User Identification by Means of Sketched Stroke Features

1
User Identification by Means of Sketched Stroke
Features

• Brian David Eoff
• 14 December 2007

2
Introduction
My goal was to determine if additional features
that could be collected from a tablet could be
used to determine the identity of the drawer.
3
Why would you do that?
I became interested when I realized that the
tablet was able to take in a lot of information
about how the users drew - beyond x, y and
time. This information could be used to divide
strokes in a collaborative system. Or be used to
recognize a person for the purpose of
authentication.
4
Is any of this true?
Is handwriting consistent? Can I recognize a
person solely on their handwriting? Questioned
Document Examination, Forensic Handwriting
Identification. They only get to look at the ink
on the page, I get to know how the person drew it.
5
(No Transcript)
6
Preliminary Study
I created a drawing panel that was able to record
tilt and pressure. Six people were asked to
contribute a writing sample each day for three
days.
7
(No Transcript)
8
(No Transcript)
9
(No Transcript)
10
Six Features
Average Pressure, Standard Deviation of Pressure,
Average Tilt X, Standard Deviation of Tilt X,
Average Tilt Y, Standard Deviation of Tilt
Y. All these are available as part of the tablet
events in NSEvent (Objective-C Cocoa Mac OS
X) These values are not easily gotten with Java
11
(No Transcript)
12
Linear Classifier
Take each individual stroke and classify it to
who drew it. Stroke length ranged from 1 to 167.
13
(No Transcript)
14
(No Transcript)
15
Discussion
Tilt is the most important metric by far. The
only time pressure was very helpful was when
classifying User Fives strokes. User Two and Six
are very close on pressure values, hence it
wasnt usable. Using two samples for training
was necessary to even out the data.
16
Future Work
? More Data (Want to Go Beyond 3) ? Experiment
with HMMs and Bayesian Networks (Beat the basic
Linear Classifier) ? Give more consideration to
context (time and position) ? Build a
collaborative (networked) drawing panel ?
Tailoring Recognition to specific users, or types
of users (clustering).
17
Conclusion Questions
???