Laura Graves, Vineel Nagisetty, Vijay Ganesh
This repo includes two methods for removing learned data from a neural network.
Unlearning involves retraining on a modified dataset that is designed to muddy the model's understanding of sensitive data
Amnesiac Unlearning involves specific removal of the learning from batches containing sensitive data
These methods were evaluated via test accuracy, membership inference attacks, and model inversion attacks.
Amnesiac ML Evaluation.ipynb contains a notebook that tests the effect of greater levels of amnesiac unlearning on model efficacy.
Membership Inference Attack.ipynb contains a notebook that runs a Membership Inference Attack on a resnet18 model trained on CIFAR100, and evaluates the success of this attack against amnesiac unlearning.
Model Inversion Attack.ipynb contains a notebook that runs a Model Inversion Attack against a trained PyTorch model.
Amnesiac Machine Learning.pdf is a preprint paper containing our research findings and a detailed description of methods and evaluation.