/steganalysis_with_CNN_for_same_key_images

Primary LanguageLua

Steganalysis via a Convolutional Neural Network using Large Convolution Filters for Embedding Process with Same Stego Key

This project shows the code of the following paper. http://arxiv.org/abs/1605.07946

The authors of this paper are: Jean-François Couchot, Raphaël Couturier, Christophe Guyeux and Michel Salomon

This code is written with TORCH. It uses CNN in order to try to detect if an image is a cover or a stego. Currently, this code has been tested with the HUGO, J-UNIWARD and WOW steganography tools and images of size 512x512. To train the network, a GPU with a lot of memory is needed (we have used a K40), whereas to test some images a small GPU with 1Gb is sufficient.

To download the pre-trained networks, we need to go in the models directory and run download.sh.

To download many cover and stego images, we need to go in the images directory and run download.sh. These images come from the Raise database: http://mmlab.science.unitn.it/RAISE/

In the code, images are labelled 1, 2, 3 etc

Here is the description of the CNN used for the steganalysis:
alternate text
Evolution of the training and testing with HUGO (payload=0.4)
alternate text
Evolution of the training and testing with HUGO (payload=0.1)
alternate text

Other examples of training and testing are in the paper

Some examples

First you need to have cover and stego images and the pre-trained models

th train_stego.lua --cover images/cover_pgm  --stego images/stego_hugo_0.4  --start_test 1 --end_test 1000 --network models/model_hugo_04_iter41.t7 -p cuda

th train_stego.lua --cover images/cover_pgm  --stego images/stego_wow_0.1  --start_test 1 --end_test 1000 --network models/model_wow_01_iter52.t7 -p cuda

th train_stego.lua --cover images/cover_jpg  --stego images/stego_uniward_0.1  --start_test 1 --end_test 1000 --network models/model_uniward_01_iter75.t7 -p cuda --ext .jpg

th train_stego.lua --cover images/cover_pgm  --stego images/stego_hugo_0.4  --start_train 1 --end_train 4000 --start_test 7001 --end_test 8000 -p cuda  -b 100
The interesting parameters are:
  • --cover: the directory containing cover images
  • --stego: the directory containing stego images
  • --start_train: number of the first image to train
  • --end_test: number of the last image to train
  • --start_test: number of the first image to test
  • --network: use a pre-trained model
  • --end_test: number of the last image to test
  • --p cuda: to run on cuda (necessary with our pretrained networks)
  • -b: size of the batch for the training part (the size 100 seems good),
  • --ext .jpg: if you want to use jpg images