SUDS: Sanitizing Universal and Dependent Steganography

These experiments were conducted on a macOS Monterey 12.5.1 with a 2.3 GHz 8-Core Intel Core i9 processor with 16 GB 2667 MHz DDR4 of memory.

Installation

1. Download this repository (zip or clone)
2. Open a terminal and change to this working directory
3. Install necessary dependencies with one of the options below (Conda Environment, Docker Build).

---

Conda Environment (Recommended)

This creates, activates, and installs all necessary dependencies.

conda create -y -n SUDS pip python=3.8 && conda activate SUDS && pip install -r requirements.txt

OR

Docker Build

Creates a docker container.

docker build --build-arg WORKDIR_PATH=$(pwd) -t suds_image . 

sudo docker run --name suds --rm -it suds_image bash

Artifact Instructions (~1.5 hrs to run)

All models are pre-trained. Reproduce results by:

1. (IMPORTANT) Extract pre-trained models:

find models/ -name "*.tar.gz" -exec gunzip -f {} \;

2. Reproduce all Figures:

chmod +x scripts/*
./scripts/test_all.sh

3. If you would like to reproduce a specific figure instead of running all tests, see the index below and run:

python *.py

Results Index

All results are saved to the results folder. The results are indexable by research question or figure number.

Artifact	Python Script	Result Location
Table 3:	python RQ1-RQ2-stats.py	results/RQ1-RQ2-stats/all_img_stats.txt
Figure 3a:	python RQ1-RQ2-imgs.py	results/suds-pretty-picture.pdf
Figure 3b:	python RQ3-plots.py	results/noise-pretty-picture.pdf
Figure 3c:	python RQ4-plots.py	results/cifar-suds-pretty-picture.pdf
Figure 4:	python RQ5-imgs.py	results/RQ3-plots/zsize_results.pdf
Figure 5:	python make_pretty_pictures.py	results/RQ4-plots/latent_mappings_plot_compact.pdf
Table 4:	python examples/data_poison/test_data_poison.py	results/data_poison/classification_results.txt

File Descriptions

Training

suds_main.py: main suds training file (sanitizing)

custom_main_deephide.py: main udh and ddh training file (hiding)

Testing

RQ1: SUDS ability to sanitize

RQ2: SUDS comparison to Gaussian noise

RQ3: flexibility; SUDS with different latent space dimensions

RQ4: SUDS for detection

RQ5: scalability; SUDS on CIFAR-10

CASE STUDY: Data poisoning

Directories

configs: training config files for udh and ddh

examples: case studies: data poisoning

models: all pre-trained models

results: where results are saved

scripts: easy to run training and testing scripts

utils: Helper functions. Model files. Etc.

Notes

• If docker throws a disk space error, try running this prior to building:

docker system prune