Detecção de tentativa de invasão por dados sintéticos em aplicações de biometria por voz

Voice-based biometric systems are very common nowadays, especi- ally with the popularization of voice command systems and digital assistants such as Google Assistant or Alexa. An important feature of these systems is to detect the user giving a command, as it controls the access to personal or sensitive information to your profile. Thus, as in face-based biometrics, audio biometrics can be attacked by synthetic data, where recordings can be presented as real data. This work presents a model based on deep neural networks capable of detecting this invasion technique. For the training, we used real data of recordings and synthetic data generated from the original recordings. We obtained satisfactory results, mainly due to the low rate of false acceptance and high rate of F1-Score, even in different environments and noises.