This is the official code for "Stega4NeRF: Cover Selection Steganography for Neural Radiance Fields"

This code requires Python 3.

You can find the pretrained models at Stega4NeRF\modelD=1.pt.

To train a low-res lego NeRF:

python run_nerf.py --config configs/lego.txt

After training for 200k iterations (~8 hours on a single 2080 Ti), you can find the following video at Stega4NeRF\logs\blender_paper_lego1\blender_paper_lego1_spiral_200000_rgb.mp4.

To render images from different viewpoints

python render_new-viewpoints-images.py 

To train a message extractor (train one-to-one mapping of secret viewpoint image to secret Messages by overfitting): Take D=1 as an example

python train_extractor.py 

After training for 2000 iterations (~27 s on a single 2080 Ti), you can find the following model at Stega4NeRF\modelD=1.pt.

To train a classification model (Implement a disguise for message extractor):

python train_cifar10.py 

After training for 500 iterations (~5 hours on a single 2080 Ti), you can find the following model at Stega4NeRF\modelD=1.pt.

To test new perspective synthesized images (Use the correct extractor key and trained modelD=1.pt) :

python test_secret.py 

To test hybrid model performance (Use trained modelD=1.pt) :

python test_cifar10.py 

The model file “model.pt” and the data in the “data” and “logs” folders can be downloaded from Baidu Netdisk.

The link is: https://pan.baidu.com/s/1s4HAhMQhgBiwjhhHiMMJig
The extraction code is: o98s

NeRF models are used to implement Stega4NeRF.