Robust Neural Audio Watermarking with Invertible Dual-Embedding
Audio watermarking embeds messages into audio and accurately extracts the watermark even after the watermarked audio has been damaged. Compared to traditional digital watermarking algorithms, neural watermarking using neural networks has better robustness for various attacks are considered during training. However, neural watermarking methods suffer to low capacity and undesirable imperceptibility. In addition, in practical scenarios, watermarks are redundantly embedded in audio according to the demand, and the audio is also subjected to cropping and splicing, which makes the efficient locating of watermarks a problem worth exploring. We design an invertible neural network to realize a dual-embedding watermarking model for efficient locating, at the same time, we consider the impact of the attack layer on the invertible neural network in robustness training and improve the model so as to enhance the reasonableness and stability.
"(*)" means necessary steps
- (*) Modify ./data/config.yaml based on the specific location of your dataset.
- (*) Build Dataset and dump into pickle:
python3 ./data/process.py - Test your built Dataset, DataLoader:
python3 ./data/test_data.py
DataLoader provides data batch shaped [batch_size, sr * audio_limit_len] each time.
IDEAW is composed of MIHNETs, a discriminator, a balance block and an attack layer.
These networks are defined in different .py files:
- mihnet.py defines the core multi-INN of IDEAW.
- componentNet.py defines the discriminator and the balance block.
- attackLayer.py defines the attack simulate layer.
The components of MIHNET, i.e. INN block is defined in innBlock.py.
Then, in ideaw.py, these components are assembled. The IDEAW in ideaw.py also provides STFT/ISTFT/Embed/Extract operations.
Finally, solver.py provides the optimizer of IDEAW.
Note that all the configurations of the model are located in ./models/config.yaml.
- (*) Modify the training configuration ./config.yaml and the paths of the related configurations and dataset in ./train.sh.
- (*) Train the model:
./train.sh
By default, the two stages of training each account for half of the iterations.
embed_extract.py provides the embedding and extracting processes of audio watermark via the trained IDEAW.
Note that if you have used the DataParallel version of solver.py to train IDEAW, please use the DataParallel version of embed_extract.py to test the trained model. embed_extract.py showcases how to use IDEAW for watermark embedding and extraction. For simplicity, hard coding is used, and the trained IDEAW can be encapsulated based on the provided approach.
The attack types in attackLayer can be customized in ./models/attackLayer.py to enhance the specific robustness of IDEAW. The factors of attacks are in ./models/config.yaml["AttackLayer"]
The code provides 8 types of common attacks on audio, including Gaussian additive noise, lower-pass filter, mpeg3 compression, etc.
Remember to modify the configuration if more attacks are included:
./models/config.yaml["AttackLayer"]["att_num"].
Due to the company's security regulations, the final version of the experimental code is debugged and run on company's server but cannot be copied out. There may be some errors in the current version. If you find any mistakes, please point them out through issue, thanks.
If IDEAW helps your research, please cite it as,
Bibtex:
@inproceedings{li2024ideaw,
title={IDEAW: Robust Neural Audio Watermarking with Invertible Dual-Embedding},
author={Li, Pengcheng and Zhang, Xulong and Xiao, Jing and Wang, Jianzong},
booktitle={Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing},
pages={1--12},
year={2024},
publisher={Association for Computational Linguistics}
}
or with a hyperlink,
Markdown: [IDEAW](https://github.com/PecholaL/IDEAW)
Latex: \href{https://github.com/PecholaL/IDEAW}{IDEAW}