probability models for impression and pattern evidence · probability models for impression and....
TRANSCRIPT
![Page 1: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/1.jpg)
Probability Models for Impression and Pattern Evidence
Sargur Srihari
Department of Computer Science and Engineering University at Buffalo, The State University of New York
Indigent Symposium 2010
MayFlower Hotel, Washington DC
![Page 2: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/2.jpg)
ForensicsForensics
DigitalDigital
AnalogAnalog
ComputerComputer Multi-
media
Multi-
mediaImpression/
Pattern
Evidence
Impression/
Pattern
Evidence
Trace
Evidence
Trace
Evidence
GeneticGenetic OtherOther
Forensic Modalities
• Network • Video • Database
• Speech• Mobile
• Photos • Latent prints • QD: Handwriting • QD: Printers
• Pollen • Footwear • Dust• DNA
• Paint• Tire tread • Blood
• Glass• Fire-arms • Semen • Fiber• Tool marks • Hair • Arson • Saliva • Voltage
![Page 3: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/3.jpg)
Forensic Opinion
• Courts allow Latent Print Examiner to opine
on ultimate question of individualization
– Evidence is attributed to a single individual and
no other
• Three possible opinions
– The evidence
• Individualizes
– No other individual on earth
• Inconclusive
• Excludes
– Definitely not this individual 3
![Page 4: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/4.jpg)
Madrid Bombing Case Latent Fingerprint Brandon Mayfield Algerian National
on bag of detonator Identified by FBI Identified by SNP
4
![Page 5: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/5.jpg)
ProbabilityModels
ProbabilityModels
GenerativeModels
GenerativeModels
Discriminative Models
Discriminative Models
Probability ofRandomCorrespondence
Probability ofRandomCorrespondence
SimilarityMeasureSimilarityMeasure
Likelihood Ratio:Same/differentLikelihood Ratio:Same/different
FrequentistFrequentist BayesianBayesian
Measurements obtainedfromEvidence
Methods for Expressing Uncertainty
Measurement s obtained from Evidence
![Page 6: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/6.jpg)
How to Compute Uncertainty?
• Discrete Variables
– Birthdays
– DNA
• Continuous Variables
– Heights
– Pattern/Impression evidence (Ongoing Research)
• 5 or 9-point scale suggested by SWGs-- no guidelines
• New statistical models being developed
6
![Page 7: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/7.jpg)
Generative Models: Several Probabilities Random two have Some two among n
A specific birthday among nsame birthday (n=2) have same birthday
PRC nPRC Specific nPRC
Graphical Model
z ={1,0}
Inference
PRC is �= p(z=1)
z’={1,0} z’={1,0}
7
![Page 8: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/8.jpg)
Generative Model: DNA Genome: sequence of 3x109 base-pairs (nucleotides A,C,G,T) Represents full set of chromosomes
Actual Electron photomicrograph Single Chromosome: ~108 base-pairs
Genome has 46 chromosomes (22 are repeated plus XX and XY)
Large portions of DNA have no survival function (98.5%) and have variations useful for identification
TH01 is a location on short arm of chromosome 11: short tandem repeats (STR) of same base pair AATG Variant forms (alleles) different for different individuals
8
![Page 9: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/9.jpg)
Generative Model: DNA
Allele Frequency of single locus for 200 individuals
DNA profile of 13 loci: Average match probability (PRC) is 0.1 per locus, 10-13 for a profile
If database has 1 million entries, since there are 500 billion pair-wise matches, nPRC = 0.05
However specific nPRC can be much lower 9
![Page 10: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/10.jpg)
Generative Models: Fingerprints
• Fingerprints are characterized by ridges and minutiae
(x,y,_)
10
![Page 11: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/11.jpg)
Generative Model: Minutiae 1. Distribution of Minutia 3. Distribution of Minutia Location Dependency
2. Distribution of Minutia 4. Distribution of Minutia Certainty Orientation
Core point predicted Using Regression
![Page 12: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/12.jpg)
Madrid Bomber Case Revisited
• Madrid Bomber case 22
minutiae identified
• 10 were matched by three FBI
experts
• Generative Model – 12 of 19 minutiae used for specific nPRC:
1, 2, 3, 4, 7, 8, 9, 10, 11, 12, 17, and 18
– specific nPRC • World population (6.8b) = 0.16
• US population (330m) = 0.008
12
![Page 13: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/13.jpg)
Discriminative Models: Fingerprints
• Need a similarity measure for fingerprints • AFIS Minutiae Matcher • Score distributions give two probabilities • Likelihood Ratio
• Error Rates
EER EER threshold
Non-Twins 3.33% 19
Fraternal 4.88% 23
Identical 5.09% 25
![Page 14: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/14.jpg)
Discriminative Model: QD
Word shape
Letter shape
Bigram shape
Known Questioned
![Page 15: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/15.jpg)
Handwriting Features and Similarities
Distributions of Similarities
Likelihood Ratio
obtained from
distributions of
similarities
for same and different
origination of evidence
15
![Page 16: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/16.jpg)
123456789
Discriminative Model for Handwriting
. Identified as same
. Highly probable same
. Probably same LLR Value for this pair is 41.52 . Indicating same
. No conclusion � . Indicating different
. Probably different
. Highly probably different
. Identified as different
![Page 17: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/17.jpg)
Human-Machine Interface: Dialog for character image comparison
![Page 18: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/18.jpg)
Signatures: Bayesian Adaptation
Questioned SignatureKnown Signatures
Approach: Learn hyper-parameters from large data setAdapt parameters using Bayesian approach
Wide variability and small learning sets for case at hand
![Page 19: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/19.jpg)
Bayesian Adaptation Error Rate
Only Prior 18%
Bayesian Adaptation 11.5% - 14.5%
After Learning After Adaptation
Parameter �g ~ N(μg 0,� g 0 2 )
� f ~ N(μ f 0,� f 0 2 )
P(�g |Yg ) � N(�g |μg 0,� g 2 ) � N(Yg |�g ,� g
2 ) = N(μg ,� g 2 )
P(� f |Yf ) � N(� f |μ f 0,� f 2 ) � N(Yf |� f ,� f
2 ) = N(μ f ,� f 2 )
Distance
)2(
)(
2 0
2
)2(
)(
2 0
2
2 0
2
2 0
2 0
2
2 0
)(2
1 )|()|()|(
)(2
1 )|()|()|(
ff
ff
gg
gg
d
ff
ffffgf
d
gg
gggggg
edYPP dFP d
edYPP dGP d
��
μ
��
μ
�� �
���
�� �
���
+
� �
�
��
+
� �
�
��
+
=�=
+
=�=
�
� p(t /G) = p(t |�g ) � p(�g |Yg ) ��
�
� d�g = 1
2� (� g 2 + � g
2 ) e �
(t�μg )2
2(� g 2 +� g
2 )
p(t /F) = p(t |� f ) � p(� f |Yf ) ��
�
� d� f = 1
2� (� f 2 + � f
2 ) e �
(t�μ f )2
2(� f 2 +� f
2 )
![Page 20: Probability Models for Impression and Pattern Evidence · Probability Models for Impression and. Pattern Evidence. Sargur Srihari. Department of Computer Science and Engineering](https://reader031.vdocuments.us/reader031/viewer/2022022512/5ae7d0797f8b9a8b2b8f20fd/html5/thumbnails/20.jpg)
Summary and Conclusion
• Uncertainty is easily expressed for DNA
– Discrete counts
• For Impression Evidence it is much harder – Continuous features/tolerances
• New Models proposed (being validated)
– Generative
• PRC
– Fingerprints, Handwriting
– Discriminative (similarity automatically computed)
• Likelihoods/ Bayesian scores
Fingerprints Handwriting