There is something magical about our ability to recognize faces and voices: even after many years of absence, we spontaneously identify a person we pass in the street or who calls us on the phone. Of course, if he or she has grown a beard or hair and decided to change the colour of his or her eyes, or calls us in the winter with a hoarse voice, we will have more difficulty... And, on the other hand, we have all believed that we recognise a person and are mistaken. But these situations are quite rare. Today, with artificial intelligence techniques, we are able to build machines capable of automatically recognizing faces with performances almost equivalent to ours.
It took no more than that for the mayor of Nice to decide to install facial recognition portals in his city to ensure the safety of his fellow citizens during the Euro 2016 event and then, if the experiment proved successful, to extend this surveillance to the whole urban area on a permanent basis.
This has been reported on television and radio recently and has been widely reported in all media. However, while there are already many uses for facial recognition, for example in airports or on social networks, there are a number of technical and legal obstacles to the automatic identification of persons known to be terrorists in a crowd, especially at a festive event.
Face recognition algorithms are based on automatic image processing that detects areas of discontinuity (mouths, eyes, hair, etc.) and then locates points of interest (corners of the lips, nostrils, eye and eyebrow tips, etc.). Then, based on these points of interest, machine learning software trained on a large number of images of faces identifies listed individuals. The more points of interest there are, and the more images are used to train the system, the more reliable the recognition is.
Today, however, learning techniques using formal neural networks (the so-called Deep Learning - deep learning) allow the processing of hundreds of millions of images each containing very many points of interest per image (more than a thousand). This explains the breathtaking performance obtained by specialized companies such as DeepFace (Facebook), FaceNet (Google), FaceFirst, Face-Sixetc. Thus, DeepFace announced, more than six months ago, a correct recognition rate of 97.25 %, almost equivalent to ours, and FaceNet even claimed get 99.63 %!
Many applications are already using these techniques: unlocking telephones, checking identity at customs at airports or on Internet accounts, annotating photographs, checking entry to casinos, etc.
However, in order to recognize a person, one or more images of his or her face must have been recorded beforehand. Social networks such as Facebook and photo management sites such as iPhoto for Apple, or Flickr, have a large quantity of these images; they are therefore in a privileged strategic position to implement the aforementioned facial recognition applications, particularly police applications, which is not without its discomfort.
Moreover, the image legislation in force in the various countries is more or less opposed to the implementation of surveillance and counter-terrorism applications. Thus, in Europe, some countries such as the Norway prohibit any shooting that does not receive the explicit consent of the person being photographed. And, for the time being, in France, the administration is not allowed to transmit image files of people on file. In short, if a terrorist hasn't had the crazy idea of putting his photo on Facebook, it's impossible to use it for public safety. As for the United States, while there is no federal law that explicitly mentions the use of facial recognition, there are laws in some states, such as Illinois and Texas, that oppose the use of technologies to identify individuals without their prior informed consent.
In addition to these legal obstacles, there are technical limitations that make the widespread use of facial recognition techniques for city-wide security ineffective. Indeed, if a face is well recognized from the front, with good lighting conditions, it appears much more difficult to identify it from three quarters or from the side with poor lighting. Consequently, the identification of individuals in shots of crowds, for example on the public highway, remains very delicate. Moreover, the FBI's work on shots of much lower quality than social network images shows much poorer performance (about 85 % of correct recognition).
What's more, if a person wears a wig, a fake beard, dark glasses and scratches himself, it becomes almost impossible to recognize him. In the United States, a movement called CV Dazzle... explains how to reroute software facial recognition with a little makeup. In this regard, how can you imagine that in a festive event like theEuro 2016The Nice town hall prevents fans from painting their faces in the colours of their favourite team and thus preventing automatic identification? It should be remembered that, very recently, no software was able to automatically make the connection between the man in the hat filmed at Brussels airport on 22 March 2016 and the images of Mohamed Abrini taken in November 2015, just before the attacks in Paris.
In short, the face recognition software that was supposed to ensure the safety of the population will make it possible to track all those whose images are listed on social networks, but not the individuals on file, for legal reasons. Moreover, those who really want to go unnoticed will make the recognition systems fail by grimacing, while the others will be tracked ...
Jean-Gabriel GanasciaProfessor, Artificial Intelligence, Cognitive Science, Pierre and Marie Curie University (UPMC) - Sorbonne University