Home Affairs thankful Australia's diversity allows for improved facial recognition

Armed with multiple algorithms and training images based on a multicultural society, Home Affairs believes it gets a bronze medal in facial recognition.
Written by Chris Duckett, Contributor

When Australia's Department of Home Affairs is training its facial recognition systems, it is using multiple algorithms and a diverse collection of imagery to avoid any bias in the system.

"We tune our algorithms all the time against a wide-ranging algorithm, we are very fortunate in Australia that we are a multicultural society, we get people coming in and out of the country from various nationalities," Acting Deputy Secretary of Intelligence and Capability at the department Joe Franzi told Senate Estimates on Tuesday.

"We don't use just one algorithm, we use a number of algorithms."

Franzi said the size of the data set used to train the system was key to avoiding bias, and cited the case of an unnamed Chinese system that performed well against Asian faces but struggled with people of other races.

Based on working with the US National Institute of Standards and Technology, Franzi claimed Home Affairs' algorithms are in the top three.

Responding to figures from South Wales Police that showed its NEC facial recognition system had produced a high number of false positives at the 2017 UEFA Champions League Final, Franzi said many systems struggle with situations like stadiums and crowded events.

"Some of these mass surveillance-type facial recognition solutions are very much in the development phase -- but people get starry eyed and they buy them," he said.

"They perform really well when you go to the conference ... and it works really well in a controlled environment, but once you actually move into a real world environment ... then you get all the issues around different lighting, angles of faces, where are cameras, do people have hats on, glasses on, [hoodies] -- a whole range of things."

Home Affairs has been using facial recognition in a Canberra trial to allow passengers to walk through the terminal from their flight without producing their passport.

"For the numbers of people coming through our airports, I want them to walk seamlessly down -- off the A380 -- and, in time, and we're not far off this, with facial recognition on the move, people's passports will stay in their pocket," Minister for Home Affairs Peter Dutton said in February.

The department is also using facial recognition in its face-matching system that will link up databases of photos stored in various federal and state agencies.

Last month, Home Affairs dismissed calls for the system to require a warrant to access, stating the requirement would "effectively prevent government agencies from using the services, or obtaining the benefits of the services".

"Obtaining a warrant is a resource-intensive process, both for the applicant agency and for the issuing authority hearing the application," Home Affairs wrote.

"The time involved in preparing, reviewing, and granting a warrant application to use services would significantly delay, and in some circumstances undermine, law-enforcement and national security investigations; impede operational activity, including the prevention of criminal acts; and divert resources from investigations."

Earlier this year, the department claimed that its "hub-and-spoke" topology for the face-matching service was helpful in preventing breaches, as well as describing a "moat" surrounding its gateways.

On Monday, the department revealed it had spent AU$5.5 million so far from its AU$10 million budget on integration work in creating the Dutton-led enforcement superministry.

"There were some IT issues that were a function of the fact that government agencies are regrettably -- and this is longstanding; it goes back 20 years -- still on different networks," departmental secretary Michael Pezzullo said.

"When you federate two systems, you've got to set up Citrix and other connection arrangements. That's not particularly specific to Home Affairs; it happens whenever you have a MOG -- a machinery-of-government change."

Related Coverage

Editorial standards