Code and data for SELM research project.

1. Introduction
2. Installation
3. Encrypt Something
4. Decrypt Something
5. Experiments
6. Cryptanalysis

Install torch (CUDA):

pip install torch==1.10.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

Do this first before installing requirements.txt because that will install a CPU-only torch.

Install packages:

pip install -r requirements.txt

Install the intrinsic package, used for efficient intrinsic dimension operations:

cd intrinsic
python setup.py develop
cd ..

Initialize relics/ (the experiment directory):

relic init

Get a key:

python -c 'import secrets; print(secrets.randbits(32))'

Encrypt with your key:

python encrypt.py --key KEY --int-dim 10000 data/examples/advil.txt

Use the key to decrypt:

python decrypt.py --key KEY advil.bin

To run a new experiment, define a new .toml file in experiments/ with whatever configuration options you want. src/config.py shows all the different options that can be changed.

.toml files can contain lists for parameters; when they do, an experiment for each value in the list is generated. For example, experiments/gpt2/wikipedia/0-4-concat.toml has two lists: one for learning_rate and intrinsic_dimension. This means there are actually 20 experiments in here: 2 learning rates * 10 intrinsic dimensions.

To run the experiments:

python -m src.experiments.run experiments/templates/paper/what-can-we-encrypt-v4.toml

If you are running out of GPU memory, you can use model parallelism to split the Fastfood transform and the GPT2 model onto separate GPUs:

CUDA_VISIBLE_DEVICES=0,2 MODEL_PARALLELISM=1 python -m src.experiments.run experiments/gpt2/examples/medium.toml

You can pass entire directories or just individual .toml files to src.experiments. Results will be saved to relics/.

If you stop an experiment and run it again, any trials that are finished in relics/ will not be run again.

Unzip the provided data:

unzip relics.zip

Play the security game on the original algorithm with an SVM:

python -m src.paper.security svm original feature-fn 500 --ratio 0.8 --quiet

Play the security game on the distribution-regularized variant with an SVM:

python -m src.paper.security svm distribution-reg feature-fn 500 --ratio 0.8 --quiet

Try to implement stronger attacks! Look in src/attacking/ for the model files and add your own.