Police facial recognition biased against certain groups

Study ‘sheds light on a concerning in-built bias’ says APCC

A Home Office study by on the facial recognition system used by UK police forces has found that it is significantly less reliable with certain ethnic groups and genders.

The study of the Police National Database’s retrospective facial recognition (FR) technology tool was carried out by National Physical Laboratory (NPL) on behalf of the Home Office and the Office of the Policing Chief Scientific Adviser.

Retrospective facial recognition uses photos taken by CCTV, for example, and attempts to match them with existing images in a database.

NPL researchers compared images of faces with 180,000 images contained in the a police database of custody photos using the Cognitec FaceVACS-DBScan ID v5.5 facial recognition (FR) system, checking for accuracy and false positives.

They tested the system’s performance on two key indicators.

The researchers also looked at rankings of possible facial matches produced by the system under different threshold settings, and studied differences in results for white, Asian (South and East Asian) and black groups.

At a face-match threshold setting of 0.9, TPIR was significantly higher for Asian subjects (97.9%) compared to white subjects (91.3%) and black subjects (86.9%). In other words, Asian people in the database were most likely to have their identities matched.

At a slightly lower threshold setting, FPIR was significantly lower for white subjects (0.04%) compared to Asian (4.0%) and black subjects (5.5%). In other words, black subjects were most likely to be falsely identified as being present in the police database.

There were differences within the categories too. black Male subjects had a lower FPIR (0.4%) than black female subjects (9.9%), and older subjects (age 42-76) had a lower FPIR (0.1%) compared to younger age groups.

The findings suggest that a person’s likelihood to appear on a list of potential suspects depends on their ethnicity and gender.

The Association of Police and Crime Commissioners (APCC), said the study calls into question the rollout of facial recognition by police forces, commenting that it “sheds light on a concerning in-built bias.”

“This has meant that, in some circumstances, it is more likely to incorrectly match black and Asian people than their white counterparts. The language is technical but, behind the detail, it seems clear that technology has been deployed into operational policing without adequate safeguards in place.”

The APCC acknowledged that work was ongoing to improve the systems, and that they could be valuable in tackling crime, but said that given their intrusive nature more thorough assessment was required.

“These technologies are increasingly invasive and sophisticated. If they are to gain the trust and support of the public, particularly those groups historically mistrustful of the police, then full transparency is vital.”

Emily Keaney, deputy commissioner for data watchdog the Information Commissioner's Office (ICO), called for “urgent clarity on this matter so we can assess the situation and consider our next steps.”

Earlier this year, the Ada Lovelace Institute and privacy groups called on the government to urgently tighten regulation around facial recognition technology, amid growing concerns over its unregulated use by police forces and private companies across the UK.