Reconstrução de imagens faciais a partir de modelos profundos

Abstract

Nowadays, face recognition systems are used for secure access in personal devices, banking applications, border control and identification of fugitives from justice. Due to the social and economic impacts involved in such a critical applications, face recognition must meet strict performance requirements, including low error rates and resistance to attacks. In this work, we study the vulnerabilities of VeriLook face recognition system on type of attack, namely, template reconstruction attack. Our method uses the generative approach to reconstruct the binary VeriLook templates, without take into account any other information. The proposed reconstruction was evaluated in a real scenario by using the VeriLook comparison algorithm. The attack was successful using a FAR of 0.1%. However, the accuracy of this attack was low. Further studies should be carried out to assess the need to encrypt the templates. We also show with the experiments that the facial pose influences on the reconstruction.