A Machine Unlearning Method Based on Multiple GAN Models

A machine unlearning method based on multiple GAN models generates synthetic data, inverts the class labels of the forget set, and fine-tunes a pretrained model, outperforming SOTA models and standard unlearning classifiers. The method has two stages: the GAN consists of two pairs of generators and discriminators, which generate synthetic data for the retain set and the forget set respectively; then the pretrained model labels the classes and inverts the labels of the forget set. Experiments on CIFAR-10 use membership inference attacks (MIA) to evaluate the model after unlearning.