P300 in Detection of Deception - Northwestern University: Psychology

advertisement

Countermeasures to P300based Guilty Knowledge

Tests of Deception

J.Peter Rosenfeld, Matt

Soskins,Joanna Blackburn,

& Ann Mary Robertson

Northwestern University.

Supported by DoDPI

Countermeasure issues:

 Among the problems with both the ANS-based CQT and CIT raised by the report of the National Research Council of the National

Academy of sciences ( National Research Council, 2003 ) is the potential susceptibility of all ANS-based methods to countermeasures (CMs). As stated by ( Honts, Devitt, Winbush, &

Kircher, 1996, p. 84 ), ‘‘Countermeasures are anything that an individual might do in an effort to defeat or distort a polygraph test.’’ The National Research Council report went on to state that

‘‘Countermeasures pose a serious threat to the performance of polygraph testing because all the physiological indicators measured by the polygraph can be altered by conscious efforts through cognitive or physical means’’ ( National Research Council, 2003, p.

4 ).

ERPs to the rescue?

 Deception researchers all hoped and indeed expected that when the

P300 Event-Related EEG Potential was introduced as the dependent index of recognition in a CIT ( Farwell & Donchin, 1991 ; Rosenfeld,

Angell, Johnson, & Qian, 1991 ; Rosenfeld et al., 1988 ), the CM issue would be resolved. For example, the eminent inventor of the

GKT / CIT, ( Lykken, 1998, p. 293 ), suggested about CMs to P300

CITs: ‘‘Because such potentials are derived from brain signals that occur only a few hundred ms after the GKT alternatives are presented… it is unlikely that countermeasures could be used successfully to defeat a GKT derived from the recording of cerebral signals.’’ ( Ben-Shakhar & Elaad, 2002, expressed a similar view.

) All this optimism, as shown below, turned out to be misplaced.

Some History (earliest publications)

 Rosenfeld et al., 1987,1988,1991

 Farwell and Donchin, 1991

 Allen, Iacono, & Danielson, 1992

 Johnson and Rosenfeld, 1992

 Since we were there at beginning, why do we challenge as late as 2003-4 with countermeasures? (1) It’s about time ….

2) Farwell’s web page, claiming 100% accuracy:

Stimuli used in 3-SP:

 (1)Probes (P or R in figures): Items which subject is suspected of knowing (e.g., murder weapons). Subject denies(lies by pressing ‘ NO ’ ).

 (2)Irrelevants (I or W in figures): Items of which subject has no knowledge and denies, honestly, by pressing ‘ NO ’ .

 (3) Targets (TR) Items: to which subject presses ‘ YES ’ .

(Benchmark P300).

Irrelevant Items

We ultimately knew we could beat the test…..

 In the ordinary, un-countered 3-stimulus protocol, the subject is instructed to make unique responses to explicitly assigned targets which are readily executed with the typical result that large target P300s are evoked, since these targets are also rare and additionally, meaningful , due to their unique button requirement.

(Rareness and meaningfulness are the major antecedents for P300; Johnson,

1986.)

IF….the subject can follow an experimenter’s instruction to respond uniquely to an experimenterchosen irrelevant (an explicit target)…

…….. then the subject could also covertly define some (or all) irrelevants for himself as implicit targets to which he could make unique responses.

These originally irrelevant but now secret targets would also elicit large P300s so that one could no longer depend on the probe P300 amplitude to reliably exceed that of the irrelevant P300.

* The larger probe P300 is, of course, what ordinarily makes the diagnosis of possession of concealed information.

How P300 amplitude is supposed to catch Liars:

1)P>I (‘BAD’)

2)P-TR corr

>P-I corr(‘BC-AD’)

1)P=I

2)P-I corr

>P-TR corr

Whither R-TR correlation if there are latency differences?

Probe P3 Target P3

Nothing should happen to bootstrapped amplitude difference test (BAD) but bootstrapped cross-correlation test (BC-AD) should fail.

Experiment 1, based on

Farwell & Donchin (1991) :

--6 Different Probes

(“multiple probe protocol”)

--Innocent, Guilty, and Countermeasure(CM)

Groups

--Countermeasure: Associate various latent responses to different categories (jewelry type, drawer color, operation name, etc.), all irrelevant members of the category.

General Instructions….

Mock crime scenario

Press “Yes” to Targets (on list)

Press “No” to all other stimuli (Possibly guilty probes and Irrelevants).

More simply….

 Probe Target I1 I2 I3 I4 ring bracelet necklace watch broach tiara

 pink brown yellow purple red blue

 donkey tiger lion cow pig horse

 etc., etc…... (only half the matrix here.)

 All these are shuffled, presented in random order, involving 4 repetitions of each item.

What are the covert countermeasures for the 6 categories of 6 probes?

 1) Jewelry category……….micro right finger wiggle

 2) drawer lining category…. “ left “ “

 3) owner’s name category…. “ “ toe “

 4) operation name category…” right “ “

 5) location of item category…… Imagine professor slaps you

 6) desktop category……………Do Nothing

 I.e., make irrelevants into relevant targets.

Guilty group: Probe(R)

> Irrelevant (W).

R > W

Guilty Group: TR vs R

Both have P300

Innocent Group: R vs W

Both lack P300

Innocent Group: TR vs R

TR towers over P (R)

CM Group : R vs W

No difference

P(R) vs I (W)

CM Group: Tr vs R

Target > Probe

Results

, Exp. 1: CM works, and analysis method matters:

Diagnoses of Guilty

Amplitude Difference (BAD) method,p=.1

Guilty Group Innocent Group

CM Group

9/11(82%) 1/11(9%) 2/11(18%)

Cross-Correlation(BC-AD) Method, p=.1

6/11(54%) 0/11(0%) 6/11(54%)

--with “Off the street” subjects; see next 2 slides.

Guilty Subject, probe(R) vs irrelevant(W)

R >> W: subject clearly guilty

…but clear latency shift in

TR and R P300s

..so BC-AD fails,

BAD catches the S.

Experiment 2: (“One probe protocol”) Rosenfeld et al., 1991).

 100% simpler to conceptualize:

 There are 6 items, repeated 30-40 times randomly, consisting of..

 A probe, subject maybe guilty,

 A target to force attention (not really necessary but allows us to test Farwell cross-correlation method vs. ours.)

 4 irrelevants.

Standard autobiographical oddball paradigm.

 6 dates are presented, one is the birth date.

 5 other dates with no personal significance, one is a target, four are irrelevant.

 Subject is feigning head injury, the inability to recognize birth date 50% of the time, so he presses yes and no to all dates on 50-50 basis.

Design:

 One group of advanced Ss run in 3 successive weeks . (Also Control group.)

 Week 1: All naïve(about CM) and guilty

(of having birth date and recognizing it despite overt lies.

 Week 2: Use explicit countermeasure (to be explained).

 Week 3: Repeat first week, stop the CM.

What’s the CM?

 When you see the first Irrelevant, do micro left finger wiggle.

 When you see second Irrelevant, do micro right toe wiggle.

 When you see third irrelevant, imagine me slapping your face.

 When you see fourth Irrelevant, do nothing.

 I.e., make irrels. into rel. targets.

Results, week 1(guilty):

Probe(R) vs.Irrelevant (W)

As usual, R>>W

Week 1: Probe(R) and

Target(TR):

Both have nice P3

Week 2: Explicit CM, R v W

Not so different anymore...

Week 3: 5/12 test beaters(effortless), R v W:

Surprise! R = W

Week 3: 5/12 test beaters,

R v TR--”classic defeats”:

..

and TR >> R as with innocents .

Exp. 2; quantitative.

 WK BAD BC-AD

 no CM 12/13(.92) 9/13(.69)

 CM 6/12(.50) 3/12(.25)

 no CM 7/12(.58) 3/12(.25)

 (Control group: nothing much happened over

3 weeks of repeating week 1.)

RTs for 3 weeks; week 1= week 3, proving CM not used in week 3 .

1100

1000

900

800

700

600

500

400

0 1 2

WEEK

3 4

PROBE

TARGET

IRRELEV

Conclusions, bottom lines..

 6-probe/ 3S protocol beat-able, and the

6 probe combination lacks a real rationale anyway. (Lykken wouldn’t like?)

 1-probe/3S protocol may be explicitly beat-able, but the very slow Irrelevant RT distribution will raise suspicions. 1 probe per run is more Lykkenable.

 BUT ---1-probe paradigm after CM practice is beat-able, period.

Farwell (SPR ‘08) didn’t agree:

But at the meeting, his letter, not he, showed up:

Cogn Neurodynamics

DOI 10.1007/s11571-012-9230-0

 Brain fingerprinting field studies comparing P300-MERMER and P300 brainwave responses in the detection of concealed information

 Lawrence A. Farwell • Drew C. Richardson

• Graham M. Richardson

 Pub. On line Dec 2012

Includes full “$100,000

Reward” (CM) Study.

 But: How does he know the CMs are really done??!!

 There are NO (Zip) Reaction Time data.

 And the ERPs do not suggest CMs are being done.

Labkovsky & Rosenfeld

(2011): Real CM effects on RT

Controls (???)

When CMs (2004) are done:

really

So we can forget Farwell.

(Every one else has…)

The 3SP is vulnerable to CMs, no doubt about it.

What to do?

 Go to a new paradigm—the Complex Trial

Protocol (Rosenfeld et al., 2008)

Download