Protecting Biometric Templates Using

0 downloads 0 Views 548KB Size Report
biometric template has not been tampered and whether it came from the right person. ... Nikos [6] have combined lattice and block-wise image watermark-.

Protecting Biometric Templates Using Authentication Watermarking ChunLei Li1,2 , Bin Ma1 , Yunhong Wang1 , and Zhaoxiang Zhang1 1

2

School of Computer Science and Engineering, Beihang University, 100191, Beijing, China School of Electronic and Information Engineering, Zhongyuan University of Technology, ZhengZhou 450007, China

Abstract. In this paper, we propose a novel scheme for protecting biometric templates using salient region-based authentication watermarking. Firstly, a novel multi-level authentication watermarking scheme is proposed, which is used to verify the integrity of biometric image. Secondly, the PCA features of these biometric images is used as information watermarks to recover the tampered image. Authentication bits and information bits are adaptively embedded into the biometric images based on the result of salient region detection. Experimental results show that our proposed scheme can detect the tampered region, and recover the biometric features while keeping recognizing quality. Keywords: Salient region, hierarchical, authentication watermarking, biometric templates.

1

Introduction

With the widely use of verification systems based on biometrics, it creates a demand for ensuring the security and integrity of biometrics data. Cryptography, steganography and watermarking can be used to enhance the biometrics data security and secrecy. Cryptography method is only used to verify whether the biometric template has not been tampered and whether it came from the right person. Digital watermarking which is opposed to pure cryptographic tools can enable us to localize tampered or damaged areas. Protecting biometric templates based on digital watermarking have been a research focus [1][2][3][5][6][7]. Tuan [7] proposed a new remote multi-modal biometric authentication framework based on fragile watermarking. A facial image is used as a host image to embed other numeric biometrics features, then transfer multi-biometrics over networks to a server for authentication. Ahmed [1] proposed a phase-encodingbased digital watermarking technique for fingerprint template protection and verification. Nikos [6] have combined lattice and block-wise image watermarking techniques to maintain image quality, along with cryptographic techniques to embed fingerprint templates into facial images and vice-versa protecting the biometric template. G. Qiu et al. (Eds.): PCM 2010, Part I, LNCS 6297, pp. 709–718, 2010. c Springer-Verlag Berlin Heidelberg 2010 

710

C. Li et al.

Unfortunately, the original biometric image may be distorted by the watermark, which will decrease the recognition rate in biometric identification systems. In order to reduce this effect, we should embed less information in region of background (ROB) which is susceptible to the effect of watermarks, while embed more information in the regions of salient (ROI). In this paper, a novel scheme for protecting biometric templates using salient region-based authentication watermarking is proposed. We embed the authentication watermark over the image, which is used to protect the whole image, and embed the information bits into the ROI for recovering the damaged biometric features.

2

Salient Region Detection

Salient region detection is a basic part of many theories among computer vision. The detection result is a salience map which represents the rich-information area of the object. There is tremendous engineering applications, e.g in compression and recognition, determination of suitable locations for embedding watermarks. In this paper, the salient region is detected using the method proposed by [4]. Firstly, a feature map R : [n]2 → R is generated. Then a fully-connected directed graph GA is obtained by connecting every node of the lattice M , labeled with 2 two indices (i, j) ∈ [n] , with all other n − 1 nodes. The directed edge from node(i, j) to node (p, q) will be assigned a weight w1 ((i, j), (p, q))  d(i, j)  (p, q) · F (i − p, j − 1) 2

(1)

2

where F (a, b)  exp(− a2σ+b2 ) and d(i, j) = |M (i, j) − M (p, q)|. σ is a free parameter. Thus, the weight of the edge form node (i, j) to node (p, q) is proportional to their dissimilarity and their closeness in the domain of M . The author defined a Markov chain on GA . In the end, a normalized activation map GBV SM is generated through concentrating activation into a few key locations on activation maps A , and it is shown in Fig.1(b). Through modulating the threshold, we can segment the image into the regions of salient (ROI) and background region (ROB)by the given ratio. The result with 50% ROI is shown in Fig.1(c). The segmented result can be represented by a binary saliency map MSR = {mi ∈ 0, 1|i = 1, 2, · · · , Np }. if mi = 1, the corresponding pixel is in the salient region. Otherwise, this pixel is in regions of background.

3

Authentication Watermarking and Biometric Protection

In this section, a novel multi-level authentication watermarking scheme is proposed. Based on the proposed authentication watermarking scheme, tamper detection and self-recovery of biometric templates are presented. In order to preserve the recognizing quality, the watermarks are adaptively embedded into the biometric images based on the result of salient region detection.

Protecting Biometric Templates Using Authentication Watermarking

(a) Original image

(b) GBVSM map

711

(c) Salient region

Fig. 1. Salient regions (ROI) of face image

3.1

Salient Region Map (SRM) Generation

According to the method described in Section 2, a saliency region map MSR = {mi ∈ {0, 1}|i = 1, 2, · · · Np } is generated, which is used to depict whether a pixel is in salient region. The face image is split into the image blocks with size mb × nb . And for image block Xi , the number of pixels located in a salient region can be calculated according to the following equation: NSR =

m b ×nb

mj

(2)

j=1

If the number NSR is above (mb × nb )/2, the block is in a regions of salient (ROI). Otherwise the blocks belong to regions of background (ROB). Finally, a saliency map for image blocks M bsr = {mbj |j = 1, 2, · · · Nb } is generated, where Nb is the number of image blocks. 3.2

Watermark Generation

Firstly, a cryptographic hash function MD5 is used to generate the hash code Ci of block Xi . We take the first Nbp bits from Ci , and Nbp is the number of pixels in an image block. The hash code Ci is used as authentication watermark to verify the integrity of the image. As a kind of biometrics, Eigen-face coefficients can re-construct the face image, which are used as hidden information. If the original image suffers from malicious tamper, the reconstructed face will be used as a second source of authenticity, either automatically or by a human in a supervised biometric application [1]. An Eigen-face coefficient Bm (m = 1, 2, · · · , T ) can be converted to a binary stream using 4 bytes per coefficient, corresponding 32 bits, where T is the number of the coefficient of a biometric image. We extent the Eigen-face coefficients periodically b is the number of blocks b /T  × T where N to a number stream with size Q = N

712

C. Li et al.

located in the ROI. Then the block information watermark Bl (l = 1, 2, · · · , Q) is obtained. For any Bl , it is converted to a binary stream bl (l = 1, 2, · · · , 32), and redundancy embedded into the ROI of the original image. The embedded b /T  copies. Therefore, if one or Eigen-face coefficient of a face image has N more copies are tempered by any chance, we have the chance to recover the face image from the other copies. 3.3

Watermark Embedding

In order to improve the accuracy of tamper localization, the image is split into image blocks with size 4 × 4. And to satisfy the security, four user keys k1 ,k2 ,k3 , and k4 are adopted to embed the dual watermarks in a more secure way. The embedding procedure is shown in Fig.2, and it can be described as following: Step 1. Compute the salient map M Bsr for image X ∗ which 4LSBs are set to zero as described in Section 3.1. Step 2. Generate a random position matrix M of size Np × r, where r = 3 and p is the number of pixels in an image block, each element in M is an integer in interval [1, 4], and the elements among the same column are not equal. Step 3. Partition the image into image blocks with size 4×4 . Generate the auk , )k = thentication watermark embedded position sequence f k = (I1k , I2k , · · · , IN b Biometric feature B

Original image

Bit-stream

Salient region extractor

A salient map Mbr

mb j

Y

1?

N

X i' 16  bits hash code come from four blocks

M i ' [0]

k1

Random Matrix M

k2

Position sequence f wk ( k

k3

Insert Ci k Into X i '

Yi'

Insert bi Into X i'

Yi '

X i' M i ' [1]

M i ' [2]

1, 2,3, 4)

Position sequence fc

Fig. 2. Embedding process

Yi'

Protecting Biometric Templates Using Authentication Watermarking

713

1, 2, 3, 4 using one secret key, and Nb is the number of image blocks. And generate Eigen-face coefficient embedded position sequence f Coe = (I1 , I2 , · · · , IQ ) using k3 . Step 4. In order to resist vector quantization (VQ) codebook attack and collage attack, our method is block-wise dependency, by embedding the hash-code of an image block into another four image blocks. For each image block Xi , repeat Step 4.1, Step 4.2 and Step 4.3. Step 4.1 Compute the 16-bits hash code Ci of block Xi according to the above method. And split the hash code into four 4-bits parts Ci = {Ci1 , Ci2 , Ci3 , Ci4 }, i = 1, 2, · · · , Nb . Embed the four subsets into the other four image block. For each subset Cik , the embedding position is decided by the position sequence f k = k ), k = 1, 2, 3, 4. (I1k , I2k , · · · , IN b Step 4.2 For each pixel in the image block Xi , an authentication bit is embedded into it. Instead of embedding the authentication bit into LSB of the pixel. The embedding position is decided by the value of the corresponding M [k][1](k = 1, 2, · · · , Nbp ), where Nbp is the number of pixels in an image block.  which is obtained after embedStep 4.3 Embed Eigen-face coefficients into X ding the authentication watermark, obtaining the final watermarked face image Y . An Eigen-face coefficient is embedded into an image block selected by the position sequence f Coe , and the block is located into the salient region (ROI). For i , two bits of biometric feature each pixel x ji , (j = 1, 2, · · · , Nbp ) in the block X are embedded into it according the M [k][2] and M [k][3] . 3.4

Tamper Detection and Self-recovery

Firstly, we give the process of tamper detection and localization, which is based on a 3-level hierarchical structure, it is shown in Fig.3. And it is briefly described as follows: Step 1. Generate a random position matrix M of size Nbp × r, where r = 3 as we have done in Step 2 of the watermark embedding procedure. And generate k )k = 1, 2, 3, 4 using secret key k2 as four random positions f k = (I1k , I2k , · · · , IN b done in Step 3.1 of the watermark embedding procedure. Step 2. Split the test image Y ∗ into image block Yi∗ , i = 1, 2, · · · , Nb . For each image block, repeat Step 2.1 and Step 2.2. Step 2.1 Generate the Hash value Ci∗ using MD5 Hash function for each image block Yi∗ . And partition the hash code into four parts Ci = ∗ ∗ ∗ ∗ {Ci1 , Ci2 , Ci3 , Ci4 }, i = 1, 2, · · · , Nb , each part has 4-bits. Step 2.2. Extract the four parts of embedded authentication watermark  Cfkk , k = 1, 2, 3, 4 from the other four image blocks. i

1-L detection: Get the block mark D1 = {d1i |i = 1, 2, · · · , Nb } according to the comparison between the generated authentication watermark with the extracted authentication watermark.  ∗  0, Cik = Cik ∀k = 1, 2, 3, 4 1 di = (3) 1, elsewise

714

C. Li et al.

Yi*

k1

Partition H into

H ( m, n, Yi* )

*

*

*

*

Ci1 , Ci2 , Ci3 , Ci4

Random Matrix M &RPSDUH

k2

Position sequence f 1, f 2 , f 3 , f 4

Y f*i1

Y f*i2

7KUHH /HYHO GHWHFWLRQ

Extract watermark

Y f*i3

'

'

i

i

C1'f 1 , C 2f 2 , C 3f 3 , C 4'f 4

Y f*i4

i

i

Fig. 3. Watermark extraction and verification procedure





That is to say, any of the four Cik equals its corresponding Cik , then d1i = 0, the image block keep un-tampered. Otherwise, d1i = 1 to indicate image block suffering from tampering. 2-L detection: Count the block number in its 9-neighbourhood whose di = 1, if the number is larger than or equal to 4, then d2i = 1 indicates the block had suffered from tampering.  d2i

=

 0, (d1i ≤ 4) 1, elsewise

(4)

where N9 (d1i ) represent the 9-neighbourhood of d1i . 3-L detection: Check the following four triples of neighboring block situation (N, NE, E), (E, SE, S), (S, SW, W), (W, NW, N). And N represents the block in north direction, NE represents the block in Northeast direction, E represents the block in east direction, and so on. If all blocks in any of the four triples are marked as invalid by 2-L detection, then mark as the block invalid. Then, we recover the biometric feature based on the result of tamper detection according to the following procedure. Step 1. According to the verification result, partition image blocks into two sets: tampered blocks and un-tampered blocks. Mark the tampered blocks ”1”, and mark un-tampered blocks ”0”. Step 2. Compute the salient map M Bsr for image Y ∗ with 4LSBs set to zeros as described in Step 1 of the watermark embedding procedure.

Protecting Biometric Templates Using Authentication Watermarking

715

Step 3. Generate Eigen-face coefficients embedding position sequence f Coe = (I1 , I1 , · · · , IQ ) using k3 . Step 4. Extract the Embedded Eigen-face coefficients from image block located in the regions of salient (ROI) using position sequence f Coe . An Eigen-face coefficient is extracted from an image block Yi∗ . Step 5. For each (Mj1 , Mj2 , · · · , MjQ ), where j = 1, 2, · · · , T , T is the number of coefficients of a biometric template, Q is the number of copies for each coefficient, a coefficient can be recovered. If the element of Mjk , (k = 1, 2, · · · , Q equals ’0’, the corresponding image blocks suffered from malicious tamper, we discard the extracted coefficients. We extract the coefficients from the image blocks whose mark is ’1’, then a voting-scheme is adopted to reduce the probabilities of false recover.

4

Experimental Results

As shown in Fig.4, the 3-L authentication results are demonstrated under tamper ratios is 70%. From the figure, we can conclude that our proposed hierarchical authentication watermarking has super localization accuracy, while keeping low false detection rate, even when the tamper ratio is very high. In order to evaluate the effect of watermark on the face recognition, we conduct a series of experiments on FERET face database. FERET database can be classified into two sets: query sets and reference sets. In our experiment, we select FA as reference sets, and select FB as query sets. Each set has 1191 face images, and they are scaled as 128 ×152. In experiment, the reference sets are kept unchanged, while query sets are embedded using the three methods: Biometric image embedded into information bits over the whole image, biometric image embedded information bits into ROB of the biometric images, and biometric image embedded information bits into ROI of the biometric images; are named as ’Wm1’, ’Wm2’, ’Wm3’, respectively. The original biometric image is marked as ’Original’. Then, we extract different features from these images. For the PCA feature, we select 95% energy (152 coefficients), for Gabor feature, with 4 scales and 8 directions, and the LBP feature, using the method described as [8]. NN (nearest neighbor) classifier is used as our classifier. Rank curve analysis was used in the FERET evaluations as one basis for algorithm comparison. It provides a method of to analyze recognition rates of

(a) Original

(b) Tamper

(c) 1-Level

(d) 2-Level

Fig. 4. Detection result (70%)

(e) 3-Level

716

C. Li et al.

an algorithm recognition rank. Although this analysis is simple it can provide interesting information not apparent in a rank one recognition rate. Figure.5 shows the recognition result.

Fig. 5. Recognition using different methods for watermarked image

As shown in Fig.5, if PCA and Gabor feature are adopted to evaluate our algorithm, the recognition rate of the four kinds of biometric images are nearly equal, which shows that the watermark has less affect on PCA and Gabor features of biometric image. PCA and Gabor features are the appearance feature of the biometric image, which are less affected by noise; the watermark can be described as the noise added on the image. Thus, the recognition rate of the four kind of biometric images are nearly equal based on PCA and Gabor feature. However, when the LBP feature is adopted to evaluate our algorithm, the recognition rate of ”Wm1’, ’Wm2’ and ’Wm3’ is lower than ’Original’, and ’Wm3’ is higher than ’Wm1’ and ’Wm2’, which shows that LBP which represents the local texture feature is affected by the watermark embedded into it, hence our proposed scheme can keep ”recognizing quality” to certain extent. As the Eigen-face coefficients, can be used to reconstruct the face image, the recovered Eigen-face coefficients can also be used in verification system to reconstruct the Eigen-face. The reconstructed Eigen-face is then used as a second source of authenticity. In Fig.6, we depict the recovered recognition ratio under

Protecting Biometric Templates Using Authentication Watermarking

Fig. 6. Recognition using PCA

(a) Original

(b) 10%

(c) 20%

(d) 30%

(e) 40%

(f) 50%

(g) 60%

(h) 70%

Fig. 7. Eigen-face recovery

717

718

C. Li et al.

tamper ratio [0.1 0.7], the interval is 0.1, and the number of embedded coefficients for an image is 128. We can see that the recognition rate nearly keep unchanged when the tamper ratio is below 0.4. Fig.7 gives the reconstructed Eigen-face under this conditions. the Eigen-face can be successfully reconstructed when the tamper ratio is below 0.6.

5

Conclusions

A novel scheme for protecting biometric templates using salient region-based authentication watermarking is proposed in this paper. Firstly, a novel hierarchical authentication watermarking scheme is proposed, demonstrate a good performance on tamper detection even when the tamper ratio is up to 0.7. Secondly, a self-recovery algorithm of biometric template is presented based on the same scheme, the recovered PCA coefficients can also be used for recognition system, and the reconstruct face can be used as a second source of authenticity. Finally, the effect of watermark for the biometric image is shown, providing guidance for biometric watermarking algorithm. Experimental results shows that the proposed hierarchical authentication watermarking scheme has superior tamper localization, and can efficiently recover the tampered biometric features while keeping ”recognizing quality” to some extent. Acknowledgments. This work is funded by the National Natural Science Foundation of China (No. 60873158), the National Basic Research Program of China (No. 2010CB327902), the Fundamental Research Funds for the Central Universities, and the Opening Funding of the State Key Laboratory of Virtual Reality Technology and Systems.

References 1. Ahmed, F.: Fingerprint reference verification method using a phase-encoding-based watermarking technique. Journal of Electronic Imaging 17(1), 011010 (2008) 2. Jain, A.K., Uludag, U.: Hiding Biometric Data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1494–1498 (2003) 3. Jain, A.K., Uludag, U., Hsu, R.-L.: Hiding a face in a fingerprint image. In: International Conference of Pattern Recognition, p. 30756 (2002) 4. Harel, J., Koch, C., Grid, P.P.: Graph-Based Visual Saliency. In: NIPS, pp. 545–552 (2006) 5. Kim, W.-G., Lee, H.: Multimodal biometric image watermarking using two-stage integrity verification. Signal Process. 89(12), 2385–2399 (2009) 6. Komninos, N., Dimitriou, T.: Protecting biometric templates with image watermarking techniques. In: ICB, pp. 114–123 (2007) 7. Tuan Hoang, D.T.: Remote multimodal biometric authentication using bit prioritybased fragile watermarking. In: ICPR, pp. 1–4 (2008) 8. Ahonen, T., Hadid, A., Pietika, M.: Face Description with Local Binary Patterns: Application to Face Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(12) (2006)

Suggest Documents