105
International Journal of Hybrid Intelligent Systems 9 (2012) 105–121 DOI 10.3233/HIS-2012-0150 IOS Press
A new hybrid fuzzy biometric-based image authentication model Sherin M. Youssefa, Yasser El-Sonbatyb and Karma M. Fathallaa,∗
a Department of Computer Engineering, College of Engineering and Technology, Arab Academy for Science and Technology, Alexandria, Egypt b College of Computing and IT, Arab Academy for Science and Technology, Alexandria, Egypt
Abstract. Watermarking is one of the most known techniques for authentication, tampering detection, privacy control, etc. Concerning privacy protection of patients’ medical records, efforts have been devoted to guarantee the confidentiality of data and medical images during storage and transmission via an untrustworthy channel. Our developed watermarking system, aims at fulfilling such demands through the utilization of a biometric authentication code (sender physician’s iris code), encrypted patient data and a fuzzy-based Region-of-Interest (ROI) segmentation algorithm. In this paper, a new hybrid DCT fuzzy biometric-based watermarking scheme has been introduced for privacy protection and source verification of medical and non-medical images. The proposed scheme integrates forward DCT transform with enhanced fuzzy-based ROI regions segmentation. To comply with the imperceptibility and high image quality requirements, the coefficients selection decision depends on perceptual visibility threshold estimation, based upon characteristics of the human visual system (HVS). The inclusion of a just-noticeable distortion (JND) profile, computed for DCT coefficients, is proved to outperform the traditional DCT model, with the major contributions of a new formula for luminance adaptation adjustment and the incorporation of block classification for contrast masking. A block classification has been utilized to differentiate edge regions and thus effectively avoid over-estimation of JND in the selected regions. Moreover, several enhanced fuzzy-based clustering models have been suggested for extraction of ROI regions in medical patterns, which aim at increasing the robustness to noise and yield more homogeneous regions with less spurious blobs. The proposed hybrid fuzzy-based ROI extraction scheme integrates the effect of the local neighborhood and allow it to influence the membership value of each pixel. A new Hybrid FCM (H-FCM) algorithm is proposed, which integrates spatial information with a 2D adaptive noise removal SS-FCM model. Experiments have been conducted to verify the proposed model. Several attacks have been applied to the proposed scheme and the experiments revealed promising results in terms of visual quality and extracted watermark distortion. Keywords: Watermarking, fuzzy techniques, copyright protection, DCT, HVS characteristics
1. Introduction Watermarking is a technique of embedding identification codes, called watermarks, into cover (host) media. Watermarking is used for the protection of intellectual property, data integrity, and data authentication [1,2]. Several factors have lead to the increasing ∗ Corresponding author: Karma M. Fathalla, Department of Computer Engineering, College of Engineering and Technology, P.O. 1029.Arab Academy for Science and Technology, Alexandria, Egypt. Tel.: +20 0122 7612961; Fax: +20 03 4299496; E-mail:
[email protected].
1448-5869/12/$27.50 2012 – IOS Press. All rights reserved
interest in the field of watermarking. These factors include the prevalent use of the internet and communication networks, the immense advancements in digital techniques, the wide availability of consumer electronic devices and the involvement of digital technologies in every aspect of our lives. The extensive attention given to the field of watermarking may be due to that it provides adequate solutions for many applications such as protection of intellectual property rights, copy control, tamper detection, source verification, etc. [2]. Extensive attention over the past decade has been directed to automated personal identification based on biometrics. Several factors have led to a rapid devel-
106
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
opment of multimedia watermarking systems based on biometrics including the recent advances of information technology and the increasing requirement for security. The human iris which is the part between the pupil and the white sclera provides an extraordinary texture with many interlacing minute characteristics such as freckles, coronas, stripes, etc. These visible characteristics are unique to each subject thus it can be efficiently used for personal identification, The iris is more stable and reliable for identification compared to other biometric features (such as face, voice, etc.). Another practically important advantage of iris-based personal identification systems is that they are noninvasive to their users since the iris is an externally visible organ. All these desirable properties (i.e., uniqueness, stability, and noninvasiveness) make iris-based watermarking a particularly promising solution to security in the near future. One of the important issues in the management of patients’ medical records is the privacy protection of medical images. A set of standards for privacy protection of health data were issued as part of Health Insurance Portability and Accountability Act (HIPAA), produced by the federal government. The HIPAA directs health providers to guarantee confidentiality and integrity of individually identifiable health information. The widespread of digital technology an the wide availability of networks facilitated the easy storage, maintenance, retrieval and sharing of medical data. This helps promote high quality care for patients. On the other hand, it poses a great threat on privacy of patients’ information. Constant efforts are being spent to provide security solutions [3–5] to ensure (i) confidentiality, (ii) integrity and (iii) authentication The classification of watermarking techniques can be based on several criteria. One of the most important criteria is the embedding domain, whether the watermark is embedded in the spatial or transform domain. Some of the most common transform domains are Discrete Wavelet Transform (DWT), Discrete Cosine Transform (DCT) and Discrete Fourier Transform (DFT) [6,9]. Another criterion is whether the scheme is non-blind which means that the receiver needs the original image for extraction, semi-blind where there is a need for a key or blind which implies that the extraction can be done only using the watermarked image. This paper is organized as follows. Several watermarking schemes that were lately presented are given in Section 2. Section 3 discusses the application scenarios at which the proposed watermarking system can be used. A detailed description of the system’s modules
is presented in Section 4. Section 5 displays the experimental results of our system and a final conclusion is given in Section 6. 2. Related work A watermarking algorithm was proposed in [7] using balanced multi-wavelet transform. It achieves simultaneous orthognality and symmetry. The embedding scheme is image adaptive based on a well established model. The local properties of the host image control the embedding strength of the watermark. The suggested embedding scheme is influenced by principles of spread spectrum communications. In [8] a watermarking scheme was illustrated to serve copyright protection and integrity verification using discrete cosine transform DCT. Quad tree decomposition is used for image segmentation to find visually significant regions of images. The embedding process is done in significant regions of images using quantization index modulation (QIM). Another technique for copyright protection and authentication of multimedia data is illustrated in [6]. The technique is based on Chinese Remainder Theorem (CRT). The watermarking process is in the DCT domain. The watermark is embedded in the DC and three low frequency AC coefficients per block. This helps attain robustness against JPEG compression and several other attacks Watermark embedding in the DFT domain is utilized in [10]. A circularly symmetric watermark is embedded around the mid frequency region of DFT domain magnitude The watermark is centered at the image center which provides robustness against geometric rotation attack. To reduce visible artifacts, neighborhood pixel variance masking is employed. In the field of medical image watermarking, huge efforts have been paid to provide robust watermarking systems. Pegah et al. [11] presented a novel scheme for protecting patient’s medical records and tracing illegal distribution of medical images. The embedding process is done in the wavelet domain. Particle swarm algorithms and genetic algorithms principles are applied to improve system’s performance. The visual quality of images is degraded to prevent its use directly. This is achieved by the varying strength of the watermark. The algorithm was proved to be reliable for tracing colluders. In several techniques of compression, encryption and watermarking are combined to attain medical data se-
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
curity, confidentiality and integrity. DICOM data is used as the watermark after applying Huffman compression and RC4 encryption. Manual selection of the image’s important regions (Region of Interest ROI) is performed to be able to embed the watermark in the non-ROI. The image centroid is calculated based on image moment theory. Then the image is scanned in a spiral manner where the pixels used for embedding are selected using a homogeneity measure. Multiple watermarks are employed in to protect patient information and aid tampering detection and integrity check. ROI separation is achieved by thresholding and the ROI is embedded twice in the host image. First it is embedded in the third decomposition level of wavelet transform and then it is embedded as a fragile watermark in the spatial domain. Furthermore the physician’s digital signature and electronic patient record are also embedded in the vertical and horizontal wavelet decomposition subbands. 3. Application scenarios The uses of watermarking systems imply several features that must be available in the watermarking scheme design. The design requirements are as follows: A. Imperceptibility: The Human Visual System (HVS) should not recognize a difference between the watermarked and the original image. In other words, the watermarked images have to maintain a high level of quality to preserve its commercial value. B. Security: The watermarked image should resist any manipulations by unintended users such as modification, copy and deletion. C. Robustness: The watermarked image when transmitted can encounter attacks, unintentional signal processing or noise that it must endure. The watermark has to be still extractable out within certain acceptable quality even under the previous conditions. D. Capacity: This requirement describes the pay load which how much data can be and must be embedded in image to represent its uniqueness. Our system is designed to meet the requirements of two application scenarios. The first is copyright protection of general images; the other is source verification and privacy control of medical images.
107
3.1. General application scenario The protection of general images has been extensively studied due to the increased demand of protecting digital images. The widespread of digital technologies and communication systems raised the urge to develop new techniques to protect the copyright and intellectual properties of authors. Such systems impose several design criteria such as robustness to attack and imperceptibility. Our developed watermarking system (as will be illustrated in the following sections), aims at fulfilling such demands through the utilization of a binary logo image, DCT domain embedding scheme and the computation of Just Noticeable Difference (JND) profile for the DCT coefficients. The DCT allows an image to be broken up into different frequency bands namely the high, middle and low frequency bands thus making it easier to to choose the band in which the watermark is to be inserted. The literature survey reveals that mostly the middle frequency bands are chosen because embedding the watermark in a middle frequency band does not scatter the watermark information to most visual important parts of the image i.e. the low frequencies and also it do not overexpose them to removal through compression and noiseattacks where high frequency components are targeted. Just-noticeable distortion (JND) gives us a promising way to model the perceptual redundancy. JND refers to the maximum distortion which cannot be perceived by the human eyes. Knowledge on JND no doubt can guide the image processing algorithms and systems. JND can be adopted in designing visual quality evaluation metric for images and videos , which are consistent with the HVS properties to achieve higher coding efficiency. The watermarking process for copyright protection is similar to that illustrated in Fig. 1. 3.2. Medical application scenario For decades, inter-physician medical consultation played an important role in diagnosis. The spread of digital technologies and communication networks lead to the growth of teleconsulting and telediagnosis systems. Teleconsulting is an emerging area of work with promising prospective. It is specifically useful with elderly patients, the need of international consultation, cases of emergency and disasters . . . etc. The field of teleconsultation entails several requirements such as security, privacy control, source authentication and high image quality. Our developed watermarking system,
108
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
Fig. 1. Medical watermarking system modules.
aims at fulfilling such demands through the utilization of a biometric authentication code (sender physician’s iris code), encrypted patient data and a ROI segmentation algorithm. The choice of the sender’s iris signature corresponding to iris texture as an authentication code was based on the criteria for assessing the suitability of any trait for use in biometric authentication. The criteria involves universality, uniqueness, permanence, measurability, performance, acceptability and circumvention. Encrypted patient will involve patient metadata and the suspected diagnosis. The inclusion of such data will result in a self contained record within the medical image and help protect patients’ privacy. As will be described in the following sections, new fuzzy-based ROI segmentation algorithms are designed and used for identification of the areas which need to be left intact for further inspection by the receiving physician. Figure 1 shows the system modules and watermarking process flow for the application of teleconsulting. 4. Proposed watermarking system A detailed description of the system’s modules will be given in this section. The presented system’s sender’s section can be divided into three main modules: Watermark generation and encoding module, fuzzy-based ROI segmentation module and embedding module. The proposed model integrates forward DCT transform with enhanced fuzzy-based ROI regions segmentation. The proposed scheme merges partial supervision with spatial locality to increase conventional FCM’s robustness. To comply with the imperceptibility and high image quality requirements, the coefficients selection de-
cision depends on perceptual visibility threshold estimation, based upon characteristics of the human visual system (HVS). A just-noticeable distortion (JND) profile is computed for the DCT coefficients. It is proved to outperform the DCT model, with the major contributions of a new formula for luminance adaptation adjustment and the incorporation of block classification for contrast masking. A more accurate base threshold was based on the HVS visibility threshold for digital images. A block classification has been utilized to differentiate edge regions and thus effectively avoid overestimation of JND in the said regions. Moreover, several enhanced fuzzy-based clustering models have been suggested for extraction of ROI regions in medical patterns. The receiver’s section comprises the extraction process described below and a matching module using a matching measure described in Section 5. 4.1. Systems watermarks generation and encoding A binary image is used as a copyright logo to be embedded in our first application scenario. The image is handled bit by bit and no further processing is applied on the image. The sender’s physician iris code is used for the purpose of source verification. In the generation step of the iris code, we assume the availability of an iris image specifically taken for the purpose of iris recognition. The iris code generating procedure can be divided into three steps: segmentation, normalization and feature encoding. Firstly, an automatic segmentation algorithm was presented, which would localize the iris region from an eye image and isolate eyelid, eyelash and reflection areas. Automatic segmentation was achieved through the use of the circular Hough transform for localizing the
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
(a)
109
(b)
(c)
Fig. 2. (a) Sample iris image (b) segmented iris (c) normalized iris.
iris and pupil regions, and the linear Hough transform for localizing occluding eyelids. Thresholding was also employed for isolating eyelashes and reflections. Next, the segmented iris region was normalized to eliminate dimensional inconsistencies between iris regions. This was achieved by implementing a version of Daugman’s rubber sheet model, where the iris is modeled as a flexible rubber sheet, which is unwrapped into a rectangular block with constant polar dimensions. Finally, features of the iris were encoded by convolving the normalized iris region with 1D Log-Gabor filters and phase quantizing the output in order to produce a bit-wise biometric template. Figure 2 shows a sample of a segmented iris. Patient data that can be used to assist in diagnosis are encrypted to protect the privacy of the patients and increase the system security. The encryption technique applied is a combination of the two classical approaches namely substitution and transposition. Hill Cipher is used for the purpose of substitution encryption. In the Hill cipher, linear transformation is utilized to create the cipher text from the plaintext. The cipher text row vector Y is computed as: Y = XK mod m
(1)
where X is the plaintext row vector, K is an n × n key, matrix where kij ∈ Zm in which is ring of integers modulo m where m is a natural number that is greater than one. Then, the resulted ciphertext row vector is encoded into alphabets of the main plaintext. The value of the modulus m can be optionally selected. The ciphertext Y is decrypted as: X =YK
−1
mod m
(2)
All operations are performed over Zm . The key matrix K should be invertible or equivalently the gcd (det K (mod m), m) should be 1. A double transposition scheme is employed to enhance the encryption robustness against attacks. 4.2. ROI segmentation In the medical application scenario, an ROI extraction scheme is needed to preserve image quality and to keep the ROI intact. The ROI of the image is segmented using several variants based on FCM clustering. A semi-supervised FCM algorithm (SS FCM) was introduced by [12] which allows the incorporation of expert knowledge into the FCM model. An expert defines a set of crisp labeled pixels that guide the clustering process. We denote the partially labeled pixels as: P = {p11 , . . . , p1n1 , p21 , . . . , p2n2 , pc1 , . . . , pcnc |pu1 , . . . , punu } = pl ∪ pu
(3)
The superscripts represent the class numbers with maximum c classes specified by the training data, u represents unlabeled pixels and ni denotes the number of pixels belonging to class i. The total number of pixels equals n where n = nl + nu . The fuzzy c-partition matrix of P has the following form: u U = {uuik } U u = {ui k u } | (4) labeled unlabeled U u is initialized randomly while the membership values in U l are hard labeled beforehand. The provided
110
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
σ2 =
1 I(i, j)2 − μ2 nm
(7)
i,j∈N B
where i and j belong to n × m neighborhood around each pixel of image I. An estimate for each pixel in filtered image FI is computed as follows: σ2 − ω (I(r, c) − μ) (8) σ2 where ω is the average of all the local estimated variances. Another adopted approach, is to create weightedsum intensity image using LAWS level mask [13]. It is used as the input image to SS FCM clustering algorithm creating LAWS SS FCM variant. The weightedsum image (WI) is computed by convolving the original N × M intensity image (I) by LAWS 5 × 5 normalized level mask (NL5L5). The 2D convolution is given by: F I(r, c) = μ +
Fig. 3. A Brief summary of SS FCM algorithm.
ulik are used to compute an initial set of “well seeded” cluster prototypes u01→3 nl
k=1 vi,0 = nl
(ulik )m plk
k=1
(ulik )m
WI(r, c) = (5)
Several parameters need to be set such as the maximum allowable number of iterations Tmax , m which controls the fuzziness of the resulting partition and ε which is the minimum difference required between iterations. The algorithm can be summarized as shown in Fig. 3. If n1 = 0 or w = 0, it reduces to conventional FCM. The applied modifications aim to increase robustness to noise. Also, yield more homogeneous regions with less spurious blobs. The presented work was also motivated by the extensive study of incorporating spatial information into the FCM clustering model and its promising results. The modifications applied all aim at including the effect of the local neighborhood and allowing it to influence the member ship value of each pixel. The first modification was to apply 2D adaptive noise removal filtering as a preprocessing step to the semi-supervised FCM (SS FCM) clustering process. Two-Dimensional low pass Wiener filter ... was used to produce W SS FCM variant. Wiener filter performs pixel-wise adaptive filtering based on statistics estimated from a local neighborhood of each pixel. It is specifically useful in case of an intensity image that has been degraded with constant power additive noise. 2D Wiener filter estimates the local mean and variance around each pixel as in Eqs (6) and (7): 1 I(i, j) (6) μ= nm i,j∈N B
r+2
c+2
N L5L5(i, j) ∗ I(i, j) (9)
i=r−2 j=c−2
Where
⎡
1 ⎢4 1 ⎢ ⎢6 NL5L5 = 36 ⎢ ⎣4 1
4 16 24 16 4
6 24 36 24 6
4 16 24 16 4
⎤ 1 4⎥ ⎥ 6⎥ ⎥, 4⎦ 1
r = 0,1,2, . . . N, c = 0,1,2, . . .. M and ∗ denotes 2D convolution. The resultant image pixel values represent a weighted summation of the values of the neighborhood pixels in 5 × 5 cardinality. The influence of the neighborhood pixels is inversely proportional to their distance to the center pixel in a radial manner. This approach targets reducing the effect of noisy pixels on segmentation but it also leads to blurring of some details. Moreover, one of the promising methods of incorporating spatial locality S FCM was presented by [14]. We integrated this spatial model with the SS FCM model into a Hybrid FCM (H FCM) algorithm. The main change to the previous SS-FCM is the update function of the membership values. The function accounts for the membership values of the neighborhood pixels. A spatial function defined as: uij (10) suik = j∈N B(pk )
NB (pk ) represents a square 5 × 5 window centered at pixel pk .
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
111
Fig. 5. DCT block classification and scanning order.
Fig. 4. A Brief summary of H FCM algorithm.
The spatial function is included within the membership function using the following formula: uuik suik u u j=1 ujk sjk
uuik = c
(11)
The clustering process comprises two passes. In the first pass the membership function uuik is calculated as in Eq. (5) in the spectral domain. The spatial function is incorporated into uuik in the second pass as in Eq. (10) to reflect the mapping of the pixels into the spatial domain. A summary of the described H FCM is shown in Fig. 4. After the convergence of the clustering process, a defuzzification step is required to provide hard labels to the clustered data points (pixels). Maximum membership method is used to obtain our segmented image. 4.3. Embedding processes After segmentation of the medical image ROI, we apply forward DCT transform to the non-ROI and background pixels. To comply with the imperceptibility and high image quality requirements, the coefficients selection decision depends on perceptual visibility threshold estimation, based upon characteristics of the human visual system (HVS). A JND profile is computed for the DCT coefficients using the model described in [15]. Zhang et al. [15], presented an enhanced scheme for estimating just-noticeable distortion (JND). It is proved to outperform the DCTune model, with the major contributions of a new formula for luminance adaptation adjustment and the incorporation of block classification for contrast masking. A more accurate base threshold was based on the HVS visibility threshold for digital images. The HVS visibility thresholds exhibit an ap-
proximately parabolic curve versus gray levels. Moreover, block classification has been utilized to differentiate edge regions and thus effectively avoid overestimation of JND in the said regions. Experiments have been conducted on different images and the associated subjective tests revealed improved performance of the proposed scheme over the DCTune model for luminance adaptation (especially in dark regions) and masking effect in edge regions. The selection of the DCT coefficients is based on a criterion driven from the generated JND profile. Embedding Scheme: Given: Cover Image I, size N × M and watermark wm of length k – Divide I into 8 × 8 blocks and apply forward DCT to obtain DCT coefficients C(n, i, j) – For all C(n, i, j), C(n, i, j) ∈ Middle Frequency subband: Compute JND(n,i,j), MxD(n,i,j) and MnD(n,i,j) where n: block index, i, j are coefficient indices within a block, JND(n,i,j) represents Just Noticeable Difference threshold, MxD(n,i,j) represents the difference between C(n, i, j) and Max(n,i,j) which is the maximum coefficient of a local window of size k × k centered at C(n, i, j) and MnD(n,i,j) represents the difference between C(n, i, j) and Min(n,i,j) which is the minimum coefficient of a local window of size k × k centered at C(n, i, j). – Repeat until k = length(wm) – Scan the middle frequency coefficients C(n, i, j) in the order specified in Fig. 5 – Embed watermark bits based on the following condition: if wm(k) = 1 if MxD(n,i,j) + Δ JND(n,i,j) Set C(n, i, j) = Max (n,i,j) + Δ Update Key. Next k
112
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
Sagittal
Coronal
Transversal
Fig. 6. Brain sections.
else Next C(n,i,j). if wm(k) = 0 if MnD(n,i,j) + Δ JND(n,i,j) Set C(n, i, j) = Min(n,i,j) − Δ Update Embedding Key Next k else Next C(n,i,j). – Apply Inverse DCT to obtain Watermarked Image.
of the proposed watermarking technique has been compared with Patra et al. watermarking model [6] and Badran et al. medical watermarking technique. All the experiments were run on Intel Core i5 2.53 GHz processor with 4.0 GB RAM. Experiments have been conducted on various data sets. Results are compared using different performance measures (as shown below).
4.4. Extraction process
5.1. Test datayset
Given: Watermarked Image I ∗ , size N × M and Embedding Key
The suggested modifications on standard FCM clustering are compared to evaluate their relative efficiencies and decide the most suitable algorithm for ROI segmentation in the medical application scenario. The proposed clustering techniques are evaluated using datasets which are divided into three categories, Real MR Images, MR phantoms and lung CT scans. The test set comprises hundred real MR images of patients with brain tumor which is the ROI in the experiments. Additionally, pathology free MR phantoms [16–18] of twenty subjects. A degree of abstraction was applied to segment the phantoms into three layers (clusters) this was obtained by combining several types of tissues into a single layer. The ROI comprises the gray matter and white matter. The remaining tissues are deemed as non ROI and the background pixels as BG. Three sections namely transversal, sagittal and coronal sections are taken for each subject, as shown in Fig. 6. The phantoms are simulated using BrainWeb simulator [16– 18]. The main advantage of experimenting with MR phantoms is the availability of a segmentation ground truth to be able to judge the segmentation accuracy. The images used for testing are 256 × 256 grayscale images. 50 Lung CT scan images of size 512 × 512 are used in the testing phase. The used images contain nodular lesions considered as our ROI. The datasets used to evaluate the performance of the proposed watermarking techniques are divided into
– Divide I into 8 × 8 blocks and apply forward DCT to obtain watermarked DCT coefficients C* (n,i,j). – For all C*(n,i,j), C*(n,i,j) ∈ Middle Frequency subband: Compute MxD*(n,i,j) and MnD*(n,i,j) – Using the given embedding key: if MxD*(n,i,j) = 0 Set wm*(k) = 1 Next k if MnD*(n,i,j) = 0 Set wm*(k) = 0 Next k 5. Experimental results Experiments have been carried out to validate the efficiency of the proposed watermarking techniques. Also, the studied modifications on standard FCM clustering are evaluated and compared to be able to select the best algorithm for ROI segmentation in medical application scenario. In this section, the performance of the proposed modified FCM clustering models has been compared with FCM, S FCM and SS FCM clustering algorithm. Also, the characteristics of the proposed technique are illustrated with several data sets. The performance
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
113
(a)
(b)
(c)
Fig. 7. Segmentation results of H FCM on sample images.
three categories: General benchmark images Real MR Images and lung CT scans. Three general benchmark images (Lena, Baboon and Airplane) were used to verify the systems performance in the first application scenario. To validate systems efficiency with medical images, 50 real MR images of patients with brain tumor as ROI were used. In In addition to, 50 Chest CT images which contain nodular lesions that are considered as ROI. The images used for testing are 512 × 512 grayscale images [19]. The application scenarios described demand different watermarks to be embedded. The watermark used for copyright protection is a binary image of a panda logo of size 64 × 64. In the medical application scenario, the generated various iris code lengths are used and the patient data is 50 characters long. Figure 7 illustrates sample of original cover images. A wide set of experimental biometric watermarks, of various sizes and various structural characteristics were used in our experiment. A sample is demonstrated in Fig. 8. 5.2. Performance measures Several validation functions are used to evaluate the clustering algorithms performance. Two of the most commonly used functions are partition coefficient index (pcIndex) and partition entropy index (peIndex). They measure the degree of fuzziness of partitions. They are defined as: n c u2ij pcIndex = k=1 i=1 (12) n
and
n
c
[uij log uij ] (13) n Another measure that is used is the Xie and Beni validity function [20]. It is based on the feature structure which is an advantage over the previous methods. It is defined as: n c − k=1 i=1 uik pk − vi 2 xbIndex = (14) n ∗ min{vj − vi 2 } peIndex =
−
k=1
i=1
i=j
uik presents the degree of association or membership function of the i-th data point to the k-th cluster. Segmentation accuracy (SA) is the most direct measure to evaluate clustering algorithms as segmentation algorithms. It can be only applied when the ground truth of segmentation is known. SA is defined as: #correctly classified pixels SA = (15) Total # pixels In order to verify watermarking system’s performance, two main aspects were evaluated. First, the cover image visual quality degradation has to be assessed. Peak Signal to Noise Ratio (PSNR) is used to compare the quality between two images in dB. It is given by: PSNR = 10 log10
1 MN
M N i=1
2552
j=1
(16)
[I(i, j) − I ∗ (i, j)]2
where I denotes the cover host image of size N × M and I ∗ represents the watermarked image.
114
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
Fig. 8. Segmentation results of the illustrated algorithms under Gaussian(b) and Salt&Pepper (c). Table 1 Average results of FCM and its variations applied on the MR phantoms dataset pcIndex 0.9269 0.9289 0.9600 0.9299 0.9250 0.9730
FCM SS FCM S FCM W SS FCM LAWS SS FCM H FCM
peIndex 0.0619 0.0607 0.0454 0.0661 0.0637 0.0405
xbIndex −0.0821 −0.0812 −0.0486 −0.0810 −0.0819 −0.0390
The second performance measure used is Tampering Assessment Function (TAF) to calculate the difference between the embedded and the extracted watermark. Given a watermark of length k, TAF in percentage is defined as: k
∗ TAF(%) = 1/k wm(k) ⊕ wm (k) × 100 (17) i=1
where wm(k) and wm*(k) represent the original and extracted watermarks at position (k) respectively, and ⊕ is an exclusive-OR operator. 5.3. Results Different experiments have been carried out to test and validate the proposed model using the described data sets. The performance indicators described above were used to provide statistical evaluation of performance. Given the specified measures, better performance is indicated by maximization of pcIndex as well as segmentation accuracy and minimization of peIndex and xbIndex. All experimental results are averaged over
Table 2 Average results of FCM and its variations applied on the real brain images dataset FCM SS FCM S FCM W SS FCM LAWS SS FCM H FCM
pcIndex 0.8992 0.9216 0.9424 0.9097 0.8997 0.9534
peIndex 0.0934 0.0816 0.0458 0.0804 0.0863 0.0437
xbIndex −0.1066 −0.1029 −0.0699 −0.1026 −0.1067 −0.0690
Table 3 Average results of FCM and its variations applied on the lung CT Scans dataset FCM SS FCM S FCM W SS FCM LAWS SS FCM H FCM
pcIndex 0.9003 0.8807 0.9504 0.8959 0.8940 0.9608
peIndex 0.0750 0.0739 0.0476 0.0729 0.0767 0.0424
xbIndex −0.1563 −0.1650 −0.0918 −0.1613 −0.1658 −0.0876
5 runs. Experimental study has been applied for parameter settings. After experimental analysis, the parameters Tmax, m, wk , nl and ε (described in Section 4) are set to 100, 2, 1, 150 and 0.0001, respectively. In Tables 1, 2 and 3, the performance of the proposed algorithms is compared to standard FCM, SS FCM and S FCM for different datasets as specified. As observed from the tables, there is slight variation in the performance; however, on average the following observations can be made. The application of adaptive filtering in W SS FCM leads to slight enhancement in compared to standard FCM and SS FCM mechanisms. LAWS SS FCM has resulted in some deteriora-
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
115
Table 4 Average segmentation accuracy of FCM and its variations applied on different MR phantoms FCM SS FCM S FCM W SS FCM LAW SS FCM H FCM
Set-1 traversal 0.8676 0.8698 0.8719 0.8690 0.8686 0.8848
Set-1 coronal 0.8650 0.9242 0.9290 0.9252 0.9239 0.9445
Set-1 sagittal 0.9324 0.9345 0.9272 0.9349 0.9306 0.9579
Average 0.8883 0.9095 0.9094 0.9097 0.9077 0.9291
Fig. 9. Average segmentation accuracy vs different gaussian noise variances.
tion in fuzziness and compactness of the resulting partitions; this is due to the blurring effect of the applied mask. The results also show a promising improvement in performance for H FCM algorithm over all the other algorithms. H FCM offers about 5.9% increase in pcIndex and around 43% decrease in both peIndex and xbIndex. H FCM provides a small but yet considerable advancement in performance over S FCM. This can be explained by the fact that the hard labeled pixels reduce the fuzziness and increase the compactness of
partitions. The inclusion of spatial distribution within H FCM allows the effect of the training pixels to propagate. Table 1 illustrates the results of the discussed algorithms when applied on MR phantoms. S FCM and H FCM show the best results. H FCM outperforms S FCM with an average improvement of 10.2% with the used indicators. Table 2 presents the discussed algorithms results on Real Brain MR images. The ranking of the algorithms is nearly the same as the MR
116
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
Fig. 10. Average segmentation accuracy vs different salt& pepper densities.
Fig. 11. Sample of cover images (a) brain mr images (b) chest ct scans (c) benchmark images.
phantoms results with slight degradation in the relative performance of H FCM. Table 3 illustrates the results applied on Lung CT scans. H FCM still exhibits the highest performance. It manifests good performance as it overcomes the problem of equating cluster sizes. Table 4 demonstrates the segmentation accuracy of the explained algorithms on various MR phantoms sections. Segmentation accuracy shows different results in terms of ordering the relative performance of the clustering algorithms. H FCM still presents the highest performance with 4.6% increase in average accuracy over FCM while S FCM segmentation accuracy fell slightly below the proposed algorithms. In the pro-
posed algorithms, the cluster centers are better defined than in case of S FCM for two reasons. All of the proposed algorithms present better results in terms of accuracy relative to conventional FCM. First, initialization is based on prior knowledge unlike random initialization. The second reason is that during center update the labels of the training pixels are unaltered thus allow centers to remain “well seeded”. Also, the provided supervision reduces the tendency of unsupervised FCM to equate cluster volumes. The discussed results revealed that H FCM provides the best performance within the presented algorithms. Figure 9 shows the results of applying H FCM. Figure 9(a) shows real MR image with brain tumor,
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
117
Fig. 12. A sample of segmented irises of different textures and sizes.
(a)
(b)
(c)
Fig. 13. (a)Cover images, (b) watermarked images without attack (c) watermarked image with 10% cropping attack.
Fig. 9(b) MR phantom and Fig. 9(c) Chest CT scan. Figure 9(c) displays the extraction of nodular lesion in the image, it is clear that H FCM does not suffer from the limitation of equating the cluster sizes since it extracted a very small ROI. It is also evident through visual analysis that it provides highly acceptable results. Further investigations were conducted to validate its robustness against noise. MR Phantoms were used in the coming section of our results. Two types of noise were added which are Gaussian noise and salt & pepper with varying strength. Figure 10 shows the segmentation results of the illustrated algorithms when adding different types of noise. Figure 10(b) shows results when adding uniform Gaussian noise with variance 0.1 while Fig. 10(c) when adding 10% salt and pepper noise. The ground truth is provided based on
the given segmented images in to enable comparison. Visually, it is clear that H FCM provides the highest performance compared to other algorithms. Standard FCM and Spatial FCM give very poor results with salt and pepper noise. Figure 11 displays the segmentation accuracy against different levels of Gaussian noise. It is evident H FCM performance is stable with increasing levels of noise while FCM and SS FCM accuracy is degraded rapidly. S FCM manifests similar results to these of H FCM with almost 0.2% worsening in results. Since the added Gaussian noise is uniform with coherent gray levels, the consideration of spatial locality affects positively the algorithms’ output. The addition of salt and pepper noise reduces remarkably the efficiency of S FCM as illustrated in Fig. 12. while H FCM shows noticeable robustness till 8% salt and pepper noise. Salt and Pepper noise can be considered as spike noise relative to the gray levels of MR phantoms. S FCM treats salt noise as a separate cluster thus exhibit low segmentation accuracy. In H FCM clustering centers are guided by expert help thus avoid the effect of randomness and floating centers position which lead to S FCM results. Different experiments have been carried out to test and validate the proposed watermarking model using the described data sets. The performance indicators described above were used to provide statistical evaluation of performance. The quality of the extracted watermark is evaluated using TAF and the quality of the watermarked image is assessed using PSNR. A lower TAF value would indicate that the extracted watermark is more similar to the original watermark and a higher PSNR indicates lower degradation. All experimental
118
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
(a)
(b)
Fig. 14. Comparison with Patra et al. experimented over benchmark image datasets (a) TAF vs Different Attacks (b) PSNR vs. Different Attacks. (Cropping 10%, Brightening 110%, Sharpening by 50%, Rotating the image by 2◦ , Scaling the image 512>256>512, Gaussian noise v 0.001, Salt and pepper d 0.01, Speckle noise v 0.04)
(a)
(b)
Fig. 15. Comparison with Badran et al. experimented over the medical datasets (a) TAF Vs Different Attacks (b) PSNR Vs Different Attacks. (Cropping 10%, Brightening 110%, Sharpening by 50%, Rotating the image by 2◦ , Scaling the image 512>256>512, Gaussian noise v 0.001, Salt and pepper d 0.01, Speckle noise v 0.04).
results are averaged over 5 runs and present an average of the results of the described datasets within each application. Experimental study has been applied for parameter settings. After experimental analysis, the parameter Δ is set to 3. For the parameter α is set to 1 to provide a compromise between PSNR and TAF. The system’s performance is presented against several attacks and in case of no attack. The studied attacks are (i) cropping of a block size of 10% in the middle of the watermarked image; (ii) brightening the watermarked image to 110%; (iii) sharpening the watermarked image by 50% (iv) rotating the image by 2◦ (v) scaling the image 512 > 256 > 512 (vi) adding
gaussian noise to the entire watermarked image with variance 0.001 (vii) adding salt and pepper noise to the entire watermarked image with density 0.01 (viii) adding speckle noise to the entire watermarked image with variance 0.04. Moreover, the performance is evaluated for varying JPEG compression quality factors. The TAF values of the extracted watermarks are evaluated against different levels of cropping and brightening attacks and different payload sizes. Figure 13 presents sample images after watermarking without attack and after cropping attack with the PSNR values shown. The proposed watermarking technique is compared to Patra et al. [6] in Fig. 14. Figure 14(a) illustrates the
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
119
Fig. 16. TAF vs various JPEG compression Q factor.
Fig. 17. PSNR vs various JPEG compression Q factor.
Fig. 18. TAF vs different cropping percentages.
Fig. 19. TAF vs different brightening percentages.
TAF value for the same attacks. The proposed technique displays higher robustness to attacks compared to [6] with considerable enhancement except in the case of speckle noise. This can be explained by the fact that the embedding is done in the middle frequency coefficients which provides more robustness against attacks. Figure 14(b) demonstrates the PSNR of watermarked image against different attacks. It is evident that the proposed technique imposes lower distortion to the watermarked image compared to [6] due to the incorporation of the JND-DCT model [15] into the proposed framework. It presents rather stable performance against the described attacks. The minimum PSNR reached is 21.92dB with speckle noise where the performance deteriorated compared to Patra et al. [6] with a slight drop of approximately 2dB. For the medical application scenario, the model’s results are evaluated against Badran et al. The proposed algorithm unlike Badran et al. algorithm is semi-blind
and is independent on parameter α. Figure 15(a) displays the TAF values for both algorithms and Fig. 15(b) shows the PSNR averaged over the medical datasets. The proposed algorithm exhibits better performance in terms of both watermarked image visual quality and the extracted watermark distortion. Cropping and scaling attacks manifest different results as the proposed algorithm performance is noticeably lower due to the DWT localization characteristic. The block based DCT transform provides better localization than global DCT transform this explains that the gap in performance is not large. The highest TAF value of the algorithm is 14.68% in case of speckle noise at which the watermark is still recognizable. Any attack is detected through the TAF values; this is required for our applications to be able to take any needed actions. Figure 16 presents the TAF value of the extracted watermarks against varying JPEG compression quality factors (Q factors) from 50 to 90.
120
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model
6. Conclusions
Fig. 20. PSNR Vs Different Payloads.
The algorithm’s performance is relatively robust with a maximum of 14.09% with Q factor 50. The performance is consistent on the described datasets with slight improvement in favor of the medical dataset (CT Scans and MR Images). Patra et al. [6] model exhibits the best performance due to embedding in the low frequency coefficients and the compliance with the JPEG compression zigzag scan order. The proposed algorithm provides higher robustness to JPEG compression relative to Badran et al. scheme. It is manifested by about 30% lower TAF with varying Q factors. Figure 17 displays the PSNR of the datasets against varying JPEG compression Q factors. The system’s performance is stable until Q factor 70 with acceptable average PSNR of 37.28 dB. Figuer 18 shows the TAF values against varying cropping attack. The results clarifies abrupt decline in performance with increasing percentages of cropping attack. Application Scenario 2 manifests slightly better performance than Scenario 1. Figure 19 illustrates a different pattern of gradual degradation in performance relative the brightening ratios. Scenario 2 still maintains lower distortion levels of the extracted watermark compared to Scenario 1. Figure 20 examines the response of system to varying payload sizes in terms of PSNR The system demonstrates high values of PSNR The medical dataset has higher PSNR when compared to general benchmark images. A difference of approximately 10 dB in case of 10 Kbits payload in favor of medical dataset is shown. The results show acceptable levels of degradation in terms of watermarked image quality and the matching between the extracted watermark and the embedded watermark.
In this paper, a novel multipurpose watermarking system is proposed. The described system aims to satisfy several specifications of general watermarking systems such as imperceptibility, capacity and robustness against attacks. The inclusion of a perceptual model based on HVS characteristics helped to satisfy the imperceptibility requirement. The capacity demand was considered in the selection of the parameter Δ through experimental analysis. The embedding of the watermark in the DCT middle frequency coefficients using zig-zag order aspired to achieve robustness. Moreover, an additional constraint with medical images was to separate ROI to prevent any distortion to it. H FCM was used to attain this task. Also, the utilization of biometrics based personal watermarking methods has added an emphasis on security. The provided results illustrate promising results compared to the other algorithms used for comparison. References [1]
I.J. Cox, J. Kilian, F.T. Leighton and T. Shamoon, Spread Spectrum Watermarking for Multimedia, IEEE Transactions on Image Processing 6 (1997), 1673–1678. [2] L. Masek, Recognition of Human Iris Patterns for Biometric Identification, Bachelors, 2003. [3] F. Cao, H.K. Huang and X.Q. Zhou, Medical image security in a HIPAA mandated PACS environment, Computerized Medical Imaging and Graphics: The Official Journal of the Computerized Medical Imaging Society 27 (2003), 185–196. [4] G. Coatrieux, H. Maitre, B. Sankur et al., Relevance ofWatermarking in Medical Imaging, in: IEEE EMBS Conference on Information Technology Applications in Biomedicine, Arlington, VA.Nov 2000. [5] E.F. Badran, M. Sharkas and O. Attallah, Multiple watermark embedding scheme in wavelet-spatial domains based on ROI of medical images, in: National Radio Science Conference, 2009, pp. 1–8. [6] J.C. Patra, J.E. Phua and C. Bornand, A novel DCT domain CRT-based watermarking scheme for image authentication surviving JPEG compression, Digital Signal Processing 20 (2010), 1597–1611. [7] L. Ghouti, A. Bouridane, M.K. Ibrahim and S. Boussakta, Digital Image Watermarking using Balanced Multiwavelets, IEEE Transactions on Signal Processing 54 (2006), 1519– 1536. [8] A. Phadikar, S.P. Maity and B. Verma, Region based QIMdigital watermarking scheme for image database in DCT domain, Computers & Electrical Engineering 37 (2011), 339–355. [9] V.M. Potdar, S. Han and E. Chang, A survey of digital image watermarking techniques, in 3rd IEEE International Conference on Industrial Informatics. INDIN 2005. [10] V. Solachidis and I. Pitas, Circularly Symmetric Watermark Embedding in 2-D DFT Domain, IEEE Transactions on Image Processing 10 (2001), 1741–1753.
S.M. Youssef et al. / A new hybrid fuzzy biometric-based image authentication model [11]
[12] [13] [14]
[15]
P. Fakhari, E. Vahedi and C. Lucas, Protecting patient privacy from unauthorized release of medical images using a bioinspiredwavelet-basedwatermarking approach, Digital Signal Processing 21 (2010), 433–446. A.M. Bensaid, L.O. Hall, J.C. Bezdek and L.P. Clarke, Partially Supervised Clustering for Image Segmentation, Pattern Recognition 29 (1996), 859–871. K.I. Laws, Textured image segmentation, Phd, Los Angles, California, 1980. K.-S. Chuang, H.-L. Tzeng, S. Chen et al., Fuzzy c-means Clustering with Spatial Information for Image Segmentation, Computerized Medical Imaging and Graphics 30 (2006), 9– 15. X.H. Zhang, W.S. Lin and P. Xue, Improved estimation for just-noticeable visual distortion, Signal Processing 85 (2005), 795–808.
[16]
[17] [18]
[19] [20]
121
R.C. Ra´ul, F.U. Claudia and T.B.G.d. J, Data Hiding Scheme forMedical Images, in 17th International Conference on Electronics, Communications and Computers (CONIELECOMP’07), 2007, pp. 32–37. B. Aubert-Broche, A.C. Evans and L. Collins, Anewimproved version of the realistic digital brain phantom, NeuroImage 32 (2006), 138–145. B. Aubert-Broche, M. Griffin, G.B. Pike et al., Twenty New Digital Brain Phantoms for Creation of Validation Image Databases, IEEE Transaction on Medical Imaging 25 (2006), 1410–1416. https://public.cancerimagingarchive.net/ncia. X.L. Xie and G.A. Beni, Validity Measure for Fuzzy Clustering, IEEE Transactions on Pattern Analysis and Machine Intelligence 13 (1991), 841–847.
