Computational Intelligence - February 2017 - 23

For the Type-II detector (second lines in
In automated border control machines, watchlist
Table 4), the results show the desired Doddington's categorization as the numbers
technologies are radically different from those known
resemble closest to the selected 2.5% (Figure
in homeland security such as Entry-Exit systems for
3). There is an even split between the 'Goat'
visitors, like U.S. VISIT and the European Union (EU)
and 'Wolf/Lamb' categories because the chosen percentile is the same for both categories.
SmartBorders, or those used in forensics.
There is a special case when a traveler is
considered both a 'Goat' and 'Wolf/Lamb.'
Using the Type-III detector (third lines in
Table 4), different decision values will result in different disA single-phase Doddington's detector assumes a cooperatributions of the Doddington's categories for a given database.
tive traveler who provides his/her biometric traits according
In this study, we examined the decision value 0.5 to 0.3. The
to the special rules of machine-human interactions
Type-III detector indicates (a) certain travelers are classified
(Figure 4). In the two-phase approach, we start to use a
into multiple Doddington's categories and (b) a high percentDoddington's detector before the traveler reaches the acquiage of 'Goat' travelers for the LFW database. This is predicted
sition station (the first phase). The obtained preliminary
to be caused by low genuine scores as a result of the quality
results, or pre-assessment, of a non-cooperative traveler are
of images in the LFW database. In a border crossing scenario,
then used for improving the reliability of the risk assessment
this means that only the Type-III detector will generate a realfor the cooperative traveler.
istic Doddington's landscape. High non-cooperative behavior
Figure 4 shows how the Doddington's detector can be
of travelers is detected at the waiting line of the border conintegrated into the automated border control machine. At
trol machine.
Phase I, face images of a non-cooperative (walking or waiting)
We observe that the results of the Type-I, II, and III Dodtraveler are processed in the identification mode (with respect
dington's detector are different, and that their impact on the
to the watchlist) aiming at his/her risk pre-assessment. At Phase
watchlist and e-border performance will also be different.
II, the cooperative traveler provides his/her passport/ID and
face photo which is processed in verification mode aiming at
the traveler's final risk assessment taking into account the preV. Two-Phase Watchlist Inference
assessment results.
In this section, we show how single-phase watchlist inference can
be improved. It was demonstrated experimentally in the previous
section that there is always a risk of misidentification, using a bioB. Doddington's Likelihood and Evidence
metric-enabled watchlist check, that impacts the e-border perThe traveler risk pre-assessment scenario for formalization
formance. This is the main motivation to develop a more reliable
(Figure 4) is as follows: Given a set of face images of a nonwatchlist check called two-phase watchlist inference.
cooperative traveler, determine his/her Doddington's category
using the watchlist check. This problem addresses a classic
Bayesian profile (calculation of a posterior probability given an
A. Two-Phase Traveler Risk Assessment
initial (prior) probability, likelihood, and evidence).
The concept of pre-assessment is well known in operational
Let AG and AI be average genuine and imposter scores,
forensic science [3], [64]. The idea is based on a model
embodying the likelihood ratio as a measure of the weight of
respectively, calculated based on the face matching proceevidence. The aim is to estimate the expected likelihood ratio
dure. Let us define Doddington's likelihood as the multibefore performing any test. In our approach, pre-assessment
plication of the probabilities of independent constituting
corresponds to the first phase of traveler risk assessment.
components of the Doddington's likelihood of genuine

Table 4 Watchlist check experiment for facial traits in Doddington's metric using various types of detectors (I, II, and III) and
databases: FRGC database (left plane), ORL database (central plane), and the LFW database (right plane).
frgC DaTabase
D

'goaT'

n/a

0.0000

'lamb'

'Wolf'

n/a

0.0264

iii

0.5

0.0035

0.0000

iii

0.4

0.0546

0.0000

iii

0.3

0.1972

0.0000

orl DaTabase
'sheeP'

'goaT'

0.0370

0.9630

0.000

'lamb'

'Wolf'

0.0264

0.9472

0.050

0.9965

0.050

0.000

0.0000

0.9454

0.100

0.025

0.0000

0.8028

0.150

0.075

lfW DaTabase
'sheeP'

'goaT'

0.0750

0.925

0.0000

'lamb'

'Wolf'

'sheeP'

0.0500

0.900

0.0253

0.950

0.8898

0.2474

0.0000

0.1102

0.000

0.900

0.8946

0.3299

0.0138

0.1023

0.000

0.825

0.8977

0.4250

0.6707

0.0415

0.0223

0.9777

0.0253

0.9506

note: D denotes the type of detector (i, ii, or iii); T denotes the threshold values (n/a, 0.5, 0.4, or 0.3).

February 2017 | Ieee ComputatIonal IntellIgenCe magazIne

Table of Contents for the Digital Edition of Computational Intelligence - February 2017