Presentation of Master

1
Presentation of Masters thesis

• Gait analysis Is it possible to learn to walk
  like someone else?
• Øyvind Stang

2
Introduction

• Definition of biometrics The science and
  technology of measuring and analyzing biological
  data. (http//searchsecurity.techtarget.com)
• 2 categories Behavioural and non-behavioural
• Behavioural Keystroke, voice, gait.
• Non-behavioural Fingerprints, face, iris.
• Impersonation is a well-known problem.

3
Gait

• The gait is a feature that is different from
  person to person.
• Because of this, it may be used as a biometric.
• The aim of gait authentication is to look at
  different features in a persons gait, and based
  on these, analyze whether they belong to Person
  X or not.

4
Gait cycleJain et al. Biometrics Personal
Identification in Networked Society (1999)
5
Gait

• 3 main categories of gait authentication.
• Image based gait authentication To use (a)
  camera(s) to capture images of a walking person,
  and then analyzing these images, looking for
  certain features.
• Floor-sensor based gait authentication.
• Accelerometer based gait authentication To use a
  sensor containing an accelerometer, which
  measures the acceleration in three directions,
  and then analyze the gait based on this
  acceleration data.

6
Problem (and relevant questions)

• How easy or difficult is it to learn to
  impersonate someones gait?
• If it is easy, what does that say about the
  security of gait authentication?
• Are some peoples gait more difficult to learn
  than others? gt Sheep.
• Are some people better impersonators than others?
  gt Wolves.

7
Previous work

• Robustness of biometric gait authentication
  against impersonation attack by Davrondzhon
  Gafurov, Einar Snekkenes, and Torkjel Søndrol.
• Accelerometer based.
• Distance metric The Cycle Length Method.
• Their null-hypothesis (H0) Deliberately trying
  to imitate another person will give results.
• Results p-value0.0005, i.e. too little evidence
  to support the hypothesis.

8
Prototype

• Created a prototype that reads acceleration data
  from a (ZSTAR) sensor.
• The acceleration data is then plotted in a
  coordinate system as 4 graphs, i.e. the x-graph,
  the y-graph, the z-graph, and the r-graph.
• The r-graph is the resultant graph, where each
  plot is calculated using the following formula

9
(No Transcript)
10
Prototype

• The prototype reads and plots gait data
  continually in 5 seconds before it stops.
• Created 5 gait templates of different degrees of
  difficulty (each lasting 5 seconds).
• Template A Two slow steps. Rather trivial.
• Template B A few more steps. Also rather
  trivial.
• Template C The authors natural gait.
• Template D Fast and shuffling steps.
  Difficult.
• Template E Slow, oscillating steps. Difficult.

11
Prototype

• When the program starts, the 4 graphs from one of
  the templates are plotted in the coordinate
  system.
• When we give instructions to the program to start
  reading the acceleration data, it reads from the
  sensor, and plots the incoming data in the same
  coordinate system.
• After it has read and plotted in 5 seconds, it
  stops, and the correlation between the templates
  r-graph, and the users r-graph is calculated.

12
Prototype

• A score between 0 and 100 is given, which is
  based on this correlation value.
• Correlation between 2 datasets A(a1,,an) and
  B(b1,,bn) (Pearsons r)
• In order to get a score between 0 and 100, the
  absolute value of the correlation coefficient is
  multiplied with 100.

13
The Experiment

• On the authentication lab on GUC.
• 13 participants, all men, but of different weight
  and height.
• The coordinate system was displayed on a big
  screen, so the participants could see the
  template graphs while they were walking towards
  it.
• They attempted to imitate each template 15 times.

14
The Experiment

• The participants did not see the actual gait, but
  were given a simple explanation at the beginning
  of each template.
• The aim was to see if their scores had a positive
  increase from the beginning (attempt no 1) to the
  end (attempt no 15).
• The score from each attempt was displayed in a
  pop-up box after the attempt was completed.

15
After one attempt, the screen looked e.g. like
this
16
Results

• Linear regression Finding a linear function,
  ymxb, that fits to the data.
• Tells us whether the tendency in data is
  increasing (by having a positive m) or decreasing
  (by having a negative m).
• We used Linear regression in order to analyze the
  progression from attempt no 1 to attempt no 15.

17
Template A m0,089 (5,08 degrees)
18
Template B m0,041 (2,37 degrees)
19
Template C m0,051 (2,90 degrees)
20
Template D m0.036 (2.05 degrees)
21
Template E m0.075 (4.30 degrees)
22
Analysis of results

• In all 5 templates, there is a increase in the
  scores from the 1st to the 15th attempt.
• The increase is not too large.
• Some participants scored generally high, but had
  a small increase in the scores. (Bad?)
• Some participants scored generally low, but had a
  large increase in the scores. (Good?)

23
A new attempt to analyze the results

• Since Template C contained the authors natural
  gait, it was interesting to see how good he
  managed to score when trying to walk like
  himself.
• Template C gt 150 attempts.
• The median value was 50.73 points, i.e. the
  author scores above 50 points half of the times.
• How many and how often did the participants
  manage to exceed 50 points?
• Threshold 50 pts.

24
Template of times
A Never 9/13 1 time 4/13 30.8
B Never 7/13 1 time 2/13 2 times 2/13 5 times 1/13 6 times 1/13 46.2
C Never 6/13 1 times 3/13 2 times 2/13 3 times 1/13 9 times 1/13 53.8
D Never 8/13 1 time 4/13 3 times 1/13 38.5
E Never 10/13 2 times 2/13 5 times 1/13 23.1
25
Conclusion

• It seems rather easy to learn to walk like
  someone else. Many participants (20-60) managed
  to exceed the authors median score.
• If our conclusion turns out to be true, then gait
  authentication should not be used as the only
  authentication technique.
• The risk of impersonation will then be too large.

26
What must be considered?

• Wolves and sheep?
• Few participants?
• Few natural templates?
• Too little variation between the participants?
• Other distance metrics (algorithms)? Our
  conclusion is not necessarily true for all
  algorithms.
• The graphs were not shifted before the
  correlation was calculated.

27
Further work

• A bigger experiment with more (natural)
  templates.
• Involving a camera.
• Improved visual interactive feedback.
• Sound based feedback.
• Difference between different groups.
• The issue of wolves and sheep.