Download - Large Scale Applications
Large Scale Applications
Dr. Pushkin Kachroo
Large Scale Applications
• To reduce # of matches needed– Filtering: using non-bio like sex, age
etc.– Binning: using additional biometric
Size
• Small: 1000 (approx. stored samples)
• Medium: 10K• Large: 100K• Very Large: 1M• Extremely Large: 10M
Examples
Voter Registration
Enrollment Authentication
National ID
EnrollmentAuthentication
FingerprintFAR(1) FRR
1 Finger 10-5 10-3
2 Fingers 10-10 10-3
Enrollment False Positives
• Quadratic ~ [m x FAR(1)] x m• For one finger…10^8 gives 10^11• Two finger it gives 10^6
Enrollment False Negatives
• FNR ~ 10-3
Authentication Accuracy
• False Rejects: FRR x m = 105 for 108 size
• False Accepts: 10-5
Matcher Sizing
• m(t): Database size• J(t): Match throughput; #matches
needed/time• E(t): Enrollment Rate• A(t): Authentication Rate
m(t)
Matcher Throughput
tEdEtmt
avg 0
)()(
tEtmtEtJ avge2)()()(
avga AtJ )(
Matcher Throughput-2
m(t)
0
02
)()()(ttmEA
tttEAtJtJtJ
avgavg
avgavgea
)(tJ
t0
National ID
st
)(tJ
t0
Voter Registration/Election
Total # of Matches
mtJdtJdttJ aavg
t
t
avg
t
t
aa
00
)(
Verification on Election Day
Enrollment
)(22
)()( 2
22
0
2
0
mOmtE
tdtEdttJ eavgt
t
avg
t
t
ee
Cost
• Incremental Cost : Sampling, printing ID, etc.; Linear
• Exception Handling : Cost of Resolving False Positives; quadratic
m(t)
mihmh avg )(
Total Cost
m(t)
32
2
)1()1(
)()1(
)(])1([1)(
miFARmFARhm
mmihFARm
mhmFARmmmCost
avg
avg
Even for i = 0 we get:
12
1
)()1()1(
)()1()(
mhFARifmFARh
mhFARifmmCost
avgavg
avg
Cost-Plot
m(t)
)1(
1ˆ
FARhm
avg
Cost(m)
m