BIOMETRIC AUTHENTICATION USING BRAIN RESPONSES TO

VISUAL STIMULI

Andr

e Z

uquete, Bruno Quintela and Jo

ao Paulo Silva Cunha

IEETA, University of Aveiro, Campus Univ. de Santiago, 3810-193 Aveiro, Portugal

Keywords:

Biometric authentication, Electroencephalograms, Visual evoked potentials.

Abstract:

This paper studies the suitability of brain activity, namely electroencephalogram signals, as raw material for

conducting biometric authentication of individuals. Brain responses were extracted with visual stimulation,

leading to biological brain responses known as Visual Evoked Potentials.

We evaluated a novel method, using only 8 occipital electrodes and the energy of differential EEG signals,

to extract information about the subjects for further use as their biometric features. To classify the features

obtained from each individual, we used a one-class classiﬁer per subject and we tested four types of classiﬁers:

K-Nearest Neighbor, Support Vector Data Description and two other classiﬁers resulting from the combination

of the two ones previously mentioned. After testing these four classiﬁers with features of 70 subjects, the

results showed that visual evoked potentials are suitable for an accurate biometric authentication.

1 INTRODUCTION

This document presents a study on the suitability

of induced electroencephalograms (EEGs) for imple-

menting high-quality, practical biometric authentica-

tion systems. EEGs are impossible to forge because

they reﬂect the inner self of a person, and they are

likely to be different from person to person when per-

forming similar mental activities. However, EEGs

are complex and noisy signals, being affected by dif-

ferent brain activities and other body activities as

well. Thus, we conducted our study with EEG signals

measured in particular scenarios, namely with visual

stimulations leading to very focused brain activities

known as Visual Evoked Potentials (VEP). To the best

of our knowledge, this is the ﬁrst work on EEG-based

authentication using VEPs, though some works exist

on EEG-based identiﬁcation using VEPs and EEG-

based authentication using other brain activity stimuli

(e.g. speciﬁc imagining tasks).

A biometric authentication system has four funda-

mental requirements (Jain et al., 2000):

• Universality: it should be possible to use the sys-

tem with all persons.

• Uniqueness: the system should be able to sepa-

rate different persons with a reasonably low fail-

ure probability.

• Constancy: biometric characteristics of the per-

sons should remain fairly constant for a reason-

able time (months, years).

• Collectability: biometric values should be easy to

obtain, easy to quantify and cause no discomfort.

Considering the ﬁrst requirement (universality), we

believe that only a small percentage of people could

not use the presented EEG authentication procedure.

As we used the perception of simple drawings for trig-

gering EEG signals, people with severe visual impair-

ments or blindness cannot be authenticated.

Considering the second requirement (uniqueness),

we did an empirical observation of the separation of

individuals among a limited population of 70 people

for which we had several EEG samples. Therefore,

we have no proof that it will work on other popula-

tions, but we cannot as well anticipate any reason for

not working.

Considering the third requirement (constancy),

our study is still limited. Our authentication sys-

tem uses images to trigger brain cognitive activities,

which are then measured and classiﬁed. Cognitive

activities may be affected by several factors, such

as stress, fatigue, medication, alcohol ingestion, etc.,

some of them with natural daily variations. How-

ever, the raw EEG data used was collected from a set

of people at a particular measurement session, thus

not reﬂecting daily variations or even variations along

the required time spans (months or years). Neverthe-

103

Zúquete A., Quintela B. and Cunha J. (2010).

BIOMETRIC AUTHENTICATION USING BRAIN RESPONSES TO VISUAL STIMULI.

In Proceedings of the Third International Conference on Bio-inspired Systems and Signal Processing, pages 103-112

DOI: 10.5220/0002750101030112

 SciTePress

less, in our study we concluded that EEGs collected

in a row upon many similar visual stimuli are con-

stant enough for implementing an authentication sys-

tem based upon them, which is a good starting point.

Considering the fourth requirement (collectabil-

ity), the current EEG measurement technology raises

many problems. As EEG signals are very low-power,

EEGs measurement must be done with special care to

increase signal-to-noise ratios.

Finally, electrodes must be placed always in the

same scalp location, an issue usually solved by using

EEG helmets. We anticipate that the actual technolog-

ical problems for EEG measurement may disappear in

a near future, for instance, by using sensors under the

scalp, thus we do not see it as a deﬁnitive barrier to

the use of EEGs for biometric authentication. Nev-

ertheless, in our study we made an effort to facilitate

its deployment, both with the current technology or

with future solutions. More speciﬁcally, we tried to

get the best authentication results with the minimum

possible set of electrodes (or EEG channels), all of

them located in the occipital area of the brain, were

the relevant EEG signals are to be measured.

For this study we did not obtain our own EEG

samples from people. Instead, we used a public data

set

containing EEG signals of 70 individuals, ac-

quired with 64 electrodes after their visual stimula-

tion. After an initial period of evaluation, we found

that 8 channels, located in the occipital area, where

cognitive workload is more relevant, were enough to

achieve acceptable authentication results.

For authenticating people using VEP features we

used personal one-class classiﬁers (OCCs). These

classiﬁers get as input the VEP features of the per-

son being classiﬁed and produce a TRUE/FALSE out-

put value. We used two different OCCs in order to

study which one would produce better authentication

results: K-Nearest Neighbor (KNN) and Support Vec-

tor Data Description (SVDD) (Tax, 2001). After test-

ing both classiﬁers, we also tested a two classiﬁca-

tion architectures combining both KNN and SVDD.

These combined classiﬁers, that we nicknamed OR

and AND , produce outputs after computing a logic

function of the outputs of each individual classiﬁer.

The results, obtained with personal OCCs of the

four types, showed that VEPs can be used as a bio-

metric data for authentication systems, producing re-

sults with high correctness, namely low false positive

and false negative ratios. The results also showed that

correctness is fairly stable for all evaluated subjects,

an important requirement of biometric authentication

systems.

Hosted in http://kdd.ics.uci.edu

2 ELECTROENCEPHALOGRAMS

EEG signals are electric signals gathered in the scalp

of an individual and result from the combination of

signals from two different sources: (i) close-by cere-

bral activity (ii) and non-cerebral origins, such as

eye motion, eye blinking and electrocardiac activity,

called artifacts.

EEG signals are usually decomposed in several

frequency bands. Each band contains signals asso-

ciated with particular brain activities (Basar et al.,

1995). The standard EEG frequency bands are: 0.5–

3.5 Hz (δ), 4–7 Hz (θ), 8–15 Hz (α), 15–30 Hz (β),

30-70 Hz or around 40 Hz (γ). This last one, γ band,

has been related both to gestalt perception (Keil et al.,

1999) and to cognitive functions such as attention,

learning, visual perception and memory.

For each particular brain activity there is one par-

ticular area that produces stronger electrical activity

in one of the previously referred frequency bands;

similarly, artifact manifestations are more relevant in

some parts of the scalp than in others. Consequently,

EEG signals are multi-channel signals, where each

channel corresponds to a speciﬁc scalp electrode lo-

cation. In this study we will consider only the oc-

cipital area of the scalp, which is known to provide

stronger electrical signals in the γ band in response

to visual stimulation and perception of pictures (W.

Lutzenberger and F. Pulvermllera and T. Elbertb and

N. Birbaumer, 1995; Tallon-Baudry et al., 1998; Gru-

ber et al., 2002).

2.1 Visual Evoked Potentials (VEPs)

Visual evoked potentials (VEPs) are brain activity re-

sponses to visual stimuli, which may comprise differ-

ent components, such as color, texture, motion, ob-

jects, readability (text vs. non-text), etc. Each of

these components has impact in the spacial dispersion

of the VEP through the scalp, being observed differ-

ently in each EEG channel and in different frequency

bands. Therefore, for focusing the VEP production

and analysis, the set of visual stimuli must be coher-

ent, i.e., it should stimulate always the same brain ar-

eas.

Several research works (see Section 3) were pre-

viously conducted for achieving individual identiﬁca-

tion using VEPs produced upon the presentation of

images from the Snodgrass and Vanderwart picture

set (Snodgrass and Vanderwart, 1980). This standard

set of 260 pictures was conceived for experiments in-

vestigating differences and similarities in the cogni-

tive processing of pictures. The pictures are black-

and-white line drawings executed according to a set

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

104

of rules that provide consistency of pictorial represen-

tation.

Various studies (Bas¸ar, 1980; Bas¸ar et al., 1987;

Galambos, 1992) showed that VEPs recorded from

the human scalp contain a train of short latency

wavelets in the γ band, precisely time locked to the

stimulus and lasting approximately 100 ms. Further-

more, a more recent study showed that perception of

pictures from the Snodgrass & Vanderwart picture set

induced highly synchronized neural activity, in the

γ band, between posterior electrodes (Gruber et al.,

2002).

3 RELATED WORK

Poulos et al. (Poulos et al., 1998; Poulos et al., 1999)

proposed a method to distinguish an individual from

the rest using EEG signals. They performed a para-

metric spectral analysis of α band EEG signals by ﬁt-

ting to them a linear all-pole autoregressive model.

The coefﬁcients of the ﬁtted model were then used as

features for the identiﬁcation component. In (Poulos

et al., 1998) the identiﬁcation component was built

with computational geometric algorithms; in (Pou-

los et al., 1999) they changed it to a neural network,

namely for a Kohonen’s Linear Vector Quantizer (Ko-

honen, 1989). The cerebral activity was recorded

from subjects at rest, with closed eyes, using only one

channel and during three minutes.

Although the goal of Poulos et al. was person

identiﬁcation using his brain activity, in (Poulos et al.,

1999) they experimented classiﬁcation of a person as

one of a ﬁnite set of known persons. In the tests they

recorded 45 EEG features from each of 4 individu-

als (the X set) and one EEG feature from each of 75

individuals (the non-X set). The neural network was

trained using 20 features from each X member and 30

features from non-X members. Then the system was

used to classify the remaining 25 features of each X

member and the 45 features from the remaining non-

X members. This process was repeated for all the 4 X

members, attaining a correct classiﬁcation score be-

tween 72% and 84%.

Using VEPs and signals in the γ band to per-

form subject identiﬁcation was proposed by Pala-

niappan (Palaniappan, 2004) and followed on his

posterior studies (Palaniappan and Mandic, 2005;

Ravi and Palaniappan, 2005b; Ravi and Palaniappan,

2005a; Ravi and Palaniappan, 2006; Palaniappan and

Mandic, 2007). In all these works is used the same

dataset of VEPs, recorded from 40 individuals and

comprising a 61-channel EEG for 30 VEPs triggered

by pictures chosen from the Snodgrass & Vanderwart

set.

These six subject identiﬁcation studies are all sim-

ilar; they mainly differ in ﬁltering and classiﬁcation

components. First VEP signals are processed to re-

move artifacts and noise caused by other background

brain activities not related with the VEP. Next they

are ﬁltered with a pass-band, digital ﬁlter in to iso-

late signals from the γ band. Then, for each of the 61

channels is computed its spectral power and normal-

ized with the energy values from all the 61 channels;

the 61 resulting values form a feature array. These

features are then used to perform subject identiﬁca-

tion using a classiﬁer with as many output categories

as the number of individuals used to train it; in this

case there were 40 individuals, thus the classiﬁer has

40 different outputs. In the experiments, half of the

features from each individual were used to train the

classiﬁer and the other half for testing the correctness

of its output. The tested correctness of all these ap-

proaches is somewhat similar, ranging from 85.59%

up to 99.62%.

For ﬁltering, in (Palaniappan, 2004; Ravi and

Palaniappan, 2005a; Ravi and Palaniappan, 2006)

a Butterworth ﬁlter was used, while in (Ravi and

Palaniappan, 2005b; Palaniappan and Mandic, 2005;

Palaniappan and Mandic, 2007) an elliptic FIR ﬁlter

was used (in the latter the lower pass-band thresh-

old was lower, 20 Hz). For classifying, in (Ravi

and Palaniappan, 2005b; Palaniappan and Mandic,

2005; Ravi and Palaniappan, 2006; Palaniappan and

Mandic, 2007) was used an Elman back-propagation

neural network (Elman, 1990), in (Palaniappan, 2004)

a back-propagation multi-layer perceptron, in (Ravi

and Palaniappan, 2005a; Ravi and Palaniappan, 2006)

a simpliﬁed fuzzy ARTMAP (Kasuba, 1993) and

in (Ravi and Palaniappan, 2005a) a KNN.

Some attempts were made to reduce the num-

ber of channels used in these VEP-based approaches.

In (Palaniappan and Mandic, 2007) Davies-Bouldin

Indexes (Davies and Bouldin, 1979) were used to or-

der the channels according to their relevance. Cor-

rect identiﬁcation results using the most relevant DBI-

oriented channels gave 13.63% with 1 channel, about

50% with 6 channels and 99.0% with 50 channels.

There are already several studies on authentica-

tion with EEG signals, but all them use different ap-

proaches (Marcel and de R. Milln, 2007; Sun, 2008;

Palaniappan, 2008). The table below resumes some

of their differences.

BIOMETRIC AUTHENTICATION USING BRAIN RESPONSES TO VISUAL STIMULI

105

(Marcel and de

R. Milln, 2007) (Sun, 2008) (Palaniappan, 2008)

EEG channels α, β α, β, γ α, β, γ

Electrodes 8 15 8

Feature array 96 8 11

elements (12 freqs./channel) (CSP reduct.) (PCA reduct.)

Tested subjects 9 9 5

In (Marcel and de R. Milln, 2007), authors collected

EEGs from subjects performing 3 mental activities:

imagination of movements with the right or left hand

and imagination of words beginning with the random

letter. Features’ classiﬁcation uses Gaussian Mixture

Models and Maximum A Posteriori model adaptation.

The correctness results were satisfactory but not very

conclusive, because the number of evaluated subjects

was too small (we used 70). A drawback of the clas-

siﬁcation approach is that it relies on a generic EEG

model, which may not exist or requires training with

EEGs from many people.

In (Sun, 2008), authors used 15 signals from the

same dataset used in (Marcel and de R. Milln, 2007),

raw feature reduction with common spatial patterns

(CSP) and using multi-task learning to evaluate the

advantage regarding single-task learning.

In (Palaniappan, 2008), authors collected EEGs

from subjects performing 5 imagined activities: noth-

ing in particular (baseline activity), mathematical

multiplication, geometric ﬁgure rotation, letter com-

position and visual counting. Feature arrays are ini-

tially composed by 18 channel spectral powers, 27

inter-hemispheric channel spectral power differences

and 18 entropy values (yielding the non-linearity of

channel signals). Features are then reduced to 11 el-

ements using Principal Component Analysis (PCA).

Features’ classiﬁcation uses a two-stage authentica-

tion process using maximum and minimum threshold

values stored in personal proﬁles. Like in the previous

article, correctness results were satisfactory but even

less conclusive, due to the extremelly small number

of evaluated subjects (only 5).

All these three works used imagined activities to

focus EEG-signals; we used VEPs instead. The ad-

vantages of VEPs is that they do not require any effort

from the subjects being authenticated, as VEPs occur

without any sort of human control. Furthermore, we

did an evaluation with a larger population (roughly an

order of magnitude more) than all these works, there-

fore our results yield a more trustworthy evaluation of

the universality and uniqueness requirements. Finally,

we did not use more electrodes than any of them, thus

we do not require a more complex EEG acquisition

setup.

Finally, some studies have been done with multi-

modal biometrics involving EEG signals (Riera et al.,

2008). The main advantage of this approach regard-

ing the simple EEG authentication is that one can re-

duce the number of electrodes (only 2 were used).

4 AN AUTHENTICATION

SYSTEM USING VEP

As previously stated, our goal was to build an authen-

tication system based only in occipital VEP EEG sig-

nals gathered by a small number of electrodes. Note

that, authentication is different from identiﬁcation: an

identiﬁcation system gives the identity of the subject

being evaluated, while an authentication system gives

a yes/no answer whether or not the subject being eval-

uated is who he claims to be.

The VEP-based identiﬁcation systems developed

by Palaniappan et al. are also not directly usable as

authentication systems. These systems were designed

for identifying members of a set X of N subjects, hav-

ing N possible output classiﬁcations. When these sys-

tems are used by other non-X subjects, these will be

identiﬁed as someone belonging to X. Thus, a non-X

person being authenticated only has to guess the erro-

neous identity the system gives to him, in order to get

an authentication match.

Therefore, a new architecture is required to use

EEG patterns for authenticating individuals. We pro-

pose a new one where we merge part of the contri-

butions of the previously referred systems with some

new ideas introduced by us.

4.1 Personal Classiﬁers

Our key design principle is to analyse EEG patterns in

the γ band, namely VEPs in occipital area of the brain,

with one classiﬁer per individual, and not a classiﬁer

for all individuals. Furthermore, we used an OCC for

each personal classiﬁer, which is the correct type of

classiﬁer for an authentication scenario. Thus, when

a subject claims to be X, we use Xs OCC classiﬁer to

evaluate the correctness of the claim.

OCCs may have many inputs to handle the fea-

tures obtained from subjects, but always two possi-

ble output responses: TRUE or FALSE. Each per-

sonal OCC is trained only with inputs provided by

its owner. When the individual being evaluated is the

owner of the classiﬁer, the output should be TRUE;

otherwise, the output should be FALSE. Other outputs

are errors, either false negatives or false positives, re-

spectively.

As previously referred, the goal for this new archi-

tecture was to use a reduced number of EEG channels.

In the limit we would like to use only one channel,

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

106

just like in the work of Poulos et al.. However, unlike

the approach described in (Palaniappan and Mandic,

2007), we have not tried to detect the “best” channels

(the ones with less correlation) from a set of measured

features. Instead, we chose speciﬁc channel locations

in the occipital area of the scalp and we ran authenti-

cation tests with them to ﬁnd out the set of channels

providing the highest authentication quality.

4.2 Authentication Process

Our authentication process is formed by three main

components: (i) EEG signal acquisition, (ii) feature

array extraction and (iii) feature array classiﬁcation.

First VEP EEG signals are acquired from electrodes

placed in the subject’s scalp. Then the feature array

extractor processes raw EEG samples from C chan-

nels in order to extract a biometric measure of the

subject: a feature array with C

energy values. Finally,

this feature is processed by the OCC of the subject be-

ing authenticated, either to train the OCC or to get a

TRUE or FALSE authentication outcome.

4.3 Description of the Data Set

As previously explained, we did not collect EEG sig-

nals for this study. Instead, we used a public data set

registered for conducting other EEG studies, namely

the genetic predisposition of people to alcoholism.

Thus, it was not in any way specially gathered for au-

thenticating people.

The data set is composed by EEG signals

recorded from 70 individuals, both alcoholic and non-

alcoholic, while exposed to short latency (300ms) vi-

sual stimuli. Each individual completed a total num-

ber of 45 trials corresponding to the visualization of

45 pictures from the Snodgrass & Vanderwart picture

set. EEG signals were acquired by 64 electrodes (61

actives + 3 reference), placed in the individuals’ scalp,

hardware ﬁltered with a 0.1–50Hz passband and mea-

sured at a sampling rate of 256 samples per second.

For building our authentication system we considered

only 8 occipital channels from the 64 available in the

dataset — channels PO3, PO4, POZ, PO7, PO8, O1,

O2 and OZ.

Individuals were asked to recognise the pictures

as soon as they were presented in a CRT screen, lo-

cated 1 meter away from individuals’ eyes. Each

picture was presented only for 300 ms, separated by

blank screen intervals of 5.1 seconds. After each pic-

ture presentation, only 1 second of EEG signal was

recorded, corresponding to the VEP occurrence inter-

val.

4.4 Feature Array Extraction

VEP signals, which are raw EEG signals with 1 sec-

ond measured after the presentation of the stimuli im-

ages, are the source data for for the biometric authen-

tication process of each individual. The feature ex-

traction procedure from these signals is detailed be-

low.

Detection of Artifacts. First, VEP signals contain-

ing artifacts are discarded. We considered only arti-

facts produced by eye blinking, which are the most

common and intrusive ones. Detection of eye blink-

ing artifacts is achieved with an amplitude thresh-

old method: VEP signals with magnitude above 50

µV are assumed to be contaminated (Sivakumar and

Ravindran, 2006) so they are discarded .

EEG γ-band (30-50Hz) Frequency Filtering. The

resulting artifact-free VEP signals are ﬁltered with a

30-50 Hz pass-band, using a 10

order Butterworth

digital ﬁlter. The non-linearity of this ﬁlter was can-

celled by using forward and reverse ﬁltering. The re-

sulting signal has zero phase distortion and an am-

plitude multiplied by the square of the amplitude re-

sponse of the ﬁlter. After ﬁltering, the 20 ﬁrst and 20

last output samples are discarded, because they do not

represent a properly ﬁltered signal.

Signal Composition. For computing feature arrays

we use C original EEG signals plus differential sig-

nals resulting from the subtraction of pairs of the C

EEG signals. Thus, features include C

= C +





sig-

nals, which in our case, for C = 8, means that C

= 36.

By computing differential signals from the sub-

traction of pairs of EEG signals we expect to provide

to classiﬁers information about the phase of the EEG

signals and not just information about their ampli-

tudes (energies). Phase shifts between subtracted si-

nusoidal signals with equal frequency and amplitude

produce non-null signals with an energy that is a func-

tion of the phase shift. Therefore, we included the

energy of differential signals in the features because

it could denote phase shifts between EEG channels,

thus more information about the subjects.

These differential signals are somewhat similar to

the ones used in (Palaniappan, 2008) but with two

main differences: (i) we compute the energy of dif-

ferential signals, while they compute differences be-

tween powers of different signals and (ii) we produce

a differential signal from all pairs of signals, while

they only compute differential powers between signal

on different hemispheres. Thus, we are able to evalu-

ate phase shifts on differential signals and we produce

BIOMETRIC AUTHENTICATION USING BRAIN RESPONSES TO VISUAL STIMULI

107

more information that may help to differentiate sub-

jects.

Energy Calculation and Normalization. The en-

ergy of original and differential signals is computed

with the Parseval’s spectral power ratio theorem:

E(s) =

∑

n=1

where s

is the n-th sample of signal s and N is the

total number of samples in the signal. In our case

N = 216, because we discard 40 samples of the 256

measured in 1 second of VEP after the γ ﬁltering

stage.

Finally, feature values are computed by normaliz-

ing the energy feature array. For this normalization

we divide all array values by the maximum among

them. This way, we get features with elements in the

[0, 1] interval.



1· · ·C



E [1· · ·C

]

max (E [1· · ·C

])

4.5 The Feature Classiﬁer

The feature classiﬁer is formed by independent,

personal classiﬁers; so, for authenticating someone

claiming to be X, we use the personal classiﬁer of X,

or the classiﬁer owned by X. Each personal classiﬁer

is formed by an OCC, providing two different outputs

(or classiﬁcations): TRUE and FALSE.

One-Class Classiﬁcation is a type of classiﬁcation

where we deal with a two-class classiﬁcation problem

(target and outlier) but we only need to provide infor-

mation to the classiﬁer about the target class. During

an OCC train, the boundary between the target class

and all other possible outlier classes is estimated from

the target class data only. In our authentication goal,

the target class is the classiﬁer owner while the outlier

class represents all other individuals.

In our study we used two types of OCCs in order

to check which one would produce better authentica-

tion results: KNN with k=1 and SVDD with a Ra-

dial Basis Function kernel (Tax, 2001). We also tested

two other OCC architectures, combining the outputs

the KNN with SVDD. The goal of the combinations

was to evaluate if there was any advantage in combin-

ing them in order to complement their individual cor-

rectness. The OR combination uses arithmetical av-

erages, and the AND geometrical averages. For sim-

plicity, we will call the ﬁrst a OR KNN-SVDD and the

second a AND KNN-SVDD.

We also found out that each classiﬁer should be

trained with single features from its owner, but should

be used for authentication with average tests F fea-

tures, obtained from the visualization of F images of

the subject. A possible explanation of this fact is the

following. Perception activities performed by indi-

vidual’s brain are not exactly the same for all visual

stimuli, resulting in different VEP features. By train-

ing the classiﬁer with as different as possible VEP

features from its owner, we improve its ability to rec-

ognize them in the future, disregarding possible noise

occurrences. On the other hand, by averaging VEP

features during authentication processes, we reduce

the probability of presenting to the classiﬁer features

from its owner too different from the ones it was

trained with.

During the training of each classiﬁer, we have to

provide a rejection fraction threshold that will be used

to establish acceptance or rejection ratios. Low re-

jection fraction values lead classiﬁers to produce low

false negative and high false positive ratios, while

high rejection fraction values lead classiﬁers to pro-

duce high false negative and low false positive ratios.

The choice of the best rejection fraction threshold im-

plies a balance between security (low false positive

ratio) and comfort for the individuals engaged in a

correct authentication process (low false negative ra-

tio).

5 EXPERIMENTATION

The number of feature arrays used per subject was

variable, both from start (in the data set) and further-

more after eye blink detection. Therefore, we decided

to test classiﬁers with ﬁxed numbers of features

and train classiﬁers with the maximum possible

number of features until a given maximum. This

is a conservative approach, since some classiﬁers

may not have enough features to be properly trained.

Nevertheless, we did not observe abnormal errors in

such classiﬁers.

Thus, to train each personal classiﬁer we used no

more than 30 features of its owner. For testing each

personal classiﬁer, we used 15 features of its owner

and 15 features from each of the other 69 subjects,

which makes a total of 1050 test features. Note

that each classiﬁer had never “seen” the test feature

before. The test features of each individual were used

alone or averaged in pairs or trios. The number of

features evaluated per classiﬁer is presented in the

table below.

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

108

Composition Features evaluated per classiﬁer

of features From the owner Total

Single features 15 70 × 15 = 1050

Pairs of features





= 105 70 ×





= 7350

Trios of features





= 455 70 ×





= 31850

We run authentication tests with all the proposed

four classiﬁers, in order to verify which one of them

is more suitable for our authentication system. In the

tests we tried also to assess the impact of two conﬁg-

uration parameters for the overall correctness of the

authentication system: OCC rejection fraction thresh-

old and classiﬁcation of multiple, averaged features.

5.1 Overall Evaluation Results

The overall biometric authentication results of the 70

classiﬁers for three rejection fraction thresholds and

different combinations of features are summarized in

Table 1. The values presented are the mean and stan-

dard deviation obtained from 10 independent tests

with the 70 OCCs, each one of them using different

features from the owner (to train and test his classi-

ﬁer) and from outliers (to test). The graphics of Fig. 1

show the values of these 10 independent tests per per-

sonal classiﬁer, but only for the combined OR KNN-

SVDD classiﬁer, using single and trios of features.

Table 1: Mean and standard deviation (inside parenthesis)

of correctness results for the four OCC classiﬁers obtained

in 10 independent classiﬁcation tests. Columns labeled 1,

2 and 3 represent tests using singular features and average

combinations of pairs and trios of features, respectively.

features = 1 features = 3

rejection fraction = 0.2rejection fraction = 0.5rejection fraction = 0.7

Figure 1: Average individual classiﬁcation results of the combined OR KNN-SVDD classiﬁer, obtained after 10 independent

tests. The upper (green) curve in each graphic shows the average correct owner classiﬁcations per classiﬁer, while the lower

(red) curve shows the average false positives per classiﬁer. The vertical line under each average value shows the maximum

and minimum values observed in the 10 tests.

Composition Features evaluated per classiﬁer

of features From the owner Total

Single features 15 70 × 15 = 1050

Pairs of features





= 105 70 ×





= 7350

Trios of features





= 455 70 ×





= 31850

We run authentication tests with all the proposed

four classiﬁers, in order to verify which one of them

is more suitable for our authentication system. In the

tests we tried also to assess the impact of two conﬁg-

uration parameters for the overall correctness of the

authentication system: OCC rejection fraction thresh-

old and classiﬁcation of multiple, averaged features.

5.1 Overall Evaluation Results

The overall biometric authentication results of the 70

classiﬁers for three rejection fraction thresholds and

different combinations of features are summarized in

Table 1. The values presented are the mean and stan-

dard deviation obtained from 10 independent tests

with the 70 OCCs, each one of them using different

features from the owner (to train and test his classi-

ﬁer) and from outliers (to test). The graphics of Fig. 1

show the values of these 10 independent tests per per-

sonal classiﬁer, but only for the combined OR KNN-

SVDD classiﬁer, using single and trios of features.

Rejection Owners Correctness(%) Outliers acceptance (%)

fraction 1 2 3 1 2 3

KNN

0.2 78.7 (13.1) 90.6 ( 7.5) 95.1 ( 5.3) 5.2 (1.3) 5.6 (1.6) 6.4 (1.9)

0.5 50.1 (15.9) 65.3 ( 9.3) 74.1 ( 8.6) 1.9 (0.7) 2.2 (0.9) 2.3 (1.3)

0.7 31.3 (18.5) 46.1 (11.3) 66.3 ( 9.8) 0.9 (0.4) 1.1 (0.5) 1.3 (0.7)

SVDD

0.2 76.1 (12.8) 95.2 ( 4.9) 98.5 ( 3.5) 5.7 (1.8) 8.5 (2.1) 10.1 (3.2)

0.5 58.5 (17.4) 88.3 ( 8.7) 93.7 ( 5.7) 2.8 (1.6) 4.4 (1.8) 5.1 (2.7)

0.7 44.2 (20.5) 77.7 (12.3) 85.3 ( 9.8) 1.7 (1.6) 2.6 (1.7) 3.6 (1.9)

AND

0.2 83.3 (12.1) 96.4 ( 6.1) 99.0 ( 3.0) 4.5 (5.6) 6.0 (7.4) 6.8 (8.2)

0.5 60.4 (16.8) 85.6 (12.4) 92.8 (11.3) 1.3 (2.2) 1.8 (2.9) 1.9 (3.0)

0.7 37.8 (18.7) 68.4 (18.1) 79.4 (19.2) 0.4 (0.9) 0.6 (1.1) 0.6 (1.4)

0.2 83.8 (11.0) 96.5 ( 6.0) 99.1 ( 3.5) 4.7 (5.7) 6.2 (7.6) 6.8 (8.4)

0.5 59.7 (17.2) 85.7 (13.1) 92.8 (11.5) 1.2 (1.9) 1.7 (2.8) 2.0 (3.2)

0.7 38.7 (18.2) 69.8 (18.1) 80.5 (18.1) 0.4 (0.8) 0.6 (1.2) 0.6 (1.3)

Table 1: Mean and standard deviation (inside parenthesis)

of correctness results for the four OCC classiﬁers obtained

in 10 independent classiﬁcation tests. Columns labeled 1,

2 and 3 represent tests using singular features and average

combinations of pairs and trios of features, respectively.

The results show that the rejection fraction thresh-

old used while training classiﬁers had the expected

impact on authentication results: for low rejection

values the classiﬁer provides a correct classiﬁcation of

The results show that the rejection fraction thresh-

old used while training classiﬁers had the expected

impact on authentication results: for low rejection

values the classiﬁer provides a correct classiﬁcation of

its owner (low false negative ratio) but can be mislead

by many other individuals (high false positive ratio),

while for higher rejection values the correct classiﬁ-

cation of owners decreases but the same happens to

the wrong acceptance of other individuals.

Comparing KNN and SVDD, we can conclude

that both have advantages and disadvantages: KNN

gives lower outliers acceptance ratios (false posi-

tives), while SVDD gives higher owners’ correctness

ratios (true positives).

The combined OR and AND KNN-SVDD clas-

siﬁers also have advantages and disadvantages when

compared with the isolated OCCs. In general, they

decrease the false positive ratio and most times (67%)

they increase the owners’ correctness ratio. However,

they have a noticeable tendency to increase the stan-

dard deviation of the results, being thus less assertive

than the isolated OCCs. Comparing the two combined

KNN-SVDD classiﬁers, the results show that they are

quite similar, but the OR combination is slightly bet-

ter.

Finally, these results clearly show that the qual-

ity of the authentication increases when we use com-

binations of features instead of singular features. In

absolute value, the owners’ correctness gain is much

higher that the loss in the false positive ratio.

5.2 Evaluation of Individual Classiﬁers

From the graphics of Fig. 1 we can conclude that av-

erage classiﬁcation results are fairly stable for all the

considered subjects. Therefore, with these tests we

have reasons to believe that a biometric authentication

system using EEGs may be suitable for a large major-

ity of the population. Note that the evaluated sub-

jects already include a group of people (alcoholics)

that may have visual cognition problems and that was

not noticeable in the authentication results.

A good indicator about an OCC performance is

the plot of its receiver operating characteristic (ROC)

curve. A ROC curve is calculated with several tests of

the classiﬁer with different rejection fraction thresh-

olds applied to target objects and shows the percent-

age of true positives in order to the percentage of the

false positives during each test. Thus, ROC curves

are useful to assert the effect of the rejection fraction

threshold in tuning the correctness of the OCC.

The OCC with the best performance is the one that

simultaneously maximizes true positive ratios and

minimizes false negative ratios. This performance

can be measure by calculating the area under curve

(AUC). This way, the OCC with the higher AUC is

assumed to be the OCC with best performance.

Figure 2 shows the ROCs of the 70 individual clas-

siﬁers and their average AUC values for each of the

four OCC types considered; these ROC curves where

obtained with feature trios. These results clearly show

that for evaluating trios of features the best OCC is the

combined OR KNN-SVDD, while the worse is KNN.

BIOMETRIC AUTHENTICATION USING BRAIN RESPONSES TO VISUAL STIMULI

109

features = 1 features = 3

rejection fraction = 0.2rejection fraction = 0.5rejection fraction = 0.7

Figure 1: Average individual classiﬁcation results of the combined OR KNN-SVDD classiﬁer, obtained after 10 independent

tests. The upper (green) curve in each graphic shows the average correct owner classiﬁcations per classiﬁer, while the lower

(red) curve shows the average false positives per classiﬁer. The vertical line under each average value shows the maximum

and minimum values observed in the 10 tests.

6 CONCLUSIONS

We presented in this paper a novel method for au-

thenticating individuals using their brain activity. The

EEG signals used were VEPs, i.e., brain responses

associated to visual stimuli. In the described system

we used EEG signals acquired with only 8 electrodes

placed on occipital area of the brain, which is associ-

ated to visual and cognitive perception.

The authentication system presents several im-

provements over other previous works in the area of

subject identiﬁcation using VEPs. First, We used a

reduced number of electrodes (8) and we placed them

in the scalp area where EEG signals have more cor-

relation with the stimuli. Second, we used the differ-

ences between pairs of the 8 EEG signals to create

other signals (differential signals) that provide extra

information to classiﬁers.

Third, feature arrays with the energy of original

and differential EEG signals are classiﬁed using per-

sonal classiﬁers.

Forth, we used OCC personal classiﬁers instead

of other types of classiﬁers (e.g. neural networks) or

generic classiﬁers.

Fifth, OCCs are trained with single owner features

but provide better results when tested with average of

features instead of single ones.

Regarding other related systems, we used VEPs,

which are effortless for subjects, while others used

more complex and annoying brain stimuli, such as ac-

tivity imagining, we obtained satisfactory results with

a population one order of magnitude larger then the

other proposals (70 vs. 5 or 9 subjects), and we did

not use more electrodes (only 8). Therefore, our sys-

tem has clear advantages regarding the collectability

requirements.

Average results obtained with authentication tests

with 70 individuals, using a public VEP data set,

showed that authentication with EEGs is possible and

may be used in future applications. The ratios of

owner’s correctness and false positives are fairly sta-

ble for the tested population, which is a positive indi-

cation for the universality and uniqueness of the pro-

cess.

A fundamental requirement of biometric authenti-

cation was not evaluated in this document: constancy.

In fact, we assumed that the public data set, per sub-

ject, was collected in a very short time, thus it is not

possible to take any conclusions about the constancy

of VEPs in other scenarios. Since several factors may

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

110

KNN, avg. AUC = 0.9817 SVDD, avg. AUC = 0.9913

combined AND KNN-SVDD, avg. AUC = 0.9962 combined OR KNN-SVDD, avg. AUC = 0.9965

Figure 2: ROC of the 70 individual classiﬁers and their average AUC for each type of OCC considered and using feature trios.

affect VEPs signiﬁcantly, such as stress and fatigue,

the constancy of VEP-based biometric authentication

must be addressed by future research.

REFERENCES

Bas¸ar, E. (1980). Relation between EEG and Brain Evoked

Potentials. In EEG-Brain Dynamics. Elsevier, North-

Holland biomedical Press, Amsterdam.

Bas¸ar, E., Rosen, B., Bas¸ar-Eroglu, C., and Greitschus, F.

(1987). The associations between 40 Hz-EEG the

Middle Latency Response of the Auditory and Evoked

Potential. The Int. J. of Neuroscience, 33(1-2):103–

117.

Basar, E., Basar-Eroglu, C., Demiralp, T., and Schrmann,

M. (1995). Time and Frequency Analysis of the

Brain’s Distributed Gamma-Band System. IEEE Eng.

in Medicine and Biology, 14:400–410.

Davies, D. and Bouldin, D. (1979). A Cluster Separation

Measure. IEEE Trans. on Pattern Analysis and Mach.

Intelligence, 1(2):224–227.

Elman, J. L. (1990). Finding structure in time. Cognitive

Science, 14(2):179–211.

Galambos, R. (1992). A comparison of certain gamma band

(40-Hz) brain rhythms in cat and man. In Induced

Rhythms in the Brain, pages 201–216. Birkhuser,

Boston, MA, USA.

Gruber, T., Mller, M. M., and Keil, A. (2002). Modulation

of Induced Gamma Band Responses in a Perceptual

Learning Task in the Human EEG. J. of Cognitive

Neuroscience, 14(5):732–744.

Jain, A., Hong, L., and Pankanti, S. (2000). Biometric Iden-

tiﬁcation. Comm. of the ACM, 43(2):90–98.

Kasuba, T. (1993). Simpliﬁed Fuzzy ARTMAP. AI Expert,

pages 18–25.

Keil, A., Mller, M. M., Ray, W. J., Gruber, T., and Elbert, T.

(1999). Human Gamma Band Activity and Perception

of a Gestalt. The J. of Neuroscience, 19(16):7152–

7161.

Kohonen, T. (1989). Self-organization and associative

memory: 3rd edition. Springer-Verlag New York, Inc.

Marcel, S. and de R. Milln, J. (2007). Person Authentication

Using Brainwaves (EEG) and Maximum A Posteriori

Model Adaptation. IEEE Trans. on Pattern Analysis

and Mach. Intelligence, 29(4):743–748.

Palaniappan, R. (2004). Method of identifying individuals

using VEP signals and neural network. IEE Proc. -

Science Measurement and Technology, 151(1):16–20.

Palaniappan, R. (2008). Two-stage biometric authentication

method using thought activity brain waves. Int. J. of

Neural Systems, 18(1).

Palaniappan, R. and Mandic, D. P. (2005). Energy of Brain

Potentials Evoked During Visual Stimulus: A New

Biometric? In Proc. of the Int. Conf. on Artiﬁcial Neu-

ral Networks: Formal Models and Their Applications

(ICANN 2005), LNCS 3697, pages 735–740. Springer.

Palaniappan, R. and Mandic, D. P. (2007). EEG Based

Biometric Framework for Automatic Identity Veriﬁ-

cation. J. of VLSI Signal Processing, 49:243–250.

BIOMETRIC AUTHENTICATION USING BRAIN RESPONSES TO VISUAL STIMULI

111

Poulos, M., Rangousi, M., and Kafetzopoulos, E. (1998).

Person identiﬁcation via the EEG using computational

geometry algorithms. In Proc. of the 9th European

Signal Processing (EUSIPCO’98), pages 2125–2128,

Rhodes, Greece.

Poulos, M., Rangoussi, M., Chrissikopoulos, V., and Evan-

gelou, A. (1999). Person Identiﬁcation Based on Para-

metric Processing of the EEG. In Proc. of the 6th

IEEE Int. Conf. on Electronics, Circuits and Systems

(ICECS), pages 283–286.

Ravi, K. V. R. and Palaniappan, R. (2005a). Leave-one-out

Authentication of Persons Using 40 Hz EEG Oscilla-

tions. In Proc. of the Int. Conf. on ”Computer as a

tool” (EUROCON 2005), Belgrade, Serbia & Mon-

tenegro.

Ravi, K. V. R. and Palaniappan, R. (2005b). Recognising

Individuals Using Their Brain Patterns. In Proc. of the

3th Int. Conf. on Information Tech. and Applications

(ICITA’05). IEEE Computer Society.

Ravi, K. V. R. and Palaniappan, R. (2006). Neural network

classiﬁcation of late gamma band electroencephalo-

gram features. Soft Computing, 10:163–169.

Riera, A., Soria-Frisch, A., Caparrini, M., Grau, C., and

Rufﬁni, G. (2008). Unobtrusive Biometric System

Based on Electroencephalogram Analysis. EURASIP

J. on Advances in Signal Processing, 2008.

Sivakumar, R. and Ravindran, G. (2006). Identiﬁcation

of Intermediate Latencies in Transient Visual Evoked

Potentials. Academic Open Internet J., 17.

Snodgrass, J. G. and Vanderwart, M. (1980). A Standard-

ized Set of 260 Pictures: Norms for Name Agreement,

Image Agreement, Familarity and Visual Complexity.

J. of Experimental Psychology: Human Learning and

Memory, 6(2):174–215.

Sun, S. (2008). Multitask learning for eeg-based biomet-

rics. In 19th Int. Conf. on Pattern Recognition (ICPR

2008), Tampa, Florida, USA.

Tallon-Baudry, C., Bertrand, O., Peronnet, F., and Pernier,

J. (1998). Induced γ-Band Activity during the Delay

of a Visual Short-Term Memory Task in Humans. The

J. of Neuroscience, 18(11):4244–4254.

Tax, D. M. J. (2001). One-class classiﬁcation; Concept-

learning in the absence of counter-examples. PhD

thesis, Delft University of Technology, Delft, Nether-

lands.

W. Lutzenberger and F. Pulvermllera and T. Elbertb and

N. Birbaumer (1995). Visual stimulation alters local

40-Hz responses in humans: an EEG-study. Neuro-

science Letters, 183(1-2):39–42.

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

112