Facial recognition at the international level: technical and legal aspects

Facial recognition is a technology derived from computer vision, aimed at identifying or authenticating a person based on their facial features. First experimented with in the 1960s, it has grown considerably with the emergence of artificial intelligence (AI) and machine learning. AI encompasses algorithmic systems capable, for a defined purpose, of perceiving their environment, analyzing structured or unstructured data, reasoning, and deciding on the most appropriate action, with or without the use of machine learning (according to the definition given by the OECD).

Facial recognition is becoming increasingly widespread in the United States and China, which are among the leading countries in this field. It is also being deployed in the Middle East and in certain European countries, where its use is sparking debate and controversy. In France, its adoption is currently the subject of legal, ethical, and societal discussions.

There are two main approaches to facial recognition that impact implementation and use:

• facial verification;

• facial recognition.

Facial verification seeks to confirm an identity, i.e., to determine whether two different images correspond to the same subject. It answers the question: "Are you who you say you are?" This approach is used, for example, for biometric authentication.

Facial recognition aims to determine the specific identity of the subject and answers the question: "Who are you?" Facial recognition can be classified into two categories: close-range facial recognition, which identifies a face from a finite and known set of faces previously stored in the database, and open-range facial recognition, which determines whether the face belongs to one of the individuals already stored in the database (known face), or whether it belongs to a person who is not in the database (unknown face).

In general, facial recognition is based on four main steps:

1. detection of the individual in the image;

2. preprocessing: the face is "realigned"; this step involves making the face compliant so that it can be analyzed later;

3. the generation of feature vectors;

4. Classification: feature vectors are compared in order to test whether two faces correspond to the same person.

The main models that have contributed to developments in the field of facial recognition are Eigenfaces, Alexnet, Deepface, FaceNet, Arcface, Face, and AdaCos.

The purpose of this article is to analyze the uses, legal frameworks,...