Authenticating computer access based on keystroke dynamics using a probabilistic neural network

Full text

Page 1

AUTHENTICATING COMPUTER ACCESS BASED ON KEYSTROKE
DYNAMICS USING A PROBABILISTIC NEURAL NETWORK

Kenneth Revett1, Florin Gorunescu2, Marina Gorunescu2, Marius Ene2 Sérgio Tenreiro de
Magalhães3, and Henrique M. D. Santos3

1Harrow School of Computer Science, University of Westminster, London, UK
revettk@westminster.ac.uk

3 Universidade do Minho Department of Information SystemsCampus de Azurem
4800-058 Guimaraes, Portugal
{psmagalhaes, hsantos} @dsi.uminho.pt

2Department of Mathematics, Biostatistics and Computer Science, University of Medicin and
Pharmacy of Craiova, Romania
{fgorun, mgorun, enem}@umfcv.ro

Abstract: Most computer systems are secured using a login id and password. When computers are connected to
the internet, they become more vulnerable as more machines are available to attack them. In this paper, we
present a novel method for protecting/enhancing login protection that can reduce the potential threat of internet
connected computers. Our method is based on and enhancement to login id/password based on keystroke
dynamics. We employ a novel authentication algorithm based on a probabilistic neural network. Our results
indicate that we can achieve an equal error rate of less than 5%, comparable to what is achieved with hardware
based solutions such as fingerprint scanners and facial recognition systems.


1. Introduction

The traditional method for authentication in
most information systems that are computer
based is the login/password (e.g. class C security
minimum). There are countless reports on how
such class C protected devices have been
breached, resulting in financial losses and a
dimunition in the faith people have when
transacting business over the internet (Peacock,
2004). Researchers and industry alike have
sought ways to enhance the security of
computers that currently house major financial
and scientific data throughout the world. A new
industry has been created – with the sole
purpose of providing enhanced protection from
piracy – the biometrics industry. Currently, there
are two major forms of biometrics: those based
on physiological attributes such as fingerprints,
iris, and retinal scanners and behavioural
biometrics: based on voice recognition,
signature verification and keystroke dynamics
(see Jain, 2001, 2003 for a nice review).
Physiological biometrics is based on the
notion of what we are – we each possess unique
fingerprints (even identical twins differ in their
fingerprint patters) and are therefore thought to
be spoof proof. But current literature reports
indicate that fingerprints can be spoofed (Jain,
2003, Peacock, 2004). Even though fingerprint
scanners are becoming more reliable and
cheaper to acquire, they still are subject to noise
and wear and tear. They require replacement
approximately once a year and are difficult to
place on remote access systems, such as a home
computer used in a credit card purchase over the
internet. Iris scanners and more noise tolerant,
but are certainly more expensive than fingerprint
scanners. In addition, they are (or at least appear
to be) more intrusive – a very important factor in
a biometric. Any biometric solution that is to be
used on the internet (and therefore accessed by
potentially 100s of millions of users) must be
effective and yet very unobtrusive. In the ideal
case, we wish to provide a secure site that
provides no hint that it is being heavily

Page 2

protected. Alerting users to the fact that sites on
the internet are subject to attack instills
suspicion – the exact opposite sentiment the
originators had for the internet. It was designed
to provide a medium
information could be
environment that was purely based on academic
freedom. In today’s world, the ideals can still
exist, but may require a medium that is fortified
with enhanced security features – the ethos of
biometrics.
In this paper, we present evidence that
keystroke dynamics is a viable biometric – that
provides security on par with physiological
methodologies. In addition, it provides
enhanced security in a very low-profile manner
that is acceptable by the majority of users –
virtually everyone is
authentication details such as a login id and
password. The remaining question is whether
keystroke dynamics – the typing style –
measured in terms of keystroke/keypress
duration and keyboard latency, combined with
state-of-the-art machine learning techniques are
sufficiently robust to provide the ability to
discriminate between an authentic and imposter.
If evidence can be shown that this is indeed the
case – then behavioural biometrics – specifically
keystrokes dynamics may provide a unique and
effective solution to user authentication that can
be used on personal machines as well as internet
based applications.
In this paper, we will describe the basics of
keystroke dynamics, followed by a machine
learning algorithm – a probabilistic neural
network on a dataset consisting of authentic and
imposter login attempts. We then present some
key results of our work and a brief discussion
and conclusion section.

2. Keystroke dynamics

Keystroke dynamics is a class of behavioural
biometrics that captures the typing style of a
user. By typing style, we mean it examines how
long it takes to type the login id/password, how
long we depress a key, and how long we take to
type successive keys – this is called a digraph.
Figure 1 illustrates the classic example of the
where
exchanged
scientific
in an
used to entering
Figure 1. The concept of a digraph – and the
various combinations that can be extracted and used
for biometric authentication. In this aprticular
example, the digraph is based on the character
sequence ‘no.’













data that can be extracted by entering two keys
on a standard keyboard. By collecting all
possible digraphs from the login id/password –
one can develop a model of how the person
types these credentials. Usually, there is an
enrollment exercise, where the user is asked to
enter his/her login id/password until a steady
value for each digraph is obtained. Once this
data has been collected, a reference ‘signature’
is obtained for this user. The reference is then
used on subsequent login attempts – a user with
that particular login id/password combination
has their keystroke dynamics extracted and then
compared with a stored reference value. If they
are within a prescribed tolerance limit – the user
is authenticated. If not – then the system can
decide whether to lock up the workstation – or
take some other suitable action. When devising
such a biometric solution – there is always a
trade off between being overly stringent –
rejecting every attempt to login in and being
overly lenient – allowing imposters to access the
computer. The former is usually reported as a
measure of false rejection – a type I error and
the later a false acceptance or type II error.
Another measure – called the cross over error
rate (CER) - sometimes referred to as the equal
error rate (EER) is also reported – they provide a
measure of how sensitive the biometric is at
balancing ease of use for the authentic user
while at the same time reducing the imposter

Page 3

Figure 2. The cross-over error rate is indicated as the
intersection between the FAR v FRR – when
measured against a threshold.

%
FA R FR R
C E R
T hreshold



access rate. All extant biometric systems yield a
trade-off between these two measures – those
that reject imposters effectively (low FAR) are
usually accompanied by a high FRR and vice
versa. Figure 2 depicts a typical plot of
FAR/FRR and indicates the CER point – where
the two plots intersect. The critical research
issue is ‘how can we lower the CER?’ In the
next section, we describe the historical context
in which keystroke dynamics has developed
highlighting advances
methodological approaches to this critical
question.

2.1. Background

In 1980 Gaines (Gaines, 1980) presented a
report of his work to study the typing patterns of
seven professional typists. The small number of
volunteers and the fact that the algorithm is
deducted from their data and not tested in other
people later, results on a lower confidence on the
FAR and FRR values presented. But the method
used to establish a pattern was a breakthrough: a
study of the time spent to type the same two
letters (digraph), when together in the text. Since
then, many algorithms based on Algebra and on
Probability and Statistics have been presented.
Joyce Gupta presented in 1990 (Gupta, 1990) an
algorithm to calculate a value that represents the
distance between acquired keystroke latency
times and correspondent times previously stored.
made in the
In 1997 Monrose and Rubin use the Euclidean
Distance and probabilistic calculations based on
the assumption that the latency times for one-
digraph exhibits a
(Monrose, 1997). Later, in 2000, they also
present an algorithm for identification, based on
the similarity models of Bayes, and in 2001 they
present an algorithm that uses polynomials and
vector spaces to generate complex passwords
from a simple one, using the keystroke pattern
(Monrose, 2001).
Various fuzzy logic algorithms have been
applied – mapping the variability in ones typing
patterns to a fuzzy concept. For instance,
Hussein et al (Hussien, 1989, de Ru et al., 1997
) use a combination of fuzzy clustering
algorithms - obtaining an error rate of
approximately 5-10% - depending on the
number of samples they acquired per login
id/password combination. Another study
(Tapiador, 1999) employed a fuzzy rule set in
order to classify login id/password combinations
with somewhat better success than Hussein –
although they report only their preliminary
results.
Techniques based on neural networks have
been explored – focusing on ART-2 and multi-
layer perceptrons
backpropagation algorithm. For instance,
Obadiat provides data that suggests that the
error rate can be reduced to approximately 2.4-
4.2%, depending on the exact pre-processing
performed using a non-standard neural network
(Obadiat, 1997). Sung et al., has also applied
neural networks
backpropagation) to
generating error rates on the order of 2-4%
(Sung, 2006).
Other machine learning approaches, based on
support vector machines (SVM) have been used
to address the classification problem presented
by keystroke dynamics. De Oliveira et al (de
Oliveira, 2005) have applied SVM to a small
keystroke dataset and compare their results to
standard neural network technology. The
authors claim that the SVM classifier is more
efficient and at least as accurate as neural
network technologies. Sung et al. have also
Normal Distribution
trained with the
(using
keystroke
standard
dynamics,

Page 4

applied SVM to this domain, reporting an error
rate of approximately 8-10% (Sung et a., 2006).
Lastly, Revett et al. have used the rough sets
induction algorithm to extract rules that form
models for predicting the validity of a login
id/password attempt (Revett et al., 2005). The
results indicate that the error rate can be as low
as 2-4% in many cases.
The algorithms cited are a small example of the
many approaches used to find adequate
keystroke dynamics
convenient CER. Many others could also be
referred, all with different evaluation methods,
different number of users involved (usually a
limited number of users), different number of
keystrokes required to enroll the system and
different number of repetitive operations
required to authenticate and/or identify the user.
This diversity in the algorithm parameters and in
the evaluation method makes the task of
comparing their results a very difficult one.
Furthermore, there is, in this subject, no concept
of what is a representative data sample. The
same algorithm presents different results when
tested with different volunteer groups. The only
way to compare two algorithms is to test it
against the same group.
Envisaging wide scale applications, like web-
based applications (where this method is not
executable now) one must consider the results
only if the test user group’s size is considerably
large. In this application domain one must
remember that the
necessary to execute the algorithm is a critical
factor. This is one of the driving forces behind
the approach we have adopted in this paper – we
have implemented a probabilistic neural network
(PNN) to classify whether a given login
id/password combination
authentic user or an imposter. A PNN is a
highly accurate classifier and is very efficient for
small datasets – here the dataset entails the
details for an authentic user and a set of
imposters selected randomly. In these
circumstances, we will provide evidence that
this machine learning algorithm is as accurate as
the leading results and is more efficient in many
ways. We describe the basis of the PNN
algorithm in the next section.
algorithms with a
computational effort
belongs to the
2.1. Probabilistic Neural Networks
The PNNs are basically classifiers (Specht,
1988). The general classification problem is to
determine the category membership of a
multivariate sample data (i.e. a p-dimensional
random vector x) into one of q possible groups
Ω , i = 1, 2,…, q, based on a set of
measurements. If we know the probability
density functions (p.d.f.) fi(x), usually the
Parzen-Cacoulos or
classifiers:
i
Parzen like p.d.f.
2
/22
1
1
p
1
m
( )exp
(2 )
π
2
i
m
j
i
p
j
i
xx
f x
σσ
=
⎛
⎜
⎜
⎝
⎞
⎟
⎟
⎠
−
=⋅⋅−
∑
(1)
the a priori probabilities hi = P(
occurrence of patterns from categories
the loss (or cost) parameters li associated with
all incorrect decisions given Ω =
according to the Bayesian decision rule, we
classify x into the category
li hi fi(x) > lj hj fj(x) holds true. The standard
training procedure for PNN requires a single
pass over all the training patterns, giving them
the advantage of being faster than the feed-
forward neural networks (Specht, 1988).
Basically, the architecture of PNN is limited
to three layers: the input/pattern layer, the
summation layer and the output layer. Each
input/pattern node forms a product of the input
pattern vector x with a weight vector Wi and
then perform a nonlinear operation, that is
exp[ () ( )/(2
ii
WxWx
−−−
both x and Wi are normalized to unit length),
before outputting its activation level to the
summation node. Each summation node receives
the outputs from the input/pattern nodes
associated with a given class and simply sums
the inputs from the pattern units that correspond
to the category from which the training pattern
was selected,
exp[ (
i
W
−
∑
The output nodes produce binary outputs by
using the inequality:


exp[ () (
i
ii
WxWx
−−−
∑
exp[ () (
j
jj
WxW
−−−
∑

related to two different categories
i
Ω ) of
Ω and
i
i
Ω , then,
i
Ω if the inequality
2
)]
τ
σ
(assuming that
2
) ()/(2 )]
ii
xWx
τ
σ−−
.
2
)/(2
)/(2
x
)]
2
)]
τ
σ
σ
> (2)

τ
i
Ω and
j
Ω .

Page 5

The key to obtaining a good classification
using PNN is to optimally estimate the two
parameters of the Bayes decision rule, the
misclassification costs
probabilities. In our practical experiment we
have estimate them heuristically. Thus, as
concerns the costs parameters, we have
considered them depending on the average
distances Di, inversely proportional, that is li =
1/Di. As concerns the prior probabilities, they
measure the membership probability in each
group and, thus, we have considered them equal
to each group size, that is hi = mi. As in our
previous work, we employed an evolutionary
technique based on the genetic algorithm to find
the smoothing parameters ( see Gorunescu et al,
2005 for implementation details). In the next
section, we describe the experimental methods,
with a brief description of the dataset.

3. Methods

The dataset we examined consisted of a group
of 50 subjects (all university students in a
computer science department) – 20 acting as
authentic users and the balance (30) acting as
imposters. We asked the users to enter a login
id/password of their choice (limit 15 characters
for each) with an enrollment of 10 trials. We
utilised only their digraph latencies and scan
code of the characters contained within their
login id/passwords. The data samples were
collected over a 14-day period, throughout
various periods of the day. We maintained a
running average of digraph values – where the
oldest sample of 10 was replaced – leaving a set
of 10 most recent login ids/password attempts.
We invited 30 students to act as imposters,
requesting that they attempt to hack into
someone’s account by giving them their account
holders login id/passwords. They were given 2
days to attempt over 500 trials of logging into all
20 authentic accounts and the results were
recorded. We then used our PNN to generate a
classifier that would be able to discriminate
between an authentic user and an impostor –
using subsets of the data thus obtained. We
cross validated our data in that we sampled with
and the prior
replacement until all datapoints were used in the
classification and we report the average results
from these experiments. The particular version
of the PNN we employed in this paper was the
same as that employed in previous work
(Gorunescu et al., 2005, Revett et al., 2005). We
also applied a modified version of our PNN
algorithm, that used separate smoothing factors
for each class (authentic and imposter). We
report both results in this work – and found that
using a separate smoothing factor provided
consistently better results.

4. Results

We first describe an experiment where we
examined which division used in the PNN gave
us the best classification accuracy to determine
which division provided the best accuracy. We
selected random samples for training and testing
(70/30 in this case) and applied our PNN
algorithm to these random samples. The data in
Table 1 indicate that the classification accuracy
was essentially independent on the number of
divisions employed for this dataset. Please note
that the modified PNN algorithm yielded
consistently higher results than one that
employed the same smoothing factor for both
classes – those results are not presented.

Table 1. The classification results from the
application of the modified PNN as a function of the
number of divisions used. The overall accuracy was
0.099 (9.9%) for this experiment.

Divisions TRAINING
error
10 0.0107
200.0001
300.0000
400.0003
50 0.0100
60 0.0009
70 0.0087
80 0.0917
900.0093
1000.0101
TEST error
0.1153
0.0884
0.0942
0.0923
0.0961
0.1153
0.0903
0.1003
0.1076
0.0923

Page 6

In Table 2 below, we present the classification
results after completely training the PNN on the
entire dataset, reporting the FAR and FRR as a
function of division values.

Table 2. FAR/FRR values as a function of the
division level (the same values reported in Table
1). Note the values must be multiplied by 100 to
give percentages. The values in the last row of
the right-most columns are the averages of their
respective columns.


Division
points acceptance
10 0.0483
20 0.0192
30 0.0576
40 0.0576
50 0.0576
60 0.0001
70 0.0576
80 0.0481
90 0.0288
100 0.0480
0.0422


The results presented in Table 2 indicate that our
FAR/FRR is on the order of 4% - with a total
error rate of 8.1% approximately.

5. Conclusions

We have successfully applied our modified
Specht PNN to difficult biomedical datasets and
have obtained accuracy levels comparable to
other more traditional methods. In this study,
we have employed our modified PNN to a small
dataset of login id/password digraph samples.
The modified classifier performed better than
the standard PNN algorithm by approximately
20% (data not shown). Once trained, the
classifier was able to minimise the FAR without
significantly compromising the FRR (both were
approximately 4%). These results are
comparable to traditional neural network
approaches as well
False False
rejection
0.0481
0.0197
0.0376
0.0566
0.0483
0.0021
0.0598
0.0483
0.0312
0.0427
0.0394
as more ‘modern’
approaches such as SVM. The computational
time required to train the PNN was minimal – on
the order of 2 minutes on a standard Pentium IV
desktop computer with low-to-typical amounts
of memory (128 MB) and processing speed (1
GHz).
It must be noted that these results were
obtained without any data pre-processing. We
simply collected the data, selected a random
subset for training (70%) and 30% for testing.
This algorithm is time efficient when login
id/password credentials
authentication purposes. It is a well known fact
that the training phase of the PNN algorithm
begins to degrade in terms of time efficiency
when the sample numbers are large. But in this
area of application, where we have a relatively
small number of samples for training (on the
order of 10-50) – and can select an equal number
of testing samples, training performance is not
an issue. This is in contrast to other techniques
such as the backpropagation algorithm that
requires a substantial number of training data in
order to generate accurate classification. These
advantages make the PNN a very suitable
candidate for a novel machine learning
algorithm in the context of keystroke dynamics
authentication.
We will continue to pursue the use of PNNs for
this particular domain - focusing on larger
datasets to see how the training time scales with
the number of users in the system. In addition,
we may consider some unsupervised pre-
processing of the data such as self-organised
maps. It is important to remember that if this
system is going to be used as an on-line
authentication algorithm – data pre-processing
must be kept to a minimum if it is to operate in
an unsupervised manner – which may be a
critical aspect of an on-line verification system.
One must decide how much human effort must
be spent in this process. Reduction of the pre-
processing stage also tends to reduce the extent
of overfitting the data.

5. References:

de Oliveira , M. VS, E. Kinto, Hernandez,
E.D.M, & de Carvalho,
are used for
T.C., User

Page 7

Authentication Based on Human Typing
Patterns with Artificial Neural Networks and
Support Vector Machines,

de Ru, W.G. and Eloff, J.. "Enhanced Password
Authentication through Fuzzy Logic". IEEE
Expert, Vol.12, No.6, Nov/Dec, pp.38-45, 1997.

Hussien, B., Bleha, S. & McLaren, R., . An
application of fuzzy algorithms in a computer
access security system. Pattern Recognition
Letters,9:39--43,1989.

Gaines, R. et al, Authentication by keystroke
timing: Some preliminary results. Rand
Report R-256-NSF. Rand Corp, 1980.

Gorunescu, F, Gorunescu, F,, El-Darzi, E,
Gorunescu, S., & Revett K. A Cancer Diagnosis
System Based on Rough Sets and Probabilistic
Neural Networks, First European Conference on
Health care Modelling and Computation,
University of medicine and Pharmacy of
Craiova, pp 149-159, 2005.

Jain, R. Bolle and
"Introduction to Biometrics". In "Biometrics.
Personal Identification in Networked Society".
A.Jain, R.Bolle, S.Pankanti (Eds.). pp.1-41.
Kluwer Academic Publishers., 2003

Joyce, R. and Gupta, G., Identity authorization
based on keystroke latencies. Communications
of the ACM. Vol. 33(2), pp 168-176, 1990.

Magalhães, S. T. and Santos, H. D., 2005, An
improved statistical
algorithm, Proceedings of the IADIS MCCSIS
2005.
Monrose, F. and Rubin, A. D., Authentication
via Keystroke Dynamics. Proceedings of the
Fourth ACM Conference on Computer and
Communication Security. Zurich, Switzerland,
1997.
Monrose, F. et al, Password Hardening based on
Keystroke Dynamics. International Journal of
Information Security. 2001.

SBC 2005
Sharath Pankanti.
keystroke dynamics
Obaidat, M.S. & Sadoun, B. A Simulation
Evaluation Study of neural network techniques
to Computer User Identification, Information
Sciences 102, 239-258, 1997.

Peacock, A. et al, Typing Patterns: A Key to
User Identification. IEEE Security and Privacy.
September/October 2004.

Revett K. Magalhaes, S. & Santos, H.,
Developing a Keystroke Dynamics Based Agent
Using Rough Sets, The 2005 IEEE/WIC/ACM
International Joint Conference
Intelligence and Intelligent Agent Technology
Workshop on Rough Sets and Soft Computing
in Intelligent Agents and Web Technology,
Compiegne, France, 19-22 September, 2005

Revett,K., F. Gorunescu, M. Gorunescu, E. El-
Darzi, M. Ene, A Breast Cancer Diagnosis
System: A Combined Approach Using Rough
Sets and Probabilistic Neural Networks,
Proceedings Eurocon2005 –IEEE International
Conference on “Computer as a tool”, Belgrade,
Serbia, November 21-24, pp, 1124-1127, 1-
4244-0049-X/05/$20.00 ©2005 IEEE, 2005.

Specht,D.F. “Probabilistic neural networks for
classification mapping or associative memory”.
Proceedings IEEE International Conference on
Neural Networks, 1, 1988, 525-532.
on Web
Sung, K.S. & Cho S., GA SVM Wrapper
Ensemble for Keystroke Dynamics
Authentication, International Conference on
Biometrics, Hong Kong, pp 654-660, 2006.

Tapiador, M. & Siguenza, J.A., Fuzzy Keystroke
Biometrics on Web Security. AutoID ’99
Proceedings. Workshop
Identification Advanced Technologies. IEEE. Pp
133-136., 1999.

on Automatic