Substantial improvements to biometric spoof attack detection possible in short term
Stergios Papadakis from The Johns Hopkins University Applied Physics Laboratory delivered a presentation during the International Face Performance Conference (IFPC) on Wednesday, revealing the results of the Odin program’s first three years of attempting to defeat biometric spoof attacks.
Part of the test and evaluation team at IARPA’s Odin project, Papadakis started his presentation by describing the goal of the program.
Odin aims at developing biometric presentation attack detection (PAD) technologies to enable biometric security systems to detect when someone is attempting to disguise their biometric identity.
The program pillars are essentially three. The first one relates to the ability of these technologies to detect both known and unknown attacks.
They should also have the ability to operate at relevant true/false detections rates (0.2 percent), and they should have biometric recognition capabilities at the level of existing technologies.
Three years of biometrics’ testing
According to Papadakis, the first phase of the Odin program saw a number of government-controlled tests (GCT). It was dated May 2018 and analyzed 392 unique subjects, together with 601 presentation attack instruments and 162,442 biometric trials.
The second phase, conducted in May 2019, saw 382 unique subjects, 508 presentation attack instruments, and 156,312 biometric trials.
The following phase, whose trials ran in November 2019, saw 699 unique subjects, 977 presentation attack instruments, and 244,789 biometric trials.
According to Papadakis, the COVID pandemic has slowed down planned testing for 2020 but estimated unique subjects for this year are 2000, together with 300 presentation attack instruments.
The scientist continued its webinar by mentioning that all presentation attacks tested in the Odin program are described in the open literature, including academic journals, press releases, hacker sites, and YouTube.
Different biometrics equals different PADs
In order to test as many combinations of biometric PAD technologies as possible, the Odin team tested a variety of polymeric finger PAs.
To test face PAs, various masks were used, as well as different degrees of makeup provided by make-up artists who were specifically trained to challenge the algorithms.
Eye PAs were the most challenging to simulate, Papadakis explained, and the Odin algorithms were trained with fake, plastic eyes, as well as cosmetic contact lenses.
According to the researcher, conclusions from three years of development have demonstrated that substantial improvements in presentation attack detection are feasible in the short term.
Also, the experiment has proven that makeup PAs can be a challenge for systems with low false matching rates, which could be fooled by individuals wearing makeup without any intention of carrying out a presentation attack.
Moving forward, Papadakis said the Odin program will continue, trying to identify demographic correlations with PA detection, and examining whether set conclusions can be inferred by specific attackers’ methods.
More information about the Odin program can be found at this link here.
biometric identification | biometric liveness detection | biometrics | biometrics research | facial recognition | fingerprint recognition | IARPA | International Face Performance Conference (IFPC) | iris recognition | spoof detection