CS772 Project: Zero-Shot Machine Unlearning

Overview

This repository contains the work done as part of the CS772 course project during the 6th semester.

Project Team

• Krish @flintmarko
• Siddharth @sid-kal
• Shubham @SP1029
• Labajyoti @labajyoti21
• Ashutosh @ashutoshkr4458

Approach

The project builds upon the work presented in the seed paper "Zero-shot Machine Unlearning" by V.S. Chundawat et al.

1. Error Minimization-Maximization Noise

2. Gated Knowledge Transfer

Repository Structure

• deep-inversion.ipynb: Generating images of MNIST numbers from the trained model using Deep Inversion technique.
• entropy.ipynb: Implementing the entropy criteria for filtering generated data points.
• no-attn-student.ipynb: Removing Attention from student loss.
• deep-retrain.ipynb: Retraining a new model using the images generated by Deep Inversion.

Code tweaked from

• https://github.com/ayushkumartarun/zero-shot-unlearning
• https://github.com/NVlabs/DeepInversion/tree/master