OneHotConv

This is an implementation of the OneHot CNN for JPEG steganalysis proposed in this paper.

Data

Dataset preparation is not part of this script. Make sure your data follows the following structure:

DATA-PATH
└───QF100
    └───COVER
    │      └───TRN
    │      └───VAL
    │      └───TST
    │
    └───STEGO_PAYLOAD
           └───TRN
           └───VAL
           └───TST

How to use

python3 train_lit_model.py --version {experiment name} --gpus {num gpus} --data-path {data path root} --stego-scheme {stego scheme name} --payload {payload}

WIP

• Fix training with AMP fp16
• Enable different DCT domain and Spatial domain backbones
• Update to pytorch lightning 1.0

Dependecies

Python 3.5+, pytorch 1.4+ and dependencies listed in requirements.txt.

References

Please consider citing our paper if you find this repository useful.

@article{9091221,
  author={Y. {Yousfi} and J. {Fridrich}},
  journal={IEEE Signal Processing Letters}, 
  title={An Intriguing Struggle of CNNs in JPEG Steganalysis and the OneHot Solution}, 
  year={2020},
  volume={27},
  number={},
  pages={830-834},
  doi={10.1109/LSP.2020.2993959}}