Vulnerabilities in Binary Face Template
Yi C Feng
Department of Computer Science
Hong Kong Baptist University
ycfeng@comp.hkbu.edu.hk
Pong C Yuen
Department of Computer Science
Hong Kong Baptist University
pcyuen@comp.hkbu.edu.hk
Abstract
It is generally believed that original face image is hard
to be reconstructed from its binary face template and there-
fore, binary face template is secure. In this paper, we find
that binary template is not secure and can be attacked by a
masquerade attack. A masquerade face image, which may
not ”look” similar with the original image, is constructed
from binary templates and then employed to access the sys-
tem. Experimental results show that the successful attack-
ing rate on CMU-PIE and FRGC databases are 100% and
77.14% in around 8 and 33 seconds in a personal computer,
respectively.
1. Introduction
In developing a practical biometric recognition system,
the security and privacy of the biometric templates stored in
databases or smartcards are one of the major concerns. Po-
tential attacks to biometric templates have been presented
and summarized in [3]. Also, if a biometric template is
stored in the database unprotected, it is possible to recon-
struct the original biometric data from the stored biometric
template [7, 9, 4, 5]. Therefore, biometric template secu-
rity [2, 3] is one of the most crucial issues in deploying a
practical biometric recognition system.
To protect biometric templates, original biometric tem-
plate is not stored, but a transformed/encrypted form of
the original biometric template. Three major approaches,
namely biometric cryptosystem [3], transform-based [2, 3]
and hybrid, have been proposed. In the biometric cryptosys-
tem approach, the original reference template is encrypted
with error-correcting codes, and then stored in database. In
the transform-based approach, the original biometric tem-
plate is transformed into a new domain via a one-way func-
tion for protection. Parameters of the transformation func-
tion are defined by a user-specific key or password to gain
cancelability capability. The hybrid approach [16] com-
bines the advantages of both approaches.
Binary template [10, 11, 12, 13, 14, 15] is involved in all
three approaches. For the biometric cryptosystem approach,
the encryption process requires finite input. If the original
template lies in infinite space (e.g. a face template repre-
sented as a real-valued vector), a binarization process can
be employed to transform it into a binary template (which
lies in a finite field), and then input to the encryption pro-
cess. For the transform-based approach, the binarization
process can be treated as a one-way transform. Other than
encrypting the binary templates in biometric cryptosystems,
the transform-based approach directly stores the binary tem-
plate without encryption. It benefits from two advantages:
1) the recognition performance can be preserved as encryp-
tion may degrade the performance, 2) Biometric cryptosys-
tems can only output decisions but no matching scores.
Thus a biometric cryptosystem can only be used in verifi-
cation, while the transform-based approach can be used in
identification. Algorithms [10, 11, 12, 13, 14, 15] trans-
forming raw face image into its binary face template, have
been proposed. It is claimed that the raw face image cannot
be reconstructed from its binary template.
There are two main smart attacking approaches to bio-
metric systems, namely hill-climbing attack and masquer-
ade attack. Hill-climbing attack [7, 8, 9] modifies the
raw input biometric data iteratively based on the output
matching score from the system for accessing the system.
However, since the binarization process always involves a
thresholding/quantization procedure, which strongly elimi-
nates the contributions to matching scores of small modifi-
cations in each iteration, hill-climbing attack becomes less
effective. Masquerade attack [4, 5, 6] reconstructs a raw
biometric from the template stored in database, which is
then used for accessing the system. It has been shown that a
reconstructed face image can be generated from a biometric
template and can successfully access the system.
Since the hill-climbing attack is not efficient for binary
template, we follow the masquerade attack. However, ex-
isting masquerade attack algorithms are case-specific and
cannot apply to binary face template. As such, we de-
velop a new masquerade attack algorithm. The proposed
algorithm follows the typical assumptions in masquerade
978-1-4673-1612-5/12/$31.00 ©2012 IEEE 105