On the consistency of the biometric menagerie for irises and iris matchers

Abstract

The biometric menagerie is useful in identifying the troublesome users within a biometric recognition system. In order to maximize the benefits of the menagerie classifications, it is imperative that the classifications remain constant for each subject. Irises present one of the unique scenarios for classification since each iris represents the same subject but the two irises are independent of each other. We have taken the ICE 2005 iris image dataset [8] and applied three different iris recognition algorithms to it. For each algorithm, we classified the subjects within the biometric menagerie and studied the consistency of the classifications across algorithms. We also broke the dataset into subsets by left and right iris and studied the consistency of the classifications between irises. Our results have shown that the biometric menagerie classifications are algorithm dependent and dependent on which iris is chosen. One-third of the population was classified as a weak user by only a single algorithm and a quarter of the population had irises with non-matching classifications, one of which was a weak user classification. These two subsets of the population represent all the potentially weak users in the population but the subjects cannot be considered weak due to the disagreement between the algorithms and the mismatched classifications of the two irises.In order to use the biometric menagerie effectively, one algorithm must always be used for all recognitions and modalities must be kept in disjoint datasets to reliably label weak users.

The Biometric Menagerie – A Fuzzy and Inconsistent Concept

“…Table 3 presents the results obtained in these six tests performed to find the marginal wolf-templates. As seen in Table 3, each test points out to a different marginal wolftemplate (which is an experimental result that agrees to those presented in [4] for the wolves detected in ICE database [3]).…”

“…After [7] and [4], this is the third paper that analyses the partitioning of the iris code space extracted for a certain database (University of Bath Iris Image Database, UBIID, [10] in our case) as a Fuzzy Biometric Menagerie showing that the extensions of the concepts wolf, lambs, sheep and goats have very unsharp and unstable (non stationary) boundaries. Moreover, the membership of a user to these categories can be more often expressed as a degree (as a fuzzy value) rather than as a crisp value.…”

Iris recognition with 4 or 5 fuzzy sets

“…Therefore, the assertion set AS is the sub-part of CP formed with data provided by correctly enrolled u sers. Hence, the formu lae (13) and (14) should be cons idered on the assertion set only.…”

“…Actually, the fact is both BM and TA ( [2], [6]- [9], [13], [14], [32], [33]) are two annoying symptoms illustrating intrinsic limitations of the statistical model of iris recognition [4], [5]. When it happens, the increase in the recognition error rates does not need new names like BM o r TA nor naive justifications (some users are somehow specialin case of BM; or biometric templates are "ageing"in case of TA).…”