Adding morph image to anti-fraud tool defends against morphed images

Two researchers say they have achieved “promising results” in one-upping the triplet-loss technique used to spot fraudulent face biometrics with a quadruplet-loss process.

The scientists, both working in Portugal, added a fourth image to the conventional three-image comparison technique used to spot facial recognition exploits including morphing. Their work has been posted on the Arxiv pre-peer review server.

Face morphing poses a significant threat to face biometrics, according to NIST computer scientist Mei Ngan during a presentation at the 2020 International Face Performance Conference.

Morphing software is known to inadvertently throw telltale image artefacts in images, but the flaws are being reduced or eliminated with manual post-processing.

In addition to proposed quadruplet-loss, anti-morphing efforts are evolving, too.

For the uninitiated, a triplet-loss process automatically compares three images – the anchor image, a positive image (a different image of the same face, for example) and a negative image. The negative image is clearly not related to the anchor image.

The AI algorithm succeeds by figuratively minimizing the distance between the anchor and positive and maximizing the distance between the negative and the first two image.

The quadruplet-loss process adds, not unexpectedly, a fourth facial image. The new image is a morph of the other images.

The researchers say they are studying how their technique, which they call a benchmark utility, with other facial recognition software and networks.

Article Topics

biometrics  |  biometrics research  |  face biometrics  |  face morphing  |  fraud prevention