fingerprints: recognizing 75 billion patternsbiometrics.cse.msu.edu/presentations/aniljain... ·...
TRANSCRIPT
Anil Jain
Michigan State University
October 23, 2018
http://biometrics.cse.msu.edu/
Fingerprints:
75 Billion-Class Recognition Problem
Cumins and Midlo, Finger Prints, Palms and Soles, Dover, 1961
Friction Ridge Patterns
2
Dermatoglyphics. Derma: skin; Glyphs: carving
3
Fingerprints
Each finger, including those of identical twins, has a different pattern
0.73 x 0.84 in.
75-Class Recognition Problem
• 7.5 billion x 10 = ~75 billion pattern classes
• 350 K births/day; no. of classes keeps increasing
4
First Encounter with Fingerprints
Ratha, Rover, Jain, "An FPGA-Based Point pattern Matching Processor with application to Fingerprint Matching", CAMP `95,
Sun SPARCstation host
~100 MHz CPU; 512 MB RAM
• Fingerprint matching on Sun SPARC: 70 matches/sec
• 100-times speedup on Splash 2 FPGA @1 MHz clock
16 Xilinx 4010s as PEs (512 KB memory)
5
Fingerprint Formation• Ridge formation starts at 1 or 2 focal points and spreads
over the fingertip
• Localized ridge units merge to form ridges at ~10.5
weeks of gestational age
• Fingerprints possess genotype & phenotype properties
L. S. Penrose and P. T. Ohara. The development of the epidermal ridges. Journal of Medical Genetics. 1973
M. Okajima. Development of dermal ridges in the fetus. Journal of Medical Genetics. 1975 6
Fingerprint Milestones
300
B.C.
1839 1869 1883 1900 1905 1924 1972 19991963 2001 2003 2008 2013 2014 20171858 2018
Fingerprint as a personal mark
A Chinese deed of sale with a fingerprint
Early use of fingerprints for Civil applications
(Bengal, India)
Habitual Criminals Act“What is wanted is a means ofclassifying the records ofhabitual criminal, such that assoon as the particulars of thepersonality of any prisoner(whether description,measurements, marks, orphotographs) are received, itmay be possible to ascertainreadily, and with certainty,whether his case is in theregister, and if so, who he is”
First use of fingerprints in
British criminal case
Bertillonage invented
Galton / Henry fingerprint system adopted by Scotland Yard
• Seventeen classes
• Whorl (double loop), loop (left and right)
& arch cover 99% of fingerprints
Delta Core
US Congress authorizes DOJ to collect fingerprints and arrest information
Identimat: First commercial use of biometrics
Trauring publishes paper on fingerprint matching in Nature Goldstein et al. publish face recognition paper in Proc. IEEE (1971)
FBI inaugurates full operation
of “IAFIS”
State AFIS
State AFIS
State AFIS
State AFIS IAFIS
Forensics Other operations
Criminal booking
9/11 terrorist attacks lead to govt. mandates to use biometrics in
regulating intl. travel
US-VISIT TouchIDFaceID
Apple Pay
Supreme Court upholds the constitutional
validity of Aadhaar
“Aadhaar gives dignity to the marginalized. Dignity
to the marginalized outweighs privacy,” Justice
Sikri
Aadhaar
FBI Next Generation Identification
7
Drivers: Lack of Trust
•Osmania University (OU) enhanced the
exam fee in all the affiliated colleges by
Rs. 100 per semester for
implementation of biometric attendance
system. Times of India, Oct 22, 2018
• No end to JNTU-H fake certification
verifications. HANS INDIA, Times of India, Oct 12, 2018
8
Enablers: Fingerprint Readers
1892Juan VucetichInk and Paper
1990Optical sensor
1990Capacitive sensor
9
Enablers: Processors, RAM, Algorithms
Courtesy: James Blanchard, Michigan State Police
1960s
1989
• 725K tenprints
• 15K matches/sec
2017• 4M tenprints
• 1M matches/sec
Ridge
Voting
Fingerprint Enhancement
Hong, Wang and Jain, IEEE Trans. PAMI, 1999
Fingerprint Representation
Level-1 Level-2 Level-3 Minutiae Pores and incipient ridges
Ending
Bifurcation
Orientation Field
Singular Points
Deltas
Cores Pores
Incipient Ridge
Ridge flow and pattern type
12
Template: A compact representation of fingerprint features
Minutiae Extraction
Input Image
Ridge Thinning
Minutiae Detection
Ridge Flow Ridge Filter
PostprocessingExtracted Minutiae
13
Minutiae Descriptors
•Ridge Flow-based Descriptor
• Ridge flow values in the minutiae neighborhood
•Neighboring minutiae-based Descriptor
• Set of minutiae in a local neighborhood
Flow-based Descriptor Minutiae-based DescriptorMinutia Neighborhood
14
Enrolled fingerprint
Fingerprint Comparison
Similarity = 0.9Query fingerprint
15
How to Align?
Gallery Fingerprint
Query Fingerprint
16
Jain, et al. An Identity Authentication System Using Fingerprints, Proc. IEEE, 1997
Freq
uen
cy
Similarity score
Recognition Performance
ROC curve
1
1
17
Threshold determines tradeoff between FAR & FRR
System Requirements
100K visitors/day to Disney Park, Orlando18
• Usability
• Fast verification to maintain throughput
• Low error rates, especially FRR
• Day/night operation
• Robust to finger condition: wet, dry,..
• Return on investment
• Embedded system
• Template encryption
Aadhaar:
World’s Largest Biometric System
19
121 crore (1.21 billion) individuals have been enrolled
De-duplication:
Limited Capacity of Fingerprints
Aadhaar Authentication
21
Daily Authentication Transactions
“100% successful authentication NOT possible,” UIDAI CEO admits in SC
https://uidai.gov.in/aadhaar_dashboard/auth_trend.php
State of the Art Performance
• Authentication: TAR of 99.9% @FAR = 0.001%
• Retrieval (search)• Plain prints: 99.3% (100K background gallery)
• Latent prints: 67.2% (70.2% with image + markup)
C. Watson, et al.. Fingerprint Vendor Technology Evaluation, NISTIR, 2012
M. Indovina, et al.. ELFT-EFS Evaluation of Latent Fingerprint Technologies: Extended Feature Sets NISTIR, 2012
Rolled prints Plain prints Latent prints
23
Sources of Error
No. of false minutiae = 27No. of false minutiae = 7No. of false minutiae = 0
24
489 368 6 329 77
29 11 12 21 20
Query
Sources of ErrorGenuine comparisons
Imposter comparisons
Intra-finger variations and Inter-finger similarity
Accuracy
Scale
Usability
Unusable
Hard to Use
Easy to Use
Transparent to User
101
103
105
107
90%
99%
99.99%
99.999%
What’s Next?
Fingerprint Recognition is not solved!
Scalability
• Assume one billion users
• Identification Performance
• False Negative Identification Rate (FNIR): user is
enrolled, but not retrieved
• False Positive Identification Rate (FPIR): user is not
enrolled, but an identity is returned
• Identification & verification performances are related
FPIR = 1 – (1 – FMR)N ; FNIR = FNMR
• Suppose for N = 109 enrollment, we require
FNIR = 0.0001%; FPIR = 0.001%
Require a matcher: FMR ≈ 10-12 %; FNMR = 0.0001%
27
Capacity & Persistence
• Uniqueness: How many different
individuals can we recognize?
• Permanence: Does the recognition
performance change over time?
6-digit code:106 unique PINs
PINs do not become “stale”
Prob. of False Correspondence• "Two like fingerprints would be found only once every
1048 years” (Sc. Am, 1911)
• Prob. of a fingerprint with n minutiae and another with
m minutiae sharing q minutiae
(a) M=52
m=n=q=26
P = 2.40 x 10-30
(b) M=52
m=n=26, q=10
P = 5.49 x 10-4
M = A/C
Pankanti, Prabhakar and Jain, “On the Individuality of Fingeprints”, IEEE PAMI, 2002
29
Persistence
Yoon and Jain, "Longitudinal Study of Fingerprint Recognition", PNAS, 2015
• Database: fingerprints of 20K convicts with an
average of 8 arrests over a span of 12 years
• Longitudinal model showed: Fingerprint accuracy (i)
is stable over 12 years, (ii) accuracy depends more
on fingerprint quality than time gap
30
Spoof Attacks
Chugh, Kai and Jain, "Fingerprint Spoof Buster: Use of Minutiae-centered Patches", IEEE TIFS, 2018
Requirements: TAR = 98% @FAR = 0.2%
Fingerprint Obfuscation
32
Fingerprint of Gus Winkler (1933) before and after alteration
Template Protection
33
Original Fingerprint ImageISO Fingerprint TemplateReconstructed Fingerprint ImageSimilarity Score = 460 (VeriFinger)
Fingermarks (Latent Prints)
34
Madrid Train Bombing (2004)
Partial print on a duffel bag Brandon Mayfield’s prints in file
35
Automated Latent AFIS
Acquisition Cropping Enhancement Minutiae Comparison
Reference database
Feedback
290 71 70 48
Candidate list 36
Kai and Jain, IEEE PAMI, 2018
Successful MatchLatent
EnhancedLatent
Mated Rolled
MatedRolled
# Matched minutiae = 13
Similarity score = 38
# Matched minutiae = 2
Similarity score = 3
Infant Fingerprints
Digital Persona U.are.U (500 ppi) MSU Match in Box (1900 ppi) Custom NEC Reader (1270 ppi)
Jain, Arora, Cao, Best-Rowden, Bhatnagar, Fingerprint recognition of young children, IEEE TIFS, 2017
Engelsma, Cao, Jain, Fingerprint Match in Box. IEEE BTAS, 2018
Engelsma, Cao, Jain, Fingerprint Match in Box. IEEE PAMI, 2018
Right thumb of a 3 month old infant captured with 500,1270 & 1900 ppi readers
Fingerprint Match in Box
(a) (b) (c)
A low cost ($200), open source, 1900 ppi compact (10 cm cube) fingerprint reader with embedded spoof
detector, feature extractor, and matcher with 1:100K search; thumbprint of 3-month-old baby
Privacy Concerns
40
Security v. Privacy
Summary
• Fingerprint based transactions are used by
hundreds of millions of citizens worldwide
• Applications: mobile phone unlock, social
benefits disbursement, border crossing,
forensics; new applications are emerging
• Challenges: sensor design, image quality,
robust & accurate solution, privacy, security
• It’s all about lack of TRUST
Security v. Privacy
44