Windowed DMD as a Microtexture Descriptor for Finger Vein ...eprints.mdx.ac.uk/19488/1/WIFS2015_Tirunagari.pdf · Windowed DMD as a Microtexture Descriptor for Finger Vein Counter-spoofing

Middlesex University Research Repository

An open access repository of

Middlesex University research

http://eprints.mdx.ac.uk

Tirunagari, Santosh and Poh, Norman and Bober, Miroslaw andWindridge, David (2015) Windowed DMD as a microtexture descriptor

for finger vein counter-spoofing in biometrics. In: 2015 IEEEInternational Workshop on Information Forensics and Security (WIFS),

16-19 Nov 2015, Rome, Italy.

http://dx.doi.org/10.1109/WIFS.2015.7368599

Supplementary material

Available from Middlesex University’s Research Repository athttp://eprints.mdx.ac.uk/19488/

Copyright:

Middlesex University Research Repository makes the University’s research available electronically.

Copyright and moral rights to this thesis/research project are retained by the author and/or other copyright owners. The work is supplied on the understanding that any use for commercial gain is strictly forbidden. A copy may be downloaded for personal, non-commercial, research or study without prior permission and without charge. Any use of the thesis/research project for private study or research must be properly acknowledged withreference to the work’s full bibliographic details.

This thesis/research project may not be reproduced in any format or medium, or extensive quotations taken from it, or its content changed in any way, without first obtaining permissionin writing from the copyright holder(s).

If you believe that any material held in the repository infringes copyright law, please contact the Repository Team at Middlesex University via the following email address:

[email protected]

The item will be removed from the repository while any claim is being investigated.

Windowed DMD as a MicrotextureDescriptor for Finger Vein

Counter-spoofing in Biometrics

Santosh Tirunagari, Norman Poh, Miroslaw Bober & David Windridge

University of Surrey, Guildford, Surrey, United Kingdom GU2 7XH

Finger Vein Biometrics

• Authentication system that matches thevascular patterns in an individual's finger.

• Blood vessel patterns are unique to eachindividual, as are other biometric data such asfingerprints or the patterns of the iris.

2

How it works

3

Importance

4Source: http://slate.me/1Bmmay5

Finger vein Spoofing - Background

5

Print attacks

Source: http://bit.ly/1QKJT4c

How to counter spoof ?

• Look for the cues and artefacts that differentiate valid from the spoof.

• Our Hypothesis:

– Cues that differ light reflection properties.

– micro-level artefacts that differ in quality.

• How to identify these cues and artefacts?

– Thanks to texture based methods.

6

Texture methods

7

Robust to misalignmentMicro texture

LocalGlobal

Requires preciselocalisation

vs

vs

Band pass filtering

FrequencySpatial

Texture methods

8

Texture methods Spatial / Frequency Local / Global

Windowed-Dynamic Mode Decomposition

Spatial local

Discrete Wavelet Transform

Frequency Global

Discrete CosineTransform

Frequency Global

Histogram of Gradients Spatial Global

Filters Spatial Local

Local Binary Patterns Spatial Local

Our proposal

DMD – Facial counterSpoofing

1...N Video frames

from a single video1...(N-1) Dynamic

modes

Dynamic Mode

Decomposition

1st DMD mode

(phase angle = 0)

Mode selection

LBP histogram

features

Spoof/attack

Valid access

SVM - Classifier 9

Santosh Tirunagari, Norman Poh, David Windridge, Aamo Iorliam, Nik Suki, and Anthony TS Ho. Detection of face spoofing using visual dynamics. Information Forensics and Security, IEEE Transactions on,10(4):762–777, 2015.

10

How DMD works?

.........

I AI A2I An-2I An-1I

A A2 An-1

Solve for eigenvalues and vectors of AFind the unknown matrix A

Generally using Arnoldi approximations.

How about images ?

• Our Proposal – Windowed DMD

• Research questions:

– If DMD can capture principle movements videos then would W-DMD capture texture gradients from images?

– What would be the effect of texture gradients on classification performance ?

– How effective is the W-DMD compared to plethora of existing descriptors ?

11

Our proposal – Windowed DMD

12

1 2 3 4 5 6 . . . n

1 2 3 4 5 6 . . . n

1 2 3 4 5 6 . . . n

1 2 3 4 5 6 . . . n

1_c1 1_c2

2_c1 2_c2

3_c1 3_c2

4_c1 4_c2

DMD

DMD

DMD

DMD

1_c1 2_c1 3_c1 . . (n-2)_c1

1_c2 2_c2 3_c2 . . (n-2)_c2

W-DMD(c1)

W-DMD(c2)

W-DMD on full finger vein images

13

W-DMD (C2)

W-DMD (C1)

Real Spoof

W-DMD on cropped finger vein images

Real Spoof

14

W-DMD (C1)

W-DMD (C2)

Classification framework

15

Texture Methods Cropped FullLBP 1x531 1x531

DWT 1x36 1x70DCT 1x400 1x400HoG 1x81 1x81

Entropy 1x138 1x270STD 1x138 1x270

Range 1x138 1x270W-DMD 1x3330 1x6550

W-DMD+LBP 1x531 1x531

Minimum Intersection Kernel

Texture feature dimensions

Dataset

• IDIAP’s Fingervein Spoofing Dataset

16

Protocol Training set Development set

Test set

full 120 120 200

cropped 120 120 200

Evaluation

17

• Equal Error rate based on F-ratio

• Larger F-ratio => higher separability.

• Measured even when no error is observed.

• F-ratio = [ µC − µI / σC + σI] • Where C is real and I is spoof and µ is mean and σ is

standard deviation.

Norman Poh and Samy Bengio. How do correlation and variance of base-experts affect fusion in biometric authentication tasks? Signal Processing, IEEE Transactions on, 53(11):4384–4396, 2005.

Experimental Hypotheses

• Which DMD components?

• Comparisons with other methods?

18

Exp 1: Selection of the W-DMD component

19

1,5

9 2,1

4

0,0

8

3,1

5

7,8

1

1,4

1

1,8 2,0

9

EER (F)% F-RATIO EER (F)% F-RATIO

CROPPED FULL

W-DMD (C1) W-DMD (C2)

Exp 2. Results - EER(F)(%)

20

15,26

15,11

5,72

5,77

5,34

1,74

3,04

1,89

2,98

1,59

8,7

3,7

1,27

1,04

0,61

0,46

0,22

0,16

0,1

0,08

0 2 4 6 8 10 12 14 16 18

Entropy

DCT

Range

STD

RAW

HoG

LBP

DWT

W-DMD+LBP

W-DMD

Full Cropped

F-ratio

21

1,02

1,03

1,57

1,57

1,61

2,11

1,87

2,07

1,88

2,14

1,35

1,78

2,23

2,31

2,5

2,6

2,84

2,95

3,09

3,15

0 0,5 1 1,5 2 2,5 3 3,5

Entropy

DCT

Range

STD

RAW

HoG

LBP

DWT

W-DMD+LBP

W-DMD

Full Cropped

Conclusions

• Limitations – Size of the feature vector.

• Applied W-DMD on finger vein images for valid and print attacks from 110 clients (240 (training) + 240 (development) + 400(testing)).

• Significance of the W-DMD + SVM pipleline -effectively detect the spoof samples.

• The results were promising in tackling the print attack challenge.

22

Acknowledgements: Department of Computer Science & Centre for Vision Speech and Signal Processing. University of Surrey, Guildford, UK.

23

Thank you

Any Questions ?

Paper ID: 63{s.tirunagari, n.poh}@surrey.ac.uk