Non-conventional biometrics using DOST features for secure cloud deployment

Abstract

The current decade has been experiencing a lot of research opportunities and challenges in the domain of biometric security. Simultaneously, Internet-of-Things (IoT) is also gaining seed functionality in early aspects of human lives. Cloud computing is the central thematic node in these two areas. In this work, we have proposed a novel biometric authentication scheme which is not based on conventional minutiae features, rather it is based on the frequency domain information of the fingerprint image. Input fingerprint is subjected to suitable quick pre-processing and then the discrete orthonormal Fourier transformation (DOST) features are extracted. Through suitable feature selection, chosen feature points are given to the classification stage where the recognition is accomplished using the standard AdaBoost-RF (AdaBoost Random Forest) algorithm. An overall accuracy of 98.5% has been obtained on a $k$ -fold cross-validation ( $k=5$ ) measure. The result obtained is compared with that of the four other state-of-the-art methodologies. On this, the proposed method outperforms the others in terms of accuracy and time of computation.

Keywords

Cloud computing fingerprint biometric security pattern recognition

1. Introduction

Biometric authentication along with cloud support is the most demanding task as of now [1, 2, 3]. It is because of the tremendous need of their application in various sectors. Biometric authentication has its importance and role in several sectors like information technology, national security concerns, AADHAAR applications, secure cashless banking transactions, shopping applications and many more. The ease of implementations and uses also making the digital biometric authentication system more popular. Biometric authentication refers to the task of digitized recognition of a persons identity through any of his biological traits. These traits may be fingerprint, face, palm print, periocular region, iris etc. This task comes under pattern recognition domain. Among all the biometric processes, the fingerprint recognition task is the most popular and widely used method. Computerized input of the fingerprint of a person is taken and it is used for further use to identify and re-identify the same person time to time. It helps in many secure and smooth authenticated processes, especially which involves important transactions.

Figure 1.

General overview of a fingerprint recognition system.

Fingerprints are the unique ridges present on one’s finger, for which these have been used since many a centuries as a genuine authentication trait. Mutation of these traits is not possible. They are unchangeable throughout one’s lifetime. According to a study, the chances of identical fingerprints among two different individuals is just only 1 in 2 $\times$ 10 ${}^{15}$ approximately [4]. The block diagram of a simple computerized fingerprint recognition system has been shown in Fig. 1. There are three vital phases, namely, Input and pre-processing [5], feature extraction with selection and finally recognition through classification. The steps involved in pre-processing differs in nature among different schemes. In general, it deals with identification of core-points, image binarization, and other morphological operations like thinning, thickening etc. This is followed by the feature extraction phase where key points are extracted from the pre-processed image. As of now, many a work have been reported that emphasize on the minutiae characteristics of the finger print [6]. Some of the conventional algorithms have focused on discrete cosine transform (DCT) based features and fast Fourier transform (FFT) based features for the purpose [7]. We, have proposed a discrete wavelength transform (DWT) based scheme for the purpose of fingerprint recognition which proves to be more efficient than these conventional ones.

2. Background

Among the existing techniques, three of the handpicked methods which are necessary for the proposed work point of view have been discussed below. The first two methods are considered for a neck-to-neck comparison with the proposed scheme in terms of the rate of accuracy. The third concept discussed here forms the basis of the proposed work.

2.1 Support vector machine (SVM)

It is a tool specifically used for pattern classification. It uses a hyperplane of $n-1$ dimensions for classifying the data points distributed over $n$ dimensional space. The objective of SVM is to design the most feasible hyperplane which can maximally achieve the classification task. It does so by specific learning strategy as it is a supervised learning method.

•
Kernel: Concept of linear algebra is followed to transform a classification problem into the hyperplane theme. The dot product of input ( $I$ ) and the support vector ( $I_{i}$ ) is computed using the basis:

$\displaystyle g(I)=X(0)+\sum b_{i}\{I,I_{i}\}$ (1)

where, $b_{i}$ and X(0) are the coefficient which are obtained from the training data input during learning process. The most frequently used kernels are polynomial kernel* and the exponential kernel. These polynomial kernel is defined as:

$\displaystyle\textit{kernel}_{I,I_{i}}=1+\sum\{I,I_{i}\}^{constant}$ (2)

And, the exponential kernel is defined as:

$\displaystyle\textit{kernel}_{I,I_{i}}=e(-\gamma\sum\{(I-I_{i})^{2}\})$ (3)
•
Regularization ( $r$ ): The regularization parameter ( $r$ ) informs about the rate of avoidance of the misclassification during SVM training phase. If the value of $r$ is large, then small margin size has to be taken for the hyperplane and vice-versa.
•
Gamma ( $\gamma$ )*: It decides the effect of individual training iterations. The lower value of $\gamma$ tells that the training is not so good so far and hence more number of training iterations are to be made. Similarly, a high value of $\gamma$ indicates towards the convergence of the learning phase.

Overall, the SVM has been well utilized as an efficient supervised classifier for numerous tasks [8, 9]. Choosing this for the proposed work is thus justified.

Figure 2.
General overview of the proposed system.

2.2 k-means clustering

This is an unsupervised classifier first introduced in the year 1967 [10]). Pre-specified number of clusters or groups of data points are generated out of the given set of data points based on their similarity measures corresponding to individual centroids of each clusters. Similar data points are kept in same clusters subject to their similarity measure less than aprior threshold value. In case, the similarity measure of a data point exceeds the threshold, then it becomes the centroid of a newly formed cluster. The objective here is to minimize the distance measure which is given as:

$\displaystyle\textit{distance}=\sum_{u=1}^{U}\sum_{v=1}^{V}(\textit{data}_{u}-% \textit{centroid}_{v})^{2}.$

2.3 Adaptive boosting (AdaBoost)

This was first introduced by the Godel prize winners Freund and Schapire. Generally the adaBoost is used in combination with a number of weak classifiers to improve the final integrated efficiency. Conceptually, the weighted sum of output of a classifier is projected as the final output through this method. This is adaptive in nature because it boosts the weak classifiers gradually so that they tilt towards the correct classification of wrongly classified patterns. For the assignment of weight to the learners, a simple approach has been adopted. The classifier with less than 50% rate of preliminary classification accuracy are assigned weight as ‘0’ and those with less than 50% are assigned with a negative weight value. The adaBoost-RF (Random forest) has been chosen as a tool in the proposed work. This is because of the fact that, it provides an overall good rate of recognition accuracy with faster learning mechanism.

3. Proposed scheme

The propose scheme comprises of two distinct phases, namely, phase-1 that is for feature extraction feature selection, and phase-2 that is for recognition by classification. Generic methods are adopted for cropping the images for extracting the region of interests (ROI’s) from the input fingerprint images. These ROI’s are nothing but the input to the first phase of our proposed scheme. A complete sketch of the proposed scheme is illustrated in Fig. 2.

3.1 Phase-I

This phase utilizes the two-dimensional discrete orthonormal Stockwell transformation (2-D DOST). Stockwell transform (S-transformation) is a multiple scaling method that extracts the features right from pixel levels from an input fingerprint image. It takes into account the time and frequency description of the 2-D image signal. However, one of its major limitation is that, it consumes high computation time due to feature-redundancy. To mitigate this issue, DOST technique was introduced. For a given image $F(i,j)$ having dimension $L\times L$ , the mathematical deductions of dyadic samples can be given as under:

1) 1)
Apply two-dimensional Fourier transformation (2D-FT) to the input image $F(i,j)$ to compute the samples, i.e. $G(x,y)\leftarrow F(i,j)$ . 2.
2)
Divide the $G(x,y)$ and identify the count of points (p) thereby.
3)
$\textit{Result}\leftarrow\sqrt{p}\times G(x,y)$ .
4)
Compute the inverse of $R e s u l t$ for getting the required DOST of the input. This can be given as:

$\displaystyle D(i^{\prime},j^{\prime},x_{u}),y_{v})\leftarrow\frac{1}{\sqrt{2^% {p_{i}+p_{j}-2}}}$ $\displaystyle\times\sum\limits_{i=-2^{p_{i}-2}}^{i=2^{p_{i}-1}}\sum\limits_{j=% -2^{p_{j}-2}}^{j=2^{p_{j}-1}}G(x+y_{i},y+v_{j})$ $\displaystyle e^{2\pi m\left(\left(\frac{ii^{\prime}}{2^{p_{i}-1}}\right)+% \left(\frac{jj^{\prime}}{2^{p_{j}-1}}\right)\right)}$ (4)

where, $x_{u}=2^{p_{i}-1}+2^{p_{i}-2}$ and $y_{v}=2^{p_{j}-1}+2^{p_{j}-2}$ are the frequencies along the axis and it’s normal axis respectively.where, $x_{u}=2^{p_{i}-1}+2^{p_{i}-2}$ and $y_{v}=2^{p_{j}-1}+2^{p_{j}-2}$ are the frequencies along the axis and it’s normal axis respectively.
5)
The resultant is a voice-image with the same dimension as the input.

The DOST has been computer as per the steps mentioned below:

5) 1)
Randomly a pixel with coordinates $(I,J)$ is selected from the image,
2)
The voice-image ( $V_{\textit{img}}$ ) value is calculated for all the points along the frequency ordered as ( $p_{i},p_{j}$ ) for the respective location $(I,J)$ as:

$V_{\textit{img}}\bigg{[}\frac{I}{\textit{count}}\times 2^{p_{i}-1},\frac{J}{% \textit{count}}\times 2^{p_{j}-1}\bigg{]}$
3)
A regional space is built with dimension $2*\log\textit{count}_{i}$ in a repetitive manner.

There exist both $+$ ve and $-$ ve components from the DC component The frequency domain contains positive and negative DC-components, i.e. from ( $I_{x},J_{y}$ ) $=$ (0, 0) towards the Nyquist-frequency $\textit{Freq}_{\textit{Nyquist}}$ , $(I_{x},J_{y})=(\textit{count}/2,\textit{count}/2)$ . Hence, the DC-components from frequency domain representation are having a mapping to the frequency coefficient space. Like this, a region-of-interest (ROI) is generated with $\textit{count}\times\textit{count}$ DOST values. All these values are added to the feature-vector ( $V_{f}$ ). The concatenation of all the feature vectors are kept in the feature matrix ( $M_{f}$ ). Overall process involved in this phase are represented in Algorithm 3.1.

indent=2em Computation of features[1] Initialize $M_{f}\leftarrow[\hbox{},\hbox{}]$ , $V_{f}\leftarrow\hbox{}$ , $\textit{count}\leftarrow 1$ , and $w_{1}\leftarrow 1,w_{2}\leftarrow 1$ Set the number of feature points to be extracted from the region-of-interest as $F_{\textit{count}}\leftarrow\textit{count}$ $n=1$ to count Crop the $\textit{ROI}_{n}$ $\textit{ROI}_{n}\leftarrow\textit{Dimension}(\textit{ROI}_{n},\textit{count)}$ $M_{f}(n)[1:\textit{count},1:\textit{count}]\leftarrow$

$\textit{DOST}(\textit{ROI}_{n})$ $a=1$ to count $b=1$ to count $V_{f}(n)[w_{1},1]\leftarrow M_{f}(n)[a,b]$ $w_{1}\leftarrow w_{1}+1$ Set $w_{1}\leftarrow 1$ $w_{2}=1$ to count $M_{f}[w_{2},n]\leftarrow V_{f}(n)[w_{2},1]$

indent=2em Random forest with ada-boost[1] Initialize all the weight values ( $M_{1}(w)$ ) in the weight matrix $W$ to be $\frac{i}{m}$ , where $m$ is the total number of samples. $i=1$ to |iterations| $j=1$ to |trees| Create an array $A_{j}$ with $M_{i}(w)$ $Y_{j}=\textit{bootstrapping(sample)}$ $\textit{Forest}_{j}=\textit{maketree}(\textit{sample},A_{j})$ Through hypothesis $H$ Catch hypothesis $H(\textit{sample}_{n})$ Calculate the error value for the hypothesis as $\textit{error}_{i}=\sum{M_{n}(w)}$ Set error constant as $E_{c}=\frac{1}{2}\log\frac{1-\textit{error}_{i}}{\textit{error}_{i}}$ Update the weight matrix using $\textit{error}_{i}$ Update hypothesis Update decision as per sign of the final hypothesis
3.2 Phase-II

Table 1
Comparing the proposed scheme with other benchmark schemes

Method	Training size	Training time (sec.)	Test samples	Test time per sample (sec.)	Accuracy (%)
Gabor filter	5000	1000	4000	15	83.50
F-ratio	2000	500	1500	15.5	86
Skeletal convexity	2000	500	1500	10	89.50
RF	2000	500	1500	8	93.75
K-means	NA	200	2000	12	82.50
Proposed scheme with ADBRF	5000	1000	4000	5.5	98.50

This phase deals with the recognition by classification strategy. For this, the most popular classifier namely AdaBoost algorithm has been used. The AdaBoost is used for a weak learning for the purpose. Random forest works as ensemble classifier which was proposed by Breiman. This results in sufficiently good classification so far as the rate of accuracy is concerned. Random forest applies bagging whereby the constituent classifier is built via the input patterns. A number of features are selected which are used to make decision at certain nodes of decision tree structure. The rate of error is directly dependent on the count of the input feature sets. So, this is basically a collection of several tree prediction tools where the structures rely on the feature points with independent sampling and distribution among the trees on the forest being un-changed. The training set is used for creating the trees randomly, and a hypothesis for classification is made which is based on majority voting strategy. (Algorithm 3.1) AdaBoost algorithm is the well known boosting scheme. It considers a number of weak-classifiers with more error rate and makes an integrated hypothesis where the error rate reduces to minimal [11, 12]. For the case of binary-classification, it input the training sample set ( $D=I_{c},J_{c};c=1,2,\dots,\textit{count}$ ). Here, $I_{c}$ is the array of features and $J_{c}$ contains corresponding labels (binary (0, 1)). In our case, one class is the fingerprint of person-A and rest samples are fingerprints of NOT-A person. In the steps, the training sample is alloted by a weight which tells the probability that whether it should be considered for the classification training or not. The steps assigns initial uniform weights to each of training samples. The weak classifier gets trained iteratively based on these weighted samples. For all these rounds, hypothesis with lesser error rates are formed along with putting an increment to these weights. Next, the highly weighted samples are chosen up to the next level. Finally, a linear combination is done out of these levels.

Figure 3.

Samples from the dataset.

Figure 4.

ROC plot comparison of the proposed scheme with others.

4. Simulation and result analysis

For validating the proposed scheme, a sample set of 5000 fingerprints have been considered (Samples shown in Fig. 3). This samples belongs to the CASIA database which specifically contains the index fingerprints of 500 distinct persons.All these images are 8-bit gray-level BMP files and their resolution is 328 $\times$ 356. The fingerprint sample images are re-dimensioned to a size of 64 $\times$ 64 and subjected to the phase-1 of our proposed scheme. So, a total of 64 $\times$ 64 DOST feature points are extracted out of each sample ROI. A feature-matrix is formed out of these feature-arrays. As per the steps of Algorithm 2, recognition task is accomplished. For the validation, a $k$ -fold cross-validation is used ( $k=5$ ). The random forest used here constitute a total of 20 trees with depth $=$ 2. It is observed, through a pilot experiment that the performance is optimum with 20 number of trees for the forest.

The ROC-curves (Receiver operating characteristics curve) so obtained is compared with that of two other classifiers, namely, SVM (support vector machine) and k-NN (k nearest neighbor). This is shown in Fig. 4. The maximum value for AUC (area under curve) is 98.5. A snapshot of the real-time use of the simulated software as a product of the proposed scheme is shown in Fig. 5.

Figure 5.

Snapshot of the implementation of the proposed scheme through our interface.

Table 2

Comparing the proposed scheme with other benchmark schemes

SL no.	Particular	Selection	Remark
1	Service	SaaS	Programmed device with biometric sensors (fingerprint)
2	Deployment	Private cloud	PaaS as a future work
3	Resource	Local server vis-a-vis Amazon web	Full porting to Amazon web cloud under negotiation
2	Data management	Text-captcha	Future work focuses on multi-tenancy

In the Table 1, a comparison has also been presented. It is well observed that, the proposed scheme is satisfactorily outperforming the others in terms of the overall rate of accuracy.

5. Additional phase (deployment of data onto cloud)

As the desired final phase, the data have been uploaded onto the cloud service. This is done carefully by adopting the genuine guidelines which are standardized by the industry. These are mentioned in the list given below:

•
Ensuring NULL risk with error-free exit of the process,
•
Data identity management,
•
Data protection provision,
•
Privacy policy preserver,
•
Cloud security provision,
•
Cloud service agreement provision.

The detail description of the deployment strategy have been presented in Table 2 with justified remark whenever applicable.

However, this phase is now at its incubation stage and it is still under analysis and synthesis process. The table here presents preliminary information about it subject to future scopes of full deployment.
6. Conclusion

In this paper, an efficient method has been suitably proposed along with it’s constituent algorithm for recognition of fingerprints followed by secure deployment onto the cloud. This work makes a successful attempt in real time capturing and recognition of fingerprints. For this, the DOST features have been considered without considering the conventional minutiae features that leads to costly computational complexity. Finally, hardware interfacing has been carried out successfully. Satisfactory recognition rate of 98.5% has been obtained using the proposed scheme.

References

Singh

S.P.

Ayub

and Saini

J.P.

, Literature survey on different type of fingerprint recognition, in: 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom), 2016, pp. 3748–3755.

Razak

A.H.A.

and Taharim

R.H.

, Implementing gabor filter for fingerprint recognition using verilog hdl, in: International Conference on Signal Processing and Its Application, 2009, pp. 423–427.

Baldini

and Steri

, A survey of techniques for the identification of mobile phones using the physical fingerprints of the built-in components, IEEE Communications Surveys & Tutorials (99) (2017), 1–10.

Jin

Lim

M.-H.

Teoh

A.B.J.

Goi

B.-M.

and Tay

Y.H.

, Generating fixed-length representation from minutiae using kernel methods for fingerprint authentication, IEEE Transactions on Systems, Man, and Cybernetics: Systems 46(10) (2016), 1415–1428.

Gonzaleza

R.C.

Woods

R.E.

and Eddins

S.L.

, Digital image processing using matlab. 1st Indian Reprint: Pearson Education, 2004.

Liu

and Cao

, Minutiae extraction from level 1 features of fingerprint, IEEE Transactions on Information Forensics and Security 11(9) (2016), 1893–1904.

Neethu

Sreelakshmi

and Sankar

, Enhancement of fingerprint using fft–fft–n filter, Procedia Computer Science 46 (2015), 1561–1568.

Corte

and Vapnik

V.N.

, Support-vector networks, Machine Learning 20(3) (1995), 273–297.

Ben-Hur

Horn

Siegelmann,

, and Vapnik

V.N.

, Support vector clustering, Journal of Machine Learning Research (2) (2001), 125–137.

10.

MacQueen

J.B.

, Some methods for classification and analysis of multivariate observations, in: Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability, Berkeley, University of California Press, 1967, pp. 281–297.

11.

Freund

S.R.

, Experiments with a new boosting algorithm, in: ICML, 1996, pp. 148–156.

12.

Friedman

T.R.

and Hastie

, Additive logistic regression: A statistical view of boosting (with discussion and a rejoinder by the authors), Ann Stat 28(2) (Feb 2000), 337–407.