Defend Data Poisoning Attacks on Voice Authentication

With the advances in deep learning, speaker verification has achieved very high accuracy and is gaining popularity as a type of biometric authentication option many scenes our daily life, especially growing market web services. Compared to traditional passwords, "vocal passwords" are much more convenient they relieve people from memorizing different passwords. However, new machine learning attacks putting these voice systems at risk. Without strong security guarantee, attackers could access legitimate users' accounts by fooling neural network (DNN) based recognition models. In this paper, we demonstrate an easy-to-implement data poisoning attack system, which can hardly be captured existing defense mechanisms. Thus, propose robust method, called Guardian, convolutional network-based discriminator. The Guardian discriminator integrates series novel techniques including bias reduction, input augmentation, ensemble learning. Our approach able distinguish about 95% attacked normal accounts, effective than approaches with only 60% accuracy.