Cancelable Biometrics

In the past ten years, lot of work has been carried out in the cancelable domain owing to the increase in online biometric-based authentication. Recently, a taxonomy of CB techniques [56] is proposed, which divides it into six major categories, as shown in Figure 2.3. The concept of CB was first coined by Ratha et al. [71] in the year 2001. Later he suggested [73] three varied transformation functions (cartesian, polar, and surface folding) on fingerprint images. The sole aim of these transformation functions was to distort original feature vectors such that it is computationally infeasible and difficult to retrieve original feature vectors. Later studies [24], however, demonstrate potential security threats in this scheme, but it opens up new avenues for researchers in the cancelable domain. Since then, several works have been carried out in this field.

Table 2.4 demonstrates cancelable key approaches on fingerprint trait. Apart from fingerprints, iris is one of the most used and recognised biometric modalities. Thus, securing iris is also as much important as fingerprints. In Table 2.5, promising cancelable iris techniques along with its pros and cons are discussed. Apart from that, few works have been carried out on studying cancelability on multimodal biometrics. Like Chin et al. [14] proposed a template protection technique by fusing fingerprints and palmprint features on the basis of the user-specific key. In another notable work, Barrero et al. [30] proposed a bloom filter-based approach for protected face, finger-vein, and iris features. Very recently, a random distance method-based template protection technique [42] is proposed for protecting multiple templates that include the face, palmprint, palm-vein, and finger-vein. Recently, deep features extracted from finger knuckle modality have been secured via BioHashing technique, but the proposed approach [83] is not able to maintain the inexorable security-performance trade-off. Apart from traditional biometric traits, electrocardiogram (ECG) is emerging as a promising biometric trait in many authentication and verification applications. In Ref. [19], the authors investigated cancelable ECG biometrics using BioHash. In Ref. [11], the authors obtained excellent identification performance on highly compressed ECG data using Hadamard transform, but this could not achieve non- invertibility. Also, the applications of the compressive sensing theory for ECG have been investigated for compression [16]. So far, there have been very few prior works on cancelable ECG biometrics that deal with the issue of performance deterioration induced due to cancelable schemes and validation for CB criteria.

Deep Learning-Based Cancelable Techniques

With recent advances in Al and deep learning, an array of biometric-based authentication systems demonstrate outrageous performance and present unique security and privacy concerns. One of the pioneering works in this domain is performed by Talreja et al. [90]. In their work, they have proposed a secure multi-biometric system that uses a deep neural network and error-correcting codes. They have proposed two architectures: (i) fully connected architecture and (ii) bi-linear architecture for generating cancelable templates. In another notable work [37], highly discriminative facial features are learned via deep learning-based frameworks, which are further hashed using SHA-3, a well-known cryptographic technique. In Ref. [82], the authors proposed a novel CNN Network (FDFNet) for the extraction of the discriminative finger dorsal features. Then, BioHashing was used to hash the features extracted from each finger dorsal. In Ref. [3], a cancelable multi-biometric face recognition method was presented in which multiple CNNs extracted deep features from the face, eyes, nose, and mouth regions. In Ref. [67], the authors incorporated a classic

Taxonomy of cancelable biometric techniques. (Image is taken from [56].)

FIGURE 2.3 Taxonomy of cancelable biometric techniques. (Image is taken from [56].)


Key Cancelable Approaches on Fingerprints





Ratha et al. [73]

Pioneer work in this domain

Easy to implement

Pre-aligned images required

Jin et al. [33]

Random projection-based technique named BioHashing. It projects biometric feature to random space. By taking inner product of tokenised random vector with fingerprint features.

High performance

Performance degradation in stolen token scenario, prone to similarity based attack [21]

Lee et al. [49]

Rotational and translational invariant features are extracted from each minutiae

First alignment free cancelable template

Only theoretical justification of non-invertibility and revocability. No experimental validation. Unlinkability is not studied.


et al. [5]

Geometrical properties are explored to extract distinguish features from minutiae templates

Non-invertibility without loss of discriminative power, alignment free, low time complexity

Security analysis missing

Yang et al. [103]

Both local (distance, angle) and global (orientation, frequency) features of minutiae are explored to form non-invertible template

Non-invertible and unlinkable templates

Templates are unrevocable

Zhang et al. [106]

MCC was used for generating cancelable templates

Non-invertible, revocable, and unlinkable templates

Rigorous security analysis missing

Ferrara et al. [25]

KL transformation on MCC for generating cancelable templates named as P-MCC


Sandya and Prasad [80]

feature level fusion of fingerprint structures

ensures non- invertibility and revocability

Cross database attacks not studied

Arjona et al. [6]

A two factor fingerprint matching scheme that combines fingerprint identifier, i.e., protected MCC with device identifier, i.e., physically unclonable function generated from static random access memories


discriminability, non-invertibility, revocability, and unlinkability

Only one dataset considered, i.e.. FVC2002

deep learning approach into a BioCapsule-based facial authentication system to enhance recognition accuracy. In Ref. [105], a novel privacy-preserving finger-vein recognition system is developed based on binary decision diagram and multi-layer extreme learning machine paradigm. The proposed system ensures the safety of original finger-vein templates by ensuring non-invertibility and revocability.


Key Cancelable Approaches on Iris





Chin et al. [13]

Secure iris features coined as S-Iris encoding is proposed by iterating inner product between pseudo-random number and 1 -D log Gabor iris features

Pioneer work

Security analysis missing

Zuo et al. [108]

Two salting-based approaches proposed named as GRAY SALT and BIN SALT


Deciding strength of noise pattern added to original iris template is quite challenging

Pillai et al. [68]

Cancelable iris template generation based on sectored random projections

sectored random projections was used for mitigating the performance degradation due to eyelids and eyelashes

Performance degradation in case of stolen token scenario [12]

Ouda et al. [66]

Bioencoding a template protection technique is proposed that extracts consistent bits from IrisCodes and further encoded by using randomly generated binary Codewords

Simple implementation and can be integrated with existing systems

Non-invertability is compromised when encoding factor is stolen [46].

Rathgeb et al. [74]

Bloom filter-based cancelable iris template

High System Performance

Prone towards cross matching- based attacks [31]

Lai et al. [48]

Cancelable iris templates based on indexing first one hashing technique. The proposed framework is based on Min-Hashing and further strengthened by using modulo threshold function and P-order Hadamard product

Rigorous security analysis

testing on single dataset

CASIA-V3 iris dataset

Umer et al. [96]

To improve the security of existing BioHashing technique two different tokens were used (i) User dependent (ii) User independent for generating cancelable templates

Evaluated on extensive dataset

Unlinkability is not studied





Locality Sensitive Hashing is used for generating cancelable iris codes coined as Locality Sampled Codes

Extensive experimentation and security analysis


algorithm-based similarity attack not evaluated [21]

Deep Learning versus Non-deep Learning Cancelable Techniques

With the usage of deep learning techniques in the cancelable domain, biometric feature extraction becomes less time consuming as compared to traditional feature extraction methods. Naive feature extraction methods often require pre-processing and parameter tuning according to the dataset in consideration, while a generalised trained deep model works quite well w'ith different datasets (that belongs to the same biometric modality) without much tuning. On the other hand, non-deep learning-based techniques can even work on small dataset in the resource-constrained environment, but for training deep networks, large computational capability and the large dataset are required. It should be noted that usage of deep learning techniques in generating cancelable templates is in the infancy state as not much work has been carried out so far in this domain. Thus, deciding the supremacy between the two techniques at this point is not fair without harnessing the full advantages of deep learning-based techniques in generating cancelable templates.

< Prev   CONTENTS   Source   Next >