Facial recognition software has a higher rate of incorrect matches between two photos for Asian and black people relative to white people, a United States government study has found.
The evidence of bias against minorities in the software comes as its use is set to expand at airport security checkpoints in Asia, Europe and the United States.
The US Transportation Security Administration (TSA) and US Customs and Border Protection (CBP) have been testing facial recognition technology at airports across the US, expecting it will become the preferred method to verify a passenger’s identity.
The National Institute of Standards and Technology researchers studied the performance of 189 algorithms from 99 manufacturers representing most of the industry. Some algorithms performed better than others, they concluded, meaning that it’s likely the industry can correct the problems.
The institute found that US-developed algorithms had the highest rate of incorrect matches, or false positives, for American Indians.
Researchers found a higher rate of false identifications of black women when matching their photos to an FBI database. Higher rates of mismatches increase the chance that a person could be falsely accused, the institute said.
Patrick Grother, a computer scientist at the institute and the report’s primary author, said some mismatches can be resolved with a second attempt, such as checking a passport. However, he said, “a false positive in a one-to-many search puts an incorrect match on a list of candidates that warrant further scrutiny”.
The TSA has been testing facial recognition software at airport checkpoints in Los Angeles and Las Vegas in the United States, matching live images of passengers to the photos on their identity documents. Participation in those test runs was strictly voluntary.
“TSA’s facial recognition system will be for passenger identification and to determine the appropriate level of screening only,” spokesman Mark Howell said. “TSA understands the variety of concerns related to facial recognition match performance and takes this issue seriously.”
CBP uses the technology to screen arriving international passengers at 16 US airports and exiting international passengers at 26 US airports. Additionally, CBP uses it for entry at cruise ports in New York, New Jersey, Florida and Washington state.
Both agencies plan to expand the use of the technology to screen passengers. It could speed up the process, allowing for shorter waiting times for passengers and enabling security personnel to better focus their attention.
“Facial recognition is going to replace that interaction with the TSA officer,” said Brian Jackson, a security researcher at the Rand Corporation, a policy research organisation.
Grother’s team used an assortment of more than 18 million images of about 8.5 million people culled from files of the US State Department, the US Department of Homeland Security and the FBI.
The research adds to concern about the accuracy of the technology, which has potential implications beyond airport security.
Last year, the American Civil Liberties Union used software developed by Amazon to match photos of all 535 members of US Congress against a database of 25,000 publicly available arrest photos.
The software incorrectly matched 28 lawmakers with photos of people who had been arrested. Nearly 40 per cent of those false matches were non-white lawmakers, though minorities make up only 20 per cent of Congress.
Patricia Cogswell, the TSA’s acting deputy administrator, told US lawmakers in October that the agency wanted “a very highly probable match” in its developing biometric screening process.
“We are not matching you against mugshots,” she told a US House of Representatives subcommittee on homeland security. “If you don’t match, we go back to the regular process.”
Though there may be lingering concerns about the accuracy of facial recognition technology, travellers may decide it’s worth the trade-off if they can save a few minutes.
“Opting out will always have a time cost,” Jackson said. “There’s certainly a convenience benefit to it.”
Push will get a person almost anywhere- except through a door marked “pull.”