Machine Unlearning and Privacy in AI Models

The field of machine learning is moving towards developing more robust and privacy-preserving models. Recent research has focused on machine unlearning, which enables models to forget specific data or concepts without requiring full retraining. This area has seen significant advancements, with techniques such as adaptive-lambda subtracted importance sampled scores and teleportation-based defenses being proposed to improve the efficiency and effectiveness of machine unlearning. Noteworthy papers in this area include Adapter Shield, which presents a unified framework with built-in authentication for preventing unauthorized zero-shot image-to-image generation, and Towards Irreversible Machine Unlearning for Diffusion Models, which proposes a novel method for making diffusion models forget specific training data. Additionally, research on continual unlearning and low-resource language adaptation has shown promising results, with methods such as Distill, Forget, Repeat and Source-Shielded Updates demonstrating the ability to remove unwanted concepts and adapt models to new languages while preserving performance.

Sources

Adapter Shield: A Unified Framework with Built-in Authentication for Preventing Unauthorized Zero-Shot Image-to-Image Generation

Teleportation-Based Defenses for Privacy in Approximate Machine Unlearning

Adaptive-lambda Subtracted Importance Sampled Scores in Machine Unlearning for DDPMs and VAEs

Distill, Forget, Repeat: A Framework for Continual Unlearning in Text-to-Image Diffusion Models

Grokked Models are Better Unlearners

Towards Irreversible Machine Unlearning for Diffusion Models

Adapting Large Language Models to Low-Resource Tibetan: A Two-Stage Continual and Supervised Fine-Tuning Study

RapidUn: Influence-Driven Parameter Reweighting for Efficient Large Language Model Unlearning

Mitigating Catastrophic Forgetting in Target Language Adaptation of LLMs via Source-Shielded Updates

Built with on top of