A recent study by the Commerce Department's National Institute of Safety Standards and Technology (NIST) found that facial-recognition technology is more likely to error on African American, Asian and Native American subjects than it is on Caucasians.
The results are likely to boost concerns among privacy advocates about bias in facial-recognition systems deployed at U.S airports and other points of entry and exit.
To date, U.S. Customs and Border Protection has introduced biometric facial-recognition technology for U.S. entry, exit or both at 25 U.S. airports. The biometric kiosks are deployed at gates of international flights per Congressional mandate in order to advance the federal government's goal of more thoroughly tracking visa overstays by foreign nationals. At those gates, photos are taken of flyers and then matched to passport photos that the Department of Homeland Security keeps on file. U.S. citizens can opt out of the biometric identity check and instead have their passport manually verified by a gate agent, but visitors must participate in the scans.
The NIST study, which was unveiled on Dec. 19, evaluated 189 software algorithms from 99 developers. For the study, the NIST used 18.3 million images of 8.5 million people. The photographs were provided by the State Department, the Department of Homeland Security and the FBI.
For one-to-one matching, in which the software confirms one photo matches a different photo of the same person in a database, false positives were 10 to 100 times more likely for African Americans and Asians than for Caucasians. Among U.S.-developed algorithms, false positives were even more common for Native Americans.
For Asians, however, the dramatic differences compared with Caucasians disappeared on Asian-developed algorithms. That, said Patrick Grother, the report's primary author, suggests there is a connection between an algorithm's performance and the data used to train it.
"These results are an encouraging sign that more diverse training data may produce more equitable outcomes, should it be possible for developers to use such data," Grother said.
Arguably more alarming than the false positive disparities in one-to-one matches was the study's finding relating to African American women in one-to-many matches, in which it is determined whether the person in a photo has a match in a database, such as an FBI database. The NIST team found higher rates of one-to-many false positives for African American women than other demographics, an issue that could make them more vulnerable to false accusations.
Still, the NIST emphasized that various algorithms performed differently.
"Such distinctions are important to remember as the world confronts the broader implications of face-recognition technology's use," the NIST said.