Vieses em algoritmos de reconhecimento facial: um mapeamento sistemático da literatura.

Abstract

Facial recognition is a very popular tool in applications, mainly used for authentication. Due to its popularity, the question arose as to whether it is fair in recognizing faces of people of different genders and ethnicities. Therefore, a systematic mapping was developed to analyze bias in facial recognition algorithms. Studying several scientific articles, it was possible to identify that commercial systems present bias, evidencing that the database used by the algorithm is not representative enough for people of different nationalities. This was demonstrated through several experiments, indicating that the most efficient solution to mitigate bias is a balanced database. This article serves as a basis for future studies to seek practical and efficient solutions.