DSec_project

TorchAttack: https://github.com/Harry24k/adversarial-attacks-pytorch

White Box Attack with ImageNet: https://github.com/Harry24k/adversarial-attacks-pytorch/blob/master/demos/White%20Box%20Attack%20(ImageNet).ipynb

Notebooks:

Evaluation: https://colab.research.google.com/drive/1-SQPaTivdOtxq514WGOnofiUCxCcnWb3?usp=sharing (Add in git folder)

Documentation:

Kickoff Presentation: https://docs.google.com/presentation/d/1nWF6kdSFXgJHggLFSHd_fG8sAkv-H3cqojQD-QzYeyw/edit#slide=id.p

Final Presentation: https://docs.google.com/presentation/d/1PXsY8NgJm-avOCrAOOV4VSBgxBZptYRxYlFXMR7fk_k/edit#slide=id.p

Final Paper: https://www.overleaf.com/1311294973kbsxvtkcddxq

Data

Data should be downloaded from this link 'https://www.kaggle.com/c/alaska2-image-steganalysis/data' and stored in '/alaska2-image-steganalysis' folder. This folder contains following directories and files: - Cover/ contains 75k unaltered images meant for use in training. - JMiPOD/ contains 75k examples of the JMiPOD algorithm applied to the cover images. - JUNIWARD/contains 75k examples of the JUNIWARD algorithm applied to the cover images. - UERD/ contains 75k examples of the UERD algorithm applied to the cover images. - Test/ contains 5k test set images. These are the images for which you are predicting. - sample_submission.csv contains an example submission in the correct format.

Scripts

1. steganogan_image_generation.py: This script creates steganoGAN stego from the cover images stored in this directory '/alaska2-image-steganalysis/Cover/'. It encodes randomly created alphanumeric strings of random length between 10 to 50 into all the cover images using encoder of the steganoGAN model and saves the encoded images here '.alaska2-image-steganalysis/SteganoGAN/'.

2. adversarial-attack.py: This scripts does the following tasks:

   • Loads only validation dataset of cover and all stegos.
   • Loads the model and weight '/best-checkpoint-143epoch.bin' for the independent steganalyzer.
   • Gets the prediction of the steganalyzer for these images.
   • Performs 8 different adversarial attack for all 4 stegos and cover images.
   • Gets the prediction of the stegos and cover after they go through all the attacks attacks.
   • Saves all the predictions as '/Evaluation/adv_attack/submission.csv' for further performance analysis.

3. save_steganogan_adv_atk_images.py: This script attacks the steganoGAN stegos '/alaska2-image-steganalysis/SteganoGAN/' for a small subset of the validation set (1000 images) and saves the original and all images created by 8 adversarial attacks in this directory 'alaska2-image-steganalysis/SteganoGAN_Attk/'.

4. decode_secret.py: It loads the images of original stegos and attacked stegos from '/alaska2-image-steganalysis/SteganoGAN_Attk/' and uses steganoGAN decoder to retrieve the secret and saves the secrets in this file '/Evaluation/secrets.csv'.

To reproduce the results, one should run the four scripts in the same order.

Notebooks

1. DSec_evaluation.ipynb: This notebook creates evaluation results of the experiments.

Environment

We need two python environments because steganogan works only on pytorch 1.0.0. But other libraries and models need newer version.

1. Steganogan Environment: Only steganogan_image_generation.py needs to be run in this environemnt with python 3.6. The only package to install is steganogan by running the command: pip install steganogan. It will install all necessary libraries.

2. Other Environment: All other scripts would run in this environemnt. The requirements are:

   • python 3.6
   • pytorch 1.5.0
   • torchattacks 2.12.1
   • scikit-learn 0.23.2
   • opencv 3.4.2
   • scikit-image 0.17.2
   • pandas 1.1.5
   • numpy 1.19.2
   • albumentations 0.5.2
   • matplotlib 3.3.3
   • torchvision 0.6.0
   • efficientnet-pytorch 0.7.0