Practical Hidden Voice Attacks against Speech and Speaker Recognition Systems

Hadi Abdullah, Washington Garcia, Christian Peeters,

Patrick Traynor, Kevin R. B. Butler, and Joseph Wilson.

The Hidden Audio Commands are designed to be difficult for humans to understand but easy for Voice Processing Systems (VPS) to interpret. Below are samples of the commands that have been successfully tested against VPSs.

NOTE: Due to the following effect, audio will become immediately understandable once the text is known due to this psychoacoustic effect . To ensure that reviewers do not fall victim to this psychoacoustic effect, we do not reveal the contents of this message.

For more information, please check out our paper.

To cite our work:

@INPROCEEDINGS{abdullah2019practical,

title= {{Practical Hidden Voice Attacks against Speech and Speaker Recognition Systems}},

author={Abdullah, Hadi and Garcia, Washington and Peeters, Christian and Traynor, Patrick and Butler, Kevin and Wilson, Joseph},

booktitle={Network and Distributed System Security Symposium (NDSS)},

year={2019}

}

Audio 1:

phva_demo.mp4

Perturbation Method:

  1. Time Domain Inversion

Audio 2:

Audio2.wav

Model: Azure

Perturbation Method:

  1. Random Phase Generation

Audio 3:

Audio3.wav

Model: Kaldi-DNN

Perturbation Method:

  1. Time Domain Inversion

  2. High Frequency Addition

Audio 4:

Audio4.wav

Model: Wit

Perturbation Method:

  1. Time Domain Inversion

  2. Time Scaling