Codes & Data

LLM Unlearning: Privacy-Preserving Unlearning for Trustworthy AI

Project Overview

LLM Unlearning is an open-source project focused on enabling unlearning techniques in Large Language Models (LLMs). With the rise of AI applications and increasing concerns about data privacy, this project introduces exact and approximate unlearning methods designed to make AI models privacy-preserving, trustworthy, and ethical.

Key Features

• Exact and Approximate Unlearning: Methods for forgetting specific data efficiently, while preserving model performance.
• Privacy-Preserving AI: Secure mechanisms to allow models to forget sensitive data and protect user privacy.
• Trustworthy AI: Promotes building ethical models that offer transparency and fairness in data processing.

Project 1: DP2Unlearning

Github Link: DP2Unlearning GitHub

Paper: DP2Unlearning: An Efficient and Guaranteed Unlearning Framework for LLMs
The DP2Unlearning project focuses on advanced techniques for unlearning within LLMs, offering an efficient and guaranteed framework for LLM unlearning. Navigate to the DP2Unlearning project directory to explore and simulate the results. You can also develop and adapt the methods to your own ideas and research needs.

Project 2: UnReL

Github Link: UnReL GitHub

Paper: UnReL: Efficient Exact LLM Unlearning via Targeted Relearning
The UnReL project presents an efficient exact unlearning approach for Large Language Models based on targeted relearning. The project provides the implementation and resources needed to reproduce the proposed method and explore its application to efficient and exact LLM unlearning. Navigate to the UnReL project directory to explore and reproduce the results, or adapt the implementation for your own research.