This repository is our implementation of ADG.

Provably Secure Generative Linguistic Steganography (ACL 2021 findings)
Siyu Zhang, Zhongliang Yang, Jinshuai Yang, Yongfeng Huang
https://arxiv.org/abs/2106.02011

• PyTorch
• Numpy
• Gensim

1. Download training data from Google Drive or 百度网盘 (提取码：udm4).
2. Place the datasets inside the data/ folder.

You can either train from scratch or directly download checkpoints.

python run_lm.py

The checkpoint files will be saved into models/ folder.

1. Download checkpoint files from Google Drive or 百度网盘 (提取码：7xcr).
2. Place the checkpoint files inside the models/ folder.

python adg-stega.py

The generated stegotext will be saved into stego/ folder.

cd measure/er
python measure_er.py

Results will be saved into measure/er/er.json.

cd measure/kld1
python measure_kld1.py

cd measure/kld2
python dataset_remove_unk.py
python build_vec.py
python measure_kld2.py

Results will be saved into measure/kld2/kld2.json.

If you find our repository useful, please consider citing our paper:

KLD2 is largely based on sentence2vec.