Presentation on theme: "Biometrics II Performance Measure Example Face Verification"— Presentation transcript:
1 Biometrics II Performance Measure Example Face Verification Gabor FunctionsIris VerificationHandwriting VerificationGenetic FingerprintBiometric StandardsFuture of Biometrics
2 EER - ExampleTesting a biometric system for various thresholds: Each time, 100 tests were performed, by comparing two patterns. In 40 cases, the two patterns came from the same person, in 60 cases from different persons.The table shows the result.TA: True AcceptanceTR: True RejectionFA: False AcceptanceFR: False Rejection
3 Determine Equal-Error-Rate Best threshold is about 37%, EER 29%.
4 Weighted FAR/FRRhowever, now we think that a false rejection is more worse than a false acceptancereason: it is less likely to have an intruderwe can give weights to the ratesexample: a false rejection we weight with 0.8, a false acceptance with 0.2to get the new equal-error-rate, we multiply both curves with the weights
5 Weighted EER Now, best threshold is about 18%, EER 10%. times 0.8
7 ROC: Model CasesROC is a model for the statistical distribution of correct and wrong answers of a system, depending on a parameter (here: threshold)in worst case, both distributions overlap completelyin typical case, there is some overlapin best case, there is no overlapwe monitor, how TAR and FAR grow together, when the parameter goes from 0 to a maximum valueboth start at (0%,0%) and end at (100%,100%)in worst case, both rates are always the samein best case, first TAR goes from 0 to 100, then, while TAR is 100, FAR goes from 0 to 100 (all in %)
8 Receiver Operator Characteristic best caseworst caseour example
9 From finger to facesFor fingerprint verification, the so-called minutiae played a majorrole. Is there something similar for faces?
11 Face geometryThe minutiae in faces are so-called fiducial points (like position of eyes, nose, mouth). Complex image processing algorithms are able to find these points in face images with some success, thus giving the means for a biometric template.
12 Face geometry: Problems Each detection is prone to failures.Strong requirements on the conditions during capture of the face image (illumination, background, head pose, face expression, obstacles like glasses, coverage by hair...).Detections may depend on each other, and the processing time might be high.The definition of such a position is not precise (which point is the nose?).
13 Another one... Eigenfaces Eigen-faces: an image is considered to be a point in a (very!) high-dimensional vector-space, with each pixel position giving an independent direction, and the intensity at this point the coordinate.Then it is assumed that the set of all face images is a linear subspace of the vector-space of all images, and that there is a corresponding projection operator. (btw later this was shown to be wrong)This projection is a linear operator – thus, it has eigen-vectors. The Eigen-vectors are derived for a set of training images, and any new face image is mapped into this sub-space. This gives some components, from which the correspondence of the new face image to the known ones can be calculated.
14 EigenfacesA typical set of basis vectors in Eigenface-space, and a decompositionof a new face by such components.
15 Eigenfaces: ProblemsBetter suited to identification than verification.Very high demands on the face image capture, including a correct face aspect.Performance rather poor, and can be hardly improved – but there are many works, Fisher faces asf.Such a face subspace does not really exist.Not much specificity for faces – any image class (cars or so) can be used as well.Last not least: there was some political pressure to make this method “famous”...
16 Face mesh: 3rd approachFor the template, a mesh is adapted to the surface of the face.Later on, the “effort” is measured to make this mesh matching with the newly presented face.This method is also suitable for 2D images.Problems are obvious, esp. the imprecision in the placement, and the high computational effort.
17 Iris: Part of a face!In 2002, a journalist of Geo journal tried to find out about the fate of a girl from Afghanistan, shown on the cover page of the June 1985 issue. He succeeded and could prove the identity of Sharbat Gula.Left: Cover 1985, Sharbat Gula at the age of 17. Right: Herself, 2002.
18 The persistence of iris pattern. For this proof, Daugman's iris verification algorithm was used.
19 Iris recognitionDerivation of a specific bit-pattern from the iris image.
21 Iris Pattern: Using Gabor wavelets A 2D-Gabor function is the product of a sine wave and a Gaussian.
22 Gabor filterTaking various frequencies and orientations, a set of such Gabor functions can be used as a base.The iris pattern is decomposed into such a base, giving a binary pattern. This pattern can be used as a biometric template.
23 Iris CodeDecomposition by Gabor wavelets.Iris code.
24 Handwriting?Handwriting examination is still posing a big challenge to computer science. In this case, not just the values of features of a handwriting are person-specific, but also the existing features itself.
25 Dynamic signature verification However, the dynamic pattern of a handwriting (esp. the signature) is suitable for a stable biometric feature that can be used as a biometric template.Such a pattern contains the time functions of pen position, pressure and tilt. A classical method like Dynamical Time Warping (aka Dynamic Programming) can be used for matching.
26 DNA Fingerprint IPCR method developed by Kary Mullis in 1983 gave a great stimulus to the field of DNA analysisNot just for forensic cases, DNA fingerprint is also used for parental tests, detection of inherited diseases, cloning, e.g. of fossile DNA, and derivation of gender (young chicken etc.)
27 Structure of DNADNA:CGATTTGA...ATTG ATGATG...ATGATG.... GCTTAGACCT“Junk-DNA” - also containssome repeating sequences.The number of repetitions isperson-specific (can changewith age).Genetic code, only 1-2%of whole DNA.
28 DNA Fingerprint roughly: primerprimerDNA:CGATTTGA...ATTG ATGATG...ATGATG.... GCTTAGACCTPolymerase Chain Reaction (PCR): simulates the doublingof DNA strands. Primer cut off a part of the DNA, and DNApolymerase repeatedly force a doubling of the DNA. Once thereis sufficient material, its molecular weight is derived with theAgarose gel electrophoresis method.
29 Genetic vs. real fingerprint The genetic code is composed of 8 numbers only – but it is not a human id. It can only be used for comparison.Relatives have similar genetic codes, the GC of twins is identical.Estimation is fast today, but still some effort and costs.Proper treatment of the sample is very important. There is an estimated invalidity of about 1% of all tests due to wrong treatment.In smaller community, probability of the same pattern may go down to 1:30000.
31 Why Biometric Standards? needed for application developmentintegration of biometrics in other productsvendor-independent solutionscomparability among different providers of biometric solutionsglobal use of biometric security techniques (e.g. passport control at the airport)storing data in biometric databases in standard formatreproducing the origin of a biometric templateprotection against forged data
32 Who makes standards? proposal of groups from industry NIST: National Institute of Standards and TechnologyANSI: American National Standards InstituteISO: International Standard OrganizationOASIS: Organization for the Advancement of Structured Information Standardsalso: IEEE, ICAO (Intl. Civil Aviation Organization)national: DIN (Deutsche Industrienorm)
33 CBEFF: Common Biometrics Exchange Formats Framework NISTIR 6529-A, 2004(Standard Biometric Header)SBHSBH Security OptionsIntegrity OptionsCBEFF Header VersionPatron Header VersionBiometric TypeBiometric SubtypeBiometric Data TypeBiometric PurposeBiometric Data QualityBiom. Creation DateValidity PeriodCreatorIndex (e.g. in database)...(Biometric Data Block)BDBbiometric templateif needed, data can be structuredoptional parametersoptional encryptedcan follow a different sub-standard(Signature Block)SBdepends on SBH Security OptionsMAC: Message Authentication Code, or:Digital Signature
34 Example SBH:Biometric Type No Information Given ‘000000’Multiple Biometrics Used '000001'Facial Features '000002'Voice '000004'Fingerprint '000008'Iris '000010''Retina '000020'Hand Geometry '000040'Signature Dynamics '000080'Keystroke Dynamics '000100'Lip Movement '000200'Thermal Face Image '000400'Thermal Hand Image '000800'Gait '001000'Body Odor '002000'DNA '004000'Ear Shape '008000'Finger Geometry '010000'Palm Print '020000'Vein Pattern '040000'Foot Print ‘080000’
36 BioAPIjoint standard activity of ISO and the International Electrotechnical Commission (IEC) under their Joint Technical Committee 1 (JTC1), Subcommittee SC37 Biometricsofficially called ISO/IEC BioAPI , from 1 May 2006inofficially called BioAPI 2.0 (2.1 also exists)initiated by the BioAPI consortium (BioAPI 1.0 and 1.1)
37 What is specified in BioAPI modular architecture and interfaces for integration of biometric solutionsgiven by header declarations in computer language Cmodules for:software for capturing devicessupport of image processing, feature extraction etc.compressionarchivingretrievalapplication modi:personal useenrollment checkphysical accesssupport multiple biometrics and "telebiometrics"
38 ExampleBIR: Biometric Information Record, inherits the structure of CBEFF
39 The Future of Biometrics An answer from the cinema...
40 Biometrics in the media: Star Trek Inventions CommunicatorTricorderSliding DoorWireless EarpieceBiometricsPortable MemoryTablet & StylusPersonal ComputerGPS
41 Biometrics in the Movies Fingerprints don't lie (1951): fake fingerprintsDracula 2000 (2000): handprint, voice & eye scan spoofedThe Mad Magician (1954): fingerprint identification (?)Hollow Man (2000): fingerprint + voiceprint.2001: A Space Odyssey (1968): voice recognitionMission Impossible 2 (2000): retinal, facial & voice recognition.Diamonds Are Forever (1971): fake fingerprintX-Men (2000): eye scanned access control.Blade Runner (1982): reading the eye to detect replicantFrequency (2000): fingerprint.Star Trek II The Wrath of Khan (1982): retinal recognition to open the Genesis project file.Planet of the Apes (2001) hand recognition to access the Oberon control room.Never say never again (James Bond, 1983): Eyeball replacement to access nuclear weapon.Ocean's Eleven (2001): fingerprint access control of the Bellagio safe.Wasabi (2001): handwritten signature.Star Trek III The Search for Spock (1984): Kirk voice identification to engage the Enterprise auto-destruction.Replicant (2001): face recognition access control.Bad Company (2002): laptop retina protected.Beverly Hills Cop II (1987): superglue fumes for latent printMinority report (2002): iris recognition, eyeball replacement.Licence To Kill (1989): gun with palm recognitionThe Bourne Identity (2002): palm-print identification to access safe.Back to the Future II (1989): fingerprint reader to enter McFly house, Biff paying the taxi.X-Men 2 (2003): eye scan, hand access, voice check.Paycheck (2004): face recognition to select the closest candidate.(La Femme) Nikita (1990): fingerprint match.The Bourne Supremacy (2004): HP iPAQ h5000 series / fingerprint.Judge Dredd (1995): biometrically protected gun.I, Robot (2004): voice id, side palm access control.Seven (1995): fingerprints.Catwoman (2004): lips matching.Barb Wire (1996): retinal scan.Mr & Ms Smith (2005): Voice recognition to access Mr Smith's computer.Mission Impossible (1996): fingerprint, voiceprint, retinal scan.The Island (2005): face & fingerprint recognition of a clone.Critical decision (1996): terrorist recognized using voiceprint.xXx: State of the Union (2005): hand recognition of the president.Air Force One (1997)Les chevaliers du ciel (2005): fingerprint recognition.Alien 4 Resurrection (1997): breath recognition, which is further spoofed with spray.Ocean's Thirteen (2007) Fingerprint recognition (FingerChip) to access the Greco computer room.Face/Off (1997): voice recognition, fingerprint protected jail.Fantastic Four: Rise of the Silver Surfer (2007) hand and fingerprint recognition to access labs.Gattaca (1997): DNA analysis.Men in Black (1997): removing fingerprints to become anonymous.Tomorrow never dies (1997): face recognition, fingerprint protected safes.The Creeps (1997): latent fingerprint.Antitrust (2000): dust for fingerprints on the daycare computer keyboard.Charlie’s Angels (2000): iris, fingerprint protected safe, voice-identification software.
42 Vision of Biometricsthey mostly show a realistic version of biometrics, and not so much “science fiction” (exception “Gattaca”)the purpose is the same as in reality: access restriction to physical spaces, identification of persons, securing, control of persons etc.actually, not much innovations are shown, rather reference to an “archetype” is givenneutrality: it is used by the good guys as well as by the bad guys