There was first Alicem, a small smartphone application presented by the Ministry of the Interior during the summer, and an interview granted to the World by Cédric O, Secretary of State Digital, Monday, October 14. Two events that suggest that the government is looking to move up a gear in facial recognition in France.

Cédric O, in fact, announced that he wanted to create an instance of "supervision and evaluation" of the possible uses of facial recognition. He also called for more experiments in this area to help manufacturers to make their solutions more reliable and efficient.

The promises of facial recognition

With Alicem, the Ministry of the Interior is no longer experimenting. The services of Christophe Castaner would be ready in November, according to Bloomberg, to deploy this application that will allow holders of a biometric passport or an electronic residence permit to create a digital identity and identify themselves on the Internet thanks to to facial recognition. France would thus become the first European country to offer such a service.

In the eyes of privacy advocates, Alicem is a first (small) step towards a Chinese-style system, where facial recognition is ubiquitous. Even the National Commission for Informatics and Liberties (Cnil) has expressed strong reservations about this application which, under cover of simplification of the administrative procedures, raises questions about the protection of the biometric data thus collected and their use.

More generally, associations like La Quadrature du Net denounce a climate in France where the authorities would be quick to succumb to the sirens of facial recognition without sufficiently taking into account the ethical problems raised by this technology. In February 2019, Valérie Pécresse, president of the Île-de-France region, advocated the use of facial recognition in public transport to fight "against the terrorist threat". Police would like to be able to use algorithms to identify people quickly in huge databases such as the file of wanted people, reports Le Monde. As for the mayor of Nice, Christian Estrosi, he does not miss an opportunity to declare his love for facial recognition.

80 % effective ?

But beyond ethical issues, the current craze for facial recognition may also seem technologically premature. In 2017, London police conducted a large-scale experiment to test the reliability of "intelligent video surveillance" at the Notting Hill Carnival. The failure was resounding: the machine had resulted in the arrest of an innocent person and was wrong on 35 occasions. "But that was two years ago, and in the meantime technology has improved a lot," said Josh Davis, facial recognition expert at the University of Greenwich in London, who was contacted by France 24. He noted that Last year, a new test was conducted with slightly more convincing results.

In France, experts believe that facial recognition should be 80% effective to be taken seriously, says Le Monde. "We're not there yet, we have to do more experiments," says a police officer interviewed by the newspaper.

But even this figure is to be taken with tweezers. "It all depends on the context, 80% is not enough if you have to check every report," says Jean-Luc Dugelay, a teacher at the Eurocom engineering school and a specialist in image processing. He gives the example of the airport where tens of thousands of people are likely to go through the mill of a facial recognition system. Even with a reliability rate of 80%, it still leaves thousands of cases where the alarm may be triggered wrongly, or otherwise not at all.

Behind this percentage often cited as an example, there are also very different realities. "When conditions are under control, facial recognition works very well," notes Jean-Luc Dugelay. That is, if parameters such as lighting, position and distance to the camera, or any makeup, are similar between the time of the registration of the face and its verification, the algorithms are hard to fool. For example, there is no point in growing a beard or wearing sunglasses.

Database security

But "it becomes more complicated when it comes to video surveillance biometrics," said the French expert. The comparison is more hazardous between photos in a file and images captured in the street where people can have a part of their face hidden, to be too far from the eye of Big Brother or in a poorly lit place. "We also know for example that in the current state, most facial recognition systems are struggling when the photo used for identification is more than six years old," says Davis.

Another limitation is the training to which the algorithms were submitted before taking the action. "Facial recognition is nothing without a database of images," says Jean-Luc Dugelay. Millions of faces must be fed to the machine so that she can learn to distinguish facial features with precision. The sample must also be representative of all types of face, all skin colors or different forms of eye. "In this area, Europe is lagging behind other regions such as Asia or America, where databases of large data accessible are more important," said the specialist. The European legislative framework for the use of such a database is also stricter than in China, for example.

In a note of April 2016, the National Gendarmerie also pointed out the dangers of building such biometric files in terms of computer security. "The protection of facial recognition systems [against piracy, Ed] is a major issue because you can replace a PIN or a card number that is stolen, but, like the usurpation of identity, one can easily imagine the prejudice of a diversion of digitized face ", write the authors of this document. Cyber ​​criminals could, for example, replace the face of a terrorist on a blacklist by that of an illustrious unknown.

Facial recognition technology is improving rapidly, and "all the limits mentioned may have been overcome in six months or a year," notes Davis. But in the meantime, its reliability is perfect only in very specific circumstances. In other cases, the only point that is 100% sure is that deploying a face recognition system inevitably encroaches on privacy.