ABSTRACT access into locations that house sensitive

ABSTRACT

Biometric technology, previously used primarily by government
agencies, is becoming increasingly more prevalent in private industry
organizations as a more effective method in which to protect private
information and ensure access to sensitive information is restricted. These
restrictions could involve access to networks, specific programs or
applications and access into locations that house sensitive information.
Biometric technology is designed to recognize an individual person based on
physical or behavioral characteristics. Physical characteristics include
fingerprints, palm prints and facial or iris recognition while behavioral
characteristics, typically learned or acquired traits, can include voice
recognition, the gait of how an individual walks or how an individual signs
their name. The technology involved in biometrics is used for gathering,
modeling, classifying, and evaluating that information and to make run time
decisions on authenticating identification. While there is a vast selection of
different technologies available for the varying types of biometric information
as well as the algorithms that can be employed, this paper will focus on the general
technology utilized for four common types of biometric information gathered, those
being voice, fingerprint, face and iris recognition, emphasizing how the
technology is used to collect the information, model and analyze the
information and ultimately the methodologies or algorithms used for allowing
the system to make an effective and accurate decision for identification.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

 

 

 

Introduction

In the past, biometric technology was
used primarily in the areas of government for surveillance and by law
enforcement in criminal investigations, but increasingly, in today’s world of
advancing technologies and broader use of technology in our every day lives, we
are facing more and more potential threats of having our personal information accessed
or stolen by other individuals with the intent of using that information for
their own personal gain, biometric technology is seeing an increased use in
private organizations for security access measures into locations, buildings
and information systems. Biometric identification is the process of identifying
a specific individual based on unique biological traits. These traits can be
physiological such as in fingerprints, palm prints, distinct features of the
face and the patterns of the iris or they can be behavioral or learned traits
such as the gait in which a person walks, a signature or how a person speaks.

While the individual technical
components and algorithms used for biometric technology may differ depending on
the biological trait being evaluated, the technical process is similar and
typically involves four steps. First, there must be a capture process where the
sample being evaluated is captured either with an image or audio recording. The
second step is feature extraction, a process of identifying the unique features
of the sample. Third, is enrollment, a process in which algorithms are used to
create computer representations of the samples which are then stored and used
for comparison. Then the last step is verification or the process of comparing
the processed sample against other similar samples previously stored in a
database.

There are a multitude of biometric
technologies being manufactured by a variety of vendors and since each has
their own proprietary design and algorithms being used, to cover each of them
would be well beyond the scope of this paper. Instead, the focus for this paper
will be on four of the most common biological traits used in identification,
being voice or speaker recognition, fingerprint recognition, facial recognition
and iris pattern recognition, providing a general overview of the technological
process with some examples of the most common feature extraction processes and
algorithms used.

Speaker or Voice
Recognition

One of the most common forms of
biometric technology is speaker recognition or voice biometrics. An advantage
to this form of biometric technology as opposed to other forms, is that it does
not require any additional hardware for the user interface as is required in
biometric technology utilizing fingerprint, face or iris recognition.
“Automatic speaker recognition is a technology that identifies a speaker, based
on his/her speech, using a computer system” (Li, 3). Speaker recognition is
essentially a form of pattern recognition which involves feature extraction,
speaker modeling and a classification decision strategy.

Speaker recognition or voice
recognition can be classified as either text-dependent or text-independent. In
text-dependent recognition, comparison is done against a pre-defined fixed text
such as passwords or credit card numbers. Text-independent does not compare
against any pre-defined text and allows for more flexibility in a speaker or
voice recognition system, but should be noted that it is not as accurate as a
text-dependent system. A speaker recognition systems is typically either done as
identification or verification tasks. With identification, this is a one to
many comparison, identifying the person based on data contained in a speaker
database. In the verification task, this is a one to one comparison, comparing
to ensure the person is the person they are claiming to be. While the
comparison type may differ, these two tasks share the same technology
challenges of speaker modeling and classification.

The framework involved in a speaker
recognition system is that of enrollment and then verification (Figure 1).

 

Figure
1 – Enrollment and Verification Model.  

Source:
Li, Liyuan, et al. Advanced
Topics in Biometrics.

During enrollment, speech samples are provided wherein a
feature extraction process creates speech feature vectors. These vectors are
processed by vector machines creating speaker models which are used to
characterize a person based on the individual speaking features.  These speaker models are then stored in a
speaker database. 

During the verification process, the
same feature identification process used in enrollment is utilized and then
compared against the models present in the speaker database. Feature
identification is accomplished by converting the signals of speech input into a
series of feature vectors and steps to eliminate any noise also in the signal,
which is accomplished with spectral analysis and cepstral feature normalization
using popular acoustic features such as the Mel-frequency cepstral coefficients
(MFCC).  “Low-level acoustic feature
extraction is also a data-reduction process that attempts to capture the
essential characteristics of the speaker and provides a more stable and compact
representation of the input speech than the raw speech signals” (Li, 6). There
are many processing steps that occur during spectral analysis which include
removing high-frequency sections as the speech signals pass through a low-pass
filter and then separating the signals into time frames where the signal is
considered stationary. Then the signals are passed through voice activity
detection (VAD) components, to discern if the signal does in fact contain a
human voice.  Properties of the human
auditory system are then used as part of the filters of log-energy outputs for
center frequencies and bandwidths when determining the calculation of the MFCC
(Figure 2).

            Figure
2 – MFCC Calculation.   

Source: Li, Liyuan, et al. Advanced Topics in
Biometrics.

Then in order for the system to make a decision if the person
speaking is in fact who they claim to be, the comparison must include a score
normalization algorithm which applies a scale to the likelihood of the score
distribution to different speakers. Score normalization allows for
consideration to be given to undesired variations. Normalization can be
performed in different ways, but better results have been shown using cohort
normalization. Likewise, different forms of cohort normalization can be used
such as Z-norm, T-norm or H-norm, with their differences being the method in
which they estimate the bias and scale factors. This normalization score is
then compared against an identified decision threshold, thus allowing the
system to either accept or reject the individual’s claim. It bears noting, that
this is just one example of techniques used in voice or speaker recognition
systems. There are many other techniques and features that can be applied
depending on the robustness desired including Prosadic and token features,
Gaussian mixture modeling, joint factor analysis, modeling session variability
and nuisance attribute projection.

 

 

Fingerprint
Recognition

            Fingerprint
based identification is one of the oldest of the biometric technologies. Since
fingerprints are known to be unique with series of ridges and furrows,
identification can be determined by analyzing the patterns of those ridges and
furrows, as well as the minutiae points which are specific characteristics
along the ridge, either at the bifurcation or ending.  There are many different methods used to
identify fingerprints which includes the traditional minutiae matching, pattern
matching, moire fringe patterns and ultra-Sonics. “The most popular and widely
used techniques extract information about minutiae from a fingerprint and store
that information as set of points in the two-dimensional plane” (Ling,
278). 

            One of the
most important aspects of utilizing biometric technology in fingerprint
identification is capturing a quality image of the fingerprint. There are many
options to choose from to actually capture the image which we will not go into
detail about here, but a critical step to be included is an enhancement
algorithm to improve the quality of the image and the clarity of ridges and
furrows for the minutiae extraction (Figure 3).

                        Figure
3 – Fingerprint Enhancement

                        Source:
Vacca, John R. Biometric Technologies
and Verification Systems.

Other functionalities s are employed and are necessary in the
fingerprint identification process which vary depending on the technology
vendor. But feature extraction is a core functionality which includes
algorithms to determine the center of the fingerprint around which the loops,
swirls and arches of a fingerprint are located. Biometric technology vendors
may each use a different methodology and typically guard the algorithms used as
proprietary information. However, most involve some sort of reference point
detection algorithm.  “A reference point
to be used as the basis for texture features of fingerprints… the features for
verification are then extracted with respect to the position of the reference
point” (Li, 152).  One of the most widely
used methods of reference point detection is the Method of Orientation Field
Estimation (MOFE), which searches the texture field for the point of high
gradient.

            Identification
decisions in fingerprint biometric technology are then made by comparing the
minutiae or ridge patterns or reference points depending on the methodology
used by the vendor and the algorithms chosen. Some may use simple pattern
matching of minutiae or ridges (Figure 4), while others may use correlation-based
matching techniques utilizing reference points and algorithms to determine a
minimum relative point of entropy.

                        Figure 4 –
Fingerprint Minutiae Pattern Matching

                        Source: Vacca, John R. Biometric Technologies
and Verification Systems.

 

As with most other biometric
technologies, corrective considerations also must be accounted for due to lower
image quality or slightly rotated images. One such method of correction uses a
combination of Gabor feature-based extractions with detailed wavelet features
of the fingerprint image. In this method, the initial reference point is
detected, a cross correlation is determined with the reference and test image,
applying the algorithms for determining the approximate wavelet features and
the Gabor features, then determine the match using the combined Gabor-wavelet
features. However, there are studies have revealed some corrections cannot be
accounted for which produces a higher rate of errors in results and that is
human aging. In one such study, “the performance of groups was shown to be
different with an increased error rate for elderly fingerprints…average
minutiae counts were similar across ages, but elderly fingerprints exhibited
many false minutiae” (Fairhurst, 155).

Facial Pattern
Recognition

            Facial
recognition is a biometric technology in which identification is made by
utilizing distinct features of a person’s face. 
“The three main parts of the face that usually do not change are some
primary targets: the upper sections of the eye sockets, the area surrounding
the cheekbones, and the sides of the mouth” (Vacca, 95). But facial recognition
is one of the most challenging of the biometric technologies, “the primary
problem with the collected images is poor geometry: specifically pose, size,
crop and distortion” (Elsworth, 4).  As
with most other biometric technologies it involves first capturing the image of
a person’s face, processing or identifying and extracting features using
algorithms, enrollment or storage of the processed image and finally verification
or making comparisons against a database of other facial images to determine if
a match can be found. Capturing the raw images can typically be done with most
video or camera technology widely available. Processing, enrollment and
verification algorithms will vary depending on the vendor of the technology.

Facial recognition is most commonly
done by one of two methods, Eigenface or Eigenfeature. In the Eigenface method,
the image taken is changed to both light and dark areas, then points from the two
images are used for comparison. The Eigenfeature method takes a similar
approach except it chooses certain features and calculates the distance between
them such as the distance between the eyes, width of the nose and the depth of
the eye sockets. Line edge mapping (LEM) in facial recognition is a new
approach and is an enhanced variation of the Eigen methods. In LEM recognition,
images are “thinned” to become edge images which show the boundaries or
contours at which a significant change occurs. Once this has occurred, a line
edge map of the face can be generated by applying a polygonal line fitting
process (Figure 5).

                        Figure
5 – Line Edge Map

                                Source: Li, Liyuan, et al. Advanced Topics in
Biometrics

When using LEM, the basic unit used is the line segment and
the line segment Hausdorff distance (LHD) measure. “LHD has better
discriminative power because it can make use of the additional attributes of
line orientation and line-point association, i.e., it is not encouraged to
match two lines with large orientation difference, and all the points on one
line have to match to points on another line only” (Li, 75). Error correction
algorithms are also included in the LHD to account for lighting conditions and
shifting feature points due to rotation of the facial image. It is worth
noting, that as with fingerprint recognition, the natural aging process can
present challenges with the facial recognition process. “The changes in
appearance of a face with respect to age progression is contributed to by both
genetic and environmental factors such as exposure to ultraviolet light, food
habits, climate (arid vs. wet or cold vs. hot), smoking, and exercise. The
changes in appearance due to these factors are mainly attributed to the shape
and texture variations of the face” (Fairhurst, 101). These challenges are
typically resolved, at least by private organizations, by keeping up to date
comparison samples periodically as people age. 
The LHD is built on a vector representing the distance between two line
segments accounting for adjustments of angle distance, parallel distance and
perpendicular distance (Figure 6).

                                    Figure
6 – Vector algorithm

                                                Source:
Li, Liyuan, et al. Advanced
Topics in Biometrics

Weight values are then applied to balance the displacement
distance. Finally, the number of corresponding lines is compared to determine a
final match. “The number of corresponding lines between two identical images
should be larger than that between two images of different objects” (Li, 78).

Iris Pattern
Recognition

One of the biometric technologies
considered to be the most accurate is that of iris pattern recognition. The
iris is the visible part of the eye surrounding the pupil. “It is a muscular
structure that controls the amount of light entering the eye, with intricate
details that can be measured, such as striations, pits and furrows” (Vacca,
73). Mathematical algorithms are then created using these measurable features
which are then stored and used for identification. Iris pattern recognition is
considered to be one of the most accurate is due to its stability because iris
patterns are set at infancy and remain unchanged throughout a person’s lifetime
and therefore not affected by environment or age as can occur with other
biometric methods.

            Iris
pattern recognition technology is considered an opt-in technology, meaning the
user is required to cooperate with the technology by having an image of the
iris taken using visible and infrared light (Figure 6).

Figure
6 – Image of Human Iris

Source: Vacca, John R. Biometric Technologies
and Verification Systems.

Once the image of the iris is taken, it is then processed
into what is commonly referred to as Iriscode where based on the mathematical
algorithms, the measurable features of the iris are mapped into many different
vectors or phasors. These phasors are then converted and stored as hexadecimal
representations in 512 byte templates allowing for relatively fast search and
match speed.

            The
comparison process of iris recognition is usually done by comparing the Hamming
distance of biometric representations. This process takes into account the fact
that during the feature extraction process (i.e. capturing the image), that
some bits in the representation may be unreliable and therefore should be
ignored in the comparisons, which are stored as an m-bit string called the mask
and so the comparison of two biometric representations considers the mask of
each respective image being compared. The Hamming distance computation is shown
below in Figure 7.

Figure 7 – Hamming
Distance Computation. 

Source: Ling Ngo, David Chek, et al. Biometric
Security.

The calculated Hamming distance
is then compared to a specified threshold with the comparisons considered to be
a match if below the threshold.

            During
the comparison process, considerations are taken into account that the iris
representations may be misaligned. This could be a result of a person tilting
their head while the image of the iris is being captured. “To account for this,
the matching process attempts to compensate for the error and rotates a
biometric representation by a fixed amount to determine the lowest distance”
(Ling, 277). This is accomplished by representing the Iriscode as
two-dimensional bit arrays with circular rotation shifts to the left and right
a fixed number of times being applied to each row. It is then the minimum
Hamming distance from each of the rotations that is compared to the threshold
to determine if the representations are a match.

Conclusion

Biometric
technology has seen an increase in usage as a more secure method of protecting
personal and confidential information whether it be access to a particular
location where the information is housed or to the actual systems storing the
information. This increase is not only due to the increasing potential threats
but also to the advancements that have been made in biometric technology making
it easier to use, implement and more widely available to private industry,
especially the four most common biometric technologies of speaker or voice
recognition, fingerprint recognition, face recognition and iris pattern
recognition.

            Although
each biometric system may be unique in the types of technical components,
algorithms and data storage techniques that are used, they fundamentally all
utilize four basic steps of capture, feature extraction, enrollment and
verification. First, the sample to be verified must be captured. In the case of
voice or speaker recognition, this could be any device that can accept and
record audio. For fingerprint, face and iris pattern recognition, any device
that can capture an image. After the samples are collected, one of the most
important and complicated functions of biometric technology is performed,
feature extraction. In this process, the unique features of the biometric
samples are identified taking into consideration background noise (applies to
both audio recordings and video imaging), lighting and clarity of the sample.
The exact method of extraction will depend on the unique type of biometric
sample as well as the manufacturer of the technology. Algorithms are then
applied in the enrollment process so that the identified features can be
transformed into computer code and efficiently stored for comparison. As with
feature extraction, the algorithms applied in enrollment are dependent on the
biometric being analyzed as well as the manufacturer. However, many do use more
commonly known algorithmic expressions or modified versions of them such as the
Mel-frequency cepstral
coefficients in voice recognition, the Hausdorff distance in facial recognition
or the Hamming distance in iris pattern recognition. Then finally, based on the
feature extraction and enrollment processes used, a suitable threshold is
chosen for comparing the sample to the target database for the biometric
technology to make a run time decision if the identify can be verified.

            As technology in general continues to advance
biometric technology does as well, improving the efficiency and accuracy of
current biometric authentication methods and providing the opportunity to
introduce new methods, such as thermal facial recognition and vein pattern
recognition. Likewise, new technologies and methodologies are being
investigated for protecting biometric data from vulnerabilities. “Two of
vulnerable points are the security of biometric data at database and security
of biometric data at communication channel between two modules of biometric
based system” (Rohit, 5). Once such method is to watermark biometric data at
the communication channel of the biometric system. Likewise, new methodologies
are being introduced that can account, to some degree, for the natural aging
process. These advancements in technology and methodology will help continue
the trend of making biometric technology and security systems a reliable and
secure method for identity authentication.