Color QR Pattern-Driven Cancelable Biometric Fingerprint System

In recent days, due to the high-speed inventions made by telecom industry with ever-increasing voice and data traffic will present new challenges to security and privacy protection [1]. The fraudulent activities in the wireless industry are emerging steadily and stealing identities in many ways. Due to its advanced analytics, unique information, and security level, biometric technologies are largely trusted to deter this fraud. However, with the rapid adaptation of biometric systems in a wide range of applications [2], there is also a growing concern that this may compromise the biometric trails, which is nothing but the anonymity of individuals.

In general biometric characteristics are immutable [3], and the leakage of biometric details in any one place resulting in permanent biometric compromise and remain unreliable forever.

Though cipher transformation [4] of these biometric details before creating the database can solve this problem, the key management problems associate with any block ciphers may reveal entire information [5]. And in the Fuzzy fault scheme [6] where all biometric information is encoded into some polynomials result in the high false rejection rate. On the other, some generic alternatives like cryptographic hash techniques [7] are not applicable for a wide range of biometric inputs and not support bio-cryptosystems. In order to overcome this generic problem that arises in any biometric systems cancelable biometrics was introduced where the original biometric information is canceled and replaced through some transformations and makes the same biometric data different for different application and reduce the impact the information leakages in biometric technologies. Since in biometric systems, well-approximated templates were used for biometric authentication to narrow down the penalty gap that exists between trials and real-time inputs, it is always motivated to transform the generated biometric template using some one-way function which producesthe cancelable biometric templates.

The work is structured systematically as follows. Section II gives an overview of various cancelable biometric systems and discusses the cancelable template transformation for a proven security system and also outlines the motivation for the proposed work. Section III discusses detailed texture analyzes, feature extraction, and the proposed template generation approach. Section IV discusses QR pattern generation and template transformation used for the cancelable biometric system. Performance and security analysis are given in Section V. Finally. Section VI concludes the work.

3.1 Texture analyzes, feature extraction and feature selection

In most cases,for texture analysis,a standard texture classification technique the local binary pattern operator has been used. Here, texture patterns are compared with some marginal value, which generates ternary patterns to accommodate all sorts of dynamic textures, as shown in Eq.(1), (2) and (3) where localized properties have been acquiring with the capability to setout the most discriminative patterns for template prototype modeling.

$RMPc\ =\ \sum\limits_{N=0}^{N-1}{T(x)\ *\ }{{2}^{N}}$        (1)

where,

RMPc= Resultant Matrix Pixelcode value in decimal

N= Number of pixels in the neighbourhood of central pixel

T(x) = Threshold function for LBP

$T(x)\ =\ \left\{ \begin{align}   & 1\quad if\ x\ge \,0 \\  & 0\quad if\ x\,<\ 0 \\ \end{align} \right.$      (2)

T(x) = Threshold function for LTP upper and lower

$T(x)\ =\ \left\{ \begin{align}  & \,1\quad if\ np>s \\  & \,0\,\,\,if\ np>d\And \And \,np<s \\  & -1\ \ if\ np<d \\ \end{align} \right.$        (3)

Variable x= np-cp

np = neighborhood pixel value to the central pixel

cp = central pixel of the each decomposed spatial square block

variable s = cp+tk and d = cp-tk, where tk = threshold value

Here two different characteristics are focused during feature extraction, namely spatial and morphological characteristics.The GLCM features are extracted [17] to formulate biometric templates which includes various strategies to exploit the texture property details as follows: Energy, Entropy, homogeneity, correlation values for  Orientation measure; Autocorrelation, Prominence, Shade, Dissimilarity, Homogeneity, probability values for Displacement measure; Sum average, Sum variance, Sum entropy, Difference variance, Difference entropy, Information measure of correlation for structural variant measure; Difference normalized, Moment normalized values for statistical measures.And shape features are considered for exploiting the structural appearance of the object of interest.Finally, feature selection approaches are used, which aid in decreasing the size of the template without giving rise to drastic degradation in performance metrics. Here potentially useful texture feature subset are selected from the GLCM feature set extracted by using Eq. (4) to (23) based on its variance measure, which gives prominent discriminations among various classes.

$Autocorrelation=\sum\limits_{i,j=0}^{M-1}{ij({{G}_{ij}})}$          (4)

$Contrast\text{ }=\sum\limits_{i,j=0}^{M-1}{{{G}_{ij}}(i-j)\mathop{^{2}}^{{}}}$       (5)

$Correlation=\sum\limits_{i,j=0}^{M-1}{{{G}_{ij}}\frac{(i-\mu )(j-\mu )}{{{\sigma }^{2}}}}$       (6)

$\mu =\sum\limits_{i,j=0}^{M-1}{ij({{G}_{ij}})}$    (6.1)

$Peakcorrelation=\sum\limits_{i,j=0}^{M-1}{{{G}_{ij}}\frac{(ij)-(\mu i.\mu j)}{{{\sigma }^{2}}}}$   (7)

${{\sigma }^{2}}=\sum\limits_{i,j=0}^{M-1}{{{G}_{ij}}{{(i-\mu )}^{2}}}$     (7.1)

$\mu i=\sum\limits_{i,j=0}^{M-1}{i({{G}_{ij}})}$      (7.2)

$\mu j=\sum\limits_{i,j=0}^{M-1}{j({{G}_{ij}})}$     (7.3)

Prominence = sgn (Z) |Z|^1/4     (8)

$Z=\sum\limits_{i,j=0}^{M-1}{\frac{{{(i+j-2\mu )}^{4}}({{G}_{ij}})}{4{{\sigma }^{4}}{{(1+c)}^{2}}}}$       (8.1)

Shade = sgn (X) |X|^1/3    (9)

$X=\sum\limits_{i,j=0}^{M-1}{\frac{{{(i+j-2\mu )}^{3}}({{G}_{ij}})}{{{\sigma }^{3}}{{(\sqrt{2(1+c)})}^{3}}}}$     (9.1)

$Dissimilarity=\sum\limits_{i,j=0}^{M-1}{|i-j|({{G}_{ij}})}$    (10)

$Energy =\sum\limits_{i,j=0}^{M-1}{{{({{G}_{ij}})}^{2}}}$   (11)

$Entropy =\sum\limits_{i,j=0}^{M-1}{-\ln ({{G}_{ij}}){{G}_{ij}}}$      (12)

$Homogeneity=\sum\limits_{i,j=0}^{M-1}{\frac{{{G}_{ij}}}{1+{{(i-j)}^{2}}}}$      (13)

Probability = MAX_i,j(G_ij) for all i,j    (14)

$Variance =\sum\limits_{i,j=0}^{M-1}{{{G}_{ij}}(i-\mu )\mathop{^{2}}^{{}}}$      (15)

$Sumaverage=\sum\limits_{i=2}^{2M}{i.{{G}_{x+y}}(i)}$     (16)

$Sum\operatorname{var}iance=\sum\limits_{i=2}^{2M}{{{G}_{x+y}}(i).{{(i-\text{Se})}^{2}}}$    (17)

$Sumentropy(Se)=-\sum\limits_{i=2}^{2M}{{{G}_{x+y}}(i).\log {{\{{{G}_{x+y}}(i)\}}^{{}}}}$      (18)

$Difference\operatorname{var}iance=\sum\limits_{x,y=0}^{M-1}{\operatorname{var}({{G}_{x-y}})}$      (19)

$Differenceentropy=-\sum\limits_{i=0}^{M-1}{({{G}_{x-y}}(i).\log \{{{G}_{x-y}}(i)\})}$       (20)

$Infmcorr=\frac{P{{H}_{XY}}-P{{H}_{XY}}1}{\max (P{{H}_{X}},P{{H}_{Y}})}$       (21)

$P{{H}_{XY}}=-\sum\limits_{i,j=0}^{M-1}{{{G}_{ij}}\log ({{G}_{ij}})}$       (21.1)

$P{{H}_{XY}}1=-\sum\limits_{i,j=0}^{M-1}{{{G}_{ij}}\log (G{{x}_{i}}G{{y}_{j}})}$       (21.2)

$Differencenormalized=\sum\limits_{i,j=0}^{M-1}{\frac{{{C}_{ij}}}{1+{{(i-j)}^{{}}}}}$     (22)

${{C}_{ij}}=\sum\limits_{i,j=0}^{M-1}{\frac{{{G}_{ij}}}{\sum{{{G}_{ij}}}}}$     (22.1)

$Momentnormalized=\sum\limits_{i,j=0}^{M-1}{\frac{{{C}_{ij}}}{1+{{(i-j)}^{2}}}}$    (23)

where,

Elements i and j are intensities between 0 and grey level-1( Spatial coordinates)

G_ij= Grey Level Co-occurrence Matrix(Normalized symmetrical GLCM)

M = Number of grey levels or dimension of the co-occurrence matricx

µ= GLCM mean with respect to i and j

µ_i= GLCM mean with respect to i

µ_j= GLCM mean with respect to j

Sgn(x)= sign of real number

Infmcorr = Information measure correlation

PH_X and PH_Y denote the entropy values of G_x and G_y respectively.

G_x+y and G_x-yare sum and different distribution functions respectively

σ 2= variance

C= The Correlation feature

3.2 Template generation

The biometric cryptosystem can improve accuracy with improved security and privacy. Here, the biometric level fusion is combined with template level fusion to provide higher security than all other unimodal fusion. This Unimodal system fuse information at the feature level, and different sets of biometric texture features generate different feature spaces. To make this non-uniform feature sets into convenience for application deployments and to regulate the inter and intra-class variations feature dimensions also optimized [18] based on the principal component (PCA) analyses, as shown in Figure 1. As compared to mutual content reduction using Independent Component Analysis (ICA) and discriminative subset selection using Linear Discriminant Analysis (LDA) proposed hierarchical PCA transform helps to reduce confusion metrics as well as dimensions.

Here texture features are extracted from three different texture regions along with shape features during feature extraction, not only increase the template size also carry redundant information. Moreover, it is always difficult to obtain discrimination with redundant data. PCA transform is used to select attributes by considering both consistency and correlative measures in a hierarchical manner to accomplish this task.

In this brief, the ultimate goal is to validate the performance metrics of the ternary threshold bounded texture model over a wide range of the fingerprint biometric system, as shown in Figure 2, which includes FVC2002 DB2, FVC2004 DB1, and FVC2006 DB2 to carry experimental results for performance validation. The data sets include 100 biometrics from each fingerprint data set, where each class comes with different levels of non-linear distortion, textural exploration, and image quality to prove the improved recognition accuracy. It was noted due to its distortion, morphological characteristics are inconsistent with  FVC2004 DB1, and FVC2006 DB2 offers the least relevant spatial details. But the different levels of attributes with PCA based feature attribute selection and transformation metrics of color variant QR transformation based irreversible transformations for cancelable template generation the associated degradation in overall system accuracy is minimized irrespective of the nonlinearity of the input object classes. In order to validate this, the proposed cancelable biometric is tested on a variety of fingerprint databases with moderate quality and texture details, as shown in Table 2.

From the obtained experimental results, the proposed texture model outperforms the state-of-the-art BioHashing LBP model [19], as shown in Table 3 in terms of template size. The recognition accuracy of QR color variant cancelable templates is compared in terms of Equal Error Rate(EER) measures, as shown in Table 3. As a result, it is well proved that using color driven QR pattern-based cancelable biometric template with appropriate texture computation as shown in Figure 3 and the GLCM feature set as shown in Figure 4 achieved better data retention with no significant changes as compared to Discrete Fourier Transform driven random projection model [20] and BioHashing function [19] in terms of the EER (%) matching rate performed over the FVC2004 DB1 database.

From the experimental results, it was found that PCA based feature selection optimally reduced the dimension set as shown in Figure 5 and also proved that through QR pattern and color conversion process as shown in Figure 6 not degrade the system accuracy significantly, and it is proved that the proposed QR pattern-based cancelable method attains better performance as shown Table 4.

In the cancelable biometric system, the transformed templates must adhere to act in accordance with security,revocation,and diversity requirements. The performance metrics of any template transformations are validated based on two key factors as follows: (i) The one-way transformation should not change the statistical characteristics of the input biometric template with any number of transformations; (ii) The performance metrics during authentication should be high with improved in variance level over input biometric trait changes.

Table 2. Parameters cogitation for the validation

2.png

Figure 2. The fingerprint biometric samples from the databases: (a) FVC2002 DB2, (b) FVC2004 DB1, (c) FVC2006 DB2

Table 3. Performance metrics of dimensionality reduction and feature level fusion on texture features

3.png

Figure 3. The classification results used for the texture computation

Table 4. Performance comparison of different cancelable biometric approaches from state of the art on the FVC2004 DB1 database (EER in %)

4.png

Figure 4. GLCM feature set

5.png

Figure 5. The Principal component analyzes of GLCM feature variants from each texture feature image set

6.png

Figure 6. The cancelable biometrics through the conversion of templates into a color QR pattern