SPACE: Source-free Proxy Anchor Concept Erasure for MLLMs

TL;DR

As Multimodal Large Language Models (MLLMs) face privacy challenges, existing machine unlearning methods often require access to sensitive data, which is not always available. SPACE, or Source-free Proxy Anchor Concept Erasure, is introduced as a novel framework that enables unlearning without direct access to the original data.

Key Takeaways

• Utilize SPACE for effective unlearning in MLLMs without original data access.
• Implement TPAS to select relevant proxy anchors for concept erasure.
• Apply DCSI to maintain model performance while erasing sensitive information.
• Conduct experiments to validate SPACE's effectiveness across various datasets.