schroederdewitt / perfectly-secure-steganography

Contains open source code for the paper "Perfectly-secure Steganography using Minimum Entropy Coupling"

Instructions

Pre-requisites

Set up a python3.8 virtual environment using requirements.txt. Please set up a WandB account and project to be used for the logging.

For reproducible model forwards, set environment variable CUBLAS_WORKSPACE_CONFIG=:4096:8:

Running experiments

We supply a numerically stabilised python/numpy implementation of MEC that has quadratic runtime with maximum dimension of both p and q. Please note that, for the results quoted in our paper, we use a highly optimised C++ implementation, so speed results from this code may not be comparable for iMEC.

We will release our optimised implementations at a later stage, please contact [e-mail] for further information.

Running iMEC experiments

For running iMEC at block size 10 on GPT-2 (90%) use the following:

python3 main.py 
--block-size 10
--group-name <WANDB_GROUP_NAME>
--method imec 
--medium meteor
--medium-entropy-loss-threshold 0.9
--medium-top-k 0 
--message-mode randombits
--model-device cuda:0
--name <WANDB_EXPERIMENT_NAME>
--stop-after-n-trajectories 100
--wandb-project <WANDB_PROJECT>
--wandb-entity <WANDB_ENTITY>

For GPT-2 (topk 40) set

--medium-entropy-loss-threshold 0.0
--medium-top-k 40

For WaveRNN set

--medium wavernn

For UNIF set

--medium random
--medium-top-k 40
--medium-entropy-loss-threshold 0.0

Running Meteor experiments

For running Meteor at precision 32 on GPT-2 (90%) use the following:

python3 main.py 
--group-name <WANDB_GROUP_NAME>
--meteor-is-sort 1
--meteor-precision 32
--method meteor 
--medium meteor
--medium-entropy-loss-threshold 0.9
--medium-top-k 0 
--message-mode randombits
--model-device cuda:0
--name <WANDB_EXPERIMENT_NAME>
--stop-after-n-trajectories 100
--wandb-project <WANDB_PROJECT>
--wandb-entity <WANDB_ENTITY>

To instead run Meteor:reorder, set

--meteor-is-sort 1

Running Arithmetic coding experiments

For running Arithmetic coding at precision 16 on GPT-2 (90%) use the following:

python3 main.py 
--group-name <WANDB_GROUP_NAME>
--meteor-is-sort 1
--meteor-precision 16
--method arithmetic
--medium meteor
--medium-entropy-loss-threshold 0.9
--medium-top-k 0 
--message-mode randombits
--model-device cuda:0
--name <WANDB_EXPERIMENT_NAME>
--stop-after-n-trajectories 100
--wandb-project <WANDB_PROJECT>
--wandb-entity <WANDB_ENTITY>