Facial recognition in security and the problem of bias

Yet the technology has consistently sparked controversy – previously, we have seen how its use in schools has opened up debate around its ethics, particularly in how its data is stored. Here, we wanted to take a look at another key consideration surrounding the subject – that of bias.

The prejudice problem

When implementing any new technology it’s important to understand the limitations it may have. In the case of facial recognition, the question of racial prejudice is for many a point of concern.

Previous research into the tech actually led to the discovery of inaccurate readings – largely targeting ethnic minority groups. A prime example is Rekognition, Amazon’s image and facial recognition service, which wrongly identified 28 members of Congress as criminals, with the majority of these false matches being people of colour.

Similarly, a study by the National Institute of Science and Technology found that facial recognition systems misidentified minority groups up to 100 times more than Caucasian males. The problem is also seen in the matter of identifying people’s gender, with a 2019 study revealing facial-recognition  is less effective on women, and WatchGuard research showing the tech misidentified women 18% more often than men.

Such examples showcase the limitations of certain AI technologies, and reinforces the need to understand equipment before deploying it in an operational environment, as well as ensure relevant training and procedures have been implemented addressing such limitations. So, user beware…manufacturers are not overly incentivised to highlight limitations or error rates with their products, leaving the responsibility with the purchaser to suitably inform themselves.

How can we address this?

While facial recognition technology can be an incredible tool, it is not perfect.  Whether intentionally or not, an over-reliance on it can actually detract from existing security and investigative practices required for a holistic security outcome.

Furthermore, organisations implementing this technology need to do so with careful consideration to the outcomes, including unintended consequences – as opposed to hasty, ubiquitous deployment based on a superficial understanding of the tech and potential outcomes.

Transparent and ethical reporting of product limitations would be a significant step in the right direction – though companies may be hesitant to disclose this information. Regulations over providing these details could be a viable solution, and some have suggested governments oversee the datasets used to build these tools – working to ensure they’re also representative of minority groups.

This idea is already being seen in action, with the EU recently proposing a regulatory framework for high-risk AI systems such as facial recognition to reflect ‘all relevant dimensions of gender, ethnicity and other possible grounds of prohibited discrimination’.

Important steps are already being taken to safeguard the technology against bias, and there is optimism that as it continues to develop, instances of prejudice will eventually be eliminated altogether – though we must always assume a level of error in any technology that we use.

As stated earlier, the key with any technology is to understand its capabilities and limitations, as well as its appropriateness for the intended application and whether it’s supported by suitable training, operating procedures and oversight.

With all this in place, there is little stopping facial recognition from becoming a safe and effective part of our security systems – we just need to ensure ethical practice is maintained as the technology becomes a more common part of our lives.