1. Designing a Multi-Biometric System to Fuse Classification Output of Several Pace University Biometric Systems
Leigh Anne Clevenger, Laura Davis, Paola Garcia Cardenas, Onenetta Labeach, Vinny Monaco and James Ng
Seidenberg School of Computer Science and Information Systems
Pace University, White Plains, NY
Student/Faculty Research Day, May 2, 2014

2. Abstract
Some high-level biometric systems combine, or fuse, several biometrics to increase performance over that of an individual biometric system. This project investigates how to combine, at the classification output level, several Pace University biometric systems.

3. Overview Introduction Multi-Biometric Fusion
Pace University Biometric Systems
Methodology
Proposed Fusion Approaches
Summary of Results
Conclusions
Future Work

4. Introduction Enhancement of Security Systems
2008 Higher Education Opportunity Act In order to verify the identity of students taking online tests, the U.S Government enacted this act requiring colleges and universities to implement stricter control technologies for online systems
There are different traits for biometric systems:
Physical traits Face, fingerprint, iris, retina, hand geometry, hand vein, palm print, DNA, and teeth.
Behavioral traits Voice, signature, gait, keyboard typing patterns, and mouse movements

5. Introduction Enhancement of Security Systems
Biometric security requirements from Jain et al. include:
Universality – each person has characteristic
Distinctiveness – characteristic different between two persons
Permanence – characteristic invariant over time
Collectability – characteristic measurable
Performance – analysis speed of biometric system
Acceptability – people will tolerate use of system
Circumvention – how easily can the system be fooled

6. Introduction Why a Multi-Biometric System is Needed
To enhance overall performance of multiple biometric traits over the performance of an individual trait
To identify online test takers authentication more efficiently
Usage of behavioral traits:
Continuous authentication
Non-Intrusive
Cost-Effective
Process authentication over networks

7. Multi-Biometric Fusion Literature Review
Researchers have applied:
General vector fusing algorithms at the
Feature level
Score fusing algorithms at the
Classification level Feature classification data are vectors of biometric input data. These vectors are combined by a classifier to produce score output used for authentication.

8. Multi-Biometric Fusion Block diagram for a generic multi-biometric system using classification output level fusion

9. Pace University Biometric Systems
Pace University Biometric System (PBS)
Pace University Keystroke Biometric System (PKBS)
Pace University Systems used in Online Testing
Pace Classifier and Simulated Authentication Process

10. Methodology We investigated many fusion approaches
An accepted fusion approach would:
Analyze our PBS score output data from online testing sessions
Work with data from sources other than online testing sessions, for example Android data
We narrowed the approaches down to just a few for our experimental purposes

11. Methodology Experimental Data
Four different biometric traits were collected by Pace University’s existing frontend system and backend classifier:
Mouse Motion
Click
Scroll
Keystroke
Outputs from online examinations
14 subjects taking 10 examinations each with 10 questions and 20 minutes duration. Sample (m) data of keystroke, click, motion and scroll were actively accumulated from each session.

12. Methodology Classification Results
FAR of the click trait increases as the sample data (m) increases. FRR on the other hand, decreases as more samples are added

13. Methodology Classification Results
The keystroke FRR and FAR results are similar to the click results; however the FRR seem to be higher when there are 100 samples and only starts decreasing significantly when it is close to the 200 samples.

14. Methodology Classification Results
FRR line
The motion error rate is one of the traits with the better results. When up to 50~60 samples are used, the FRR and FAR are very low. The FAR only starts increasing significantly after 200 samples.

15. Methodology Classification Results
This graph shows how the scroll trait is very inconsistent. It is believed that this phenomenon happens due to the fact that users do not always interact with the scrolling functionality, forcing the inconsistency of the graph.

16. Methodology Classification Results
Motion ROC
In the ROC curve we can see the FAR and FRR percentages for all the four traits. Motion trait is the closest to zero.

17. Proposed Fusion Approaches Chair-Varshney Decision Fusion Rule
The local decisions (n) are combined to obtain a global decision; this combination is done by a decision fusion center (DFC):
When this rule was applied in “Decision Fusion for Multimodel Active Authentication,” using the parallel decision fusion architecture, the global decision returned better results (lower error rate) than that of an individual decision performing at its best on its own -which is exactly the type of results we are trying to obtain for this research.

18. Proposed Fusion Approaches Six Classifier Fusion Strategies
These six classifier fusion strategies are theoretically studied by L. Kuncheva, and they are as follows:
Minimum
Maximum
Average
Median
Majority Vote
Oracle

19. Summary of Results
Motion and keystroke biometric traits were the top two biometrics traits with the best performance in these studies
Simply concatenating the feature vectors from each biometric yielded an EER of 3.9%. This is worse than the single best performer, mouse motion

20. Conclusions
Motion data was the most accurate biometric trait / Scroll data was the most inaccurate due to different factors.
Proposed methods of examination for classification output level fusion:
The Chair-Varshney decision fusion rule approach
The six classifier fusion strategies (minimum, maximum, average, median, majority vote and oracle)

21. Future Work
Application of the proposed classification output fusion methods compared with the single measurement performances of the four individual biometric traits
Comparison to previous EER results
For any of these methods to be successful, we will need to yield an EER of no more than 1.139%, which is just 0.001% less than the EER of our most successful single biometric trait, motion

22. Thank you for your attention!
Questions?
Thank you for your attention!

23. Methodology Classification Results

24. Pace University Biometric Systems Pace University Biometric System (PBS)
PBS has a generic, robust backend system designed to accommodate various biometric frontend feature inputs like keystroke, mouse, stylometry and voice
PBS relies on databases of approved signature templates that were enrolled and stored earlier
PBS analyses the signature inputs against the approved data stored on the template database to verify identify and authenticate users

25. Pace University Biometric Systems Pace University Keystroke Biometric System (PKBS)
PKBS consists of two logging components and a single backend for extracting features and classification
Web-based loggers are loaded into any webpage and captures all keystroke and mouse movements, while buffering and transmitting them to a server
Cross-platform native logger is written in Java

26. Pace University Biometric Systems Pace University Systems used in Online Testing
A no-hurry system is used to authenticate user
All keystroke data from the online testing is used
To detect unauthorized users the system has to act in a minute or less before any harm is done
Keystroke data is captured using a Java applet that uses the PC Windows clock to record key press and release times in millisecond format

27. Pace University Biometric Systems Pace Classifier and Simulated Authentication Process
Implements the kNN (k-Nearest Neighbor) classifier
Used to transform a multi-class (polychotomy) problem into a dichotomy model, this is, a problem that only involves two classes. These classes are:
within-person (authenticated)
between-person (not authenticated)
Simulated Authentication Process
User’s keystroke sample requiring authentication is first converted into a feature vector. The difference between this feature vector and an earlier-obtained enrollment feature vector from this user is computed, and the resulting difference vector is classified as one of the two previous classes.