Spotting adversarial samples for ASV by neural vocoders

Introduction

• This repo is used for reproducing the main result of "Adversarial Sample Detection for Speaker Verification by Neural Vocoders"
• Paper: Arxiv
• Audio demo: Demo
• Video: Youtube

Installation

git clone https://github.com/HaibinWu666/spot-adv-by-vocoder.git
pip install -r requirements.txt
cd ParallelWaveGAN
pip install -e .

Usage

1.Train ASV

• Prepare the data

cd speaker_verification/examples/VoxCeleb/verification
set the voxceleb1_path、voxceleb2_path、musan_path、rirs_path in run.sh (voxceleb1_path and voxceleb2_path should be formated as voxceleb1_path/dev/wav/idxxx and voxceleb2_path/dev/aac/idxxx;)
set the stage in run.sh to 0 to build soft links
bash run.sh
set the stage in run.sh to 1 to format data
bash run.sh

• Train

You can skip this step and use our pretrained model (speaker_verification/pretrained_model/)

set the stage in run.sh to 2
bash run.sh
the model is saved in exp/

• Evaluate

set the stage in run.sh to 3
set the ckpt_path
bash run.sh

2. Generate adversarial samples

• Prepare data

set the stage in run.sh to 0
set the voxceleb1_path
bash run.sh

• Attack

set the stage in run.sh to 1
set the voxceleb1_path
bash run.sh

• Evaluate

set the stage in run.sh to 2
bash run.sh

• Prepare trial files

set the stage in run.sh to 3
bash run.sh

3. Train vocoder

You can skip this step and use our pretrained model in ParallelWaveGAN/pretrained_model/

• Prepare the data and train

cd ParallelWaveGAN/egs/ljspeech/voxceleb1
set the voxceleb1_path in run.sh
set the stage in run.sh to 2
bash run.sh

4. Resynthesis the wav

• Use vocoder to resynthesis the wav

For adversarial audio
cd ParallelWaveGAN
set model_dir and data_dir in run_audio_generation.sh
set model_dir=pretrained_model/train_nodev_ljspeech_parallel_wavegan.v1.long 
set data_dir=../speaker_verification/examples/VoxCeleb/attack/data/adv_data_epsilon15_it5
bash run_audio_generation.sh
For clean audio
cd ParallelWaveGAN
set model_dir=pretrained_model/train_nodev_ljspeech_parallel_wavegan.v1.long
set data_dir=../speaker_verification/examples/VoxCeleb/attack/data/clean
bash run_audio_generation.sh

• Use Griffin-Lim to resynthesis the wav

For adversarial audio
cd Griffin-Lim
set data_root in run.sh=../speaker_verification/examples/VoxCeleb/attack/data/adv_data_epsilon15_it5
bash run.sh
For clean audio
cd Griffin-Lim
set data_root in run.sh=../speaker_verification/examples/VoxCeleb/attack/data/clean
bash run.sh

5. Result illustration

cd speaker_verification/examples/VoxCeleb/attack
set stage to 5 in run.sh
bash run.sh

Citation

If you think this work helps your research or use the code, please consider citing our paper. Thank you!

@inproceedings{wu2022adversarial,
  title={Adversarial Sample Detection for Speaker Verification by Neural Vocoders},
  author={Wu, Haibin and Hsu, Po-Chun and Gao, Ji and Zhang, Shanshan and Huang, Shen and Kang, Jian and Wu, Zhiyong and Meng, Helen and Lee, Hung-Yi},
  booktitle={ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={236--240},
  year={2022},
  organization={IEEE}
}

Reference

• The implementation of ParallelWaveGAN is from https://github.com/kan-bayashi/ParallelWaveGAN
• https://github.com/thuhcsi/torch_speaker
• https://github.com/clovaai/voxceleb_trainer