AgeGen Bio Track: Continuous Mouse Behavioral Biometrics-Based Age

and Gender Proﬁling in Online Education Platforms

Aditya Subash

1 a

, Insu Song

1 b

, Ickjai Lee

2 c

and Kyungmi Lee

2 d

College of Science and Engineering, James Cook University, Singapore

College of Science and Engineering, James Cook University, Cairns, Australia

Keywords:

Mouse Behavioral Biometrics, Age Classiﬁcation, Gender Classiﬁcation, Online Education Platforms.

Abstract:

Mouse behavioral biometric-based authentication systems have attracted signiﬁcant attention as they are con-

sidered a more secure alternative to conventional online assessment fraud detection systems. This is attributed

to their ability to continuously authenticate users non-intrusively by analyzing their distinctive mouse operat-

ing behavior. Most behavioral biometric-based research studies focus on predicting user identity as the primary

objective for online assessment fraud detection. However, they do not consider predicting other user-centric

parameters like age and gender. Furthermore, there is a need to identify the best segmentation approach and

mouse behavior feature set for age and gender classiﬁcation. We propose the AgeGen Bio track system, a

continuous mouse behavioral biometric-based age and gender tracking system for online education platforms.

To accomplish this, we ﬁrst collect novel mouse behavior data with user demographic information. We then

evaluate the efﬁcacy of different segmentation approaches, feature sets, and machine learning models for age

and gender classiﬁcation. Experimental results show that the random forest algorithm paired with the three

mouse-movement segmentation approach and user characteristic feature set are the best approaches that need

to be incorporated into the system, as they achieved promising results.

1 INTRODUCTION

Recently, the education sector has advanced from of-

ﬂine to online settings over the past decade due to ad-

vancements in information technology (IT), resulting

in widespread accessibility and proliferation of on-

line education throughout the world (Garg and Goel,

2022). Several higher educational institutions (HEIs)

now offer online courses as part of blended or fully

online education (Garg and Goel, 2022; Wei et al.,

2021). Like ofﬂine education, assessments are in-

tegral to any online educational curriculum and are

often organized to evaluate learning outcomes, sub-

ject application, and knowledge retention (Garg and

Goel, 2022). Despite the advantages of online assess-

ments, a signiﬁcant challenge is the prevalence of as-

sessment fraud or academic dishonesty, often facili-

tated by the misuse of digital technologies (Blau and

Eshet-Alkalai, 2017; Susnjak and McIntosh, 2024).

https://orcid.org/0000-0003-0196-6256

https://orcid.org/0000-0002-9835-0070

https://orcid.org/0000-0002-6886-6201

https://orcid.org/0000-0003-3304-4627

These technologies make it straightforward to commit

online assessment fraud by offering paid services and

tools that assist students in credential sharing, fake

identity matching, and plagiarism (Susnjak and McIn-

tosh, 2024; Noorbehbahani et al., 2022). The current

methods to prevent online assessment fraud, includ-

ing conventional authentication approaches, are one-

time, non-repudiable, intrusive, and expensive (Sid-

diqui et al., 2021; Subash et al., 2024).

Mouse behavioral biometric-based authentication

systems have become a more popular and secure al-

ternative to online assessment fraud detection. This is

due to their cost effectiveness and ability to continu-

ously verify user identity based on observable mouse

movement behaviors (Zheng et al., 2011), which is

inherently more challenging to spoof or replicate,

making the approach more robust. Most behav-

ioral biometric-based research studies focus on pre-

dicting user identity (ID) as the primary objective

for security-related applications. However, they do

not consider predicting other user-centric parameters,

such as age and gender. Despite many studies in the

ﬁeld, there is no research and datasets for mouse be-

havioral biometric-based age and gender recognition

Subash, A., Song, I., Lee, I. and Lee, K.

AgeGen Bio Track: Continuous Mouse Behavioral Biometrics-Based Age and Gender Proﬁling in Online Education Platforms.

DOI: 10.5220/0013138500003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 383-393

ISBN: 978-989-758-737-5; ISSN: 2184-433X

383

in online education platforms. Furthermore, we must

identify the best segmentation approach and mouse

behavior feature set for our speciﬁc application.

We propose the AgeGen Bio track system, a con-

tinuous mouse behavioral biometric-based age and

gender tracking system for online education plat-

forms. To accomplish this, we ﬁrst collect novel

mouse behavior data with user demographic infor-

mation. We then evaluate the efﬁcacy of differ-

ent segmentation approaches, feature sets, and ma-

chine learning models for age and gender classiﬁca-

tion. Speciﬁcally, we will evaluate several machine

learning (ML) approaches, including Decision Tree

(DT), Random Forest (RF), Naive Bayes (NB), Sup-

port Vector Machine (SVM), Logistic regression (LR)

and k-Nearest Neighbor (k-NN). The main contribu-

tions of the paper are as follows.

• An AgeGen Bio Track system capable of contin-

uously monitoring user age and gender in online

education platforms.

• Newly collected task-speciﬁc mouse behavior

data with user-centric information from a case

study for our experimentation.

• Comprehensive study comparing several segmen-

tation approaches, feature sets, and machine

learning models to identify the most effective ap-

proaches that can be integrated into the system.

2 BACKGROUND

This section will highlight the importance of includ-

ing age and gender prediction in mouse behavioral

biometrics-based online assessment fraud detection.

According to previous research (Van Balen et al.,

2017), men and women differ in physical dimensions

as described by anthropometric data recorded from

several sources. These attributes can signiﬁcantly im-

pact the maneuverability of peripheral (mouse) de-

vices across 2-dimensional surfaces, as it requires

coordination between the arms, wrist, and ﬁngers

(Van Balen et al., 2017). Research conducted by (Fr-

yar et al., 2012) indicates signiﬁcant differences in

arm lengths between males and females across several

age groups. These variations could lead to differences

in mouse movements (Van Balen et al., 2017).

Furthermore, several research studies regarding

motor behavior indicate that men tend to move faster

with less accuracy than women (Barral and Deb

2004; Rohr, 2006). For example, (Rohr, 2006) con-

ducted a study by requesting subjects to participate in

a mouse-pointing task that required them to click tar-

gets of various sizes across the midline of a device.

According to the study, women showed greater accu-

racy and slower deceleration time than men during the

ballistic component of mouse movement (Van Balen

et al., 2017).

Similarly, (Jim

enez-Jim

enez et al., 2011) also in-

vestigated the effect of age and gender on motor be-

havior. A total of 246 participants (123 males and 123

females) were recruited from seven age groups for the

study. Participants were required to perform a set of

physical and computer tasks. Parameters such as ﬁn-

ger tapping frequency, movement time, walking time,

and visual reaction time were measured for analysis.

Results indicate that age and gender play a signiﬁ-

cant role in motor behavior. Furthermore, the speed

of motor performance was found to be better in men

(Jim

enez-Jim

enez et al., 2011).

Studies have also conﬁrmed signiﬁcant differ-

ences between typing speeds of different age groups

due to generational differences. A study by (Pentel,

2017) conﬁrmed that participants in the age group 16-

19 were faster at typing than other age groups.

The aforementioned studies clearly indicate that

age and gender characteristics are essential in under-

standing distinctive user behaviors, making them im-

portant parameters that must be included in current

mouse behavioral biometric-based online assessment

fraud detection systems.

3 RELATED WORK

This section will critically analyze, summarize, and

present ﬁndings on several previous studies on be-

havioral biometric-based age and gender prediction.

Speciﬁcally, we will summarize the several datasets,

data collection strategies, features, and AI methodolo-

gies implemented in current research. For this pur-

pose, we review several papers from well known pub-

lishers, including IEEE, Science Direct, and Springer.

Given the variety of behavioral biometrics studies,

we will focus on reviewing articles related to age and

gender classiﬁcation using keystroke or mouse behav-

ior biometrics.

3.1 Datasets and Data Collection

Strategies

According to our review, datasets are classiﬁed into

1) Task-speciﬁc and 2) unconstrained datasets. Task-

speciﬁc datasets collect user behavior based on prede-

termined mouse operation tasks (Zheng et al., 2011;

Van Balen et al., 2017). Meanwhile, unconstrained

datasets collect mouse behavior data by continuously

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

384

monitoring users while they perform their daily ac-

tivities without any constraints (Zheng et al., 2011;

Van Balen et al., 2017). These types of data are col-

lected to perform static or dynamic authentication.

We will now describe the various types of data

used in research. This includes brieﬂy describ-

ing public and novel datasets used in behavioral

biometric-based age group and gender prediction.

3.1.1 Novel Datasets

1. Van Balen et al. (2017) collected mouse behavior

data from 94 participants in a controlled envi-

ronment for gender classiﬁcation. Participants

were required to perform a speciﬁc predeter-

mined task involving identifying and clicking

certain targets of different sizes located in 16

possible locations. Once the target is clicked, a

new target will be displayed. Each participant

performed several practice trials consisting of

combinations of target size, target distance,

and approach angle. After practice sessions,

participants were then required to perform

four blocks of 64 movement trials, with each

block containing random sequences of two trials

for each combination of target locations and sizes.

2. Tsimperidis et al. (2017, 2018, 2021) gathered

unconstrained keystroke behavior data to pre-

dict age, gender, and educational qualiﬁcations

(Tsimperidis et al., 2020). The data was captured

using a key logger (IRecU) while participants

engaged in their daily activities. Participants

were also requested to provide demographic and

educational details alongside the keystroke data.

According to the authors, keystroke behavior

data was collected from four age groups: 18-25,

26-35, 36-45, and 46+.

3. Tsimperidis et al. (2015) collected keystroke be-

havior data from 24 subjects using a key-logging

application while participants were typing a ﬁxed

text of 850 characters, twice on a laptop and a

desktop. In addition to keystroke behavior data,

gender, and left/right-handedness information

was also obtained (Tsimperidis et al., 2015).

4. Idrus et al. (2013, 2014) collected static keystroke

behavior data from 110 participants for age, gen-

der, and handedness classiﬁcation. According to

the study, participants belonged to two different

countries, namely, France and Norway. Two key-

boards with various keyboard layouts, AZERTY

and QWERTY, are used during data collection.

The data collection procedure involved partic-

ipants writing ﬁve common phrases 20 times

each. Furthermore, they were required to write

the phrases ten times with one hand and ten times

with two hands. According to the author, this

data will be publicly available. Data acquisition

was conducted using software developed using

the publicly available GREYC software.

5. Pentel (2017, 2019) collected uncontrolled online

key-stroke and mouse behavior data for age group

and gender classiﬁcation. User behavior data

was collected from six sources, including the

school’s internal management system, feedback

questionnaires, testing environments, and con-

trolled experiments. Data was acquired using a

JavaScript key-logging tool integrated into all six

sources. According to the author, user behavior

data was collected between 2011 and 2017 from

several different age groups. Additional data such

as screen resolution, device type (laptop, desktop,

mobile devices), and operating system were also

collected.

6. Kolakowska et al. (2016) collected keystroke

and mouse behavior from 42 participants in a

completely uncontrolled environment. Out of

42, 9 were females, and 33 were males. Data

acquisition involved developing a browser plug-in

that recorded relevant keystrokes and mouse be-

havior data while participants performed several

activities in the browser. According to the study,

different plug-in versions were made for Chrome

and Opera browsers. Furthermore, user behavior

data was collected from various age groups and

peripheral devices (mouse, touchpad, trackpoint,

touchpad).

3.1.2 Public Datasets

1. Buriro et al. (2016) used a publicly available

keystroke dataset (TDAS) for age, gender, and

operating hand estimation. The dataset col-

lected keystroke data from 150 participants across

six distinct age groups. Keystroke behavior

data was collected while participants typed two

PINs, one 4-digit (5560) and one 16-digit PIN

(137966662480852), using a Samsung Galaxy

Tablet. Furthermore, this dataset was collected in

three different user-tagged location

2. Studies also implemented the publicly available

GREYC keystroke behavior dataset for gender

AgeGen Bio Track: Continuous Mouse Behavioral Biometrics-Based Age and Gender Proﬁling in Online Education Platforms

385

Table 1: Datasets implemented by behavioral biometric-based age and gender research studies.

Author Modality Participants Environment Public Type Purpose

Van Balen et

al., 2017

Mouse 94 Controlled No Task Speciﬁc Gender

Classiﬁcation

Buriro et al.,

2016

Keystroke 150 Semi-

Controlled

Yes Task Speciﬁc Age/Gender/

Hand

Classiﬁcation

Tsimperidis et

al., 2017

Keystroke - Uncontrolled No Unconstrained Age

Classiﬁcation

Tsimperidis et

al., 2018

Keystroke 75 - No Unconstrained Gender

Classiﬁcation

Tsimperidis et

al., 2015

Keystroke 24 Uncontrolled No Unconstrained Gender

Classiﬁcation

Fairhurst and

Da Costa-

Abreu, 2011

Keystroke 133 Controlled Yes Task Speciﬁc Gender

Classiﬁcation

Giot and

Rosenberger,

2012

Keystroke 133 Controlled Yes Task Speciﬁc Gender

Classiﬁcation

Idrus et al.,

2014

Keystroke 110 Controlled Yes Task Speciﬁc Age/Gender/

Hand

Classiﬁcation

Idrus et al.,

2013

Keystroke 110 Controlled Yes Task Speciﬁc Gender/Hand

Classiﬁcation

Pentel, 2017 Keystroke

/Mouse

1519 Uncontrolled Yes Unconstrained Age/Gender

Classiﬁcation

Tsimperidis et

al., 2021

Keystroke 118 Uncontrolled No Unconstrained Age/Gender

Classiﬁcation

Pentel, 2019 Keystroke 7119 Uncontrolled Yes Unconstrained Age

Classiﬁcation

Kolakowska

et al., 2016

Keystroke

/Mouse

42 Uncontrolled No Unconstrained Gender

Classiﬁcation

prediction. This dataset contains data from 133

participants who were required to type a prede-

termined password several times. In addition to

keystroke data, gender data was also acquired.

Among 133 participants, 98 are male and 35

are female (Fairhurst and Da Costa-Abreu, 2011;

Giot and Rosenberger, 2012).

From our review (Table 1), it is evident that re-

search on keystroke behavior biometrics has seen a

marked increase, particularly in the context of pre-

dicting age and gender, in comparison to mouse be-

havioral biometrics. Furthermore, there are very

few publicly available datasets for mouse behavioral

biometric-based age and gender prediction.

3.2 Behavior Biometric Features

We will discuss the various features employed in be-

havioral biometric-based age and gender classiﬁca-

tion. Given that our background review encompasses

multiple biometric modalities, we will categorize and

emphasize the features according to each modality.

Speciﬁcally, we will focus on keystroke and mouse

behavior-based features for age and gender predic-

tion.

3.2.1 Keystroke Behavior Features

After data collection, raw data such as key code

(Tsimperidis et al., 2017; Tsimperidis et al., 2018;

Syed Idrus et al., 2014; Idrus et al., 2013; Pentel,

2017; Pentel, 2019), action type (key press or release)

(Tsimperidis et al., 2017; Tsimperidis et al., 2018;

Syed Idrus et al., 2014; Idrus et al., 2013; Pentel,

2017; Pentel, 2019; Kolakowska et al., 2016), the date

the action took place, and timestamp are collected and

(Tsimperidis et al., 2017; Tsimperidis et al., 2018;

Syed Idrus et al., 2014; Idrus et al., 2013; Pentel,

2017; Pentel, 2019; Kolakowska et al., 2016) are used

for feature extraction.

Using this data, dwell (hold) time (Buriro et al.,

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

386

Table 2: AI-based approaches used in behavioral biometric-based age and gender classiﬁcation.

Author DL ML Fusion DBC

Van Balen et al.,

2017

No LS-MR + LR No No

Buriro et al.,

2016

NN, DNN NB, SVM, RF No No

Tsimperidis et al.,

2017

ANN (MLP) OneR, SVM,

BFDT, DT, NB,

No No

Tsimperidis et al.,

2018

MLP, RBFN SVM, RF, NB No No

Tsimperidis et al.,

2015

No NB, SVM, RF No Manhattan,

Euclidean

Fairhurst and

Da Costa-Abreu,

2011

No KNN, NB, DT DCS-LA,

Majority

Voting, Sum

Giot and Rosen-

berger, 2012

No SVM Score,

Template

Idrus et al., 2014 No SVM Majority

Voting,

Score

Idrus et al., 2013 No SVM No No

Pentel, 2017 No LR, SVM, KNN,

DT, RF

No No

Tsimperidis et al.,

2021

RBFN SVM, SL, NB,

BNC

No No

Pentel, 2019 No SVM, RF, FT, LR No No

Kolakowska et

al., 2016

NN BNC, RT, DT,

AdaBoost

No No

2016; Fairhurst and Da Costa-Abreu, 2011; Pentel,

2017; Tsimperidis et al., 2021), release-press latency

(Buriro et al., 2016; Fairhurst and Da Costa-Abreu,

2011; Giot and Rosenberger, 2012; Syed Idrus et al.,

2014; Pentel, 2017), release-release latency (Buriro

et al., 2016; Fairhurst and Da Costa-Abreu, 2011;

Giot and Rosenberger, 2012; Syed Idrus et al., 2014),

press-release latency (Buriro et al., 2016; Fairhurst

and Da Costa-Abreu, 2011; Giot and Rosenberger,

2012; Syed Idrus et al., 2014), press-press latency

(Buriro et al., 2016; Fairhurst and Da Costa-Abreu,

2011; Giot and Rosenberger, 2012; Syed Idrus et al.,

2014; Tsimperidis et al., 2021), and a vector combi-

nation of all the previous latencies (Giot and Rosen-

berger, 2012; Syed Idrus et al., 2014; Idrus et al.,

2013), are extracted and used for age and gender clas-

siﬁcation.

Detailed analysis also reveals that studies fre-

quently extract the most common diagram patterns

and compute the average and standard deviation val-

ues of n-graph-latencies and hold times for analysis

(Tsimperidis et al., 2017; Tsimperidis et al., 2018;

Pentel, 2017; Pentel, 2019). Additional features ex-

tracted include the relative frequency of corrective

keys (DEL) (Pentel, 2017; Kolakowska et al., 2016),

average time (Pentel, 2017), correctness (ratio be-

tween the number of keystrokes and the number of

characters in the ﬁnal text) (Pentel, 2017), percent-

age use of special character keys (Kolakowska et al.,

2016), pauses between words (Pentel, 2019), and sta-

tistical measurements of (average, standard deviation,

maximum, minimum, variance, mode, and range) of

typing speed, hold time, and press-press latency were

also computed (Kolakowska et al., 2016).

Studies were also found to integrate gender classi-

ﬁcation into keystroke authentication systems for en-

hanced performance (Giot and Rosenberger, 2012).

Feature selection algorithms, such as Information

Gain (IG), and oversampling techniques, including

SMOTE, have also been implemented for analysis

(Buriro et al., 2016; Tsimperidis et al., 2018; Tsim-

peridis et al., 2021).

3.2.2 Mouse Behavior Features

Raw mouse behavior data, such as timestamp

(Van Balen et al., 2017; Pentel, 2017), coordinates

space (x, y) (Van Balen et al., 2017; Pentel, 2017), ac-

AgeGen Bio Track: Continuous Mouse Behavioral Biometrics-Based Age and Gender Proﬁling in Online Education Platforms

387

tion type (Van Balen et al., 2017; Kolakowska et al.,

2016), and location/size of the target (Van Balen et al.,

2017), are recorded and used for feature extraction.

Using this data, several different types of mouse

behavior features, such as temporal, spatial, and accu-

racy metrics, were calculated (Van Balen et al., 2017).

These metrics can again be subdivided into sev-

eral types of features, including reaction time (RT),

peak velocity (PK), time to peak velocity (TPV),

duration of ballistic movement (DB), shape of ve-

locity proﬁle (SV), proportion of ballistic move-

ment (PB), number of movement corrections (NC),

time to click (TC), hold time (HT), movement time

(MT), path length (PL), path length to best path ratio

(PLR), task axis crossings (TXC), movement direc-

tion changes (MDC), orthogonal movement changes

(MDC), movement variability (MV), absolute error

(AE), horizontal error (HE), vertical error (VE), ab-

solute horizontal error (AHE), and absolute vertical

error (AVE) (Van Balen et al., 2017). Additional

attributes such as distance, angle, velocity, move-

ment, acceleration, action, and direction-based fea-

tures were also implemented by current research stud-

ies (Pentel, 2017; Kolakowska et al., 2016).

3.3 Machine Learning Approaches

This section highlights the several AI methods used

in mouse behavioral biometric-based age and gender

analysis. On analysis, we ﬁnd that approaches, in-

cluding logistic regression (LR) (Pentel, 2017; Pentel,

2019), support vector machine (SVM) (Buriro et al.,

2016; Tsimperidis et al., 2017; Tsimperidis et al.,

2018; Tsimperidis et al., 2015), random forest (RF)

(Buriro et al., 2016; Tsimperidis et al., 2018; Pen-

tel, 2017), k-nearest neighbors (KNN) (Fairhurst and

Da Costa-Abreu, 2011; Pentel, 2017), OneR (Tsim-

peridis et al., 2017), best ﬁrst decision tree (BFDT)

(Tsimperidis et al., 2017), rotation forest (RT) (Ko-

lakowska et al., 2016), AdaBoost (Kolakowska et al.,

2016), simple logistics (SL) (Tsimperidis et al., 2017;

Tsimperidis et al., 2021), decision tree (DT) (Tsim-

peridis et al., 2017; Tsimperidis et al., 2015; Fairhurst

and Da Costa-Abreu, 2011; Pentel, 2017), Bayesian

network classiﬁer (BNC) (Tsimperidis et al., 2021;

Kolakowska et al., 2016), and na

ıve Bayes (NB)

(Buriro et al., 2016; Tsimperidis et al., 2017; Tsim-

peridis et al., 2018; Tsimperidis et al., 2015) are some

of the popular ML approaches implemented in the

ﬁeld. A few studies also use a combination of ML ap-

proaches for classiﬁcation. For example, Van Balen et

al. (2017) implements a combination of least-squares

multiple regression (LS-MR) and LR for gender clas-

siﬁcation.

Upon further analysis, several studies also uti-

lize distance-based classiﬁers (DBC) and various fu-

sion techniques (Tsimperidis et al., 2015; Fairhurst

and Da Costa-Abreu, 2011), which include Manhat-

tan distance, Euclidean distance, dynamic classiﬁer

Selection based on local accuracy (DCS-LA), ma-

jority voting, sum-based methods, template informa-

tion, and score-based fusion approaches (Tsimperidis

et al., 2015; Fairhurst and Da Costa-Abreu, 2011;

Giot and Rosenberger, 2012; Syed Idrus et al., 2014).

Despite the popularity of ML approaches, deep

learning (DL) approaches were also found to be im-

plemented in the ﬁeld. For example, (Buriro et al.,

2016; Kolakowska et al., 2016) both implement a

novel neural net (NN) and deep neural network ar-

chitecture (DNN) for behavioral biometric-based user

characteristics (age, gender, or operating handed-

ness) classiﬁcation. Similarly, (Tsimperidis et al.,

2017; Tsimperidis et al., 2018) implements MLP for

keystroke behavioral biometric-based age and gen-

der prediction. In addition to the conventional DL

approaches applied previously, radial basis function

networks (RBFN) (Tsimperidis et al., 2018; Tsim-

peridis et al., 2021) have also been implemented. Fur-

ther investigation shows that the research studies also

rely on meta-algorithms, such as AdaBoost, multi-

boot, random-correction-code, exhaustive-correction-

code, and rotation forest, to boost classiﬁer perfor-

mance (Tsimperidis et al., 2018; Tsimperidis et al.,

2021).

Table 2 summarizes the AI approaches and eval-

uation criteria (EC) implemented for analysis. Based

on the information presented in Table 2, we ﬁnd that

the ML approaches are more popular for behavioral

biometric-based age and gender classiﬁcation despite

the availability of advanced DL approaches (Table 2).

Further analysis shows that SVM is the most popular

ML approach for behavioral biometric-based age and

gender prediction.

4 DATA COLLECTION

PROCEDURE

Our background analysis conﬁrms that there are no

publicly available mouse behavior datasets for contin-

uous age and gender classiﬁcation in online education

platforms (Table 1). Hence, we collect novel mouse

behavior data for our speciﬁc application.

Before we pursue data collection, we need to un-

derstand what type of mouse behavior data needs to

be collected for accurate age and gender classiﬁca-

tion. This is realized through our comprehensive

background analysis, which reveals that prior re-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

388

Figure 1: Methodology used in our study for mouse behavioral biometric-based age and gender classiﬁcation.

search gathers mouse behavior data by instructing

participants to complete predeﬁned tasks in a con-

trolled environment or by observing their daily activi-

ties in an uncontrolled setting (Van Balen et al., 2017;

Pentel, 2017; Pentel, 2019).

Since our primary objective is to perform age and

gender classiﬁcation in online education platforms,

we ﬁnd that the former data collection method is more

suitable for our experimentation. Hence, we collect

novel mouse behavior data while participants engage

in an online assessment game, which consists of sev-

eral assessment-like tasks, including clicking the tar-

get, MCQ, drag-drop, and matching tasks.

The rationale for incorporating various tasks in the

data collection procedure is derived from our back-

ground analysis. Our investigation shows that sev-

eral mouse behavioral biometric-based authentication

studies collect several varieties of mouse event data

(Siddiqui et al., 2021; Subash et al., 2024; Zheng

et al., 2011), including mouse movements, clicks,

drag, and scroll events for analysis. Therefore, we in-

cluded four simple but different tasks in our data col-

lection procedure to collect the same variety of mouse

behavior data.

Data acquisition was done with the help of a

web application developed using HTML, CSS, and

JavaScript. Necessary data is recorded using the

mouse event listener method. In addition to this,

we also collect screen dimension information using

existing JavaScript methods. The description of the

tasks is as follows:

1. MCQ Task: Contains four simple general

knowledge questions that participants answer by

selecting the correct choice among four choices.

The subsequent question is displayed only when

the current question is answered correctly. Once

a choice is selected, the participant must click

the submit button. If the choice is incorrect, the

participants are shown a prompt indicating they

must answer the question again. In other words,

the user can rectify their answer until the correct

choice is selected.

2. Click the Target Task: This task involves clicking

a target (button element) that alters its position

each time it is clicked. In total, the target changes

position nine times.

3. Drag-Drop Task: Requires participants to drag an

image of an animal into the correct drop-box con-

taining the label of the animal’s category (mam-

mal, amphibian, reptile, ﬁsh, or bird). There are

ﬁve images and ﬁve drop-boxes. If the partici-

pant drags the image in-to the correct dropbox,

the background is changed to green, indicating a

proper response. Furthermore, the scaled version

of the image will be displayed within the box. If

the participant drags an image into the in-correct

dropbox, the images are returned to their original

positions, and the background is changed to red

for a short time frame. After which, the back-

ground color returns to its original state. Similarly

AgeGen Bio Track: Continuous Mouse Behavioral Biometrics-Based Age and Gender Proﬁling in Online Education Platforms

389

Table 3: Comparison of 3-, 5-, and 10-MM segments for age classiﬁcation.

Segmentation Method Approaches Acc (%) Pre (%) Rec(%)

3-MM

SVM 56.41 65.61 50.61

LR 57.57 56.99 53.26

NB 56.95 55.96 52.37

KNN (k= 6) 63.06 62.77 60.80

RF* 76.61 76.46 75.83

DT 67.76 67.41 66.44

5-MM

SVM 56.24 61.96 51.19

LR 53.40 52.25 52.15

NB 49.03 51.42 51.20

KNN (k= 6) 59.28 58.52 57.23

RF 59.89 61.63 61.19

DT 55.43 57.28 56.87

10-MM

SVM 56.96 61.94 54.12

LR 54.94 54.21 53.61

NB 54.94 54.31 53.01

KNN (k= 6) 59.19 59.89 57.46

RF 54.34 56.92 55.80

DT 55.95 56.29 56.28

to MCQ tasks, the participants are allowed to

correct their mistakes.

4. Matching Task: This is the ﬁnal task that the par-

ticipants perform. They must identify four pairs

of matching images (country ﬂags) among eight

images displayed on the screen. If the selected

images do not match, they are shown brieﬂy and

restored to their original state.

Data was collected from 20 participants recruited

from Sanjay Gandhi College of Education, Ben-

galuru, India. As part of the data collection process

for this study, all participants were required to pro-

vide informed consent prior to their involvement. The

process was carefully designed to ensure compliance

with ethical guidelines and maintain transparency re-

garding the nature of the research and the use of col-

lected data.

All participants were required to perform MCQ,

drag-drop, and matching tasks ten times each. In ad-

dition to mouse behavior data, we collected user de-

mographic (age group and gender) information via a

pre-participation questionnaire. Among the 20 par-

ticipants, ten are female and ten are male, distributed

across two age groups: 18-22 and 23-27 years of age.

During task engagement, raw data, such as times-

tamp, screen height, screen width of the content area,

screen coordinates (X, Y), event action types, element

on which the event was performed, offset X, and Y are

collected for further feature extraction. Mouse behav-

ior data is received individually for each user and task

type in JSON format.

5 EXPERIMENTAL RESULTS

This section will present the experimental results, us-

ing which we identify the best segmentation method,

ML model, and features set to be integrated into the

AgeGen Bio Track system for continuous age and

gender classiﬁcation.

As mentioned before, we will compare several dif-

ferent ML models, including DT, RF, NB, LR, SVM,

and kNN. Training and testing will be performed us-

ing the hold-out approach, where 90% will be used

for training, and 10% will be used for testing. Before

performing segmentation, model training, and evalu-

ation, we will extract features per mouse event and

remove any unwanted values and columns. We then

pre-process the data by replacing outliers using the

average values of the attributes and performing nec-

essary segmentation methods. The overall working

methodology is illustrated in Figure 1.

5.1 Best Segmentation Method and ML

Model for Age and Gender

Classiﬁcation

Before training the ML models, we apply segmenta-

tion as a preprocessing technique to logically group

mouse behavior data into meaningful blocks of infor-

mation. These segments are then used to compute

aggregate values of the attributes, such as minimum

(min), maximum (max), average (Avg), and standard

deviation (std) (Subash et al., 2024). For our exper-

imentation, we will perform segmentation according

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

390

Table 4: Comparison of 3-, 5-, and 10-MM segments for gender classiﬁcation.

Segmentation Method Approaches Acc (%) Pre (%) Rec(%)

3-MM

SVM 58.79 59.45 58.45

LR 58.54 58.53 58.54

NB 55.12 58.49 54.41

KNN (k= 6) 62.14 62.57 61.91

RF* 75.76 75.75 75.76

DT 69.41 69.51 69.47

5-MM

SVM 52.58 54.05 53.14

LR 53.60 56.66 54.30

NB 55.12 59.15 54.27

KNN (k= 6) 57.36 57.42 57.13

RF 50.55 53.72 51.46

DT 54.82 54.78 54.78

10-MM

SVM 59.79 60.36 60.27

LR 58.78 58.53 57.69

NB 52.72 55.21 54.26

KNN (k= 6) 59.39 61.10 60.38

RF 52.52 49.82 49.91

DT 55.15 54.85 54.84

to the procedure mentioned in (Subash et al., 2024),

which uses the n-mouse movement (MM) segmen-

tation approach. This approach groups a predeﬁned

number (n) of mouse movement events into a single

segment for analysis (Siddiqui et al., 2021; Subash

et al., 2024).

However, based on our background analysis, the

n-value varies according to the study performed (Sid-

diqui et al., 2021; Subash et al., 2024). In other words,

there is no standard n value for analysis. Therefore, to

understand the effect and identify the best MM seg-

mentation approach, we test n in three scenarios: 3, 5,

and 10. For evaluation, we extracted 40 features iden-

tiﬁed in our previous work (Subash et al., 2024) using

data only from the ﬁrst attempt. After feature extrac-

tion, we perform a train-test split (90/10), min-max

normalization, and label encoding.

Based on the results (Tables 3 and 4), we conclude

that 3-MM segmentation is the best-performing seg-

mentation approach for our speciﬁc application. Fur-

thermore, RF is the best performing ML algorithm

evaluated under the 3-MM segmentation approach,

achieving more than 75% accuracy, precision, and re-

call. It is also observed that SVM, DT, RF, LR, NB,

and KNN achieve similar performance when imple-

menting the 5- and 10-MM segmentation approach.

A signiﬁcant increase in performance is detected in

RF and DT when we compare 3-, 5- and 10-MM seg-

mentation approaches.

5.2 Identifying Best Feature Set for Age

and Gender Classiﬁcation

In the previous section, we identiﬁed the best-

performing segmentation approach and ML model.

Our experimentation conﬁrms that the 3-MM seg-

mentation approach evaluated on the RF model out-

performs other approaches considered in this study.

In this section, we will improve the performance

of the RF model by implementing feature selection

approaches. Speciﬁcally, we will compare two ap-

proaches, univariate and recursive elimination feature

selection (FS) algorithms, to determine the best fea-

ture set for age and gender classiﬁcation. For our

study, we perform FS after Min-Max normalization

as the attributes have different SI units depending on

the feature extracted. Furthermore, we selected the

top 20 features for our analysis.

Table 5: Comparison of different FS approaches for age

classiﬁcation.

ML FS Approach Acc

(%)

Pre

(%)

Rec

(%)

RF Original

Feature Set

76.61 76.46 75.83

RF Univariate FS 72.71 72.62 71.55

RF Recursive

Elimination

80.76 80.83 79.98

From this experiment, it can be conﬁrmed that

the RFE-selected feature set (Table 7) is suitable for

mouse behavioral biometric-based age and gender

AgeGen Bio Track: Continuous Mouse Behavioral Biometrics-Based Age and Gender Proﬁling in Online Education Platforms

391

classiﬁcation in online education platforms. Based on

the performance achieved (Tables 5 and 6), it is ev-

ident that the recursive feature elimination approach

gives the best feature set for age and gender classiﬁ-

cation. It is noticed that there is a 4-5% increase in

performance in all evaluation criteria when we com-

pare the performance between the RFE-selected fea-

tures and the original feature set.

Table 6: Comparison of different FS approaches for gender

classiﬁcation.

ML FS Approach Acc

(%)

Pre

(%)

Rec

(%)

RF Original

Feature Set

75.76 75.75 75.76

RF Univariate FS 72.95 72.94 72.94

RF Recursive

Elimination

79.91 79.91 79.92

Table 7: User characteristic feature set selected for age and

gender classiﬁcation.

Aggregate Values Attributes

Average (Avg) Co-ordinates (x, y),

timestamp, angle to path

tangent, jerk

Minimum (Min) Co-ordinates (x, y),

vertical velocity,

horizontal velocity,

acceleration, timestamp

Maximum (Max) Co-ordinates (x, y),

vertical velocity,

acceleration, jerk,

distance, angular

velocity, timestamp

Stand-alone value Elapsed Time

6 CONCLUSION

It is essential to identify user characteristics, such

as age and gender, to improve online assessment

fraud detection systems that safeguard online edu-

cation platforms. This paper proposes AgeGen Bio

Track: a continuous mouse behavioral biometric-

based age and gender classiﬁcation model for this pur-

pose. To accomplish this, we collected novel task-

speciﬁc mouse behavior data while participants en-

gaged in an online assessment game, using which we

identiﬁed the best segmentation approach, machine

learning model, and feature set for continuous age

and gender classiﬁcation. Our investigation indicates

that our segmentation approach with the RF algorithm

and user characteristic feature set attains satisfactory

performance of 80% in all evaluation criteria. The

overall performance achieved by our proposed ap-

proach indicates positive results in mouse behavioral

biometric-based age and gender classiﬁcation for our

speciﬁc application. Our ﬁndings and comprehensive

background analysis also support further research in

the ﬁeld and suggest that user age and gender param-

eters can be fused for behavioral biometric-based au-

thentication to enhance performance.

REFERENCES

Barral, J. and Deb

u, B. (2004). Aiming in adults: Sex and

laterality effects. Laterality (Hove), 9(3):299–312.

Blau, I. and Eshet-Alkalai, Y. (2017). The ethical dis-

sonance in digital and non-digital learning environ-

ments: Does technology promotes cheating among

middle school students? Computers in human behav-

ior, 73:629–637.

Buriro, A., Akhtar, Z., Crispo, B., and Del Frari, F. (2016).

Age, gender and operating-hand estimation on smart

mobile devices. In 2016 International Conference

of the Biometrics Special Interest Group (BIOSIG),

pages 1–5. Gesellschaft fuer Informatik.

Fairhurst, M. and Da Costa-Abreu, M. (2011). Using

keystroke dynamics for gender identiﬁcation in social

network environment. In 4th International Conference

on Imaging for Crime Detection and Prevention 2011

(ICDP 2011), pages P27–, Stevenage. IET.

Fryar, C. D., Gu, Q., and Ogden, C. L. (2012). Anthropo-

metric reference data for children and adults: United

states, 2007-2010. Vital and health statistics. Series

11. Data from the National Health Survey, (252):1–

48.

Garg, M. and Goel, A. (2022). A systematic literature

review on online assessment security: Current chal-

lenges and integrity strategies. Computers & Security,

113:102544.

Giot, R. and Rosenberger, C. (2012). A new soft biomet-

ric approach for keystroke dynamics based on gen-

der recognition. International Journal of Information

Technology and Management, 11(1-2):35–49. PMID:

44062.

Idrus, S. Z. S., Cherrier, E., Rosenberger, C., and Bours, P.

(2013). Soft biometrics database: A benchmark for

keystroke dynamics biometric systems. In 2013 In-

ternational Conference of the BIOSIG Special Inter-

est Group (BIOSIG), pages 1–8. Gesellschaft f

ur In-

formatik e.V. (GI).

Jim

enez-Jim

enez, F. J., Calleja, M., Alonso-Navarro, H.,

Rubio, L., Navacerrada, F., Pilo-de-la Fuente, B.,

Plaza-Nieto, J. F., Arroyo-Solera, M., Garc

ıa-Ruiz,

P. J., Garc

ıa-Mart

ın, E., and Ag

undez, J. A. (2011).

Inﬂuence of age and gender in motor performance in

healthy subjects. Journal of the neurological sciences,

302(1):72–80.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

392

Kolakowska, A., Landowska, A., Jarmolkowicz, P., Jar-

molkowicz, M., and Sobota, K. (2016). Automatic

recognition of males and females among web browser

users based on behavioural patterns of peripherals us-

age. Internet research, 26(5):1093–1111.

Noorbehbahani, F., Mohammadi, A., and Aminazadeh, M.

(2022). A systematic review of research on cheating

in online exams from 2010 to 2021. Education and

information technologies, 27(6):8413–8460.

Pentel, A. (2017). Predicting age and gender by keystroke

dynamics and mouse patterns. In Adjunct Publication

of the 25th Conference on User Modeling, Adaptation

and Personalization, pages 381–385, New York, NY,

USA. ACM.

Pentel, A. (2019). Predicting user age by keystroke dy-

namics. In Silhavy, R., editor, Artiﬁcial Intelligence

and Algorithms in Intelligent Systems, pages 336–343,

Cham. Springer International Publishing.

Rohr, L. E. (2006). Gender-speciﬁc movement strategies

using a computer-pointing task. Journal of motor be-

havior, 38(6):431–137.

Siddiqui, N., Dave, R., and Seliya, N. (2021). Continu-

ous authentication using mouse movements, machine

learning, and minecraft. arXiv.org.

Subash, A., Song, I., Lee, I., and Lee, K. (2024). Mouse

dynamics-based online fraud detection system for on-

line education platforms. In Yang, X.-S., Sher-

ratt, S., Dey, N., and Joshi, A., editors, Proceedings

of Ninth International Congress on Information and

Communication Technology, pages 257–269, Singa-

pore. Springer Nature Singapore.

Susnjak, T. and McIntosh, T. (2024). Chatgpt: The

end of online exam integrity? Education sciences,

14(6):656–.

Syed Idrus, S. Z., Cherrier, E., Rosenberger, C., and Bours,

P. (2014). Soft biometrics for keystroke dynamics:

Proﬁling individuals while typing passwords. Com-

puters & security, 45:147–155.

Tsimperidis, I., Arampatzis, A., and Karakos, A. (2018).

Keystroke dynamics features for gender recognition.

Digital investigation, 24:4–10.

Tsimperidis, I., Katos, V., and Clarke, N. (2015).

Language-independent gender identiﬁcation through

keystroke analysis. Information and computer secu-

rity, 23(3):286–301.

Tsimperidis, I., Rostami, S., and Katos, V. (2017). Age

detection through keystroke dynamics from user au-

thentication failures. International journal of digital

crime and forensics, 9(1):1–16.

Tsimperidis, I., Yoo, P. D., Taha, K., Mylonas, A., and

Katos, V. (2020). R2bn: An adaptive model for

keystroke-dynamics-based educational level classiﬁ-

cation. IEEE transactions on cybernetics, 50(2):525–

535.

Tsimperidis, I., Yucel, C., and Katos, V. (2021). Age

and gender as cyber attribution features in keystroke

dynamic-based user classiﬁcation processes. Elec-

tronics (Basel), 10(7):835–.

Van Balen, N., Ball, C., and Wang, H. (2017). Analy-

sis of targeted mouse movements for gender classi-

ﬁcation. EAI endorsed transactions on security and

safety, 4(11):153395–15.

Wei, X., Saab, N., and Admiraal, W. (2021). Assessment of

cognitive, behavioral, and affective learning outcomes

in massive open online courses: A systematic litera-

ture review. Computers and education, 163:104097–.

Zheng, N., Paloski, A., and Wang, H. (2011). An efﬁ-

cient user veriﬁcation system via mouse movements.

In Proceedings of the 18th ACM conference on Com-

puter and communications security, pages 139–150,

New York, NY, USA. ACM.

AgeGen Bio Track: Continuous Mouse Behavioral Biometrics-Based Age and Gender Proﬁling in Online Education Platforms

393