði; jÞ Z 32; Lði; jÞ ¼ > > > 1; if gði; jÞ Z 10; > > > : 0; otherwise:

ð7Þ

After obtaining the maps of T and L, the map table of maximum number of data-embeddable bits of g is obtained: Map= min(T, L). And the value range of each element in Map is 0–4. The map tables of red, green and blue channels of ‘flower’ cover image are shown in Fig. 5.

directly. Since only the least significant bits of pixels are changed, it is visually imperceptible by human. Nevertheless, it is vulnerable to even a slight image manipulation and can be analyzed easily by statistical steganalysis methods such as histogram-based analysis (Westfeld and Pfitzmann, 2000). To overcome the shortcomings of LSB, the secret images are embedded into the middle-significant-bit plane (MSB) of the cover image in this paper. That is, the middle significant bits of pixels are altered instead of the least bit plane (s). Considering the HVS model, the binary sequences are embedded as follows: ( Cp ði; jÞ ¼ (

3.3. Secret images hiding Cq ði; jÞ ¼ 3.3.1. Secret images conversion Because the secret image is a difference result, the values of some pixels may be negative. For hiding conveniently, the secret images are converted to binary sequences before hiding. The detailed process of conversion of a secret image is described as follows: Step 1: Divide the secret image into two parts, sign part and digital part. For the sign part, it is represented by a binary bit plane. If the sign is positive, the relative value of bit is assigned to zero; otherwise, one is assigned. Step 2: Represent each value of digital part with binary bits. Almost all value can be represented by six bits and only very few values might take up seven bits. For accordant hiding, only the last six bits are preserved in case some values take up seven bits. Thus, each pixel of secret image is represented by seven bits, the first bit is sign and the other six bits are digital. For the sake of clarity, two examples of conversion are shown as following: 13/1

001101

44/0

101100

Step 3: Connect the seven bits of each pixel orderly. Thus, the secret image with the size of m  n can be converted to binary sequence B= (BkA{0,1}, k= m  n  7). This conversion style might result in more or less loss of some secret images since some values of digital parts may take up seven bits and that only the last six bits are preserved for hiding. However, this loss is very minor (through statistical analysis, the average mean square error values of palmprint and iris are 0.25 and 0.18, respectively) and cannot affect the identification performance, which will be proved by the multimodal identification accuracy in Section 5. 3.3.2. Secret images embedding The basic idea of least significant bit (LSB) (Bender et al., 1996) is to embed secret message into the least bit plane of cover image

251

Cp ði; jÞ;

Mapði; jÞ ¼ 3 and BðkÞ ¼ 0;

Cp ði; jÞ;

Mapði; jÞ ¼ 3 and BðkÞ ¼ 1;

Cq ði; jÞ;

Mapði; jÞ ¼ 4 and BðkÞ ¼ 0;

Cq ði; jÞ;

Mapði; jÞ ¼ 4 and BðkÞ ¼ 1;

ð8Þ

where CR,p and CR,q present the p and q bit planes of cover image respectively. B is an embedded binary sequence. In particular, the palmprint and iris secret images are hidden into the red and blue channels, respectively, since the human visual system is less sensitive to these channels in RGB color space. Different cover images possess different map tables for embedding. If the number of bits of red or blue channels is not enough for embedding, the remainder binary sequences are embedded into the green channel. An example of our proposed hiding method is shown in Fig. 6. The extraction process of secret images is the reverse of embedding, and the extracted binary sequence is lossless. Fig. 7 shows an extracted example of biometric images. The binary sequences are extracted as follows. ( BðkÞ ¼

0; 1;

if Mapði; jÞ 4 2 and Sði; jÞ ¼ Cði; jÞ; if Mapði; jÞ 42 and Sði; jÞ a Cði; jÞ;

ð9Þ

where C and S are the cover image and stego-image, respectively and Map is the map table of cover image. By executing the above procedure, two binary sequences can be extracted from the three channels of stego-image. For one binary sequence, each seven bits are composed as a group. The first bit of each group is recorded as a sign bit. The remaining six bits are converted to decimal form as digital part. Combine the sign part and digital part to obtain the secret image (shown in Fig. 7(b)). The reconstructed image of each modality can be obtained using the secret key. According to t~ k and the positions of related regions Yk, the reconstructed sub-images P0 k are generated using Eq. (4). Together the secret images with reconstructed images to form the complete biometric images (shown in Fig. 7(d)).

Fig. 5. (a)–(c) map tables of three channels of cover image: (a) MapR, (b) MapG and (c) MapB.

ARTICLE IN PRESS 252

M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

Fig. 6. Hiding of secret images. (a) Secret images, (b) cover image, and (c) Stego-image (PSNR =49.84).

Fig. 7. Biometric images extraction. (a) Stego-images, (b) secret images, (c) reconstructed images, and (d) recovered biometric images.

Fig. 8. ROIs of palmprints from two individuals in PolyU database.

Fig. 9. ROIs of iris images from two individuals in CASIA database.

Fig. 10. The five cover images: (a) Lena, (b) F16, (c) Peffer, (d) Flower and (e) Horse.

4. Multimodal biometric identification The objective of data hiding algorithm is to provide security to a biometric identification system without compromising the quality

and features of the transmitted biometric images. The accuracy of multimodality biometrics is taken as a metric to validate the robustness of the proposed hiding approach. The databases and identification algorithms are presented in the following subsections.

ARTICLE IN PRESS M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

253

4.1. Biometric database

Fig. 11. The PSNR vales of EXP 1 and EXP 2.

Two databases, applied widely in the field of biometrics, are adopted to test the proposed approach. One is the palmprint database from Hong Kong Polytechnic University (http://www4. comp.polyu.edu.hk/ biometrics) and the other is the iris database from CASIA (http://www.cbsr.Ia.ac.cn/nglish/IrisDatabases. asp). In our study, 756 palmprints from 108 individuals are used to evaluate the identification performance. Each individual provides seven images, four images used for training and the others for testing. The size of original palmprint is 384  284 pixels. Region of interest (ROI) of each original palmprint is cropped with size 64  64. Fig. 8 shows the ROIs of two individuals. The CASIA database contains 756 iris images from 108 individuals. Each individual provides seven images; the numbers of training and testing samples are the same with palmprint. The size of each image is 320  280. The ROI of iris image is segmented using our pervious proposed user-specific method (Qi et al., 2008). Some ROIs with size 64  128 are shown in Fig. 9.

Fig. 12. Five types of cropping attacks. (a) Horizontal, (b) vertical, (c) circle, (d) Left-top and (e) Right bottom.

Fig. 13. Attacked stego-images and recovered biometric images.

ARTICLE IN PRESS 254

M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

4.2. Identification algorithms In the stage of feature extraction, 2-D principal component analysis (2DPCA) (Yang et al., 2004) is used to extract the features of palmprint. Gabor filters (Zhang et al., 2002) and statistical entropy are adopted to extract iris features. The detailed process is described in Qi et al. (2008). In the decision stage, given a pair of palmprint and iris images of an unknown user, the features are first extracted and then matched to all templates in the biometric databases. For the user i in the palmprint template database, there are n matching scores (Pti, t =1,2 ,y, n, i =1,2 ,y, N), where n is the number of templates of one user and N is the number of users. The smallest score measured by Manhattan distance is recorded as the final palmprint matching result. Similarly, this process is also applied to the corresponding iris image; the matching scores are recorded as (Iti, t= 1,2 ,y, n, i= 1,2 ,y, N. The final degree of similarity between the unknown user and the user i is expressed as the summation of the two smallest scores, which is defined as minimal distance rule (MDR): n i nN Si ¼ minnt ¼ 1 ðPti Þ=maxnN j ¼ 1 ðPj Þ þ mint ¼ 1 ðIt Þ=maxj ¼ 1 ðIj Þ;

i ¼ 1; 2; . . . ; N:

ð10Þ The minimum of Si indicates the maximal similarity of the training sample and the testing sample. And they are considered coming from the same class.

testing samples need 324 cover images. Therefore, the average values of PSNR are computed for each type of stego-image. Fig. 11 shows the comparison results of two approaches. In general, the image quality is acceptable if the PSNR value is greater than 35. That is, the embedded information is invisible to human eyes. As can be seen from Fig. 11, all PSNR values of both approaches are larger than 35, which indicates that both approaches achieved very good imperceptibility. Furthermore, all PSNR values of five stego-images are larger than 48 using our proposed hiding approach and the average PSNR value of these five types of stego-images is 49.36, which displays more outstanding imperceptibility than EXP 2 of 46.78. The perfect imperceptibility of the proposed methods can be attributed to the following two aspects. First, by adopting correlation analysis based on PLS and PSO, the energies of residual palmprint and iris are much less than the original palmprint and iris by more than 22 and 27 times on average, respectively. Moreover, most bits of secret images are zero. So, the number of changed pixel values of cover image is less in the process of hiding. Second, the hiding capacity of EXP 1 is less than EXP 2 by 1/8 since each pixel of the secret images of EXP 1 is represented by seven bits. Therefore, EXP 1 exhibits much better imperceptibility than EXP 2. Table 1 The integrity comparisons of EXP 1 and EXP 2. Attacks (parameters)

5. Experiments

Palmprint PSNR1

The proposed hiding approach embeds the multimodal biometric images into the cover image for network-based biometric identification. There are five color images (512  512) shown in Fig. 10 used as cover image database (not limited to these five images in the real applications) to test the performance of the proposed approach. Three of them are from the common images used in information hiding (Chan and Cheng, 2004; Tsai and Wang, 2007), while the others from image retrieval database COREL (http://www.corel.com). Two series of comparison experiments are designed to demonstrate the predominant performance of the proposed approach. One is our proposed correlation analysis based hiding approach (called EXP 1). In experiments, the parameters of correlation analysis and hiding are set as k=4, p =3, and q= 4. And the other is embedding the original palmprint and iris into cover image directly (called EXP 2). That is, no correlation analysis procedure is considered in EXP 2. Each pixel value of biometric image is represented by eight binary bits to form the binary sequence. The processes of embedding and extraction are similar to EXP 1. Section 5.1 gives the performance comparisons in the view of imperceptibility of EXP 1 and EXP 2. The robustness comparisons are illustrated in Section 5.2. 5.1. Imperceptibility After the watermark is embedded into the original image, the peak signal to noise ratio (PSNR) is used to evaluate the quality of the stego-image: 1 0 B 255  255  3  M  N C C B PSNR ¼ 10  log10 B C; M P N A @P 2 ½CR;G;B ði; jÞSR;G;B ði; jÞ

ð11Þ

i¼1j¼1

where C and S are the cover image and stego-image, respectively. Because the carrier is selected randomly from the cover image database for a pair testing palmprint and iris, 324 (108  3) pair

Iris PSNR2

PSNR1

PSNR2

Gaussian filtering (standard deviation) 0.300 57.57 0.325 34.87 0.350 23.52 0.375 19.06 0.400 16.66

59.15 30.48 20.31 16.29 14.14

53.12 28.39 19.79 15.82 13.73

36.40 18.67 12.55 9.73 8.16

Gaussian noise (mean, variance) (0,0.0000010) 28.13 (0,0.0000015) 24.36 (0,0.0000020) 22.24 (0,0.0000025) 20.84 (0,0.0000030) 19.85

24.75 21.17 19.20 17.96 17.05

26.48 22.68 20.53 19.11 18.09

19.25 16.04 14.26 13.07 12.21

Impulse noise (density) 0.08 0.09 0.10 0.11 0.12

25.89 25.30 24.79 24.23 23.89

22.59 22.05 21.57 21.14 20.72

24.23 23.65 23.12 22.64 22.20

17.33 16.85 16.43 16.03 15.66

Resize (scale) 1.15 1.20 1.25 1.30 1.35

25.74 19.33 78.95 22.48 18.24

22.33 16.40 83.41 19.05 15.18

23.00 15.87 80.01 23.31 14.76

15.07 9.70 72.38 12.43 8.75

Cropping (area) Horizontal (15.23%) Vertical (15.23%) Circle (19.62%) Left-top (25%) Right-bottom (25%)

22.93 22.63 31.70 16.44 68.68

20.82 19.53 26.70 14.89 62.27

20.97 21.06 15.27 26.98 29.03

14.66 14.67 13.26 25.03 12.01

Table 2 The identification results without attacks. Modalities

CIR (%)

Palmprint Iris Palmprint+ Iris

98.77 97.22 100

ARTICLE IN PRESS M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

255

Fig. 14. CIR comparisons of different attacks: (a) Gaussian filtering attacks, (b) Gaussian noise attacks, (c) Impulse noise attacks, (d) Resize attacks and (e) Cropping attacks.

48.96

39.36

24.84

23.67

Fig. 15. Biometric images extracted from different stego-images.

5.2. Robustness First, the robustness of hiding is evaluated by the integrity of biometric images after attacking the stego-images. Five common types of frequency and geometric attacks are performed, including Gaussian filtering with 3  3 kernel, Gaussian noise, impulse noise, resize, and cropping. For each type of attack, the different attacked degrees are tested. For the cropping attacks, there are

many cropping types. Three cropping factors (area, shape and position) are considered in our work and the five types of cropping are shown in Fig. 12. The integrity of extracted biometric images is also evaluated by PSNR. Fig. 13 shows some attacked results. The first column is the attacked stego-images with attacked types. The numbers between brackets are the corresponding attacked degree. The detailed attacked parameters can be found in Table 1. The second

ARTICLE IN PRESS 256

M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

and the fourth columns are the biometric images extracted using EXP 1. And the remainder two columns are extracted using EXP 2. The values below the biometric images are the corresponding PSNR. The comparisons of integrity of palmprint and iris with various attacks are shown in Table 1, where PSNR1 and PSNR2 are corresponding PSNR values extracted using EXP 1 and EXP 2, respectively. Seen from the values of the two approaches, this table demonstrates that the EXP 1 provides better robustness than EXP 2. Next, under the stego-images being attacked, we adopt the performance of multimodality identification to evaluate the robustness of hiding. The applicable information hiding technique should be resilient to some common attacks, especially when it is applied to protect biometric images. The comparisons of correct identification rate (CIR) from both unimodal systems (palmprint or iris modality) and the multimodal biometric system (palmprint and iris modality) without attacks are listed in Table 2. Seen from the results, it is clear that the identification accuracy based on the palmprint outperforms the iris identification. It can also be seen that the fusion of palmprint and iris improves the accuracy and the CIR can reach 100%. Then, the effect of attacks on the extracted palmprint and iris images for multimodality identification is evaluated. Fig. 14 shows the identification results with various attacks. As seen from Fig. 13, EXP 2 failed to thwart the kinds of attacks as a whole. On the contrary, EXP 1’s are very stable in the whole process of experiments. The attacks of Gaussian filtering and Gaussian noise nearly have no effect on CIR with small attacked degree, and the CIR of EXP 1 can reach 96.3% when the density of noise reaches 3  10  6. The CIRs of EXP 1 are almost invariable with the increase of density of impulse noise from 8% to 12%. Obviously, the CIRs of EXP 2 decrease rapidly with increasing attacked degree for the above-mentioned three types of attacks. Both CIRs of EXP 1and EXP 2 are not affected when the scaling is equal to 1.25 because of the nearest interpolation. For the cropping attacks, the CIR curves are variable obviously with the manner of cropping. The CIRs of EXP 1 can reach 99.69%, 99.38% and 97.53% for the horizontal, vertical and right-bottom cropping types, respectively. However, all CIRs are lower than 85% of EXP 2 with cropping attacks. In summary, the results in Table 1 and Fig. 14 demonstrate that our proposed approach is robust to some common attacks and outperforms the approach without correlation analysis. After correlation analysis, the principal energy parts of biometric images can be reconstructed by cover image and are not hidden with the secret image resulting they are secure and intact. Only the residual energy parts (secret images) are hidden into the cover image and the damage to these parts has a lower effect on the restoration of biometric images compared with principal energy parts. Therefore, the importance and effectiveness of correlation analysis are proved again by robustness evaluation.

6. Conclusions and future work A novel image hiding approach based on correlation analysis and human visual system for biometric identification has been proposed in this paper. For the sake of security, steganography technique is adopted for secret transmission. Compared with existing methods, the correlation between the biometric images and the cover image is analyzed by partial least squares (PLS) and particle swarm optimization (PSO), aiming to make use of the abundant information of cover image to represent the secret images. For each transmission, the secret images are hidden into a cover image selected randomly. Furthermore, the MSB and HVS

model are adopted for hiding. It is worth noting that each biometric image is transmitted by two parts: reconstructed coefficients and residual image. Each part obtained by the attacker is useless for misdeed or identification. Extensive experimental results show that our proposed approach can resist some common attacks and accomplish the secret transmission task effectively and efficiently for networked biometric identification. In the attacking experiments, we have found that the same attack to stego-images formed with different cover image hiding the biometric images from the same individual, the quality of extracted biometric images is distinguishing obviously. Fig. 15 shows the extracted biometric images with PSNR values from the stego-images submitted to Gaussian filtering with standard deviation of 0.325. So, the intrinsic correlation between biometric and cover image will be studied in the future work, which might enhance the robustness of the hiding algorithm.

Acknowledgments This work was supported by the Key Project of MOE (No. 109052), the Training Fund of NENU’S Scientific Innovation Project (No. NENU-STC07018), the Fund of Jilin Provincial Science & Technology Department (No. 20070322), and the NSFC (No. 10871037). References Bender W, Gruhl D, Morimoto N. Techniques for data hiding. IBM Systems Journal 1996;35:3–336. Chang CC, Kieu TD, Wu WC. A lossless data embedding technique by joint neighboring coding. Pattern Recognition 2009;42:1597–603. Chan CK, Cheng LM. Hiding data in images by simple LSB substitution. Pattern Recognition 2004;37:469–74. Eyadat Mohammad, Vasikarla Shantaram. Performance evaluation of an incorporated DCT block-based watermarking algorithm with human visual system model. Pattern Recognition Letters 2005;26:1405–11. Hedieh Sajedi, Mansour Jamzad: Cover selection steganography method based on similarity of image blocks, IEEE 8th International Conference on Computer and Information Technology Workshops (2008) 379-84. Jain AK, Uludag U, Hsu RL. Hiding a face in a fingerprint image, in: proceeding of the International Conference on Pattern Recognition 2002; 3: 756–9. Kennedy, J, Eberhart, R.: Particle Swarm Optimization. Proceeding of the 1995 IEEE International Conference on Neural Network 1995; 1942-8. Kermani Z. A robust steganography algorithm based on texture similarity using Gabor filter. IEEE signal processing and Info. Technology 2005:578–82. Khan Muhammad Khurram, Zhang Jiashu, Tian Lei. Chaotic secure content-based hidden transmission of biometric templates. Chaos, Solitons and Fractal 2007;32:1749–59. Kutter M, Winkler S. A vision-based masking model for spread-spectrum image watermarking. IEEE Transactions on Image Processing 2002:16–25. Lee IH, Tsaia WH. Data hiding in grayscale images by dynamic programming based on a human visual model. Pattern Recognition 2009;42:1604–11. Mohamed Mostafa Abd Allah. Embedded biometric data for a secure authentication watermarking. SPPR 2007:191–6. Noore A, Singh R, Vatsa M, Houck MM. Enhancing security of fingerprints through contextual biometric watermarking. Forensic Science International 2007;169:188–94. Oliveira MV, Schirru R. Applying particle swarm optimization algorithm for tuning a neuro-fuzzy inference system for sensor monitoring. Progress in Nuclear Energy 2009;51:177–83. Qi HY, Zheng D, Zhao JY. Human visual system based adaptive digital image watermarking. Signal Processing 2008;88:174–88. Qi M, Lu YH, Li JS, Li XL, Kong J. User-Specific Iris Authentication Based on Feature Selection. CSSE 2008:1040–3. Ratha NK, Connell JH, Bolle RM. An analysis of Minutae Matching Strength. Proceedings of the Third International Conference on Audeo-Based Biometric Person Authentication 2001:223–8. Ratha NK, Connell JH, Bolle RM. Secure data hiding in wavelet compressed fingerprint images. Proceedings of the 2000 ACM Workshop on Multimedia 2000:127–30. Ratha NK. Hiding Biometric Data. IEEE Transaction on Pattern Analysis and Machine Intelligence 2003;25:1494–8. Shen JJ, Hsu PW. A robust associative watermarking technique based on similarity diagrams. Pattern Recognition 2007;40:1355–67. Sun TH. Applying particle swarm optimization algorithm to roundness measurement. Expert Systems with Applications 2009;36:3428–38.

ARTICLE IN PRESS M. Qi et al. / Journal of Network and Computer Applications 33 (2010) 247–257

Tsai Y Yu, Wang CM. A novel data hiding scheme for color images using a BSP tree. Journal of System and Software 2007;80:429–37. Vatsa, M, Singh, R, Mitra, P, Noore, A. Comparing robustness of watermarking algorithms on biometrics data, Proceedings of the Workshop on Biometric Challenges from Theory to Practice-ICPR Workshop 2004; 5-8. Vatsa M, Singh R, Noore A. Feature based RDWT watermarking for multimodal biometric system. Image and Vision Computing 2009;27:293–304. Vatsa M, Houck MM, Morris K. Robust biometric image watermarking for fingerprint and face template protection. IEICE Electron Express 2006;3: 23–28. Wang RZ, Tsai YD. An image-hiding method with high capacity based on best-block matching and k-means clustering. Pattern Recognition 2007;40:398–409.

257

Wold, H. Partial least squares, In S. Kotz and N. L. Johnson, editors, Encyclopedia of the statistical Sciences 6 (1985) 581–91. Westfeld A, Pfitzmann A. Attacks on steganographic systems. Lecture Notes in Computer Science 2000;1768:61–75. Yang J, Zhang D, Frangi AF, Yang JY. Two-dimensional PCA: a new approach to appearance-based face representation and recognition. IEEE Transaction on Pattern Analysis and Machine Intelligence 2004;26:131–7. Zhang, JG, Tan, TN, Ma, L. Invariant Texture Segmentation via Circular Gabor Filters, International Conference on Pattern Recognition, 2002. p. 901–4. Zhou JZ, Fang RC, Li YH, Zhang YC, Peng B. Parameter optimization of nonlinear grey Bernoulli model using particle swarm optimization. Applied Mathematics and Computation 2009;207:292–9.