1
International Workshop on Biometrics and Forensics (IWBF)
Lisbon, April 4-5, 2013
Introducing a LR-based
identification system in
forensic practice:
opportunities and challenges
Prof. Christophe Champod
IC 1106
Objectives of this presentation
• Expose my views on the opportunities andchallenges a probabilistic (LR-based) assessment of the weight of evidence may pose in traditional forensic identification areas.
• Focus on one area: fingerprint comparison. But the considerations will apply to other areas.
• More emphasis is put on the perspective of the
fingerprint examiner and organisations than on the perspective of the engineer / computer scientist.
2
A few themes
1.
Managing the change from “
fact
” to
“
opinion
”.
2.
Managing the arrival of
probabilistic
models, backing up expert’s opinion
and conflict resolution.
3.
Q/A tools
and saying more from
“
inconclusive
” cases.
3
From
fact
to
opinion
• Saks MJ, Koehler JJ. The individualisation Fallacy in Forensic Science Evidence. Vanderbilt Law Review. 2008;61:199-219.
• Saks MJ, Faigman DL. Failed Forensics: How Forensic Science Lost Its Way and How It Might Yet Find It. Annual Review of Law and Social Science. 2008;4(1):149-71.
• Cole SA. Forensics Without Uniqueness, Conclusions Without Individualization: the New Epistemology of Forensic Identification. Law Probability and Risk. 2009;8(3):233-255.
• State of Maryland v. Bryan Rose (2007) Circuit Court for Baltimore County, Case No. K06-545. Judge Susan Souder: “The 100 percent certainty expressed by Mr. Meagher in this case, as well as in other forum, and others has been persuasively questioned by some academics and defense counsel. The absolute certainty has been proved to be wrong in the past.”
Individualization: NAS report
• Difficult to report?– p.112: Rather than claiming absolute truth, science approaches truth either through breakthrough discoveries or incrementally, by testing theories repeatedly.
– P.142: Although there is limited information about the accuracy and reliability of friction ridge analyses, claims that these analyses have zero error rates are not scientifically plausible.
– P.184:The concept of individualisation is that an object found at a crime scene can be uniquely associated with one particular source. By acknowledging that there can be uncertainties in this process, the concept of “uniquely associated with” must be replaced with a probabilistic association, and other sources of the crime scene evidence cannot be completely discounted.
5 National Research Council, Strengthening Forensic Science in the United States: A Path Forward. Washington, D.C.: The National Academies Press, 2009.
The challenge
“Radical skepticism of all possible assertions of uniqueness is not justified, and the optimal format for explaining the logical impact of a match is not self-evident. But it is clear that one
way or another, a less absolutist and more nuanced theory of
identification is essential if forensic scientists are to adhere to scientific precepts and to contribute to the just resolution of criminal cases.”
6 Kaye DHScience Evidence: Listening to the Academies. Brooklyn Law . Probability, Individualization, and Uniqueness in Forensic Review. 2010; 75: 1163-1185.
Recent changes in reporting
7
“Individualization of an impression to one
source is the
decision
that the likelihood the
impression was made by another (different)
source is
so remote
that it is considered as a
practical impossibility
”
SWGFAST Document #10 Standard for Examining Friction Ridge Impressions and Resulting Conclusions, Ver. 2.0,
http://www.swgfast.org/documents/examinations-conclusions/121124_Examinations-Conclusions_2.0.pdf
Just an Opinion ?
8 A. Campbell, The Fingerprint Inquiry Report.
Edinburgh: APS Group Scotland, 2011.
2.45 - The
decision
whether or not
a mark can be individualised is
potentially a complex one calling
for a series of subjective
judgments on the part of the
examiner. The
decision
is one of
3
Opinionization
We should require more than mere
ipse dixit
S.A. Cole, The 'Opinionization' of Fingerprint Evidence, BioSocieties. 3 (2008) 115-113. “To begin with, there is a sense in which ‘opinion’ can become an all-encompassing shield that deflects all accountability.”
9
Recent changes in reporting
10
“Individualization of an impression to one
source is the
decision
that the likelihood the
impression was made by another (different)
source is
so remote
that it is considered as a
practical impossibility
”
SWGFAST, Standards for Examining Friction Ridge Impressions and Resulting Conclusions, ver. 1.0.
That opens a brand new door on
…
11
http://leadershipfreak.files.wordpress.com/ 2010/03/open-door.jpg
probabilities
Greener grass, blue sky, really?
Your
statement
never mention that
adverse probability, why ?
I would
not
expect this
on another individual ?
What do you mean by
A Decision
?
A
likelihood
so remote ?
Practical impossibility
?
Just an Opinion ?
13 A. Campbell, The Fingerprint Inquiry Report.
Edinburgh: APS Group Scotland, 2011.
38.24 - What matters
more than
the choice of language
(whether
the witness says that he is
‘confident’, ‘sure’, ‘certain’ or ‘in
no doubt’) is the
transparency
of
the opinion.
Transparent decision process
Weight of the forensic findings (MP < 10-15)
ID decision – An adventitious match is a practical impossibility Utility All fingers on Earth “Leap of faith” (D. Stoney) Decision
1/all fingers LR for FP comparison 99.99…%
14 A. Biedermann, S. Bozza, and F. Taroni, "Decision theoretic properties of forensic identification: Underlying logic and argumentative implications," Forensic Science International, vol. 177, pp. 120-132, 2008.
How that decision is taken?
Expert Working Group on Human Factors in Latent Print Analysis, Latent Print Examination and Human Factors: Improving the Practice through a Systems Approach. Washington DC: U.S. Department of Commerce, National Institute of Standards and Technology, 2012, p. 68 Framework (priors) Earth population paradigm Evidence Two generic questions forming a likelihood ratio Update (posteriors) Require both the priors and the evidence Decision on the ID or Exclusion
Based on the posterior probabilities and an utility function
15
What do you know about:
16
Relative frequencies
of level
1 features (general pattern)
and their use ?
Relative occurrence of
level 2 features
(
minutiae
) ?
Level 3 (
pores
) ?
And it goes beyond that…5
Requirement for training
17
Regardless of the availability of statistical models to assess fingerprint comparison, there is an urgent need for fingerprint experts
to understand and to be able to articulate in court and in their statements the probabilistic nature of their decisions.
Equilibrium in the
spectrum
of knowledge
• Years of experience • Unstructured data collection from uncontrolled casework • Relevant systematic studies (published or documented) • Structured portfolio of cases • Proficiency and collaborative testingProbabilistic vs unfettered opinions
18
Arrival of probabilistic models
• Help assign the weight of evidence to the wholeconfiguration without decomposing the contribution of its individual minutiae.
• The more recent efforts have been successfully presented to the Royal Statistical Society: C. Neumann, I. W. Evett, and J. Skerrett, "Quantifying the weight of evidence from a forensic fingerprint comparison: a new paradigm," Journal of the Royal Statistical Society, vol. 175, pp. 371-415 (with discussion), 2012.
19
Neumann & al. (2012)
8 C. Neumann, I. W. Evett and J. Skerrett
(a)
(b)
Fig. 2. Illustrations of the radial triangulation used to orderminutiae(the variables listed in Table 1 are indicated in grey in (a);!, ‘directions’ of theminutiae): the type of theminutiais not indicated since it is a dis-crete variable; depending on distortion, radial triangulation ofminutiaeconfigurations may produce different ordered sets; the differences between (a) and (b) are presented by using thinner lines
unique, rotation invariant and robust to fingerprint distortion. For a given configuration the method requires the determination of a unique centre for the polygon, defined by the arithmetic
mean of the Cartesian co-ordinates of theminutiae. The polygon is then defined by connecting
theminutiae, which become the vertices, in clockwise order with respect to the centre of the
polygon (Fig. 2). Eachminutiais connected to two contiguousminutiaeand also to the centre
of the polygon, hence creating a series of contiguous triangles, which all share one vertex located
in the centre of the polygon. Providing that theminutiaeare always considered in the same order
with respect to the centre, the capture of information for a given configuration is invariant to the orientation of the fingerprint on the image. Furthermore, the circular acquisition of the minutiaerenders the acquisition process independent of the choice of startingminutia. Table 1
provides the notation for the data that are extracted from theminutiae. Note that the directions
of theminutiaeare recorded along the supporting ridge for ridge ending, and along the bisector
of the splitting ridges for a bifurcation (Fig. 2).
4.2. Organization of minutiae features
For a configuration ofk minutiaethere are 5kfeatures that are recorded in a feature array. A
fea-ture array that is recorded from a randomly selectedminutia, which is subsequently denoted 1,
may be written (following the notation in Table 1) w.k,1/=.
{δi,σi,θi,αi,τi,},i=1, 2, . . . ,k/:
We usewin this context because we may be considering a marky, printxdatabase member
zor a simulated mark (for the notation, see later). Note thatw.k,1/is an ordered set. The array
that would be created by starting at the next (clockwise)minutia, which is denoted 2, would be
w.k,2/=.{δ i,σi,θi,αi,τi,},i=2, 3, . . . ,k, 1/: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48
Quantifying Weight of Evidence from Fingerprint Comparison 21
3 4 5 6 7 8 9 10 11 12 −15 −10 −5 0 5 10 15 20 25 Log 10 LR (H d ) Number of minutiae 3 4 5 6 7 8 9 10 11 12 −15 −10 −5 0 5 10 15 20 25 Log 10 LR (H p ) Number of minutiae (b) (a)
Fig. 5. Distributions for LR.Hp/and LR.Hd/from the large validation study, for configurations of sizes kD3–12: (a) LRs whenHpis true; (b) LRs whenHdis true
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 20
To be part of the accreditation scope, regardless of its scientific merits, it will require operational validation
How LRs are assigned (generic)
21 J. Abraham, C. Champod, C. Lennard, and C. Roux, "Spatial Analysis of Corresponding Fingerprint Features from Match and Close Non-Match Populations," Forensic Science International, in press.
LR=P s | Genuine
(
)
P s | Impostor
(
)
The conditioning propositions are not so trivial as Genuineand Impostor
Forensic Error Rates
22 J. Abraham, C. Champod, C. Lennard, and C. Roux, "Spatial Analysis of Corresponding Fingerprint Features from Match and Close Non-Match Populations," Forensic Science International, in press.
RMED=3.4% RMEP=3.2% What is an acceptable error rate? We should confirm them through an operational validation Independent sets of marks and prints of known sources are required for an operational validation
Models can be used in two ways:
23
41.32 […] to provide background data
to assist fingerprint examiners with their evaluation of marks and to enable them to express the strength of their
conclusion in a transparent and verifiable manner.
A. Campbell, The Fingerprint Inquiry Report. Edinburgh: APS Group Scotland, 2011.
Backing up examiners ?
24
We need to precisely define how these “experts” will operate.
A set of SOPs need to be drafted before any operational implementation
To embrace it, I need to understand and
trust the model I need to be able to explain its meaning and limitations
I may want to identify regardless of the number given by the
model Models are
inevitable in the future
As in DNA, probabilities will be asked by both prosecution and defence
7
Pilot joint project
25
The
situation
we want to handle
26
My opinion is that the mark has been identified
to the right thumb of Mr X.
The probability for the mark to originate from someone else is so small
that I consider it to be a
practical impossibility.
We submitted your case a statistical analysis through the University XYZ, Prof S. Tat
The LR obtained is 1.8e+6, that amounts to a match probability of 5.6e-7
How do you get from a match probability of 5.7e-7 to an identification?
Expert 1 (LR = 5.8e10
5)
27
Expert 2 (LR = 9.6e10
10)
Conflict resolution procedures
29 Decision First examiner Statistical Model Second examiner (verifier)A quick look at the task complexity
Understanding the concept of “sufficiency” in friction ridge examination
C. Neumann, C. Champod, M. Yoo, G. Langenburg, T. Genessay, NIJ award - 2010-DN-BX-K267
Results presented at the annual meeting of the American Academy of Forensic Science, Feb 2013.
30
Understanding “Sufficiency”
• 15 comparisons:– 12 pairs latent/control prints from same source – 3 pairs latent/control prints from different sources
• Information captured through a web-based software
designed to support the ACE process (Picture Annotation Software – PiAnoS –
https://ips-labs.unil.ch/pianos/index.html) • Approximately 600 examiners contacted
– 145 completed first comparison – 123 completed all 15 comparisons
31
PiAnoS'
9
PiAnoS'
33 Minutiae annotated Minutiae annotated Change of size of displayed minutiae markers Change of size of displayed minutiae markersPiAnoS'
34PiAnoS'
35 120 4 21 127 11 120 17 7 3 126 34 7 93 29 6 98 73 13 45 43 19 66 2 106 19 100 4 21 13 9 103 11 73 40 76 8 38 72 18 33 28 5 90Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
ID EXC INC ID EXC INC ID EXC INC
Decisions following comparison
N umb er of re sp on da nt s
04.04.13
Trial 8 (same source)
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
Categorical Towards Neutral Categorical Towards Neutral Categorical Towards Neutral Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
Categorical Towards Neutral Categorical Towards Neutral CategoricalTowards Neutral Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
Categorical Towards Neutral CategoricalTowards Neutral Categorical Towards Neutral Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
Categorical Towards Neutral CategoricalTowards Neutral Categorical Towards Neutral
Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
CategoricalTowards Neutral CategoricalTowards Neutral CategoricalTowards Neutral
Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
CategoricalTowards Neutral CategoricalTowards Neutral CategoricalTowards Neutral
Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 12 (different sources)
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
Categorical Towards Neutral Categorical Towards Neutral Categorical Towards Neutral Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
CategoricalTowards Neutral Categorical Towards Neutral Categorical Towards Neutral Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
Categorical Towards Neutral CategoricalTowards Neutral Categorical Towards Neutral Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
Categorical Towards Neutral CategoricalTowards Neutral Categorical Towards Neutral
Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
CategoricalTowards Neutral CategoricalTowards Neutral CategoricalTowards Neutral
Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Trial 01 same source Trial 02 different sources Trial 03 same source
Trial 04 same source Trial 05 same source Trial 06 same source
Trial 07 same source Trial 08 same source Trial 09 different sources
Trial 10 same source Trial 11 same source Trial 12 different sources
Trial 13 same source Trial 14 same source Trial 15 same source
Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading Guiding correctly Neutral Misleading
CategoricalTowards Neutral CategoricalTowards Neutral CategoricalTowards Neutral
Level of confidence R el ia bi lit y of co ncl usi on Decision in analysis NV VID VEO Conclusion Exclusion Identification Inconclusive
Conflict resolution procedures
39 Decision First examiner Statistical Model Second examiner (verifier)
Other usage: Q/A Tool ?
40 J. Abraham, C. Champod, C. Lennard, and C. Roux, "Spatial Analysis of Corresponding Fingerprint Features from Match and Close Non-Match Populations," Forensic Science International, in press.
11
Other usage: Suitability tool?
41 Image from the FIGS group, Adrian Gardner
The expected LR associated with the 5-6 level 2 features is 12’540.
Is the mark suitable for further comparison,…, for identification? Is that a complex case?
Models can be used in two ways:
42
41.33 […] the possible application of probabilistic analysis to comparisons that examiners would otherwise consider to be inconclusive, the objective being to produce evidence that could be used in court of the probability of a match.
A. Campbell, The Fingerprint Inquiry Report. Edinburgh: APS Group Scotland, 2011.
“Unable to exclude” cases
43 A. Campbell, The Fingerprint Inquiry Report.
Edinburgh: APS Group Scotland, 2011. 38.84 - Before a finding of ‘unable to exclude’ is led in evidence, careful consideration will require to be given to (a) the types of mark for which such a finding is meaningful and (b) the
proper interpretation of the finding. An examiner led in evidence to support such a finding will require to give a
careful explanation of its limitations.
Saying more than “inconclusive”
44
Again, we need to precisely define the scope of usage
The statistics may convey more weight than it deserves
Very useful source of additional information, either as evidence or for intelligence purposes
We don’t want to mislead anyone
Already a reality for the Dutch NFI
Value for comparison
Searchable on IAFIS
Value for identification
Diminishing match probabilities (or increasing LR) IDEN
T IF IC A T O N D EC ISI O N 45
Risks of misleading may outweigh benefits
Helps with the decision making Intelligence tool Allows transparency but marginal impact on decision making [1 – 1/10-3] [1/10-3-1/10-9] [<10-9]
Need more resources Need more resources Actual resources
Complex Non-complex
Managing the change
46
Educate Judiciary and field officers
Train fingerprint examiners Define the
exact scope
Define the SOPs
Conduct operational validation Agree on acceptable error rates Conflict resolution procedures
Contact details
Prof. Christophe ChampodEcole des sciences criminelles / Institut de police scientifique Batochime / Quartier Sorge
CH-1015 Lausanne
Tel: +41 (0)21 692 46 29 Fax: +41 (0)21 692 46 05 E-mail: christophe.champod@unil.ch