SLMIA-SR

This repository contains the code for the paper "SLMIA-SR: Speaker-Level Membership Inference Attacks against Speaker Recognition Systems", which has been published at the 31st Network and Distributed System Security Symposium (NDSS 2024).

[paper]

Environment setup

conda create -n SLMIA-SR python=3.7.11
conda activate SLMIA-SR
sh install_requirements.sh

Install ffmpeg according to the instructions in ffmpeg install instructions

Dataset preparation

Download VoxCeleb-2 and LibriSpeech from VoxCeleb-2 download link and LibriSpeech download link. You should have the following dataset structure:

• $data_dir/vox2/dev/aac/spk_dir/chap_dir/*.m4a
• $data_dir/vox2/test/aac/spk_dir/chap_dir/*.m4a
• $data_dir/ls/train-clean-100/spk_dir/chap_dir/*.flac
• $data_dir/ls/train-clean-360/spk_dir/chap_dir/*.flac
• $data_dir/ls/train-other-500/spk_dir/chap_dir/*.flac
• $data_dir/ls/test-clean/spk_dir/chap_dir/*.flac
• $data_dir/ls/test-other/spk_dir/chap_dir/*.flac
• $data_dir/ls/dev-clean/spk_dir/chap_dir/*.flac
• $data_dir/ls/dev-other/spk_dir/chap_dir/*.flac

Run python -W ignore prepare_dataset.py --data_dir $data_dir --dataset vox2 --n_jobs 20 to convert the audio file into the pytorch tensor.

Run python split_dataset.py --data_dir $data_dir --dataset vox2 to split the dataset as specified by the Table-III in the paper.

SRS training

Run sh train_SRS.sh $data_dir vox2 to train both the target and shadow SRSs on the dataset VoxCeleb-2.

Extract membership inference features

Setting-1 (TABLE V in the paper)

Run python extract_feature.py -dataset vox2 -data_dir $data_dir LSTM_GE2E

Setting-2 (TABLE VI in the paper)

Run python extract_feature.py -dataset vox2 -data_dir $data_dir -num 10 -imposter_num 20 -imposter_voice_num 10 -voice_chunk_splitting -partial_utterance_n_frames 320 -imposter_VCS LSTM_GE2E

Train and evaluate the attack model

Setting-1 (TABLE V in the paper)

Run python train_evaluate_attack_model.py -system_type LSTM_GE2E -dataset vox2 -fpr 0.1

Setting-2 (TABLE VI in the paper)

Run python train_evaluate_attack_model.py -system_type LSTM_GE2E -dataset vox2 -num 10 -imposter_num 20 -imposter_voice 10 -voice_chunk_splitting -imposter_VCS -partial_utterance_n_frames 320 -fpr 0.1

Feel free to change the dataset to ls and fpr to 0.2 since this dataset has less than 1,000 speakers in each splitted set.

If our work or code is useful for you, consider citing our paper as follows:

@inproceedings{SLMIA-SR,
  author       = {Guangke Chen and
                  Yedi Zhang and
                  Fu Song},
  title        = {SLMIA-SR: Speaker-Level Membership Inference Attacks against Speaker Recognition Systems},
  booktitle    = {Proceedings of the 31st Annual Network and Distributed System Security (NDSS) Symposium},
  year         = {2024},
}