Experiment instructions

This is a step-by-step guide for reproducing the experiments in the paper "Evaluating Machine Unlearning via Epistemic Uncertainty". All experiment configurations are pre-defined in the provided shell scripts.

Installation

1. Install python 3.9 and the latest versions of conda and pip on your machine.
2. Install mlflow using pip:

pip install mlflow==1.23.0

3. All requirements will be automatically installed on the first call of one of the scripts below. Therefore, the first run might take a little bit extra time.

Prepare pre-trained models for experiments

To reproduce the experiments in the paper run:

./mnist_training.sh
./cifar_training.sh

Additional model trainings can be run via:

mlflow run . -e training -P seed={SEED} -P dataset={DATASET} -P learning_rate={LR} -P epochs={EPOCHS} -P batch_size={BS} -P output={OUT};

You can find the usage help by running:

python run_training.py -h

Note: Each batch update will be stored for Amnesiac Unlearning. Depending on the model and the dataset this might lead to a large memory usage!

Compute efficacy of initial models

To reproduce the experiments in the paper run:

./mnist_initial.sh
./cifar_initial.sh

You can compute the efficacy for additional initializations by running:

mlflow run . -e initial_efficacy -P seed={SEED} -P dataset={DATASET} -P target={TARGET_CLASS} -P percentage={P} -P output={OUT}

You can find the usage help by running:

python run_initial_efficacy.py -h

If you want to check the initial efficacy for a specific pre-trained model, make sure to provide the same random seed!

Perform membership inference attack on pre-trained models

To reproduce the experiments in the paper run:

./mnist_attack.sh
./cifar_attack.sh

You can perform additional membership inference attacks on the pre-trained models by running:

mlflow run . -e attack_pre_trained -P seed={SEED} -P model={MODEL} -P dataset={DATASET} -P target={TARGET_CLASS} -P percentage={P} -P output={OUT}

You can find the usage help by running:

python run_attack_pre_trained.py -h

Perform forgetting algorithms

You can find the usage help for all forgetting algorithms by running:

python run_forgetting.py -h

Retraining

To reproduce the experiments in the paper run:

./mnist_retraining.sh
./cifar_retraining.sh

You can perform additional retrainings by running:

mlflow run . -e forgetting -P seed={SEED} -P model={MODEL} -P dataset={DATASET} -P target={TARGET_CLASS} -P percentage={P} -P forgetting=retraining -P learning_rate={LR} -P epochs={EPOCHS} -P batch_size={BS} -P output={OUT}

Make sure to use the same hyper-parameters and random seed as for the original training!

Fisher Forgetting

To reproduce the experiments in the paper run:

./mnist_fisher.sh
./cifar_fisher.sh

You can perform additional Fisher Forgetting experiments by running:

mlflow run . -e forgetting -P seed={SEED} -P model={MODEL} -P dataset={DATASET} -P target={TARGET_CLASS} -P percentage={P} -P forgetting=fisher -P alpha={A} -P output={OUT}

Amnesiac Unlearning

To reproduce the experiments in the paper run:

./mnist_amnesiac.sh
./cifar_amnesiac.sh

You can perform additional Amnesiac Unlearning experiments by running:

mlflow run . -e forgetting -P seed={SEED} -P model={MODEL} -P dataset={DATASET} -P target={TARGET_CLASS} -P percentage={P} -P forgetting=amnesiac -P history={BATCH_UPDATES} -P output={OUT}

Create plots

To plot the experimental results as seen in the paper run:

python create_plots.py -r {PATH_TO_EXPERIMENTS} -o {OUT}

Note that you won't obtain the exact same plots as in the paper, since we adjusted the axes' limits for nicer presentation.

Inspect experiments using mlfow

Finally, if you want to examine any of your experiments in detail, use the mlflow dashboard. Run mlflow ui in the directory containing mlruns. This should be the project directory by default. Afterwards, open 127.0.0.1:5000 in your browser.