This repository is the core implementation of the black-box audio adversarial attack approach proposed by our paper (Electronic Music Assassin: Towards Imperceptible Physical Adversarial Attacks against Black-box Automatic Speech Recognitions).
1.Clone this repo.
git clone https://github.com/yzslry/Electronic-Music-Assassin.git
cd Electronic-Music-Assassin
2.Create a virtual environment running Python 3.8 interpreter of later.
3.Install the dependencies.
pip install -r requirements.txt
1.Register the target ASR cloud services provided and fill in the relevant information in the account.py.
2.Create the music folder and add the song you would like as carrier audios to it in wav format with the sample rate of 16000
3.Use the cloud text-to-speech service to generate audios of the target attack commands in wav format with the sample rate of 16000 and place them under the command folder
4.For attacks in the digital world, you can obtain the AEs by executing the following code：
python attack_digital.py --speech-file-path target_TTS_path --music-file-path ./music --attack-target [tencentyun,aliyun,iflytec,google,azure] --sample-num successful_AEs_number
Successful AEs will be saved in folder success_digital_samples
5.For attacks in the physical world, you can obtain the AEs by executing the following code：
python attack_physical.py --speech-file-path target_TTS_path --music-file-path ./music --sample-num AEs_number
AEs will be saved in folder physical_samples and you need the test wheather they are successful in the physical devices.
You can see the AEs we used in our user study on human auditory perception in the folder AEs. AEs for both the physical and digital worlds are included in it. To minimize the impact of player quality and ambient noise on the actual perception, we recommend listening to these demos using headphones.
To facilitate comparison, we split the target commands into two groups, one half for comparison in the physical world and the other half for comparison in the digital world.
We first show five groups of AEs on Google Assistant in the digital world:
digital_group1.mp4
digital_group2.mp4
digital_group3.mp4
digital_group4.mp4
digital_group5.mp4
Next we present five groups of AEs in the physical world:
physical_group1.mp4
physical_group2.mp4
physical_group3.mp4
physical_group4.mp4
physical_group5.mp4
Then, we show some videos taken during the attack in the physical world:
demo0.mp4
demo1.mp4
demo2.mp4
demo3.mp4
demo4.mp4
Finally, we present a demo that automatically wakes up Google Assistant and successfully executes an attack：