International. A sample of commercial facial recognition algorithms seems to suggest that facial biometric systems currently rely too heavily on certain attributes to just as easily differentiate between different people who share race and gender categories.
Biometric researchers from the Maryland Test Facility (MdTF) and the U.S. Department of Homeland Security's Directorate of Science and Technology have published a paper exploring the problem with five commercial algorithms of unnamed facial recognition, with a commercial algorithm of iris recognition as a control.
The test was conducted by John Howard, Yevgeniy Sirotin and Jerry Tipton of MdTF and Arun Vemury of S&T.
His paper on "Quantifying the Degree to which Race and Gender Characteristics Determine Identity in Commercial Facial Recognition Algorithms," suggests that those facial recognition algorithms that look for large databases of homogeneous images tend to result in disparate treatment based on race and gender. The research builds on the scientist's previous work, which showed that images of people from the same demographic tend to score more similar to each other than when compared to people from a different demographic, a phenomenon they call "broad homogeneity."
Analysis of the main components used by biometric algorithms shows that most variations between vectors or facial appearance are not related to race or gender, although about 10 percent are. This means that it should be possible for algorithms to more consistently avoid offering unfair precision differentials.
"In addition, the separation between paired and unpaired score distributions reconstructed exclusively with PCs (main components) that do not group individuals by race and gender was reduced only modestly, suggesting that CFRAs (commercial facial recognition algorithms) can maintain acceptable performance even when ignoring face characteristics associated with race and gender," write the authors of the report.
Components without significant grouping based on demographic data account for 62 percent of the total score variation for facial recognition algorithms.
The researchers also cite recent research suggesting that demographic attributes can be removed from images and still be effective at matching facial recognition algorithms.
Human review may help, the researchers suggest, but ultimately, the technology must continue to evolve.
"The development of demographically blinding CFRAs that explicitly ignore facial features associated with race and gender will help maintain equity as the use of this technology increases," the report says. "We believe that developing such algorithms and demonstrating fairness, including shrinking the demographic pool, should be a focus of attention for companies selling facial recognition technology."
ID4Africa CEO Dr. Joseph Atick has asked Facebook to use its resources to help resolve any questions about whether facial recognition algorithms can be developed to meet high standards of accuracy without a significant difference in performance with subjects based on race and gender.
The full report can be read here.
