How to Control Acceptance Threshold for Biometric Signatures with Different Confidence Values?

1

HOW TO CONTROL ACCEPTANCE THRESHOLD FOR BIOMETRIC SIGNATURES WITH DIFFERENT CONFIDENCE VALUES?Yasushi Makihara( 槇原 靖 ) , Md. Altab Hossain, Yasushi Yagi( 八木 康史 )

大阪大学ICPR 2010

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwan

2

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

n

INTRODUCTION

Biometrics-based verification Quality measure

False Acceptance Rate(FAR) False Rejection Rate(FRR) Receiver Operating Characteristics (ROC) curve

3

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

Receiver Operating Characteristics (ROC) curve

4

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

ROC curve

5

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

ROC curve

6

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

Simplified example High confidence (right side) Low confidence(left side)

7

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

Simplified example High confidence (right side) Low confidence(left side)

8

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

FAR FRR

Error rate Acceptance rate

9

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

Gradient

Lower error gradient accepted samples are positive samples

Higher error gradient accepted samples are negative samples

Middle error gradient positive and negative samples in the accepted

samples are balanced

10

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

Implementation

(distance, quality measure)

Weight (ith positive sample for kth quality measure control point)

11

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

nADAPTIVE ACCEPTANCE THRESHOLD CONTROL

Implementation Gaussian kernel-based non-parametric PDF

estimation

Optimal approximation coef. of regularization term

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

n

EXPERIMENTS

Test data

12

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

n

EXPERIMENTS

Simulation data

13

14

VC

Lab, D

ept. o

f Com

pute

r Scie

nce

, NTH

U, Ta

iwa

n

CONCLUSION & DISCUSSION

Outperforms the previous methods in terms of the ROC curve, particularly under a lower FAR or FRR tolerance condition

With the assumption that distributions of distance and quality measures are consistent in the training and test sets, the optimality is not guaranteed in case where the distributions are in consistent.