Universal Adversarial Perturbations Generative Network for Speaker Recognition

In this paper, we demonstrated the existence of the universal adversarial perturbations~(UAPs) for the speaker recognition systems. We proposed a generative network to learn the mapping from the low-dimensional normal distribution to the UAPs subspace, then synthesize the UAPs to perturbe any input signals to spoof the well-trained speaker recognition model with high probability. Experimental results on TIMIT and LibriSpeech datasets demonstrate the effectiveness of our model.

Cite us

If you use this code or part of it, please cite us!
Jiguo Li, Xinfeng Zhang, Chuanming Jia, Jizheng Xu, Li Zhang, Yue Wang, Siwei Ma, Wen Gao, "universal adversarial perturbations generative network for speaker recognition" arxiv, IEEE Xplore.

Prerequisites

• linux
• python 3.5 (not test on other versions)
• pytorch 1.3+
• torchaudio 0.3
• librosa, pysoundfile
• json, tqdm, logging

data preparing

download dataset and pretrained model

• you can download TIMIT dataset from the offficial site or here
• download the pretrained speaker recognition model from here and save the model into ./output/SincNet_TIMIT, this model is released by the author of SincNet.
• you can download LibriSpeech dataset from here. you can download our reproducable speaker recognition model for LibriSpeech model from here.
• you can download the pretrained models for our universal attacker from here

preprocess the data

prepare the data, split the audio data into slices:

python ./common/prepare_dataset.py --type speaker --data_root TIMIT_DATA_ROOT --output_dir TIMIT_DATA_ROOT/processed

training the attacker model

In the project root directory, run

python ./train_generator.py --output_dir ./output/timit_generator --no_dist --noise_scale 1 --norm_factor 1000 --norm_clip 0.01 --speaker_factor 1 --wlen 200 --frame_dim 3200 --margin 10 --data_root DATAROOT --dataset timit --mul 0.1 --epoch 50

An example of the script:

python ./universal_attacker/train_generator.py --output_dir ./output/generator/timit_generator_norm2000_mul01_wlen800_margin10_epoch50_clip001_nolrdecay_fixdmse --speaker_model ./output/SincNet_TIMIT/model_raw.pkl --speaker_cfg ./config/timit_speaker_generator.cfg --noise_scale 1 --norm_factor 2000 --mul 0.1 --norm_clip 0.01 --speaker_factor 1 --wlen 800 --frame_dim 12800 --margin 10 --epoch 50 --data_root /data00/jiguo/dataset/TIMIT_lower --dataset timit --num_workers 8 --test_output ./output/generator/timit_generator_norm2000_mul01_wlen800_margin10_epoch50_clip001_nolrdecay_fixdmse_test --no_dist

The training for TIMIT dataset will cost about 2 day on a 1080ti GPU.

project home page

More results can be found on our UAPs for speaker recognition project main page.

Acknowledgement

Thanks to the valuable discussion with Jing Lin. Besides, thanks to the open source of SincNet, Pytorch-Kaldi, PESQ.

Please feel free to contact me (jiguo.li@vipl.ict.ac.cn, jgli@pku.edu.cn) if you have any questions about this project. Note that this work is only for research. Please do not use it for illegal purposes.