Facial recognition technology needs human intervention, hand coding, experts tell Ethics committee

Facial recognition technology (FRT) needs human oversight to avoid egregious errors in the legal system according to panelists who testified before the House of Commons’ Ethics committee as it studies the impacts of its use and impact. At a hearing Monday, University of York psychology professor Rob Jenkins told the committee that one of the benefits of having human intervention as part of the FRT is that glaring errors can be fished out before being acted upon. Jenkins noted that the United Kingdom’s Surveillance Camera Code of Practice states that FRTs “should always involve human intervention before decisions are taken that affect an individual adversely” – a model of human oversight which Australia is similarly adopting. “It’s important to have a human safeguard but the fact that human face recognition is not infallible also means that we should expect humans to introduce errors to those systems if they are given the decision and that’s the result of cognitive biases that we always carry with us,” he said. Facial recognition technology has already been under scrutiny in Canada. In Feb. 2021, British Columbia’s Office of the Information and Privacy Commissioner – in conjunction with the federal Office of the Privacy Commissioner (OPC) – found that software company Clearview AI engaged in “illegal mass surveillance” when it scraped some three billion images of faces and corresponding biometric identifiers off of publicly accessible information on the internet such as social media. Clearview AI’s software compares images with billions of pictures scraped from public websites on the internet, providing the searcher with information including a person’s name, address, phone number and occupation. Its database is reportedly used by more than 600 law enforcement agencies around the world. In 2020, some 31 civil society organizations called for a ban on law enforcement use of the technology, after the NDP called for a moratorium earlier in the year. Jenkins said that expectations of facial recognition technology’s accuray are formed based on people’s real-life experiences of facial recognition in everyday life. However, he argued that experience is misleading when it comes to security and forensic settings, because humans are only “extremely good” at identifying familiar faces whereas humans are “surprisingly bad” at recognizing faces generally. Security, witness testimony, and forensics involve the latter, he said. “It’s essential that we evaluate FRT in the context of unfamiliar face recognition by humans. This is partly because the current face recognition infrastructure relies on unfamiliar face recognition by humans, making human performance a relevant comparison,” he said. As an example, he said, in many security and forensic applications of FRT automated database search delivers a candidate list, but the final face identity decisions are made by human operators. Jenkins noted that human oversight does create an “upper limit” on what FRT could achieve because even if FRT were 100 per cent accurate, human error would be “baked in the system.” He said those errors can be mitigated by ensuring targeted recruitment of experts, evidence-based training, and the transparent disclosure of the strengths and limitations of the technology. Jenkins said the error rates of human facial recognition in unfamiliar scenarios, such as passport identification, is 10 per cent. However, there is “very little understanding” on the error rate of computer-based systems. The error-rate data for computer-based systems is based on how the technology operates in ideal situations – in which the conditions for both the image and the setting in which the match is being conducted are very good. Princeton University computer science graduate researcher Angelina Wang told the committee that facial recognition tasks are accomplished by a model that has been trained using machine learning rather than hand coded rules which is a “pertinent technical problem” that should prevent FRT from being deployed. With hand coded rules, she explained, models could be trained to learn that two people are likely to look the same if they have the same coloured eyes. Machine learning involves giving the model a very large dataset of faces with annotations and instructed to learn from it, such as images of the same person typically collected on platforms with homogenous working populations. She highlighted previous arguments made before the committee that these models can create racial and gender biases and said these challenges are theoretically surmountable. However, there still remain two problems that science cannot yet address – "brutalness" and interpretability. Wang said interpretability poses a problem because sometimes the engineers and researchers who built the technology do not understand how it makes its decisions, which can pose problems if an individual wants to contest a classification made about them. Brutalness, she said, refers to the ways the technology can break down or allow bad actors to circumvent the models, such as by putting a pair of glasses on a face to misidentify an individual. “FRTs are an incredibly cheap surveillance device to deploy and that makes it very dangerous because of how quickly it can proliferate. Our faces are such a central part of our identities and generally do not change over time so this kind of surveillance is very concerning,” Wang said. – Reporting by Jenna Cocullo at jcocullo@thewirereport.ca and editing by Michael Lee-Murphy at mleemurphy@thewirereport.ca.

Facial recognition technology needs human intervention, hand coding, experts tell Ethics committee

Broadcast | April 4, 2022

Unlock all the Canadian telecom, broadcasting and digital media news you need.

FREE TRIAL

SUBSCRIBE

Inside The Wire Report

Information

Hill Times Publishing

Get in Touch With Us

Facial recognition technology needs human intervention, hand coding, experts tell Ethics committee

Broadcast | April 4, 2022

Unlock all the Canadian telecom, broadcasting and digital media news you need.

FREE TRIAL

SUBSCRIBE

Related Stories

Federal government invests $7 million in high-tech firms in Ottawa

AI researchers call on governments to adopt safeguards ahead of elections

Four cyber bills dead on the order paper as Parliament prorogued

Feds launch AI compute strategy

Reuse & Permissions

Inside The Wire Report

Information

Hill Times Publishing

Get in Touch With Us