STARFAST: a Wireless Wearable EEG Biometric System based on the ENOBIO Sensor H.S.PAVAN KUMAR
C.ABDUL KHALEEK
Madina engineering college
Madina engineering college
ABSTRACT—we present a wearable,
Rate (FAR) of 0.82%. The obtained
wireless biometry system based on the new
performance measures improve the results
ENOBIO
of similar systems presented in earlier work.
4-
channel
electrophysiology
recording device developed at Star lab.
I. INTRODUCTION
Features
from
The present paper introduces a method to
and
authenticate people from their physiological
electrocardiogram (ECG) recordings have
activity, concretely the combination of ECG
proven to be unique enough between
and EEG data. We call this system
subjects for biometric applications. We
STARFAST (STAR Fast Authentication
show here that biometry based on these
bio-Scanner
recordings offers a novel way to robustly
modalities are already being exploited
authenticate or identify subjects. In this
commercially for person authentication:
paper, we present a rapid and unobtrusive
voice recognition, face recognition and
authentication method that only uses 2
fingerprint recognition are among the more
frontal electrodes and a wrist worn electrode
common modalities nowadays. But other
referenced to another one placed at the ear
types of biometrics are being studied as
lobe. Moreover, the system makes use of a
well: ADN analysis, keystroke, gait, palm
multistage
extracted
electroencephalogram
Test).
Several
biometric
architecture,
which
print, ear shape, hand geometry, vein
improved
system
patterns, iris, retina and written signature
performance. The performance analysis of
[33]. Although these different techniques for
the system presented in this paper stems
authentication exist nowadays, they present
from an experiment with 416 test trials,
some problems. Typical biometric traits,
where an Equal Error Rate (EER) of 0% is
such as fingerprint, voice and retina, are not
obtained after the EEG and ECG modalities
universal, and can be subject to physical
fusion and using a complex boundary
damage (dry skin, scars, loss of voice...). In
decision. If a lineal boundary decision is
fact, it is estimated that 2-3% of the
used we obtain a True Acceptance Rate
population is missing the feature that is
(TAR) of 97.9% and a False Acceptance
required for authentication, or that the
demonstrates
fusion
(EEG)
to
provided biometric sample is of poor
problem of entity authentication in body
quality. Furthermore, these systems are
area sensor network (BASN) for m-Health.
subject of attacks such as presenting a
Physiological signals detected by biomedical
registered deceased person, dismembered
sensors have dual functions: (1) for a
body part or introduction of fake biometric
specific medical application, and (2) for
samples. It also compares the performance
sensors in the same BASN to recognize each
of the system with the Active Two system
other by biometrics. A feasibility study of
from Biosemi. The Active Two system is
proposed entity authentication scheme was
state of the art commercial equipment that is
carried out on 12 healthy individuals, each
proven to have very good performance and
with 2 channels of photoplethysmogram
has been used in many studies. A list of
(PPG) captured simultaneously at different
some works can be found in [1]. So far 3
parts of the body. The beat-to-beat heartbeat
demonstration
been
interval is used as a biometric characteristic
developed using the system. EOG based
to generate identity of the individual. The
Human Computer Interface (HCI), EEG and
results of statistical analysis suggest that it is
ECG based Biometry for Authentication
a possible biometric feature for the entity
(presented at pHealth as a poster) and an
authentication of BASN.
applications
have
EEG based Sleepiness Prediction system for
Advantages
Since
every
drivers. In all cases the applications are
living and functional person has a recordable
designed around the same 4 channel system.
EEG/ECG signal, the EEG/ECG feature is
ENOBIO has been developed with the help
universal. Moreover brain or heart damage
of the SENSATION FP6 IP 507231.
is something that rarely occurs, so it seems
With the evolution of m-Health, an
to be quite invariant across time. Finally it
increasing number of biomedical sensors
seems very difficult to fake an EEG/ECG
will be worn on or implanted in an a
signature
individual in the future for the monitoring,
biometric system.
diagnosis, and treatment of diseases. For the
system
optimization of resources, it is therefore
characteristics:
necessary to investigate how to interconnect
friendliness, fast operation and low cost. The
these sensors in a wireless body area
perfect biometric trait should have the
network, wherein security of private data
following characteristics: very low intra-
transmission is always a major concern. This
subject variability, very high inter-subject
paper proposes a novel solution to tackle the
variability, very high stability over time and
or
to
should
attack
an
EEG/ECG
An ideal biometric present
100%
the
following
reliability,
user
universality. In the next section we show the
four channels plus electrical reference can
general
be recorded. The hardware is power with a
architecture
performance
of
the
and
the
system
global
we
have
developed.
rechargeable
Li-Ion
battery
for
long
duration. The communication of ENOBIO with the Laptop is done via ZigBee to USB
SYSTEM
DESCRIPTION
AND
SPECIFICATIONS:
Ag/AgCl or optionally dry. We can see in
ENOBIO is a wearable, wireless, 4channel,
all-digital
receiver. The electrodes can be active
electrophysiology
Figure 1 ENOBIO recording EEG data and transmitting them wirelessly to the laptop to
recording system that has been optimized for
be processed for an on-line application.
dry electrodes. The system has been
METHODS AND RESULTS:
designed in a modular way to allow
In this paper, we present a rapid and
researchers to build systems tailored to their
unobtrusive
needs. Robust, snap-in cables of any length
method that only uses 2 frontal electrodes
allow any combination of EEG, EOG and
(for EEG recording) with another electrode
ECG in a single wearable system without
placed
loss of signal quality. ENOBIO employs
recording)and all 3 referenced to another
three fundamental enabling aspects to bring
one placed at the right earlobe. Moreover the
electrophysiology into daily life:
system makes use of a multi-stage fusion
- Active digital electrodes digitize
on
authentication/identification
the
left
wrist
(for
ECG
architecture, which has been demonstrated
the signal on site in order to reduce
to
environmental noise.
EEG/ECG recording device is ENOBIO [23,
improve
system
performance.
The
24, 32, 35], a product developed at - Modular design allows users to customize
STARLAB BARCELONA SL. It is a
a single system for various applications
wireless 4 channel (plus the common mode) device with active electrodes. It is therefore
- Wireless communication allows wearable
quite unobtrusive, fast and easy to place.
recording of data away from the lab or
Even though ENOBIO can work in dry
controlled environments. The precision of
mode, in this study conductive gel has been
the signal that can be acquired with
used (see fig. 1).
ENOBIO is 1-3 uVpp to 1 mV from 0.1 to 100 Hz, and the output is a digital signal at 256 Samples/s per channel, 16 bit. Up to
decision.
The
enrolment
and
the
authentication test follow a common path until the feature extraction. Then the red arrows indicates the enrolment path and the green ones the authentication path. Briefly, the preprocessing module segments the recorded takes into 4 seconds epochs. From Figure1: ENOBIO sensor embedded in a
each one of those epochs we extract 5
cap. We can see the 4 channel inputs (red
different
pins). In the picture, only two channels are
coefficients, fast Fourier transform, mutual
plugged. The data are transmitted wirelessly
information, coherence and cross correlation
to the laptop. The general schema used is
[10, 11, 12, 19, 20, 21]. The classifier used
described below: Fig. 2 General scheme of
in both the signature extraction module and
the method. The data acquisition module is
in the recognition module is the Fisher
the software that controls the ENOBIO
Discriminant Analysis [13], with 4 different
sensor in order to capture the raw data.
discriminant functions
Remember that 4 channels are recorded: 2
linear, quadratic and diagonal quadratic).
EEG channels placed in the forehead, 1
During the enrolment process, we need four
ECG channel placed in the left wrist and 1
3-minutes takes, and these are used to select
electrode placed in the right earlobe for
the best combinations of features and
referencing the data. At this point the data is
classifiers, and then to train the classifiers
separate in EEG data and ECG data and sent
for each subject. This new feature is what
to two parallel but different biometric
we call ‘personal classifier’ approach, and it
modules for EEG and ECG. (the difference
improves the performance of the system
is not shown in the scheme for simplicity).
considerably. The authentication test takes
For the signature extraction module, four 3-
are 1-minute long. The preprocessing and
minutes
the feature extraction are equivalent to the
takes
are
needed.
Once
the
features:
auto-regression
(linear,
diagonal
signatures are extracted, they are both stored
enrolment case.
in the database for further retrieval when an
loads the trained classifiers for each
authentication process takes place. Then the
modality
recognition
independent score is then provided for each.
module
provides
an
(ECG
The recognition module and
authentication score for both modalities and
INITIAL
SYSTEM
finally the fusion module provides the final
BENCHMARKING:
EEG)
and
an
For the signal analysis we compared the signal recorded with ENOBIO and Biosemi simultaneously. The electrodes of every system where placed in Fp1 of 1 cm between them. We analyzed the signal referenced to left ear lobe. We recorded spontaneous EEG, 1 minute with eyes open And 1 minute with the eyes closed for 2 subjects, using wet Ag/AgCl electrodes for
Figure 3: Spectral density of the signal
both systems. Figures 2 to 5 show the power
recorded with ENOBIO and Biosemi in fp1
spectral density of the signals recorded with
referenced to left ear lobe. Sample was 1
ENOBIO (blue line) and Biosemi (red line).
minute of spontaneous EEG with eyes close.
We can see a very good match in the
DEMONSTRATION
spectrums of both systems. Just a small peak
APLICATIONS:
of 50 Hz appears for ENOBIO for subject 1.
Tree different applications have been built on ENOBIO system. In this section we make a fast overview of them. A EOG based Human
Computer
Interface.
Star
lab
developed a system to track eye movements. When we look at a scenario, we focus our objects of interest changing our view with fast eye movements. These movements are detected by the system and their amplitude And direction extracted from the EOG Figure 2: Power spectral density of the
signal. The signal is acquired by ENOBIO
signal recorded with ENOBIO and Biosemi
and sent wirelessly to the ENOBIO server
in fp1 referenced to left ear lobe. Sample
installed in a laptop. This raw signal can be
was 1 minute of spontaneous EEG with eyes
sent via TCP/IP to any application, in that
open.
case to the EOG HCI system. The application processes the data in real time and outputs a vector that represents the movement on the screen. The
System is set on a cap as shown in Figures 1
himself), 350 impostor situations (when an
and 8.
enrolled subject claims to be another subject from the database) and 16 intruder situations (when a subject who is not enrolled in the system claims to be a subject belonging to the database). Once the EEG and the ECG biometrics results are fused, using a complex boundary decision (dotted line 2 in
Figure 4: EOG-HMI system. 3 electrodes are placed on the forehead; the 4th one and the electrical reference are placed in the ear lobes with a clip. Figure 7 shows a graphical representation of the output of a movement.
fig. 3), we can obtain an ideal performance, that is True Acceptance Rate (TAR) = 100% and False Acceptance Rate (FAR) = 0%. If a linear boundary decision is used (line 1 in fig. 3), we obtain a TAR = 97.9% and a FAR = 0.82%. The results are summarized in table I. Bidimensional decision space. An ordinate represent the ECG probabilities and abscises the EEG probabilities. Red crosses represent impostor/intruder cases and green crosses represent legal cases. Two decision functions are represented this system has been tested as well to validate the initial
Figure 5: The vector of the picture
state of users, and has been proved sensitive
represents the eye gaze movement of a user
enough to detect it. If a subject has suffered
of the EOG Human Computer Interface. The
from sleep deprivation [28], alcohol intake
absolute error in the prediction for the
[26, 27] or drug ingestion when passing an
movements is 5, 9 % with a standard
authentication
deviation of 6.6 for the X axe, and 9, 9%
performance decreases. This fact provides
with a standard deviation of 13, 1 for the Y
evidence that such a system is able to detect
axe. Then the fusion module [22] provides a
not only the identity of a subject but his state
final
as well.
decision
authentication.
about In
order
the to
subject test
the
performance of our system we use 48 legal situations (when a subject claims to be
TABLE I
test,
the
authentication
validate in a continuous way that the person supposed to be tele-present in an audiovisual interactive space is actually the person that is supposed to be. This could facilitate the personalization of the reaction of the virtual
CONCLUSION
AND
FUTURE
WORK: These
results
show
that
the
authentication of people from physiologic data can be achieved using techniques of machine learning. Concretely it shows that the fusion of two (or more) independent biometric
modules
increases
the
performance of the system by applying a fusion stage after obtaining the biometric scores. This result shows that processing the different physiological modalities separately on
different
processing
modules,
and
introducing a data fusion step, the resulting performance can be increased. Applying a very similar approach, we could easily adapt the system to do emotion recognition [29, 30, and 34] from physiological data, or develop a Brain Computer Interface, just starting from different ground truth data. From our point of view, this easy to extend feature of our system is the more interesting part of our study along with the ‘personal classifier’
approach
which
improves
considerably the performance of the system. The system described could have different applications for Virtual Reality. It can
environment [31], or secure interactions that guarantee the authenticity of the person behind. Although the system described here was oriented towards person identification, its performance has been significantly modified by introducing physiological data obtained in altered states such as the ones resulting from sleep deprivation. This indicates that such a system could be used to extract dynamically changing, such as physiological activity related to mood or to intense cognitive activity. For example, it could be used to extract the features described (sleep deprivation, alcohol intake), but also information about basic emotions (see, for example, [34]). The dynamic extraction of such features could be used to evaluate the response of people in virtual environments, as well as adapt the behavior of such environments to the information extracted dynamically. The advantage that such a system would present compared to currently existing approaches is that –by having built a decisional system that is completely modular- we could estimate each class independently of the other classes obtained, which would allow us to have a modular approach to define the interaction
between the human and the machine. Finally
[4] Mohammad G. et al. (2006) Person
we would like to highlight the usefulness of
identification by using AR model
using a single system for user state
For EEG signals. Proc. 9th International
monitoring
Conference on Bioengineering
and
user
personalization.
Assuming that the system is worn anyway
Technology (ICBT2006), Czech Republic,5
for physiological feedback, this same system
pp.
can personalize your experience because it
[5]
knows who you are without any interaction
electroencephalogram as a biometric.
required by the user.
Proc. Canadian Conf. On Electrical and
ENOBIO is an electrophysiology recording
Computer Engineering, pp.
wireless system. The quality of the signal is
1363-1366.
proven to be as good as the signal that can
[6] Poulos M. et al. (1998) Person
be acquires with state of the art wired
identification via the EEG using
equipment. Different applications can be
Computational
built on the ENOBIO system. We presented
Proceedings of the Ninth
a biometric application, a Human Machine
European Signal Processing, EUSIPCO’98,
Interface and a Sleepiness prediction system.
Rhodes, Greece.
REFERENCES
September1998, pp. 2125-2128.
[1] Eischen S., Luctritz J. and Polish J.
[7] Poulos M. et al. (1999) parametric
(1995) Spectral analysis of EEG
person identification from EEG
From Families. Biological Psycholology,
Using computational geometry. Proc. 6th
Vol. 41, pp. 61-68.
International Conference on
[2] Hazarika N., Tsoi A. and Sergejew A.
Electronics,
(1997) nonlinear considerations
(ICECS’99),v. 2, pp. 1005-1008.
in
EEG
signal
Classification.
IEEE
Transactions on signal processing, Vol. 45, pp.829-836. [3] Marcel S., Mill J. (2005) Person authentication using brainwaves (EEG) and maximum a posteriori model adaptation. IDIAP Research Report 05-81, 11 pp.
Para
jape
R.
et
al.(2001)The
geometry
Circuits
and
algorithms.
Systems
