EXE - A Survey On Methods and Challenges in EEG Based Authentication

Computers & Security 93 (2020) 101788
Contents lists available at ScienceDirect
Computers & Security

journal homepage: www.elsevier.com/locate/cose
A survey on methods and challenges in EEG based authentication

Amir Jalaly Bidgoly a,∗, Hamed Jalaly Bidgoly b, Zeynab Arezoumand a
a
Department of Information Technology and Computer Engineering, University of Qom, Iran
b
School of Electrical and Computer Engineering, University of Tehran, Iran
a r t i c l e i n f o a b s t r a c t
Article history: EEG is the recording of electrical activities of the brain, usually along the scalp surface, which are the
Received 13 October 2019 results of synaptic activations of the brain’s neurons. In recent years, it has been shown that EEG is
Revised 21 January 2020
an appropriate signal for the biometric authentication and has important features such as resistance to
Accepted 10 March 2020
spoofing attacks and impossibility to use under pressure and coercion states. In this paper, the state-
Available online 11 March 2020
of-the-art methods in EEG based authentication are reviewed. This review includes a number of aspects
Keywords: such as the various tasks that the user required to perform during the authentication, devices and avail-
EEG able datasets, the preprocessing procedures and the classification methods used in the EEG biometric au-
Biometric factor thentication. Both shallow and deep classification methods are reviewed in this paper. The study shows
Authentication that the deep learning approaches which are used in the past few years, although still require further
User identification research, have shown great results. Moreover, the paper summarizes the works to address the open chal-
Pattern recognition
lenges of this area. The EEG authentication challenges have been discussed from a variety of points of
Survey
view, including privacy, user-friendliness, attacks, and authentication requirements such as universality,
permanency, uniqueness, and collectability. This paper can be used as a preliminary plan and a roadmap
for researchers interested in EEG biometric.
© 2020 Elsevier Ltd. All rights reserved.
1. Introduction • Biometric factors such as fingerprints or iris may be forged or

imitated. For instance, researchers show that an intruder can
In computer science and cryptography protocols, authentication simply make a fake fingerprint by just having a picture of one’s
is defined as the act of confirming the claimed identity of a sub- finger (Galbally et al., 2013). However, unlike the other biomet-
ject. Authentication is a critical tool to manage access to physical ric factors, EGG is not exposed to the intruder which makes it
or digital resources such as buildings, rooms or computing devices. uncapturable and thus, hard to be forged (Ruiz-Blondet et al.,
One of the most important approaches for user authentication is 2016).
the use of unique biological characteristics of the human being. • EEGs are emotional state dependent and stress or fear changes
Different biometric factors are provided for this purpose, which en- the normal brain waves’ patterns regardless of the activity. As a
compass almost many aspects of human being which can be mon- result, EGG biometric cannot be invoked under force and coer-
itored and processed by the computer systems, such as fingerprint, cion situations, while in the other authentication methods, even
palm, iris, voice, etc. In particular, researchers have special atten- in the challenge-based response ones, an adversary may access
tion to the use of biometric signals such as EEG, ECG, and EMG as to a system by force or threatening the user.
an appropriate tool of biometric authentication (Abbas et al., 2015; • In the other biometrics, one may use a dead (and warm) body
Blasco et al., 2016). These signals have distinguished advantages to authorize his/her access to the system. However, for EGG bio-
over other biometrics; for example, they are more robust against metric, a dead brain does not produce EEG signals. Hence, EEG
spoof attacks. itself guarantees that the user is alive and authenticating by his
Among the biometric signals, “electroencephalogram” (EEG) has own.
gained more interest. EEG biometric has unique advantages which
make it prominent between other biometric factors, such as: EEG based authentication has many applications in many fields
of computer science (Yang et al., 2018) and it can be used as a
tool to enhance the security of computer systems. For instance,
Klonovs et al. reported the development of a mobile application
∗
Corresponding author. that increases the security of smartphones through implementing
E-mail address: jalaly@qom.ac.ir (A. Jalaly Bidgoly). two-step authentication, by combining traditional methods of au-
https://doi.org/10.1016/j.cose.2020.101788
0167-4048/© 2020 Elsevier Ltd. All rights reserved.
2 A. Jalaly Bidgoly, H. Jalaly Bidgoly and Z. Arezoumand / Computers & Security 93 (2020) 101788
Fig. 1. Number of documents published about EEG based authentication.
thentication and an EEG based one (Klonovs et al., 2013). Apart In Section 3, the properties of the EEG as an authentication tool
from conventional authentication applications, some applications are studied. The available datasets in this field are reviewed in
have gained particular interest to use this biometric factor. “Brain Section 4. The Section 5 represents the various tasks that can be
computer interface” (BCI) is an outstanding example here (Lin et al., performed during EEG recording. Section 6 looks at how the sig-
2017; Ramzan and Shidlovskiy, 2018). BCI is a try to make direct nals can be enhanced by preprocessing. Shallow and deep models
communication gateway between a human brain and a device. The in EEG biometric are discussed in Sections 7 and 8, respectively.
human may control an electronic device not only by sending ex- Section 9 discusses about the security and privacy of EEG biomet-
plicit commands but also by his or her brainwaves. Integrating BCI ric. Finally, the paper ends with a review of the open challenges
and EEG based authentication makes these applications not only and the conclusion accordingly in Sections 10 and 11.
execute the user commands but also recognize his/her identity be-
fore executing. Since BCI is based on brainwaves, EEG biometric is 2. Background of EEG
the best candidate as an authentication factor in this application.
EEG may also have other security usages than authentication. As EEG is the recording of electrical activities of the brain, usu-
an example, researchers have shown that these signals can be used ally along the scalp surface. These electrical activities are results of
to generate a cryptographic key (Nguyen et al., 2017; Ravi et al., ionic current flows caused by the synchronized synaptic activation
2008). The generated key is unique for each person and task and of the brain’s neurons (Niedermeyer and da Silva, 2005) and re-
is more robust to attacks as compared to other encryption systems veal as rhythmic voltage fluctuations that range from 5 to 100μV
(Ravi et al., 2008). in amplitude and between 0.5 and 40 Hz in frequency. Analyzing
Considering the advances in the development of portable med- the dominant frequencies and amplitude of EEG waveforms in the
ical devices, including EEG headset, as well as the unique advan- different parts of brains can provide some clues about the physi-
tages of EEG, this biometric gains significant interest in recent cal or the mental state of the person(Bickford, 1987; Sanei, 2013).
years. Fig. 1 represents documents published in the context of EEG Based on the frequency, brain waves are classified into the five fre-
based authentication. Rapidly growing publications in recent years quency bands as follows(Sanei and Chambers, 2013):
show the increasing interest of the researchers in this field. Be- Delta (1 − 4Hz) is the slowest and usually the highest wave-
sides, the industries have shown their interest in using it in their form in amplitude. The delta band is observed in babies and
products. For instance, trending companies, like NeuroSky, are do- during deep sleep in adults;
ing research (Eeg headsets and the rise of passthoughts, 2017) Theta (4 − 8Hz) is observed in the children, drowsy adults and
and designing sensors Eeg - electroencephalogram - bci in this during memory recalling(Klimesch, 1996). The amplitude of
field. Moreover, researchers are working on EEG wearable systems Theta waves is normally less than 100μV;
using gears from Emotiv and NeuroSky, and even Google Glass Alpha (8 − 12Hz) is usually the dominant frequency band and
(Moore, 2016). appears during relaxed awareness or when eyes are closed.
An EEG based authentication system requires several steps. At Focused attention or relaxation with opening eyes reduces
first, the user should perform a specific task while the system the amplitude of the Alpha band. These waves are normally
records his/her EEG signals. The signals then have to be prepos- less than 50μ V;
sessed to enhance its quality. After preprocessing, the enhanced Beta (12 − 25Hz) is associated with thinking, active concen-
signals are used to train a classifier. Traditionally, shallow classi- tration and focused attention. Also, executing body move-
fiers such as “Support Vector Machine” (SVM) and “k-nearest neigh- ments or seeing other’s body movements increases Beta
bors” (KNN) have been used for this aim, but in recent years many power(Zhang et al., 2008). The amplitude of Beta waves is
of them are replaced by deep learning approaches. To use shallow normally less than 30μV;
models, it is necessary to extract different features of the signals, Gamma (over 25Hz) is observed during multiple sensory pro-
and then the model can be trained on them; while in deep learn- cessing(Read and Innis, 2017). Gamma patterns have the
ing approach, the feature extraction is not required as an extra step lowest amplitude.
and the model can be directly trained on the preprocessed data.
In this paper, the state-of-the-art research in different aspects In addition to the function-frequency relations, it is also be-
of biometric authentication systems has been reviewed. The pa- lieved that each part of the brain is responsible for doing a spe-
per continues as follows. In the next section, the characteristics of cific physical or mental function. Consequently, for studying brain
brain signals and also methods of data acquisition are presented. waves of each desired task, not only analyzing the dominant fre-
quency but also recording from the corresponding part of the brain
A. Jalaly Bidgoly, H. Jalaly Bidgoly and Z. Arezoumand / Computers & Security 93 (2020) 101788 3
Table 1
Functions associated to different parts of the brain(Demos, 2005).
Region Local channels Functions
Frontal Lobe Fp1, FP2, FPz, Pz, F3, F7, F4, F8 Memory, concentration, emotions.
Parietal Lobe P3, P4, Pz Problem Solving, attention, grammar, sense of touch.
Temporal Lobe T3, T5, T4, T6 Memory, face recognition, hearing, word recognition, social clues.
Occipital Lobe O1, O2, Oz Reading, vision.
Cerebellum — Motor control, balance.
Sensorimotor Cortex C3, C4, Cz Attention, mental processing, fine motor control, sensory integration.
1. Knowledge factor: the information which just the user knows,

2. Ownership factor: the object which just the user has/owns, and
3. Biometric factor: the physiological or behavioral characteristics
which just the user is.
The last one includes the biological aspects of a human, cover-
ing a wide range of elements from fingerprint to biometric signals
and DNA. The biometric factors have many unique advantages over
the other authentication factors. For instance, they are not required
to carry with, they cannot be stolen like the ownership factors, and
they are not required to be memorized like the knowledge factors.
Hence, the biometric factors bring the most comfort authentication
methods for the users.
Every biometric factor should have a set of requirements to be
appropriate for the authentication purpose. These requirements are
1) universality, 2) permanency, 3) uniqueness, and 4) collectabil-
ity (Armstrong et al., 2015; Jain et al., 2004). Universality means
that each biometric feature, which is considered as an authentica-
Fig. 2. Electrode placement in the 10–20 standard and the corresponding brain tion factor, should be capable of being used for every person (not
lobes under each electrode.
only for a group of people with a specified set of assumptions).
Permanency indicates that the biometric factor should be stable
should be more attended. Functions that are associated to each over time. Uniqueness requires that the biometric factor be dis-
part of brain, are summarized in Table 1 (Demos, 2005). tinguishable for every different user. Finally, collectability refers to
To capture the EEG signals, low impedance electrodes are at- the requirement that each biometric factor should be easily col-
tached to the scalp to pick up electric potentials generated by the lected from the user. In literature, it seems that EEG satisfies all
brain activities. The electrodes can be placed via conductive gel, the requirements. It could easily be collected from everyone by a
namely wet electrode, or be directly contacted with skin, namely low-cost EEG headset, and a set of features can be extracted from
dry electrode. Typically, dry electrodes are easier to attach, however, it which is unique for each individual. Also, papers show that the
they are more sensitive to the motion artifacts. There are some features are quite stable over time. In the Section 10, we will re-
standards to locate and label the electrodes on the scalp; e.g., 10– visit these requirements and discuss the state-of-the-art and chal-
20 standard(Demos, 2005). In the 10–20 standard, the electrodes lenges of the research in achieving them.
are placed at 10% and 20% points along lines of longitude and lat- As mentioned before, constructing an EEG based authentication
itude, respectively, and labeled according to the attached lobe. Be- system requires several steps. Usually, the user should be calm
sides, odd numbers are assigned to the electrodes on the left hemi- and perform a pre-defined task with a specific protocol. These
sphere, and even numbers are assigned to the electrodes on the tasks are defined in such a way that bring the brain into action,
right one; see Fig. 2. but do not require movement of the other parts of the body. In-
There are both off-the-shelf clinical and commercial products stances of these tasks are singing a song in mind, imagining exer-
in the market for recording EEG signals. The commercial products cise of a sport, thinking of a particular keyword, or simply resting.
are often designed as a headset for ease of use. These EEG head- The authentication system records the user EEG signals through a
sets are not only capable of capturing user’s EGG, but also offer- portable and commercially available EEG headset. These signals are
ing developer toolkits that can detect emotions, attention or sim- initially pre-processed to enhance their quality and then used to
ple BCI commands which makes them much easier to use in prac- train a classifier. Please note that we should distinguish between
tice. Emotiv and Neurosky companies are the leading ones in these “user identification” and “user authentication” (which often the pa-
products. As an example, “Emotiv Epoc+” headset (Fig. 3a) can cap- pers mistakenly used them instead of each other). The identifica-
ture AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8 and AF4 tion process deals with the question of who the user is, but the
channels using 14 embedded wet electrodes with 2048Hz sampling authentication method is to prove or disprove the user’s claimed
rate. “Mindwave” headset (Fig. 3b) is another sample provided by identity by considering a given statistical property. In the former,
Neurosky. These devices have been used in various papers to man- each user is modeled as a separate class and the model tries to
ually collect the required EEG data. identify the user’s class by receiving his/her EEG records. In the
latter, a binary classifier is used only for admission or rejection
3. EEG as a biometric authentication factor of the claimed identity. In this case, the model receives an indi-
vidual identity and the recorded EEG signals and then predicts if
The authentication process is usually performed based on some the records belong to the given identity. In both cases, an initial
previously known and unique information about the user which recording of the EEG should be used to train the classifier, namely
are named as “authentication factors”. Generally, there are three “registration phase”. After the initial training of the model, the user
types of authentication factors: will be able to authenticate/identify himself to the system in the
Fig. 3. Commercial EEG headsets.
Fig. 4. The process of EEG-based authentication.
“enrollment phase”. To this end, he must repeat exactly the task 4.1. Physionet EEG motor movement/imagery dataset
that he performed during the initial recording. This new record
is then processed and classified by the model and based on the Physionet EEG Motor Movement/Imagery dataset
classifier type, the identity of the user is detected in the identifica- (Goldberger et al., 20 0 0) is one of the most popular and publicly
tion or confirmed in the authentication. In other words,the model available datasets produced by the BCI20 0 0 system (Schalk et al.,
determines that who is enrolled in the identification or who that 2004). This dataset includes EEG records from 109 healthy vol-
gives statistical property is the same as the claimed identity in the unteers who performed various Motor/Imagery tasks. EEG signals
authentication. The overall process of authenticating using EEG is were recorded by using 64 channels with the sampling rate
presented in Fig. 4. of 160Hz. Each subject performed 14 experimental runs: two
one-minute baseline runs (one with eyes opened, one with eyes
closed), and three two-minute runs of each of the four following
4. Available datasets tasks: open and close left or right fist, imagine opening and
closing left or right fist, open and close both fists or both feet, and
In this section, the public datasets that are mostly used in the imagine opening and closing both fists or both feet.
EEG authentication are reviewed.
Due to the number of subjects and sampling channels, as well used for EEG authentication in some works (Delpozo-Banos et al.,
as the diversity of tasks, this dataset attracts the attention of re- 2015; Pham et al., 2015; Vahid and Arbabi, 2016; Wilaiprasitporn
searchers and has been used in many works including (Delpozo- et al., 2019). EEG samples were collected from 32 healthy subjects
Banos et al., 2015; Kang et al., 2018; Keshishzadeh et al., 2016; Kim using 32 channels with 512 Hz sampling rate. For each one, 40
and Kim, 2019; Lan Ma et al., 2015; Sun et al., 2019; Thomas and one-minute videos were played, where each one rises a specific
Vinod, 2018; Wang and Najafizadeh, 2016), to name but a few. emotion of the subject such as pride, joy, satisfaction, hope, sad-
ness, and fear.
4.2. BCI competition III dataset
4.7. Other datasets
This dataset, which itself is made up of a set of different
datasets, is provided by the cooperation of a team of universi- Aside from the datasets mentioned above, there are other avail-
ties and laboratories such as Graz University of Technology, Uni- able data collections in this field that are less well-known. Keirn
versity OF TÜBINGEN and Berlin BCI group. Datasets include EEG and Aunon dataset (Keirn and Aunon, 1990) is a small and old
records of one to five subjects performing different tasks. For in- database containing EEG from 5 individuals (4 males and 1 female
stance, IIIa as one of these datasets is the EEG records of 3 sub- between the ages of 21 and 48), which is recorded by 6 chan-
jects (named K3b, K6b, and L1b) that were recorded with 250Hz nels and the sampling rate of 250 Hz. The tasks defined in this
sampling rate using the Neuroscan amplifier with 62 EEG channels dataset are interesting and unique. Each user had to do five men-
(60 electrodes + 2 reference electrodes). In the experiments, the tal tasks, including silently rest, solving a complex multiplication
subjects performed 60 trails of the imagination of the movement without physical motion, silently imagining the rotation of a three-
of 4 body parts according to a cue, including left hand, right hand, dimensional object around an arbitrary axis, visualizing numbers
foot, and tongue. These datasets are used by various research (Bao on an imaginary backboard, and writing a mail to a friend in mind.
et al., 2009; Hu, 2009; 2010; Jian-feng, 2009; Nguyen et al., 2017; Several datasets have also been introduced in recent years.
Xiao and Hu, 2010). Sockeel et al. (2016) published a dataset including 11 subjects’
EEG recorded by a 62 channels BrainAmp cap in 2016. In 2017,
4.3. BCI competition IV datasets Kaur et al. (2017) have published a dataset containing brain sig-
nals of 60 different subjects while listening to four different types
In continue to BCI competitions, the fourth BCI competition of music including devotional, electronic, classical and rock. Their
was held in 2008 with 5 different datasets. These datasets include main goal was to examine the impact of different emotions on EEG
more subjects and tasks. For example, Dataset 2a (Brunner et al., based identification systems.
2008), as one of these datasets, contains EEG signals from 9 sub-
jects recorded using 22 EEG channels and 3 EOG channels with the 5. EEG acquisition protocols
sampling rate of 250Hz. Subjects performed four different tasks in-
cluding the imagination of the movements of the left hand, the In general, protocols used in the EEG recording can be catego-
right hand, both the legs and the tongue. The experiments were rized into three groups: resting states, mental tasks, and tasks with
recorded in two sessions on different days for each subject. Each an external stimulus. The selected protocol can affect the proce-
session contains 6 runs that were separated by short breaks. These dure or the accuracy of authentication. For instance, the resting
datasets have also been used by various papers (Lawhern et al., states or the mental tasks do not require any extra equipment but
2018; Schirrmeister et al., 2017). EEG recording devices; while, tasks followed by external stimuli
4.4. UCI KDD EEG database
The records of this dataset (Begleiter et al., 2002) are obtained

by 64 electrodes at a sampling rate of 256 Hz. The number of
subjects was 122 including normal and alcoholic males. In experi-
ments, each person was exposed to visual stimuli, including black
and white images from the Snodgrass and Vanderwart collection
(Snodgrass and Vanderwart, 1980). The experiments included two
stimulants S1 and S2, each showing for 300 ms and separating
within 1.6 seconds. The subjects were asked to determine if S1
complies with S2 or not. In some cases, only one trigger was pro-
vided. Liew et al. (2015) and Delpozo-Banos et al. (2015) are some
examples of papers which are used this dataset in their works.
4.5. Australian EEG database
Australian EEG database (Hunter et al., 2005) is a result of an

11-year study (1991–2002) on EEG signals of 40 patients (20 males
and 20 females) at John Hunter Hospital. The EEG was recorded
with open eyes and closed eyes for about 20 min, using 23 elec-
trodes at a sampling rate of 167 Hz. This dataset is used in sev-
eral research (Hunter et al., 2005; Nguyen et al., 2012; 2013; Phung
et al., 2014).
4.6. DEAP
The DEAP dataset (Koelstra et al., 2012) is a famous and well-

known dataset in emotion detection studies, however, it is also Fig. 5. Distribution of different EEG acquisition protocols in the EEG authentication.
require some devices to create the appropriate stimulation. On the peak that appears between 200 − 700ms after an infrequent stim-
other hand, simple tasks like resting states are easily affected by ulus(Polich, 2007). It’s worth mentioning that the RSVP protocol is
environment noise and artifacts, whereas mental tasks and tasks also a very common paradigm in the BCI applications(Wang et al.,
followed by external stimuli create a higher “signal-to-noise ratio” 2018). The other common paradigm between the BCI and the au-
(SNR). This can be achieved by distinguishing “Event Related Poten- thentication protocols, is periodic visual stimuli leading to “Steady-
tial” (ERP) which is an electrical potential generated in the brain in State Visual Evoked Potentials” (SSVEP) that is, a periodic visual
response to any kind of sensori-motor event, external stimulus or stimulus, in the frequency range of 4 to 60 Hz, can induce en-
mental task(Luck, 2014). Besides, the ERP is time synchronized to hanced stationary periodic oscillations with the same frequency
the event and lasts to one second. In comparison to other non- and its higher harmonics in the brain waves recorded over vi-
biometric authentication methods, ERPs can be categorized into sual areas(Phothisonothai, 2015; Piciucco et al., 2017). This stim-
“challenge-based response” methods in which user is asked to re- ulus may be generated by a panel of flickering LEDs. Rate of in-
spond to a system command(Idrus et al., 2013). The other factor creasing in the power spectrum of brain frequencies is not only
that should be considered in the EEG acquisition protocol is that, corresponding with stimulation frequency, intensity and duty cy-
which part of the brain engages more to execute the desired task. cle of flickering, but also varies in different individuals. Despite the
For instance, ERP responses to the visual stimuli are mostly ob- simplicity, SSVEPs are useful tools because of the excellent SNR and
served in the occipital lobe, visual cortex, and central region. As a robustness against artifacts.
result, with some compromising, the electrodes can be reduced to The drawback of all VEPs is the need for external devices to cre-
Oz, O1, and O2. In the following, the protocols mainly used in the ate the stimuli and the time synchronization between the stimuli
EEG authentication are described. and the EEG recording. However, it is shown that VEP components
are very stable over time(Gaspar et al., 2011) and they can satisfy
5.1. Resting states the permanency condition of biometric authentications. Besides, in
SSVEP, there is no need to synchronize recording.
Resting states are the most popular protocol in acquiring EEG
signals(Di et al., 2019; Kim and Kim, 2019; La Rocca et al., 2014;
Nakamura et al., 2017; Poulos et al., 1998; Schons et al., 2018; Waili 5.3. Acoustic stimuli
et al., 2019; Wang and Najafizadeh, 2016). In this protocol, the in-
dividual is asked to be completely relaxed, normally sit on a chair, Acoustic Stimulus is another type of ERP that is emerged by lis-
in a quiet environment and then EEG signal is recorded. Although tening to a piece of music or a special tone; though, it is not as
both closed eyes, namely “REC”, and opened eyes, namely “REO”, common as VEP. In Kaur et al. (2017), four different genres of mu-
are used in this protocol, there are differences in the dominant fre- sic were played for individuals in which they induce different emo-
quency band and the more efficient EEG channels. In REO, the best tions and interests. The individuals were also asked to provide mu-
results are achieved from the central region, while with REC, the sic preference which acts as a personal identification mechanism.
parietal region does. Besides, α band is dominant band(Barry et al., In Frank et al. (2017), individuals were asked to determine three of
2007; La Rocca et al., 2014) during REC. The popularity of this pro- their favorite songs. Then, these songs were used to create clips to
tocol lies in its simplicity. Besides, it does not require any extra represent an audio pass-thought.
requirements or instructions, however, the environment should be
quiet and the individual should have no mental activity or preoccu- 5.4. Mental task
pation; otherwise, the results are deflected. Nevertheless, the rest-
ing state can be used as a baseline EEG activity for the other tasks In the mental tasks, the individual is asked to imagine some
in the preprocessing stage. physical body movements or do mental activities. For instance, in
Marcel and Millan (2007), Hu (2010), Chuang et al. (2013) and
5.2. Visual stimuli Das et al. (2018), the individual should move hand or foot imagi-
narily or do both imaginary and physical body movement(Alyasseri
Visual stimuli, which are also known as “Visual Evoked Poten- et al., 2018; Sharma and Vaish, 2016). It is shown that imagi-
tial” (VEP), is a group of ERPs that are caused by external visual nary tasks lead to better results than similar physical ones. In
stimuli. VEPs, in turn, cover a wide range of visual stimuli. In Ruiz- Chuang et al. (2013) and Johnson et al. (2014), the individual was
blondet et al. (2014) and Gui et al. (2015), the individual is asked asked to silently sing a favorite song, counting numbers in mind
to silently read some unconnected texts including words (e.g. BAG, and concentrate on a desired thought as a “pass-thought”. Doing
FISH), pseudo-words (e.g. MOG, TRAT), acronyms (e.g., MTV, TNT), mental multiplication, visualizing the rotation of a given geometric
illegal strings (e.g., BPW, PPS) or instances of their own names. In figure around an axis, counting numbers in mind and composing a
Zuquete et al. (2010) and Abo-Zahhad et al. (2016), a sequence of letter to a friend are the other instances of mental tasks considered
pictures, which each one is followed by a blank screen, are dis- in Kumari and Vaish (2016). Imagine a person, a gesture or taste of
played on a screen for 1–2 s and the individuals are asked to rec- a food(Frank et al., 2017) and generation of words beginning with
ognize the pictures as soon as they are presented. In some cases, randomly selected letter(Marcel and Millan, 2007) are other sets of
individuals are additionally asked to respond to a rare target pic- mental tasks that individuals are asked to perform.
ture, by concentrating on or pressing a button, among a large se- In comparison to the visual stimuli, although mental tasks do
quence of non-target pictures. The target picture can be a spe- not require special devices for stimulation, they still need some
cial geometric shape(Das et al., 2015; 2016b), numbers among all simple equipment to generate a clue for the individual to initial-
characters(Das et al., 2016a; 2016b), moving objects versus sta- ize the task.
tionary objects(Zhang et al., 2018), or a picture already selected
by own individual as a password(Liew et al., 2018). “Rapid Serial
Visual Presentation” (RSVP) is another protocol in which the se- 5.5. Multi-Tasks
quence of infrequent target/frequent non-target pictures are shown
sequentially with a high rate, e.g., 2–10 Hz, to minimize the train- There are some protocols that the EEG is recorded with more
ing duration (Chen et al., 2016; Zeng et al., 2019; Zhang et al., than one type of stimulus. For instance, individuals are asked
2018). RSVP, in particular, elicits P300 ERP, that is a special positive to watch short music videos that induce different emotional
states(Chen et al., 2019; Pham et al., 2015; Vahid and Arbabi, Laplacian filter is used as a spatial high pass filter that enhances
2016). The fusion of both EEG and EOG signals is another multi- localized activities while suppressing the diffusion ones. Laplacian
model protocol that is used to improve the accuracy of classifica- is calculated by subtracting the sum of weighted potential of the
tion(Bhateja et al., 2019). neighborhood electrodes from the current electrode potential as
the following(McFarland et al., 1997):
6. Preprocessing vrj
l r j∈Si di j
Similar to any other classifications, the EEG authentication re- v =v −
i i 1
, (3)
j∈Si di j
quires some preprocessing on the raw EEG signals to enhance
the quality. Preprocessing methods can be categorized into three where Si is set of neighborhood electrodes of ith electrode and dij
groups described in the following. is the distance between the electrodes of i and j.
6.1. Frequency domain filtering

6.3. Time domain filtering
As mentioned in Section 2, the brain wave frequencies range
from 0.5 to 40Hz. As a result, a band-pass filter is usually ap- We categorize some preprocessing methods that are applied di-
plied to the raw signals to filter the lower and higher frequen- rectly on the time series of EEG, as the time domain filters; though,
cies. To aim this, “Butterworth”, as an anti-aliasing infinite impulse they may have common characteristics with other domains. In the
response filter, is the most common band-pass filter used in the following, these preprocessing methods are briefly introduced.
EEG preprocessing(Alyasseri et al., 2018; Frank et al., 2017; Singh Ensemble averaging is a simple, but effective, technique that is
et al., 2015; Thomas et al., 2017). “Chebyshev” filter is the other applied to EEG to not only reduce the noise measurements but
common filter used in the literature(Chen et al., 2016; Zeng et al., also to enhance ERP signals. In general, averaging of multiple tri-
2019). In comparison to Butterworth, Chebyshev filter has sharper als, recorded with the same protocol, reduces the noise power. In
cut-off frequency with more ripples. It’s worth mentioning that, particular, if EEG is recorded after a stimulus presentation, averag-
Butterworth is also used to separate different EEG bands in the ing of multiple time-locked ERP responses attenuates the baseline
feature extraction stage. In some cases, that one desires to keep brain activities whereas the peaks of ERPs are amplified. This is a
higher frequencies, a “Notch” filter, with stopband frequency of 50 common preprocess to increase the SNR in the ERP-based authen-
or 60Hz, is applied to the signals to remove the power line arti- tications(Das et al., 2016b; Frank et al., 2017; Gui et al., 2015). An-
facts(Abdullah et al., 2010a; Delpozo-Banos et al., 2015); however, other method to enhance ERPs is baseline removal. Here, the EEG
it may add some phase distortion. signals are initially measured in the rest state, before presenting
stimulus. Then, the mean of this measurement is used to remove
6.2. Spatial domain filtering baseline activities from the ERPs in either time domain(Chen et al.,
2019) or frequency domain(Koelstra et al., 2012). Removing the
When sufficiently large numbers of electrodes are employed, baseline by considering EEG by its own is also common in the
spatial filters can be used to reduce noise, detect and remove body literature. In this type, the EEG is normalized to a dimensionless
artifacts, and enhance localized activities. The most common spa- zero-mean signal with variance one(Das et al., 2016b; Ruiz-Blondet
tial filters used in the EEG biometric are “Common Average Refer- et al., 2016; Zhang et al., 2018). Thresholding is another preprocess
encing” (CAR)(Das et al., 2016a; 2015; Maiorana and Campisi, 2018; method in which spikes above a certain threshold, usually 50μV,
Maiorana et al., 2016b; Piciucco et al., 2017), “Independent Compo- are bypassed. Thresholding is used to remove eye-blinking or mus-
nent Analysis” (ICA) (Arnau-Gonzalez et al., 2017; Maiorana et al., cle reflex artifacts (Frank et al., 2017; Sundararajan et al., 2015).
2016b; Tangkraingkij et al., 2010; Wang and Najafizadeh, 2016),
and Laplacian filter (Kostílek and Št’astnỳ, 2012; Marcel and Mil-
lan, 2007). 7. Shallow classification
In CAR, the measured potential at each electrode is subjected to
the mean of the entire electrodes (McFarland et al., 1997): Shallow methods can be considered as two-level processing;
extracting some distinguishing features and then applying a clas-
1 i
M
sification method on the extracted features. However, by consid-
c u = vu − v, (1)
M ering the type of signal, the features and classifiers can be com-
i=1
pletely different in various applications. In the EEG biometric, the
where vu is the raw measured potential of electrode u and M is the input signals are the time series of electrical potential fluctuations
number of all electrodes. It makes the signal robust to inappropri- recorded from different locations of brain. As a result, features that
ate reference choices in mono-polar recordings or unexpected ref- can be extracted in any of time domain, frequency domain or spa-
erence variations tial domains, are mainly useful; e.g., coefficients of Auto Regres-
ICA is one of the blind source separation techniques in which sive Model in the time domain, Power Spectral Density in the fre-
components identified as eye movements/blinking, background quency domain, Wavelet coefficients in the time-frequency domain
noise, and muscle artifacts can be removed. Indeed, the raw EEG or common spatial coherence in the spatial-frequency domain. Sim-
data is considered as a linear combination of multiple sources, in- ilarly, the applied methods should be able to classify these types
cluding both brain and non-brain activities and external noises, of features. In the following, the most common feature extraction
and is modeled as follows(Maiorana et al., 2016b): and classification methods in the EEG authentication are reviewed.
It is worth mentioning that in any level, the available methods
s = W x, (2)
can be utilized singly or fused together to improve the results.
where x is the measured vector, s is separated sources and W is In the feature-level, the features extracted from different methods
the transformation matrix. “Canonical Correlation Analysis”(Koike- are fused together directly or by means of some feature reduction
Akino et al., 2016) and “Empirical Mode Decomposition” (EMD) tools such as Canonical Correlation analysis. In the decision-level,
(Kumari and Vaish, 2016) are the other source separation tech- the matching score of different classification methods are fused to-
niques. gether to enhance confidence level.
7.1. Feature extraction
Accuracy (%)
“Autoregressive Model” (AR) is one of the most common meth-
99.87
92.9
99.8
ods to extract features in the EEG authentication. In AR model, a
100
100
88
97
90
95
98
95
78
time series is modeled as a weighted summation of previous Q
samples(La Rocca et al., 2014):
Fusion of Manhattan, Euclidean and cosine distances

Q
x[n] = − aq x[n − q] + w[n], (4)
Fusion of Euclidean distance, SVM and LDA.

q=1
where w[n] is the white noise and {aq } is set of AR coefficients.

Consequently, by considering the EEG signal as a dynamic time
series, the measured potential activity of each channel can be
modeled by an AR. Then, set of coefficients of all channels is
Guassian Mixture Model
used as the feature vector(Abo-Zahhad et al., 2015; Bhateja et al.,

2019; Keshishzadeh et al., 2016; La Rocca et al., 2014; Maiorana
Euclidean distance
cross correlation
cross correlation
and Campisi, 2018; Maiorana et al., 2016a; Pham et al., 2014;

2015). AR coefficients can be classified by “Support Vector Ma-
HMM/SVM
chine” (SVM) (Keshishzadeh et al., 2016), “Artificial Neural Net-

Method
works” (ANN) (Bhateja et al., 2019), “Linear Decremental Analy-

HMM
ANN
ANN
ANN
LDA
sis” (LDA) (Abo-Zahhad et al., 2015) or similarity based classifiers

(Maiorana et al., 2016a). AR extracts EEG features in the time do-
AR, MFCC, bump representation
main.
In some cases, the time series of EEG signals are directly used
AR, PSD, Spectral coherence
to generate a “template”, rather than modeling by AR. The template

Gamma band Wavelet
contains a short period of the processed signal of one or multiple

channels(Chuang et al., 2013; Das et al., 2016a; Gui et al., 2015;
Kumar et al., 2017; Ruiz-Blondet et al., 2016). The processing in-
cludes averaging over trials or filtering frequency bands. Although
Template
Template
template
Features
Wavelet
Wavelet
template generation is significantly less complex than AR model-

wavelet
ing, it requires more storage space and consequently, it may not be

PSD
PSD
AR
practical on a large scale dataset or people. Template can be classi-

fied by “Hidden Markov Model” (HMM) (Kumar et al., 2017), cosine
Number of channels
distance (Das et al., 2016a) or cross correlation(Ruiz-Blondet et al.,

2016). Obviously, the template represents the EEGs in the time do-
main. Another simple, but compact, representation of time domain
is the statical characteristics of the time series; e.g., mean, stan-
dard deviation or RMS of EEG signal per channel(Kaur et al., 2017;
26
19
26
14
19
20
8
6
1
1
4
Kumar et al., 2017).

“Power Spectral Density” (PSD) is another common feature de-
Number of subjects
scribing the EEG signals in the frequency domain. Indeed, PSD is

the power distribution of different frequencies or different bands.
PSD is usually estimated by means of Discrete Fourier Transform
and Welch Method(Di et al., 2019) or Bartlett Method(Pham et al.,
2014). Here, the feature vector is the augmentation of PSD of
32
37
17
10
10
50
50
20
60
45
5
all selected channels(Ashby et al., 2011; Di et al., 2019; Maio-

rana et al., 2016a; Pham et al., 2015) and it can be classified by
Imaginary/ Physical movement
REC, REO, mental computation
SVM(Ashby et al., 2011), similarity-based methods(Maiorana et al.,

2016a) or LDA(Di et al., 2019).
Regarding the non-stationary nature of EEG signals, “Wavelet
Shallow classification works on the EEG Authentication.
Transform” (WT) is also widely used as a powerful feature extrac-

Imaginary Task
tor in both time-frequency domains. WT decomposes a signal to a

Acoustic ERP
Eye blinking
weighted summation of functions, namely wavelet, that each one

REO, REC
REO,REC
REO,REC
is a shifted and scaled copy of a basic function, namely mother

Task
wavelet. Each wavelet approximates the signal with a lower res-

VEP
VEP
VEP
VEP
olution. By using Wavelet decomposition, the set of wavelets’ coef-

ficients per channel is defined as the feature vector(Alyasseri et al.,
Maiorana and Campisi (2018)
2018; Kaur et al., 2017; Sharma and Vaish, 2016) and it can be
Ruiz-Blondet et al. (2016)
Ruiz-Blondet et al. (2017)

Marcel and Millan (2007)
Sharma and Vaish (2016)

Abo-Zahhad et al. (2015)
classified by HMM(Kaur et al., 2017), ANN(Alyasseri et al., 2018) or

Maiorana et al. (2016a)
SVM(Kaur and Singh, 2017). In rather cases, the statical properties

Waili et al. (2019)
Kaur et al. (2017)
of wavelets’ coefficients are also used as features; e.g. mean, stan-

Gui et al. (2014)
Gui et al. (2015)
Di et al. (2019)
dard deviation, RMS and entropy(Abdullah et al., 2010b; Kaur and

Singh, 2017).
The other less common features introduced in the literature are
Study
Table 2
“mel-frequency cepstrum coefficients” (MFCCs) (Maiorana and Camp-

isi, 2018; Piciucco et al., 2017) in the frequency domain, “sort time
Fourier transform (STFT) (Phothisonothai, 2015) in the time-spatial Model” (GMM)(Marcel and Millan, 2007) are the other classification
domain and “common spatial coherence”(Singh et al., 2015) in the methods that are used in the EEG authentication.
spatial-frequency domain. Similar to the feature extraction step, again, the classification
Finally it’s worth mentioning that in many studies, multiple results of different methods can also be fused to increase confi-
methods are used simultaneously to extract features in different dence level(Abo-Zahhad et al., 2016; Kaur and Singh, 2017; Maio-
domains; e.g. combination of AR and PSD in Ashby et al. (2011), AR rana et al., 2016a).
and statical properties in Kumar et al. (2017), AR, PSD, entropy and
mutual information in Hu (2010), entropy, mutual information and 8. Deep classification
statical properties in Kumari and Vaish (2016), and AR and MFCC in
Maiorana and Campisi (2018). In these cases, the different type of With the rapidly increasing interest and development of deep
features are fused directly or processed by some features’ reduc- learning, many fields of computer science have evolved. The deep
tion tools such as “Principle Component Analysis” (PCA)(Kaur and learning methods are capable of feature extraction from the raw
Singh, 2017; Palaniappan, 2006) or “Canonical Correlation Analysis” data. Along with other areas, deep learning has been gained sig-
(CCA) (Bhateja et al., 2019; Kumari and Vaish, 2016). nificant interest in many various aspects of EEG processing such as
motor imagery classification (Tabar and Halici, 2016), EEG decod-
7.2. Classification methods ing and visualization (Schirrmeister et al., 2017), EEG-based emo-
tion recognition (Yin et al., 2017), and last but not the least EEG-
Similarity-based methods are the simplest and the least com- based authentication. Deep learning approaches have significantly
plex classification approach in which two vectors are compared increased the accuracy of EEG biometric authentication systems to
and labeled as the same class if their similarity score is greater over 99% even for a set of over 100 people (Sun et al., 2019).
than a pre-defined threshold. Indeed, a feature vector generated Most of the deep EEG biometric methods proposed in the pa-
in the enrollment phase is compared with the feature vectors pers are on the bases of “Convolutional Neural Networks” (CNN)
generated in the registration phase. Besides, it’s mainly used in models. Das et al. (2018), for example, showed that CNN mod-
the authentication, rather than identification. The more common els are able to reach 99.3% accuracy in the EGG user identifi-
similarity-based methods include “Euclidean distance”(Fraschini cation. They used a CNN model with four convolutional layers
et al., 2015; Gui et al., 2015), “cross correlation”(Ruiz-Blondet et al., and two max-pooling layers. Wu et al. (2018) similarly, reached
2016; Thomas et al., 2017), “cosine distance”(Das et al., 2015), “Man- the precision of 97% for EGG collected during a 30-day inter-
hattan distance”(Maiorana et al., 2016a), and “Dynamic Time Wrap- val (which is one of the EEG biometric challenges). To over-
ping” (Das et al., 2016b). These methods can be used to directly come the need of deep learning approaches to a huge amount
compare two templates(Das et al., 2015; Ruiz-Blondet et al., 2016) of data, Schons et al. (2018) used data augmentation. By creating
or the features extracted from different domains(Maiorana et al., overlapped time windows, they were able to increase the train-
2016a; Piciucco et al., 2017). ing samples to several hundred times. Their proposed method
“Support Vector Machine” is a linear classifier that defines a significantly decreased EER compared with the previous works
hyper-plane to separate features of different classes, mostly bi- (Fraschini et al., 2015; Yang et al., 2018). Using the adversarial in-
nary ones, in such a way that maximizes the distances of each variant representation learning approach, Zhang et al. (2018) pro-
class from the hyper-plane. Regarding the binary nature of the posed an adversarial convolutional network for the EEG biomet-
authentication, SVM is widely used in the literature as both lin- ric, which can distinguish individual characteristics from the sig-
ear (Ashby et al., 2011; Di et al., 2019; Kaur and Singh, 2017; nals recorded during different sessions. In their proposed model,
Keshishzadeh et al., 2016; Kumar et al., 2017) and nonlinear clas- a CNN with four convolutional layers for representing data and
sifier(Vahid and Arbabi, 2016). SVM can be used to classify fea- two fully connected networks as identifier and adversary were
tures extracted from the time domain(Kumar et al., 2017) or the used. Chen et al. (2019) proposed a new convolutional neural net-
frequency domain(Di et al., 2019). work with global spatial and local temporal filters called GSLT-
“Linear Decremental Analysis” (LDA) is another linear classifier CNN. These works are just some instances to name, and there
that can reduce feature dimensions as well by accounting class la- exist other similar works in the literature. For example, Lan Ma
bels. Therefore, LDA is used where there are so many features with et al. (2015), Schons et al. (2018), and Wu et al. (2018) proposed
possibility of correlation(Di et al., 2019; Jayarathne et al., 2016) CNN models, respectively, with two, three, and two different con-
or the features are extracted from different sources(Abo-Zahhad volutional layers in their works.
et al., 2015; 2016). LDA can be used to classify AR features(Abo- The important point in using CNN models is how to model
Zahhad et al., 2015) or PSD features(Di et al., 2019). the input data. In most cases, the input is considered as a two-
“Hidden Markov model” (HMM) is widely used to model signals dimensional matrix, one dimension is the sampling channels, and
with a temporal dynamic nature, like as EEG. The HMM can be the other dimension is samples collected per channel. Some meth-
simply defined as a finite set of hidden states with transition prob- ods, such as Wu et al. (2018), have chosen a different approach. He
abilities from each state to another one. It is used to directly model defined the input as the matrix of C × F where C is the channels
time series(Kumar et al., 2017) or to classify features extracted in and F is the voltages averaging in non-overlapping time-windows.
the time-frequency domains(Kaur et al., 2017; Maiorana and Camp- Convolutional neural networks have very good performance
isi, 2018). with static data, such as photos or a part of the signal, but
“Artificial Neural Network” (ANN) is a nonlinear classifier that are not designed to extract information and time features in
tries to mimic the structures of the human neural system. ANN the time series such as the EEG signals. Hence, various works
contains a hidden layer with a nonlinear characteristic that lets it have been done to use other deep learning methods including
model nonlinear data with more degrees of freedom. Therefore, it “Recursive Neural Networks” (RNNs) models for the EEG biomet-
can be used not only for the authentication, but also for the classi- ric authentication. RNNs and their family, “Long Short-Term Mem-
fication(Abdullah et al., 2010a; Alyasseri et al., 2018; Bhateja et al., ory” (LSTM) and “Gated Recurrent Unit” (GRU), are good tools for
2019; Hu, 2010; Sharma and Vaish, 2016; Waili et al., 2019). extracting time attributes in sequences. Sun et al. (2019) pro-
“K-Nearest Neighbors” (KNN) (Chuang et al., 2013; Singh et al., posed a new neural network based approach combining CNNs
2015), “Support Vector Data Description” (SVDD) (Pham et al., 2015), and LSTMs, called ”1D-Convolutional Long Short-Term Memory” (1D-
“Random Forest” (RF)(Kaur and Singh, 2017), and “Gaussian Mixture Convolutional-LSTM) for the EEG authentication. This model is
more accurate than either CNN or LSTM and is also able to main-
tain its performance even with the reduction of channels. The au-
thors showed that 1D-Convolutional-LSTM with 16 channels gives
0.71% higher accuracy than the CNN system with 64 channels.
Accuracy (%)
Wilaiprasitporn et al. (2019) have also proposed a combination of
CNN and RNN networks. Moreover, they reviewed two types of
99.62
99.30
99.17
97.60
99.58
99.30
89
96
97
94
88
RNN networks, including LSTMs and GRUs. Compared to LSTMs,
GRUs are faster to train and require fewer data to generalize, even
Augmentation
though they have comparable performance and accuracy. The ac-
curacy obtained in this paper for the CNN-GRU and CNN-LSTM is
99.17% and 98.23%, respectively, on the DEAP dataset.
Yes
The works on the EEG authentication based on the deep learn-
No
No
No
No
No
No
No
No
No
No
ing approaches are summarized in Table 3. It should be mentioned
FC LSTM/GRU
that despite the excellent performance of deep learning approaches
in the EEG biometric, these methods are not free of problems.
For example, Ozdenizci Özdenizci et al. (2019) showed that these
methods do not yet have acceptable performance across different
FC
FC
FC
FC
FC
FC
FC
FC
FC
FC
sessions.
2
2
3
4
1
2
2
3
1
1
5
conv.
conv.
conv.
conv.
conv.
conv.
conv.
conv.
conv.
conv.
conv.
Layers
9. Security and privacy of the EEG biometric
4
3
2
3
2
3
3
5
2
4
4
CNN+LSTM +GRU
Although EEG based authentication is perceived as a leading
technique, it still faces several security and privacy challenges. An
CNN+LSTM
Base model
EEG biometric can be broken down by an adversary if attacker
could generate an input EEG signal which its features are close
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN
enough to a target user. The attacker has three different ways of
doing this: 1) imitate the victim’s brain, 2) brute force the system,
3) sniffing the victim’s EEG signals and try to reproduce it. Imitat-
Sample duration
ing the victim’s brain means that the attacker somehow becomes
aware of the victim’s thoughts during the registration phase (for
example, in the task of imagining the body movement, he knows
0.5 s
0.6 s
12 s
10 s
which part of the body, the user imagines to move) and tries to
3s
1s
1s
1s
6s
1s
1s
imitate those thoughts to fool the system and bypass the authenti-
cation mechanism. Johnson et al. (2014) had shown that the exis-
No. of channels
tence of an imitating attacker in an environment increases the false

acceptance rate of the system from 2% to 5%. The error rate might
still seem low enough at the first glance, but considering the very
16
16
64
64
28
64
14
64
17
16
limited population size of the study (only 15 subjects), this arises
5
a warning that error may increase unacceptably in larger popula-

No. of subjects
tions. In another study, it was shown that in the authenticating

by the imaginary sport task, the system error would increase from
0% to 17%, if an imitating attacker exists in the environment who
157
109
100
109
is aware of the users’ selected sport. This research clearly demon-

32
15
15
23
10
10
40
strates the sensitivity of the imitation attacks’ success rate to the

chosen task.
Imaginary/Physical movement
Imaginary Body movement
Although the studies have not yet proved the high probabil-
ity of imitation attacks succeeding, by considering the very few
studies investigating the imitating attacks (including Johnson et al.,
Deep classification works on the EEG authentication.
2014; Sohankar et al., 2015), it is still too early to claim the robust-
Virtual Driving
ness of EEG biometric against these type of attacks. In the current

situation, solutions such as locking the user account after several
REC, REO
REC/REO
failed attempts are sufficient to prevent the imitating attacker.

RSVP
RSVP
RSVP
RSVP
Task
VEP
VEP
The second method that an adversary may use to break the sys-
tem, is the brute-force attack which is testing every possible EEG
as the input with the hope of eventually finding the correct one.
Arnau-Gonzalez et al. (2017)
Wilaiprasitporn et al. (2019)
The probability of an authentication factor breaking through the

Özdenizci et al. (2019)
brute-force is usually measured by the “entropy” metric. Suppose

Lan Ma et al. (2015)
Schons et al. (2018)
Zhang et al. (2018)
the adversary has a method to generate as many fake EEG sig-

Chen et al. (2019)
Mao et al. (2017)
Sun et al. (2019)

Das et al. (2018)
Wu et al. (2018)
nals as required, the entropy of EEG biometric is defined as the

binary logarithm (log2 n) of the average number of the required
EEG generated by the adversary to break the authentication sys-
Study
Table 3
tem. Sadeghi et al. (2017) showed that EEG biometric entropy is

at best 83-bits. This entropy is obtained extracting features using
PSD and classifying by a Naive Bayes Classifier. The entropy of other
methods was significantly lower in their study. For example, the
entropy of AR feature extractor and SVM classifier were estimated
Table 4 advent of recent methods such as GANs in synthetic data genera-

Entropy of different biometrics.
tion has made similarity check alone not enough to prevent these
Biometric Study Entropy (bits) attacks. Piplani et al. (2018) showed that the EEG based authenti-
Finger vein Krivokuća et al. (2020) 4.2–19.5 cation system might be easily tricked into accepting the synthetic
Retina Arakala et al. (2009) 16.7 data generated by a GAN model. These weaknesses in EEG biomet-
Voice K. Inthavisas (2012) 18–30 ric require serious consideration and proposing defensive methods
Iris Hao et al. (2006) 44 to strengthen authentication.
Fingerprint Li et al. (2012) 48
Given the weakness of the liveness and similarity testing meth-
Face Feng and Yuen (2012) 75
Iris Kanade et al. (2009) 94 ods, it seems that the only way to prevent these attacks is to pre-
Gait Hoang et al. (2015) 50–139 vent intruders from sniffing the users’ EEG signals by approaches
EEG Sadeghi et al. (2017) 83 such as encryption transmitting channels and EEG records. How-
EEG Bajwa and Dantu (2016) 82
ever, reviews of the apps on the market show that they do not
only meet these security recommendations, but also the program
reveals many security vulnerabilities, even more than those dis-
to be only 36 bits. In another similar research (Nguyen et al., 2019), cussed above. Xiao et al. (2019) explored the possibility of 5 at-
the EEG entropy is approximated by 82 bits based on 18 selected tacks on the apps in NeuroSky App store. The attacks include the
channels. To have a better view of the superiority of EEG biomet- sniffing and recording the signals emitted from the victim’s head-
ric over some well-known biometric, the entropy of some common set device. The analysis showed that from 156 apps publicly avail-
biometrics are summarized in Table 4. It is worth mentioning that able in the store, all are vulnerable to this attack. The apps trans-
the entropy of human chosen passwords is 20–22 bits(Wang et al., mit the EGG signals through the “over-the-air” (OTA) transmission
2017). The studies on EEG entrophy considered only shallow clas- protocol between the brainwave headset and an RF dongle which
sification methods. Given that recent works have paid much more can be sniffed by a proximate adversary. Moreover, from 31 free
attention to deep learning approaches, the claims of these studies apps on the store, all are vulnerable to at least one type of remote
need to be further explored in the future research considering the attacks.
deep learning methods. Sniffing and accessing the users’ recorded EEGs do not only
Besides the entropy of the EEG biometric, there are two other threaten system security but can also violate user privacy. EEG
points considering with brute-forcing EEG biometric: 1) how an records, especially in resting states, reveal much information about
adversary can generate artificial EEG signals, 2) how the system the user. The current research shows that EEG can be used to pre-
can differentiate artificial machine generated signals and human dict viewed images, transfer brain waves to texts, monitor sleep,
recorded signals. Generating artificial EEG is a subclass of artificial and reveal personal information such as sex, age, and user’s ill-
signal generation. Traditionally, it can be achieved by some meth- nesses or addictions (Bickford, 1987; Gondesen et al., 2019; Höller
ods such as autoregression, or statistical simulation or physiologi- and Uhl, 2018; Xiao et al., 2019). It can even be used to inference
cal modeling of EEG, however, they often demonstrate poor gener- the user’s current emotion (Aggarwal et al., 2018) or activities such
alization performance (Nik Aznan et al., 2019). With the develop- as reading, blinking, solving math problems, listening to music, re-
ment of machine learning methods and in particular deep learn- laxing, thinking of an item, and watching videos (Ni et al., 2017;
ing approaches in recent years, special attention has been paid to Xiao et al., 2019). Given these challenges, prior to the commercial
the use of these methods for artificial data generation. The most use of EEG biometric, a solution must be found to prevent the dis-
important of these methods are “Generative Adversarial Networks” closure of personal information through the brain’s signals while
(GANs) (Creswell et al., 2018). GANs provide a way to generate allowing the user to be authenticated. This issue will be discussed
new data with the same statistics as the training set. A variety of further in the next section as one of the challenges.
works have been done in recent years to produce fake EEG signals
with GAN (for instance, (Hartmann et al., 2018; Nik Aznan et al., 10. Open challenges
2019; Panwar et al., 2019; Piplani et al., 2018)) that have shown
a very good performance. These tools have made it easier for the Many researchers have used EEG as a biometric authentication
adversary to attack EEG based authentication systems. In return, method so far, though these methods are often implemented only
the authentication systems should provide a way to distinguish in the controlled environment, and there are still several chal-
machine-made and human-generated signals. The solution is the lenges ahead of their applications in real-life scenarios (Yang and
liveness test method which ensures that input data is originating Deravi, 2017). For instance, the biometric authentication factor
from a live human being, and it is not artificial data. Unfortunately, requirements such as universality and uniqueness are still not
the EEG signals are believed to have an intrinsic liveness property fully satisfied. This section reviews these challenges from different
due to the point that they can only be obtained when a live hu- points of view including biometric factor requirements (i.e. uni-
man subject is present. Clearly, given the success of recent meth- versality, permanency, uniqueness, and collectability), user-friendly
ods such as GAN for generating fake EEG signals, this assumption and user privacy.
is not realistic (Sohankar et al., 2017). Some liveness test meth-
ods such as reflection or response to an external stimulus may be 10.1. Universality
applicable here, but as far as we know there is no liveness test
designed specifically for the specification of EEG signals. A universal EGG based authentication system should be capa-
Last but not the least attacking method, is attempting to reuse ble of being used for almost all people. Current research have
a sniffed target’s EEG signals to log in the system as the target. studied either a limited set of individuals with predefined as-
Replay attack by using the exact sniffed EEG can be easily pre- sumptions or a publicly available dataset that is believed to be
vented by rejecting a duplicated signal since it is not plausible that biased (Gondesen et al., 2019). The subjects considered in these
the recorded biometric signals of a person are exactly the same works were all healthy and almost young. The evaluation of ex-
at two different sessions. Even adding white noise to the sniffed isting unhealthy or affected subjects in the EEG authentication
signal can be detected using a similarity check in which each sig- systems is still an important and open discussion (Del Pozo-
nal should have a threshold of difference with the previous ones Banos et al., 2014). Currently, there exist few works which have
(Roberts, 2007; Sadeghi et al., 2017; Sohankar et al., 2015). But the studied a dataset including healthy and alcoholic users (e.g.
Nguyen et al., 2012), but as far as we know, these works arenot

only very limited, but also has not gone beyond alcoholic users.
Hence, it can be argued that the EEG biometric has not met uni-
versality, yet.
Another issue that must be considered for universality is the
use of feature extractor models instead of classifiers. In all re-
search, a classifier is learned on the recorded EEG of a set of sub-
jects, and then this model is used to distinguish between those
subjects. This approach is not suited to have a global EEG biometric
model since in this approach, at least one EEG sample of involved
subjects is needed to train the model, and when a new person is
added to the system, the whole model should be re-trained from
scratch. In deep learning approaches, a deep model can act as a
feature extractor instead of a classifier. A feature extractor model
can be used to extract the feature vector of any subject (even
those who are not visited in the learning phase), and then the sub-
ject is identified to the corresponding class by vector similarity-
based algorithms (e.g. cosine similarity). Among the papers, only
Schons et al. (2018) have tried to propose an EEG feature extractor
model using a CNN, although they did not employ the feature vec-
tors directly, and used it as the input of a fully connected neural
network which ultimately makes their approach similar to a classi-
fication model. The authors believe that the existence of a feature
extraction model that does not need to be re-trained for each per-
son’s entry is essential for the universal use of EEG biometric. Fig. 6. Distribution of participant size in the EEG based authentication research.
Another issue should be considered in the universality, is lan-
guage dependencies in ERPs with linguistic stimuli, e.g. visualized
a word of play a song. Up to the best knowledge of the authors,
there is no research directly investigating the effects of different of individuals. However, current studies often consider a very lim-
languages on the EEG authentication. However, there are clues that ited and small set of populations between 4 and 50 subjects which
processing another language, especially in low proficiency users, cannot prove the uniqueness of the EEG biometric in practice. The
engages different and more parts of the brain and with higher co- distribution of participant size in the literature is represented in
herence than the first language. In adults, the left hemisphere is Fig. 6. As it is represented, more than 80% of works are on less
mostly involved in the native language processing, whereas in pro- than 50 subjects, and just a few works have studied their pro-
cessing the second/foreign language, the right one is. Besides, it is posed methods for more than 100 subjects. Also, it has been ar-
shown that in the users with higher proficiency in the second lan- gued that the identification ability of EEG biometric would be de-
guage, linguistic processing is shifted to the left hemisphere and creased by increasing the number of participants (Boubakeur et al.,
with a lower coherence in both hemispheres(Reiterer et al., 2009; 2017; Tangkraingkij et al., 2010). Tangkraingkij et al. (2010) showed
2011). Furthermore, it seems that harder language leads to more that the accuracy of the system may even be reduced up to 9% only
coherence(Cheung et al., 2009). As a result, the authors believe by increasing participants by ten. We believe that EEG biometric
that if words or song are going to be used in the ERPs, it should cannot be yet claimed to be a universally unique biometric factor
be in the native language; otherwise, EEG pattern over scalp may such as fingerprint.
change by aging or improving proficiency.
10.2. Permanency
10.4. Collectability
Early studies on the EEG biometric authentication have only
been performed on recorded data from a series of consecutive ses- From the collectibility point of view, many of the challenges
sions, thus this question has still not been well answered whether have been solved; however, this ability has not yet reached ma-
the patterns of the EEG change over the long term. The perma- turity. Current works generally require specific conditions during
nency of EEG biometric has been challenged just in recent papers recording EEG and it should be done in a relaxed and coopera-
(La Rocca et al., 2014; 2014; Maiorana and Campisi, 2018; Maiorana tive manner. For example, in the sampling of the resting state with
et al., 2016a). Maiorana et al. (2016a) has argued that there are closed eyes, any sound in the environment will disrupt the func-
very few studies in which the data were acquired during different tioning of the system. These conditions are not available in the
sessions on different days, and for those who have been record- uncontrollable environment. As a result, the use of EEG biometric
ing data over different days, often the distance between sessions authentication systems in practice is still with challenges.
were only a few weeks. He later in his recent work (Maiorana and On the other hand, the channels required for sampling should
Campisi, 2018) examined the EEG records which have been ac- be reduced to the extent possible. The number of channels re-
quired during a three-year period. This paper is the best on the quired in the papers is about 20 on average, which is more than
permanency of EEG biometric, however, this field is still in the the number of electrodes in most commercially available devices.
early stage and needs further research. Just a 17% of works have used less than 5 channels. This chal-
lenge is particularly true in the deep learning approaches, which
10.3. Uniqueness currently have the best possible accuracy. The number of channels
required in these works is 33 channels on average, which seems
In order to ensure that the biometrics are unique, it must be to be difficult to be used in practice. Hence, more research is yet
shown that they can uniquely identify an individual in a large set required to be made to reduce the number of channels.
10.5. User privacy records should be protected using methods such as encrypting the
transmission channel or hashing the stored data. Liveness test is
As mentioned before, EEG records may be used to predict the another way to deal with these attacks. Despite the importance
private information of the users. To protect the users’ privacy, un- of liveness test, it is rarely discussed in the literature, and since
necessary access to stored EEG records should be restricted. One a dead person lacks biometric signals, it is assumed that brain sig-
may suggest to protect the storage of recorded EEG through en- nals can only be received from a living person. As mentioned in the
cryption with data cryptography algorithms. There exist some pa- previous section, this assumption is not true. We believe that cur-
pers on EEG encryption (e.g. Lin et al., 2014; Nguyen et al., 2017), rently one of the most important security challenges of EEG bio-
however, we believe encryption does not solve this challenge, since metric is to provide a liveness test method.
if an attacker gains the control of the system and thus the key
of encryption, he can recover encrypted EEG records. These sys-
10.7. User-friendly
tems should employ a similar approach to password-based authen-
tication. These systems never store the passwords itself but the
While recording brain signals, users need to perform a specific
hash of passwords. The hashes could only be used for proving the
task under defined conditions. The tasks may not be a pleasure for
claimed identity, but does not reveal any more information. Con-
some users. Chuang et al. (2013) suggested that an EEG biomet-
ventional hashing functions cannot be used in the EEG-based au-
ric system can be built independently of tasks. They have shown
thentication since one person’s EEG cannot be exactly the same
that tasks do not have much impact on the system accuracy, thus
on two different sessions, and even a small change in the input
the system does not require to be dependent on a specific task.
of these hash functions will result in completely different outputs.
Moreover, they provided the users with a questionnaire that re-
This leads to losing the ability to verify the claimed information at
vealed some tasks are difficult or not pleasing to the users. Conse-
the time of authentication. For this aim, there are other protection
quently, the authors suggested that, in order to increase the user-
methods for biometrics which are known as biometric template
friendliness of the system, each user may separately choose the
protection methods including feature transformation, key binding
task of interest. Apart from this, this topic has not been addressed
(such as fuzzy vault and fuzzy commitment), and key generation
by other research so far and should be looked at in the future.
(such as secure sketch and fuzzy extractor) (Jain et al., 2008). These
methods make it computationally impossible for an adversary to
obtain the original biometric feature values, while the system is 11. Conclusion
still able to verify the claimed biometric at the time of authenti-
cation. However, as far as we know, no research has done yet on This paper provides an overview of authentication and identi-
the using of these methods in the EEG biometric. The only related fication methods using brain signals. The paper first reviews the
paper is the work done by Singandhupe et al. (2017) in which the characteristics of brain signals and their requirements as an au-
fuzzy extractor is used to extract a key to secure communication thentication tool. In continue, available public datasets and differ-
between a person and the drone under control. The paper only fo- ent protocols for data acquisition are studied. Among the selected
cused on the security of communication and it has not discussed datasets, Physionet and DEAP are the most common and useful
the feasibility of using the proposed method in the general EEG datasets due to the number of subjects, channels, and diversity of
based authentication and larger populations. the tasks. Also, these datasets, in total, covered a wide range of
Similar to other biometrics, the EEG biometric features also EEG acquisition protocols, including rest states, music videos (both
need to be protected from revealing personal information, al- visual/acoustic ERPs) and physical/mental tasks. Although the rest
though this has not yet been addressed in the literature. There states are simpler and more common in practice, ERP based tasks
are plenty of formal works support the template protection meth- create higher SNR. However, they require some equipment for time
ods for biometric authentication factors. For example, He and synchronization between initiation of stimulus and EEG recording.
Wang (2014) proposed a biometrics-based authentication scheme The next reviewed topic is the most common preprocessing
using elliptic curve cryptography and fuzzy extractor. They also methods of EEG signals. The appropriate preprocessing methods
proved the completeness of the proposed scheme using the BAN are not only dependent on the type and trials of tasks, but also
logic. In addition, there are several mature works that suggest us- on the selected feature extraction method. For instance, ensemble
ing multi-factor authentication for increasing both user security averaging can only be applied in ERP based tasks. Or, frequency fil-
and privacy. For instance, Wang and Wang (2016) defined the se- tering is required whenever time-domain features or templates are
curity model of a two-factor authentication and then formally ver- going to be extracted. Removing artifacts is another common pre-
ified its required properties. This work provides a good theoretical process, however, some researches have kept bio-artifacts, i.e. eye
background for this topic. In other works (Wang et al., 2015), an blinking, since they contain biometric data as well(Bhateja et al.,
enhanced scheme with both user anonymity and provable security 2019). Moreover, in this paper, different models of classifications
is suggested. The strength of this work is that it tries to achieve that are used in literature for EEG authentication are also reviewed.
perfect security and privacy with the least additional communica- These methods are divided into two categories of shallow and deep
tion or computation cost. classifications. However, the use of the deep learning approach in
this area has begun in recent years and there are still not many
10.6. Security papers on that, but this approach has well-proven in this field and
surpasses shallow methods. Shallow classification is a two-stage
As mentioned in Section 9, the spoofing attack is the main se- process in which the features should be extracted from prepro-
curity challenge of EEG biometirc. Currently, the most important cessed EEG data in the first stage and then, they can be classified
method of spoofing attack against EEG biometric is to use machine in the second stage. The correct selection of feature extractor and
learning approach to generate synthetic EEG signals. The advances classifier affects the result accuracy, directly. Although AR and PSD,
in the GAN models make it easy for the attackers to generate fake as the extracted features, and similarity-based and SVM methods,
EEG signals similar to a victim’s EEG, with the sufficient numbers as classifiers, are more common than the others, due to the lack
EEG samples. There are main defense techniques: 1) prevent the of a standard benchmark, one cannot explicitly determine which
attackers from accessing the users’ EEG, 2) providing a liveness methods are better in overall. Deep learning approaches, on the
test. For preventing the attacker to access the target’s EEG, the EEG other hand, the model can be directly trained on the preprocessed
data, and indeed, it, in turn, extracts distinguishable features dur- Chen, Y., Atnafu, A.D., Schlattner, I., Weldtsadik, W.T., Roh, M.C., Kim, H.J., Lee, S.W.,
ing the training phase. Therefore, in the deep approaches, only se- Blankertz, B., Fazli, S., 2016. A high-security EEG-based login system with RSVP
stimuli and dry electrodes. IEEE Trans. Inf. Forensics Secur. 11 (12), 2635–2647.
lecting the appropriate network, the number of layers and their Cheung, M.-c., Chan, A.S., Sze, S.L., 2009. Increased theta coherence during chinese
types matters. Currently, CNN is the most network used in the reading. Int. J. Psychophysiol. 74 (2), 132–138.
deep EEG authentication. It’s worth mentioning that this paper is Chuang, J., Nguyen, H., Wang, C., Johnson, B., 2013. I think, therefore I am: Usability
and security of authentication using brainwaves. In: Lecture Notes in Computer
the first work that reviewed deep learning methods in EEG authen- Science (including subseries Lecture Notes in Artificial Intelligence and Lecture
tication. Furthermore, the paper discussed the open challenges in Notes in Bioinformatics), 7862, pp. 1–16. LNCS
this field by summarizing existing work. These challenges are ad- Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B., Bharath, A.A.,
2018. Generative adversarial networks: an overview. IEEE Signal Process. Mag.
dressed in different categories including universality, permanency,
35 (1), 53–65.
uniqueness, collectability, privacy, security, and user-friendliness. Das, R., Maiorana, E., Campisi, P., 2016. EEG biometrics using visual stimuli: a longi-
The challenges are required to be addressed in future works. tudinal study. IEEE Signal Process. Lett. 23 (3), 341–345.
Das, R., Maiorana, E., Campisi, P., 2018. Motor imagery for eeg biometrics using con-
volutional neural network. In: 2018 IEEE International Conference on Acoustics,
Declaration of Competing Interest Speech and Signal Processing (ICASSP). IEEE, pp. 2062–2066.
Das, R., Maiorana, E., La Rocca, D., Campisi, P., 2015. EEG biometrics for user recog-
nition using visually evoked potentials. In: Lecture Notes in Informatics (LNI),
The authors declare that they have no known competing finan- Proceedings –Series of the Gesellschaft fur Informatik (GI) P-245, pp. 303–310.
cial interests or personal relationships that could have appeared to Das, R., Piciucco, E., Maiorana, E., Campisi, P., 2016. Visually evoked potentials for
influence the work reported in this paper. EEG biometrie recognition. In: 2016 1st International Workshop on Sensing, Pro-
cessing and Learning for Intelligent Machines, SPLINE 2016 – Proceedings.
Del Pozo-Banos, M., Alonso, J.B., Ticay-Rivas, J.R., Travieso, C.M., 2014. Elec-
References troencephalogram subject identification: areview. Expert Syst. Appl. 41 (15),
6537–6554.
Abbas, S.N., Abo-Zahhad, M., Ahmed, S.M., Abbas, S.N., Abo-Zahhad, M., Delpozo-Banos, M., Travieso, C.M., Weidemann, C.T., Alonso, J.B., 2015. EEG biometric
Ahmed, S.M., 2015. State-of-the-art methods and future perspectives for per- identification: a thorough exploration of the time-frequency domain. J. Neural
sonal recognition based on electroencephalogram signals. IET Biom. 4 (3), Eng. 12 (5).
179–190. Demos, J.N., 2005. Getting Started With Neurofeedback. WW Norton & Company.
Abdullah, M.K., Subari, K.S., Loong, J.L.C., Ahmad, N.N., 2010. Analysis of effective Di, Y., An, X., He, F., Liu, S., Ke, Y., Ming, D., 2019. Robustness analysis of identifica-
channel placement for an eeg-based biometric system. In: 2010 IEEE EMBS Con- tion using resting-state EEG signals. IEEE Access 7 (c), 42113–42122.
ference on Biomedical Engineering and Sciences (IECBES). IEEE, pp. 303–306. Eeg - electroencephalogram - bci. http://neurosky.com/biosensors/eeg-sensor/. Ac-
Abdullah, M.K., Subari, K.S., Loong, J.L.C., Ahmad, N.N., 2010. Analysis of the eeg cessed: 14 June 2019.
signal for a practical biometric system. World Acad. Sci. Eng. Technol. 68, Eeg headsets and the rise of passthoughts, 2017, http://neurosky.com/2017/02/
1123–1127. eeg- headsets- and- the- rise- of- passthoughts/ (Accessed 14 June 2019).
Abo-Zahhad, M., Ahmed, S.M., Abbas, S.N., 2015. A new EEG acquisition protocol for Feng, Y.C., Yuen, P.C., 2012. Binary discriminant analysis for generating binary face
biometric identification using eye blinking signals. Int. J. Intell. Syst.Appl. 7 (6), template. IEEE Trans. Inf. Forensics Secur. 7 (2), 613–624.
48–54. Frank, D., Mabrey, J., Yoshigoe, K., 2017. Personalizable neurological user authenti-
Abo-Zahhad, M., Ahmed, S.M., Abbas, S.N., 2016. A new multi-level approach to cation framework. In: 2017 International Conference on Computing, Networking
EEG based human authentication using eye blinking. Pattern Recognit. Lett. 82, and Communications, ICNC 2017, pp. 932–936.
216–225. Fraschini, M., Hillebrand, A., Demuru, M., Didaci, L., Marcialis, G.L., 2015. An
Aggarwal, S., Aggarwal, L., Rihal, M.S., Aggarwal, S., 2018. Eeg based participant in- EEG-based biometric system using eigenvector centrality in resting state brain
dependent emotion classification using gradient boosting machines. In: 2018 networks. IEEE Signal Process. Lett. 22 (6), 666–670.
IEEE 8th International Advance Computing Conference (IACC), pp. 266–271. Galbally, J., Marcel, S., Fierrez, J., 2013. Image quality assessment for fake biometric
Alyasseri, Z.A.A., Khader, A.T., Al-Betar, M.A., Papa, J.P., ahmad Alomari, O., 2018. detection: application to iris, fingerprint, and face recognition. IEEE Trans. Image
Eeg-based person authentication using multi-objective flower pollination algo- Process. 23 (2), 710–724.
rithm. In: 2018 IEEE Congress on Evolutionary Computation (CEC). IEEE, pp. 1–8. Gaspar, C.M., Rousselet, G.A., Pernet, C.R., 2011. Reliability of ERP and single-trial
Arakala, A., Culpepper, J.S., Jeffers, J., Turpin, A., Boztaş, S., Horadam, K.J., McK- analyses. Neuroimage 58 (2), 620–629.
endrick, A.M., 2009. Entropy of the retina template. In: International Conference Goldberger, A.L., Amaral, L.A., Glass, L., Hausdorff, J.M., Ivanov, P.C., Mark, R.G., Mi-
on Biometrics. Springer, pp. 1250–1259. etus, J.E., Moody, G.B., Peng, C.K., Stanley, H.E., 20 0 0. Physiobank, physiotoolkit,
Armstrong, B.C., Ruiz-Blondet, M.V., Khalifian, N., Kurtz, K.J., Jin, Z., Laszlo, S., 2015. and physionet: components of a new research resource for complex physiologic
Brainprint: assessing the uniqueness, collectability, and permanence of a novel signals.. Circulation 101 (23), e215–e220.
method for erp biometrics. Neurocomputing 166, 59–67. Gondesen, F., Marx, M., Gollmann, D., 2019. Eeg-based biometrics. In: Biomet-
Arnau-Gonzalez, P., Katsigiannis, S., Ramzan, N., Tolson, D., Arevalillo-Herrez, M., ric-Based Physical and Cybersecurity Systems. Springer, pp. 287–318.
2017. Es1d: a deep network for eeg-based subject identification. In: 2017 IEEE Gui, Q., Jin, Z., Xu, W., 2014. Exploring eeg-based biometrics for user identification
17th International Conference on Bioinformatics and Bioengineering (BIBE). and authentication. In: 2014 IEEE Signal Processing in Medicine and Biology
IEEE, pp. 81–85. Symposium (SPMB). IEEE, pp. 1–6.
Ashby, C., Bhatia, A., Tenore, F., Vogelstein, J., 2011. Low-cost electroencephalogram Gui, Q., Jin, Z., Xu, W., Ruiz-Blondet, M.V., Laszlo, S., 2015. Multichannel eeg-based
(EEG) based authentication. In: 2011 5th International IEEE/EMBS Conference on biometric using improved RBFneural networks. In: 2015 IEEE Signal Processing
Neural Engineering. IEEE, pp. 442–445. in Medicine and Biology Symposium (SPMB), pp. 1–6.
Bajwa, G., Dantu, R., 2016. Neurokey: towards a new paradigm of cancelable bio- Hao, F., Anderson, R., Daugman, J., 2006. Combining crypto with biometrics effec-
metrics-based key generation using electroencephalograms. Comput. Secur. 62, tively. IEEE Trans. Comput. 55 (9), 1081–1088.
95–113. Hartmann, K. G., Schirrmeister, R. T., Ball, T., 2018. EEG-GAN: generative adversarial
Bao, X., Wang, J., Hu, J., 2009. Method of individual identification based on elec- networks for electroencephalograhic (EEG) brain signals. arXiv:1806.01875.
troencephalogram analysis. In: Proceedings – 2009 International Conference on He, D., Wang, D., 2014. Robust biometrics-based authentication scheme for multi-
New Trends in Information and Service Science, NISS 2009, pp. 390–393. server environment. IEEE Syst. J. 9 (3), 816–823.
Barry, R.J., Clarke, A.R., Johnstone, S.J., Magee, C.A., Rushby, J.A., 2007. Eeg differences Hoang, T., Choi, D., Nguyen, T., 2015. Gait authentication on mobile phone using
between eyes-closed and eyes-open resting conditions. Clin. Neurophysiol. 118 biometric cryptosystem and fuzzy commitment scheme. Int. J. Inf. Secur. 14 (6),
(12), 2765–2773. 549–560.
Begleiter, H., Porjesz, B., Litke, A., Zhang, X.L., Wang, W., 2002. Event related poten- Höller, Y., Uhl, A., 2018. Do eeg-biometric templates threaten user privacy? In: Pro-
tials during object recognition tasks. Brain Res. Bull. 38 (6), 531–538. ceedings of the 6th ACM Workshop on Information Hiding and Multimedia Se-
Bhateja, V., Gupta, A., Mishra, A., Mishra, A., 2019. Artificial Neural Networks Based curity. ACM, pp. 31–42.
Fusion and Classification of Eeg/eog Signals. In: Information Systems Design and Hu, J.F., 2009. New biometric approach based on motor imagery EEG signals. In:
Intelligent Applications. Springer, pp. 141–148. FBIE 2009 – 2009 International Conference on Future BioMedical Information
Bickford, R., 1987. Electroencephalography. In: Adelman, G. (Ed.), Encyclopedia of Engineering, pp. 94–97.
Neuroscience. Cambridge, pp. 371–373. Hu, J.F., 2010. Biometric system based on EEG signals by feature combination. In:
Blasco, J., Chen, T.M., Tapiador, J., Peris-Lopez, P., 2016. A survey of wearable bio- 2010 International Conference on Measuring Technology and Mechatronics Au-
metric recognition systems. ACM Comput. Surv. 49 (3), 43. tomation, ICMTMA 2010, 1, pp. 752–755.
Boubakeur, M.R., Wang, G., Zhang, C., Liu, K., 2017. Eeg-based person recognition Hunter, M., Smith, R.L., Hyslop, W., Rosso, O.A., Gerlach, R., Rostas, J.A.,
analysis and criticism. In: 2017 IEEE International Conference on Big Knowledge Williams, D.B., Henskens, F., 2005. The Australian EEG database. Clin. EEG Neu-
(ICBK). IEEE, pp. 155–160. rosci. 36 (2), 76–81.
Brunner, C., Leeb, R., Müller-Putz, G., Schlögl, A., Pfurtscheller, G., 2008. BCI Compe- Idrus, S.Z.S., Cherrier, E., Rosenberger, C., Schwartzmann, J.-J., 2013. A review on au-
tition 2008–Graz Data Sset A. Technical Report. thentication methods. Aust. J. Basic Appl. Sci. 7 (5), 95–107.
Chen, J.X., Mao, Z.J., Yao, W.X., Huang, Y.F., 2019. EEG-Based biometric identification Jain, A.K., Nandakumar, K., Nagar, A., 2008. Biometric template security. EURASIP J.
with convolutional neural network. Multimed. Tools Appl.. Adv. Signal Process. 2008, 113.
Jain, A.K., Ross, A., Prabhakar, S., 2004. An introduction to biometric recognition. Maiorana, E., La Rocca, D., Campisi, P., 2016. On the permanence of EEG signals for
IEEE Trans. Circt. Syst. Video Technol. 14 (1), 4–20. doi:10.1109/TCSVT.2003. biometric recognition. IEEE Trans. Inf. Forensics Secur. 11 (1), 163–175.
818349. Maiorana, E., Solé-Casals, J., Campisi, P., 2016. EEG signal preprocessing for biometric
Jayarathne, I., Cohen, M., Amarakeerthi, S., 2016. Brainid: development of an recognition. Mach. Vis. Appl. 27 (8), 1351–1360.
eeg-based biometric authentication system. In: 2016 IEEE 7th Annual Informa- Mao, Z., Yao, W.X., Huang, Y., 2017. Eeg-based biometric identification with deep
tion Technology, Electronics and Mobile Communication Conference (IEMCON). learning. In: 2017 8th International IEEE/EMBS Conference on Neural Engineer-
IEEE, pp. 1–6. ing (NER), pp. 609–612.
Jian-feng, H., 2009. Multifeature biometric system based on EEG signals. In: Pro- Marcel, S., Millan, J.d.R., 2007. Person authentication using brainwaves (EEG) and
ceedings of the 2nd International Conference on Interaction Sciences: Informa- maximum a posteriori model adaptation. IEEE Trans. Pattern Anal. Mach. Intell.
tion Technology, Culture and Human. ACM, pp. 1341–1345. 29 (4), 743–748.
Johnson, B., Maillart, T., Chuang, J., 2014. My thoughts are not your thoughts. In: McFarland, D.J., McCane, L.M., David, S.V., Wolpaw, J.R., 1997. Spatial filter selection
Proceedings of the 2014 ACM International Joint Conference on Pervasive and for eeg-based communication. Electroencephalogr. Clin. Neurophysiol. 103 (3),
Ubiquitous Computing: Adjunct Publication. ACM, pp. 1329–1338. 386–394.
K. Inthavisas, D.L., 2012. Secure speech biometric templates for user authentication. Moore, S.K., 2016. “Brainprint” biometric id hits 100% accuracy [news]. IEEE Spectr.
IET Biom. 1. 46–54(8) 53 (6). 14–14
Kanade, S., Camara, D., Petrovska-Delacrtaz, D., Dorizzi, B., Nakamura, T., Goverdovsky, V., Mandic, D.P., 2017. In-ear eeg biometrics for feasible
2009. Application of biometrics to obtain high entropy and readily collectable real-world person authentication. IEEE Trans. Inf. Foren-
cryptographic keys. World Acad. Sci. Eng. Technol. 52, sics Secur. 13 (3), 648–661.
330. Nguyen, D., Tran, D., Ma, W., 2019. A study on combing EEG signals and crytography
Kang, J.-h., Chang, Y., Kim, S.-p., 2018. Neurocomputing electroencephalographic fea- for bitcoin security. Aust. J. Intell. Inf.Process. Syst. 34.
ture evaluation for improving personal authentication performance. Neurocom- Nguyen, D., Tran, D., Sharma, D., Ma, W., 2017. On the study of EEG-based
puting 287, 93–101. cryptographic key generation. In: Procedia Computer Science. Elsevier B.V.,
Kaur, B., Singh, D., 2017. Neuro signals: a future biomertic approach towards user pp. 936–945.
identification. In: 2017 7th International Conference on Cloud Computing, Data Nguyen, P., Tran, D., Huang, X., Sharma, D., 2012. A proposed feature extraction
Science & Engineering-Confluence. IEEE, pp. 112–117. method for EEG-based person identification. Int. Conf. Artif.Intell..
Kaur, B., Singh, D., Roy, P.P., 2017. A novel framework of EEG-based user identi- Nguyen, P., Tran, D., Le, T., Huang, X., Ma, W., 2013. Eeg-based person verification
fication by analyzing music-listening behavior. Multimed. Tools Appl. 76 (24), using multi-sphere SVDDand UBM. In: Pacific-Asia Conference on Knowledge
25581–25602. Discovery and Data Mining. Springer, pp. 289–300.
Keirn, Z.A., Aunon, J.I., 1990. A new mode of communication between man and his Ni, Z., Yuksel, A.C., Ni, X., Mandel, M.I., Xie, L., 2017. Confused or not confused?:
surroundings. IEEE Trans. Biomed. Eng. 37 (12), 1209–1214. Disentangling brain activity from eeg data using bidirectional LSTM recurrent
Keshishzadeh, S., Fallah, A., Rashidi, S., 2016. Improved EEG based human authen- neural networks. In: Proceedings of the 8th ACM International Conference on
tication system on large dataset. In: 2016 24th Iranian Conference on Electrical Bioinformatics, Computational Biology,and Health Informatics. ACM, New York,
Engineering, ICEE 2016, pp. 1165–1169. NY, USA, pp. 241–246.
Kim, D., Kim, K., 2019. Resting state eeg-based biometric system using concatena- Niedermeyer, E., da Silva, F.L., 2005. Electroencephalography: Basic Principles, Clini-
tion of quadrantal functional networks. IEEE Access 7, 65745–65756. cal Applications, and Related Fields. Lippincott Williams & Wilkins.
Klimesch, W., 1996. Memory processes, brain oscillations and EEGsynchronization. Nik Aznan, N.K., Atapour-Abarghouei, A., Bonner, S., Connolly, J.D., Al Moubayed, N.,
Int. J. Psychophysiol. 24 (1), 61–100. Breckon, T.P., 2019. Simulating brain signals: creating synthetic eeg data via
Klonovs, J., Petersen, C.K., Olesen, H., Hammershøj, A., Hammershoj, A., 2013. ID neural-based generative models for improved SSVEP classification. In: 2019 In-
proof on the go: development of a mobile EEG-based biometric authentication ternational Joint Conference on Neural Networks (IJCNN), pp. 1–8.
system. IEEE Spectr. 8 (February), 81–89. Özdenizci, O., Wang, Y., Koike-Akino, T., Erdoğmuş, D., 2019. Adversarial deep learn-
Koelstra, S., Mühl, C., Soleymani, M., Lee, J.S., Yazdani, A., Ebrahimi, T., Pun, T., Ni- ing in eeg biometrics. IEEE Signal Process. Lett. 26 (5), 710–714.
jholt, A., Patras, I., 2012. DEAP: a database for emotion analysis; using physio- Palaniappan, R., 2006. Electroencephalogram signals from imagined activities: a
logical signals. IEEE Trans. Affect. Comput. 3 (1), 18–31. novel biometric identifier for a small population. In: International Conference
Koike-Akino, T., Mahajan, R., Marks, T.K., Wang, Y., Watanabe, S., Tuzel, O., Orlik, P., on Intelligent Data Engineering and Automated Learning. Springer, pp. 604–611.
2016. High-accuracy user identification using eeg biometrics. In: 2016 38th An- Panwar, S., Rad, P., Quarles, J., Huang, Y., 2019. Generating eeg signals of an RSVP
nual International Conference of the IEEE Engineering in Medicine and Biology experiment by a class conditioned Wasserstein generative adversarial network.
Society (EMBC). IEEE, pp. 854–858. In: 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC).
Kostílek, M., Št’astnỳ, J., 2012. Eeg biometric identification: repeatability and influ- IEEE, pp. 1304–1310.
ence of movement-related eeg. In: 2012 International Conference on Applied Pham, T., Ma, W., Tran, D., Nguyen, P., Phung, D., 2014. Multi-factor EEG-based user
Electronics. IEEE, pp. 147–150. authentication. Proc. Int. Jt. Conf. Neural Netw. 4029–4034.
Krivokuća, V., Gomez-Barrero, M., Marcel, S., Rathgeb, C., Busch, C., 2020. Towards Pham, T., Ma, W., Tran, D., Tran, D.S., Phung, D., 2015. A study on the stability of eeg
measuring the amount of discriminatory information in finger vein biometric signals for user authentication. In: 2015 7th International IEEE/EMBS Conference
characteristics using a relative entropy estimator. In: Handbook of Vascular Bio- on Neural Engineering (NER). IEEE, pp. 122–125.
metrics. Springer, pp. 507–525. Phothisonothai, M., 2015. An investigation of using SSVEP for EEG-based user au-
Kumar, P., Saini, R., Pratim Roy, P., Prosad Dogra, D., 2017. A bio-signal based frame- thentication system. In: 2015 Asia-Pacific Signal and Information Processing As-
work to secure mobile devices. J. Netw. Comput. Appl. 89 (January), 62–71. sociation Annual Summit and Conference (APSIPA). IEEE, pp. 923–926.
Kumari, P., Vaish, A., 2016. Feature-level fusion of mental task’s brain signal for an Phung, D.Q., Tran, D., Ma, W., Nguyen, P., Pham, T., 2014. Using Shannon entropy as
efficient identification system. Neural Comput. Appl. 27 (3), 659–669. eeg signal feature for fast person identification.. In: ESANN, 4, pp. 413–418.
La Rocca, D., Campisi, P., Scarano, G., 2014. Stable EEG features for biometric recog- Piciucco, E., Maiorana, E., Falzon, O., Camilleri, K.P., Campisi, P., 2017. Steady-state
nition in resting state conditions. Commun. Comput. Inf. Sci. 452, 313–330. visual evoked potentials for eeg-based biometric identification. In: 2017 Inter-
Lan Ma, Minett, J.W., Blu, T., Wang, W.S.-Y., 2015. Resting state EEG-based biomet- national Conference of the Biometrics Special Interest Group (BIOSIG). IEEE,
rics for individual identification using convolutional neural networks. In: 2015 pp. 1–5.
37th Annual International Conference of the IEEE Engineering in Medicine and Piplani, T., Merill, N., Chuang, J., 2018. Faking it, making it: fooling and improving
Biology Society (EMBC), pp. 2848–2851. brain-based authentication with generative adversarial networks. In: 2018 IEEE
Lawhern, V.J., Solon, A.J., Waytowich, N.R., Gordon, S.M., Hung, C.P., Lance, B.J., 2018. 9th International Conference on Biometrics Theory, Applications and Systems
Eegnet: a compact convolutional neural network for eeg-based brain–computer (BTAS), pp. 1–7.
interfaces. J. Neural Eng. 15 (5), 056013. Polich, J., 2007. Updating p300: an integrative theory of p3a and p3b. Clin. Neuro-
Li, P., Yang, X., Qiao, H., Cao, K., Liu, E., Tian, J., 2012. An effective biometric cryp- physiol. 118 (10), 2128–2148.
tosystem combining fingerprints with error correction codes. Expert Syst. Appl. Poulos, M., Rangoussi, M., Kafetzopoulos, E., 1998. Person identification via the eeg
39 (7), 6562–6574. using computational geometry algorithms. In: 9th European Signal Processing
Liew, S.-H., Choo, Y.-H., Low, Y.F., Mohd Yusoh, Z.I., 2018. EEG-based biometric au- Conference (EUSIPCO 1998). IEEE, pp. 1–4.
thentication modelling using incremental fuzzy-rough nearest neighbour tech- Ramzan, Q., Shidlovskiy, S., 2018. Evolution of the Brain Computing Interface
nique. IET Biom. 7 (2), 145–152. (BCI) and proposed electroencephalography (EEG) signals based authentication
Liew, S.-H., Choo, Y.-H., Low, Y.F., Yusoh, Z.I.M., 2015. Identifying visual evoked po- model. MATEC Web Conf. 155, 01006.
tential (vep) electrodes setting for person authentication. Int. J. Adv. Soft Com- Ravi, K.V.R., Palaniappan, R., Eswaran, C., Phon-Amnuaisuk, S., 2008. Data encryp-
put. Appl. 7 (3), 85–99. tion using event-related brain signals. In: Proceedings – International Confer-
Lin, C.-F., Shih, S.-H., Zhu, J.-D., 2014. Chaos based encryption system for encrypting ence on Computational Intelligence and Multimedia Applications, ICCIMA 2007,
electroencephalogram signals. J. Med. Syst. 38 (5), 49. 1, pp. 540–544.
Lin, C.-T., Liu, Y.-T., Wu, S.-L., Cao, Z., Wang, Y.-K., Huang, C.-S., King, J.-T., Chen, S.-A., Read, G.L., Innis, I.J., 2017. Electroencephalography (EEG). Am. Cancer Soc., pp. 1–18.
Lu, S.-W., Chuang, C.-H., 2017. Eeg-based brain-computer interfaces: a novel Reiterer, S., Pereda, E., Bhattacharya, J., 2009. Measuring second language proficiency
neurotechnology and computational intelligence method. IEEE Syst. Man Cy- with eeg synchronization: how functional cortical networks and hemispheric
bern. Mag. 3 (4), 16–26. involvement differ as a function of proficiency level in second language speak-
Luck, S.J., 2014. An Introduction to the Event-Related Potential Technique. MIT Press, ers. Second Lang. Res. 25 (1), 77–106.
Cambridge, MA, USA. Reiterer, S.M., Pereda, E., Bhattacharya, J., 2011. On a possible relationship between
Maiorana, E., Campisi, P., 2018. Longitudinal evaluation of EEG-Based biometric linguistic expertise and EEG gamma band phase synchrony. Front. Psychol. 2,
recognition. IEEE Trans. Inf. Forensics Secur. 13 (5), 1123–1138. 334.
Roberts, C., 2007. Biometric attack vectors and defences. Comput. Secur. 26 (1), Thomas, K.P., Vinod, A.P., 2018. EEG-Based Biometric authentication using gamma
14–25. band power during rest state. Circt. Syst. Signal Process. 37 (1), 277–289.
Ruiz-blondet, M., Khalifian, N., Armstrong, B.C., Jin, Z., Kurtz, K.J., Laszlo, S., 2014. Thomas, K.P., Vinod, A.P., Robinson, N., 2017. Online biometric authentication using
Brainprint: identifying unique features of neural activity with machine learning. subject-specific band power features of EEG. In: Proceedings of the 2017 Inter-
In: Proc. 36th Annual Conf. of the Cognitive Science Society, pp. 827–832. national Conference on Cryptography, Security and Privacy. ACM, pp. 136–141.
Ruiz-Blondet, M.V., Jin, Z., Laszlo, S., 2016. Cerebre: a novel method for very high ac- Vahid, A., Arbabi, E., 2016. Human identification with eeg signals in different emo-
curacy event-related potential biometric identification. IEEE Trans. Inf. Forensics tional states. In: 2016 23rd Iranian Conference on Biomedical Engineering and
Secur. 11 (7), 1618–1629. 2016 1st International Iranian Conference on Biomedical Engineering (ICBME).
Ruiz-Blondet, M.V., Jin, Z., Laszlo, S., 2017. Permanence of the cerebre brain biomet- IEEE, pp. 242–246.
ric protocol. Pattern Recognit. Lett. 95, 37–43. Waili, T., Johar, M.G.M., Sidek, K., Nor, N.M., Yaacob, H., Othman, M., 2019. Eeg based
Sadeghi, K., Banerjee, A., Sohankar, J., K. S. Gupta, S., 2017. Geometrical analysis of biometric identification using correlation and MLPNN models.. Int. J. Online Eng.
machine learning security in biometric authentication systems. In: 2017 16th 15 (10), 77–90.
IEEE International Conference on Machine Learning and Applications (ICMLA), Wang, D., Cheng, H., Wang, P., Huang, X., Jian, G., 2017. Zipf’s law in passwords. IEEE
pp. 309–314. Trans. Inf. Forensics Secur. 12 (11), 2776–2791.
Sadeghi, K., Sohankar, J., Banerjee, A., Gupta, S.K., 2017. A novel spoofing Wang, D., Wang, N., Wang, P., Qing, S., 2015. Preserving privacy for free: efficient and
attack against electroencephalogram-based security systems. In: 2017 provably secure two-factor authentication scheme with user anonymity. Inf. Sci.
IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & 321, 162–178.
Trusted Computed, Scalable Computing & Communications, Cloud & Big Wang, D., Wang, P., 2016. Two birds with one stone: two-factor authentication with
Data Computing, Internet of People and Smart City Innovation (Smart- security beyond conventional bound. IEEE Trans Dependable Secure Comput 15
World/SCALCOM/UIC/ATC/CBDCom/IOP/SCI). IEEE, pp. 1–6. (4), 708–722.
Sanei, S., 2013. Adaptive Processing of Brain Signals. John Wiley & Sons. Wang, Y., Najafizadeh, L., 2016. On the invariance of EEG-based signatures of indi-
Sanei, S., Chambers, J., 2013. EEG Signal Processing. Wiley. viduality with application in biometric identification. In: Proceedings of the An-
Schalk, G., McFarland, D.J., Hinterberger, T., Birbaumer, N., Wolpaw, J.R., 2004. nual International Conference of the IEEE Engineering in Medicine and Biology
Bci20 0 0: a general-purpose brain-computer interface (BCI) system. IEEE Trans. Society, EMBS, 2016, pp. 4559–4562.
Biomed. Eng. 51 (6), 1034–1043. Wang, Z., Healy, G., Smeaton, A.F., Ward, T.E., 2018. A review of feature extraction
Schirrmeister, R.T., Springenberg, J.T., Fiederer, L.D.J., Glasstetter, M., and classification algorithms for image rsvp based BCI. In: Signal Processing and
Eggensperger, K., Tangermann, M., Hutter, F., Burgard, W., Ball, T., 2017. Machine Learning for Brain-Machine Interfaces, pp. 243–270.
Deep learning with convolutional neural networks for EEG decoding and Wilaiprasitporn, T., Ditthapron, A., Matchaparn, K., Tongbuasirilai, T., Banluesombat-
visualization.. Hum. Brain Mapp. 38 (11), 5391–5420. kul, N., Chuangsuwanich, E., 2019. Affective eeg-based person identification us-
Schons, T., Moreira, G.J.P., Silva, P.H.L., Coelho, V.N., Luz, E.J.S., 2018. Convolutional ing the deep learning approach. IEEE Trans. Cognit. Dev.Syst..
network for EEG-based biometric. In: Lecture Notes in Computer Science (in- Wu, Q., Zeng, Y., Zhang, C., Tong, L., Yan, B., 2018. An EEG-based person authentica-
cluding subseries Lecture Notes in Artificial Intelligence and Lecture Notes in tion system with open-set capability combining eye blinking signals. Sensors 18
Bioinformatics) 10657 LNCS, pp. 601–608. (2), 1–18.
Sharma, P.K., Vaish, A., 2016. Individual identification based on neuro-signal using Xiao, D., Hu, J., 2010. Identification of motor imagery EEG signal. In: 2010 Inter-
motor movement and imaginary cognitive process. Opt.-Int. J. Light Electron national Conference on Biomedical Engineering and Computer Science, ICBECS
Opt. 127 (4), 2143–2148. 2010.
Singandhupe, A., La, H.M., Feil-Seifer, D., Huang, P., Guo, L., Li, M., 2017. Se- Xiao, Y., Jia, Y., Cheng, X., Yu, J., Liang, Z., Tian, Z., 2019. I can see your brain: inves-
curing a UAV using individual characteristics from an eeg signal. In: 2017 tigating home-use electroencephalography system security. IEEE Internet Things
IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, J. 6 (4), 6681–6691.
pp. 2748–2753. Yang, S., Deravi, F., 2017. On the usability of electroencephalographic signals for bio-
Singh, B., Mishra, S., Tiwary, U.S., 2015. EEG based biometric identification with re- metric recognition: a survey. IEEE Trans. Hum. Mach. Syst. 47 (6), 958–969.
duced number of channels. In: International Conference on Advanced Commu- Yang, S., Deravi, F., Hoque, S., 2018. Task sensitivity in EEG biometric recognition.
nication Technology, ICACT 2015-August, pp. 687–691. Pattern Anal. Appl. 21 (1), 105–117.
Snodgrass, J.G., Vanderwart, M., 1980. A standardized set of 260 pictures: norms for Yin, Z., Zhao, M., Wang, Y., Yang, J., Zhang, J., 2017. Recognition of emotions using
name agreement, image agreement, familiarity, and visual complexity. J. Exp. multimodal physiological signals and an ensemble deep learning model. Com-
Psychol.. put. Methods Programs Biomed. 140, 93–110.
Sockeel, S., Schwartz, D., Pélégrini-Issac, M., Benali, H., 2016. Large-scale functional Zeng, Y., Wu, Q., Yang, K., Tong, L., Yan, B., Shu, J., Yao, D., 2019. Eeg-based iden-
networks identified from resting-state EEG using spatial ICA. PLoS One. tity authentication framework using face rapid serial visual presentation with
Sohankar, J., Sadeghi, K., Banerjee, A., Gupta, S.K.S., 2015. E-BIAS: a pervasive optimized channels. Sensors 19 (1), 6.
EEG-based identification and authentication system. In: Proceedings of the Zhang, F., Mao, Z., Huang, Y., Lin, X., Ding, G., 2018. Deep learning models for
11th ACM Symposium on QoS and Security for Wireless and Mobile Networks, eeg-based rapid serial visual presentation event classification. J. Inf. Hiding Mul-
pp. 165–172. timed. Signal Process. 9 (1), 177–187.
Sohankar, J., Sadeghi, K., Banerjee, A., Gupta, S.K.S., 2017. Systematic analysis of live- Zhang, Y., Chen, Y., Bressler, S.L., Ding, M., 2008. Response preparation and inhibi-
ness detection methods in biometrie security systems. In: 2017 IEEE Smart- tion: the role of the cortical sensorimotor beta rhythm. Neuroscience 156 (1),
World, Ubiquitous Intelligence Computing, Advanced Trusted Computed, Scal- 238–246.
able Computing Communications, Cloud Big Data Computing, Internet of People Zuquete, A., Quintela, B., Silva Cunha, J.P., 2010. Biometric authentication using brain
and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), responses to visual stimuli. In: International Conference on Bio-inspired Sys-
pp. 1–6. tems and Signal Processing, pp. 103–112.
Sun, Y., Lo, F.P., Lo, B., 2019. EEG-Based user identification system using 1D-con-
volutional long short-term memory neural networks. Expert Syst. Appl. 125, Amir Jalaly Bidgoly received his BSc degree in computer engineering (software)
259–267. from University of Kashan (Kashan, Iran), in 2005, MSc degree in computer engi-
Sundararajan, A., Pons, A., Sarwat, A.I., 2015. A generic framework for EEG-based neering (software) from Iran University of Science and Technology (IUST) in 2009,
biometric authentication. In: 2015 12th International Conference on Information and PhD in computer engineering (software) in Isfahan University (Isfahan, Iran) in
Technology-New Generations. IEEE, pp. 139–144. 2015. He is currently an Associated Professor in Department of Computer Engineer-
Tabar, Y.R., Halici, U., 2016. A novel deep learning approach for classification of eeg ing and Information Technology in University of Qom. His research interests include
motor imagery signals. J. Neural Eng. 14 (1), 016003. System Security, Social Security, Trust and Reputation Systems, and Formal Verifica-
Tangkraingkij, P., Lursinsap, C., Sanguansintukul, S., Desudchit, T., 2010. Personal tion.
identification by EEG using ICA and neural network. In: International Confer-
ence on Computational Science and Its Applications. Springer, pp. 419–430.

EXE - A Survey On Methods and Challenges in EEG Based Authentication

Uploaded by

Copyright:

Available Formats

EXE - A Survey On Methods and Challenges in EEG Based Authentication

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

EXE - A Survey On Methods and Challenges in EEG Based Authentication

Uploaded by

Copyright:

Available Formats

Computers & Security 93 (2020) 101788

Contents lists available at ScienceDirect

Computers & Security

A survey on methods and challenges in EEG based authentication

1. Introduction • Biometric factors such as ﬁngerprints or iris may be forged or

Fig. 1. Number of documents published about EEG based authentication.

Region Local channels Functions

1. Knowledge factor: the information which just the user knows,

Fig. 3. Commercial EEG headsets.

Fig. 4. The process of EEG-based authentication.

4.4. UCI KDD EEG database

The records of this dataset (Begleiter et al., 2002) are obtained

4.5. Australian EEG database

Australian EEG database (Hunter et al., 2005) is a result of an

The DEAP dataset (Koelstra et al., 2012) is a famous and well-

6.1. Frequency domain ﬁltering

7.1. Feature extraction

Fusion of Manhattan, Euclidean and cosine distances

Fusion of Euclidean distance, SVM and LDA.

where w[n] is the white noise and {aq } is set of AR coeﬃcients.

used as the feature vector(Abo-Zahhad et al., 2015; Bhateja et al.,

and Campisi, 2018; Maiorana et al., 2016a; Pham et al., 2014;

chine” (SVM) (Keshishzadeh et al., 2016), “Artiﬁcial Neural Net-

works” (ANN) (Bhateja et al., 2019), “Linear Decremental Analy-

sis” (LDA) (Abo-Zahhad et al., 2015) or similarity based classiﬁers

to generate a “template”, rather than modeling by AR. The template

contains a short period of the processed signal of one or multiple

template generation is signiﬁcantly less complex than AR model-

ing, it requires more storage space and consequently, it may not be

practical on a large scale dataset or people. Template can be classi-

distance (Das et al., 2016a) or cross correlation(Ruiz-Blondet et al.,

Kumar et al., 2017).

scribing the EEG signals in the frequency domain. Indeed, PSD is

all selected channels(Ashby et al., 2011; Di et al., 2019; Maio-

REC, REO, mental computation

SVM(Ashby et al., 2011), similarity-based methods(Maiorana et al.,

Transform” (WT) is also widely used as a powerful feature extrac-

tor in both time-frequency domains. WT decomposes a signal to a

weighted summation of functions, namely wavelet, that each one

is a shifted and scaled copy of a basic function, namely mother

wavelet. Each wavelet approximates the signal with a lower res-

olution. By using Wavelet decomposition, the set of wavelets’ coef-

Ruiz-Blondet et al. (2017)

Sharma and Vaish (2016)

classiﬁed by HMM(Kaur et al., 2017), ANN(Alyasseri et al., 2018) or

SVM(Kaur and Singh, 2017). In rather cases, the statical properties

of wavelets’ coeﬃcients are also used as features; e.g. mean, stan-

Gui et al. (2015)

dard deviation, RMS and entropy(Abdullah et al., 2010b; Kaur and

“mel-frequency cepstrum coeﬃcients” (MFCCs) (Maiorana and Camp-

tence of an imitating attacker in an environment increases the false

a warning that error may increase unacceptably in larger popula-

tions. In another study, it was shown that in the authenticating

is aware of the users’ selected sport. This research clearly demon-

strates the sensitivity of the imitation attacks’ success rate to the

ness of EEG biometric against these type of attacks. In the current

failed attempts are suﬃcient to prevent the imitating attacker.

The probability of an authentication factor breaking through the

brute-force is usually measured by the “entropy” metric. Suppose