EXPERIMENTAL REPORTING

Important experimental observations made for the various template transformation approaches under conventional and deep learning-based categories are summarised in Tables 7.1 and 7.2. The databases and modalities under evaluation are also mentioned. For many approaches multiple modalities and databases are used for experimentation. Also various parameter selections are defined. Due to brevity of space, results are reported for the best parameters defined in the manuscript and for selected modalities. In the case of DNN-based transformation scheme, comparison of original templates are performed using many DNN architectures, and performance is reported in the transformed domain.

TABLE 7.1

Experimental Observations for Conventional Template Transformation Approaches

Category

Technique

Modality

Database

Performance

in

Transformed

Domain

Base Line Performance

Random

Projection

BioHashing [8]

Fingerprint

Face

Palmprint

Knuckleprint

FVC2002 DB1 ORL Private PolyU FKP

  • 0% EER 0% EER 0% EER
  • 25.9% EER

EER

30.5% EER

Multispace

Random

Quantisation

[П]

Face

FERET

7.09% EER

4.52% EER

Multispace Random Projection [12]

Face

ORL

25.77% EER

25.11% EER

User-dependent

Multi-state

Discretisation

[13]

Fingerprint

FVC2002 DB1

3.42% EER

14.84% EER

RP with vector translation [14]

Face

FERET+AR+

Aging+PIE

18.68% EER

17.54% EER

Random

Convolution

Random Kernel [17]

Face

CMU-PIE

100% RI

100% RI

BioConvolving

[18]

Signature

MYCT

7.95% EER

6.33% EER

Curtailed

circular

convolution

[19]

Fingerprint

FVC2002 DB1, DB2, DB3

2%, 2.3%, 6.12% EER respectively

Random

Noise

BioPhasor [20]

Palmprint

PolyU

0.13% EER

-

Gray Salting [21]

Iris

MMU 1

95.6% GAR

98% GAR

XOR-based salting [22]

Palmprint

CASIA

0.55% EER

0.50% EER

Random

Mapping

Table indices [23]

Iris

CASIA-V1

0.37% EER

0.28% EER

Index-of-Max

[24]

Iris

CASIA-v3

0.54% EER

0.38% EER

Random Slope-Vl [251

Palmprint

PolyU

0.48% EER

0.42% EER

Random Distance [26]

Palmprint

CASIA

0.53% EER

0.50%

TABLE 7.1 (Continued)

Experimental Observations for Conventional Template Transformation Approaches

Category

Technique

Modality

Database

Performance in Transformed Domain

Base Line Performance

Non-invertible

Polar transforms [32]

Fingerprint

FVC2002

DB1.DB2

5,93%, 4,0% EER

5.41%, 2.82% EER

Bloom filters [33,35]

Iris

CASIA-v3

1.49% EER

-

Random Permutation Maxout Transform [37]

Face

AR. FERET

6.95%, 3.65% EER

8.53%, 4.52% EER

TABLE 7.2

Experimental Observations for Neural Networks Based Template Transformation Approaches

Category

Technique

Modality

Database

Performance in Transformed Domain

Base Line Performance

Random

Projection

Deep Belief Network [42]

Finger-

vein

FV_NET64

91.2% GAR

91.8%. GAR

Local binary convolution neural networks [43]

Finger

Knuckle

PolyU FKP

0.125% FAR

FDFNet (FC & LBC) & RP [44]

Finger

Knuckle

PolyU FKI

0.002% EER (minor knuckle)

VGG-Face CNN& RP [45]

Face

CMU-PIE

99.95%. GAR

-

Deep Secure Quantisation (DSQ) [46]

IRIS

CASIA-v4

EER < 1%

Random

Convolution

BAM and linear convolution [47]

IRIS

CASIA-IrisV3-

Interval

3.56% EER

1.78% EER

Multiple deep CNN networks and Bio-Convolving [48]

Face

FERET. LFW. PaSc

  • 97.14%,
  • 98.93%,
  • 97.38%

accuracy

Random

Noise

Stacked autoencoder and chaotic matrices [49,50]

General

TABLE 7.2 (Continued)

Experimental Observations for Neural Networks Based Template Transformation Approaches

Category

Technique

Modality

Database

Performance in Transformed Domain

Base Line Performance

Random

Mapping

DNN and maximum entropy binary codes (MEB) [51]

Face

PIE, YALE. Multi-PIE

1,14%, 0.71%, 0,90% EER.

DNN and error correcting codes (ECC) [52]

Face & Iris

Casia-Webface & CASIA-Iris- Thousand1

99,99% GAR

Deep Table-based Hashing (DTH) [54]

Face

YouTube Faces -1-Face Scrub

0.0048%i

-

Non-Invertible and other Transforms

SAN and permutation [55]

Face

Celeb A, MCT. LF

f 39,3%, 39.2%, 72.5%, EER

19.7%, 8.0%, 33.4%, 16.9%> EER

DNN-based random block scrambling [56]

Face

-

96.72% RI

Binary Decision Diagram (BDD) and multi-layer extreme machine learning [57]

Finger-

vein

SDUMLA.

MMCBNU_6000,

UTFVP

93.09%, 98.70%, 98.61% CIR

REAL-LIFE CHALLENGES FOR APPLICATIONS OF CANCELABLE BIOMETRIC SYSTEMS

In spite of sufficient proof-of-concept, the present situation lags a necessary proof- of-work concept. The above section provides an exhaustive set of template transformation approaches and their results on matching performance. The contribution of DNN is towards improved matching performance. However, only evaluation on matching performance is not sufficient for designing a system that meets the practical implementation scenario. In order to meet the gap between research and reality certain implementation challenges must be addressed. This section highlights the drawback of implementing cancelable system which limits its important applications:

a. Multiple Identity Registrations Scenario: The cancelable system enrols a user T only on the basis of its transformed PI, generated using a user-specific key Kj. There may be cases when a same person may enrol again with different key Kr In that case the system outs a new pseudo-identity PL It becomes imperative to identify the person re-enrolling with the same biometric but different key. This is challenging as the system enrols diverse templates of the same person as two entirely different pseudo-identities. An attacker may fool the system this way to mask him and have multiple access accounts or even obtain multiple keys. A possible solution to this may be multi-level non- invertible transforms, where all users are transformed using the same key at the first level and then transformed using different user-specific keys at the second level. While it becomes difficult to differentiate between templates at the first level conventionally, there is tremendous scope for neural networks for learning to improve this learning process.

b. Generating Random Keys and Their Mixing: Another important aspect is related with generating random keys (AD) and fixing the range of values to which they belong. For most salting techniques, determining the range and distribution of random keys is important. If the random keys range is pre-dominating, it will have more effect while generating the transformed template. The range and distribution of the values generated for random key must be in accordance with the extracted feature vector. The specified range for a proposed work must not be used directly as it entirely depends on the type of feature extraction technique and the biometric trait to which it belongs. Some recent works also suggest the use of one biometric to generate random inputs to be used for feature distortion, like brain signals, or voice [58]. Feng et al. (2018) have generated techniques for generating and revoking brain passwords for head gear devices [59]. Again DNN forms the backbone of defining such extraction and their mixing here.

c. Insider Attack Scenario: The cancelable systems enhance user privacy by storing only transformed identities that are revocable and do not reveal any information about the original template due to non-invertible nature of the transform. However, the system still remains susceptible to insider attacks, where a malicious insider may uplift the transformed template from database to intercept the system. This can be prevented by using secret sharing techniques like [60,61]. The transformed pseudo-identity may be divided into two or more shares, which are distributed over multiple database servers and user-token. In that case, a malicious insider will only have a share which does not reveal the actual referenced transformed pseudo-identity until the remaining shares are available.

d. Designing the Storage over User-Specific Token: As a deviation to the name, the user-specific token must not store the entire key, but may be a hash or index to it. If the entire key a stored on the user-token it becomes an easy target for the attackers to read the information. Also a multi-secret sharing scheme like [61,62] may be useful here which shall input both transformed template and key to output distributed shares.

e. Re-enrollment Scenario in Case of Token Compromise: The cancelable systems allow user to re-enroll if the template is compromised. However, if the token is compromised then generation of new key is easy. User may re-enroll to cast the effect of the changed key. This is an important design issue and need clever tricks that allow only issuance of new token without inputting the biometric again. Some works like [62] address this issue by again applying secret sharing and designing a separate enrolment, authentication, and revocable modules. Still it remains an open issue and design challenge to be addressed.

f. Entire Database Compromise Scenario: If the entire database is compromised, then it will require re-enrolment of all users, which is the main limiting factor to practical implementation of the system. Again solutions w'hich transcends over multi-level transformation or cascading non-invertible and invertible mappings in presence of user-specific key might be useful; but multi-level transformation may have reduction in performance. Again a secret sharing-based approach might be helpful in this case. The entire database is transcended into three or more shares stored over distributed database servers and user token. If one of the database server is compromised it shall not reveal any original information and can be traced back to generate new shares without user re-enrolment [62]. This designing also needs more improvisations in future.

g. Exhaustive Evaluation for Performance: Most of the performance evaluations are either performed in the worst- and best-case scenarios. In the worst-case scenario, the discriminability of the transformation is analysed by assigning the same transformation key to each user. The then generated transformed templates are matched against each other to measure the performance in the transformed domain. It is expected to be comparable to the original, yet a little degradation is observed. In the best-case scenario, the templates are transformed by assigning different user-specific key to each user, which significantly increases the inter-user variations to give almost 0% false accept rate. The practical implementations need more detailed analyses on matching scores to set the system threshold for a match or nonmatch. The actual testing to qualify is the combination of both worst and best case testing. The combination can be defined as, initially generate a set of reference transformed templates by assigning each user a different user- specific keys say, КК„ (best case). Then generate probe templates by first transforming all users u„...un using key К,, to match against reference template of user Uj. This outputs a set of genuine and impostor matching scores for all users against user u,. By repeating this process for all users, one may be able to map the overall scores obtained for genuine and impostor analysis. Segregation can be easily defined over the overall set of genuine and impostor mappings to set the system threshold.

 
Source
< Prev   CONTENTS   Source   Next >