MINEX 04 - NIST Visual Image Processing Group
MINEX 04 - NIST Visual Image Processing Group
MINEX 04 - NIST Visual Image Processing Group
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Minutiae InteroperabilityExchange Test(<strong>MINEX</strong> <strong>04</strong>)C. L. Wilson<strong>NIST</strong>
Exchange of Fingerprint Data<strong>Image</strong>s or TemplatesFactors To Consider:•Accuracy•<strong>Processing</strong> time•Size on card
Minutiae Example
Fingerprint Matching1. Enrollment2. Subsequentattempts atidentityverification3. Matching = ? MatchNon-Match
Types of Templates• Proprietary templates– Individual vendor’s representation of images• Standard templates: INCITS 378 format– MIN:A templates• codes minutiae coordinates (x, y), angle (θ), type,& quality– MIN:B templates• MIN:A data plus ridge count, core, and deltainformation
Largest Biometric Test to Date…• 4 datasets:– POEBVA, DHS2, POE, and DOS• Number of Samples– >60,000 matched fingerprint pairs– >120,000 non-match fingerprints• 14 vendors– Six participants in MIN:B testing• 4.4 billion comparisons resulting in >45GB of scores & >1 terabyte of data in total
VendorsA. Cogent SystemsIncorporatedB. Dermalog IdentificationSystems GMBHC. Bioscrypt IncorporatedD. Sagem MorphoIncorporatedE. NeurotechnologijaF. InnovatricsG. NEC CorporationH. TechnoimagiaCorporationI. Identix IncorporatedJ. Biologica SistemasK. SPEX ForensicsL. Secugen CorporationM. NITGen CorporationN. Cross MatchTechnologies
<strong>MINEX</strong> Questions• Do standard templates give accuracycomparable with proprietary (imagebased)implementations?• Can template data be generated andmatched by different vendors withoutincrease in error rates?
Detection Error Tradeoff Curves
Types of Tests• Single- v. Two- Finger• Proprietary v. Native• Native v. Interoperable (Scenario 1)• Scenarios 2,3,&4• Four datasets of different quality
Results 1• The best proprietary templates aresuperior to the MIN:A & MIN:B templatesin accuracy.• The enhanced MIN:B template performedsimilarly to the basic MIN:A template.
Proprietary and Native Performance for Single-Finger Test at FMR=0.01 for POEBVA data0A B C D E F G H I J K L M NVendorFNMR0.020.160.140.120.10.080.060.<strong>04</strong>0.02FNMRProprietaryMIN:AMIN:B0.0180.0160.0140.0120.010.0080.0060.0<strong>04</strong>0.0020A B C D E F G H I J K L M NVendorProprietaryMIN:AMIN:B
Proprietary and Native Performance for Two-Finger Test at FMR=0.01 for POEBVA data0.080.060.<strong>04</strong>0.020ProprietaryA B C D E F G H I J K LMN MIN:AMIN:BVendorFNMR0.020.018Proprietary0.016MIN:A0.014MIN:B0.0120.010.0080.0060.0<strong>04</strong>0.0020A B C D E F G H I J K L M NVendor
Results 2• As with most recent tests (by <strong>NIST</strong> andothers), the error rates between matchingalgorithms vary by at least an order ofmagnitude.• Two-finger authentication with standardtemplates can achieve the accuracy ofsingle-finger authentication withproprietary templates.
Interoperability Example 110.90.80.70.60.50.40.30.20.10A B C D E F G H I J K L M N D (#1)E (#8)J (#14)Enrollment Template Generator0.020.0180.0160.014FNMRs @ FMR=0.01 for…• Scenario 1, Two-Finger, POEBVA data• Vendors of Rank 1, 8, & 14 forauthentication templategeneration/matchingv.All enrollment template generatorsD (#1)E (#8)J (#14)FNMR0.0120.010.0080.0060.0<strong>04</strong>0.0020A B C D E F G H I J K L M NEnrollment Template Generator
Interoperability Example 210.9E (#1)FNMRs @ FMR=0.01 for…FNMR0.80.70.60.50.40.30.20.10AA (#8)H (#14)BCD EF GHI J KLMNAuthentication TemplateGenerators/Matchers• Scenario 1, Two-Finger, POEBVAdata• Vendors of Rank 1, 8, & 14 forenrollment template generationv.All authentication templategenerators/matchersFNMR0.020.0180.0160.0140.0120.010.0080.006E (#1)A (#8)H (#14)0.0<strong>04</strong>0.0020A B C D E F G H I J K L M NAuthentication Template Generators/Matchers
Results 3• The leading vendors in templategeneration are not always the leaders inmatching and vice-versa.– Some template generators produce standardtemplates that are matched more accuratelythan others. Some combination of templatesfail completely.– Some matchers compare templates moreaccurately than others.
Finding the Largest Interoperable <strong>Group</strong>FNMRs at FMR=0.01for Scenario 1, Two-Finger, POEBVA data
Finding the Largest Interoperable <strong>Group</strong>FNMRs at FMR=0.01for Scenario 1, Two-Finger, POEBVA data
Finding the Largest Interoperable <strong>Group</strong>CEB 0.0072 0.0073F 0.0032 0.0103N 0.0086 0.0094FNMRs at FMR=0.01for Scenario 1, Two-Finger, POEBVA data
Largest <strong>Group</strong> for 2-Finger 2POEBVA s.t. . themax FNMR≤0.01 @ FMR=0.01
Interoperable Template Generatorsand MatchersDatasetCriterion(s.t. FNMR≤0.01@ FMR=0.01)Value ofCriterionTemp.Gen’s# Temp.Matchers#POEBVA group max 0.0087A,B,C,D,E,F,G,NPOEBVA group mean 0.0094 A,B,C,D,E,F,G,K,L,M,N8 A,B,C,D,F,G 611 A,B,C,D,E,F,G,I,L,M,NDHS2 group max 0.0081 B,D,N 3 A,D 211DHS2 group mean 0.0095 B,D,KN4 A,C,D,F,G 5
Results 4• Certification of an interoperable group ofproducts requires some priorspecification of the required accuracy.– More products will interoperate when theaccuracy requirement is low and vice versa.– More products can be certified if the group’smean error rate is below a threshold than iftheir worst interoperable pair is used forcertification.
Results 5• Performance is sensitive to the quality ofthe dataset.– Applies to both proprietary and interoperabletemplates.– Two higher quality datasets (POEBVA andPOE) provide reasonable interoperability.Two lower quality datasets (DOS and DHS2)do not.
For more information• See the report onlinehttp://fingerprint.nist.gov/minex<strong>04</strong>/• ContactCharles Wilsoncwilson@nist.gov(301) 975-2080
Proprietary and NativePerformanceFNMR0.010.0090.0080.0070.0060.0050.0<strong>04</strong>0.0030.0020.001Proprietary (1)MIN:A (1)MIN:B (1)Proprietary (2)MIN:A (2)MIN:B (2)0A B C D E F G H I J K L M NVendor