THE IMPACT OF AUDIO WATERMARKING ON AUDIO ANTI-SPOOFING COUNTERMEASURES

This repository provides the official implementation for our paper: THE IMPACT OF AUDIO WATERMARKING ON AUDIO ANTI-SPOOFING COUNTERMEASURES . Explore the interactive demo and visualizations here.

What’s inside

Watermark-Spoofing Dataset (WSD) — scripts and label files to generate our watermarked evaluation and training sets.
Knowledge-Preserving Watermark Learning (KPWL) — scripts to reproduce the KPWL framework and experiments.

Watermark-Spoofing Dataset (WSD)

We derive the WSD from public corpora and benchmark it against the original datasets:

ASVspoof 2019 (LA) — training split used for model training. Download: https://datashare.is.ed.ac.uk/handle/10283/3336
ASVspoof 2021 (LA & DF) — used for evaluation.

-LA: https://zenodo.org/record/4837263 -DF: https://zenodo.org/record/4835108
In-the-Wild — used for evaluation. Download: https://deepfake-total.com/in_the_wild

Typical setup: train on ASVspoof 2019 LA (train); evaluate on ASVspoof 2021 LA/DF (eval) and In-the-Wild.

The data generation utilities live in data_generation/. Follow that folder’s README to reproduce WSD and its variants (seen/unseen watermark splits).

Installation (KPWL)

# 1) Clone
git clone https://github.com/Alphawarheads/Watermark_Spoofing.git
cd Watermark_Spoofing

# 2) (Optional) Use the bundled Fairseq snapshot
unzip fairseq-a54021305d6b3c4c5959ac9395135f63202db8f1.zip

# 3) Conda env
conda create -n kpwl python=3.7 -y
conda activate kpwl

# 4) PyTorch (adjust CUDA/ROCm wheels as needed)
pip install torch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1

# 5) Fairseq (editable)
cd fairseq-a54021305d6b3c4c5959ac9395135f63202db8f1
pip install --editable ./
cd ..

# 6) Project requirements
pip install -r requirements.txt

Notes: our main experiments were trained on 4 GPUs/DCUs.

Pre-trained wav2vec 2.0 XLS-R (300M)

Download XLS-R weights from: https://github.com/pytorch/fairseq/tree/main/examples/wav2vec/xlsr

Evaluation

The commands below produce three score.txt files (one per evaluation set). These scores are later used to compute EER (%).

Tip: Always use the full flags --eval_track and --eval_output (do not abbreviate to --eval).

LA 2021

Original (non-watermarked):

TRACK=LA
PROTOCOLS=/public/home/qinxy/.../ASVspoof2021.LA.cm.eval.trl.txt
DATABASE_PATH=/public/home/qinxy/AudioData/Antispoofing/ASVspoof2021/ASVspoof2021_LA_eval/flac/
EVAL_OUTPUT=/public/home/qinxy/.../test2/${COMMENT}_eval_ori_LA21.txt
CUDA_VISIBLE_DEVICES=0,1,2,3 \
torchrun --standalone --nproc_per_node=4 \
    main_multinodes.py --track="$TRACK" --is_eval --eval \
  --model_path models/model_4dcu_kpwl.pth \
  --protocols_path="$PROTOCOLS" \
  --database_path="$DATABASE_PATH" \
  --eval_output="$EVAL_OUTPUT"

Watermark-Spoof (Seen):

TRACK=LA
PROTOCOLS=/public/home/qinxy/zhangzs/watermarking/The-A-Files-master/protocols/tagged/LA21/protocol_wm_75_only.txt
DATABASE_PATH=/public/home/qinxy/AudioData/Antispoofing/LA21_phase1/
EVAL_OUTPUT=/public/home/qinxy/.../test2/${COMMENT}_eval_wm_seen_LA21.txt

Watermark-Spoof (Unseen):

TRACK=LA
PROTOCOLS=/public/home/qinxy/zhangzs/watermarking/The-A-Files-master/protocols/tagged_phase2/LA21/split/protocol_wm_75_only.txt
DATABASE_PATH=/public/home/qinxy/AudioData/Antispoofing/LA21_phase2/
EVAL_OUTPUT=/public/home/qinxy/.../test2/${COMMENT}_eval_wm_unseen_LA21.txt

See eval.sh for the full set of evaluation scripts.

Test with our pretrained models:

Merge scores and evaluate

_wm_ in a filename = scores produced on watermarked audio.
_ori_ in a filename = scores produced on original (non-watermarked) audio.

# If auto_merge_scores is not in the same file, import it:
# from your_module import auto_merge_scores

TRACK = "ITW"     # one of {"LA21", "DF21", "ITW"}
comment = "ori"   # label used in your score filenames

outs = auto_merge_scores(
    track=TRACK,
    score_wm_path=f"/public/.../test2/{comment}_train_eval_wm_SLS_{TRACK}.txt",
    score_ori_path=f"/public/.../test2/{comment}_train_eval_ori_SLS_{TRACK}.txt",
)

print("Output Results:")
for p in outs:
    print(" -", p)

See the README in generate_data/ for details on how the protocols are constructed.

Training

Baseline pretraining (original)

CUDA_VISIBLE_DEVICES=0,1,2,3 \
torchrun --standalone --nproc_per_node=4 \
  main_multinodes.py \
  --track=DF --lr=0.000001 --batch_size=5 --loss=WCE \
  --num_epochs=50 --comment 4dcu_baseline

KPWL fine-tuning

CUDA_VISIBLE_DEVICES=0,1,2,3 \
torchrun --standalone --nproc_per_node=4 \
    main_kpwl.py \
  --eval_track=DF --lr=5e-7 --batch_size=5 --num_epochs=2 \
  --model_path best_model.pth \
  --comment model_4dcu_kpwl

Acknowledgments & Third-Party Code

This repository builds upon and incorporates modified components from the following open-source projects:

SLSforASVspoof-2021-DF — https://github.com/QiShanZhang/SLSforASVspoof-2021-DF.git

The-A-Files — https://github.com/pawel-kaczmarek/The-A-Files.git

Portions of the code are adapted from these repositories. All third-party code remains under its original licenses; please refer to the upstream LICENSE files and retain all applicable copyright notices. If any attribution or licensing detail is incomplete, let us know and we will correct it promptly.

Contact / Citation If you use this repository or WSD/KPWL in your research, please cite the paper and link back to this repo.

Watermark_Spoofing

Most Popular