Biometric Authentication using Online Signatures

1
Biometric Authentication using Online Signatures

• Berrin Yanikoglu
• Faculty of Engineering Natural Sciences
• Sabanci University
• (joint work with Alisher Kholmatov)

2
Biometrics

• Utilization of physiological characteristics or
  behavioral traits
• for identity verification of an individual.
• Face, Fingerprint, Iris, Retina, Voice, Hand
  Geometry,
• Signature, Keystroke Dynamics,

Biometrics can not be - Forgotten, Lost,
Stolen, - Copied?
3
Biometrics

• Which one?
• Unique? retina, iris, fingerprint,
  face, signature, hand, voice...
• Copied easily? retina, iris, face, fingerprint,
  signature, hand, voice...
• Acceptable? face, signature, voice, hand,
  fingerprint, iris, retina...
• Cheap/easy? face, signature, voice, hand,
  fingerprint, iris, retina

4
Signature Verification
Does this signature belong to the claimed ID?
5
Signature Verification

• Static (offline)
• Image of the signature is the input
• Widely used
• Applications bank check clearing

3/8/2002 Dear
John, ........... ....................... ........
........................... ......................
............. ..................
Best regards,

• Dynamic (online)
• Pressure-sensitive tablets used to capture the
  image and dynamic properties of the signature
• More unique, more difficult to forge
• Applications building entrance, credit card
  processing, added security to laptops/PDAs...

6
On-Line vs. Off-Line
On-line
Off-line
7
Input

• 100-200 samples/sec
• 1000 pixels/inch resolution
• - x,y coordinates of the signature trajectory
• on pen down
• - time stamp at each sample point
• records pen-up time
• pressure at each point
• - pen tilt at each point

8
Input
9
Common Preprocessing Techniques

• Resampling
• to reduce number of points
• to obtain equidistant points capturing the shape
• Gaussian filter to smooth signature
• local speed, accelaration, jerk (third
  derivative)
• Critical points (trajectory change, start/end
  pts.) extraction
• positions retained before resampling
• propagated to smoothed signature by interpolation

10
Critical Points
11
Preprocessing

• Preprocessing removes variability across
  different signatures of the user, making the job
  of verification easier
• Preprocessing also removes properties peculiar to
  a particular user
• Preprocessing should be carefully designed taking
  into considerations the hardware to be used and
  circumstances it will be used.

12
Features

• Features can be divided in to two categories
• Global properties of the whole signature. (Ex.
  total writing time, bounding box, average pen
  speed, number of strokes, Fourier
  descriptors...).
• Local properties that refer to a position within
  the signature (local curvature, speed,
  acceleration...).
• Global features are faster to compute and
  compare,
• but usually have higher error rate.

13
Local Features

• Spatial features
• absolute coordinates w.r.t the center of the
  signature
• x y offset btw. two consecutive points
• sine cosine of the angle with x axis
• curvature
• grey values in the 9x9 neighbourhood...
• Dynamic features
• absolute speed and relative speed (absolute speed
  normalized by the average signing speed) at each
  point
• acceleration, pressure, pen tilt...

14
Signature Trajectory
15
Matching

• Feature matching (global)
• Each signature is described by a set of global
  features
• Comparison to another signature (equal number of
  features) using various pattern recognition
  techniques
• Point-by-Point matching (local)
• Each signature is described by a set of local
  features extracted at each sampling point
• sequence of feature vectors.
• Comparison to another signature (different number
  of features) via
• Dynamic time warping
• Hidden Markov Models
• String matching

16
Dynamic time warping

• Goal find the best non-linear alignment between
  two sequences, such that the total distance is
  minimized.

17
Dynamic Time Warping
Linear in the length of the two signatures.
18
Dynamic time warping
Dist (p1,p2) p1-p2 if p1-p2 gt
threshold 0
otherwise
19
Verification

• Compare the test signature (Y) to the reference
  signatures (Xi)
• belonging to the claimed identity.

• Y

x2
xT
x3
x1
x5
x4
20
Which Distance?

• Typically, the distance to the nearest reference
  signature or the distance to a template
  signature, was chosen, in an ad-hoc manner, to
  classify the signature as genuine or forgery.

21
Enrollment

• 5-10 Reference Signatures
• After cross alignment, compute the
• Average Minimum Distance to closest neighbor
  mmin
• Average Maximum Distance to farthest neighbor
  mmax
• Average Distance to Template
• m templ
• Variance of Distances

22
Proposed Method(How to use these distances?)

• use these distances (min, max, template),
  normalized by the corresponding profile averages
  or transformed to a z-scores, as features
• dmin (Y) / mmin (RID)
• f(Y) dmax (Y) / mmax (RID)
• dtempl (Y) / mtempl(RID)
• where Y is the test signature and RID is the
  reference set belonging to the claimed ID.
• - find the separating boundary using pattern
  classification techniques
• Linear Classifier with PCA
• Bayes classifier
• Support Vector Machines

23
Feature Space(Distances normalized by
corresponding profile averages)
24
Performans Evaluation

• False Acceptance Rate (FAR)
• False Rejection Rate (FRR)

• Equal Error Rate (ERR)

25
Our Systems Performance

• Input
• 100 samples/sec
• only x,y coordinates on pen down
• Data
• 100 persons
• DS1 182 Genuine Signatures
• DS2 313 Skilled Forgery Signatures
• DS3 124 Genuine Signatures
• Results
• EER 1.46
• (Signatures were collected over the period of 6
  months after the enrollment to the system.)

26
Sample Signatures
27
Our Systems Performance
28
Experiments
29
State-Of-The-Art Performance Results
results are on random forgeries
we won the first place in the First International
Signature Verification Competition (SVC 2004)!
30
Conclusions

• The problem is approached as a 2-class
  classification problem, using standard pattern
  recognition techniques
• results around 1.4 equal error rate
• Forgery qualities vary widely
• standard tests and common databases are starting
  to become available
• Signature verification has its place among other
  biometrics
• even though it has inherent limitations

31
(No Transcript)
32
Questions Comments

• berrin_at_sabanciuniv.edu