AudioAuth

A dual-watermarking framework for robust audio integrity verification and source attribution.

---

AudioAuth embeds two complementary watermarks into audio using frequency-partitioned encoding: a model watermark (fixed pattern in even frequency bands) for source attribution, and a data watermark (dynamic payload in odd frequency bands) for fine-grained temporal tamper localization.

Installation

git clone https://github.com/pujariaditya/AudioAuth.git
cd AudioAuth
pip install -e .

Quick Start

from pathlib import Path
import torch
from AudioAuth import Config, load_models
from AudioAuth.processor import load_audio

# Load config and model
config = Config.from_sources(yaml_file=Path("configs/train_stage3.yml"))
model, _ = load_models(config, checkpoint_path=Path("checkpoints/best.pth"), device="cuda")
model.eval()

# Load audio and create watermark message
audio = load_audio("input.wav", sample_rate=16000).unsqueeze(0).unsqueeze(0).to("cuda")  # [1, 1, T]
# Create watermark message (16 model bits + 16 data bits)
model_bits = torch.ones(1, 16, device="cuda")                    # fixed model signature
data_bits  = torch.randint(0, 2, (1, 16), device="cuda").float() # dynamic payload
msg = torch.cat([model_bits, data_bits], dim=1)                   # [1, 32]

# Embed watermark
with torch.no_grad():
    out = model(audio, msg, phase="audio_sample")

watermarked = out["watermarked_signal"]       # [1, 1, T] — watermarked audio
detected    = out["watermarked_detector_output"]  # [1, 32]  — 16 model + 16 data bits
locator     = out["watermarked_locator_output"]   # [1, 1, T] — per-sample watermark presence

Training

AudioAuth uses a 3-stage training pipeline:

# 1. Prepare dataset
python scripts/prepare_dataset.py --audio-dir /path/to/audio --output-dir data/

# 2. Stage 1: Base watermarking
python train.py --cfg-path configs/train_stage1.yml

# 3. Stage 2: Attack robustness
python train.py --cfg-path configs/train_stage2.yml

# 4. Stage 3: GAN fine-tuning
python train.py --cfg-path configs/train_stage3.yml

For multi-GPU training, use torchrun --nproc_per_node=N train.py --cfg-path .... Set resume_from in each config to the previous stage's checkpoint.

Audio Samples

Compare original and watermarked audio — the watermark is imperceptible.

Sample	Original	Watermarked	SI-SNR (dB)
sample_001	Play	Play	22.46
sample_002	Play	Play	21.50
sample_003	Play	Play	20.00

Citation

@article{pujari2026audioauth,
  author    = {Pujari, Aditya and Rattani, Ajita},
  title     = {AudioAuth: A Dual-Watermarking Framework for Robust Audio Integrity and Source Attribution},
  journal   = {IEEE Transactions on Biometrics, Behavior, and Identity Science},
  year      = {2026},
  month     = {01},
  pages     = {1-15},
  doi       = {10.1109/TBIOM.2026.3679274}
}

License

MIT License. See LICENSE for details.