Zero-shot Machine Unlearning using GANs

Abstract

This thesis addresses the crucial task of machine unlearning, which involves the removal of specific data from trained machine-learning models to comply with privacy regulations and enhance data quality. With the rapid advancements in AI and the extensive use of machine learning models in various applications, efficient unlearning methods are increasingly urgent. Traditional approaches, such as retraining models from scratch, are impractical due to high resource consumption and time constraints.

The thesis proposes two innovative techniques: MuGAN and zMuGAN. MuGAN, short for Machine unlearning using GANs, is designed for scenarios with limited access to the original training dataset. It uses Generative Adversarial Networks (GANs) to capture the data distribution during the model's initial training and generate synthetic data for unlearning upon receiving such a request. Similarly, zMuGAN, short for zero-shot Machine unlearning using GANs, addresses situations where no access to the training data is available at all. It utilizes a GAN-based model inversion technique to approximate the original dataset and facilitate unlearning through an impair-repair unlearning process. Both techniques are evaluated on image classification tasks, particularly class forgetting, highlighting the sensitive nature of image data. The proposed methods effectively preserve model utility while ensuring effective unlearning.

The primary contribution of this thesis is proposing robust and efficient solutions for machine unlearning for scenarios with restricted data access. By leveraging GANs and innovative unlearning processes, MuGAN and zMuGAN offer significant advancements in the field, addressing the urgent need for practical and scalable unlearning techniques.