Relationship-Aware Safety Unlearning for Multimodal LLMs
This addresses safety failures in generative multimodal models for AI deployment, though it appears incremental as it builds on existing unlearning and concept-erasure approaches.
The paper tackles the problem of multimodal LLMs generating unsafe content when benign concepts are combined in specific relationships, proposing a relationship-aware safety unlearning framework that suppresses unsafe object-relation-object tuples while preserving benign uses, achieving improved safety metrics with minimal performance degradation.
Generative multimodal models can exhibit safety failures that are inherently relational: two benign concepts can become unsafe when linked by a specific action or relation (e.g., child-drinking-wine). Existing unlearning and concept-erasure approaches often target isolated concepts or image-text pairs, which can cause collateral damage to benign uses of the same objects and relations. We propose relationship-aware safety unlearning: a framework that explicitly represents unsafe object-relation-object (O-R-O) tuples and applies targeted parameter-efficient edits (LoRA) to suppress unsafe tuples while preserving object marginals and safe neighboring relations. We include CLIP-based experiments and robustness evaluation under paraphrase, contextual, and out-of-distribution image attacks.