Fooling End-To-End Speaker Verification With Adversarial Examples
Обман сквозной системы верификации диктора с помощью состязательных примеров
2018-04-01
SCID: 54.1/j9xppduc
Discuss with AI
adversarial examplesblack-box attacksend-to-end deep neural networksspeaker verificationwhite-box attacks
Figures from the paper
Abstract (AI)
Automatic speaker verification systems are increasingly used as the primary means to authenticate costumers. Recently, it has been proposed to train speaker verification systems using end-to-end deep neural models. In this paper, we show that such systems are vulnerable to adversarial example attacks. Adversarial examples are generated by adding a peculiar noise to original speaker examples, in such a way that they are almost indistinguishable, by a human listener. Yet, the generated waveforms, which sound as speaker A can be used to fool such a system by claiming as if the waveforms were uttered by speaker B. We present white-box attacks on a deep end-to-end network that was either trained on YOHO or NTIMIT. We also present two black-box attacks. In the first one, we generate adversarial examples with a system trained on NTIMIT and perform the attack on a system that trained on YOHO. In the second one, we generate the adversarial examples with a system trained using Mel-spectrum features and perform the attack on a system trained using MFCCs. Our results show that one can significantly decrease the accuracy of a target system even when the adversarial examples are generated with different system potentially using different features.
Key Findings
1
Adversarial waveforms sounding like speaker A can cause verification systems to accept them as being spoken by speaker B.
2
Attack effectiveness transfers across feature representations: examples generated using Mel-spectrum features can significantly reduce accuracy in systems using MFCCs.
3
Black-box attacks transfer across systems trained on different datasets, including adversarial examples generated on NTIMIT and tested against YOHO-trained systems.
4
End-to-end deep speaker-verification systems are vulnerable to adversarial examples that remain nearly indistinguishable from original speech to human listeners.
5
White-box attacks successfully target end-to-end networks trained on either YOHO or NTIMIT datasets.
Research Object
end-to-end deep neural speaker verification systems
Research Subject
vulnerability of speaker verification systems to adversarial example attacks, including attack transferability across training datasets, models, and acoustic features
Publication Details
Publication Date
2018-04-01
Journal
Publisher
ISSN
Cited by
181
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest
References available in scid.ai7
Deep Residual Learning for Image Recognition2016
ImageNet classification with deep convolutional neural networks2017
Neural Machine Translation by Jointly Learning to Align and Translate2014
Intriguing properties of neural networks2013
Front-End Factor Analysis for Speaker Verification2010
End-to-end text-dependent speaker verification2016
Deep neural network-based speaker embeddings for end-to-end speaker verification2016