WaveGuard Defense

Code for our USENIX 21 paper WaveGuard: Understanding and Mitigating Audio Adversarial Examples .

Audio Examples from paper Audio Examples

Requirements

pip install -r requirements.txt

Also install Deepspeech following the same instructions as in https://github.com/carlini/audio_adversarial_examples to evaluate the defense.

Running the defense

Running the defense on a directory of wav files (sampled at 16KHz):

python Defender/defender_multiple.py --in_dir <PATH TO DIR WITH WAV FILES> --out_base <PATH TO OUTPUT DIR> --defender_type DEFENDER_TYPE --defender_hp DEFENDER_HYPERPARAMETER;

Defender type can be lpc, mel_heuristic, filter_power, quant, downsample_upsample. defender_hp corresponds to number of lpc coeffecients, mel bins, quantization bits, downsampling rare for lpc, mel_heuristic, quant, downsample_upsample respectively.

Evaluating the AUC

The contents of --in_adv can be generated using past works on audio adversairal examples( 1, 2 ) by applying these attacks on the directory of benign audio examples --in_orig. The contents defended directories --in_orig_def, --in_adv_def need to be generated using one of our defenses described above. Then use transcribe_deepspeech.py to generate transcriptions from the deepspeech model for each directory. Then run below command to evaluate the AUC:

python evaluate_detector.py --in_orig <DIR CONTAINING ORIGINAL UNDEFENDED AUDIO> --in_orig_def <DIR CONTAINING ORIGINAL DEFENDED AUDIO> --in_adv <DIR CONTAINING ADVERSARIAL UNDEFENDED AUDIO> --in_orig <DIR CONTAINING ADVERSARIAL DEFENDED AUDIO>

Citing our work

@inproceedings{hussain2021waveguard,
  title={WaveGuard: Understanding and Mitigating Audio Adversarial Examples},
  author={Hussain, Shehzeen and Neekhara, Paarth and Dubnov, Shlomo and McAuley, Julian and Koushanfar, Farinaz},
  booktitle={USENIX Security 21},
  year={2021}
}