Discriminative binary feature learning and

Pattern Recognition 77 (2018) 289–305

Contents lists available at ScienceDirect

Pattern Recognition journal homepage: www.elsevier.com/locate/patcog

Discriminative binary feature learning and quantization in biometric key generation Amir Anees, Yi-Ping Phoebe Chen∗ Department of Computer Science and Information Technology, La Trobe University, Melbourne, Australia

a r t i c l e

i n f o

Article history: Received 3 May 2017 Revised 6 November 2017 Accepted 16 November 2017 Available online 20 November 2017 Keywords: Equalized local binary pattern Facial features Biometric key generation Quantization Recognition rate Statistical analysis and robustness

a b s t r a c t Biometric key generation based on facial features is far more complicated than face recognition, the problem being that there is no relevant information saved in databases for matching facial features in the key generation system, contrary to the face recognition system. In this paper, first, we develop an efficient unified framework for generating stable, robust and secure cryptography keys based on facial features, without the need to save information related to facial features in the database. Second, the facial features are extracted using a proposed equalized local binary pattern which shows promising results when simulated on standard face databases. Third, to cater for variations and provide flexibility in error tolerance, we propose a quantization scheme which not only cater for the variations, it also aids in providing security and reducing the size of the features. Fourth, a secure key generation mechanism is developed based on the facial features in which keys can be periodically updated. Fifth, the robustness and security of the generated keys are evaluated on a set of standard statistical tests comprising three requirements: randomness, weak biometric privacy and strong biometric privacy. Lastly, comparing our approach with several state-of-the-art methods, it exhibited superior performance. © 2017 Elsevier Ltd. All rights reserved.

1. Introduction There has been an enormous and exponential growth in multimedia along with its applications over the last three decades [1–3]. The main driving factor for this growth is the Internet, which offers a low cost means of communication. However, its drawback relates to security issues. To secure this means of communication, tremendous efforts have been exerted in the field of cryptography [4]. Although the origins of cryptography go back thousands of years [5], this subject has attracted a large amount of attention from mathematicians and cryptographers in the last few decades. A cryptographic system or cryptosystem encompasses two elements: an algorithm and a key [6]. At transmission, the cryptographic algorithm with a cryptographic key is applied to confidential data to transform it into an unreadable format and at the receiver, it is reinstated to its original form by applying an inverse algorithm with the same key. It is assumed that the cryptographic algorithm is public and the key is private or only known by authenticated users [7]. Therefore, once a key is compromised, the whole cryptosystem will then be compromised. The advanced encryption standard (AES) [8] is a benchmark block cryptosystem which is still unbreak∗

Corresponding author. E-mail addresses: [email protected] (A. Anees), [email protected] (Y.-P.P. Chen). https://doi.org/10.1016/j.patcog.2017.11.018 0031-3203/© 2017 Elsevier Ltd. All rights reserved.

able and widely used across all security applications. It has three versions, each with a block size of 128 bits, but with three different key lengths: 128, 192 and 256 bits. Keys of these lengths are very hard to remember on the go. Customers often save these keys on a local hard drive, only protected by a four- or six-digit password which can easily be cracked. Researchers have argued that although confidential data have been secured by sophisticated cryptosystems, the cryptosystem itself is not fully protected due to the vulnerability of the key. This e-problem can be addressed by replacing the six-digit password with a biometric template, such as the face, iris, fingerprint or voice [9]. This is known as a biometric identification problem with two stages: enrollment of biometric templates and key release at the time of identification. In the enrollment process, first, the biometrics Ub of a user are extracted to give a biometric signal Us , then from Us unique biometric features Fu are extracted using biometric sensor Ef in the presence of unwanted noise N. The unique biometric features of an object are saved in a database D that corresponds to its specific key. At the time of key release, the biometric features of the object are extracted again and matched with the one saved in the database using matcher Mf . If these features match each other, the key will be released. Biometric identification provides much better security than a six-digit password, as shown in Fig. 1. Also, instead of releasing the key, the key can be generated on the go using biometrics. The key is not saved on

290

A. Anees, Y.-P.P. Chen / Pattern Recognition 77 (2018) 289–305

Fig. 1. Key release or key generation by matching the biometric features with templates saved in the database; an extension of the biometric identification or recognition problem as shown above (also known as biometric key binding in the literature). A harder task is depicted in the lower part of the figure for generating constant keys on the go using biometric features. Also, the compromise of auxiliary information does not result in any security leak.

a local hard drive but can be generated on the go via biometric features. This can be considered as a biometric recognition problem which also has two stages: one is the enrollment of biometric templates as in biometric identification and the second is key generation. After the biometric features of an object match with the one in the database, its corresponding key will be generated using those biometric features.

2.

3.

1.1. Motivations In both of these systems, biometric identification and recognition (also known as biometric key binding in the literature), which are more or less the same in a given problem context, lack the original motivation of generating a key on the go. Just as a key which is secured by a weak password can be compromised, so too can the biometric template saved in the database. Hence, once the biometric template is compromised, the key which is based on that template will also be compromised. Research on biometric template security [10] has been conducted to address this problem, however the existing proposals make the whole system much more complex. To generate cryptography keys on the go using biometrics, it is necessary that no or irrelevant information related to biometric features and keys should be saved in the database to ensure that any compromise of this irrelevant information does not result in a security leak, as depicted in Fig. 1. In parallel, generating a constant key from variant biometric features of the same object is a very difficult task. For instance, face features of an object will vary with variations in pose, illumination, expression and low image resolution. In biometric key binding, this can be done with an error tolerance, however constant key generation using biometrics without any relevant information leads to zero error tolerance, making it a very difficult task. 1.2. Contributions of this work In this paper, we develop a framework for generating constant and robust cryptography keys using facial features. The proposed work is effective and efficient as there is no need for the face templates to be stored in the database at the time of key generation. Also, the variations are catered for by the proposed quantization technique. The contributions of this paper are summarized as follows: 1. We develop an effective and efficient unified framework for generating cryptography keys. The generated keys have strong

4.

5.

cryptographic characteristics such as robustness, randomness, security and stability. We propose the Equalized Local Binary Pattern (ELBP) for the extraction of face features with the new histogram equalization technique. A 3-bit quantization technique is developed to cater for variations in face features. The quantization technique not only caters for the variations, it also provides security and reduces the size of the features as well. It also provides greater security than the 2-bits but at the expense of more computational complexity. Some auxiliary information is recorded at the time of enrollment which will be used at the time of key generation. It is later argued that this auxiliary information does not cause any security leak. We develop a secure key generating mechanism based on facial features, such that, for a face feature F and its corresponding key K(F), the key generated for an another face F , i.e. K(F ) will be distinctly different from K(F), given F ≈ F . To evaluate the correctness and security of key generation, we analyze it using a set of standard statistical tests. These statistical tests are applied for the very first time on the cryptography generated keys to the best of our knowledge.

The rest of the paper is organized as follows: Section 2 presents the related work, Section 3 presents feature learning based on the equalized local binary pattern, Section 4 explains the proposed 2and 3-bit quantization techniques and the mechanism for generating cryptography keys, Section 5 presents the simulation results on the recognition rate, Section 6 presents the security and statistical analysis for the generated keys and Section 7 concludes the manuscript. 2. Related work In this section, we briefly review two related topics: (1) biometric key generation, and (2) face representation and feature learning for face recognition systems. 2.1. Biometric key generation The notion of using biometrics for generating cryptography keys without the need for a biometric template was first proposed by Bodo in 1994 [11]. This technique was supported by Janbandhu and Siyal [12] by integrating biometric data with the public-key infrastructure to generate a private key. These methods, however, were


not successful due to two limitations. First, these methods require biometric samples of a same user to be highly consistent which is very difficult because of the variations in biometric samples caused by the environmental and physiological factors. Secondly, only one key is associated with a biometric sample of a user. If the key is compromised then the biometric sample associated with that key is useless, especially in the case where the periodic updating of the cryptographic key is required. To address these issues, many methods which do not require biometric templates or keys to be stored in the database have been proposed in the literature [13–16]. Biometric key generation based on semisupervised data clustering is proposed in [13]. This scheme tries to optimize a niching memetic algorithm to effectively and simultaneously model both intra- and inter-user variations. In [14], the authors proposed a bit allocation algorithm for biometric discretization to allocate bits dynamically to every feature element based on a Binary Reflected Gray code. Their scheme is based on a combination of bit statistics (reliability measure) and signal-to-noise ratio (discriminability measure) in performing feature selection and bit allocation procedures. Template-free biometric-key generation by means of fuzzy genetic clustering is developed in [15]. The proposed method was evaluated on handwritten signature data and compared with related methods, and the results were very promising. A slightly different approach to binary biometrics using an analytic framework to estimate the performance curves under a Gaussian assumption is proposed in [16]. Biometric based cryptographic key generation from faces is developed in [17]. The developed method employs the entropy-based feature extraction process in which variations are catered for with Reed–Solomon error correcting codes. The technique was evaluated using 3D face data and is shown to reliably produce keys of suitable length for 128bit AES. In [18], a novel face hashing method with feature fusion for biometric cryptosystems is proposed. The proposed method is based on Helper Data Schema (HDS) and is used as the input parameter of cryptographic key generating algorithms to produce a renewable biometric crypto key. However, the presented method does not have any security analysis. To enhance security and privacy in biometrics, a changeable biometrics for appearance-based face recognition is proposed in [19]. The performance of the cases are compared when PCA and ICA coefficient vectors are used for verification and when the coefficient vectors are transformed. A US patent [20] for cryptographic key generation using the biometric data of fingerprints is proposed. Based on the features of the fingerprint, a message is created and a cryptographic key is generated by applying a message digest function to the message. Other methods of generating stable keys without the need for biometric templates are in [21,22], which extract the statistical features from the biometric samples of a user, create feature spaces, map each feature space into a binary string and concatenate all the binary strings to finally generate a cryptography key. These methods do not require biometric templates and they largely ignore the interuser variations, and as a result, they are unable to generate discriminative keys for a larger set of databases. 2.2. Face representation and feature learning Traditionally, a face recognition system comprises four stages: face detection, pre-processing, face representation and face matching [23]. The latter two stages are the most important in any face recognition system. The distinguishing features are extracted to make the face images more separable in face representation, while in face matching, effective and efficient classifiers are designed to identify different face patterns. The execution of a face recognition system is notably affected by face representation compared to face matching. Different face features extracted in the face representation of the same person vary significantly with variations

291

in illumination, pose, occlusion, expression, background and resolution. These variations decrease the probability of matching the face samples of the same person and increase the probability of matching the face samples of different persons, causing a major challenge for face recognition systems. There has been broad work on face representation in the literature, which can be grouped into two classes: holistic feature representation and local feature representation. Holistic representation is the most commonly utilized in face recognition. It depends on the lexicographic requesting of raw pixel qualities to yield one vector for every image, taking in an element subspace to preserve the measurable data of face images. Common holistic techniques are PCA [24] and LDA [25]. In contrast, local feature extraction alludes to portraying just a locale/part of the image by utilizing some transformation standard or particular estimations with the end goal that the last result depicts the fundamental image content in a way that ought to yield a unique solution at whatever point the same substance is experienced. Typical local features methods are LBP [26] and Gabor wavelets [27]. In addition to face representation, feature learning is a key process, a set of techniques that learns a feature, a transformation of an unrefined image input into a representation that can be successfully put through in machine learning. Extensive work has been conducted on feature learning in recent years, some of this being: sparse autoencoders [28], denoising auto-encoders [29] and restricted Boltzmann machine [30]. However, these real-valued features have several disadvantages when compared with binary codes, for example, they have slow computational speed, they consume a lot of memory and are not robust against local variations. In addition to holistic and local feature representations, deep learning methods [31–33] and sparse coding methods [34] for feature learning are becoming popular. In [31], an approach for mixing a de-identified face locale with its unique background is displayed with the end goal of finishing the procedure of face deidentified. It has been appeared in trial comes that the face deidentification conveys high security assurance inside the face district while mixing the de-identified face area with its unique background. A hybrid deep learning architecture for privacy-preserving mobile analytics is presented in [32]. To provide privacy based on defined privacy constraints, the authors relied on the siamese network and machine learning approaches. The experimental results confirmed the strong privacy in parallel with effectiveness for emotion detection, gender classification and face recognition. For privacy-preserving deep learning, a practical system for multiple parties is developed in [33]. The developed framework allows multiple parties to collaborate on learning an accurate neural network for the same given objective. Incorporating convolutional sparse coding from naturally occurring point trajectory corpora, a method for learning an over-complete basis is proposed in [34]. The proposed method is used to increase the probability that the Reduced Isometry Property condition holds for a broad class of point trajectories and camera motions. The authors have also proposed a one inspired objective for trajectory reconstruction. In this paper, we propose the Equalized Local Binary Pattern (ELBP) for face representation and feature learning which will be used in generating stable cryptography keys.

3. Feature learning and unified framework The facial features used for generating cryptography keys are extracted through the proposed improved version of the Local Binary Pattern (LBP) [35,36] called the Equalized Local Binary Pattern (ELBP). ELBP can resist gray scale variations in dark and bright images and can sustain the effect of sparse noise addition.

292


Fig. 2. (a)–(c) Different LBP patterns: (a) the pattern is invariant with respect to gray scale addition if the resulting value lies between [0 255], (b) the LBP pattern changes as the resulting value exceeds the mentioned range or (c) changes due to the addition of sparse noise, (d) gray scale variations in a face image by 60, the difference between the histogram features of two gray variant images is approximately 13% and (e) small sparse noise addition in a face image, the difference between these two histogram features being approximately 45%.

3.1. LBP and its drawback LBP transforms an image or image section into array of codes by encoding the signs of the pixel difference between the center pixel and its neighbor pixels. It was initially proposed for texture classification, however it has been recently used in other applications as well. For LBP8, 1 , let us consider a patch of an image containing 9 pixels in a 3 × 3 matrix. The center pixel is denoted as hc , the neighbor pixels are denoted as hp and the pixel differences are denoted as d p = h p − hc , where p = 0, 1, 2, . . . , 7. The LBP code, Cp is formed by concatenating bit codes, bp , given as [36]:

Cp =

7

b p ∗ 2 p,

(1)

p=0

where

bp =

1

i f d p ≥ 0,

0

i f d p < 0.

(2)

It was observed in [36] that most of the LBP codes in LBP8, 1 comprise fundamental characteristics of texture. These codes with fundamental characteristics are called uniform codes or patterns. The uniformity U is defined as the number of bit-wise circular transitions from 1 to 0 or vice versa. An LBP is a uniform pattern if U ≤ 2, such as “00111000” and “11100000” which have U = 2. There are 58 uniform patterns in LBP8, 1 , given as [36]:

7

C pu

=

p=0

bp ∗ 2p

P+1

i f U (LBP8,1 ) ≤ 2, otherwise,

(3)

where

U (LBP8,1 ) = |b7 − b0 | +

7

|b p − b p−1 |.

(4)

p=1

LBP is invariant to a constant gray scale change in pixel values. For instance, an addition of integer 20 to a 3 × 3 pixel matrix will

not change its LBP as shown in Fig. 2(a). However, in the case of bright or dark 3 × 3 pixel matrices, an addition/subtraction of a small number can change its LBP as depicted in Fig. 2(b). In this case, a small integer value of 20 is added to a bright pixel matrix. As a result, some of the pixels experience change in their bit codes, resulting in a different LBP pattern. Also, an addition of small sparse noise can change the pattern. A small image patch of 3 × 3 matrix usually has high autocorrelation, therefore an addition of a sparse noise as small as integer 5 can change its pattern significantly, as shown in Fig. 2(c). The effects of constant gray scale variation and the addition of a small sparse noise with the respective differences to the original face images are shown in Fig. 2(d) and (e), respectively. It is demonstrated that the differences created by these two effects are large enough to categorize their affected images as different from the original face images. The effect of noise addition is addressed in [37] by employing a uniform pattern with Noise Resistant LBP (NRLBP). NRLBP encodes the binary codes bp into three states rather than the two states of 1 and 0. The third state “Uncertain” is assigned to bp if dp is small. The uncertain bit X is then assigned as either 1 or 0 forming a uniform pattern. NRLBP shows good results in the presence of noise but the computational complexity is very high and therefore is not suitable in some applications such as generating cryptography keys on the go. Also, in addition to security, cost and performance are two main factors considered in designing cryptosystems. Key scheduling or key generation in cryptosystems has, at most, 500 gate equivalents because most computational complexity is reserved for the construction of cryptosystem’s algorithm. Therefore, it is an essential requirement to keep the computational complexity of key generation as low as possible to synchronize with the hardware platform. 3.2. Proposed equalized LBP (ELBP) We have observed that a large variation in the gray scale of an image changes its LBP pattern to either full dark spot (0 0 0 0 0 0 0 02 ) or bright spot (111111112 ). Also, if the image initially has the pixel


293

f1 denotes 0 0 0 0 0 0 0 02 , f58 denotes 111111112 and f59 denotes the non-uniform pattern (P + 1). The set of 59 equalized pattern fea tures F = [ f1 , f2 , . . . , f59 ] are calculated based on the values of features of f1 , f58 and f59 . These three features are calculated first as follows:

⎧ max( fi ) i f fk > max( fi ), ⎪ ⎪ ⎪ ⎨ ∀i = [2, 3, . . . , 57], ∀k = [1, 58, 59], fk = ⎪ ⎪ ⎪ ⎩ fk otherwise.

(5)

The rest of 56 features will be calculated as follows:

⎧ f − max( fi ) i f fk > max( fi ), ⎪ ⎨ k ∗ ∀i = [2, 3, . . . , 57], ∀k = [1, 58, 59], fk = ⎪ ⎩ 0

otherwise.

fi = fi +

fi

57

j=1

Fig. 3. Proposed equalized LBP (ELBP): (a) LBP histogram of an image, (b) LBP histogram of the same image when noise (with variations in gray scale and sparse noise) is added. In this case, two uniform (0 0 0 0 0 0 0 02 , 111111112 ) and one nonuniform features are to be equalized to the maximum value of the rest of the features. The remaining values of these three features are distributed among the other features according to the other features’ original distributions. (c) ELBP of the image whose LBP histogram is shown in (a) and (d) ELBP of the noisy image whose LBP histogram is shown in (b). The ELBP histogram of the noisy image is almost similar to the ELBP histogram of the original image.

values near the extreme range of [0 255], then even a small variation in gray scale can shift its LBP pattern to either a dark or bright spot. Moreover, the addition of sparse noise in an image mostly changes the LBP pattern from uniform (U ≤ 2) to non-uniform (U > 2). Therefore, it is seen that the gray scale and noise variations mostly contribute to three LBP patterns: 0 0 0 0 0 0 0 02 , 111111112 and the non-uniform pattern (P + 1, as mentioned in Eq. (3)). We propose that the LBP patterns should be equalized before features can be extracted. Let the 59 LBP pattern features be denoted in a set F = [ f1 , f2 , . . . , f59 ], where f1 , f2 , . . . , f58 are uniform LBP patterns,

(6)

×

fj

∗

fk

,

∀i = [2, 3, . . . , 57].

(7)

k=1,58,59

The proposed equalized technique for LBP features is shown in Fig. 3. The pattern features are only equalized if the values of fk exceed the maximum value of the rest of the features, where k = [1, 58, 59]. In this case, fk will be assigned a maximum value from

fi , being left with fk∗ = fk − max( fi ) values, where i = [2, 3, . . . , 57].

These remaining values fk∗ will then be added evenly to fi accord

ing to their original distribution resulting in fi . The equalization and LBP feature learning has much less computational complexity compared to NRLBP. Also, the results of the ELBP features in the presence of noise and gray scale variations are much better than LBP, as described in Section 5. These equalized features F are then used in the construction of a cryptography key. 3.3. Proposed framework The proposed framework comprises two algorithms, enrollment and key generation, as depicted in Fig. 4. The steps for the enrollment algorithm are defined as follows: 1. The l facial biometric samples of an object B1 , B2 , . . . , Bl are input into the enrollment algorithm. These are varied samples of an object, with variations mainly in pose.

Fig. 4. A top level block description of the proposed unified framework comprising two algorithms: enrollment and key generation.

294


2. The eyes of a biometric sample are assigned weights m as they contain more useful information than the rest of the face. The value of m should be greater than or equal to 1. In our case, m has assigned a value of 2.2 for the simulations described later. 3. Uniform LBP pattern features F = [ f1 , f2 , . . . , f59 ] are extracted for both the face and weighted eyes and a histogram of these features is then created. 4. These LBP pattern features are then equalized to obtain equal ized features F = [ f1 , f2 , . . . , f59 ] using the proposed equalization technique mentioned in Eqs. (5)–(7). 5. Steps 2–4 are repeated for each biometric sample and the mean of these features Fm = [ f1m , f2m , . . . , f59m ] is calculated (a detailed description of how to obtain the mean features is depicted in Algorithm 1 in Section 5). 6. The proposed quantization technique is applied on the mean features, taking input Fm and giving outputs of a set of quan tized features Fq and a perturbation vector P. Fq will have no use in the enrollment stage and will be discarded. At the time of key generation, the following steps are required: 1. One facial biometric sample B1 of an object will be taken as input. 2. Steps 2–4 remain the same as described in the enrollment algorithm. 3. The proposed quantization technique is applied on the equal ized features, taking inputs of F and P and giving the output of balanced features Fb . 4. From Fb , the proposed key generation mechanism is applied to obtain a stable cryptography key Ki . The quantization procedure required at enrollment and key generation is described in the next section with a description of the proposed key generation mechanism. 4. Quantization and key generation mechanisms The basic purpose of quantization for facial features is to cater for the variations. However, interestingly, it does not only remove the variations of the features, it also enhances the security of the system. In this section, 2-bit and 3-bit quantization schemes are developed. The developed quantization schemes are much less complex and much more effective than the existing quantization schemes presented in [38,39]. With quantization, extra information is also recorded, known as the perturbation vector and will be used at the time of key generation. This extra information is responsible for the security of the system. The more information which is recorded, the more secure the system will be. However, an enhancement in the security of the system with quantization also brings some complications, i.e. the more information, the more complex the system will be. Furthermore, the proposed quantization process can also be applied to other biometric features as well such as iris patterns, fingerprints, handwritten signatures, etc. to cater for the variations which are a result of the different versions of a same biometric. 4.1. Two-bit quantization scheme The 2-bit quantization procedure is depicted in Fig. 5. In this procedure, a 2-bit auxiliary information is recorded for each feature, hence this is why it is called a 2-bit quantization scheme. Quantization in the enrollment and the key generation stages are described as follows: 4.1.1. Enrollment Let us suppose fim represents one of scalar equalized mean features from set

Fm

and

fiq

is a quantized value of

fim .

Let ϕ be

Fig. 5. Illustration of 2-bit quantization procedure.

the maximum allowable change in fim due to some variations. The quantization interval is then set to ϕ (ϕ ) can take any value depending upon the application).

1. The scalar equalized mean feature fim is quantized to fiq according to the following equation:

i f (n − 1 )ϕ ≤ fim < nϕ ,

fiq = n

(8)

where, n = 1, 2, . . . is the quantized interval and a positive integer. 2. With the quantization, a 2-bit auxiliary information is also recorded for each fim in a two bit binary vector ρi = [ρi1 ρi2 ] according to the following rule:

[ ρi 1

ρi2 ] = ( fiq − 1 ) mod 4.

(9)

3. The auxiliary information ρ i recorded for each feature is concatenated one by one into perturbation vector P1 × 118 which will be used in the adjustment of the features in the key generation algorithm.

P = {ρ1 ||ρ2 || . . . ||ρ59 }, where

|| represents concatenation operator.

4.1.2. Key generation In key generation, the scalar features of an object are extracted and equalized. For simplicity and convenience, let fim be the equalized feature. are cal1. The quantized feature f and its perturbation vector ρ i

iq

culated according to Eqs. (8) and (9). 2. The scalar feature fim extracted at the time of key generation may be distorted due to some variations. To cater for these vari ations, the feature fim is modified to fim∗ with the help of P according to the following rules:

ϑi =

⎧ ⎪ ⎨0

i f [ρi1 ρi2 ] = ([ρi1

ρi2 ] − 1 ) mod 4,

1 ⎪ ⎩

i f [ρi1 ρi2 ] = ([ρi1 otherwise.

ρi2 ] + 1 ) mod 4,

2

⎧ f +ϕ ⎪ ⎪ ⎨ im fim∗ = fim −ϕ ⎪ ⎪ ⎩ fim

if

ϑi = 0,

if

ϑi = 1,

otherwise.

(10)

(11)


295

Fig. 6. (a) 3-bit quantization procedure, (b) Case I: Value distorted from f im = 36 to fim = 43, (c) Case II: Value distorted from f im = 36 to fim = 69, (d) Case III: Value distorted from f im = 36 to fim = 13 and (e) Case IV: Value distorted from f im = 36 to fim = 74.

3. The modified feature fim∗ is then quantized according to Eq. (8) to get fiq∗ . This quantized feature fiq∗ is variation free

if the feature fiq was originally varied in the specified range of ± ϕ.

will be used in the adjustment of the features in the key generation algorithm.

P = {ρ1 ||ρ2 || . . . ||ρ59 }, where

|| represents concatenation operator.

4.2. Three-Bit quantization scheme A 3-bit quantization scheme provides greater security but also increases the computational complexity of the system. The 3-bit quantization procedure is depicted in Fig. 6(a). In this procedure, a 3-bit auxiliary information is recorded for each feature, which is why it is called a 3-bit quantization scheme. Quantization in the enrollment and key generation is described as follows: 4.2.1. Enrollment Let us suppose fim represents one of the scalar equalized mean

features from set Fm and fiq is a quantized value of fim . Let ϕ be

the maximum allowable change in fim due to some variations. The quantization interval is then set to ϕ (ϕ can take any value depending upon the application). fim

[ ρi 1

ρi2 ρi3 ] = ( fiq − 1 ) mod 8.

i are cal1. The quantized feature fiq and its perturbation vector ρ culated according to Eqs. (8) and (12). 2. The scalar feature fim extracted at the time of key generation may be distorted due to some variations. To cater for these vari ations, the feature fim is modified to fim∗ with the help of P. First i as follows: the value for ϑi is determined based on ρ i and ρ

fiq

1. The scalar equalized mean feature is quantized to according to the Eq. (8). 2. For each fim , a 3-bit auxiliary information is calculated and recorded in a three bit binary vector ρi = [ρi1 ρi2 ρi3 ] according to the following rule:

4.2.2. Key generation In key generation, the scalar features of an object are extracted and equalized. For simplicity and convenience, let fim be the equalized feature.

(12)

3. The auxiliary information ρ i recorded for each feature is concatenated one by one into perturbation vector P1 × 177 which

ϑi =

⎧ 0 ⎪ ⎪ ⎪ ⎪ ⎪ 1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨2 3

⎪ ⎪ ⎪ 4 ⎪ ⎪ ⎪ ⎪ ⎪ 5 ⎪ ⎪ ⎪ ⎩ 6

i f [ρi1 ρi2 ρi3 ] = ([ρi1

ρi2 ρi3 ] − 1 ) mod 8,

i f [ρi1 ρi2 ρi3 ] = ([ρi1

ρi2 ρi3 ] − 2 ) mod 8,

i f [ρi1 ρi2 ρi3 ] = ([ρi1

ρi2 ρi3 ] − 3 ) mod 8,

i f [ρi1 ρi2 ρi3 ] = ([ρi1

ρi2 ρi3 ] + 1 ) mod 8,

i f [ρi1 ρi2 ρi3 ] = ([ρi1

ρi2 ρi3 ] + 2 ) mod 8,

i f [ρi1 ρi2 ρi3 ] = ([ρi1

ρi2 ρi3 ] + 3 ) mod 8,

otherwise. (13)

296

A. Anees, Y.-P.P. Chen / Pattern Recognition 77 (2018) 289–305 i , the value for fim∗ is Finally, having the values of ϑi , ρ i and ρ modified according to the following rule:

fim∗ =

⎧ fim + ϕ ⎪ ⎪ ⎪ ⎪ ⎪ fim + 2 ∗ ϕ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ fim + 3 ∗ ϕ fim − ϕ ⎪ ⎪ ⎪ ⎪ fim − 2 ∗ ϕ ⎪ ⎪ ⎪ ⎪ ⎪ fim − 3 ∗ ϕ ⎪ ⎪ ⎪ ⎩

fim

if

ϑi = 0,

if

ϑi = 1,

if

ϑi = 2,

if

ϑi = 3,

if

ϑi = 4,

if

ϑi = 5,

ϑi = 5. Now using Eq. (14), the modified value of scalar feature fim∗ is obtained as, fim∗ = fim − 3 ∗ ϕ = 69 − 30 = 39.

(14)

otherwise.

3. The modified feature fim∗ is then quantized according to ∗ Eq. (8) to get fiq . This quantized feature fim∗ is variation free

if the feature fim was originally varied in the specified range of ± (3∗ ϕ ). Similarly, the proposed work can be extended for four-bit quantization scheme that provides much more security and robustness to the extracted features, however with the cost of more computational complexity. 4.3. Illustrative example To have a better insight into the quantization procedure, let us take an example of 3-bit quantization in different scenarios. Let us suppose in the enrollment algorithm, an extracted scalar mean equalized feature has a value of fim = 36 with ϕ = 10. The quantized value and the auxiliary bits obtained using Eqs. (8) and (12) are fiq = 4 and ρi1 = 0, ρi2 = 1, ρi3 = 1. Now consider the following cases in which at key generation, the value of scalar mean equalized feature is changed due to some variations. Case I Consider in this case, the received value of the scalar mean equalized feature is changed to fim = 43 (as shown in Fig. 6(b)). The auxiliary bits then obtained using Eq. (12) are ρi1 = 1, ρi2 = 0, ρi3 = 0. Now according to Eq. (13),

[ρi1 ρi2 ρi3 ] = ([ρi1

therefore

ρi2 ρi2 ] + 1 ) mod 8,

1 0 0 = (0 1 1 + 1 ) mod 8 = (4 ) mod 8 = 4, therefore

Finally, the modified quantized value for the modified scalar feature is obtained using Eq. (8), i.e. fiq∗ = 4.

which is a correct value. Case III Consider in this case, the received value of the scalar mean equalized feature is changed to fim = 13 (as shown in Fig. 6(b)). The auxiliary bits then obtained using Eq. (12) are ρi1 = 0, ρi2 = 0, ρi3 = 1. Now according to Eq. (13),

[ρi1 ρi2 ρi3 ] = ([ρi1

ρi2 ρi2 ] − 2 ) mod 8,

0 0 1 = (0 1 1 − 2 ) mod 8 = (1 ) mod 8 = 1, therefore

ϑi = 1. Now using Eq. (14), the modified value of scalar feature fim∗ is obtained as, fim∗ = fim + 2 ∗ ϕ = 13 + 20 = 33.


which is a correct value. Case IV Consider in this case, the received value of the scalar mean equalized feature is changed to fim = 74 (as shown in Fig. 6(b)). The auxiliary bits then obtained using Eq. (12) are ρi1 = 1, ρi2 = 1, ρi3 = 1. Now according to Eq. (13),

ϑi = 6. Now using Eq. (14), the modified value of scalar feature fim∗ is obtained as, fim∗ = fim = 74.

ϑi = 3. Now using Eq. (14), the modified value of scalar feature fim∗ is obtained as, fim∗ = fim − ϕ = 43 − 10 = 33.

Finally, the modified quantized value for the modified scalar feature is obtained using Eq. 8, i.e.


which is a not a correct value showing that, in this case the variation cannot be corrected.

fiq∗ = 4.

4.4. Comparison with previous quantization schemes

which is a correct value. Case II Consider in this case, the received value of the scalar mean equalized feature is changed to fim = 69 (as shown in Fig. 6(b)). The auxiliary bits then obtained using Eq. (12) are ρi1 = 1, ρi2 = 1, ρi3 = 0. Now according to Eq. (13),

The developed quantization schemes are much less complex and much more effective than the existing quantization schemes [38,39]. Table 1 shows the comparative computational time (ms) which demonstrates the computational complexity and corresponds to the cost required. The comparative analysis of distortion tolerant (levels) which demonstrates the effectiveness of proposed framework applied on three different quantization schemes (2-, 3- and 4-bit) is also presented in Table 1. Comparing our approach with the state-of-the-art methods, it exhibited superior performance.

[ρi1 ρi2 ρi3 ] = ([ρi1

ρi2 ρi2 ] + 3 ) mod 8,

1 1 0 = (0 1 1 + 3 ) mod 8 = (6 ) mod 8 = 6,

A. Anees, Y.-P.P. Chen / Pattern Recognition 77 (2018) 289–305 Table 1 Comparative computational time (ms) and distortion tolerant (levels) analysis of proposed framework with [38,39] applied on three different quantization schemes (2-, 3- and 4-bit). Analysis

Computational time (ms)

Distortion tolerant (Levels)

Quantization bits

2

3

4

2

3

4

Ref. [38] Ref. [39]

3 3

5 5

8 8

1.5 1.5

1.5 1.5

1.5 1.5

Equalized LBP

1.2

3

5

1.5

3.5

7.5

the false positive rate to zero. To achieve this, the margin of difference between different biometric samples of the same subject should be as minimal as possible. Although this may decrease the possibility of recognizing an individual, i.e. the true positive rate may be reduced if the variations in its biometrics exceed the predefined threshold, it is expected that the variations would be low in constant circumstances. 5.1. Collecting mean of features

4.5. Key generation module There are 59 quantized features obtained in which 58 are uniform and one is non-uniform. The 56 features will be used for key generation ignoring ‘0 0 0 0 0 0 0 0’, ‘11111111’ and the non-uniform feature. For simplicity and convenience, we write quantized fea ture fiq∗ as fi . Let F be the set of 56 quantized equalized features, i.e. F = [ f1 , f2 , . . . , f56 ]. The features are converted into binary bits having k bits in each feature, where k is a variable that has a different number of bits for different features depending upon their values. For instance, a quantized feature with a value of 21 will have five binary bits of 10,101 while a quantized feature having a value of 15 will have four binary bits of 1111. For a face image of size 512 × 512, the average number of quantized intervals per uniform feature is approximately 400 with ϕ = 10. Therefore, the average value for k for each feature will be 9, and the overall length of the F in binary form will be far more than the 256 bits. Let p( f, w ) represent the bit-wise rotation on f for w bits. The vector F is modified to Fn by applying the bit-wise rotation operation on each feature and concatenating all the bits as follows:

F = [ p( f1 , w1 )|| p( f2 , w2 )|| . . . || p( f56 , w56 )]1×56 .

(15)

Fn = p(F , w f ).

(16)

where, wi = mod ( fi , ki ), i ∈ {1, 2, . . . , 56}, w f = mod (s f , sk ), s f = 56 56 i=1 f i , sk = i=1 ki and || represents the concatenation operator. The first 256 bits of Fn are extracted and placed in a vector Fnk . The key K will finally be generated of length 256 bits by applying bit-wise rotation operation on Fnk as follows:

K = p(Fnk , wk )1×256 .

297

(17)

where, wk = mod (s f , 256 ). The generated key K is strong with respect to cryptographic characteristics. For example, the key generated for the face feature F, i.e. K(F) is distinctly different from another key K(F ) generated for the face F , given F ≈ F ; this is the case in twins. The next two sections are based on simulations conducted on two aspects of this work: one is the ability to recognize an individual’s facial features as unique and different from the others and second is the robustness of key generation. 5. Simulation results on recognition rates The application of facial features recognition considered here is cryptography key generation which does not require much robustness. This is in contrast to a security application where the subject has to be recognized in different environments with variations in pose, illumination, expression and low image resolution. In our application, the features of a subject will be extracted in a constant illumination with minimal variations in pose and expression at the time of enrollment and key generation. However, the mandatory requirement for this application is to ensure that the same key is not generated from two or multiple different subjects, i.e. reducing

To achieve our goal, the mean features of the biometric samples are taken as input in the enrollment algorithm. At enrollment, a perturbation vector of a subject is achieved by the quantization of the mean features of its biometric samples, as shown in Fig. 4. However, only those biometric samples whose difference is less than the specified threshold compared to their reference biometric sample will be considered. Let us consider that there are M biometric samples of a subject and Ir is their reference biometric sample. The ELBP features of Ir are extracted first using function uni f orm_lbp. The features of the other biometric samples will then be extracted one by one and will be considered for contribution into the mean. The mean difference abs(ai − bi )/2 of 59 features and the maximum of these differences is recorded, where ai is a feature of Ir , bi is a feature of other biometric samples and i = [1, 2, . . . , 59]. If the maximum value of these differences exceeds the specified threshold th, then that biometric sample will not be considered. This threshold th is initially given as the difference between the ELBP histograms of two test images of the subjects. Let us consider the maximum difference between any two test images of a same subject is 10 and the minimum difference between any test image of this subject and any test image of other subjects is greater than 10. In this case, the true positive rate for this subject will be 100% and the false positive rate is 0. However, this is an ideal case, and in reality, the true positive rate might be less than the maximum value and the false positive rate might be greater than the minimum value. Our goal is to optimize th to get the best of both these worlds. The quantization threshold quan_th for a specific subject is also calculated through the mean difference of its valid biometric samples, as explained in Algorithm 1. This quantization threshold is actually equal to the quantization interval ϕ that will be used in the key generation algorithm as well as for testing the recognition rate (specificity). 5.2. Simulations considering quantization threshold & mean features (sensitivity vs specificity) The simulations considering the quantization threshold & mean features are conducted on the AT&T face database [40] which comprises 400 images of 40 subjects with 10 different images for each subject and on the face recognition database of the University of Essex [41] which has 20 images for each of the 153 subjects, making a total of 3060 images. Fig. 7 depicts a small sample of the AT&T face database showing the test images of three subjects considered in the enrollment algorithm with a low value of th. The tick symbol shows that a test image passes the enrollment algorithm and will be considered in the mean features, while a test image with the cross symbol will be discarded. This process certainly decreases the recognition rate with low th but at the same time decreases the probability of generating the same key from the test images of different subjects. Table 2 shows a confusion matrix of 10 subjects with 20 images for each subject taken from the face recognition database of the University of Essex. It can be seen that the recognition rate is on the low side but there are no test images being recognized as a different subject. For instance, in subject two, 19 test images are recognized correctly while one is not;

298


Fig. 7. A small sample of test images of three subjects taken from the AT&T face database [40]. The sample images are compared with their reference images in the enrollment algorithm with a low value of th. The tick symbol shows that a test image passes the enrollment algorithm and will be considered in the mean features while a test image with the cross symbol will be discarded.

Algorithm 1 Enrollment procedure: The test images of a subject will only contribute to the mean if and only if their difference is less than the specified threshold of the reference test image. Inputs: Test images of a subject Il (1×M ) , Reference test image Ir , Threshold th, LBP Uniform function, uni f orm_lbp(. ). Output: Mean vector of 59 features Fm ( 1 × 59 ), quantization threshold quan_th. 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13:

lbp_vec (1 : 59, 1 ) = uni f orm_lbp(Ir ) temp_vec (1 : 59, 1 ) = lbp_vec (1 : 59, 1 ) j=1 for i ← 2 : M do temp_vec (1 : 59, 2 ) = uni f orm_lbp(Il (i )) for k ← 1 : 59 do temp_vec (k, 3 ) = abs((temp_vec (k, 1 )− temp_vec (k, 2 ))/2 ) Endfor max_val = max(temp_vec (1 : 59, 3 )) if max_val < th then j = j + 1; lbp_vec (1 : 59, j ) = temp_vec (k, 2 )

23:

Endif Endfor for k ← 1 : 59 do Fm (1, k ) = mean(lbp_vec (k, 1 : j )) temp_vec (k, 3 ) = max(lbp_vec (k, 1 : j ) temp_vec (k, 4 ) = min(lbp_vec (k, 1 : j )) temp_vec (k, 5 ) = mean(lbp_vec (k, 1 : j ) mx_me = temp_vec (k, 3 ) − temp_vec (k, 5 ) me_mn = temp_vec (k, 5 ) − temp_vec (k, 4 ) quan_th = max(mx_me, me_mn )

24:

Endfor

14: 15: 16: 17: 18: 19: 20: 21: 22:

however, this one test image is also not being recognized as a different subject. The key that is generated from this test image will be unique from the keys generated from all the subjects, including its own subject. The key however is incorrect and hence the subject has to input its biometric sample again in order to generate its correct key. The results for the AT&T face database [40], on the other hand, are, however, more interesting and have large variations in expression and pose. Fig. 8(a) shows a graph of quantization threshold quan_th achieved at the enrollment stage against specified threshold th. The impact of th is on the true positive rate (sensitivity),

Table 2 Confusion matrix of 10 subjects with 20 test images for each subject taken from the face recognition database of the University of Essex [41]. The sample images are tested in the enrollment algorithm with a low value of th. The recognition rate is on the low side but there are no test images being recognized as a different subject. Predicted class

Actual class

No.

1

2

3

4

5

6

7

8

Acc.

1 2 3 4 5 6 7 8

20 0 0 0 0 0 0 0

0 19 0 0 0 0 0 0

0 0 18 0 0 0 0 0

0 0 0 17 0 0 0 0

0 0 0 0 18 0 0 0

0 0 0 0 0 19 0 0

0 0 0 0 0 0 20 0

0 0 0 0 0 0 0 18

100% 95% 90% 85% 90% 95% 100% 90%

where increasing its value also increases the true positive rate and the impact of quan_th is on the true negative rate (specificity), where increasing its value also increases the true negative rate. It was expected that Fig. 8(a) would show that with an increase of th the value of quan_th will also increase, which it does up to the value of 80. However, increasing th from this point does not further increase quan_th. As a result, the true negative rate also stop decreasing at a point where th equals 80 and then almost remains stable throughout the values, as depicted in Fig. 8(b). In this case, considering the specificity for quan_th (specificityst ), the recognition rate for both sensitivity and specificity are quite high after th = 120. The graph of the specificity for th (specificityqt ) is also plotted in Fig. 8(b) which shows excellent results for the true negative rate. However, the true negative rate decreases exponentially after th = 160 but the true negative rate for (specificityqt ) remains stable and can be deployed accordingly. 5.3. ROC and rank one recognition rate The final recognition results for the proposed method considering specificityqt in comparison to other related work in the form of rank curves are shown in Fig. 9. Fig. 9(a) (b) compares the results of the proposed method with five other methods [24,26,58,59] using the University of Essex [41] and the AT&T face databases [40] respectively, while Fig. 9(c) compares the results using the extended Yale Face Database B [42] which contains 16,128 images of 28 human subjects in 9 poses and 64 illumination conditions. The rank one recognition rates for the widely used FERET database


299

Fig. 8. Simulated results on the AT&T face database [40], (a) quantization threshold quan_th against specified threshold th; quan_th stops decreasing at th = 80, (b) recognition rate of sensitivity and specificity, both are high where quan_th became stable.

Fig. 9. ROC curves; comparison analysis of the proposed method with the existing work, applied on (a) University of Essex face database [41], (b) AT&T face database [40] and (c) extended Yale face database B [42].

[43] as a comparison with the existing work is listed in Table 3. The FERET dataset consists of 13,539 face images of 1565 subjects who are diverse across age, gender, and ethnicity. We followed the standard FERET evaluation protocol [43], where six sets including the training, fa, fb, fc, dup1, and dup2 were constructed for evaluation. All face images in these six sets are aligned and cropped into 128 × 128 pixels according to the provided eye coordinates. Besides the binary feature, deep learning is one of the growing research fields which show promising results in image processing, specifically in face recognition. Deep learning is the application of artificial neural networks (ANNs) for learning multiple tasks having more than one hidden layer. For the comparison, we take the LFW face database [60] which consists of 13,233 images of 5749 subjects in which 1680 subjects have two or more images. For evaluation, we adopt the unsupervised settings which have no training data. Instead of taking the mean of features, we use the face images for sensitivity and specificity. Let F1 and F2 be the features of image 1 and 2 respectively and f(F1 , F2 ) is the function that returns the similarity of these two features. We consider a threshold σ such that if f(F1 , F2 ) ≤ σ then both features belong to the same subject and vice-versa. By varying σ , we calculate the values of sensitivity and specificity and eventually the area under curve which is the accuracy rate. Furthermore, we cropped the images from the LFW database into the size of 64 × 64 pixels consisting of face only. It is worth mentioning here that the slight difference in the size of face images does not impact the results of accuracy

Table 3 A comparison of rank-one recognition rates (percent) with the state-of-the-art feature descriptors and the standard FERET evaluation protocol.

∗

Method

fb

fc

dup1

dup2

LBP [26] LBP + WPCA [26] LGBP [44] LGBP + WPCA [44] LVP [45] LGT [46] HGGP [47] HOG [48] DT-LBP [49] LDP [50] GV-LBP-TOP [51] DLBP [52] GV-LBP [51] LQP + WPCA [53] POEM [54] POEM + WPCA [54] s-POEM+WPCA [55] DFD [56] DFD + WPCA [56] LBP weighted [26]

93.0 98.5 94.0 98.1 97.0 97.0 97.6 90.0 99.0 94.0 98.4 99.0 98.1 99.8 97.0 99.6 99.4 99.2 99.4 97.0

51.0 84.0 97.0 99.0 70.0 90.0 98.9 74.0 100.0 83.0 99.0 99.0 98.5 94.3 95.0 99.5 100.0 98.5 100.0 79.0

61.0 79.4 68.0 83.8 66.0 71.0 77.7 54.0 84.0 62.0 82.0 86.0 80.9 85.5 77.6 88.8 91.7 85.0 91.8 66.0

50.0 70.0 53.0 85.0 50.0 67.0 76.1 46.6 80.0 53.0 81.6 85.0 81.2 78.6 76.2 85.0 90.2 82.9 92.3 64.0

Equalized LBP

98.6

84.2

76.3

69.4

The results of other methods are from the original papers.

300

A. Anees, Y.-P.P. Chen / Pattern Recognition 77 (2018) 289–305 Table 4 Accuracy comparison of different face feature representation methods on the LFW database in unsupervised settings. Method

Recognition rate (%)

DeepFace [61] LARK unsupervised [62] MRF-Fusion-CSKDA [63] Spartans [64] Equalized LBP

95.92 78.30 98.94 94.28 81.37

∗

The results of other methods are from the original papers.

rates. Table 4 compares the accuracy of different methods with the proposed work on the LFW database in unsupervised settings. It can be observed that the results of the proposed work are on the slightly lower side but very competitive with the benchmark work. Furthermore, the low recognition rate is due to the significantly lower computational complexity of the proposed work compared to the other work. 5.4. Resistant to sparse noise ELBP can resist the effect of the addition of sparse noise. One of the most common types of sparse noise is salt and pepper in which a certain number of pixels in an image are replaced with dark and bright spots. Fig. 10(a) shows the effect of salt and pepper noise with different density levels. Also, instead of replacing this with a full dark or bright spot, a certain gray level can be added to a certain number of pixels. Fig. 10(b) shows the effect of adding random gray levels with different percentage intensities of pixel values. At the fourth level of noise density and intensity, the noisy images of salt and pepper and gray addition are hardly recognizable. We conducted ROC experiments on the AT&T face database [40] for four levels of density for salt and pepper noise and four levels of percentage intensity for gray additional noise to show the impact of percentage of noise and gray scale variations during the proposed equalization. The results of these two experiments with different percentages are illustrated in Fig. 11. The recognition rates in Fig. 11(a) for noise density up to 0.050 are 100% while it decreases when the noise level is increased. Similarly, the recognition

Fig. 10. Images with sparse noise, (a) salt and pepper noise (replacement with either dark or bright spot) with different densities level, (b) salt and pepper noise (addition of random gray levels) with different percentage of intensities.

rates in Fig. 11(b) for noise intensity are very high, being up to 7.5% and then decreases afterward. 5.5. Computational time The computational time of our proposed method is calculated and compared with the other existing binary operators. Our hardware configuration comprises a 3.4 GHz CPU and a 24GB RAM. The results are listed in Table 5. The computational time is slightly higher than the LBP binary operator due to the additional proposed step of equalizing the three LBP binary features. However, the computational complexity is much less than the NRLBP and ENRLBP. 6. Security analysis for key generation Two major characteristics are used to evaluate biometric key generation: correctness and security. The correctness of the proposed framework was evaluated through ROC and the recognition rates in the previous section. The security of biometric key genera-

Fig. 11. ROC curves of AT&T face database [40] injected with (a) salt and pepper noise with four noise densities, 0.025, 0.050, 0.075, 0.100, and (b) gray addition with four noise percentage intensities, 2.5, 5, 7.5, 10.

A. Anees, Y.-P.P. Chen / Pattern Recognition 77 (2018) 289–305 Table 5 Computational time (ms) comparison of different face feature representation methods. Method

Feature dimension

Time

LBP [26] DFD [56] SIFT [57] CBFD [23] NRLBP [37] ENRLBP [37] Equalized LBP

3776 50,176 8192 32,0 0 0 3776 3776 3,776

22.9 1,511.2 63.7 227.3 30.96 36.67 23.1

Table 6 NIST statistical tests to check the randomness of the proposed key generation. The results show that the proposed key generation passes the randomness statistical tests. Statistical test

Decision

Frequency (Mono bit) test Frequency test within a block (Blocks sizes: 3, 4, 5, 6, 7, 8) The runs test Tests for the longest-run-of-ones in a block (Blocks size: 8) The binary matrix rank test (4 Matrices, Rows: 8, Columns: 8) (16 Matrices, Rows: 4, Columns: 4) The discrete fourier transform (Spectral) test The non-overlapping template matching test (Template length = 4, Blocks = 2, 4, 8) The overlapping template matching test (Template length = 4, Blocks = 4, 8) Maurer’s universal statistical test The approximate entropy test The cumulative sums (Cusums) test

Passed Passed Passed Passed Passed Passed Passed Passed Passed Passed Passed Passed

tion is evaluated following three major benchmark security conditions: (1) Basic security requirements, (2) Entropy analysis and (3) Guessing attacks. 6.1. Basic security requirements The three basic security requirements presented in [65] are: randomness, weak biometric privacy and strong biometric privacy. These security requirements are tested on the proposed method and the statistical analysis are detailed in the following subsections. 6.1.1. Key randomness The key generated through the system module should appear as random to an adversary which has access to the auxiliary information. To check the randomness, we applied the NIST statistical test suite [66]. This test suite includes 15 statistical tests which were originally designed for a larger stream of binary bits (usually > 10 0 0 0), however these tests can also be applied on shorter streams as well, as in our case where the binary cryptography key has a length of 256 bits. The results of these tests on our proposed key generation are listed in Table 6 showing that our proposed key generation passes the randomness tests. 6.1.2. Weak biometric privacy (relevance of perturbation vector P) The adversary should not be able to extract any relevant information about biometrics from the given auxiliary information. The only auxiliary information used at the time of key generation, apart from the user’s biometrics is the perturbation vector P . This information can be saved in the database of a local hard disk or can be carried on a smart card. It is important to demonstrate that no relevant information can be deduced about the user’s biometrics from his perturbation vector in case it is stolen or lost. Let ρ i1

301

represent the four bit auxiliary information for a quantized fea ture fiq , i.e. ρi1 = 1111. The quantized feature can then be stated

and can have any value of fiq = 4n − 1, where, n = 1, 2, 3, . . .. It

is impossible to predict the exact and correct value of f iq based on its ρ i1 and therefore it cannot be reverse engineered from ρ i1 to fiq . Also, to show the resistance of P against a brute force attack, let consider a simple example by taking a face image of size 512 × 512. The number of ELBP codes made in this face image is equal to 510 × 510 = 260,100 (subtracting the first & last rows and columns). As the vast majority, sometimes over 90 percent, of all 3 × 3 ELBP patterns present in the observed textures belongs to uniform patterns [36], therefore, the number of uniform ELBP codes in the given example of face image is 260,100 ∗ 0.90 = 234,090. There are 58 uniform ELBP codes, so each code can have 4036 (234090/58) number of average uniform ELBP codes. Furthermore, if we consider the quantization interval, ϕ set to 10, then there are 403 different quantized intervals. Therefore, the total number of combinations of four bit auxiliary information in whole P is (4 ∗ 403 − 1 )58 = 1.03 × 10186 . However, ELBP histograms are similar among faces from different subjects, and the search space can be considerably reduced. Therefore, assuming there are only 30 different and unique quantization intervals instead of (4 ∗ 403 − 1 ), the total number of combinations of four bit auxiliary information in whole P is now 3058 = 4.71 × 1085 . The total number of combinations is still large enough that a modern computer will take more than 1010 years to check all the combinations demonstrating a strong resistance against the brute force attack. 6.1.3. Strong biometric privacy The adversary should not be able to extract any relevant information about biometrics from the generated key. The key generating mechanism based on facial features is effective and secure, such that, for a face feature F and its corresponding key K(F), the key generated for another face F , i.e. K(F ) is distinctly different from K(F), given F ≈ F . To test this, we used avalanche analysis. The avalanche effect alludes to an attractive property of cryptographic algorithms, typically block ciphers. The avalanche effect is obvious if, when an input is changed marginally (for example, flipping a single bit) the output changes fundamentally (e.g., half the output bits flip). In this work, we use avalanche analysis to check the robustness of the generated key, changing a single bit in the feature and observing the change in its corresponding key. Statistically, the avalanche effect can be measured through the number of pixel change rate (NPCR) and unified average change intensity (UACI), given as [67]:

i

NPCR =

D (i ) × 100%, N

1 |K1 (i ) − K2 (i )| UACI = N 2M − 1

(18)

× 100%,

(19)

i

where K1 and K2 are the two cryptography keys obtained by the difference in a single bit of face features, N is the length of the cryptography key, M is the number of bits considered and D(i) is defined as

D (i ) =

0

i f K1 (i ) = K2 (i ),

1

i f K1 (i ) = K2 (i ).

NPCR implies the change rate of the quantity of the binary bits of the cryptography key when a single bit of face features is adjusted and the UACI measures the normal power of contrasts between the face features and the cryptography key. The desired minimum value of NPCR should be 50%. These analyses are applied on the test images of different databases. Eight cases are considered for each group of test images, varying M from 1 to 8 bits.

302


Table 7 Comparative UPCR and NACI analysis of proposed framework applied on three test images of different databases. The analysis is done for eight different cases considering the number of bits M varying from 1 to 8. Analysis

NPCR(%)

M

Image 1

Image 2

Image 3

Image 1

UACI(%) Image 2

Image 3

1 2 3 4 5 6 7 8

60.55 81.25 89.06 90.63 91.80 93.75 92.97 90.63

45.70 67.19 80.86 92.19 91.80 96.09 92.97 96.88

48.83 71.88 79.69 92.19 89.84 96.09 92.97 96.34

60.55 35.42 32.48 30.31 25.58 28.94 26.50 29.64

45.70 33.33 29.97 29.69 30.30 23.07 26.91 27.34

48.83 37.76 38.17 34.90 28.35 32.14 33.37 31.61

Fig. 12. Average entropy of 193 different generated keys for different cases of Bs . The average entropy of the generated keys is very good for cases when Bs = 1, 2, 4 and significantly decreases from the ideal scenario after Bs = 4, however, it does not depict the vulnerability of our proposed system, and the number of symbols in these cases is simply way too low to have high entropy.

The values of the NPCR and UACI analysis for all eight cases for the proposed framework are shown in Table 7, demonstrating that these values are well above the specified threshold.

6.2. Entropy analysis Entropy is a measure of unpredictability of the state, or equivalently, of its average information content. In a binary source, it represents the minimum expected number of binary bits per symbol required to communicate that source. In our case, the length of generated keys is 256 binary bits. We have calculated the entropy of our generated keys for different cases of number of bits per symbol, Bs . For instance, if we consider one bit per symbol then there are 256 symbols in the key, the ideal entropy in this case is 1. Similarly, if we consider four bits per symbol then there are 64 symbols in the key, and the ideal entropy in this case is 4. Fig. 12 illustrates the average entropy of 193 different generated keys for different cases of Bs . The ideal entropy for these different cases is also shown in this figure. The average entropy of the generated keys is very good for cases when Bs = 1, 2, 4. The entropy significantly decreases from the ideal scenario after Bs = 4, however, it does not depict the vulnerability of our proposed system; the number of symbols in these cases is simply way too low to have high entropy. The number of symbols against Bs is also shown in this figure, plotted on the other side of the y-axis.

6.3. Guessing attacks The adversary can use many ways to guess or break the authentic keys. In this section, we examine different kinds of guessing attacks and the average number of guesses the adversary has to make in these attacks to obtain the authentic keys. 6.3.1. Exhaustive search attack In this case, the adversary has no knowledge or information about the key generation system, the algorithm, the auxiliary information or the feature space. The adversary has to use an exhaustive search for the authentic key. The average number of guesses depends upon the length of the generated key. If L is the length of the generated key, then the number of guesses by the exhaustive search is Gesa ≈ 2L−1 . It is the same as the key space considered in encryption and is a well-known result. 6.3.2. Authentic key statistics attack In this case, the adversary has the statistics of the generated authentic keys and then tries to guess the keys using frequency analysis of the authentic keys. Let y j , ( j = 1, 2, . . . , N ) be the N subjects and kj be their N corresponding keys. Since it is possible that two subjects generate the same authentic key (false positive), let k j , ( j = 1, 2, . . . , M ) be the M unique keys with their frequencies fj such that f1 ≥ f2 ≥ . . . ≥ fM . In this case, the average number of guesses the adversary has to make is Gaksa = M j=1 j ∗ f j which directly depends on the number of unique authentic keys generated, i.e. the more unique keys there are, more guesses will be needed. In our work, the number of unique keys generated is almost equivalent to the number of subjects considered and is enough to resist this attack, as supported by our claim regarding stability and robustness under the assumption that only biometric samples whose difference is less than the specified threshold compared to their reference biometric sample will be considered. 6.3.3. Device key statistics attack In this case, the adversary has the statistics of the key generation device or system. The adversary can test different subjects to generate their respective keys and can test the key generation system by collecting all the unique keys. Let y j , ( j = 1, 2, . . . , N ) be the N subjects and kj be their N corresponding keys. Let us

consider a case for subject s1 . Also, let ki j , ( j = 1, 2, . . . , Mi ) be the Mi unique keys with their frequencies fi j , i = 1, 2, . . . , N, j = 1, 2, . . . , Mi such that fi1 ≤ fi2 ≤ . . . ≤ fiMi . Of these Mi unique keys,

one is the authentic key for subject s1 . Let it be Kip , where 1 ≤ p ≤ Mi . Let Tp be an event that happens when this pth is the unique authentic key for subject s1 . This is tested for all the subjects and the count of frequencies of this event Tp is recorded. Let Fp be the frequencies of these events, where 1 ≤ p ≤ max(Mi ). In this case, the average number of guesses the adversary has to make max(M ) is Gdksa = j=1 i j ∗ Fj . Clearly, the number of guesses depends upon the frequencies of events Fj and the authentic key can be hacked by considering the frequencies in non-increasing manner with a low number of guesses. To avoid this, the number of unique keys and the key space should be large enough to resist this attack which is the case in our proposed work. 7. Conclusion We have developed a unified framework for generating cryptography keys based on facial features which is efficient and the first of its kind. There are three different modules in the proposed framework: learning the facial features, the quantization of the facial features and key generation. The facial features are extracted


through the proposed binary feature learning operator. The application of facial feature recognition considered here is cryptography key generation which does not require much robustness. This is in contrast to a security application where the subject has to be recognized in different environments with variations in pose, illumination, expression and low image resolution. In our application, the features of a subject are extracted in a constant illumination with minimal variations in pose and expression at the time of enrollment and key generation. There are no facial templates saved in the database, therefore, to cater for inter-user variations and to generate discriminative cryptography keys, we developed a 3-bit quantization technique to quantize the extracted binary features. The developed quantization technique not only cater for the variations, it also enhances the security of the system, which is a significant benefit. The quantized features are then used to generate stable keys such that, for any two very similar facial biometrics, their generated keys are distinctively different from each other. Like all new proposals, we strongly encourage the analysis of our framework before its immediate deployment. The proposed work is a general framework intended for the direct generation of biometric keys using biometric data, and thus, can be extended to be applied to other biometric modalities as well, such as iris patterns, fingerprint, handwritten signature, etc. Also, biometric features from multiple biometrics’ data can be used to generate even more secure keys, but at the expense of more computational complexity.

References [1] Y. Rong, S. Xiong, Y. Gao, Low-rank double dictionary learning from corrupted data for robust image classification, Pattern Recognit. 72 (2017) 419–432. [2] H. Lee, Y.-P. P. Chen, Cell morphology based classification for red cells in blood smear images, Pattern Recognit. Lett. 49 (2014) 155–161. [3] S.C. Cheng, K.Y. Cheng, Y.-P. P. Chen, GHT-based associative memory learning and its application to human action detection and classification, Pattern Recognit. 46 (11) (2013) 3117–3128. [4] F.-J. G-Serrano, A. N-Vázquez, A. A-Martín, Training support vector machines with privacy-protected data, Pattern Recognit. 72 (2017) 93–107. [5] A. Ibraham, Al-kindi: the origins of cryptology: the arab contributions, Cryptologia 16 (2) (1992) 97–126. [6] F. Ahmed, A. Anees, V.U. Abbas, M.Y. Siyal, A noisy channel tolerant image encryption scheme, Wireless Pers. Commun. 77 (4) (2014) 2771–2791. [7] A. Anees, A.M. Siddiqui, F. Ahmed, Chaotic substitution for highly autocorrelated data in encryption algorithm, Commun. Nonlinear Sci. Numer. Simul. 19 (9) (2014) 3106–3118. [8] J. Daemen, V. Rijmen, The Design of Rijndael: AES - The Advanced Encryption Standard, Springer, 2002. [9] M. Jakobsson, S. Taveau, The Case for Replacing Passwords with Biometrics, in: Mobile Security Technologies, 2012, pp. 1–6. [10] A.K. Jain, K. Nandakumar, A. Nagar, Biometric template security, EURASIP J. Adv. Signal Process. 8 (2008). [11] A. Bodo, Method for producing a digital signature with aid of a biometric feature, 1994, German Patent DE 4243908A1. [12] P.K. Janbandhu, M.Y. Siyal, Novel biometric digital signatures for internet-based applications, Inf. Manage. Comput. Secur. 9 (5) (2001) 205–212. [13] W. Sheng, S. Chen, G. Xiao, J. Mao, Y. Zheng, A biometric key generation method based on semisupervised data clustering, IEEE Trans. Syst. Man Cybern. 45 (9) (2015) 1205–1217. [14] M.H. Lim, A.B.J. Teoh, K.A. Toh, An efficient dynamic reliability-dependent bit allocation for biometric discretization, Pattern Recognit. 45 (5) (2012) 1960–1971. [15] W. Sheng, G. Howells, M.C. Fairhurst, F. Deravi, Template free biometric key generation by means of fuzzy genetic clustering, IEEE Trans. Inf. Forensics Secur. 3 (2) (2008) 183–192. [16] E.J.C. Kelkboom, G.G. Molina, T.A.M. Kevenaar, R.N.J. Veldhuis, W. Jonker, Binary biometrics: an analytic framework to estimate the performance curves under gaussian assumption, IEEE Trans. Syst. Man Cybern. A, Syst. Hum. 40 (3) (2010) 555–571. [17] B. Chen, V. Chandran, Biometric based cryptographic key generation from faces, in: Digital Image Computing Techniques and Applications, 9th Biennial Conference of the Australian Pattern Recognition Society on, 2007, pp. 1–8. 3–5 Dec. [18] Z. Zeng, P.A. Watters, A novel face hashing method with feature fusion for biometric cryptosystems, in: Universal Multiservice Networks, 2007. ECUMN ’07. Fourth European Conference on, 2007, pp. 1–6. 14–16 Feb.

303

[19] M. Jeong, C. Lee, J. Kim, J.-Y. Choi, K.-A. Toh, J. Kim, Changeable biometrics for appearance based face recognition, in: Biometric Consortium Conference, 2006 Biometrics Symposium: Special Session on Research at the, 2006, pp. 1–5. 19 Sept.–21 Aug. [20] V. Bjorn, Cryptographic key generation using biometric data, 20 0 0, U.S. Patent no. 6,035,398. [21] A. Kumar, D. Zhang, Hand geometry recognition using entropy based discretization, IEEE Trans. Inf. Forensics Secur. 2 (2) (2007) 181–187. [22] A.B.J. Teoh, W.K. Yip, S. Lee, Cancellable biometrics and annotations on biohash, Pattern Recognit. 41 (6) (2008) 2034–2044. [23] J. Lu, V.E. Liong, X. Zhou, J. Zhou, Learning compact binary face descriptor for face recognition, IEEE Trans. Pattern Anal. Mach. Intell. 37 (10) (2010) 2041–2056. [24] M. Turk, A. Pentland, Eigenfaces for recognition, J. Cogn. Neurosci. 3 (1) (1991) 71–86. [25] P.N. Belhumeur, J. Hespanha, D.J. Kriegman, Eigenfaces vs. fisherfaces: recognition using class specific linear projection, IEEE Trans. Pattern Anal. Mach. Intell. 1 (7) (1997) 711–720. [26] J. Kittler, A. Hilton, M. Hamouz, J. Illingworth, 3d assisted face recognition: A survey of 3d imaging, modelling and recognition approachest, in: Proc. Eur. Conf. Comput. Vis., 2004, pp. 469–481. [27] C. Liu, H. Wechsler, Gabor feature based classification using the enhanced fisher linear discriminant model for face recognition, IEEE Trans. Image Process 11 (4) (2002) 467–476. [28] Y. Bengio, P. Lamblin, D. Popovici, H. Larochelle, Greedy layer-wise training of deep networks, in: Proc. Adv. Neural Inf. Process. Syst., 2007, pp. 153–160. [29] S. Rifai, P. Vincent, X. Muller, X. Glorot, Y. Bengio, Contractive auto-encoders: Explicit invariance during feature extraction, in: Proc. Int. Conf. Mach. Learn., 2011, pp. 833–840. [30] G.E. Hinton, S. Osindero, Y.-W. Teh, A fast learning algorithm for deep belief nets, Neural Comput. 18 (7) (2006) 1527–1554. [31] Z. Sun, L. Meng, A. Ariyaeeinia, X. Duan, Z.-H. Tan, Privacy protection performance of de-identified face images with and without background, 39th International Convention onInformation and Communication Technology, Electronics and Microelectronics (MIPRO), 2016. [32] S.A. Osia, A.S. Shamsabadi, A. Taheri, H.R. Rabiee, N.D. Lane, H. Haddadi, A hybrid deep learning architecture for privacy-preserving mobile analytics, 2017, ArXiv:1703.02952. [33] R. Shokri, V. Shmatikov, Privacy-preserving deep learning, in: Conference on Computer and Communications Security (CCS), ’15 Proceedings of the 22nd ACM SIGSAC, 2015, pp. 1310–1321. [34] Y. Zhu, S. Lucey, Convolutional sparse coding for trajectory reconstruction, IEEE Trans. Pattern Anal. Mach. Intell. 37 (3) (2015) 529–540. [35] T. Ojala, M. Pietikäinen, D. Harwood, A comparative study of texture measures with classification based on featured distribution, Pattern Recognit. 29 (1) (1996) 51–59. [36] T. Ojala, M. Pietikäinen, T. Maenpaa, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns, IEEE Trans. Pattern Anal. Mach. Intell. 24 (7) (2002) 971–987. [37] J. Ren, X. Jiang, J. Yuan, Noise-resistant local binary pattern with an embedded error-correction mechanism, IEEE Trans. Image Process. 22 (10) (2013) 4049–4060. [38] F. Ahmed, M.Y. Siyal, V.-U. Abbas, A secure and robust hash-based scheme for image authentication, Signal Process 90 (5) (2010) 1456–1470. [39] F. Ahmed, A. Anees, Hash-based Authentication of Digital Images in Noisy Channels, in: Robust Image Authentication in the Presence of Noise, 2015, pp. 1–42. [40] F. Samaria, A. Harter, Parameterisation of a stochastic model for human face identification, in: Proceedings of 2nd IEEE Workshop on Applications of Computer Vision, 1994. [41] D. Hond, L. Spacek, Distinctive descriptions for face processing, in: 8th BMVC, 1997, pp. 320–329. [42] P.J. Phillips, H. Moon, S.A. Rizvi, P.J. Rauss, From few to many: illumination cone models for face recognition under variable lighting and pose, IEEE Trans. Pattern Anal. Mach. Intell. 23 (6) (2001) 643–660. [43] P.J. Phillips, H. Moon, S.A. Rizvi, P.J. Rauss, The FERET evaluation methodology for face-recognition algorithms, IEEE Trans. Pattern Anal. Mach. Intell. 22 (10) (20 0 0) 1090–1104. [44] W. Zhang, S. Shan, W. Gao, X. Chen, H. Zhang, Local gabor binary pattern histogram sequence (LGBPHS): a novel non-statistical model for face representation and recognition, in: Proc. IEEE Int. Conf. Comput. Vis., 2005, pp. 786–791. [45] X. Meng, S. Shan, X. Chen, W. Gao, Local visual primitives (LVP) for face modelling and recognition, in: Proc. Int. Conf. Pattern Recog., 2006, pp. 536–539. [46] Z. Lei, S.Z. Li, R. Chu, X. Zhu, Face recognition with local gabor textons, in: Proc. Int. Conf. Adv. Biometrics, 2007, pp. 49–57. [47] B. Zhang, S. Shan, X. Chen, W. Gao, Histogram of gabor phase patterns (HGPP): a novel object representation approach for face recognition, IEEE Trans. Image Process. 16 (1) (2007) 57–68. [48] E. Meyers, L. Wolf, Using biologically inspired features for face processing, Int. J. Comput. Vis. 76 (1) (2008) 93–104. [49] D. Maturana, D. Mery, A. Soto, Face recognition with decision tree-based local binary patterns, in: Proc. 10th Asian Conf. Comput. Vis., 2010, pp. 618–629. [50] B. Zhang, Y. Gao, S. Zhao, J. Liu, Local derivative pattern versus local binary pattern: face recognition with high-order local pattern descriptor, IEEE Trans. Image Process. 19 (2) (2010) 533–544.

304


[51] Z. Lei, S. Liao, M. Pietikainen, S.Z. Li, Face recognition by exploring information jointly in space, scale and orientation, IEEE Trans. Image Process. 20 (1) (2011) 247–256. [52] D. Maturana, D. Mery, A. Soto, Learning discriminative local binary patterns for face recognition, in: Proc. IEEE Int. Conf. Automatic Face Gesture Recog. Workshops, 2011, pp. 470–475. [53] S.U. Hussain, T. Napoleon, F. Jurie, Face recognition using local quantized patterns, Proc. Brit. Mach. Vis. Conf. (2012) 1–12. [54] N.-S. Vu, A. Caplier, Enhanced patterns of oriented edge magnitudes for face recognition and image matching, IEEE Trans. Image Process. 21 (3) (2012) 1352–1365. [55] N.-S. Vu, Exploring patterns of gradient orientations and magnitudes for face recognition, IEEE Trans. Inf. Forensics Secur. 8 (2) (2013) 295–304. [56] Z. Lei, M. Pietikainen, S.Z. Li, Learning discriminant face descriptor, IEEE Trans. Pattern Anal. Mach. Intell. 36 (2) (2014) 289–302. [57] D.G. Lowe, Distinctive image features from scale-invariant key-points, Int. J. Comput. Vis. 60 (2) (2004) 91–110. [58] B. Moghaddam, T. Jebara, A. Pentland, Bayesian face recognition, Pattern Recognit. 33 (20 0 0) 1771–1782. [59] R. Beveridge, D. Bolme, M. Teixeira, B. Draper, The CSU Face Identification Evaluation System User’s Guide: Version 5.0., Comput. Sci. Dept., Colorado State Univ., Denver, CO, 2003.

[60] E. L-Miller, G. B. H., A.R. Chowdhury, H. Li, G. Hua, Labeled Faces in the Wild: A Survey, in: M. Kawulok, M.E. Celebi, B. Smolka (Eds.), Advances in Face Detection and Facial Image Analysis, Springer, 2016, pp. 189–248. [61] Y. Taigman, M. Yang, M. Ranzato, L. Wolf, Deepface: Closing the gap to human-level performance in face verification, IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) (2014) 1–8. [62] H.J. Seo, P. Milanfar, Face verification using the LARK representation, IEEE Trans. Inf. Forensics Secur. 6 (4) (2011) 1275–1286. [63] S.R. Arashloo, J. Kittler, Class-specific kernel fusion of multiple descriptors for face verification using multiscale binarised statistical image features, IEEE Trans. Inf. Forensics Secur. 9 (12) (2014) 2100–2109. [64] F. J-Xu, K. Luu, M. Savvides, Spartans: single-sample periocular-based alignment-robust recognition technique applied to non-frontal scenarios, IEEE Trans. Image Process. 24 (12) (2015) 4780–4795. [65] L. Ballard, S. Kamara, M.K. Reiter, The practical subtleties of biometric key generation, in: Proceedings of the 17th Annual USENIX Security Symposium, 2008, pp. 61–74. [66] A. Rukhin, et al., A Statistical Test Suite for Random and Pseudorandom Number Generators for Cryptographic Applications, National Institute of Standards and Technology, 2010. [67] Y. Wu, J.P. Noonan, S. Agaian, NPCR and UACI randomness tests for image encryption, Cyber Journals: Multidisciplinary Journals in Science and Technology, Journal of Selected Areas in Telecommunications (JSAT) (2010) 31–38.


305

Amir Anees received his BS degree in Electrical Engineering from HITEC University, Taxila-Cantt Pakistan in 2011 and MS degree in Electrical Engineering from Military College of Signals, National University of Sciences and Technology, Pakistan in 2014. He is currently working toward the Ph.D. degree in computer science at La Trobe University, Melbourne, Australia. His research interests include image encryption, image hashing and chaos-based encryption. Yi-Ping Phoebe Chen is Professor and Chair at the Department of Computer Science and Computer Engineering, La Trobe University, Melbourne Australia. She was Head of Department of Department of Computer Science and Computer Engineering, La Trobe University from Sep 2010 to April 2012. Prof Chen is the Chief Investigator of ARC Centre of Excellence in Bioinformatics. Phoebe received her BInfTech degree with First Class Honors and PhD in Computer Science (Bioinformatics) from the University of Queensland. Before she joined La Trobe, Phoebe was Associate Professor (Reader) in Deakin University from Dec 2003 to April 2010. She worked as a Associate Lecturer/Lecturer/Senior Lecturer in Queensland University of Technology from Jul 1999 to Nov 2003. Professor Chen has been doing multi-discipline researches for more than 15 years. She is currently working on knowledge discovery technologies and is especially interested in their application to genomics and biomedical science. Her research focus is to find effective solutions for integrating, visualizing, analyzing and mining big data, complex structures and functions, for scientific and biomedical applications. She has been working in many emerging areas such as bioinformatics, pattern recognition, health informatics, multimedia, and databases. She has published over 200 research papers, many of them appeared in top journals and conferences such as IEEE Transactions on Biomedical Engineering, Pattern Recognition, IEEE Transactions on Information Technology in Biomedicine, Molecular Systems Biology, Aging Cell, Nucleic Acids Research, Journal of Medical Genetics, RNA, BMC Genomics, BMC Bioinformatics, Current Drug Metabolism, Data Mining and Knowledge Discovery, IEEE Transactions on Knowledge and Data Engineering, Information Systems, SIGMOD, and ACM Transactions. She is steering committee chair of Asia-Pacific Bioinformatics Conference (founder) and International conference on Multimedia Modeling.