250 likes | 378 Views
A Novel, Countermeasure-proof, P300-Based Protocol for Detection of Concealed Information. J.Peter Rosenfeld, Michael Winograd, Elena Labkovsky, Ann Ming Lui Department of Psychology Institute for Neuroscience Northwestern University.
E N D
A Novel, Countermeasure-proof, P300-Based Protocol for Detection of Concealed Information J.Peter Rosenfeld, Michael Winograd, Elena Labkovsky, Ann Ming Lui Department of Psychology Institute for Neuroscience Northwestern University
Previous P300 DD protocols used Separate Probe(P),Irrelevant(I) and Target(T) trials. • 80% to 95% correct detection rates….but…. *Rosenfeld et al. (2004) and Mertens, Allen et al. (2007):These methods are vulnerable to Counter-measures (CMs) via turning I’s into covert T’s.
Results from Rosenfeld et al. (2004): Farwell-Donchin paradigm (BAD and BCAD are 2 analysis methods.) Diagnoses of Guilty Amplitude Difference (BAD) method,p=.1 Innocent Group Guilty Group CM Group 1/11(9%) 9/11(82%) 2/11(18%) Cross-Correlation(BC-AD) Method, p=.1 0/11(0%) 6/11(54%) 6/11(54%)
Results (hit rates) from Rosenfeld et al. (2004): Rosenfeld paradigm WeekBAD*BC-AD* 1: no CM 12/13(.92) 9/13(.69) 2: CM 6/12(.50) 3/12(.25) 3: no CM 7/12(.58) 3/12(.25) *Note: BCD and BAD are 2 kinds of analytic bootstrap procedures.
In the “Complex Trial Protocol,” P/I and T/NT decisions are separated: • Two stimuli per trial, 1.2-1.5 s apart. • The first is P or I presented in white font. • The second is the same P or I presented either in T color (green) or one of 4 non-T colors (red, yellow, etc.) • (The second could also be T and NT numbers, or whatever.)
4 STIMULUS TYPES STIMULUS TYPE NUMBER PROBABILITY • Probe Target 30 .09 • Probe non-Target 30 .09 • Irrelevant Target 30 .09 • Irrelevant non-Target 240 .73 • All Probes 60 .18 • Oddballs 90 .27 • One probe and 4 irrelevants • P (T/P) = .5… vs… P (T/I) =.11 (Confound?)
DESIGN: as in ’04 paper, Exp. 2: 3 weeks in one group • Week 1 Naïve • Week 2 CM • Week 3 Repeat Week 1 • One block with one probe type, but category varied and counterbalanced across subjects/weeks: 1)mother’s first names 2) family surnames 3) home towns • Main Study plus near replication. • Innocent Control Group for FPs.
Countermeasures in Week 2: Also as in ’04 paper Left finger press to Irrelevant # 1 Left toe wiggle to Irrelevant # 2 Right toe wiggle to Irrelevant # 3 Imagine Prof slaps you for Irrelevant#4 All these are done covertly so that operator cannot detect them.
P & I Individual RTs in CT Protocol (Flat liner at bottom did not beat test.)
Statistical tests within each subject: • T/F-tests comparing Week 1 RT or RT Variance versus Week 2 are all p<.01, and < .001 in the one subject who beat the test in Week 2. • Thus CM use is detectable.
Two* possible (P-I) tests: • 1)Traditional: Probe versus mean of all Irrelevants, P vs I-All. 2)Probe versus Maximum Irrelevant P vs I-max (“simple hit”) or Probe versus I-max not associated with elevated RT (“RT-screened Hit”). * at .9 or .95 confidence levels.
Main Study. Within-subject correct detections of guilty subjects based on bootstrap comparison of probe P300 against the average of all irrelevant P300s over 3 weeks. • WEEKHit Rate [Hit Rate] • Week 1 (no CM): 11/12 (92%) [12/12*( 100%)] • Week 2 (CM): 10/11 (91%) [11/12* (92%)] • Week 3 (no CM): 11/12 (92%) [12/12* (100%)] • Main Study: False positive(FP) group. • Confidence=.9 Confidence=.95 • TestFPsHitsA’FPsHitsA’ • Iall .08 .92 .95 0 .92 .98 • Imax 0 .92 .98 0 .92 .98
Main Study: Simple and RT-qualified diagnoses (at confidence = .9, Probe vs. Imax (or RT-qualified Imax) across 3 weeks (n values in parentheses). CM use also shown. Week 1 (12)Week 2 (11)Week 3 (12) Simple hits .92 .73 .92 Hits/RT qualified .92 .91 .92 CM use 0.0 1.0 0.0
Near Replication: Within-subject correct detections (“Hits”) of guilty subjects based on bootstrap comparison (at 2 confidence levels) of probe(P) P300 against the average of all irrelevant P300s (I-All) over weeks, and against the largest irrelevant P300 (I-Max). • CONFIDENCE LEVEL: 0.90 • WEEKP vs I-All: Hits, [FPs], A’P vs I-Max: Hits, [FPs] A’ • 1: 12/12 (100%),[8%] .91 11/12 ( 92%), [0%] .98 • 2: 12/12 (100%) 11/12 ( 92%) * • 3: 9/10 (90%) 7/10 (70%) • CONFIDENCE LEVEL: 0.95 • WEEK P vs I-All: Hits, [FPs], A’P vs I-Max: Hits, [FPs], A’ • 1: 11/12 (92%),[0%] .98 11/12 ( 92%), [0%] .98 • 2: 12/12 (100%) 11/12 ( 92%) * • 3: 9/11 (82%) ** 8/11 (73%)**
CTP Mock Crime Study: Preliminary Results. Note Target= 11111. (Mike Winograd’s study)
Conclusions (& Why ? ) • The complex trial protocol is CM-resistant and accurate in the CIT Context. *The S1 involves no classification or decision, unlike older protocols, whose target classification task is removed, leaving all resources devoted to probe/irrelevant recognition. *CMs force more attention to first stimulus increased probe (& Irrel) P300s.
NEXT? • We need to extend CTP to our hybrid CQT screening protocol (Rosenfeld et al., 1991.) • We need to try 3-4 blocks a session, each with different probe category. • CTP should be even better with more Irrelevants.
A Novel, Countermeasure-proof, P300-Based Protocol for Detection of Deception (DD). jp-rosenfeld@northwestern.edu