Integrating Information

Integrating Information

Dr. Pushkin Kachroo

Integration

Matcher 1

Matcher 2

Integration Decision

Match

No Match

B1

B2

Expanding a Biometric

Multiple Matchers

Multiple Biometrics

Multiple Fingers

Multiple Samples

Multiple Sensors

One Finger

Multiple Tokens

Coupling

Sensor 1 Sensor 2

Process 1 Process 2

Integration

Match Decision

Sensor 1 Sensor 2

Process 1 Process 2

Integration

Match Decision

Tightly Coupled Loosely Coupled

Boolean Combinations

Biometric aBiometric b

Accept/RejectAND

Accept/RejectBiometric aBiometric b

babORa FRRFRRFRR }{

bababORa FARFARFARFARFAR

OR

babANDa FARFARFAR }{

bababANDa FRRFRRFRRFRRFRR

Boolean: Convenience/Security

Biometric aBiometric b

Accept/RejectAND

Accept/RejectBiometric aBiometric b

baba FRRFRRFRRFRR

OR

baba FARFARFARFAR

FAR FRR

OR

AND

ba FARFAR

ba FRRFRR baFARFAR

baFRRFRR

Improve Convenience:Lower FRR (OR)

Improve Security:Lower FAR (AND)

Filtering-Binning

Penetration Rate: Ppr: The fraction of database being matched on average

Binning Error Rate: Pbe

• Filtering using non-biometric, e.g. using last name. (P,B)

• Binning using biometric, e.g. some whorl pattern (B,B’)

TradeoffTradeoff

Filtering Error-Negative Identification

• Adding Pn for subject dn to negative identification prescribes narrowing down on a smaller set of biometric template =>

Since we are comparing over a smaller set, the chance of false positives goes down. However, false negatives goes up because you might say the person is not in the database (looking at the smaller set) when the person might be in the full database.

Filtering Error-Positive Identification

• The probability that a person is who she/he says she/he is equals the probability of a match between stored biometric template and a newly acquired biometric sample. This match probability does not change if additional knowledge or possession is supplied.

Dynamic Authentication

• Example: Conversational biometric….allows for natural filtering by asking knowledge information during conversation; could include possession; while speaker recognition is taking place.

Boolean: Score Level Integration

1T as

bs

2T

ANDOR

OR

Normal Distribution

1T as

bs

2T

Accept

Reject

TssssG baba )1(),(

Tssss

ssGba

bbaa

b

b

a

aba

22

22

22),(

Normal Distribution: Problems

• Covariance Matrix is assumed to be diagonal; okay for disparate biometrics but not for similar ones e.g. two fingers.

• Gaussian gives non-zero probability to negative scores.

Distance based

),(1),( mm BBsBBDist

2

2

1exp

2

1)(

m

mEdd

Prob

)),((

)),((

mm

mm

BBDist

BBDistEE

B and Bm are templates from the same biometric

)1)(|()|(

)|()|(

mmmm

mmm PcohortBPperB

PperBBper

ProbProb

ProbProb

per means person

Degenerate Cases

Ts

ssGa

aba

2),(

ba

Ts

ssGb

bba

2),(

ba

as

bs

ROC based Methods

as

bs

Match Mismatch

),(),( baba ssFRRssF ),(1),( baba ssFARssG

Compare to…

)(spn

)(spm

TFNM FM