BioCanCrypto: An LDPC Coded Bio-Cryptosystem on Fingerprint Cancellable

Template
2021 IEEE International Joint Conference on Biometrics (IJCB) | 978-1-6654-3780-6/21/$31.00 ©2021 IEEE | DOI: 10.1109/IJCB52358.2021.9484391

Xingbo Dong and Zhe Jin Leshan Zhao

School of Information Technology Chinese University of HongKong, Shenzhen
Monash University, Malaysia [email protected]
{xingbo.dong, jin.zhe}@monash.edu

Zhenhua Guo
Shenzhen International Graduate School
Tsinghua University, Shenzhen
[email protected]

Abstract name, ID, and badge) and the associated proof method (e.g.,
MD5 based password) should be present to the system.
Biometrics as a means of personal authentication has However, a physical card is easy to be lost, stolen, or forged.
demonstrated strong viability in the past decade. How- The password is easy to be lapsed in memory [1], and may
ever, directly deriving a unique cryptographic key from bio- suffer from illegal access risks in case of highly possible il-
metric data is a non-trivial task due to the fact that bio- legal key sharing. Hence, those conventional methods are
metric data is usually noisy and presents large intra-class far from meeting the needs of reliability and safety.
variations. Moreover, biometric data is permanently as- One alternative way to alleviate the issues above is to
sociated with the user, which leads to security and pri- apply biometrics, i.e., identification based on who you are.
vacy issues. Cancellable biometrics and bio-cryptosystem The human biological traits used for recognition systems
are two main branches to address those issues, yet both are generally unique, robust, collectible, natural, and re-
approaches fall short in terms of accuracy performance, liable. The system can be accessed in a non-repudiation
security, and privacy. In this paper, we propose a Bio- manner by utilizing biometric data without presenting the
Crypto system on fingerprint Cancellable template (Bio- ID and password. Unfortunately, directly applying biomet-
CanCrypto), which bridges cancellable biometrics and bio- rics as a key is a challenging task due to: 1) the noisy and
cryptosystem to achieve a middle-ground for alleviating the fuzzy characteristics of biometrics. For example, the iris
limitations of both. Specifically, a cancellable transforma- code, which is proposed to represent human iris features
tion is applied on a fixed-length fingerprint feature vector [2], was found that most iris codes generated from human
to generate cancellable templates. Next, an LDPC cod- iris contain 10%-30% error [3]. The fingerprint is proved
ing mechanism is introduced into a reusable fuzzy extrac- to be user-unique according to Pankanti et al. [4], but also
tor scheme and used to extract the stable cryptographic key suffers from intra-class variations due to the sensor noise
from the generated cancellable templates. The proposed and unstable acquisition process (e.g., placement of the fin-
system can achieve both cancellability and reusability in ger on the sensor) [5]; 2) biometric data is strongly linked
one scheme. Experiments are conducted on a public fin- with the user, the compromise of biometric data will lead to
gerprint dataset, i.e., FVC2002. The results demonstrate security and privacy concerns.
that the proposed LDPC coded reusable fuzzy extractor is
effective and promising. Hence, biometric template protection (BTP) schemes are
proposed to tackle the above issues [6–9]. Among exist-
ing BTP schemes, biometric cryptosystem (BC) [10, 11] is
1. Introduction a prominent BTP technique devised for key management,
which can be divided into two categories, namely key bind-
Personal authentication is mandatory by a wide variety ing and key generation. For the former, the biometric data
of applications such as bank counter and smartphone un- is bound to a key (secret) to generate the Helper Data (HD)
locking. To validate the request, a user identifier (e.g., user- during enrollment. The secret generated during enrollment

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 04,2025 at 11:51:08 UTC from IEEE Xplore. Restrictions apply.
 can be deemed as an identifier of the user. The HD should ther strengthen the performance of crFE, a key redundancy
neither leak information about biometric nor secrets, and mechanism and a most consistent hash code extraction are
it is stored as public information. When a genuine query introduced into the crFE.
biometric is present, the secret would be retrieved but in- The main contribution of this work can be summarized
feasible for an adversary. For the key generation scheme, as:
the HD is generated directly from biometric data. A ran-
dom key or identifier would be generated on the fly based • A fingerprint bio-cryptosystem on the cancellable tem-
on the query biometric and the HD at the query time. Those plate is proposed in this paper to generate a stable cryp-
BC schemes combine biometrics with a cryptographic key. tographic key. Since the proposed bio-cryptosystem is
Hence, it could be the best option to replace conventional working on the cancellable template generated by can-
passwords and gain advantages of both biometrics and cryp- cellable hashing, the system can achieve both cancella-
tosystems. bility and reusability.
Fuzzy Extractor [12] is a typical case of key generation • A variant of reusable fuzzy extractor, namely coded
category. However, the original fuzzy extractor can only reusable fuzzy extractor (crFE), is devised. Specifi-
generate one key from one biometric input, limiting the de- cally, with the novel usage of LDPC, the most consis-
ployment of the fuzzy extractor. For example, users may tent hash codes extraction and a key redundancy mech-
want to create multiple keys from their same fingerprint and anism, the performance of crFE outperforms the state-
use those unique keys in different applications. To solve this of-the-arts bio-cryptosystems.
issue, Boyen et al. proposed a reusable fuzzy extractor con-
cept in 2004 [13]. A fuzzy extractor is reusable if it remains • The performance of the proposed scheme is evaluated
secure even when a user enrolls the same or correlated bio- on the public fingerprint dataset FVC2002. The results
metric data multiple times. Later in 2016, Canetti et al. suggest that the proposed framework is effective and
proposed a reusable fuzzy extractor (rFE) based on “digi- applicable.
tal locker” and it made no assumptions about how multiple
biometric data are correlated [14]. Readers are directed to 2. Related Work
[13, 14] for the details of the reusable fuzzy extractor. 2.1. Fuzzy extractor
However, the reusable fuzzy extractor still has some lim-
Generally, a biometric cryptosystem encrypts the orig-
itations. Firstly, like most existing fuzzy extractors, binary
inal templates to helper data, and error-correcting coding
or integer format biometric features are required. In case
methods are used to handle the intra-class variance. The
of non-binary format features, such fuzzy extractor will be
input of most biometric cryptosystems requires input in fi-
not suitable anymore. On the other hand, the cancellable
nite fields. A biometric cryptosystem either binds a key
property is usually missing in the reusable fuzzy extractor
to a biometric or generates a key from the biometric and
scheme. What is worth highlighting is the difference be-
stores the helper data in the database instead of the bio-
tween reusability and cancellability. The former focuses on
metric data. Retrieving the key or the original biometric
generating a unique security key from the same biometric
data should be computationally hard based on the helper
input, while the latter focuses on generating a revocable, un-
data, which means there is no or limited information leak-
linkable, and non-invertible template based on cancellable
age about the key and biometrics from the helper data. In a
transformation.
biometric cryptosystem, the key is released or reconstructed
We proposed a biometric cryptosystem to generate a only if the genuine’s biometric data is presented for veri-
cryptographic key from a cancellable fingerprint template fication. Several popular biometric cryptosystem schemes
to address the above issues. The system consists of are proposed in the literature, such as fuzzy vault[18] and
two stages, namely the fixed-length fingerprint cancellable fuzzy commitment[19] under the key binding group, secure
template generation stage and the key extraction stage. sketch and fuzzy extractor [12] under the key generation
Specifically, to generate fixed-length fingerprint features, group.
Minutiae-Cylinder-Code (MCC) fingerprint descriptor [15] Fuzzy Extractor is a typical scheme for key generation.
is firstly generated from fingerprint minutia points, then Dodis et al. proposed the concept of fuzzy extractor first in
transformed into a fixed-length vector according to the [12]. Fuzzy extractor consists of a pair of randomized pro-
method proposed by Jin et al. [16]; Subsequently, the fixed- cedures, “generate” (Gen) and “reproduce” (Rep). Specifi-
length fingerprint feature is hashed into integer vectors by cally,
a cancellable ranking based hashing function [17]. To gen-
erate the cryptography key, a reusable fuzzy extractor with • Gen(w) −→ (P, R): Take a biometric input w ∈ W
Low-density Parity-check (LDPC) encoding, denoted as a (where W can be regarded as the distribution on cer-
coded reusable fuzzy extractor (crFE), is proposed. To fur- tain biometric modality). Then, extract a random string

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 04,2025 at 11:51:08 UTC from IEEE Xplore. Restrictions apply.
 r ∈ {0, 1}` and a public helper string P . Based on 2.2. Bio-cryptosystem Based on Cancellable Tem-
[20], P is termed as Aux Data (AD), and Key is termed plates
as pseudonymous identifier (PI).
Few works have been done on the fingerprint to combine
• Rep(w0 ∈ W, P ) −→ R0 : Take a biometric query bio-cryptosystem with cancellable templates. In this sub-
w0 ∈ W 0 (where W 0 can be regarded as another cap- section, the latest related works, which are not limited to
ture of previous biometric modality). If the distance fingerprints, are reviewed.
between w and w0 is within a tolerance distance t, then In [22], a finger-vein cancellable bio-cryptosystem was
Rep(w0 , P ) = r with probability 1 − γ, where γ > 0 proposed for healthcare application authentication and data
is the error. encryption. Specifically, the healthcare data is encrypted by
a secret key; then, a cancelable biometrics scheme is applied
The correctness of fuzzy extractor requires that r can be on the finger-vein features to generate the cancellable tem-
reproduced when w ∈ W and w0 ∈ W are close enough in plate. Subsequently, a fuzzy commitment scheme is used
distance. The fuzzy extractor’s security requires that R is to bind with the secret key, and the final generated help
random when W has enough entropy. data is stored on the smart card. To decrypt the data, query
However, original fuzzy extractor can only generate one finger-vein features are transformed by the same cancelable
key from one biometric input, which limits the deploy- biometrics scheme, then the secret is released based on the
ment of the fuzzy extractor. For example, users may want cancelable template and stored help data.
to generate multiple keys from his same fingerprint, and In [23], a cancelable fuzzy vault system was proposed on
use those distinct keys in different applications. To solve Delaunay triangle group-based fingerprint features. Specif-
this issue, Boyen et al. proposed a reusable fuzzy extrac- ically, a cancellable polar transformation is applied to each
tor in 2004 [13]. A fuzzy extractor is reusable if it re- Delaunay triangle group. The non-invertible polar transfor-
mains secure even when a user enrolls the same or cor- mation unit is a triangle instead of a single minutia. Hence it
related values multiple times. Specifically, assuming that can avoid non-linear distortion. In [24], another cancellable
w, w1 , . . . , wρ are fingerprint features from the same fin- fuzzy vault was proposed. A non-invertible template is gen-
ger, (P, R) = {Pi , Ri }i∈{1,2,...,ρ} can be generated from erated by applying a periodic transformation on the biomet-
Gen (wi ) → (Pi , Ri ), and given the public helper data and ric template. The generated template is then used as the
key, i.e., {Pi , Ri }i∈[ρ] , [ρ] := {1, 2, . . . , ρ} and P , R still fuzzy vault input, strengthening the fuzzy vault with the
meet the security requirement. cancellability property. In [25], to adopt fuzzy commitment
In 2016, Canetti et al. constructed a reusable fuzzy ex- on fingerprint, a binary length-fixed feature generation was
tractor (rFE) in [14] based on “digital locker” [21], and it proposed. Minutiae triplets are used in the scheme. Hence
made no assumptions about how multiple biometric data the alignment of fingerprint minutiae is avoided. Besides,
are correlated. Besides, it achieved computational security various error correction codes were explored in this scheme.
under the existence of digital lockers. Specifically, at enroll In [26], a biometric cryptosystem based on cancelable
stage in Canetti’s reusable fuzzy extractor scheme, three pa- templates generated by discrete Fourier transform (DFT)
rameters, i.e., the key k, the nonce c and random mask w, and random projection was proposed. In the proposed sys-
are generated randomly. The biometric input x is firstly fil- tem, polar grid-based fingerprint features are transformed
tered by the mask w based on AND operation denoted as using the DFT and random projection, creating a non-
v = x AN D w. Then vector v is hashed by SHA256 with invertible template. A bit-toggling strategy is also uti-
salting parameter t, hence
the L L a cipher can be generated lized to further inject noise into the generated template to
by the SHA256(v, t), is the exclusive or (XOR). Fi- strengthen template security.
nally, the k is output as a cryptography key, while the tuple In [27], a bio-cryptosystem on cancelable fingerprint
(cipher, w, c) is stored in the system as the help data. At templates was proposed. Firstly, cancellable templates are
query stage, the same mask is applied on the biometric in- generated from fingerprints, then a combined template is
put x0 , i.e., v 0 = x0 AN D w. SHA256 is used to hash the v 0 created, and finally, the cryptographic key is generated from
with the same parameter t, and then XORed with the stored the combined template based on a thresholding mechanism.
cipher, to reproduce the final key k. It is worth highlighting In [28], a cancelable multi-biometric authentication based
that the original rFE from [14] locks up the generated ran- on a bit-wise encryption transform and fuzzy extractor was
dom k using parts of w in an error-tolerant way, instead of proposed on iris and face. In this scheme, the first biometric
applying a secure sketch on w. However, a reusable fuzzy template is used to generate a random string as a secret key
extractor normally requires the binary input, which limits s by a fuzzy extractor. Subsequently, the second biometric
the utilization of such schemes. On the other hand, a per- template is transformed into a protected template by a can-
formance drop is usually expected when applying the fuzzy cellable bit-wise encryption transform which uses the s as
extractor in case of poor discriminative features. the encryption key. The transformed template is stored in

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 04,2025 at 11:51:08 UTC from IEEE Xplore. Restrictions apply.
 User Device the projection matrix P ∈ RNt ×d is generated as the
MCC Fixed-length LDPC Fuzzy
Representation Features Extractor Key Sign(Key) eigen-vectors of K, where d is the output dimensions.
Given one MCC descriptor query instance Ωq , match-
Enroll Template Stored in System
Help Data Matching ing scores vi = sim (Ωq , Ωt (i)) is computed for all
Query training samples Ωt (i) where 1 ≤ i ≤ Nt . Then
MCC
Representation
Fixed-length
Features
LDPC Fuzzy
Extractor rKey Sign(rKey) v is transformed with the same kernel function into
v = f (v). The fixed-length fingerprint vector x can
Figure 1. Block diagram of the proposed framework. be generated by x = vP ∈ Rd , where v ∈ RNt . In
this paper, we fixed the fingerprint vector x with length
299, i. e., d = 299, as the default parameter settings
the database for matching. from [16].
Generally, those methods utilized cancellable transfor-
mation to generate a binary template, and then a bio- 3. Cancellable hashing. To protect the generated fixed-
cryptosystem is established on those templates. It is noted length fingerprint vector x, and generate binary tem-
that only a few related works have been done on finger- plate, a cancellable transformation function, i.e.,
prints. Besides, the performance drop is usually observed; index-of-max hashing, proposed in [17] is adopted.
for instance, the FRR can be 6.43% with FAR=2% in [26], Specifically, the index-of-max hashing of one hash
there is still a lot of room to improve. code is defined as: h(x; W ) = arg max wiT x ∈ R,
1≤i≤q
where wi ∈ Rd ∼ N (0, Id ) , i = 1, 2...q, and W =
3. Proposed method [w1 , w2 , . . . , wq ]T ∈ Rd×q . Thus, the cancellable tem-
It is inevitable to experience a performance drop when plate is generated as h = {hi |i = i, . . . , m}, where
applying a BTP algorithm in a biometric system. The situ- hi ∈ {1, 2, · · · , q} by repeating above steps m times.
ation becomes worse when both cancellable transform and The template is in integer format, which fits the subse-
bio-cryptosystem are integrated into one system. We pro- quent fuzzy extractor input requirement.
pose an LDPC coded bio-cryptosystem on cancellable fin- To be noted that the cancellable transformation and fuzzy
gerprint template, which bridges cancellable biometrics and extractor mandates a fixed-length feature vector that re-
bio-cryptosystem together to achieve a middle-ground with quires high discriminative power to preserve the perfor-
benefits from both sides. The LDPC coding scheme is also mance. Hence m = 512, q = 16 is applied as the default
utilized to preserve the performance of the fuzzy extractor. parameters in the cancellable hashing.
An overview of the system is shown in Figure 1. The pro-
posed scheme is composed of a fixed-length fingerprint bi- 3.2. LDPC coded reusable fuzzy extractor
nary template generation stage and key extractor stage. We After the fixed-length cancellable templates are ob-
first briefly describe fixed-length fingerprint template gen- tained, the next task is to construct a biometric cryptosystem
eration and then the LDPC coded reusable fuzzy extractor. to generate a stable cryptography key. A Low-Density Par-
3.1. Fixed-length fingerprint vector generation ity Check (LDPC) coded reusable fuzzy extractor, crFE, is
proposed based on existing reusable fuzzy extractor (rFE)
The method from [17] is adopted to generate the fixed- from [14], which is designed based on “digital locker” and
length fingerprint cancellable templates. This is because it makes no assumptions about how multiple biometric data
the method in [17] is one of the latest state-of-the-art can- are correlated.
cellable transformations, which can lead to compact integer We noticed that the fuzzy extractor would lead to a sig-
hash codes with high accuracy performance. It consists of nificant performance drop. Generally, error correction code
the minutiae descriptor MCC extraction stage [15], fixed- (ECC) is applied in fuzzy commitment scheme to handle
length feature generation stage based on kernel transforma- the intra-class variations; hence, introducing error correc-
tion [16], and the cancellable transformation stage based on tion code (ECC) shall improve the performance. Among
index-of-max hashing[17]. Specifically: different ECCs, LDPC is adopted as it brings several ad-
1. MCC extraction stage. Minutia Cylinder Code de- vantages as suggested from [29]: 1) index-of-max hashing
scriptors Ω are generated from extracted minutiae reduces the amount of stored data, which can benefit the
points. utilization of LDPC; 2) LDPC codes have greater error cor-
rection capabilities than classical algebraic codes, and the
2. Fixed-length feature generation. Given a set of MCC relative error-correcting capability is almost constant as the
training samples Ωt = {Ω(i) | i = 1, . . . , Nt } where code length increases for a fixed rate; 3) LDPC codes allow
Nt is the total number of Ωt , a kernel matrix K ∈ to reduce the size of the code representation, by exploiting
RNt ×Nt is computed with a kernel function f (·). Then the sparse nature of parity check matrix H in LDPC.

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 04,2025 at 11:51:08 UTC from IEEE Xplore. Restrictions apply.
 Algorithm 1 Key Generation utilized in our scheme.
Input: Firstly, the most consistent hash codes are extracted
Biometric template x, from the generated cancellable template to minimize the
LDPC encoder Φ. hash codes intra-class variation. Specifically, J templates
Output: per user are generated, denotes as {hji |j = 1, · · · , J, i =
key k, help data HD = {cipher, w, t} 1, · · · , m}, the consistent hash codes index is identified
1: k = rand(), t = rand(), w = rand() if the hash code value of the corresponding index have
2: ek = Φ(k) aPsubstantial probability to be the same, i.e., H(i) =
3: v = x AND w −p(hji ) log(p(hji )) for a given i, top-k index will be col-
4: digest = sha256(v, t) lected as the mask to extract the most consistent hash codes
ĥ ∈ {1, 2, · · · , q}k for each user, where k ≤ m.
L
5: cipher = digest ek
Next, the generated hash codes ĥ is divided into c-blocks
k
where c = K , K is the input length of the crFE. To fur-
Algorithm 2 Key Reproduce
ther improve the performance, r different non-overlapping
Input: blocks will be utilized to extract the identical key, hence
Biometric query x0, obtain the key redundancy. We denote r as the redundancy
HD = {cipher, w, t}, factor. Finally, rc keys are extracted in total and concate-
LDPC decoder Ψ. nated together to form the final key.
Output:
Reproduced k0 3.4. Summary of the proposed framework
1: v0 = x0 AND w
To wrap up this section, we highlight the main character-
2: digest0 = sha256(v0, t)
L istics of the proposed scheme as below:
3: plains = digest0 cipher
4: k0 = Ψ(plains) 1. In our proposed scheme, the extracted key can be can-
celed by using the newly generated random projection
matrix W in the index-of-max hashing, in case of the
An LDPC code is specified by a parity-check matrix H compromising of the key. The proposed scheme also
containing mostly 0s and a low density of 1s. The rows of inherits the unlinkability, invertibility from the index-
the H represent the equations while columns represent the of-max hashing. The reader can refer to [17] for the
bits in the codeword, i.e., code symbols. LDPC is accompa- detailed description of the cancellable characteristics.
nied with three parameters (n, dc, dr), where n is the code-
word length, dc is the the number of 1’s in each column, and 2. In our proposed scheme, different and uncorrelated
dr is the number of 1’s in each row, where dr > dc > 3. keys can be generated by utilizing the reusable prop-
In our crFE scheme, LDPC coding is introduced to the erty of rFE. Meanwhile, this can also be done by gener-
original reusable fuzzy extractor to achieve better perfor- ating a new random projection matrix W . This brings
mance. The main difference between LDPC coded fuzzy more flexibility to the system.
extractor and the original reusable fuzzy extractor is the 3. Since our proposed bio-cryptosystem is established on
LDPC encoding/decoding of the key k. At enroll stage, the cancellable template; it provides extra layer protec-
the randomly generated key is encoded by LDPC, and tion to the biometric template, i.e., the original biomet-
then AND with the same random mask w, followed by ric data is protected by cancellable transformation and
the sha256 hashing to generate the digital locker. Subse- bio-cryptosystem. As suggested by [30–32], ranking-
quently, an XOR operation is performed with the encoded based hashing irreversibility is overestimated, which
key to generate the final cipher, and the help data consists may incur risks. While in our scheme, the cancellable
of cipher, random mask w and salt t is stored in the sys- template is not stored in the system but used to gen-
tem, denoted as HD = {cipher, w, t}. The LDPC coded erate the key, which can avoid such a threat. Besides,
reusable fuzzy extractor is shown in Algorithm 1 and Algo- the compromising of key and help data will not lead to
rithm 2. direct disclosure of raw biometric data.
3.3. Key Generation from cancellable template by
4. Experiments and Results
crFE
4.1. Datasets and protocol
To make full use of the rich information from the gen-
erated cancellable template h ∈ {1, 2, · · · , q}m , and the In this experiment, public fingerprint datasets, i.e.,
error correction capability of LDPC, several strategies are FVC2002 (DB1, DB2, and DB3) [33] are used. Because

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 04,2025 at 11:51:08 UTC from IEEE Xplore. Restrictions apply.
 most well-known bio-cryptosystems adopted this dataset, Table 1. Verification performance on FVC2002 (%).
EER FRR/zeroFAR
hence a consistent performance comparison can be made Methods
DB1 DB2 DB3 DB1 DB2 DB3
in Section 4.5. For FVC2002, each sub-dataset consists of Fixed-length feat. 0.12 0.59 2.71 0.8 1.0 9.5
100 users and 8 samples per user. The minutiae template Cancellable Hashing 0.94 1.45 5.41 1.6 7.3 19.3
is extracted according to the ISO-compliant format by Ver-
iFinger SDK1 . Official MCC SDK2 is used to generate the Table 2. Key retrieval performance of rFE on FVC2002 datasets
MCC descriptor. with different r (k = 512, c = 32,(%)).
key DB1 DB2 DB3
To generate fixed-length fingerprint features, as de- r # keys
length (bits) FRR FAR FRR FAR FRR FAR
scribed in [16], 1st to 3rd samples of each identity are used 1 32 4096 98 0 93 0 99 0
as training samples to generate the projection matrix P, the 2 16 2048 50 0 50 0 75 0
rest of the samples (i.e., 4th –8th , 100*5 samples in total) 4 8 1024 12 0 16 0 33 0
8 4 512 0 0 5 0 12 0
of each identity are used to generate the fixed-length finger-
print vectors, as well as the subsequent hash codes. Within
this subset of data, the fixed-length fingerprint vector’s ac-
(128 bits) to achieve the best processing efficiency. Corre-
curacy performance and the following hash codes are mea-
sponding to the crFE 128bits input, the LDPC parameters
sured by Equal Error Rate (EER) based on the FVC2002
(n, dc, dr) is fixed as n = 128, dc = 3, dr = 16, under this
testing protocol [17, 33].
setting, the corresponding key length is 106 bits instead of
To extract consistent hash codes from the cancellable
128 bits in rFE. As for the index-of-max hashing, the de-
templates, 4th –6th samples (J = 3) are utilized to extract
fault parameter, i.e., m = 512, q = 16, suggested by [17] is
the consistent hash code location masks for each user as
adopted.
aforementioned. To simulate the key generation and key
In the evaluation experiment, k = {64, 128, 256, 512} is
reproduction processes in the crFE, the 7th sample is used
adopted as the consistent hash codes length (top-k), hence
as the enroll sample, and a key is generated for each user
number of blocks is determined correspondingly as c =
according to algorithm 1. The 8th sample is used as a k
query sample to reproduce the key by algorithm 2. The 16 = {4, 8, 16, 32}. Subsequently, r = {1, 2, 4} is adopted
as the the redundancy factor settings.
performance of the proposed overall biometric cryptosys-
To explore the effect of different r value on the key re-
tem is evaluated by False Reject Rate (FRR) with respect to
trieval performance, we fix other parameters as k = 512 and
False Accept Rate (FAR). FAR and FRR are determined by
c = 512
16 = 32, and the specific FRR and FAR of rFE under
checksum comparison between k and k0. For a key retrieval
r = {1, 2, 4} is shown in Table 2. As a result, suggests,
system, zero FAR and lower FRR are expected.
larger r will certainly achieve better performance. In detail,
4.2. Fingerprint fixed-length features verification the FRR drops from 98% to 0% when r increases from 1
performance to 8. This is due to the redundancy of the key generated
from the hash blocks, which can better avoid information
In this subsection, we first evaluate the accuracy of distortion.
the fixed-length representation and then the cancellable Besides the redundancy factor r, the consistent hash code
hash codes. The specific accuracy in terms of EER and length k is also explored on rFE and crFE. The key retrieval
FRR/zeroFAR is shown in Table 1. Based on the results, performance with different k is shown in Table 3.
the EER of the fixed-length features can achieve 0.12% on
When r = 4, as observed from the Table 3, k = 128, c =
FVC2002 DB1. However, after the cancellable transforma-
8 can achieve a overall good performance on both rFE and
tion, the performance drops to 0.94%. This is unavoidable
crFE. A larger k will certainly lead to a smaller FAR, but
due to the random hashing projection. The FRR under zero
will increase FRR as longer hash codes will contain more
FAR also shows similar performance degradation. The in-
noise information. When r = 8, Table 3 suggests that k =
consistency of FARs over different datasets is not supervis-
512, c = 32 achieves the best accuracy. It is also consistent
ing. This is due to multiple factors impacted such as the
with the result from r = 4 that larger k leads to smaller
natural characteristics of dataset e.g. distributions, the al-
FAR but higher FRR.
gorithm of cancellable biometrics, and fixed-length feature
Regarding the performance obtained by crFE, the result
vectors and so on so forth.
from Table 3 suggests that crFE can achieve better accuracy
4.3. Key retrieval performance of crFE compared with rFE. In detail, crFE can achieve FRR=2%
with zero FAR on DB1, FRR=4% with 1% FAR on DB2,
In this subsection, the accuracy of crFE is evaluated. As and FRR=15% with 0% FAR on DB3. The counterpart rFE
for the crFE, the input length is fixed as K = 16 bytes achieves FRR=2% with 0% FAR, FRR=7% with 1% FAR
1 https://www.neurotechnology.com/ on DB2, and FRR=14% with zero% FAR on DB3 respec-
2 http://biolab.csr.unibo.it/research.asp?selObj=82 tively. However, the performance gain is not observed from

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 04,2025 at 11:51:08 UTC from IEEE Xplore. Restrictions apply.
 Table 3. Key retrieval performance of rFE/crFE on FVC2002 datasets with different k and r (%).
rFE crFE
k c # keys r key length DB1 DB2 DB3 key length DB1 DB2 DB3
(bits) FRR FAR FRR FAR FRR FAR (bits) FRR FAR FRR FAR FRR FAR
64 4 1 128 0 8 4 9 8 10 106 1 10 4 11 8 8
128 8 2 256 2 0 7 1 14 0 212 2 0 4 1 15 0
4
256 16 4 512 2 0 8 0 22 0 424 5 0 10 0 23 0
512 32 8 1024 12 0 16 0 33 0 848 11 0 17 0 33 0
128 8 1 128 0 12 2 11 5 14 106 0 15 1 12 8 19
256 16 2 8 256 0 2 3 2 4 2 212 0 4 2 3 6 2
512 32 4 512 0 0 5 0 12 0 424 2 0 5 0 14 0

Table 4. Time cost per query in seconds under different settings Table 5. Comparison with other biometric cryptosystems.
(r = 8). Database type Key length Accuracy Performance
Only Bio-cryptosystem
k c # keys Time Cost per Query Li et al. [34] FVC02 DB2 8 coefficients FRR@zeroFAR= 7
128 8 1 1.0520s Li et al. [25] FVC02 DB2 50 bits FRR@zeroFAR=4.85
256 16 2 2.0004s FVC02 DB1 DB1:FRR@zeroFAR=8.6
Liu et al. [35] 3 coefficients
FVC02 DB2 DB2:FRR@zeroFAR=16
512 32 4 4.0396s Bio-cryptosystem and cancellable biometrics Combined
FVC02 DB1 DB1:FRR@zeroFAR=34
Yang et al. [23] 7 coefficients
FVC02 DB2 DB2:FRR=23,FAR=0.02
Best accuracy ( 16bits):
the result in Table 3, the possible reason is that the larger FVC02 DB1
Alam et al. [26] 16 - 128 bits DB1:FRR=6.43,FAR=2.06
FVC02 DB2
redundancy will cause some noise to lead to poor perfor- DB2:FRR=7.71,FAR=3.15
Proposed BioCanCrypto scheme
mance. rFE FVC02 DB1 DB1:FRR@zeroFAR=2
256 bits
(k = 128, r = 4) FVC02 DB2 DB2:FRR=7,FAR=1
4.4. Time cost evaluation crFE FVC02 DB1
212 bits
DB1:FRR@zeroFAR=2
(k = 128, r = 4) FVC02 DB2 DB2:FRR=4,FAR=1
To evaluate the time efficiency of the crFE driven HD rFE FVC02 DB1
512 bits
DB1:FRR@zeroFAR=0
wallet system, the crFE driven HD wallet system is imple- (k = 512, r = 8) FVC02 DB2 DB2:FRR@zeroFAR=5
crFE FVC02 DB1 DB1:FRR@zeroFAR=2
mented with Python 3.0, and the time cost under different 424 bits
(k = 512, r = 8) FVC02 DB2 DB2:FRR@zeroFAR=5
settings is evaluated based on a cloud Ubuntu server with
Intel Xeon CPU (3.2GHz, 128 core), 256GB RAM. The ex-
periment is carried out on FVC2002 DB1. The time cost is both can lead to a higher level protection mechanism and
tabulated in Table 4. extra protection layer. In this paper, a bio-cryptosystem
The data from Table 4 suggests that more time is needed is established on the cancellable template. Specifically,
for crFE as the consistent hash code length k increases. to generate a cryptographic key from fingerprints by the
crFE takes a longer time than expected to complete the key fuzzy extractor, the state-of-the-art fingerprint descriptor,
retrieval; this is because LDPC needs extra time to decode i.e., MCC, is firstly generated from the fingerprint minu-
the key, and hence prolongs the process. However, the pro- tiae points and then transformed into fixed-length by a ker-
cessing time is still acceptable under k = 128, r = 8, as nelized function, followed by a cancellable hashing. The
it can complete the process within around 1s. Besides, the final generated template is in integer format, which can be
processing time can still be optimized by tuning the LDPC directly used in the fuzzy extractor.
parameter or using specifically designed hardware.
To improve the fuzzy extractor’s accuracy performance,
4.5. Comparison with the state-of-the-art several measures are taken: 1) encoding the key by LDPC;
2) consistent hash code extraction; 3) key redundancy
Finally, a comparison between the proposed biometric mechanism. LDPC is a linear error-correcting block code
cryptosystem and the state-of-the-art is presented in Table suitable in our settings to correct errors in very noisy bio-
5. We can see from the comparison that the best accuracy metric data. Note that using of LDPC is optional and de-
on DB1 and DB2 is achieved by rFE(k = 512, r = 8) and pending on practical settings. The most consistent hash
crFE(k = 128, r = 4), respectively. The results suggest that codes are extracted to reduce intra-variations, which can
our proposed biometric cryptosystem outperforms others in further improve the performance. At last, the hash code vec-
terms of ZERO FAR and key length. tor is divided into sub-blocks, and a key redundancy mech-
anism is utilized to further boost the performance. Based on
5. Conclusion the experiment result, the time cost can be done in seconds.
Cancellable biometrics and bio-cryptosystem are two Meanwhile, better accuracy is achieved.
main branches to protect biometric data. A combination of The proposed scheme can be applied in the case of key

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 04,2025 at 11:51:08 UTC from IEEE Xplore. Restrictions apply.
 management scenarios such as the blockchain wallet pri- ing fixed-length representation from minutiae using kernel methods
vate key generation, which can provide effectiveness to the for fingerprint authentication. IEEE Transactions on Systems, Man,
and Cybernetics: Systems, 46(10):1415–1428, 2016.
user. The future direction is to adopt the proposed bio- [17] Z. Jin, J. Y. Hwang, Y. Lai, S. Kim, and A. B. J. Teoh. Ranking-Based
cryptosystem on other biometric modalities and optimize Locality Sensitive Hashing-Enabled Cancelable Biometrics: Index-
the execution efficiency, accuracy and security further. of-Max Hashing. IEEE Transactions on Information Forensics and
Security, 13(2):393–407, Feb. 2018.
[18] A. Juels and M. Sudan. A fuzzy vault scheme. Designs, Codes and
Acknowledgment Cryptography, 38(2):237–257, 2006.
[19] A. Juels and M. Wattenberg. A fuzzy commitment scheme. In Pro-
The work was partially supported by the Natural Science Founda- ceedings of the 6th ACM Conference on Computer and Communica-
tion of China (NSFC) (No. 61772296), Shenzhen Fundamental Research tions Security, CCS ’99, pages 28–36, New York, NY, USA, 1999.
Fund (No. JCYJ20170412170438636), and the Graduate Research Stu- ACM.
dent Global Research Experience (GRE) Travel Grant provided by School [20] ISO/IEC 24745:2011 - Information technology – Security techniques
of Information Technology, Monash University Malaysia. – Biometric information protection.
[21] R. Canetti and R. R. Dakdouk. Obfuscating point functions with
multibit output. In Annual International Conference on the The-
References ory and Applications of Cryptographic Techniques, pages 489–508.
Springer, 2008.
[1] J. Yan, A. Blackwell, R. Anderson, and A. Grant. Password mem- [22] W. Yang, S. Wang, J. Hu, G. Zheng, J. Chaudhry, E. Adi, and
orability and security: Empirical results. IEEE Security & privacy, C. Valli. Securing mobile healthcare data: a smart card based can-
2(5):25–31, 2004. celable finger-vein bio-cryptosystem. IEEE Access, 6:36939–36947,
[2] J. Daugman. How iris recognition works. In The essential guide to
2018.
image processing, pages 715–739. Elsevier, 2009. [23] W. Yang, J. Hu, and S. Wang. A delaunay triangle group based fuzzy
[3] S. Simhadri, J. Steel, and B. Fuller. Cryptographic authentication
vault with cancellability. In 2013 6th International Congress on
from the iris. In International Conference on Information Security,
Image and Signal Processing (CISP), volume 3, pages 1676–1681.
pages 465–485. Springer, 2019.
IEEE, 2013.
[4] S. Pankanti, S. Prabhakar, and A. K. Jain. On the individuality of
[24] T. K. Dang, Q. C. Truong, T. T. B. Le, and H. Truong. Cancellable
fingerprints. IEEE Transactions on pattern analysis and machine
fuzzy vault with periodic transformation for biometric template pro-
intelligence, 24(8):1010–1025, 2002.
tection. IET Biometrics, 5(3):229–235, 2016.
[5] U. Uludag, S. Pankanti, and A. K. Jain. Fuzzy vault for fingerprints.
[25] P. Li, X. Yang, H. Qiao, K. Cao, E. Liu, and J. Tian. An effective
In International Conference on Audio-and Video-Based Biometric
biometric cryptosystem combining fingerprints with error correction
Person Authentication, pages 310–319. Springer, 2005.
codes. Expert Systems with Applications, 39(7):6562–6574, 2012.
[6] V. M. Patel, N. K. Ratha, and R. Chellappa. Cancelable Biometrics:
[26] B. Alam, Z. Jin, W.-S. Yap, and B.-M. Goi. An alignment-free cance-
A review. IEEE Signal Processing Magazine, 32(5):54–65, Sept.
lable fingerprint template for bio-cryptosystems. Journal of Network
2015.
and Computer Applications, 115:20–32, 2018.
[7] M. Sandhya and M. V. N. K. Prasad. Biometric Template Protec-
[27] A. Sarkar and B. K. Singh. Cryptographic key generation from can-
tion: A Systematic Literature Review of Approaches and Modali-
celable fingerprint templates. In 2018 4th International Conference
ties. In R. Jiang, S. Al-maadeed, A. Bouridane, P. D. Crookes, and
on Recent Advances in Information Technology (RAIT), pages 1–6.
A. Beghdadi, editors, Biometric Security and Privacy: Opportuni-
IEEE, 2018.
ties & Challenges in The Big Data Era, Signal Processing for Secu- [28] D. Chang, S. Garg, M. Hasan, and S. Mishra. Cancelable multi-
rity Technologies, pages 323–370. Springer International Publishing, biometric approach using fuzzy extractor and novel bit-wise encryp-
Cham, 2017. tion. IEEE Transactions on Information Forensics and Security,
[8] E. Chandra and K. Kanagalakshmi. Cancelable biometric template
15:3152–3167, 2020.
generation and protection schemes: A review. In 2011 3rd Interna- [29] M. Baldi, M. Bianchi, F. Chiaraluce, J. Rosenthal, and D. Schipani.
tional Conference on Electronics Computer Technology, volume 5, On fuzzy syndrome hashing with ldpc coding. In Proceedings of the
pages 15–20, Apr. 2011. 4th International Symposium on Applied Sciences in Biomedical and
[9] K. Nandakumar and A. K. Jain. Biometric Template Protection:
Communication Technologies, pages 1–5, 2011.
Bridging the performance gap between theory and practice. IEEE [30] X. Dong, Z. Jin, and A. T. B. Jin. A genetic algorithm enabled
Signal Processing Magazine, 32(5):88–100, Sept. 2015. similarity-based attack on cancellable biometrics. In 2019 IEEE
[10] C. Rathgeb and A. Uhl. A survey on biometric cryptosystems and
10th International Conference on Biometrics Theory, Applications
cancelable biometrics. EURASIP Journal on Information Security,
and Systems (BTAS), pages 1–8. IEEE, 2019.
2011(1):3, 2011. [31] X. Dong, Z. Jin, A. B. J. Teoh, M. Tistarelli, and K. Wong.
[11] K. Xi and J. Hu. Bio-cryptography. In Handbook of Information and
On the security risk of cancelable biometrics. arXiv preprint
Communication Security, pages 129–157. Springer, 2010.
arXiv:1910.07770, 2019.
[12] Y. Dodis, L. Reyzin, and A. S. F. Extractors. How to generate strong
[32] H. Wang, X. Dong, Z. Jin, A. B. J. Teoh, and M. Tistarelli.
keys from biometrics and other noisy, data april 13. EUROCRYPT,
Interpretable security analysis of cancellable biometrics using
2004.
constrained-optimized similarity-based attack. In Proceedings of the
[13] X. Boyen. Reusable cryptographic fuzzy extractors. In Proceedings
IEEE/CVF Winter Conference on Applications of Computer Vision,
of the 11th ACM conference on Computer and communications se-
pages 70–77.
curity, pages 82–91. ACM, 2004.
[33] D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar. Handbook of
[14] R. Canetti, B. Fuller, O. Paneth, L. Reyzin, and A. Smith. Reusable
fingerprint recognition, 2nd edition. Springer Science & Business
fuzzy extractors for low-entropy distributions. In Annual Interna-
Media, 2009.
tional Conference on the Theory and Applications of Cryptographic
[34] P. Li, X. Yang, K. Cao, X. Tao, R. Wang, and J. Tian. An alignment-
Techniques, pages 117–146. Springer, 2016.
free fingerprint cryptosystem based on fuzzy vault scheme. Journal
[15] R. Cappelli, M. Ferrara, and D. Maltoni. Minutia cylinder-code:
of network and computer applications, 33(3):207–220, 2010.
A new representation and matching technique for fingerprint recog-
[35] E. Liu and Q. Zhao. Encrypted domain matching of fingerprint
nition. IEEE transactions on pattern analysis and machine intelli-
minutia cylinder-code (mcc) with l1 minimization. Neurocomputing,
gence, 32(12):2128–2141, 2010.
259:3–13, 2017.
[16] Z. Jin, M.-H. Lim, A. B. J. Teoh, B.-M. Goi, and Y. H. Tay. Generat-

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 04,2025 at 11:51:08 UTC from IEEE Xplore. Restrictions apply.