On the Repeatability of EEG Features in a Biometric Recognition

Framework using a Resting State Protocol

Daria La Rocca

, Patrizio Campisi

and Gaetano Scarano

Department of Applied Electronics, Universit`a degli Studi “Roma Tre”, via Della Vasca Navale 84, I-00146 Roma, Italy

DIET, Sapienza Universit`a di Roma, via Eudossiana 18, I-00184 Roma, Italy

Keywords:

EEG, Biometrics, Repeatability, Resting.

Abstract:

In this paper the feasibility of the electroencephalogram (EEG) as biometric identiﬁer is investigated with

focus on the repeatability of the EEG features employed in the proposed framework. The use of EEG within

the biometric framework has already been introduced in the recent past although it has not been extensively

analyzed. In this contribution we infer about the invariance over time of the employed EEG features, which

is one of the most relevant properties a biometric identiﬁer should possess in order to be employed in real life

applications. For the purpose of this study we rely on the “resting state” protocol. The employed database

is composed by healthy subjects whose EEG signals have been acquired in two different sessions. Different

electrodes conﬁgurations pertinent to the employed protocol have been considered. Autoregressive statistical

modeling using reﬂection coefﬁcients has been adopted and a linear classiﬁer has been tested. The obtained

results show that a high degree of repeatability has been achieved over the considered interval.

1 EEG AS BIOMETRICS

From the beginning of the 20th century, EEG analysis

has been mainly used in medicine to study brain dis-

eases like Alzheimer, epilepsy, Parkinson and many

others. Speciﬁcally, EEG signals, acquired by means

of scalp electrodes, sense the electrical activity re-

lated to the ﬁring of speciﬁc collections of neu-

rons responding to a variety of cognitive tasks such

as audio or visual stimuli, real or imagined body

movements, imagined speech, etc. The most rele-

vant cerebral activities falls in the range of [0.5,40]

Hz. In details ﬁve wave categories have been identi-

ﬁed, each associated to a speciﬁc bandwidth and to

speciﬁc cognitive tasks: delta waves (δ) [0.5,4]Hz

which are primarily associated with deep sleep, loss

of body awareness, and may be present in the wak-

ing state; theta waves (θ) [4,8]Hz which are asso-

ciated with deep meditation and creative inspiration;

alpha waves (α) [8,13]Hz which indicate either a re-

laxed awareness without any attention or concentra-

tion; beta waves (β) [13,30]Hz usually associated to

active thinking; gamma waves (γ) [30,40]Hz usually

used to locate right and left movements.

In the last decades, the brain activity, registered by

means of EEG, has been heavily employed in brain

computer interfaces (BCI) (Dornhege et al., 2007)

and more recently in brain machine interface (BMI)

(Carmena, 2012) for prosthetic devices. In the last

few years EEG signals have also been proposed to be

used in biometric based recognition systems (Campisi

et al., 2012).

EEG signals present some peculiarities, which are

not shared by the most commonly used biometrics,

like face, iris, and ﬁngerprints. Speciﬁcally, brain sig-

nals generated on the cortex are not exposed like face,

iris, and ﬁngerprints, therefore they are more privacy

compliant than other biometrics since they are “se-

cret” by their nature, being impossible to capture them

at a distance. This property makes EEG biometrics

also robust against the spooﬁng attack at the sensor

since an attacker would not be able to collect and feed

the EEG signals. Moreover, being brain signals the

result of cognitive processes, they cannot be synthet-

ically generated and fed to a sensor, which also ad-

dresses the problem of liveness detection. Also, the

level of universality of brain signals is very high. In

fact people with some physical disabilities, prevent-

ing the use of biometrics like ﬁngerprint or iris, would

be able to get access to the required service using

EEG biometrics. However, the level of understanding

of the physiological mechanisms behind the genera-

tion of electric currents in the brain, not yet fully got,

makes EEG a biometrics at its embryonic stage. Nev-

419

La Rocca D., Campisi P. and Scarano G. (2013).

On the Repeatability of EEG Features in a Biometric Recognition Framework using a Resting State Protocol.

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, pages 419-428

DOI: 10.5220/0004339104190428

 SciTePress

Table 1: State of the art on EEG biometrics using a resting state protocol.

Paper Protocol # Subj # Ch. Features

(Poulos et al., 1999) CE 4 1 AR (6th)

(Campisi et al., 2011) CE 48 3 AR (6th)

(Paranjape et al., 2001) CE, OE 40 1 AR (21th)

(Palaniappan and Patnaik, 2007) CE 5 6 AR (6th)

(Zhao et al., 2010) CE 10 1 AR (6th) +

PSD

(Abdullah et al., 2010) CE, OE 10 4 AR (21th)

(Mohammadi et al., 2006) OE 10 2-3 AR (11th)

(Nakanishi et al., 2009) CE 23 1 FFT

ertheless, some preliminary, but promising, results

have already been obtained in the recent literature, see

for example (Poulos et al., 1999), (Brigham and Ku-

mar, 2010), (Riera et al., ), (Marcel and del R. Millan,

2007), (Campisi et al., 2011) where a review on the

state of the art of EEG biometrics is also given, and

(La Rocca et al., 2012). Due to the early stage of re-

search dealing with EEG as biometrics, currently, the

deployment of convenient and accurate EEG based

applications in real world are limited with respect to

well established biometrics like ﬁngerprints, iris, and

face. However the brain electrical activity has already

shown some potentials to allow automatic user recog-

nition. Answers to practical and theoretical ques-

tions addressed for the development of a usable sys-

tem can be found in (Su et al., 2010) where promis-

ing results are obtained from the implementation of

a portable EEG biometric framework for applications

in real world scenarios. Improvements in EEG signal

acquisition and technological advances in the use of

wireless and dry sensors, easy to wear and robust with

respect to noise (Debener et al., 2012) could represent

the cue for outlining guidelines for practical systems

implementation.

In Table 1, an extensive although not exhaustive

list of research studies which have already been pub-

lished using a resting state acquisition protocol, either

closed eyes (CE) or open eyes (OE), is provided. It is

evident that the database dimension is quite limited in

almost all of these contributions. This is also due to

the lack of a public EEG database suitably collected

for the biometric recognition purpose, where acquisi-

tions and protocols would be designed according to

the speciﬁc requirements. In fact most of the works in

this ﬁeld test the implemented techniques on datasets

recorded in BCI contexts. Moreover the issue of the

repeatability of EEG biometrics in different acquisi-

tion sessions has never been systematically addressed

in any of the aforementioned contributions and it has

never received the required attention from the scien-

tiﬁc community. Nevertheless, its understanding is

propaedeutic towards the deployment of EEG biomet-

rics in real life. Although in some referred works dif-

ferent acquisition sessions have been provided, they

were considered to assort a single dataset where ran-

domly selected EEG segments were used for training

or testing a classiﬁcation algorithm for the recogni-

tion purpose. On the other hand, in (Brigham and

Kumar, 2010) the session-to-session variability was

tested on a dataset of 6 subjects performing imagined

speech. The entire set of 128 channels was used to

extract features, and results show a decreasing perfor-

mance when considering sessions temporally apart,

which led to assess that the imagined speech EEG

does not show to have a reliable degree of repeata-

bility. Therefore, in this paper we further speculate

on the use of EEG as a biometric characteristic by

focusing on the analysis of repeatability of its fea-

tures, thus starting ﬁlling a gap in the existing liter-

ature. More in details, we rely on two simple acquisi-

tion protocols, namely “resting states with eyes open”

and “resting states with eyes closed” to acquire data

from nine healthy subjects in two acquisition sessions

separated in time. Different conﬁgurations for the

number of electrodes employed and for their spatial

placement have been taken into account. Speciﬁcally,

sets of three and ﬁve electrodes have been considered

to acquire the signals, and several frequency bands

have been analyzes. The so acquired signals, after

proper preprocessing, are then modeled using autore-

gressive stochastic modeling in the feature extraction

stage. Linear classiﬁcation is then performed. The pa-

per is organized as follows. The acquisition protocol

is detailed in Section 2, and in Section 3 the template

extraction procedure is described. In Section 4 classi-

ﬁcation is performed, while experimental results are

given in Section 5. Finally conclusions are drawn in

Section 6.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

420

FPZ

CPZ

CP6

CP4

CP2

CP1

CP3

CP5

TP7

FCZ

POZ

PO3 PO1

PO7

PO8

PO4PO2

AFZ

AF3 AF1

AF4AF2

AF8

AF7

FP2

FP1

FT7

FC4

FC2

FC1

FC3

FC5

TP8

FT8

FC6

Figure 1: Scalp electrodes positioning in the employed pro-

tocol according to an extension of the standard 10-20 mon-

tage.

2 EMPLOYED ACQUISITION

PROTOCOL

Nine healthy volunteers have been recruited for this

experiment. Informed consent was obtained from

each subject after the explanation of the study, which

was approved by the local institutional ethical com-

mittee. During the experiment, the participants were

comfortably seated in a reclining chair with both arms

resting on a pillow in a dimly lit room properly de-

signed minimizing external sounds and noise in or-

der not to interfere with the attention and the relaxed

state of subjects. The subjects were asked to per-

form one minute of “resting state with eyes open” and

one minute of “resting state with eyes closed” (Barry

et al., 2007) in two temporally separated sessions,

from 1 to 3 weeks distant from each other, depending

on the subject. Brain activity has been recorded us-

ing a BrainAmp EEG recording system operating at a

sampling rate of 200 Hz. The EEG was continuously

recorder from 54 sites positioned according to the In-

ternational 10-20 system as shown in Figure 1. Such

conﬁguration is not meant to be a user convenient so-

lution, but allows to investigate about a proper elec-

trode placement, able to catch distinctive features ac-

cording to the employed protocol. Before starting the

recording session, the electrical impedance of each

electrode was kept lower than 10 kOhm through a

dedicated gel maximizing the skin contact and allow-

ing for a low-resistance recording through the skin.

After the EEG recording sessions, the EEG signals

have been band pass ﬁltered in the band [0.5,30] Hz.

In addition, a preprocessing has been applied to im-

prove the signal-to-noise ratio (SNR), as described in

Section 3.1.

3 TEMPLATE EXTRACTION

The template is generated by considering the sig-

nals acquired by a properly chosen set of electrodes.

We have tested different acquisition conﬁgurations.

Speciﬁcally, sets of three and ﬁve electrodes have

been employed in our tests to understand, at a ﬁrst

stage, the proper number of electrodes to employ, lim-

iting it, and at a later stage, the proper electrodes po-

sitioning to use in order to capture repeatable and sta-

ble features, if present. The signals so acquired are

preprocessed as described in Section 3.1 in order to

perform denoising and to select the proper subbands.

Then, the EEG signals in the selected subbands are

AR modeled as described in Section 3.2. The tem-

plate is obtained by concatenating the reﬂection coef-

ﬁcients vectors related to the different channels in the

sets under analysis. Speciﬁc brain rhythms are mainly

predominant in certain scalp regions during different

mental states. Therefore, we expected a certain vari-

ability of recognition performance spanning the entire

scalp through speciﬁc conﬁgurations of electrodes,

and considering the closed or open eyes condition,

being different the capability to catch distinctive fea-

tures.

3.1 Preprocessing

Before performing feature extraction, each acquired

raw EEG signal has been processed as described in

the following. Neural activity reﬂected in resting state

EEG signals shows to contain frequency elements

mainly below 30Hz. Hence, a decimation factor has

been applied to the collected raw signals, after ﬁlter-

ing them through an anti-aliasing FIR ﬁlter. A sam-

pling rate of S

= 60Hz was selected to retain spectral

information present in the four major EEG subbands

referring to the resting state (δ [0.5,4]Hz, θ [4,8]Hz, α

[8,13]Hz and β [13,30]Hz). The γ subband [30,40]Hz

is not considered, given that it is known not to be rel-

evant in a resting condition. A further stage of zero-

phase frequency ﬁltering was applied to discriminate

the different EEG rhythms. The single δ, θ, α and β

subbands and their combinations (frequency compo-

nents from 0.5Hz up to 30Hz, and from 0.5Hz up to

14Hz) have been considered in our experiments.

A spatial ﬁlter has been then applied to the ac-

quired signals. When sufﬁciently large numbers of

OntheRepeatabilityofEEGFeaturesinaBiometricRecognitionFrameworkusingaRestingStateProtocol

421

electrodes are employed, potential at each location

may be measured with respect to the average of all po-

tentials, approximating an inactive reference. Speciﬁ-

cally, a common average referencing (CAR) ﬁlter has

been employed in the herein proposed analysis by

subtracting the mean of the entire C

= 54 electrodes

montage (i.e. the common average) from the channel

c of interest, with c = 1,2,··· ,C

, at any one instant,

according to the formula:

CAR

[n] = V

[n] −

∑

j=1

[n], (1)

where V

[n] is the potential between the j-th elec-

trode and the reference electrode, for the user u, with

u = 1, 2, ··· ,U. CAR ﬁltering has been employed

to reduce artifacts related to inappropriate reference

choices in monopolar recordings (Schwartz and An-

drasik, 2003) or not expected reference variations, as

well as to provide measures as independent as pos-

sible from the recording session. This results in an

increased signal-to-noise ratio, since artifacts related

to a single reference electrode are better controlled,

as showed in (McFarland et al., 1997), where authors

compared spatial ﬁlter methods with a conventional

ear reference in an EEG-based system.

A set of instances to be used for the training and

the testing stages has been obtained from the signal

segmentation. A range from 1 up to 3 seconds of

EEG frame length has been spanned stepwise, in or-

der to best characterize each user brain signal for the

identiﬁcation purpose. The one second frame length

has been experimentally selected as it has shown to

best catch distinctive features of users’ EEG segments

for the recognition purpose. This can be observed in

Figure 2, where best performance is achieved both

for sets of three and ﬁve electrodes, considering one

second EEG segments, in the band δ ∪ θ ∪ α ∪ β =

[0.5,30]Hz shown to be the best performing. These

results refer to 10 order AR modeling and best sets

of three and ﬁve channels, and show averaged perfor-

mance obtained training the classiﬁer on each acqui-

sition session and testing it on the other one. Such

framework has been employed to increase the num-

ber of trials used to study the repeatability of EEG

biometrics in terms of recognition performance over

the investigated period.

In this stage an overlap interval between adjacent

frames was set to increase the sample size. Overlap-

ping percentages of 25%, 50% and 75% have been

tested. Subsequently the DC component jointly to

the linear trend has been removed from each EEG

segment. The so obtained data-set have been further

processed to extract the distinctive features from each

user brain signal, as described below.

3.2 Modeling and Feature Vector

After the preprocessing stage, detailed in Section 3.1,

each acquired signal is modeled as a realization of an

AR stochastic process. A realization x[n] of an AR

process, of order Q, can be expressed as:

x[n] = −

∑

q=1

Q,q

· x[n− q] + w[n] (2)

where w[n] is a realization of a white noise process

of variance σ

, and a

Q,q

are the autoregressive coefﬁ-

cients. The well known Yule-Walker equations (Kay,

1988), which allow calculating the Q coefﬁcients, can

be solved recursively, employing the Levinson algo-

rithm and introducing the concept of reﬂection coef-

ﬁcients. Speciﬁcally:

(

Q,q

= a

Q−1,q

+ K

· a

Q−1,Q−q

, q = 1, ··· ,Q− 1

= σ

Q−1



1− K



(3)

where the factor K

is the so-called reﬂection coefﬁ-

cient of order Q which is calculated as follows (Kay,

1988):

= −

[Q] +

Q−1

∑

q=1

[q] · a

Q−1,Q−q

/σ

Q−1

(4)

where the generic R

[m] is the signal autocorrelation

function, deﬁned as R

[m] = E {x[n]x[n− m]}, for all

m ≥ 0.

Among the possible estimation approaches, the

Burg method (Kay, 1988) estimates the reﬂection co-

efﬁcients K

, for q = 1,...,Q, operating directly on

the observed data x[n] rather than estimating the au-

tocorrelation samples R

[m]. Therefore, the Burg’s

reﬂection coefﬁcients, which have been shown in

(Campisi et al., 2011) to achieve better performance

than the most commonly employed AR coefﬁcients,

are here employed.

Given the generic user u, and the generic channel

c, let us indicate with ζ

(u,c)

the vector, of length Q,

composed by the AR model reﬂection coefﬁcients K

for q = 1,...,Q, using the Burg method:

(u,c)

= [K

(u,c)

,··· , K

(u,c)

]

. (5)

The model order Q has been selected according to

the Akaike Information Criterion (AIC) to minimize

the information loss in ﬁtting the data. It can be ob-

served in Figure 3(a), that the AIC(Q) function, aver-

aged among subjects and channels, reaches minimum

plateau zone for values of Q from 6 to 12. The feature

vector x for the user u is obtained by concatenating

the AR coefﬁcients vectors related to the signals ob-

tained from the channels in the set under analysis. The

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

422

1 2 3

96.5

97.5

98.5

99.5

100

Frame Length (s)

Performance %

3 ch.

5 ch.

Figure 2: Classiﬁcation results in % obtained for the best

performing set of three (P7-Pz-P8) and ﬁve channels (Cz-

CP5-CPz-CP6-Pz), considering AR order Q = 10 and dif-

ferent values of frame length (T

0 10 20 30

0.2

0.4

0.6

0.8

AR Order (Q)

AIC(Q)

6 8 10 12

100

AR Order (Q)

Performance %

3 ch.

5 ch.

Figure 3: a) AIC function, averaged on all subjects and

channels, for the frequency band [0.5,30] Hz. b) Classiﬁ-

cation results in % obtained for the best performing set of

three (P7-Pz-P8) and ﬁve channels (Cz-CP5-CPz-CP6-Pz),

considering T

= 1s and different AR orders Q.

10 AR order has been experimentally selected since it

has shown to best ﬁt the EEG data for the recognition

purpose, as it can be observed in Figure 3(b), where

correct classiﬁcation percentage is reported consider-

ing one second EEG segmentation. Averaged results

are shown, obtained training on each session and test-

ing on the remaining one, and considering the best

performing sets of three and ﬁve channels.

4 CLASSIFICATION

The classiﬁer we propose estimates the class (user

identity) to which the observed feature vector x be-

longs to by means of a linear transformation

(G) =

G, where the transformation matrix G is obtained

by minimizing the mean square error (MMSE) thus

obtaining:

G = argmin

∑

i=1

·E

x|H



−

y(Γ)]

−

y(Γ)]



(6)

where H

indicates the hypothesis x belongs to the i-th

class, with i = 1,2,··· ,N. Here, assuming the hypoth-

esis H

holds, the vector y

= [0,...,0,1,0,... ,0] with

the unique one in the i-th position, indicates the class

i x belongs to, while

(G) represents its estimation.

denotes the prior probability that x belongs to the

i-th class. It can be easily shown that the employed

optimization criterion given in (6) brings to the nor-

mal equations:

· G = R

(7)

where R

= E{xx

} is the auto-correlation matrix

for the elements of the feature vector x, while R

∑

i=1

·E

x|H

{x} · y

turns out to be the matrix whose

columns are the probabilistically averaged condi-

tional mean values of the observations x.

4.1 Dataset

As pointed out in Section 3, different sets of N

= 3,5

channels, to acquire the signals from which the fea-

ture vectors x is extracted, have been considered for

both the employed protocols. Given a chosen set of

channels, each of the signals so acquired has been pre-

processed, as described in Section 3.1, segmented into

frames, and modeled by resorting to the reﬂection

coefﬁcients of an AR model of order Q. Therefore,

considering, for each user, EEG signals of duration

of 60s, segmented into frames of 1, 2 and 3s, with

an overlap factor of 75%, a number of N

= 237, 117

and 77 frames has been obtained respectively, each of

which is represented by the feature vector x of Q×N

elements.

Such a set of feature vectors has been collected for

each of the two temporally separated recording ses-

sions, 1-3 weeks distant from each other, and each

protocol, i.e. closed and open eyes resting condi-

tions. It is worth pointing out that the vectors used

in the training stage and in the recognition stage have

been obtained from the two different acquisition ses-

sions in order to infer about the repeatability over the

considered interval of the EEG features for the ac-

quired dataset and the employed acquisition proto-

cols. Hence, we applied the classiﬁcation algorithm

selecting the train and the test datasets without shuf-

ﬂing the EEG frames belonging to different sessions,

as performed in other works with user recognition

aims. In order to achieve our goal, each one of the

two sessions has been sequentially considered for the

training dataset while the remaining session has been

used to obtain the test dataset, thus obtaining two cou-

ples of temporally separated datasets, (training set,

recognition set) to train and test the classiﬁer. This

kind of validation framework has been provided just

to encrease the statistical signiﬁcance of the results.

They show that we can’t assess a perfect simmetry of

changes over time, but that the features keep stable

OntheRepeatabilityofEEGFeaturesinaBiometricRecognitionFrameworkusingaRestingStateProtocol

423

over the considered interval (1-3 weeks). Results of

each test are shown in subsequent columns of Tables

3, and 5 where each set has been acquired at a differ-

ent time.

4.2 Training

The training stage consists in the estimation of the

matrix G in (7) computed as G =

−1

, where

the matrices R

and R

are estimated through Mon-

teCarlo runs, considering equal prior probabilities P

for all the classes (users identities) to distinguish be-

tween. The estimation was obtained performing the

following two sample averages:

∑

i=1

∑

m=1

m,i

∑

i=1

∑

m=1

m,i

(8)

where x

m,i

is the m-th observed feature vector be-

longing to the i-th class, with M being the num-

ber of instances of x for each class, and y

[0,...,0,1,0,...,0]

, with the unique 1 in the i-th po-

sition. The considered matrices can be simply up-

graded in case of enrollment of N

′

new users, sum-

ming the related matrices x

m,i

to R

, and adding

new columns i to R

given by

M(N+N

′

)

∑

m=1

m,i

where i = N + 1, ...,N + N

′

. To avoid failures and

to control accuracy in the estimation of R

−1

, the sin-

gular value decomposition based pseudoinversion has

been used for the matrix inversion.

4.3 Recognition

In the recognition stage, a linear transformation is ap-

plied to each of the M × N observations from the test

dataset. For the i-th user a score vector

was ob-

tained for each instance of x

in the dataset applying

the discrimination matrix G to x

m,i

= G· x

m,i

(9)

with m = 1,...,M. Subsequently, the M score vectors

related to each tested user were summed together to

reduce the misclassiﬁcation error, obtaining

∑

m=1

m,i

. (10)

Finally the estimation of the index representing the

user identity is obtained locating the maximum of the

score vector

= [ ˆy

(1),..., ˆy

(N)]

according to the

criterion

i = arg

maxy

(l). (11)

5 EXPERIMENTAL RESULTS

An extensive set of experiments has been carried out

to test the repeatability of EEG features in the con-

sidered interval. Repeatability and stability represent

properties of paramount importance for the use of

EEG biometrics in real life systems. For this purpose

we have selected two simple tasks to be performed by

a set of users, and a classiﬁcation problem has been

set up, where the training set and the one to be used in

the recognition stage have been chosen belonging to

temporally separated sessions 1-3 weeks distant from

each other.

More in detail, given the “resting state” acquisi-

tion protocols here considered and the 54 employed

channels shown in Figure 1, we have considered dif-

ferent subsets of acquisition channels in order to ﬁnd

the best performing spatial arrangements of the elec-

trodes while minimizing their number. Although the

considered acquisition technique doesn’t result user

convenient, not being this the focus of the paper, in a

preliminary study, as this is, it allows to detect on the

scalp the brain rhythms which privide the best distinc-

tive features, according to the employed protocol. In

order to achieve this goal we have considered sets of

three and ﬁve electrodes, listed in Tables 3 and 5.

Template extraction has been performed as de-

scribed in Section 3, by ﬁrst preprocessing the EEG

signals, which includes decimation with sampling

rates S

= 60 Hz, CAR spatial ﬁltering, segmentation

into frames of T

s with an overlapping factor O

be-

tween consecutive frames, and eventually band pass

ﬁltering in order to analyze the subbands δ, θ,α and

β, which are the ones interested by the “resting state”

protocols, and some of their combinations. A value of

= 75% has been here employed since we have ex-

perimentally proven it is able to guarantee good per-

formance as it provides an adequate sample size to

assort the training and recognition datasets. Then the

so obtained frames are modeled using an AR model,

whose tested orders Q = {6,8,10, 12} have been esti-

mated by means of the AIC function (see Figure 3(a)).

Value of T

= 1s and Q = 10 have been selected as

they showed to best characterize the users’ EEG for

the recognition task (see Figure 2.)

The template is then obtained by concatenating

the reﬂection coefﬁcients of the signals acquired by

means of the electrodes set under analysis, thus gen-

erating feature vectors of length 3Q,5Q for the sets of

three, and ﬁve electrodes respectively.

In Tables 3 and 5 the results obtained for sets of

three and ﬁve electrodes when using the MMSE clas-

siﬁer, described in Section 4, are given for both em-

ployed protocols CE and OE. It is worth pointing out

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

424

Table 2: Classiﬁcation results in % for CE protocol, obtained using the acquisition session t for training and the acquisition

session r for recognition, with t,r = S

and t 6= r, for the subband δ ∪ θ∪ α ∪β = [0.5,30] Hz, for sets of three electrodes.

For each test t → r 2 results are provided, considering 75% of the training dataset while 25% (ﬁrst column) and 75% (second

column) of the test dataset.

Closed Eyes

Spatial ﬁltering (CAR) No spatial ﬁltering

Electrodes S

→ S

Fp1 Fpz Fp2 51.66 57.57 65.03 69.43 59.87 70.18 65.31 73.42

AF3 AFz AF4 63.43 65.40 66.15 70.79 76.89 91.00 79.23 87.15

F7 Fz F8 50.54 57.76 73.89 82.09 70.32 71.26 78.86 88.65

F3 Fz F4 53.26 59.54 66.34 67.84 59.45 65.54 66.71 69.39

F1 Fz F2 60.81 66.43 74.03 82.56 73.14 85.98 72.95 76.14

FC3 FCz FC4 73.61 81.15 80.45 91.09 77.12 91.14 84.29 94.05

T7 Cz T8 68.26 74.59 70.56 77.12 65.64 59.77 63.90 69.29

C3 Cz C4 78.15 82.51 74.82 85.56 78.57 93.48 84.29 87.39

C1 Cz C2 78.43 88.98 81.81 94.19 80.78 87.01 92.50 99.91

TP7 CPz TP8 65.78 75.34 62.82 57.01 75.06 79.75 80.97 85.61

CP3 CPz CP4 63.62 66.85 59.12 72.95 69.25 71.26 80.03 87.25

P7 Pz P8 93.44 100 94.56 99.62 95.50 100 97.47 100

P5 Pz P6 89.69 99.62 93.25 99.06 80.54 87.01 92.31 100

P3 Pz P4 79.00 79.89 86.08 89.22 67.84 67.93 70.84 78.48

P1 Pz P2 65.07 70.60 62.82 70.79 63.90 63.85 63.06 69.48

PO3 POz PO4 70.98 69.06 77.64 80.87 68.07 81.20 74.07 85.51

O1 POz O2 69.57 74.82 70.70 67.79 69.67 78.43 66.24 70.75

Table 3: Classiﬁcation results in % for OE protocol, obtained using the acquisition session t for training and the acquisition

session r for recognition, with t,r = S

and t 6= r, for the subband δ ∪ θ∪ α ∪β = [0.5,30] Hz, for sets of three electrodes.

For each test t → r 2 results are provided, considering 75% of the training dataset while 25% (ﬁrst column) and 75% (second

column) of the test dataset.

Open Eyes

Spatial ﬁltering (CAR) No spatial ﬁltering

Electrodes S

→ S

Fp1 Fpz Fp2 57.48 68.03 67.65 64.65 43.46 49.79 55.41 56.59

AF3 AFz AF4 56.12 56.96 47.12 56.17 53.87 59.21 69.39 73.98

F7 Fz F8 63.57 63.48 67.28 70.18 62.17 67.14 69.67 74.54

F3 Fz F4 66.10 68.17 72.25 70.28 65.78 73.00 60.76 69.01

F1 Fz F2 66.01 66.67 67.23 70.98 54.62 58.37 67.28 69.10

FC3 FCz FC4 83.97 87.15 78.57 87.11 66.99 68.45 67.56 65.96

T7 Cz T8 87.06 83.68 84.11 89.59 76.79 77.03 75.25 80.97

C3 Cz C4 80.08 90.53 78.20 80.87 76.84 81.58 69.48 77.92

C1 Cz C2 72.39 73.65 73.65 82.37 65.45 65.17 69.85 81.20

TP7 CPz TP8 53.26 53.91 58.74 66.29 52.23 55.09 58.09 62.59

CP3 CPz CP4 63.24 72.53 62.96 65.17 75.43 82.00 60.29 63.90

P7 Pz P8 55.32 56.54 59.82 64.70 59.17 60.85 51.24 47.30

P5 Pz P6 53.21 54.99 62.59 64.32 65.07 66.67 53.59 54.71

P3 Pz P4 53.73 57.99 68.59 76.32 79.93 85.33 67.74 76.65

P1 Pz P2 55.79 51.34 59.45 66.76 68.40 74.87 56.82 60.29

PO3 POz PO4 51.76 49.41 55.37 52.23 56.02 63.01 56.26 60.76

O1 POz O2 52.13 48.76 54.85 54.15 54.81 56.02 56.49 61.46

OntheRepeatabilityofEEGFeaturesinaBiometricRecognitionFrameworkusingaRestingStateProtocol

425

Table 4: Classiﬁcation results in % for CE protocol, obtained using the acquisition session t for training and the acquisition

session r for recognition, with t,r = S

and t 6= r, for the subband δ ∪ θ ∪ α∪ β = [0.5,30] Hz, for sets of ﬁve electrodes.

For each test t → r 2 results are provided, considering 75% of the training dataset while 25% (ﬁrst column) and 75% (second

column) of the test dataset.

Closed Eyes

Spatial ﬁltering (CAR) No spatial ﬁltering

Electrodes S

→ S

FCz T7 Cz T8 CPz 86.31 89.26 82.75 91.80 76.09 76.23 78.67 76.65

FCz C5 Cz C6 CPz 91.05 97.47 88.98 91.14 88.51 94.23 93.34 96.20

FCz C3 Cz C4 CPz 84.90 93.48 83.50 92.73 85.47 97.37 88.05 93.86

FCz C1 Cz C2 CPz 86.69 99.25 84.76 96.67 89.64 98.12 90.67 100

Cz TP7 CPz TP8 Pz 81.67 88.89 80.08 78.62 81.81 88.89 88.98 100

Cz CP5 CPz CP6 Pz 93.44 96.30 95.22 100 94.98 100 96.62 100

Cz CP3 CPz CP4 Pz 93.11 96.30 81.53 89.45 91.28 100 88.94 95.97

Cz CP1 CPz CP2 Pz 96.30 100 93.81 98.03 93.67 95.08 91.65 100

CPz P7 Pz P8 POz 95.78 98.87 94.75 97.84 93.91 100 94.70 97.23

CPz P5 Pz P6 POz 87.62 90.39 91.37 96.39 83.36 94.51 91.70 98.69

CPz P3 Pz P4 POz 81.15 87.11 84.95 87.48 79.51 87.34 84.20 87.29

CPz P1 Pz P2 POz 69.29 75.01 68.82 73.09 65.31 62.96 64.70 72.95

Pz PO3 PO4 O1 O2 68.92 76.00 77.64 94.47 63.48 67.60 77.31 86.12

CPz Pz POz O1 O2 73.23 80.54 76.47 82.51 74.96 75.57 79.75 84.62

Table 5: Classiﬁcation results in % for OE protocol, obtained using the acquisition session t for training and the acquisition

session r for recognition, with t,r = S

and t 6= r, for the subband δ ∪ θ ∪ α∪ β = [0.5,30] Hz, for sets of ﬁve electrodes.

For each test t → r 2 results are provided, considering 75% of the training dataset while 25% (ﬁrst column) and 75% (second

column) of the test dataset.

Open Eyes

Spatial ﬁltering (CAR) No spatial ﬁltering

Electrodes S

→ S

FCz T7 Cz T8 CPz 84.01 79.93 88.05 88.89 79.75 78.90 82.42 87.67

FCz C5 Cz C6 CPz 80.26 75.76 73.98 78.62 70.89 72.90 64.60 70.60

FCz C3 Cz C4 CPz 87.29 88.89 84.29 85.51 73.89 73.56 69.43 73.46

FCz C1 Cz C2 CPz 78.76 81.15 75.95 79.93 69.62 73.42 73.42 80.83

Cz TP7 CPz TP8 Pz 66.39 66.67 59.54 68.78 60.24 59.07 57.06 63.15

Cz CP5 CPz CP6 Pz 68.12 68.21 64.98 66.57 76.75 79.42 71.82 74.87

Cz CP3 CPz CP4 Pz 77.78 86.45 69.48 73.00 74.12 76.84 68.07 71.12

Cz CP1 CPz CP2 Pz 84.58 87.72 71.92 76.51 68.96 71.59 65.73 68.26

CPz P7 Pz P8 POz 67.04 76.14 66.15 69.06 71.68 75.81 61.51 63.90

CPz P5 Pz P6 POz 59.68 59.68 63.95 67.09 62.96 66.67 58.42 56.17

CPz P3 Pz P4 POz 63.43 68.87 65.64 67.09 75.81 77.78 76.23 80.97

CPz P1 Pz P2 POz 65.26 66.48 55.70 57.95 68.82 77.40 61.28 59.49

Pz PO3 PO4 O1 O2 50.45 51.20 53.12 45.94 56.21 55.56 57.34 55.93

CPz Pz POz O1 O2 59.21 59.68 63.48 63.29 55.56 55.56 54.99 58.27

that the signals employed to obtain the templates to

be used in both the recognition and the training stage

are disjoint in time. Therefore two different combi-

nations of training (t) and recognition (r) sessions,

(t, r) with t,r ∈ {S

} and t 6= r, have been em-

ployed in order to infer about the repeatability, over

the considered interval, of the EEG features under

analysis, for a real usability of an EEG-based biomet-

ric system. Such kind of tests varying the sequence

of sessions in the recognition framework are provided

to validate the results about repeatability of the fea-

tures over the interval under analysis. The results for

the different tests performed are reported separately,

not expecting to make assumptions on symmetry of

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

426

Table 6: Classiﬁcation results in % for both CE and OE protocols, obtained using the same acquisition session S for training

and for recognition, with S = S

, for the subband δ∪ θ∪ α ∪ β = [0.5,30] Hz, for the best performing sets of three and ﬁve

electrodes. Results are provided considering 75% of the dataset for training while 25% of the dataset for recognition.

Closed Eyes Open Eyes

Spatial ﬁltering (CAR) No spatial ﬁltering Spatial ﬁltering (CAR) No spatial ﬁltering

Electrodes S

→ S

P7 Pz P8 96.81 100 98.03 100 95.22 96.11 94.56 93.25

Cz CP5 CPz CP6 Pz 98.87 100 100 100 100 100 97.00 91.80

changes over time. From Tables 3 and 5, it is possible

to note that triplets of channels allow achieving about

same performancethen conﬁgurations employing sets

of ﬁve channels. This is due to a good spatial localiza-

tion achieved by conﬁgurations of only three sensors,

which allow to well capture the underlying phenom-

ena, reducing the problem dimensionality. Moreover,

in Tables 3 and 5 results are shown considering the

band F = δ ∪ θ∪ α∪β = [0.5, 30]Hz which is the one

that allows obtaining the best results, and considering

a preprocessing including CAR ﬁltering or not.

Provided performance refers to a cross-validation

framework, obtained selecting for each user 75%

of feature vectors x related to cyclically subsequent

frames from the training dataset, while 25% and

75% from the test dataset, as reported in subse-

quent columns of Tables 3 and 5. Numerical results

are obtained averaging over 230 independent cross-

validation runs to improve the statistical analysis. As

previously pointed out, recognition tests have been

carried out keeping independency between the train-

ing and the test datasets, acquired in different ses-

sions, for the classiﬁcation purpose. This aspect is

highlighted in Tables 3 and 5 denoting with S

→ S

the result achieved training on S

and testing on S

It should be noticed that applying the CAR ﬁlter

in the preprocessing stage doesn’t yield a general im-

provement in the performance for all employed sets

of channels and protocols, while it appears to pro-

vide best results for some selected channels (FC3-

FCz-FC4, C3-Cz-C4, T7-Cz-T8) for sets of 3 chan-

nels in the OE protocol. This is likely to be due to ar-

tifacts which more affected the open-eyes condition,

removed by the spatial ﬁltering. As regards differ-

ences between the two employed protocols it is evi-

dent, by observing the reported results, that in this ex-

periment the CE protocol provides best performance

considering the adopted EEG feature extraction for

the recognition task, achieving 100% of correct clas-

siﬁcation, for example using channels P7-Pz-P8 and

75% of the test feature vectors for each user in the

cross-validation framework. This has been observed

to be due to being the open-eyes signal more affected

by the eyes movement artifacts. It is also worth to

be noticed that the combination of channels affected

in a different way the recognition results for CE and

OE protocols. In fact, referring to sets of three chan-

nels the parietal region has proven to best perform in

CE condition, while the central region achieved best

results in OE condition. This is in agreement with

the fact that in resting state with eyes closed the dom-

inant brain rhythm α can be detected mainly in the

posterior area of the scalp, while it is attenuated when

opening eyes. Moreover it has been observed, indi-

vidually analyzing the extracted brain rhythms, that

in CE the α band most contributed to the best per-

formance obtained combining all bands ([0.5, 30]Hz).

Repeatability over the considered interval of the ana-

lyzed EEG features can be inferred by observing that

users enrolled in a session have been recognized in a

different one, disjoint in time from 1 to 3 weeks. Be-

sides, it is also evident that by swapping the training

and recognition roles of the session datasets, that is by

considering (t,r) or (r,t), quite coherent performance

are obtained. Finally Table 6 shows results obtained

training and testing the classiﬁer on the same session.

It should be noticed that very high correct recogni-

tion rate is achieved considering just 25% of the test

dataset (100% for CE and S

), while a greater num-

ber of feature vectors for each user are needed in the

inter-sessions framework. This evidence proves the

importance of speculating about the stability and re-

peatability over time of EEG features for biometric

systems.

6 CONCLUSIONS

In this paper the problem of repeatability over time of

EEG biometrics, for the same user, within the frame-

work of EEG based recognition, has been addressed.

Simple “resting state” protocols have been employed

to acquire a database of nine people in two differ-

ent sessions separated in time from 1 to 3 weeks,

depending on the user. Although the dimension of

the database employed is contained, we would like

to stress out that this contribution represents the ﬁrst

systematic analysis on the repeatability issue in EEG

biometrics. As such, this contribution paves the road

to more reﬁned analysis which would include more

OntheRepeatabilityofEEGFeaturesinaBiometricRecognitionFrameworkusingaRestingStateProtocol

427

sessions separated in time as well as different ac-

quisition protocols. Extensive simulations have been

performed by considering different sets of electrodes

both with respect to their positioning and number.

In summary in our analysis a very high degree of

repeatability over the considered interval has been

achieved with a proper number of electrodes, their

adequate positioning and by considering appropriate

subband related to the employed acquisition protocol.

REFERENCES

Abdullah, M., Subari, K., Loong, J., and Ahmad, N. (2010).

Analysis of effective channel placement for an eeg-

based biometric system. In Biomedical Engineering

and Sciences (IECBES), 2010 IEEE EMBS Confer-

ence on, pages 303 –306.

Barry, R., Clarke, A., Johnstone, S., Magee, C., and Rushby,

J. (2007). Eeg differences between eyes-closed and

eyes-open resting conditions. Clinical Neurophysiol-

ogy, 118:2765–2773.

Brigham, K. and Kumar, B. V. (2010). Subject identiﬁca-

tion from electroencephalogram (eeg) signals during

imagined speech. In Proceedings of the IEEE Fourth

International Conference on Biometrics: Theory, Ap-

plications and Systems (BTAS’10).

Campisi, P., La Rocca, D., and Scarano, G. (2012). Eeg for

automatic person recognition. Computer, 45(7):87–

89.

Campisi, P., Scarano, G., Babiloni, F., De Vico Fallani,

F., Colonnese, S., Maiorana, E., and L., F. (2011).

Brain waves based user recognition using the eyes

closed resting conditions protocol. In IEEE Interna-

tional Workshop on Information Forensics and Secu-

rity (WIFS’11).

Carmena, J. M. (2012). Becoming bionic. IEEE Spectrum,

49(3):24–29.

Debener, S., Minow, F., Emkes, R., Gandras, K., and

de Vos, M. (2012). How about taking a low-cost,

small, and wireless eeg for a walk? Psychophysiol-

ogy, 49(11):1617–1621.

Dornhege, G., del R. Mill´an, J., Hinterberger, T., McFar-

land, D., and M´oller, K.-R., editors (2007). Towards

Brain-Computing Interfacing. MIT Press, Cambridge,

MA.

Kay, S. (1988). Modern Spectral Estimation. Theory and

Applications. Prentice-Hall.

La Rocca, D., Campisi, P., and Scarano, G. (2012). Eeg

biometrics for individual recognition in resting state

with closed eyes. In Proceedings of the International

Conference of the Biometrics Special Interest Group

(BIOSIG), pages 1–12.

Marcel, S. and del R. Millan, J. (2007). Person authentica-

tion using brainwaves (eeg) and maximum a posteri-

ori model adaptation. IEEE Transactions on Pattern

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

McFarland, D., McCane, L., David, S., and Wolpaw, J.

(1997). Spatial ﬁlter selection for EEG-based commu-

nication. Electroencephalography and Clinical Neu-

rophysiology, 103(3):386–394.

Mohammadi, G., Shoushtari, P., Molaee Ardekani, B., and

Shamsollahi, M. B. (2006). Person Identiﬁcation by

Using AR Model for EEG Signals. In Proceedings of

World Academy of Science, Engineering and Technol-

ogy, volume 11, pages 281–285.

Nakanishi, I., Baba, S., and Miyamoto, C. (2009). Eeg

based biometric authentication using new spectral fea-

tures. In ISPACS 2009, pages 651–654.

Palaniappan, R. and Patnaik, L. M. (2007). Identity veriﬁca-

tion using resting state brain signals. In Quigley, M.,

editor, Encyclopedia of Information Ethics and Secu-

rity, pages 335–341. IGI Global.

Paranjape, R., Mahovsky, J., Benedicenti, L., and Koles,

Z. (2001). The electroencephalogram as a biometric.

In Canadian Conference on Electrical and Computer

Engineering, pages 1363–1366.

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

gelou, A. (1999). Parametric person identiﬁcation

from the eeg using computational geometry. In The

6th IEEE International Conference on Electronics,

Circuits and Systems, 1999. ICECS ’99., volume 2,

pages 1005–1008.

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

Rufﬁni, G. Unobtrusive biometric system based on

electroencephalogram analysis. EURASIP J. Adv. Sig-

nal Process.

Schwartz, M. and Andrasik, F. (2003). Biofeedback: A

Practitioner’s Guide. Guilford Press.

Su, F., Xia, L., Cai, A., Wu, Y., and Ma, J. (2010). Eeg-

based personal identiﬁcation: from proof-of-concept

to a practical system. In Pattern Recognition (ICPR

2010), 20th International Conference on, pages 3728–

3731.

Zhao, Q., Peng, H., Hu, B., Liu, Q., Liu, L., Qi, Y., and

Li, L. (2010). Improving individual identiﬁcation in

security check with an eeg based biometric solution.

In Proceedings of the 2010 international conference

on Brain informatics, BI’10, pages 145–155, Berlin,

Heidelberg. Springer-Verlag.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

428