Relationship-Aware Safety Unlearning for Multimodal LLMs

Researchers propose a new framework for improving safety in multimodal AI models by targeting unsafe relationships between objects rather than removing entire concepts. The approach uses parameter-efficient edits to suppress dangerous combinations while preserving benign uses of the same objects and relations.

• →Current AI safety approaches often cause collateral damage by removing entire concepts rather than specific unsafe relationships.
• →The new framework targets object-relation-object (O-R-O) tuples to identify and suppress unsafe combinations like 'child-drinking-wine'.
• →Parameter-efficient LoRA edits are used to modify model behavior without broad destructive changes.
• →The approach preserves benign uses of objects while eliminating harmful relationship combinations.
• →Testing includes robustness evaluation against paraphrase, contextual, and out-of-distribution attacks.