da/sec scientific talk on Biometrics
Topic: Bayesian Decision Theory:
Evaluation of binary Decision Systems
by Andreas Nautsch
D19/2.03a, April 12, 2018 (Thursday), 12.00 noon
Keywords — Visual Performance, Risk Minimization, Score Calibration, Biometric Verification
Performance estimation is crucial to the assessment of novel algorithms and systems. The prototype domain is biometric verification as a binary decision: accepting or rejecting a biometric identity claim, i.e. granting or denying access. Depending on the application e.g., online banking transaction or theme park access, the operating point of an application is dependent on security and convenience requirements. Therefore, the Bayesian deciosion framework is a formalism, denoting thresholds in terms of prior probabilities, e.g. of facing an attack, and costs, e.g. of rejecting a genuine identity claim. Conventional performance assessment in biometrics addresses the competing Type I and Type II error rates by plotting either error rates against another, which is beneficial as different systems output different score scales. The Bayesian decision framework enables the unification of score scales by transforming system outputs into a well-defined probabilistic interpretation, in which scores represent so-called „likelihood ratios“: the relative support of an „accept“ decision over a „reject“ decision given the evidence.
For the purpose of increasing evaluation transparency, these likelihood ratio scores are summarized by so-called verbal scales, which are used in forensics in order to provide a quantification to laymans by a forensic examiner. Generalizing the scope to binary decision making, verbal score scales are reflected for decision making, i.e. scores lead to an „accept“ decision, when the support of a likelihood ratio exceeds the prior and cost demands. Conventional error rate plots are enhanced by interrelating digestible likelihood ratios as verbal bands are color encoded in error tradeoff diagrams, such as receiver operating characteristic (ROC) or detection error tradeoff (DET) plots. Further implications towards visual performance and formalized decision requirements will be addressed in the presentation.