It can be difficult to tell if a pair of unfamiliar faces match, presenting potential opportunities for identity fraud. One new form of fraud exploits this with fake IDs that contain a photo produced by combining images of two different faces – this “face morph” may look sufficiently like these two people to serve as ID for both.
New research highlights the cunning of face morph IDs, suggesting that both humans and smartphone software find it difficult to distinguish face morph images from photos of the two faces contributing to the morph.
David Robertson from the University of York and colleagues challenged both human viewers and smartphone face recognition software to identify face morphs as distinct from the two faces combined in the morph. Both were required to detect whether a pair of photos matched or not, with the pair sometimes comprising the morph along with one of the contributing faces.
They found that we are surprisingly bad at identifying face morphs. Initially, human viewers were incapable of distinguishing a 50-50 morph photo from its contributing photos a “worryingly high” 68 percent of the time. However, after simply briefing the viewers to look out for manipulated, “fraudulent” images, the error rate dropped greatly to 21 percent. Meanwhile, the smartphone software achieved similar results to briefed human viewers, with an error rate of 27 percent. The authors note, however, that these rates are still significantly higher than the usual error rates when comparing two photos of entirely different people.
Participants in the study were unlikely to be as motivated, as trained or as skilled at spotting fraudulent photos as real world professionals, and real criminals could use more sophisticated morphing techniques. Nonetheless, the findings indicate that humans and smartphones may not be naturally adept at identifying face morphs, a weakness that could be exploited by fraudsters.
Research Article: Robertson DJ, Kramer RSS, Burton AM (2017) Fraudulent ID using face morphs: Experiments on human and automatic recognition. PLoS ONE 12(3): e0173319. doi:10.1371/journal.pone.0173319
Image Credit: A.M. Burton, CC-BY 4.0