1. IMPAIRED-USER INPUT SCENARIOS FOR KEYSTROKE BIOMETRIC AUTHENTICATION
ANALYZING
IMPAIRED-USER INPUT SCENARIOS FOR KEYSTROKE BIOMETRIC AUTHENTICATION
Gonzalo Perez
J. Vinnie Monaco
Advisor – Dr. Charles Tappert
Monday, September 17, 2018

2. KEYSTROKE BIOMETRIC RATIONALE
According to
ROY MAXION
a research professor of computer science at Carnegie Mellon
“MOTIONS THAT ARE PERFORMED NUMEROUS TIMES, ARE GOVERNED BY MOTOR CONTROL, NOT DELIBERATE THOUGHT. THAT IS WHY SUCCESSFULLY MIMICKING KEYSTROKE DYNAMICS IS PHYSIOLOGICALLY IMPROBABLE.”

3. THE STUDY EXAMINES HOW TO BETTER ANALYZE ABNORMAL TYPING BEHAVIOR
FOCUS OF STUDY
THE STUDY EXAMINES HOW TO BETTER ANALYZE ABNORMAL TYPING BEHAVIOR
IMPAIRED
DISTRACTED
One-handed, typing behavior may not be performed numerous times and may not necessarily be governed by motor control.
One handed typing behavior has been found to be erratic when compared to standard two-handed typing behavior, and this study attempts to shed light as to how to better authenticate distracted or impaired typing scenarios.

4. OTHER STUDIES TO FURTHER STRENGTHEN KEYSTROKE BIOMETRICS INCLUDE:
OVERALL GOAL
TO STRENGTHEN CURRENT KEYSTROKE BIOMETRIC SYSTEMS BY DEVELOPING A MODEL WHICH WILL ACCOUNT FOR VARIOUS IMPAIRED OR
DISTRACTED INPUT SCENARIOS
OTHER STUDIES TO FURTHER STRENGTHEN KEYSTROKE BIOMETRICS INCLUDE:
Emotion
Word Quality
Short Structured Tests
Arbitrary Long-text
Stylometry

5. OBJECTIVES OBJECTIVE 1 OBJECTIVE 2 OBJECTIVE 3 OBJECTIVE 4
To better understand how a keystroke biometric system handles users that have been impaired or distracted
Collect data by simulating a quiz to capture arbitrary free-text input from users entering data using various scenarios
Run various experiments to determine which model best authenticates users
Continuously analyze results from experiments and modify parameters to improve EER’s

6. IMPAIRED USER REQUIREMENTS BIOMETRIC RATIONALE
Research will focus on constrained user input
BIOMETRIC SYSTEMS NEED TO CONSIDER USER REQUIREMENTS AS USERS MAY
HAVE SOME PHYSIOLOGICAL AND MEDICAL FACTORS THAT AFFECT THE
USABILITY AND EFFICIENCY OF BIOMETRICS
VISUALLY IMPAIRED SUBJECTS
May suffer from Aniridia, absence of an iris
Person may be blind
Person may have eye tremors
HEARING IMPAIRMENTS
Subject may not be able to hear instructions that are needed for a biometric system
Speech may be affected due to hearing loss

7. IMPAIRED KEYSTROKE BIOMETRIC RATIONALE
A traditional keystroke biometric system would require the user to always input keystrokes in a normal state. (using both hands)
In a normal setting, users may type using one hand only if they are on the phone or drinking a cup of coffee or perhaps one hand or arm is injured.
TO INTRODUCE VARIOUS KEYSTROKE INPUT METHODS IN ORDER TO STRENGTHEN THE VALIDITY OF THE KEYSTROKE BIOMETRIC SYSTEM
THE STUDY WILL ANALYZE VARIOUS KEYSTROKE INPUT METHODS TO DETERMINE IF
THE USER CAN STILL BE AUTHENTICATED OR PROVIDE A NEW MODEL.

8. DATA CAPTURE- PARTICIPANTS
EXPERIMENT FEATURED 81 STUDENTS WHICH SUCCESSFULLY ENROLLED
Entered data as part of a simulated quiz
The quiz format encouraged users to enter arbitrary long-text input responses
Various scenarios were introduced
Both Hand Normal
Left Hand Only
Right Hand Only

9. HARDWARE
90%
SINCE THE QUIZZES WERE TAKEN TOWARDS THE END OF CLASS SESSION, 90% OF USERS UTILIZED AN HP VMWARE PLATFORM USING A STANDARD HP KEYBOARD
Some of the users could not complete the exam in class and had to use their laptop or desktop at home. The system prompts the users to identify which system they were using to enter their keystrokes.

10. DPS DISSERTATION DATA CAPTURE MOODLE

11. Every question in each exam was unique
MOODLE DPS QUIZ
STUDENTS WERE ASKED TO
Log into a Moodle learning platform which is an open source alternative from blackboard
Complete three exams, each exam asking them to answer five questions related to the content that was being covered in the introductory computer science course
Every question in each exam was unique
Keystrokes were logged by a JavaScript event logging framework which was embedded into a Moodle learning platform

12. DATA CAPTURED
81 Users entering at least 100 CHARACTERS for every question in each scenario listed below
Both Hands B
Left Hand Only L
Right Hand Only R

13. FIRST ITERATION EXPERIMENTS & RESULTS
AFTER CAPTURING THE KEYSTROKES FROM USERS THROUGH VARIOUS NORMAL AND IMPAIRED SCENARIOS, WE BEGAN TO RUN SIMULATIONS:
B train – B test
B train – L test
B train – R test
Initially, we were hoping to find decent results with this experiment and fine tune the dataset in order to provide
a novel method to authenticate one hand only typing

14. 38% FIRST ITERATION EXPERIMENTS & RESULTS (cont’d)
TABLE 1- FIRST ITERATION RESULTS
TRAIN DATA
TEST DATA
FEATURES
EER (%)
BOTH
Both
All
3.3
Left
38.04
Right 38
Our results were not encouraging with this method as B Train, L and R Test gave us EER’s in the upper 38% range. Typing one handed proved significantly alter a user’s typing behavior and as a result gave us very poor EER rates
38%

15. SECOND ITERATION EXPERIMENTS & RESULTS
SECOND EXPERIMENT
L train – L test
R train – R test
As a result from our initial findings, we realized that one handed typing behavior was SIGNIFICANTLY DIFFERENT than the typing behavior of two handed typing
Next we decided to experiment to determine if one handed typing behavior was so erratic, that it would be difficult to authenticate with the same one handed test sample

16. SECOND ITERATION EXPERIMENTS & RESULTS (cont’d)
TABLE 2- SECOND ITERATION RESULTS
TRAIN DATA
TEST DATA
FEATURES
EER%
LEFT
Left
All
13.96
RIGHT
Right
15.61
We were pleased to see the results of the single handed train and test data.
The EER rates were relatively low, in the mid-teens which concludes that user one handed samples do have a conclusive pattern that can be analyzed and authenticated with a keystroke biometric system with relative efficacy.
However, we wanted to try another experiment to determine if we could improve the error rate to a lower number if possible.

17. THIRD ITERATION EXPERIMENTS & RESULTS
AFTER CAPTURING THE KEYSTROKES FROM USERS THROUGH VARIOUS NORMAL AND IMPAIRED SCENARIOS, WE BEGAN TO RUN SIMULATIONS:
B train, L test, L Features
B train, R test, R Features
L train, L test, L Features
R train, R test, R Features
With the intent of lowering EER rates further, we wanted to experiment by filtering features which would better authenticate impaired users.
We created the feature sets for left/right sides by filtering the linguistics features to those that contain keys on each side of the keyboard

18. THIRD ITERATION EXPERIMENTS & RESULTS (cont’d)
TABLE 3- THIRD ITERATION RESULTS
TRAIN DATA
TEST DATA
FEATURES
EER%
BOTH
Left
35.85
Right
36.79
LEFT
22.75
RIGHT
26.64
Much to our dismay, the left/right feature filter actually worsened the results of our testing. The initial hypothesis was that if a user is typing with one hand, they would perform more natural typing behavior on the segment of the keyboard with the one hand that they were typing with.
Our results did not align with this hypothesis and the reason could be related to omission of the segments. OMITTING A SEGMENT of the keyboard excludes many features of the keylogger system which DEGRADED, not improved the results of the experiment.

19. FOURTH ITERATION EXPERIMENTS & RESULTS
THE GOAL OF A FOURTH ITERATION WAS TO
Combine some of the datasets in order to exclude the need for a system to initiate a detector function and then engage various fallback procedures in order to authenticate a user.
Therefore, we combined all of the samples into one experiment which included approximately 1200 DATA POINTS PER USER which needed to be split into 5 SAMPLES. The experiment was so large that it required two days to complete.

20. FOURTH ITERATION EXPERIMENTS & RESULTS (cont’d)
TABLE 3- FOURTH ITERATION RESULTS
TRAIN DATA
TEST DATA
FEATURES
EER%
BOTH-LEFT-RIGHT
Both-left-right
All
12.41
Both
4.86
Left
15.82
Right
15.74
The results were very encouraging as the EER’s were with the standard margin of error when comparing the training and testing conditions separately.
Fewer assumptions are made with this method
The method does require that B, L, R samples be collected during the enrollment phase.
System doesn’t need to know whether a sample is one-handed when testing.
Avoids requiring a detector and fallback procedure for one-handed samples.

21. DISCUSSIONS
Each iteration provided valuable information which assisted us in expanding and developing the research
Initially, we expected to find patterns between the both hand sample and the one handed sample which could have been identified, isolated and matched accordingly
One handed samples were too erratic and could not be matched with decent rates using our tools
Keyboard segmentation actually worsened results
Combining B+L+R, and testing across all scenarios proved to be the best approach that would authenticate users and provide a seamless test implementation process

22. CONCLUSION 01
The major contribution of this research study was to provide a novel approach to authenticate impaired users of a keystroke biometric system 02
The research is an important step towards creating a more robust keystroke biometric system and is also an essential topic that must be considered when designing any biometric system, albeit physical 03
Furthermore, our novel approach of combining datasets consisting of various scenarios and then subsequently testing across single scenarios can be an approach to consider for other behavioral biometric systems

23. BIOMETRIC IDENTIFICATION COMPETITION PAPER SUBMITTED TO ICB201504 02
Participants competed against each other to obtain the highest classification accuracies and submitted classification results through an online system
A paper on one-handed keystroke biometrics which was based from our research was submitted and accepted to the International Conference on Biometrics (ICB 2015) in Phuket, Thailand 01 4 1
Competition participants designed classification models trained on the normally-typed samples in an attempt to classify an unlabeled dataset that consists of normally-typed and one-handed samples
We provided our unlabeled dataset and 9 teams from all over the world competed for the top spots 3 2 03

25. ICB 2015 – Phuket, Thailand