80% accuracy in facial recognition database is a positive match for Delhi Police: Report

In a test by the ACLU, 80% accuracy was considered a mixed result, with several US Congressmen falsely matching in criminal mugshots

An 80 per cent accuracy in facial recognition databases was considered a positive match by the Delhi Police, which deployed the technology to identify accused in major clashes in the city, RTI records show. The RTI records, obtained by the Internet Freedom Foundation and reported by The Indian Express, showed that the police conduct “empirical investigation” for facial matches that have an accuracy of over 80 per cent before starting legal action.

In cases where the accuracy is below 80 per cent, the police consider it a “false positive” which is subject to “due verification with other corroborative evidence”, the records show.

Read | AI is using fake data to learn to be less racist

The police has used this technology to identify suspects in the 2020 riots, the clashes at Red Fort in 2021 during the farmers’ protest and the Jahangirpuri riots earlier this year, the report said.

Experts raised concern over the benchmark used by the police for legal action on the basis of database matches. “Even if the technology does not give a sufficient enough result, the Delhi Police could continue to investigate anyone who may have gotten a very low score. Thus, any person who looks even slightly similar could end up being targeted, which could result in targeting communities who have been historically targeted,” Anushka Jain, associate counsel at IFF said.

In the international scene, 80 per cent accuracy was considered a mixed result, with a test in 2018 by the American Civil Liberties Union (ACLU) on Amazon’s Rekognition system, falsely associating 28 images of US Congress members with mugshots in a criminal database.

In response, Amazon had said that “while 80 per cent confidence is an acceptable threshold for photos of hot dogs, chairs, animals or other social media use cases, it wouldn’t be appropriate for identifying individuals with a reasonable level of certainty”.

Activists contend that the usage of technologies like facial recognition in the absence of a basic data protection framework could potentially pose risks for communities that have been historically targeted. The police also declined to reveal whether they have made any arrests of alleged perpetrators using the technology, citing a section of the RTI Act that exempts them from disclosing information that could impede investigations, the report said.