Towards Certified Unlearning for Deep Neural Networks

The code is associated with Towards Certified Unlearning for Deep Neural Networks (ICML 2024).

1.Environment

Experiments are performed on an Nvidia RTX A6000 with Cuda 11.3.

Notice: Cuda is enabled for default settings.

2.Usage

We have three datasets for experiments, namely MNIST, CIFAR-10, and SVHN. Refer to Appendix D for the hyperparameter settings.

2.1 Training the Original Model

Run

python train.py

Hyperparameter settings can be found in the Appendix. After running this code file, an original model will be saved in the ./model/ directory.

2.2 Obtaining the Unlearned Model

Run

python unlearn.py

Hyperparameter settings can be found in the Appendix. After running this code file, an unlearned model will be saved in the ./model/ directory.

2.3 Testing the Unlearned Model

Change the PATH_unl variable in the test_unlearn.py file to the path of the unlearned model to be evaluated.

Run

python test_unlearn.py

For the Membership Inference Attack evaluation, change the PATH_unl variable in the test_unlearn.py file to the path of the unlearned model to be evaluated.

Run

python attack.py