Transferbility_of_AEs

This is the github page for Section VIII.B (Audio Tranferability Experiments) of the paper "SoK: A Modularized Approach to Study the Security of Automatic Speech Recognition Systems", we put our audio adversarial examples introduced in the paper.

Experiment 1: Transferability of the AEs based on end-to-end ASR model

In folder Transferbility_of_AEs_based_DeepSpeech_v0.1.0_command_01/02/03/04/05) contains adversarial samples generated based on DeepSpeech v0.1.0 with its' decoding results. Specifically, *.mp4 is the decoding demo on Microsoft STT Webpage, *.txt is the decoding text on Deepspeech/V0.2.0 and Deepspeech/V0.3.0, *.json is the decoding json result on Microsoft STT API.

The commands injected in adversarial examples (AEs) are as follows.

jzy-adv.wav - open the website
msn-adv.wav - turn off the light
wye-adv.wav - where is my car
zdc-adv.wav - how is the weather
zet-adv.wav - open the front door

Experiment 2: Transferability of the AEs based on classical ASR model

In folder Transferbility_of_AEs_based_Kaldi_ASpIRE_Chain_Model_command_01/02) contains adversarial samples with its' decoding results on Microsoft STT API and Google STT API.

The commands injected in adversarial examples (AEs) are as follows.

WAA-google-ASPIRE-Sample-0-64.wav - navigate to my home
WAA-google-ASPIRE-Sample-0-76.wav - open youtube

Experiment Code

Thanks to the excellent works of Yakura et al ( Github Link ) and Chen et al ( Github Link ), our work had been greatly facilitated.