Wearing a mask no longer protects against facial recognition
An engineer from the Japanese group NEC tests a system allowing facial recognition despite wearing a mask, in Tokyo, January 6, 2021. REUTERS - KIM KYUNG-HOON
Text by: Nicolas Sanders Follow
7 min
This is an unexpected effect of the Covid-19 pandemic.
With the generalization of the wearing of a mask, facial recognition systems have become panicked.
Difficult to identify a quidam when three-quarters of his face are hidden.
But the sector quickly got up to speed with new algorithms.
The health crisis is even creating new markets.
Publicity
Read more
In July 2020, the masks did not only protect against Covid-19, but also against Artificial Intelligence.
At that time, a study by the National Institute of Standards and Technology (NIST) found that the algorithms of facial recognition programs were not sophisticated enough to identify people. who wore a mask, whatever its standard or shape, both paper and cloth.
The work of NIST is authoritative in this area.
The FRVT (Face Recognition Vendor Test, test between facial recognition solutions) is the benchmark in which any company wishing to exist in this still nascent market must appear.
According to
the results compiled by the American agency
from a hundred or so software programs among those most often used in the world, their error rate climbed from 5% to 50% when wearing the anti-Covid mask.
Without the mask, the error rate of facial recognition solutions is around 0.3%.
The study authors also noted that the higher the mask was worn on the face, the less reliable the algorithm.
Ditto with black masks, which generate more errors than blue or white surgical masks.
Most specialists agree that there is ultimately nothing surprising about this.
A facial recognition algorithm works by measuring the distance between three essential features which are the eyes, nose and mouth.
By concealing these last two landmarks, the mask which occupies about 70% of the face greatly complicates the measurement.
Outsmart an algorithm with a piece of paper
Would a pair of ski goggles and a piece of paper be enough to thwart the very high technicality of Artificial Intelligence algorithms?
One main factor in NIST's assessment was the rate of false mismatches.
That is, the number of matching faces that escape the algorithm.
But according to the very latest study released by NIST last November, the coronavirus has stung developers of facial recognition software.
"
Some developers submitted algorithms after the pandemic showing significantly improved accuracy and are now among the most accurate in our test,
" say the researchers.
And for good reason, the error rate would have now returned to a much lower level, a little below 5%.
In a highly competitive context - with the key to very lucrative markets - where each company is competing to offer the most efficient software, the players in facial recognition have from the start of the pandemic focused their attention on the means to overcome the generalized wearing of the mask.
"
The health crisis is a life-size laboratory for facial recognition allowing artificial intelligences to improve themselves
" analyzes Olivier Tesquet, journalist at Télérama, specialist in new modes of surveillance and author of the book
"State of technological emergency"
to be published on 4 February.
This is confirmed by Laurent Lepetit, head of technology development at id3, a French company which has developed various identity verification products: "
All our teams have been diverted from their usual tasks to move forward on new software
".
Like some of its competitors, id3 has managed in a few months to develop a solution allowing facial recognition despite wearing a mask.
First, the system identifies whether or not the person has a mask, explains Laurent Lepetit.
If this is the case, dedicated software is implemented to verify the identity, "
a peri-ocular system which takes into account only the visible part of the face, namely the eyes
".
"We are moving from a safe market to a sanitary market"
Detecting whether people are wearing a mask is not in itself the objective of the service sold by id3, but other market players have not hesitated to take a position in this niche.
This is the case of CyberLink, a Taiwanese company present around the world with a solution called FaceMe.
“
In October 2020, we launched FaceMe Health, an independent facial recognition module that verifies the wearing of the mask, measures the subject's body temperature and identifies them,
” explains Lara Gerhard, the company's marketing manager.
"
What is very striking with the health crisis,
Judge Olivier Tesquet,
is that companies yesterday already well established in secure markets are creating new products sold with other purposes, we are moving from a secure market to a market sanitary
”.
The terrorist threat, often cited to legitimize facial recognition systems, has been joined by the threat of the virus.
"
We are sliding from security, the first of freedoms which is a very powerful lever for generalizing and trivializing equipment in surveillance technology, towards the first health of freedoms which is an even more powerful lever to trivialize these tools
" continues Olivier Tesquet .
Currently, in France as in Europe,
facial recognition practiced in real time
on public roads is not authorized.
Only a posteriori recognition has been granted since 2012, in particular by the police, with the criminal record processing file (TAJ).
But whether it is authentication or identification, it is clear that facial recognition is a system that is still developing under the regime of experimentation.
"
Local authorities have the opportunity to try experiments, to try their luck, with behind the possibility that a court or the CNIL will come and tell them to stop everything
" warns Olivier Tesquet.
Israeli software to identify S files
Indeed, if it was previously mandatory to receive the prior consent of the CNIL to launch this type of experiment, it is no longer necessary.
The commission intervenes only in appeal.
An experiment was thus carried out during the 2019 edition of the Nice Carnival with Israeli software designed to identify supposed S files. In the context of this test, the participants were all volunteers.
During the first confinement in March and April 2020, the Cannes city hall used the services of the company Datakalab to check whether or not the inhabitants wore a mask on the public highway.
An “experiment” also carried out in the capital by the same company in the RER station Châtelet-Les Halles.
The stated objective was to count people with a mask, and not to identify them formally, but the CNIL put a stop to the Parisian experiment.
Facial recognition systems allow data to be processed remotely and potentially without the data subject being informed.
Because this "contactless" technology raises unprecedented questions relating to social choices, the CNIL called in 2018 for a major democratic debate on this subject.
Absent from the recent bill on global security adopted by the Assembly last November, facial recognition in France will have to find its legislative framework if possible before 2024 and the Olympic Games organized in Paris.
This is because the stakes are high, with a market valued at several billion euros.
Newsletter
Receive all the international news directly in your mailbox
I subscribe
Follow all the international news by downloading the RFI application
google-play-badge_FR
Digital
On the same subject
United States: in Boston, the reliability of facial recognition questioned
International report
China: facial recognition worries civil society
United States: after IBM, Amazon bans police from using facial recognition