Syed Tazeem

Machine Unlearning using Data Obfuscation

This capstone project explores the concept of Machine Unlearning, a critical emerging area in AI focused on enabling models to forget specific data samples without retraining from scratch. We propose a novel approach using data obfuscation techniques to support unlearning across various deep learning architectures, including Simple MLPs, Convolutional Neural Networks (CNNs), pre-trained CNNs, and Transformer-based models.

We evaluate our methods on the MNIST and CIFAR-10 datasets to demonstrate efficacy across both grayscale and colored image classification tasks. The project includes performance comparisons of model accuracy, retention degradation, and forgetting success post-obfuscation. Our findings indicate that data obfuscation can be an effective and scalable strategy for enforcing unlearning while preserving overall model utility.