Manufacturing foils for police lineups with an artificial

Manufacturing foils for police lineups with an artificial face synthesizer Caitlin Grist & Colin Tredoux Department of Psychology University of Cape Town AP-LS March 2013

How to choose fillers for lineups?

Selection strategies • Match to suspect – foils are chosen to meet some degree of subjective physical similarity [usually facial similarity] to the suspect • Match to description of perpetrator – foils are chosen to match a number of the descriptors given of the perpetrator by the eyewitness

Selection strategies - issues • Match to suspect – how much similarity? – who decides? on what basis? • Match to description – person descriptions are usually vague – default descriptors can be a problem

Automating match-to-suspect • A decade working with synthetic faces – ID, an eigenface compositor • Recently developed a lineup generator tool – can specify number of foils required and level of similarity

Experiments • Experiment 1 – are the faces believed to be realistic? • Experiment 2 – can we control suspect-foil similarity so that it has convincing results? • Experiment 3 – test lineups for fairness, using mock witness methods

Experiment 1 • 49 student participants • Each shown 25 real faces and 25 artificial faces • Asked to rate as either real or fake • Additional questions to determine reasons for differential ratings, if any – e. g. how natural is the face? rate the picture quality

Results Testing against chance performance Fake Mean SD df t calc r 0. 47 0. 49 48 -0. 43 -0. 06

Results Testing against chance performance Mean SD df t calc r Fake 0. 47 0. 49 48 -0. 43 -0. 06 Real 0. 67 0. 47 48 2. 53 0. 34

Results Mean Real Fake t p d Realness 5. 77 4. 83 4. 8 <. 05 . 41 Picture Quality 5. 71 4. 77 5. 0 <. 05 . 43 Naturalness 5. 67 4. 8 4. 5 <. 05 . 38

Discussion • Greater than chance accuracy when the faces were real • No difference from chance when faces were artificial • Real faces rated higher than artificial faces on all additional questions • Effect sizes small to moderate

Experiment 2 • Using a face recognition paradigm: – Three suspects chosen to vary in similarity to perpetrator [3 perpetrators; sampling factor] – Five lineups generated for each suspect, of varying average foil-suspect similarity – Perpetrator present and perpetrator absent lineups used – i. e. Perp-suspect-sim X suspect-foil-sim

Example of high-suspect similarity lineup 15

Example of low-suspect similarity lineup 16

Exp 2, Method cont. • 150 student participants • Viewed target face (2 s) • Distractor task (3 mins) • Shown lineup and asked to choose person they saw, if present • Each participant viewed 3 lineups

Results

Experiment Three • Mock witness procedure used to assess bias and effective size of sub-sample of lineups used in Exp 2 [only looked at P-P lineups; needs to be re-done for all lineups] – Modal description generated for each of the 3 perpetrators from 6 ‘pilot’ witnesses – 125 student participants • Each presented with a description of target face 19 • Shown corresponding lineup and asked to choose

Results 20

Discussion • Higher levels of suspect-foil similarity result in greater fairness – protects against poor witnesses • But Study 2 shows high levels of suspect-foil similarity suppress Hits – this is a trade-off that we have to decide at a different level of discussion 21

Conclusions • The current research provides some promise for a lineup generator • Artificial faces pass an important test of ‘detectability’ • We are able to manipulate similarity of suspect to foils in predictable ways, and that makes me think we can address the optimal similarity problem

Fin