RAZOR: Ratio-Aware Layer Editing for Targeted Unlearning in Vision Transformers and Diffusion Models

Researchers introduce RAZOR, a new framework for efficiently removing sensitive information from AI models like CLIP and Stable Diffusion without requiring full retraining. The method selectively edits specific layers and attention heads in transformer models to achieve targeted 'unlearning' while preserving overall performance.

• →RAZOR enables efficient removal of undesirable content from vision transformers and diffusion models without full retraining.
• →The framework identifies critical layers and attention heads that contribute most to target data forgetting while preserving useful knowledge.
• →Testing on CLIP, Stable Diffusion, and vision-language models shows accurate forgetting across identity, style, and object erasure tasks.
• →The method operates significantly faster than conventional unlearning techniques while maintaining model performance.
• →RAZOR offers a practical solution for model safety and compliance in transformer-based vision applications.