UK facial recognition bias report reveals Home Office using outdated Cognitec software

Cognitec released its FaceVACS-DBScan 5.5 software for biometrics matching at scale in 2020. A biometric accuracy and bias assessment of the software by the UK’s National Physical Laboratory for the Home Office was released last Thursday, with worrying results.

But facial recognition has evolved substantially since 2020. At the time, the biometrics industry was grappling with the revelations spurred by Joy Buolamwini’s Gender Shades, which prompted NIST to launch its assessments of demographic differentials at the tail end of 2019. And Cognitec’s face biometrics software has since been updated several times. The current version is 5.9, but Home Office’s Police National Database uses 5.5, which is why that version was tested.

Cognitec Marketing Manager Elke Oberg told Biometric Update in an email that multiple upgrades since the release of FaceVACS-DBScan 5.5 have improved its matching accuracy and significantly minimized bias.

The NPL evaluation showed statistically significant differentials between several demographics, including higher error rates for women, older people and those with darker skin. The Information Commissioner’s Office and Association of Police and Crime Commissioners have concerns about the facial recognition system’s governance.

FaceVACS-DBScan ID 5.9 uses the B16 algorithm, while version 5.5 used the B12 algorithm. Oberg notes in an email that the face finding technology used in the older algorithm was completely replaced in the B14 update.

Cognitec has also submitted several algorithms for evaluation by the U.S. National Institute of Standards and Technology. The latest algorithm from the company submitted to NIST FRTE 1:N and FRTE Demographic Effects is Cognitec_007.

Oberg describes that as a C series algorithm, explaining “We develop A,B, and C algorithms that differ in template size, speed and memory footprint, catering to different use cases and platforms.”

That means that “In regards to accuracy and bias measures, cognitec_007 would be somewhere in between B12 and B16. The B16 algorithm used in FaceVACS-DBScan ID 5.9 in 2025 is a further upgrade from the cognitec_007 algorithm.”

Perhaps most damning for the Home Office is the presence of all past Cognitec submissions to the Demographic Effects evaluation, showing clearly the gains that have been made since half a decade ago. Between a 2018 update and the submission of cognitec_007 in 2024, the company’s false match rate for the demographic group with the most errors, West African females 65 years and over, fell from 1.9 percent to 0.5 percent.

The police force in New South Wales, Australia discontinued use of a Cognitec algorithm earlier this year that it licensed in 2011 and never updated. No official reason was given for the change, giving rise to speculation that it was due to the legacy algorithm’s demographic differentials in performance or cost. NSW Police opted for an in-house adaptation of open-source software instead.

The time may have arrived for facial recognition vendors or regulators to insist on contractual time frames for the use of their technology.

Article Topics

algorithms  |  biometric bias  |  biometric matching  |  biometrics  |  Cognitec  |  facial recognition  |  Home Office Biometrics (HOB)  |  National Physical Laboratory