Lions, Swans, Monkeys – A New Approach to Biometric Menagerie, Stanislav Sartasov, Lanit-Tercom / SPbSU, CEE-SECR 2017

Biometric system users’ classification based on their recognition quality is an important issue when developing and exploiting such systems. Existing approaches based on a biometric menagerie concept are described and their limitations are shown. A new classification using biometric system errors is proposed and its feasibility is demonstrated.

Transcript

1. October 2017, St. Petersburg Software Engineering Conference Russia October 2017, St. Petersburg Software Engineering Conference Russia Lions, Swans, Monkeys – A New Approach to Biometric Menagerie Stanislav Sartasov, SPSU
2. What is biometric system?
3. What is biometric system? Biometric DB Matcher Feature extractor Scanner Protocol Access Denied Access granted
4. Genuine and impostor scores G-score I-score
5. Genuine and impostor scores I G Number of occurences Similarity score 0 1
6. Genuine and impostor scores I G Probability density distribution Similarity score 0 1
7. Genuine and impostor scores I G Probability density distribution Similarity score 0 1
8. Biometric system performance – classic approach • Single threshold • All available values • Users are equiprobable
9. Biometric system performance – classic approach I G Probability density distribution Similarity score TEER FRR FAR
10. Receiver Operating Characteristic
11. Biometric system performance – classic approach I1 I2 G1 G2 Probability density distribution Similarity score T1 T2
12. Biometric menagerie - related works • G. Doddington et al., 1998 • Users with high I-scores or low G-scores exist I Lamb Goat G Probability density distribution Similarity score
13. Biometric menagerie - related works N. Yager, T. Dunstone, 2007, 2010 Image from N. Yager and T. Dunstone. 2010. The Biometric Menagerie. IEEE Transactions on Pattern Analysis and Machine Intelligence 32, 2 (2010), 220–230. DOI: http://dx.doi.org/10.1109/tpami.2008.291
14. Biometric menagerie - related works • H. Zheng at al. 2015 • A database with EER=0% • Both Doddington and Yager-Dunstone menageries found • High I-scores and low G-score do not necessarily result in errors
15. Proposed approach I G Similarity score Probability density distribution M(I) M(G) Lowest G Highest I
16. Proposed approach = ≥ → − ≤ → − < < → − = tg 2 5 = tg 3 10 = sgn max( ) − min( ) < 0 → − ℎ ≥ 0 → −
17. Results
18. Results
19. How to use it? Biometric DB Menagerie classifier Feature extractor Scanner
20. Conclusions • It works  – Users can be reliably separated into bins – Generic framework • needs to be adjusted – WIP • Stanislav.Sartasov@gmail.com