A Multispectral Sensor for Fingerprint Spoof Detection
Jan 1, By: Robert K. Rowe
Biometric sensors are devices
that can collect information capable of performing an automatic
determination of a person's identity. Common types of biometric
technologies include facial recognition, iris recognition, and
fingerprint sensing. Although not yet ubiquitous, biometric
sensors—and particularly fingerprint sensors—are poised to become
much more prevalent in many people's lives for both commercial and
government applications.
Recent commercial introductions
of fingerprint technology include an IBM laptop computer with a
built-in fingerprint sensor, and a line of fingerprint-enabled
computer products from Microsoft including a keyboard and mouse. The
goal of both products is to offer an alternative to user names and
passwords as a means to authorize access to computers, networks, and
services.
In the aftermath of 9/11, the
federal government and other governing bodies around the world have
turned to biometrics to verify the identity of foreign travelers
entering and leaving the U.S. and other countries. Furthermore, as
networked digital devices become more prevalent in warfare, there is
a growing movement toward biometric sensing to ensure that the
person using a particular device is authorized to do so. Fingerprint
sensors are among the leading candidate technologies for these and
many other security applications.
Conventional fingerprint sensors read the
superficial friction ridge patterns of the skin on the fingertips.
Common sensor types include capacitive, radio frequency, thermal,
and optical arrays. Although each sensing modality is fundamentally
different from the others, each generates an image that
distinguishes between points of contact with the sensor (fingerprint
ridges) and points where there is a gap (fingerprint valleys).
A common problem with current
fingerprint sensors is that they are an easy target for
"spoofing"—the art of using artificial samples to imitate real and
authorized fingerprint patterns. While this security breach is a
concern for all biometric technologies, the problem is particularly
pronounced for fingerprint sensors because people leave copies of
their fingerprints on most objects they touch throughout the day. A
quick search on the Internet can provide a motivated individual with
enough information to convert a latent fingerprint into an effective
spoof using familiar materials that can be bought at a hobbyist
supply shop.
Figure 1. In a typical optical
fingerprint sensor based on total internal reflectance (TIR),
light enters from the left of the prism and is diffusely
reflected from the right-side facet to uniformly illuminate
the sample surface. The imager detects bright regions where
air gaps are located and darker areas where skin or other
material of appropriate refractive index is in contact with
the glass (inset).
|
The reason most fingerprint sensors are
susceptible to spoofing is that little or no information is acquired
specific to the object touching the sensor. As long as the sensor
can produce a pattern sufficiently close to that of the enrolled
fingerprint, authorization is granted. As an example, consider a
common type of optical fingerprint sensor based on total internal
reflectance (TIR), as shown in Figure 1. An image is formed when a
material with an appropriate index of refraction contacts the sensor
surface. This material can be human skin, but it can also be
silicone, gelatin, or a variety of other substances.
Furthermore, superficial
fingerprint patterns may be worn, damaged, or simply hard to read
due to the skin's surface conditions—too wet or too dry, for
example. Collecting fingerprints of good quality from older people
is particularly difficult because their skin is likely not to be
very supple. The resulting poor-quality images, even if collected
from a properly authorized person, can lead to authorization
rejections. The system administrator might then set the sensor's
security threshold to a more lenient setting to reduce the number of
false negatives. This relaxed threshold further exacerbates the
susceptibility of a sensor to spoof attacks.
A very powerful way to address
these security concerns is by collecting a multispectral image of
the skin directly below the surface fingerprint. Living human skin
has certain unique optical characteristics due to its chemical
composition, which predominately affects optical absorbance
properties, as well as its multilayered structure, which has a
significant effect on the resulting scattering properties. By
collecting images generated from different illumination wavelengths
passed into the skin, different subsurface skin features may be
measured and used to ensure that the material is living human skin.
When such a multispectral sensor is combined with a conventional
fingerprint reader, the resulting sensing system can provide a high
level of certainty that the fingerprint originates from a living
finger.
Figure 2. The key components of a
multispectral imager include illumination sources (LEDs of
various wavelengths), an imaging array (silicon CCD or CMOS),
and a crossed polarizer arrangement that emphasizes light that
has undergone multiple scatter events in the
skin.
|
The key components of a multispectral
imager suitable for imaging fingers are shown in Figure 2 The light
sources are LEDs of various wavelengths spanning the visible and
short-wave IR region. Crossed linear polarizers may be included in
the system to reduce the contribution of light that undergoes a
simple specular reflection to the image, such as light that is
reflected from the surface of the skin. The crossed polarizers
ensure that the majority of light seen by the imaging array has
passed through a portion of skin and undergone a sufficient number
of scattering events to have randomized the polarization. The
imaging array is a common silicon CMOS or CCD detector.
In general, the optical
resolution requirements for this application of a multispectral
imager are relatively modest. Because of the highly scattering
nature of skin, the multispectral imager need not have a greater
resolution than that used for the conventional fingerprint image,
typically 250–1000 pixels/in. Assuming a nominal 1 in. sensing
surface, a readily available VGA or 1.3 megapixel array provides
adequate resolution for most applications.
Figure 3. This conceptual layout
illustrates an optical fingerprint sensor that combines a
conventional TIR-based imager and a multispectral imager. Both
can view a finger
simultaneously.
|
Figure 3 shows the layout of an optical
fingerprint sensor that combines a conventional total internal
reflection (TIR) fingerprint reader with a multispectral imager. As
can be seen, both sensors can view a finger placed on the sensing
surface without interfering with each other. The multispectral
imager can thus provide significant new biometric information
without requiring any different or additional actions on the part of
the user.
Figure 4. The print on the left of a dry
fingertip was taken with a conventional TIR imager. On the
right is the same finger, but imaged with a multispectral
sensor. The latter image was collected using five wavelengths
(475, 500, 560, 576, and 625 nm). It is shown here as a
pseudo-color representation of the first three factors
produced by a decorrelation-stretching technique operating on
the original five image
planes.
|
An example of a conventional fingerprint
image and a multispectral image of the same finger is illustrated in
Figure 4. In this case, the skin of the subject's finger is
relatively dry, causing a noticeable deterioration in the contrast
and continuity of the lines in the conventional fingerprint image.
In contrast, the multispectral pseudocolor image shows spectral and
spatial features that are well defined and consistent with a living
finger. In addition, the fingerprint image is observable in the
multispectral data, which can be used to further authenticate the
conventionally collected fingerprint pattern as well as to augment
missing or poorly defined portions of the conventional fingerprint.
Figure 5. Multispectral image data can
clearly discriminate between a living finger and an
ultra-realistic spoof. The graphs on the left show how similar
the spectral content of each image is to that expected for a
genuine finger.
|
A highly realistic artificial finger made
by Alatheia Prosthetics (Brandon, MS) was one of a number of
different spoof samples used to test a multispectral imager's
ability to discriminate between real fingers and spoofs. Figure 5
shows the results of a multivariate spectral discrimination
performed to compare the consistency of the spectral content of a
multispectral image of a real finger with both a second image of a
real finger and a prosthetic replica of the same finger. The
imager's ability to distinguish between the two sample types is
clear.
Lumidigm is currently working
with partners to integrate the multispectral technology into
conventional optical fingerprint sensors. The resulting product is
scheduled to become available later this year for applications
including Homeland Security and commercial physical access. One
anticipated benefit is a level of security and usability that goes
far beyond today's fingerprint sensors and that will pave the way to
broader adoption of fingerprint sensors as the biometric of choice.
This material is based on work
supported by the Air Force Research Laboratory, Rome, NY, under
contract number FA8750-04-C-0190. Any opinions, findings,
conclusions, or recommendations expressed herein are those of the
author and do not necessarily reflect the views of the Air Force
Research Laboratory.