UnGAN: Machine Unlearning Strategies through Membership Inference

Аннотация

As regulatory and ethical demands for data privacy and the right to be forgotten increase, the ability to effectively unlearn specific data points from machine learning models without retraining from scratch becomes paramount. Machine unlearning aims to efficiently eliminate the influence of certain data points on a model. We propose the UnGAN, a novel approach to machine unlearning that leverages Generative Adversarial Networks (GANs) to address the growing need for efficient and reliable data removal from trained models. UnGAN proposes a unique unlearning strategy through membership inference, where a discriminator network is trained to identify whether a given input was part of the model's training set. The discriminator is a three-layer fully connected network employing ReLU activation functions, receiving inputs from the output of the model undergoing unlearning and the class label. This architecture enables the discriminator to learn the membership status of data points with high precision, thereby guiding the unlearning process.