www.biosec.org
Project’s
Rationale
In December 2003 the BioSec project, co-funded by the VI Framework Programme of
the European Commission, started its two years life cycle with the purpose to
consolidate and progress research on biometric technologies.
Biometric technologies provide
automated means to identify a subject according to physiological or
psychological features. These technologies have a high potential to guarantee
the identity of the individuals and they appear as the key technologies to provide
trust and secure access to private information. Nevertheless, this technology
is not completely consolidated or accessible with the required features in
security, trust and confidence to the public nowadays. This is due to several
factors such as the lack of standards defining the communication and data
models to guarantee the security and integrity of the exchanged information,
the lack of maturity of the technology with the consequent vulnerability to
fake and security attacks, the lack of technological solutions to guarantee the
privacy and ownership of the biometric data, or the incompatibility among
biometric components, devices, systems, and data available in the market. All
these factors contribute to a low usability and acceptance by the citizens. BioSec has the challenge to solve these problems by working
in several clearly defined action lines.
BioSec is the project that
will bring deployment of a European-wide approach to biometric technologies for
security applications. It has the support of the European Commission and a
consortium, led by Telefónica I+D, with twenty
three organizations from nine countries (
It will provide a ‘digital me’ for
the European citizen. From the technological point of view, BioSec
will provide improved performance to: novel 3D face and hand method,
noise-cancellation based voice verification method, as well as with emphasis on
multimodal biometrics, including face-voice and iris-finger combinations used
together with advanced classification methodology. Fake-resistive methods will
be developed for high security fingerprint and iris verification.
In order to put the biometric
technologies in the final user’s hands –the citizen– and increase their
confidence in them, the results of the research activities to be carried out in
BioSec are directed to produce personal biometric
storage on smart cards and personal privacy portable devices. New ID tokens
will be addressed on BioSec objectives. Interoperable
token devices, integrating biometric information (of any modality that is in BioSec agenda) in a secure and tamper-free solution will
enable a new generation of devices that will transform the user privacy concept
into a real concept, under user control.
As BioSec
is a Research Technology and Development project on pan-European scope, it has
defined an efficient strategy for the wider dissemination of its results and
achievements. Several communication channels have been established to be
present at different levels: creation of an observatory on those standardization
activities related to the biometric technologies, creation of a web site with
updated information on the project progress (www.biosec.org), organization of
bi-annual open workshops to present and discuss the project achievements,
participation in scientific conferences and events, publication of articles and
white papers, etc.
Research
Objectives
BioSec is an Integrated Project (IP) where Biometrics
and security play together to leverage trust and confidence in a wide spectrum
of everyday applications. Partners from nine countries constitute a critical
mass in the Biometric area (including big companies, biometric HW/SW producers,
prestigious universities and SMEs).
Recent
investigations carried out in
Performance
evaluation will be conducted throughout the whole project in a sound and
statistically relevant way, under a specifically-created committee control.
Significant efforts will be devoted to contributions to standards, co-operation
with existing organizations (such as NoE), and to
dissemination of the results.
BioSec activities include development of
demonstrators in various scenarios (such as airport security and remote
authentication) for which important final users have been already involved.
Other technology deployments will be made during the project lifetime, driven
by the first successful research results; the participation of big industrial
partners (with big customers) ensures a very large catchment
area.
A
strong management with well established roles and responsibilities will ensure
proper decisions to be promptly taken in order to successfully drive the
project towards its ambitious objectives.
Our Contribution
Improving discriminant analysis for face verification
A novel algorithm that can be used to boost the performance of
face authentication methods that utilize Fisher’s criterion is presented. The
algorithm is applied to matching error data and provides a general solution for
overcoming the “small sample size” (SSS) problem, where the lack of sufficient
training samples causes improper estimation of a linear separation hyper-plane
between the classes. Two independent phases constitute the proposed method.
Initially, a set of locally linear discriminant models is used in order to
calculate discriminant weights in a more accurate way than the traditional
linear discriminant analysis (LDA) methodology. Additionally, defective
discriminant coefficients are identified and re-estimated. The second phase defines
proper combinations for person-specific matching scores and describes an
outlier removal process that enhances the classification ability. Our technique
was tested on the M2VTS and XM2VTS frontal face databases. Experimental results
indicate that the proposed framework greatly improves the authentication
algorithm’s performance.
Description of Algorithm
We propose a framework for using discriminant techniques that can overcome the small sample size problem, after face verification matching features are extracted by an elastic graph matching algorithm for face verification. Elastic graph matching is based on the analysis of a facial image region and its representation by a set of jets which are local descriptors (i.e. feature vectors) extracted at the nodes of a sparse grid:
,
where 
denotes the output of a local
operator applied to image
at the 
scale or the pair (scale, orientation), 
defines the pixel coordinates and 
denotes
the dimensionality of the feature vector. The grid nodes are either evenly
distributed over a rectangular image region or placed on certain facial
features (e.g., nose, eyes, etc.) called fiducial
points. The core of the elastic graph matching is to translate and deform the
reference graph on the test image in order to find the correspondences of the
nodes of the reference graph on the test image. That is accomplished by
minimizing the same cost function that employs node jet similarities and
preserve the node relationships.
The algorithm that was developed in order to enhance the face verification performance of an arbitrary elastic graph matching algorithm can be summarized in the following two steps:
- Weigh the similarity values at
the nodes of the elastic graphs by coefficients that derive a multiple
local discriminant analysis scheme.
- Weigh the classification scores
that are derived by matches done on multiple reference images of the same
subject by using discriminant analysis.
The first
step of our algorithm is to find a person specific discriminant transform of
the new local similarity values and create the total similarity measure. Let 
be a column vector comprised by the
new local similarity values at every node. The algorithm should learn a
discriminant function 
that is person specific and form the
total similarity measure between faces:
.
During this
stage of the algorithm, traditional LDA was found to sometimes degrade the
classification performance because of the SSS problem where
the dimensionality of the samples is larger than the number of available
training samples. A
novel alternative method was used to define the discriminant weighting vector 
that is used to form a total
similarity, or distance, measure between the reference face and a test person,
as such:
.
Specifically,
was calculated by
implementing the following procedure:
·
For claims related to each reference person 
, grid
nodes that do not possess any discriminatory power are identified and
discarded.
· In order to provide better estimation to Fisher’s linear discriminant, the between-class scatter matrix is redefined so as to accommodate the prior probabilities of how well the mean of each class - client or impostor - was estimated. Thus, the so called weighted between-class scatter matrix is formed.
·
In order to give remedy to the
SSS problem each matching vector is broken down to a number of smaller
dimensionality vectors of equal lengths. Thus, 
subsets are formed and
separate LDA processes
are carried out. For the XM2VTS data, the optimum value for the length of these
subsets is found and used. In addition, discriminant coefficients that are
badly estimated due to a small training set, or due to any large non-linearity
between corresponding subsets, are identified and re-estimated in an iterative
fashion, if needed.
·
Each of the weight vectors
produced, having smaller dimensionality than the original feature vectors, is
normalized so that their within class variance equals to one. This
normalization step enables the proper merging of all the subset weight vectors
to form the overall weight vector .
During the second step of the algorithm that was
developed, the set of matching scores that correspond
to each person’s reference photos is used in a supplementary discriminant
analysis step. Once again, the modified Fisher’s criterion that uses the
weighted between-class scatter matrix is used. In particular, person specific
weights, 
, are calculated for the 
scores, corresponding
to reference person
images, for the problem that is formulated as such:
.
In addition, this step is complemented by an outlier removal process related to the number of the minimum impostor matching scores in the training set of each reference person. Successively, the final verification decision that is a weighted version of the sorted matching scores is produced.
Evaluation Process
The XM2VTS database contains 295 subjects, 4 recording sessions and two shots (repetitions) per recording session. The data is divided in three different sets the training set, the evaluation set and the test set. The training set is used to create client and impostor models for each person. The evaluation is used to learn the thresholds. In order to evaluate the DEGM algorithm the ‘Configuration I’ experimental setup of the XM2VTS database was used. For this configuration the training set has 200 clients, 25 evaluation impostors and 70 test impostors. The training set of the Configuration I contains 200 persons with 3 images per person. The evaluation set contains 3 images per client for genuine claims and 25 evaluation impostors with 8 images per impostor. Thus, evaluation set gives a total of 3 × 200 = 600 client claims and 25 × 8 × 200 = 40.000 impostor claims. The test set has 2 images per client and 70 impostors with 8 images per impostor and gives 2 × 200 = 400 client claims and 70 × 8 × 200 = 112.000 impostor claims. For the M2VTS database the Brussels protocol was implemented that employs the ‘leave-one-out’ and ‘rotation’ estimates, and a total of 5,328 client claim tests and 5,328 impostor claim tests (1 client or impostor x 36 rotations x 4 shots x 37 individuals) were carried out.
Results
Let us call the combination of the morphological elastic graph
matching, EGM, and the weighting approach that makes up for the first phase of
the proposed algorithm, as is described in the previous section as LEGM.
Moreover, let LLEGM be the second phase of our algorithm. In order to evaluate
the performance of these methods the False
Acceptance (FAR) and False Rejection (FRR) rate measures are
used. Figure 1-A shows a critical region of the ROC curves for the raw EGM data
using (4), classical LDA (7) applied on the raw EGM data, LEGM and LLEGM
evaluated on the M2VTS database. Figure 1-B shows the same corresponding ROC
curves when the algorithms were evaluated on the XM2VTS evaluation set and
Figure 1-C the corresponding ones for the XM2VTS test set. Results are
presented in logarithmic scales. In addition, Table 1 shows the equal error
rates (EER) for each algorithm, a
common face authentication evaluation measure that is specified as the point
where FAR and FRR are identical. The evaluation tests on the two databases show
that for both FAR and FRR, LEGM is indisputably a better performer
than either EGM or LDA while LLEGM almost always provides additional
improvement to the classification ability of LEGM.
Figure 1: ROC curves. A: M2VTS; B: XM2VTS – evaluation set; C: XM2VTS – test set.
|
Method |
EER (%) |
||
|
M2VTS |
XM2VTS – Evaluation Set |
XM2VTS – Test Set |
|
|
EGM |
6.06 |
7.33 |
8.51 |
|
LDA |
8.94 |
5.32 |
5.75 |
|
LEGM |
4.37 |
4.67 |
2.51 |
|
LLEGM |
4.17 |
3.90 |
2.46 |
Table 1: Evaluation results for each
process.
BioSec Partners:
Telefónica Investigación
y Desarrollo S.A.U. (Spain); Sisäasiainministeriö
(Ministry of the Interior Finland); Finnair Plc
(Finland); Tampereen Yliopisto:
University of Tampere (Finland); VTT Electronics
(Finland); Biometrika SRL (Italy); Universita Di Bologna (Italy); Etra I+D (Spain); Universidad Carlos III de Madrid (Spain);
Universidad Politecnica de Madrid (Spain); Universitat Politecnica de Catalunya (Spain); Ibermática
S.A. (Spain); Universität zu
Köln (Germany); MediaScore Gesellschaft
für Medien- und Kommunikationsforschung mbH
(Germany); Siemens AG (Germany); Giesecke & Devrient Gmbh (Germany); ATMEL Grenoble S.A. (France); Centre for Research and Technology
Hellas (Greece); Expertnet A.E. (Greece); Aristotle
University Of Thessaloniki (Greece); VCON Telecommunications Ltd. (Israel); Naukowa i Akademicka
Siec Komputerowa: NASK
(Poland); Katholieke Universiteit
Leuven (Belgium).