Identity Verification and Fraud Detection During Online Exams with a
Privacy Compliant Biometric System
M. A. Haytom
1,2
, C. Rosenberger
1
, C. Charrier
1
, C. Zhu
2
and C. Regnier
2
1
Normandie Univ., UNICAEN, ENSICAEN, CNRS, GREYC, 14000 Caen, France
2
TestWe, 75003 Paris, France
Keywords:
Personal Data, Biometric Authentication, Privacy Protection, Machine Learning.
Abstract:
Distant learning is an alternative solution to education when the learner is far from the school or cannot attend
courses for professional or medical reasons. The main objective of this work is to design a smart application
of remote exams, using a multibiometric system combining face with deep learning and keystroke dynamics
to verify the identity of the learner. Privacy protection is consider in this work as an important issue because
many personal data are processed in the proposed solution. We consider in this paper experiments under real-
life conditions to identify abnormal behaviours with confidence indicators. We show the system ability to
make the correct decision while preserving learner’s privacy.
1 INTRODUCTION
Today, with the rapid growth of online courses and ex-
ams, more and more students or professionals are ap-
plying for distance learning. Certainly, distance learn-
ing offers several advantages: the opportunity to take
a training course from a distant establishment, arrange
teaching for the personal and professional life of the
learner, have great organizational flexibility and re-
duce accommodation and transportation costs. It al-
lows to have the same training level obtained in real
classrooms. Yet, secure tools must be used to safely
access the exams. It is, thus, essential to control the
student access and environment during an exam. On-
line assessment and supervised tests address the issue
of student verification and the environment in which
access to unauthorized documents and resources is a
major problem.
Data processing and environmental control must
respect the principles defined by the General Regula-
tion on the protection of personal information (Carey,
2018). The information often collected in a regula-
tory context can make the verification of a person a
safe and secure step. Higher education encourages
to fight cheating more effectively, schools are so very
interested in the quality and integrity of online proc-
toring. There is strong evidence that cheating has in-
creased in classrooms as it has been shown that 70%
of students cheat during their schooling (Chauvel, ).
Various possible attacks that threaten the privacy of
individuals can also be observed. The student may be
faced with an identity theft, whether hacking his/her
account, or attempting to perform fraudulent actions.
It is therefore mandatory to count and identify them
to define the necessary requirements for the manage-
ment of exams and the protection of personal data.
The need for reliable user authentication solu-
tions has increased with growing security concerns
and rapid advances in networking and communica-
tion. Biometrics refers precisely to a computer tech-
nology that stimulates the physical presence of an in-
dividual, it conceives its identity and its movements
as sources of dangers and risks. It is described as the
science of recognizing an individual based on their
physical or behavioral characteristics. This technol-
ogy is an excellent candidate for verifying the identity
of an individual. Biometrics are divided into two cate-
gories: 1) physical and 2) behavioural. Physical anal-
ysis may include the geometry of the hand, the retina,
the characteristics of the iris and so on. Behavioural
analysis helps to verify a person’s identity by examin-
ing measurable activity, such as gesture recognition,
keystroke dynamics and so on.
There are two main contributions in this work: 1)
proposal of an original authentication solution based
on facial recognition and keystroke dynamics for on-
line assessments and the integration of a security sys-
tem to detect fraud during online exams and 2) per-
Haytom, M., Rosenberger, C., Charrier, C., Zhu, C. and Regnier, C.
Identity Verification and Fraud Detection During Online Exams with a Privacy Compliant Biometric System.
DOI: 10.5220/0009874104510458
In Proceedings of the 17th International Joint Conference on e-Business and Telecommunications (ICETE 2020) - SECRYPT, pages 451-458
ISBN: 978-989-758-446-6
Copyright
c
 2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
451

forming experiments in real conditions to validate the
proposed approach (Fig. 1). The paper is organized as
follows. In section 2, we present the related works on
the identity verification and fraud detection solutions
during online examinations. We present the proposed
system in section 3. Section 4 is dedicated to exper-
imental results. We finally conclude in section 5 and
give some perspectives of this work.
2 RELATED WORKS
In the literature, we can only find few studies that
deal with remote exam management. Duchatelle et
al. conducted a real-life remote monitoring experi-
ment at the University of Caen Normandie (P. Beust,
2016) allowing students in different countries to take
this exam. According to the obtained results, several
students have accepted to participate in the experi-
ment. Half of the candidates go as far as creating an
account on the platform of the organism that manages
the exams. As part of this experiment, 31 participants
completed the test and almost a third of the candi-
dates could not pass the test due to technical prob-
lems. This experience remains among the first online
distance exams using biometric-based authentication.
Recently, authors in (Arnautovski, 2019) proposed a
face recognition system for online examinations. Yet,
the proposed system provides no privacy protection
and has not been tested in real conditions.
2.1 Keystroke Dynamics
Many applications nowadays require personal data
about the user, where the proof of identity is used to
provide a secure service and to guarantee that a per-
son is authorized to access an online multimedia ser-
vice. In addition, it is necessary to convince users that
the proposed solution is credible and that the informa-
tion is processed reliably. Without a doubt, keystroke
dynamics could play a necessary role to improve the
identity verification of individuals. This biometric so-
lution has a very high degree of acceptability when
compared to other modalities such as iris or gesture
recognition.
Keystroke dynamics represents the analysis of
keyboard interactions (for example the flight time be-
tween two pressed keys), recorded during the use of a
laptop by the user. The collected information describe
the time when the key is pressed and the time when
the key is released when the learner uses his keyboard.
Researchers have carried out numerous studies in this
field, it is possible to explore these information for
the authentication of an individual or profiling pur-
pose in order to recognize the gender or the age cate-
gory according to the password or the edited text with
a recognition rate close to 75% (Giot et al., 2012;
Idrus et al., 2013). In fact, it is possible to combine
keystroke data with other biometric modality in order
to have a more secure verification system. In the first
place, we have integrated keystroke dynamics into our
remote exam management system as a basic modality
to increase the level of security and to have a robust
identity verification process.
2.2 Facial Recognition
Facial recognition systems have become an useful
tool in many fields of application. It remains an ac-
ceptable solution for many people because it is less
expensive and easier to use. Facial recognition sys-
tems are well-known computer vision applications
whose purpose is to verify or identify a person. There
are many approaches for facial recognition in the lit-
erature, such as: morphological elastic graph match-
ing (Kotropoulos et al., 2000), scale-invariant feature
transform (Bicego et al., 2006) and so on.
However, testing in a non-controlled environment
(variable lighting, variable orientation of the person
towards the camera, moving person, etc.) remains
problematic, since the verification of users is carried
out with high uncertainty especially during the acqui-
sition phase. Certainly, many other methods to ver-
ify the identity of a person such as Deep Learning
based approach exist (Liu et al., 2015). To address the
critical environmental issue, the research activity fo-
cused on the analysis of human perception in order to
mathematically describe the processes that lead to the
recognition of actions, in order to mimic to the recog-
nition process of faces and objects by human beings.
Today, there are several works and scientific publica-
tions based on artificial intelligence for user verifica-
tion and identification (Liu et al., 2015; Parkhi et al.,
2015).
Face recognition is a constantly evolving field,
continually improving. Identity verification through
deep learning has recently received a lot of atten-
tion from the research and industry community and
is starting to be applied in a variety of areas, mainly
for security. We can find a study that handles the
authentication based on a Convolution Neural Net-
work (Parkhi et al., 2015), researchers compute VGG-
Face CNN descriptors using CNN implementation
based on the VGG-Very-Deep-16 CNN architecture
and they evaluate it on a large databases (Labeled
Faces in the Wild and the YouTube Faces). Deep
learning remains an effective technique and works
much better than many facial recognition methods.
SECRYPT 2020 - 17th International Conference on Security and Cryptography
452

Figure 1: Multimodal biometric system with an automated monitoring.
The combination of keystroke dynamics and fa-
cial recognition has also been studied. Unlike (Gupta
et al., 2015) and (Giot et al., 2010), our approach is
based on a deep learning solution for faces and then
the use of a different protocol of keystroke dynamics
for free text field. In addition, the use of biometric and
non-biometric data to form a model allowing identity
verification and fraud detection in a remote examina-
tion.
3 PROPOSED APPROACH
In this paper, we propose the integration of both facial
recognition solution and keystroke into the remote ex-
amination management system as a multimodal solu-
tion to increase the security level of the application.
In fact, for the faces, we improved the pre-processing
phase to have a better presentation for data analysis
before using deep learning process to extract the char-
acteristics of a learner. In fact, our objective is to
set up a combination of several existing techniques
in a remote examination management system to have
a reliable and efficient solution. The biometric signa-
tures are part of a lot of so-called sensitive data. So,
we implement security measures and data protection
processes by obtaining an encrypted template in the
form of a secure code. The biohashing algorithm has
been used to ensure the confidentiality of the data ex-
changed since the storage of biometric templates is a
major problem.
Moreover, we propose the use of a fraud detection
model to avoid frauds and identity theft problems
during remote exams. In this work, we conducted
an experiment in real conditions in order to validate
the proposed solution. Several fraud scenarii have
been considered; direct communication with a person
nearby, the use of unauthorized documents, the use of
a second screen smartphone and/or tablet to connect
to the internet or sound transmission as sound effects
with fingers or hands or even leaving the room to
have answers, etc. The steps of the proposal solution
are:
Authentication
1. Keystroke dynamics process
2. Face recognition process
3. Compute for each feature vector its Biocode
Fraud detection
1. Perform a continuous verification
2. Create fraud detection model
3.1 Identity Verification
3.1.1 Keystroke Dynamics Process
For the first biometric modality, we carry out a con-
tinuous verification throughout the exam session for
each learner. First of all, when the user writes a text
using his keyboard, each word represents a signature.
Words with more than two letters are processed and
analyzed. From several time sequences, we extract
the learner’s template where: 1) the time interval be-
tween pressed and released key, 2) the time interval
between two consecutive pressures, 3) the time inter-
val between a release and the pressing of the next key
and 4) the overall duration of a series of characters.
Unlike several state of the art studies, the learner’s
hesitation time was taken into account with the erro-
Identity Verification and Fraud Detection During Online Exams with a Privacy Compliant Biometric System
453

neous data. The hesitation times to write a text re-
veal a unique data related to the person and his real
way of using his keyboard, which is not the case for
passwords. For each keystroke, the time intervals are
recorded and concatenated with the duration to con-
stitute a vector of characteristics.
Keystroke dynamics can be affected by a number
of variables (how well the learner prepared for differ-
ent exams, and how familiar he/she is with the topic
of each question within the same exam, etc), to avoid
this, during the enrollment the user is invited to write
a text of a page, the hesitation time is taken into ac-
count with a delay of 3 seconds maximum. Depend-
ing on the security requirements of the application,
the KD is used in this context only for texts at least
one page long.
3.1.2 Facial Recognition Process
Face recognition systems involve two important
phases, the enrollment and the authentication. The
first step concerns the user enrollment: enrollment
means the acquisition of the biometric input data, the
detection of region of interest and the extraction of
features to define a template of each genuine user. The
genuine model is stored as a reference after applying
some quality requirements (Saad et al., 2012). For
face recognition based systems, the concerned bio-
metric data is the face image, and the extracted fea-
tures are stored in the user device memory.
To achieve the enrollment phase in a reliable man-
ner on a massive scale, a familiarization session was
planned to get started with the software one day be-
fore the exam. During this simulation the learner is
invited to capture his image, this image will be used
in order to make a verification during the real session
of the exam.
On the first hand, to authenticate users, we start
by capturing images using the student laptop camera.
A pre-processing step has been used to enhance the
input data. Once the face image has been acquired,
we use a face detection tool to crop and normalize
the faces. Next, we extract features using the Convo-
lutional Neural Network (CNN) model (Parkhi et al.,
2015) to obtain a template representing the signature
of the user. With the biometric model, students can
access online courses and exams depending on the
threshold used for the verification process. First,
we compare the reference model with the signature
acquired during authentication and we compute a
score distance.
Quality Evaluation
The no-reference image quality assessment algorithm
BLIINDS2 (Saad et al., 2012) has been used to
evaluate the quality of the image, using a natural
scene statistical model of discrete cosine trans-
form coefficients. The approach relies on a simple
Bayesian inference model to predict image quality
scores based on certain extracted features. During
the pre-processing step, if the image quality metric
is lower than the threshold, the acquired image is not
saved, otherwise the image is recorded.
Adjusting Intensity
The values of the intensity of the image have been
adjusted to improve the contrast of the output image
by saturating the higher side and the lower side by 1
percent of all pixel values.
Face Detection
In this part, Viola-Jones Cascade Object Detector has
been used to detect the faces (Viola and Jones, 2001).
Indeed, the calculation is based on the integration of
the pixels of the image as well as the use of pseudo-
Haar features. After getting the size and the position
or the spatial coordinates of the region of interest.
The output is cropped including all the pixels of the
image inside the rectangle. The image is resized by
specifying the number of rows and columns to get
a matrix of two dimensions (resolution 224 × 224
pixels).
VGG Face - Convolution Neural Network
CNN architectures were mainly used for character
recognition tasks. They are proven very effective
in areas such as image recognition and classification
(Parkhi et al., 2015). The model was formed on 2.6
million images over more than 2,600 people. The ac-
curacy of the system has been evaluated at 98.95%,
compared to DeepFace and FaceNet system. To ex-
tract the feature vector, VGG pre-training model has
been applied to the input image (resolution 224 ×224
pixels), the output data represents the user pattern or
the unique signature of the learner.
3.1.3 Privacy Protection
The general principle of BioHashing is to generate
a binary code called Bio-Code, which will be used
during the enrollment phase and to verify the identity
of the learner. It can be obtained from the biometric
data (the template of the user) and a random num-
ber. In the literature, we can find many studies using
Biohashing for many biometrics modalities such as
fingerprints (Belguechi et al., 2010; Belguechi et al.,
2012), keystroke dynamics (Migdal and Rosenberger,
2018) and so on.
In this work, we use the BioHashing algorithm in
order to protect the VGG face vector and keystroke
SECRYPT 2020 - 17th International Conference on Security and Cryptography
454

signature computed from an user for privacy reasons.
Face images and keystroke data are not stored for the
same reasons. This algorithm consists firstly in pro-
jecting the native biometric data on an orthogonal ba-
sis generated from the random number. The resulting
dimension corresponds at most to the dimension of
the primary representation of the biometric data. The
goal of this step is to hide the biometric data in a part
of the multidimensional space. The use of an orthogo-
nal base ensures the conservation of the similarity re-
lations between the templates (BioCodes), while kep-
ping good recognition result. The second step aims to
quantify this result using a determined threshold. In
our study, for face the size of the generated biocode
is 512 bits. The choice of the threshold and the size
of the bioCode guarantees the non-inversion of the
process. In addition, this two-factor process (Finger-
Code and Seed-key) ensures that it is not possible to
retrieve the native biometric data given the BioCode.
3.2 Fraud Detection
In our study, we use face detection, analysis of events
using images and facial recognition by performing a
calculation on a four seconds duration window. In
fact several parameters were used: the keyboard ac-
tivity, the keystroke dynamics and the acoustic signal.
On the other hand, we perform an analysis for each
type of information, then we use statistical solutions
such as the histogram in order to distinguish honest
learners and learners who cheat during the exam. For
example, for face detection, the histogram allows to
represent the distribution of the numbers of detection
continuously until the end of the test. The first step
consists in comparing the data acquired from a person
authorized to take the examination with and without
fraud attempts to the information collected from a per-
son not authorized to take the examination. When the
amplitude of the signal is higher than the threshold at
a given moment, we consider that the user cheated. To
set the threshold, some machine learning tools have
been used in order to create a fraud detection model
by automatically setting the threshold value.
The threshold could have a significant impact on
performance. During this study we note that, the ob-
servations are well separated between a person who
cheats and an honest learner. In fact, the use of sev-
eral characteristics (data from different sensors) dras-
tically reduce the impact of the threshold on perfor-
mance.
We trust the output provided by the camera, even
if in a way, a user can rely on techniques similar to
those provided by malware to produce a fake video
stream. In fact, the basic exam management solution
blocks access to internet during the examination with
a secure virtual environment, and the learner can’t
manipulate other applications before sending the copy
of the exam and closing the application.
3.2.1 Fraud Model
Here, we propose to use a machine learning method to
detect the fraud automatically and creating a fraud de-
tection model to manage online proctoring. The data-
set for training contains normal cases and frauds. The
model was created using four supervised classifica-
tion algorithms: 1) the Decision Tree, 2) the Logis-
tic regression, 2) the k-nearest neighbors scheme and
4) the Naive Bayes classifier after pre-processing and
re-sampling to obtain the user indicators. Indicators
were obtained after launching a continuous analysis
for each user.
In fact, a set of biometrics and non biometrics data (25
learners) have been used to train the model, each sam-
ple contains; 1) face detection analysis, 2) the key-
board event (by counting the number of keys), 3) the
distributions of image analysis, 4) three distributions
based on the amplitude of the sound signal by vary-
ing the amplitude thresholds 5) the keystroke dynam-
ics and 6) the feature vectors of the deep parameter
of facial recognition every four seconds, each signa-
ture of face was obtained after running the process
of deep learning. To form our detection model, data
from various scenarios were used. We start the train-
ing by using a class with 703 characteristics, for each
user 150 instances, making a total of 3750 instances.
The model was trained to count cheating attempts and
identify frauds automatically.
4 EXPERIMENTAL RESULTS
4.1 Experimental Protocols
4.1.1 Experiment 1
In this work, biometric data (face images) from 30
users working in the GREYC Laboratory and TestWe
Company were collected using a laptop computer.
This number of participants is not so high, but is in
the same range as published studies (see section 2).
Biometric information is collected from individuals
via an integrated camera of a ”ThinkPad S440” com-
puter. Intel Core i7-4510U computer with a CPU of
2.00 and 2.60 [GHz] was used during the examina-
tion.
This first experiment remains a preliminary study
to have the feelings of the users, evaluate the pro-
Identity Verification and Fraud Detection During Online Exams with a Privacy Compliant Biometric System
455

posed method and the system ability to make the right
decision. Each user is invited to take the exam re-
motely with a graphical user interface (QT applica-
tion). More precisely, the sample is recorded and
it is the same for each user. Experiments were car-
ried out with two different protocols: the first proto-
col consists of passing the test in an uncontrolled en-
vironment for ten participants and the second proto-
col twenty people have passed the test with the same
device in a controlled environment. The constraints
of the uncontrolled environment are: the variation of
light in the examination room, the face pose and the
presence of posters and images in the background of
the learners. For this preliminary study, the database
contains a number of 30 people and represents a class
of examination, for each participant 720 images are
collected making a total of 21600 samples.
4.1.2 Experiment 2
We have created an application with a text editor to
collect keystroke data. The application allows us to
collect the flight time between each pressed key as
well as the release time of the keys. Two sessions
were carried out (the same number of participants of
the previous experience); during the first session, par-
ticipants were asked to write one page of text and then
write ten sentences before sending the data. The sec-
ond session consists of launching a second application
(keylogger) for a period of one day in order to col-
lect as much data as possible. The number of words
recorded for each learner varies with an average of
1200. We take the minimum value (150 samples) in
order to process the data of all participants. A stand-
by button is used when a participant enters personal
information (password, PIN, etc.).
4.1.3 Experiment 3
A third experiment in real conditions was carried out
with ENSICAEN students to simulate fraud attempts
in a remote examination. Several fraud scenarios have
been considered using unauthorized supports (smart-
phone, blackboard, second computer, etc.) or the help
of a friend nearby (direct or indirect communication)
or even leaving the room to seek answers. The over-
all duration is 10 minutes for each simulation. User
interface application (QT C++) has been used to col-
lect the data, the information collected are the audio
signal, the sequence of images then the flight time be-
tween each used key.
Table 1: Evaluation in an uncontrolled and in a controlled
environment (Equal Error Rate).
Template uncontrolled env controlled env
Ref (aver) FaceC BioC FaceC BioC
10 0.531 0.292 0,0099 0.001
20 0.529 0.26 0,0088 0.002
30 0.444 0.24 0,0069 0.001
40 0.35 0.254 0,0057 0.001
50 0.316 0.282 0.0065 0.001
4.2 Performance Evaluation
We consider that the conditions related to the sensors
are fulfilled. Some common error rates in biomet-
rics will be used such as Equal Error Rate (EER) with
computes the system configuration to have a tradeoff
between the FAR (false acceptance rate) and the FRR
(false rejection rate) values.
4.2.1 Face Recognition
We consider the database n1 composed of 720 im-
ages (resolution 224 × 224 pixels) for 10 people. The
faceCode of each user is generated according to the
method presented in section 3.1, with deep learning.
Once this is done, the biohashing has been used, 720
faceCodes are available for each person, which means
7200 faceCodes with a length of 512 bits. After ran-
dom projection and quantization, 7200 BioCodes are
generated. For each person, the average of several
FaceCodes is kept as a reference, the other FaceCodes
are used to evaluate the performance of the system.
For the first scenario, the reference FaceCodes
was chosen after calculating the average of the first
ten FaceCodes. Consecutively, the number of sam-
ples increased from 10 to 50 with a step of 10. On the
other hand, for the bioCodes, the reference templates
were chosen after calculating the average in a consec-
utive way. The average of several BioCodes is used as
a reference, the other BioCodes are also used for the
comparison. The BLIINDS2 quality assessment algo-
rithm (Saad et al., 2012) was not used to increase the
false acceptance rate during the pre-processing phase.
The EERs obtained (Table 1) are apparently far from
being the best (with errors related to non detection)
compared to the results of the literature. However,
these values allow us to define a basic system perfor-
mance in an uncontrolled environment.
In the second scenario, the process of evaluating
the image quality was used, which reduces the num-
ber of false acceptances. In the table (Table 1), we see
that in a controlled environment, the error rates de-
crease. The EER value of Facecode is greater than the
EER value of the BioCode in both scenarios. Indeed,
SECRYPT 2020 - 17th International Conference on Security and Cryptography
456

Table 2: Evaluation of two sessions (EER).
ref template BioHash Native
(average 1-10) min max min max
Keylogger 0.019 0.16 0.002 0.004
Normal sess 0.045 0.079 0.006 0.022
Table 3: The results of the fusion of faceCode scores and
secure keystroke signature (EER).
ref BioCode KD merged
template EER EER score
10 0.0012 0,0813 0,0813
20 0.0018 0,0738 0,07378
30 0.0009 0,0722 0,07225
40 0.0009 0,0859 0,08592
50 0.00095 0,0880 0,08803
without the constraints related to the environment and
the quality of the input data, after using bioHashing
the error rates are close to zero. However, it must at
least the average of the first forty FaceCodes to find
an optimal reference template.
4.2.2 Keystroke Dynamics
The same principle to secure the face data was applied
to protect the keystroke dynamics templates. Based
on the results of our experiment, we see that the EER
values vary between 0.2% and 2% for native data, and
between 4 % and 22 % with bioHashing (table 2, key-
logger and normal session). This is due to the fact
that the size of the characteristic vector is very small.
On the other hand, to secure more and more the bio-
metric data, we increase the size of the keycode (the
keystroke signature) before applying bioHashing by
adding relevant information related to the learner. We
note that the test based on the keylogger gives lower
error rates than the normal session since this test al-
lows participants to use their keyboard in a natural
way.
4.2.3 Multimodal Fusion
The fusion during the generation of scores induces
several problems such as non-homogeneous scores,
various distributions and vectors of different sizes. In
our study, we obtain an error rate of 8% with the av-
erage of the first 10 templates (Table 3, the results
of the merger of scores), a considerably reduced er-
ror percentage compared to fusion in the extraction of
characteristics. Indeed, for the reference template, the
average of 10 to 50 templates was used for the signa-
ture of the face and the keystroke dynamics.
4.2.4 Fraud Detection Model
The idea here is to highlight the classification tools to
create a model that will be used to detect fraud dur-
ing the online exam. To detect fraud, multiple infor-
mation of 25 participants from different sensors were
used to generate the model. The target is: ”fraud at-
tempt” or ”honest learner”. We chose all the biometric
and non-biometric data available from these authors
that were sufficiently numerous (with a 10 minutes
duration exam), which were sufficient to carry out an
analysis and determine the fraud attempts.
Table 4: Evaluation results (Cross validation, k = 20).
Model AUC CA Precis Recall
kNN 0.999 0.995 0.995 0.995
Tree 1 1 1 1
NB 0.987 0.962 0.962 0.962
LR 1 1 1 1
In order to establish the automated detection sys-
tem, a set of data was used for each user (set of char-
acteristics: face detection, facial recognition, video
activity, keystroke dynamics, sound signal and key-
board events). We use different supervised classifi-
cation approaches to build the model (Decision Tree
(DT), logistic Regression (LR), k-nearest neighbors
(kNN) and Naive Bayes (NB)).
A confusion matrix has been used to evaluate the
proposed fraud detection method. After setting the
output variable (”honest learner” and ”fraud attempt”)
and comparing the detection accuracy of different ma-
chine learning algorithms by adjusting the parameter
of the reliability estimation method (cross validation).
we obtain a precision close to 100% with the Logistic
Regression and the Decision Tree using all the charac-
teristics, (Table 4). These two methods allow to iden-
tify almost all fraud attempts (detecting 2100 with the
DT and 2099 fraud attempts with the LR, see Tables
5). After carrying out the experiments, the DT and
RL learning methods provide the best performance.
The DT and the RL remain a relevant solution for the
classification and the detection of fraud in a remote
exam. To conclude, we can integrate these two learn-
ing solutions into our biometric system to detect an
anomaly or the unusual behaviour using a fraud de-
tection model in order to improve online assessment.
Table 5: Confusion matrix - Tree and LR.
Tree (predicted) LR (predicted)
honest fraud honest fraud
honest (act) 1650 0 1650 0
fraud (act) 0 2100 1 2099
Identity Verification and Fraud Detection During Online Exams with a Privacy Compliant Biometric System
457

5 CONCLUSION AND
PERSPECTIVES
This work addresses the general problem of the pro-
tection of the privacy of a user during a remote exami-
nation. It seeks to define the different forms it can take
fraud, as well as the expected properties of a secure
biometric anti-cheat system that respects the privacy
of learners. To conclude, the integration of biomet-
rics into distance learning systems will help teachers
to effectively control student authentication, course
tracking, provide certificates in an automated manner,
analyse student behaviour during exams, the valida-
tion of the certificates of success as well as the detec-
tion of fraud with a high level of accuracy. The pro-
posed solution is effective against identity theft and
fraud attempts. Indeed, this system is able to detect
fraud automatically, it respects the privacy and confi-
dentiality of the data exchanged and solves an impor-
tant part of a major problem.
Concerning perspectives, we intend to evaluate
the system performance after merging several modal-
ities (face, keystroke dynamics, gaze, gestures and so
on) then, create a fraud model using a larger database.
In addition, the different scores can be combined to
generate a confidence index on the identity of the
learner during the examination. After all, other so-
lutions could be included to detect spoofing attacks.
REFERENCES
Arnautovski, L. (2019). Face recognition technology in the
exam identity authentication system-implementation
concept. Proceedings of Papers, page 50.
Belguechi, R., Cherrier, E., and Rosenberger, C. (2012).
Texture based fingerprint biohashing: Attacks and ro-
bustness. In 2012 5th IAPR International Conference
on Biometrics (ICB), pages 196–201. IEEE.
Belguechi, R., Rosenberger, C., and Ait-Aoudia, S. (2010).
Biohashing for securing minutiae template. In 2010
20th International Conference on Pattern Recogni-
tion, pages 1168–1171. IEEE.
Bicego, M., Lagorio, A., Grosso, E., and Tistarelli, M.
(2006). On the use of sift features for face authentica-
tion. In 2006 Conference on Computer Vision and Pat-
tern Recognition Workshop (CVPRW’06), pages 35–
35. IEEE.
Carey, P. (2018). Data protection: a practical guide to UK
and EU law. Oxford University Press, Inc.
Chauvel, C. 70% des
´
el
`
eves trichent pendant leur scolarit
´
e.
Giot, R., Hemery, B., and Rosenberger, C. (2010). Low
cost and usable multimodal biometric system based on
keystroke dynamics and 2d face recognition. In 2010
20th International Conference on Pattern Recogni-
tion, pages 1128–1131. IEEE.
Giot, R., Rosenberger, C., and Dorizzi, B. (2012). Re-
connaissance du genre par analyse de dynamique de
frappe au clavier sur texte libre.
Gupta, A., Khanna, A., Jagetia, A., Sharma, D., Alekh, S.,
and Choudhary, V. (2015). Combining keystroke dy-
namics and face recognition for user verification. In
2015 IEEE 18th International Conference on Com-
putational Science and Engineering, pages 294–299.
IEEE.
Idrus, S. Z. S., Cherrier, E., Rosenberger, C., and Bours, P.
(2013). Soft biometrics for keystroke dynamics. In
International Conference Image Analysis and Recog-
nition, pages 11–18. Springer.
Kotropoulos, C., Tefas, A., and Pitas, I. (2000). Morpho-
logical elastic graph matching applied to frontal face
authentication under well-controlled and real condi-
tions. Pattern Recognition, 33(12):1935–1947.
Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep
learning face attributes in the wild. In Proceedings
of the IEEE international conference on computer vi-
sion, pages 3730–3738.
Migdal, D. and Rosenberger, C. (2018). Towards a personal
identity code respecting privacy.
P. Beust, V. Cauchard, I. D. (2016). Enseigner autrement.
Parkhi, O. M., Vedaldi, A., and Zisserman, A. (2015). Deep
face recognition.
Saad, M., Charrier, C., and Bovik, A. C. (2012). Eval-
uation de la qualit
´
e des images sans r
´
ef
´
erence
par mod
´
elisation statistique. COmpression
et REpr
´
esentation des Signaux Audiovisuels
(CORESA).
Viola, P. and Jones, M. (2001). Rapid object detection us-
ing a boosted cascade of simple features. In Proceed-
ings of the 2001 IEEE computer society conference on
computer vision and pattern recognition. CVPR 2001,
volume 1, pages I–I. IEEE.
SECRYPT 2020 - 17th International Conference on Security and Cryptography
458