MMUnlearner introduces a pioneering framework for selective knowledge removal in multimodal large language models (MLLMs) while preserving essential capabilities.

• Enables targeted erasure of visual patterns associated with specific entities
• Preserves corresponding textual knowledge already encoded in the model
• Maintains model performance on unrelated tasks and concepts
• Offers a practical solution for privacy and security concerns in multimodal AI

Why It Matters: As MLLMs become increasingly integrated into business applications, the ability to selectively remove sensitive visual information without retraining addresses critical security requirements and regulatory compliance issues.

MMUnlearner: Reformulating Multimodal Machine Unlearning in the Era of Multimodal Large Language Models