Read on Twitter
Thread:
Today, NIST released a study on facial recognition of masked faces, finding face masks increased error rates from as low as 0.3% to between 5% & 50%. More usefully, NIST's study provides some handy tips on how to make FR less likely to ID you. https://www.nist.gov/news-events/news/2020/07/nist-launches-investigation-face-masks-effect-face-recognition-software 1/
Today, NIST released a study on facial recognition of masked faces, finding face masks increased error rates from as low as 0.3% to between 5% & 50%. More usefully, NIST's study provides some handy tips on how to make FR less likely to ID you. https://www.nist.gov/news-events/news/2020/07/nist-launches-investigation-face-masks-effect-face-recognition-software 1/
Critically, NIST's 1-to-1 match study found only a modest change to false positives (incorrectly ID as someone else) and a large increase in false no-matches, either because FR couldn’t find any face or found no match (a false negative). So how can we get more no-matches? 2/
Tip 1: Cover that nose! The more of the nose a mask covers, the lower the algorithm’s accuracy. Masks that reach near the eyes have a false non-match rate 36 times higher for the median algorithm. And a nose bridge can also help reduce slippage and glasses fogging. Win-win! 3/
The NIST study found the low position (mask below the nose) allows significantly more FR accuracy than a high position nose covering mask. And it is less effective at protecting against the spread of coronavirus. 4/
Tip 2: Go wide! Full-face-width masks give false negative rates two times higher than rounder masks (typical with many N95s). If you need both N95 and FR protection, consider adding a stylish bandana on top of your round N95. 5/
Tip 3: Go black! Most algorithms have higher error rates in black masks over light-blue masks (typical for a surgical mask). Plus black matches with every outfit. 6/
NIST did not test textured or decorated masks, but notes a potential tradeoff: Decorated masks may induce higher facial localization errors, but could also serve as a "soft biometric.” Even a full mask can become identifiable if unique and worn consistently. 7/
NIST also didn't look at eye-protection, like glasses and goggles, but probably it doesn’t make FR any easier. When exercising your 1st A rights, goggles can help protect your eyes from both covid and tear gas, so you might want them anyway. And a balaclava. 8/
Note NIST studied 89 existing FR algorithms, but not new ones designed for masks. NIST’s next report will evaluate the performance of 'mask-enabled’ algorithms, and begin to look at accuracy for different demographic groups (a big problem with FR). https://www.wsj.com/articles/facial-recognition-software-suffers-from-racial-bias-u-s-study-finds-11576807304 9/
To be clear, presenting these as “tips" is my spin. NIST is studying facial recognition effectiveness “in collaboration with” DHS and CBP, who see masks as a problem to be solved, not as a protection for your civil liberties. https://venturebeat.com/2020/02/06/u-s-homeland-security-has-used-facial-recognition-on-over-43-7-million-people/ 10/
To learn more about facial recognition, check out @EFF’s Street-level Surveillance project. https://www.eff.org/pages/face-recognition -fin-
