Cancelable Biometrics: A Review

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 1

Vishal M. Patel, Member, IEEE, Nalini K. Ratha, Fellow, IEEE, and

Rama Chellappa, Fellow, IEEE

Abstract

Recent years have seen an exponential growth in the use of various biometric technologies for trusted

automatic recognition of humans. With rapid adaptation of biometric systems, there is a growing concern

that biometric technologies may compromise privacy and anonymity of individuals. Unlike credit cards

and passwords, which can be revoked and reissued when compromised, biometrics are permanently

associated with a user and cannot be replaced. In order to prevent the theft of biometric patterns, it

is desired to modify them through revocable and non-invertible transformations to produce Cancelable

biometric templates. In this paper, we provide an overview of various cancelable biometric schemes for

biometric template protection. We discuss the merits and drawbacks of available cancelable biometric

systems and identify promising avenues of research in this rapidly evolving ﬁeld.

Index Terms

Biometrics, cancelable biometric templates, biohashing, salting, random projections, biometric tem-

plate protection.

I. INTRODUCTION

Biometrics refers to the physiological or behavioral characteristics of an individual. Many physical

characteristics, such as face, ﬁngerprints and iris and behavioral characteristics such as voice, gait

and keystroke dynamics, are believed to be unique to an individual. Hence, biometric analysis offers

a reliable solution to the problem of identity veriﬁcation. Recent developments in sensing and computing

technologies have made biometric systems more affordable and as a result they are easily embedded in a

Vishal M. Patel is with the Center for Automation Research, UMIACS, University of Maryland, College Park, MD 20742

(e-mail: [email protected]).

Nalini K. Ratha is with IBM Watson Research Center, Hawthorne, NY 10532 USA (e-mail: [email protected]).

Rama Chellappa is with the Department of Electrical and Computer Engineering and the Center for Automation Research,

UMIACS, University of Maryland, College Park, MD 20742 (e-mail: [email protected]).

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 2

variety of smart consumer devices such as mobile phones and tablets. Despite the widespread deployment

of biometric systems in various applications, the use of biometrics raises several security and privacy

concerns as outlined below [1].

1) Biometrics is not secret: The knowledge-based authentication methods totally rely on secrecy. For

instance, passwords and cryptographic keys are known only to the user and hence secrecy can

be maintained. In contrast, biometrics such as voice, face, signature and even ﬁngerprints can be

easily recorded and potentially misused without the user’s consent. Face and voice biometrics are

vulnerable to being captured without the user’s explicit knowledge.

2) Biometrics cannot be revoked or cancelled: If a biometric can be presented by a human being

who is one of the enrolled users, many biometrics security issues will be different. For example,

biometric-based authentication systems will not have to deal with spoofed biometrics and also replay

attacks on biometric systems. If a hacker gets access to the biometrics samples and has the ability

to present it to the system at choice emulating a human presence, there will be no trust associated

with the biometrics. In this scenario, we say that the biometrics has been compromised forever.

Passwords, crypto-keys and PINs can be changed if compromised. When tokens such as credit

cards and badges are stolen, they can be replaced. However, biometrics is permanently associated

with the user and cannot be revoked or replaced if compromised.

3) Cross application invariance and cross-matching: It is highly encouraged to use different passwords

and tokens in traditional authentication systems. However, biometrics-based authentication methods

rely on the same biometrics. If a biometric template is exposed once, it is compromised forever.

If a biometric template is compromised in one application, then the same method can be used to

compromise all applications where the biometric is used. Furthermore, since the same biometrics

is used across all applications and locations, the user can be potentially tracked if one or more

organizations collude and share their respective biometric databases.

4) Persistence: While relative robustness over time is a boon for biometrics it can also be a big

challenge from a privacy point of view when it needs to be changed. The uniqueness contained in

them is still the same even though the signal as well the template can look different.

Regarding privacy violations, cross-matching and inability to revoke a biometric are two major issues.

A simple approach would be to use standard encryption techniques such as hash functions or encryption

to enhance the privacy. Hash functions have been used to protect biometric templates in which one way

functions are used to compute a digest. Even though these functions are almost impossible to invert,

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 3

they produce signiﬁcantly different digest even with minor changes in the input. In practice, all biometric

templates change with environmental conditions. For instance, face and iris biometrics are signiﬁcantly

affected by illumination variations. Therefore, these functions can not be used directly in practice despite

being theoretically very strong as they apply only to exact data. Furthermore, when data are encrypted,

they need to be decrypted to carry out matching. This creates a possible attack point to get access to the

decrypted templates.

In order to overcome the vulnerabilities of biometric systems, both biometrics and crypto research

communities have addressed some of the challenges. Several biometric template protection schemes

have been proposed in the literature [2], [3], [4], [5], [6], [7], [8]. In particular, Cancelable biometrics

[3], [4], [5], [9] has gained a lot of interest in recent years. In this method, instead of storing the

original biometric, it is transformed using a one way function. The transformation can be applied either

in the original domain or in the feature domain. It was shown that this way of constructing biometric

templates has the desired properties of cancelable biometric templates [3], [4], [5]. In particular, it provides

revocability since a compromised biometric can be re-enrolled using another transformation. It preserves

privacy since it is computationally difﬁcult to recover the original biometric from a transformed one. It

prevents cross-matching between databases since each application uses a different transformation. And

it does not degrade the accuracy of a matching algorithm as the statistical characteristics of features are

approximately maintained after transformation. This allows one to use existing matching algorithms.

There are also some closely related but not equivalent biometric template protection schemes based on

cryptosystems [10] that have been studied extensively. These methods combine cryptographic keys with

transformed versions of the original biometric templates to generate secure templates. In these methods,

some public information, known as helper data, is generated. Depending on how the helper data is

used, biometric cryptosystems can be broadly classiﬁed into key binding and key generation systems. In

the key generation systems, both the helper data and the key are directly generated from the biometric

templates, while in the key binding systems, the helper data are obtained by combining the key with the

biometric template. Examples of key binding systems include fuzzy commitment [11] and fuzzy vault

[6]. Key generation schemes based on secure sketches [7] have also been proposed in the literature. In

the biometric cryptosystems, the level of security depends on the amount of information revealed by the

helper data. Other methods for biometric template protection include distributed source coding [12] and

fuzzy extractors [13]. A review of biometric cryptosystems can be found in [10], [8], and [14].

Our goal in this paper is to survey recent available approaches for designing cancelable biometric

templates, discuss their advantages and limitations, and identify areas still open for exploration. Fur-

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 4

thermore, we will discuss possible ways of attacking cancelable biometric systems. Development of

cancelable schemes for biometric template protection is crucial as biometric systems are beginning to

proliferate into the core physical and information infrastructure of our dynamic society.

Rest of the paper is organized as follows. Section II reviews various recent cancelable biometric

template protection methods. Attacks against cancelable biometric systems are discussed in Section III.

Finally, Section IV concludes the paper with a brief summary and discussion.

II. CANCELABLE BIOMETRIC TEMPLATES

In this section, we review a number of recent strategies for generating cancelable biometric templates.

In these methods, a function that is dependent on some parameter is used to generate protected biometric

templates. The parameter of the function is used as the key. Figure 1 shows the basic concept of cancelable

biometric template-based on non-invertible transformations.

Fig. 1: Block diagram of a cancelable biometric system.

A. Non-invertible Geometric Transforms

One of the earliest methods for generating cancelable biometric templates was based on non-invertible

geometric transformations. The idea is to morph the original biometric templates by applying signal

domain or feature domain transformations [3], [4], [5]. Figures 2(a) and (b) show examples of these

transformations applied in the signal domain and in the feature domain, respectively for face and ﬁnger-

print biometrics. Three different transformations were proposed for ﬁngerprint biometric in [4], [5].

These transformations are the Cartesian transformation, the polar transformation and the functional

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 5

transformation. Prior to applying these transformations, the images are registered by ﬁrst estimating

the position and orientation of the singular points (core and delta) and expressing the minutiae positions

and angles with respect to these points. Registration is an integral part of this method.

(a) (b)

Fig. 2: Illustration of non-linear transformation applied to face and ﬁngerprint biometrics [4]. (a) Feature

domain transformation for ﬁngerprint biometrics. Each minutiae (feature) position is transformed using a

non-invertible function y = f(x). The minutiae position x

is mapped to y

= f(x

). If we know y

, the

inverse mapping is a many-to-one transformation. x

, x

are all valid inverse mappings

to y

. (b) Illustration of cancelable biometrics for face recognition. The face is distorted in the original

pixel (signal) domain prior to feature extraction. The distorted version does not match with the original

face, while the two instances of distorted faces match among themselves.

For the cartesian transformation, the minutiae positions are measured in rectangular coordinates with

reference to the position of the singular point by aligning the x-axis with its orientation. The coordinate

system is divided into cells of ﬁxed size. The transformation consists of changing the cell positions [4],

[5]. Note that this transformation is not a simple permutation as the condition of irreversibility requires

that cells are mapped to the same cell.

In the polar transformation method, the minutiae positions are measured in the polar coordinate with

reference to the core position. The angles are measured with respect to the core orientation. As a result,

the coordinate space is divided into polar regions. The non-invertible transform consists of changing the

polar wedge positions. The minutiae angles also change with differences in the wedge positions before

and after transformation [4], [5].

One of the limitations of both polar and Cartesian transformations is that they are unstable in the sense

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 6

that a small change in minutiae position in the original ﬁngerprint can lead to a large change in minutiae

position after transformation if the point crosses a sharp boundary [4]. As a result, various functions giving

a locally smooth transformation of the minutiae positions were introduced in [4], [5]. The transformation

is modeled using a vector valued function

F (x, y) whose phase determines the direction of translation

and the extent of translation is given by the magnitude |

F | or alternately another vector valued function

G(x, y). One such function proposed in [4], [5] is an electric potential ﬁeld parameterized by a random

distribution of charges. The magnitude and phase of this function are given by

F | =



i=1

(z − z

)

|(z − z



Φ(x, y) =

arg

i=1

(z − z

)

|(z − z

where z = x + iy is the position vector and the random key K = [z

, z

, · · · , z

, q

, · · · , q

]

determines the position and magnitude of the charges. The transformation is given by

= x + K|

G(x, y)| + K cos(Φ

(x, y))

= y + K|

G(x, y)| + K sin(Φ

(x, y))

= mod (θ + Φ

(x, y) + Φ

rand

, 2π).

See [4] for more examples of various transformations and their analysis in terms of non-invertibility and

attack strength. This method was later extended in [15] so that it does not require the registration of

images. However, the approach in [15] exhibits lower veriﬁcation rates than [4].

In a related work, [16] proposes a mesh warping-based approach for generating cancelable iris tem-

plates. In this method, the iris texture is re-mapped according to a distorted grid mesh laid over it.

Distortions are speciﬁed by a key which offsets each vertex in the original mesh by some amount.

Speciﬁcally, a regular grid is placed over the texture in which the vertices are then randomly displaced

using the key as seed to a random number generator.

B. Random Projections

Another non-invertible transformation that is widely used for generating cancelable biometric templates

is based on random projections [17], [18]. In these methods, the extracted feature x ∈ R

from a biometric

is projected onto a random subspace A ∈ R

n×N

with n < N . Here, each entry a

i,j

of A is an independent

realization of a random variable. This process is described as follows

y = Ax, (1)

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 7

where y is the n dimensional random projection vector. Since we are embedding N dimensional feature

vectors in a space of a lower dimension n, for any biometric recognition to be effective, it is important

that the relative distances between any two points in the feature space be preserved in the output random

space. This is essentially characterized by the Johnson-Lindenstrauss (JL) lemma [19].

Lemma 1: For any 0 <  < 1 and any integer p, let n be a positive integer such that n ≥

4 ln(p)



/2−

Then, for any set S of p = |S| data points in R

, there is a map f : R

→ R

such that, for all

x, y ∈ S,

(1 − )kx − yk

≤ kf (x) − f(y)k

≤ (1 + )kx − yk

. (2)

This lemma essentially states that, a set S of points in R

can be embedded into a lower-dimensional

Euclidean space R

such that the pairwise distance of any two points is approximately maintained. In

fact, it can be shown that f can be taken as a linear mapping represented by an n × N matrix A whose

entries are randomly drawn from certain probability distributions [19]. This in turn implies that it is

possible to change the original form of the data and still preserve its statistical characteristics useful for

recognition.

In recent years, various improvements in the proof and the statement of the JL lemma have been

made (see [20] and [21] for more details). In fact, it has been shown that given any set of points S, the

following are some of the matrices that will satisfy (2) with high probability, provided n satisﬁes the

condition of the Lemma 1 [21]:

• n×N random matrix A whose entries a

i,j

are independent realizations of Gaussian random variables

i,j

∼ N (0,

• Independent realizations of ±1 Bernoullie random variables

i,j











√

with probability

−

√

with probability

• Independent realizations of related distributions such as

i,j











with probability

0 with probability

−

with probability

A random projection-based cancelable biometric method for iris recognition was proposed in [17].

Applying the random projections directly on the iris images usually degrades the performance due to

the following reasons. First of all in real iris images, despite good segmentation algorithms, there will

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 8

still be some outliers due to specular reﬂections, eye lashes and eyelids. Also, different parts of the iris

have different quality. By taking a linear transformation of the entire vector, one combines the good iris

regions as well as the outliers and thereby corrupts the data. To deal with this, [17] proposes Sectored

Random Projections (SRP) in which random projections are applied separately on each sector and the

resulting transformed vectors are concatenated to form the cancelable template. As a result, outliers can

corrupt only the corresponding sector and not the entire iris vector.

Fig. 3: An overview of the SRP method [17].

Figure 3 shows an overview of this method [17]. The enrollment system extracts the iris pattern of

the user, computes the Gabor features, applies a different random projection for each application and

transfers the new pattern to the application database. Note that even if the transformed pattern and the

key (i.e. the projection matrix) are stolen, the user’s iris pattern cannot be generated from them due to the

dimension reduction caused by the projection. Also even if a hacker steals the user’s iris pattern either

from the client system or using a hidden scanner, without knowing the random projection he/she cannot

generate the transformed patterns required by the application. During the veriﬁcation stage, the application

obtains the iris image and the random projection matrix from the user, computes the transformed pattern

and compares it with the ones in its database. In case, the random projection matrix or the transformed

patterns are compromised, one can create a new random projection matrix and obtain a new transformed

pattern which can be updated into the application database. Instead of the user providing the random

matrix during veriﬁcation, the application can generate and store it along with the cancelable template in

its database. Though this will be an easier scheme for the user to operate, it is less secure as a hacker can

get both the random projection matrices and the transformed patterns by breaking into the application

database.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 9

The approach of SRP [17] was later extended in [18] using sparse representation-based classiﬁcation.

It was shown that the sparsity patterns that one obtains before and after applying random projections are

similar. As a result, cancelable biometric templates can be directly used for authentication rather than the

original ones without degrading the performance of a sparse representation-based classiﬁcation algorithm.

C. Cancelable Biometric Filters

Motivated by the success of the correlation ﬁlter-based methods in pattern recognition and computer

vision applications [22], a random convolution method for generating cancelable biometric templates

was proposed in [23]. The idea is to encrypt biometric templates using random user speciﬁc convolution

kernels. The training images are convolved with a random convolution kernel. The seed used to generate

the random convolution kernel is used as the PIN. The convolved training images are then used to generate

a Minimum Average Correlation Energy (MACE) biometric ﬁlter. This encrypted ﬁlter is stored and used

for authentication. Figure 4(a) shows the enrollment stage using this method.

(a) (b)

Fig. 4: Correlation ﬁlter-based approach to cancelable biometrics [23]. (a) Enrollment stage for encrypted

ﬁlters. (b) Authentication stage using encrypted MACE ﬁlters.

During the recognition stage, the user presents the PIN and the encrypted ﬁlter which is used to

generate the convolution kernel. This random convolution kernel is convolved with the test face images

presented by the user. The convolved test images are cross-correlated with the encrypted MACE ﬁlter

and the resulting correlation outputs are used to authenticate the user. Figure 4(b) shows authentication

stage for this method.

It was shown that convolving the training images with any random convolution kernel prior to building

the MACE ﬁlters used for biometric recognition does not change the resulting correlation output [23].

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 10

As a result, the recognition accuracy is maintained. Furthermore, different cancelable biometric templates

can be generated from the same biometric by simply changing the convolution kernels.

Other correlation-based cancelable biometric methods include correlation invariant random ﬁltering

(CIRF) [24], [25] which was shown to have almost the same accuracy as the conventional ﬁngerprint

veriﬁcation based on the chip matching algorithm.

D. BioConvolving

Another convolution-based approach for generating cancelable biometric templates was recently pro-

posed in [26]. This method is applicable to any biometric whose template can be represented by a set of

sequences. In this method, each transformed sequence f

(i)

[n], i = 1, · · · , F, is obtained from the corre-

sponding original sequence r

(i)

[n], i = 1, · · · , F, which represents a generic discrete sequence of length N

belonging to the original biometric template. In particular, a number (W −1) of different integer values d

between 1 and 99 are randomly selected, ordered in ascending order such that d

> d

j−1

, j = 1, · · · , W,.

These numbers are arranged in a vector d = [d

, · · · , d

]

, where d

and d

are set to 0 and 100,

respectively. Here, the vector d represents the key of the transformation. The original sequence r

(i)

[n] is

divided into W non-overlapping segments r

(i)j,N

[n] of length N

= b

− b

j−1

(i)j,N

[n] = r

(i)

[n + b

j−1

], n = 1, · · · , N

, j = 1, · · · , W, (3)

where



100



, j = 1, · · · , W. (4)

A transformed sequence f

(i)

[n], n = 1, · · · , K, is then obtained through the linear convolution of the

sequences r

(i)j,N

[n], j = 1, · · · , W as

(i)

[n] = r

(i)1,N

[n] ∗ · · · ∗ r

(i)W,N

[n]. (5)

Each original sequence r

(i)

[n], i = 1, · · · , F undergoes the same decomposition before applying convolu-

tions. As a result, the length of the transformed sequences is equal to K = N − W + 1. A normalization

is applied to make the transformed sequences zero mean and unit standard deviation. Different templates

can be generated from the original biometric template by simply changing the size or the values of the

parameter key d. Figure 5 shows an example of a feature transformation where W = 3 [26]. See [26]

for more details on different ways of generating transformed sequences, invertibility analysis and their

application in signature-based authentication.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 11

Fig. 5: Sequence transformation using the BioConvolving approach [26], where W = 3.

E. Bloom Filters

Recently, Bloom ﬁlter-based cancelable biometric template protection schemes were proposed in [27],

[28], [29]. A Bloom ﬁlter is essentially a space-efﬁcient probabilistic data structure representing a set

in order to support membership queries. In particular, alignment-free cancelable iris biometric templates

based on adaptive Bloom ﬁlters were introduced in [27] in which the generic adaptive Bloom ﬁlter-

based transform is applied to binary feature vectors of different iris recognition algorithms. It was shown

that such a method can enable template protection, compression of biometric data, and computationally

efﬁcient biometric identiﬁcation. Furthermore, rotation-invariant Bloom ﬁlter-based transform can provide

a high level of security while maintaining recognition accuracy [27].

F. Knowledge Signatures

Voice-based cancelable biometric templates using knowledge signatures were proposed in [30]. The

idea is based on a group signature scheme which allows members of a group to sign messages on the

group’s behalf such that the resulting signature does not reveal their identity. They consider voiceprint as

the knowledge of the user and the user’s voiceprint transmitted to the template which isn’t the original

feature, but a signature of knowledge. Legitimate signatures can not be generated without factorizing

a large integer and the original feature. As a result, an individual’s privacy can be protected. We refer

readers to [31] and [30] for more details on knowledge signatures and their uses in generating cancelable

biometric templates for voiceprints.

G. BioHashing Methods

BioHashing methods are essentially an extension of random projection. In BioHashing [9], [32], [33],

[34], [35], [36], [37], [38] feature extraction method such as wavelet transform is ﬁrst used to extract

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 12

the biometric feature x ∈ R

from the input biometric data. Using a user speciﬁc Tokenized Random

Numbers (TRN), n orthogonal pseudo-random vectors, b

∈ R

, i = 1, · · · , n, are generated, where

n ≤ N. Then, the dot product of the feature vector and all the random vectors is calculated. Finally, a

binary discretization is applied to compute n bit BioHash template as

c = Sig

− τ

, (6)

where Sig(·) is deﬁned as a signum function and τ is an empirically determined threshold. Eq. (6)

only applies to a user who holds the user-speciﬁc random vectors b

∈ R

, i = 1, · · · , n, and thus

the formulation can be extended to introduce an ensemble of random subspaces, where each subspace

represents different individual k. The resulting BioHash is given as

= Sig

− τ

, k = 1, · · · , g, (7)

where g is the total number of users in the system. Finally, the BioHash code is compared by the Hamming

distance for the similarity matching. Figure 6 shows the progression of BioHshing [33]. The BioHashing

framework is demonstrated to be a one-way transform, hence providing a high degree of security to the

biometric and external factors. A detailed statistical analysis of the BioHashing framework in terms of

random multispace quantization operations can be found in [9].

Fig. 6: Overview of BioHashing [33].

H. Random Permutations

Another common approach for generating cancelable biometric templates is based on random permu-

tation of features. In [39], two such methods were proposed for generating cancelable iris templates. The

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 13

ﬁrst method, namely GRAY-COMBO, transforms the Gabor features by circularly shifting and adding

rows at random. BIN-COMBO, the second method, applies similar transformations on the iris codes

by random shifting and XOR-ing. As pointed out by the authors, these methods gradually reduce the

amount of information available for recognition. Since these methods employ linear transformations on

the Gabor feature vectors, they are also sensitive to outliers in the form of eyelids, eye lashes and specular

reﬂections. [18] proposes to overcome this limitation by dividing the feature into different regions and

permuting them randomly in a dictionary. Without prior knowledge of the locations of sectored features in

a dictionary, it is impossible to perform recognition. A similar approach was also proposed in [16] where

each block of the target texture is mapped to a block from the source texture. In this method, a re-mapping

of blocks instead of a permutation is performed, as it is not reversible. Source blocks which are not part

of the mapping are not contained in the transformed texture. As a result, it is impossible to reconstruct the

original iris texture. Another permutation-based cancelable method for ﬁngerprint biometric was presented

in [40]. This method permutes a binary vector obtained from ﬁngerprint features and stores them in the

database. During authentication, the binary vector obtained from the ﬁngerprints of the user are permuted

using the key provided by the user and matched with the database.

In these methods, key security is essential for protecting privacy of individuals. One of the advantages

of these methods is that since permutations are merely rearranging the feature vector, authentication

accuracy is not affected by these operations.

I. Salting Methods

One of the simplest ways of generating cancelable biometric templates is by simply mixing in a

totally artiﬁcial pattern. The mixing patterns can be pure random noise, a random pattern or a synthetic

pattern. Two such salting methods were proposed in [39] for iris recognition namely GRAY-SALT and

BIN-SALT. These methods add random patterns or synthetic iris patterns to the Gabor features and

iris codes, respectively. Unlike GRAY-COMBO and BIN-COMBO permutation-based methods, they

do not suffer from the problem of outlier ampliﬁcation and reduction of useful area. However, it is

difﬁcult to decide the relative strength of the noise patterns to be added. Adding very strong patterns

will reduce the discriminative capacity of the original iris patterns and hence lead to lower recognition

results. Adding weaker patterns can lower the non-invertibility property, making it easier to extract useful

information about the original iris biometric from the transformed patterns. Also, if the added patterns

are compromised, the original iris patterns could be extracted from the transformed patterns by a simple

subtraction operation.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 14

J. Hybrid Methods

Several biometric template protection approaches make use of both cryptosystems and cancelable

biometrics [41], [42]. One such hybrid system was proposed in [41] for face biometric. They introduced

biotoken which is “the revocable identity token produced by applying a revocable transform to biometric

data, such that identity matching is done in the encoded/revocable form” [41], [42]. Speciﬁcally, this

approach combines the ideas of transformation of data, robust learning measures and encryption of

biometric data. The method essentially separates the data into two parts, the fractional part which is

retained for local distance computation and the integer part which is encrypted. It was shown that for

face biometric this method signiﬁcantly improved the performance of the PCA and LDA algorithms. This

work was later extended for ﬁngerprints in [42].

K. Summary of Cancelable Biometric Template Protection Schemes

The cancelable biometric template protection schemes reviewed in this paper can be broadly divided

into two main categories as shown in Figure 7 - methods that require a special matcher and methods that

can work with the existing matchers. These schemes can be further classiﬁed into two categories, namely,

registration free methods and methods that require good registration of biometric samples. Finally, these

methods can be further divided into two types of schemes - schemes that work with the original biometric

samples (denoted as signal) and schemes that work with the features extracted from the biometric signals

(denoted as feature).

Among the methods that require a special matcher and good registration of biometric samples, Bioto-

kens [41], [42] is a signal-based method and BioConvolving [26], salting, PalmHashing [37] and PalmPha-

sor [37] methods are feature-based methods. On the other hand, correlation-based MACE ﬁlter approach

[23] is a signal-based, registration-free method that requires a special matcher. Combo [39], block-

remapping [16], image warping [16], non-invertible transforms [4] and dynamic RP [43] methods fall

under signal-based methods that require registration and can work with the existing matchers. Whereas

permutations [18], RP [17], BioHashing [35], [9] and PalmHashing [34] methods are feature-based that

can work with the existing matchers and require good registration. Registration-free methods that can

work with existing matchers include minimum distance graph [44] and curtailed circular convolution [45]

methods which are signal-based and a registration free approach proposed in [15] which is a feature-based

method.

Furthermore, Table I summarizes key cancelable biometric template protection approaches in terms of

their performances on various biometric datasets. Note that the performances of different methods are

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 15

Fig. 7: Categorization of cancelable biometric template protection schemes.

reported in terms of False Rejection Rate (FRR), Equal Error Rate (EER), rank-1 Recognition Rate (RR),

Genuine Accept Rate (GAR), and False Acceptance Rate (FAR).

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 16

TABLE I: Key cancelable biometric template protection schemes.

Method Biometric Dataset (Subjects) Performance Remarks

Non-invertible transforms [4] Fingerprint IBM-99 (188) FRR: ∼ 35, 15, 15 -

Image warping [16] Iris CASIA Iris V3 (396) EER: 1.6 - 6 -

Random projections [17] Iris MMU1 dataset (100) RR: 97.7 -

Biometric ﬁlters [23] Face CMU PIE (65) RR: 100 -

BioConvolving [26] Online signature MCYT (330) EER: 6.33 - 15.40 -

BioHashing [35] Fingerprint FVC 2002 (100) EER: ∼ 0 FAR: 0

PalmHashing [34] Palmprint Palmprint dataset (50) EER: 0-0.222 FAR: 0

BioHashing [9] Face FERET (1196) EER: 0.002 - 7.51 -

GRAY-COMBO [39] Iris MMU1 dataset (100) GAR: ∼ 0.995 FAR : 10

−4

BIN-COMBO [39] Iris MMU1 dataset (100) GAR: ∼ 0.965 FAR : 10

−4

Block re-mapping [16] Iris CASIA Iris V3 (396) EER: 0.2 - 1.6 -

GRAY-SALT [39] Iris MMU1 dataset (100) GAR: ∼ 1 FAR : 10

−4

BIN-SALT [39] Iris MMU1 dataset (100) GAR: ∼ 0.995 FAR : 10

−4

Atom permutations [18] Iris ND-IRIS-0405 (356) RR: 99.17 -

Biotokens [42] Fingerprint FCV 2000-4 (100) EER: 0.012-0.086 Hybrid method

Biotokens [41] Face FERET (1196) EER: 0.9997 Hybrid method

Dynamic random projections [43] Fingerprint FVC2002DB2-A (800) EER: ∼ 0.05 -

PalmHash code [37] Palmprint PolyU dataset (7752) EER: 0.38 2D PalmHash code

PalmPhasor code [37] Palmprint PolyU dataset (7752) EER: 0.32 2D PalmPhasor code

Minimum Distance Graph [44] Fingerprint FVC2002-DB1a,b (100) EER: 0.0227 -

Curtailed circular convolution [45] Fingerprint FVC2002-DB1,2,3 (100) EER: 0.02, 0.03, 0.0612 -

III. ATTACKS AGAINST CANCELABLE BIOMETRIC TEMPLATES

A generic biometric system consists of a sensor, a feature extraction module, a biometric template

database, a matcher module, and an application device which is driven by the matcher’s response.

Researchers have identiﬁed different points of attacks in a biometric system as shown in Figure 8. The

attacks can come in various forms such as: trojan horse attack, front end attack, phishing and farming

attacks, back end attack and communication channel attack. The unauthorized access to raw biometric

templates is among the most serious threats to users’ privacy and security. Some of the attacks can be

averted using cancelable biometric systems while some of them are extremely difﬁcult to detect. See [46]

for more details on different types of attacks.

The cancelable biometric systems can be attacked by exposing the parameter (key) of the transformation

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 17

Fig. 8: Possible attack points in a generic biometric system [46], [8].

being applied to biometric templates. In the case when the transformation is invertible, the original

biometric can be reconstructed. In this case, security is in the secrecy of the key. If the transformation

is not invertible, then an attacker can try to approximately recover the original biometric templates. For

instance, it was shown in [47], [48], [49] that face images can be restored from encrypted templates.

Attacks against the cancelable system using non-invertible transforms [4] are proposed in [50]. It was

argued that when multiple transformed templates are generated from the same original template, they

can be cracked by a method known as Attack via Record Multiplicity. In particular, given a transformed

template, an attacker can ﬁnd the inverse solutions by inverting the transformation. Due to many-to-one

property of transform functions, there may be several solutions out of which is the original solution. The

attacker can come up with a way to to pick out the right solution. A similar dictionary attack method is also

proposed in [51] to recover the original templates from the cancelable templates. Also, convolution-based

cancelable biometric systems’ [26], [23] security depends on how well blind deconvolution algorithms

are able to recover the original biometric templates.

Several vulnerabilities in BioHashing-based systems have also been investigated [52], [53], [54], [55].

One of the major limitations of BioHashing methods is their low performance when attackers are in

possession of secret key. To deal with this problem, [52] proposed an improved BioHashing method

which is more robust than the original BioHashing method [32], [33]. In [53], it was shown that even

without having a genuine users’ private random vectors, a preimage of a BioCode can be easily calculated

from a lost BioCode. As a result, an attacker can gain an illegal access to a system. It was observed

that simple data dimension reduction and discretization as is done in most BioHashing methods may

be vulnerable to preimage attacks. Similarly, a new way to approximate the original biometric feature

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 18

from the transformed template in a cancelable biometric scheme was recently proposed in [54]. Their

method is based on a genetic algorithm which essentially determines the optimal value of a criterion by

simulating the evolution of a population and survival of best ﬁtted individuals [54]. It was shown that a

genetic algorithm can allow an intruder to recover a biometric template, similar to the original template,

under some realistic assumptions.

In a related work, [43] analyzed the security concerns over random projection-based cancelable systems

[49], [56] and proposed a dynamic random projection method to alleviate these concerns by forming a

non-linear projection process which relates the random matrix’s assembly to the biometric feature vector

itself. The dynamic random projection method greatly increases the computational complexity to apply

inversion attacks in the token stolen cases. Furthermore, it was shown that this method does not degrade

the biometric performance compared with the ﬁxed matrix-based random projection [43].

In recent years, several biometrics protection schemes have been proposed in the literature that attempt

to protect the privacy of biometric templates without using a key [57], [58], [59], [60]. For instance, a

visual cryptography method is introduced in [58] which decomposes a biometric image into two noise-

like images, called sheets, that are stored in two different databases. During the authentication, the two

sheets are overlaid to create a temporary image for matching. One of the limitations of this method is

that it requires two separate databases to work together, which may not be practical in some applications.

Another method for protecting ﬁngerprint biometric combines two ﬁngerprints from two different ﬁngers

to generate a new template [57]. For authentication, two query ﬁngerprints are required and a two-stage

matching process is proposed for matching the two query ﬁngerprints against a combined template. One

of the advantages of this method is that by using the combined template, the complete minutiae feature

of a single ﬁngerprint will not be compromised when the database is stolen [57].

In order to deploy a biometric template protection system, one needs to investigate the security strength

of the template transformation technique and deﬁne metrics that facilitate security evaluation. Towards

this end, six different evaluation metrics were deﬁned in [49]. Furthermore, the security of BioHashing

and cancelable ﬁngerprint templates were analyzed based on these metrics. It was reported that both

these schemes are vulnerable to intrusion and linkage attacks because it is relatively easy to obtain either

an approximation of the original biometric template in the case of BioHashing or a preimage of the

transformed template in the case of cancelable ﬁngerprints.

In a related work, [61] presents several evaluation criteria, metrics and testing methodologies for as-

sessing biometric template protection algorithms. In particular, criteria such as accuracy of the recognition

algorithm, throughput, storage requirements, performance degradation of a biometric template protection

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 19

scheme, diversity and error rate of failing to generate a protected template are discussed in detail. These

deﬁnitions will help researchers in designing robust biometric template protection schemes.

IV. CONCLUSIONS AND FUTURE DIRECTIONS

Often we use standard information security tools such as encryption or secure hashing methods to

protect the biometric content. There are two issues with this approach. First, as the biometrics data

(image, template) constantly change with every sample acquisition, the encrypted biometrics has to be

decrypted for matching. If it is decrypted, that opens an opportunity for the hacker to attack at the

output point of the decryption. If a secure hash function is used, the matching of the secure hashes

is totally useless as biometrics signals never reproduce exactly. While the hash will be best in terms

of privacy, the biometrics matching will not ever produce the positive authentication result. Cancelable

biometrics is inspired by this approach but handles biometric variability. The transformation management

in cancelable biometrics is equivalent to key management in information security. For example, a part

of the transform can be retained by the user, another part can stay with the authentication system. Until

the two come together, the biometrics authentication can’t take place. But the keys in encryption or hash

functions are derived totally differently than the cancelable biometrics transform. Secondly, because of

the special construction, the matching of the cancelable biometrics signal or template is carried out in

the transformed domain. In fact, the original biometrics signal is not required to be retained as both

enrollment and authentication is carried out using the transformed biometrics.

This article presented a review of recent developments in such template protection schemes which

included non-invertible transform-based methods, BioHashing and hybrid methods. There are several

challenges to be overcome before successfully designing a cancelable biometric system. Below we list a

few.

1) In order for the transform to be repeatable, the biometric signal must be positioned in the same

coordinate system each time. This requires that an absolute registration be done for each biometric

signal acquired prior to the transformation. Registration-free cancelable biometric systems have

also been proposed in the literature [15], [44], [62], [45]. However, some of these methods do

not perform well in practice. For instance, a registration-free construction of cancelable ﬁngerprint

biometric templates [15] exhibits lower veriﬁcation performances than the one proposed in [4]

which requires registration. More robust registration free non-invertible transform and BioHashing

methods are needed.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 20

2) The recently introduced theory of compressive sampling allows one to reconstruct the original

signal from a few random measurements provided that certain conditions are met. Many cancelable

biometric template protection systems make use of random projections [17], [18], [32], [33].

It remains an interesting problem to study the vulnerability of such cancelable systems using

compressive sampling.

3) In the past few years, we have witnessed an exponential growth in the use of mobile devices such

as smartphones and tablets. Most mobile devices use passwords, pin numbers, or secret patterns

for authenticating users. As long as the device remains active, there is no mechanism to verify that

the user originally authenticated is still the user in control of the device. As a result, unauthorized

individuals may improperly gain access to personal information of the user if the password is

compromised. Active authentication systems deal with this issue by continuously monitoring the

user identity after the initial access has been granted. Examples include systems based on screen

touch gestures [63], gait recognition [64], and device movement patterns (as measured by the

accelerometer) [65]. Development of cancelable active authentication systems is a nascent area of

research.

4) Blind deconvolution is an extremely ill-posed problem in which one attempts to recover the original

signal from convolved outputs without the explicit knowledge of the convolution kernel. Recent

advances in signal processing community have shown that one can approximate the convolution

kernel directly from the observations. These methods exploit some underlying structure of signals

such as sparsity. It remains to be seen whether convolution-based cancelable systems are robust to

these blind deconvolution methods.

5) Most cancelable biometric template protection schemes have been evaluated on small and mid-

size datasets consisting of hundreds and thousands of samples. However, in order to really see the

signiﬁcance and impact of various biometric template protection schemes, they need to be evaluated

on large-scale datasets containing millions of samples.

6) As the research community advances biometric template protection schemes, third party evaluation

for security attacks and evaluation of the revocable methods are needed. Some efforts are being

made [66], however, more standardization efforts are needed to establish guidelines and procedures

for testing and evaluating various cancelable biometric systems.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 21

AUTHOR BIOGRAPHIES

Vishal M. Patel (Ph.D., UMD, 2010) is a member of the research faculty at the University of Maryland

Institute for Advanced Computer Studies (UMIACS). His research interests are in signal processing,

computer vision and machine learning with applications to biometrics and object recognition. Dr. Patel

was a recipient of the ORAU postdoctoral fellowship in 2010. He is a member of the IEEE, Eta Kappa

Nu, Pi Mu Epsilon, and Phi Beta Kappa.

Nalini K. Ratha (Ph.D., MSU 1996) is a Research Staff Member at the IBM Thomas J. Watson Research

Center, Yorktown Heights, New York where he leads the biometrics research efforts in building efﬁcient

biometrics systems. In addition to more than 80 publication in peer reviewed journals and conferences,

and co-inventor on 12 patents, he has co-edited two books on biometrics recognition. He is a Fellow of

IEEE, Fellow of IAPR and Senior Member of ACM. His current research interests include biometrics,

computer vision, pattern recognition and special purpose architecture for computer vision systems.

Rama Chellappa (Ph.D., Purdue, 1981) is a Professor and Chair of Electrical and Computer Engineering

(ECE) and an afﬁliate Professor of Computer Science at the University of Maryland (UMD), College

Park. He is also afﬁliated with the Center for Automation Research, the Institute for Advanced Computer

Studies (Permanent Member) and a Minta Martin Professor of Engineering. He is a Fellow of IEEE,

IAPR, OSA, AAAS, ACM and AAAI. His current research interests are clustering, 3D modeling from

video, image and video-based recognition of objects, dictionary-based inference, and domain adaptation.

REFERENCES

[1] N. K. Ratha, “Privacy protection in high security biometrics applications,” in Ethics and Policy of Biometrics, ser. Lecture

Notes in Computer Science, A. Kumar and D. Zhang, Eds. Springer Berlin Heidelberg, 2010, vol. 6005, pp. 62–69.

[2] G. I. Davida, Y. Frankel, and B. J. Matt, “On enabling secure applications through off-line biometric identiﬁcation,” IEEE

Symposium on Security and Privacy, pp. 148–157, May 1998.

[3] N. K. Ratha, J. H. Connel, and R. Bolle, “Enhancing security and privacy in biometrics-based authentication systems,”

IBM Systems Journal, vol. 40, no. 3, pp. 614–634, 2001.

[4] N. Ratha, S. Chikkerur, J. Connell, and R. Bolle, “Generating cancelable ﬁngerprint templates,” IEEE Transactions on

Pattern Analysis and Machine Intelligence, vol. 29, no. 4, pp. 561 –572, April 2007.

[5] R. M. Bolle, J. H. Connel, and N. K. Ratha, “Biometrics perils and patches,” Pattern Recognition, vol. 35, no. 12, pp.

2727–2738, 2002.

[6] A. Juels and M. Sudan, “A fuzzy vault scheme,” Designs, Codes and Cryptography, vol. 38, no. 2, pp. 237–257, Feb.

2006.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 22

[7] Y. Sutcu, Q. Li, and N. Memon, “Protecting biometric templates with sketch: Theory and practice,” IEEE Transactions on

Information Forensics and Security, vol. 2, no. 3, pp. 503–512, Sept 2007.

[8] A. K. Jain, K. Nandakumar, and A. Nagar, “Biometric template security,” EURASIP Journal on Advances in Signal

Processing, vol. 2008, pp. 113:1–113:17, Jan 2008.

[9] A. Teoh, A. Goh, and D. Ngo, “Random multispace quantization as an analytic mechanism for biohashing of biometric and

random identity inputs,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, no. 12, pp. 1892–1901,

Dec. 2006.

[10] U. Uludag, S. Pankanti, S. Prabhakar, and A. Jain, “Biometric cryptosystems: issues and challenges,” Proceedings of the

IEEE, vol. 92, no. 6, pp. 948–960, June 2004.

[11] A. Juels and M. Wattenberg, “A fuzzy commitment scheme,” in ACM Conference on Computer and Communications

Security. New York, NY, USA: ACM, 1999, pp. 28–36.

[12] S. Draper, A. Khisti, E. Martinian, A. Vetro, and J. S. Yedidia, “Using distributed source coding to secure ﬁngerprint

biometrics,” in IEEE International Conference on Acoustics, Speech and Signal Processing, vol. 2, April 2007, pp. II–

129–II–132.

[13] Y. Dodis, L. Reyzin, and A. Smith, “Fuzzy extractors: How to generate strong keys from biometrics and other noisy data,”

in Advances in cryptology-Eurocrypt 2004. Springer Berlin Heidelberg, 2004, pp. 523–540.

[14] C. Rathgeb and A. Uhl, “A survey on biometric cryptosystems and cancelable biometrics,” EURASIP Journal on Information

Security, vol. 2011, no. 3, pp. 1–25, 2011.

[15] S. Chikkerur, N. Ratha, J. Connell, and R. Bolle, “Generating registration-free cancelable ﬁngerprint templates,” in IEEE

International Conference on Biometrics: Theory, Applications and Systems, Sept 2008, pp. 1–6.

[16] J. Hmmerle-Uhl, E. Pschernig, and A. Uhl, “Cancelable iris biometrics using block re-mapping and image warping,” in

Information Security, ser. Lecture Notes in Computer Science, P. Samarati, M. Yung, F. Martinelli, and C. Ardagna, Eds.

Springer Berlin Heidelberg, 2009, vol. 5735, pp. 135–142.

[17] J. K. Pillai, V. M. Patel, R. Chellappa, and N. K. Ratha, “Sectored random projections for cancelable iris biometrics,” in

IEEE International Conference on Acoustics Speech and Signal Processing, March 2010, pp. 1838–1841.

[18] ——, “Secure and robust iris recognition using random projections and sparse representations,” IEEE Transactions on

Pattern Analysis and Machine Intelligence, vol. 30, no. 9, pp. 1877–1893, Sept. 2011.

[19] W. Johnson and J. Lindenstrauss, “Extensions of lipschitz maps into a hilbert space,” in Contemporary Mathematics, 1984,

pp. 189–206.

[20] S. Dasgupta and A. Gupta, “An elementary proof of a theorem of johnson and lindenstrauss,” Random Structures &

Algorithms, vol. 22, no. 1, pp. 60–65, Jan 2003.

[21] D. Achlioptas, “Database-friendly random projections: Johnson-lindenstrauss with binary coins,” Journal of Computer and

System Sciences, vol. 66, no. 4, pp. 671–687, Jun 2003.

[22] B. V. K. V. Kumar, A. Mahalanobis, and R. D. Juday, Correlation Pattern Recognition. Cambridge University Press,

2005.

[23] M. Savvides, B. Kumar, and P. Khosla, “Cancelable biometric ﬁlters for face recognition,” in International Conference on

Pattern Recognition, vol. 3, Aug 2004, pp. 922–925 Vol.3.

[24] K. Takahashi and S. Hirata, “Cancelable biometrics with provable security and its application to ﬁngerprint veriﬁcation,”

IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, vol. 94-A, no. 1, pp. 233–

244, 2011.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 23

[25] S. Hirata and K. Takahashi, “Cancelable biometrics with perfect secrecy for correlation-based matching,” in Advances in

Biometrics, ser. Lecture Notes in Computer Science, M. Tistarelli and M. Nixon, Eds. Springer Berlin Heidelberg, 2009,

vol. 5558, pp. 868–878.

[26] E. Maiorana, P. Campisi, J. Fierrez, J. Ortega-Garcia, and A. Neri, “Cancelable templates for sequence-based biometrics

with application to on-line signature recognition,” IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems

and Humans, vol. 40, no. 3, pp. 525–538, May 2010.

[27] C. Rathgeb, F. Breitinger, C. Busch, and H. Baier, “On the application of bloom ﬁlters to iris biometrics,” IET Journal on

Biometrics, vol. 3, no. 4, pp. 207–218, 2014.

[28] C. Rathgeb and C. Busch, “Cancelable multi-biometrics: Mixing iris-codes based on adaptive bloom ﬁlters,” Computers

& Security, vol. 42, no. 0, pp. 1 – 12, 2014.

[29] C. Rathgeb, F. Breitinger, and C. Busch, “Alignment-free cancelable iris biometric templates based on adaptive bloom

ﬁlters,” in IAPR International Conference on Biometrics, June 2013, pp. 1–8.

[30] W. Xu, Q. He, Y. Li, and T. Li, “Cancelable voiceprint templates based on knowledge signatures,” in International

Symposium on Electronic Commerce and Security, Aug 2008, pp. 412–415.

[31] J. Camenisch and M. Stadler, “Efﬁcient group signature schemes for large groups,” in International Cryptology Conference

on Advances in Cryptology. London, UK, UK: Springer-Verlag, 1997, pp. 410–424.

[32] A. B. J. Teoh and C. T. Yuang, “Cancelable biometrics realization with multispace random projections,” IEEE Transactions

on Systems, Man, and Cybernetics, Part B, vol. 37, no. 5, pp. 1096–1106, 2007.

[33] A. B. Teoh, Y. W. Kuan, and S. Lee, “Cancellable biometrics and annotations on biohash,” Pattern Recognition, vol. 41,

no. 6, pp. 2034 – 2044, 2008.

[34] T. Connie, A. Teoh, M. Goh, and D. Ngo, “Palmhashing: a novel approach for cancelable biometrics,” Information

Processing Letters, vol. 93, no. 1, pp. 1 – 5, 2005.

[35] A. Teoh, D. N. C. Ling, and A. Goh, “Biohashing: two factor authentication featuring ﬁngerprint data and tokenised random

number,” Pattern Recognition, vol. 37, no. 11, pp. 2245 – 2255, 2004.

[36] A. Kong, K.-H. Cheung, D. Zhang, M. Kamel, and J. You, “An analysis of biohashing and its variants,” Pattern Recognition,

vol. 39, no. 7, pp. 1359 – 1368, 2006.

[37] L. Leng and J. Zhang, “Palmhash code vs. palmphasor code,” Neurocomputing, vol. 108, no. 0, pp. 1 – 12, 2013.

[38] L. Leng, A. B. J. Teoh, M. Li, and M. K. Khan, “Analysis of correlation of 2dpalmhash code and orientation range suitable

for transposition,” Neurocomputing, vol. 131, no. 0, pp. 377 – 387, 2014.

[39] J. Zuo, N. Ratha, and J. Connell, “Cancelable iris biometric,” International Conference on Pattern Recognition, pp. 1–4,

2008.

[40] F. Farooq, R. Bolle, T.-Y. Jea, and N. Ratha, “Anonymous and revocable ﬁngerprint recognition,” in IEEE Conference on

Computer Vision and Pattern Recognition, June 2007, pp. 1–7.

[41] T. Boult, “Robust distance measures for face-recognition supporting revocable biometric tokens,” in International

Conference on Automatic Face and Gesture Recognition, April 2006, pp. 560–566.

[42] T. Boult, W. Scheirer, and R. Woodworth, “Revocable ﬁngerprint biotokens: accuracy and security analysis,” in IEEE

Conference on Computer Vision and Pattern Recognition, June 2007, pp. 1–8.

[43] B. Yang, D. Hartung, K. Simoens, and C. Busch, “Dynamic random projection for biometric template protection,” in IEEE

International Conference on Biometrics: Theory Applications and Systems, Sept 2010, pp. 1–7.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 24

[44] P. Das, K. Karthik, and B. C. Garai, “A robust alignment-free ﬁngerprint hashing algorithm based on minimum distance

graphs,” Pattern Recognition, vol. 45, no. 9, pp. 3373 – 3388, 2012.

[45] S. Wang and J. Hu, “Design of alignment-free cancelable ﬁngerprint templates via curtailed circular convolution,” Pattern

Recognition, vol. 47, no. 3, pp. 1321 – 1329, 2014.

[46] N. Ratha, J. Connell, and R. Bolle, “An analysis of minutiae matching strength,” in Audio- and Video-Based Biometric

Person Authentication, ser. Lecture Notes in Computer Science, J. Bigun and F. Smeraldi, Eds. Springer Berlin Heidelberg,

2001, vol. 2091, pp. 223–228.

[47] A. Luong, M. Gerbush, B. Waters, and K. Grauman, “Reconstructing a fragmented face from a cryptographic identiﬁcation

protocol,” in IEEE Workshop on Applications of Computer Vision, Jan 2013, pp. 238–245.

[48] A. Adler, “Sample images can be independently restored from face recognition templates,” in IEEE Canadian Conference

on Electrical and Computer Engineering, vol. 2, May 2003, pp. 1163–1166 vol.2.

[49] A. Nagar, K. Nandakumar, and A. K. Jain, “Biometric template transformation: A security analysis,” in Proceedings of

SPIE, Electronic Imaging, Media Forensics and Security XII, vol. 7541, 2010, pp. 75 410O–75 410O–15.

[50] F. Quan, S. Fei, C. Anni, and Z. Feifei, “Cracking cancelable ﬁngerprint template of ratha,” in International Symposium

on Computer Science and Computational Technology, vol. 2, Dec 2008, pp. 572–575.

[51] S. Shin, M.-K. Lee, D. Moon, and K. Moon, “Dictionary attack on functional transform-based cancelable ﬁngerprint

templates,” ETRI Journal, vol. 31, no. 5, pp. 628–630, 2009.

[52] A. Lumini and L. Nanni, “An improved biohashing for human authentication,” Pattern Recognition, vol. 40, no. 3, pp.

1057 – 1065, 2007.

[53] Y. Lee, Y. Chung, and K. Moon, “Inverse operation and preimage attack on biohashing,” in IEEE Workshop on

Computational Intelligence in Biometrics: Theory, Algorithms, and Applications, March 2009, pp. 92–97.

[54] P. Lacharme, E. Cherrier, and C. Rosenberger, “Reconstruction attack on biohashing,” in International Conference on

Security and Cryptography, 2013.

[55] R. Belguechi, E. Cherrier, and C. Rosenberger, “Texture based ﬁngerprint biohashing: Attacks and robustness,” in IAPR

International Conference on Biometrics, March 2012, pp. 196–201.

[56] X. Zhou and T. Kalker, “On the security of biohashing,” in Proceedings of SPIE, Electronic Imaging, Media Forensics

and Security II, vol. 7541, 2010, pp. 75 410Q–75 410Q–8.

[57] S. Li and A. Kot, “Fingerprint combination for privacy protection,” IEEE Transactions on Information Forensics and

Security, vol. 8, no. 2, pp. 350–360, Feb 2013.

[58] A. Ross and A. Othman, “Visual cryptography for biometric privacy,” IEEE Transactions on Information Forensics and

Security, vol. 6, no. 1, pp. 70–81, March 2011.

[59] A. Othman and A. Ross, “On mixing ﬁngerprints,” IEEE Transactions on Information Forensics and Security, vol. 8, no. 1,

pp. 260–267, Jan 2013.

[60] ——, “Privacy of facial soft biometrics: Suppressing gender but retaining identity,” in European Conference on Computer

Vision Workshops, vol. 8926, 2015, pp. 682–696.

[61] K. Simoens, B. Yang, X. Zhou, F. Beato, C. Busch, E. Newton, and B. Preneel, “Criteria towards metrics for benchmarking

template protection algorithms,” in IAPR International Conference on Biometrics, March 2012, pp. 498–505.

[62] N. Zhang, X. Yang, Y. Zang, X. Jia, and J. Tian, “Generating registration-free cancelable ﬁngerprint templates based on

minutia cylinder-code representation,” in IEEE International Conference on Biometrics: Theory, Applications and Systems,

Sept 2013, pp. 1–6.

May 14, 2015 DRAFT

IEEE SIGNAL PROCESSING MAGAZINE, VOL. X, NO. X, MONTH 20XX 25

[63] H. Zhang, V. M. Patel, M. E. Fathy, and R. Chellappa, “Touch gesture-based active user authentication using dictionaries,”

in IEEE Winter Conference on Applications of Computer Vision, 2015.

[64] M. Derawi, C. Nickel, P. Bours, and C. Busch, “Unobtrusive user-authentication on mobile phones using biometric gait

recognition,” in International Conference on Intelligent Information Hiding and Multimedia Signal Processing, Oct 2010,

pp. 306–311.

[65] A. Primo, V. V. Phoha, R. Kumar, and A. Serwadda, “Context-aware active authentication using smartphone accelerometer

measurements,” in IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2014, pp. 98–105.

[66] S. Rane, “Standardization of biometric template protection,” IEEE MultiMedia, vol. 21, no. 4, pp. 94–99, Oct 2014.

May 14, 2015 DRAFT