27-08-2014, 12:53 PM
MIXING FINGERPRINTS FOR TEMPLATE SECURITY AND PRIVACY
MIXING FINGERPRINTS.pdf (Size: 1.32 MB / Downloads: 34)
ABSTRACT
Securing a stored fingerprint image is of paramount importance because a compromised fingerprint cannot be easily revoked. In this
work, an input fingerprint image is mixed with another fingerprint
(e.g., from a different finger), in order to produce a new mixed image that obscures the identity of the original fingerprint. Experiments on the WVU and FVC2000
datasets show that the mixed fingerprint can potentially be used for
authentication and that the identity of the original fingerprint cannot
be easily deduced from the mixed fingerprint. Further, the mixed
fingerprint can facilitate in the generation of cancelable templates
1. INTRODUCTION
Preserving the privacy of the stored biometric template (e.g.,fingergerprint image) is necessary to mitigate concerns related to data
sharing and data misuse [10]. This has heightened the need to impart privacy to the stored template, i.e., to de-identify it in some
way. De-identifying biometric templates is possible by transformformed in this way is referred to as a cancelable template since it
can be “canceled” by merely changing the transformation function
[3] [17]. At the same time, the transformed template can be used
during the matching stage within each application while preventing
cross-application matching. Further, the transformation parameters
can be changed to generate a new template if the stored template is
deemed to be compromised.
There has been a vast amount of work done in generating a
cancelable fingerprint template [16].1
In this study, we consider
the problem of mixing two fingerprint images in order to generate
a new cancelable fingerprint image. The mixed image incorporates
characteristics from both the original fingerprint images, and can be
used in the feature extraction and matching stages of a biometric
system.
2. MIXING FINGERPRINTS: THE PROPOSED
APPROACH
The ridge flow of a fingerprint can be represented as a 2D Amplitude
and Frequency Modulated (AM-FM) signal [13]:
I(x, y) = a(x, y) +b(x, y)cos(Ψ(x, y)) +n(x, y), (1)
where I(x, y) is the intensity of the original image at (x, y), a(x, y)
is the intensity offset, b(x, y) is the amplitude, cos(ψ(x, y)) is
the phase and n(x, y) is the noise. Based on the Helmholtz
Decomposition Theorem [5]the phase can be uniquely decom
2.1 Fingerprint Decomposition
Since ridges and minutiae can be completely determined by the
phase, we are only interested in Ψ(x, y). The other three parameters in Equation (1) contribute to the realistic textural appearance
of the fingerprint. Before the decomposition task, the phase Ψ(x, y)
must be reliably estimated; this is termed as demodulation.
2.1.1 Vortex demodulation
The objective of vortex demodulation [12] is to extract the amplitude b(x, y) and phase Ψ(x, y) of the fingerprint pattern. First, the
DC term a(x, y) has to be removed since the failure to remove this
offset correctly may introduce significant errors in the demodulated
amplitude and phase [12]. To facilitate this, a normalized fingerprint
image, f(x, y), containing the enhanced ridge pattern of the fingerprint (generated by the VeriFinger SDK3
) is used. From Equation
(1), f(x, y) = I(x, y)−a(x, y) ≃ b(x, y)cos(Ψ(x, y)).
2.1.3 Helmholtz Decomposition
The Helmholtz Decomposition Theorem [5] is used to decompose
the determined phase Ψ(x, y) of a fingerprint image into two phases.
The first phase, ψc is a continuous one, which can be unwrapped,
and the second is a spiral phase,ψs, which cannot be unwrapped but
can be defined as a phase that exhibits spiral behavior at a set of
discrete points in the image. The Bone’s residue detector [1] [5]
is first used to determine the spiral phase ψs(x, y) from the demod
2.2 Fingerprint Pre-alignment
To mix two different fingerprints after decomposing each fingerprint
into its continuous component cos(ψc(x, y)) and spiral component
cos(ψs(x, y)), the components themselves should be appropriately
aligned. Previous research has shown that two fingerprints can be
best aligned using their minutiae correspondences. However, it is
difficult to insure the existence of such correspondences between
two fingerprints acquired from different fingers. In this paper, the
components are pre-aligned to a common coordinate system prior
to the mixing step by utilizing a reference point and an alignment
line. The reference point is used to center the components. The
alignment line is used to find a rotation angle about the reference
point. This angle rotates the alignment line to make it vertical. The
two phase components of each fingerprint are rotated by the same
angle
2.2.1 Locating a reference point
The reference point used in this work is the northern most core point
of extracted singularities. For plain arch fingerprints or partial fingerprint images, Novikov et al.’s technique [15] [18], based on the
Hough transform, is used to detect the reference point
2.3 Mixing Fingerprints
Let F1 and F2 be two different fingerprint images from different
fingers, and let ψci(x, y) and ψsi(x, y) be the pre-aligned continuous
and spiral phases, i = 1,2. As shown in Figure 1, there are two
different mixed fingerprint image that can be generated, MF1 and
MF2:
MF1 = cos(ψc2 +ψs1),
MF2 = cos(ψc1 +ψs2).
(6)
The continuous phase of F2 (F1) is combined with the spiral phase of
F1 (F2) which generates a new fused fingerprint image MF1 (MF2)
3. EXPERIMENTS AND DISCUSSION
The performance of the proposed fingerprints mixing approach was
tested using two different datasets. The first dataset was taken
from the West Virginia University (WVU) multimodal biometric
database [2]. A subset of 300 images corresponding to 150 fingers
(two impressions per finger) was used. The second dataset was the
FVC2000 DB2 fingerprint database containing 110 fingers with 8
impressions per finger (a total of 880 fingerprints). The VeriFinger
SDK was used to generate the normalized fingerprint images and
the matching scores. Also, an open source Matlab implementation
[11] based on Hong et al.’s approach [8] was used to compute the
orientation and frequency images of the fingerprints.