COMPARATIVE PERFORMANCE ANALYSIS: CANCELABLE BIOMETRICS

In this section, a comparative performance between well-known state-of-the-art studies has been made. For this, the ideal fingerprint and iris biometric traits under the same testing protocols and evaluation parameters have been selected.

Fingerprint Cancelable Biometrics: As a common thumb rule in CB, the matching of biometric traits takes place in the transformed domain so that original information can be protected. Thus, we consider those fingerprint studies only which performed matching in the transformed domain and used FVC datasets. Each of FVC dataset contains 800 fingerprint images with eight images per finger. To make a better understanding, four state-of-the-art fingerprint approaches [35,36,43,100] are considered. The results given in Table 2.7 reveal that ranking-based hashing approaches are a better choice in the cancelable domain for fingerprint images.

Iris Cancelable Biometrics: In this case, efforts are made to make a comparison between CNN-based approaches and traditional cancelable approaches. The results were evaluated on standard datasets like MMU, IITD, CASIA, and UBIRIS iris datasets, and the results are tabulated in Table 2.8. It can be observed that deep learning-based cancelable approaches are a better choice in the cancelable domain

TABLE 2.7

Comparative Analysis of Cancelable Biometrics over Fingerprint Datasets

Approaches

Database

Accuracy (CRR %)

Accuracy (EER %)

RGHE [36]

FVC2000DB1

99.22

1

RGHE [36]

FVC2000DB2

100

0.5

PHT [1001

FVC2000DB1

-

1

PHT [1001

FVC2000DB2

-

2

PHT [1001

FVC2004DB2

-

13.2

PR-NNLS [43]

FVC2000DB1

98.34

2.48

PR-NNLS [43]

FVC2000DB2

97.01

1.51

PR-NNLS [43]

FVC2004DB2

96.34

7.44

URP-IOM [35]

FVC2000DB1

-

0.20

URP-IOM [35]

FVC2000DB2

-

0.88

URP-IOM [35]

FVC2004DB2

-

3.08

TABLE 2.8

Comparative Analysis of Cancelable Biometrics over Iris Datasets

Approaches

Database

Accuracy (CRR %)

Accuracy (EER %)

LSC [781

IITD

-

1.4

/*, [96]

IITD

100

0.008

Л» l%]

MMU

100

0.006

CNN-RP [87]

IITD

98.66

0.12

CNN-RP [87]

MMU

95.57

0.15

RD [77]

CASIAv3-I

-

0.42

RD [77]

CASIAv4-T

-

2.07

Morton filter [69]

IITD

-

«0

Morton filter [69]

CASIA-V4

-

«0

for fingerprint images. However, it is important to note that deep learning architecture in the above-mentioned studies was only utilised for iris feature extraction. There is no end-to-end deep learning architectures that are retrieved from the literature on cancelable iris biometrics.

CONCLUSIONS AND FUTURE PROSPECTIVE OF DEEP LEARNING IN BIOMETRICS

In order to leverage true benefits of deep learning in the biometric domain, voluminous challenging datasets are needed that can represent real-world scenarios. Currently available biometric datasets - although some of them contain a large number of images like MS-Celeb, FaceNet, and WebFaces - are far from representing the true world population. Another point of concern is that biometric-based models should be designed in such a manner that can be implemented in real-world situations at a reasonable cost with high computational speed. As we know, high computational cost is associated with deep neural networks. Thus, selecting the right kind of deep learning-based architecture with high accuracy and low' computational cost is a challenging issue in the biometric domain. A possible solution is to design a network with comparable efficiency but w'ith a less computational cost like the once designed by Felix et al. [39]. Here, an effective surrogate convolutional layer based on the domain knowledge is designed that affords sufficient parameter saving as compared to the traditional standard convolutional layer and thus can be deployed in a resource- constrained environment also.

With large-scale deployments, unimodal biometric systems often suffer from challenging issues like intra-class variations, non-universality, and many more. In such scenarios, multimodal biometric systems serve as a rescue measure that integrates multiple biometric modalities and thus helps in improving the recognition rate. It should be noted that some work has been carried out in biometric fusion using machine learning techniques, but it suffers from challenging issues, as depicted in Ref. [7]. Currently, a deep learning-based architecture that can amalgamate feature representation and aggregation from multiple biometric traits simultaneously is needed. Moreover, in biometrics, security is a prime concern in order to gain public trust and confidence in it. Mainly biometric samples need to be protected from adversarial attacks, template attacks, and presentation attacks.

 
Source
< Prev   CONTENTS   Source   Next >