Using Automated Facial Recognition to Select Fillers for Eyewitness Identification Lineups

Abstract

The current recommendation for selecting fillers for an eyewitness identification lineup is to use the eyewitness' description of the culprit (match-to-description). I argue that this recommendation is inadequate given eyewitness' limited recall ability and expanding police resources. Instead, I contend that fillers should be selected based on their match to the appearance of the suspect (match-to-suspect). Specifically, in this dissertation, I suggest using automated facial recognition (AFR) software and the similarity ratings generated to select match-to-suspect fillers. Although researchers have noted problems with match-to-suspect fillers, using objective assessments of similarity offers resolution to those problems. In Study 1, AFR was used to build lineups with three different levels of suspect-filler similarity, resulting in three similarity lineup conditions: low, moderate, and high. The purpose of Study 1 was to determine the suspect-filler similarity level which would result in the greatest discriminability. Participants made an identification decision from a simultaneous lineup (Study 1a; N = 1509) or a sequential lineup (Study 1b; N = 1416). Comparison of full ROC curves identified that discriminability did not significantly differ between the three similarity conditions using either the simultaneous or the sequential presentation. However, follow-up expected utility analyses identified that the high similarity lineup should be preferred to the low and moderate similarity lineups under assumptions of low base rates of suspect guilt and high costs of mistaken identifications. In Study 2 (N = 1044), I compared the high similarity AFR lineup to a match-to-description lineup presented simultaneously. The purpose of Study 2 was to examine how lineups built using AFR performed in comparison to the current recommended match-to-description filler selection strategy. Comparison of full ROC AUTOMATED FACIAL RECOGNITION TO SELECT FILLERS iii curves identified that discriminability did not significantly differ between the two conditions. Though, follow-up expected utility analyses identified that the AFR match-tosuspect lineup should be preferred over the match-to-description lineup across low, moderate, and high base rates of suspect guilt and under assumptions of high costs of mistaken identifications. Overall, this research suggests that non-human objective assessments of similarity may be the preferred filler selection strategy over the recommended match-to-description strategy, and that identifying an optimal level of filler similarity may be possible.AUTOMATED FACIAL RECOGNITION TO SELECT FILLERS iv ACKNOWLEDGMENTS Most importantly, I would like to thank my supervisor, Dr. Joanna Pozzulo, for your support, guidance, advice, and positivity throughout the years. Since the start of my Master's degree 7 years ago, you have done nothing but express the utmost confidence in me and that is, undoubtably, a major reason why I have reached this milestone today.You have provided me with so so many opportunities over t...