Transcript Chapter_6
Identification System Errors Guide to Biometrics – Chapter 6 Handbook of Fingerprint Recognition - 1.4 Presented By: Chris Miles Extending to Identification ● How do we extend our numerical models for verification errors for identificatation? ● FNMR – False Non Match Rate ● FMR – False Match Rate ● What Issues are presented Identification System ● Maintains a database of enrolled users ● Tries to match input against the database ● Positive Identification == Negative Identification ● Output – List of best matches – Ideally just the true identity – Best Match – yes/no in the list Example ● Casino using face detection to identify people on the Nevada Gaming Commission's black list – http://gaming.nv.gov/loep_main.htm – Basis for other government biometrics systems ● N = the number of people on the list ● M = number of people through the casino daily ● Calculate FNMRN and FMRN Matching System ● ● Parallel version of your favorite verification algorithm Attempt to match all users against the database FNMRN ● ● ● The chance of being falsely rejected is the same as verification Chance of not matching against your template – chance of matching someone else's template Assuming no FMR, FNMRN = FNMR FMRN ● ● ● FMRN = Chance of matching someones template ^ number of templates FMRN = 1 – (1 – FMR)N Number of daily false matches = M * FMRN = M (1 - (1-FMR)N) Accuracy Scales Worse then Computation ● ● The chance of being falsely accepted rises exponentially with the number of templates Suppose algorithm is 99.99% accurate – 100 people in the database – Each has 8 templates – 10,000 people through the casino a day ● FMRN = 1 - .9999800 = 0.076 ● FMRN * 10000 = 768 False accepts a day Winnowing ● ● ● True identification is exponentially hard, so generally we compromise and just return a list of probable matches. Input -> System -> List of Candidate Matches A second system, biometric or a human supervisor, then tries to identify the user from the new List / Database of candidates ● Candidates -> Second System -> Identity ● “Passing the buck” so to speak Who's on the list? ● ● Threshold – Apply a threshold to the similarity metric – similarity > threshold -> On the list Rank – ● Take the K most similar templates Hybrid – Take the K most similar templates so long as there similarity > threshold Weaknesses ● Threshold – ● ● If several users kind of match the input, but not quite, a threshold based system would return nothing Rank – Impostor -> List of bad matches – Solution: Generic Impostor Model -> Additional Template representing a non-match situation, if a user matches this -> returns nothing. Hybrid – Strengths of both techniques cover the weaknesses Hybridization Ideas ● ● Adjust K based upon how many are above the threshold Adjust the threshold based upon the distribution of similarities Multiple Templates ● ● Example had multiple templates per individual Input might match mutiple templates from one person ● Only one might need to be in the list ● Domain Dependent Characterizing Identification ● FNMR and FMR ~= Reliability and Selectivity ● Reliability ● – 1 - FRR – How often we correctly identify someone who is in the database Selectivity – K – Rel or – (m-1) FAR – Number of incorrect matches returned RSC, ROC, RPC Curves ● These curves show the compromises involved ● ROC Compromises between FAR and FRR rate – ● ● Should the vending machine take my ripped dollar and someone elses forgery? RPC Curves – If google returned more results it would be less likely to miss relavant ones – Would include more irrelevant results however RSC Curves Three systems ● Theshold Based – Previous Example ● Rank-Based identification ● Rank-order statistics ● Rank Probability Mass Function Threshold System Errors ● ● Similar to previous example only returns a list of individuals above the threshold Errors – FARM = m * FAR * (1-FAR)m-1 - Falsely Match one individual – Ambiguous answer -> List has length > 1 – P(Ambiguous) = 1 – [1 – (m+1) * FAR](1 - FAR)m-1 – FRRM = 1 - (1 – FRR) * (1 – FAR) m-1 ~= FRR Rank Based System Errors ● ● ● Only works in very restricted close world scenarios (No Impostors) Only one error – Misidentification by the correct user being ranked below another Analyze probabilistic distribution of ranks – Rank Probability Mass Function