(will be inserted by the editor)

Biometric Identification through Eye-Movement

Patterns

Akram Bayat · Marc Pomplun

the date of receipt and acceptance should be inserted later

Abstract This paper describes how to identify unique individual readers using
their eye-movement patterns. A case study including forty participants was con-
ducted in order to measure eye movement during reading. The proposed biometric
method is developed based on an informative and stable eye-movement feature
set that gives rise to a high performance multi-class identification model. Multiple
individual classifiers are trained and tested on our novel feature set consisting of
28 features that represent basic eye movement, scan path and pupillary character-
istics. We combine three high-accuracy classifiers, namely Multilayer Perceptron,
Logistic, and Logistic Model Tree using the average of probabilities as the combina-
tion rule. We reach an overall accuracy of 95.31% and an average Equal Error Rate
(EER) of 2.03% and propose a strategy for adjusting decision thresholds that de-
creases the false acceptance rate to 0.1%. Our approach dramatically outperforms
previous methods, making it possible for the first time to build eye-movement
biometric systems for user identification and personalized interfaces.
Keywords Biometric identification · Equal error rate · Eye movement · Pattern
recognition

1 Introduction

A biometric system is essentially a pattern recognition process that operates by

acquiring biometric data from an individual, extracting a feature set from the
acquired data, and comparing this feature set against the template set in the
database (Jain, Ross, and Prabhakar, 2004).
A biometric system typically operates in one of two modes: verification or
identification. In the identification task, an unknown biometric sample is compared

A. Bayat
Department of Computer Science, University of Massachusetts Boston, Morrissey Boulevard,
Boston, MA, USA.
E-mail: akram@cs.umb.edu
M. Pomplun
E-mail: marc@cs.umb.edu
2 Akram Bayat, Marc Pomplun

to a whole database of known individuals, and the best matching template is

selected. On the other hand, the verification task consists of verifying whether
the claimant is the one who he or she claims to be (Bednarik et al., 2005). In this
work, we consider our biometric implementation to operate in identification mode,
which means that its task is to determine the user’s identity or decide that the
user is unknown.
Biometric systems based on physical or behavioral attributes of a person such
as face, voice and fingerprint for personal recognition have been widely used over
the last decade. Using eye movement in biometric systems is an important yet
challenging research area with many applications in health care, smart environ-
ments, security and personalized interfaces. The eyes in particular offer a variety
of physical (iris) and behavioral (eye movements) properties that make them ideal
for the purposes of biometric identification, due to their high specificity and the
complex mechanical logistics of reproduction (Komogortsev and Khan, 2008). For
this reason, there is an increased tendency toward integration of eye movement
biometrics into the standard identification methods in order to enhance the level of
security. However, the error rates produced by current methods of eye movement
analysis are higher than those of accepted biometric systems. For example, the
EER of 10.8% and Rank-1 identification rate of 51% by Rigas and Komogortsev
(2014a) on a large database of 200 subjects yielded one of the best results among
all existing eye movement-driven biometric methods.
In the context of biometric identification, we aim to develop a model that is
capable of identifying a user by implementing various learning techniques to clas-
sify the user’s eye movement patterns for the best match. Our identification task
is formulated as a supervised classification problem, whose training data contains
eye movements of forty human subjects that were recorded while they performed
reading activities in laboratory conditions.
Our objective differs from earlier studies as follows: First, we provide an iden-
tification method by capturing normal user activity, yet are able to obtain sig-
nificantly higher performance compared to previous studies. The reading process
presents an instance of high usability, as it is a very common activity. However,
the effects of learning and text content in reading tasks make it very challenging
to achieve the same results by task repetition for the same person. For this reason,
many previous studies used designs for their eye movement biometrics algorithms
that require the display of specific stimuli and cannot be adapted to general task
performance.
Second, our feature set consists of features that have not been widely studied
in the context of reading experiments. The features allow us to reliably distinguish
individuals based on differences in eye-movement patterns. The features have been
evaluated and selected by considering two experiments to decrease the effect of the
text content on the identification procedure.
Last but not least, we apply a multi-class classification methodology which
constructs a decision threshold for each class. We propose a new algorithm of
adjusting decision threshold probabilities of each class that improves identification
performance compared to an equal error rate (EER) based method.
The rest of the paper is organized as follows: Section 2 gives an overview over
the experiment design and data. Section 3 describes the extracted features. The
classification methods and performance evaluation are described in Sections 4 and
Biometric Identification through Eye-Movement Patterns 3

5, respectively. Related work is summarized in Section 6, and we conclude the

paper in Section 7.

2 Experimental Design

In this section, we describe an experimental design which provides our identifica-

tion procedure with robustness against the content of the texts that are presented
during experiments.

2.1 Design Criteria

In order to best capture an actual user’s eye-movement behavior during a reading

experiment, two important design criteria must be considered: first, if subjects
read an identical set of passages, it is likely that the identification algorithm will
be overfitted to specific feature characteristics of texts used in the experiment. Sec-
ond, if each subject reads a set of passages with different contents, it is possible
that the identification algorithm captures differences in texts, rather than differ-
ences in subjects. Therefore, we need to consider the experiments in a way that
the robustness of the identification procedure against the contents of the texts is
guaranteed. For this purpose two different experiments are taken into account to
investigate the text content effect on identification performance. In the first exper-
iment, different passages are used for different individuals, whereas in the second
experiment, the same set of passages is presented to all subjects. We use two data
sets (Dataset I and Dataset II), previously collected by Attar et al. (2016), that are
compatible with our experimental design. Datasets I and II meet the requirements
of our first and second experiment, respectively.

2.2 Experimental Setup

All screens were presented on a 22-inch View-Sonic LCD monitor with a refresh
rate of 75 Hz and resolution of 1024 x 768 pixels. Eye movements were monitored
with an SR Research EyeLink-2k system with a sampling frequency of 1000 Hz.
The passages that were used for data collection contain general topics (food,
science, health, history) taken from Washington Post News. Each passage has
between 230 and 240 words. The passages are easily readable texts in order to
decrease the influence of subjects’ prior knowledge on the experiments.
Forty native English speakers (25 female) with an average age of 20.4 years
(SD=5.35) and with normal or corrected to normal vision participated in the ex-
periments. For the first experiment with twenty participants (Dataset I), each
subject read six passages which were different for each subject. In the second ex-
periment that resulted in Dataset II, the other twenty subjects read six passages
which were identical for all subjects. Every passage in both experiments was di-
vided into 3 screens that added up to a total of 18 screens for six passages. The
font color for the text was black and the background color was grey.
4 Akram Bayat, Marc Pomplun

3 Feature Extraction

This section describes our feature extraction method, which consists of scan path
and pupillary response analysis. We aim to discriminate individuals by their visual
behavior during the reading task. The visual behavior is represented via a feature
set reflecting the dynamics of eye movement patterns.
It is important to choose the candidate features that provide the highest level
of specificity and noise tolerance. We attempt to choose features that represent the
patterns of eye movements that are observed during reading. Most of the features
used in this work are based on global processing, whereas word-by-word processing
is not specifically analyzed.
By considering the properties of eye-movement patterns in the reading task, we
extract features that hold promise as physiological and behavioral characteristics.
Moreover, the designed features should be less influenced by the content of the
particular texts used in the experiment. For instance, the frequency of the words
determines the likelihood that they are fixated. However, it is difficult to find
features that are text independent since eye-movement patterns are reactions to
the text stimulus. Besides, we use texts that require little prior knowledge in
order to decrease the influence of this knowledge on the features. Furthermore, all
subjects are at very similar age. For this reason, eye-movement patterns are not
significantly influenced by the readers’ age differences.
The raw data collected by the eye tracker system for each subject contains
various activities during reading, including fixations, saccades and blinks. Each of
these activities as well as the current gaze coordinates and pupil diameter were
measured at a temporal resolution of one millisecond.
A scan path is defined as the trajectory of eye movements over a screen, con-
taining a number of fixations and saccades with the number, duration, order and
placement of the fixations and saccades varying among individuals (Phillips and
Edelman, 2008). Analyzing the scan path of each subject during reading can lead
to measurable characteristics that are distinct for each subject. By analyzing a
scan path, we compute features that can considerably support biometric charac-
teristics. The extracted features are categorized into four groups: fixation features,
saccadic features, pupillary response features and spatial reading features.

3.1 Fixation Features

Different types of fixation based features are defined as follows:

Scan path length expresses the efficiency of reading which is computed as
the sum of the Euclidean distances between all consecutive fixations in the scan
path. Only fixations within the reading screen are considered for computing this
length, so any fixation outside of the screen is discarded.
Spatial Fixation count average is the number of fixations within the scan
path divided by the number of lines of text being read. The line spacing, font size
and formatting of text presentation are identical on all screens.
Fixation rate is the average number of fixations per second. It is computed
as the number of fixations over the sum of all fixation and saccade durations in a
scan path.
Biometric Identification through Eye-Movement Patterns 5

Fixation Speed is the number of fixations in a scan path over the total time
needed for reading a screen.
Average fixation duration is measured as the sum of fixation durations over
the number of fixations.

3.2 Saccadic Features

We define basic and complex saccadic features, which are saccade related metrics. If
a blink happens during any saccade, that saccade is removed from the computation
of baseline and complex features.
Average saccade duration is computed as the sum of saccade durations
over the total number of saccades.
Average horizontal (vertical) saccade amplitudes of at least 0.5 de-
grees are measured as the sum of horizontal (vertical) saccade amplitudes greater
than 0.5 degrees over the total number of saccades with horizontal (vertical) ampli-
tudes greater than 0.5 degrees. Horizontal and vertical saccade amplitudes indicate
between-word and between-line saccades, respectively (Holland and Komogortsev,
2011a).
Average saccade horizontal (vertical) Euclidean length is the sum of
horizontal (vertical) Euclidean distances between fixation locations over the total
number of saccades.
Average saccade velocity is defined as the sum of Euclidean norm of the
horizontal and vertical velocities over the total number of saccades in a scan path.
The horizontal (vertical) saccade velocity is defined as the velocity with
which eyes move horizontally (vertically) from a fixation point to another. A very
simple, fast and accurate way to compute saccade velocity is to use a two-point
central difference. If these two points are considered as two adjacent fixation points
(we discard all data points between two fixation points), the signal-to-noise ratio
will significantly decrease. A more robust way to compute the horizontal saccade
velocity(VHS ) and vertical saccade velocity (VV S ) is to use the following formulas
which are designed for a 1000 Hz sampling rate of data collection:
n
1X
VHS = VH (k) (1)
n
k=0

n
1X
VV S = VV (k) (2)
n
k=0

where
x([k + 3]T ) − x([k − 3]T )
VH (k) = (3)
6T
y([k + 3]T ) − y([k − 3]T )
VV (k) = (4)
6T
where x and y are the coordinates of a sample point within a saccade, T is the
sampling interval (1ms), and k is the index for discretized time, i.e. k = 0, 1, 2, ..., n.
Average peak velocity is the sum of peak velocities over the total number of
saccades in a scan path, where the peak velocity is defined as the highest velocity
reached between any two consecutive samples during the saccade.
6 Akram Bayat, Marc Pomplun

Average ratio of peak velocity over saccade velocity is the sum of

all ratios of peak velocity over saccade velocity divided by the total number of
saccades in a scan path.
Inflection count is the number of saccades in which the horizontal or vertical
velocity has a different sign than the preceding saccade. This value can indicate
direction of attention in progressing through the text (Holland and Komogortsev,
2012).

3.3 Pupillary Response Features

Pupillary response features consist of features that reflect changes in pupil size
during reading activity. The fact that the magnitude of pupil dilation is a function
of processing load or mental effort has long been known in neurophysiology. The
various changes in pupil diameters in different participants that result from their
reactions to the same reading task can be considered as dynamic features. In order
to model these changes, the standard deviation, average rate of pupil size change,
and difference between the minimum and maximum pupil size in a scan path are
computed.
The standard deviation of pupil diameter is applied as follows:
v
u N
u1 X (MF ixation (i) − MScanpath )2
σ= t (5)
N (MScanpath )2
i=1

Where MScanpath and MF ixation , respectively, are mean pupil diameters that are
observed during one scan path and during each fixation in that scan path, and N
is the number of fixations within the scan path.
The average rate of pupil size change, VP , is measured as follows:
n−2
1 X P ([k + 2]T ) − P ([k − 2]T )
VP = (6)
n−3 4T
k=2

where n is the number of fixations within a scan path, P is the pupil size, and T
is the sampling interval (1ms) of data collection.
The difference between minimal and maximal pupil diameter in each
scan path is measured as another pupil feature.

3.4 Spatial Reading Features

Spatial reading features reveal eye movement behavior in terms of efficiency. For
comprehending a sentence and a passage, readers must establish word order. It
means that their gaze moves to upcoming words in the text when they become
relevant for sentence comprehension. We define a progressive reading procedure as
moving forward toward the next words on a line of text. We define a saccade as a
progressive saccade if the saccade angle deviates from this direction by less than
20 degrees. Saccades that move the eyes in other directions are not considered to
belong to a progressive, efficient reading procedure. However, a saccade in opposite
Biometric Identification through Eye-Movement Patterns 7

Table 1 Classification accuracies for different classifiers.

Dataset Number of Subjects Classifier Accuracy

Logistic 97.57%
Dataset I 20 Multilayer Perceptron 96.53%
Logistic Model Tree 96.88%
Random Forest 96.87%
Logistic 92.76%
Dataset II 20 Multilayer Perceptron 92.50%
Logistic Model Tree 93.03%
Random Forest 90.08%
Logistic 93.04%
Dataset I & II 40 Multilayer Perceptron 91.53%
Logistic Model Tree 90.92%
Random Forest 89.26%

direction, landing on the next line of the text, is also counted as a progressive
saccade. Then we define the following features in this reading space:
Average horizontal (vertical) forward saccade velocities are measured
as the sum of progressive horizontal (vertical) saccade velocities over the total
number of progressive saccades.
Average absolute forward saccade velocities are the sum of absolute val-
ues of progressive saccade velocities over the total number of progressive saccades.

4 Classification

In this section, we describe our classification algorithm for our biometric user iden-
tification. Dataset I and Dataset II contain eye-movements data of forty subjects.
Each subject has read six passages with each passage being presented across three
successive screens. A feature vector is extracted for each screen that consists of the
scan path features in that screen. In this way, each subject has 28 feature vectors.
Feature extraction and classification are performed on Dataset I, Dataset II and
a combination of these datasets.

4.1 Individual Classifiers

Multi-class classification algorithms which consider each class label as a unique

subject are implemented using several classifiers including the Multilayer Percep-
tron, Logistic, and decision tree based classifiers (Logistic Model Tree and Random
Forest). Classifiers are trained and tested using a 10-fold cross validation method
on the set of extracted features. We primarily evaluate the performance of these
classifiers using their accuracies. The summary results are presented in Table ??.
The Logistic model outperforms the other classifiers for every dataset. The Mul-
tilayer Perceptron is the second most accurate classifier; its good performance
is in line with prior research on biometric identification (Rigas, Economou, and
Fotopoulos, 2012a). Considering both datasets, the best performance is achieved
8 Akram Bayat, Marc Pomplun

Table 2 Classification accuracies for different classifiers.

Dataset Number of Subjects Classifier Accuracy

Average of Propabilities 97.92%
Dataset I 20 Majority Voting 97.92%
Average of Propabilities 96.52%
Dataset II 20 Majority Voting 95.17%
Average of Propabilities 95.31%
Dataset I & II 40 Majority Voting 94.40%

with Dataset I as expected. Obtaining the highest accuracy when using Dataset
I reflects higher recognition power due to the effects caused by differences in the
texts that are used for that experiment.

4.2 Combined Classifier

Combining multiple good classifiers can improve accuracy, efficiency and robust-
ness over single classifiers. The idea is that different classifiers may offer com-
plementary and diverse information about patterns to be classified, allowing for
potentially higher classification accuracy (e.g., Bayat, Pomplun, and Tran, 2014).
We use the vote classifier to combine the optimal set of classifiers from the previous
section by selecting a combination rule. The average of probabilities is considered
as our combination rule that returns the mean of the probability distributions
for each of the single classifiers. It is found that for our biometric identification
model the average of probabilities yields a better result than other methods such
as majority voting (94.40%). The combination consisting of the Logistic, Multi-
layer Perceptron and Logistic Model Tree classifiers yields the highest accuracy in
all three datasets. The results of our combined classifiers are listed in Table ??.
The accuracy decreases from 97.92% to 96.52% when considering data collected
with the same set of passages for all subjects (Dataset II). The accuracy takes
another drop to 95.31% when combining Dataset I and Dataset II. The reason for
combining two datasets is to reduce the error rate exhibited by the content of the
text that is used for identification. The small changes in accuracy rates for differ-
ent datasets and the behavioral nature of our feature set suggest that our feature
set and classifiers can well capture unique characteristics of different individuals
by using their eye-movement patterns. For the remainder of this paper, the model
constructed by using the combination of Datasets I and II will be used as our
classifier.

5 Identification Performance

Evaluating the predictive performance of models is a vital step in model develop-

ment. Such evaluation assists in determining the suitability of a model for specific
applications (Pearce and Ferrier, 2000). Simple classification accuracy is often a
poor metric for measuring performance in biometric systems, which makes it nec-
essary to assess the performance of the predictive model by other characteristics
Biometric Identification through Eye-Movement Patterns 9

Fig. 1 FAR and FRR Errors for the multiclass classifier with 95.31% accuracy and decision
threshold with probability of 0.5. The maximum value of FAR and FRR over all classes is
0.62% and 33.33%, respectively.

such as a receiver operating characteristic (ROC) and equal error rate (EER).
An ROC curve is used to show proportions of correctly and incorrectly classified
predictions over a wide and continuous range of decision threshold levels.

In biometric systems, from the user’s point of view, an error of accuracy occurs
when the system identifies an invalid user or when the system fails to identify a
valid user. The associated error rates are called False Acceptance Rate (FAR)
and False Rejection Rate (FRR), respectively, which are the most commonly used
metrics for identification problems. EER is the rate at which both FAR and FRR
are equal. The value of the EER can be easily obtained from the ROC curve. In
general, lower EER indicates higher accuracy.

Our final combined classifier is a probabilistic multi-class classifier that yields

an instance probability where an instance is a feature vector that represents a user’s
eye-movement scan path and its relevance probability is a value that represents the
predicted degree to which an instance belongs to a given user. This probabilistic
classifier uses a decision threshold with a probability of 0.5 for each class to produce
a discrete classifier that reaches the accuracy value of 95.31%.

The general strategy in evaluating the FAR and FRR errors for multi-class
classifiers is to reduce the problem of multiclass classification to multiple binary
classifications in which each class has its own value of FAR and FRR errors. Fig-
ure ?? shows the values of FAR and FRR errors for all forty classes. The combined
multi-class classifier achieves 95.31% probability of identification at average 0.134%
FAR and 6.051% FRR.
10 Akram Bayat, Marc Pomplun

Fig. 2 FAR and FRR versus decision threshold for two-sample binary classes; the EER value
serves as the decision threshold. In panel a) FAR= FRR in Threshold ' 0.5. In panel b) FAR
6= FRR in Threshold ' 0.5.

5.1 Optimizing the Performance of Identificationl

The combined classifier that obtains the accuracy of 95.31% uses a decision thresh-
old with a probability of 0.5 to map from instances to predicted classes. However,
it does not guarantee providing us with the optimal values of FAR and FRR er-
rors. Figure ?? shows the ROC graphs for two sample binary classes (two users)
in which FAR and FRR are plotted on the vertical axis and decision threshold is
plotted on the horizontal axis. Figure ??a shows an EER value of 0.0015 for both
FAR and FRR related to the decision threshold value of 0.49 for a sample binary
class that guarantees 0.15% EER or less for this class.
Selecting the EER value as the decision threshold is often a good choice for
identification applications. In Figure ??b, the FAR and FRR errors are not optimal
at a threshold value of 0.5, because the value of FAR (0.277) at a threshold of 0.5
is fairly high for an identification problem; it means that the classifier would fail
at 27.7 percent of attempts to identify a valid user.
ROC analysis and EER based decision making are commonly employed in two-
class classification problems because they are easy to analyze and visualize. For
our multi-class classifier we need to deal with more than one decision threshold
and make multi-decisions for our multi-class predictive model.
It is important to know that with more than two classes, ROC analysis be-
comes too complex to be managed. In our multi-class model, with 40 classes, the
confusion matrix becomes a 40×40 matrix containing the 40 correct classifications
(the major diagonal entries) and 1560(402 − 40) possible errors (the off-diagonal
entries). Instead of managing trade-offs between FAR and FRR, we have 1560
errors (Fawcett, 2006).
Biometric Identification through Eye-Movement Patterns 11

Fig. 3 EER vs. decision threshold for our proposed multiclass model.

One method for handling n classes is to consider each class as an independent

binary problem and to produce n different ROC graphs, one for each class. It
means that for any class in the set of all classes, the ROC graph plots classification
performance only for that class as the positive class and all other classes as the
negative class. In this way, we will have 40 decision thresholds and 40 EER values
for our 40-class identification model that are represented in Figure ??.
As a result, our classification model achieves an EER value no greater than
6.1% and an average EER rate of 2.03% for identifying users based on their eye-
movement biometrics. As can be seen, by adjusting thresholds for having equal
error rates of FAR and FRR, we obtain a lower average rate of FRR (2.03% <
6.051%) but a greater rate of FAR (2.03% > 0.134% ).

5.2 Optimal Differential Error Rate (ODER)

An identification of users is possible using our classifier in conjunction with a

threshold adjusting method to reach the EER. However, we propose a new criterion
(ODER) to tune the decision thresholds to achieve a desired tradeoff between FAR
and FRR suitable for our identification application. The reason behind designing
our criterion is that while EER is a common measure for making decisions about
the thresholds, it does not guarantee the optimal error rates of FAR and FRR for
our application. For instance, for one of our sample binary classes with an EER
of 3.2%, a threshold is set to a low value of 0.05. A better value of the FAR rate
(.91%) is achieved by considering a higher threshold of 0.33; however, this causes
an increase in the FRR rate (10.9%). In our identification problem, FAR should
be preferably very low in order to avoid a high number of false identifications.
Therefore, we optimize the operating point for the threshold on each binary class by
a new criterion which is different than the commonly recommended EER operating
12 Akram Bayat, Marc Pomplun

Table 3 Error rate of identification for different threshold choosing criteria.

Threshold Error Maximum Minimum Average
Adjusted Threshold EER 6.1% 0% 2.03%
0.5 FAR 0.62% 0% 0.134%
0.5 FRR 33.33% 0% 6.051%
Adjusted Threshold ODER FAR 0.47% 0% 0.1%
Adjusted Threshold ODER FRR 16.67% 0% 3.9%

point. Our objective of this optimization is to obtain a compromise between a very

small FAR and a small FRR instead of having FAR = FRR. By setting a higher
threshold value, a lower FAR can be obtained. The proposed criterion (ODER) is
described in Algorithm 1. This algorithm minimizes FAR as follows:

min(F AR), given (F AR − F RR) ≤ 10−1 (7)

The results of implementing ODER is listed in Table ?? and illustrated in Fig-

Algorithm 1 The ODER gives the optimal values of FAR and FRR and their
corresponding threshold value for our identification problem
min F AR ← 1
loop . for all values of thresholds in each class
if | F AR − F RR |≤ 10−1 then
if F AR ≤ min F AR then
min F AR ← F AR
end if
end if
end loop

ure ??. ODER outperforms the EER by decreasing FAR from 2.03% to 0.1%,
while the FRR increase from 2.03% to 3.9%. In order to precisely compare EER
and ODER criteria for deciding the value of the decision threshold, we compute
the relative changes in FAR and FRR for each criterion. In this way, in ODER,
FAR decreases by 95.07% and FRR increases by only 47.9%. This result suggests
that the ODER algorithm outperforms the EER-based threshold decision making.

6 Related Work

The idea of using biometric identification by itself or by fusing it with other stan-
dard identification methods has been studied for many years. Finger prints, face,
DNA, voice and gait are some of the physiological and behavioral characteristics
that have been used in biometric identifiers. In recent years, the potential of eye-
movement tracking as a biometric has been investigated as an additional way that
can be integrated with other biometrics. The involuntary eye movements could
reflect the underlying anatomical organization which is unique for each individual.
Bednarik et al. (2005) present a first step towards using eye movement as
a biometric identification. An identification rate of 60% by using dynamics of
Biometric Identification through Eye-Movement Patterns 13

Fig. 4 Comparing average FAR and FRR for each of three different criteria (Decision thresh-
old = 0.5, EER, and ODER).

pupil diameters is obtained. Based on that, they suggest an integration of the

eye-movement based identification into general video-based biometric systems.
Rigas, Economou, and Fotopoulos (2012b) monitor individuals’ allocation of
attention while they observe various face images. They use a non-parametric mul-
tivariate graph-based measure for the comparison of eye-movement signals that
yields 30.0% EER and 70.2% Rank-1 identification rate on a pool of 15 subjects.
Kinnunen, Sedlak, and Bednarik (2010) capture a subject’s eye movements
while he is watching a movie and use short-term eye-gaze direction to construct
feature vectors which are modeled using Gaussian mixtures. They achieve equal
error rates ranging from 28.7% to 47.1% depending on the size of the training
dataset.
In the work of Holland and Komogortsev (2013), complex eye movement fea-
tures serve as the basis for the Complex Eye Movement Behavior biometrics, where
eye movement features are represented as statistical distributions. The best results
for a database of 32 subjects are a minimal EER of 16.5% and a maximum Rank-1
identification rate of 82.6%.
Rigas and Komogortsev (2014b) present a method based on Fixation Density
Maps (FDMs) for representing overt visual attention as a biometric template. They
use a large database of 200 subjects with the resulting best EER of 10.8% and
Rank-1 identification rate of 51%.
Cantoni et al. (2015) attempt to distinguish individuals by the way they gaze
in sixteen different images. Using density and duration of fixations as their main
features, they report an EER of 27.06%.
14 Akram Bayat, Marc Pomplun

Liang et al. (2015) present a video-based biometric identification model in

which overt visual attention features, such as acceleration, geometric, and mus-
cle properties, are obtained from gaze data and employed as biometric traits to
recognize people. Experiments were carried out using a Backpropagation neural
network and a Support Vector Machine. In an experiment with five subjects, an
identification accuracy of 82% is achieved.
All previous work mentioned in the Introduction uses different visual tasks
rather than a reading task as employed in the present study. While there have been
many research studies using eye-movement tracking while subjects are reading a
text (Rayner, 1998); Vitu et al., 1995), only a few studies consider it as a method
of cooperation between the user and the system for biometric identification.
The work of Holland and Komogortsev (2011b) is one of the most successful
studies in biometric identification via eye-movement tracking in reading. Scan path
characteristics, representative of the observer’s visual attention strategies during
reading, were investigated. Eye movements were recorded for 32 subjects over four
recording sessions. The best reported result yielded an EER of 27%.

7 Conclusions

In this paper, identification accuracy of up to 95.31% and an average EER rate

of 2.03% were obtained from analyzing readers’ eye movements. The data were
acquired from 40 subjects under laboratory conditions. A new set of features was
taken into account and different performance metrics were used for evaluating
identification performance. Combining the three best classifiers using the average
of probabilities method turned out to be the best classifier for our biometric iden-
tification problem, outperforming all individual classifiers in terms of accuracy.
We further showed that our multi-class probabilistic classifier can be improved
in performance by selecting good decision thresholds that guarantee an average
equal error rate of 2.03%, which is significantly lower than the EERs achieved by
the best eye-movement classifiers in the literature. Our findings indicate that the
incorporation of the proposed features, classification methods, and optimization
of decision thresholds can lead to a significant improvement in biometric accuracy
and robustness in identification through reading tasks. The unique characteristics
of eye movements and promising results of our identification method make this
biometric an excellent approach for various applications such as personalized user
interfaces or integration with other identification methods in order to enhance the
level of security. While previous work already suggested that eye movements could
be used for such purposes, the drastic increase in accuracy achieved by our tech-
nqiue makes this paradigm relevant for actual applications for the first time. For
future work, we plan to extend our biometric identification task in several ways.
First, we intend to collect data from more users of various ages. Second, we plan
to consider eye-movement similarities among subjects and exploit interdependence
among various classes. The goal of this line of reserach is to increase the robustness
of our method for a wide range of users while requiring as few eye movements as
possible for reliable identification.
Biometric Identification through Eye-Movement Patterns 15

Acknowledgements The authors thank Ms. Nada Attar in the Visual Attention Laboratory
at the University of Massachusetts Boston for providing them with data for evaluating the
eye-movement classifiers presented in this study.

References

1. Jain, A.K., Ross, A. and Prabhakar, S. (2004). An introduction to biometric recognition.

Circuits and Systems for Video Technology, IEEE Transactions on, 14(1), pp.4-20.
2. Bednarik, R., Kinnunen, T., Mihaila, A. and Frnti, P. (2005). Eye-movements as a biomet-
ric.In Image analysis (pp. 780-789). Springer Berlin Heidelberg.
3. Komogortsev, O.V. and Khan, J.I. (2008). Eye movement prediction by Kalman filter with
integrated linear horizontal oculomotor plant mechanical model. In Proceedings of the 2008
symposium on Eye tracking research & applications (pp. 229-236). ACM.
4. Holland, C. and Komogortsev, O.V. (2011).Biometric identification via eye movement scan-
paths in reading. In Biometrics (IJCB), 2011 International Joint Conference on (pp. 1-8).
IEEE.
5. Phillips, M.H. and Edelman, J.A. (2008). The dependence of visual scanning performance
on saccade, fixation, and perceptual metrics. Vision research, 48(7), pp.926-936.
6. Bayat, A., Pomplun, M. and Tran, D.A. (2014). A study on human activity recognition
using accelerometer data from smartphones. Procedia Computer Science, 34, pp.450-457.
7. Pearce, J. and Ferrier, S. (2000). Evaluating the predictive performance of habitat models
developed using logistic regression. Ecological modelling,133(3), pp.225-245.
8. Fawcett, T. (2006). An introduction to ROC analysis. Pattern recognition letters, 27(8),
pp.861-874.
9. Bednarik, R., Kinnunen, T., Mihaila, A. and Frnti, P. (2005). Eye-movements as a biometric.
In Image analysis (pp. 780-789). Springer Berlin Heidelberg.
10. Rigas, I., Economou, G. and Fotopoulos, S. (2012). Biometric identification based on the
eye movements and graph matching techniques. Pattern Recognition Letters, 33(6), pp.786-
792
11. Kinnunen, T., Sedlak, F. and Bednarik, R. (2010). Towards task-independent person au-
thentication using eye movement signals. In Proceedings of the 2010 Symposium on Eye-
Tracking Research & Applications (pp. 187-190). ACM.
12. Cantoni, V., Galdi, C., Nappi, M., Porta, M. and Riccio, D. (2015). GANT: Gaze analysis
technique for human identification. Pattern Recognition, 48(4), pp.1027-1038.
13. Rayner, K. (1998). Eye movements in reading and information processing: 20 years of
research. Psychological bulletin, 124(3), p.372.
14. Vitu, F., ORegan, J.K., Inhoff, A.W. and Topolski, R. (1995). Mindless reading: Eye-
movement characteristics are similar in scanning letter strings and reading texts. Perception
& Psychophysics, 57(3), pp.352-364.
15. Silver, D.L. and Biggs, A. (2006). Keystroke and Eye-Tracking Biometrics for User Iden-
tification. In IC-AI (pp. 344-348).
16. Holland, C.D. and Komogortsev, O.V. (2012). Biometric verification via complex eye move-
ments: The effects of environment and stimulus. In Biometrics: Theory, Applications and
Systems (BTAS), 2012 IEEE Fifth International Conference on (pp. 39-46). IEEE.
17. Attar N., Wu, C., Sia, D. and Pomplun, M. (2016). A Deeper Understanding of Optimal
Viewing Position Using Eye Fixations and Character Recognition on Text-Viewing and Read-
ing Tasks. ACM ETRA: 2016 Symposium on Eye Tracking Research & Applications(pp.
209-212).
18. Holland, C.D. and Komogortsev, O.V. (2013), June. Complex eye movement pattern bio-
metrics: Analyzing fixations and saccades. In Biometrics (ICB), 2013 International Confer-
ence on (pp. 1-8). IEEE.
19. Rigas, I. and Komogortsev, O.V. (2014). Biometric recognition via probabilistic spatial
projection of eye movement trajectories in dynamic visual environments. Information Foren-
sics and Security, IEEE Transactions on, 9(10), pp.1743-1754.
20. Attar, N., Fomenky, P., Ding, W. and Pomplun, M. (2016): Improving Cognitive
Load Level Measurement through Preprocessing Psychological Data by Random Sub-
space Method. IEEE 2nd International Conference on Human Computer Interactions
(ICHCI16)(pp. 1-6).