Dynamic Authentication by Typing Patterns TJHSST Senior Research Project Proposal Computer Systems Lab 2009-2010

Dynamic Authentication by Typing Patterns
TJHSST Senior Research Project Proposal

Computer Systems Lab 2009-2010
Luke Knepper
luke@lukeknepper.com
October 22, 2009
Abstract
This project will analyze and test the accuracy of dynamic typing
pattern authentication methods. The program will generate a dy-
namic set of text that the user will be prompted to type, and then
it will feed the user’s typing characteristics through neural network
structures. Experimentation will be done to determine the most ac-
curate neural network structures and data collection conditions. This
process will be an improvement on normal typing pattern authenti-
cation techniques, which use static passwords rather than dynamic
text.
Keywords: authentication, security, typing patterns, neural net-
works
1 Introduction
1.1 Background
Current authentication techniques span all three tiers of security:
• Tier 1 - Identification (usernames)
• Tier 2 - Knowledge and Possession (passwords, ID cards, security ques-

tions, etc.)
1
• Tier 3 - Skills and Capabilities (captchas, voice recognition)
In the online world, usernames can be stolen, passwords and security ques-
tions can be guessed, and captchas can be cracked. The process of analyzing
typing patterns, a tier 3 security method, has been around for decades, but
only recently has been put to use commercially. Currently, commercial prod-
ucts only analyze the user’s typing patterns when the user is typing their
password, a simple, static word. Any computer hacker can easily write a
keylogger to record the user’s keystrokes when typing this word and then
simulate that process to gain access to the user’s account.
1.2 Purpose
Dynamically generating content for the user to type and then analyzing that
content will make hacking the algorithm considerably harder, especially when
more advanced typing characteristics are used. This project will study the ac-
curacy of authenticating users by their typing characterisitics using dynamic
text blocks.
1.3 Procedure
To refine the neural network, an online system will be created to collect
massive amounts of typing data from volunteers. This data will be used to
train the different neural network structures in order to determine the most
accurate structure for this purpose. Variations that will be tested include
the use of backpropagation to train the network, genetic techniques to breed
the weight vectors, and a Hebbian learning system to do so efficiently. To
refine the user set-up and log-in processes, a series of tests will be completed
using a sample of user volunteers on multiple sittings. The users will create
dummy accounts and then attempt to log in to their own accounts and the
accounts of others. The accuracy of different set ups will be measured.
1.4 Hypothesis
Based on research, the hypothesis is that a multi-layered genetically-bred
neural network using averages of mid-sized typing data will be the best trade-
off for accuracy and efficiency, but that this set-up will not offer the accuracy
needed to be a stand-alone authentication system, and will instead have to
2
work in conjunction with older techniques, such as passwords and security
questions.
2 Background
There is currently a patent (US 6151593) for an authentication scheme by
typing pattern analysis. This method reads in the time between keystrokes
for a user when typing their password and then trains a three-layered neural
network to this combination. It does not allow for a dynamically-generated
content to be used, and does not test the different lengths of passwords. This
is the most basic application of typing techniques for authentication, and this
project will extend beyond these simple methods.
Multiple typing pattern log-in software packages exist, such as Psylock,
but they all have the same weakness as the patent above: they rely on
a static password instead of dynamic typing content and therefore can be
easily hacked.
An independent team of researchers, headed by Peacock et al., tested the
effect of many variations, including neural network set-up, password length,
acceptance stringency, data used, and function used. They found the most
effective neural network structure from their tests was to use a set-up where
many independent neural networks are trained on different cores (i.e. par-
allel processing) using randomly generated starting weight vectors. During
the training, the best weight vectors are picked and created using genetic al-
gorithms. They found the smaller (more stringent) acceptance ranges came
up with a good amount of false alarms (when it didn’t let the correct user
in, happened 22% of the time) but also minimized break-ins (when the in-
correct user was let in, happened 3% of the time). They also found the
most effective password length was 7 characters, a mid-sized password (the
longer passwords had no break-ins but many false alarms, and the shorter
passwords had many break-ins). They concluded that a linear evaluation
function was more effective than a quadratic function and that averaging
was more effective than counting each training run. They suggest their re-
sults can be improved (75% success, 22% false alarm and 3% break-in for
their best algorithm).
Another team working under L. Maisuria compared the accuracy of neural
networks compared to cluster algorithms. A multi-layered perceptron-based
neural network which learned on the Hebbian learning theory was used, as
3
were ten different metrics to compare the clusters for the clustering.
They tested the different algorithms by recruiting twenty volunteers to par-
ticipate in three different sittings. In the first sitting, they all trained their
neural networks by typing in their password sixty times. In the next sittings,
they attempted to log in to their accounts and break into the accounts of
others. The sittings were spaced out by one week.
The study found that the clustering methods were slightly more accurate
than the neural networks in rejecting impostors. They only found an av-
erage of 80% to 90% accuracy in rejection rate, not enough to comprise a
stand-alone security system but certainly good enough to be used in con-
junction with traditional methods. They found that all keystrokes should be
measured, including the beginning and end strokes to the enter key, for the
highest accuracy. They found that allowing impostors to observe the users
typing before attempting to break in to their accounts had little effect on the
accuracy.
3 Design Criteria
The final version will have two stages:
• The set-up phase, where users will be prompted to type a longer amount
of text with which to train the system.
• The log-in phase, where users will be prompted with a dynamically

generated text block in order to gain access to their account
At completion, this project should be composed of the following:
1. Neural Network
• The neural network should dynamically adapt to allow for an in-

crease or decrease in layers, easily allowing for a multi-layered
neural network structure.
• The weight vector will be trained by starting with a population of
randomly-generated vectors, training them slightly via back prop-
agation, then breeding them to form a new generation, training
them again, and repeating this process until an optimal vector is
created. This process should be done across multiple processors
or cores to speed it up.
4
2. Test Data Collection
The test data will be collected via an online application, which will
prompt subjects to enter in a paragraph and then store their keystroke
information for future experimentation.
3. GUI
The GUI will be optimized for the optimal size found by the previ-
ous testing and experimentation, whether this be one word, sentence,
paragraph, or a combination of those.
4. Authentication
The application will determine what other tiers of authentication are
needed in order to have the most accurate system. It is predicted to
include a username and possibly a password as well, in addition to the
passage which the user is prompted to type.
4 Procedure
The development phase will be separated into three parts:
• Proof of Concept
• Neural Network Development
• Input Optimization
4.1 Proof of Concept

A simple proof-of-concept was completed in October ’09. The program
prompts two users to both type a sentence and uses their data to train a
simple single-layer neural network. It then prompts the users with a third
sentence and instructs one of them, whose identity is unknown to the com-
puter, to type the third sentence. It runs the final data through the trained
neural network and determines which user typed the third sentence. It has
been tested twenty times with different users and has a 90% accuracy to
this date, proving that this concept does work. This simple implementation
leaves much to be desired.
5
4.2 Neural Network Development
In order to test the accuracy of different neural network structures, large
amounts of test data must be collected. A Java applet will be constructed
and posted online to collect this data. The applet will consist of a text box
in which the user can type and a text area which will display messages to
the user to prompt them to type certain text. They will be prompted for
different kinds of text, including a word (repeated times), a sentence, and
a short paragraph. The typing data will be collected and stored on to the
server.
This data will then be used in order to optimize the neural network. An
automated testing algorithm will be created which will repeatedly train a
given neural network type with one set of typing data, and then run typing
data from the same person and from others through the network to com-
put the accuracy. The program will repeat this process for many different
typing pattern sets, and then compute the average accuracy of the given net-
work type. Different networks will be tested to determine which is the most
accurate.
4.3 Input Optimization

Using the data collected in the previous stage, the automated testing system
will be modified to determine the most accurate data size. Instead of varying
the neural network structure, now the corpus size will be varied. Different
sizes include single word, sentence, and short paragraph. It has been shown
that a bigger corpus increases accuracy, but is less efficient. By comparing
the accuracy and efficiency of the different lengths, the experimenter can
determine the optimal length.
The consistency over time will then be tested using a small group of
about twenty volunteers. The test will be broken up into three sittings, simi-
lar to the research discussed above. The first sitting will have the volunteers
creating dummy accounts in the optimized system by repeatedly typing dy-
namically generated text. In the next sittings, spaced out by one week and
one month, the subjects will attempt to log in to their accounts and into
the accounts of others (simulating impostors). The accuracy in allowing a
correct log-in and rejecting an impostor will be computed to determine if the
method holds up over periods of time.
6
References
[1] Cho, S. and Han, D. ”Apparatus for authenticating an individual based
on a typing pattern by using a neural network system.”
http://www.freepatentsonline.com/6151593.html
[2] Peacock, A. et al. ”Typing Patterns: A Key to User Identification.”

http://www2.computer.org/portal/web/csdl/doi/10.1109/MSP.2004.89
[3] Maisuria, L. ”A COMPARISON OF ARTIFICIAL NEURAL NET-

WORKS AND CLUSTER ANALYSIS FOR TYPING BIOMETRICS
AUTHENTICATION.”

Dynamic Authentication by Typing Patterns TJHSST Senior Research Project Proposal Computer Systems Lab 2009-2010

Uploaded by

Copyright:

Available Formats

Dynamic Authentication by Typing Patterns TJHSST Senior Research Project Proposal Computer Systems Lab 2009-2010

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dynamic Authentication by Typing Patterns TJHSST Senior Research Project Proposal Computer Systems Lab 2009-2010

Uploaded by

Copyright:

Available Formats

Dynamic Authentication by Typing Patterns

TJHSST Senior Research Project Proposal

• Tier 1 - Identification (usernames)

• Tier 2 - Knowledge and Possession (passwords, ID cards, security ques-

• The log-in phase, where users will be prompted with a dynamically

• The neural network should dynamically adapt to allow for an in-

• Neural Network Development

4.1 Proof of Concept

4.3 Input Optimization

[2] Peacock, A. et al. ”Typing Patterns: A Key to User Identification.”

[3] Maisuria, L. ”A COMPARISON OF ARTIFICIAL NEURAL NET-

You might also like